Artificial Intelligence in Defect Classification Tasks Using the Example of Chromatographic Analysis of Dissolved Gases in Power Transformers

In this paper, an approach to solving classification problems is considered using the example of the analysis of dissolved gases in a power transformer. The approach was verified by an input dataset consisted of the gases concentration values in the transformer oil, the excess of which indicates a specific defect in the transformer as a result of its operation in an abnormal mode. The processes of exploratory data analysis and input data preprocessing were shown. To improve the performance of machine learning and neural network models, 10 standard statistical features describing the distribution of a random variable were extracted from the input data set, and a correlation matrix was calculated using the Pearson correlation criterion to eliminate multicollenarity. The following classification algorithms were used: gradient boosting, Decision tree, Random Forest, k-nearest neighbors, and artificial neural networks). By training the models and testing them on a preprocessed dataset that none of the models had previously used, we obtained a prediction result with high accuracy. Using the prediction distribution, a confusion matrix was compiled and the accuracy of the algorithms was estimated.