A Machine Learning (Gramian Angular Field + SVR-PSO) Model for In-Process Cutting Tool Wear Assessment Using Multisensor Heterogeneous Data Fusion

Abstract

Abstract. This research developed a tool wear assessment system using sensor data fusion and machine learning. A transition in the dominant tool wear mechanism from abrasive to adhesive or diffusion is induced by variations in cutting parameters, including cutting speed, feed rate, and depth of cut. Scanning electron microscopy (SEM)/energy-dispersive X-ray (EDX) analysis confirmed severe tool wear and workpiece material transfer at high speeds (120 m/min, 0.15 mm/rev), while low speeds (60 m/min, 0.05 mm/rev) promoted stable abrasion, highlighting the critical need for parameter optimization. The optimal machining parameters, including a cutting speed of 97 m/min, a feed rate of 0.115 mm/rev, and a depth of cut of 0.38 mm were identified through response surface methodology (RSM). To develop a tool wear predictive model, fresh experiments were conducted using an optimal set of parameters. Tool vibration signals and machined surface textures were acquired as inputs, and tool flank wear was measured as the target parameter. Tool vibration signals were transformed from time series to polar coordinates using the Gramian angular field (GAF) to capture spatial information. Features were extracted through Gabor wavelet transform (GWT) and selected using kernel-based principal component analysis (KPCA). These features were fed to different machine learning algorithms, including support vector regression (SVR), random forest (RF), decision tree (DT), least squares (LS), and ridge regression (RR), with an optimal 85:15 data split. The SVR model achieved 76.21% accuracy with this split. SVR hyperparameters were optimized using grid search CV, randomized search CV, particle swarm optimization (PSO), differential evolution technique (DET), and harmony search optimization (HSO, with SVR-PSO achieving the best results: accuracy of 96.17%, mean absolute error (MAE) of 0.0436, and root mean square error (RMSE) of 0.00358. Experimental validation showed the SVR-PSO algorithm achieving prediction accuracy of 97.95%, MAE of 0.0115, and RMSE of 0.0158.