An end-to-end framework based on acoustic emission for welding penetration prediction

The adverse effects of inadequate welding penetration on the service performance of welded assemblies have raised significant concerns. In this study, we explore the benefits of acoustic emission (AE) sensors for monitoring the laser-arc hybrid weld penetration state during welding processes. The underlying mechanisms of AE signal generation under different penetration states are elaborated, and the feasibility of using time-frequency domain features for monitoring penetration state is validated. Building upon these findings, we propose an end-to-end framework named WAENet. This framework incorporates an automated optimization module for signal processing and time-frequency domain feature extraction, as well as an improved modular convolutional neural network (CNN) for recognition. To enhance the CNN's performance, techniques such as grouped convolution, depth-wise separable convolution, and global average pooling are employed. Furthermore, the training process of the CNN also considers the involvement of hyperparameters related to signal processing and feature extraction, as proposed in this article. WAENet achieves an average accuracy of 99.62 % in identifying the penetration states, which outperforms other models that use alternative feature extraction or classification techniques for comparison. These studies expand the application scope of AE and CNN in the field of intelligent welding.