A novel adversarial domain adaptation transfer learning method for tool wear state prediction

Abstract

Tool wear prediction is of critical importance to maintain the desired part quality and improve productivity in machining. The traditional tool wear prediction based on deep learning mostly considers the same type of tools under the same working conditions. It assumes that the collected data obey the same distribution and that the training data labels are sufficient, which has significant limitations in practical machining applications. In this paper, a novel adversarial domain adaptation transfer learning was proposed to predict the tool wear state of end milling tools under different working conditions, including the laboratory and actual industrial machining conditions. Firstly, the dual-path deep residual shrinkage network was used to extract the tool wear multiscale sensitive features from the spindle vibration signals. Then, a balance parameter was added to the traditional adversarial domain adaptation model, which can dynamically and quantitatively evaluate the relative importance of marginal and conditional distribution. Thus, the alignment of the source and target tool feature space was realized by dynamically learning domain invariant representations. Finally, the proposed method was verified on an 8 mm and 2 mm tool wear states prediction. Compared with different transfer learning methods, the superiority of the proposed dynamic adversarial domain adaptation method was proved.