Conventional single-ended traveling wave fault location methods, which commonly depend on partial features of traveling wavefronts, may lead to fault location failure, especially for weak-signal faults such as high impedance or zero-crossing faults and close-in faults. To tackle it, this paper presents a fault location method according to the extracted panoramic features of traveling wave full waveform (TWFW) in time-frequency domains. Firstly, two rules of the TWFW features are probed, that is, the wavefront arrival sequence is varied from fault sections, as well as the frequency distribution of wavefronts is strongly related to the precise fault distance. Ergo, the mapping relationships between the TWFW and fault distance are qualitatively confirmed in order to demonstrate the uniqueness of the TWFW subsequently. Next, a LeNet-5-based convolution neural network (CNN) model is constructed to quantitatively evaluate such mapping relationships. In this model, various TWFW features will be extracted to the convolution channels when the CNN parameters are optimized adaptively, and the mapping relations can be formed in light of these sensitive features to estimate the fault distance. Finally, a Grad-CAM visualization method is deployed in the case study, and the accuracy along with the robustness of the proposed method in various fault conditions can be validated consequently.
Read full abstract