Separation and Recognition of Partial Discharge Signals from Multiple Sources Using Encoder-Decoder based Neural Network
Kotaro Matsuyama, Yasutomo Otake · 2024
Partial Discharge (PD) detection is important method to diagnose electric power equipment and find hidden dangers in it. There may exist multiple PD sources in the equipment, which could seriously affect the accuracy of subsequent analysis. To separate the multi-source signal, clustering method and BSS method were applied. However, these methods have limitations such as requiring multiple sensors and not being usable in pulse overlapped case. In this paper, we introduced Conv-TasNet, which was trained to decompose the mixed signals and reconstruct the original signal. This method requires single sensor and can be available for overlapped case. We have obtained two types of partial discharge signals in laboratory and mixed them to simulate a multi-source PD signal. The signals were generated to simulate the case of overlapped and non-overlapped pulses, respectively. The mixed signals entered Encoder-Decoder based Neural Network and was separated into single-source signal. The separation accuracy of multi-PD signals was evaluated by SI-SNR and these values were 25.22dB and 15.67dB, respectively. Furthermore, the defect types of separated signals were identified using CNN model. The accuracy of classification was almost 100%. These result shows that our proposed method can sufficiently separate multi-source PD signals.