Robust Tool Wear Prediction using Multi-Sensor Fusion and Time-Domain Features for the Milling Process using Instance-based Domain Adaptation

Tool wear prediction is a significant task in milling, offering several benefits including cost reduction, improved quality, and enhanced productivity. However, predicting a tool wear is challenging due to the inherent uncertainty of the milling process and the types of data that can be used for prediction. Further, limited availability of labeled training data in the target domain makes it challenging to train models precisely and reduces their predictive performance. Thus, present study tackles this issue with a novel TrAdaBoost Regressor (instance-based domain adaptation) approach with real-time machining data. TrAdaBoost leverages information from the labeled source domain to improve predictions in the target domain, effectively utilizing the available labeled data and unlabeled target data. The TrAdaBoost Regressor is the combination of adaptive boosting and instance-weighting for the source and target domain. Hence, it is implemented to optimize predictive performance and enhance generalizability of a model across varying machining parameters. Real-time machining data is acquired and processed through sequence of steps including feature extraction, scaling, and feature selection. The selected features are used for wear prediction with TrAdaBoost Regressor through various base estimators and their performance is evaluated using different evaluation metrics. Thus results shows that, TrAdaBoost Regressor with RFR gives the highest R2 score in the range of 0.989-0.999 during tool wear prediction for the features selected using SFS with RFR. Also, the proposed approach addresses the challenges of covariate shift and data scarcity in tool wear prediction and prove its adaptability during tool wear prediction for new unlabeled data.