Multimodal Mutual Neural Network for Health Assessment of Power Transformer
Zhikai Xing, Yigang He
Abstract
Health assessment technology has a significant role in monitoring the status of a power transformer. However, machine learning has not been exploited effectively for health assessment due to the limitation of multimodal heterogeneity data. This article presents a multimodal mutual neural network to evaluate the health assessment of the power transformer. The multimodal data contain the dissolved gas in oil and infrared images. To capture the significant information from the multimodal data, we use the one-dimensional convolution neural network to extract the dissolved gas feature and deep residual squeeze-and-excitation neural network to extract the infrared images feature. And then, the multimodal mutual attention method extracts the significant information from dissolved gas feature and infrared images feature. Finally, the health assessment is output by Softmax. The dataset contains 41 power transformers with 1000 days operation. Through the experimental results, the presented method obtains a high classification accuracy and the accurate health assessment of the power transformer. Moreover, the proposed method can provide decision basis for actual operation and maintenance. The limitation of the proposed method is the large training time.