A novel bearing fault diagnosis approach combining F-VMD with CNN-BiLSTM and cross-attention

Abstract

To enhance the reliability and accuracy of bearing fault diagnosis under challenging operating conditions, this study proposes a method that integrates a combined fast Fourier transform (FFT) and variational mode decomposition (VMD) preprocessing framework with a convolutional neural network (CNN)-bidirectional long short-term memory (BiLSTM)-cross-attention network. In the preprocessing stage, FFT is applied to the raw signal to extract its global frequency spectrum, while VMD decomposes the signal into multiple time-domain intrinsic mode components; the subsequent fusion of these two types of features constructs a dual-view feature space that simultaneously captures transient impulsive dynamics and steady harmonic patterns, thereby enriching the discriminative information for fault recognition. Then, the CNN module extracts local spatial features from the fused representation, the BiLSTM captures bidirectional temporal dependencies, and a cross-attention mechanism is further employed to enhance cross-feature interactions and promote collaborative feature learning. Experimental validation on the Case Western Reserve University dataset and a self-collected dataset demonstrates that the proposed method achieves over 98% classification accuracy, and maintains robust performance under Gaussian white noise and simulated mechanical noise. These results confirm its excellent diagnostic accuracy, robustness, and generalization capability, offering effective support for intelligent condition monitoring.