Digitalization and Intelligentization
LUO Yun, LI Zhanguo, FU Longxia, WANG Daoyi, ZHANG Xinzhong, LI Yaohua, CHENG Liang, JIANG Xia
To address the difficulty of fault diagnosis in thermal power equipment with multi-parameter coupling and gradual changes, a fault early warning method based on least absolute shrinkage and selection operator (LASSO) regression feature selection and bidirectional long-short term memory (BiLSTM) multivariate regression prediction was proposed. Taking a coal mill in a 1 000 MW power unit as the research subject, feature parameters such as mill current, outlet pressure, and inlet-outlet differential pressure were selected to represent blockage faults. LASSO regression was employed to select the feature variables, and a multivariate regression prediction model was established based on the BiLSTM algorithm. According to the variation mechanism of the feature parameters during mill blockage and the predicted values of the model, a mill blockage fault index was constructed. Finally, the warning threshold was determined using the kernel density estimation method, enabling mill blockage fault warnings. Actual data analysis shows that when the coal mill is operating normally, the average relative error of the BiLSTM multivariate regression prediction model is 1.13%. Compared with the traditional error back-propagation (BP) neural network and support vector regression (SVR) model, it has higher accuracy and the ability to predict the trend of parameter change. When the coal mill is operating abnormally, this method can detect operational abnormalities earlier than the multivariate state estimation technique (MSET) algorithm model, enabling early fault warning under variable operating conditions of the coal mill.