Multi-modal denoised data-driven milling chatter detection using an optimized hybrid neural network architecture.

Chatter, a type of self-excited vibration, deteriorates surface quality and reduces tool life and machining efficiency. Chatter detection serves as an effective approach to achieve stable cutting. To address the low accuracy in chatter detection caused by the limitations of both one-dimensional temporal and two-dimensional image modal information, this study proposes a multi-modal denoised data-driven milling chatter detection method using an optimized hybrid neural network architecture. A data denoising model combining Complementary Ensemble Empirical Mode Decomposition (CEEMD) and Singular Value Decomposition (SVD) is established. The Ivy algorithm is employed to optimize the hyperparameters of CEEMD-SVD. Multi-modal data features of different machining states are then obtained using time-frequency domain methods and Markov transition field methods. Sensitivity analysis of time-frequency domain features is conducted using Pearson correlation coefficient analysis. A hybrid neural network model (DBMA) for chatter detection is constructed by integrating dual-scale parallel convolutional neural networks, bidirectional gated recurrent units, and multi-head attention mechanisms. The Ivy algorithm is utilized to optimize the hyperparameters of DBMA. The t-SNE algorithm is employed to visualize features extracted from different network layers of the chatter detection model. Results demonstrate that effective denoising of machining signals and the use of multi-modal data can significantly improve the accuracy of state detection. Compared with other methods, the proposed model exhibits superior stability and robustness.