In thermal power plants

the operating environment of coal mills is harsh and faults in coal mills occur frequently. Early warning of the coal mill failures and evaluating remaining usable time of the devices under relevant fault conditions is of great significance to improving the safety of thermal power plants. This paper proposes a novel coal mill fault prediction technology based on semi-supervised learning method. Firstly

the DBSCAN clustering is applied to divide the history data of the coal mill into normal states and fault states.Different class labels are assigned to different states and the remaining usable time of the equipment is counted.Then

the random forest classifier is used to establish a fault prediction model for the coal mill according to the class label sequence

the state of the coal mill operating data is predicted

and the fault category and corresponding equipment remaining usable time are determined. Finally

this method is applied in a thermal power unit by using the actual running data of the coal mills

and the prediction result is compared with that of the early warning models of other classification algorithms such as k-nearest neighbor algorithm

naive Bayes and linear discriminant analysis. The experimental results show that

this method is superior to other early warning models

which can accurately mark the different development stages of the mill faults and give the remaining available time of the mills more accurately.