An Intelligent Multiscale Spatiotemporal Fusion Network Model for TCM

In the process of milling, accurate and reliable monitoring of tool condition monitoring (TCM) is essential to ensure machining quality and efficiency. Among the current data-driven approaches, machine-learning-based supervised methods suffer from cumbersome feature extraction, while deep learning is gradually gaining popularity in TCM due to its adaptive feature extraction capability. However, the traditional deep-learning prediction methods also have the problems of low efficiency and poor interpretability during TCM. Therefore, a novel multiscale spatial–temporal fusion network (MSFNet) model is proposed in this article. By designing multiscale residuals and parallel spatial–temporal feature extraction modules, the efficient shallow feature extraction and noninterference spatial–temporal feature fusion are realized, respectively. In addition, based on expert knowledge, this article adopts a filtering preprocessing algorithm and a smoothing postprocessing algorithm to address the impact of abnormal data and sudden changes of predicted value on the prediction performance of the model, which enhance the robustness of the prediction. After experimental verification, the results show that compared with other deep-learning methods, the proposed MSFNet achieves lower root mean square error (RMSE) and mean absolute error (MAE), which significantly improves the accuracy of tool wear prediction.