Data-Driven Approaches for Machining Efficiency Through Generalizable Tool Wear Prediction and Power Consumption Analysis

Abstract

Industry 4.0 and smart manufacturing have generated large amounts of manufacturing data, creating new opportunities to use data-driven methods to improve machining efficiency. This dissertation examines machining efficiency through two connected aspects: machine condition monitoring and power consumption. Machine condition monitoring, particularly tool wear monitoring, affects efficiency by influencing tool life, surface quality, downtime, and machining cost. Power consumption is also important because it is an indicator of machining efficiency and contributes to the energy required under different machining conditions. Although machine learning-based tool wear prediction has received increasing attention, many existing models are developed under limited experimental conditions and cannot generalize well to new conditions. In addition, many data-driven machining studies on power consumption focus primarily on prediction accuracy while offering limited explanation of how underlying mechanisms, such as tool health, process stability, and thermo-mechanical loading, influence power consumption. This dissertation addresses these gaps through three connected contributions. First, a systematic literature review examines recent developments in machine learning-based tool wear prediction, including model types, input features, target variables, datasets, generalizability levels, and the use of workpiece- and tool-related information. The review identifies limited generalizability as a major gap, primarily due to heavy reliance on sensor-based features, limited attention to workpiece material properties and tool characteristics, and a lack of comprehensive datasets that cover diverse workpiece materials, tool types, coolant conditions, and cutting parameters. Second, building on this gap, a data-driven framework using few-shot learning is developed to improve the generalizability of flank wear (VB) prediction across seven workpiece materials. A Gaussian process regression (GPR) is compared with artificial neural networks, random forests, and k-nearest neighbors using a leave-one-workpiece-out evaluation with limited fine-tuning data. The GPR model outperformed the benchmark models, achieving R² ≥ 0.90 for four new materials and 0.81 for another, while two materials showed lower accuracy with R² values of 0.68 and 0.60. SHAP analysis identified ultimate tensile strength, hardness, cutting speed, and feed rate as the most important predictors. Third, partial least squares structural equation modeling (PLS-SEM) is applied to explain how thermo-mechanical loading, process stability, and tool health contribute to power consumption in milling. This contribution moves beyond simple prediction by modeling the direct and indirect relationships among these machining mechanisms. The results show that thermo-mechanical loading has the strongest overall effect on power consumption, both directly and through its impact on process stability. Tool health has a significant effect on power consumption indirectly through process stability, rather than through a direct effect. This dissertation contributes to machining efficiency research by connecting three levels of analysis: reviewing the current state of tool wear prediction and identifying major gaps, developing a more generalizable flank wear prediction model that addresses key gaps identified in the systematic literature review, and explaining how tool health and other underlying machining mechanisms affect power consumption as an indicator of efficiency. The findings show that incorporating workpiece material properties, tool characteristics, and cutting conditions can improve the generalizability of tool wear prediction, while PLS-SEM analysis reveals how tool health and other underlying mechanisms affect power consumption. These outcomes provide practical guidance for developing more generalizable tool wear prediction models and improving machining efficiency.