Fault diagnosis in railway transmission systems is critical for operational safety, yet existing methods often fail to address complex, system-level fault combinations, especially those unseen during training. This paper introduces the Uncertainty-Aware Heterogeneous Blind Deconvolution (UncertainHBD) ensemble network, an end-to-end framework for reliable system-level diagnosis. First, we construct an ensemble of heterogeneous neural blind deconvolution (HBD)-based submodels to extract robust component-level features from tri-axis vibration signals. Second, a novel prototype-similarity-based reliability method is proposed to distinguish known (in-distribution) and unknown (out-of-distribution) fault states. For known faults, the model provides a high-confidence system-level diagnosis. For novel unseen combinations, it integrates component-level predictions from submodels with a calibrated reliability score. This dual-path approach delivers high diagnostic accuracy for known fault conditions while robustly addressing previously unseen scenarios. Experimental evaluations on the BJTU‑RAO bogie datasets confirm the proposed method’s effectiveness and reliability.

System-Level Fault Diagnosis,Blind Deconvolution,Heterogeneous Neural Network,Reliability Estimation