Vibration Signal Processing And Intelligent Fault Diagnosis Of Hydroelectric Generator Units

With the continuous adjustment of energy strategy in China, the modern energy system is running into a safe, clean, efficient and low-carbon era. On one hand, conventional hydropower and pumped storage power will have a new development opportunity, with the hydroelectric installment capacity increasing rapidly. On the other hand, to compensate the impact from the integration of new energy resources including wind power, solar power, ocean energy and so on, hydroelectric power will undertake more responsibility for peak load regulation and frequency modulation. Meanwhile, the operation and management of hydroelectric power is expected to be more effective in order to solve the problems of wind abandoning, water abandoning and solar abandoning. As the key equipment of the hydropower transforming process, hydroelectric generator units (HGU) are becoming more and more complex and large-scale, which makes the coupling of different parts stronger, the nonlinearity and non-stationary of vibration signals increasing, and the mapping relation between faults and symptoms more complex. Then, the traditional condition monitoring and fault diagnosing techniques can no longer meet the analysis requirements of HGU in the new situation. To promote the operational stability of HGU, it is urgent to explore novel theories and methods, for instance novel signal analysis and fault diagnosis methods to enhance the analytical precision of condition monitoring and fault diagnosing, with the running data of HGU from the monitoring system. Therefore, in order to solve the key scientific issues in engineering practice of vibration signal analysis and fault diagnosis for HGU, the non-stationary signal process of HGU is taken as the entry point in this paper, and advanced signal process methods are introduced to extract the frequency components of fault signal submerged in the background noise. Then, the feature extracting method based on blind parameter identification of time series model is deeply researched, thus to improve the identifiability of different fault symptoms. Meanwhile, to solve the problem that traditional fault diagnosis methods are easy to over-studying when dealing with uneven and unbalanced distributed samples, a novel fault diagnosis method based on fuzzy adaptive threshold decision rules is proposed by blending the classification capacity of support vector data description and the neighbourhood depiction superiority of k nearest neighbor theory. Furthermore, to find the related omen at early stage of fault progression, a prediction model based on aggregated ensemble empirical mode decomposition and support vector regression is established to forecast the state tendency of HGU precisely. The main contents and innovative achievements in this paper are as follows:(1) Due to the effect of strong noisy background and complicated electromagnetic interference, the frequency components of fault signal is submerged in the background noise, resulting in that the signal gathered from the monitoring system is hard to reflect the real operating state of HGU. To solve this problem, a novel de-noising method for vibration signal of HGU was proposed based on dual decomposition and correlation analysis, which integrates the capability of high frequency noise suppressing by singular value decomposition and the ability of adaptive signal processing by variational mode decomposition, and the method includes two stages of pre-filtering and correlation de-noising. Firstly, the vibration signal is decomposed by singular value decomposition and the mean filtering strategy is employed to select the effective singular values with which the vibration signal is reconstructed, thus the pre-filtering is achieved. Then, the reconstructed signal is decomposed into a collection of components by variational mode decomposition. Subsequently, normalized autocorrelation functions of all mode components are calculated, after which the energy focusability indexes are deduced and applied to select the effective components, with the accumulation of which the de-noising of the vibration signal is accomplished. The proposed method is employed to denoise a simulated noisy vibration signal, and quantitative analysis of correlation coefficient and signal-to-noise ratio shows that the noise mixed in the signal can be filtered effectively. Finally, the successful application in denoising a measured vibration signal from a large hydroelectric generating unit attests the practicality and effectiveness of the proposed method.(2) Influenced by complex changing conditions and multiple coupled incentives, the mapping relation between vibration fault symptoms and fault types of HGU is difficult to represent, restricting the further promoting of fault diagnosis precision. To extract the fault symptoms which can effectively characterize the fault information, a novel feature extraction method for non-stationary faults was proposed based on blind parameter identification of multivariate autoregressive model established with mode components, by combining the ability of non-stationary signal processing by variational mode decomposition with the superiority of time-series model in identifying the parameters blindly of complex dynamic systems. In the proposed method, the original signal is firstly decomposed into a collection of mode components by variational mode decomposition, thus to make up the deficiency of autoregressive models in dealing with nonlinear and non-stationary signals. Then, all the components are applied to construct a multivariate autoregressive model, after which the fault feature vector is constituted with the parameters by identifying the model. And the vector can reflect the changing rules of the mathematical model contained within the dynamic system in respects of structure and parameters, i.e., it can characterise the fault information of HGU well. In the experimental study, the proposed method is used to extract fault features and combined with support vector machine to achieve fault classification. And the experimental result notarizes the effectiveness of the proposed method. Finally, the engineering practical value is verified by feature extraction of cavitation signal from the No.2 unit in Baishan hydropower station.(3) Aiming to solve the problem that the distribution of fault samples may be uneven or unbalanced in fault diagnosis for HGU, this paper proposed a novel fault diagnosis model based on weighted support vector data description and fuzzy adaptive threshold decision rules by combining support vector data description with k nearest neighbor theory. Take the sample weight into consideration, the model achieves the diverse treatment for different samples in the training stage. And the weight is assigned with local density and size-based weight, while local density of each data point is obtained with k nearest neighbor approach and size-based weight of each data point is computed according to the proportion of classes, thus to effectively reduce the effect on support vector data description model from the distribution of samples during the training stage. Meanwhile, in consideration of that the classification precision is low in the fuzzy recognition area of multiclass hyperspheres or when dealing with classification problem with uneven samples, the decision rules based on fuzzy threshold of relative distance and k nearest neighbor are structured to further enhance the classification ability of support vector data description. Numerical simulation and fault diagnosis experiments show that the proposed model is effective and feasible.(4) The running state change of HGU is accompanied with the fault evolving process from quantitative to qualitative variation, while capturing the relevant symptoms timely in early phase of fault development is the precondition for fault diagnosis. For this purpose, the feasibility of state tendency forecast for HGU was studied in this paper, after which a prediction model based on aggregated ensemble empirical mode decomposition and support vector regression was proposed. In the method, the state variable sequence is firstly decomposed into a collection of mode components by ensemble empirical mode decomposition, after which all the components are analyzed in regard to the frequency and energy conditions. Then an aggregation strategy based on the similarity of frequency and energy conditions is structured to obtain the refactored components. Subsequently, the phase space matrix in accordance with each refactored component is deduced by phase space reconstruction and homologous optimal support vector regression forecasting model is established to predict the matrix. Later, the ultimate forecasting values of state tendency can be determined by accumulating the forecasting values of all refactored components. Refactoring the mode components with the proposed aggregation strategy, the prediction time is saved effectively, and the effect of illusive components on the overall prediction result is deduced. Finally, the vibration trend prediction experiments for the No.l unit in Ertan hydropower station are performed and achieve satisfactory application results.