Key Technologies of Intelligentized Welding Manufacturing: Welding Arc Acoustic Sensing and Monitoring Technology 9811520011, 9789811520013

Citation preview

Na Lv · Shanben Chen

Key Technologies of Intelligentized Welding Manufacturing Welding Arc Acoustic Sensing and Monitoring Technology

Key Technologies of Intelligentized Welding Manufacturing

Na Lv Shanben Chen •

Key Technologies of Intelligentized Welding Manufacturing Welding Arc Acoustic Sensing and Monitoring Technology

123

Na Lv School of Electronic Information and Electrical Engineering Shanghai Jiao Tong University Shanghai, China

Shanben Chen School of Materials Science and Engineering Shanghai Jiao Tong University Shanghai, China

ISBN 978-981-15-2001-3 ISBN 978-981-15-2002-0 https://doi.org/10.1007/978-981-15-2002-0

(eBook)

© Springer Nature Singapore Pte Ltd. 2020 This work is subject to copyright. All rights are reserved by the Publisher, whether the whole or part of the material is concerned, specifically the rights of translation, reprinting, reuse of illustrations, recitation, broadcasting, reproduction on microfilms or in any other physical way, and transmission or information storage and retrieval, electronic adaptation, computer software, or by similar or dissimilar methodology now known or hereafter developed. The use of general descriptive names, registered names, trademarks, service marks, etc. in this publication does not imply, even in the absence of a specific statement, that such names are exempt from the relevant protective laws and regulations and therefore free for general use. The publisher, the authors and the editors are safe to assume that the advice and information in this book are believed to be true and accurate at the date of publication. Neither the publisher nor the authors or the editors give a warranty, expressed or implied, with respect to the material contained herein or for any errors or omissions that may have been made. The publisher remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. This Springer imprint is published by the registered company Springer Nature Singapore Pte Ltd. The registered company address is: 152 Beach Road, #21-01/04 Gateway East, Singapore 189721, Singapore

Acknowledgements

This work is supported by the National Natural Science Foundation of China under the grant nos. 51975367, 51075268, Startup Fund for Youngman Research at SJTU (SFYR at SJTU) no. 18X100040049 and Shanghai Sciences and Technology Committee under grant no. 09JC1407100, People’s Republic of China. The author wishes to acknowledge the relative study works finished by Dr. Fang Gu, Dr. Liu Lijun, Dr. Fan Ding, Dr. Ma Yuezhou, Dr. Wang Jifeng, Dr. Chen Bo, Dr. Chen Huabin, Dr. Xu Yanling, Dr. Zhang Zhifen, Dr. Chen Chao and Mr. Zhao Liangqiang, Miss. Zhang Huanhuan, Miss. Yo Bo and so on.

v

Contents

. . . . . . .

1 1 2 3 4 4 5

....... ....... .......

9 9 9

.......

12

.......

14

1

Multi-source Information of Arc Welding Dynamic Process . 1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1.2 Acoustic Signal of Welding Dynamic Process . . . . . . . . . 1.3 Classification of Welding Acoustic Signal . . . . . . . . . . . . 1.4 Feature and Application Value of Arc Sound Signal . . . . . 1.5 Influencing Factor of Arc Sound Signal . . . . . . . . . . . . . . 1.6 Microphone Array Measurement of Arc Sound . . . . . . . . .

2

Acoustic Mechanism and Arc Sound Source Modeling for GTAW Welding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1 Basic Arc Sound Source Model . . . . . . . . . . . . . . . . . . 2.1.1 Weld Pool Oscillation Model of Arc Sound . . . 2.1.2 Analysis and Validation of Sound Source for GTAW . . . . . . . . . . . . . . . . . . . . . . . . . . . 2.1.3 Influence of Welding Parameters on Arc Sound Signal During Dynamic Welding . . . . . . . . . . .

3

. . . . . . .

. . . . . . .

Feature Extraction and Analysis of Arc Sound Signal with Dynamic Welding Process . . . . . . . . . . . . . . . . . . . . . . . . . 3.1 Welding Operation System Setup . . . . . . . . . . . . . . . . . . . . 3.2 Arc Sound Signal Preprocessing Method . . . . . . . . . . . . . . . 3.2.1 Removal of DC Component . . . . . . . . . . . . . . . . . . . 3.2.2 Noise-Removal Process . . . . . . . . . . . . . . . . . . . . . . 3.2.3 Region of Interest (ROI) Selected . . . . . . . . . . . . . . . 3.2.4 Window Function Addition . . . . . . . . . . . . . . . . . . . 3.2.5 Analysis of Arc Sound Signal in Short Time Domain . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3.2.6 Frequency-Domain Analysis of Arc Sound Signal . . . 3.2.7 Time–Frequency Analysis of Arc Sound Signal . . . . . 3.2.8 Relationship Between Sound Feature and Penetration of Weld Pool . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

. . . . . . .

. . . . . . .

. . . . . . .

17 17 21 22 22 24 24

... ... ...

25 26 29

...

31

. . . . . . .

. . . . . . .

vii

viii

4

5

6

7

Contents

Channel Generation Mechanism and Modeling for Arc Sound Signal During GTAW . . . . . . . . . . . . . . . . . . . . 4.1 Generation Mechanism of Arc Sound Channel . . . . . . . . . . . 4.2 Arc Sound Channel Modeling Based on Cepstrum Coefficient . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.2.1 A Mathematical Model Based on Cepstrum Analysis . 4.2.2 Cepstrum Coefficients Model of the Arc Sound Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4.3 Arc Sound Channel Modeling Based on Linear Predictive Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .

... ...

37 37

... ...

39 40

...

42

...

43

....

51

....

53

. . . . .

. . . . .

56 57 58 60 64

......

67

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

67 68 68 70

. . . .

. . . .

. . . .

. . . .

. . . .

. . . .

71 71 72 74

......

81

......

81

......

83

......

86

Prediction Model Establishment Based on Arc Sound Feature Array . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.1 Welding Penetration Recognition Analysis Using Arc Sound Signal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.2 Prediction Model Using BP_AdaBoost Neural Network for GATW Welding . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3 Classification Model Using Hidden Markov Model . . . . . . . 5.3.1 Wavelet Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . 5.3.2 Modeling Based on Hidden Markov Model . . . . . . . 5.3.3 Comparing with Traditional Model . . . . . . . . . . . . . Relationship Modeling Between Weld Pool Collapse and Welding Penetration Based on Analyzing Arc Sound Features . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.1 Welding Experimental Design and Arc Sound Signal Collection . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.2 Processing of Arc Sound Signal for Weld Pool Collapse . 6.2.1 De-DC Component Processing . . . . . . . . . . . . . . 6.2.2 Denoising . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3 Accuracy Verification Test for Prediction Model Based on Arc Sound Feature . . . . . . . . . . . . . . . . . . . . . . . . . . 6.3.1 The Piecewise Linear Fitting for Arc Length . . . . 6.3.2 Validation of Arc Length Prediction Model . . . . . 6.4 Prediction Experiment of Welding Pool Collapse . . . . . . Real-Time Control of Welding Penetration via Arc Sound Signal for GTAW Welding . . . . . . . . . . . . . . . . . . . . . . . . . 7.1 Design of Real-Time Processing Software for Arc Sound Signal During GTAW Welding . . . . . . . . . . . . . . . . . . . 7.2 Arc Height Tracking Control Experiment via Arc Sound Signal of GTAW Welding . . . . . . . . . . . . . . . . . . . . . . . 7.3 Welding Penetration Control Experiment via Arc Sound Signal of GTAW Welding . . . . . . . . . . . . . . . . . . . . . . .

. . . . .

. . . . .

Contents

8

9

Microphone Array Technology in Welding Dynamic Process Monitoring . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.1 Establishment of Microphone Array Acquisition System . . . . 8.2 Research of Blind Signal Separation in Welding Dynamic Process . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.2.1 FastICA Blind Signal Separation Algorithm . . . . . . . 8.2.2 FastICA Blind Signal Separation Results . . . . . . . . . . 8.3 Arc Sound Signal Analyzed by Dynamic Welding Process . . 8.3.1 Feature Extraction of Dual-Microphone . . . . . . . . . . . 8.3.2 Establish Linear Fitting Model of Dual-Microphone . . 8.3.3 Establish Linear Fitting Model of Dual-Microphone . . 8.3.4 Improvement of Prediction Model for Dual-Microphone . . . . . . . . . . . . . . . . . . . . . . . . . . . 8.4 Welding Dynamic Process Monitoring via Microphone Array 8.4.1 Time Delay Estimation Theory . . . . . . . . . . . . . . . . . 8.4.2 Inspection and Localization of Welding Defects . . . . 8.4.3 Welding Defects Localization and Recognition . . . . .

ix

... ...

97 97

. . . . . . .

. . . . . . .

. . . . . . .

99 100 101 104 104 106 108

. . . . .

. . . . .

. . . . .

111 114 114 115 120

Multi-source Information Fusion Between Welding Arc Sound and Other Welding Dynamic Processes . . . . . . . . . . . . . . . . . . . . . . 123

10 Summary and Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 125 Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127

Chapter 1

Multi-source Information of Arc Welding Dynamic Process

1.1 Introduction Nowadays, the research on intelligent and automatic welding process information monitoring has been getting more concern, within the rise of intelligent manufacturing industry [1–4]. However, when automatic and intelligent technology is applied in welding industrial production process, the open-loop control method mainly adopts input parameters to control the output characteristics. The output characteristics cannot be fed back to the input parameters. This leads to the application of the automatic welding technology scope only confined to a single-structure, shape-fixed welding. For those welding specimens with complex shape and unpredictable environment, experienced welding workers still need to adapt to the changing environment independently and obtain high welding quality. As the material connection technology has been extended to aerospace, nuclear power equipment, underwater equipment and other areas, these special environments require higher welding quality and precision of equipment. Therefore, more researchers pay attention to the study and application of closed-loop control of real-time welding quality. Meanwhile, the development of related hardware and software technologies of high-speed sensing technology, data acquisition technology and signal processing technology, as well as the extensive use of intelligent control algorithm, have laid a theoretical foundation for the research of arc welding process information extraction and quality control. The most important step for the realization of intelligent welding technology is to obtain and analyze the dynamic information of the welding process, and to correspond, quantify and model the characteristics of all kinds of information with the weld pool state. The real-time online closed-loop control of welding quality could be achieved through the analysis of arc information [5, 6]. The welding process itself is a transition process of solid–liquid–solid phase caused by arc heat source, which contains the sound, light, electricity and other information. The information is corresponding to the dynamic change process of welding pool. Therefore, many studies have been done in the field of dynamic information acquisition and feature extraction

© Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_1

1

2

1 Multi-source Information of Arc Welding Dynamic Process

algorithm for many years, including the visual image information, arc information, arc sound information, spectral information, pool oscillation information and the temperature information of weld pool. The researches of various sensors have characterized the statement of dynamic welding process from different perspectives. Each of them has its advantages and limitations. In all, the welding quality control based on multi-information can be considered as an imitation of human behavior process. It mainly refers to the operation process of the professional welder, including the acquisition of welding process information, such as the shape of weld pool and the arc sound signal. The information is processed and analyzed by the human brain. The welding condition is judged to be stable or unstable. In order to achieve automatic welding quality control, it is crucial to obtain the multi-information with high precision and high frequency.

1.2 Acoustic Signal of Welding Dynamic Process Professional welders perceive and control the dynamic formation process of weld joint by observing the shape and size of the pool. As the literature analysis of visual sensing [7–9] suggests, it is feasible to simulate the welder’s eyes to obtain the relevant information. However, limited by current size of visual sensor and the adaptability of image processing algorithm, various factors will affect the image collection in practical applications, such as smoke, metal evaporation and splashing. From the perspective of bionics, the welder could operate the welding dynamic process only using the eyes and ears. Considering this assumption, many researchers started to pay attention to the acoustic signal generated during welding process. As the best assistant of visual information, acoustic signal has been considered as valuable information because of its abundant information, good instantaneity and non-contact. In order to find out whether acoustic signal plays a role in welding monitoring, some scientists in Canada designed special experiments to prove the decisive significance of sound signal in determining the quality of welding process [10]. The experiment is conducted by putting on the headset for the welder, playing music to him or delaying the welding acoustic into the welder’s ears. The experiment system is shown in Fig. 1.1. The welding quality fluctuates significant while the acoustic signal is delayed. The results showed that the acoustic signal played an important role in the quality discrimination of welders. It also illustrated that visual and auditory signals had some complementary functions in judging the welding process. Combining these two sensors will pave the way for the comprehensive realization of the intelligent automatic welding process. The sound itself is a pressure wave produced when an object vibrates. It alternately compresses and pushes molecules of the surrounding propagation medium. In the arc welding process, there are three sound sources of acoustic signal: welding arc, welding shielding gas and welding equipments (such as welding machine, wire feeder and other peripheral equipment). In literature [11, 12], the scientists proved that the acoustic signal generated by shielding gas and welding equipment is very

1.2 Acoustic Signal of Welding Dynamic Process

3

Fig. 1.1 Welder psycho-acoustic experiment

weak compared with the one generated from the arc itself through welding process. Therefore, this illustrated that the noise signal would not affect the analysis of acoustic signal. That is why the researchers would like to name the acoustic signal of welding with “the arc sound signal”. In all, the research concerning sound source of welding has grown greatly alongside with tremendous progress, however, the analysis was narrowed to rational analysis with no formula validation. Further studies about sound source of welding arc can pay more attention to the mathematical model and physical model establishment.

1.3 Classification of Welding Acoustic Signal Welding arc sound is a non-stationary random signal generated by air shock in the welding process. It results from the energy change of welding arc. It has an accurate and sensitive corresponding relationship to the change of welding quality and the dynamic state of weld pool. According to different welding method, the generation mechanisms of arc acoustic signal are different due to the different welding heat sources. The analysis method of arc acoustic signal should be imparity from each other. At present, welding heat sources can be divided into several categories, including heat and pressure, and so on. Among them, the arc energy is the main source for arc welding method. For different welding methods, it is difficult to directly use the same signal processing method to evaluate the welding quality online. For the arc welding, there are two types of welding technology. One is called penetrating welding method, including laser welding and plasma arc welding. Owing to the high heat concentration in the welding process, a small hole will form during the welding process, and the arc sound signal will have a significant change during the penetration. It makes it easier to recognize the acoustic signal, so the research

4

1 Multi-source Information of Arc Welding Dynamic Process

on real-time monitoring of welding penetration status based on sound signal has been accomplished well. The second one is called arc welding, such as argon arc welding and gas metal arc welding. Because the energy is not as concentrated as laser welding, there is no obvious characteristic change of arc acoustic signal. The acoustic signal is generally produced by welding arc, so it is called arc sound signal. Most of the researches focused on the feature extraction and modeling, which were used to predict and identify the dynamic changes of welding process [13, 14]. There is no evidence that real-time welding quality monitoring based on arc sound signal has been realized. That is the problem which has been tried to be solved in this monograph.

1.4 Feature and Application Value of Arc Sound Signal The welding process itself is a heating process, which contains the vibrating plasma atmosphere and the vibration inside the liquid pool. As the rapidly changing signals generally have a lot of inertia, it is necessary to analyze the signal statistics to get its signal distribution. It can be seen from previous studies that arc signal is a nonlinear time-varying vibration signal, and its statistical analysis pays more attention to its physical significance. The arc sound signal is closely related to the stability of welding arc. The factors affecting the arc stability include the changes of welding parameters and the type of weld droplet transition, which can be reflected by the arc sound signal. The application research of arc sound signal has been accomplished for many years, involving the analysis of sound source, the analysis of the relationship between arc sound signal and welding spatter and welding parameters. Most of the studies divided the sound signal into pulse signal and disturbance signal in the research of establishing the recognition model for welding quality based on arc sound signal, such as Refs. [15, 16]. The transfer function of peak current and arc sound signal was established. It is found that the pulse signal comes from the short circuit between the electrode and the parent material, and some splashes impact the molten pool. The molten metal will detonate from the electrode. The sudden ignition of the arc causes a sharp rise in the temperature and expansion of the surrounding protective gas, which causes the vibration of the surrounding gas. The disturbance signal mainly comes from the arc explosion and molten pool variation in physical and geometric dimensions. In a word, the application research of arc sound signal is to realize the real-time monitoring for welding process quality.

1.5 Influencing Factor of Arc Sound Signal The sound itself is a kind of mechanical wave. The arc sound signal during welding process is similar to the oscillating waveform in view of its welding characteristics. During the GTAW welding process, the pulsation of plasma flow between tungsten

1.5 Influencing Factor of Arc Sound Signal

5

electrode and welding workpiece, and the intrinsic vibration frequency of molten pool will be diffused out in the form of sound waves. The plasma flow inside the arc itself is the basic sound source to generate vibration. When the motion of the vibration source causes the particles in the surround arc atmosphere to deviate from their normal position and spread from near to far, the arc sound signal waveform is finally formed. From the perspective of arc energy, arc sound formation is a phenomenon that a part of the mechanical vibration energy generated by sound source is propagated by the elastic medium around it in the way of wave; that is, the process of transforming mechanical energy into acoustic energy. In view of its mechanism, the arc sound signal has a significant correspondence with the changes of various parameters and dynamic characteristics in the welding process. At present, the research scope of arc sound signal on the dynamic information change during welding process mainly includes the influence of welding process parameters, such as welding voltage, arc, welding speed, type and speed of protective gas; and the influence of arc sound signal on welding arc length, welding spatter process change, dynamic deformation and defect generation. In many studies, the relationship between arc sound signal and various parameters have been summarized, such as corresponding relationship with welding voltage, welding current. As a heat source, welding arc contains a large amount of plasma inside, which will generate intense vibration under the action of high temperature to form the arc sound signal. S = K (I

dI dV +V ) dt dt

(1.1)

S represents the sound pressure, V represents the voltage of the arc, the current passing through the arc is labeled as I and K is the constant parameter. Some scientists had studied the arc sound source model of TIG welding, the relationship between arc sound signal and weld parameter, and arc length [17, 18]. The results showed that the arc sound signal could reflect the physical character of welding dynamic process. It is useful in the welding quality monitoring via sensor technology. Among the electrical feature of welding, the power of arc showed the most impact to the arc sound signal and arc energy. It can be concluded from Fig. 1.2 that the arc sound energy could be used for the recognition of welding penetration state.

1.6 Microphone Array Measurement of Arc Sound The earlier researches of welding arc sound were using a single microphone. However, there are a lot of interference source in the actual welding production environment, such as human and machine operation voice, and the background noise. The collected signal is a mixed signal which is difficult to distinguish the effective information from the noise information. Then the scholars came up with the idea of

6

1 Multi-source Information of Arc Welding Dynamic Process

Fig. 1.2 The relationship between arc sound energy and penetration state pressure

introducing the array into the welding process monitoring. Using the excellent positioning ability and independent source, separation ability of the microphone array can make it easier to realize welding dynamic process monitoring. The application of sound source positioning technology of microphone array in industry [19, 20] is mainly reflected in the realization of robot’s positioning and tracking of vibration source, as well as the noise detection of engines in automobiles and motorcycles and large machinery. Among them, the robot sound source positioning and tracking is a very important development field. In the field of welding, microphone array sound source localization technology is still in the exploratory stage, the team of Prof. Luo Zhen from Tianjin University is the pioneer in this field. Dr. Ao sansan used eight microphones to form an array with a structure of 2 × 4 to collect the system, as shown in Fig. 1.3. Through blind source separation algorithm, signal noise reduction and feature extraction were carried out for the laser welding process. According to the special environment of welding process, there are couple of problems in application of microphone array sound source positioning technology in the welding process: (1) the sound signal is relatively rough, while for the small size welding defects, location is a delicate process; (2) the existing sound source positioning algorithm comes from the communication field, and the signal in the communication field is usually a single frequency, while the welding arc sound is a multi-frequency and complex signal; (3) there are a few researches that studied the welding sound positioning based on microphone array, and the positioning results are relatively rough and need to be verified. Microphone array sound source positioning is more common in the field of robot research, and it is relatively rough, and the number of array elements is at least 8, and in some researches being more than 64;

1.6 Microphone Array Measurement of Arc Sound

7

Fig. 1.3 Schematic diagram of microphone array position

(4) the location of defects can be corresponding through the time axis according to the position of signal mutation in the straight welding seam; the studied positioning seems like not meaningful; however, it has research value for the complex shape workpieces welding monitoring process.

Chapter 2

Acoustic Mechanism and Arc Sound Source Modeling for GTAW Welding

2.1 Basic Arc Sound Source Model In order to achieve the detection of welding defects, it is necessary to analyze the mechanism of sound source generation [21]. The welding arc sound generation can be divided into two parts: one is a semi-transparent near-conical cavity enclosed by the protective atmosphere between tungsten electrode and welding test plate, which is excited by the sound source generated by the change in arc energy, and then sound source generation formed after the modulation of sound channel. The second is produced by resonance of liquid metal under the action of various arc forces, surface tension Fe , supporting force Fac , N, gravity Fa and shear force σ1 . The mechanism diagram of sound source generation is shown in Fig. 2.1. There are three sources of sound in gas flow: monopole sound source, dipole sound source and quadrupole sound source. The welding arc sound source has been proved to be quadrupole sound source because of the arc forces that lead to the turbulent flow of air. The sound pressure of quadrupole sound source is shown in Eq. 2.1. The arc sound source model can be divided into two parts [22]. The first type model is generated by arc power, as in Eq. 2.2 (Fig. 2.2). Fac = pd =

cfQb2 cos θ cos(θ + 90◦ )e−jkr 2r 3 Fac = UI

(2.1) (2.2)

2.1.1 Weld Pool Oscillation Model of Arc Sound The second type of sound source model is generated from oscillation model. The model between intrinsic frequency of the molten pool resonance and the size, shape © Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_2

9

10

2 Acoustic Mechanism and Arc Sound Source Modeling for GTAW Welding

Fig. 2.1 The sound source model mechanism

Fig. 2.2 Quadrupole sound source model of welding arc sound

of the molten pool is derived by using the two-dimensional film wave model. Each point on the membrane vibrates up and down in a direction whose displacement is a function of time and position. Let the surface density of film be σ and tension is f , then the force in z direction will be F Z (x, y) = f (

∂ 2u ∂ 2u + 2) ∂x2 ∂y

(2.3)

2.1 Basic Arc Sound Source Model 1 ∂ 2 u(x,y,t) c2 ∂t 2

11

=

∂ 2 u(x,y,t) ∂x2

+

∂ 2 u(x,y,t) ∂y2

(2.4)

Variables separation u(r, θ, t) = z(r, θ )e−iωt , 1 ∂z ∂ 2z 1 ∂ 2z + + 2 2 + k 2z = 0 2 ∂r r ∂r r ∂θ

(k = ω/c)

(2.5)

In the penetration state, the bottom of the molten pool is no longer supported by the workpiece metal, and the entire molten pool presents obvious vibration. The internal motion of the molten pool is more complex than that of the non-molten pool. After full penetration of the workpiece, the aggregate shape of the upper surface of the pool does not change much, so the arc force changes very less with time. The gravity of molten pool increases steadily, which is consistent with the continuous and stable heat source applied to the workpiece in the welding process. The supporting force of the solid curved surface wall on the molten pool is opposite to that of the other four forces. As the molten pool grows, the supporting force of the solid curved surface wall on the molten pool decreases. In the approximate derivation, the entire molten pool can be regarded as an equivalent “thin film”, and the intrinsic frequency of the oscillation of molten pool in the molten state can be directly obtained by using the membrane wave model. Similar to the case without penetration, in the state of penetration, there are different modes of molten pool oscillation, mainly as shown in Fig. 2.3, k = 2.405 ∗ 2/Deq = 4.81/Deq . In fact, the width of the molten pool back is smaller than the front, and the 2 = 21 (Dt2 + Db2 ). In addition, the surface tenequivalent diameter Deq is given by Deq  sion f in c = f /σ should be replaced by 2γ , the areal density σ is replaced by hρs , where h is the thickness of the sheet and ρs is the solid metal density, thus realizing the  mass equivalence of “cylinder” element. To sum up, k = 4.81/Deq , c = 2γ /hρs , k = ω/c, f = ω/2π Fig. 2.3 Pool oscillation mode

12

2 Acoustic Mechanism and Arc Sound Source Modeling for GTAW Welding

 γ −1 f = 1.083 D hρs eq

(2.6)

Thus, the arc sound signal is influenced by both gas oscillation effect and molten pool oscillation effect. In order to inspect the welding defect, it is essential to extract the effective information [21].

2.1.2 Analysis and Validation of Sound Source for GTAW The arc acoustic signal of pulsed argon tungsten arc welding is closely related to welding voltage and welding current. Further interception of signals around 0.1 s and observation of their correlation shows that arc acoustic signals are composed of periodic and two ringing transient signals, and the occurrence position of arc acoustic signals is exactly corresponding to the rising and falling edges of welding voltage signals. The results show that the acoustic source excitation of arc acoustic signal mainly comes from the change of arc itself. When the peak current and the base current alternate, the change of arc causes the change of arc acoustic. At the same time, the welding power P can be calculated according to the welding voltage and welding current, and the result is shown in Fig. 2.4. It is observed and more obvious in the comparison between arc sound and power that the acoustic source excitation of arc sound comes from the change of arc energy. Figure 2.5 shows arc acoustic waveform in different short time areas. It can be seen that in 4 s the arc acoustic signal alternates between pulse peak value and pulse base value for a period of 500 ms. For arc acoustic signals within 400 ms, the duty cycle of peak signal and base signal is 1:1. Moreover, within 30 ms the arc sound wave is composed of peak and base period which shaped like a ring. The distance between the two peak signals period of 4 ms, and a large peak signal and another set of smaller peak signal for a

(b)

0 0

Power P/(w)

0.5

500

1000

1500

2000

2500

0.5

0 0

Sound s/(v)

Sound s/(v)

Power P/(w)

(a)

5

0.2 0 40

x 10

4

60

80

100

120

0.4 0.2 0

10

Sampling number

0.4

4500 5000 5500 6000 6500 7000 7500 Sampling number

Fig. 2.4 The comparison of weld power and arc sound in short time period

2.1 Basic Arc Sound Source Model

Sound s/(v) Sound s/(v) Sound s/(v)

Fig. 2.5 The arc sound waveform figure in different short time period

13

0.5 0 -0.5

4

1

2

3

0.1

0.2

0.3

0

0.5 0 -0.5

0

0.4

0.5 0 -0.5

9.625ms

4ms 0.005 0

0.015

0.01

0.02

0.025

0.03

Weld time t/(s)

period of 4 ms, is spaced 9.625 ms. These peak signals occur at the same periodic position, and each small period contains a small peak signal and a large peak signal, which is not obvious in the base peak signal. A section of arc acoustic signal sequence was extracted for cepstrum analysis, and the results of cepstrum analysis of arc acoustic signal were shown in Fig. 2.6. As can be seen from the results, after cepstrum analysis, obvious periodic characteristics could be seen, with peak values of 14, 29, 44 and 59 ms. A series of peaks occurred nearby, and the time interval between the two peaks was 15 ms. Since the cycle was 15 ms, the frequency was 1/cycle = 66.67, approximately 70 Hz, indicating that these small excitation signals were consistent with the welding frequency. In view of the pitch period estimation results, this series of peaks is the pulsed current excited by the sound. When amplifying cepstrum signal, it is found that except the peak value represented by the pulse current, there is no periodic repetition, which indicates that there are no other obvious harmonics in the sound signal except the

0.5 0 -0.5

Amplitude f/(v)

0

0.5

1 1.5 Sampling number

0.1 14ms

29ms

0.02

0.04

0.05 0 0

2 x 10

59ms

Time t/(s)

0.08

2 1 0 0

1

6

44ms

0.06

Amplitude f/(v)

(b)

0.1

Amplitude f/(v)

Sound s/(v)

(a)

2 Frequency Fs/(Hz)

4

3 x 10

4

0.2 0.1 0 0

14ms 29ms

0.02

0.04

44ms

59ms

0.06

0.08

0.1

Time t/(s)

Fig. 2.6 The comparison of the cepstral analysis with sound pressure and DCT frequency feature. a the cepstral feature and sound pressure, b the cepstral and DCT feature

14

2 Acoustic Mechanism and Arc Sound Source Modeling for GTAW Welding

harmonic of the current sound, indicating that the excitation mechanism of the sound signal by the molten pool oscillation is different from the current. This means that the sound produced by the pool oscillation may not be in the form of a fundamental frequency multiple of its intrinsic frequency, but in some single frequency [23].

2.1.3 Influence of Welding Parameters on Arc Sound Signal During Dynamic Welding Sound itself is a kind of mechanical wave, and for the sound signal in the welding process due to its welding characteristics, there is also a wave of oscillation. For the GTAW welding process, the plasma pulsation between the tungsten electrode and the welding workpiece and the intrinsic vibration frequency of the weld pool will be diffused in the form of sound waves. The plasma flow in the arc itself is the basic sound source that generates vibration. When the motion of the vibration source causes the particles in the surrounding arc atmosphere to deviate from the normal position and spread from near to far, the arc acoustic signal waveform is eventually formed. From the perspective of arc energy, arc acoustic formation is a phenomenon that part of the mechanical vibration generated by the sound source is propagated out by the elastic medium around it by means of waves, that is, the mechanical energy is transformed into acoustic energy. In view of its mechanism, arc acoustic signal has a significant relationship with the variation of various parameters and dynamic characteristics during welding process. The research on the influence of arc acoustic signal on the dynamic information of welding process mainly includes the influence on various welding process parameters, such as welding voltage, arc, welding speed, protection gas type and flow rate, and the influence of arc acoustic signal on welding arc length, welding spatter process, dynamic deformation and defects. Many studies summarize that the arc sound signals with various process parameters, such as the corresponding relationship between welding voltage and welding current, for different welding methods are based on arc as heat source. The heat source produces intense vibration and the sound generates from the plasma arc change. The arc sound signal has close relationship with the arc characteristic [23, 24]. The study shows that arc sound pressure can be expressed as follows: S = K(I

dI dV +V ) dt dt

(2.7)

2.1 Basic Arc Sound Source Model

15

Fig. 2.7 The relationship between weld speed, arc length and arc sound pressure

In the formula, S is sound pressure, V represents the voltage of arc, I represents the current passing through arc, and K is a constant (Fig. 2.7).

Chapter 3

Feature Extraction and Analysis of Arc Sound Signal with Dynamic Welding Process

The welding arc acoustic signal is similar to the vibration signal. Different types of feature extraction and analysis can be carried out according to the collection method. As the audible sound was chosen for this study, the collected system and feature extraction were done according to this principle.

3.1 Welding Operation System Setup An automatic control system is designed for the welding experiments based on arc acoustic signal fusion control of pulse GTAW welding process. The system is shown in Fig. 3.1, which is composed of four parts: welding system; motion control system; visual system and acoustic system [23]. Welding system: According to the characteristics of research object and relevant requirement of experiment, the system mainly includes AC/DC welding power supply, wire feeding machine, control box, water and gas protection device. The system could achieve the welding arc on and arc off through the control circuit and industrial personal computer. Considering the further study of welding quality control, several parameters should be adjustable during the welding process, such as peak pulse current Ip , pulse base current Ib , wire feed speed vb and welding speed v. The peak pulse current is the most effective factor for the welding penetration. The analog voltage from 0 to 10 V is the output by the acquisition card. It is amplified by DC amplifier circuit to realize the control of welding current and wire feeding speed, so as to realize the communication between computer and welding machine. It is necessary to calibrate the actual welding parameter with the analog parameters, as shown in Fig. 3.2. The results of calibration are linear characteristic. The linear equation between peak welding current I and setting voltage U 1 is U1 =

I − 24.4 52.1

© Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_3

(3.1) 17

18

3 Feature Extraction and Analysis of Arc Sound Signal … Welding current

Arc sensor A/D

Control circuit

Host Computer H

Conditioner

Welding machine

Wire Feeder

Ignition switch

Voltage

Step motor

D/A

Microphone Visual sensor

Motion mechanism

Welding current (A)

500 400 300 200 100 0

0

2

4 6 Voltage(V)

8

Wire feeding speed(mm/s)

Fig. 3.1 GTAW welding system and sensing system 50 40 30 20 10 0

0

(a) Welding current curve

2

4 6 Voltage(V) (b) Wire feeding speed curve

8

Fig. 3.2 The relation between the analog signal and welding parameters

The linear equation between wire feed speed and setting voltage is U2 =

Vf + 4.4 6.91

(3.2)

Motion control system: The specific motion control system is achieved by stepper motor. The angular displacement is converted into planar displacement through the

3.1 Welding Operation System Setup

19

pulse driving form, and the output current driving mechanism is combined with the driving controller to realize the motion in positive and negative directions [25]. Visual system: In order to study the generation and change of welding penetration state, it is necessary to set up the visual sensing system from different aspects. A three-way vision sensor is designed to detect the image from the weld pool, the arc shape and the penetration state of back side [26, 27] (Fig. 3.3). The hardware equipment includes digital WAT-902H HD industrial camera, M2514-MP telephoto lens, Daheng CG400 image acquisition card and industrial computer. After the calibration of the camera, the image is as shown in Fig. 3.4. Acoustic system: The audio sensing system is responsible for the acquisition of acoustic signal during the welding process. The arc sound signal of welding is collected at 36 kHz sampling rate and 12 bits precision. It is picked by MP201, an omnidirectional capacitance microphone, of which the frequency response range is from 20 to 20 kHz, the sensitivity is 50 mV/Pa and the dynamic range is >146 dB. The microphone is settled at 75° horizontal angle to the workpiece. The arc sound

Fig. 3.3 The visual sensing system: a The photograph of visual sensor, b light path diagram

Fig. 3.4 The image of three-dimensional visual sensor

Front-end weld seam Backside weld seam Weld width

Weld pool

20

3 Feature Extraction and Analysis of Arc Sound Signal …

Table 3.1 Major parameters of audio sensor

Type

MP201

Diaphragm and canning material

Nickel, nickel alloy

Open circuit sensitivity

−26 ± 2 dB (50 mV/Pa)

Frequency response

20 Hz to 20 kHz

Polarization voltage

0V

Dynamic range

>146 dB

Background noise

0

(4.9)

n0.2 V for acoustic pressure. The selected region is shown in Figs. 6.2 and 6.3. 1

Arc sound pressure s(v)

Fig. 6.2 Selected region of arc sound signal

0.5 0 -0.5 -1 1

2

3

4

Sample point

1

Arc sound pressure s(v)

Fig. 6.3 Extracted signal

0.5 0 -0.5 -1 1000

2000

Sample point

5

x 10

3000

5

70

6 Relationship Modeling Between Weld Pool Collapse …

6.2.2 Denoising The original arc sound signal contains many noises, which are environmental noise and disturbance noise. The inevitable environmental noise mainly generates from the electrical equipment like welding equipment and transformer. The pulse interference noise is a kind of random signals produced by arc itself, influenced by welding parameter and the high-frequency signal during ignite arc. The unstable arc is produced by alternating pulses. The pulse interference noise will disturb the actual arc sound signal. Denoising method is aimed at removing the noise and extracting effective information from original signal. The wavelet packet transform is used to extract rich non-stationary information from original signal. The moving average method is used to eliminate the pulse interference from arc sound. A new way of reducing noise of arc sound signal is proposed in this paper. The specific procedure is as follows: 1. x(n) is the original signal, n∈(1,M), three-layer Daubechies wavelet is first used to decompose the arc sound signal. The signal is divided into high-frequency coefficient and low-frequency coefficient. 2. Choose the best threshold to optimize the coefficients both in high frequency and low frequency of wavelet packet. 3. Reconstruct the arc sound signal with optimized coefficient, then get the processed arc sound signal x’(n). 4. Select an appropriate window style and size “N” for x’(n) to smooth the data. To proving the experimentation, N is set to be 100 as window size. The processed signal of arc sound signal S’(n) is replaced by means of 100 neighboring points like the following algorithm: x  (i) =

N 

S  (i)/N

(6.2)

i=1

S  (i) = x  (i) i = 1, 2, 3 . . . n

(6.3)

5. By that analogy, the new processed arc sound signal S’(n) is generated with the same length of original signal but with less noise in it. The results of different denoising algorithms are shown in Fig. 6.4. It concludes from the results that: (a) different denoising method has favorable denoising effect on account of specific noise. The wavelet packet analysis is used to reduce the environmental noise. The moving average denoising method is used to reduce the pulse interference noise. (b) From Fig. 6.4b, c, it can be seen that there is less environmental noise in pulse GTAW process, so the wavelet packet analysis is not obvious than the moving average denoising method in noise reduction. The moving average denoising method could remove more pulse interference noise in arc sound signal. (c) According to special characteristic of arc sound signal, a new way of noise reduction based on wavelet packet and moving average denoising method is

6.2 Processing of Arc Sound Signal for Weld Pool Collapse

(b)

0.5

Arc sound pressure s(v)

Arc sound pressure s(v)

(a)

0.45 0.4 0.35 0.3 0.25 0.2 0

1000

2000

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0

3000

1000

2000

3000

Sample point

Sample point

(c)

(d) 0.5

Arc sound pressure s(v)

Arc sound pressure s(v)

71

0.45 0.4 0.35 0.3 0.25 0.2 0

1000

2000

Sample point

3000

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0

1000

2000

3000

Sample point

Fig. 6.4 The signal processing results of some denoising algorithms: a original signal, b the wavelet packet analysis, c the moving average denoising, d the moving average denoising based on wavelet packet analysis

proposed. The result of new denoising method is shown in Fig. 6.4d. It reduced the environmental noise as well as pulse interference noise.

6.3 Accuracy Verification Test for Prediction Model Based on Arc Sound Feature 6.3.1 The Piecewise Linear Fitting for Arc Length Considering the arc volume variation at different arc length situation, in order to improve the arc length prediction accuracy of the linear model based on arc sound signal, the piecewise linear fitting method is proposed to set different prediction model for different arc length variation region. The arc length is changed between the range from 3 to 6 mm according to specific character of pulse GTAW. It is divided into two linear prediction models of 3–4 mm and 4–5–6 mm arc length. Different arc heights correspond to different linear fitting coefficients. The prediction accuracy has

72

6 Relationship Modeling Between Weld Pool Collapse …

(b) 0.5

Arc sound pressure s(v)

Arc sound pressure s(v)

(a) 0.45 0.4 0.35 0.3 0.25 0.2 0

2000

3000

Arc sound pressure s(v)

0.45 0.4 0.35 0.3 0.25 0.2 1000

2000

3000

Sample point

0.3 0.25 0.2

4000

2000

1000

3000

4000

Sample point

(d)

0.5

0

0.4 0.35

0

4000

Sample point

(c) Arc sound pressure s(v)

1000

0.5 0.45

0.5 0.45 0.4 s=0.0460h+0.1075

0.35 0.3 0.25 0.2 3

3.5

4

Arc length h(mm)

Fig. 6.5 The linear fitting process when the arc height is changed from 3 to 4 mm: a extracted arc sound signal, b the moving average denoising based on wavelet packet analysis, c the linear fitting result, d the final relational model between the arc height and the arc sound pressure

improved. The linear fitting results of 3–4 mm model are shown in Fig. 6.5. After the denoising and linear analyzing, the linear model between arc height and arc sound is expressed as s = 0.046h + 0.1075. In order to make the linear model more adaptable, same procedures are implemented several times under the same condition. The average of each coefficient is set to be the final linear model, like s = 0.04045h + 0.12445. The results are shown in Fig. 6.6. The fitting results of 4–5–6 mm linear model are shown in Figs. 6.7 and 6.8. The final linear model is s = 0.0119h + 0.2173.

6.3.2 Validation of Arc Length Prediction Model In order to verify the piecewise linear fitting effect, the prediction models are used to forecast the arc length via arc sound signal. The proof test is implemented to verify the prediction accuracy of the piecewise linear model, in which the arc length

6.3 Accuracy Verification Test for Prediction Model Based on Arc …

(b) 0.5

Arc sound pressure s(v)

Arc sound pressure s(v)

(a)

73

0.45 0.4 s=0.0460h+0.1075

0.35 0.3 0.25 0.2 3

3.5

4

0.5 0.45 0.4 s=0.0349h+0.1414

0.35 0.3 0.25 0.2 3

3.2

Arc length h(mm)

3.4

3.6

3.8

4

Arc length h(mm)

Arc sound pressure s(v)

(c) 0.5 0.45 0.4 0.35

s=0.04045h+0.12445

0.3 0.25 0.2 3

3.2

3.4

3.6

3.8

4

Arc length h(mm)

Fig. 6.6 The final relational model between the arc height and the arc sound pressure when the arc height is changed from 3 to 4 mm: a arc height from 3 to 4 mm #1, b arc height from 3 to 4 mm #2, c the final result of linear fitting

gradually changes to 6–5–4–3 mm. The original arc sound signal is first processed as per the above procedures. The results are shown in Fig. 6.9b, c. Then it is divided into two parts as shown in Fig. 6.9d, according to the variation arc length like 6–5–4 mm and 4–3 mm. Different linear fitting models are used to predict different range of arc length, using s = 0.04045h + 0.12445 to predict arc length changes from 4 to 3 mm and using s = 0.0119h + 0.2173 to predict arc height changes from 6 to 5 and then to 4 mm. The prediction results and errors of different linear fitting models are shown in Figs. 6.10, 6.11 and 6.12. It can be concluded that the prediction results have more errors at the beginning of the welding process; however, at the end of the welding process, the precision accuracy becomes more precise. By analyzing the welding characteristic of pulse GTAW and generating mechanism of arc sound signal, the larger errors are produced by instability of welding. Because the arc sound signal is generated from air vibration in the sound channel system which is composed of shielding gas and welding plate. At the beginning of welding process, the sound channel system has not formed a stable cavity, so the arc sound collected by microphone is not accurate. As the sound channel is becoming more stable, the prediction accuracy

74

6 Relationship Modeling Between Weld Pool Collapse …

(b) 0.5

Arc sound pressure s(v)

Arc sound pressure s(v)

(a) 0.45 0.4 0.35 0.3 0.25 0.2 0

4000

Sample point

0.5 0.4 0.35 0.3 0.25 0.2 2000

4000

Sample point

0.3 0.25 0.2 0

2000

6000

4000

Sample point

(d)

0.45

0

0.4 0.35

6000

Arc sound pressure s(v)

Arc sound pressure s(v)

(c)

2000

0.5 0.45

6000

0.5 0.45 0.4 0.35

s=0.0138h+0.2062

0.3 0.25 0.2 4

4.5

5

5.5

Arc length h(mm)

6

Fig. 6.7 The linear fitting process when the arc height changes from 4 to 5 and then to 6 mm: a extracted arc sound signal, b the moving average denoising based on wavelet packet analysis, c the linear fitting result, d the final relational model between the arc height and the arc sound pressure

could be improved. At the latter half of welding process, the predicted value is larger than the true value because of the collapsing of weld pool. The specific error value is shown in Table 6.2. The final average error of linear fitting model is 0.580487 mm. It is proved that arc sound signal could achieve the real-time control of arc length changes in pulsed GTAW.

6.4 Prediction Experiment of Welding Pool Collapse After the validation of the arc length prediction model, the piecewise model is used to forecast the surface height as well as the sag depression. Prof. Zhang Yuming had proved that the weld pool depth was related to the penetration status of weld pool. The sag depression could reflect the weld penetration and also the surface forming state. The testing experiment is designed to check the prediction effect of precise linear model on forecasting the sag depression through acoustic signal [33]. The

6.4 Prediction Experiment of Welding Pool Collapse

(b) Arc sound pressure s(v)

Arc sound pressure s(v)

(a)

75

0.5 0.45 0.4 0.35

s=0.0138h+0.2062

0.3 0.25 0.2 4

4.5

5.5

5

0.5 0.45 0.4 s=0.0100h+0.2284

0.35 0.3 0.25 0.2

5

4.5

4

6

5.5

6

Arc length h(mm)

Arc length h(mm)

Arc sound pressure s(v)

(c) 0.5 0.45 0.4 s=0.0119h+0.2173

0.35 0.3 0.25 0.2 4

4.5

5

5.5

Arc length h(mm)

6

Fig. 6.8 The final relational model between the arc height and the arc sound pressure when the arc height changes from 4 to 5 and then to 6 mm: a arc height of 4–5–6 mm #1, b arc height of 4–5–6 mm #2, c the final result of linear fitting

schematic of this experiment is shown in Fig. 6.13. L represents the arc height and r represents the depth of weld penetration. The welding experiment is implemented under the same welding parameters above; however, the gap between the two pieces of weld plate is increased to make more sag depression during the welding process. Also, the arc length is set as 4 mm during the whole welding process. The variations in arc length generate only the sag depression of weld pool. The measuring tool is shown in Fig. 6.13b. The arc sound signal is processed by extraction and denoising method as explained above. The results are shown in Fig. 6.14. It can be concluded that after removing the environmental noise and pulsed interference noise, the acoustic signal has less undesired signal and an uptrend, as shown in Fig. 6.14d. The range of sound pressure is coincident to the linear model of arc length from 4–5–6 mm. The sound pressure of this is implemented according to s = 0.0119h + 0.2173. The predicted result is calculated as follows: Arclength = L + r

(6.4)

76

6 Relationship Modeling Between Weld Pool Collapse …

(b) Arc sound pressure s(v)

Arc sound pressure s(v)

(a) 1 0.5 0 -0.5 -1 2

4

6

8

Sample point

0.5 0 -0.5

0.3 0.25 0.2 2000

4000

6000

6000

8000

Sample point

(d)

0.35

4000

2000

5

0.4

0.15

-1

x 10

Arc sound pressure s(v)

Arc sound pressure s(v)

(c)

1

8000

0.4

Arc height 6-5-4mm

0.35 0.3 0.25

Arc height 4-3mm

0.2 0.15

2000

4000

6000

8000

Sample point

Sample point

Fig. 6.9 The specific scheme of arc sound signal processing under different arc height of 6–5–4– 3 mm: a original signal, b the extracted signal, c the denoising signal, d the final relational model between the arc height and the arc sound pressure 10

Forcast arc length Arc length The error of arc length

The error of arc length h(mm)

8 6 4 2 0 -2 -4 -6

0

100

200

300

400

500

600

700

800

900

Sample point

Fig. 6.10 The error of linear fitting model for changing arc height from 4 to 3 mm

6.4 Prediction Experiment of Welding Pool Collapse

77

10

Forcast arc length Arc length The error of arc length

The error of arc length h(mm)

8 6 4 2 0 -2 -4 -6

0

1000

2000

3000

4000

5000

6000

7000

Sample point

Fig. 6.11 The error of linear fitting model for changing arc height from 6 to 5 and then to 4 mm

The error of arc length h(mm)

10

Forcast arc length Arc length The error of arc length

8 6 4 2 0 -2 -4 -6

0

1000

2000

3000

4000

5000

6000

7000

8000

Sample point

Fig. 6.12 The error of linear fitting model for changing arc height from 6 to 5 and to 4 and then to 3 mm

Table 6.2 The error of prediction arc height for three different arc height changes

Error(mm) arc height

Emax

Emin

Emean

6–5–4

3.66444

−6.006

−0.223254

4–3

1.46689

−0.976108

−0.580487

6–5–4–3

2.66444

−6.006

−0.580487

It means the predicted arc length contains the real arc height L and the sag depression. The results are shown in Fig. 6.15. The real measured arc length is gradually increasing as the penetration state changed from partial penetration to full penetration. It is becoming smoothly when the penetration is stable to full penetration. By

78

6 Relationship Modeling Between Weld Pool Collapse …

(a) The geometric description of weld joint

(b) The measuring tool forweld pool depth

Fig. 6.13 The schematic of experiment for forecasting the weld pool depth

(b)

1 0.8 0.6 0.4 0.2 0

(c)

1

2

3

Sample point

4

5 4 x 10

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0

2

4

Sample point

x 10

4

0.8 0.7 0.6 0.5 0.4 0.3 0.2 0

(d) Arc sound pressure s(v)

Arc sound pressure s(v)

Arc sound pressure s(v)

Arc sound pressure s(v)

(a)

1

2

3

4

Sample point

5 4 x 10

0.5 0.45 0.4 0.35 0.3 0.25 0.2 0

2

4

Sample point

x 10

4

Fig. 6.14 The specific scheme of arc sound signal processing under 4 mm arc height: a the extracted signal, b the denoising signal by removing environmental noise, c the denoising signal by removing the pulsed noise, d the final relational model between the arc height and the arc sound pressure

Real arc length AL(mm)

Prediction arc length h(mm)

6.4 Prediction Experiment of Welding Pool Collapse

79

8 6 4

2mm Error

2 0

0

1

2

4

3

6

5

Sample point

x 10

4

8 6 4 2

Penetration 0

100

50

150

Pulse number

Fig. 6.15 The comparison of prediction arc length and measured arc length

the influence of the accumulation of filler metal, the measured position of arc height is a little higher than the true position, so real measured arc height is a little shorter than the actual arc height of weld pool. The 2 mm gap is generated by the change of penetration status. The measured arc length is below the 4 mm first. Then as the penetration gets more deep, the arc length increases to more than 5 mm. However, for the prediction arc length, because it is calculated through the linear model which is setting up at full penetration state, it is not fitting the prediction model in partial penetration. The prediction arc length is about 6 mm, which is more than the real arc height. For the full penetration state, the prediction effect is more accurate than at the beginning. It can be used to predict the surface height of weld pool.

Chapter 7

Real-Time Control of Welding Penetration via Arc Sound Signal for GTAW Welding

The arc acoustic signal of pulsed argon tungsten arc welding is closely related to the penetration state of welding process, and is also closely related to various welding dynamic changes. The rule can be obtained by feature extraction and classification, and it is better to predict different penetration states. However, there are still many problems from offline analysis to online real-time monitoring. Owing to the complexity of the welding process, the real-time online monitoring of the fusion state in the welding process requires that the model has a good robustness and adaptability.

7.1 Design of Real-Time Processing Software for Arc Sound Signal During GTAW Welding The realization of a real-time signal processing system consists of many related parts and thus constitutes a system. The development of a real-time signal processing system requires the following steps, including process design, hardware design, software design, system integration and testing. The first step is to validate the algorithm, focusing on full validation of all possible data types. From this point of view, this step does not need the speed of the algorithm to be very fast, but needs the algorithm to be able to adapt to the changes of a variety of conditions, where good adaptability is to build a good foundation for the subsequent real-time processing system. Therefore, a software system for arc acoustic signal extraction and penetration status identification of pulsed argon tungsten arc welding (GTAW) is designed, which can be used for algorithm verification of real-time signal processing and monitoring system. In view of the realization of functions of each part and the influence of modular design idea, the system is mainly divided into the following modules: data loading module, data preprocessing module, feature extraction module, image processing module, fusion state classification and recognition module and arc length prediction and control module. The main functions of these modules are:

© Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_7

81

82

7 Real-Time Control of Welding Penetration via Arc Sound …

(1) Data loading module: Load files of welding process information, including arc acoustic signal files and image information files, through the interface window. (2) Data preprocessing module: According to the research in Chap. 3, the collected arc acoustic signal contains a large amount of noise information, which seriously affects the effective features. In order to study the correspondence between arc acoustic signal and weld penetration state, it is necessary to remove all kinds of noise information. Considering the real-time performance of the real-time algorithm, in order not to affect the calculation speed, the envelope of the pretreatment module, the DC component of arc acoustic signal, the arc acoustic denoising processing part and the extraction part of the region of interest were included. (3) Feature extraction module of data: Feature extraction function of arc acoustic signal is realized. It includes all characteristic information mentioned in Chap. 3, including time-domain characteristics of arc acoustic signal, such as mean value, energy, standard deviation, covariance, root mean square, kurtosis factor and skewing factor, and the frequency-domain characteristics, namely the arc sound signal frequency-domain amplitude value of the discrete cosine transform. When setting up an adjustable switch according to the needs of the user we can choose a different area of the arc sound signal in frequency-domain feature extraction. When analyzing the defective welding arc sound signal, we can choose to have apparent change not in frequency-domain analysis for the area in order to make the software has better practicability. The spectrum characteristics of arc acoustic signal are also extracted. The extraction part of arc acoustic channel parameter characteristics, as described in Chap. 4, extracts the cepstrum characteristics of arc acoustic signals. Linear prediction and analysis of LPCC coefficient characteristics and layer spectrum characteristic parameters, and so on, are the parameters of fusion characteristics for subsequent modeling of pattern recognition. (4) Data of the image processing module: In order to further validate based on the control effect of arc sound signal, we need an auxiliary information with more intuitive judgment to identify the effect of arc sound signal characteristics. So we need to use the image visual information for authentication information. This software includes an image processing algorithm based on template matching, which extracts the characteristic value of the back weld width from the collected image information. The back weld width can more directly reflect the weld penetration state, and it can better judge the recognition effect of the model by matching it with the identification result of arc acoustic signal. (5) Penetration status of classification module: Extraction of arc sound signal based on the above multi-dimensional feature vector group, fusion state recognition of recognition and classification prediction: in the fifth chapter research content, the article mainly studied two kinds of model algorithm, including the BP neural network regression forecasting model recognition part classification and regression BP_AdaBoost neural network prediction model. The two kinds of recognition tool can implement different parameters under the condition of the modeling process, realize the recognition of the characteristics of different dimension penetration state and calculate the identification error of the results. Considering the recognition efficiency, the neural network model and SVM model have been established based on

7.1 Design of Real-Time Processing Software for Arc Sound …

83

Fig. 7.1 The feature extraction result of pulse GTAW arc sound signal

multi-information features. It finally achieved the fusion state recognition during the automatic welding process. (6) Arc length prediction control module: According to the analysis in Chap. 6, it can be seen that arc acoustic signal has a linear relationship with arc length and collapse under welding seam. Therefore, a predictive arc length control module is designed in this software to realize arc acoustic signal preprocessing, denoising and linear fitting. It can predict arc length by linear fitting formula and calculate its corresponding prediction error. It also includes piecewise linear fitting of arc acoustic signal and error analysis of the fitting results. The arc length prediction part based on arc acoustic signal can predict and analyze arc acoustic characteristic signals according to the requirements of users and the different types of data, respectively, for the change in the first-order and second-order arc lengths (Fig. 7.1).

7.2 Arc Height Tracking Control Experiment via Arc Sound Signal of GTAW Welding Based on the above research, the control system of experiment is implemented by Visual C++ 6.0. Figure 7.2 shows the specific scheme of experimental verification. The monitoring of arc sound is implemented at pulse peak period. While detecting the rising edge of pulse point A, the arc sound signal is collected after 60 ms delay. Every 5000 points x(i), i = 1, 2, 3 . . . n are collected in each pulse peak period. In order to get effective information, a threshold value is set as 0.2 to extract the most relative

84

7 Real-Time Control of Welding Penetration via Arc Sound …

(b)

T4=250ms T4=200ms T3=150ms T2=100ms T1=50ms

A B C

Peak level Background level D EF

O

t (s)

Arc sound pressure s(v)

U (V)

(a)

1 0.5 0 -0.5 -1

Tb

Tp

1000

T

2000

3000

Sample point

Fig. 7.2 The specific scheme of experimental verification: a the wave appearance of the welding current, b the denoising signal by removing environmental noise

signal. The average arc sound pressure S  (i) is calculated under 5000 points. Then, the extracted signal is fitting according to the linear model of arc length changes. For this experiment, the fitting is implemented based on s = 0.0119h + 0.2173. S(i) = x(i) > 0.2

(7.1)

S  (i) = S(i), i = 1, 2, 3 . . . n

(7.2)

h 1 (i) = (S  (i) − b1 )/a1 or h 2 (i) = (S  (i) − b2 )/a2

(7.3)

Figure 7.3 shows a program flow diagram of arc length control based on arc sound signal. After processing the arc sound signal, the deviation data and rectifying voltage are updated every 500 ms. The PID controller is chosen for the experiment on closed-loop system, and considering the complexity and dynamics of welding process, different controller parameters should be proposed for different penetration state of welding pool. So a new PID controller with segmented self-adaption is proposed, which can select PID controller parameters automatically during the arc length tracking process [33]. The formulas of the rectifying voltage are shown in Eqs. (7.6) and (7.7). u stands for the rectifying voltage, h is the offset of arc length and K I , K P and K D are the PID controller parameters. The specific calculation process is as follows: First, while the offset is too small (|h(t)| ≤ 0.2549 mm), the rectifying voltage is calculated by Eq. (7.1), which means there is no need to adjust the position of torch. Because the error is too small to affect the penetration state of weld pool, so it is not necessary to change the welding torch position. Second, while 0.2549 < |h(t)| < 3.782 mm, the Eq. (7.2) is used to calculate the rectifying voltage. The K P , K I and K D are 0.9, 0.3 and 0.015 in this paper, respectively.

7.2 Arc Height Tracking Control Experiment via Arc Sound …

85

Start Initialize Data& Welding Parameters Detect the Rising Edge A?

Welding Parameters Setting

Succeed?

Y

Delay 60ms

N

Y

Arc on Welding Sound Processing and Control Arc off

N

N

Arc Sound Collection

Adaptive PID Controller

Arc Sound Processing

Rectifying Voltage Output

Deviation 3.782 mm), the rectifying voltage is calculated by Eq. (7.3). Because the arc length offset is too large to get a good welding quality, so it is necessary to adjust it immediately. u(t) = K P × h(t)

(7.5)

For the system, it also has an upper limit of 10 V which is the limit of data acquisition card. If the rectifying voltage is over the limit, it will be set as 10 V.

86

7 Real-Time Control of Welding Penetration via Arc Sound …

⎤⎡ K ⎤ P ⎢ ⎥ ⎣ ⎦ u(t) = h(t) h(t)h(t) − h(t − 1) ⎣ K I ⎦ j=0 KD ⎡ ⎤ t  = ⎣h(t) h(t)h(t) − h(t − 1)⎦ K ⎡

t 

(7.6)

j=0

⎧ ⎪ ⎨ 0 0 K = KP ⎪ ⎩ KP

T |h(t)| ≤ 0.2549 mm 0 T 0.2549 < |h(t)| < 3.782 mm KI KD T |h(t)| > 3.782 mm 00

(7.7)

Based on the analysis above, two kinds of experiments have been designed to prove the accuracy of self-adaptive PID controller. One is achieving the arc length control while the welding torch is changing from 4 to 6 mm. Another is achieving the arc length control while the welding torch is fixed; however, the workpiece has an arch shape which changes the welding arc length. The results of monitoring effect at two different working situations are shown in Figs. 7.4 and 7.5. It can observed from Fig. 7.4a that the self-adaptive PID controller could adjust the position of welding torch while arc length is changing from 4 to 6 mm. The adjusting speed is quick enough to maintain a stable welding process. From Fig. 7.4b, the prediction error will vibrate between positive error and negative error as the plate being arch shaped, because the original welding path is set as two stable arc lengths. The prediction error is about 0.5 mm, which is good enough for achieving arc length control. In conclusion, the monitoring of arc length through linear model of arc sound signal can be achieved for real-time while regulating the welding process. The experiment has proved that the arc sound signal is good enough for monitoring of arc length, and is also capable of achieving the welding quality control through arc length monitoring.

7.3 Welding Penetration Control Experiment via Arc Sound Signal of GTAW Welding After feature extraction and recognition of arc sound signal via BPANN model, three penetration states have already been identified. In order to achieve the online control of welding penetration via arc sound signal, a special controller need to be designed on the platform of Al alloy through pulsed GTAW welding process. The program flow is shown in Fig. 7.6. After the initialization of system parameters and setting the welding parameters, the collection has been started along with the welding process. As per the pulse characteristic of pulsed GTAW, the welding current has the pulsed waveform, as shown in Fig. 7.7. According to the analyses above, the

7.3 Welding Penetration Control Experiment via Arc Sound …

87

(a) Original welding path

6

Arc length h(mm)

4 2

Actual motion path

Deviation value

0

Deviation voltage

-2 0

10

30

40

50

60

70

Sample point

(b) 8

Original welding path

6

Arc length (mm)

20

Actual motion path

4

Deviation value

2 0 -2 -4

Deviation voltage 0

100

50

150

Sample point

Fig. 7.4 The monitoring results for two different workpieces: a the flat plate of arc length change, b the arched plate of stable arc length

2mm

Welding path

Welding path 2mm Fig. 7.5 The real product photo for two different workpieces

88

7 Real-Time Control of Welding Penetration via Arc Sound …

Start Initialize Data& Welding Parameters Detect the Rising Edge A?

Welding Parameters Setting

Y

Delay 60ms

N

Succeed?

N

Penetration state Output

Arc Sound Collection

Y

Arc on

Penetration=2 (full penetration)

Arc Sound Feature extracting and regression

Welding Sound Processing and Control

Y

N N

Arc off

BPNN classification

Y

Current output

End Fig. 7.6 The program flow diagram of welding penetration control

(b) T4=250ms T4=200ms T3=150ms T2=100ms T1=50ms

A B C

Peak level Background level D EF

O

t (s)

Tb

Tp T

Sound pressure s/(v)

U (V)

(a)

1 0.5 0 -0.5 2.5

3

Weld time s/(s)

Fig. 7.7 a The waveform of welding current, b collected signal of arc sound

3.5 4 x 10

7.3 Welding Penetration Control Experiment via Arc Sound …

89

collection should be implemented only in the ROI region, so the operation is set to detect the rising edge A of welding current and collect the arc sound signal at 60 ms delay. It will collect 5000 points each time during the pulse peak domain and then the feature extraction and regression of arc sound feature have been accomplished. The features are then fed as input into the BPANN model to forecast the penetration state and the predicted result is compared with the standard penetration parameter. According to the comparison results, three recognition rules of penetration state have been defined as: First, while the recognition result is in the range of (1.5 ≤ |state| ≤ 2.5), this state is considered as full penetration and the welding current is constant. Second, while the recognition result is in the range of (0.5 < |h(t)| < 1.5), this state is considered as partial penetration and it is necessary to increase the heat input to get more penetration. The welding current will be increased as follows: I (t) = I (t − 1) + 5

(7.8)

Third, while the recognition state is in the range of (2.5 < |h(t)| < 3.5), this state is considered as excessive penetration and the heat input need to be decreased for maintaining good welding quality. The welding current is as follows: I (t) = I (t − 1) − 5

(7.9)

While the prediction result is out of the standard range, the welding current maintains the original value. I (t) = I (t − 1)

(7.10)

Based on the comprehensive consideration above, a new piecewise function controller based on BPANN model is proposed to achieve the online monitoring of welding quality based on the acoustic signal. The specific formula is shown in (7.11): ⎧ ⎨ I (t − 1) + 5 I (t) I (t) ⎩ I (t − 1) − 5

0.5 < |h(t)| < 1.5 1.5 < |h(t)| < 2.5 3.5 < |h(t)| < 4.5

(7.11)

Then in order to verify the control effect of the special controller, two kinds of trials have been designed on different workpieces. First, the welding experiments are carried out on constant welding parameter; for other welding, trials are implemented on different shape of workpieces which have added the controller to the control system. The welding quality will be influenced by the shape of workpiece due to the different shape of workpiece leads to uneven heat input, and increasing the heat will affect the penetration of welding pool. Thus, this kind of trials could be used to test the monitoring effect of controller.

90

7 Real-Time Control of Welding Penetration via Arc Sound …

Fig. 7.8 T-shape welded plate under constant welding current and the arc sound energy

It can be concluded from the T-shape workpiece shown in Fig. 7.8 that the welding width is generally getting wider while decreasing the weld plate. During the whole welding process, the weld joint generally changed from partial penetration to full penetration, and then becomes excessive penetration due to the uneven heat dissipation. The heat input becomes the large in the end of welding, so the energy of arc sound signal will have a sharp falling after the general increasing trend. It is obvious that the welding quality is unstable under the constant welding parameters due to the shape of the welding plate. Thus, it is necessary to achieve automatic adjustment of welding parameter during the T-shape workpiece GTAW welding according to the variation of welding penetration. The special piecewise controller based on BPANN is used in T-shape, dumb-bell shape and gradient type workpiece in order to test the control effect. The verification trials are implemented for the same welding parameters on the collected acoustic signal during the welding process. Then after the preprocessing and feature extraction of arc sound signal, the current is adjusted according to the rules of controller. Figure 7.9 shows the result of verification experiment and Fig. 7.10 showed the energy variation trend of arc sound signal under the regulation of controller. Figure 7.11 shows the rectification of peak current during the control experiment of

Fig. 7.9 Welded workpiece under BPANN-PW controller

7.3 Welding Penetration Control Experiment via Arc Sound …

150

Energy E/(J)

Fig. 7.10 Energy of arc sound signal with BPANN-PW controller

91

100 50

20

40

60

80

100

80

100

Sampling number

280

Peak current I/(A)

Fig. 7.11 The rectification of peak current in welding experiment with BPANN-PW controller

260 240 220 200 20

40

60

Sampling number

T-shape workpiece. It can be concluded that the energy of arc sound was vibrated at a certain range due to the adjustment of BPANN piecewise (BPANN-PW) controller as shown in Fig. 7.10. According to the experiment of constant welding parameter, the arc sound energy presents a rising trend causing the variation of heat input. The peak current is rectified at each step and presents the falling trend because in order to decrease the influence of heat input it is important to lower the peak current. As the results showed, the peak current indeed decreased along with the rectification of BPANN-PW controller, which means this controller is suitable for our welding monitoring and control. In addition, it can be seen from the welding plate shown in Fig. 7.9 that the weld joint is maintained a certain width and height, and the welding penetration state is good and even for the welding process. Thus, this special BPANN-PW controller could achieve online welding penetration control for T-shape workpiece and the control effect could get good welding penetration and welding quality [34] (Fig. 7.12). Then in order to verify the control effect, the same verification experiments have been implemented on the dumb-bell shape and gradient-type workpiece. First, the experiments have been implemented under constant welding parameters. Owing to the variation in heat conduction, the welding process could not get good penetration in dumb-bell shape and gradient-type workpiece, and the results are shown in Figs. 7.13a and 7.14a. Then the verification experiments are also implemented on these two

92

7 Real-Time Control of Welding Penetration via Arc Sound …

Fig. 7.12 The flow diagram of online control of penetration state via arc sound signal

Fig. 7.13 Dumb-bell workpiece welded under constant current and BPANN-PW controller. a Topside and backside of workpiece with constant welding current. b Topside and backside of workpiece with BPANN-PW controller

7.3 Welding Penetration Control Experiment via Arc Sound …

93

Fig. 7.14 Energy of arc sound and rectify of peak current in welding experiment with BPANN-PW controller of dumb-bell shape workpiece

kinds of workpiece with the BPANN-PW controller. The flow diagram of the online control of penetration state via arc sound signal is shown in Fig. 7.12. The arc sound signal will be processed after all the procedures above is followed and finally used to achieve the online monitoring of penetration state. It can be directly observed from Figs. 7.13b and 7.15b that the welding quality has been greatly improved by the special controller and the welding current has been controlled at a certain range which is also sensitive to the change of penetration state for different welding joint (Fig. 7.16). The result showed that the welding current always had an unstable area during the starting arc period and then it was generally changed according the analysis rule of piecewise controller. While the recognition value was full penetration, then the welding current maintained the original current, and when the recognition value was partial penetration the welding current should increase 5 A at every step of controlling. On the contrary, the welding current should decrease 5 A every 5000

94

7 Real-Time Control of Welding Penetration via Arc Sound …

Fig. 7.15 Gradient-type welded workpiece under BPANN-PW controller. a Topside and backside of workpiece with constant welding current. b Topside and backside of workpiece with BPANN-PW controller

Fig. 7.16 Energy and rectify of peak current in welding experiment with BPANN-PW controller of gradient type workpiece

7.3 Welding Penetration Control Experiment via Arc Sound …

95

collecting sample while the recognition value was excessive penetration. After the adjustment, the process will continue from the beginning and the new arc sound signal is collected. In conclusion, the online monitoring system could achieve the collection of arc sound signal and feature extraction. The self-design BPANN-PW controller could control the penetration state of GTAW welding process via acoustic signal.

Chapter 8

Microphone Array Technology in Welding Dynamic Process Monitoring

Many researches have been performed in acoustic sensor; however, arc sound of welding itself is a complex feature. It is hard to identify the pure sound signal to achieve utilizing control-based acoustic signal. Most of the studies were implemented on single microphone which could not catch overall information during the welding process. Thus, we decided to consider using more microphones to collect welding sound signal.

8.1 Establishment of Microphone Array Acquisition System The schematic diagram of the monitoring system is shown in Fig. 8.1. It consists of six parts: the robotic system, the vision sensor system, acoustic sensor system, welding system, welding exhaust system and the computer [35]. The vision sensor and the dual-microphone are connected directly to the computer and control by a computer software. The vision system is composed of a CCD and optical system. The result of vision sensor is considered as observation and compared with the analysis of sound signal. The robotic system is ARC Mate 100iC, a six-degree industry robot manufactured by FANUC with a Lincoln Electric Power Wave F355i and AutoDrive 4R90. The welding exhaust system is provided by Ozone Pollution Technology. The acoustic system consisted of two ultimate microphones for professional USB recording. The Yeti Pro microphone utilizes a high-quality 192 kHz/2-bit analogto-digital converter to send remarkable audio fidelity directly into computer. The specification of microphone is shown in Table 8.1. The experiment is carried out using the 4 mm GS 250 steel through pulsed MIG welding. The parameters are shown in Table 8.2. In order to get the relevance between arc sound signal and distance, a series of special experiments were designed and implemented on the GS 250 steel. The height of weld gun is maintained at 4 mm and the welding path is from starting point A to ending point B. The changing path © Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_8

97

98

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Welding system

Welding system

Welding machine Exhaust fan

Welding Robot Welding process control PC

Noise generation Signal collection

Signal collection

Workpiece Microphone #1

Image collection

Microphone #2

Visual sensor Fig. 8.1 The schematic diagram of the experimental system Table 8.1 The specification of microphone Performance

MP201

Power required/consumption

5 V 500 mA (USB)/48 V DC phantom power (analog)

Sample rate

22–192 kHz

Bit rate

24 bit

Capsules Polar patterns Frequency response

3 Blue-proprietary 14 mm condenser capsules Stereo, cardioid, bidirectional, omnidirectional 20–20 kHz

Sensitivity

4.5 mV/Pa (1 kHz)

Max SPL Impedance Power output (RMS) THD Signal to noise Dimensions (extend in stand) Weight (microphone)

120 dB (THD:0.5%1 kHz) >16 ohms 130 mW 0.009% 114 dB 4.72 (12 cm) × 4.92 (12.5 cm) × 11.61 (29.5 cm) 1.2 lbs (0.55 kg)

8.1 Establishment of Microphone Array Acquisition System Table 8.2 Experiment conditions and parameters

99

Parameter type

Value

Parameter type

Value

Impulse frequency (Hz)

2

Wire diameter φ (mm)

1

Feedback arc voltage (V)

23.3

TRIM

0.55

Feedback arc current (A)

120.7

Ar air flow L (l/min)

12

WFS (IPM)

280

Arc height (mm)

3–6 mm

Welding speed V (mm/s)

6

Material type

GS 250

Fig. 8.2 The schematic diagram of the experimental system

is shown in Fig. 8.2. The distance of migration is ±10 mm away from the welding seam center. Two acoustic sensors collected the arc sound signal simultaneously and sent the data to the computer. The program is running through VC++ and MATLAB.

8.2 Research of Blind Signal Separation in Welding Dynamic Process There were splash sound and background interference noise in welding site in addition to arc sound; that is to say, the actual welding environment is a complex environment which distributes multiple sources in space. This means that the sound signal collected by a single microphone is a mixed sound signal containing but being

100

8 Microphone Array Technology in Welding Dynamic Process Monitoring

not equivalent to our interested welding arc sound signal, so it is necessary to use microphone array technology based on multi-channel mixed sound signal collected by microphone array. Blind signal separation algorithm can be used to separate out welding-related sound signals which we are interested in, thus helping to realize quality monitoring of dynamic welding process.

8.2.1 FastICA Blind Signal Separation Algorithm Blind signal separation is a classic problem in the field of signal processing— meaning that original signals are separated from multiple observed signals without the knowledge of the sound sources and the mixed model. So far, there are many ways to solve such problems, but independent component analysis (ICA) method is an important and well-developed way to solve this problem, and FastICA algorithm is currently the most widely used method. FastICA blind signal separation algorithm is a fixed-point algorithm based on the criterion of negative entropy maximization, which was first proposed by Aapo Hyvärinen in 1999 and belongs essentially to a neural network algorithm. The basic idea of FastICA algorithm is projection tracking, using fixed-point iterative method to obtain projection vectors, and ensuring projections that observed signals are projected in the direction of projection vector with non-Gaussian maximum. Figure 8.3 is the diagram of FastICA algorithm, and the approximate maximum of negative entropy is taken as independent criterion of separated signals. At the same time, a nonlinear link g is introduced in the algorithm, because simply maximizing negative entropy results in divergence to infinity, and the introduction of nonlinear link can overcome it. Meanwhile, g is monotonically reversible, so the statistical independence of y is equivalent to the statistical independence of u. The nonlinear function g has been determined according to the following principle, assuming that p(ui ) and p(yi ) denote the probability density function of ui and yi , respectively. They satisfy the following relation:

Fig. 8.3 FastICA algorithm diagram

8.2 Research of Blind Signal Separation in Welding Dynamic Process

p(yi ) =

∂u i p(u i ) p(u i ) =  ∂ yi gi (u i )

101

(8.1)

According to peak power constrained maximum entropy theorem, we can observe that when yi obeys uniform distribution, the output entropy can be the largest. In order to ensure yi satisfies uniform distribution in the range of [0, 1], that is, p(yi ) = 1, then there must be: 

p(u i ) = gi (u i )

(8.2)

Therefore, g needs to satisfy the constraint in (8.2), that is, the derivative of g should be equal to the probability distribution of u. But it is not necessary to strictly satisfy the above constraints. When g is close to the cumulative distribution function of the original signal s, the algorithm can get a good separated result. In practice, the following functions are used in the algorithm: g1 (u) = tanh(a1 ∗ u)

(8.3)

g2 (u) = u ∗ exp(−a2 ∗ u 2 /2)

(8.4)

g3 (u) = u 3

(8.5)

where g1 is suitable for coexistence of sub-Gaussian and super-Gaussian distributions in source signals, g2 is suitable for separating super-Gaussian mixed signals and g3 is suitable for separating sub-Gaussian mixed signals. Speech signals usually satisfy super-Gaussian distributions, so this paper choice is nonlinear function g2 to realize FastICA separation process. In this paper, FastICA toolbox under MATLAB platform is used to implement blind signal separation process. The algorithm flowchart is shown in Fig. 8.4.

8.2.2 FastICA Blind Signal Separation Results In actual welding process, there was serious splash, and the specific performance was that the bursting sound of droplets was obvious, and welding arc sound produced by plasma resonance seemed to be much lower and deeper, that is to say, there were two kinds of sound closely related with welding behaviors in actual welding environment, as well as background noise. Therefore, we need to use FastICA blind signal separation algorithm to distinguish these sound sources in order to facilitate features analysis of sound signals subsequently. Figure 8.5 shows the time and frequency domain analysis [fast Fourier transform (FFT)] results of the first observed signal. It can be seen from the time-domain waveform that the mixed signal contains welding arc sound and splash sound. In

102

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.4 FastICA algorithm flowchart

Fig. 8.5 Time and frequency domain analysis of first observed signal

8.2 Research of Blind Signal Separation in Welding Dynamic Process

103

order to further compare these two different sounds, we extract a single arc sound signal and splash sound signal to analyze, and the comparison results are shown in Fig. 8.6. Figure 8.6a shows that the splash sound signal exhibits a sharp oscillation character in time domain compared with arc sound signal. Figure 8.6b shows that the high-frequency energy of splash sound is significantly enhanced, and the difference in energy distribution in frequency domain further reflects the significant difference in vocal mechanism between these two sounds. From the above analysis, we can conclude that FastICA algorithm successfully recovers splash sound from the mixed signals in pulsed MAG welding process,

(a) Time domain comparison

(b) Frequency domain comparison Fig. 8.6 Comparison between arc sound signal and splash sound signal

104

8 Microphone Array Technology in Welding Dynamic Process Monitoring

but does not recover the welding arc sound, which may be caused by that frequency domain. Energy distribution of arc sound is broad, and its non-Gaussian characteristic is not significant enough. The frequency-domain energy distribution of splash sound is concentrated due to its violent oscillation characteristic, so its non-Gaussian characteristic is obvious, which is more likely to be recovered under the same algorithm (Figs. 8.7 and 8.8).

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process The arc sound generated in MIG welding process belongs to a kind of vibration signal. The features of dual-microphones also have periodicity and statistics features just like the vibration signals. The time-domain feature is widely used in signal process because of its direct visual and easy to get [23].

8.3.1 Feature Extraction of Dual-Microphone Considering the direct-current coupling pattern, the DC bias would disturb the collection effect. Preprocessing needs to be implemented first before feature extraction. m s(i) (8.6) s(i) = s(i) − n=0 m Then, a rectangular window (n = 15,120) was added to the original signal in order to decrease the influence of impulse oscillation. The feature could be a more accurate expression for the changing welding path. For arc sound signal in MIG welding process, the statistical characteristics are the most favorite and widely used features in signal analysis. Six time-domain features were chosen as the evaluation criterion for the arc sound signal. They are standard deviation, mean sound, covariance, root-mean-square, log energy and arc energy, as described in the following Eq. (8.7).   n n N −1  1  1 2 |xi |; En = e= xn (m); xrms =  xi2 ; n i=1 n m=0 i=1   n  1  (xi − x)2 ; Le = log(E) Sd =  n−1 i=1

(8.7)

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process Fig. 8.7 Time and frequency domain analysis of observed and separated signals

(a) Time domain waveforms of observed signals

(b) Frequency domain results of observed signals

(c) Time domain waveforms of separated signals

(d) Frequency domain results of separated signals

105

106

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.8 Time and frequency domain analysis of splash sound signal

The processing results are shown in Fig. 8.9. The features could reflect the changing position of sound source which presents the vibration of welding path. While the weld gun moves from starting point A to ending point B, microphones show different changing pattern. The difference indicated that there was a strong relationship between arc sound energy with distance.

8.3.2 Establish Linear Fitting Model of Dual-Microphone The welding path was changing from −10 to 10 mm during the whole welding process. When combining the arc sound feature with welding deviation ±10 mm, it can be concluded that the arc sound energy had linear relationship with the welding deviation. Then the linear fitting was implemented to dual-microphone separately. 

S1 (i) = k1 a + b1 S2 (i) = k2 a + b2

(8.8)

The fitting results shown in Fig. 8.10 present that microphone 1 was progressively increasing along with the welding path moving from −10 to 10 mm. The microphone 2 was decreasing progressively along with the welding deviation. The linear fitting model of dual-microphone for monitoring the welding path is: 

S1 (i) = 2.9827a(i) + 90.0119 S2 (i) = −2.49676a(i) + 98.7005

(8.9)

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process

(a)

(b)

0.12

200 Microphone 1 Microphone 2

0.11

Energy

150

0.1

STD

107

0.09

100 50

0.08 0

(c)

50 Sample

0 0

100

20

(d)

-3

2

Microphone 1 Microphone 2

x 10

40 60 Sample

80

100

10 5

Log energy

Mean sound

1 0 -1

20

40

60

80

(e)

Microphone 1 Microphone 2 0

x 10

20

40

60

100

0.25 Microphone 1 Microphone 2

0.2 10

RMS

Covariance

80

Sample

(f)

-3

12

-15

100

Sample

-5 -10

Microphone 1 Microphone 2

-2 0

0

8 6 4 0

40

60

Sample

0.1 0.05

Microphone 1 Microphone 2 20

0.15

80

100

0

0

50

100

Sample

Fig. 8.9 The feature extraction of dual-microphone: a standard deviation, b arc sound energy, c mean sound, d log energy, e covariance, f root-mean-square

In order to make the model more flexible and adaptable, five sets of experiments were conducted under the same welding conditions. The statistical features were extracted from the arc sound signal for both microphones. The arc sound energy was chosen to be the one to implement the linear fitting. The linear fitting results are obtained separately for each acoustic sensor, and the results are shown in Fig. 8.11.

108

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.10 The linear fitting of arc sound energy

160

Arc energy

140

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

s=-2.49676a+98.7005

120 100 80 60 s=2.9827a+90.0119

40 -10

-5

0

5

10

Deviation value a(mm)

According to these five linear fitting models, the new model for the monitoring welding path based on dual-microphone could be calculated as follows: ⎧ ai1 + ai2 + ai3 + ai4 + ai5 ⎪ ⎨a = n + b + b b ⎪ i1 i2 i3 + bi4 + bi5 ⎩b = n

(8.10)

The final model for the dual-microphone is: 

S1 (i) = 2.339109a2 (i) + 82.17477 S2 (i) = −2.483544a2 (i) + 93.89946

(8.11)

8.3.3 Establish Linear Fitting Model of Dual-Microphone In order to prove the monitoring effect of this linear model, some verification tests were implemented in this system. Under the same welding condition, the welding path was from −10 to 10 mm and arc sound signal were collected and analyzed by dual-microphone, as shown in Fig. 8.12. Two sets of arc sound feature presented the different changing trend of dual-microphone. Then, the relationship between arc sound energy with welding deviation could be deducted from 4 to 6. The prediction path of welding could be calculated from: ⎧ S1 (i) − 82.17477 ⎪ ⎨ a2 (i) = 2.339109 93.89946 − S2 (i) ⎪ ⎩ a (i) = 2 2.483544

(8.12)

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process

(a)

200

200

Arc energy

150 100

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

180

Microphone 1 Microphone 2

Energy

109

160

s=-2.32659a+82.855

140 120 100 80

50

60

0 0

20

40

60

80

40 -10

100

Sample

(b)

Arc energy

Energy

100

100

50

50

0

20

40

60

80

s=1.882145a+74.07975

0 -10

100

Sample

(c)

140

80

100

Arc energy

120

Energy

100

60

Microphone 1 Microphone 2

40 20

0

20

40

60

5

10

80

20 -10

100

s=-1.523285a+89.67535

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

60 40

s=2.77028a+80.5148

-5

0

5

10

Deviation value a(mm) 160

140

140

Arc energy

120 100 80 60 40

0

80

Sample

(d)

-5

Deviation value a(mm)

120

Energy

10

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

s=-2.4272a+90.7955

150

150

Microphone 1 Microphone 2

0

20

40

60

80

120 100 80

s=2.9827a+90.0119

40 -10

100

Arc energy

150 100 50

150

60

Sample

5

10

80

100

s=-3.643885a+107.47095

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

100 50

40

0

200

Microphone 1 Microphone 2

20

-5

Deviation value a(mm)

200

0

Micro #1 Fit curve for Micro #1 Micro #2 Fit curve for Micro #2

s=-2.49676a+98.7005

60

Sample

(e)

0

5

200

Microphone 1 Microphone 2

0

0

Deviation value a(mm)

200

Energy

s=1.62381a+72.2538

-5

0 -10

s=2.43661a+94.0136

-5

0

5

10

Deviation value a(mm)

Fig. 8.11 The linear fitting of arc sound energy and fitting results: a Test 1, b Test 2, c Test 3, d Test 4, e Test 5

110

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.12 The linear fitting of arc sound energy

200

Microphone 1 Microphone 2

Energy

150 100 50

0

0

60

40

20

Sample

80

100

The predicted results of two acoustic sensors are shown in Figs. 8.13 and 8.14. The blue points in Fig. 8.13 were the predicted deviation which showed obviously linear characteristic just like the welding path. The red points were the absolute predicted error of microphone 1 which ranged from −15.3139 to 15.42 mm. Mean error of microphone 1 was 6.5814 mm. For microphone 2, the absolute error ranged 30

Fig. 8.13 The predicted results of microphone 1

20

Distance

10 0 -10 Predicted deviation Real deviation Error

-20 -30 -40

0

5

10

15

20

25

30

35

Time

40

Fig. 8.14 The predicted results of microphone 2

Distance

20 0 -20

Predicted deviation Real deviation Error

-40 -60

0

5

10

15

20

Time

25

30

35

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process

111

from −8.198 to 3.554 mm and the mean error was −2.5084 mm. The prediction accuracy of microphone 2 was better than microphone 2. Because the microphone 1 was influenced by welding exhaust fan, two prediction models were separately established without any internal relationship. The predicted accuracy of dual-microphone was different and separated from each other during the monitoring of welding path due to the modeling method. The modeling method based on separated acoustic sensor was the same as the single sensor in monitoring the welding process. Also, the predictive effect of these two models was not good enough to monitor the welding path. So further improvement need to be done to the arc sound signal.

8.3.4 Improvement of Prediction Model for Dual-Microphone In order to highlight the advantages of dual-microphone, it is essential to combine the data from two microphones. As the arc sound signals were simultaneously acquired during the welding process, so the arc sound could be expressed by arc sound energy of two microphones in 3D vision just like in Fig. 8.15. It can be concluded that five sets of arc sound energy had the similar changing trend while presented with arc sound based on two acoustic sensors. When 3D arc sound is projected onto two-dimensional axis, one position in the line welding line corresponded to two microphones and also each position had five values for one microphone under the same welding environment, such as x 1 , x 2 , x 3 , x 4 and x 5 . In order to get relative feature of the arc sound energy for the two acoustic sensors, the mean value of these five points were calculated as: x(1,2) =

x1 + x2 + x3 + x4 + x5 5

(8.13)

Fig. 8.15 The 3D vision of arc sound through dual-microphone

Energy 2

200 150 100 50 10

150 100 5

0

Time

50 -5

-10

0

Energy 1

112

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.16 Two-plane projection graph: a Microphone #1, b Microphone #2

The synthetic arc sound of above five sets of data is shown in Fig. 8.17. After calculating separately for each sensor, the two-dimensional data was projected onto 3D axis with three parameters (Fig. 8.16). It can be concluded from the 3D graph that the arc sound could be described as space curve function based on two arc sound energies. The general equation of space curve is: 





y = β 0 + β 1 x1 + β 2 x2

(8.14)

The arc sound feature was fitting as the general space curve in order to get a prediction model for monitoring the welding path. After fitting the arc sound signal we get: y = −11.2439 + 0.2819x1 − 0.1264x2 Fig. 8.17 Three-plane projection graph

(8.15)

8.3 Arc Sound Signal Analyzed by Dynamic Welding Process

113

Fig. 8.18 The prediction result of 3D monitoring model

where βˆ0 = −11.2439, βˆ1 = 0.2819 and βˆ2 = −0.1264, the confidence interval of βˆ0 , βˆ1 , βˆ2 were, respectively, [−13.4794 −9.0084], [0.2681 0.2957] and [−0.1395 −0.1132]. On comparing the results, the arc sound signal also implemented binary linear regression. The fitting model was y = −18.1589 + 0.6621x1 − 0.2375x2 − 0.0025x12 + 0.0004x22

(8.16)

For the binary linear model, βˆ0 = −18.1589, βˆ1 = 0.6621, βˆ2 = −0.2375 and βˆ3 = 0.0004. Then, the verification tests were conducted to these models. The prediction results are presented in Fig. 8.18. It was obvious that the predicted value was close to the real value than the separated single linear model. For two 3D prediction models, the mean absolute error of one variable linear model was 0.6708 mm, the maximum error was 5.6703 mm and the minimum error was 0.3458 mm. For the binary linear model, the mean absolute error was 1.1901 mm, the maximum error was 3.652 mm and the minimum error was 1.9575 mm. These results showed that the prediction effect of one variable linear model was better than binary model. The one variable prediction model was suitable for our monitoring of welding path.

114

8 Microphone Array Technology in Welding Dynamic Process Monitoring

8.4 Welding Dynamic Process Monitoring via Microphone Array 8.4.1 Time Delay Estimation Theory According to the inverse square law of signal propagation, the signal receiving model of the dual-microphone array can be expressed as: xi (n) = s(n − τi )/di + n i (n)

(8.17)

In this formula, s(n − τi ) represents acoustic signal, n i (n) is additive white noise, i = 1, 2, di , τi represent the distance and time delay between sound source and ith microphone, respectively. The cross-correlation function of two-channel signals x1 (n) and x2 (n) can be expressed as: R12 (τ ) = E(x1 (n)x2 (n − τ ))

(8.18)

Substituting Eq. (8.17) into Eq. (8.18) we get: R12 (τ ) = α1 α2 E(s(n − τ1 )s(n − τ2 − τ )) + α1 E(s(n − τ1 )n 2 (n − τ )) + α2 E(s(n − τ2 − τ )n 1 (n)) + E(n 1 (n)n 2 (n − τ ))

(8.19)

α1 and α2 are the attenuation factors determined by propagation distance. In most cases, the noise signals are independent of each other, and formula 8.19 could be simplified as: R12 (τ ) = α1 α2 E(s(n − τ1 )s(n − τ2 − τ )) = α1 α2 R SS (τ − (τ1 − τ2 ))

(8.20)

R SS is the cross-correlation function of sound source signal. It can be concluded from the properties of the autocorrelation function that while τ = (τ1 − τ2 ), R12 (τ ) has the maximum value. The value of the two-way observation signal when the cross-correlation function obtains the maximum value is the time delay value. In the actual environment, the existence of noise and reverberation will weaken the maximum value of Eq. (8.20). When the noise and reverberation are relatively serious, multiple peaks will appear which seriously affects the accuracy and precision of time delay estimation. Generalized cross-correlation (GCC) is the solution to this problem. It is to weight the signal in the power spectrum domain, highlight the signal and suppress the noise, so as to sharpen the peak value of the correlation function at the time delay, expressed as:

8.4 Welding Dynamic Process Monitoring via Microphone Array

115

∞ RGCC (τ ) =

ψg (ω)φ12 (ω)eiωr dω

(8.21)

−∞

Here, ψg (ω) is the weight function, φ12 (ω) is the power spectrum of R12 (τ ), and ω is the angular frequency.

8.4.2 Inspection and Localization of Welding Defects After the inspection of sound source, the next step is to analyze the welding defect. The analysis experiments are designed in large current and low welding speed in order to generate welding defect. A. Feature Analysis of Welding Sound for Welding Defect Three weld seams are selected to analyze the characteristic changes of sound signals during the welding process. First, in actual welding process, when burn-through defects happened, the obvious changes could be heard, mainly in two aspects: (1) welding arc tones significantly changed; and (2) splash sound intensity significantly reduced. It was also noted that: (1) the changing times of welding arc tones corresponded to the number of welding burn-through defects; (2) where the splash intensity weakened did not correspond to burn-through defects. Therefore, the changes of welding arc tones and splash intensity are chosen as the two focus. B. Arc Tones Variation During Welding Process There are three typical sound signal states in actual welding process, namely, the first state—not burn-through with strong splash, the second state—not burn-through with no or light splash, and the third state—burn-through. In order to find out the specific changes of arc tones during the welding process, it is necessary to analyze and compare these three typical sound signal states. Take the first observed signal of the first weld seam as an example. We take 8000 sampling points of three typical states to analyze. The time-domain waveforms of three typical states are shown in Fig. 8.19. The sound pressure amplitude of splash signal is significantly higher than that of arc sound signal, and the variation in time-domain waveform of burn-through state is obviously different from that of the other states. Since the variation in signal waveform is mainly represented by frequency domain energy distribution, we need to further analyze three typical signals in frequency domain. The FFT results of three typical signals are shown in Fig. 8.20. It can be seen that the low-frequency band energy distribution of burn-through sound signal is significantly different from that of the other two states, which is consistent with the previous conclusion, that is, the arc tones suddenly become more dull; the energy of the band above 2000 Hz is relatively weak, especially compared with the first typical signal state (the energy distribution of splash signal is mainly concentrated in high frequency), which is also consistent with the previous analysis, namely, when

116

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.19 Three typical states of welding sound signals

Fig. 8.20 FFT results of three typical signals

burn-through defect occurred, the splash intensity significantly reduced. The lowfrequency band energy distribution of three typical signals is shown in Fig. 8.21. It can be observed that 500–1000 Hz band energy of burn-through state is significantly higher than that of the other two states. The time-domain waveform of three burnthrough state signals and the energy distribution of 0–4000 Hz band are shown in Fig. 8.22. The signal characteristics of three burn-through defects are consistent, and the energy of 500–1000 Hz band is significantly enhanced. In order to further verify that the selected 500–1000 Hz band energy can identify welding burn-through defects, this paper analyzes the short-term energy of the first observed signal of three weld seams through windowing, using Hamming window

8.4 Welding Dynamic Process Monitoring via Microphone Array

117

Fig. 8.21 Low-frequency energy distribution of three typical signals

Fig. 8.22 The time and frequency domain waveform of three welding sound signals in burn through

of 3000 sampling points. The results of 500–1000 Hz band short-term energy analysis are shown in Fig. 8.16. It can be clearly found that when burn-through defects occurred, the energy of 500–1000 Hz band suddenly increased. Further analysis indicates that the abrupt changing point of 500–1000 Hz band energy corresponds to the end position of burn-through defect. This phenomenon can be explained by the fact that at the end of burn-through defect, a closed cavity is formed in weld seam, and there is a resonance effect of arc in this cavity, resulting in significant changes in arc tone (Fig. 8.23). C. Arc Tones Variation During the Welding Process According to the conclusion of section III, we know that FastICA blind signal separation algorithm can successfully separate out the splash sound signal. So we can

118

8 Microphone Array Technology in Welding Dynamic Process Monitoring

Fig. 8.23 Short intensity variation during the welding process

analyze the relation between the change of splash intensity and welding burn-through defects based on the results of blind signal separation. The intensity changes of the splash signal (separated signal ic1) are analyzed from four aspects, such as short-term energy, standard deviation, kurtosis and skewness. Short-term energy directly reflects splash intensity, and it is calculated as follows: En =

N −1 

xn2 (m)

(8.22)

m=0

where xn (m) is the sound pressure value at the mth sampling point of the nth frame and N is the window length (3000 in the text). The standard deviation, as an indicator of the dispersion degree of data distribution, can be used to measure the fluctuation level of welding sound signal, which is calculated as follows:  σn =

N −1 1  [xn (m) − μn ]2 N − 1 M=0

(8.23)

where μn is the mean of the nth frame signal. Kurtosis can be used to describe the steepness of data distribution compared with normal distribution: more than 3 is relatively steep, and less than 3 is relatively flat. The bigger the absolute value of the peak, the greater the difference between the data distribution and normal distribution. Kurtosis can be calculated by the following formula:

8.4 Welding Dynamic Process Monitoring via Microphone Array

Kurtosis =

N −1 1  [xn (m) − μn ]4 /σn4 N − 1 m=0

119

(8.24)

Skewness can be used to measure the symmetry of the data distribution compared with normal distribution; bigger than 0 indicates that at the right side of data distribution there are many extreme values; less than 0 means that at the left side there are many extreme values. The bigger the absolute skewness value is, the greater the skewness of the data distribution is compared to normal distribution. Skewness can be calculated by the following formula: Skewness =

N −1 1  [xn (m) − μn ]3 /σn3 N − 1 m=0

(8.25)

In this paper, a short-time analysis function has been proposed to realize shorttime energy analysis, including short-time energy, standard deviation, kurtosis and skewness of splash signal. The results of analysis are shown in Fig. 8.24. In contrast with three burn-through defects in first weld seam, it is found that the short-time energy, standard deviation and kurtosis distribution change obviously in the corresponding positions, but the change of skewness is less obvious. Here, we choose short-term energy as an important characteristic, which is more directly correlated with splash intensity.

Fig. 8.24 Short-term analysis result of splash signal

120

8 Microphone Array Technology in Welding Dynamic Process Monitoring

8.4.3 Welding Defects Localization and Recognition According to the analyses above, the process of inspection of welding defect can be concluded as shown in Fig. 8.25. The welding defect will be judged in two ways after preprocessing and feature extraction. When the mutation happened, there are two different processed algorithms accomplished. First, the recognition of welding defect has been conducted by BP_AdaBoost. The features are set to be the input of the classification model. The structure of BP_AdaBoost has been introduced in another paper [14–16]. Secondly,

Fig. 8.25 Process of inspection of welding defects

8.4 Welding Dynamic Process Monitoring via Microphone Array

121

Fig. 8.26 The results of welding defect localization

according to the time delay and energy ratio, calculate the formula and get the prediction result from least squares robust regression equation just as the localization of sound source. In order to verify the prediction results, the experiments have been conducted under big current and low welding speed to generate failure welding. The prediction results are shown in Fig. 8.26. The results showed that the position of welding defect could be inspected correctly. The prediction number of welding defects also corresponds to the actual welding process. However, the prediction of welding defect is more accurate at the end of the generation than at the beginning. The next step is trying to optimize the recognition algorithm in order to judge the beginning of the defect generation, at least about the size of defect. This new way still gets a good effective result in welding defect analysis and localization. Combined with the classification model of welding defect, this method could provide the judgment about the type of defect and location of the welding defect.

Chapter 9

Multi-source Information Fusion Between Welding Arc Sound and Other Welding Dynamic Processes

As welding is a highly nonlinear and time-varying complicated thermal process accompanied by strong current and intense emission of sound, arc light and heat, the single sensor applied in the above reports can only obtain part information reflecting the penetration state instead of the comprehensive information. Not to mention that the sensors are likely to be influenced by various factors such as arc disturbance, spatter and noise. Therefore, many researchers attempted to utilize multisensory to implement welding process monitoring. Chen et al. [52] used multi-sensor information fusion technology in pulsed GTAW to obtain different information about the welding process. As for the application of multi-sensor for welding penetration monitoring, some researchers had done a series of studies about multi-sensory data fusion of optical and acoustic sensors. Zhang and Chen [54] implemented seam penetration identification using improved support vector machine based on fusion of sound, voltage and spectrum signals. Wu et al. [55] extracted acoustic features and visual features as the input of deep belief network (DBN) used for identifying the VPPAW penetration state. Traditional optical and acoustic signal process and feature extraction methods applied in the above reports need precise calculation and large amounts of memory of computers, such as average filtering, Fourier transform, wavelet decomposition and so on, which are impossible for human beings. A novel methodology for real-time penetration monitoring of aluminium alloy in pulsed GTAW was proposed using pattern-based visual and acoustic feature extraction and ensemble learning method (Fig. 9.1). The proposed methodology can predict penetration well through imitating welder’s information-perceiving and decision-making process to visual-acoustic signal during pulsed GTAW process, which can be regarded as a promising technique for intelligent welding. In future, not only welding penetration, welding defect are worth to be investigated. Pattern-based feature extraction method and machine learning method might be an effective tool for real-time weld defect monitoring, which

© Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_9

123

124

9 Multi-source Information Fusion Between Welding Arc …

Fig. 9.1 The result of acquired visual and acoustic information

will enhance welding quality effectively during automatic welding process. Intelligent welding technology is a hot topic and worth to be paid more attention. There are lots of works that have studied in this area and more works are willing to be done about intelligent welding.

Chapter 10

Summary and Conclusions

The arc sound signal had been studied for many years. The research scope is ranging from mechanism to control system. The SJTU have been doing a lot of works about the online welding quality monitoring. Many fundamental research have been implemented about the characteristics of welding sound, the processing of welding sound signals and the pattern recognition of the welding sound signal and welding penetration. All of these studies provide the technical basis for achieving the welding automation using acoustic signal. The automatic welding can only be achieved in laboratory. For industrial application, the arc sound signal still cannot offer stable control accuracy for the monitoring system. The online GTAW welding monitoring still has some problems to solve. Therefore, to achieve the monitoring of GTAW welding quality using arc sound signal will be the research topic in the future.

© Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0_10

125

Bibliography

1. Wu L, Chen S (2000) Intelligent welding technology. National Defense Industry Press, Beijing 2. Chen SB, Wu J (2009) Intelligentized technology for arc welding dynamic process. LNEE, vol 29. Springer, Berlin 3. Chen S, Lv N (2014) Research evolution on intelligentized technologies for arc welding process. J Manuf Process 16:109–122 4. Tarn TJ, Chen SB, Chen XQ (2015) Robotic welding, intelligence and automation. In: The advances in intelligent systems and computing. Springer 5. Du QY, Chen SB, Lin T (2006) Inspection of weld shape based on the shape from shading. Int J Adv Manuf Technol 27(7):667–671 6. Lee SK, Na SJ (2002) A study on automatic seam tracking in pulsed laser edge welding by using a vision sensor without an auxiliary light source. J Manuf Syst 21(4):302–315 7. Fan C, Lv F, Chen S (2009) Visual sensing and penetration control in aluminum alloy pulsed GTA welding. Int J Adv Manuf Technol 42:126–137 8. Ma H, Wei S, Li L, Lin T, Chen S (2011) Mixed logical dynamical model of the pulsed gas tungsten arc welding process with varied gap. Proc Inst Mech Eng I-J Syst Control Eng 225(3):270–280 9. Xu YL, Fang G, Chen SB, Zou JJ, Ye Z (2014) Real-time image processing for vision-based weld seam tracking in robotic GMAW. Int J Adv Manuf Technol 73:1413–1425 10. Tam J, Huissoon J (2005) Developing psycho-acoustic experiments in gas metal arc welding. In: Proceedings of the IEEE international conference on mechatronics & automation, Niagara Falls, Canada, pp 1112–1117 11. Saini D, Floyd S (1998) Investigation of gas metal arc welding sound signature for on-line quality control. Weld J 77(4):172–179 12. Tam J (2005) Methods of characterizing gas metal arc welding acoustics for process automation. University of waterloo 13. Grad L, Grum J (2004) Feasibility study of acoustic for on-line monitory in short circuit gas metal arc welding. Int J Mach Tool Manuf 44:551–561 14. Cudina M, Prezelj J (2003) Evaluation of the sound signal based on the welding current in the gas-metal arc metal welding process. Proc Inst Mech Eng 217(5):483–494 15. Sun A, Kannatey-Asibu E, Gartner M (1999) Sensor systems for real-time monitoring of laser weld quality. J Laser Appl 11(4):153–168 16. Kang MJ, Rhee S (2001) The statistical models for estimation the amount of spatter in the short circuit transfer mode of GMAW. Weld J 80(1): l–8 17. Wang JF, Chen B, Chen HB, Chen SB (2009) Analysis of arc sound characteristics for gas tungsten argon welding. Sens Rev 29(3):240–249 18. Sánchez Roca A, Carvajal Fals H, Blanco Fernández J et al (2009) Stability analysis of the gas metal arc welding process based on acoustic emission technique. Weld Int 23(3):173–180 © Springer Nature Singapore Pte Ltd. 2020 N. Lv and S. Chen, Key Technologies of Intelligentized Welding Manufacturing, https://doi.org/10.1007/978-981-15-2002-0

127

128

Bibliography

19. Ao SS (2012) Theory and method of laser welding process quality detection based on microphone array. Tianjin University 20. Ishi CT, Chatot O, Ishiguro H et al (2009) Evaluation of a MUSIC-based real-time sound localization of multiple sound sources in real noisy environments. IEEE, pp 2027–2032 21. Zhao LQ, Wang JF, Lin T et al (2010) Analysis of GTAW molten pool oscillation model for aluminium alloy. J Shanghai Jiaotong Univ (s1):92–94 22. Wang JF, Yu HD, Qian YZ et al (2011) Feature extraction in welding penetration monitoring with arc sound signals. Proc Inst Mech Eng B-J Eng Manuf 225(B9):1683–1691 23. Lv N (2014) Research on penetration feature recognition and real-time control of aluminum alloy pulsed GTAW based on arc acoustic signal. Shanghai Jiaotong University 24. Lv N, Chen S (2011) Investigation on acoustic signals for on-line monitoring of welding. In: Robotic welding, intelligence and automation. Springer, Berlin, pp 235–243 25. Xu Y, Fang G, Lv N, Chen S, Zou J (2015) Computer vision technology for seam tracking in robotic GTAW and GMAW. In: Robotics and computer integrated manufacturing, vol 32, pp 25–26 (IF:1.84) 26. Zhang T, Wu MH, Zhao YZ, Chen SB (2014) Motion planning for a new-model obstaclecrossing mobile welding robot. Ind Robot-Int J 41:87–97 27. Zhang Z, Chen S (2014) Real-time seam penetration identification in arc welding based on fusion of sound, voltage and spectrum signals. J Intell Manuf 1–12 28. Lv N, Zhong JY, Chen HB, Lin T, Chen SB (2014) Real-time control of welding penetration during robotic GTAW dynamical process by audio sensing of arc length. Int J Adv Manuf Technol 74:235–249 29. Lv N, Xu YL, Fang G, Zhao H, Chen SB (2015) Mechanism analysis and feature extraction of arc sound channel for pulse GTAW welding dynamic process. In: Robotic welding, intelligence and automation. Springer 30. Lv N, Xu Y, Zhong J, Chen H, Wang J, Chen S (2013) Research on detection of welding penetration state during robotic GTAW process based on audible arc sound. Ind Robot: Int J 40(5):474–493 31. Lv N, Xu YL, Fang G, Yu XW, Chen SB (2016) Research on welding penetration state recognition based on BP-Adaboost model for pulse GTAW welding dynamic process. In: 2016 IEEE workshop on advanced robotics and its social impacts (ARSO). IEEE, pp 100–105 32. Na L, Ji-yong Z, Hua-bin C, Shan-ben C, Ji-feng W (2013) Penetration feature extraction and modeling of arc sound signal in GTAW based on wavelet analysis and hidden Markov model. In: 2013 IEEE international symposium on industrial electronics (ISIE), pp 1–6 33. Lv N, Zhong J, Chen H, Lin T, Chen S (2014) Real-time control of welding penetration during robotic GTAW dynamical process by audio sensing of arc length. Int J Adv Manuf Technol 74(1–4):235–249 34. Lv N, Xu YL, Li SC, Chen SB (2017) Automated control of welding penetration based on audio sensing technology. J Mater Process Technol 250:81–98 35. Lv N, Fang G, Zhao H, Chen S, Zou J (2016) Real-time monitoring of welding path in pulse metal-inert gas robotic welding using a dual-microphone array. Int J Adv Manuf Technol 2016:1–14 36. Zhang ZF, Chen XZ, Chen HB, Zhong JY, Chen SB (2014) Online welding quality monitoring based on feature extraction of arc voltage signal. Int J Adv Manuf Technol 70:1661–1671 37. Zhang T, Wu MH, Zhao YZ, Chen XQ, Chen SB (2014) Optimal motion planning of mobile welding robot based on multivariable broken line seams. Int J Robot Autom 29:215–223 38. Yang CD, Zhong JY, Chen YX, Chen HB, Lin T, Chen SB (2014) The realization of no back chipping for thick plate welding. Int J Adv Manuf Technol 74:79–88 39. Yang CD, Zhang HJ, Zhong JY, Chen YX, Chen SB (2014) The effect of DSAW on preheating temperature in welding thick plate of high-strength low-alloy steel. Int J Adv Manuf Technol 71:421–428 40. Chen YX, He YS, Chen HB, Zhang HJ, Chen SB (2014) Effect of weave frequency and amplitude on temperature field in weaving welding process. Int J Adv Manuf Technol 75:803– 813

Bibliography

129

41. Lv N, Xu Y, Zhang Z, Wang J, Chen B, Chen S (2013) Audio sensing and modeling of arc dynamic characteristic during pulsed Al alloy GTAW process. Sens Rev 33(2):141–156 42. Zhang ZF, Yu H, Lv N, Chen SB (2013) Real-time defect detection in pulsed GTAW of Al alloys through on-line spectroscopy. J Mater Process Technol 213:1146–1156 43. Yu H, Xu Y, Lv N, Chen H, Chen S (2013) Arc spectral processing technique with its application to wire feed monitoring in Al-Mg alloy pulsed gas tungsten arc welding. J Mater Process Technol 213(5):707–716 44. Yu H, Chen H, Xu Y, Zhang Z, Chen S (2013) Spectroscopic diagnostics of pulsed gas tungsten arc welding plasma and its effect on weld formation of Al-Mg alloy. Spectrosc Lett 46(5):350– 363 45. Yu H, Ye Z, Chen S (2013) Application of arc plasma spectral information in the monitor of Al-Mg alloy pulsed GTAW penetration status based on fuzzy logic system. Int J Adv Manuf Technol 68(9–12):2713–2727 46. Zhang Z, Chen X, Chen H, Zhong J, Chen S (2013) Online welding quality monitoring based on feature extraction of arc voltage signal. Int J Adv Manuf Technol 1–11 47. Ye Z, Fang G, Chen S, Dinham M (2013) A robust algorithm for weld seam extraction based on prior knowledge of weld seam. Sens Rev 33(2):125–133 48. Zhang HJ, Cai CB, Yu ZS, Chen SB (2013) Control of root pass stress by two-sided arc welding for thick plate of high strength steel. Rev Adv Mater Sci 33:298–304 49. Xu Y, Zhong J, Ding M, Chen S (2013) The acquisition and processing of real-time information for height tracking of robotic GTAW process by arc sensor. Int J Adv Manuf Technol 65(5– 8):1031–1043 50. Xu Y, Lv N, Zhong J, Chen H, Chen S (2012) Research on the real-time tracking information of three-dimension welding seam in robotic GTAW process based on composite sensor technology. J Intell Robot Syst 68(2):89–103 51. Chen X-Z, Huang Y-M, Chen S-B (2012) Model analysis and experimental technique on computing accuracy of seam spatial position information based on stereo vision for welding robot. Ind Robot: Int J 39(4):349–356 52. Chen B, Wang J, Chen S (2010) Prediction of pulsed GTAW penetration status based on BP neural network and D-S evidence theory information fusion. Int J Adv Manuf Technol 48(1–4):83–94 53. Chen B, Chen S (2010) Multi-sensor information fusion in pulsed GTAW based on fuzzy measure and fuzzy integral. Assem Autom 30(3):276–285 54. Zhang Z, Chen S (2017) Real-time seam penetration identification in arc welding based on fusion of sound, voltage and spectrum signals. J Intell Manuf 28(1):207–218 55. Wu, D, Huang Y, Chen H, He Y, Chen S (2017) VPPAW penetration monitoring based on fusion of visual and acoustic signals using t-SNE and DBN model. Mater Des 123:1–14