Multimodal Fault Diagnosis of Rotating Machinery Using Dynamic Fusion and Improved DS Theory

Noting that single mode fault diagnosis methods have limitations in complex fault detection, and single fault diagnosis models have low accuracy and robustness, this article proposes a multimode fault diagnosis method for rotating mechanical equipment based on feature-level dynamic fusion and decision-level improved Dempster–Shafer (DS) evidence theory. First, by introducing time-domain, frequency-domain, and time-frequency domain data, three fault diagnosis submodels: convolutional neural network, long short-term memory network, and random forest are constructed for extracting and processing three different modal input signals. Second, a dynamic fusion strategy is developed during feature-level fusion to gradually optimize the features of different modalities, to improve the robustness and accuracy of fault diagnosis. Preliminary fault classification results for each submodel are obtained through different feature-level fusion methods. Third, the improved DS evidence theory is used to fuse the fault classification information of each submodel, and the evidence is corrected from both the performance and fault type aspects of the submodels, to obtain the final fault diagnosis result after fusion. Finally, the superiority of the proposed method is verified through experiments.