.png){kind=icon}
.png){kind=icon}
Return
Fault classification in power distribution systems based on limited labeled data using multi-task latent structure learning
.png){kind=icon}
OA
AI
DOI：10.1016/j.scs.2021.103094.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
A significant issue for fault classification in power distribution systems is limited fault data for training classifiers to identify power failure types for remediation. Measurement data from power systems are mostly unlabeled without specified fault types, and labeled data with confirmed fault types are very limited, posing challenges to training classifiers with sufficient accuracy. Existing fault classification methods to deal with small labeled samples explore latent structures between labeled and unlabeled data. However, this line of methods has inaccurate assumptions on the relationship between unlabeled and labeled data and suffers from accuracy loss when dealing with very limited data that are labeled. This paper proposes a novel latent structure learning under a multi-task learning framework to supplement information and deal with the challenges in limited labeled data for fault classification. The proposed method not only takes advantage of the latent structure in unlabeled data that are not effectively utilized but also overcomes the limitations of latent structure learning by preventing classifiers from being overfitted to unlabeled data. The method was validated by an experimental study from distribution-level phasor devices in a hardware-in-the-loop testbed compared with state-of-the-art fault classification algorithms. The method is also demonstrated for the robustness against measurement noise.
Keywords:
Fault diagnosis
Power distribution systems
Distribution-level phasor measurement unit (D-PMU)
Machine learning
Multi-task latent structure learning
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
IF:
12
Papers:
7.8K
Citations:
5.2W
