Data Driven Prediction Of The Performance Parameters Of Aero-engine Gas Path

Diagnosis and monitoring of aero-engine faults is a significant field of aero-engine research.Timely prevention and diagnosis of aero-engines faults can not only reduce the overall maintenance costs and keep the aircraft running smoothly,but also reduce the occurrence of accidents and effectively avoid human and financial losses.Accurate prediction of real-time performance parameters is important for aero-engine faults diagnosis and monitoring.The components of aero-engine include intake port,compressor,combustion chamber,tail nozzle and so on.Each component is interlocked and constitutes an extremely complex structural system.Once some parts fail,the whole system will not work properly.One of the main manifestations of aero-engine failure is the misalignment of compressor physical speed,so accurate prediction of compressor physical speed is very important for aero-engine faults diagnosis.Traditional prediction model based on mathematical mechanism has strong nonlinearity and uncertainty.With the progress of science and technology,the internal structure of aero-engine becomes more complex,and the accuracy of modeling is more and more difficult to guarantee.Since the 21^st century,in the wave of big data,the industry has formed a trinity of data-driven strategy of acquisition,modeling and decision.In the research of modeling process,machine learning algorithm can effectively avoid the difficulty of solving mechanism model by using only experimental data without knowing the internal structure of the system.In this paper,the test data of a turbofan engine provided by Shenyang Engine Design Institute of Aviation Industry of China are taken as the research object.Three machine learning algorithms based on the test data are used to model and predict the compressor physical speed in the gas path components of an aero-engine in real-time.In the process of model determination,an improved adaptive inertia updating weighted particle swarm optimization algorithm is used to optimize the model parameters.The results show that the improved particle swarm optimization algorithm overcomes the problems of poor adaptability and slow convergence of the traditional particle swarm optimization algorithm.In order to solve the problem of incomplete sample information,a spatial reconstruction algorithm based on sparse auto-encoder is used to change the dimension of the sample,and to describe the working conditions of aero-engine completely.Rolling learning-prediction technology is used to establish the input vector of the model.The results show that the accuracy of model prediction can be improved by using sparse auto-encoder to properly increase the dimension of low-dimensional samples.In the process of verification,three popular machine learning algorithms are adopted:such as random forest,support vector machine and kernel extreme learning machine,and their analysis and derivation are given.Finally,the regression accuracy of the three algorithms is evaluated by grouping validation.The results show that compared with support vector machine,random forest algorithm enhances the generalization of the model and avoids over-fitting.Compared with kernel extreme learning machine,random forest algorithm increases the number of samples,reduces the dimension of samples and enhances the learning ability of the model for high-dimensional data and avoids under-fitting.