Performance Degradation Prediction And Fault Diagnosis For Aero-Engines Based On Kernel Adaptive Filtering

Prognostics and health management(PHM)is the key technology to upgrade the maintenance strategy and enhance the safety,reliability and affordability of aero-engines.As the essential parts of PHM system,diagnostic and prognostic techniques focus on fault detection,health assessment and prediction such as to facilitate maintenance decision-making procedure.This dissertation studies the gas path fault diagnosis and performance degradation prediction for aero-engines based on kernel adaptive filtering(KAF)within the data-driven framework.The main research work and innovations are as follows:Given that a single predictive model is unable to capture the underlying dynamic characteristics of complex time series completely,an online learning approach based on KAF with the combination of global and local kernel functions is proposed.In this approach,multiple prediction models utilizing different types of kernel functions are constructed individually and combined in a weighted-sum form to yield an ensemble model.The weights assigned to each prediction model can be tuned in an online mode by Kalman filter according to the prediction error of the ensemble model.Simulation results demonstrates that the proposed ensemble prediction model is capable of approximating the dynamic process of complex systems with higher accuracy than the single model.A prognostic approach combining particle filtering(PF)and KRLS algorithm,which attempts to quantify the uncertainty of degradation process is proposed.An integrated single dimensional health index(HI)that can describe the degradation process of the engine is first derived by fusing a set of sensor signals based on principal component analysis.Then,the constructed HI is treated as hidden degradation state of an aero-engine and SSM is utilized to characterize the evolution of HI over time.The state transition equation that describes the degradation progression is trained by KRLS algorithm and can be updated recursively.Particle filtering is utilized to estimate the degradation state and further offer the possibility distribution function of the remaining useful life(RUL).A case study validates the effectiveness of the proposed prognostic method and the results shows the prediction uncertainty is reduced continuously by incorporating the information acquired from the monitored data.Traditional sparsification methods discards the redundant data to curb the growth of the network generated by KRLS algorithm,which may result in accuracy loss.To solve this problem,a sparse KRLS based on redundant sample modification(RSM-SKRLS)algorithm where redundant data is employed to update the coefficients of the existing network and the relatively more important data are used for updating the network structure is proposed.The novel algorithm is able to achieve a compact network without sacrificing the accuracy performance.Furthermore,a RUL prediction method based on RSMSKRLS algorithm and fuzzy clustering is proposed.Fuzzy clustering algorithm is used to separate the whole training dataset into several subsets.Sub-models for RUL prediction is trained by RSM-SKRLS algorithm based on each training subset such that the diversity of sub-models is guaranteed.The weights of sub-models are set as fuzzy membership values of testing sample to each training subset.Simulation results show that the prediction accuracy of the proposed model is higher than the single model.To address the problem that the network generated by KRLS algorithm keeps growing with the accumulation of training samples,a SDA-KRLS(KRLS with structure dynamic adjustment)algorithm is proposed.The basic idea of this method is that the contribution to minimizing the cost function is employed to quantify the importance of kernel units and acts as a criterion to determine whether a new kernel unit is important enough to be added into the existing network and which old kernel unit needs to be pruned to restrict the network size to a fixed value.By doing so,the network structure is optimized adaptively.The aero-engine gas path component fault pattern classifier that obtains excellent accuracy performance within the whole flight envelope range is designed based on SDA-KRLS algorithm.It is worthy noting that the model complexity is moderate,which facilitates the novel diagnosis method’s application in online settings.The approximation function trained by KRLS is essentially a static feed-forward neural network,which limits its implementation in temporal tasks.In order to overcome this drawback,a KRLS with dynamic reservoir(DR-KRLS)algorithm is proposed by incorporating a dynamic reservoir into KRLS algorithm.The reservoir that belongs to a recurrent neural network is considered as a temporal function that maps the history of time series into a reservoir state space and the spatial relationship between the reservoir state and the target output is approximated by KRLS algorithm.With the utilization of the reservoir,the novel method makes a significant improvement in the capability of modeling dynamic systems.A sensor fault diagnosis system based on DR-KRLS algorithm is developed.The sensor signal prediction model with online adaptability is trained by DR-KRLS algorithm.Given that the predictions of sensor signals can be given in the form of a possibility distribution,the fault threshold is determined automatically according to Chebyshev’s Inequality for sensor fault detection.Simulation results shows that the presented method can detect the sensor drift and bias faults quickly and reconstruct the sensor signals with high accuracy.