A method and system for acoustic measurement of bolt axial force in wind turbine generators

Abstract

This invention provides an acoustic measurement method and system for the axial force of wind turbine bolts, relating to the field of electronic digital data processing technology. The method includes: acquiring the original acoustic signature of the bolt; performing preprocessing on the original acoustic signature to obtain a denoised acoustic signature signal set; performing feature extraction on the denoised acoustic signature signal set to obtain an acoustic signature feature dataset; performing standardization and principal component analysis dimensionality reduction on the acoustic signature feature dataset to obtain a standardized feature vector set; constructing a CNN-LSTM fusion prediction model and performing training; inputting the standardized feature vector set into the trained CNN-LSTM fusion prediction model to perform training and obtain the predicted axial force value; performing error verification and state determination; and outputting the error verification result and state determination result. This invention solves the problems of existing technologies, such as the need to disassemble bolts, low detection accuracy, and inability to adapt to complex operating conditions during wind turbine bolt axial force detection.

Description

Technical Field

[0001] This invention relates to the field of electronic digital data processing technology, and in particular to an acoustic measurement method and system for the axial force of bolts in a wind turbine generator. Background Technology

[0002] Wind turbines operate under harsh conditions such as strong winds, vibrations, and alternating loads. Key components like blades, nacelles, and towers are mostly connected by bolts. The axial force of these bolts directly provides preload, suppressing slippage, loosening, and fatigue damage, thus preventing connection failures. However, if the bolt axial force weakens or becomes overloaded, it can lead to a collapse in the reliability of the connections, causing serious malfunctions such as component detachment, equipment shutdown, or even complete turbine overturning. Regarding methods for detecting bolt axial force: traditional contact methods, such as torque wrenches and strain gauges, require bolt disassembly, which is inefficient and prone to damaging components, failing to meet the needs of large-scale inspections. Traditional non-contact methods mostly focus on defect detection; however, because they do not establish a quantitative correlation between acoustic signatures and axial force, detection errors generally exceed ±8%, resulting in low measurement accuracy. Existing acoustic measurement methods have not optimized feature extraction and modeling logic for the complex operating conditions of wind turbines, such as strong winds and vibrations, making it difficult to effectively filter environmental interference and extract axial force correlation information from acoustic signatures, thus limiting their practicality.

[0003] Therefore, there is an urgent need to provide a sound measurement method and system that has high quality of bolt acoustic signal acquisition, can resist interference signal processing, and integrates multiple data feature extraction and intelligent prediction. Summary of the Invention

[0004] The purpose of this invention is to provide an acoustic measurement method for the axial force of wind turbine bolts, specifically an acoustic measurement method for the axial force of wind turbine bolts based on an axial force prediction model constructed from acoustic fingerprint signals. This method is applicable to the non-destructive testing of the axial force of high-strength bolts in key parts of wind turbine towers, nacelles, hubs, etc. It solves the problems in the prior art that require bolt disassembly, have low testing accuracy, and cannot adapt to complex operating conditions during wind turbine bolt axial force testing.

[0005] The technical solution of the present invention:

[0006] According to one aspect of the present invention, an acoustic measurement method for the axial force of bolts in a wind turbine generator is provided, comprising:

[0007] Based on an AI-powered magnetic wall-climbing robot, the original sound signature of the bolt was obtained;

[0008] Preprocessing is performed on the original voiceprint to obtain a denoised voiceprint signal set;

[0009] Perform feature extraction on the denoised voiceprint signal set to obtain the voiceprint feature dataset;

[0010] Standardize and perform principal component analysis to reduce the dimensionality of the voiceprint feature dataset to obtain a standardized feature vector set.

[0011] Build a CNN-LSTM fusion prediction model and perform training;

[0012] The standardized feature vector set is input into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value;

[0013] Obtain historical detection values, and perform error verification and status determination based on historical detection values ​​and axial force prediction values;

[0014] Output error verification results and status determination results.

[0015] Furthermore, the acquisition of the original soundprint of the bolt based on the AI-powered magnetic wall-climbing robot specifically includes:

[0016] AI magnetic climbing robot walks on the surface of a wind turbine equipped with an electromagnetic adsorption mechanism;

[0017] The AI ​​magnetic climbing robot attaches a high-frequency piezoelectric acoustic sensor to the head of a bolt using a retractable robotic arm at its end, while controlling the contact pressure to be 0.2–0.3 MPa.

[0018] Under the condition that the ambient wind speed does not exceed 8m / s, the original acoustic text of the bolt is obtained by sampling at a frequency of 400kHz, with a single acquisition time of 1–3 seconds, and continuous acquisition for 3 times.

[0019] Furthermore, the frequency response range of the high-frequency piezoelectric acoustic sensor is 20kHz–200kHz, the sensitivity is not less than 80mV / Pa, and the protection level is IP65; the positioning accuracy of the telescopic robotic arm is ±0.5mm.

[0020] Furthermore, the preprocessing of the original voiceprint to obtain a denoised voiceprint signal set specifically includes:

[0021] Four-level wavelet decomposition is performed using the db6 wavelet basis, combined with a soft thresholding denoising algorithm. ,in, The standard deviation of noise. The signal length is used to filter out the 50Hz and 100Hz components using an adaptive notch filter, and the signal segment with an energy ratio of not less than 90% is selected to obtain the denoised voiceprint signal set.

[0022] Furthermore, the aforementioned voiceprint feature dataset includes at least time-domain features, frequency-domain features, and time-frequency-domain features. The time-domain features include at least peak value, effective value, kurtosis, skewness, and impulse factor. The frequency-domain features include at least spectral peak value, center frequency, 50–150 kHz frequency band energy, and frequency standard deviation. The time-frequency-domain features include at least wavelet entropy, marginal spectral peak value, and Hilbert-Huang transform energy.

[0023] Furthermore, the principal component analysis dimensionality reduction specifically retains the top 5 principal components with a cumulative contribution rate of ≥95%.

[0024] Furthermore, the CNN-LSTM fusion prediction model is configured with a CNN sub-network and an LSTM sub-network. The LSTM sub-network contains two LSTM units, and the CNN sub-network contains three convolutional layers and two max-pooling layers. In the CNN sub-network, the kernel size of each convolutional layer is 3×1, the number of filters is 32, 64, and 128 respectively, and the activation function is ReLU. The kernel size of the max-pooling layer is 2×1. In the LSTM sub-network, each LSTM unit has 64 hidden units and is configured with a gating mechanism.

[0025] Furthermore, the training dataset input to the CNN-LSTM fusion prediction model contains 1800 labeled samples, including at least M24, M30, and M36 bolts with an axial force range of 50–350 kN. The labeled values ​​of the labeled samples are obtained by a standard axial force tester with an accuracy of ±1%.

[0026] Furthermore, the CNN-LSTM fusion prediction model, during training, uses the root mean square error as the loss function, Adam as the optimizer, and an initial learning rate of 0.001, which decays to 0.0001 after 500 training iterations. It also employs an early stopping strategy, terminating training when the root mean square error of the validation set fails to decrease for 20 consecutive iterations.

[0027] According to another aspect of the present invention, an acoustic measurement system for the axial force of a wind turbine bolt is provided, comprising:

[0028] The AI ​​magnetic wall-climbing robot integrates a visual recognition module, an electromagnetic adsorption mechanism, a retractable robotic arm, and a high-frequency piezoelectric acoustic sensor.

[0029] The signal preprocessing module is used to perform preprocessing on the original voiceprint to obtain a denoised voiceprint signal set;

[0030] The feature extraction module is used to perform feature extraction on the denoised voiceprint signal set to obtain the voiceprint feature dataset;

[0031] The feature dimensionality reduction module is used to perform standardization and principal component analysis on the voiceprint feature dataset to obtain a standardized feature vector set.

[0032] The model building module is used to build a CNN-LSTM fusion prediction model and perform training.

[0033] The model prediction module is used to input the standardized feature vector set into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value;

[0034] The error verification and status determination module is used to obtain historical detection values ​​and perform error verification and status determination based on the historical detection values ​​and the axial force prediction values.

[0035] The results output module is used to output error verification results and status determination results.

[0036] Based on the above technical features, the beneficial effects of the present invention are as follows:

[0037] (1) The method provided by the present invention can achieve non-destructive, high-precision and automated detection of the axial force of high-strength bolts in key parts such as towers, nacelles and hubs without disassembly or shutdown by constructing an acoustic pattern-axial force prediction model.

[0038] (2) The method provided by the present invention can collect high-quality bolt acoustic signals through AI magnetic wall-climbing robot. By fusing time domain, frequency domain and time-frequency domain multi-dimensional features, and combining principal component analysis dimensionality reduction and CNN-LSTM fusion prediction model, a precise mapping relationship between acoustic features and bolt axial force can be established. Attached Figure Description

[0039] Figure 1 A flowchart illustrating the method of this invention;

[0040] Figure 2 A schematic diagram illustrating the principle of the CNN-LSTM fusion prediction model in the method of this invention is provided.

[0041] Figure 3 The time-domain waveform of the original voiceprint signal used in the method of this invention;

[0042] Figure 4 The time-domain waveform of the speaker signal after wavelet denoising provided by the method of the present invention;

[0043] Figure 5 The frequency domain spectrum of the original voiceprint signal used in the method of this invention;

[0044] Figure 6 Frequency domain spectrum of the speaker signal after wavelet denoising provided by the method of the present invention;

[0045] Figure 7 A bar chart showing the results of voiceprint feature parameter extraction in the method of this invention;

[0046] Figure 8 This invention provides a line graph of the contribution rate of principal components in the dimensionality reduction of principal component analysis in the method.

[0047] Figure 9 This invention provides a convergence curve diagram for the training of the CNN-LSTM fusion prediction model in the method.

[0048] Figure 10 This invention provides a scatter plot comparing the predicted and measured values ​​of bolt axial force in the method.

[0049] Figure 11 A schematic diagram of the system structure is provided for this invention. Detailed Implementation

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0051] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0052] Example 1

[0053] Please refer to the above as well. Figures 1-11 An embodiment of the present invention provides an acoustic measurement method for the axial force of bolts in a wind turbine generator, the method comprising:

[0054] Based on an AI-powered magnetic wall-climbing robot, the original sound signature of the bolt was obtained;

[0055] Preprocessing is performed on the original voiceprint to obtain a denoised voiceprint signal set;

[0056] Perform feature extraction on the denoised voiceprint signal set to obtain the voiceprint feature dataset;

[0057] Standardize and perform principal component analysis to reduce the dimensionality of the voiceprint feature dataset to obtain a standardized feature vector set.

[0058] Build a CNN-LSTM fusion prediction model and perform training;

[0059] The standardized feature vector set is input into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value;

[0060] Obtain historical detection values, and perform error verification and status determination based on historical detection values ​​and axial force prediction values;

[0061] Output error verification results and status determination results.

[0062] To further clarify the present invention, the following description will incorporate actual testing of the axial force of the nacelle bolts of a 2.0MW wind turbine at a certain mountain wind farm:

[0063] First, based on the AI-powered magnetic wall-climbing robot, the original sound signature of the bolt was obtained, including:

[0064] The AI ​​magnetic climbing robot is equipped with an electromagnetic adsorption mechanism to walk on the surface of a wind turbine. The electromagnetic adsorption mechanism can ensure stable walking on the surface of the wind turbine.

[0065] The AI ​​magnetic wall-climbing robot uses a retractable robotic arm at its end to attach a high-frequency piezoelectric acoustic sensor to the head of a bolt, and controls the contact pressure to be 0.2–0.3 MPa. The positioning accuracy of the retractable robotic arm is ±0.5 mm.

[0066] Under the condition that the ambient wind speed does not exceed 8m / s, the original acoustic text of the bolt is obtained by sampling at a frequency of 400kHz, with a single acquisition time of 1–3 seconds, and continuous acquisition for 3 times.

[0067] It should be noted that for every 10kN change in axial force, the acoustic frequency shifts by about 5kHz. In this embodiment, the acoustic sensor adopts a high-frequency piezoelectric design, that is, a high-frequency piezoelectric acoustic sensor with a frequency response range of 20kHz-200kHz. It can cover the acoustic frequency fluctuation range caused by the change in bolt axial force, and its sensitivity is not less than 80mV / Pa, and its protection level is IP65.

[0068] It should be noted that the sampling rate is set to 400kHz, which is twice the highest response frequency of the high-frequency piezoelectric acoustic sensor, thus satisfying the Nyquist sampling theorem and avoiding signal aliasing. The acquisition time is set to 2s / acquisition, 400kHz×2s, meaning that 800,000 sampling points are generated in a single acquisition. The acquisition process is executed continuously in 3 sessions, with an interval of 1s between each session, which can reduce random noise interference. Specifically, the first acquisition eliminates the impact signal at the moment of contact with the robotic arm, and the second and third acquisitions are used for subsequent data processing to ensure data stability.

[0069] Next, preprocessing is performed on the original voiceprint to obtain a denoised voiceprint signal set, specifically including:

[0070] Firstly, such as Figures 3-6 As shown, the original acoustic signature signal is decomposed into four levels using the db6 wavelet basis. Each level yields approximation coefficients (i.e., low-frequency signals) and detail coefficients (i.e., high-frequency noise). Considering the interference characteristics in the wind turbine environment, soft thresholding is applied to the detail coefficients. The threshold calculation formula is as follows: In the formula, σ is the noise standard deviation, which is estimated by the detail coefficient of the 4th layer. N is the number of sampling points, preferably 800000, which can improve the signal-to-noise ratio to more than 40dB.

[0071] Secondly, the 50Hz and 100Hz components are filtered out by an adaptive notch filter, thereby achieving vibration filtering in the original acoustic signature.

[0072] Third, the signal segment with an energy ratio of no less than 90% is intercepted, and invalid signals generated by the contact of the high-frequency piezoelectric acoustic sensor and the extension and retraction of the telescopic robotic arm are removed, thereby realizing the signal interception in the original acoustic pattern.

[0073] Fourth, obtain the denoised voiceprint signal set.

[0074] In some embodiments, after processing the original voiceprint, taking a 50Hz / 100Hz robot as an example, the measured threshold is 0.023V. The amplitude of motor vibration noise is reduced from 0.18V to below 0.02V, and the amplitude of wind noise in the 1-5kHz range is reduced from 0.09V to below 0.01V. Vibration noise and wind noise can be effectively suppressed. Furthermore, according to the MATLAB signal analysis tool, the signal-to-noise ratio of the denoised voiceprint signal is increased from the original 22dB to 42dB, the signal waveform smoothness is improved by 65%, and the 50-150kHz frequency band signal associated with bolt axial force is undistorted.

[0075] Next, feature extraction is performed on the denoised voiceprint signal set to obtain the voiceprint feature dataset, specifically, such as... Figure 7 As shown, the voiceprint feature dataset includes at least time-domain features, frequency-domain features, and time-frequency-domain features. These features encompass at least 12 feature parameters: the time-domain features include at least peak value, effective value, kurtosis, skewness, and impulse factor. As the axial force increases, the bolt micro-deformation becomes more significant, and the time-domain feature value increases linearly with the increase of axial force. The frequency-domain features include at least spectral peak value, center frequency, 50–150kHz frequency band energy, and frequency standard deviation. When the axial force changes, the bolt's natural frequency shifts, and the spectral peak position moves towards higher frequencies as the axial force increases, increasing the proportion of frequency band energy. The time-frequency-domain features include at least wavelet entropy, marginal spectral peak value, and Hilbert-Huang transform energy (HHT). When the axial force is overloaded, the bolt structure stability decreases, the wavelet entropy decreases (<0.8), and the HHT energy distribution is concentrated in the high-frequency band.

[0076] In some embodiments, time-domain features are obtained by calculating the time-domain signal in the center of the denoised speaker signal using the MATLAB signal processing toolbox.

[0077] Firstly, the maximum signal value is calculated based on the peak value, using the following formula: In some embodiments, the measured value is 1.27V, which indicates that the greater the axial force, the more significant the micro-deformation of the bolt, and the higher the peak value of the acoustic signal.

[0078] Secondly, the root mean square value is calculated based on the effective value, using the following formula: In the formula For the first The voltage value at each sampling point, measured in some embodiments as 0.39V, reflects the average energy of the signal and is linearly positively correlated with the axial force, with a correlation coefficient of 0.92.

[0079] Third, the fourth-order central moment is calculated based on kurtosis, using the following formula: In the formula, σ is the mean and σ is the standard deviation. In some embodiments, the measured value is 3.1, which indicates that the kurtosis is close to the normal distribution value of 3 under normal axial force, and will be greater than 4.0 when the axial force is overloaded.

[0080] Fourth, the third-order central moment is calculated based on skewness, using the following formula: In some embodiments, the measured value is 0.11, reflecting the symmetry of the signal waveform. When the axial force is abnormal, the absolute value of the skewness will be greater than 0.5.

[0081] Fifth, based on the pulse factor, the calculation formula is as follows: The measured value is 3.26, reflecting the signal pulse characteristics. When the axial force decays, the pulse factor will drop below 2.5.

[0082] In some embodiments, based on the time-domain features of the antecedent of the denoised speaker signal, a Fast Fourier Transform (FFT, 8192 points) is performed to convert it into a frequency-domain signal, and then frequency-domain features are extracted. Specifically:

[0083] First, obtain the peak frequency of the spectrum, which is the frequency corresponding to the maximum value of the frequency domain signal. In some embodiments, it is measured to be 126kHz, which indicates that for every 20kN increase in axial force, the peak frequency shifts to the higher frequency direction by about 10kHz.

[0084] Secondly, the frequency corresponding to the centroid of the spectral energy is calculated based on the center frequency, using the following formula: In the formula, For frequency The corresponding power, measured at 99kHz in some embodiments, characterizes a linear shift in the center frequency as the axial force changes.

[0085] Third, calculate the energy in the 50-150kHz frequency band, that is, the proportion of energy in this frequency band to the total energy. In some embodiments, the measured value is 83%. This frequency band is the core frequency band related to bolt axial force, and its proportion is ≥80% under normal axial force.

[0086] Fourth, calculate the frequency standard deviation (Freq_Std). The frequency standard deviation reflects the dispersion of the spectral distribution. The formula is: In some embodiments, the measured 24kHz frequency indicates that the standard deviation is greater than 30kHz when characterizing axial force anomalies.

[0087] In some embodiments, time-frequency domain features are extracted based on wavelet transform and Hilbert-Huang transform (HHT), specifically as follows:

[0088] Firstly, regarding wavelet entropy, the signal is decomposed into 5 levels of wavelet and the information entropy of each scale coefficient is calculated. In some embodiments, the measured value is 0.86, which indicates that when the axial force is overloaded, the stability of the bolt structure decreases and the wavelet entropy drops below 0.8.

[0089] Secondly, for the Hilbert-Huang transform, the signal is decomposed into eight intrinsic mode functions (IMFs) through empirical mode decomposition (EMD), and the sum of the energies of each IMF is calculated. In some embodiments, the measured value is 1.22 J, which indicates a positive correlation with the axial force. For every 10 kN increase in axial force, the energy increases by approximately 0.05 J.

[0090] Third, for the marginal spectrum peak value, the maximum value of the HHT marginal spectrum is taken, which is 0.082W / Hz in some embodiments, reflecting the degree of energy concentration in the time and frequency domain. When the axial force is normal, the marginal spectrum peak value is ≥0.07W / Hz.

[0091] Next, standardization and principal component analysis (PCA) dimensionality reduction are performed on the voiceprint feature dataset to obtain a standardized feature vector set. Specifically, the 12 voiceprint feature parameters covered in the voiceprint feature dataset are successively standardized by Z-score and reduced by principal component analysis. The top 5 principal components with a cumulative contribution rate of not less than 95% are retained, and the min-max is normalized to the [0, 1] interval to form a 5-dimensional standardized feature vector, thus obtaining a standardized feature vector set.

[0092] In detail, Z-score standardization is used to make the mean 0 and the standard deviation 1, thereby eliminating dimensional differences, such as the time-domain feature unit being V and the frequency-domain feature unit being kHz. Then, the feature covariance matrix (12×12 order) is calculated, and the eigenvalues ​​and eigenvectors are solved by the Jacobi iteration method. The eigenvalues ​​are sorted from largest to smallest, and the principal components with a cumulative contribution rate ≥95% are selected. Specifically, the components retained by principal component analysis for dimensionality reduction are: PC1, PC2, PC3, PC4, and PC5. Among them, PC1 is a feature that comprehensively represents the time-domain intensity, including peak value and effective value; PC2 is a feature that represents the frequency-domain distribution, including spectral peak value and center frequency; PC3 is a feature that represents the time-frequency domain energy, including Hilbert-Huang transform energy and marginal spectral peak value; PC4 is a feature that represents the signal complexity, including wavelet entropy and kurtosis; and PC5 is a feature that represents the stability, including frequency standard deviation and impulse factor.

[0093] In some embodiments, such as Figure 8 As shown, the formed 5-dimensional standardized feature vector, the experimental results show that the contribution rates of the first 5 principal components are 42.3%, 25.8%, 15.2%, 8.7%, and 4.2%, respectively, and the experimental vector is: [0.63, 0.59, 0.72, 0.48, 0.66]. The experimental cumulative contribution rate reaches 96.2%, which can meet the dimensionality reduction requirements, and the dimension is reduced by 58.3%, which can significantly reduce the computational load of the subsequent model.

[0094] Then, a CNN-LSTM fusion prediction model is constructed. This model is trained using a labeled dataset to establish a quantitative mapping between voiceprint features and axial force, such as... Figure 2 As shown, the specific process includes:

[0095] First, the input layer of the CNN-LSTM fusion prediction model receives 5-dimensional normalized feature vectors from the normalized feature vector set, which are then converted into a three-dimensional tensor of (1, 5, 1) through the Reshape layer.

[0096] Secondly, a CNN sub-network is constructed. The CNN sub-network is used to extract spatial correlation information of voiceprint features, such as the distribution pattern of frequency domain features. Specifically, the CNN sub-network contains 3 convolutional layers and 2 max pooling layers. The kernel size of each convolutional layer is 3×1, and the number of filters is 32, 64, and 128 respectively. The activation function is ReLU. The kernel size of the max pooling layer is 2×1. The first convolutional layer extracts the local correlation of features, such as the co-transformation of PC1 and PC2. The second and third convolutional layers deepen the feature abstraction. The pooling layer reduces the number of parameters through downsampling (from 128 feature maps to 32), thereby avoiding overfitting.

[0097] Third, an LSTM sub-network is built. The LSTM sub-network is used to learn the temporal changes of voiceprint features, such as adapting to voiceprint data collected multiple times and capturing the dynamic relationship between axial force and features. Specifically, the LSTM sub-network contains two layers of LSTM units. Each layer of LSTM units has 64 hidden units and a dropout coefficient of 0.2, which can prevent overfitting during training. It is also equipped with a gating mechanism (input gate, forget gate, output gate) to determine the temporal dependency of the feature vector sequence collected three times in a row. Furthermore, it can improve the model's generalization ability by capturing the trend of feature changes between multiple collections.

[0098] Fourth, the output layer of the CNN-LSTM fusion prediction model protects a fully connected layer. The fully connected layer is activated by the Linear activation function, which can output the predicted value of bolt axial force. The model weights are initialized using the He normal distribution to avoid gradient vanishing.

[0099] Next, the training dataset for the CNN-LSTM fusion prediction model is constructed as follows:

[0100] Firstly, the training dataset was obtained by collecting data from three mainstream bolt specifications: M24, M30, and M36. Each specification had 15 axial force levels, with a specific axial force range of 50-350 kN, covering the commonly used axial force range for wind turbine bolts. Twenty sets of acoustic signature data were collected for each level, resulting in 900 sets of labeled data for each bolt specification, for a total of 1800 sets of labeled data across the three specifications. The labeled values ​​were obtained using a standard axial force tester with an accuracy of ±1%. Furthermore, in some embodiments, 900 sets of original data and 900 sets of data augmentation data were used, with the augmentation method involving the addition of ±3% random noise.

[0101] Secondly, data augmentation was performed on the training dataset. Specifically, the dataset was expanded to 1800 sets by adding ±3% random noise, thereby simulating slight disturbances to the wind turbine. Then, it was divided into training set, validation set, and test set in a 7:2:1 ratio, i.e., 1260 sets for training, 360 sets for validation, and 180 sets for test.

[0102] Next, the training and optimization of the CNN-LSTM fusion prediction model are performed, specifically as follows:

[0103] First, the initial learning rate is set to 0.001, the number of iterations is 800, and the root mean square error (RMSE) is used as the loss function, with the following formula:

[0104]

[0105] Secondly, This is the actual axial force. To predict axial force, n is the number of samples; thus, the parameters of the CNN-LSTM fusion prediction model are trained.

[0106] Third, the Adam optimizer is used, and the optimizer contains... , Furthermore, the learning rate is dynamically adjusted, decaying to 0.0001 after 500-800 training iterations, and an early stopping strategy is employed, terminating training when the root mean square error of the validation set fails to decrease for 20 consecutive iterations. This process completes the algorithm optimization of the CNN-LSTM fusion prediction model, such as... Figure 9 As shown, Figure 9 The horizontal axis represents the number of iterations, and the vertical axis represents the root mean square error.

[0107] Fourth, after the CNN-LSTM fusion prediction model is trained, the RMSE on the test set is ≤1.2kN, corresponding to a detection error of ≤±3%. For example, in some embodiments, after training, the RMSE of the model on the test set is 1.08kN, and the mean absolute error (MAE) is 0.82kN, which meets the accuracy requirements.

[0108] Next, the standardized feature vector set is input into the trained CNN-LSTM fusion prediction model for training to obtain the axial force prediction value. The input model uses the acoustic signature feature vector of the bolt to be detected. Specifically, the preprocessed acoustic signature signal undergoes feature extraction to obtain a standardized vector, which is then input into the trained CNN-LSTM fusion prediction model to output the axial force prediction value. .

[0109] Next, historical detection values ​​are obtained. Based on these historical detection values ​​and the predicted axial force values, error verification and status determination are performed. Specifically, historical detection values ​​refer to the verified predicted axial force values ​​recorded when the same bolt was previously tested using this method. In detail, the error verification is a two-stage verification: in the first stage, if the predicted axial force value... If the deviation from the historical test value of the bolt is ≤±3%, the result will be output directly; if the axial force prediction value is in the secondary verification... If the deviation from the historical test value of the bolt is greater than ±3%, the acoustic fingerprint signal is re-collected twice, and the average of the three predicted values ​​is calculated. If the deviation still exceeds the standard, it is marked as "pending re-inspection". The status judgment is based on the design axial force marking of the bolt status. Specifically, within ±5% of the design value is "normal", within ±5% to ±8% is "warning", and in the "warning" state, a re-inspection is required within one week. Exceeding ±8% is "abnormal", and in the "abnormal" state, the machine is immediately shut down for processing. It should be noted that the design value comes from the specified value in the technical documents during the wind turbine generator manufacturing process, and this specified value is used as the design value.

[0110] Finally, the error verification results and status determination results are output.

[0111] In some embodiments, such as Figure 10 As shown, during inference, the CNN-LSTM fusion prediction model first inputs a 5-dimensional feature vector into the input layer of the model. The CNN layer extracts spatial features (such as the linear correlation between PC1 and PC2), and the output (1, 1, 32) feature tensor is fed into the LSTM layer. After learning the temporal pattern, a (1, 64) hidden state vector is generated. Finally, it is mapped to the axial force prediction value through a fully connected layer. For example, with the measured input feature vector [0.63, 0.59, 0.72, 0.48, 0.66], the model inference time is 0.03s, and the output axial force prediction value is 279.5kN.

[0112] Example 2

[0113] Based on the same inventive concept as the acoustic measurement method for the axial force of a wind turbine bolt in Embodiment 1 above, such as Figure 11 As shown, this invention also provides an acoustic measurement system for the axial force of wind turbine bolts, specifically including: an AI magnetic climbing robot, a signal preprocessing module, a feature extraction module, a feature dimensionality reduction module, a model building module, a model prediction module, an error verification and state determination module, and a result output module. The AI ​​magnetic climbing robot integrates a visual recognition module, an electromagnetic adsorption mechanism, a retractable robotic arm, and a high-frequency piezoelectric acoustic sensor. The signal preprocessing module is used to preprocess the original acoustic signature to obtain a denoised acoustic signature signal set. The feature extraction module is used to extract features from the denoised acoustic signature signal set to obtain an acoustic signature feature dataset. The feature dimensionality reduction module is used to perform standardization and principal component analysis dimensionality reduction on the acoustic signature feature dataset to obtain a standardized feature vector set. The model building module is used to build a CNN-LSTM fusion prediction model and perform training. The model prediction module is used to input the standardized feature vector set into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value. The error verification and state determination module is used to obtain historical detection values ​​and perform error verification and state determination based on the historical detection values ​​and the axial force prediction value. The result output module is used to output the error verification result and the state determination result.

[0114] The specific example of the acoustic measurement method for the axial force of a wind turbine bolt in Embodiment 1 above is also applicable to the acoustic measurement system for the axial force of a wind turbine bolt in this embodiment. Through the detailed description of the acoustic measurement method for the axial force of a wind turbine bolt above, those skilled in the art can clearly understand the acoustic measurement system for the axial force of a wind turbine bolt in this embodiment. Therefore, for the sake of brevity, it will not be described in detail here.

[0115] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered illustrative and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the scope of the invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0116] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An acoustic method for measuring the axial force of bolts in a wind turbine generator, characterized in that, include: Based on an AI-powered magnetic wall-climbing robot, the original sound signature of the bolt was obtained; Preprocessing is performed on the original voiceprint to obtain a denoised voiceprint signal set; Feature extraction is performed on the denoised voiceprint signal set to obtain the voiceprint feature dataset; Standardize and perform principal component analysis to reduce the dimensionality of the voiceprint feature dataset to obtain a standardized feature vector set. Build a CNN-LSTM fusion prediction model and perform training; The standardized feature vector set is input into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value; Obtain historical detection values, and perform error verification and status determination based on historical detection values ​​and axial force prediction values; Output error verification results and status determination results; Specifically, the preprocessing of the original voiceprint to obtain a denoised voiceprint signal set includes: A 4-level wavelet decomposition based on the db6 wavelet basis is performed, combined with a soft thresholding denoising algorithm, with a threshold λ=σ. 2lgN, where σ is the noise standard deviation and N is the signal length. At the same time, the 50Hz and 100Hz components are filtered out by an adaptive notch filter, and the signal segment with an energy ratio of not less than 90% is truncated to obtain the noise-reduced voiceprint signal set. The aforementioned voiceprint feature dataset includes at least time-domain features, frequency-domain features, and time-frequency-domain features. The time-domain features include at least peak value, effective value, kurtosis, skewness, and impulse factor. The frequency-domain features include at least spectral peak value, center frequency, 50–150 kHz frequency band energy, and frequency standard deviation. The time-frequency-domain features include at least wavelet entropy, marginal spectral peak value, and Hilbert-Huang transform energy. The aforementioned principal component analysis dimensionality reduction specifically retains the top 5 principal components with a cumulative contribution rate of ≥95%. The CNN-LSTM fusion prediction model is configured with CNN subnetworks and LSTM subnetworks. The LSTM subnetwork contains 2 LSTM units, and the CNN subnetwork contains 3 convolutional layers and 2 max pooling layers. In the CNN subnetwork, the kernel size of each convolutional layer is 3×1, the number of filters is 32, 64, and 128 respectively, and the activation function is ReLU. The kernel size of the max pooling layer is 2×1. In the LSTM subnetwork, each LSTM unit has 64 hidden units and is configured with a gating mechanism.

2. The method according to claim 1, characterized in that, The aforementioned acquisition of the original sound signature of the bolt based on the AI-powered magnetic wall-climbing robot specifically includes: AI magnetic climbing robot walks on the surface of a wind turbine equipped with an electromagnetic adsorption mechanism; The AI ​​magnetic climbing robot attaches a high-frequency piezoelectric acoustic sensor to the head of a bolt using a retractable robotic arm at its end, while controlling the contact pressure to be 0.2–0.3 MPa. Under the condition that the ambient wind speed does not exceed 8m / s, the original acoustic text of the bolt is obtained by sampling at a frequency of 400kHz, with a single acquisition time of 1–3 seconds, and continuous acquisition for 3 times.

3. The method according to claim 2, characterized in that, The frequency response range of the high-frequency piezoelectric acoustic sensor is 20kHz–200kHz, the sensitivity is not less than 80mV / Pa, and the protection level is IP65; the positioning accuracy of the telescopic robotic arm is ±0.5mm.

4. The method according to claim 1, characterized in that, The training dataset input to the CNN-LSTM fusion prediction model contains 1800 labeled samples, including at least M24, M30, and M36 bolts with an axial force range of 50–350 kN. The labeled values ​​of the samples are obtained by a standard axial force tester with an accuracy of ±1%.

5. The method according to claim 4, characterized in that, The CNN-LSTM fusion prediction model described above uses the root mean square error as the loss function during training, with Adam as the optimizer and an initial learning rate of 0.001, which decays to 0.0001 after 500 training iterations. It also employs an early stopping strategy, terminating training when the root mean square error of the validation set fails to decrease for 20 consecutive iterations.

6. An acoustic measurement system for the axial force of bolts in a wind turbine generator, implementing the method described in any one of claims 1 to 5, characterized in that, include: AI magnetic wall-climbing robot, integrating a visual recognition module, an electromagnetic adsorption mechanism, a retractable robotic arm and a high-frequency piezoelectric acoustic sensor; The signal preprocessing module is used to perform preprocessing on the original voiceprint to obtain a denoised voiceprint signal set; The feature extraction module is used to perform feature extraction on the denoised voiceprint signal set to obtain the voiceprint feature dataset; The feature dimensionality reduction module is used to perform standardization and principal component analysis on the voiceprint feature dataset to obtain a standardized feature vector set. The model building module is used to build a CNN-LSTM fusion prediction model and perform training. The model prediction module is used to input the standardized feature vector set into the trained CNN-LSTM fusion prediction model to perform training and obtain the axial force prediction value; The error verification and status determination module is used to obtain historical detection values ​​and perform error verification and status determination based on the historical detection values ​​and the axial force prediction values. The results output module is used to output error verification results and status determination results.

Images