Fan blade running state monitoring system and method based on voiceprint recognition

Abstract

The invention discloses a system and a method for monitoring the running state of a fan blade based on voiceprint recognition. The method comprises the following steps: acquiring voiceprint data and environment data during running of the fan blade through a data acquisition module; performing noise reduction processing by adopting an improved VMD method, and adaptively optimizing a decomposition modal number and a penalty factor through fuzzy entropy and an energy ratio; constructing a plurality of branch input data according to the voiceprint data; inputting the two types of branch data into a pre-trained multi-branch model, carrying out weighted fusion on similar results by combining a voiceprint signal-to-noise ratio and a defect confidence difference, and outputting a state result of a defect type and confidence; and when the confidence exceeds a threshold value, integrating the environment data, the blade parameters and the GAP features to generate positioning analysis data, determining a defect position through a defect positioning model, and triggering an alarm. According to the method, the feature quality is improved through adaptive noise reduction, the feature fusion effect is optimized through dynamic weight, the defect recognition accuracy is enhanced through confidence fusion, efficient positioning is achieved, and the monitoring efficiency and reliability are remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation equipment monitoring, and in particular to a system and method for monitoring the operating status of wind turbine blades based on voiceprint recognition. Background Art

[0002] With the rapid development of renewable energy, wind power, as a key component of clean energy, is playing an increasingly critical role in the global energy mix. As the core equipment in wind power systems, the operational reliability of wind turbines directly impacts power generation efficiency and operation and maintenance costs. Among wind turbine components, blades are key components that withstand complex aerodynamic loads. Exposure to high-speed rotation and harsh environmental conditions over extended periods makes them susceptible to failures such as cracks and corrosion, which can affect normal wind turbine operation and even cause major safety incidents. Therefore, real-time and effective monitoring of wind turbine blade operating status has become a crucial tool for the wind power industry to ensure equipment safety and improve operation and maintenance efficiency.

[0003] Existing technologies for monitoring the condition of wind turbine blades often ignore the correlation between various defect states and individual voiceprint features, resulting in an ineffective fusion of multiple voiceprint features, affecting the accuracy of defect results and, in turn, lowering the accuracy and efficiency of the condition monitoring method. Therefore, further improvements are needed in the voiceprint recognition-based method for monitoring the operating condition of wind turbine blades. Summary of the Invention

[0004] The purpose of the present invention is to provide a wind turbine blade operating status monitoring system and method based on voiceprint recognition, so as to solve the technical problem that the existing technology does not fully consider the correlation between multiple defect states and each voiceprint feature, resulting in poor multi-voiceprint feature fusion effect and decreased defect recognition accuracy, which in turn causes low accuracy and efficiency of the status monitoring method.

[0005] To achieve the above object, the technical solution provided by the present invention is:

[0006] A method for monitoring the operating status of a fan blade based on voiceprint recognition comprises the following steps:

[0007] S1. Acquire voiceprint data and temperature and humidity environmental data of the wind turbine blades during operation, perform noise reduction preprocessing on the voiceprint data using an improved VMD method to obtain noise-reduced voiceprint data; extract MFCC features, GFCC features, and Mel spectrogram time-frequency images from the noise-reduced voiceprint data;

[0008] S2. Calculate feature weights based on the defect category importance, defect feature sensitivity, and feature stability of the wind turbine blade, fuse the MFCC features and GFCC features to obtain branch input data 1; and use the Mel spectrogram time-frequency image as branch input data 2;

[0009] S3. Input the branch input data 1 and the branch input data 2 into the multi-branch result generation model respectively to obtain the output results corresponding to the branch labels, calculate the confidence weights according to the voiceprint signal-to-noise ratio and the confidence difference of the defect category, perform weighted fusion on the confidences of the same defect category, and obtain a status result including the defect category and its corresponding confidence;

[0010] S4. Obtain monitoring results including defect location information based on the status results and temperature and humidity environmental data. When the confidence level exceeds the defect threshold, trigger an alarm and locate the defect location.

[0011] To optimize the above technical solutions, specific measures taken also include:

[0012] The improved VMD method comprises:

[0013] Define the decomposition mode number K and penalty factor α, and set the minimum and maximum decomposition mode numbers;

[0014] Decompose the voiceprint data into K modal MTs in descending order according to the number of decomposition modes. k , and calculate the average fuzzy entropy value MHS of all modes under different decomposition mode numbers avg ;

[0015] An objective function corresponding to the decomposition mode number is constructed to find the optimal decomposition mode number. The objective function satisfies the following formula:

[0016]

[0017] Among them, k represents the number corresponding to the mode under the decomposition mode number;

[0018] Construct a penalty factor optimization function to find the optimal penalty factor. The penalty factor optimization function satisfies the following formula:

[0019]

[0020] Where j represents the decomposition mode number K opt The numbers corresponding to the lower mode, <, > represent the inner product operator symbols;

[0021] Distinguish noise-dominated modes from signal-dominated modes based on the fuzzy entropy and energy proportion of each mode;

[0022] The dominant mode of the signal is reconstructed to obtain the noise-reduced voiceprint data.

[0023] The feature weight calculation includes:

[0024] Construct the feature weight calculation function MTQF (QZX, QTM, TW), where QZX is the importance of the defect category, QTM is the sensitivity of the defect feature, and TW is the feature stability;

[0025] Substitute the defect category importance QZX, defect feature sensitivity QTM and feature stability TW into the function to calculate the MFCC feature weight;

[0026] The GFCC feature weight is obtained according to the difference between the sum of the feature weights and the MFCC feature weight.

[0027] Generating a status result according to branch input data includes:

[0028] Get the branch input data and output results corresponding to each branch label;

[0029] Divide the historical branch input data and its corresponding historical output results into training set, validation set, and test set, and perform data preprocessing;

[0030] Select two machine learning models as the base models of the multi-branch;

[0031] The basic model is trained using the training set corresponding to each branch, and the model hyperparameters are adjusted on the validation set corresponding to each branch to obtain the pre-trained model of each branch.

[0032] The performance of each pre-trained model is verified on the test set, and finally integrated to obtain a multi-branch result generation model.

[0033] The construction of the multi-branch result generation model includes:

[0034] Get the historical branch input data and its corresponding historical output results corresponding to the branch label;

[0035] Input each branch label and the corresponding branch input data into the multi-branch result generation model, and output the corresponding output result including the defect category prediction and its confidence;

[0036] Extracting the signal-to-noise ratio (SNR) of the voiceprint data and the confidence difference (QLZC) of different branches for the same defect category;

[0037] Construct the confidence weight function ZQF(SNR, QLZC);

[0038] Calculate the confidence weight of branch label one and the confidence weight of branch label two respectively;

[0039] The confidence levels of different branches of the same defect category are weighted and fused according to the calculated weights to generate the final status result including the defect category and its corresponding confidence level.

[0040] Generating monitoring results according to status results and environmental data includes:

[0041] Obtaining status results, environmental data, fan blade parameters, and branch input data within a time window;

[0042] Extracting a GAP feature of branch input data one from the multi-branch model;

[0043] If the defect category confidence in the status result is greater than the corresponding defect confidence threshold, a fan blade defect alarm signal is generated, and the environmental data, fan blade parameters, branch input data 1 and GAP features are integrated to generate positioning analysis data, which is input into the pre-trained defect localization model to obtain a monitoring result including defect location information;

[0044] Otherwise, no action is taken.

[0045] The defect location model obtains historical location analysis data and corresponding monitoring results, and divides the historical location analysis data and its corresponding historical monitoring results into a training set, a validation set and a test set for preprocessing;

[0046] Select an AI model as the base model, train it using the training set, and adjust the hyperparameters on the validation set to obtain a pre-trained model;

[0047] By verifying the performance of the pre-trained model on the test set, a trained defect localization model is finally obtained; the defect localization model is used to input localization analysis data and output defect location information.

[0048] As another important technical solution, the present invention also provides a wind turbine blade operating status monitoring system based on voiceprint recognition, comprising:

[0049] The data acquisition module is used to obtain the soundprint data generated by the wind turbine blades during operation through the data acquisition device, and extract the MFCC features, GFCC features and Mel spectrogram time-frequency images from them, and simultaneously obtain environmental data including temperature and humidity;

[0050] The data analysis module includes a data integration unit and a result generation unit;

[0051] The data integration unit calculates feature weights based on defect category importance, defect feature sensitivity, and feature stability, and constructs several branch input data based on the voiceprint data;

[0052] The result generating unit is used to generate a status result according to the branch input data, and generate a monitoring result and an alarm signal according to the status result and the environmental data;

[0053] The early warning module is used to issue an alarm prompt;

[0054] The data acquisition module, data analysis module and early warning module are connected in sequence.

[0055] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned method for monitoring the operating status of wind turbine blades based on voiceprint recognition is implemented.

[0056] The present invention also proposes a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the above-mentioned method for monitoring the operating status of wind turbine blades based on voiceprint recognition.

[0057] Compared with the prior art, the present invention has the following beneficial effects:

[0058] 1. This application constructs several branch input data based on voiceprint data; generates status results based on the branch input data; generates monitoring results based on the status results and environmental data, constructs a multi-branch result generation model, and dynamically fuses multiple features of different branches to make feature expression more effective. At the same time, the results of different branches are dynamically fused in the decision-making stage to make the final defect identification result more accurate and reasonable. At the same time, when defects exist, the defect position is located in a timely and effective manner, thereby improving the efficiency of the status monitoring method.

[0059] 2. This application achieves effective noise reduction of voiceprint data by improving the VMD algorithm, thereby maximizing the degree of noise reduction. At the same time, it implements adaptive selection of the number of decomposition modes and penalty factors, avoiding over-decomposition or under-decomposition caused by artificially setting the number of decomposition modes and penalty factors, improving the accuracy of noise and signal separation, and providing accurate data support for subsequent multi-feature extraction.

[0060] 3. This application considers several factors that affect MFCC features and GFCC features, constructs a feature weight calculation function for MFCC features, and dynamically adjusts the corresponding feature weights when MFCC features and GFCC features are fused, so that the fused features can monitor several defect types, thereby improving the accuracy of defect identification.

[0061] 4. This application obtains the status results corresponding to different feature data through a pre-trained multi-branch result generation model, and dynamically adjusts the confidence fusion weights between the results based on the relationship between the status results and the multiple branches, and independently calculates the weights for each defect category to solve the differences in noise sensitivity of different categories and improve the accuracy and robustness of blade defect detection results. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1This is a flow chart of the method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to the present invention.

[0063] Figure 2 A flow chart is generated for the results of the wind turbine blade operating status monitoring method based on voiceprint recognition of the present invention.

[0064] Figure 3 This is a schematic diagram of the principle of the fan blade operating status system based on voiceprint recognition of the present invention. DETAILED DESCRIPTION

[0065] The above contents of the present invention are further described in detail below in the form of specific implementation methods, but this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments. All technologies implemented based on the above contents of the present invention belong to the scope of the present invention.

[0066] like Figure 1 As shown, the present invention provides a method for monitoring the operating status of wind turbine blades based on voiceprint recognition, comprising:

[0067] S1. Acquire voiceprint data and environmental data including temperature and humidity during the operation of the wind turbine blades, perform noise reduction preprocessing on the voiceprint data using an improved VMD method to obtain noise-reduced voiceprint data; and extract MFCC features, GFCC features, and Mel spectrogram time-frequency images from the noise-reduced voiceprint data.

[0068] S2. Calculate feature weights based on defect category importance, defect feature sensitivity, and feature stability, fuse the MFCC features and GFCC features to obtain branch input data 1; and use the Mel spectrogram time-frequency image as branch input data 2;

[0069] S3. Input the branch input data 1 and the branch input data 2 into the multi-branch result generation model respectively, calculate the confidence weight according to the voiceprint signal-to-noise ratio and the confidence difference of the defect category, perform weighted fusion on the confidence of the same defect category, and obtain a status result including the defect category and its corresponding confidence;

[0070] S4. Obtain monitoring results including information on the location of defects based on the status results and environmental data. When the confidence level exceeds the defect threshold, trigger an alarm and locate the defect location.

[0071] In some implementations, voiceprint data and environmental data are acquired; the voiceprint data refers to the voiceprint generated by the wind turbine blades during operation, including environmental noise and the like.

[0072] Several branch input data are constructed based on the voiceprint data; the branch input data refers to the input data corresponding to several branch models when blade defect monitoring is performed.

[0073] Generate status results based on branch input data; status results refer to the results of defect recognition, including defect categories and their corresponding confidence levels.

[0074] The monitoring results are generated based on the status results and environmental data. The monitoring results refer to the status of the wind turbine blades, including information on the location of defects when defects exist.

[0075] In this embodiment, several branch input data are constructed based on the voiceprint data, including:

[0076] Get voiceprint data and defect category importance.

[0077] The denoised voiceprint data is obtained by performing a denoising preprocessing operation on the voiceprint data; the denoising preprocessing operation is achieved by an improved VMD method.

[0078] Extract MFCC features, GFCC features and Mel spectrogram time-frequency images corresponding to the voiceprint data.

[0079] Branch input data 1 is obtained by fusing MFCC features and GFCC features with corresponding feature weights; the Mel spectrogram time-frequency image is used as branch input data 2. Feature weights are calculated based on the importance of defect categories.

[0080] This embodiment extracts several features of voiceprint data, and effectively fuses them according to the relationship between the features to construct several branch data. It considers multivariate features as key parameters for wind turbine blade status monitoring, thereby improving the accuracy of the status monitoring method.

[0081] The noise reduction preprocessing operation in this embodiment is implemented by an improved VMD method, including the following steps:

[0082] Get voiceprint data.

[0083] Calculate the fuzzy entropy MHS corresponding to the voiceprint data.

[0084] Define the decomposition mode number K and penalty factor α.

[0085] Set the minimum number of decomposition modes and the maximum number of decomposition modes; the minimum number of decomposition modes and the maximum number of decomposition modes are set based on experience. In this embodiment, the minimum number of decomposition modes and the maximum number of decomposition modes are set to 2 and 10 respectively, that is, when performing noise reduction processing, the voiceprint data is divided into 2 to 10 modes.

[0086] Decompose the voiceprint data into K modal MTs in descending order according to the number of decomposition modes. k .

[0087] Calculate the average fuzzy entropy MHS corresponding to all modes under the decomposition mode number avg .

[0088] In some embodiments, the current number of decomposition modes is 5, that is, the voiceprint data is decomposed into 5 modes, the fuzzy entropy corresponding to each mode is calculated, and then the average value operation is performed to obtain the average fuzzy entropy corresponding to the decomposition mode number of 5.

[0089] Construct the objective function corresponding to the decomposition mode number; the objective function satisfies the following formula:

[0090]

[0091] Among them, k represents the number corresponding to the mode under the decomposition mode number; the purpose of the objective function is to find the optimal decomposition mode number.

[0092] Construct a penalty factor optimization function that satisfies the following formula:

[0093]

[0094] Where j represents the decomposition mode number K opt The number corresponding to the lower mode; the purpose of the penalty factor optimization function is to find the optimal penalty factor, and the objective function is expressed as maximizing the orthogonality between modes; <, > represent the inner product operator symbols, which are used to measure the correlation between modes.

[0095] The optimal decomposition mode number K is obtained by solving the objective function opt ; Solve the optimization function through the particle swarm algorithm to obtain the optimal penalty factor α opt ;

[0096] In some embodiments, it is determined whether the fuzzy entropy and energy proportion corresponding to several modes are all within the noise range; the energy proportion represents the ratio of the energy of each modal component to the total energy of the original signal; and the noise range is set based on experience.

[0097] In this embodiment, the noise range corresponding to the fuzzy entropy is set to a range greater than 1.2, and the energy proportion is set to a range less than 5%; that is, when the fuzzy entropy is greater than 1.2 and the energy proportion is less than 5%, the corresponding mode is set to the noise-dominated mode.

[0098] If yes, the mode is considered noise-dominated; if no, the mode is considered signal-dominated.

[0099] The dominant mode of the signal is reconstructed to obtain the noise-reduced voiceprint data.

[0100] This embodiment achieves effective noise reduction of voiceprint data by improving the VMD algorithm. While maximizing the noise reduction effect, it has the ability to automatically select the optimal number of decomposition modes and penalty factors. It effectively avoids the over-decomposition or under-decomposition problems that may be caused by manually setting parameters in traditional methods, and significantly improves the separation accuracy between noise and useful signals, thereby providing a more accurate and reliable data foundation for the subsequent multi-feature extraction process.

[0101] In this embodiment, the feature weight is calculated based on the importance of the defect category, including:

[0102] Obtain the defect category importance QZX, defect feature sensitivity QTM, and feature stability TW; the defect category importance QZX refers to the degree of influence of the defect on the wind turbine equipment. The defect category importance QZX is set based on experience. In this embodiment, crack defects, corrosion defects, and deformation defects are considered. Considering that crack defects have the greatest influence on the wind turbine and deformation defects have the least influence on the wind turbine, the defect category importances corresponding to crack defects, corrosion defects, and deformation defects are set to 0.5, 0.3, and 0.2, respectively, in this embodiment.

[0103] The defect feature sensitivity QTM refers to the sensitivity of the defect category to the MFCC feature and the GFCC feature. The specific value is set based on experience. In this embodiment, the defect feature sensitivity of the crack defect on the MFCC feature is set to 0.9.

[0104] Feature stability TW refers to the stability of MFCC features and GFCC features. It is calculated in real time by the inverse of the variance of MFCC / GFCC features within the time window, reflecting the volatility of MFCC / GFCC features in the time dimension.

[0105] The feature weight calculation function MTQF(QZX, QTM, TW) is constructed through the nonlinear relationship between the defect category importance QZX, defect feature sensitivity QTM and feature stability TW and the feature weights corresponding to the MFCC features;

[0106] The frequency band sensitivity function satisfies the following formula:

[0107]

[0108] Among them, c represents the number corresponding to the defect category; the difference between MFCC features and GFCC features is amplified by the exponential function, and the corresponding weight of the dominant feature is strengthened.

[0109] Substitute the defect category importance QZX, defect feature sensitivity QTM and feature stability TW into the feature weight calculation function to obtain the feature weight corresponding to the MFCC feature.

[0110] The feature weight corresponding to the GFCC feature is obtained by performing a difference calculation between the sum of the feature weights and the feature weight corresponding to the MFCC feature; the sum of the feature weights is set according to experience. In this embodiment, the sum of the feature weights is set to 1, that is, the sum of the weight coefficients is considered to be 1.

[0111] This embodiment comprehensively considers various factors that affect the extraction of MFCC and GFCC features and constructs a weight calculation function for MFCC features. During the feature fusion process, the weight distribution of MFCC and GFCC features can be dynamically adjusted so that the fused features can more effectively reflect the changing characteristics of various defect types, thereby improving the accuracy and adaptability of defect recognition.

[0112] In this embodiment, generating a status result according to branch input data includes:

[0113] Obtain branch input data, branch labels and voiceprint data; the branch input data includes branch input data one and branch input data two; the branch labels include branch label one and branch label two; branch input data one is used to input the input data in the model corresponding to branch label one; branch input data two is used to input the input data in the model corresponding to branch label two.

[0114] Input the branch input data (branch input data one or branch input data two) into its corresponding branch model (branch label one or branch label two) to obtain the output result of the branch; the multi-branch result generation model is constructed through a machine learning model.

[0115] Extract the defect category and defect confidence corresponding to the output result.

[0116] Extract the voiceprint signal-to-noise ratio (SNR) corresponding to the voiceprint data.

[0117] The confidence levels corresponding to the same defect category are weighted and fused according to the confidence weights to obtain the status result; the confidence weights are generated according to the voiceprint signal-to-noise ratio.

[0118] In some embodiments, a multi-branch outcome generation model is constructed using a machine learning model, including:

[0119] Several historical branch input data corresponding to branch labels and their corresponding historical output results are divided into training data, verification data and test data corresponding to the branch labels; the training data, verification data and test data corresponding to the branch labels are preprocessed accordingly to obtain training set, verification set and test set; the ratio between the training set, test set and verification set is 7:2:1.

[0120] Select two machine learning models as base models; select two machine learning models as basic sub-models; in this embodiment, use the CNN-LSTM model as the sub-model of branch input data one; use the ResNet-18 model as the sub-model of branch input data two.

[0121] The corresponding basic models are trained through their respective training sets, and the learning rates and other hyperparameters are adjusted on their respective validation sets to obtain their respective pre-trained models.

[0122] By verifying the respective pre-trained models on their respective test sets, we finally obtain the input branch labels and their corresponding branch input data, and output a multi-branch result generation model that is the output result corresponding to the branch label.

[0123] The confidence weight in this embodiment is generated based on the voiceprint signal-to-noise ratio, including:

[0124] Obtain the voiceprint signal-to-noise ratio (SNR) and defect category confidence difference (QLZC). The defect category confidence difference refers to the degree of difference between the confidence levels corresponding to the same defect type under the branch label.

[0125] The confidence weight function ZQF(SNR, QLZC) is constructed through the nonlinear relationship between the voiceprint signal-to-noise ratio and the defect category confidence difference and the confidence weight corresponding to the defect category in branch label one; among them, the confidence weight function is a Sigmoid function.

[0126] The confidence weight function satisfies the following formula:

[0127]

[0128] In some embodiments, γ is represented by an adjustment factor, and γ is set to control the difference sensitivity, γ>0, and the specific value is set according to experience. In this embodiment, γ is set to 2.0; BSNR is represented by a standard signal-to-noise ratio, and the specific value is set according to experience. In this embodiment, BSNR is set to 20dB. It is believed that when SNR exceeds 20dB, the time domain features are more reliable under high signal-to-noise ratio, and therefore tend to favor the state result corresponding to branch label one; in other embodiments, BSNR can also be set to 15dB.

[0129] Substitute the voiceprint signal-to-noise ratio and the defect category confidence difference into the confidence weight function to calculate the confidence weight corresponding to the defect category in branch label one.

[0130] The confidence weight corresponding to the defect category in branch label 2 is obtained by performing a difference calculation between the sum of the confidence weights and the confidence weight corresponding to the defect category in branch label 1; the sum of the confidence weights is set based on experience, and in this embodiment, the sum of the confidence weights is set to 1.

[0131] In some embodiments, a crack defect is detected, wherein the confidence level corresponding to the crack defect in branch label 1 is 0.92, and the confidence level corresponding to the crack defect in branch label 2 is 0.62; the voiceprint signal-to-noise ratio (SNR) is 18 dB; at this time, the confidence weight of the crack label corresponding to branch label 1 is ZQF C (0.9, 0.3) = 0.574; then the confidence weight of the crack label corresponding to branch label 2 is 1-0.574 = 0.426.

[0132] like Figure 2 As shown, in this embodiment, the monitoring results are generated based on the status results and environmental data, including:

[0133] Obtain status results, environmental data, fan blade parameters and branch input data within the time window; the time window refers to the time range for collecting voiceprint data; fan blade parameters refer to the parameters of the fan blades during operation, such as fan speed; environmental data includes temperature and humidity, etc.

[0134] Extract the GAP feature of branch input data in the multi-branch result generation model; the GAP feature refers to the feature map obtained after passing through the global average pooling layer in the ResNet-18 network model.

[0135] Extract the defect categories and their corresponding confidence levels from the status results.

[0136] Determine whether the confidence level is greater than the corresponding defect confidence threshold. The confidence threshold is set based on experience. In this embodiment, the defect confidence threshold corresponding to the crack defect is set to 0.65, and the defect confidence threshold corresponding to the corrosion defect is set to 0.7. This is because the severity of the crack defect is higher than that of the corrosion defect. Therefore, the defect confidence threshold corresponding to the crack defect is set to be more sensitive.

[0137] Yes, a fan blade defect alarm signal is generated.

[0138] The environmental data, blade parameters, branch input data, and GAP features are integrated into positioning analysis data; the positioning analysis data is the data required for analyzing the location of defects when defects exist.

[0139] The positioning analysis data is input into the defect positioning model to obtain the monitoring results; the defect positioning model is constructed through an artificial intelligence model; the monitoring results include the coordinates of the defect location when the defect occurs.

[0140] No, do nothing.

[0141] This embodiment determines whether a defect exists by comparing the final confidence levels obtained for different defect categories with their corresponding defect confidence thresholds. When a defect is detected, a pre-trained defect localization model is further used to accurately locate the defect position. If no defect is detected, the localization step is skipped, thereby realizing an asynchronous processing mechanism for condition monitoring and defect localization. This method effectively reduces unnecessary computational overhead and improves the overall efficiency and recognition accuracy of the wind turbine blade condition monitoring system.

[0142] The defect location model in this embodiment is constructed using an artificial intelligence model, including:

[0143] Obtain some historical positioning analysis data and its corresponding historical monitoring results;

[0144] Several historical positioning analysis data and their corresponding historical monitoring results are divided into training data, verification data and test data; data preprocessing is performed on the training data, verification data and test data to obtain training set, verification set and test set; the ratio between the training set, test set and verification set is 7:2:1.

[0145] Select an artificial intelligence model as the basic model; preferably, the artificial intelligence model includes an Encoder-Decoder model, etc.;

[0146] Train the basic model using the training set, and adjust the learning rate and other hyperparameters on the validation set to obtain the pre-trained model;

[0147] By verifying the pre-trained model on the test set, we finally obtain the input positioning analysis data and output the defect positioning model as the monitoring result.

[0148] like Figure 3 As shown, another embodiment of the present application provides a wind turbine blade operating status monitoring system based on voiceprint recognition, including: a data acquisition module, a data analysis module and an early warning module; the data acquisition module is electrically connected and / or communicatively connected to the data analysis module; the data analysis module is electrically connected and / or communicatively connected to the early warning module;

[0149] The data acquisition module obtains the soundprint data generated by the wind turbine blades during operation through the data acquisition equipment, and extracts MFCC features, GFCC features and Mel spectrogram time-frequency images from it, and also obtains environmental data including temperature and humidity.

[0150] In some embodiments, the data acquisition device includes several sensors, such as a temperature sensor, a humidity sensor, and a microphone array.

[0151] The data analysis module includes a data integration unit and a result generation unit.

[0152] The data integration unit calculates feature weights based on the importance of defect categories, defect feature sensitivity, and feature stability, and constructs several branch input data based on the voiceprint data;

[0153] A result generating unit, configured to generate a status result including a defect category and its corresponding confidence level according to the branch input data, and generate a monitoring result including defect location information and an alarm signal according to the status result and environmental data;

[0154] The early warning module is used to issue an alarm based on the alarm signal, which includes an alarm signal of a defect in a wind turbine blade.

[0155] In another embodiment, the present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned method for monitoring the operating status of wind turbine blades based on voiceprint recognition is implemented.

[0156] In another embodiment, the present invention provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the above-mentioned method for monitoring the operating status of wind turbine blades based on voiceprint recognition.

[0157] In the embodiments disclosed herein, computer storage media may be tangible media that may contain or store programs for use by or in conjunction with an instruction execution system, device, or apparatus. Computer storage media may include, but are not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or equipment, or any suitable combination of the foregoing. More specific examples of computer storage media may include electrical connections based on one or more wires, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0158] Some of the data in the above formula are calculated by removing the dimensions and taking their numerical values. The formula is a formula that is closest to the actual situation obtained by software simulation of a large amount of collected data; the preset parameters and preset thresholds in the formula are set by technical personnel in this field according to actual conditions or obtained through simulation of a large amount of data.

[0159] The working principle of this application is as follows: by acquiring voiceprint data and environmental data; constructing several branch input data according to the voiceprint data; generating status results according to the branch input data; generating monitoring results according to the status results and environmental data, constructing a multi-branch result generation model, and dynamically fusing multiple features of different branches to make feature expression more effective. At the same time, the results of different branches are dynamically fused in the decision-making stage to make the final defect identification result more accurate and reasonable; at the same time, when there are defects, the defect position is located in a timely and effective manner, thereby improving the efficiency of the status monitoring method and avoiding the problem that the existing technology often ignores the degree of correlation between multiple defect states and each voiceprint feature, resulting in the failure of effective fusion of multiple voiceprint features, affecting the accuracy of the defect results, and thus leading to the problem of low accuracy and efficiency of the status monitoring method.

[0160] The above description is only a preferred embodiment of the present invention and does not constitute any form of limitation to the present invention. Any simple modification, equivalent replacement and improvement made by any technician familiar with the profession to the above embodiment without departing from the scope of the technical solution of the present invention and based on the technical essence of the present invention shall still fall within the scope of protection of the technical solution of the present invention.

Claims

1. A method for monitoring the operating status of wind turbine blades based on voiceprint recognition, characterized in that: The following steps are involved: S1. Acquire voiceprint data and temperature and humidity environmental data of the wind turbine blades during operation, perform noise reduction preprocessing on the voiceprint data using an improved VMD method to obtain noise-reduced voiceprint data; extract MFCC features, GFCC features, and Mel spectrogram time-frequency images from the noise-reduced voiceprint data; S2. Calculate feature weights based on the defect category importance, defect feature sensitivity, and feature stability of the wind turbine blade, fuse the MFCC features and GFCC features to obtain branch input data 1; and use the Mel spectrogram time-frequency image as branch input data 2; S3. Input the branch input data 1 and the branch input data 2 into the multi-branch result generation model respectively to obtain the output results corresponding to the branch labels, calculate the confidence weights according to the voiceprint signal-to-noise ratio and the confidence difference of the defect category, perform weighted fusion on the confidences of the same defect category, and obtain a status result including the defect category and its corresponding confidence; S4. Obtain monitoring results including defect location information based on the status results and temperature and humidity environmental data. When the confidence level exceeds the defect threshold, trigger an alarm and locate the defect location.

2. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 1, characterized in that: The improved VMD method comprises: Define the decomposition mode number K and penalty factor α, and set the minimum and maximum decomposition mode numbers; Decompose the voiceprint data into K modal MTs in descending order according to the number of decomposition modes. k , and calculate the average fuzzy entropy value MHS of all modes under different decomposition mode numbers avg ; An objective function corresponding to the decomposition mode number is constructed to find the optimal decomposition mode number. The objective function satisfies the following formula: Among them, k represents the number corresponding to the mode under the decomposition mode number; Construct a penalty factor optimization function to find the optimal penalty factor. The penalty factor optimization function satisfies the following formula: Where j represents the decomposition mode number K opt The numbers corresponding to the lower mode, <, > represent the inner product operator symbols; Distinguish noise-dominated modes from signal-dominated modes based on the fuzzy entropy and energy proportion of each mode; The dominant mode of the signal is reconstructed to obtain the noise-reduced voiceprint data.

3. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 1, characterized in that: The feature weight calculation include: Construct the feature weight calculation function MTQF (QZX, QTM, TW), where QZX is the importance of the defect category, QTM is the sensitivity of the defect feature, and TW is the feature stability; Substitute the defect category importance QZX, defect feature sensitivity QTM and feature stability TW into the function to calculate the MFCC feature weight; The GFCC feature weight is obtained according to the difference between the sum of the feature weights and the MFCC feature weight.

4. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 1, characterized in that: Generating a status result according to branch input data includes: Get the branch input data and output results corresponding to each branch label; Divide the historical branch input data and its corresponding historical output results into training set, validation set, and test set, and perform data preprocessing; Select two machine learning models as the base models of the multi-branch; The basic model is trained using the training set corresponding to each branch, and the model hyperparameters are adjusted on the validation set corresponding to each branch to obtain the pre-trained model of each branch. The performance of each pre-trained model is verified on the test set, and finally integrated to obtain a multi-branch result generation model.

5. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 4 is characterized in that: The construction of the multi-branch result generation model includes: Get the historical branch input data and its corresponding historical output results corresponding to the branch label; Input each branch label and the corresponding branch input data into the multi-branch result generation model, and output the corresponding output result including the defect category prediction and its confidence; Extracting the signal-to-noise ratio (SNR) of the voiceprint data and the confidence difference (QLZC) of different branches for the same defect category; Construct the confidence weight function ZQF(SNR, QLZC); Calculate the confidence weight of branch label one and the confidence weight of branch label two respectively; The confidence levels of different branches of the same defect category are weighted and fused according to the calculated weights to generate the final status result including the defect category and its corresponding confidence level.

6. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 1, characterized in that: Generating monitoring results according to status results and environmental data includes: Obtaining status results, environmental data, fan blade parameters, and branch input data within a time window; Extracting a GAP feature of branch input data one from the multi-branch model; If the defect category confidence in the status result is greater than the corresponding defect confidence threshold, a fan blade defect alarm signal is generated, and the environmental data, fan blade parameters, branch input data 1 and GAP features are integrated to generate positioning analysis data, which is input into the pre-trained defect localization model to obtain a monitoring result including defect location information; Otherwise, no action is taken.

7. The method for monitoring the operating status of wind turbine blades based on voiceprint recognition according to claim 6, characterized in that: The defect location model obtains historical location analysis data and corresponding monitoring results, and divides the historical location analysis data and its corresponding historical monitoring results into a training set, a validation set and a test set for preprocessing; Select an AI model as the base model, train it using the training set, and adjust the hyperparameters on the validation set to obtain a pre-trained model; By verifying the performance of the pre-trained model on the test set, a trained defect localization model is finally obtained; the defect localization model is used to input localization analysis data and output defect location information.

8. A wind turbine blade operating status monitoring system based on voiceprint recognition, characterized in that: include: The data acquisition module is used to obtain the soundprint data generated by the wind turbine blades during operation, and extract the MFCC features, GFCC features and Mel spectrogram time-frequency images from it, while also obtaining temperature and humidity environmental data; Data analysis module, including data integration unit and result generation unit; The data integration unit is used to calculate feature weights based on the importance of defect categories, defect feature sensitivity, and feature stability of the wind turbine blades, and to construct a plurality of branch input data based on the voiceprint data; The result generating unit is used to generate a status result including a defect category and its corresponding confidence level according to the branch input data, and generate a monitoring result including defect location information and an alarm signal according to the status result and environmental data; The early warning module is used to issue an alarm according to the alarm signal generated by the result generation unit.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for monitoring the operating status of wind turbine blades based on voiceprint recognition as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the method for monitoring the operating status of a wind turbine blade based on voiceprint recognition according to any one of claims 1 to 7.

Images