Research On Operational Condition Monitoring And Intelligent Fault Diagnosis Of Reversible Pump Turbine

At present,under the goals of realizing the "emission peak" and the "carbon neutrality",China is accelerating the construction of a safe,efficient,clean and low-carbon new energy structure.Pumped storage is one of the most mature large-scale energy storage methods at present,which is very important in the construction of new energy structure and power system.As an important equipment in the pumped hydro energy storage power station(PHESPS),once the operational condition of the reversible pump turbine(RPT)is abnormal,if minor,the stabilities and economic benefits of the power grid will be affected,and if severe,the safety production accident of the PHESPS may be caused.In order to improve the intelligent construction of PHESPSs in China,it is urgent to carry out the operational condition monitoring and intelligent fault diagnosis of the RPT to improve the operational reliability,safety and stability of the unit.Because the RPT is disturbed by complex and multi-faceted coupling factors during the operation,the signal has strong nonlinearity and non-stationarity,and the previous methods are difficult to meet the analysis requirements.To this end,the advanced signal processing,deep learning and artificial intelligence algorithms are combined in the present paper to study the operational condition monitoring,feature extraction of signal,fault diagnosis and recognition,and nonlinear signal prediction of the RPTs.The characteristics of flow-induced vibrations and pressure fluctuations during typical transient processes of the RPT are revealed,and the analysis process and model suitable for the signal de-noising,vibrational fault diagnosis and recognition,and nonlinear vibrational signal prediction of RPT are proposed,which can provide some new ideas for improving the safe operational level of the unit.The main contents and innovations of the research work in the present paper can be summarized as follows:(1)The systematic and comprehensive studies on the characteristics of flow-induced vibrations and pressure fluctuations of the RPTs during the start-up processes of generating and pumping modes are carried out.Firstly,as for the start-up process of generating mode,the amplitudes and physical sources of vibrations at the top cover under different rotational speed conditions are explained in combination with the pressure fluctuation signals at several typical monitoring points.Three rotational speed partitions are proposed to illustrate the characteristics of flow-induced vibrations at the top cover.It is revealed that the vibrations at the top cover are affected by the internal fluid flow of the unit(the vortex ropes in the draft tube and rotor-stator interaction(RSI)in the vaneless space)and the mechanical aspects(the mass imbalance of rotor).At the same time,it is found that the main vibrational excitation sources of the upper and lower brackets of the unit are the mechanical aspects(the mass imbalance of rotor),and the influences of fluid flow on them can be ignored.Secondly,as for the start-up process of pumping mode,the propagation laws and characteristics of the pressure fluctuations of the unit are emphatically studied.The results show that,although the pressure fluctuations in the vaneless space maintain the largest during the whole start-up process of pumping mode,its physical sources are quite different in different time periods.During the stop to synchronous condenser pump(SCP)mode,the strongest pressure fluctuation originates from the resonance phenomenon between the blade passing frequency(BPF)that gradually increases during the speeding-up process of the unit and the random flow of a fixed frequency within a specific rotational speed region.During the SCP to pumping mode,the strongest pressure fluctuation originates from the strong backflows during the pressure construction process of the runner chamber.During the steady pumping mode,due to the large guide vane opening and the shrinkage of the vaneless space,the RSI intensity between the rotational impeller and the stationary guide vanes significantly enhances with both the BPF and RSI frequency dominant.(2)The de-noising of the shaft displacement signal and the shape recognition of the shape orbit of the RPT are investigated.Firstly,in the signal denoising part,the de-noising process of secondary decomposition is proposed in the present paper,which combines the wavelet threshold de-noising method and the variational mode decomposition(VMD)to effectively separate the important characteristic frequencies in the shaft displacement signal.Through setting the threshold of cross-correlation coefficient,the non-noise dominant modes based on VMD are reconstructed so as to realize the de-noising of the shaft displacement signal.This process is used for de-noising of the simulated and measured shaft displacement signals of RPT,and the evaluation indexes of signal de-noising are introduced to verify the effectiveness of the proposed process.Secondly,the clear shape of shaft orbit is obtained by further synthesizing the de-noised shaft displacement signals.Then,combining with the characteristics of pressure fluctuations in the draft tube,the evolutions of three typical shapes of shaft orbits and their permutation entropy analysis results with the loads are expounded.Finally,the special permutation entropy eigenvector is constructed using the mean and maximum values of permutation entropy of the shaft orbit,and it is further input into the intelligent classifier extreme learning machine for training and recognition.In the two cases of randomly setting the proportion of samples of the training set to 50%and 75%,the average recognition rate of 10 times of the model to the samples of test set can reach 96.67%and 100%,respectively,and the calculation time of two testing method is 0.01 s,indicating that good shape recognition rate of shaft orbit and high efficiency are achieved.(3)The complex flow states and vibrational faults of RPT are diagnosed and recognized.Aiming at the low recognition rate of traditional shallow learning algorithms in the classification and recognition of complex signals,the deep belief network(DBN)in the deep learning algorithms is introduced into this field in the present paper.And a fault diagnostic model that combines multidimensional eigenvectors based on VMD and DBN is proposed.In this model,the eigenvector is formed using the multinomial time-domain and frequency-domain eigenvalues of the signal,as well as the energy ratio and permutation entropy of each mode component based on VMD to represent different types of flow states and vibrational faults.And the multidimensional eigenvectors are input into the DBN to complete the diagnosis and recognition of flow states and vibrational faults.Through the recognition and classification of the measured pressure fluctuation signals in the vaneless space and vibrational fault signals at the top cover of the RPT in China,in the two cases of randomly setting the proportion of samples of the training set to 50%and 75%,the average recognition rate of 10 times of the model to the samples of test set of the eigenvectors of pressure fluctuation signals in the vaneless space can reach 97.64%and 98.89%,respectively,and the average recognition rate of 10 times of the model to the samples of test set of the eigenvectors of vibrational fault signals at the top cover can reach 99.57%and 100%,respectively,which are far higher than the traditional shallow learning algorithms.The recognition results also verify the effectiveness of the proposed model in the present paper.(4)The complex nonlinear vibrational signal of the RPT is predicted.Compared with the conventional hydro-turbines,the actual operational conditions of the RPTs are more complex and changeable,and their vibrational signals are more nonlinear and non-stationary.In the present paper,a prediction model that combines the VMD with strong ability of signal decomposition and the long short-term memory(LSTM)neural network in the deep learning algorithm is proposed.Firstly,to minimize the nonlinearity and non-stationarity in the signal,the original vibrational signal is decomposed using VMD to obtain several relatively stationary mode components.Then,after each mode component is normalized,the LSTM neural network is used for modeling and prediction in turn.Finally,the multi-step prediction of the nonlinear vibrational signal of the RPT is successfully realized through denormalization of the predicted data points and signal reconstruction.The model is used to predict the vibrational signal of stator frame during the transient process of shutdown from pumping mode of the RPT in China.The root mean square error,mean absolute error and mean absolute percentage error in the prediction results are 0.8359 μm,0.6670μm and 13.17%,respectively,which are lower than other mainstream prediction models.And the predicted signal waveform is closest to the measured signal waveform,which indicates that the prediction model has certain engineering application value.