This item was submitted to Loughborough's Research Repository by the author. Items in Figshare are protected by copyright, with all rights reserved, unless otherwise indicated.

The real time analysis of acoustic weld emissions using neural networksThe real time analysis of acoustic weld emissions using neural networks

PLEASE CITE THE PUBLISHED VERSION

PUBLISHER

European Institute for the Joining of Materials

VERSION

AM (Accepted Manuscript)

PUBLISHER STATEMENT

This work is made available according to the conditions of the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0) licence. Full details of this licence are available at:https://creativecommons.org/licenses/by-nc-nd/4.0/

LICENCE

CC BY-NC-ND 4.0

REPOSITORY RECORD

Burge, K.L., T.J. Harris, Raymond Stroud, and John McCardle. 2019. “The Real Time Analysis of AcousticWeld Emissions Using Neural Networks”. figshare. https://hdl.handle.net/2134/26282.

J. R. McCardle 1997

The International Conference on the Joining of Materials - JOM

6th International Conference

Helsingor, Denmark, April 5 - 7, 1993

pp 60 - 67

J. R. McCardle 1997

THE REAL TIME ANALYSIS OF ACOUSTIC WELD EMISSIONS USINGNEURAL NETWORKS

J.R. McCardle B.Sc., K.L. Taylor-Burge B.Sc., M.A., R.R. Stroud M.Tech., Ph.D.,T.J. Harris B.Sc., Ph.D.

Neural Applications Group, Department of Design, Brunel University, United Kingdom

Synopsis.

Artificial Neural Networks (ANNs) are becoming an increasingly viable computing toolin control scenarios where human expertise is so often required. The development ofsoftware emulations and dedicated VLSI devices is proving successful in real worldapplications where complex signal analysis, pattern recognition and discrimination areimportant factors.

An established observation is that a skilled welder is able to monitor a manual arcwelding process by subconsciously changing the position of the electrode in response toan adverse change in audible process noise. Expert systems applied to the analysis ofchaotic acoustic emissions have failed to establish any salient information due to theinabilities of conventional architectures in processing vast quantities of erratic data at realtime speeds.

This paper describes the application of a hybrid ANN system, utilising a combination ofmultiple ANN architectures and conventional techniques, to establish system parameteracoustic signatures for subsequent on line control.

Introduction.

The Neural Applications Group at Brunel University, UK, has aimed its research at thedevelopment of autonomous products and control systems utilising state of the arttechnologies. One area of continuing study has been the development of a fullyintegrated, hybrid control system for automated welding processes; in particularindustrial semi-automated submerged arc welding (SAW).

The successful application of Artificial Neural Networks (ANNs) to interpret ultrasonicechoes of the welding head in real time has yielded a system capable of seam trackingand monitoring weld penetration, highlighting the ANNs computational power whendealing with erratic, noise polluted data [1].

The relationship between the process variables of a weld and acoustic emission (AE),both audible and inaudible, has been the subject of various much research. It is anestablished observation that a skilled manual welder is able to intrinsically position theelectrode and vary the arc length in response to adverse fluctuations of process noise.

Recent attempts to establish salient relationships between AEs and process parametershave utilised conventional digital signal processing techniques and expert systems toclassify the data [2]. The inability of these systems to respond correctly to novel anderratic inputs was compounded by speed limitations in manipulating vast quantities ofdata in real time, causing data explosions and consequently negative results [3].

I

J. R. McCardle 1997

The aim of this research is the development of a hybrid system incorporatingconventional DSP methods and ANN techniques to classify features within acoustic datarelating to weld parameter fluctuations, in real time. It is envisaged that the acousticmethod will complement the existing ultrasonic array in ultimately providing a singlesystem able to monitor and control all the process variables.

The Weld Parameters

The creation of an optimum submerged arc weld is dependent upon six primaryparameters;

weld plate preparation (fit up)position of the electrode (seam tracking)rate of travel along the seamconsistency of gravitational feed of the granular fluxweld voltageweld current (determined by the rate of feed of the sacrificial electrode).

The nature of raw acoustic data is such that emissions could contain clues relating to allthese parameters and eventual investigation will be given to achieving all possibleacoustic signatures for diagnostic purposes. The primary objective, however, is theidentification of signal patterns associated with changes in weld voltage. This method,when run in conjunction with the existing ultrasonic weld penetration monitor, wouldrender a system capable of maintaining an optimum balance between voltage and currentand hence achieve on line weld stability.

Data Acquisition

The transducer element consisted of three omni-directional electret condensermicrophones (ECMs). The ECMs selected exhibited uniform frequency responsecharacteristics suitable for infra, audible and ultrasound detection. The operation ofECMs negate the possible influence of induced noise created by the magnetic flux of thewelding arc [4]. The transducer was mounted 300mm from the welding head togetherwith suitably noise shielded pre-amplification.

The signals were then subjected to a series of eight active bandpass filters within theaudible range, a low pass filter to isolate infrasound (break frequency 15Hz) and a highpass filter (break frequency 201(Hz) to emphasise ultrasonic frequencies.

The data manipulation was achieved with the TMS 320C30 system board hostingHypersignal Workstation software. The transient capture sampling frequency was set to40kHz and 100ksamples enabling 2.5 sec of data capture at a 16 bit resolution. Sampleswere taken of ambient conditions as well as singular and combined readings of weldingancillary equipment such as the mains transformer, fume extractor fan and the kineticcontrol system. This provided sound intensity levels and features which could beidentified within the final weld recordings.

Three welding runs were then initiated, with preset parameters and ten readings takenfrom each. The initial weld parameters were set at the default, the second run with lowvoltage and the final with high voltage in relation to weld current as shown in table 1.

2

J. R. McCardle 1997

V IWeld Run 1 (default) 41.1 820AWeld Run 2 29.7 785AWeld Run 3 61.2 815A

Table 1.Weld Voltage & Current Settings

The captured data was then subject to a 1024 point FFT analysis and a contour mapobtained as shown in Figs 2, 3 & 4. Further manipulation enabled a power spectraanalysis by performing an rms of the 715 I-F1 frames over the 2.5 sec capture period.This effectively removed transient noise pollution and yielded the most prevalent featuresin the frequency domain. Real time frequency and power spectrums were then monitoredto gauge consistency over a complete weld run of 1 Metre lengths.

The Neural Classification Approach.

ANNs have been successfully implemented in scenarios where high speed signalprocessing of noisy or corrupted data has thwarted expert systems [2, 3, 5, 6]. Thedifference in the two systems stems from their architectures. The conventional computerrelies on a single powerful processor passing synchronised data, whereas, neural systemscontain many simple processors (neurons) heavily interconnected. This enables theparallel processing necessary to deal with complex signals at high speed. The advantagesof networks lie in their ability to generalise and have consequently proven themselves inpredictive systems requiring the assessment of novel data [7].

Observations of biological neural systems have shown that different neurons andtopologies perform different functions. Similarly, differing ANN architectures andlearning algorithms exhibit different characteristics of cognition and recognition. Thesecharacteristics can be exploited to great effect if the correct paradigm is selected for thesituation. Pattern recognition problems are relieved by the extraction of salient features,by pre-processing, on which a pattern classifying network can base its considerations [8].

The well established ANN architecture of Back Error Propogation [9], often refered toas the 'work-horse' of neural computing, could be exploited in the area of signalprocessing. They have been successfully employed in bandpass functions and signalenhancement and could be used in pre-processing raw acoustic signals prior to DSPoperation.

The basis of the signal pattern classifier for this research is the Kohonen Self -Organising Feature Map [10]. It is a two layered network of fully connected neurons thatcan self organise, from a random starting point, a topological map showing the naturalrelationships of the patterns used in its training data sets. This is achieved by thecompetitive activation of an output neuron dictated by the learning algorithm. Theoutput neurons are arranged in a matrix so that each has theoretical neighbours. Theactivation threshold of neighbouring neurons are adjusted proportionally to the distancein vector space from the winning neuron. In this way each neuron is inhibited orencouraged to activate when presented with the next training set. The training continues

3

)SAF DevimaticSAW

• I• I• 1>

IECMs 1 Pre-amp

Ultrasolic arrayC'

AN

TMS320C30DSP

Active Filters

N

Process I IBM PS/2 70 IS

J. R. McCardle 1997

until a state of equilibrium occurs or an acceptable level of accuracy is reached. Thissystem has proven successful in the interpretation of speech phonemes [11] and thediagnosis of lung infections via AEs [12]. In the same manner it is anticipated that thisparadigm is capable of isolating similarities and grouping the signals common to unstablewelds and separating them from stable inputs. Further sub-divisions within thetopological map would include the remaining parameters which influence unstable welds.

Fig.1 illustrates the experimental set up. Research in this field of application continues.

Figure 1Experimental set up

Results

So far research has concentrated on establishing usable characteristics in the frequencydomain for introduction to a suitably tailored Kohonen classifier network.

Figures 2, 3 & 4 illustrate frequency contour mappings of the recorded data for weldruns 1, 2 & 3 (see table 1) respectively, from which the following observations weremade:-

1. The vertical readings indicate noise from gas leakage through the flux. Theconcentration of this noise being most apparent on the low voltage run.

2. The high voltage and default runs show a consistent frequency feature in the14kHz to 16kHz range. .

3.The frequency intensities within this bandwidth are greater in the high voltageweld.

4. There is a substantial absence of low frequencies, up to 4kHz, on the defaultweld setting.

Figures 5, 6 & 7 illustrate the comparisons of real time freeze frame PF1 analysis takenfrom a continuous weld run. The signals tend to enforce the observations made from thefrequency contour map. The salient features are further enhanced with the real timepower spectra analysis giving a 1024 point FFT 25 frame rms. An example for a highvoltage setting is shown in Fig. 8.

4

Fuel: deFweldu FFT Length: 1024Overlap: 500

Fs: 40.06 kHz, Win: Hamm Framosizo: 640

== mw -5.000 dB/COLOR,

c:

F3 IflPF7 PF:MFT

1.1.0 15Fp Up ELY-LE

11,4-ael

C

ON 0. 000-2. 499CONTOUR Time spanned:

SPECTROGRAPH ICdB/COLOR, CONTOUR ON

ANAL_ YS I SCD

CD_J

C

Fuel: weldluau

Fs: 40.06 kHz. Win: Hamm

FFT Longth: 1024Ouorlap: 500Framesizo: 640

-5.000 dn/coLon, CONTOUR OH

ANALYS I STime spanned: 0.000-2.499 S

F3 I-CLPF2 PR.31/TU 1)0 11.3

UFLF.0 ClA AUT

•444:

C. G. L. Or

Fuel: woldhuau FET Length: 1024Ou.rlap: 500

Fs: 40.06 kHz. Win: Hamm Fre.m.5siro: 640

0#60W

SPECTROGRAPH I C• • •

1,0S4 %AM swa sew was wow

J.12.McCardle1997

SPECTROGRAPHIC ANALYSIS

Figure 2.Frequency Contour Map, Default Weld.

Figure 3.Frequency Contour Map, Low Voltage.

Figure 4.

Frequency Contour Map, High Voltage.

5

0 000 2.0001...H.FREO"Ne. 4--

F11.: $4.1=11‘.....0 FFT Lomatl,: 1024F11.2: actfuroldu Fr.:m.1Di... 040SaNmp.1ima Froclu.,.w: 40.00 1,1-4= Trio Trace 1

0

C. , , B: 4— =1..-F11.1: rr7 X024

a.C.4.1du Frames 1.. : 040From.u.n.w : 40.00 Tr-10

—3.000o-

- 100. 20.031,H.

riE SPEC T' RUM FINAL_EF0 .=

=-71

0 000 2.0101061-1.FRE0

2.0001,H. /01N.: 2.0.00H=FREM

J. R. McCardle 1997

RE — T I ME SPEC T RUM ANAL R

Figure 5.Comparative Freeze Frame 141-1', Low Voltage weld & Default Weld

Figure 6.Comparative Freeze Frame 1-1- 1, High Voltage Weld & Default Weld

FRE — SF=•E -r IRI_J0

—100.07000

C , , :FFT L....Nat.,: 1024

640Froaqu..,...d: 40.00 RH= Trla Trace: 1

Figure 7.Comparative Freeze Frame FFT, High Voltage & Low Voltage Welds

6

J. R. McCardle 1997

REAL — TIME SPECTRWM ANALY=R.0=e+P

Filais w.laIntome.0 FF-7 1_4noths 1024e40

Sam4lir,0 F-rociu.noW . .40.00 RHz Tr- So Tr-as...1

Figure 8.Power Spectra, High Voltage. (1024 point FFT, 25 frame averaging.)

Conclusions

Conventional DSP methods utilising the TMS320C30 are capable of providing real timesignals, in the frequency domain, which yield distinguishable features associated withnon-optimum weld voltage settings for SAW.

It is feasible that a Kohonen Feature Map could be generated to indicate stable andunstable weld states from successive on-line power spectra.

Complex signal processing and classification of noise corrupted data often require thecombination of conventional digital and neural techniques.

The research and development of weld automation continues, with the application ofANN technology, toward the provision of a fully integrated welding system.

References

[1] Stroud,R.R, Harris, T.J, Taylor-Burge, K.L,.'Neural Networks in Automated WeldControl', TWI, Gateshead, UK, 1991.

[2] Chawla, K.S, Norrish, J, 'Real Time Quality Monitoring Using Analysis of the ArcSound', TWI, Cambridge, UK, 1992.

[3] BRITE EuRam project - BREU 0415 BE-4006-90.

[4] International Institute of Welding, 'The Physics of Welding', Pergamon Press, 1984.

[5] Vu, V, et al, 'Time Encoded Matrices as Input Data to Artificial Neural Networks forCondition Monitoring Applications', COMADEM '91, Southampton, UK, July 1991.

[6] O'Brien, J.C, et al, 'Can Neural Nets Work in Condition Monitoring ?', COMADEM'92, Senlis, France, July 1992.

7

J. R. McCardle 1997

[7] Tanaka, T, et al,' Trouble Forecasting by Multi-Neural Networks on ContinuousCasting Process of Steel Production', Proc 1st ICANN, Espoo, Finland, June 1991.

[8] Harris, T.J, 'Neural Networks and their Application to Diagnostics and Control',COMADEM '92, Senlis, France, July 1992.

[9] Hecht - Nielson, R, 'Neurocomputing', Addison - Wesley Publishing Company, 1991.

[10] Dayhoff, J, 'Neural network Architectures. An Introduction.', Van NostrandReinhold, 1990.

[11] Kohonen, T, 'The Neural Phonetic Typewriter', Computer 21(3) 11-22, 1988.

[12] Kallio,K, et al, 'Classification of Lung Sounds by Self Organising Feature Maps', 1stICANN, Espoo, Finland, 1991.

[13] Hyperception, 'Hypersignal Workstation', Advanced Signal Processor Environment,Revision 2.0, Feb 1991, Software and User Manual.

8