| In the last decade,the demand for air transport has surged and competition between airlines is fierce,which makes aero-engine safety become an important issue.In the maintenance process of aero-engines,the maintenance time is determined according to the threshold value of engine gas path performance degradation in routine maintenance operations such as washing.If the routine maintenance is too frequent,it will increase the maintenance cost of the engine.If the routine maintenance cycle is too long,the engine may have a safety hazard.In order to achieve efficient and economical daily maintenance,accurate short-term prediction of engine performance is required.In the long-term operation of an engine,accurate maintenance plan and spare engines allocation plan are the guarantee of economic and safe operation,which is supported by accurate engine gas-path performance trend,and the long-term prediction technology of engine performance must guarantee the stability and robustness of its trend.In the process of engine maintenance decision making,accurate fault location is a core problem.Because of the multi-source and timing of signals,fault diagnosis technology is also the key technology in the decision of engine maintenance scheme.The above study provides the single machine maintenance time and maintenance scheme decision,but as a result of fleet size is large,fleet maintenance decision not only needs to ensure the single engine performance,but also needs to consider the economy of fleet maintenance and operation,so a multi-level optimization model is built,which provides comprehensive maintenance decision for the airlines.In addition,due to the high purchasing and holding costs of aero-engines,many airline bases have close collaborative relationship with each other,to guarantee the aircraft mission and reduce the operation cost,and the cooperation network model with a high robustness will become more and more urgent as the number of bases expands and the deployment frequency increases.Aiming at the above problems,this paper intends to focus on the following research work.In order to ensure the accuracy and efficiency of the short-term prediction of civil aero-engine performance,it is necessary to establish a short-term prediction model with strong noise resistance,high model accuracy and efficiency.Due to the high efficiency of the extreme learning machine model,the good generalization performance and the good noise resistance of the random forest model,it is proposed to use the extreme learning machine and the random forest integration method to establish the engine gas path performance short-term prediction model.Considering that working conditions such as Mach number and flight height may affect the value of engine air path performance parameters,the method of working condition separation is proposed to analyze the processing scheme of engine gas path performance prediction under multiple working conditions,and to improve the accuracy of short-term performance prediction model.In order to obtain the main trend of long-term changes in the performance of civil aero-engines in service,the problems of overfitting and underfitting in the existing regression prediction model should be focused on.A fitting sensitivity model of the prediction model to trainning samples is built to control the fitting degree of the regression model to the original samples.The approximation model is built to be insensitive to samples that are far from the clustering center and sensitive to samples that are close to the clustering center,which adjusts the fitting precision of the prediction model to get the main performance trend.Due to the large noise of the gas path performance data of civil aero-engine,strong nonlinear relationship between the original monitoring data and poor separability,in order to improve the accuracy of fault diagnosis,it is necessary to remove the noise in the samples and extract the characteristic information that can represent the essence of faults and normal samples.In addition,the nature of engine failure is small probability events,and small size sample pattern recognition is also a problem that must be solved.In this paper,we will improve the stacked auto-encoder network.By using the denoising function of the stacked auto-encoder network,the inter-layer module to remove noiswe is added,and realize denoising while extracting features.In order to expand the number of fault samples,the function of generating new samples in genetic algorithm evolution is used to expand the fault feature set.In the relatively complete feature space,the accuracy of fault diagnosis will be improved by pattern recognition method.In order to solve the maintenance decision problem of fleet,this paper takes the availability of fleet as the upper optimization goal,the maintenance cost of single engine as the lower optimization goal,and the engine reliability as the main constraint,and establishes a two-layer maintenance decision optimization model.In such large-scale optimization problems,there is an exponential relationship between the number of variables and the size of the model solution space.Therefore,when the number of engines increases,the model solution efficiency is an important index for model evaluation.In addition,since the solution space of such a model is disconnected,all non-inferior solution spaces must be obtained in order to obtain the optimal solution.Inspired by the rapid filling of nutrient solution area during yeast propagation,this paper intends to study the solution method of maintenance decision model based on yeast propagation search.By means of solution migration and multi-dimensional propagation,all disconnected solution spaces are found and the solution search efficiency is improved,which supports the efficient solution of the above model.As mentioned above,in order to reduce the operation cost,it is necessary to establish an efficient and robust engine base cooperative network,among which the difficulty is to ensure that the deployment activities of spare engines will not cause the chain reaction of lacking engine.This paper intends to establish a base cooperative network based on scale-free network,considering the spare engine number,connectivity degree and engine transit time between bases,and establish a connection probability function between base nodes to obtain a base cooperative network with heterogeneous structure.At the same time,according to the real-time number of spare engines,the allocation path of searching the required spare engine is optimized based on the Q learning algorithm,which can find an engine that can be deployed quickly and not make the chain reaction of lacking of engine in other base.These key technologies involve the basic physical knowledge of civil aero-engines and the prediction,pattern recognition and optimization methods based on deep learning,which have great theoretical significance.At the same time,in the application of civil aero-engine intelligent maintenance,these technologies can improve the accuracy of civil aero-engine gas path performance short-term prediction and long-term prediction,improve the accuracy of civil aero-engine gas path fault diagnosis,improve the efficiency of the fleet maintenance decision model,and improve the efficiency of the spare engine management,which has great application significance. |
PDF Full Text Request