Multi-Branch Residual Multiscale CNN Based Power Transformer Fault Diagnosis on Vibration Signal

Status of the transformer is highly associated with safe and stable operation of the whole power system. Since the vibration signal is generated along with operation of the transformer and can indicate change of equipment operation status in real-time, fault diagnosis based on the vibration signal is a feasible method to monitor status of the transformer. In this paper, an end-to-end multi-branch-attention-multiscale CNN (MAMCNN) framework is proposed based on a one-dimensional convolutional neural network, in which multi-branch inputs, multiscale residual learning, and attention mechanism-guided multi-branch fusion techniques are integrated to identify states of the 220 kV transformer. To address the problem of small samples for transformer fault diagnosis, the proposed method first tests on a public dataset of rolling element bearing vibration. Results show that MAMCNN still has good differentiation for fault features under strong noise and fluctuating operating conditions without denoising, and has high accuracy and stability for state identification. MAMCNN is then applied to 220 kV transformer fault diagnosis based on vibration signals, and results exhibit high accuracy, rapid and stable convergence in identifying four transformer states.