Research on defect diagnosis and fault prediction algorithm of water wall of utility boiler

Traditional detection methods rely on manual experience, so it is difficult to accurately monitor the state of water wall in real time, and the existing fault prediction algorithms are insufficient in accuracy and reliability. Therefore, this paper proposes an LSTM-Prophet hybrid model of multimodal time series feature fusion and attention enhancement, which combines online learning mechanism and uncertainty quantification. The model collects data from multi-source sensors, and realizes real-time diagnosis and prediction of water cooling wall state through the structure of data layer, feature layer, model layer and application layer. In the feature layer, the physical mechanism features and data-driven features are integrated to enhance the model expression ability. In the model layer, the diagnosis module adopts the multi-label classification network based on Attention-LSTM, and the prediction module uses the method based on Prophet-Transformer, and introduces Monte Carlo Dropout to quantify the prediction uncertainty. The experiment takes the boiler water wall system of a 600MW supercritical coal-fired power plant as the object, and uses the historical operation data from 2018 to 2023 and the real-time monitoring data in 2024 to verify it. The results show that the accuracy of this method in defect diagnosis is 94.7%, the macro F1 score is 92.3%, and the reasoning time is 9.5 ms/sample, which is better than the traditional threshold method, SVM and CNN-LSTM. In fault prediction, MAE is 0.12mm, RMSE is 0.15mm, and the interval coverage rate reaches 95.1%, which is better than Prophet and ARIMA. The model also shows good generalization ability of small samples, online learning ability and engineering application advantages, which provides effective support for real-time monitoring and maintenance of water wall of utility boiler.