A Novel Remaining Useful Life Prediction Method for Milling Tool Based on Transfer Hierarchical Vision Transformer

Remaining useful life (RUL) prediction of milling tools based on deep learning can effectively decrease the negative impact of late tool replacement. However, variations in monitoring data for different milling tools with the same type will restrict the generalizability of the methodology to engineering applications. Aiming at the issues mentioned above, an RUL prediction method for milling tools based on a transfer hierarchical vision transformer (THVIT) is proposed in this study. An innovative domain adaptation (DA) method based on the source domain sub-feature benchmark is utilized to decrease the distribution discrepancy between the depth features extracted by the hierarchical vision transformer (HVIT). Validated by milling tool monitoring data collected from a 5-axis machining center, the mean absolute error of THVIT is reduced by 14.21% in comparison to HVIT. Compared with the classical DA approach represented by the maximum mean discrepancy, the root mean square error of THVIT is decreased by 12.82%. The experimental results demonstrate that the proposed method can effectively guarantee the accuracy of RUL prediction.