Abstract: Unlike single parts compound faults, fault information in multiple parts compound faults is dispersed throughout numerous domains and features are related to one another, resulting in a high number of redundant or irrelevant features in the feature set. A three-stage hybrid feature selection strategy was proposed to overcome this problem by screening sensitive features from feature sets and improving fault classification accuracy. The fault features were first evaluated using four filter models, and the evaluation findings were then weighted and ranked depending on the classification error rate. Finally, based on the weighted sorting results, three heuristic search strategies were merged to select the optimum sub-set. Through a set of gear-bearing compound fault data sets containing 11 fault categories, the result shows that the proposed method may drastically reduce the dimension of the feature set while also boosting classification accuracy.