Intelligent tool wear monitoring based on multi-channel hybrid information and deep transfer learning

Effective tool wear monitoring (TWM) is essential for maintaining high quality and efficiency of cutting operations, preventing defective parts, and minimising economic losses. However, current research on TWM is mostly concentrated on specific working conditions, which limits its application. This study proposes a variable-condition TWM method based on multi-channel hybrid information and deep transfer learning. Multi-channel hybrid information is formed by combining multidimensional cutting force data and multidimensional process information. Based on this, a residual network (ResNet) is trained to obtain a TWM model that could predict the tool wear under multiple working conditions. Once variable working conditions arise (e.g., a change of tool or workpiece), the established ResNet model can be fine-tuned using a small amount of multi-channel hybrid information. These models were validated using datasets from a local experiment and NASA. The prediction results of the local dataset show that the model using hybrid information has a maximum tool wear prediction error of 9 µm and 2.7 µm under multiple and variable working conditions, respectively, which is much smaller than the prediction error of the model using only cutting force data. Moreover, fine-tuning the multi-channel hybrid information yields better prediction performance than using only sensor data for the TWM model. These results were demonstrated using the NASA dataset.