SUMMARY

A practical, accurate, and generalizable system for monitoring tool condition during a machining process would enable advancements in manufacturing process automation, cost reduction, and efficiency improvement. Previously proposed systems using various individual machine learning (ML) models and other analysis techniques have shown low generalizability to new machining and environmental conditions, as well as a common reliance on expensive or intrusive sensory equipment which hinders their industry adoption. While ensemble ML techniques show significant advantages over individual models in terms of performance, overfitting reduction, and generalizability improvement, they have only begun to see limited applications within the field of tool condition monitoring (TCM). To address the research gaps which currently surround TCM system generalizability and optimal ensemble model configuration for this application, nine ML model types, including five heterogeneous and homogeneous ensemble models, are employed for tool wear classification. Sound, spindle power, and axial load signals are utilized through the sensor fusion of practical external and internal machine sensors. This original experimental process data is collected through tool wear experiments using a variety of machining conditions. Four feature selection methods and multiple tool wear classification resolution values are compared for this application, and the performance of the ML models is compared across metrics including k-fold cross validation and leave-one-group-out cross validation. The generalizability of the models to data from unseen experiments and machining conditions is evaluated, and a method of improving the generalizability levels using noisy training data is examined. T-tests are used to measure the significance of model performance differences. The extra-trees ensemble ML method, which had never before been applied to signal-based TCM, showed the best performance of the nine models.