A method for denoising a sound and vibration signal of a hydraulic turbine, a storage medium and a monitoring system

Abstract

The application discloses a water turbine acoustic vibration signal denoising processing method, a storage medium and a monitoring system, belongs to the field of signal processing and water turbine mechanical state monitoring, and solves the problems of existing methods, such as dependence on artificial experience, easy loss of weak fault characteristics under strong noise, and unstable effect. The application first collects acoustic vibration signals and performs EEMD decomposition, relies on a correlation coefficient to screen effective modes to complete initial denoising; a sample entropy and spectral kurtosis comprehensive fitness function is constructed, a genetic algorithm is adopted to adaptively optimize SVD embedding dimension and singular value reservation order; and optimal parameters are used to implement SVD secondary denoising to obtain a pure signal. Layered denoising can effectively suppress environmental and structural noise, get rid of the limitation of artificial parameter adjustment, and jointly evaluate two indexes to give consideration to denoising and fault characteristic reservation; global optimization of the algorithm avoids local optimization, is suitable for adapting to non-stationary acoustic vibration signals under variable working conditions, can significantly improve the signal-to-noise ratio and weak fault identification degree, is easy to be programmed and integrated into a water turbine online monitoring system, and reduces the risk of missed diagnosis and misdiagnosis of a unit fault.

Description

Technical Field

[0001] This invention relates to the field of signal processing and mechanical condition monitoring technology, and in particular to a method for noise reduction processing of acoustic and vibration signals of a water turbine, a storage medium, and a monitoring system. Background Technology

[0002] Hydropower turbines are the core energy conversion equipment in hydroelectric power plants. These units operate continuously under complex conditions of varying head, flow rate, and load. During operation, issues such as fluid cavitation, blade wear, guide vane failure, and bearing damage generate characteristic acoustic and vibration pulse signals. The acoustic and vibration signals generated by hydropower turbines during operation naturally possess nonlinear and non-stationary characteristics. During on-site data acquisition, multiple interference noises, including water flow turbulence, mechanical impact, powerhouse structural conduction, and electromagnetic radiation, are superimposed. This causes weak fault impact characteristics representing early-stage damage to be completely submerged by background noise, making them impossible to directly identify. This easily leads to missed or misdiagnosed faults, shortening the unit's operation and maintenance cycle and equipment lifespan. Therefore, noise reduction processing of hydropower turbine acoustic and vibration signals under strong noise conditions and accurate extraction of early weak fault characteristics are core technical challenges that urgently need to be overcome in the field of hydraulic unit condition monitoring and fault diagnosis.

[0003] Currently, mainstream noise reduction technologies in the industry are divided into three main approaches: transform domain noise reduction, adaptive mode decomposition (EMD) noise reduction, and matrix decomposition (MDD) noise reduction. Each approach has mature research and engineering application solutions, but they all have inherent limitations in the complex acoustic and vibration signal scenarios of hydro turbines. Existing representative solutions employ CEEMDAN (Complete Ensemble Empirical Mode Decomposition with Adaptive Noise) + Singular Entropy (SVD), SVD pre-processing + VMD autocorrelation screening, and EMD + cross-entropy SVD combined noise reduction approaches. While these solutions achieve signal-noise separation to some extent, they still suffer from drawbacks such as strong reliance on manual parameters, a single evaluation dimension, insufficient adaptability, and poor weak fault retention. The following sections will elaborate on these shortcomings in conjunction with general noise reduction algorithms and existing technical solutions: I. Inherent Defects of Various Basic Noise Reduction Algorithms 1. Wavelet threshold transform domain noise reduction The core principle is to perform multi-scale frequency band decomposition using wavelet basis and reconstruct the signal after threshold filtering of high-frequency detail coefficients. However, the noise reduction effect relies entirely on manually selecting the wavelet basis, the number of decomposition levels, and the soft and hard thresholds, without any adaptive matching rules. For broadband fault impact signals from water turbines, it is prone to problems such as excessive removal of high-frequency fault features and residual low-frequency noise, resulting in poor noise reduction stability.

[0004] 2. EMD series adaptive decomposition algorithms (EMD / EEMD / CEEMDAN) EMD relies on extreme value envelope interpolation to adaptively decompose the signal, but it suffers from modal aliasing, endpoint distortion, and ambiguity in the physical meaning of each component. EEMD can alleviate aliasing by adding noise multiple times and averaging, but the effective IMFs after decomposition can only be manually identified and screened, and residual noise in the signal cannot be completely eliminated. CEEMD improves the decomposition precision by progressively adding noise, but the computational load increases exponentially, and the decomposition process is prone to generating false modes without actual physical meaning, interfering with fault feature identification. VMD achieves frequency domain separation through variational constraints and has excellent anti-aliasing theoretical performance, but the number of decomposed modes K and the penalty factor α need to be manually set. Parameter imbalance can lead to under-decomposition or over-decomposition, directly losing fault frequency band information.

[0005] 3. SVD Matrix Decomposition and Noise Reduction The high-dimensional Hankel matrix is ​​reconstructed using a one-dimensional temporal phase space, and noise filtering is achieved by singular value truncation. The noise reduction effect is determined by two core parameters: the embedding dimension and the singular value retention order. Existing general solutions lack a global adaptive optimization mechanism and can only rely on repeated manual calculations and parameter adjustments. The noise reduction effect fluctuates greatly under varying load and noise intensity conditions of the turbine, resulting in insufficient robustness.

[0006] 4. Limitations of traditional single evaluation indicators Most existing noise reduction solutions only use one of the following indicators to evaluate the noise reduction quality: signal-to-noise ratio, single entropy value, or kurtosis. Pursuing only the signal-to-noise ratio can easily lead to over-smoothing and the erasure of weak fault pulses. Relying solely on kurtosis to enhance the impact will retain a large amount of random clutter, making it impossible to simultaneously address noise suppression and fault feature retention. When processing early weak fault signals, it is very easy to lose key diagnostic information.

[0007] II. Technical Shortcomings of Existing Comparative Patent Solutions 1. CEEMDAN-PE-SVD - Singular Entropy Noise Reduction For example, CN118410278A discloses an adaptive noise reduction method for vibration signals based on CEEMDAN-PE-SVD-singular entropy. This method uses CEEMDAN decomposition + permutation entropy (PE) to screen components, and then determines the SVD truncation order through singular entropy. The defects are as follows: (1) CEEMDAN has huge computational overhead, and the real-time computing pressure of the embedded monitoring equipment is high; (2) It only relies on permutation entropy to screen IMFs, without combining the original signal correlation to remove noise modes, and a large number of low-correlation clutter components enter the subsequent SVD process; (3) The SVD embedding dimension is fixed and there is no adaptive optimization logic, resulting in poor phase space reconstruction fit; (4) The evaluation only relies on the single dimension of singular entropy, which cannot quantify the overall messiness of the signal and has limited weak fault retention capability.

[0008] II. SVD pre-processing + VMD autocorrelation EFI index noise reduction For example, CN106446829A discloses a method for noise reduction of vibration signals of hydropower units based on SVD and VMD modal autocorrelation analysis. This scheme first performs SVD pre-filtering, then VMD decomposition, and selects effective modes through energy concentration index (EFI). The defects are as follows: (1) The noise reduction order is SVD first and then VMD. The original noisy signal is first truncated by matrix, which is very easy to lose weak fault energy in advance; (2) The key parameters of VMD are all set manually, without adaptive optimization methods; (3) The components are judged only by autocorrelation energy concentration, without quantifying the strength of frequency domain impact characteristics. Low amplitude cavitation and wear characteristics of the turbine are easily judged as noise and removed; (4) There is no multi-index joint optimization mechanism throughout the process. Parameter matching relies on manual experience, and the adaptability to working conditions is weak.

[0009] 3. EMD-Cross-Entropy-SVD Noise Reduction For example, CN115574922A discloses a method and system for noise reduction of vibration signals of hydropower units based on cross-entropy. This scheme uses basic EMD decomposition, cross-entropy to screen components, and then SVD for secondary noise reduction. The defects are as follows: (1) The original modes of basic EMD are severely mixed, and EEMD is not used to improve the decomposition stability, resulting in a large amount of noise and fault components mixed together; (2) Cross-entropy only measures the difference between the component and the original signal distribution, and cannot characterize the frequency domain fault impact intensity; (3) The SVD embedding dimension and singular truncation order are fixed, and there is no global optimization process, so the noise reduction effect is unstable under complex noise conditions; (4) Only the single index of entropy is used for evaluation, and it is impossible to simultaneously balance the signal complexity and fault characteristic amplitude.

[0010] A comprehensive analysis of general noise reduction algorithms and the aforementioned existing technical solutions reveals several shortcomings in current turbine acoustic and vibration noise reduction technologies: First, key parameters for modal decomposition and SVD reconstruction rely on manual adjustment, lacking an adaptive global optimization strategy, resulting in poor versatility across different units and operating conditions. Second, component screening and singular value truncation rely solely on single entropy and correlation-based indices, failing to simultaneously address signal clutter suppression and fault impact feature enhancement. Third, the decomposition and noise reduction architecture is unreasonable; some solutions involve matrix truncation followed by frequency domain decomposition, potentially leading to premature loss of early, weak fault energy. Fourth, high-precision decomposition algorithms such as CEEMD and VMD have excessively high computational costs, making them difficult to adapt to on-site online real-time monitoring equipment. Fifth, the lack of a layered, step-by-step noise reduction approach means that single-stage noise reduction cannot effectively separate multiple layers of clutter, including fluid noise, structural noise, and electromagnetic interference, resulting in limited improvement in fault identification under strong noise conditions.

[0011] Therefore, the industry urgently needs a hierarchical adaptive noise reduction scheme that uses EEMD for primary coarse noise reduction, genetic algorithm for adaptive optimization of SVD dual key parameters, and sample entropy + spectral kurtosis for joint optimization of multiple indicators. This scheme can eliminate dependence on manual parameters, fully preserve the early weak fault characteristics of the turbine in a multi-layered noise environment, and improve the robustness of noise reduction and the accuracy of fault diagnosis. Summary of the Invention

[0012] The technical problem to be solved by this invention is to provide a method, storage medium and monitoring system for noise reduction of acoustic and vibration signals of water turbines, which overcomes the defects of existing noise reduction methods for water turbines, such as parameters relying on human experience, poor adaptability to operating conditions, easy loss of weak fault features under strong noise, and unstable noise reduction effect. It achieves adaptive optimization of noise reduction parameters, and while suppressing environmental and structural noise, it completely preserves the weak impact features of early faults of water turbines, thereby improving the robustness of signal noise reduction and the reliability of fault diagnosis under complex operating conditions.

[0013] To achieve the above technical objectives, the present invention adopts the following technical solution: This invention provides a method for noise reduction processing of acoustic and vibration signals from a water turbine, comprising the following steps: Step 1: Collect the original acoustic and vibration signals under the operating state of the turbine, perform ensemble empirical mode decomposition on the original acoustic and vibration signals, and output multi-order intrinsic mode function components and residual components; ensemble empirical mode decomposition is performed by superimposing white noise on the original signal multiple times and performing empirical mode decomposition separately, and taking the average of all decomposition results to weaken the traditional EMD mode mixing defect.

[0014] Step 2: Calculate the correlation coefficient between each intrinsic mode function component and the original acoustic signal. Select effective modal components according to the preset correlation coefficient threshold. Superimpose and reconstruct the selected effective components to obtain the first noise-reduced signal. The correlation coefficient calculation logic is as follows: calculate the mean of the original signal and the single-order modal component respectively. Calculate the difference between the sampling points of the two types of signals and their respective means. Multiply and accumulate the corresponding differences to obtain the covariance term. Calculate the sum of squares of the two sets of differences respectively. Multiply the sums of squares and take the square root to obtain the standard deviation product term. Divide the covariance term by the standard deviation product term to obtain the correlation coefficient of the corresponding modal component. Only retain the components with a correlation coefficient greater than the preset threshold as effective intrinsic mode function components and remove the noise-dominant mode.

[0015] Step 3: Construct a comprehensive fitness function by combining sample entropy and spectral kurtosis, and use a genetic algorithm to adaptively optimize the two key parameters of the required embedding dimension and singular value retention order for singular value decomposition, and output the globally optimal parameter combination; Among them, the sample entropy calculation logic is as follows: a uniform similarity tolerance is set, and the probability of signal matching patterns with the length as the base length and the base length plus one is calculated respectively. The natural logarithm of the two probabilities is divided and the opposite number is taken. The smaller the value, the stronger the signal regularity, which is used to measure the signal noise complexity. Spectral kurtosis calculation logic: Extract the signal amplitude spectrum and calculate the mean of the amplitude spectrum. Calculate the difference between each amplitude spectrum data and the mean. Calculate the fourth power cumulative value and the second power cumulative value of the difference respectively. Divide the fourth power cumulative value by the square of the second power cumulative value to obtain the spectral kurtosis. The higher the value, the more significant the frequency domain fault impact characteristics. The overall fitness function construction logic is as follows: set weight coefficients to balance sample entropy and spectral kurtosis, and use the product of the weight coefficients and spectral kurtosis minus the current denoised signal sample entropy as the fitness value; Genetic algorithm execution flow: Initialize population size, number of iterations, crossover probability, mutation probability, define the value range of embedding dimension and singular value retention order, and randomly generate the initial population with two-dimensional parameter combinations; construct the phase space matrix for each set of parameters and complete singular value decomposition to reconstruct the temporary denoised signal, and substitute it into the fitness function to calculate the quality of individuals; retain high-quality individuals according to fitness, and sequentially perform crossover recombination and random parameter mutation operations to update the population; iterate until the maximum number of iterations is reached or the fitness converges, and output the optimal embedding dimension and optimal singular value retention order.

[0016] Step 4: Using the optimal parameters obtained in Step 3, perform secondary singular value decomposition denoising on the primary denoised signal to output the final clean denoised signal. This involves three steps: Step 4.1: Use the optimal embedding dimension to reconstruct the phase space of the first-order denoised signal and build a high-dimensional phase space trajectory matrix; Step 4.2: Perform singular value decomposition on the phase space trajectory matrix to obtain a set of singular values ​​sorted from high to low energy; Step 4.3: Select the principal singular value component corresponding to the optimal singular value retention order before sorting to reconstruct the signal, discard the remaining low-energy noise singular value components, and obtain the final denoised signal.

[0017] Furthermore, the present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for noise reduction processing of acoustic and vibration signals of a water turbine.

[0018] Furthermore, the present invention also provides a turbine condition monitoring system, comprising a signal acquisition module, a signal processing module, and a fault analysis module; the signal acquisition module acquires the original acoustic and vibration signals of the turbine; the signal processing module is equipped with a processor and runs the above-mentioned turbine acoustic and vibration signal noise reduction processing method to complete signal noise reduction; the fault analysis module identifies turbine faults and determines the unit's operating status based on the noise-reduced signal.

[0019] The present invention provides a method, storage medium, and monitoring system for noise reduction processing of acoustic and vibration signals from a water turbine, which has the following beneficial effects: 1. This invention employs a first-stage preprocessing noise reduction structure combining EEMD with correlation coefficient screening, which can adaptively decompose signals, extract effective components, alleviate the mode aliasing problem of traditional EMD, reduce noise interference on effective modes, and improve the effectiveness and stability of the initial noise reduction. 2. This invention quantifies the correlation between each IMF component and the original acoustic-vibration signal through correlation coefficients, automatically distinguishing between effective fault modes and pure noise modes, eliminating the need for manual subjective judgment of mode selection and reducing human error in signal processing. 3. This invention introduces a genetic algorithm to synchronously and globally adaptively optimize the SVD embedding dimension and singular value preservation order, abandoning the traditional manual parameter selection mode based on experience, eliminating the subjectivity and uncertainty of parameter selection, adapting to various operating conditions and turbine signals with different noise intensities, and improving the algorithm's robustness and adaptability to operating conditions. 4. This invention constructs a comprehensive fitness evaluation system based on two indicators: sample entropy and spectral kurtosis. It evaluates the noise reduction effect from two dimensions: signal complexity and frequency domain impulse characteristics. Unlike single evaluation indicators, this system simultaneously achieves noise suppression and fault feature retention, avoiding the over-filtering of weak fault features. 5. This invention sets weighted coefficients to balance two evaluation indicators: sample entropy and spectral kurtosis. The weights of these indicators can be flexibly adjusted based on the turbine model and fault type, broadening the algorithm's applicability and improving its universal adaptability to various types of turbine equipment. 6. The genetic algorithm of this invention fully executes selection, crossover, and mutation operations. The mutation mechanism randomly perturbs the parameters to broaden the search range, effectively avoiding the defect of getting trapped in local optima and ensuring that the globally optimal parameter combination is obtained. 7. This invention employs a two-tiered architecture of coarse EEMD noise reduction and fine GA-SVD noise reduction, which peels away environmental noise and structural noise in stages, achieving multi-gradient noise reduction from time-frequency decomposition to matrix reconstruction, significantly improving the signal-to-noise ratio and fault feature extraction capability under strong noise environments. 8. This invention utilizes sample entropy to quantify the complexity of signal noise and constrains irregular noise redundancy during the parameter optimization stage, effectively suppressing clutter interference generated by fluid disturbances and unit structural vibrations. 9. This invention utilizes spectral kurtosis to characterize the strength of frequency domain impulse features. The parameter optimization process actively amplifies the frequency domain peak value of the fault pulse, preserving and enhancing the modulation characteristics of weak early-stage faults in the turbine, thereby improving fault feature identification and stability. 10. This invention uses a genetic algorithm to match the optimal parameters for SVD phase space reconstruction, and the constructed phase space trajectory matrix closely matches the true dynamic laws of the signal, reducing the loss of useful signal information during the reconstruction stage. 11. This invention filters the principal singular value reconstructed signal based on the optimal singular value retention order, eliminates low-energy noise singular value components, and retains fault characteristic energy to the maximum extent while thoroughly filtering out noise: 12. The entire algorithm flow of this invention is logically structured, easy to program and develop, and can be directly embedded into hardware such as industrial control computers and embedded processors. It is suitable for online status monitoring scenarios of water turbines and has strong practicality for engineering implementation. 13. This invention is equipped with a computer-readable storage medium, which can completely embed the noise reduction algorithm program, allowing various monitoring devices to directly call it, reducing the secondary development cycle and development cost of the equipment: 14. This invention is used in conjunction with an integrated turbine condition monitoring system, which integrates signal acquisition, adaptive noise reduction, and fault analysis modules to achieve fully automated processing from acoustic and vibration signal acquisition to unit fault determination. 15. This invention designs a complete set of algorithm logic to address the nonlinear and non-stationary characteristics of the acoustic and vibration signals of water turbines. Compared with single noise reduction algorithms such as wavelet thresholding and VMD, the noise reduction effect under complex operating conditions of variable load and variable flow is less volatile, and the output results are more stable. 16. This invention eliminates the need for repeated manual adjustments to wavelet basis, decomposed mode number, penalty factor, and other parameters, significantly reducing on-site debugging workload, shortening signal preprocessing time in the early stages of fault diagnosis, and improving detection efficiency. 17. The layered noise reduction structure of this invention peels away multiple layers of noise interference layer by layer, purifying the original acoustic and vibration signals at multiple levels, further enhancing the signal-to-noise ratio improvement effect under complex noise conditions: 18. This invention can completely retain the early minor fault characteristics of the turbine, overcoming the shortcoming of traditional noise reduction methods that easily lose weak fault information, and reducing the probability of missing minor faults in the unit: 19. This invention can identify potential early damage to the unit, output fault warning signals in a timely manner, reduce equipment losses due to abnormal operation, and extend the service life of the entire turbine. 20. This invention relies on a genetic algorithm to automatically match the optimal processing parameters, eliminating the need for manual intervention in parameter adjustment throughout the process and reducing the reliance on the professional technical experience of operators for turbine fault diagnosis. Attached Figure Description

[0020] The present invention will be further described below with reference to the accompanying drawings and embodiments: Figure 1 This is a flowchart illustrating the technical process of the method of the present invention. Figure 2 This is a flowchart of the GA optimization process in this invention; Figure 3 This is a time-domain and frequency-domain comparison diagram of the denoised signal and the original noisy signal in Embodiment 2 of the present invention; Figure 4 This is a time-domain and frequency-domain comparison diagram of the noise-reduced signal and the original clean signal in Embodiment 2 of the present invention. Detailed Implementation

[0021] The technical solutions of the present invention will be further described below with reference to the embodiments and accompanying drawings: Example 1 like Figure 1 As shown in the figure, this embodiment provides a method for noise reduction processing of acoustic and vibration signals of a water turbine, including the following steps: Step 1: Collect raw acoustic and vibration signals during turbine operation. Ensemble Empirical Mode Decomposition (EEMD) is performed on the original acoustic vibration signal to obtain several Intrinsic Mode Function (IMF) components and residual components. Step 2: Calculate the original acoustic vibration signal and the values ​​of each IMF component. The correlation coefficient is used to filter effective IMF components based on a preset correlation coefficient threshold, and the filtered effective IMF components are used to reconstruct the first-stage noise-reduced signal. ; Step 3: Construct a comprehensive fitness function using sample entropy and spectral kurtosis, and use a genetic algorithm (GA) to adaptively optimize the key parameters of singular value decomposition (SVD) to determine the optimal embedding dimension. Order of preservation of optimal singular values ; Step 4: Based on the optimal embedding dimension Order of optimal singular value preservation For a single noise reduction signal The final denoised signal is obtained by performing singular value decomposition (SVD) and reconstruction. .

[0022] Furthermore, in step 1, the implementation method of the ensemble empirical mode decomposition (EEMD) is as follows: [The original acoustic signal is then processed]. White noise is superimposed multiple times and empirical mode decomposition (EMD) is performed on each of the multiple decompositions. The average value of the results obtained from the multiple decompositions is calculated, and finally, several IMF components and residual components are obtained.

[0023] Furthermore, in step 2, the correlation coefficient between the IMF component and the original acoustic vibration signal is calculated using the following formula: (1); In the formula, This refers to the sampling point data of the original acoustic vibration signal; The mean value of the original acoustic vibration signal; This refers to the sampling point data for a single-order IMF component; The mean of the single-order IMF components; This represents the total number of signal sampling points. For the first The correlation coefficients corresponding to the IMF components of order 1.

[0024] Furthermore, the screening rule for effective IMF components in step 2 is as follows: a correlation coefficient threshold is preset. , retain satisfaction The IMF components are selected as valid IMF components, and IMF components that do not meet the threshold conditions and are mainly noise are removed.

[0025] Furthermore, such as Figure 2 As shown, in step 3, the initialization operation of the genetic algorithm (GA) includes: setting the population size, maximum number of iterations, crossover probability, mutation probability, and defining the range of values ​​for the embedding dimension. The range of values ​​for the singular value preservation order .

[0026] Furthermore, in step 3, the initial population of the genetic algorithm uses two-dimensional vectors. express, For the embedding dimension, The order is preserved for singular values, and the parameters of all initial individuals are randomly generated within the corresponding value range.

[0027] Furthermore, in step 3, the formula for calculating the sample entropy is: (2); In the formula, To set a similarity tolerance, the length is The probability of signal matching mode; To set a similarity tolerance, the length is Signal matching pattern probability; sample entropy Used to characterize the complexity of a signal.

[0028] Furthermore, in step 3, the formula for calculating the spectral kurtosis is as follows: (3); In the formula, The signal amplitude spectrum; The mean of the amplitude spectrum; spectral kurtosis. Used to characterize the sharpness of the signal's frequency domain energy distribution and the intensity of fault impact characteristics.

[0029] Furthermore, in step 3, the expression for the comprehensive fitness function is: (4); In the formula, is a weighting coefficient used to balance the two evaluation metrics of sample entropy and spectral kurtosis; The sample entropy of the currently reconstructed signal; represents the spectral kurtosis of the currently reconstructed signal.

[0030] Furthermore, the iterative process of the genetic algorithm described in step 3 is as follows: Step 3.1: Based on the parameters corresponding to individuals in the population For a single noise reduction signal Construct the phase space matrix and perform SVD decomposition to reconstruct the temporary denoised signal. ; Step 3.2: Substitute the temporary noise-reduced signal into the comprehensive fitness function, calculate the fitness value of each individual in the population, and select the best individual in the current population; Step 3.3: Perform selection, crossover, and mutation operations on the population in sequence to complete the population update; Step 3.4: Repeat the above steps until the preset maximum number of iterations is reached or the comprehensive fitness function converges, and output the globally optimal parameter combination. .

[0031] Furthermore, in step 3.3, the selection operation involves sorting the individuals in the population according to their fitness values ​​and retaining individuals with excellent overall performance; the crossover operation involves recombining the parameters of the parent individuals according to a preset crossover probability to generate new offspring individuals; and the mutation operation involves adjusting the embedding dimension according to a preset mutation probability. Singular value preservation order By applying random perturbations, the global search range of the algorithm can be expanded.

[0032] Furthermore, step 4 specifically includes: Step 4.1: Adopt the optimal embedding dimension For a single noise reduction signal Phase space reconstruction is performed to construct a high-dimensional phase space trajectory matrix; Step 4.2: Perform singular value decomposition (SVD) on the phase space trajectory matrix to obtain several singular values; Step 4.3: Before keeping Each principal singular value component is analyzed, and the noise components corresponding to the remaining singular values ​​are discarded to complete signal reconstruction, outputting the final denoised signal. .

[0033] Furthermore, the aforementioned method for noise reduction processing of turbine acoustic and vibration signals is applied to turbine acoustic and vibration signals that have nonlinear and non-stationary characteristics and are superimposed with strong environmental noise and structural noise, in order to extract early and weak fault characteristics of the turbine.

[0034] Example 2 In another preferred embodiment, based on Embodiment 1, this embodiment provides a method for noise reduction of acoustic and vibration signals of a water turbine. This method targets the acoustic and vibration signals of early cavitation faults in a mixed-flow water turbine. The turbine operates with a rated head of 60m and variable load. The original acoustic and vibration signal is superimposed with water flow disturbance noise and unit structural vibration noise. The signal sampling frequency is 20kHz, and the sampling duration is 10s. All steps of this invention are fully executed. The specific parameters and operation procedures are as follows: Step 1: Collect the raw acoustic and vibration signals output by the acoustic sensors in the turbine's waterwheel chamber. Perform EEMD decomposition; add white noise N=80 times to EEMD, with the noise amplitude being 0.2 times the standard deviation of the original signal. After multiple EMD decompositions, take the average value to finally obtain the twelfth-order IMF component and the first-order residual component.

[0035] Step 2: Calculate the correlation coefficient between each IMF component and the original signal step by step according to formula (1), and set the correlation coefficient threshold. ; Filter out The 3rd to 9th order IMFs are used as effective components, and they are superimposed and reconstructed to obtain a first-order denoised signal v, removing the 1st and 2nd order high-frequency pure noise IMFs and the 10th to 12th order spurious mode components.

[0036] Step 3: Adaptively optimize the SVD embedding dimension using a genetic algorithm. Singular value preservation order GA initialization parameters: population size 30, maximum number of iterations 50, crossover probability 0.7, mutation probability 0.05; embedding dimension interval [5, 30], singular value order interval [2, 15]; set matching length m=2, similarity tolerance is signal standard deviation 0.15, weight coefficients... ; Calculate the overall fitness function according to formula (4): (5); Step 3.1: For each group in the population Construct the phase space matrix and reconstruct it using SVD to generate a temporary denoised signal. ; Step 3.2: Calculate the sample entropy and spectral kurtosis corresponding to each group of parameters according to formulas (2) and (3), respectively, substitute them into formula (5) to solve for the fitness value, and select the current best individual; Step 3.3: Update the population using roulette wheel selection, single-point crossover, and random perturbation mutation; Step 3.4: Iterate until the fitness converges on the 32nd iteration, and output the globally optimal parameters. .

[0037] Step 4: Apply optimal parameters Secondary noise reduction: Step 4.1: Reconstruct the phase space of the first-order denoised signal with an embedding dimension of 14 to generate a high-dimensional trajectory matrix; Step 4.2: Perform SVD decomposition on the trajectory matrix to obtain a sequence of singular values ​​arranged in descending order of energy; Step 4.3: Retain the first six principal singular values ​​to reconstruct the signal, discard the remaining noise singular values, and output the final denoised signal. .

[0038] Plot the time-frequency diagram of the denoised signal, such as... Figure 3 and Figure 4 As shown, compared with the original noisy signal, fluid clutter and structural noise are significantly suppressed, the peak value of cavitation fault impact pulse is significantly highlighted, the signal-to-noise ratio is improved by 12.8dB, and early weak cavitation fault characteristics can be directly identified.

[0039] Example 3 In another preferred embodiment, based on embodiment 1, such as Figure 1 As shown, this embodiment provides a computer-readable storage medium, which is a non-volatile computer-readable storage medium. The medium stores a computer program, and the program code fully implements the noise reduction processing method for the acoustic and vibration signals of a water turbine as described in Embodiments 1 and 2.

[0040] A computer-readable storage medium includes, but is not limited to, the following types: USB flash drive, solid-state drive (SSD), hard disk drive (HDD), read-only optical disc (CD-ROM), digital video disc (DVD-ROM), flash memory card (SD card, Secure Digital card, TF card, Trans Flash card, CF card, Compact Flash), embedded flash memory chip, programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), disk, optical disc, and cloud storage server storage partition.

[0041] Various types of storage media can be connected to the industrial control computer for turbine monitoring and embedded data acquisition terminals. The program modules within the media are divided as follows: 1. Signal Reading and EEMD Decomposition Module: Reads the acoustic and vibration time-series data collected by the sensor, performs EEMD decomposition with white noise averaging, and outputs all IMF components; 2. Correlation coefficient screening and reconstruction module: It has built-in correlation coefficient calculation logic, supports manual input or adaptive matching of correlation coefficient thresholds, automatically filters effective IMFs and generates a noise-reduced signal; 3. GA parameter optimization module: Built-in sample entropy and spectral kurtosis calculation program and comprehensive fitness function, encapsulates complete iterative logic of selection, crossover and mutation, and automatically outputs the optimal embedding dimension and singular value retention order; 4. SVD Secondary Noise Reduction and Reconstruction Module: Receives the optimal parameters output by GA, completes phase space matrix construction, singular value decomposition, principal component reconstruction, and outputs a clean noise-reduced signal; 5. Result Output Module: Stores the denoised signal, sample entropy, spectral kurtosis, and optimal parameters to the local hard drive, and supports exporting time-frequency waveforms.

[0042] When the industrial control computer processor reads and runs the program in this storage medium, it fully executes all the noise reduction steps in Example 1 without the need to write additional algorithm code; it can adapt to the acoustic and vibration data of different models of mixed-flow and axial-flow turbines, shortening the equipment secondary development cycle by 70%, and can process historical acoustic and vibration monitoring data offline in batches.

[0043] Example 4 In another preferred embodiment, based on embodiments 1 to 3, this embodiment provides a turbine condition monitoring system, including a signal acquisition module, a signal processing module, and a fault analysis module, with each module operating collaboratively: 1. Signal acquisition module: Three piezoelectric acoustic vibration sensors are installed in the turbine waterwheel chamber, top cover, and tailrace pipe, along with a signal conditioning circuit and a 20kHz synchronous acquisition card. The original acoustic vibration signals under the variable load conditions of the unit are acquired in real time, and the digital timing signals are transmitted to the signal processing module. 2. Signal processing module: Equipped with an ARM (Advanced RISC Machines) industrial processor, the processor is connected to a computer-readable storage medium as described in Example 2, and runs the noise reduction program in the medium; it executes a two-stage noise reduction process in real time, automatically completes EEMD mode screening, GA adaptive parameter optimization, and SVD secondary noise reduction, and outputs one set of clean acoustic vibration signals with noise reduction completed per second; 3. Fault Analysis Module: Includes a built-in fault feature database with kurtosis thresholds for typical faults such as cavitation, blade wear, and guide vane jamming. After receiving the noise-reduced signal from the signal processing module, it automatically extracts frequency domain impact features, compares them with database thresholds to determine fault classification, and simultaneously pushes fault warnings and stored time-frequency comparison graphs to the maintenance host computer. The fault warnings and stored time-frequency comparison graphs are shown below. Figure 3 and Figure 4 As shown.

[0044] The complete system workflow is as follows: the sensor collects the raw acoustic and vibration signals → the signal processing module calls the noise reduction algorithm in the storage medium to complete the joint noise reduction → the fault analysis module identifies early weak faults based on the noise-reduced pure signal → the host computer displays the noise-reduced waveform and diagnostic results.

[0045] Under actual on-site testing conditions, the system can reliably identify minor blade wear faults in environments with fluctuating turbine loads and strong water flow noise. The fault misdiagnosis rate has been reduced from 21.3% for the traditional single SVD noise reduction scheme to 3.7%, enabling 24-hour uninterrupted automatic monitoring of the unit.

[0046] The signal acquisition module is used to acquire acoustic and vibration signals during the operation of the turbine in real time; the signal processing module is equipped with a processor, which runs a computer program to implement the turbine acoustic and vibration signal noise reduction processing method described in Examples 1 and 2, and performs noise reduction processing on the acquired acoustic and vibration signals; the fault analysis module completes turbine fault identification and operation status determination based on the noise-reduced signal.

[0047] In a preferred embodiment, the implementation of the ensemble empirical mode decomposition (EEMD) is as follows: [The text abruptly shifts to a different topic] ...to the original acoustic signal... White noise is superimposed multiple times, and Empirical Mode Decomposition (EMD) is performed on each decomposition. The average value of the results from multiple decompositions is calculated, ultimately yielding several IMF components and residual components. This approach, through the strategy of multiple superpositions of white noise and averaging, effectively counteracts the mode aliasing phenomenon caused by signal discontinuities in EMD decomposition, making the time-frequency distribution of each IMF component more concentrated and its physical meaning clearer. The introduction of white noise forces the signal to be uniformly distributed across different scales. After multiple decompositions, the noise components cancel each other out, while the true fault characteristics, due to their inherent correlation, are preserved and amplified in the mean. This provides a cleaner and more reliable modal component basis for subsequent correlation coefficient selection, significantly improving the reliability and consistency of the first-level noise reduction.

[0048] In the preferred embodiment, the correlation coefficient between the intrinsic mode function component and the original acoustic vibration signal is calculated as follows: First, the overall mean of the original acoustic vibration signal and the overall mean of the single-order intrinsic mode function component are calculated respectively; then, the difference between each sampling point of the original acoustic vibration signal and its mean, and the difference between each sampling point of the single-order intrinsic mode function component and its mean are calculated. The two sets of corresponding differences are multiplied sequentially and summed to obtain the covariance summation term; the sum of squares of the differences between each sampling point of the original acoustic vibration signal and the sum of squares of the differences between each sampling point of the single-order intrinsic mode function component are calculated respectively. The product of the two sums of squares is multiplied and the square root is taken to obtain the standard deviation product term; finally, the covariance summation term is divided by the standard deviation product term, and the result is the correlation coefficient of the corresponding intrinsic mode function component. The above settings adopt the classic calculation framework of Pearson correlation coefficient, which can quantify the linear correlation strength between each order IMF and the original signal with a unified numerical scale. The closer the correlation coefficient is to one, the richer the original signal information carried by the mode component, and the closer it is to zero, the more noise the component is. The calculation method has a clear formula and low computational load, making it easy to execute in real time on embedded devices. It provides an objective and quantifiable criterion for subsequent automated screening of effective modes, avoiding inconsistencies caused by subjective selection based on manual observation of waveforms.

[0049] In the preferred embodiment, the selection rule for the effective intrinsic mode function (IMF) components in step 2 is as follows: a correlation coefficient threshold is pre-set. Intrinsic Mode Function (IMF) components with correlation coefficients greater than this threshold are retained as valid IMF components. The system removes Intrinsic Mode Function (IMF) components whose correlation coefficients are less than or equal to a threshold and are primarily noise. This setting, through a preset threshold, enables automated batch screening of effective modes, eliminating the need for step-by-step manual assessment of modal attributes and significantly reducing subjective bias caused by human intervention. The threshold can be flexibly adjusted according to the actual noise level; in high-noise environments, the threshold can be appropriately lowered to avoid missing weakly fault-related modes, while in low-noise environments, the threshold can be appropriately raised to more thoroughly remove residual noise components. This rule, combined with the correlation coefficient calculation method, forms a complete screening chain from quantitative assessment to automated decision-making, making the first-stage noise reduction process highly repeatable and engineering-applicable.

[0050] In a preferred embodiment, the initialization operation of the genetic algorithm (GA) includes: setting the population size, maximum number of iterations, crossover probability, and mutation probability, while defining the range of values ​​for the embedding dimension. The range of values ​​for the singular value preservation order The above settings, by pre-defining the population size and number of iterations, achieve a reasonable balance between computational resources and optimization accuracy, avoiding insufficient search space coverage due to an excessively small population or unnecessary time consumption due to too many iterations. The genetic operations of crossover and mutation probabilities provide clear behavioral guidelines, while defining the value ranges of embedding dimension and singular value preservation order constrains the search range within the physically realizable parameter space, eliminating meaningless extreme values ​​and ensuring that the algorithm searches efficiently within a reasonable range from the outset, laying a standardized foundation for obtaining the global optimum.

[0051] In a preferred embodiment, the initial population of the genetic algorithm uses individuals that include the embedding dimension. Singular value preservation order Two-dimensional parameter combination This means that the parameters of all initial individuals are randomly generated within their corresponding value ranges. This setting encodes individuals using a two-dimensional parameter combination, ensuring that each individual directly corresponds to a complete set of SVD reconstruction parameters. The low dimensionality of the search space and the simplicity of the encoding method facilitate rapid convergence of the genetic algorithm. The random generation of parameters within their value ranges ensures a uniform distribution of the initial population in the search space, avoiding the risk of premature convergence due to excessive concentration of initial individuals in a localized area. Compared to high-dimensional encoding, two-dimensional encoding significantly reduces the computational overhead of a single iteration, enabling the algorithm to run efficiently on devices with limited computing power, such as embedded processors, thus meeting the real-time requirements of online monitoring systems for hydroelectric turbines.

[0052] In the preferred embodiment, the sample entropy is calculated as follows: under a set similarity tolerance, the matching pattern probabilities of two types of signals with lengths of a specified length and lengths of a specified length plus one are statistically analyzed; the natural logarithm of the two probabilities is taken after dividing the two probabilities, and then the negative of the result is taken to obtain the final sample entropy; the sample entropy is used to characterize the complexity of the signal; in the above settings, the sample entropy quantifies the regularity of the signal by comparing the pattern matching probabilities of adjacent length sequences. The lower the entropy value, the more regular the signal and the less noise component; the higher the entropy value, the more complex the signal and the stronger the noise interference. Incorporating it into the fitness evaluation system can guide the algorithm to tend to choose SVD parameter combinations that reduce signal complexity during parameter optimization, thereby avoiding the loss of useful information due to excessive smoothing while denoising. This index has low requirements for signal length and strong resistance to noise interference, making it suitable as one of the core evaluation dimensions for the optimization of noise reduction of hydro turbine acoustic and vibration signals.

[0053] In the preferred embodiment, the spectral kurtosis is calculated as follows: First, the amplitude spectrum of the signal is extracted, and the overall mean of the amplitude spectrum is calculated. The difference between each data point in the amplitude spectrum and its mean is calculated, and the sum of the fourth power and the sum of the square power of these differences are calculated respectively. The fourth power sum is divided by the square of the square power sum, and the resulting value is the spectral kurtosis. Spectral kurtosis is used to characterize the sharpness of the signal's frequency domain energy distribution and the intensity of fault impact characteristics. With the above settings, spectral kurtosis is extremely sensitive to impact components in the frequency domain. When there are periodic impacts caused by faults in the signal, the amplitude spectrum will show sharp peaks at the fault characteristic frequencies, and the spectral kurtosis value will increase significantly. Introducing it into the fitness function allows the genetic algorithm to actively retain or even enhance the frequency domain components corresponding to these impact characteristics during optimization, preventing the SVD denoising process from filtering out weak fault pulses as noise. This index calculation only involves statistical moment operations of the amplitude spectrum, resulting in high computational efficiency and forming a complementary evaluation in both the time and frequency domains with the sample entropy.

[0054] In the preferred embodiment, step 3 involves constructing the comprehensive fitness function by selecting weighting coefficients to balance sample entropy and spectral kurtosis. The product of the weighting coefficients and spectral kurtosis is subtracted from the sample entropy corresponding to the current signal, and the result is the value of the comprehensive fitness function. This configuration unifies sample entropy and spectral kurtosis under the same optimization objective, flexibly adjusting their contribution ratios through weighting coefficients. When the weights favor sample entropy, the optimization objective focuses on reducing the overall signal complexity and suppressing irregular noise; when the weights favor spectral kurtosis, the optimization objective focuses on enhancing frequency domain impact characteristics and protecting fault pulses. This dual-index combined evaluation method overcomes the potential bias of single-index optimization, achieving the best trade-off between denoising capability and feature preservation capability in the SVD parameter optimization result, significantly improving the fault identifiability of the denoised signal.

[0055] In the preferred embodiment, in step 3.3, the selection operation involves sorting the individuals in the population according to their fitness values ​​and retaining individuals with excellent overall performance; the crossover operation involves recombining the parameters of the parent individuals according to a preset crossover probability to generate new offspring individuals; and the mutation operation involves adjusting the embedding dimension according to a preset mutation probability. Singular value preservation order Random perturbation is applied to expand the global search range of the algorithm. The above settings include selection operations to ensure the preservation and accumulation of optimal parameter combinations in each generation of the population, crossover operations to generate new parameter combinations through parent parameter recombination, increasing the coverage density of the search space, and mutation operations to randomly perturb the parameters with a certain probability, allowing the algorithm to escape the current local optimum and explore better solutions. These three operations work together to give the genetic algorithm powerful global search capabilities within a limited number of iterations. The final output SVD parameter combination can simultaneously meet the dual requirements of noise reduction and feature preservation, avoiding getting trapped in suboptimal solutions that would result in substandard noise reduction quality.

[0056] In a preferred embodiment, the turbine acoustic and vibration signal noise reduction method is applied to turbine acoustic and vibration signals exhibiting nonlinear and non-stationary characteristics, superimposed with strong environmental and structural noise, to extract early, subtle fault features of the turbine. Given that turbines are subjected to multiple sources of interference during actual operation, such as water flow turbulence, tailrace pressure pulsation, and mechanical vibration, the acoustic and vibration signals exhibit significant nonlinear and non-stationary characteristics, making them difficult to effectively handle using traditional linear filtering methods. This invention addresses these signal characteristics by employing a hierarchical strategy combining EEMD adaptive decomposition and GA-optimized SVD. This strategy can gradually peel away strong environmental and structural noise without relying on the assumption of signal stationarity, effectively separating early, subtle fault features submerged in noise. This provides a high-quality signal foundation for turbine condition monitoring and fault early warning, significantly reducing the risk of missing minor unit faults.

[0057] In summary, this invention proposes a noise reduction method, storage medium, and monitoring system for acoustic and vibration signals of water turbines. First, the acoustic and vibration signals are acquired and decomposed using EEMD. Effective modes are selected based on correlation coefficients to complete the initial noise reduction. Then, a fitness function combining sample entropy and spectral kurtosis is constructed. A genetic algorithm is used to adaptively optimize the SVD embedding dimension and singular value retention order. The optimal parameters are then used to perform a second SVD noise reduction to obtain a clean signal. This overcomes the shortcomings of existing methods, such as parameter dependence on human experience, poor adaptability to operating conditions, easy loss of weak fault features under strong noise, and unstable noise reduction effect.

[0058] At the technical architecture level, this invention combines EEMD with correlation coefficient screening and SVD optimized by genetic algorithm to construct a two-level hierarchical noise reduction architecture for hydro turbine acoustic and vibration signal processing, which differs from existing single noise reduction algorithms or simple cascaded schemes. EEMD is responsible for coarse noise reduction to remove high-frequency random noise, while GA-optimized SVD is responsible for fine noise reduction to separate structural noise. The two have clear division of labor and work progressively layer by layer. Compared with single algorithms such as wavelet thresholding and VMD, the noise reduction effect fluctuates less under varying load and flow conditions.

[0059] At the evaluation system level, this invention constructs a comprehensive fitness function by jointly using sample entropy and spectral kurtosis, and uses this function to drive a genetic algorithm to adaptively optimize key SVD parameters globally. This combination of evaluation system and optimization framework has not been publicly disclosed in the field of turbine noise reduction. Sample entropy constrains the noise reduction process from the dimension of signal complexity, while spectral kurtosis constrains it from the dimension of frequency domain impact characteristics, overcoming the one-sidedness that inevitably leads to optimization of a single index, and enabling the noise reduction result to simultaneously consider noise suppression and the preservation of weak fault features. Furthermore, the introduction of two weighted coefficient balancing indicators allows for flexible adjustment of the evaluation emphasis according to the unit model and fault type, improving the algorithm's adaptability to different equipment and fault scenarios.

[0060] At the parameter optimization level, this invention uses the Pearson correlation coefficient as the automated screening basis for each IMF component of EEMD, replacing manual subjective judgment of modality selection with quantified correlation degree. Furthermore, it fully incorporates the selection, crossover, and mutation operations of the genetic algorithm into the SVD parameter optimization process, achieving synchronous adaptive global optimization of two-dimensional parameters with both embedding dimension and singular value retention order. The mutation mechanism randomly perturbs the parameters to broaden the search range, effectively avoiding getting trapped in local optima, enabling the algorithm to obtain the globally optimal parameter combination even under complex conditions such as varying operating conditions and strong noise.

[0061] The above design completely eliminates the reliance on manual parameter tuning for the entire noise reduction process, and realizes automated decision-making from primary mode screening to secondary parameter optimization. While suppressing environmental noise and structural noise, it fully preserves the weak impact characteristics of early turbine faults, significantly improves the robustness of signal noise reduction and the reliability of fault diagnosis under complex operating conditions, and is easy to integrate into the turbine online monitoring system, reducing the risk of missed and misdiagnosed unit faults.

Claims

1. A method for noise reduction processing of acoustic and vibration signals from a water turbine, characterized in that, Includes the following steps: Step 1: Collect the original acoustic and vibration signals under the operating state of the water turbine, and perform ensemble empirical mode decomposition on the original acoustic and vibration signals to obtain several intrinsic mode function components and residual components; Step 2: Calculate the correlation coefficient between each intrinsic mode function component and the original acoustic vibration signal, filter the effective intrinsic mode function components according to the preset correlation coefficient threshold, and reconstruct the first noise-reduced signal using the filtered effective intrinsic mode function components; Step 3: Construct a comprehensive fitness function using sample entropy and spectral kurtosis, and use a genetic algorithm to adaptively optimize the key parameters of singular value decomposition to determine the optimal embedding dimension and the optimal singular value preservation order; Step 4: Based on the optimal embedding dimension and the optimal singular value preservation order, perform singular value decomposition on the first denoised signal and reconstruct it to obtain the final denoised signal.

2. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 1, the implementation method of the set empirical mode decomposition is as follows: white noise is superimposed multiple times on the original acoustic vibration signal and empirical mode decomposition is carried out respectively. The average value of the results obtained from multiple decompositions is calculated, and finally several orders of intrinsic mode function components and residual components are obtained.

3. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 2, the correlation coefficient between the intrinsic mode function component and the original acoustic vibration signal is calculated as follows: First, the overall mean of the original acoustic vibration signal and the overall mean of the single-order intrinsic mode function component are calculated respectively; then, the difference between each sampling point of the original acoustic vibration signal and its mean, and the difference between each sampling point of the single-order intrinsic mode function component and its mean are calculated. The two sets of corresponding differences are multiplied sequentially and summed to obtain the covariance summation term; the sum of squares of the differences between each sampling point of the original acoustic vibration signal and the sum of squares of the differences between each sampling point of the single-order intrinsic mode function component are calculated respectively. The product of the two sets of sums of squares is multiplied and the square root is taken to obtain the standard deviation product term; finally, the covariance summation term is divided by the standard deviation product term, and the result is the correlation coefficient of the corresponding intrinsic mode function component.

4. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 3, characterized in that, The selection rule for effective intrinsic mode function components in step 2 is as follows: a correlation coefficient threshold is preset, intrinsic mode function components with correlation coefficient values ​​greater than the threshold are retained as effective intrinsic mode function components, and intrinsic mode function components with correlation coefficient values ​​less than or equal to the threshold and which are mainly noise are eliminated.

5. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 3, the genetic algorithm initialization operation includes: setting the population size, maximum number of iterations, crossover probability, mutation probability, and defining the value range of the embedding dimension and the value range of the singular value retention order.

6. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 5, characterized in that: In step 3, the initial population individuals of the genetic algorithm are represented by a two-dimensional parameter combination that includes the embedding dimension and the singular value preservation order. The parameters of all initial individuals are randomly generated within the corresponding value range.

7. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 3, the sample entropy is calculated as follows: under the condition of setting similarity tolerance, the probability of matching patterns of two types of signals with a specified length and a specified length plus one is statistically analyzed; the natural logarithm is taken after dividing the two types of probabilities, and then the opposite number of the calculation result is taken to finally obtain the sample entropy; the sample entropy is used to characterize the complexity of the signal.

8. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 3, the spectral kurtosis is calculated as follows: First, the amplitude spectrum of the signal is extracted, and the overall mean of the amplitude spectrum is calculated; the difference between each data point of the amplitude spectrum and its mean is obtained, and the sum of the fourth power of the difference and the sum of the square power of the difference are calculated respectively; the sum of the fourth power is divided by the square of the sum of the square power, and the resulting value is the spectral kurtosis. Spectral kurtosis is used to characterize the sharpness of the signal's frequency domain energy distribution and the intensity of fault impact characteristics.

9. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, In step 3, the comprehensive fitness function is constructed as follows: two indicators, weight coefficient and spectral kurtosis, are selected. The product of the weight coefficient and spectral kurtosis is subtracted from the sample entropy corresponding to the current signal, and the result is the value of the comprehensive fitness function.

10. A method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, The iterative process of the genetic algorithm described in step 3 is as follows: Step 3.1: Based on the embedding dimension and singular value retention order parameter combination corresponding to the individual population, construct the phase space matrix of the first denoised signal and perform singular value decomposition to reconstruct the temporary denoised signal; Step 3.2: Substitute the temporary noise-reduced signal into the comprehensive fitness function, calculate the fitness value of each individual in the population, and select the best individual in the current population; Step 3.3: Perform selection, crossover, and mutation operations on the population in sequence to complete the population update; Step 3.4: Repeat steps 3.1 to 3.3 until the preset maximum number of iterations is reached or the comprehensive fitness function converges, and output the globally optimal combination of embedding dimension and singular value retention order.

11. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 10, characterized in that: In step 3.3, the selection operation is to sort the individuals in the population according to their fitness values ​​and retain individuals with excellent overall performance; the crossover operation is to recombine the parameters of the parent individuals according to a preset crossover probability to generate new offspring individuals; the mutation operation is to randomly perturb the embedding dimension and the singular value retention order according to a preset mutation probability to expand the global search range of the algorithm.

12. The method for noise reduction processing of acoustic and vibration signals of a water turbine according to claim 1, characterized in that, Step 4 specifically includes: Step 4.1: Use the optimal embedding dimension to reconstruct the phase space of the first-order denoised signal and construct a high-dimensional phase space trajectory matrix; Step 4.2: Perform singular value decomposition on the phase space trajectory matrix to obtain several singular values; Step 4.3: Retain the best singular values ​​that are ranked first. Retain the number of principal singular value components corresponding to the order of the retained singular values, discard the noise components corresponding to the remaining singular values, complete the signal reconstruction, and output the final denoised signal.

13. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, which, when executed by a processor, implements a method for noise reduction of acoustic and vibration signals of a water turbine as described in any one of claims 1 to 12.

14. A turbine condition monitoring system, characterized in that: It includes a signal acquisition module, a signal processing module, and a fault analysis module; the signal acquisition module is used to acquire acoustic and vibration signals during the operation of the turbine in real time; the signal processing module is equipped with a processor, which runs a computer program to implement the turbine acoustic and vibration signal noise reduction processing method according to any one of claims 1 to 12, and performs noise reduction processing on the acquired acoustic and vibration signals; The fault analysis module uses the noise-reduced signal to identify turbine faults and determine their operating status.

Images