Proactive Maintenance Strategy for Boiler Feed Pumps in Power Plant Industries Using Long Short-Term Memory Autoencoder-Based Anomaly Detection

Abstract

The boiler feed pump (BFP) plays a crucial role in power plants. As part of a prognostic and health management strategy, deep learning frameworks are currently being developed for anomaly detection to improve reliability. This study proposes a data-driven proactive maintenance framework based on anomaly detection using data from the supervisory control and data acquisition (SCADA) system. This research applied an anomaly detection framework built on a long short-term memory autoencoder (LSTM-AE) to monitor several critical operating parameters, including vibration, bearing temperature, lube oil temperature, and motor winding temperature, in a coal-fired steam power plant in Indonesia. Data collected in 2023 and 2024 were used to train and validate the models. The applied LSTM-AE demonstrated strong reconstruction performance for both training and validation datasets, indicating its ability to learn and reproduce the dominant temporal patterns of normal SCADA behavior. The model was able to highlight deviations from learned normal behavior, which were interpreted as early-stage anomalies. These anomaly indications were subsequently verified through field inspection, which revealed mechanical wear in several critical components, including slight pump shaft bending, excessive clearance in the impeller hub bush, impeller wear ring, balance disc, and balance seat. Although a quantitative comparison with conventional fixed-threshold or statistical baseline methods was not conducted in this study, the results demonstrate the practical feasibility of applying LSTM-AE–based anomaly detection for industrial condition monitoring using SCADA data. The proposed framework shows potential to support proactive maintenance activities, such as inspection prioritization and maintenance scheduling.