A parallel neural network method based on multi-level fusion of acoustic-vibration signals for slight and compound fault diagnosis of rolling bearings

Conventional diagnosis methods relying solely on vibration signals often fail to identify the weak fault features of slight and compound faults, particularly under noise interference. To address these limitations, this study proposes a parallel neural network method based on multi-level fusion of acoustic-vibration signal for rolling bearing fault diagnosis. First, Filter Bank (Fbank) features are extracted to enhance the representation of slight and compound faults within multimodal signals by addressing the challenges posed by weak fault features and high noise sensitivity. This approach enables the extraction of spectral information from both acoustic and vibration signals, thereby improving the feature representation capability of the proposed framework. Second, a CNN-BiGRU parallel neural network is constructed to comprehensively capture weak fault features by integrating a convolutional neural network for spatial feature extraction with a bidirectional gated recurrent unit (BiGRU) for temporal feature extraction. Finally, a multi-level fusion strategy that combines feature-level and decision-level fusion is adopted to exploit complementary information from rolling bearings, significantly improving diagnostic accuracy and overcoming the reduced reliability of slight and compound fault diagnosis caused by incomplete fault information in unimodal diagnostic methods. Experimental results on a slight and compound fault dataset demonstrate that the proposed method achieves superior diagnostic performance under various noise conditions. The proposed method achieves a diagnostic accuracy of 99.55% at SNR = 10 dB and maintains over 90% accuracy even at SNR = −10 dB, outperforming conventional diagnostic approaches.