Leak Event Diagnosis for Power Plants: Generative Anomaly Detection Using Prototypical Networks.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Signal Processing

Background:

• AI-based anomaly detection excels in many applications but struggles with limited or poor-quality training data, especially in hazardous environments.
• Deploying AI systems in facilities with constrained data collection poses significant challenges.

Purpose of the Study:

• To propose Generative Anomaly Detection using Prototypical Networks (GAD-PN) for anomaly detection using limited normal samples.
• To address the challenge of data scarcity in hazardous environments by leveraging generative models and prototypical networks.

Main Methods:

• GAD-PN integrates CycleGAN with Prototypical Networks (PNs) to learn from metadata and simulated data.
• Prototypical Networks classify normal and abnormal samples using learned prototypes from limited normal data.
• CycleGAN is used to generate synthetic anomaly data from normal data, overcoming the difficulty of collecting real anomaly samples.

Main Results:

• The GAD-PN model achieved over 90% leak detection accuracy in pipe leakage scenarios across three different environments, even with limited normal data.
• Demonstrated an average improvement of approximately 30% compared to traditional unsupervised learning models trained on limited datasets.
• The model showed adaptability to various environments with similar anomalous scenarios.

Conclusions:

• GAD-PN offers a robust solution for anomaly detection in data-constrained, hazardous environments.
• The integration of generative models and prototypical networks significantly enhances detection performance.
• This approach provides a viable method for improving safety and efficiency in industrial applications like power plants and smart factories.