Tool remaining useful life prediction considering wear state based on hybrid attention network

Abstract

Accurate prediction of the remaining useful life for the cutting tool is a key part of the predictive maintenance of computer numerical control machines. However, the wide variety of tools makes the process of modeling different tool wear regularities redundant and cumbersome. In addition, it is difficult to deal with the input characteristics of multi-sensor monitoring signals in a targeted manner. To solve the above problems, a hybrid predictive model with squeeze-and-excitation (SE) module is proposed. Combined with adaptive feature extraction based on convolutional neural network and observation based on bidirectional gated recurrent unit, accurate multivariate regression prediction is achieved. The SE module enhances the focus on crucial features. Finally, through the design of the tool wear experiment and the combination of the public dataset, the accuracy and generalization ability of the proposed model are verified under different tool types and different working conditions.