Welding defects detection based on deep learning with multiple optical sensors during disk laser welding of thick plates

Abstract

A multi-sensor system, including an auxiliary illumination (AI) visual sensor system, an UVV band visual sensor system, a spectrometer, and two photodiodes, is established to capture signals of the welding status during high-power disk laser welding. The features of visible light and reflected laser light signal were extracted by decomposing the originally captured signals into different frequency bands by wavelet packet decomposition method (WPD). The captured signal of the spectrometer mainly covers the optical wavelength from 400 nm to 900 nm, which was divided into 25 sub-bands to extract the spectrum features of the spectrometer signal by statistical methods. The features of the plume are acquired by the UVV band visual sensor system, and the features of keyhole are extracted captured from the images captured by AI visual sensor system through the digital image processing method. Based on these quantified real-time features of the welding process, a deep learning algorithm based on convolutional neural network (CNN) was developed to detect three different welding defects during high-power disk laser welding. The established deep learning model is compared with the backpropagation neural network (BP) model, and it shows higher accuracy and robustness in detecting welding defects with the multi-sensor system. Its effectiveness was also validated by four welding experiments with different welding parameters.