Fault diagnosis method for rolling bearings based on improved recurrence plot and EfficientNet

Abstract To address the challenges of manual feature extraction reliance and difficulty in identifying weak faults in rolling bearing fault diagnosis, this paper proposes a novel method based on an improved Recurrence Plot (RP) and EfficientNet. In the signal encoding stage, the method directly constructs recurrence images using the Euclidean distance matrix, termed the Continuous Distance Recurrence Matrix (CDRM)–that is, it preserves the complete continuous values of Euclidean distances between phase space trajectories as recurrence matrix elements, avoiding information loss caused by traditional binarization thresholding and enhancing the characterization of nonlinear dynamic features in vibration signals. In the network architecture, to overcome the limitation that the standard Squeeze-and-Excitation (SE) attention mechanism relies solely on global average pooling for spatial compression while neglecting multi-directional structural information, a Multi-Path Squeeze-and-Excitation (MP-SE) attention mechanism is proposed. MP-SE introduces four parallel pooling paths–global average pooling, horizontal strip pooling, vertical strip pooling, and diagonal pooling–to aggregate feature responses from different spatial dimensions, enabling the network to adaptively focus on fault-relevant multi-directional texture regions. The combination of these two innovations forms an end-to-end diagnostic framework. Experimental results on the Case Western Reserve University dataset demonstrate that the proposed method achieves a best average accuracy of 99.39% (std 0.08%) in identifying 10 fault states. Comparative analysis of encoding methods reveals that CDRM (99.39%) outperforms grayscale image encoding (92.61%), Gramian Angular Summation Field (95.72%), Gramian Angular Difference Field (96.39%), and traditional RP (97.72%). Network structure comparisons show that EfficientNet with MP-SE (99.39%) surpasses Inception (95.72%), ResNet (98.50%), DenseNet (98.56%), and MobileNet (96.83%). In cross-speed generalization experiments, EfficientNet achieved optimal performance on all cross-speed tasks. Cross-dataset generalization experiments further validate the method’s good generalization ability under different data distributions. The comprehensive experimental results validate the high efficiency and reliability of the proposed method in rolling bearing fault diagnosis.