Unsupervised learning and expanded strategy for Multi-parameter coupling oil-immersed transformer intelligent fault diagnosis system

The scarcity and imbalance of transformer fault data can result in issues such as model overfitting, reduced model representativeness, and diminished performance on test sets, thereby making it challenging to accurately assess the condition of the transformer. To address this issue, the paper investigates the similarity between the sequence of dissolved gases in transformer oil and DNA sequences. It then introduces a case augmentation method for dissolved gases in transformer oil, based on pseudo-random integration. This approach aims to provide additional data while mitigating the randomness that can lead to instability in classification accuracy. Furthermore, most current methods are unable to tackle the challenge of zero-sample identification, which involves recognizing fault conditions that are not present in the training data and for which no prior examples exist. This limitation hinders their practical application. To achieve this, the K-means++ clustering algorithm is utilized. Leveraging the characteristics of dissolved gas cases in oil, unsupervised learning and classification are performed. Additionally, a Kvalue estimation method based on the Silhouette coefficient and the Calinski-Harabaz index is proposed, which improves the efficacy and reliability of the unsupervised performance. The experiment confirmed the effectiveness of using a deep residual network (ResNet) to classify nine different transformer fault states. The results indicated that, with the same initial dataset, the pseudo-random integration method yields significantly more accurate and stable outcomes, enhancing the accuracy of transformer fault classification.