Original Gas Path Parameters-based Gas Path Anomaly Detection Of Civil Aero-engines

The reliability of the civil aero-engine is related to the safety of passenger life and property and the economic benefit of airlines.However,the innovation ability of health management and anomaly detection technology in China airport is insufficient,and the core technology is insufficient.The service provided by the aero-engine manufacturers is excessively dependent.It is very important to develop an independent gas-path anomaly detection method to improve the industrial level and national defense strength.Aero-engine anomaly mostly causes by gas-path,therefore,to realize independent civil aviation engine health management and anomaly detection method,the feature extraction and anomaly detection method of the original gas-path parameters is proposed.Firstly,the assumption is proposed that features that change the slowest with time can represent some inherent attributes of gas-path parameters to some extent.Combined with the slow feature analysis algorithm and the kernel trick,the mixed-kernel slow feature analysis algorithm is implemented by combining different kernel functions into mixed-kernel functions.The algorithm can make full use of the characteristics of the high-dimensional feature space and the advantages of different kernel functions without increasing the computation amount.The original value of the gas path parameters is mapped to the high-dimensional space by using the kernel trick.In the high-dimensional space,the features that change the slowest with time are extracted as some inherent at tributes of the gas path parameters.This inherent attribute is applied to gas-path anomaly detection to reduce the difficulty of anomaly detection and improve the accuracy of anomaly detection.The experimental results show that the features with the slowest time change can indeed represent some inherent attributes of the engine to a certain extent,and the anomaly data are more obvious in these inherent attributes,which is a gas-path feature extraction method with wi de application prospects.From the perspective of time series,this paper firstly studies the advantages and disadvantages of attention mechanism and long short-term memory neural network,and then studies the mapping relationship mining model based on attention mechanism and short and long time memory network.To improve the accuracy of the mining model mapping relations and three times to create a single parameter mapping relationship with the rest of the parameters corresponding to the mining model,the output of certain conditions the theoretical value of the gas circuit parameters,physical meaning similar to baseline gas path parameters,the gas-path doesn’t accord with the theoretical value of the parameters of the actual value can be used as a feature is used in anomaly detection,This characteristic is similar to the physical meaning of gas-path parameter deviation.In order to improve the performance of this feature in anomaly detection,a weakly supervised cross-entropy loss function under extreme sample imbalance was proposed.By combining it with a simple classifier,the learned mapping relationship of the model could be forced to better reflect the difference between normal samples and abnormal samples.Experimental results show that the gas path parameters based on mapping relation mining feature extraction method can accurately learn the mapping relationship between the gas path parameters,extracted features can reflect the health status of the engine,the loss function can eff ectively improve the difference of normal and abnormal samples,the algorithm of feature extraction effect is good,is of strong explanatory,It is a feature extraction method with wide application prospect.The advantages and disadvantages of density peak clustering algorithm are analyzed,the existing improved methods are studied,and the local density definition method based on K-nearest neighbor,the determination method of new outlier threshold,and the new sample allocation strategy is introduced for density peak clustering,and the improved density peak clustering algorithm for anomaly detection is realized.The influence of the k-nearest neighbor number on the clustering effect is analyzed.Combined with the weakly supervised label information,the weakly supervised clustering parameter adjustment strategy and the unsupervised solution are proposed,and the weakly supervised and unsupervised anomaly detection methods for engines with different number s of labels are realized.Feature extraction of the experimental results show that the abov e method can extract the characteristics of abnormal samples from different angles,the anomaly detection method can effectively detect the most known exception,at the same time have less omission and false positives,especially weak supervision cases anomaly detection effect is better,is a kind of method has wide application prospect of anomaly detection.A prototype gas path anomaly detection system for the civil aero-engine is developed.The system can directly extract the characteristics from the orig inal values of gas path parameters of multiple engines.According to the number of data labels,the unsupervised or weakly supervised anomaly detection method is selected to realize the abnormal detection of gas path p arameters.In addition,the system also integrates mature algorithms which have been studied by the research lab.The prototype system can meet the needs of aviation engine health management and anomaly detection to a certain extent and provide theoretical and technical support for it.