Abstract
As a complex thermo-physical process, the plasma arc welding (PAW) is easy to be unstable due to external interferences. Weld quality monitoring is important for intelligent robot PAW welding. Due to the different instability mechanisms, it is difficult to obtain high adaptivity and accuracy with features extracted in a single time window. In this paper, a novel feature extraction method based on sliding multiscale windows is proposed to improve model accuracy and calculation speed. A group of windows with different time widths are established to extract multiscale information. Windows slide throughout welding process and are synchronized on the timeline for feature correlation. The welding current and arc voltage are processed to extract features inside windows, including signal denoising by discrete wavelet transform (DWT) and dimension reduction by primary components analysis (PCA). Based on the feature vectors extracted from multiscale-windows, support vector machine (SVM) with radial basis function (RBF) kernel is used. The best window width is determined automatically by model training. The proposed method is used to predict weld quality for PAW in the field of shipbuilding. The results show that the model with multiscale feature extraction is helpful to improve prediction precision and recall ratio.
Highlights
Due to the complexity of welding process, it is difficult to build an intelligent stability monitoring system only with statistical or mathematical modeling methods
Statistical features were calculated in wavelet packet coefficients, and the features were used to build a convolutional neural network (CNN) model
A feature extraction method based on multiscale windows was proposed to improve the accuracy and response speed of evaluation model
Summary
Due to the complexity of welding process, it is difficult to build an intelligent stability monitoring system only with statistical or mathematical modeling methods. Shevchik et al (2020) proposed a method for real-time detection of laser welding process instabilities based on a deep artificial neural network. They used wavelet packet transform to extract features from the laser back reflected signal and acoustic emission signals. You et al (2015) used wavelet packet decomposition (WPD) method to extract features from optic signals in laser welding They exerted primary component analysis (PCA) method to perform further refining for the features. Statistical features were calculated in wavelet packet coefficients, and the features were used to build a convolutional neural network (CNN) model Their model detected blowout, undercut, and humping defects with an accuracy of more than 93%. In the studies, advanced signal analysis methods were used to extract feature information in welding signals without human involvement, and these methods performed well in many monitoring objects
Sign-in/Register to view highlights & summary 
Submitted Version (
Free)
Published Version
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call