Research On Fault Diagnosis Of Overpressure And Heating Surface Leakage Of Small Supercharged Boiler Based On Machine Learning

Supercharged boiler is the core equipment of ship steam power system,operating in harsh environment for a long time,which is a equipment with frequent failures.Once it breaks down,if it fails to diagnose and locate it in time,it will directly threaten the safety of the ship’s power system.Conventional fault diagnosis highly relies on the reliability of the expert’s experience,and often has problems such as long time-consuming and low diagnosis efficiency.Therefore,it is of great practical significance to diagnose major faults of marine supercharged boilers.Taking the small supercharged boiler as the research object,the research on the fault diagnosis of the overpressure of the steam drum and the leakage of the heating surface is carried out.First,in view of the difficulty of obtaining actual fault data,a small supercharged boiler simulation model was established through the GSE platform,and the reliability of the simulation model is ensured through stable dynamic verification,then the data on the overpressure of the drum and the leakage of the heating surface are obtained through the fault simulation.Second,in view of the poor quality of fault data caused by noise in actual data,a fault library is established through feature extraction,and the quality of the database is improved through pre-processing such as data filtering and standardization.Then,in order to solve the empirical reliability problem of conventional fault diagnosis,a research on the fault diagnosis of drum overpressure based on three supervised learning algorithms including SVC,RF and MLP is carried out.Finally,in view of the problem of lack of label data in actual data,a semi-supervised SVC model and a semi-supervised RF model are established by combining LPA with SVC and random forest.The fault diagnosis of heating surface leakage based on two kinds of semi-supervised learning is studied,and the diagnostic characteristics of the two under different ratios of missing labels are discussed.The results show that in the aspect of drum overpressure fault diagnosis based on supervised learning,the total diagnostic accuracy and recall rate of SVC,Random Forest and MLP are all higher than 0.91,and the precision rate and F1-score are higher than 0.92 and 0.9,respectively.The overall diagnosis results are relatively satisfactory.Among them,random forest is the best,with an F1-score of 0.972.In terms of boiler heating surface leakage fault diagnosis based on semi-supervised learning,the total accuracy,precision,recall and F1-scores of semi-supervised SVC and semi-supervised random forest are higher than 0.97 and 0.99,respectively.The overall diagnosis result is very good.Semi-supervised random forest performs best,whose F1-score is 0.996.Under the condition that the F1-score is not lower than 0.98,the maximum label missing ratios of semi-supervised SVC and semi-supervised random forest are 44% and 67%,respectively.The research results have certain guiding significance for the safe operation and health management of small booster boilers.