| Welding sound signal contains abundant information about dynamic welding process and welding quality.The study of the welding sound signal was originally designed to imitate the auditory perception of experienced welders.But in the actual welding process,welders use two ears to get the sound,for welders,this can not only be used as a rough basis to estimate the sound source location,but multi-dimensional welding sound information can also provide a much more adequate basis to help welders estimate welding quality and adjust welding parameters.At the same time,there are always more than one sound source and noise source in the actual welding environment,it means that in the past research the sound signal collected by the single microphone was a mixed signal,including welding arc sound,splash sound,human voice,operation sound of machine,background noise and so on.Actually speaking,it is not that accurate to treat a mixed sound signal as the welding arc sound signal to analyze the dynamic welding process.In order to verify the multi-dimensional welding sound signal contains much more dynamic information and quality information about welding,the research of low-carbon steel pulsed GMAW and aluminum alloy pulsed GTAW dynamic process monitoring based on welding sound signal was designed in this paper.Firstly,this paper established a linear microphone array acquisition system composed of three MP201 1/2 inch pre-polarized free field microphones to collect the sound signals during welding process,then the sound signals were analyzed from four aspects: sound mechanism,auditory perception,time-domain waveform and frequency-domain energy distribution.At the same time,the feasibility of ICA algorithm applied to welding process was analyzed and verified.Then the splash signal was successfully separated from the three observed signals in pulsed GMAW process based on FastICA algorithm,it appeared to be an intense oscillatory waveform from the aspect of time-domain and its frequency-domain energy distribution mainly concentrated in the high frequency band of 6000-8000 Hz.Actually,the arc sound signal was not separated out due to its less obvious non-Gaussian characteristic compared with the splash sound signal.For pulsed GTAW process,noises were effectively removed from the welding sound signals through FastICA blind signal separated algorithm.In order to realize welding dynamic process monitoring,the characteristics of welding sound signals were analyzed and extracted firstly.It did not occur splash during the process of Aluminum alloy pulsed GTAW,and the change of welding states could be estimated from the change of the loudness of sound,the analysis of sound signals was relatively simpler than low carbon steel pulsed GMAW process,so the paper mainly dealt with the characteristic analysis of pulsed GMAW process.According to the change of auditory perception in pulsed GMAW process,the analysis of characteristics was determined as two aspects: tone variation of arc and intensity change of splash.Three typical states of sound signal were extracted,from the analysis of time-domain waveform and frequency-domain energy distribution,it was found that the energy of the 500-1000 Hz band could well represent the tone variation of arc in burn-through state.At the same time,the short-time analysis of the splash sound signal obtained by FastICA algorithm showed that the short-time energy changed very obviously when defects occurred.Therefore,it was determined as a sign of the change in the intensity of the splash signal and was associated with the burn-through defects.Meanwhile,the change of splash intensity was only one aspect affected by the occurrence of defect and did not reflect its essential characteristics,so it was only treated as an auxiliary estimation factor under this paper's experimental condition.For pulsed GMAW process,when selecting the feature inputs of the defect recognition model,it was found that the ratio of the low frequency energy of the observed signal to the splash signal energy presented a high degree of sensitivity and anti-interference to the defects,which could be used as a robust monitoring feature.Thus,the 500-1000 Hz band's short-time energy of the observed signal,the short-time energy of the splash signal and the ratio of the two parameters were used to determine the burn-through defect recognition model.For pulsed GTAW process,it was found that the short-time energy and standard deviation of the first and second separated signals could characterize the dynamic change of welding quality well and could be used to establish the defect recognition model.For pulsed GMAW process,the classification accuracy of the logistic regression model could reach 99%,but the misclassification rate of the burn-through defects was as high as 66.7%,this might be caused by the serious imbalance of the positive and negative samples.In order to solve this problem,the training samples were divided into 10 categories,each of which was used to establish a model together with the negative samples,then 10 models were weighted based on the classification accuracy rate and the defects misclassification rate to finally establish an optimization model.For pulsed GTAW process,the classification accuracy of the logistic regression model was 97.5% and the misclassification rate of the defects was 11.1%.For pulsed GMAW process,a BP neural network model of a 3-8-1 structure was also used to establish the recognition model.The classification accuracy and defects misclassification rate were 99.5% and 33.3% respectively.After adopting the AdaBoost optimization algorithm,the average error rate of the weak classifiers was 0.9%,the error rate of the strong classifier was 0.5%,which showed the AdaBoost optimization algorithm improve the classification accuracy to a certain extent.In order to solve the problem of high defects misclassification rate caused by the imbalance of positive and negative samples,a BP neural network optimization model was established through resampling of defect samples,and its classification accuracy was almost 100%.For pulsed GTAW process,a BP neural network model of a 3-8-1 structure was also established,and its classification accuracy and defects misclassification rate were 95% and 18.2% respectively.Therefore,this research proved that the multi-channel welding sound signals could realize the dynamic process monitoring of welding. |
PDF Full Text Request