Study On Intelligent Fault Diagnosis And State Tendency Prediction Of Hydroelectric Generator Units Based On Time-frequency Analysis And Nonlinear Entropy

With the constant adjustment of China's energy structure,hydropower energy has become an important part of constructing a modern intelligent energy system with multi-energy supplement and supply-demand coordinated characteristics.The vigorous development of hydropower energy plays an important strategic role in promoting green and low-carb energy development,improving energy supply capacity and optimizing energy distribution in China.As the core equipment of energy conversion in hydropower stations,hydroelectric generator units are developing towards the trends of mega-scale in size,complexity and intelligence.The structure of units is becoming more complicated,and the degree of integration is getting higher.The operation safety problems of units are becoming increasingly prominent.Meanwhile,hydroelectric generator unit is a highly complex nonlinear system whose operation is affected by hydro-mechanical-electrical coupling factors.The vibration fault signal of the unit often presents non-stationary and nonlinear characteristics,the conventional state analysis and fault diagnosis methods of hydroelectric generator units have certain theoretical and engineering difficulties in accurately characterizing the complex mapping relationship between faults and symptom,and achieving accurate state assessment and fault diagnosis.Thus,it is urgent to explore new state analysis and fault diagnosis methods to improve the reliability of fault diagnosis and the accuracy of state trend prediction,and provide reasonable guidance for after-sales maintenance decision-making and predictive maintenance.Aiming at the above problems,this paper introduces advanced signal processing methods,non-linear dynamics theories and intelligent fault diagnosis schemes to solve the key issues in engineering applications such as non-stationary signal analysis and feature extraction,fault detection and fault classification,as well as trend prediction of operation state in the hydroelectric generating units.Empirical mode decomposition,multiscale permutation entropy,refined composite multiscale dispersion entropy,max-relevance and min-redundancy maximum feature extraction based on mutual information,extreme learning machine,support vector machine and random forest,etc.,are used as technical means to explore and improve the theory or application deficiencies of existing methods Based on the aforementioned techniques,the improved non-stationary vibration signal analysis of hydroelectric generator units is designed,the nonlinear feature extraction method based on feature space reconstruction and multiscale permutation entropy is proposed,a two-stage comprehensive fault diagnosis strategy for fault detection and classification is presented,and a state trend segmentation prediction model based on diffuse entropy discrimination and extreme learning machine is constructed.These results have expanded the thinking for the innovation of related fields and achieved certain engineering application value.The main contents and achievements of this paper are given as follows:(1)Aiming at the problem of end effect in EMD,the causes of end effect and its influence on signal decomposition results are analyzed in depth.An improved EMD method based on extreme learning machine and mirror extension for restraining the end effect is proposed.This method effectively combines the advantages of extreme learning machine extension and mirror extension in restraining the end effect,and completes the two-stage extension of signal extremum in the process of sifting intrinsic mode functions for each signal decomposition.In the first stage,the extreme point sequence of signal is initially extended by the excellent performance of the extreme learning regression machine in data prediction;in the second stage,the mirror extension method is utilized to extend the extreme point sequence after the initial extension,so as to avoid that the extreme point sequence is insufficient to fit the reality in the decomposition process.Finally,in the simulation signal experiment,the effectiveness of the proposed method is verified by qualitative and quantitative comparison analysis.And it is applied to the analysis of pressure fluctuation signal of draft tube of hydraulic turbine,which proves that it has certain engineering application value.(2)Considering that the mapping relationship between faults and symptoms of hydroelectric generator units under strong background noise and multi-source excitation coupling is difficult to accurately represent,a feature extraction method based on feature space reconstruction and multiscale permutation entropy is proposed to extract nonlinear feature of hydroelectric generator units.Firstly,a feature space reconstruction method based on ensemble empirical mode decomposition(EEMD)is designed to deal with the non-stationary characteristics of unit fault signals,which are prone to mode aliasing and energy leakage.Taking the energy as the measurement criteria,the method reconstructs the intrinsic modal components which are highly sensitive to fault information,and realizes the preliminary extraction of fault features.Meanwhile,combined with the advantages of multiscale permutation entropy,it can detect the dynamic mutation behavior of signals at different scales,and further divides the preliminary extracted reconstructed feature space into multiscale permutation entropy.The full representation of unit fault characteristics is thus completed.The validity of the proposed feature extraction method is conducted by the international standard fault diagnosis cases.Finally,the proposed method is successfully applied to feature extraction of cavitation signals of hydroelectric generator units,which shows its engineering practicability.(3)Aiming at the different derivation and development rules of multi-type faults in the actual operation of hydroelectric generator units,and deeply considering the actual diagnosis requirements of initial fault detection and accurate identification of different fault types and fault degrees,a fine composite multi-scale scatter entropy and fast set empirical model are proposed.Two-stage comprehensive fault diagnosis strategy for fault detection and classification of state decomposition,a two-stage comprehensive fault diagnosis strategy based on refined composite multiscale dispersion entropy and fast ensemble empirical mode decomposition is proposed.In the stage of fault detection,considering the need to distinguish the normal or fault state,the variation law of the sensitivity of fault signal and normal signal to refined composite multiscale dispersion entropy scale factor is explored,and a fault detection method based on composite multiscale dispersion entropy criterion is proposed,which realizes the quick judgment of the unit's health condition.Furthermore,if a fault is detected,considering the shortcomings of single scatter entropy in characterizing different fault types and degrees,the advantages of fast ensemble empirical mode decomposition for efficient processing of complex nonlinear signals and refined composite multiscale scatter entropy for measuring signal uncertainty or irregularity from multiple scales are proposed.A generalized multiscale feature extraction method based on fast ensemble empirical mode decomposition and refined composite multiscale permutation entropy is proposed to realize multi-dimensional and wide-area fault feature extraction.Considering the high-dimensional and redundant features of extracted features,the fault classification model based on the maximal correlation and minimal redundancy of mutual information and random forest classifier is established to realize the accurate classification of different types and degrees of faults.The practical examples of fault diagnosis show that the proposed comprehensive fault diagnosis strategy can effectively detect and classify faults,and provides an effective way to solve the problem of fault diagnosis in engineering practice.(4)With the continuous accumulation of operation time,the inevitable performance degradation and equipment failure of hydroelectric generating units,a subsection prediction model based on dispersion entropy discriminant and extreme learning machine is proposed in combination with the characteristics of unit state trend prediction to predict the trend of unit operation status and to catch abnormal features as early as possible.Considering the different prediction accuracy of extreme learning machine for diverse fluctuation trend signals,this method decomposes the different scale fluctuations or the trends of complex signal sequence representing operation states into different intrinsic mode function components step by step through EEMD.A criterion to measure the signal complexity based on dispersion entropy is established to predict the intrinsic mode function components with similar complexities together,thus to improve the prediction accuracy and reduce the prediction complexity.Eventually,the original state signal is predicted by accumulating the prediction results of each segment.The proposed prediction model has been successfully validated and applied to the trend prediction of vibration of hydroelectric generator units.