Application of unsupervised identification of dissolved gases in transformer oil based on spin coating film making process

Abstract

Listen

Translate article

Addressing the issues of low efficiency and uneven collection of dissolved gases in transformer oil leading to overfitting and poor performance of identification models, we propose a novel film-making process that integrates Gaussian process and unsupervised pre-classification to enhance the recognition efficiency of dissolved gases in transformer oil. This method not only forms a thinner and more uniform separation layer, significantly improving degassing performance and collection efficiency, but also addresses the problems of insufficient data labeling and sample imbalance by introducing the K-means++ clustering algorithm and pseudo-random integration technology, thereby enhancing model robustness and generalization ability. Moreover, the designed Gaussian Process Multi-Classification (GPMC) method employs probabilistic interpretation for result presentation, which increases the accuracy of fault identification. Experimental results show that under consistent starting conditions, the RCC and ARI indicators of our pre-classification method are close to 0.8, with the test set’s recognition rate exceeding 80 %, while the GPMC method misclassified only 2.4 % of the cases in the 1800-case dataset. These improvements make our method particularly effective for handling uncertainties and imbalances in dissolved gas cases in transformer oil, showcasing its potential for practical applications.