Early Fault Detection in Hydroelectric Power Plants Using Bi-LSTM Autoencoders and Predictive Maintenance Frameworks

Abstract

This paper proposes a predictive maintenance framework for power plants leveraging Bidirectional Long Short-Term Memory (Bi-LSTM) Autoencoders validated on data from a real-world power plant. The framework preprocesses sensor data through interpolation, and MinMax scaling. By learning temporal dependencies, the Bi-LSTM Autoencoder detects anomalies through reconstruction errors, enabling robust health monitoring. Health indices are computed using probabilistic methods and enhanced with thresholds and smoothing techniques for real-time reliability. Validation across datasets representing ideal, normal, and faulty conditions demonstrates the framework’s effectiveness in early anomaly detection and system health evaluation. Notably, this framework enables tracking the plant's state and identifying faults up to 34 hours before their occurrence, with an accuracy of 96.8% showcasing the framework’s potential to reduce unplanned outages, optimize costs, and enhance system reliability, making it invaluable for proactive monitoring in industrial applications.