Study On Hybrid Intelligent Fault Diagnosis And State Tendency Prediction For Hydroelectric Generator Units

With the continuous progress of energy structure reform in China,developing hydropower energy is strategically significant for building a clean and low-carbon energy system and promoting sustainable development of economic and society.As the core equipment in the energy conversion process of hydropower station,hydroelectric generator units are developing towards complexity,mega-scale,integration and intelligence,which bring the increaasingly inevitable problem of safe and reliable operation.Recently,the concept of “condition based maintenance”(CBM)has been put forward and applied,which provides a feasible idea for effectively reducing the risk of abnormal or fault occurrence and improving the stable operation level of units.Specifically,condition monitoring,fault diagnosis and state tendency prediction are three important parts of CBM.Because of the coupling effect of hydraulic,mechanical and electrical and other interference factors,the monitoring signals of units show obvious non-linear and non-stationary characteristics,and there is a complex mapping relationship between fault and symptoms.For this reason,traditional methods are difficult to satisfy the requirements of accurate analysis of units operation status.Therefore,based on the operation characteristics of hydropower units,the researches of novel state analysis theory and methods have important value of engineering application for improving the accuracy of fault diagnosis and trend prediction and ensuring the stable operation of units.For this purpose,this paper focuses on the key scientific issues in engineering application,such as de-noising analysis and feature extraction of complex non-stationary signals of hydropower units,hybrid fault diagnosis and state trend prediction.Taking advanced theories and methods as research basis,including fast ensemble empirical mode decomposition,deep learning,manifold learning and grey-Markov model,we aim at exploring and improving the defects of existing approaches in the relevant fields.As a result,some methods of non-stationary signal analysis and state feature extraction for hydropower generating units are designed and put forward.A multi-step progressive fault diagnosis system based on health status indetification and feature reduction is developed.Besides,we propose the state tendency prodiction model systemtically combining moving window and grey-Marlov theory.Consequently,the relevant research results provide necessary theoretical basis for the method innovation and technological progress of research paradigms,and have great promotion and application prospects and important engineering application value.The contents that call great attention and positive results of this thesis are illustrated as follows.(1)Aiming at the problem that the vibration signals which can effectively represent the actual operation conditions of hydropower units are easily submerged by strong background noise,a multi-stage de-noising method based on adaptive singular value decomposition(SVD)and intrinsic mode functions(IMFs)reconstruction is proposed.The method combines the advantages of SVD and fast ensemble empirical mode decomposition(FEEMD)in high-frequency noise suppression and adaptive signal processing.Based on the two-stage process of primary filtering and secondary denoising,the original singal can be denoised.In the first stage of filtering,the raw signal is preliminarily decomposed by SVD.Based on the analysis of the effect of effective singular value sequence on filtering performance,an adaptive selection method of singular value based on correlation analysis is developed to effectively filter high-frequency background noise.In the second stage of denoising,the signal after filtering is firstly decomposed by FEEMD.Subsequently,with the selection method of IMFs based on permutation entropy theory,the effective mode functions can be determined and reconstructed to complete the signal denoising.Based on these two-stage processing,the denosing effeciveness can be improved significantly.(2)Considering that it is difficult to accurately the mapping relationship between operation conditions and symptoms under the coupling interference of multi-source excitation,a state feature extraction method for hydropower units based on FEEMD energy entropy and hybrid ensemble auto-encoder is proposed on the basis of statistical analysis approach and deep learning technology.Aiming at the requirement of on-line identification of normal and fault conditions in engineering practice,the role of energy entropy in dynamic change behavior of nonlinear vibration signals when abnormity or fault occurs is analyzed in depth.Based on the advantages of FEEMD in dealing with complex non-linear and non-stationary signals efficiently,a health state feature extraction method of hydropower units based on FEEMD energy entropy is designed to rapidly obtain the energy entropy feature of vibration signals.Futhermore,a hybrid ensemble auto-encoder is innovatively constructed to solve the problem that single energy entropy feature is not enough to identify the specific fault type,which overcomes the limitation of shallow feature learning models,contributes to learn the robust and effective representations and meets the requirement of accurate fault pattern recognition.(3)In order to scientifically establish the feature boundary of normal and fault conditions of hydropower units,an on-line identification system for the health status of units based on the statistical analysis of energy entropy is constructed.With mathematical statistics theory,histortical sample set and the method of energy entropy,the online health status analysis of real-time samples can be realized.Furthermore,for the purpose of suppressing the information redundancy and sensitive features submergence existing in the high-dimensional fault feature space,a new feature reduction algorithm based on the idea of parametric linear mapping mode,namely,modified t-distributed stochastic neighbor embedding(M-tSNE),is proposed innovatively,which has significant advantages in improving fault diagnosis accuracy and reducing computational time.On the basis of the above theories and researches,a multi-step progressive fault diagnosis method based on energy entropy identification and deep features reduction is first proposed,which overcomes the shortcomings of the traditional single-step diagnosis model in the analysis of complex fault conditions,such as high complexity and low accuracy.Specifically,in the framework of diagnosis method,the implementation process can be simplified into two stages,including health status detection and fault type identification,which is in according with the thought of intelligent diagnosis for units in engineering pratice.(4)According to the actual requirement of the state tendency prediction for hydropower units,a trend prediction method based on moving window and Grey-Markov model is proposed due to the predictability of units' state trend.The method fully integrates Grey-Markov forecasting model,the optimization principle of grey background value and moving window based prediction mechanism,and tries to decrease forecasting errors from the perspective of model construction,prediction mechanism and residual modification.Because of the above researches,the forecasting accuracy can be effectively improved.In addition,aiming at the demand of different prediction time scales in engineering applications,three hybrid prediction models,including series-connected prediction model,parallel-connected prediction model and embedded prediction model,are constructed for the state trend forecasting of hydropower units.Meanwhile,the corresponding combination prediction procedures are designed.On balance,the proposed method not only ensures the prediction accuracy but effectively improve the computational efficiency,which can provide necessary guidance for formulating reasonable unit maintenance plans.