Chatter detection in milling process with feature selection based on sub-band attention convolutional neural network

Abstract

For fault detection using deep learning methods, feature extraction and subsequent selection are important tasks to achieve higher classification accuracy. In this study, a new methodology is proposed for extracting frequency band features for chatter detection during cutting and selecting optimal bands as the input for deep learning classifier. The vibration response during cutting process was measured on the main mechanical parts: the cutter and spindle head. Feature extraction using time-varying variance analysis was performed to quantify the fluctuation of the vibration. To improve the robustness under various operating conditions, variance is calculated for each frequency band. The sub-band attention CNN (SBA-CNN), which combines sub-band CNN (SB-CNN) and the attention layer, was proposed. The attention layer was used to evaluate the importance of each frequency band of time varying variance. The attention weights of the frequency bands were obtained, and the high-weight bands were selected as the optimal input feature of SB-CNN training for chatter detection. The performance of the proposed method was compared with existing feature selection methods and machine learning classifier models.