A novel cross-domain tool breakage monitoring method based on locality preserving joint transfer with intra-class compactness

Tool breakage monitoring (TBM) during milling is vital for improving production efficiency and ensuring product quality. The tool breakage monitoring model under the traditional machine learning framework needs to be retrained due to the lack of sufficient generalization ability in the face of new machining scenarios. Nevertheless, it is hard to obtain sufficient broken samples directly from industrial scenarios to support the monitoring model retraining. This paper proposes a novel transfer learning strategy termed locality preserving joint transfer with intra-class compactness (ICC-LPJT) to address the problem that broken samples under variable machining conditions are challenging to obtain. The key approach is to train the transfer learning model to generate a mapping matrix that not only preserves the prior distribution structure properties of the training data, but also minimizes the distribution differences between different datasets. More importantly, it is able to maximize intra-class sample compactness, thereby alleviating the inter -class edge sample overlap problem caused by complex milling process. Meanwhile, improved multiscale symbolic dynamic entropy (IMSDE) is developed to extract more distinguishable tool condition sensitive features. Eventually, a novel cross -domain tool breakage monitoring method based on ICC-LPJT and IMSDE is developed. The presented method is able to construct a cross -domain monitoring model suitable for diverse machining scenarios without any broken samples included in the target domain training data. Consequently, the proposed method is more suitable for tool breakage monitoring in industrial environments. Experimental results illustrate that the accuracy of the presented method is mainly above 95 % in the cross -domain monitoring of various spindle speeds.