This paper proposes a fault diagnosis method based on ensemble learning (EL) and dual-layer evidence fusion (DEF)

to address the prevalent challenges of class imbalance and inter-modal conflict in circuit breaker diagnosis. In the proposed method

first

the EasyEnsemble integration strategy is employed to construct modality-specific sub-classifier ensembles for each signal source. Second

the Dempster-Shafer (D-S) evidence theory is utilized to integrate the outputs of these sub-classifier ensembles

converting the predictive results of multiple base models within the same modality into a unified

high-certainty body of evidence

thereby realizing the first-layer fusion. Finally

the independent evidence bodies outputted by each modality are combined again based on D-S theory to achieve the second-layer fusion

generating the final system decision that possesses inherent fault tolerance. A comprehensive experimental platform was established for data acquisition and validation. Results demonstrate that the proposed method achieves an F1 score of 92.85%

significantly outperforming all comparison algorithms. Crucially

the method maintains a high level of diagnostic performance even under simulated sensor failure conditions

and its ability to convert physical faults into quantifiable decision alerts via the conflict coefficient validates its effectiveness and superior robustness in complex industrial environments.