Gas Path Fault Diagnosis For Aeroengines Based On Extreme Learning Machine

Aero-engine fault diagnosis,as an important part of the engine health management system,is a key way to realize engine condition-based maintenance,reduce service cost,improve equipment reliability,ensure flight safety,and complete combat mission.In this paper,a turbofan engine is taken as the research object,and the fault diagnosis method of gas path based on ELM is emphatically studied.The fault simulation method and data preprocessing method of gas path components are studied,and the component-level fault model is established by adjusting the component characteristic diagram with health parameters,and the fault simulation of each rotating component in the gas path is realized.A hybrid denoising method based on ELM and wavelet denoising is proposed,which improves the denoising effect of sensor measurement signals.Given the problem that the state estimation of the highpressure turbine is uncertain and difficult to solve due to the lack of high-pressure turbine outlet crosssection sensors in engineering applications,an analytical signal construction method based on the KELM algorithm is proposed,which is suitable for both steady-state and dynamic engine processes.Compared with the traditional data-driven estimation method,this method has a simple structure,small input dimension,few training samples and easy to expand the full envelope,thus enriching the information sources of fault diagnosis.To solve the problem of the effectiveness attenuation of historical training samples when the OSELM algorithm is applied to sensor fault diagnosis,a novel MOS-ELM algorithm considering memory mechanism is proposed,which makes the influence of training samples on online training network attenuate with time.The proposed MOS-ELM algorithm enhances the influence of recent training samples on the online training network,ensures that the online training network is more suitable for the current estimation state,and improves the online estimation accuracy.An online sensor fault diagnosis and signal reconstruction system based on the MOS-ELM algorithm is designed,which improves the accuracy of sensor analytical redundancy estimation,fault detection,and fault signal reconstruction.To solve the problem that the standard KELM algorithm is lack of sparsity in classification application,which leads to time-consuming network training and testing,a novel DKELMs algorithm is proposed.A simplified sub-learning machine considering errors of all training samples is designed,and the decentralized sub-learning machine sets are used instead of the standard KELM network structure.The classification results of each sub-learning machine are fused based on DS evidence theory.Compared with the standard KELM algorithm,the proposed DKELMs greatly improves the sparsity of the network.Especially when dealing with large-scale training samples,it greatly improves the training speed and testing speed.The fault locator and fault pattern recognizer of gas path components based on the DKELMs algorithm are designed,which greatly reduces the training time and test time and greatly improves the real-time performance of component fault diagnosis.To quantitatively estimate the unmeasurable health parameters and surge margin of engine components,an online health state estimator is studied.To solve the problem that the estimation accuracy of the OS-ELM algorithm is easily affected by abnormal training samples,a novel AWOSELM algorithm is proposed.The credibility of different training data blocks acquired sequentially is evaluated adaptively,and a recursive least square method for network parameters with weighted samples is designed.By reducing or even eliminating the adverse effects of abnormal training samples on the online training process,the proposed AWOS-ELM algorithm can improve the accuracy of online regression estimation in the presence of abnormal training samples.A health parameter and surge margin estimator based on the AWOS-ELM algorithm is designed,which improves the online estimation accuracy of the engine health parameter and surge margin under the condition of component performance degradation and sudden failure.The thrust estimation and thrust recovery control method suitable for full-envelope,degradation and fault conditions are studied.Aiming at the lack of sparsity of the KELM algorithm in dealing with regression problems under large-scale training samples,a novel IPKELM algorithm is proposed.The sample selection strategy and elimination strategy for constructing the hidden layer of the network are designed,and the training subset that contributes the most to the optimization goal is selected.Moreover,a simplified recursive solution method is designed for calculating the updated network weights.Without reducing the regression estimation accuracy of the KELM algorithm,the IPKELM algorithm greatly improves the sparsity of the network and greatly shortens the test time.The full envelope thrust estimator based on the IPKELM algorithm is designed.Compared with the KELM algorithm,the proposed IPKELM algorithm can greatly improve the real-time performance of engine thrust estimation under the conditions of steady-state operating point,acceleration and deceleration process,and component degradation in different flight cycles.A thrust recovery controller based on direct thrust control,a turbine outlet temperature limiter,and a maximum speed limiter are designed,so that the degraded or faulty engine can provide the required thrust for the aircraft as much as possible without overtemperature or overshooting.