Multi-modal deep learning bearing fault diagnosis based on time domain-angular field-phase space fusion

Abstract Bearing fault diagnosis in rotating machinery is challenged by the complex, nonlinear, and nonstationary nature of vibration signals. To address these limitations, this paper proposes a novel three-stream deep learning framework that synergistically integrates complementary representations from temporal, angular, and phase-space domains. The architecture comprises three parallel branches: a one-dimensional convolutional neural network extracts hierarchical features directly from raw vibration signals; a two-dimensional convolutional neural network captures spatio-temporal patterns from Gramian Angular Summation Field images encoding long-range dependencies; and a fully-connected network processes recurrence quantification analysis features to characterize nonlinear dynamics. A Bidirectional LSTM network subsequently performs adaptive feature fusion and captures temporal dependencies across modalities. Comprehensive experimental evaluations on multiple bearing fault datasets demonstrate that the proposed method effectively overcomes the limitations of single-domain analysis while maintaining competitive model complexity, offering a practical and interpretable solution for industrial bearing fault diagnosis applications.