Multimodal Data Fusion-Driven Fault Diagnosis of Aero-engine Bearings Using a Pseudo-Siamese Neural Network

To address the challenges of fault diagnosis for intershaft rolling bearings in aeroengines under complex operating conditions and imbalanced data, this paper proposes a fault diagnosis method based on multimodal data fusion. A three-channel pseudo-siamese neural network is developed (TrPSNN). Temporal vibration signals are processed through two branches: a simplified variational kernel CNN (1D-SVKNet) and a physics-guided CNN (1D-PGCNN), for feature extraction. A third branch uses a two-dimensional CNN (2D-GASFCNN) to extract time-frequency features from GAF images. Multimodal features are then fused in the fully connected layer. The model employs a Crested Porcupine Optimizer to tune the hyperparameters of a gated recurrent unit (CPO-GRU), which enhances convergence speed and classification performance. Experiments using the harbin institute of technology (HIT) aero-engine dataset and the case western reserve university (CWRU) bearing dataset demonstrate that the proposed method significantly outperforms single-modality and other conventional bearing fault diagnosis approaches. Even under imbalanced and few-sample conditions, the method achieves classification accuracy above 98.5%. Furthermore, ablation and adversarial noise experiments confirm that the proposed model outperforms other mainstream models in fault classification. These results demonstrate the model's superior feature extraction and transfer learning capabilities.