Abstract-Fans,as critical auxiliary equipment in thermal power plants, frequently experience abnormal operating conditionsand suffer from low detection accuracy when using single-classifier methods. To address these issues, this paper proposes an abnormal condition detection method combining the Boruta algorithm and D-S evidence theory. First, the Boruta algorithm is employed to screen critical variables, and an integrated anomaly detection model is constructed using three algorithms: SVM, Deep Forest, and LGBM. The WD-S method optimizes evidence fusion through the MK-MMD. Experimental results demonstrate that the proposed method achieves an accuracy of 99.85%, outperforming individual classifier algorithms.