Tool wear prediction based on hybrid feature selection

Abstract

Tool wear monitoring can provide information and improve productivity for tool change decisions in Computer Numerical Control (CNC) machine manufacturing. The processing of indirect signals and extracting sensitive features related to tool wear are critical. In this paper, a hybrid feature selection method is proposed and verified by improved Long Short-Term Memory (LSTM) models. Firstly, data preprocessing is performed on the acquired vibration, force, and acoustic emission (AE) signals and relevant features are extracted from multi-domains and preliminary filtered using a mixture of the SRCC and MI methods. Then, feature dimensionality is further reduced by combining dimensionality reduction methods. After the sensitive features are enhanced by the self-attention mechanism, temporal dependencies are extracted by the LSTM layer. Finally, tool wear is predicted using the fully connected layers. The experiment shows that the combined feature extraction strategy can select sensitive features more effectively and improve the prediction performance.