Deep Learning Based Power Transformer Monitoring Using Partial Discharge Patterns

Abstract

Measurement and recognition of Partial Discharge (PD) in power apparatus is considered a protuberant tool for condition monitoring and assessing the state of a dielectric system. During operating conditions, PD may occur either in the form of single and multiple patterns in nature. Currently, for PD pattern recognition, deep learning approaches are used. To evaluate spatial order less features from the large-scale pat-terns, a pre-trained network is used. The major drawback of traditional approaches is that they generate high dimensional data or requires additional steps like dictionary learning and dimensionality reduction. However, in real-time applications, interference incorporated in the measured single and multiple PD patterns may reduce the identification of the exact patterns and causes inaccurate diagnosis of equipment. The residual pooling layer is proposed in this work to overcome the drawbacks and provides fast learning. The projected algorithm consists of a residual encoding module and an aggre-gation module for spatial information preserving and order less feature generating. The advantages of the proposed work produce low dimensional data compared to other deep learning approaches. At last, the impact of random noise in the measured PD signal on recognition rate is investigated and addressed.