Intrinsic and Extrinsic Learning Framework for Multi-Equipment Incipient Fault Detection and Classification

Incipient faults (IFs) are the precursors of power equipment failures. Due to the low occurrence frequency, IF data are scarce. The scarcity of IFs leads to the difficulty of identifying IFs. Traditional methods lack the ability to learn rich and meaningful representations of IF data, especially under the circumstance of limited IF data. Besides, some methods that involve transforming waveforms into images do not yield advantages in capturing temporal relationships and analyzing waveform distortion. To address these problems, an intelligent framework called INTEL-IFD is developed. In the data process, a weighted IF Gramian matrix expression method is proposed to obtain weighted Gramian images with augmented IF characteristics for further image-based intelligent identification. To address the challenges of limited fault data, a data enhancement tool, including data expansion and image augmentation, is developed. In the modeling process, INTEL-IFD combines the advantages of self-supervised learning and Siamese networks to design a novel network structure. Since the network structure extracts waveform image features from two learning processes, the features of IFs are thoroughly mined. Based on field IF data, the IFs detection and classification accuracy of INTEL-IFD reached 0.9583 and 0.8542, respectively. The representative results demonstrate the effectiveness of INTEL-IFD in identifying IFs in distribution systems.