Identify incipient faults through similarity comparison with waveform split-recognition framework

Introduction: Incipient faults of distribution networks, if not detected at an early stage, could evolve into permanent faults and result in significant economic losses. It is necessary to detect incipient faults to improve power grid security. However, due to the short duration and unapparent waveform distortion, incipient faults are difficult to identify. In addition, incipient faults usually have a small data volume, which compromises their pattern recognition.

Methods: In this paper, an incipient fault identification method is proposed to address these problems. First, a Waveform Split-Recognition Framework (WSRF) is proposed to provide a two-step process: 1) split waveform into several segments according to cycles, and 2) recognize incipient faults through the similarity of decomposed segments. Second, we design a Similarity Comparison Network (SCN) to learn the waveform by sharing the weights of two Convolution Neural Networks (CNNs), and then calculate the gap between them through a non-linear function in high-dimensional space. Last, disassembled filters are devised to extract features from the waveform.

Results: The method of initializing weights can improve the speed and Accuracy of training, and some existing datasets like MNIST consisting of 250 handwritten numbers from different people are able to provide initial weights to disassembled filters through the adaptive data distribution method. This paper uses field data and simulation data to verify the performance of SCN and WSRF.

Discussion: WSRF can achieve more than 95% Accuracy in identifying incipient faults, which is much higher than three other methods in literature. And this method can achieve good results at different fault locations and different fault times. Which compromises their pattern recognition.