Faulted-Phase classification for transmission lines using gradient similarity visualization and cross-domain adaption-based convolutional neural network

Abstract Accurate faulted-phase classification for transmission lines is important to ensure the power systems security, and the machine learning-based methods were widely studied because of their strong generalization ability. However, these methods often require precise marking of fault occurring time. Also, existing methods face the challenges in real-world applications because they are trained using the laboratory samples. In this paper, we propose a novel faulted-phase classification model for the transmission lines. First, the gradient similarities among multi-channel electrical signals are converted to the proposed gradient similarity-based images (GS-images), which are used as the input of neural network. With the aid of this conversion, the fault features are more obvious and there is no need to mark the fault occurring time. Second, cross-domain adaption is introduced to the optimization objective of convolutional neural network (CNN). Through this adaption, the distribution discrepancies of the top-layer features extracted by the neural network between laboratory and real-world fault samples are reduced significantly, thereby increasing the model applicability in diagnosing the real-world faults. To validate the effectiveness of the proposed model, several state-of-the-art faulted-phase classification models are used for comparison. The results show that the proposed model is well-performed in accuracy, noise immunity and real-world applicability.