Anomaly Detection of Power Plant Equipment Using Long Short-Term Memory Based Autoencoder Neural Network

Area of Science:

• Engineering
• Artificial Intelligence
• Data Science

Background:

• Condition-based maintenance of power plant equipment relies heavily on anomaly detection.
• Conventional fixed threshold methods fail to provide early warnings for equipment abnormalities.
• There is a need for robust frameworks capable of learning normal operational patterns.

Purpose of the Study:

• To propose a general anomaly detection framework for power plant equipment.
• To enable early detection of equipment abnormalities for improved maintenance strategies.
• To establish a normal behavior model (NBM) for learning equipment operating patterns.

Main Methods:

• Developed a novel anomaly detection framework utilizing a long short-term memory-based autoencoder (LSTM-AE) network.
• Established a normal behavior model (NBM) to capture spatial and temporal operating variable patterns.
• Employed Mahalanobis distance (MD) for overall residual (OR) analysis and kernel density estimation (KDE) for defining a 99% confidence interval to detect deviations.

Main Results:

• The NBM demonstrated high accuracy and generalizability, with low average root mean square errors (0.026 training, 0.035 test) and mean absolute percentage error (0.027%).
• The proposed framework successfully identified abnormal operations in an induced draft fan case study.
• Real-time monitoring of the overall residual (OR) effectively detected equipment abnormalities.

Conclusions:

• The LSTM-AE-based anomaly detection framework provides an effective solution for early detection of equipment abnormalities in power plants.
• The NBM accurately models normal equipment behavior, enabling reliable anomaly identification.
• This approach significantly enhances condition-based maintenance by allowing timely intervention.