Wireless Vibration Signal Prediction of Power Transformers Based on Contrastive Adversarial Sparse Model

Wireless vibration analysis is regarded as a cost-effective and nondestructive method for monitoring the operational states of power transformers. However, influenced by noise and mutation information, vibration-based prediction methods remain in the preliminary phase for precise prediction of power transformer. To address this issue, a contrastive adversarial sparse model is proposed to predict the progress of power transformer faults in the early phases. First, a sparse transformer improves the feature extraction ability of the bidirectional recurrent gated network to reduce noise and extracts significant information from original signal. Then, the mutation information is suppressed by a contrastive adversarial method, based on which discrimination layer predicts the power transformer fault. Finally, to constrain the network, a contrastive loss based on feature matching is proposed to learn more extra information distinct from sparse transforms. Datasets are collected using wireless vibration sensors from a 10-kV real power transformer. Under an identical parameter setting, experimental results show that the proposed model outperforms comparison algorithms in terms of root mean square error (RMSE), correlation coefficient, symmetric mean absolute percentage error, and mean absolute error. Especially, the RMSE of the proposed model is 0.0242, exceeding the performance of the conventional model in power transformer fault prediction.