A physics-informed learning approach for milling stability analysis with deep subdomain adaptation network

Efficient and accurate chatter prediction is critical for selection of chatter-free process parameters to improve machining productivity and surface quality of the workpiece. However, challenges arise due to the uncertainty and inaccuracy in model parameters, which may lead to significant differences between predicted and measured stability boundaries. This study introduces a novel physics-informed learning approach for efficiently determining stability in milling based on the deep subdomain adaptation network. First, a deep subdomain adaptation network (DSAN) is developed to extract essential features that characterize the relationship between operating conditions and chatter stability, using both physical models and testing data. Subsequently, we introduce the Local Maximum Mean Discrepancy (LMMD) metric that quantifies the discrepancies in mathematical distribution between features from physical models and testing data, which are generally present and tend to be significant under conditions such as high spindle speeds, heavy cutting forces, or when flexible workpieces are involved. Following this, the LMMD loss is defined and incorporated into the established feedforward network. This loss is calculated and minimized iteratively during the network's training via backpropagation. We demonstrate that considering those discrepancies in the proposed hybrid-driven modeling enhances prediction accuracies without compromising efficiency. The proposed approach is experimentally validated by extensive milling tests and exhibits greater industrial applicability compared to previous methods.