A method for identifying the operating state of a hydraulic turbine
The turbine signal is decomposed by CEEMD and Pearson correlation coefficients, and the characteristic vector is constructed after denoising, and combined with the particle swarm algorithm to optimize the limit learning machine, the problem of poor turbine operation status recognition is solved, and efficient turbine status recognition is achieved.
Patent Information
- Application Number
- CN202210047292.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-01-17
- Publication Date
- 2024-12-17
- Estimated Expiration
- 2042-01-17
AI Technical Summary
The prior art is difficult to directly display the relationship between the turbine signal and the operating state, resulting in poor recognition of the turbine operation state.
Complementary set empirical modal decomposition (CEEMD) is used for signal decomposition, combined with Pearson's correlation coefficient, the signal dominant mode and noise dominant mode are judged, and the wavelet decomposition is decomposed again, and the arrangement entropy of the IMF component is calculated to construct the eigenvector, and the A particle swarm algorithm optimization limit learning machine (ELM) based on simulated annealing algorithm is used to establish a turbine operation state recognition model.
Through signal preprocessing and feature extraction, noise interference is reduced, the accuracy of turbine operation status recognition is improved, processing time is saved, and the classification effect of the extreme learning machine is improved.
Smart Images
Figure CN114462448B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of water turbine operating state recognition, and in particular to water turbine operating state recognition based on an improved extreme learning machine. Background Art
[0002] With the increase in electricity consumption by residents, the safety and reliability of the power grid has become an issue that society needs to pay attention to. Hydropower is a common way of power generation. As the core component of hydropower, once a turbine fails, it may lead to immeasurable consequences. Therefore, timely status prediction for turbines to ensure safe and stable operation of turbines is a difficult problem to be solved at present.
[0003] At present, many scholars at home and abroad have made good progress in identifying the operating status of turbines from three aspects: theoretical research, experimental research, and numerical simulation. Liao Weili et al. calculated the turbine model under steady and unsteady conditions, and studied the influence of the opening difference on the pressure pulsation of the entire turbine. Thomas et al. studied the real machine and added a drainage device on the basis of the real machine. The results show that under full load, the pressure pulsation generated by the drainage device is more serious. Under other load conditions, the drainage device can reduce the pressure pulsation that spreads inside the draft tube tailwater pipe. Zhang Fei, Gao Zhongxin et al. used experimental methods to study the pressure pulsation in the tailwater pipe under partial load, analyzed how the main frequency of pressure pulsation changes in the flow channel, and pointed out that the high peak pressure pulsation during the operation of the tailwater pipe under partial load directly affects the smooth operation of the equipment. Liu Xiaobing et al. used large eddy numerical simulation to study the turbine, and used the calculation results of the RANS model as the initial field of unsteady calculation.
[0004] However, these scholars did not directly demonstrate the relationship between turbine signals and turbine operating status. As a new trend, deep learning and machine learning have developed, and turbine operating status recognition methods represented by deep learning and machine learning have achieved good results in turbine operating status recognition.
[0005] It can be seen from the research of most scholars that if you want to directly show the relationship between the turbine signal and the turbine operating state and realize the turbine operating state recognition, you must go through three processes: signal preprocessing (such as wavelet transform, empirical mode decomposition, complementary set empirical mode decomposition), feature extraction, and construction of an operating state recognition model. Therefore, the present invention takes signal preprocessing, feature extraction, and construction of an operating state recognition model as the technical route, and proposes a turbine operating state recognition method. Summary of the invention
[0006] The technical problem solved by the present invention is to propose a method for identifying the operating state of a water turbine.
[0007] The technical solution adopted by the present invention to solve the technical problem is: a method for identifying the operation state of a turbine. The method comprises the following steps: 1) using complementary ensemble empirical mode decomposition (CEEMD) to decompose the turbine signal to obtain IMF components; 2) calculating the similarity between each IMF component and the turbine signal by the Pearson correlation coefficient, judging the dividing point between the signal dominant mode and the noise dominant mode, using wavelet denoising to reduce the noise dominant mode, and reconstructing the denoised signal with the remaining IMF components; 3) using complementary ensemble empirical mode decomposition to decompose the denoised signal, calculating the permutation entropy of the IMF component whose main frequency component is similar to the main frequency component of the turbine signal, and constructing a feature vector; 4) using a particle swarm optimization (SA-PSO) based on a simulated annealing algorithm to optimize an extreme learning machine (ELM) to establish a turbine operation state identification model SA-PSO-ELM.
[0008] A method for identifying the operating state of a water turbine provided by the present invention may also have such a technical feature, wherein step 1) includes the following steps: decomposing the water turbine signal using CEEMD to obtain IMF components.
[0009] A method for identifying the operating state of a turbine provided by the present invention may also have such a technical feature, wherein step 2) includes the following steps: introducing the Pearson correlation coefficient to calculate the similarity between each IMF component and the turbine signal, and determining the dividing point between the signal-dominated mode and the noise-dominated mode; calculating the correlation between the turbine signal and the IMF component through the Pearson correlation coefficient, and drawing a curve of the IMF correlation coefficient distribution as the order of the IMF component changes, and the selection criterion of the dividing point is the position where the similarity curve first changes in the opposite direction, the IMF components at and before the dividing point are noise-dominated modes, and the other IMF components are signal-dominated modes; using wavelet denoising to reduce the noise-dominated mode, and reconstructing it with the IMF components of the signal-dominated mode to obtain a denoised signal.
[0010] A method for identifying the operating state of a turbine provided by the present invention may also have such a technical feature, wherein step 3) includes the following steps: performing CEEMD decomposition on the denoised signal, selecting IMF components whose main frequency components are similar to the main frequency components of the turbine signal; calculating the permutation entropy of each IMF component, and constructing a feature vector of the turbine operating state identification model.
[0011] A method for identifying the operating state of a turbine provided by the present invention may also have such a technical feature, wherein step 4) includes the following steps: setting the parameters of SA-PSO and the parameters of ELM; calculating the fitness according to the classification results of the extreme learning machine; updating the parameters of the extreme learning machine; judging whether the optimization curve converges and whether the error meets the requirements, if it converges and meets the requirements, outputting the classification result, otherwise returning to the second step.
[0012] The beneficial effects of the present invention are as follows: since the turbine signal is decomposed by CEEMD, processing time is saved to a certain extent; since the similarity between each IMF component and the turbine signal is calculated by the Pearson correlation coefficient, the dividing point between the signal dominant mode and the noise dominant mode is judged, and the noise dominant mode is denoised by wavelet denoising, the noise component in the turbine signal is reduced, and the interference of noise on the identification of the turbine operation state is reduced to a certain extent; the denoised signal is decomposed by CEEMD, the permutation entropy of the IMF component whose main frequency component is similar to the main frequency component of the turbine signal is calculated, and the fault characteristics of the turbine signal are extracted; the weight and threshold of ELM are optimized by SA-PSO, and the accuracy of ELM is improved.
[0013] The features and advantages of the present invention will be described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 This is the effect diagram of wavelet threshold denoising.
[0015] Figure 2 This is the flow chart of the particle swarm algorithm.
[0016] Figure 3 Flowchart of the simulated annealing algorithm.
[0017] Figure 4 This is the structural diagram of the extreme learning machine.
[0018] Figure 5 This is the flow chart of the turbine operating state identification model proposed in this paper.
[0019] Figure 6 This is the appearance diagram of the tailwater pipe of the experimental device.
[0020] Figure 7 This is the decomposition result of the complementary set empirical mode decomposition of the signal at the volute inlet under normal conditions.
[0021] Figure 8 is the curve of the Pearson correlation coefficient changing with the order of IMF components.
[0022] Fig. 9This is the spectrum diagram of the signal after denoising the turbine signal at the volute inlet under normal conditions.
[0023] Fig.10 The 10 IMF components are obtained after the denoised signal is subjected to complementary set empirical mode decomposition.
[0024] Fig.11 This is the result diagram of the turbine operating state identification model proposed in this paper.
[0025] Fig.12 Result graphs for extreme learning machine and extreme learning machine optimized by particle swarm algorithm.
[0026] Fig.13 The following is the overall flow chart. DETAILED DESCRIPTION
[0027] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are mostly within the scope of protection of the present invention.
[0028] It should be noted that, under the premise of no conflict, the embodiments of the present invention and the features in the embodiments can be combined with each other.
[0029] The specific implementation platform of the method for identifying the operating status of a water turbine in the present invention is: the operating system is Windows, and the programming software is Matlab2020a.
[0030] Figure 1 This is the effect diagram of wavelet threshold denoising.
[0031] The principle of wavelet threshold denoising is: the coefficients with modulus values greater than and less than a certain threshold in the wavelet coefficients of each frequency band obtained after wavelet decomposition are processed separately, and then the transformed coefficients are subjected to wavelet inverse transform, that is, wavelet reconstruction, to obtain the denoised signal.
[0032] like Figure 1 As shown in the figure, the wavelet soft threshold function is used to denoise the noisy rectangular wave signal and compared with the hard threshold denoising, which shows that the wavelet soft threshold denoising has a better denoising effect.
[0033] Figure 2 This is the flow chart of the particle swarm algorithm.
[0034] like Figure 2As shown in the figure, first, initialize the particle speed and position; then calculate the particle fitness value; obtain the individual optimal value and the global optimal value; update the particle speed and particle position (r 1 , r2 takes a random value between 0 and 1, V id , X id are particle velocity and particle position, respectively. id , P gd are the individual optimal value and global optimal value of the particle respectively, c1 and c2 are learning factors, and W is the inertia weight);
[0035] V id =WV id +c 1 r 1 (P id -X id )+c 2 r 2 (P gd -X id )
[0036] X id =X id +V id
[0037] Finally, determine whether the end condition is met. If so, output the result and end the algorithm. Otherwise, return to the "calculate particle fitness value" step.
[0038] Figure 3 Flowchart of the simulated annealing algorithm.
[0039] like Figure 3 As shown in the figure, first, an initial solution is randomly generated and the objective function of the initial solution is calculated; then, a new solution is generated and the objective function of the new solution is calculated; the difference between the objective function of the new solution and the objective function of the initial solution is calculated; if the difference is not greater than 0, the new solution is accepted, otherwise the new solution is accepted according to the Metropoils criterion; if the maximum number of iterations is reached, it is determined whether the termination condition is met, otherwise it returns to the step of "generating a new solution and calculating the objective function of the new solution"; finally, if the termination condition is met, the algorithm is terminated, otherwise the temperature is slowly lowered, the number of iterations is reset, and the step of "generating a new solution and calculating the objective function of the new solution" is returned.
[0040] Figure 4 This is the structural diagram of the extreme learning machine.
[0041] ELM is a machine learning system based on a feedforward neural network, which has been widely used in fault diagnosis (such as mechanical fault diagnosis). The output of the extreme learning machine is determined by determining the input weights between the input layer and the hidden layer of ELM and the bias of the hidden layer. When ELM is used for classification, it only needs to set the number of neurons in the hidden layer in advance.
[0042] Figure 5 This is the flow chart of the turbine operating state identification model proposed in this paper.
[0043] like Figure 5 As shown, first, the parameters of ELM are initialized; then, the fitness is calculated from the classification results of ELM (the mean square error of the classification results is used as the fitness function); the parameters of ELM are updated; if the termination condition is met, the classification result is output and the algorithm ends, otherwise the temperature parameter is updated and the step of "calculating fitness from the classification results of the extreme learning machine" is returned.
[0044] The two learning factors of PSO are set to 2.8 and 1.3 respectively, the maximum number of iterations is 180, the initial inertia value and the inertia value when iterating to the maximum generation are 0.9 and 0.3 respectively, the maximum speed of the particle is 1, and the annealing constant inertia weight of SA-PSO-ELM is set to 0.5. The number of feature vector inputs is 175, 125 are selected as training samples, and 50 are selected as test samples. The number of neurons in the ELM input layer, the number of neurons in the hidden layer, and the number of neurons in the output layer of the model are 5, 10, and 1 respectively.
[0045] Figure 6 This is the appearance diagram of the tailwater pipe of the experimental device.
[0046] like Figure 6 As shown, with reference to the IEC60193 test standard, the test adopts a unified standard test procedure. The data comes from the tailwater pipe of a turbine of a large electrical machinery research institute in China. The tailwater pipe is used to guide the water flow in and out of the runner. It is the main flow-through component of the turbine and the main part of the turbine pressure pulsation test. The experimental data in this article comes from the signals of the seven main parts of the tailwater pipe of the turbine collected by the pressure sensor: volute inlet, bladeless area +Y, bladeless area -Y, cone pipe +Y0.3D2, cone pipe -Y0.3D2, elbow pipe outside, elbow pipe inside.
[0047] The experiment has four variable parameters for adjustment of working conditions: guide vane opening (mm), unit speed (r / min), flow rate (m 3 / s), device cavitation number. The smaller the device cavitation number, the worse the anti-cavitation performance of the turbine. When different operating parameters are set, the turbine operation state transitions from normal state to warning state, alarm state, abnormal state, and fault state in sequence. In this experiment, the parameter settings of the turbine when collecting turbine signals are shown in Table 1.
[0048]
[0049] Table 1
[0050] Figure 7It is the decomposition result of complementary set empirical mode decomposition of the signal at the volute inlet under normal conditions.
[0051] The signal of the volute inlet under normal conditions is decomposed according to the following formula (X(t), t, r, IMF k are the original signal, time, residual, and the i-th IMF component respectively):
[0052]
[0053] like Figure 7 As shown in the figure, starting from the third-order IMF component, the noise content gradually decreases, and it can be approximately considered that the dividing point between the signal-dominated mode and the noise-dominated mode is 3.
[0054] Figure 8 is the curve of the Pearson correlation coefficient changing with the order of IMF components.
[0055] According to the following formula, the similarity between each IMF component and the turbine signal is calculated and the similarity curve (COV(X,Y), ρ X , Y are covariance, variance of variable X, and variance of variable Y respectively):
[0056]
[0057] like Figure 8 As shown in Figure 3, the emergence of IMF4 causes the change trend of the similarity curve to reverse for the first time, and the cutoff point is 3. It can be seen that the result of the cutoff point estimation of the signal-dominated mode and the noise-dominated mode is correct.
[0058] Fig. 9 This is the spectrum diagram of the signal after denoising the turbine signal at the volute inlet under normal conditions.
[0059] like Fig. 9 As shown in the figure, the frequency of the turbine signal is mainly distributed in the 0-60Hz range. It can be seen that the frequency of the turbine signal is mainly distributed in the low frequency band. When extracting the features of the turbine signal, the features of the frequency of the turbine signal in the 0-60Hz range are mainly extracted.
[0060] Fig.10 The 10 IMF components are obtained after the denoised signal is subjected to complementary set empirical mode decomposition.
[0061] like Fig.10 As shown in FIG. 1 , IMF3-IMF7 components are the main frequency components of the turbine signal. Therefore, after CEEMD decomposition of the signal at the volute inlet in a normal state, the five components IMF3-IMF7 are selected for feature extraction.
[0062] Fig.11 This is the result diagram of the turbine operating state identification model proposed in this paper.
[0063] like Fig.11 As shown in the figure, all 50 test samples are predicted correctly, and the accuracy of the turbine operation status recognition model proposed in this paper is 100%.
[0064] Fig.12 Result graphs for extreme learning machine and extreme learning machine optimized by particle swarm algorithm.
[0065] like Fig.12 As shown in the figure, when the extreme learning machine predicts, samples 13-16 and sample 40 are predicted incorrectly, a total of 5 samples are predicted incorrectly, and the accuracy of the extreme learning machine is 90%; when the extreme learning machine optimized by the particle swarm algorithm predicts, all 50 samples are predicted correctly, and the accuracy of the extreme learning machine optimized by the particle swarm algorithm is 100%.
[0066] In order to quantitatively analyze the relationship between the prediction effect and the mean square error and the coefficient of determination, the same model was used for experiments, and the results are shown in Table 2.
[0067]
[0068] Table 2
[0069] As shown in Table 2, the mean square error of SA-PSO-ELM is 1.3992, which is smaller than that of ELM and PSO-ELM; the determination coefficient of SA-PSO-ELM is 0.92296, which is larger than that of ELM and PSO-ELM, that is, the highest accuracy. SA-PSO-ELM has the best turbine operation status identification effect.
Claims
1. A method for identifying the operating state of a hydraulic turbine based on an improved extreme learning machine comprises the following steps: 1) Perform complementary ensemble empirical mode decomposition (CEEMD) on turbine signals to obtain IMF components; 2) The Pearson correlation coefficient is introduced to calculate the similarity between each IMF component and the turbine signal, and the dividing point between the signal-dominant mode and the noise-dominant mode is determined. The noise-dominant mode is denoised using wavelet denoising, and reconstructed with the remaining IMF components to obtain the denoised signal; 3) Decompose the denoised signal, select the IMF components whose main frequency components are similar to the main frequency components of the turbine signal, calculate the permutation entropy of each IMF component, and construct the feature vector; 4) Using the simulated annealing algorithm-Particleswarm optimization (SA-PSO) to optimize the extreme learning machine (ELM), a turbine operation state recognition model SA-PSO-ELM was established; Step 1: Decompose the turbine signal using CEEMD to obtain the IMF components (X(t), t, r, IMF k are the original signal, time, residual, and the i-th IMF component respectively): Step 2 details are as follows: The Pearson correlation coefficient is introduced to calculate the similarity between each IMF component and the turbine signal, and to determine the dividing point between the signal-dominated mode and the noise-dominated mode; CO V(X,Y), ρX, and ρY are the covariance, the variance of variable X, and the variance of variable Y, respectively: When judging the similarity between the IMF component and the original signal, the original signal is regarded as X and the IMF component is regarded as Y; The curve of IMF correlation coefficient distribution changing with the order of IMF component is drawn, and the correlation between turbine signal and IMF component is calculated by Pearson correlation coefficient. The selection criterion of the demarcation point is the position where the similarity curve first reverses. The IMF components at and before the demarcation point are noise-dominated modes, and the other IMF components are signal-dominated modes. Wavelet denoising is used to reduce the noise-dominant mode, and the denoised signal is reconstructed with the IMF component of the signal-dominant mode.
2. According to the method for identifying the operating state of a hydraulic turbine based on an improved extreme learning machine according to claim 1, the specific details of step 3 are as follows: Perform CEEMD decomposition on the denoised signal and select the IMF components whose main frequency components are similar to those of the turbine signal; Calculate the permutation entropy of each IMF component and construct the eigenvector.
3. According to the method for identifying the operating state of a hydraulic turbine based on an improved extreme learning machine according to claim 1, the specific details of step 4 are as follows: Initialize the parameters of SA-PSO and ELM; The fitness is calculated from the classification results of the extreme learning machine; Update the parameters of the extreme learning machine; Determine whether the optimization curve converges and whether the error meets the requirements; if it converges and meets the requirements, output the classification result, otherwise return to the "calculate fitness from the classification result of the extreme learning machine" step.