LSTM-Autoencoder-Based Interpretable Predictive Maintenance Framework for Industrial Systems

The overall efficiency of thermal power plants is largely dependent on the health and performance of its various systems, of which the air pre-heaters (APHs) are one of the most crucial. The existing APH fault diagnosis techniques require labeled data for training, which is challenging to obtain in real-life scenarios. To mitigate the existing drawbacks, this study presents an unsupervised fault detection method based on LSTM-autoencoders, which gives 96% accuracy even in the absence of labeled data. Further, a Random Forest regressor has been utilized to predict the time until failure, with a Mean Absolute Error of 2 days. Explainable AI (XAI) is used to improve the interpretability of the prognosis model by shedding light on feature influences and decision-making. A root cause analysis was also performed to determine how the process parameters affect the fault occurrence in APH. The comprehensive approach to fault detection, prognosis, and root cause analysis contributes significantly to fault management and system optimization, ensuring the reliability and efficiency of industrial processes.