Comparative evaluation of feature selection methods and deep learning models for precise tool wear prediction

Tool wear prediction is essential in modern manufacturing, as it significantly affects product quality, production costs, and machine lifespan. The accuracy of predictive models relies heavily on the quality of input data and the effectiveness of feature selection, which simplifies the dataset while retaining critical information. In this study, we investigate different feature selection techniques using the PHM 2010 milling dataset, including the Pearson correlation coefficient (PCC), random forest (RF), and a hybrid method combining PCC and RF with principal component analysis (PCA). We assess the effectiveness of these feature selection methods on machine learning models, including artificial neural networks (ANN), one-dimensional convolutional neural networks (1-D CNN), and bi-directional long short-term memory networks (Bi-LSTM). The model performance is evaluated using standard metrics, such as mean absolute error (MAE), root mean square error (RMSE), and R-squared (R²). Our results show that the hybrid feature selection approach yields the best results, with the Bi-LSTM model achieving an average MAE of 2.70, RMSE of 3.27, and R-squared of 0.98 on the T1 dataset.