DBF-CNN: A Double-Branch Fusion Residual CNN for Diagnosis of Induction Motor Broken Rotor Bar
Y H Shu, Wenlei Zhang, Xiangjin Song, Guohai Liu, Qianyi Jiang
Abstract
Effective diagnosis of broken rotor bars (BRBs) is crucial for the safe operation of induction motors (IMs). The broken rotor bar feature component is small in amplitude and close in frequency to the fundamental frequency, resulting in it being easily masked by the fundamental frequency component. Due to insufficient feature extraction and utilization, most of the current deep learning-based diagnostic methods for BRBs cannot estimate the severity of faults. To solve these problems, a feature fusion residual convolutional neural network with a double branch (DBF-CNN) for BRBs in IM was proposed. First, the Hilbert transform was applied to the raw current signal to highlight the fault features. Thereafter, the residual structure was used to avoid the degradation problem caused by network deepening. Meanwhile, a double-branch structure was used to extract global and local features from the signal separately. In addition, an attention feature fusion method was proposed to solve the problem of feature loss and offset that may occur during the fusion process. Forward and reverse fusion were used for the two branches, respectively. DBF-CNN can fully utilize the features in the current signal to estimate the severity of rotor bar breakage under different load conditions. Finally, the proposed network was tested using data collected under different load conditions and fault severities. Experimental results on a publicly available dataset from the Sao Paulo University and a dataset collected by the experimental platform show that DBF-CNN is able to achieve accuracy rates of 99.86% and 99.00%. In addition, based on the results of the comparison experiments, DBF-CNN has higher accuracy and faster convergence speed than existing networks.