An Automatic Annotation and Pattern Recognition Method Based on Semisupervised ACGAN for Multisource Partial Discharge Diagnosis

The low on-site identification accuracy of data-driven partial discharge (PD) recognition methods is attributed to the overlap of features from multiple pulse sources and the inconsistent distribution of samples. To address the challenges of misreporting and misclassification inherent in traditional diagnostic approaches in this domain, an end-to-end semisupervised multisource partial discharge pattern recognition method is proposed, which could enhance the efficiency of on-site sample utilization. Initially, the enhanced auxiliary classifier generative adversarial network (ACGAN) is employed to learn from the labeled laboratory sample set, which could ensure heightened recognition accuracy on this training dataset. Subsequently, the model parameters are frozen, and the discriminator’s end-layer with a fully connected layer and Softmax activation function is utilized to perform online annotation on the unlabeled on-site samples’ phase resolved partial discharge (PRPD) spectrograms, obtaining pseudolabeled samples. Finally, the laboratory samples and pseudolabeled samples are merged to construct a new training set, and the model parameters are iteratively updated using this dataset. The final multisource partial discharge identification results are obtained through downstream fine-tuning. The results show that the proposed method can effectively solve the problem of inconsistent sample distribution through automatic labeling and improve the utilization efficiency of unlabeled samples in the field. Compared with traditional diagnostic methods, the identification accuracy rate, accuracy rate, and recall rate on the field sample set are increased by 14.89%, 11.77%, and 15.17%, respectively.