Research On Feature Extraction Of Welding Defects For Aluminum Alloy During Pulsed GTAW Process Using Multisensor-based Information Fusion

Intelligent welding is one of the most significant research topics in intelligent manufacturing area. Thus, sensing technology and sensor information processing are the key factors to achieve the intelligent and automatic welding process. Recently, miniaturization and non-contact sensors, such as arc sensor, vision sensor, sound sensor, arc spectrum sensor and so on, are increasingly applied into the real-time monitoring and controlling of welding process and quality. On the one hand, a great amount of information related to welding quality has been acquired from different kinds of sensors; on the other hand, “Big Data” of welding process has inevitably been generated. Therefore, the key issue is how to remove the massive noise and redundant information, more importantly, how to extract the effective information that can be promptly fed back and utilized in real-time monitoring of welding quality.In this paper, feature extraction, evaluation and selection, prediction and identification for welding seam defects have been carefully researched by means of multisensory fusion of arc spectrum, arc sound, arc voltage and welding image. During the research Al alloy pulsed Gas Tungsten Arc Welding(GTAW) has been chosen as the research object and real-time defect detection in dynamic welding process as the research goal.An experimental system for pulsed GTAW has been set up, which can automatically control the welding process. A multiple-signal acquisition system has been built as well in order to acquire and save the information from welding arc spectrum, arc sound, arc voltage and seam images. By means of multi-source information, the production mechanisms of typical welding defects, including surface oxidation, under penetration, over penetration, burning through and surface porosity, has been carefully analyzed, together with the singularity of different signals in time, frequency and time-frequency domain. Methodologies of signal processing and feature extraction for each kind of signal have been proposed.For spectrum signal, a methodology for feature extraction and evaluation based on spectrum band of interests(SOI) has been proposed. Eight SOIs were selected based on proposed maximum singularity rules respectively, from which three statistic features, e.g., root mean square, deviation and kurtosis have been extracted to characterize the average intensity, singularity degree and curve shape of SOI. Then, the pulse interference in the spectrum feature curve has been effectively eliminated using Coief4 as the wavelet function and 5 level decomposing. Furthermore, the sensitivity of features to welding defects has been quantified based on the proposed evaluation criteria, e.g., SNR logarithm. Besides, based on selected emission lines of H I(656.28nm) and Ar I(641.63nm), multiple features, such as the ratio of peak area, ratio of peak intensity and ratio of variance and so on, have been extracted and their sensitivities to porosity defect of welding seam have been quantitatively evaluated using Fisher criteria. At last, quick detection of porosity defect has been achieved based on STD threshold lines of spectrum features.For arc sound signal, the algorithms of feature extraction in time, frequency and time-frequency domain have been developed respectively. Firstly, an algorithm of feature extraction based on local sound signal of interests has been proposed, and the correlation between the feature and defect of under penetration and caving has been analyzed. Secondly, an algorithm of feature extraction from sound frequency segment attention based on sound signal power spectrum density has been proposed, based on different attention mechanism. The frequency band of Weltch PSD has been divided and the statistic features of sound signal are extracted. The correlation between welding defect and sound PSD frequency has been analyzed.An algorithm of feature extraction and evaluation for relative wavelet energy of sound signal has been proposed. Feature set of E(j) which characterizes the sound energy in different frequency bands has been calculated based on db3-3. Moreover, an evaluation criterion, namely, Maximum Standard Deviation between Class(MSDBC), has been proposed to evaluate the ability of features to separate multiple states of seam penetration condition, thus, effectively eliminating the redundant features. Furthermore, the sound signal was carefully analyzed from different aspect by means of wavelet package in time-frequency domain. It was found that the frequency band from 7.5k to 10 kHz has relatively high correlation with the status of welding penetration. At the end, based on the vision attention method, image feature parameters, such as ROI-1-CountRate3, ROI-2-CountRate3 and ROI-3-CountRatio have been extracted from regions of interest in the welding pool image. These parameters are capable of detecting the defects of under penetration, surface oxidation and burning through.In order to extract more effective information hidden inside the “Big Data” of welding process and select the optimum feature subset, a feature selector called Hybrid Improved fisher filter and SVM-CV wrapper(HIFSCW) has been proposed under the data driven technique. First, the improved Fisher criterion is used as the feature filter. Therein, the weight factor of feature parameter was adaptively updated based on its voting,which can protect some features with smaller Fisher value. Then, the proposed SVM-CV classification model was used as the wrapper, which is established and optimized by using 10-fold cross validation and grid searching of best parameters. Furthermore, four feature subset sections have been defined judging from the test accuracy curve, which are insufficient section, complementary section, optimum section and redundant section.A feature-level fusion model called SVM-CV was developed to predict and identify different welding penetration status. By means of the feature space selected by MSDBC criteria, the fusion model was greatly simplified. Besides, the voltage information has remedied the shortage of sound information feature to separate different welding defects, like under penetration and burning through. At last, the test accuracy of defect classification model has greatly improved from 74.19% calculated based on single-senor model to 94.31% of multi-sensor fusion model。Different degree of seam porosity defect has been artificially designed and its mechanism and singularity of signal feature have been researched. Based on the developed feature selection method, on the one hand, HIFSCW has selected the optimum feature subset; on the other hand, the SVM-CV classification model has predicted single and coupling welding defects with the accuracy of 94.72% using the optimum feature subset. Comparing with other models based on single and double sensors, multisensory fusion model detects welding defects, with better classification accuracy, stability and robustness.