The tool wear plays a crucial role in determining the quality of the workpiece. Excessive tool wear results in a decrease in machining accuracy and speed and a decline in yield. At the same time, frequent tool changes increases costs and affects the processing speed as well. Therefore, it becomes essential to accurately determine the tool wear status to plan a reasonable tool change time and even further optimize the tool design. Data-driven wear prediction based on multi-sensor signal processing proves to be a feasible solution to this problem. However, the model trained in one working condition is usually not able to fit another condition well. Sometimes the tool wear characteristic may be a different even for the same kind of tool under the same working condition, limiting the practical application of this method in the industry. Furthermore, the acquisition of the tool wear ground truth generally relies on offline detection with the microscope, which leads to low efficiency of the data labeling.

In response to the above problems, we conducted experiments on the wear prediction problem of face milling cutters based on the framework of tribological informatics. In this study, force/torque sensors, vibration sensors, and acoustic emission sensors were selected as the input sources for the training and predictions of tool wear models, recording the signals generated during the processing of the workpiece. Specifically designed filters firstly filtered the original signal to reduce the interference of non-cutting factors. A series of statistics from time and frequency domain analyses, e.g. the mean, variance, and extreme values, were then extracted as the signal features. To obtain the ground truth of the tool wear, a CMOS camera was installed on the machining platform capturing pictures of the wear position of the cutter. We proposed a machine vision-based tool wear recognition method. With the prior knowledge of the wear morphology of the face milling cutter, the method uses image preprocessing, Canny edge extraction, and region recognition to achieve in-situ tool wear acquisition and avoid frequent disassembly of the cutter. With features from multi-sensor signals and tool wear ground truth, a tool wear prediction model was then built and trained through the random forest regression method, verifying the feasibility of the data-driven wear prediction based on the multi-sensor signal.

Further, to improve the prediction accuracy under different working conditions, we proposed a new wear prediction method based on feature transfer. We determine the one-to-one linear transfer equation between each feature in the history data and feature of the preliminary data of the current tool, by minimizing the maximum mean discrepancy (MMD). The MMD is also used as the evaluation function to judge the feature transfer quality and select well- transferred features. A new prediction model can then be trained with the well- transferred features from the history data and predict tool wear status for the remaining tool lifetime. Experiments results showed that the wear prediction model trained by data from one of the conditions cannot provide good predictions under the other two working conditions. In contrast, the model trained by transferred features can significantly improve the prediction accuracy. Specifically, the coefficient of determination (R-squared) increased from 0.37 to 0.96 by feature transfer, proving the effectiveness of the tool wear prediction method based on transfer learning proposed in this paper.