Research on Transformer Oil Multi-frequency Ultrasonic Monitoring Technology Based on Convolutional Neural Network

Abstract

Facing the deficit of the effective measures to evaluate the insulation status of power transformers in service, this paper brought up a new method to estimate the physical and chemical properties of transformer oil through its transmission characteristics for ultrasonic signals under various frequencies. However, given the large volume of acquired ultrasound spectrum data by such technology, and the complexity as well as the variety of the transformer structures and conditions in which the transformer oil resides in, the interpretation of above data and the prediction of transformer health becomes enigmatic. Thus a recognition method was brought up by this paper to connect the ultrasonic spectrum to transformer oil conditions through Convolutional Neural Network. First of all, for the transformer oil test data, by using the density-based clustering method, the “ standard oil” and other “degraded oils” approaching the standard are distinguished to achieve the purpose of distinguishing the quality of the transformer oil. Then, the principal component analysis is used to reduce the dimensionality of the ultrasonic spectrum data of the transformer oil. The dimensionality classification results of the reduced dimensional ultrasonic spectrum data and transformer oil test parameters are used as the input and output data of the algorithm model. The Convolutional Neural Network is established and the model parameters are trained. The final accuracy rate of the assessment model is 92%. Finally, a transformer oil condition detection method based on multi-frequency ultrasonic spectroscopy was established.