Wavelet-based hybrid CNN-BiLSTM approach in tool wear monitoring

Recent advancements in deep learning techniques have revolutionized modeling from raw sensor data, leading to more challenges in tool wear monitoring. An effective tool wear monitoring model ensures reliable and efficient manufacturing processes. This work presents a novel approach that combines wavelet transform techniques for multiresolution analysis with a hybrid deep learning method for tool wear identification. The proposed model incorporates the Maximal Overlap Discrete Wavelet Transform (MODWT) for raw signal preprocessing, paired with a hybrid deep learning architecture that combines CNNs and BiLSTM. First, to identify the most significant information related to tool condition, the energy distribution across each MODWT decomposition level is analyzed. Higher decomposition levels play a critical role in capturing the cutting signal's medium- and low-frequency components, which are crucial for detecting significant changes in the tool wear progression. Then, the effectiveness of the proposed model was evaluated and compared to other widely used deep learning models using milling datasets. The results show that the MODWT-CNN-BiLSTM model consistently outperforms other models in terms of key performance metrics, including Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and regression coefficient (R²) across all datasets. Specifically, the proposed model achieved the lowest MAE (0.0197) and RMSE (0.0276), along with the highest R² value (0.9839), demonstrating superior accuracy and reliability. This approach underscores the effectiveness of combining wavelet-based analysis, such as MODWT, with deep learning techniques like CNNs and BiLSTM for tool wear monitoring in practical milling operations.