Civil Aeroengine Gas Path Parameter Deviations Mining And Its Application In Anomaly Detection

Civil aero-engines are major products needed to provide power for the aircraft.Anomaly detection for the aero-engine before each flight mission can avoid great economic and losse.As key parameters to evaluate the health status of gas path components and units,the method to access path parameter deviations depends greatly on origin engine manufacturers,which has become one of the most urgent problems in China’s civil aviation.To achieve the autonomous monitoring of the civil aero-engines in China,the following research work studies the method of mining the gas path parameter deviations and its anomaly detectionFirstly,according to the characteristics of the gas path parameters,a regression model for the aero-engine gas path parameter deviations based on Res-BP neural network is established by combing the Multi-layer Perceptron(MLP)with the Deep Residual Network(Res Nets).Considering that the introduction of unimportant variables into the model will not only increase the training cost,but also reduce the regression effect of the model,this paper uses the Mean Impact Value(MIV)to screen the input variables of the model,and then determines the accurate mapping relationship.The results show that the Res-BPNN regression model can be effectively applied to calculate the aero-engine gas path parameter deviations,and the regression effect is good.Secondly,when the gas path parameter deviations collected by a certain model is insufficient,it is difficult to establish the regression model based on Res-BPNN.Therefore,a knowledge transfer method of the regression model based on adversarial mechanism is proposed.In this method,the gas path features are extracted by stacking Res-BP learning block,then the domain-invariant features between source domain and target domain can be mined by stacking multi-layer multi-kernel maximum mean discrepancy(MK-MMD)adaptation layers and one domain confusion layer.After that the deviations can be obtained,and the deep domain adaptation regression model based on Res-BPNN is established.The results show that this method can further reduce the discrepancy of domain feature distribution,and get better regression effect than the traditional algorithm.Then,coupling effect between the key gas path parameter deviations is considered,it is necessary to comprehensively consider the temporal and spatial correlation of the multi-dimensional timeseries of gas path parameter deviations.Therefore,a method of feature representation and gas path anomaly detection by multi-dimensional timeseries of the gas path parameter deviation is proposed.This method first analyzes the performance law of the gas path abnormal mode in the multi-dimensional timeseries of gas path parameter deviations;secondly,it determines the timeseries sub-sequence division method and extracts the primary features of gas path states based on the gas path abnormal performance law;then,an entropy evaluation method is used to evaluate and optimize the primary features.At last,an isolation forest algorithm is used to realize early warning of abnormal gas path.Finally,research works are combined with the civil airline’s requirements for aero-engine gas path anomaly detection method,the prototype system for civil aero-engine gas path anomaly detection is designed and developed to support for civil airlines to work on aero-engine gas path anomaly detection.