A Double-Branch Improved Residual Shrinkage Network for Diagnosis of Induction Motor Broken Rotor Bar Under Small Samples

• Published in: IEEE transactions on instrumentation and measurement 2025-01, Vol.74, p.1-12
• Main Authors: Liu, Guohai, Jiang, Qianyi, Sun, Yao, Song, Xiangjin, Tang, Han, Liu, Zhengmeng, Chen, Qian
• Format: Article
• Language: English
• Subjects: Accuracy; Bars; Bidirectional gated recurrent unit (GRU); broken rotor bars (BRBs); Fault diagnosis; Feature extraction; Motors; Noise; regularization; residual shrinkage network; Rotors; small samples; Training; Transforms; Vibrations
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3502726

The broken rotor bar (BRB) fault is one of the typical faults of induction motors (IMs). The accurate fault diagnosis of BRB can effectively reduce economic losses and enhance the stability of the system. Aiming at the problems of limited data collection and the difficulty in estimating the severity of faults, a double-branch improved residual shrinkage network (DB-IRSN) for BRBs in IMs was proposed, where both transformed current and vibration signals are utilized. First, the Hilbert transform is utilized to enhance fault features in current signals. Subsequently, the current and vibration signals are input into the IRSN simultaneously to filter the noise component of the signal. Moreover, the dense connection mechanism is used to connect the residual shrinkage module to avoid the loss of effective information. Then, regularization methods are used to avoid the overfitting phenomenon under small samples, and the bidirectional gated recurrent unit (BiGRU) is used to fuse spatiotemporal features. DB-IRSN is capable of fully extracting the fault characteristics from two signals to achieve good performance in diagnosis. Finally, the experiments are carried out on two datasets obtained from Sao Paulo University and the experimental platform under different severities of BRBs and loads. When the training set accounts for 0.1, the model achieves an accuracy of 97.98% and 97.65%, respectively, with performance reaching 99% as the proportion of the training set increases. In addition, experimental results in noisy environments and unbalanced voltage (UV) indicate that the proposed DB-IRSN demonstrates satisfactory performance under small samples. Comparative experiments show that DB-IRSN has higher accuracy compared to existing networks.