An online cutting tool wear monitoring method based on the integration of DSCNN and pre-training/fine-tuning transfer learning

• In this paper, an integrated architecture combining depthwise separable convolutional neural networks (DSCNN) with the pre-training and fine-tuning (PTFT) transfer learning strategy to develop the online tool wear monitoring model (PTFT-DSCNN), in order to improve the generalization capability and the computational efficiency of online monitoring of cutting tool wear for variable operating conditions, and to reduce the dependency of the model training process on a large number of target domain datasets. • Experimental validation shows that compared to CNN, DSCNN, and DAN-DSCNN models, the PTFT-DSCNN model has more excellent performances in terms of computational accuracy and computational efficiency under variable operation conditions together with a small target domain sample size. The superior generalization capability and monitoring efficiency of the model developed in this work makes it better suited for online tool wear monitoring in complex machining scenarios. In order to improve the generalization capability and the computational efficiency of online monitoring of cutting tool wear for variable operating conditions, and to reduce the dependency of the monitoring model training process on a large number of target domain datasets, this study proposes an integrated architecture combining depthwise separable convolutional neural networks (DSCNN) with the pre-training and (PTFT) transfer learning strategy to develop the online tool wear monitoring model (PTFT-DSCNN). The DSCNNs with more efficient convolution approach are utilized as the backbone networks in the pre-training stage, which contributes to reducing the computing load and enhancing the computational efficiency of the model. In the pre-training stage, through multi-source domain adaptation of two historical condition datasets and 20% of target condition dataset, the feature distribution differences between historical and target conditions are minimized. Subsequently, the frozen network structure and layer weight parameters of the model are transferred from the pre-training stage to the target condition prediction stage, enabling a rational initialization of the target condition prediction model. After fine-tuning with target domain dataset, the computational accuracy of the monitoring model under target cutting conditions is further enhanced. Therefore, the generalization capability of the monitoring model under variable cutting conditions is enhanced. Experimental validation shows that compared to CNN, DSCNN, and DAN-DSCNN models, the PTFT-DSCNN model has more excellent performances in terms of computational accuracy and computational efficiency under variable cutting conditions together with a small target domain sample size. The superior generalization capability and monitoring efficiency of the model developed in this work makes it better suited for online tool wear monitoring in complex machining scenarios.