A data-driven time-sequence feature-based composite network of time-distributed CNN-LSTM for detecting pore defects in laser penetration welding

Abstract

The pore in laser penetration welding significantly deteriorates the mechanical property, and is an important criterion for evaluating the product quality. The intelligent diagnosis of welding can guide the optimization of process parameters to inhibit the pore formation. Considering that the signals in laser welding have time-sequence features and abundant implicitness information may cause high computational effort and information misidentify, an intelligent pore defect diagnosis method based on time–frequency feature extraction and a combined neural network of Convolutional Neural Networks (CNN) and Long short-term memory (LSTM) was proposed. Firstly, the visual signal results of vapor plume demonstrated that the pore formation was accompanied by irregular and continuous variation in vapor plume morphology during the subsequent period. Secondly, this denoising, decomposition, and restructuring of signals were performed by wavelet packet transform, and it was found that the sustaining fluctuation of frequency could localize the pore formation in the corresponding position of weld metal. Therefore, the signal was finely segmented to construct a cube time–frequency spectrogram data with the time-sequence characteristics. Finally, a combined classification model of CNN and LSTM was constructed for recognizing the temporal-spatial information of cube spectrogram data, realizing the online monitoring of pore defect. The results indicated that the proposed method was a promising solution for monitoring pore defect in laser penetration welding and improving product quality.