New fault diagnosis framework for railway switch machines based on machine learning with heterogeneous model fusion

As critical electromechanical components for modern railway operations, railway switch machines require highly accurate fault diagnosis to ensure operational safety, enhance efficiency, and prevent major failures. Current fault diagnosis methods face critical challenges in real-world industrial environments, including poor generalization ability, the scarcity of fault samples, and low diagnostic accuracy for multiple faults. To address the poor generalization ability, we propose the electro-dynamic parameter arithmetic fusion (EPAF) method to integrate multi-source sensor signals, which are subsequently encoded into image inputs via the Gramian angular field (GAF). In view of the scarcity of fault samples, an improved vector quantized variational autoencoder (IVQ-VAE) is introduced to generate high-quality synthetic samples, complemented by transfer learning (TL) for model pretraining and fine-tuning. To enhance diagnostic accuracy for multiple faults, a hybrid strategy is designed to improve the effectiveness of heterogeneous model fusion. The framework is validated using a real-world railway switch machine dataset. The experimental results demonstrate that the framework covered 15 distinct states, achieving a high diagnostic accuracy of 95.24% on the target dataset. This result validates the high robustness and practical value of the framework under complex operating conditions. • Average fusion combines multi-source data, and GAF converts it to 2D. • Multi-model fusion with transfer learning improves fault diagnosis. • A hybrid strategy is proposed and validated for multi-model fusion. • Real-world data validates the method’s effectiveness and generalization.