A Low-Cost Laser Welding Monitoring Framework Based on Depth-Wise Separable Convolution with Photoelectric Signals

Abstract

In recent years, the process monitoring based on optical radiation detection widely applied in laser welding monitoring process, such as visual cameras, spectrometers and photoelectric sensors. This study proposes a low-cost monitoring model based on a CNN module with the combination of convolution and depth-wise separable convolution (DSC) applying the industrial photoelectric sensors. This model aims to generate more effective features from the primitive signals captured by the visible light photoelectric sensor and the reflective laser photoelectric sensor, without pre-processing in advance. The DSC is applied to generate features to reveal the inherent features of welding statuses, and especially reduce the computing costs during monitoring process. The proposed model in this study acquired high accuracy with low space complexity and time complexity compared with the traditional model. The model also performs well under the limited and unbalanced welding data, indicating its good robustness. This study provides a low-cost method for real-time monitoring of laser welding process.