A tool wear condition monitoring method for non-specific sensing signals

•The proposed method has the generalization capability between different sensing channels.•A comprehensive evaluation method of model performance is proposed.•The performance difference between different sensing signals is reduced.•The limitation of the single-channel signal is overcome.•The effectiveness of the method is verified by two experimental cases. Real-time and accurate monitoring of tool wear conditions is crucial to achieving double optimization of production cost and product quality. However, the differences in the characteristics of different signals limit the ability of the monitoring model to generalize between sensing channels, which becomes an important factor limiting the promotion of the model. To solve this problem, an improved parallel residual network based on single-channel and non-specific sensing signals is proposed in this paper. The limitation of the single-channel signal with little information and poor anti-interference ability is overcome by adaptively extracting the multi-scale spatial features of the sensing signal. Hybrid dilated convolution is introduced to expand the receptive field, and then the long historical domain information is obtained. At the same time, the information dependence between layers is enhanced by introducing skip connections. These two designs ensure the perceptual generalization ability of the model. Considering the tool replacement time and the imbalance classification of labels, a comprehensive evaluation method is proposed for model performance evaluation. In addition, the variation law of tool wear in the milling process of Ti-6Al-4V thin-walled parts is investigated. Finally, the validity and transferability of the model are verified by two milling datasets with different cutting conditions. On the basis of ensuring the perceptual generalization ability of the model, the differences in model performance based on acceleration and cutting force signals are controlled within 4.5 % and 1 %, respectively, and the overall average recognition performance is 96.3 % and 92.5 %, respectively. This study provides a feasible solution for intelligent tool replacement in the actual machining environment. [Display omitted]