Power Transformer Fault Diagnosis and Prediction Based on Deep Learning

Abstract

Power transformers assume the core functions of power conversion and power flow regulation in the new power system, and its fault outage will lead to the chain risk of power grid. However, the diagnostic accuracy of traditional threshold methods and shallow machine learning is limited in the face of non‑stationary and strong‑noise1 signals, and the data distribution deviation caused by variable load conditions seriously restricts the field generalization ability of the model. To solve the above bottleneck, this paper proposes a complete technical system that integrates multi‑domain feature enhancement, dual‑branch domain‑adaptive diagnosis, and temporal adversarial prediction. In the preprocessing stage, an adaptive denoising strategy based on improved variational mode decomposition and permutation entropy is designed, and the time‑frequency‑energy 3D multi‑channel input tensor is constructed by using the synchrosqueezed S‑transform (SSST) combined with kurtosis, peak factor, and energy spectrum entropy, which effectively solves the problem of poor separability of fault features under strong noise. In the diagnosis stage, a dual‑branch adversarial domain‑adaptive convolutional network is proposed. With dilated EfficientNet as the backbone, global semantic and local sensitivity dual branches are set to extract wideband and transient features in parallel, and the dual‑domain adaptation mechanism of maximum mean discrepancy and an adversarial discriminator is introduced to achieve high‑precision diagnosis in the unlabeled target domain under variable working conditions. The experimental results show that the accuracy of the model reaches 94.7% in cross‑operating‑condition scenarios, which is 8.2 percentage points higher than that of the traditional domain‑adversarial network. In the prediction stage, a temporal generative adversarial network is constructed, which takes dilated causal convolution and self‑attention spatio‑temporal gating as its core, combines Monte Carlo dropout interval estimation guided by dynamic time warping morphological loss and discriminator confidence, and realizes probabilistic prediction of the remaining useful life (RUL), and the prediction interval coverage rate reaches 88.6%. The method in this paper has achieved significant performance improvement in the three dimensions of diagnosis accuracy, prediction reliability and computational efficiency, and provides complete technical support for the intelligent operation and maintenance of power transformers from state recognition to trend prediction.