Intelligent monitoring of porosity in laser melting deposition based on deep transfer learning

Abstract

Porosity defects represent one of the greatest challenges to the stability and reliability of parts produced by Laser Melting Deposition (LMD). Integrating sensor data with machine learning algorithms for porosity monitoring has garnered considerable attention. However, current machine learning approaches heavily rely on abundant labeled data and the assumption of independent and identically distributed (i.i.d.) data. In LMD processes, variations in process conditions could potentially affect the distribution and correlation of process data. Collecting sufficient labeled data for new processes from scratch is time-consuming and costly, and conducting experiments that cover all possible conditions to build a comprehensive dataset is highly challenging. Inspired by transfer learning, we propose a deep subdomain adaptation method based on ShuffleNetV2 (DSAS) to monitor porosity for the LMD process. First, based on the dynamic evolution of the melt pool and the physical mechanism of pore formation, local defect samples with corresponding structures were constructed, along with a source-domain porosity monitoring network. Second, for the more complex and variable conditions in the target domain, porosity monitoring was achieved using the proposed DSAS. Specifically, samples collected under the source and target domain were projected into different subdomains based on porosity levels. Domain alignment was achieved by calculating the Local Maximum Mean Discrepancy (LMMD) distance between domains. DSAS not only learns cross-domain features to adapt the model to diverse process conditions during transfer, but also alleviates the challenge of similar boundary samples in porosity monitoring by aligning features within subdomains. Experimental results demonstrate that the proposed method effectively leverages historical knowledge from relevant experimental data to monitor porosity under complex and changeable working conditions. This significantly improves model development efficiency and reduces costs in data-scarce scenarios, offering substantial value for the implementation and promotion of quality monitoring in LMD processes.