关键词: 高阻故障, 深度度量学习, 生成对抗网络, 原型网络, 三元组损失, 故障辨识

Abstract: Aiming at the difficulties in high-impedance fault(Hif) identification, such as weak fault characteristics, difficulty in constructing classification boundaries due to sample scarcity, and high misoperation rates under unknown disturbances, a new Hif identification method fusing generative adversarial networks(GAN) and deep metric learning is proposed. This method adopts time-domain data augmentation techniques to randomly perturb and inject noise into the original signal, in order to expand the fault samples and enhance the model's adaptability to noisy environments.first, a GAN adversarial training mechanism is introduced to generate simulated unknown disturbance samples, addressing the incompleteness of negative samples in training data. Second, a deep encoder based on gated recurrent Unit(GRU) is constructed to accurately capture key sequential physical characteristics of Hif, such as the zero-rest phenomenon and asymmetry. Third, an improved triplet loss function combined with an online hard sample mining strategy is adopted for end-to-end training. By incorporating GAN-generated simulated disturbances as strong negative samples, the model is forced to push unknown disturbances away from normal and fault clusters in the feature space. finally, prototypical networks are utilized to adapt to small sample scenarios, establishing standard feature archives for each category. A two-stage threshold discrimination strategy is then employed to achieve precise identification of Hif and effective rejection of unknown anomalies. Simulation verification and comparative analysis demonstrate that the proposed method not only performs well in Hif recognition accuracy,but also achieves zero error response to unknown system disturbances with the enhanced boundary defense capability of GAN, significantly improving the robustness and security of the model in open set environments.

Key words: high-impedance fault, deep metric learning, generative adversarial networks, prototypical networks, triplet loss, fault identification