An innovative Multisource Lightweight Adaptive Replayed Online Deep Transfer Learning algorithm for tool wear monitoring

Accurately monitoring tool wear during the cutting process is crucial for ensuring the precision manufacturing of components and enhancing machining efficiency. However, even under the same cutting conditions, there are differences in the feature distributions between offline data and online data, leading to poor monitoring effectiveness of traditional offline transfer learning algorithms. To effectively address this challenge, this paper proposes an innovative algorithm named MS-LARODTL (Multisource Lightweight Adaptive Replayed Online Deep Transfer Learning). In the MS-LARODTL algorithm, a lightweight convolution layer design is initially adopted, resulting in a reduction of 86.64 % in convolution time to enhance the algorithm's execution speed. Subsequently, online data features are replayed to adapt batch normalization layer parameters through a replay buffer, thereby improving the monitoring accuracy of the algorithm. Lastly, by replaying source data and online data features and leveraging Maximum Mean Discrepancy to align data features, the algorithm's monitoring accuracy is further enhanced. Experimental results distinctly demonstrate that the MS-LARODTL algorithm surpasses other multi-source transfer learning methods across all evaluation metrics on the self-constructed CNC milling dataset. These findings underscore the effectiveness and potential of the MS-LARODTL algorithm, emphasizing its importance in the design and application of tool wear monitoring algorithms.