Noise invariant partial discharge classification based on convolutional neural network. (June 2021)

• Record Type: Journal Article
• Authors: Raymond, Wong Jee Keen; Xin, Chong Wan; Kin, Lai Weng; Illias, Hazlee Azil
• Is Part Of: Measurement. Volume 177(2021)
• Publication Date: 2021-06
• Subjects: Partial discharge -- Feature extraction -- Convolutional neural network -- Pattern recognition -- Electrical insulation -- Machine learning; Weights and measures -- Periodicals; Measurement -- Periodicals; Measurement; Weights and measures; Periodicals; 530.8
• DOI: 10.1016/j.measurement.2021.109220
• Languages: English
• ISSNs: 0263-2241

Abstract

Highlights: This study uses real partial discharge measured from cable joints as the dataset. Partial discharge data was used as an image to train a convolutional neural network. 5 networks & 4 image types were tested to find the best noise invariant classifier. The system was trained using clean data but tested using data overlapped with noise. The proposed system has 16.90% better performance than traditional methods. Abstract: Partial discharge (PD) pattern recognition is essential since it can help to identify the nature of the insulation defect. Numerous machine learning models have been utilized for PD classification applications in the past. However, traditional machine learning models rely on manual feature extraction to obtain training data. They are usually trained using clean PD data measured in the laboratory but are expected to work on-site where some degree of interference or noise is expected. When tested using clean PD data, most machine learning models can easily achieve above 90% accuracy. However, when tested using PD data overlapped with noise, classification accuracy reduces significantly. In this work, the development of a convolutional neural network (CNN)-based PD classification system using transfer learning was proposed. In order to achieve a more practical performance evaluation, a modified 10-fold cross-validation procedure was used where the CNN-based PD classifier was trained using clean PD data but tested using PD data that has been overlapped by noise. The results showed that CNN-based PD classifier was able to achieve up to 16.90% higher classification accuracy under noise contamination compared to traditional machine learning with manual feature extraction. This shows that the proposed method was able to retain higher classification accuracy in the presence of noise.