Hierarchical Fault Diagnosis for Rolling Bearings Based on Convolutional Neural Network With Tree Decision Layer

To address under-utilization of the prior knowledge provided by the hierarchical fault information and low diagnostic accuracy resulting from ignoring the hierarchical relationship between fault location and fault severity, an integrated model that combines convolutional neural network (CNN) with the tree decision layer for bearing faults diagnosis is proposed. The fault location was first judged, and the fault severity was then judged based on that. The proposed integrated model leveraged both the feature extraction capability of CNN and the hierarchical structure and decision-making characteristics of decision tree. First, a shared network backbone and two task-specific branching fully connected layers were utilized to extract features related to fault location and fault severity. Second, the tree decision layer was introduced to integrate their classification results, and then the weights of the model were adjusted using a weighted hierarchical classification loss generated by the tree decision layer, so as to learn the hierarchical information of bearing fault diagnosis. Finally, the effectiveness and performance of the proposed model were validated based on the Paderborn University bearing datasets. Results show that the proposed model can achieve an average classification accuracy of 99.15%. Compared with the existing diagnostic models in the field, the proposed model can achieve more accurate fault location and severity classification.