Aero Engine Fault Diagnosis And Performance Parameter Prognosis Based On Data-driven Methods

Aero engine health management is the key to ensure flight safety and reduce maintenance costs.In this thesis,the technologies based on extreme learning machine for aero engine gas path fault diagnosis and performance parameter prognosis are studied.The main contents are as follows:Aimed at the ELM's poor stability and generalization performance,the Bayesian extreme learning machine(BELM)is employed by using Bayesian estimation method to calculate weights of ELM.On this basis,coefficient association method and weight sharing technology are applied into BELM for the purpose of improving sparsity.The sparse Bayesian extreme learning machine(SBELM)avoids the direct calculation of generalized inverse and removes redundant nodes in the hidden layers of ELM efficiently.Simulation tests indicates that the proposed SBELM simplify the model structure greatly without much accuracy loss,and the stability and generalization performance are also improved compared with basic ELM.Corresponding to aero engine gas path fault diagnosis under dynamic process,the methodology of combination of multi-layer kernel extreme learning machine(MLKELM)and hidden Markov model is proposed.The MLKELM is for feature extraction and dimension reduction,which is sourced from the auto-encoder technology in neural networks.The data matrix is mapped into a new data representation through multi-layer networks and the extraction performance is superior to original KPCA.HMM is used for pattern classification based on time series,which is suitable for nonlinear stochastic systems like aero engine.Test results show that the proposed fault diagnosis methodology brings a significant improvement both in diagnostic confidence and computational efforts in the application to a turbofan engine fault diagnosis during the dynamic process.Considering that kernel extreme learning machine(KELM)lack of sparsity in the application of aero engine performance parameter prognosis,an improved reduced KELM named DRKELM is proposed.The DRKELM incorporates traditional greedy forward learning algorithm into backward learning algorithm to gain more sparsity and enhance testing time further.The proposed methods produces satisfactory performance of regression with fewer nodes,and reduces regression consuming time from the tests on benchmark dataset.The DR-KELM application to aero engine performance parameter prognosis also demonstrates its superior performance with more sparse structure.