A Deep Learning Framework using Convolution Neural Network for Classification of Impulse Fault Patterns in Transformers with Increased Accuracy D. Dey, B. Chatterjee, S. Dalai, S. Munshi Department of Electrical Engineering Jadavpur University, India and S. Chakravorti NIT, Calicut, India ABSTRACT The paper presents a method using deep learning framework based on convolution neural network (CNN), for identification and localization of faults of transformer winding under impulse test. The results show that the proposed method outperforms the existing methods significantly. The present scheme eliminates the requirement of separate feature extraction and classification algorithms for the analysis of fault current patterns. A part of the proposed network performs feature learning and the other part classifies the features in a supervised manner. The method is computation intensive but capable of achieving very high degree of accuracy; on an average a margin of more than 7% compared to other published literature till date. Index Terms — Insulation diagnosis, impulse test, fault classification, convolution neural network (CNN), deep learning. 1 INTRODUCTION IT is well known that, being one of the most costly and functionally important equipment in electric power system, power transformers need critical care while manufacturing, especially the insulation design. The insulation strength of winding after manufacturing is assessed by performing impulse testing as per IEC -60076 (Part-IV) standards [1]. It is already established by several researchers that due to the complex nature of the transformer insulation system, accurate identification of the faults from the current responses during impulse testing requires extensive human expertise and it is also time consuming. It would therefore be appropriate to develop a method which provides accurate information regarding the type, nature and location of insulation failure. There are many papers already published by various researchers, including the research group of the authors, presenting different methods based on soft computing for this purpose [2-8]. However, there is always a scope of improvement in the accuracy of fault identification of transformers under impulse test. This paper addresses the issue of achieving higher accuracy compared to existing methodologies for identification and localization of faults of transformer winding under impulse test. The data collection and testing procedures followed in this paper are already described in details in earlier papers published by the authors [4-8]. Hence, the emphasis is given only on the new method based on deep learning using convolution neural network (CNN) for impulse fault classification, which outperforms all other existing methodologies by a wide margin, to the best of the knowledge of the authors. To make this new data analysis method for impulse fault classification of transformer as the main focus of the paper, the authors have presented the method and results in the form of a short communication. 2 DETAILED DESCRIPTION OF THE SCHEME The proposed scheme can be understood as a layered network, similar to conventional neural network with specific functions of each layer for supervised learning. In the training phase of the supervised learning using the proposed deep learning framework, the disc-to-disc faults of winding insulation (i.e. insulation failure between the discs or between turns of a transformer winding) are emulated in transformer model at different winding positions and corresponding tank- current waveforms due to applied impulse are recorded. These waveforms for specific fault conditions are taken as input dataset to the Convolution Neural Network based deep learning framework. Within different layers of the network, the various features are learnt from the training data and they are classified into classes of appropriate fault conditions. During testing phase, the trained network is probed with unknown input patterns not used during testing, and the performance of the network to correctly identify the fault class is observed. Details of the CNN topologies can be found in various sources like [9-12]. For most of the applications, these types of networks are used for image analysis. In the present work, it is Manuscript received on 4 May 2017, in final form 11 September 2017, accepted 12 September 2017. Corresponding author: D. Dey. 3894 D. Dey et al.: A Deep Learning Framework using Convolution Neural Network for Classification of Impulse Fault DOI: 10.1109/TDEI.2017.006793