A machine learning (GAF + SVR-PSO) model for in-process cutting tool wear assessment using multi-sensor heterogeneous data fusion

This research presents a machine learning-based system for accurate tool-wear assessment by fusing vibration and surface texture data. The study first identified that tool-wear mechanisms shift from stable abrasion to severe adhesive/diffusion wear as cutting parameters increase, a transition confirmed by SEM/EDX analysis. Response Surface Methodology (RSM) was employed to determine the optimal machining parameters: a cutting speed of 97 m/min, a feed rate of 0.115 mm/rev, and a depth of cut of 0.38 mm. Fresh experiments were then conducted using this optimal parameter set to generate data. Tool vibration signals were converted from time series to spatial images using a Gramian angular field (GAF), while machined surface textures were analyzed. Features were extracted via Gabor wavelet transform (GWT) and reduced using Kernel-based Principal Component Analysis (KPCA). Various machine learning models were evaluated on an 85:15 data split, with a baseline Support Vector Regression (SVR) model achieving 76.215% accuracy. The SVR's hyperparameters were subsequently optimized using multiple techniques. The SVR model optimized with Particle Swarm Optimization (PSO) delivered superior performance, achieving 96.17% accuracy, a mean absolute error (MAE) of 0.0436, and a root mean square error (RMSE) of 0.00358. Finally, experimental validation confirmed the model's robustness, with the SVR-PSO algorithm reaching a prediction accuracy of 97.95%, an MAE of 0.0115, and an RMSE of 0.0158.