An Uncertainty Decomposition-Based Automated Out-of Distribution Detection Framework- for Trustworthy Machinery Fault Diagnosis

In modern industrial systems, machinery frequently encounters out-of-distribution (OOD) samples due to unknown faults and abnormal operating conditions. However, conventional deep learning methods, which rely on point estimation and lack uncertainty quantification capabilities, struggle to provide reliable judgments under such circumstances. Although existing Bayesian deep learning methods have made progress in identifying unknown faults, their quantification accuracy and ability to distinguish between different types of anomalies under unknown conditions remain limited. To address this issue, this paper proposes an uncertainty decomposition-based automatic detection framework to enhance the model's ability to detect OOD faults. First, a dynamic KL divergence weighting mechanism is introduced to improve the model's capability to capture uncertainties at different network layers. Then, a more refined uncertainty decomposition method is designed to enhance the interpretability of uncertainty quantification. Finally, an automatic detection mechanism based on statistical analysis is developed to enable automatic identification of OOD samples. Experimental results on two datasets demonstrate that the proposed framework not only improves the diagnostic accuracy of Bayesian deep learning models but also achieves more precise uncertainty quantification, thereby enhancing the reliability of diagnostic results.