Bearing Fault Diagnosis Using One-Dimensional Convolutional Neural Network-Informer-Multi-Head Attention Mechanism Parallel Modeling

Abstract

Based on the complex characteristics of rolling bearing vibration signals, such as non-stationarity, high noise, and long-sequence dependencies, this paper proposes a parallel fault diagnosis model that integrates a one-dimensional convolutional neural network (1DCNN), Informer, and MHA. This model effectively combines the local time-frequency feature extraction capabilities of 1DCNN, the long-sequence modeling advantages of Informer based on the ProbSparse mechanism, and the dynamic fusion of multi-scale features by the multi-head attention mechanism (MHA) to achieve efficient feature representation. To enhance the model's noise immunity and feature relevance, a dual preprocessing method of variational mode decomposition (VMD) and fast Fourier transform (FFT) is introduced to strengthen the capture of fault features in non-stationary vibration signals. Comparative experimental results on the Case Western Reserve University (CWRU) standard dataset demonstrate that the proposed model exhibits significant diagnostic performance. The model achieves an average accuracy of 99.9% and an  of 0.99 under ten fault conditions, with minimal standard deviation. It outperforms several existing methods (such as CNN-LSTM, TCN, and Transformer), demonstrating its superior accuracy, stability, and robustness. Further t-SNE visualization analysis intuitively demonstrates the model's excellent classification boundaries and cluster separability in high-dimensional feature spaces. This research presents a highly accurate, efficient, and generalizable solution for intelligent fault diagnosis under complex operating conditions, offering significant reference value for promoting theoretical development and industrial applications in the field of intelligent diagnosis.