Rolling Bearing Fault Diagnosis Model Based on Phase Space Reconstruction Combined with Two-Channel Multi-Scale Feed-Forward Network

To address the issue of reduced diagnostic robustness in rolling bearings due to noise interference, this study proposes a novel PSR-MSFFN-BiGRU fault diagnosis model based on chaos theory. The one-dimensional raw signal is transformed into a two-dimensional chaotic signal u sing Phase Space Reconstruction (PSR). Following this, the MSFFN-BiGRU model is designed, integrating a parallel dual-channel feed-forward network for signal feature extraction and a bidirectional gated recurrent unit (BiGRU) to improve diagnostic performance. Experimental evaluations conducted on the bearing datasets from both CWRU and JNU indicate that the maximum Lyapunov exponents of rolling bearing fault signals are consistently greater than zero, and the correlation dimensions are non-integer values exceeding zero, confirming the chaotic nature of signals across different fault types. Under noise-free conditions, the proposed model achieves average classification accuracies of 99.97 % and 98.73 % on the CWRU and JNU datasets, respectively. Even under noisy conditions, with a signal-to-noise ratio of -8 dB, the model still maintains high accuracies of 90 % and 88.75 %, respectively. These results strongly demonstrate the model's outstanding robustness and excellent generalization capability.