Research On The Gas Path Anomaly Detection Of Civil Aero-engine Based On Deep Features

Civil aero-engine is the main power source and bleed device of civil aircraft,and its health status is very important for flight safety and economy.To improve the flight safety and economy,it is necessary to detect engine gas path anomalies timely and accurately.At present,the commonly used gas path anomaly detection methods mainly include analytical model methods and data-driven methods.Because of the uncertainly of modeling and the complexity of engine system,the accuracy of gas path anomaly detection methods based on analytical model is difficult to be guaranteed.With the development of monitoring methods and the development of big data technology,data-driven gas path anomaly detection methods have attracted more and more attention.However,due to the large number of engine models,the complex internal laws of monitoring data,numerous monitoring parameters,many interference factors in the monitoring process,the statistical features commonly used in traditional shallow machine learning methods are difficult to characterize engines' health status accurately,resulting in the low accuracy of gas path anomaly detection.Deep features are high-level abstract representations of the original data.They are good at representing the internal laws of data,reducing the dimension of data and the interference of secondary information and noise.Deep learning methods can extract deep features from original data automatically to achieve end-to-end learning and improve the effect of anomaly detection.Effective deep learning usually depends on a large amount of labeled data.However,it usually takes much cost to collect a lot of labeled data.Transfer learning can help deep learning overcome the aboved difficulities,and it can transfer the knowledge of source domain's deep features to target domain's deep features,help extract target domain's deep features or improve their generalization ability.Therefore,based on the theories of deep learning and transfer learning,this paper explores civil aero-engine gas path anomaly detection methods based on deep features.The scientific problems and key technologies of intermittent gas path anomaly detection,persistent gas path anomaly detection,gas path anomaly detection with few samples and gas path anomaly detection without labels are studied respectively.Intermittent gas path anomalies are common anomalies in engine's actual operation,In view of the short duration time of such anomalies,Quick Access Recorder(QAR)data with high sampling frequency is chosen as the data source,an engine intermittent gas path anomaly detection method based on the performance difference of two engines in the same aircraft is proposed.Aiming at the problem that the deep features extraction of QAR data is easily influenced by the variation of working conditions and external environments,the performance parameter difference of two engines in the same aircraft is used as model input.Aiming at the characteristics that QAR data contains some noise,a stacked denoising autoencoder with strong anti-noise ability is selected to perform unsupervised feature learning on the input to extract deep features.Verified by the real QAR data,the proposed method is better than the comparative methods in anomaly detection performance.In addition to intermittent gas path anomalies,another type of common anomalies in actual engine operation is persistent gas path anomalies.Aiming at the characteristics of this type of anomalies with a long duration and increasing deviation from normal,the performance parameter deviation is chosen as the data source,and an engine persistent gas path anomaly detection method that integrates lazy learning with Convolutional Neural Network(CNN)is proposed.Aiming at the problem that the deep features of performance parameter deviation extracted by CNN are not capable of representing local distribution,the lazy learning and CNN is integrated organically,take advantage of lazy learning's good use of local information,and overcome the inherent limitations of CNN model and lazy learning method.Verified by the real gas path performance parameter deviation data,the proposed method is better than single CNN model or lazy learning method in anomaly detection performance.Performance parameter deviation value are usually firmly held by engine manufacturers,once the cooperation with them is interrupted,the airlines will not be able to obtain performance parameter deviation data to detect engine persistent gas path anomalies.To improve the independent ability of the airlines to detect persistent gas path anomalies,the Aircraft Communication Addressing and Reporting System(ACARS)message is chosen as the data source,an engine persistent gas path anomaly detection method based on grouped convolutional denoising autoencoder is proposed.Aiming at the problem that the deep features of high-dimensional ACARS message data are difficult to extract effectively,this method divides all ACARS parameters into several independent variable groups according to their correlations,and extracts the deep features of each variable group's ACARS message data separately,and then fuses them to form the feature vectors.To reduce the demand for a lot of training samples,a denoising autoencoder is used to perform unsupervised feature learning on each variable group's ACARS message data,and the learned parameters are used as the parameters of convolutional feature mapping.Verified by the real ACARS message data,the proposed method is better than the comparative methods in anomaly detection performance,and the number of parameters and time cost is reduced.All the above studies assume that the gas path condition monitoring data are relatively sufficient.However,due to short introduction time or the number is small,resulting in certain types of engines' labeled gas path condition monitoring data are few.In this situation,the extracted deep features are not generalized enough.Therefore,on the basis of the above studies,an engine gas path anomaly detection method based on adaptive fine-tuning method under few samples condition is proposed.Aiming at the problem of large randomness in the source domain selection of the fine-tuning method,an source model engine selection strategy based on similarity measure is proposed;Aiming at the irrationality of the assumption that the deep features of source domain and target domain are the same in the fine-tuning method,a deep features transfer optimization method is proposed to select a layer transfer scheme adaptively;Verified by the the real engine gas path condition monitoring data,the proposed method can significantly improve the performance of engine gas path anomaly detection with few samples.For some models of engines in actual engineering,the gas path condition monitoring data are unlabeled due to the failure to purchase related services or unpredictable reasons.In this situation,the deep features can not be extracted by supervised learning methods.Therefore,on the basis of the above studies,an engine gas path anomaly detection method based on weights constrained adversarial discriminative domain adaptation under unlabeled condition is proposed.According to the analysis of gas path anomaly detection task's characteristics,adversarial discriminative domain adaptation method is selected in the framework of adversarial domain adaptation.Aiming at the problem that the gas path condition monitoring data is few,which makes it difficult for the target model's gas path anomaly detection model's parameters to converge to a good solution robustly,a customized weights constraint term is added to the generative adversarial loss function to improve the convergence of target model's parameter optimization.Real gas path condition monitoring data verifies the correctness of this engine gas path anomaly detection method under unlabeled conditions.This paper introduces the theories and methods of deep learning and transfer learning into civil aero-engine's gas path anomaly detection.It has certain theoretical significance and application value for enriching the technology system of civil aero-engine gas path anomaly detection and improving its operation and maintenance management level.The proposed methods and models also have certain reference value for the anomaly detection research of other complex equipment.