RNN-Autoencoder Approach for Anomaly Detection in Power Plant Predictive Maintenance Systems

Induced Draft Fans (IDF) and Primary Air Fans (PAF) are critical equipment in steam power plants. Anomaly detection based on machine learning models is an approach that is currently being developed for optimization and increasing the effectiveness of predictive maintenance (PDM) as well as increasing the reliability of thermal power plants. The research aims to develop a data-driven model for diagnostic and prognostic equipment, produce accurate predictions using many sensor data taken in real-time from the SCADA system, and design a PDM management framework using an anomaly detection system. This research proposes to use a combined recurrent neural network (RNN)-autoencoder approach as a "normal" behavior model (NBM) with the Mahalanobis Distance (MD) statistical method for the detection of anomalies in power plant equipment. Based on time-series input sensor data, the RNN – autoencoder is utilized to predict the behavior of the equipment in health circumstances. In contrast, the MD is used to determine the distance between the actual parameters of the equipment and its "normal" behavior prediction to determine the anomaly condition. This study examined the performance of the long short-term memory (LSTM) and gated recurrent unit (GRU) models in modeling normal behavior with hyperparameter optimization. The LSTM with the best hyperparameters had a validation loss of 5,690 x 10-4 and a validation accuracy of 93.36 percent, whereas the GRU with the best hyperparameters had a validation loss of 4.484 x 10-4 and a validation accuracy of 93.47 percent. GRU can outperform the LSTM model. The proposed framework can detect anomaly conditions in various cases of IDF and PAF equipment disturbances, both in early warning of equipment failure and during downtime conditions.