Enhancing tool wear prediction with deep learning models in additive manufacturing processes

The requirement for a discovery representation to predict the automatic breakdown of factory systems’ long-term useful existence or mechanisms has developed for manufacturers. Typical methods of Tool Wear (TW) forecasting either utilize physics-based modeling and/or a statistical method that requires significant manual feature selection and often have the added complication of dealing with real-time data, which reduces predictive accuracy and efficiency in modern manufacturing environments. In intelligent manufacturing, TW monitoring has become more crucial to improve machining efficiency. TW state can be efficiently characterized by multi-domain features; however, manual feature fusion reduces monitoring efficiency and prevents further advancements in prediction accuracy. This research proposes a new wear-predicting method using L2 Regularization optimized Dynamic Artificial Neural Network (L2RO-DANN) for multi-domain feature fusion that overcomes these deficiencies. The approach involves data pre-processing through Wavelet Transform (WT) and data augmentation, feature extraction with Short-Time Fourier Transform (STFT), and vision transformer (ViT). TW detection is performed by using L2RO-DANN for improved efficiency. The error rate and diagnosis performance findings illustrate the efficacy of the suggested framework in terms of MAE (0.0032), R² (1.0000), RMSE (0.0041), sensitivity (99.12%), specificity (99.89%), and cumulative error (0.03%).