Pyramid LSTM auto-encoder for tool wear monitoring

Tool wear monitoring is necessary to guarantee the product quality in high-speed machining. In this paper, a pyramid long short-term memory network (LSTM) based on spectral features, is proposed to extract tool wear features from multi-source monitoring signals. In order to improve the robustness of the model under varied working conditions, the unlabeled signal is trained by an auto-encoder structure to reduce the information loss. With the LSTM architecture, the model can better reflect the multi-period characteristics of long-term signals and is more suitable for the actual industrial production. The experimental validation of high-speed milling verifies the effectiveness of the proposed method.