A Case Study on Turbine Anomaly Detection Using Unsupervised Learning Methods

Turbines, which utilize hot, high-pressure gases and steam to generate electrical power, are subject to mechanical failures during operation due to various factors.If these failures are not promptly addressed, they can lead to severe component damage and even catastrophic accidents.Accordingly, power plants monitor and analyze signals such as vibration, pressure, and temperature to prevent them.With the advancement of high-efficiency and large-scale technologies, the complexity of turbine operations has increased, leading to a higher likelihood of new types of failures.To address this challenge, early warning technologies such as the Multivariate State Estimation Technique (MSET) are widely utilized for anomaly detection in power plants.However, the accuracy of MSET heavily relies on memory matrix obtained from normal operating conditions.If the memory matrix is outdated or does not reflect the latest operating conditions, the model's performance degrades significantly.To prevent this, major facilities are manually reconstructing and using training data.However, they are struggling with increased data analysis time and operational costs.To resolve this issue, this paper proposes a method to automatically update the memory matrix using a discrimination index, which consists of an anomaly index and weight.This approach ensures continuous model reliability, enhances the efficiency of power plant fault monitoring, and minimizes operational costs and maintenance efforts.