Tool wear condition monitoring across machining processes based on feature transfer by deep adversarial domain confusion network

Deep learning-based data-driven methods have been successfully developed in tool wear condition monitoring (TWCM), relying on the massive available labeled samples and the same probability distribution between training and testing data. However, these two prerequisites are often difficult to satisfy in actual industries, which results in significant performance deterioration of those methods. This paper proposes an intelligent cross-domain data-driven TWCM method based on feature transfer by a deep adversarial domain confusion network (DADCN) model. In this model, source and target feature extractors sharing the same network architecture are employed to obtain high-level representation from time-frequency spectrums of vibration signals in the different domains respectively. An independent adversarial learning mechanism is designed in domain obfuscator to learn domain-invariant feature knowledge, while the maximum mean discrepancy is applied to measure the distribution difference between different domains. A cross-domain classifier is utilized for tool wear condition monitoring across machining processes. The performances of the proposed DADCN model under two distribution measure criteria are experimentally demonstrated using six transfer tasks between laboratory and factory platforms. The results indicate that the DADCN model can improve the monitoring accuracy and exhibit distinct clustering of tool wear conditions, promoting a successful application of data-driven methods in actual industrial fields.