EARLY PREDICTION OF TUBE LEAK FAULTS IN PULVERISED COAL-FIRED BOILER USING OPTIMIZED DEEP FEED FORWARD NEURAL NETWORK

Abstract

Boiler tube leaks are a critical issue in coal-fired power plants, leading to costly unplanned shutdowns and reduced operational availability. Traditionally, detection has relied on manual inspection and reactive maintenance, limiting early intervention and increasing downtime. While previous researches have explored predictive maintenance techniques, they often employed simplistic neural network structures and small datasets, leading to unreliable early prediction of leaks. Additionally, many of these models were developed for fluidised-bed coal-fired boilers, with limited success in pulverised coal-fired boilers, and were often constrained by insufficient data from a small number of sensors. This research addresses these challenges by developing a more advanced predictive maintenance approach using deep feed forward neural networks (DFFNN) for accurate early leak prediction in pulverised coal-fired boilers. The research focuses on collecting and analysing relevant sensor data to form a comprehensive dataset, optimising the design of the DFFNN for accurate leak prediction, and evaluating the model's performance in industrial environments against existing methods. Leveraging over 61 million data points from 2012 to 2020, ten DFFNN models were constructed and thoroughly optimised. Seven key hyper-parameters including activation function, optimizer, mini-batch size, loss function, learning rate, and the number of epochs were fine-tuned through extensive hyper-parameter tuning to refine the network architectures, which comprised 6 to 9 hidden layers with up to 512 neurons per layer. Validation and prediction results were emphasised, with the models demonstrating correlations ranging from 82.8% to 99.3% across both training and testing datasets. This research also implemented a multi-model detection method to enhance leak prediction reliability by analysing simultaneous fault detections across 12 fault events from 2012-2020, as well as in the post learning period data from 2021-2022. By requiring consistent threshold crossings from at least three models over three consecutive days, this method effectively minimised false alarms and increased confidence in fault predictions, achieving early detection of confirmed leaks up to 15 days before shutdowns. The findings confirm the effectiveness of this advanced DFFNN-based method, providing a reliable foundation for future improvements in leak prediction systems for pulverised coal-fired boilers. Comparative benchmarking against established methods demonstrated the improved performance of the proposed models, offering advancements in prediction accuracy, lead time, and reliability, contributing to more effective predictive maintenance practices.