Water turbine runner fault diagnosis method based on RLMD optimized by improved masking signal method

Abstract

The invention discloses a water turbine runner fault diagnosis method based on RLMD optimized by an improved masking signal method, and relates to the technical field of water turbine fault diagnosis. An obtained original sound vibration signal is decomposed into a plurality of components by using RLMD, a Spearman correlation coefficient is used to select a component containing effective features, and an approximate entropy is used to remove a component containing a high-frequency interference noise signal. According to the method, the selected components are masked by using the MS method, and the processed PF components are reconstructed to obtain denoised signals, so that the modal aliasing phenomenon is further reduced, the redundant semaphore entering subsequent calculation is greatly reduced, the improved MS-RLMD method is more suitable for water turbine fault diagnosis, the time required for fault detection is shorter, and the recognition precision is higher. The GWO-SVM of the SVM is optimized by using the grey wolf algorithm to recognize the denoised signal, the GWO optimization algorithm is used for parameter optimization, the SVM performance is further improved, and the recognition accuracy of the GWO-SVM is 96.67%.

Description

Technical Field

[0001] The present invention relates to the technical field of hydroturbine fault diagnosis, and particularly relates to a hydroturbine runner fault diagnosis method based on an improved masking signal method-optimized RLMD. Background Art

[0002] As a clean, safe and efficient low-carbon energy source, hydropower energy is widely popular due to its mature technology and pollution-free emissions. However, in recent years, the construction of large hydropower plants (LHPs) has tended to saturate, and many regions prefer small hydropower plants (SHPs) with less ecological damage and lower economic costs. However, due to the influence of the ecological environment and geographical location, compared with LHPs, the hydroturbines of SHPs are more prone to failures.

[0003] As a key device for hydropower generation, the operating state of the hydroturbine directly affects the energy conversion efficiency. Small hydropower plants are usually built on canals or rivers with perennial water flow in remote mountainous areas. There are some fine solid particles in the water flow passing through the hydroturbine that cannot be removed by sedimentation tanks. Due to the diversity of the basins where SHPs are located, the water flow will contain different types of foreign objects, which will affect the normal operation of the hydroturbine.

[0004] Sediment erosion refers to the wear caused by sediment carried in the fluid medium on the surface of hydroturbine components, resulting in a reduction in the power generation efficiency of the hydroturbine, the generation of noise and vibration, and further shortening the service life of the unit and increasing the maintenance cost. Sediment-laden water flow will not only cause corrosion of hydroturbine components, but also cause serious deformation and rapid damage to the metal surface of the components, accompanied by cavitation. At the same time, the abrasion faults caused by external foreign objects rushing into the flow passage and colliding with hydroturbine components will also greatly shorten the life of the hydroturbine. Therefore, it is very important to monitor the sediment erosion faults and foreign object abrasion faults of the hydroturbine in a timely and effective manner.

[0005] The basis for monitoring the hydroturbine is to collect the signals of the unit. With the advantages of non-contact measurement, wide application range, simple equipment, flexible installation and easy signal acquisition of acoustic vibration signal sensors, acoustic vibration signals are used more and more widely, but there are still deficiencies in the field of fault detection and analysis of hydroelectric generating units. The reason is that the sound sources in hydropower stations are complex during actual operation. The collected original acoustic vibration signals not only contain fault signals but also couple mechanical, hydraulic and excitation noises. If the fault signal features are extracted directly from the original acoustic vibration signals while ignoring the huge background noise, it will surely cause large errors. To extract the most representative fault features and then improve the diagnostic classification accuracy, the most crucial step is to denoise the collected original acoustic vibration signals.

[0006] Since the acoustic vibration signals collected under the actual operating conditions of hydro-generator units are often non-stationary and non-linear, traditional signal processing methods are difficult to achieve good results in practical engineering applications. For this reason, the adaptive time-frequency domain decomposition method (LMD) was proposed, which decomposes the original signal into a series of product functions (PFs). Compared with other decomposition methods, LMD can not only retain the effective information of the original signal to a greater extent, but also has good performance in processing non-linear and non-stationary signals. However, LMD does not solve the problems brought by the endpoint effect and mode mixing. Therefore, LMD is optimized from three aspects: boundary conditions, envelope estimation, and screening stop criteria, and the RLMD method (robust local mean decomposition method) is proposed. Compared with LMD, RLMD has better convergence and can obtain relevant frequency modulation and amplitude modulation signals. Therefore, RLMD, as a signal preprocessing method, has been widely used in various fields. Although RLMD improves LMD and eliminates the influence of the endpoint effect, there are still certain defects in practical applications, and the mode mixing effect that is easily generated during the decomposition process has not been solved.

[0007] The masking signal (MS) method was first proposed to suppress the mode mixing phenomenon that appears during the EMD decomposition process. Some scholars have also combined the MS method with the LMD decomposition method and proposed a cyclic iteration method that combines MS and LMD multiple times. It has been verified by simulation signals that this method can effectively suppress the mode mixing effect. However, in the application of the above MS method, there is a one-to-one correspondence between the masking signal of the MS method and the signal decomposition method, which means that different masking signals need to be manually obtained for each component obtained by signal decomposition. For the complex acoustic vibration signals collected during the actual operation of the water turbine, RLMD decomposition often obtains more redundant components. Although the existing MS method can solve the mode mixing problem to a certain extent, it will bring a huge amount of calculation in practical applications. And selecting a suitable algorithm for fault identification after signal processing is also a key step, which directly affects the accuracy of fault identification. Summary of the Invention

[0008] The purpose of the present invention is to provide a water turbine runner fault diagnosis method based on RLMD optimized by an improved masking signal method, so as to solve the problems of mode mixing effect, large amount of calculation, and unsatisfactory fault identification accuracy existing in the prior art when dealing with water turbine fault signals.

[0009] To solve the above technical problems, the present invention adopts the following technical solutions: A water turbine runner fault diagnosis method based on RLMD optimized by an improved masking signal method, which is characterized by including the following steps:

[0010] S1. Collect the acoustic vibration signals of the water turbine when it operates under different working conditions;

[0011] S2. Decompose the acoustic vibration signal into multiple components using the robust local mean decomposition method, select the components containing effective features using the Spearman correlation coefficient method, and remove the components containing high-frequency interference noise signals using approximate entropy to obtain the filtered components;

[0012] S3. After masking the components of the effective features using the masking signal method, reconstruct the components to obtain the denoised signal;

[0013] S4. Randomly classify the denoised signal to obtain a training set and a validation set, input the training set into GWO-SVM for model training to obtain a diagnostic model, and then input the validation set into the diagnostic model for fault identification.

[0014] A further technical solution is that the working conditions in S1 are three types: normal operating conditions, sediment-laden water flow conditions, and foreign object impact conditions.

[0015] A further technical solution is that the specific steps of S2 are as follows:

[0016] S2-1. Perform robust local mean decomposition RLMD

[0017] After discretizing the input continuous signal to obtain x(n), and treating it as the original signal for processing; find the local maximum of x(n) to determine the symmetry point; extend the original signal using the mirror extension method to obtain x(t); assume that the local mean of all extrema is m i , the i-th local amplitude of consecutive extrema is a i , then there is:

[0018]

[0019] where n i , n i+1 are all local extreme points of x(t);

[0020] According to m i and a i , automatically select the appropriate fixed subset size k = odd(s c +3s σ ), and use the moving average algorithm to obtain the local mean function m 11 (t) and the local amplitude function a 11 (t); from m 11 (t) and a 11 (t), obtain the zero-mean signal k 11 (t) and the pure frequency modulation signal s 11 (t), and the formulas are as follows:

[0021] k 11 (t) = x(t) - m 11 (t)

[0022]

[0023] Iterate s 11 (t) as the original signal until the objective function is satisfied and then stop:

[0024] f = RMS[z(n)] + EK[z(n)]

[0025] Where:

[0026]

[0027] z(n) = a(n) - 1, i.e., the zero baseline envelope signal, N s is the total number of signal sampling points, is the arithmetic mean of z(n);

[0028] When the iteration stops, the pure frequency modulation signal function s 1n (t), the envelope signal function a1(t) are obtained, where a1(t) is shown as follows:

[0029]

[0030] Multiply the final amplitude modulation signal by the pure frequency modulation signal function to obtain the first PF component:

[0031] PF1 = a1(t)s 1n (t)

[0032] After separating PF1 from the original signal x(t), a new signal u1(t) is obtained; then take u1(t) as the original signal and repeat the whole process j times until u j (t) is a constant or a monotonic function:

[0033]

[0034] Finally, the original signal x(t) is expressed as:

[0035]

[0036] Thus, the PF components after RLMD decomposition are obtained;

[0037] S2 - 2. Calculate the Spearman correlation coefficient and approximate entropy of the obtained PF components, and screen the PF components according to the magnitudes of the Spearman correlation coefficient and approximate entropy. The screened PF components are used as the components of effective features;

[0038] Calculate the Spearman correlation coefficient of all PF components and the original signal x(t):

[0039]

[0040] where n is the number of data pairs, and d i is the difference level between two variables, and r s is the Spearman rank correlation coefficient. The closer its value is to 0, the lower the correlation between the PF component and the original signal x(t); the closer it is to 1, the stronger the correlation. Sort the PF components in descending order according to the absolute value of the calculated correlation coefficient, and define the last 40% of the components as redundant components weakly correlated with the original signal, and remove these components to reduce the computational amount of subsequent masking processing;

[0041] Calculate the approximate entropy of all PF components:

[0042]

[0043] where m is the embedding dimension, r is the threshold, and Φm(r) and Φm+1(r) are the proximity probabilities of m-dimensional and (m + 1)-dimensional embeddings respectively. The smaller the ApEn, the stronger the regularity of the component; the larger the ApEn, the more complex and chaotic the component. Sort the PF components in descending order according to the corresponding ApEn, and define the first 40% of the components as the components containing a large amount of high-frequency noise. Remove these components to reduce the noise interference.

[0044] A further technical solution is that the specific steps of S3 are as follows:

[0045] S3-1. Use the masking signal method to perform masking processing on the selected components:

[0046] Let the signal to be processed be x(t), and perform Hilbert transform on x(t) to obtain y(t):

[0047]

[0048] where λ is the integration variable;

[0049] Obtain the instantaneous frequency:

[0050]

[0051] where is the phase function;

[0052] By the energy mean method

[0053]

[0054] The format of the masking signal is:

[0055]

[0056] Average instantaneous frequency Average amplitude

[0057] S3-2. Represent the n PF components of the effective features as:

[0058] S3-3. For PF i+ and PF i- Perform RLMD decomposition again respectively, and take the first output component as pf i+ and pf i- . At this time, redefine PF i as IMS-PF i :

[0059]

[0060] Then, perform final signal reconstruction on the obtained component IMS-PF i , i = 1, 2,..., n, and the obtained reconstructed signal is as follows:

[0061]

[0062] A further technical solution is that the specific steps of the GWO-SVM are as follows:

[0063] For a given training set:

[0064] T = {[a1, y1], [a2, y2], …, [a l , y l} ∈ (Ω × Y) l

[0065] where a i ∈ Ω = R n , Ω is called the input space, and each point a i in the input space is composed of n attribute features; y i ∈ Y = {-1, 1}, i = 1, 2, …, l; find a real-valued function g(x) on R n and construct a classification function

[0066] f(x) = sign(g(x));

[0067] Introduce a kernel function, map it to a high-dimensional feature space, and transform the corresponding dual problem into:

[0068]

[0069] Let the optimal solution be Then:

[0070]

[0071] Thus, the optimal classification model obtained is:

[0072]

[0073] The grey wolf algorithm is used to optimize the parameters c and g, and the obtained optimal parameters are used to perform grid optimization and training on SVM, thereby obtaining the final GWO-SVM.

[0074] Compared with the prior art, the beneficial effects of the present invention are:

[0075] 1. RLMD is used to decompose the acquired original acoustic vibration signal into multiple components. The Spearman correlation coefficient is used to select the components containing effective features, and then approximate entropy is used to remove the components containing high-frequency interference noise signals. The MS method is used to perform masking processing on the selected components, and the processed PF components are reconstructed to obtain the denoised signal. By citing the Spearman correlation coefficient and approximate entropy as the evaluation criteria for removing noise components and selecting effective components for masking processing, the mode mixing phenomenon is further alleviated, and the amount of redundant signals entering the subsequent calculation is greatly reduced. The improved MS-RLMD method is more suitable for hydro turbine fault diagnosis, with shorter fault detection time and higher recognition accuracy.

[0076] 2. The GWO-SVM optimized by the grey wolf algorithm is used to identify the denoised signal. SVM is more proficient in processing small sample data and non-linear data. In the GWO-SVM model, the GWO optimization algorithm is used for parameter optimization, further improving the performance of SVM. The recognition accuracy of GWO-SVM is 96.67%. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 It is the process flow chart of the present invention.

[0078] Figure 2 It is the time domain and frequency domain diagrams of the simulation signal in the embodiment.

[0079] Figure 3 It is the RLMD decomposition result diagram of the simulation signal in the embodiment.

[0080] Figure 4 It is the radar diagram of the Spearman correlation coefficient of each component obtained by RLMD decomposition in the embodiment.

[0081] Figure 5 It is the frequency domain diagrams of the RLMD and IMS-RLMD decomposition results in the embodiment.

[0082] Figure 6 It is the LMD decomposition result diagram of the simulation signal in the embodiment.

[0083] Figure 7Decomposition result diagrams of LMD and IMS-LMD in the embodiments.

[0084] Figure 8 Noise reduction evaluation index diagrams of LMD, IMS-LMD, RLMD and IMS-RLMD in the embodiments.

[0085] Figure 9 Time-domain diagram of the sound and vibration signal under normal working conditions.

[0086] Figure 10 Sound signal diagram under the sand-containing working condition.

[0087] Figure 11 Sound signal diagram under the foreign object impact working condition.

[0088] Figure 12 Time-frequency domain image of the original signal.

[0089] Figure 13 RLMD decomposition result diagram of the experimental data.

[0090] Figure 14 Spearman correlation coefficient and approximate entropy of the PF component.

[0091] Figure 15 Frequency-domain diagrams of the RLMD and IMS-RLMD decomposition results of the experimental data.

[0092] Figure 16 Time-domain diagrams of the original signals under three working conditions.

[0093] Figure 17 IMS-RLMD decomposition result and reconstructed signal.

[0094] Figure 18 GWO-SVM recognition result.

[0095] Figure 19 Recognition results of the signals obtained by GWO-SVM for different noise reduction methods.

[0096] Figure 20 Cluster result diagrams of the signals obtained by different noise reduction methods.

[0097] Figure 21 Confusion matrix of different recognition methods.

[0098] Figure 22 Evaluation index diagrams of using different recognition models. Detailed implementation manners

[0099] In order to make the objectives, technical solutions and advantages of the present invention more clear and understandable, the following further details the invention in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0100] Embodiment

[0101] A method for diagnosing faults of a water turbine runner based on an optimized RLMD using an improved masking signal method, comprising the following steps:

[0102] S1. Collect the acoustic vibration signals of the water turbine during operation under different working conditions; the working conditions are three types: normal operation condition, sediment-laden water flow condition, and foreign object impact condition.

[0103] S2. Use the robust local mean decomposition method to decompose the acoustic vibration signal into multiple components, use the Spearman correlation coefficient method to select the components containing effective features, and then use approximate entropy to remove the components containing high-frequency interference noise signals.

[0104] The specific steps are as follows:

[0105] S2-1. Perform robust local mean decomposition RLMD

[0106] The RLMD decomposition process is described as follows: After discretizing the input continuous signal, x(n) is obtained and regarded as the original signal for processing; find the local maximum of x(n) to determine the symmetry point; use the mirror extension method to extend the original signal to obtain x(t); assume that the local mean of all extreme values is m i , the i-th local amplitude of consecutive extreme values is a i , then there are:

[0107]

[0108] where n i , n i+1 are all local extreme points of x(t);

[0109] According to m i and a i automatically select an appropriate fixed subset size k = odd(s c +3s σ ), and use the moving average algorithm to obtain the local mean function m 11 (t) and the local amplitude function a 11 (t); from m 11 (t) and a 11 (t) to obtain the zero-mean signal k 11 (t) and the pure frequency modulation signal s 11 (t), the formulas are as follows:

[0110] k11 r(t) = x(t) - m 11 (t)

[0111]

[0112] Take s 11 (t) as the original signal and iterate until the objective function is satisfied and stop:

[0113] f = RMS[z(n)] + EK[z(n)]

[0114] Where:

[0115]

[0116] z(n) = a(n) - 1, that is, the zero baseline envelope signal, N s is the total number of signal sampling points, is the arithmetic mean of z(n);

[0117] When the iteration stops, the pure frequency modulation signal function s 1n (t), the envelope signal function a1(t), where a1(t) is shown as follows:

[0118]

[0119] Multiply the final amplitude modulation signal by the pure frequency modulation signal function to obtain the first PF component:

[0120] PF1 = a1(t)s 1n (t)

[0121] After separating PF1 from the original signal x(t), a new signal u1(t) is obtained; then take u1(t) as the original signal and repeat the whole process j times until u j (t) is a constant or a monotonic function:

[0122]

[0123] Finally, the original signal x(t) is expressed as:

[0124]

[0125] Thus, the PF components after RLMD decomposition are obtained;

[0126] S2 - 2 Due to the influence of background noise, the RLMD decomposition result contains a large number of redundant components. Calculate the Spearman correlation coefficient and approximate entropy of the calculated PF components, and screen the PF components according to the size of the Spearman correlation coefficient and approximate entropy.

[0127] Calculate the Spearman correlation coefficient between all PF components and the original signal x(t):

[0128]

[0129] where n is the number of data pairs, and d i is the difference rank between two variables, and r s is the Spearman rank correlation coefficient. The closer its value is to 0, the lower the correlation between the PF component and the original signal x(t); the closer it is to 1, the stronger the correlation. Sort the PF components in descending order according to the absolute value of the calculated correlation coefficient, and define the last 40% of the components as redundant components weakly correlated with the original signal, and remove these components to reduce the computational complexity of subsequent masking processing;

[0130] Calculate the approximate entropy of all PF components:

[0131]

[0132] where m is the embedding dimension, r is the threshold, and Φm(r) and Φm+1(r) are the neighboring probabilities of m-dimensional and (m + 1)-dimensional embeddings respectively. The smaller the ApEn, the stronger the regularity of the component; the larger the ApEn, the more complex and chaotic the component. Sort the PF components in descending order according to the corresponding ApEn, and define the first 40% of the components as components containing a large amount of high-frequency noise. Remove these components to reduce noise interference.

[0133] S3. After masking the selected components using the masking signal method, reconstruct the components to obtain the denoised signal.

[0134] The specific steps are as follows:

[0135] S3-1. Mask the components of effective features using the masking signal method: The principle of the masking signal method (MS) is: use the method of taking the average after addition and subtraction to reduce the error caused by the moving average method, and solve the problems of mode mixing and noise.

[0136] Let the signal to be processed be x(t), and after performing the Hilbert transform on x(t), obtain y(t):

[0137]

[0138] where λ is the integration variable;

[0139] Obtain the instantaneous frequency:

[0140]

[0141] where is the phase function;

[0142] By the energy mean method

[0143]

[0144] The mask signal format is:

[0145]

[0146] Average instantaneous frequency Average amplitude

[0147] S3-2. Represent the n PF components of the effective features as:

[0148] S3-3. For PF i+ and PF i- Perform RLMD decomposition again respectively, and take the first output component as pf i+ and pf i- At this time, redefine PF i as IMS-PF i :

[0149]

[0150] Then, perform the final signal reconstruction on the obtained components IMS-PF i , i = 1, 2,..., n, and the obtained reconstructed signal is as follows:

[0151]

[0152] S4. Randomly classify the denoised signal to obtain a training set and a validation set. Input the training set into GWO-SVM for model training to obtain a diagnostic model, and then input the validation set into the diagnostic model for fault identification.

[0153] The specific steps of the GWO-SVM are as follows:

[0154] In SVM, considering both the code running speed and the recognition accuracy, the indirect one-versus-one method is adopted in this paper to construct a multi-class support vector machine; for the given training set:

[0155] T = {[a1, y1], [a2, y2], …, [a l , y l} ∈ (Ω × Y) l

[0156] where a i ∈ Ω = R n , Ω is called the input space, and each point a i in the input space is composed of n attribute features; y i∈Y = {-1, 1}, i = 1, 2, …, l; Search for R n A real-valued function g(x) on R is constructed to build the classification function

[0157] f(x) = sign(g(x));

[0158] The kernel function is introduced and mapped to a high-dimensional feature space, and the corresponding dual problem becomes:

[0159]

[0160] Let the optimal solution be Then:

[0161]

[0162] Thus, the optimal classification model is obtained as:

[0163]

[0164] The one-against-one multi-class support vector machine is suitable for small sample datasets, greatly reducing the matrix storage pressure and shortening the calculation time. However, the selection of the support vector machine kernel function has a great impact on its performance. Therefore, the grey wolf optimization algorithm is used to optimize the support vector machine, and this algorithm realizes the optimization search by simulating the behavior followed by grey wolves during group hunting.

[0165] In the GWO algorithm, α is the optimal solution, followed by β and δ, and the solutions for all other cases are grouped into ω, and the priority of the solutions of α, β, δ, and ω is from high to low. The optimization process of the GWO algorithm includes not only the internal hierarchical system of the wolf pack but also the hunting steps, such as tracking, surrounding, and attacking the prey, etc.

[0166] Surrounding the prey:

[0167]

[0168] In the formula, A and C are cooperation coefficient vectors, t is the current iteration number, X p is the current position vector of the prey, X(t) represents the position vector of the wolf pack at the iteration number t, denotes the Hadamard product, r1 and r2 are random vectors in [0, 1], and a linearly decreases from 2 to 0 during the iteration process.

[0169] Hunting, that is, during the iteration process, find the optimal α, β, δ and their positions, and then use ω to further locate. The mathematical model of this behavior is as follows:

[0170]

[0171] The step size and direction of the ω individuals in the wolf pack moving towards α, β, δ:

[0172]

[0173] Final position of ω:

[0174]

[0175] Attack the prey. When the prey stops moving, the grey wolf ends the hunting process by attacking. To simulate the process of approaching the prey, the value of a gradually decreases, so the fluctuation range of A also decreases accordingly. That is, during the iteration process, when the value of a linearly decreases from 2 to 0, the corresponding value of A also changes within [-a, a].

[0176] When the value of A is within the interval, the next position of the grey wolf can be at any position between its current position and the prey position. When -1 < A < 1, the wolf pack attacks the prey (falls into local optimum). When A > 1 or A < -1, the grey wolf moves away from the prey to explore other areas, so as to find the global optimum. At the same time, GWO has a component C (random weight) that is beneficial to global exploration, which can avoid falling into local optimum.

[0177] Use the above grey wolf algorithm to optimize the parameters c and g, and use the obtained optimal parameters to perform grid optimization and training on SVM, so as to obtain the final GWO-SVM.

[0178] To verify that the improved MS-RLMD (denoted as: IMS-RLMD) suppresses the mode mixing phenomenon in the presence of noise, a simulation signal is used for verification. The corresponding frequencies in this simulation signal are 30Hz, 100Hz, and 240Hz respectively. The time-domain and frequency-domain images are as Figure 2 shown. The signal formula is as follows: x(t) = sin(60πt) + sin(200πt) + sin(480πt) + n(t), where n(t) is Gaussian white noise with a signal-to-noise ratio of 5.

[0179] Use the RLMD method to decompose the simulation signal x(t), and the time-domain and frequency-domain of the obtained components are as Figure 3 shown. As Figure 3As shown in (b), the component PF1 obtained by RLMD decomposition contains two frequencies of 240Hz and 100Hz at the same time, and the component PF2 contains two frequencies of 100Hz and 30Hz, resulting in modal aliasing. The component PF3 has a better effect, with a higher peak only at 30Hz. However, due to noise interference, some false components appear in components PF4, PF5 and PF6. Therefore, the MS method is used to alleviate this effect, but the existing MS method calculates all components until the conditions are met, with too many iterations and too much calculation. On this basis, an improved MS method is proposed. By using the Spearman correlation coefficient and approximate entropy, n effective components with greater correlation with the original signal x(t) are selected for research, thereby reducing the amount of calculation.

[0180] Calculate the Spearman correlation coefficient of each component and display it using a radar chart, such as Figure 4 As shown. Obviously, the correlation coefficients of each component decrease in turn, among which PF4 decreases more than PF3, and is retained at 60%. The Spearman correlation coefficients of PF4, PF5 and Residual are close to 0, and the correlation with the original signal is very weak. PF4, PF5 and Residual are judged to be redundant components and removed. At the same time, the approximate entropy of PF4, PF5 and Residual is greater than PF2 and PF3, indicating that PF4, PF5 and Residual contain a large number of irrelevant noise signals. Therefore, the first three components PF1, PF2 and PF3 are selected for masking signal processing. When the average instantaneous frequency of the masking signal is selected as 30Hz to process PF1, PF2 and PF3, the three components can be processed accordingly. When the average instantaneous frequency is selected as 240Hz, the processing effect on PF1 is obviously unreasonable. When 100Hz is selected, the processing effect of PF2 is better, and the effect of PF1 is worse. Therefore, the average instantaneous frequency of the mask signal should be selected according to the actual situation. After many tests, it was found that the average instantaneous frequency was 30Hz for PF1, 100Hz for PF2, and 30Hz for PF3 to achieve better processing results. The processing results are as follows: Figure 5 (b) as shown.

[0181] The first three simulation signal components contain three different frequencies. After IMS processing, although there is still a 100Hz frequency component in PF1, it has been greatly reduced compared to before processing. Similarly, the 30Hz frequency component in PF2 has also been significantly reduced. From the above, it can be determined that the mask signal method has a certain inhibitory effect on the modal aliasing phenomenon generated in RLMD decomposition.

[0182] In order to verify the advantages of the proposed method, local mean decomposition (LMD) is introduced for comparison. Figure 6As shown, during the LMD decomposition process, serious endpoint effects occurred. Moreover, during the iteration process, the deformation at the endpoints continuously spread inwards, causing serious deformation at the endpoints of each component. Meanwhile, compared with RLMD, the LMD decomposition method has more iteration times and takes longer.

[0183] As Figure 6 (b) shows that LMD decomposes the simulation signal x(t) into 7 components, and the number of components increases compared with RLMD. Meanwhile, the component PF1 contains frequencies of 100Hz and 240Hz, and the components PF2 and PF3 simultaneously contain frequencies of 30Hz, 100Hz, and 240Hz, resulting in a more serious modal aliasing effect compared with RLMD. Therefore, it can be judged that the RLMD method is superior to the LMD decomposition method.

[0184] According to the Spearman correlation coefficient and approximate entropy of each component, the first 4 PF components are selected and processed using the masking signal method. The processing results are as Figure 7 (b) shows. From the figure, although the IMS method suppresses the frequencies of 100Hz in component PF2 and 30Hz in component PF3 to a certain extent, it reduces the frequency of 240Hz in PF1. Meanwhile, by comparing the PF3 components in Figure 7 (a) and (b), it can be found that the false signals generated by the LMD decomposition are not alleviated or eliminated by the IMS method.

[0185] To further quantitatively analyze the denoising effect, the root mean square error (RMSE), signal-to-noise ratio (SNR), improved signal-to-noise ratio (ISNR), and correlation coefficient (CC) are selected to evaluate the reconstructed signals obtained after denoising using different methods, which are expressed as follows:

[0186]

[0187] Among them, n is the signal length, x(i) is the original signal, f(i) is the noise signal, is the signal after denoising, is the mean value of x(i), is mean value.

[0188] RMSE is one of the commonly used indicators to evaluate the signal denoising effect. The smaller the RMSE value, the smaller the difference between the denoised signal and the original signal, and the better the denoising effect. The signal-to-noise ratio is one of the most important indicators to evaluate the signal quality, which represents the ratio of the signal energy to the noise energy. The larger the SNR value, the lower the percentage of noise in the signal, and the better the signal quality. ISNR is also one of the important indicators to evaluate the signal denoising method. ISNR = SNR after -SNR before, the higher the ISNR value, the better the quality of the denoised signal. CC is an important indicator to measure the similarity between the original signal and the denoised signal. The closer CC is to 1, the closer the denoised signal is to the original signal, and the better the denoising effect.

[0189] Figure 8 Shows the RMSE, SNR, ISNR, and CC of the reconstructed signals obtained by four different methods. It can be seen from the figure that the results obtained by the IMS-RLMD method have the smallest RMSE and the largest signal-to-noise ratio, ISNR, and CC. Compared with the original signal, the IMS-RLMD denoising method increases the signal-to-noise ratio by 64.5%, which indicates that after being processed by the IMS-RLMD method, the interference noise is effectively reduced and the high-frequency signals are highlighted. In summary, compared with LMD, IMS-LMD, and RLMD, the overall performance of IMS-RLMD is the best.

[0190] To verify the above method, a ZH102CRY2301 noise sensor was used to collect the acoustic signals under normal operating conditions, such as Figure 9 shown. The acoustic signals of the runner blades under the condition of sediment concentration of 0.73 kg / m 3 were collected, such as Figure 10 shown. M12 nuts were put into the water turbine generator set through the pre-reserved pipe openings in advance, and the mutation signals generated by the nuts entering the water turbine generator set together with the water flow were collected. The acoustic signals under the condition of foreign object impact were collected, such as Figure 11 shown. 50 groups were collected for each.

[0191] Randomly select one group of data as the original data, and display the noise reduction and recognition process. The time-frequency domain image is as Figure 12 shown. First, the RLMD decomposition method was used to decompose the original signal into 5 PF components and 1 residual Residual. The decomposition results are as Figure 13 shown.

[0192] The correlation coefficient and approximate entropy of each component were calculated respectively. According to the calculation results, the PF components were screened to complete the preliminary noise reduction. The calculation results are as Figure 14 shown. According to Figure 14 (a), the Spearman correlation coefficient of the components PF5 and Residual approaches 0, that is, these two components are basically uncorrelated with the original signal, so they are judged as invalid components. Since there are often huge high-frequency noises during the operation of the hydropower unit, the approximate entropy is introduced to further screen the signal components. According to Figure 14 (b), the approximate entropy of the component PF1 far exceeds that of other components, indicating that there is a lot of noise in the component PF1. Therefore, PF1 is judged as an invalid component. After screening, PF2, PF3, and PF4 are judged as valid components. The masking signal method was used to process the valid components, and the processing results are asFigure 15 as shown

[0193] Comparison Figure 15 It can be clearly seen from (a) and (b) that the masking signal method significantly suppresses Figure 15 the 58 Hz frequency originally belonging to PF3 mixed in PF2, and at the same time, the 29 Hz frequency mixed in PF3 is also suppressed to a certain extent. It is thus determined that the IMS-RLMD decomposition method greatly suppresses the false signal components caused by modal aliasing in RLMD, retains the important information in the original signal, and improves the efficiency of signal noise reduction.

[0194] Using IMS-RLMD to perform noise reduction on the 150 * 10240 data set (50 groups under normal conditions, 50 groups under sediment conditions, and 50 groups under rubbing conditions) composed of the original data respectively, a 150 * 10240 data set composed of the reconstructed signal prx i (t) is obtained after processing.

[0195] The reconstructed signal is as follows

[0196] prx i (t) = PMS - PF i1 + PMS - PF i2 + PMS - PF i3 , i = 1, 2, 3, …, 150

[0197] Randomly select one group from each of the three types of sound signals for display. As Figure 16 shown, from left to right are the time domain diagrams of the original sound signal data under normal operating conditions, sediment-laden water flow conditions, and rubbing conditions (foreign object impact).

[0198] Perform IMS-RLMD processing on the three types of sound signals respectively to obtain three IMS-PF components, and then reconstruct the signals according to the obtained components. The results are as Figure 17 shown.

[0199] To verify the effectiveness of IMS-RLMD noise reduction, select the original sample set and the sample sets after noise reduction by EMD, LMD, RLMD, and VMD methods as the control groups. The data sets obtained by different processing methods are shown in Table 1

[0200] Table 1 Reconstructed signals of six processing methods

[0201]

[0202] Note: imf ij , pf ij , PF ij , IMF ijFor the decomposed components of the EMD, LMD, VMD, and RLMD methods, IMS_PF ij is the component obtained by the IMS-RLMD method.

[0203] Time-frequency domain feature extraction is performed on the reconstructed signal dataset obtained after IMS-RLMD processing to obtain the feature dataset of the reconstructed signal. The grey wolf optimization algorithm-improved support vector machine (GWO-SVM) is used to classify and identify this feature dataset.

[0204] Randomly select 80% of each group of data as the training sample set of GWO-SVM, and the remaining 20% as the test set sample. Repeat the recognition process 50 times. The support vector machine SVM is widely used in fault recognition because it is suitable for the classification and recognition of small sample data. However, the performance of SVM is closely related to the selection of its kernel function. In this paper, the grey wolf optimization algorithm (GWO) is selected to optimize the parameters. Set the number of wolf packs to 20 and perform 100 iterations to optimize the parameters c and g. Import the training sets of the three types of signals into GWO-SVM for training. As the number of iterations increases, the GWO fitness value stabilizes at 99.11%, and at the same time, the optimal optimization parameter c best = 77.6285, g best = 0.01. Use the obtained optimal parameters to perform grid optimization and training on SVM. Finally, put the 3*10*18 test set into the trained SVM for classification and recognition. The classification results of GWO-SVM for the dataset processed by the IMS-RLMD method are as Figure 18 (a) shown.

[0205] To verify the effectiveness of the improved signal processing method IMS-RLMD, select the original signal x(t) without noise reduction and the reconstructed signals rex(t), rlx(t), rx(t), rvx(t) processed by the EMD, LMD, RLMD, and VMD methods as the control group and compare them with the classification results of IMS-RLMD. The recognition results are respectively as Figure 18 shown.

[0206] Due to the severe influence of the mode aliasing phenomenon, the EMD and LMD methods cannot accurately separate the normal operating conditions (category 1) and the sediment-laden flow conditions (category 2). Although the RLMD method has been improved based on LMD, it still cannot completely avoid the influence of mode aliasing. Due to its unique decomposition method (decomposing from high frequency to low frequency), VMD is more suitable for high-frequency noise denoising than other methods, but VMD is affected by parameters and requires manual parameterization. At the same time, VMD is not good at capturing transient features. It can be found that the VMD method cannot accurately identify the mutation signals generated by the impact conditions (category 3). Compared with the above methods, IMS-RLMD can minimize the influence of the mode aliasing phenomenon and achieve a 100% accuracy rate for 3-class recognition.

[0207] Figure 19 The recognition accuracy of the signals after denoising by different methods is shown. Among them, the recognition accuracy rate of IMS-RLMD is the highest, which is 96.67%. The recognition accuracy rates of the EMD, LMD, RLMD, and VMD methods for the original signal and the denoised signal are 66.67%, 75%, 78.3%, 86.67%, and 90% respectively. This shows that the IMS-RLMD denoising method can filter out a large amount of interference noise to the greatest extent and fully retain the characteristics of various types of signals.

[0208] To further illustrate the superiority of the proposed IMS-RLMD denoising method, the original signal x(t), the signal rx(t) obtained by denoising with the RLMD method, and the signal obtained by denoising with the IMS-RLMD method are dimension-reduced and visualized, and the confidence interval is set to 95% to draw the clustering region. The results are as Figure 20 shown. Among them, the green spheres are category 1, representing the sound signals of normal operating conditions; the red spheres are category 2, representing the sound signals of sediment-laden flow operating conditions; and the blue spheres are category 3, representing the sound signals of rubbing conditions.

[0209] Figure 20 (a) is the classification of the feature set of the original data x(t). It can be seen from the figure that for the original signal x(t) without denoising, the clustering effect of category 3 is poor, and there is serious category aliasing between category 1 and 3. The reasons are analyzed as follows: For the original signal, the huge background noise accompanying the operation of the hydropower unit causes great interference in feature extraction, thereby reducing the clustering effect. Moreover, due to the large noise, the noise signals contained in category 1 are misjudged as the mutation signals generated by rubbing represented by category 3, and some mutation signals in category 3 are also misjudged as the noise signals in category 1. This further illustrates that in the fault diagnosis of hydropower units, it is very necessary to denoise the original signal.

[0210] After denoising the original signal using the RLMD method, the clustering results of the sound signals of the three operating conditions are as Figure 20(as shown in (b)). At this time, the separation effect of the acoustic signals under three different working conditions has been significantly improved, but there is still a phenomenon of unclear separation, and the cross - influence between categories 1 and 3 has not been eliminated. It shows that although the RLMD decomposition method has a certain noise reduction effect, due to the influence of the mode mixing phenomenon, some high - frequency signals will be further enhanced, thus having a more serious impact on the classification results. Therefore, in order to better separate the rubbing sound signal from the huge noise signal, it is necessary to effectively suppress the mode mixing phenomenon generated in the RLMD decomposition. Thus, the IMS - RLMD noise reduction method is proposed. Figure 20 (c) is the clustering result of the feature data set processed by the IMS - RLMD noise reduction method. As shown in the figure, the feature data sets of all three categories have obtained good clustering effects. Compared with the data set obtained by noise reduction using the RLMD method, the IMS - RLMD effectively reduces the interference of high - frequency noise on the classification of category 1 and category 3.

[0211] The IMS - RLMD noise reduction method not only uses RLMD to suppress the endpoint effect generated in the decomposition process, but also greatly reduces the adverse effects brought by the mode mixing effect by constructing a masking signal function, thus better separating the noise signal and the rubbing fault sound signal, and increasing the recognition accuracy of GWO - SVM to 96.67%. This can further prove the importance of the masking signal (IMS) method in the recognition of fault sound signals.

[0212] To further illustrate the effectiveness of the IMS - RLMD denoising method and the applicability of this method and the GWO - SVM model, the convolutional neural network (CNN) and long short - term memory (LSTM) are introduced as control groups. The unprocessed original signal and the signal denoised by the IMS - RLMD method are input into the LSTM, CNN and GWO - SVM models for recognition. The confusion matrix of the recognition results is as Figure 21 shown.

[0213] First, the unprocessed original signal is divided into a training set and a test set according to a ratio of 8:2. Secondly, the training set is respectively input into three models of LSTM, CNN and GWO - SVM for training. Finally, the test set is input into the trained models. The confusion matrix of the classification results is as Figure 22 shown in the first column. The recognition accuracies of the above three models are 63.33%, 76.67% and 66.67% respectively. The reason may be that the CNN model is good at extracting local features of data, so the recognition accuracies of category 1 and category 2 are higher than those of the LSTM model and GWO - SVM. However, at the same time, the CNN model cannot well recognize mutation signals (category 3). In contrast, LSTM and GWO - SVM have better control of global features than the CNN model, so the recognition accuracy of category 3 is higher.

[0214] In order to further illustrate the applicability of the signal denoising method IMS-RLMD and the classification model GWO-SVM, the original signal is first denoised using the IMS-RLMS method, and then the denoised signal is input into the three models for comparison. Figure 22 The recognition accuracy of LSTM and CNN models has been improved, but these models are prone to overfitting due to the small size of the fault sample data set. At the same time, after multiple manual adjustments to the important parameters of the model, the recognition accuracy is still only 80% and 90%. In contrast, SVM is better at processing small sample data and nonlinear data. In the GWO-SVM model, the GWO optimization algorithm is used for parameter optimization, which further improves the SVM performance. The recognition accuracy of GWO-SVM is 96.67%.

[0215] In order to evaluate the performance of the above three models, accuracy, precision, recall and F1 score are selected as evaluation indicators. The definition of each evaluation indicator is as follows:

[0216]

[0217]

[0218] Among them, TP represents the number of correctly identified positive class signals. TN represents the number of correctly identified negative class signals. FP represents the number of incorrectly identified positive class signals. FN represents the number of incorrectly identified negative class signals.

[0219] Figure 22 The evaluation indicators of the three models with different signal inputs are shown. It can be seen that the overall performance of the denoised signal is better than the original signal, and this result further proves the effectiveness of the IMS-RLMD denoising method. When the three models process the original signal, the CNN model performs best, but it still cannot meet the needs. When processing the denoised signal, the evaluation indicators of GWO-SVM are better than the other two models, with accuracy, precision, recall and F1_score of 96.67%, 96.97%, 96.67% and 96.66% respectively. Therefore, the fault diagnosis scheme that uses the IMS-RLMD method for signal denoising and then uses the GWO-SVM model for identification has the best overall performance, good stability and accuracy, and better meets the needs of hydropower units under actual operating conditions.

[0220] Although the present invention has been described herein with reference to various illustrative embodiments, it should be understood that those skilled in the art can devise many other modifications and embodiments that will fall within the scope and spirit of the principles disclosed in this application. More specifically, within the scope of the disclosure, the drawings, and the claims of this application, various deformations and improvements can be made to the components or layout of the subject combination layout. In addition to the deformations and improvements to the components or layout, other uses will also be apparent to those skilled in the art.

Claims

1. A method for diagnosing turbine runner faults based on RLMD optimized by an improved masked signal method, characterized in that The steps include: S1. Collecting acoustic vibration signals of the turbine when it is running under different working conditions; S2. Use the robust local mean decomposition method to decompose the acoustic vibration signal into multiple components, use the Spearman correlation coefficient method to select the components containing effective features, and use the approximate entropy to remove the components containing high-frequency interference noise signals to obtain the screened components; S3. After masking the filtered components using a mask signal method, the components are reconstructed to obtain a denoised signal; S4. Randomly classify the denoised signal to obtain a training set and a validation set. Input the training set into GWO-SVM for model training to obtain a diagnostic model. Then input the validation set into the diagnostic model for fault identification.

2. The method according to claim 1, characterized in that: The working conditions in S1 are three types: normal operation condition, sand-laden water flow condition, and foreign body impact condition.

3. The method according to claim 1, characterized in that: The specific steps of S2 are as follows: S2-1. Perform robust local mean decomposition (RLMD) Discretize the input continuous signal to get x(n), and treat it as the original signal for processing; find the local maximum of x(n) to determine the symmetry point; use the mirror extension method to expand the original signal to get x(t); assume that the local mean of all extreme values ​​is m i , the i-th local amplitude of the continuous extreme value is a i , then: where n i , n i+1 are all local extreme points of x(t); According to m i and a i Automatically select an appropriate fixed subset size k = odd(s c +3s σ ), the local mean function m is obtained by using the sliding average algorithm 11 (t) and the local amplitude function a 11 (t); by m 11 (t) and a 11 (t) Get zero mean signal k 11 (t) and pure FM signal s 11 (t), the formula is as follows: k 11 (t)=x(t)-m 11 (t) Will s 11 (t) is used as the original signal and iterated until the objective function is satisfied: f=RMS[z(n)]+EK[z(n)] Where: z(n)=a(n)-1, that is, zero baseline envelope signal, N s is the total number of signal sampling points, z is the arithmetic mean of z(n); When the iteration stops, the pure FM signal function s is obtained. 1n (t), envelope signal function a1(t), where a1(t) is as follows: Multiplying the final AM signal with the pure FM signal function gives the first PF component: PF1=a1(t)s 1n (t) After separating PF1 from the original signal x(t), a new signal u1(t) is obtained; the whole process is repeated j times with u1(t) as the original signal until u j (t) is a constant or a monotonic function: Finally, the original signal x(t) is expressed as: Thus, the PF component after RLMD decomposition is obtained; S2-2. Calculate the Spearman correlation coefficient and approximate entropy of the obtained PF components, and screen the PF components according to the Spearman correlation coefficient and the approximate entropy; first calculate the Spearman correlation coefficient and sort the PF components from small to large according to the absolute value, and take the first 60% of the components close to 1; then calculate the approximate entropy of all components and sort them from large to small, and take the last 60% of the components.

4. The method according to claim 1, characterized in that: The specific steps of S3 are as follows: S3-1. Use the mask signal method to mask the filtered components: Assume that the signal to be processed is x(t), and after performing Hilbert transform on x(t), we get y(t): Where λ is the integral variable; Find the instantaneous frequency: in is the phase function; By energy mean method The mask signal format is: Average instantaneous frequency Average Amplitude S3-2. The n PF components of the effective features are expressed as: S3-3. To PF i+ With PF i- Perform RLMD decomposition again and take the first component of the output as pf i+ With pf i- PF i Redefining as IMS-PF i : Then the obtained component IMS-PF i , i=1,2,...,n for the final signal reconstruction, the resulting reconstructed signal is as follows:

5. The method according to claim 1, characterized in that: The specific steps of the GWO-SVM are as follows: For a given training set: T={[a1,y1],[a2,y2],…,[a l ,y l ]}∈(Ω×Y) l Where a i ∈Ω=R n ,Ω is called the input space, and each point a in the input space i , composed of n attribute features; y i ∈Y={-1,1},i=1,2,…,l;Find R n A real-valued function g(x) on , constructing the classification function f(x)=sign(g(x)); The kernel function is introduced and mapped to the high-dimensional feature space, which changes the corresponding dual problem into: Assume the optimal solution is but: So the optimal classification model is The parameters c and g are optimized using the Grey Wolf Algorithm, and the optimal parameters are used to perform grid optimization and training on the SVM to obtain the final GWO-SVM.

Images