Investigation of an Optimized Linear Regression Model with Nonlinear Error Compensation for Tool Wear Prediction

To solve the problem of insufficient accuracy in tool wear process modeling and Remaining Useful Life (RUL) estimation, this study proposes a two-stage prediction method. Firstly, a linear prediction benchmark model is constructed: Support Vector Regression (SVR) is used to preliminarily model the tool wear process, obtaining initial prediction results and their error distribution. Building on this foundation, an Autoencoder (AE) is introduced to establish a nonlinear mapping relationship for the errors, achieving effective compensation of the SVR prediction results and establishing the SVR–AE prediction model. To further enhance model performance, the Ant Colony Optimization (ACO) algorithm is utilized to optimize three key parameters: the number of training epochs, batch size, and hidden layer dimensions, ultimately establishing the ACO–SVR–AE optimization model. Experimental validation demonstrates that on the PHM2010 dataset, compared to the Support Vector Regression (SVR) and Autoencoder (AE) models, the proposed method achieves average reductions of 26.1% in Mean Squared Error (MSE) and 14.5% in Mean Absolute Error (MAE). Compared to traditional random forest and neural network models, the MSE and MAE show average reductions of 32.3% and 25.3%. By combining linear modeling with nonlinear error compensation, this method provides an integrated optimization approach to prediction tasks in complex industrial scenarios.