STUDY OF THE POSSIBILITIES OF THERMAL CONTROL OF A STEAM BOILER AS PART OF A SHIP POWER PLANT BASED ON MACHINE LEARNING

Abstract

The article discusses approaches to the use of data obtained from instrumentation for performance monitoring, fault detection, etc., which allows to reduce downtime, reduce maintenance, and reduce energy production costs. However, not all instrumentation is sensitive to fault detection. The paper proposes a methodology for selecting the optimal number and composition of a set of instrumentation necessary for effective automated detection of boiler failures based on machine learning for the analysis of steam boiler failures as part of a ship power plant. In this paper, the mRMR dimensionality reduction method and supervised SVM machine learning classifier are used to classify normal and false boiler states. The calculation results show that without implementing the selection of the optimal data set, the SVM-based classifier provides 94.1% machine learning accuracy. After eliminating irrelevant instrumentation, the performance of the machine learning classifier increased in the experiment with optimal instrumentation data. The SVM-based mRMR algorithm provides machine learning accuracy of 97.4%, which is more efficient than the approach without implementing the optimal data set selection.