Fault Diagnosis of Rotating Machinery Bearings Based on GNCNN-BIGRU and Multimodal Feature Fusion

In engineering practice, the limited availability of fault samples and the single-modal nature of data often result in insufficient feature extraction for fault diagnosis. To overcome the limitations of single-channel data interpretation, which cannot fully capture fault characteristics or achieve high diagnostic accuracy, and considering that fault signals of rotating machinery bearings in practical applications are often weak and easily obscured by noise, this paper proposes a parallel dual-channel fault diagnosis method for motor bearings. The proposed method integrates image-based and time-series signal feature extraction for fusion classification. Specifically, the Gramian Angular Difference Field (GADF) is employed to transform time-series signals into two-dimensional images, which are processed simultaneously with the original one-dimensional signals through a parallel dual-channel model. The first channel utilizes a Modified Convolutional Neural Network (MCNN) with Group Normalization (GN) to extract spatial features, while the second channel employs a Bidirectional Gated Recurrent Unit (BIGRU) to capture temporal features. Experimental evaluations on the Case Western Reserve University (CWRU) dataset reveal that the proposed method achieves an average fault diagnosis accuracy of 99.80%, demonstrating superior accuracy and robustness compared to single-channel models.