In-situ monitoring in laser powder bed fusion based on acoustic signal time-frequency synchrosqueezing transform and multi-scale spatially interactive fusion convolutional neural network

Variability and repeatability challenges in Laser powder bed fusion (LPBF) limit its industrial applications. The difficulty of effectively evaluating similar defects with varying severity poses reliability issues for in-situ monitoring based on acoustic monitoring. In the LPBF acoustic monitoring scenario, the original signal amplitude is weak, and the acoustic field coupling information is complex, so it is difficult to directly obtain high-quality and physically reasonable LPBF defect features, which leads to the difficulty of reliably monitoring LPBF defects. This study proposes a novel LPBF pore defect monitoring approach based on acoustic signal physical defect information feature extraction and interpretable deep learning. Firstly, based on the established correlation mechanism among the molten pool, defects, and signals in the LPBF process, the effectiveness of acoustic feature representation in reflecting defect information is demonstrated. Secondly, the TFSST method is proposed to enhance the expression of the most relevant and physically reasonable time-frequency features of weak amplitude LPBF defect signals. Finally, the multi-scale spatially interactive fusion convolutional neural network (MSSIF-CNN) is proposed for in-situ monitoring of LPBF defects using an ABAE monitoring system. The proposed method can recognize five LPBF defects with an accuracy of 99.12 %. In addition, the visualization results verify that the model possesses superior adaptive defect feature extraction capabilities and effectively focuses on the most relevant and physically reasonable defect characterization information. With millisecond monitoring response times and physically interpretable defect monitoring decision-making capabilities, the method offers the potential for industrialized in-situ monitoring.