关键词: 可靠性, 贝叶斯网络, 涡轴发动机, 多目标, 性能预测

Abstract: The turboshaft engine is a kind of highly complex and precise thermal machinery, which is usually used as the power source of helicopter. Its performance directly affects the reliability and safety of flight missions. Turboshaft engines require a very high level of manufacturing. Typically, a qualified turboshaft engine requires two performance parameters: power and critical section temperature. However, in practice, it is difficult to manufacture an engine that can pass a single test run, often requiring several attempts after reassembly. Accurate and effective prediction of turboshaft engine performance can help to anticipate risks, which is important to improve the reliability of the helicopter and ensure the safe completion of the mission, as well as to guide the production process to improve the qualification rate. In recent years, with the development of computer technology and the rise of artificial intelligence and big data, Bayesian networks are increasingly used in the fields of data analysis and machine learning. Bayesian networks are a new type of probabilistic graphical model that combines the advantages of probability theory and graph theory to make the complex systems under study clear and understandable and to quantify, reconstruct and reason about complex systems. In the turboshaft engine performance prediction problem studied in this paper, two performance parameter target variables need to be considered together, which requires simultaneous determination of whether the two target variables of a turboshaft engine to be predicted can satisfy their respective qualifying conditions. Considering the practical research background, this paper combines each of the three multi-objective transformation strategies with Bayesian networks to construct a multi-objective performance prediction model for turboshaft engines, extend and improve the plain Bayesian classifier, and realize the prediction and analysis of multi-objective classification problems in the Bayesian network model. In this paper, based on the collected data of a certain type of turboshaft engines, according to the manufacturer's suggestions, four attribute variables that affect the two performance indicators of turboshaft engine — power and key section temperature are extracted firstly, which are the dimensions of part 1, part 2 and part 3, and the temperature. Then, three multi-objective transformation strategies, target generation, binary association and classifier chain, are introduced and combined with Bayesian networks respectively to construct three multi-objective performance prediction models (TG_NB, BR_NB, CC_NB) based on different strategies for turboshaft engines. Finally, to compare the performance of these multi-objective performance prediction models, while considering the validity of the results. In this paper, decision tree models, logistic regression models, and random forest models (TG_DT, TG_LR, TG_RF) are developed in conjunction with the objectives. The accuracy of each model is compared and validated so that the optimal model can be selected to effectively predict the performance of the turboshaft engine. Since the TG_NB model transforms the original dataset by target generation before modelling, and the new target variable integrates the dependencies of the two old target variables, and performs the most outstandingly in all three performance indicators. Meanwhile, compared with the other two models, TG_NB is presented as a single classifier with a simple and more interpretable model, which improves the modelling speed and prediction speed. Due to the limitation of experimental data, we have not yet considered the possibility of deformation of parts under extreme temperature. In future studies, we will collect more data and explore the effect of external temperature on the part dimensions.

Key words: reliability, Bayesian network, turboshaft engine, multi-objective, performance prediction