Evidence Convolutional Model for Fault Diagnosis: Uncertainty-Aware Deep Learning With Evidence Theory

With the rapid advancement of deep learning (DL) technology, data-driven fault diagnosis methods, particularly those that leverage convolutional neural networks (CNNs) for feature extraction, have demonstrated the potential to achieve high accuracy through end-to-end learning. However, most deep learning-based models cannot estimate the uncertainty of diagnostic results, which is crucial in practical applications, especially when encountering out-of-distribution (OOD) data. To address this issue, this paper proposes an Evidence One-Dimensional Convolutional Model (EODCM), which integrates evidence theory with multilayer one-dimensional convolution (ODC) for fault diagnosis. By introducing additional uncertainty estimation capabilities into the convolutional model, EODCM not only achieves high fault diagnosis accuracy but also provides uncertainty quantification for its predictions. Specifically, the proposed method treats the predictions of the ODC as subjective evidence and imposes a Dirichlet distribution on class probabilities, enabling effective evidence collection from data during training. Additionally, a specific loss function based on evidence theory is employed to enhance the uncertainty estimation capability of the model. To further improve generalization performance, a feature fusion strategy is adopted to integrate multi-view information, enriching the feature representation of one-dimensional data. Experimental validation on a gearbox fault diagnosis case from Southeast University demonstrates that EODCM effectively estimates predictive uncertainty and reduces overconfidence by 16.90% when facing OOD data. Furthermore, the proposed method maintains robust performance under varying noise conditions, highlighting its great potential for practical applications.