A pumped storage unit fault diagnosis method

Abstract

The application provides a pumped storage unit fault diagnosis method, first, obtaining the operation data of the pumped storage unit, including non-electric quantity signal, electric quantity signal and environmental quantity signal; second, introducing time-varying sparsity constraint in SVMD decomposition to form TV-SVMD decomposition method to decompose the non-electric quantity signal; performing category weighting on the diagonal averaging in SGMD decomposition to form CW-SGMD method to decompose the electric quantity signal; performing AF filtering processing on the environmental quantity signal to reduce the complexity of the data; performing linear constraint minimum variance weighting processing on the decomposed and filtered signals according to the expert experience to focus on the key factors; then, performing FSFC feature extraction on the weighted data to obtain the key features of the weighted data which have the greatest influence on the fault diagnosis; finally, establishing a VS-DA-TMMODE fault diagnosis model to diagnose the fault of the pumped storage unit, optimizing the model parameters through the RRTO optimization algorithm to obtain the final diagnosis result.

Description

Technical Field

[0001] This invention relates to the field of pumped storage, specifically a method for diagnosing faults in pumped storage units. Background Technology

[0002] With the continuous expansion of power systems and the large-scale integration of renewable energy sources (such as wind and solar power), higher demands are placed on the stability and flexibility of the power grid. Pumped storage units, as a key energy storage and peak-shaving power source, are crucial for ensuring grid security through stable and reliable operation. Therefore, developing efficient and accurate fault diagnosis methods to promptly detect and address unit faults, reducing downtime and maintenance costs, is an urgent need for power system development.

[0003] Pumped storage hydroelectric units have complex structures and operate in harsh environments, making them prone to various malfunctions. Traditional fault diagnosis methods often rely on human experience, resulting in low efficiency and a high risk of errors. However, with the development of technologies such as big data and artificial intelligence, advanced signal processing, feature extraction, and pattern recognition techniques can be used to achieve real-time monitoring of the unit's operating status and early warning of faults. This improves the accuracy and reliability of fault diagnosis, providing strong support for the safe operation of the unit.

[0004] Currently, some fault diagnosis methods have been applied, such as vibration analysis and temperature monitoring. However, these methods often have limitations, such as poor adaptability to complex operating conditions and low sensitivity to early-stage faults. Therefore, further research and development of new fault diagnosis methods are needed, combining multiple technologies to improve the diagnostic capabilities and adaptability for unit faults. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide a fault diagnosis method for pumped storage units. This method can predict the temperature of the guide bearing of the hydro-generator unit in advance, making it convenient for operation and maintenance personnel to detect abnormal temperature points in advance and take measures to ensure the stable operation of the hydro-generator unit.

[0006] To achieve the above-mentioned technical features, the objective of this invention is as follows: a fault diagnosis method for pumped storage units, the method comprising the following steps: Step 1: Obtain the non-electrical quantity signals of the pumped storage unit, introduce time-varying sparsity constraints (TV) into the continuous variational mode decomposition (SVMD) to form the TV-SVMD decomposition method, and decompose the non-electrical quantity signals; Step 2: Obtain the electrical quantity signals of the pumped storage unit, and perform category weighting (CW) processing on the diagonal averaging in the symplectic geometric mode decomposition (SGMD) to form the CW-SGMD decomposition method to decompose the electrical quantity signals; Step 3: Acquire environmental quantity signals of the pumped storage unit and perform adaptive (AF) filtering on the environmental quantity signals; Step 4: Perform linear constraint minimum variance weighted processing on the signal data after decomposition and filtering in Steps 1-3 to further optimize the feature extraction effect and improve the accuracy of subsequent analysis; Step 5: Perform Functional Feature Selection (FSFC) feature extraction on the weighted data to obtain the key features that have the greatest impact on fault diagnosis. Step 6: Establish a multi-scale dual-axis time-series multi-mode coding (VS-DA-TMMODE) fault diagnosis model to diagnose faults in pumped storage units. Optimize model parameters using the Rapid Exploratory Random Tree Optimization (RRTO) algorithm to obtain the final diagnostic results.

[0007] Preferably, the non-electrical signals of the pumped storage unit in step 1 include vibration, sway, rotational speed, winding temperature, guide bearing temperature, and water pressure.

[0008] Preferably, step 1, which involves performing TV-SVMD decomposition on the non-electrical signals of the pumped storage unit, includes the following steps: Step 1.1: For any non-electrical input signal Assume it is divided into: ; In the formula: For the first d First mode; To exclude the first d Signals outside the first mode, i.e. residual signals; Step 1.2: Improve the Lagrangian objective function of the original SVMD by introducing a time-varying sparsity constraint function. : ; In the formula, The initial constraint strength; To adjust the parameters; The improved Lagrangian objective function is obtained: ; In the formula, For the first i One modal component; For the first i The center frequency of each modal component; The unit impulse function; These are weight parameters; The kernel function of the Hilbert transform; It is an L2 norm; n The total number of items; Step 1.3: To obtain the most ideal hypothetical result, the following constraints must be met: Ensure that each decomposed mode is compact near its center frequency, based on this... d The constraints that the first mode should satisfy are: J 1; residual signal Energy in The minimum constraint condition is satisfied at the center frequency of the effective component. J 2; By minimizing the above two constraints, we obtain the... d To distinguish between modes of different orders and avoid mode aliasing, constraints for mode differentiation are established. J 3; Signal reconstruction ensures that the decomposed modes and residual signals can completely reconstruct the original signal; Step 1.4: Transform the constraints described in Step 1.3 into an optimization problem: ; In the formula, To balance J 1. J 2 and J The parameters of 3 are obtained, and the optimization problem is solved by solving the improved Lagrangian objective function to obtain the 3rd parameter. d First mode and the updated residual signal ; Step 1.5: Repeat the modal extraction and optimization steps until the residual signal reaches the set threshold or a specified number of modes are extracted, and output the final vibration modal components: ; Similarly, all remaining non-electrical signals are also decomposed according to steps 1.1 to 1.5 above, resulting in... , ... .

[0009] Preferably, the electrical quantity signals in step 2 include voltage, current, and partial discharge.

[0010] Preferably, step 2, which involves CW-SGMD decomposition of the electrical signals of the pumped storage unit, includes the following steps: Step 2.1: Construct the trajectory matrix. For any electrical quantity signal, for a given signal... x = x 1, x 2, ..., x nAccording to Takens' embedding theorem, one-dimensional data can be reconstructed into multi-dimensional signals: ; In the formula, d For the embedding dimension; t For delay time; m = n -( d -1) t ; X The reconstructed multidimensional data matrix; x These are data points in the original one-dimensional time series data; Step 2.2: Perform symplectic geometric similarity transformation based on the trajectory matrix. X Constructing the covariance matrix A and Hamilton matrix M They are respectively: ; ; make According to the definition of the Hamiltonian matrix, the symplectic orthogonal matrix is ​​obtained. Q for: ; In the formula, B It is an upper triangular matrix; l ( A )= l ( B )= l (2) X ); Based on the properties of the Hamiltonian matrix, calculate B eigenvalues l 1, l 2, ..., l d ,get X eigenvalues s i = l i , i =1, 2, ..., d ,in: l i Sort in descending order; denoted as Q i , i =1, 2, ..., d , is a matrix A The eigenvectors corresponding to the eigenvalues ​​of , then The initial single-component matrix is ​​obtained. Z i Then the trajectory matrix Z Represented as:Z = Z 1+ Z 2+…+ Z d ; Step 2.3: After the above transformations, the dimension of the initial component matrix is ​​obtained as follows: m × d After diagonal averaging, the matrix is... Z i Transform into d Group length is n Time series, this d The sum of the time series is the original signal. For ease of calculation, the elements of the matrix are defined as follows: Z ij For any initial single component Z i ,definition: ; In the formula: This indicates taking the smaller value between m and d; This indicates taking the larger value between m and d; These are the elements of the transformed matrix; if m ≤ d, then... equal to z of the original matrix ij If m > d, then equal to z of the original matrix ji Then, by taking the diagonal average, we get: ; In the formula, the first row of the equation represents the expression for each k Starting from the top left corner of the matrix, take the first... k The formula for each of the diagonal elements is calculated; the second line of the formula represents the average of each... k Take the front The formula for the diagonal elements is calculated, and their average is given; the formula in the third row represents... n It is a matrix Z The number of columns; for each k ,from Start by taking the last diagonal element and calculating its average; Introducing categorical weighting into the diagonal average for each index q or interval k Assigning exponential decay weights To reflect the importance of different categories, the revised formula is as follows: ; get d The group of single-component signals are: ; Step 2.4: Similar Component Recombination. The components decomposed by the SGMD method are not independent of each other and contain components with the same frequency and features. It is necessary to perform similarity analysis on each component and merge them to obtain SGC. i Then from the original signal x Remove SGC i The residual signal is denoted as g h ,Right now: ; In the formula, h This represents the number of iterations. The iteration stops when the normalized mean square error (NMSE) of the residual signal and the original signal is less than a given threshold. ; In the formula, for each iteration, the residual error is calculated as the square of the difference between the predicted value and the actual value. Sum the squares of all residual errors to obtain Divide the sum of squared residual errors by the sum of the original data. The normalized mean square error (NMSE) is obtained. When the iteration stops at the threshold condition, we get: ; In the formula, N To obtain the number of SGC components; The residual components are used to obtain the final decomposition result of the current signal. Similarly, all remaining electrical quantity signals are decomposed according to steps 2.1 to 2.4, resulting in... ... .

[0011] Preferably, the environmental quantity signals in step 3 include ambient temperature, ambient humidity, and voiceprint.

[0012] Preferably, the environmental quantity signal is subjected to AF filtering processing, which includes the following steps: Step 3.1: For any environmental quantity signal, for the signal... First, calculate the autocorrelation function: ; Pick t Within a reasonable range, find the peak position of the main cycle; Step 3.2: Construct an autocorrelation filter, select an FIR filter, and use the autocorrelation function as the filter coefficients. And perform normalization: ; Step 3.3: Perform convolution filtering on the signal using the filter described above. For signal Perform convolution to obtain the filtered signal: ; result The filtered signal retains patterns with strong autocorrelation while suppressing short-term random noise. Similarly, the remaining environmental signals are also filtered according to steps 3.1 to 3.3, and the processed signals are as follows: , ... .

[0013] Preferably, step 4 specifically includes the following steps: Step 4.1: Let the principal components be... Secondary components are ~ The remaining variables are ~ ; Step 4.2: Perform linearly constrained minimum variance weighting on the above signals. The modes corresponding to the principal components and the modes corresponding to all secondary components are then subjected to linearly constrained minimum variance weighting. The remaining variables are embedded as biases into the weighting process. The process is as follows: ; In the formula, The data is weighted. As weight; For bias; As a secondary component; For the remaining variables; i Indicates an index for the data; Step 4.3: Introduce Lagrange multipliers a and b Construct the Lagrange function: ; In the formula, For variance; a The constraint that corresponds to a sum of weights of 1. b The constraint that the sum of the weights of the primary components is greater than the sum of the weights of the secondary components; Step 4.4: For each weight and bias Find the partial derivative: Solving the above system of partial derivative equations yields the optimal weights. and bias Finally, combining the above steps, we obtain the final weighted data: .

[0014] Preferably, the extraction of weighted data and inverse FSFC features includes the following steps: Step 5.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Represented as a linear combination of basis functions, using the first k Using principal components to approximate the data: ; In the formula, For the first j The modal component in the ... s Scores on each principal component; It is the first s The principal component basis functions are represented as the score matrix of all modal components. ,in p It is the number of modal components. k It is the number of principal components in each modal component, and the score vector is represented as ; Step 5.2: The FSFC method performs feature selection by minimizing the following objective function: ; In the formula, It's a category label. X ( i ) is a matrix X The i OK, B j yes B The corresponding number in the middle j Subvectors of each modal component oh j It is the feature weight. l 1 and l 2 is the regularization parameter; Step 5.3: Construct the dual problem for optimization: ; In the formula, V and Z As dual variables; and These are the dual functions of the loss function and the regularization term, respectively. Step 5.4: Introduce the Lagrange function, for Taking its partial derivative yields the weight vector for each modal component. : ; In the formula, For disciplinary parameters; Step 5.5: If If the coefficient is zero, it means that the mode is not important and can be removed; otherwise, select the mode components with non-zero residual coefficients as the extracted m feature components, and finally obtain the dataset after feature selection: , as input to the pumped storage fault diagnosis model.

[0015] Preferably, the specific process of establishing the VS-DA-TMMODE fault diagnosis model is as follows: Step 6.1: Input Data Multi-scale representations are generated after sampling: , among which is Um No. m Scale data, scale m The downsampling step size is 2, for each scale m ,have: ; In the formula, Conv represents the convolution operation; stride=2 indicates that the stride of the convolution operation is 2. Step 6.2: At the coarsest scale Channel mixing is performed, and a variational self-attention mechanism is applied: ; In the formula, Q M , K M , V M It is by The query, key, and value matrix obtained by linear projection; Step 6.3: Embed all multi-scale time series into the deep pattern set, where: ; Step 6.4: For each scale m Applying Fast Fourier Transform to extract the pre-processed material K One main frequency: ; In the formula, A For amplitude; f k For frequency; p k = T / f k The period length; Step 6.5: For each scale m Time series remodeling K Multi-resolution temporal images: ; In the formula For each time frame, a biaxial attention mechanism is applied: ; ; In the formula, Q col , K col , V col It refers to the query, key, and value of column axis attention; Q row , K row , V row It refers to the query, key, and value of the row axis attention; Step 6.6: For each scale m, sum the weighted representations of the K periods. ⊙ represents element-wise multiplication; after weighted processing, a multi-scale feature set is obtained: The feature at each scale is processed by multiple heads, and the results are integrated to obtain the final output: ; In the formula, Head m It is a linear head layer at the m-th scale; Ensemble is an ensemble method.

[0016] Preferably, the RRTO optimization algorithm optimizes the model parameters. oh and r The specific process is as follows: Step 7.1: Initialize parameters oh , r Their value ranges are [L] ω U ω ] and [L ρ U ρ Initialize the position of each individual in the population: ; In the formula, r i,ω and r i,ρ It is a random number in the range (0,1]. Step 7.2: Calculate the step size adjustment function based on the adaptive step size walking strategy. As the number of iterations t increases, dynamically adjust the parameters related to the step size. K and E The individual location will be updated accordingly: ; In the formula SStep size, r 1 is a random number between (0, 1]. C It is the step size penalty factor; Step 7.3: Calculate the difference between the optimal individual and the current individual based on the adaptive step size strategy of absolute difference: according to the current optimal position. X best and the individual's current location X i The difference is used to update the individual's position: ; In the formula, This is an adaptive adjustment factor; Step 7.4: Boundary-based adaptive step size strategy, boundary detection, and individual update method are as follows: ; In the formula, The phase angle is used to calculate the updated individual; Step 7.5: Repeat steps 7.2-7.3 until the maximum number of iterations is reached or the convergence condition is met. Finally, output the final optimal parameter combination. ,in Therefore and The objective function is a variable.

[0017] The present invention has the following beneficial effects: 1. This invention employs the SGMD decomposition method, improving the Lagrangian objective function by introducing a time-varying sparsity constraint (TV) function and setting constraint conditions. This ensures that the decomposed modes are compact, highly discriminative, and avoid aliasing, while also accurately reconstructing the original signal. Its unified framework is applicable to various mechanical quantity signals, effectively extracting fault features and improving the accuracy and reliability of fault diagnosis.

[0018] 2. This invention can accurately extract signal features by using the CW-SGMD decomposition method, improve decomposition accuracy through trajectory matrix construction and symplectic geometric similarity transformation, highlight the main features by category-weighted diagonal average, and reduce the influence of noise by similar component recombination. It is applicable to a variety of electrical signals and provides a reliable basis for fault diagnosis.

[0019] 3. This invention uses the AF filtering method to determine the filter coefficients using the autocorrelation function, which can effectively preserve the effective mode of the signal, suppress short-time random noise, and is applicable to a variety of environmental signals, providing clearer and more accurate environmental information for fault diagnosis.

[0020] 4. This invention employs a linear constraint minimum variance weighting method to rationally allocate weights, highlighting the contribution of the main components and suppressing the interference of secondary components. At the same time, it introduces bias and other variables to make more comprehensive use of signal information, improve the feature expression ability after data fusion, and thus enhance the accuracy and reliability of fault diagnosis.

[0021] 5. This invention uses the FSFC feature extraction method to filter out key feature components from weighted data. By minimizing the objective function, feature selection and weight optimization are achieved, unimportant modes are eliminated, data dimensionality is reduced, and key features are highlighted, thereby improving the input quality of the fault diagnosis model and increasing diagnostic efficiency and accuracy.

[0022] 6. This invention employs the VS-DA-TMMODE fault diagnosis model, which utilizes multi-scale representation, variational self-attention mechanism, dual-axis attention mechanism, and fast Fourier transform to effectively extract features and focus on key information, thereby enhancing the accuracy and robustness of fault diagnosis. Its multi-scale feature set and ensemble method can comprehensively analyze data, improving diagnostic reliability.

[0023] 7. This invention employs the RRTO optimization algorithm to quickly and effectively optimize model parameters. Through strategies such as adaptive step size adjustment and boundary detection, it improves optimization efficiency and accuracy. It combines multiple optimization strategies to avoid local optima and find the global optimum, thereby enhancing the performance and reliability of the fault diagnosis model. Attached Figure Description

[0024] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0025] Figure 1 This is a schematic diagram of the TV-SVMD decomposition process provided by the present invention.

[0026] Figure 2 This is a schematic diagram of the CW-SGMD decomposition process provided by the present invention.

[0027] Figure 3 This is a schematic diagram of the FSFC feature extraction process provided by the present invention.

[0028] Figure 4 This is a flowchart illustrating the RRTO algorithm provided by the present invention.

[0029] Figure 5 This is a schematic diagram of the fault diagnosis system provided by the present invention. Detailed Implementation

[0030] The embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0031] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1: A method for diagnosing faults in pumped storage units, the method comprising the following steps: Step 1: Acquire non-electrical quantity signals of the pumped storage unit, including: vibration, sway, rotational speed, winding temperature, guide bearing temperature, and water pressure; introduce time-varying sparsity constraint (TV) into the Successive Variational Mode Decomposition (SVMD) decomposition to form the TV-SVMD decomposition method, and decompose the non-electrical quantity signals; The SVMD decomposition method improves the Lagrangian objective function by introducing a time-varying sparsity constraint (TV) function and setting constraints to ensure that the decomposed modes are compact, highly discriminative, and avoid aliasing, while also accurately reconstructing the original signal. Its unified framework is applicable to various mechanical signals, effectively extracting fault features and improving the accuracy and reliability of fault diagnosis.

[0033] TV-SVMD decomposition of non-electrical signals of pumped storage units includes the following steps: Step 1.1: Taking vibration signals as an example, for the input signal... Assume it is divided into: ; In the formula: For the first d First mode; To exclude the first d Signals outside the first mode, i.e. residual signals; Step 1.2: Improve the Lagrangian objective function of the original SVMD by introducing a time-varying sparsity constraint function (TV). : ; In the formula, The initial constraint strength; To adjust the parameters; The improved Lagrangian objective function is obtained: ; In the formula, For the first i One modal component; For the first i The center frequency of each modal component; The unit impulse function; These are weight parameters; The kernel function of the Hilbert transform; It is an L2 norm; n The total number of items; Step 1.3: To obtain the most ideal hypothetical result, the following constraints must be met: Ensure that each decomposed mode is compact near its center frequency, therefore the first... d The constraints that the first mode should satisfy are: J 1; residual signal Energy in The minimum constraint condition is satisfied at the center frequency of the effective component. J 2; By applying the two minimization constraints above, we can obtain the... d To distinguish between modes of different orders and avoid mode aliasing, constraints for mode differentiation are established. J 3; Signal reconstruction ensures that the decomposed modes and residual signals can completely reconstruct the original signal; Step 1.4: Transform the constraints described in Step 1.3 into an optimization problem: ; In the formula, To balance J 1. J 2 and J The parameters of 3 are obtained, and the optimization problem is solved by solving the improved Lagrangian objective function to obtain the 3rd parameter. d First mode and the updated residual signal ; Step 1.5: Repeat the modal extraction and optimization steps until the residual signal reaches the set threshold or a specified number of modes are extracted, and output the final vibration modal components: ; Similarly, all remaining non-electrical signals are also decomposed according to steps 1.1 to 1.5 above, resulting in... , ... .

[0034] Step 2: Acquire electrical quantity signals of the pumped storage unit, including voltage, current and partial discharge. Perform class-weighted (CW) processing on the diagonal averaging in Symptotic Geometric Mode Decomposition (SGMD) to form the CW-SGMD decomposition method, and decompose the electrical quantity signals. Among them, the CW-SGMD decomposition method can accurately extract signal features, improve decomposition accuracy through trajectory matrix construction and symplectic geometric similarity transformation, highlight the main features with category-weighted diagonal average, reduce the influence of noise by similar component recombination, and is applicable to a variety of electrical signals, providing a reliable basis for fault diagnosis.

[0035] The CW-SGMD decomposition of the electrical signals of pumped storage units includes the following steps: Step 2.1: Construct the trajectory matrix. Taking a current signal as an example, for a given signal... x = x 1, x 2, ..., x n According to Takens' embedding theorem, one-dimensional data can be reconstructed into multi-dimensional signals: ; In the formula, d For the embedding dimension; t For delay time; m = n -( d -1) t ; X The reconstructed multidimensional data matrix; x These are data points in the original one-dimensional time series data; Step 2.2: Perform symplectic geometric similarity transformation based on the trajectory matrix. X Constructing the covariance matrix A and Hamilton matrix M They are respectively: ; ; make According to the definition of the Hamiltonian matrix, the symplectic orthogonal matrix is ​​obtained. Q for: ; In the formula, B It is an upper triangular matrix; l ( A )= l ( B )= l (2) X ); Based on the properties of the Hamiltonian matrix, calculate B eigenvalues l 1, l 2, ..., l d ,get X eigenvalues s i = l i , ( i =1, 2, ..., d ),in: l i Sort in descending order; denoted as Q i , ( i =1, 2, ..., d ), is a matrix A The eigenvectors corresponding to the eigenvalues ​​of , then The initial single-component matrix is ​​obtained. Z i Then the trajectory matrix Z Represented as: Z = Z 1+ Z 2+…+ Z d ; Step 2.3: After the above transformations, the dimension of the initial component matrix is ​​obtained as follows: m × d After diagonal averaging, the matrix is... Z i Transform into d Group length is n Time series, this d The sum of the time series is the original signal. For ease of calculation, the elements of the matrix are defined as follows: Z ij For any initial single component Z i ,definition: ; In the formula: This indicates taking the smaller value between m and d; This indicates taking the larger value between m and d; These are the elements of the transformed matrix; if m ≤ d, then... equal to z of the original matrix ij If m > d, then equal to z of the original matrix ji Then, by taking the diagonal average, we get: ; In the formula, the first row of the equation represents the expression for each k Starting from the top left corner of the matrix, take the first... k The formula for each of the diagonal elements is calculated; the second line of the formula represents the average of each... k Take the front The formula for the diagonal elements is calculated, and their average is given; the formula in the third row represents... n It is a matrix Z The number of columns; for each k ,from Start by taking the last diagonal element and calculating its average; Introducing categorical weighting into the diagonal average for each index q or interval k Assigning exponential decay weights To reflect the importance of different categories, the revised formula is as follows: ; get d The group of single-component signals are: ; Step 2.4: Similar Component Recombination. Since the components decomposed by the SGMD method are not independent and contain components with the same frequency and features, it is necessary to perform similarity analysis on each component and merge them to obtain SGC. i Then remove SGC from the original signal x. i The residual signal is denoted as g h ,Right now: ; In the formula, h This represents the number of iterations. The iteration stops when the normalized mean square error (NMSE) of the residual signal and the original signal is less than a given threshold. ; In the formula, for each iteration, the residual error is calculated as the square of the difference between the predicted value and the actual value. Sum the squares of all residual errors to obtain Divide the sum of squared residual errors by the sum of the original data. The normalized mean square error (NMSE) is obtained. When the iteration stops at the threshold condition, we get: ; In the formula, N To obtain the number of SGC components; The residual components are used to obtain the final decomposition result of the current signal. Similarly, all remaining electrical quantity signals are decomposed according to steps 2.1 to 2.4, resulting in... ... .

[0036] Step 3: Acquire environmental signals of the pumped storage unit, including ambient temperature, ambient humidity and acoustic signature, and perform adaptive filtering (AF) on the environmental signals; Among them, the AF filtering method uses the autocorrelation function to determine the filter coefficients, which can effectively preserve the effective mode of the signal, suppress short-time random noise, and is applicable to a variety of environmental signals, providing clearer and more accurate environmental information for fault diagnosis.

[0037] AF filtering of environmental quantity signals includes the following steps: Step 3.1: Taking temperature signals as an example, for temperature signals... First, calculate the autocorrelation function: ; Pick t Within a reasonable range, find the peak position of the main cycle; Step 3.2: Construct an autocorrelation filter, select an FIR filter, and use the autocorrelation function as the filter coefficients. And perform normalization: ; Step 3.3: Perform convolution filtering on the signal using the filter described above. For signal Perform convolution to obtain the filtered signal: ; result The filtered signal retains patterns with strong autocorrelation while suppressing short-term random noise. Similarly, the remaining environmental signals are also filtered according to steps 3.1 to 3.3, and the processed signals are as follows: , ... .

[0038] Step 4: Perform linear constraint minimum variance weighted processing on the signal data after decomposition and filtering in Steps 1-3 to further optimize the feature extraction effect and improve the accuracy of subsequent analysis; Among them, the linear constraint minimum variance weighting method, by reasonably allocating weights, highlights the contribution of the main components and suppresses the interference of the secondary components. At the same time, it introduces bias and other variables to make more comprehensive use of signal information, improve the feature expression ability after data fusion, and thus improve the accuracy and reliability of fault diagnosis.

[0039] Specifically, the steps include the following: Step 4.1: Let the principal components be... Secondary components are ~ The remaining variables are ~ ; Step 4.2: Perform linearly constrained minimum variance weighting on the above signals. The modes corresponding to the principal components and the modes corresponding to all secondary components are then subjected to linearly constrained minimum variance weighting. The remaining variables are embedded as biases into the weighting process. The process is as follows: ; In the formula, The data is weighted. As weight; For bias; As a secondary component; For the remaining variables; i Indicates an index for the data; Step 4.3: Introduce Lagrange multipliers a and b Construct the Lagrange function: ; In the formula, For variance; a The constraint that corresponds to a sum of weights of 1. b The constraint that the sum of the weights of the primary components is greater than the sum of the weights of the secondary components; Step 4.4: For each weight and bias Find the partial derivative: Solving the above system of partial derivative equations yields the optimal weights. and bias Finally, combining the above steps, we obtain the final weighted data: .

[0040] Step 5: Perform Feature Selection for Functional Classification (FSFC) feature extraction on the weighted data to obtain the key features that have the greatest impact on fault diagnosis. Among them, the FSFC feature extraction method can filter out key feature components from weighted data, achieve feature selection and weight optimization by minimizing the objective function, eliminate unimportant modes, reduce data dimensionality, highlight key features, thereby improving the input quality of fault diagnosis models and improving diagnostic efficiency and accuracy.

[0041] The weighted data and inverse FSFC feature extraction include the following steps: Step 5.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Represented as a linear combination of basis functions, using the first k Using principal components to approximate the data: ; In the formula, For the first j The modal component in the ... s Scores on each principal component; It is the first s The principal component basis functions are represented as the score matrix of all modal components. ,in p It is the number of modal components. k It is the number of principal components in each modal component, and the score vector is represented as ; Step 5.2: The FSFC method performs feature selection by minimizing the following objective function: ; In the formula, It's a category label. X ( i ) is a matrix X The i OK, B j yes B The corresponding number in the middle j Subvectors of each modal component oh j It is the feature weight. l 1 and l 2 is the regularization parameter; Step 5.3: Construct the dual problem for optimization: ; In the formula, V and Z As dual variables; and These are the dual functions of the loss function and the regularization term, respectively. Step 5.4: Introduce the Lagrange function, for Taking its partial derivative yields the weight vector for each modal component. : ; In the formula, For disciplinary parameters; Step 5.5: If If the coefficient is zero, it means that the mode is not important and can be removed; otherwise, select the mode components with non-zero residual coefficients as the extracted m feature components, and finally obtain the dataset after feature selection: , as input to the pumped storage fault diagnosis model.

[0042] Step 6: Establish a multi-scale dual-axis temporal multi-mode encoder (VS-DA-TMMODE) fault diagnosis model to diagnose faults in pumped storage units. Optimize model parameters using a rapidly-exploring random tree-based optimizer (RRTO) algorithm to obtain the final diagnostic results.

[0043] Among them, the VS-DA-TMMODE fault diagnosis model, through multi-scale representation, variational self-attention mechanism, dual-axis attention mechanism, and fast Fourier transform, can effectively extract features and focus on key information, enhancing the accuracy and robustness of fault diagnosis. Its multi-scale feature set and ensemble method can comprehensively analyze data and improve diagnostic reliability.

[0044] Based on this, the VS-DA-TMMODE fault diagnosis model is established, and its specific process is as follows: Step 6.1: Input Data Multi-scale representations are generated after sampling: , among which is Um No. m Scale data, scale m The downsampling step size is 2, for each scale m ,have: ; In the formula, Conv represents the convolution operation; stride=2 indicates that the stride of the convolution operation is 2. Step 6.2: At the coarsest scale Channel mixing is performed, and a variational self-attention mechanism is applied: ; In the formula, Q M , K M , V M It is by The query, key, and value matrix obtained by linear projection; Step 6.3: Embed all multi-scale time series into the deep pattern set, where: ; Step 6.4: For each scale m Applying Fast Fourier Transform to extract the pre-processed material K One main frequency: ; In the formula, A For amplitude; f k For frequency; p k = T / f k The period length; Step 6.5: For each scale m Time series remodeling K Multi-resolution temporal images: ; In the formula For each time frame, a biaxial attention mechanism is applied: ; ; In the formula, Q col , K col , V col It refers to the query, key, and value of column axis attention; Q row , K row , V row It refers to the query, key, and value of the row axis attention; Step 6.6: For each scale m, sum the weighted representations of the K periods. ⊙ represents element-wise multiplication; after weighted processing, a multi-scale feature set is obtained: The feature at each scale is processed by multiple heads, and the results are integrated to obtain the final output: ; In the formula, Head m It is a linear head layer at the m-th scale; Ensemble is an ensemble method.

[0045] The RRTO optimization algorithm has the advantage of quickly and effectively optimizing model parameters. Through strategies such as adaptive step size adjustment and boundary detection, it improves optimization efficiency and accuracy. It combines multiple optimization strategies to avoid local optima and find the global optimum, thereby enhancing the performance and reliability of the fault diagnosis model.

[0046] Furthermore, the RRTO optimization algorithm optimizes the model parameters. oh and r The specific process is as follows: Step 7.1: Initialize parameters oh , r Their value ranges are [L] ω U ω ] and [L ρ U ρ Initialize the position of each individual in the population: ; In the formula, r i,ω and r i,ρ It is a random number in the range (0,1]. Step 7.2: Calculate the step size adjustment function based on the adaptive step size walking strategy. As the number of iterations t increases, dynamically adjust the parameters related to the step size. K and E The individual location will be updated accordingly: ; In the formula S Step size, r 1 is a random number between (0, 1]. C It is the step size penalty factor; Step 7.3: Calculate the difference between the optimal individual and the current individual based on the adaptive step size strategy of absolute difference: according to the current optimal position. X best and the individual's current location X i The difference is used to update the individual's position: ; In the formula, This is an adaptive adjustment factor; Step 7.4: Boundary-based adaptive step size strategy, boundary detection, and individual update method are as follows: ; In the formula, The phase angle is used to calculate the updated individual; Step 7.5: Repeat steps 7.2-7.3 until the maximum number of iterations is reached or the convergence condition is met. Finally, output the final optimal parameter combination. ,in Therefore and The objective function is a variable.

[0047] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for diagnosing faults in a pumped storage unit, characterized in that, The method includes the following steps: Step 1: Obtain the non-electrical quantity signals of the pumped storage unit, introduce time-varying sparsity constraints (TV) into the continuous variational mode decomposition (SVMD) to form the TV-SVMD decomposition method, and decompose the non-electrical quantity signals; Step 2: Obtain the electrical quantity signals of the pumped storage unit, and perform category weighting (CW) processing on the diagonal averaging in the symplectic geometric mode decomposition (SGMD) to form the CW-SGMD decomposition method to decompose the electrical quantity signals; Step 3: Acquire environmental quantity signals of the pumped storage unit and perform adaptive (AF) filtering on the environmental quantity signals; Step 4: Perform linear constraint minimum variance weighted processing on the signal data after decomposition and filtering in Steps 1-3 to further optimize the feature extraction effect and improve the accuracy of subsequent analysis; Step 5: Perform Functional Feature Selection (FSFC) feature extraction on the weighted data to obtain the key features that have the greatest impact on fault diagnosis. Step 6: Establish a multi-scale dual-axis time-series multi-mode coding (VS-DA-TMMODE) fault diagnosis model to diagnose faults in pumped storage units. Optimize model parameters using the Rapid Exploratory Random Tree Optimization (RRTO) algorithm to obtain the final diagnostic results. Step 1, which involves performing TV-SVMD decomposition on the non-electrical signals of the pumped storage unit, includes the following steps: Step 1.1: For any non-electrical input signal Assume it is divided into: ； In the formula: For the first d First mode; To exclude the first d Signals outside the first mode, i.e. residual signals; Step 1.2: Improve the Lagrangian objective function of the original SVMD by introducing a time-varying sparsity constraint function. : ； In the formula, The initial constraint strength; To adjust the parameters; The improved Lagrangian objective function is obtained: ； In the formula, For the first i One modal component; For the first i The center frequency of each modal component; The unit impulse function; These are weight parameters; The kernel function of the Hilbert transform; It is an L2 norm; n The total number of items; Step 1.3: To obtain the most ideal hypothetical result, the following constraints must be met: Ensure that each decomposed mode is compact near its center frequency, based on this... d The constraints that the first mode should satisfy are: J 1; residual signal Energy in The minimum constraint condition is satisfied at the center frequency of the effective component. J 2; By minimizing the above two constraints, we obtain the... d To distinguish between modes of different orders and avoid mode aliasing, constraints for mode differentiation are established. J 3; Signal reconstruction ensures that the decomposed modes and residual signals can completely reconstruct the original signal; Step 1.4: Transform the constraints described in Step 1.3 into an optimization problem: ； In the formula, To balance J 1. J 2 and J The parameters of 3 are obtained, and the optimization problem is solved by solving the improved Lagrangian objective function to obtain the 3rd parameter. d First mode and the updated residual signal ; Step 1.5: Repeat the modal extraction and optimization steps until the residual signal reaches the set threshold or a specified number of modes are extracted, and output the final vibration modal components: ； Similarly, all remaining non-electrical signals are also decomposed according to steps 1.1 to 1.5 above, resulting in... , ... ; Step 2 involves performing CW-SGMD decomposition on the electrical signals of the pumped storage unit, including the following steps: Step 2.1: Construct the trajectory matrix. For any electrical quantity signal, for a given signal... x = x 1, x 2, ..., x n According to Takens' embedding theorem, one-dimensional data can be reconstructed into multi-dimensional signals: ； In the formula, d For the embedding dimension; τ For delay time; m = n -( d -1) τ ; X The reconstructed multidimensional data matrix; x These are data points in the original one-dimensional time series data; Step 2.2: Perform symplectic geometric similarity transformation based on the trajectory matrix. X Constructing the covariance matrix A and Hamilton matrix M They are respectively: ； ； make According to the definition of the Hamiltonian matrix, the symplectic orthogonal matrix is ​​obtained. Q for: ； In the formula, B It is an upper triangular matrix; λ ( A )= λ ( B )= λ (2) X ); Based on the properties of the Hamiltonian matrix, calculate B eigenvalues λ 1, λ 2, ..., λ d ,get X eigenvalues σ i = λ i , i =1, 2, ..., d ,in: λ i Sort in descending order; denoted as Q i , i =1, 2, ..., d , is a matrix A The eigenvectors corresponding to the eigenvalues ​​of , then The initial single-component matrix is ​​obtained. Z i Then the trajectory matrix Z Represented as: Z = Z 1+ Z 2+…+ Z d ; Step 2.3: After the above transformations, the dimension of the initial component matrix is ​​obtained as follows: m × d After diagonal averaging, the matrix is... Z i Transform into d Group length is n Time series, this d The sum of the time series is the original signal. For ease of calculation, the elements of the matrix are defined as follows: Z ij For any initial single component Z i ,definition: ； In the formula: This indicates taking the smaller value between m and d; This indicates taking the larger value between m and d; These are the elements of the transformed matrix; if m ≤ d, then... equal to z of the original matrix ij If m > d, then equal to z of the original matrix ji Then, by taking the diagonal average, we get: ； In the formula, the first row of the equation represents the expression for each k Starting from the top left corner of the matrix, take the first... k The formula for each of the diagonal elements is calculated; the second line of the formula represents the average of each... k Take the front The formula for the diagonal elements is calculated, and their average is given; the formula in the third row represents... n It is a matrix Z The number of columns; for each k ,from Start by taking the last diagonal element and calculating its average; Introducing categorical weighting into the diagonal average for each index q or interval k Assigning exponential decay weights To reflect the importance of different categories, the revised formula is as follows: ； get d The group of single-component signals are: ; Step 2.4: Similar Component Recombination. The components decomposed by the SGMD method are not independent of each other and contain components with the same frequency and features. It is necessary to perform similarity analysis on each component and merge them to obtain SGC. i Then from the original signal x Remove SGC i The residual signal is denoted as g h ,Right now: ； In the formula, h This represents the number of iterations. The iteration stops when the normalized mean square error (NMSE) of the residual signal and the original signal is less than a given threshold. ； In the formula, for each iteration, the residual error is calculated as the square of the difference between the predicted value and the actual value. Sum the squares of all residual errors to obtain Divide the sum of squared residual errors by the sum of the original data. The normalized mean square error (NMSE) is obtained. When the iteration stops at the threshold condition, we get: ； In the formula, N To obtain the number of SGC components; The residual components are used to obtain the final decomposition result of the current signal. Similarly, all remaining electrical quantity signals are decomposed according to steps 2.1 to 2.4, resulting in... ... ; The specific process for establishing the VS-DA-TMMODE fault diagnosis model is as follows: Step 6.1: Input Data Multi-scale representations are generated after sampling: , among which is Um No. m Scale data, scale m The downsampling step size is 2, for each scale m ,have: ； In the formula, Conv represents the convolution operation; stride=2 indicates that the stride of the convolution operation is 2. Step 6.2: At the coarsest scale Channel mixing is performed, and a variational self-attention mechanism is applied: ； In the formula, Q M , K M , V M It is by The query, key, and value matrix obtained by linear projection; Step 6.3: Embed all multi-scale time series into the deep pattern set, where: ； Step 6.4: For each scale m Applying Fast Fourier Transform to extract the pre-processed material K One main frequency: ； In the formula, A For amplitude; f k For frequency; p k = T / f k The period length; Step 6.5: For each scale m Time series remodeling K Multi-resolution temporal images: ； In the formula For each time frame, a biaxial attention mechanism is applied: ； ； In the formula, Q col , K col , V col It refers to the query, key, and value of column axis attention; Q row , K row , V row It refers to the query, key, and value of the row axis attention; Step 6.6: For each scale m, sum the weighted representations of the K periods. ⊙ represents element-wise multiplication; after weighted processing, a multi-scale feature set is obtained: The feature at each scale is processed by multiple heads, and the results are integrated to obtain the final output: ； In the formula, Head m It is a linear head layer at the m-th scale; Ensemble is an ensemble method.

2. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, The non-electrical signals of the pumped storage unit in step 1 include vibration, sway, rotational speed, winding temperature, guide bearing temperature, and water pressure.

3. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, The electrical signals in step 2 include voltage, current, and partial discharge.

4. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, The environmental signals in step 3 include ambient temperature, ambient humidity, and voiceprint.

5. The method for fault diagnosis of a pumped storage unit according to claim 4, characterized in that, AF filtering of environmental quantity signals includes the following steps: Step 3.1: For any environmental quantity signal, for the signal... First, calculate the autocorrelation function: ； Pick τ Within a reasonable range, find the peak position of the main cycle; Step 3.2: Construct an autocorrelation filter, select an FIR filter, and use the autocorrelation function as the filter coefficients. And perform normalization: ； Step 3.3: Perform convolution filtering on the signal using the filter described above. For signal Perform convolution to obtain the filtered signal: ； result The filtered signal retains patterns with strong autocorrelation while suppressing short-term random noise. Similarly, the remaining environmental signals are also filtered according to steps 3.1 to 3.3, and the processed signals are as follows: , ... .

6. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, Step 4 specifically includes the following steps: Step 4.1: Let the principal components be... Secondary components are ~ The remaining variables are ~ ; Step 4.2: Perform linearly constrained minimum variance weighting on the above signals. The modes corresponding to the principal components and the modes corresponding to all secondary components are then subjected to linearly constrained minimum variance weighting. The remaining variables are embedded as biases into the weighting process. The process is as follows: ； In the formula, The data is weighted. As weight; For bias; As a secondary component; For the remaining variables; i Indicates an index for the data; Step 4.3: Introduce Lagrange multipliers a and b Construct the Lagrange function: ； In the formula, For variance; a The constraint that corresponds to a sum of weights of 1. b The constraint that the sum of the weights of the primary components is greater than the sum of the weights of the secondary components; Step 4.4: For each weight and bias Find the partial derivative: Solving the above system of partial derivative equations yields the optimal weights. and bias Finally, combining the above steps, we obtain the final weighted data: .

7. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, The extraction of weighted data and inverse FSFC features includes the following steps: Step 5.1: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] Represented as a linear combination of basis functions, using the first k Using principal components to approximate the data: ； In the formula, For the first j The modal component in the ... s Scores on each principal component; It is the first s The principal component basis functions are represented as the score matrix of all modal components. ,in p It is the number of modal components. k It is the number of principal components in each modal component, and the score vector is represented as ; Step 5.2: The FSFC method performs feature selection by minimizing the following objective function: ； In the formula, It's a category label. X ( i ) is a matrix X The i OK, B j yes B The corresponding number in the middle j Subvectors of each modal component ω j It is the feature weight. λ 1 and λ 2 is the regularization parameter; Step 5.3: Construct the dual problem for optimization: ； In the formula, V and Z As dual variables; and These are the dual functions of the loss function and the regularization term, respectively. Step 5.4: Introduce the Lagrange function, for Taking its partial derivative yields the weight vector for each modal component. : ； In the formula, For disciplinary parameters; Step 5.5: If If the value is 0, it means that the mode is not important and can be removed. Otherwise, select the modal components with non-zero residual coefficients as the m extracted feature components, and finally obtain the dataset after feature selection: , as input to the pumped storage fault diagnosis model.

8. The method for fault diagnosis of a pumped storage unit according to claim 1, characterized in that, The RRTO optimization algorithm optimizes the model parameters. ω and ρ The specific process is as follows: Step 7.1: Initialize parameters ω , ρ Their value ranges are [L] ω U ω ] and [L ρ U ρ Initialize the position of each individual in the population: ； In the formula, r i,ω and r i,ρ It is a random number in the range (0,1]. Step 7.2: Calculate the step size adjustment function based on the adaptive step size walking strategy. As the number of iterations t increases, dynamically adjust the parameters related to the step size. K and E The individual location will be updated accordingly: ； In the formula S Step size, r 1 is a random number between (0, 1]. C It is the step size penalty factor; Step 7.3: Calculate the difference between the optimal individual and the current individual based on the adaptive step size strategy of absolute difference: according to the current optimal position. X best and the individual's current location X i The difference is used to update the individual's position: ； In the formula, This is an adaptive adjustment factor; Step 7.4: Boundary-based adaptive step size strategy, boundary detection, and individual update method are as follows: ； In the formula, The phase angle is used to calculate the updated individual; Step 7.5: Repeat steps 7.2-7.3 until the maximum number of iterations is reached or the convergence condition is met. Finally, output the final optimal parameter combination. ,in Therefore and The objective function is a variable.

Images