Ensemble-Based Self-Supervised Meta-Learning for Anomaly Detection in Power Systems
TL;DR
Experiments show that integrating meta-features and generated pseudo-labels achieves performance comparable to supervised learning based on the ground truth labels, demonstrating the proposed approach as a promising threshold independent solution for real-world, label-scarce industrial applications.
Abstract
Modern power systems are often equipped with advanced monitoring infrastructure that can collect vast and diverse data streams. Data-driven approaches, such as anomaly detection (AD), can be applied to these data, complementing existing monitoring practices, and supporting adaptive and intelligent decision-making. Industrial anomaly detection is challenging due to the predominance of normal operational data, rare anomalous events, and limited labeled samples. The success of unsupervised AD approaches are often dependent on the proper thresholds, which determine whether or not a sample is anomalous. The often manual process of setting the thresholds can be challenging in ensuring the accuracy, reliability, and robustness of AD in a wide range of operational scenarios. To address this challenge, we propose a self-supervised, meta-learning-based anomaly detection framework that reduces reliance on explicit threshold tuning. The method combines multiple models to exploit complementary strengths and uses reconstruction errors of an ensemble of autoencoders as meta-features to improve detection of subtle anomalies. By generating labels from meta-features, the approach supports learning in label-scarce scenarios. Experiments on different public and industrial datasets show that integrating meta-features and generated pseudo-labels achieves performance comparable to supervised learning based on the ground truth labels. The results demonstrate the proposed approach as a promising threshold independent solution for real-world, label-scarce industrial applications.