Few-Shot Learning for Anomaly Detection in Gas-Fired Power Plants Using Prototypical Networks

Abstract

This study proposes a few-shot learning (FSL) approach based on prototypical networks for anomaly detection in gas-fired power plants with limited labeled data. A dataset containing 70 labeled operational samples from five types of abnormal conditions was used. The model was trained and evaluated under a 5-way 5-shot experimental setup, with classical machine learning methods such as Random Forest (RF), Support Vector Machine (SVM), K-Nearest Neighbor (KNN), and Logistic Regression (LR) employed as comparative baselines. The proposed FSL model achieved 92.9% accuracy, 91.7% precision, 93.5% recall, and an F1-score of 92.4%, outperforming all baseline models. Experimental results demonstrate that the prototypical network can effectively learn discriminative feature representations under small-sample constraints, offering a lightweight and efficient solution for real-time anomaly detection in industrial systems.