Research On The Performance Prediction Of Commercial Aircraft Engine By Multi-source Information Fusion

Prognostics and Health Management(PHM)supports to the condition based maintenance of complex systetms well for including technical capabilities as health state prediction and synthentic processing of multi-source information.Nowadays,domestic commercial aircraft engine PHM domain has few methods for performance inference of aircraft engines,leading to poor abilities in predicting the on-wing and post-repairing performance of the component and entire system performance of engines.With commercial aircraft engine as the research objective,the problems about engine performance modeling,components degradation diagnosis,whole engine performance prediction,and post-repairing performance predictions of single and multiple engine fleets are studied.Methods of data and knowledge fused modeling of engine performance,multi-variate fused engine performance prediction,multi-source information fused engine post-repairing performance prediction,and tranfer learning based trans-fleet post-repairing performance prediction were proposed.The research contents are arranged according to the order of engine components,whole engine,and single and multiple fleets.Multi-source information from the engine designing knowledge,on-line monitoring data,maintenance workscope,and serving time was included during the inference processes,aiming at providing theorectical and technical supports to the performance prediction business of domestic airlines.The research contents are arranged as follows:As engines are complex multi-component systems,the component unidentifiability problem caused by the deficiency of observation has to be solved by fused modeling process of observed data and physical knowledge.Thus,a method based on the collaborative modeling of components and system,referred to as Trans-Layer Model Learning(TLML),was proposed.A judgment criterion characterizing the informational adequacy of the combined data and physical knowledge modeling was proposed and proved.This criterion clarified the informational adequacy of a fused set of sensor dataset and physical knowledge with regard to certain modeling task,and the proposed method trains component models with a two-layered nested alternative scheme,which can identify component response surfaces with fusion of sensor data and physical knowledge with good prevention of over-fitting.The experiment conducted on a simulation system proved the correctness of the proposed criterion,and that TLML can accurately model system components in observability deficient and noisy environment.To solve the high nonlinearity and filter diverging problems in components performance diagnosis,a method combining TLML engine modeling and modified unscented Kalman filter was proposed.Based on the analysis on engine system and component working properties,this method uses TLML to build high fidelity engine performance model with fusion of sensor observation,physical constraints,and domain knowledge,and uses the acquired model to construct system observation equation.Then,it uses moving window sampling strategy to modify traditional unscented Kalman filter and extends system observation vector,thus prevented filter divergence caused by system observation deficiency.The experiment on engine monitoring data proved that this method has sufficient diagnosis precision on component degradations,and the diagnostic results coincided well with the damage status of engine hardware structures.With regard to the problems of too short prediction range,and poor quantification ability of anticipate performance uncertainty in engine performance prediction process,a health feature parameter prediction method based on the similarity analysis of multivariate trajectory was proposed.This method evaluates the statistical distances of the objective feature trajectory to historical trajectories in high dimensional space with Chi-square and Maximum Likelihood analyses,and then uses the proposed Gaussian component descending order aggregation method to estimate the probability density functions(PDF-s)of the health feature parameters of the objective trajectory at each future time point.As validated by the experiment carried on multiple degradation trajectories from one engine fleet,the proposed method can predict the health trends of engines over hundreds of flight cycles with high precision,meanwhile estimate the PDF-s of the health feature parameters of engines with sufficient precision.To solve the problems of heterogeneous data fusion and the coupling between engine working condition and health state parameters in predicting the post-repairing performance of large engine fleets,a multi-source data fusion method was proposed.This method uses cross entropy increasing factor to regulate the order of parameters in multi-dimensional parameter series,and then uses stacked and convolutional auto-encoders respectively to extract features from engine workscope data and pre-repairing parameter series.Then,the extracted features were combined with engine using time features to build combined features for training extreme gradient boost model to predict engine post-repairing performance and estimate the importance of influence factors.The experiment performed on a whole engine fleet proved that the proposed method can fuse multi-source information from the data of three types,and the prediction results have good precision and robustness.With the samples from large fleet as reference,transfer learning techniques can solve the the problem of sample deficiency in predicting the post-repairing performance of small fleets.To solve the problem of joint transfering of feature extraction model and regression model,which is difficult for classic tranfer learning theory,weak learning machine and feature translation matrix were used as auxiliary models.This method takes the prediction tasks on large fleet and small fleet as source task and target task,and projects the features of target task into the feature space of the source task.Then,the method uses translated target feature and source feature simultaneously to train strong learning machine and predict the post-repairing performance of the small fleet.The experiment carried on two fleets with different scale proved that the proposed method effectively transferred the knowledge from the source task to the target task,thus enhanced the prediction precision of the small fleet.