Wind turbine blade acoustic fault detection method based on transfer learning

Abstract

The present application belongs to the technical field of wind power generation equipment state monitoring and intelligent fault diagnosis, and discloses a wind turbine blade acoustic fault detection method based on transfer learning, which constructs standardized samples through acoustic signal simulation and multi-dimensional data enhancement, and extracts weak fault features by using STFT-Mel spectrum; a double-scale time-frequency attention module is introduced into the light MobileNetV3 to improve the focusing and recognition ability of early crack features; through transfer learning combined with MMD domain difference and entropy minimization regularization, efficient alignment of simulated domain and measured wind field data is realized, and the generalization performance of the model under small sample conditions is enhanced; a double-flow feature fusion mechanism of transverse difference spectrum and longitudinal historical baseline self-evolution is adopted to suppress common mode noise, realize single-blade positioning, progressive degradation early warning and synchronous aging recognition. The method has small parameter quantity and low calculation cost, and is suitable for high-precision and low-cost intelligent detection of early damage of wind turbine blades.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of wind power generation equipment state monitoring and intelligent fault diagnosis, and particularly relates to a wind turbine blade acoustic fault detection method based on transfer learning. BACKGROUND

[0002] To meet the global renewable energy development and wind turbine safe operation requirements, countries have increased research and application of offshore wind power technology in the 21st century. With the increase of wind turbine capacity and the complexity of blade structure, cracks and wear and tear are prone to occur in the blade during long-term operation, which not only reduces power generation efficiency, but also affects the overall structural safety of the wind turbine. Therefore, early fault detection and state monitoring of the wind turbine blade are of great significance.

[0003] Existing blade fault diagnosis methods mainly include vibration analysis, visual detection and acoustic signal analysis. Among them, acoustic signal analysis collects the sound of the wind turbine during operation, extracts and identifies fault features, and has the advantages of non-invasion, low cost and strong real-time performance. However, the wind turbine operating environment is complex, and wind noise and mechanical noise interference is significant, making the high-frequency whistling sound amplitude of blade cracking low and easy to be covered, increasing the difficulty of fault identification.

[0004] The existing acoustic diagnosis method mainly has the following problems: first, the high-frequency whistling sound amplitude is low, and it is easy to be disturbed by environmental noise, making it difficult to identify early faults; second, most methods rely on real-time collection and online calculation, and mobile or offline deployment is difficult to achieve; third, the limited data set results in insufficient accuracy of the model under small sample and high noise conditions.

[0005] For example, the blade crack detection system based on acoustic feature analysis disclosed in the authorized announcement CN 119491796 B realizes accurate positioning of cracks through acoustic signal acquisition, material fatigue monitoring, signal separation and crack positioning modules. However, this system relies on a material fatigue database and a complex feature comparison algorithm, and has limited ability to identify early micro-cracks in a high-noise environment, has high algorithm complexity, and is not conducive to mobile or offline deployment.

[0006] For example, the patent application with the authorized announcement CN 114565006 B combines vibration and audio data to realize real-time monitoring of blade damage through deep learning and machine learning, and can be synchronized to the wind farm control center. Its limitations are that it relies on a large number of samples for model training, and the feature extraction is not sensitive enough to weak acoustic signals such as high-frequency whistling, resulting in a decrease in accuracy in an end-side or data-scarce environment.

[0007] A patent application with the application publication number CN 116631442 A discloses a wind turbine blade fault acoustic detection method and system based on deep learning. The method uses a microphone array to collect acoustic signals, and realizes fault detection by inputting the acoustic signals into a deep neural network through beamforming, high-pass filtering and multiple frequency spectrum features, while providing a client-server interaction architecture. Although the method improves the signal-to-noise ratio and automation level, the system complexity is high, the model calculation amount is large, and it is difficult to deploy in real time on the resource-limited side. At the same time, there are still limitations in identifying early cracks and high-frequency whistling characteristics under small sample conditions.

[0008] The above technical solutions generally rely on high-precision deep learning models or image, vibration and other multi-modal signal analysis, and have problems of large training data size, high computing resource consumption and high model deployment difficulty. The existing acoustic detection method lacks standardized training samples and a transferable end-to-end modeling framework, and has insufficient ability to identify acoustic features of micro cracks and early damage, which is difficult to meet the needs of on-site rapid diagnosis, single-blade fault positioning and long-term online monitoring of wind turbine units in complex operating environments. SUMMARY

[0009] The purpose of the present application is to overcome the shortcomings of the prior art, and to provide a wind turbine blade acoustic fault detection method based on transfer learning. For wind turbine operating audio signals, a multi-source training sample set combining acoustic simulation and multi-dimensional data enhancement is constructed, a Mel frequency spectrum feature mapping and a lightweight MobileNetV3 network structure are used, and a pre-trained acoustic feature model is introduced through a transfer learning mechanism. Combined with dynamic comparison between blades and self-evolution mechanism of historical health baseline of each blade, the method realizes accurate identification of early cracks and high-frequency whistling characteristics, accurate positioning of single-blade faults, three-blade synchronous attenuation identification and long-term early warning of gradual damage under small sample conditions.

[0010] To achieve the above purpose, the present application adopts the following technical solutions:

[0011] A wind turbine blade acoustic fault detection method based on transfer learning, comprising:

[0012] 1) Acoustic data simulation

[0013] The rotpsgen algorithm is used to input noise amplitude, frequency, blade coordinates, radius, speed, and number of blades to generate signals and simulate crack faults. Combined with field recordings, a mixed data set is formed.

[0014] 2) Data enhancement processing, data enhancement processing is performed on the simulated and a small amount of measured audio data to improve the robustness of the model to environmental noise, wind speed changes and other factors, and to improve the small sample and class imbalance problem.

[0015] 3) Mel-spectrogram feature extraction

[0016] Generate a spectrogram using short-time Fourier transform and 128 Mel filters, convert it into a 224x224 pixel RGB image, generate a Mel spectrogram (Mel Spectrogram), preserve acoustic features, and use it as model input;

[0017] Model construction and weight initialization

[0018] Use MobileNetV3-Large as the backbone network and load ImageNet pre-trained weights. Insert the DSTFA (Dual-Scale Time-Frequency Attention Module) module after the last reverse residual module output and replace the classification head with a fully connected layer. Freeze all feature layers except the DSTFA module, the last 1x1 convolution layer, and the classifier;

[0019] 5) Transfer learning and domain adaptation fine-tuning

[0020] Use augmented simulated samples and real samples for joint training. Fine-tune only the unfrozen module parameters based on the above model. Introduce MMD (Maximum Mean Discrepancy) and Entropy regular terms to construct a joint loss function (including cross-entropy), and use the Adam optimizer to iterate 10 times to align the features of the simulated and real domains, adapt to the wind field environment, and improve the model's generalization performance;

[0021] 6) Offline detection and result determination

[0022] Deploy the model on a portable platform, record about 10-15 seconds of three-blade running audio, generate Mel spectrograms, infer normal / abnormal results and probabilities, single-blade fault location results or early warning levels, and automatically save the original audio and inference results to the local. Suitable for offline detection in the field, no network connection is required, and accumulated data supports subsequent model updating and baseline self-evolution;

[0023] 7) Visualization and auxiliary analysis

[0024] Generate 224x224 Mel spectrograms and frequency response curves, three-blade transverse difference spectrograms, and single-blade longitudinal trend curves, support manual review, historical comparison to realize blade state trend analysis, and assist in operation and decision-making;

[0025] 8) End-to-end acoustic diagnosis process

[0026] The integrated acoustic signal simulation, data enhancement, feature extraction, model training, offline detection and visual analysis, and inter-blade dynamic comparison and historical baseline self-evolution collaborative diagnosis form an independently running end-to-end diagnosis system, which supports rapid fault identification and operation and maintenance decision-making under the condition of no real-time monitoring;

[0027] 9) Inter-blade dynamic comparison and historical baseline self-evolution collaborative diagnosis and early warning After generating the three-blade Mel spectrum in step 6), the dynamic health benchmark of the three-blade spectrum is calculated in real time, and the transverse difference spectrum is obtained by subtracting each blade from the benchmark; At the same time, each blade independently maintains its own historical health baseline, and the deviation degree of the current spectrum from the baseline is calculated to obtain the longitudinal evolution feature; The transverse difference spectrum and the longitudinal evolution feature are spliced and input into the model in step 4), and the single-blade fault positioning, early degradation warning or three-blade synchronous attenuation identification result is output; When the diagnosis is healthy for more than 5 times in a row, the baseline of each blade is slowly updated with the current data, and the update is immediately frozen when an exception occurs to prevent baseline pollution.

[0028] The beneficial effects of the present application are:

[0029] The present application first constructs a physically consistent, sufficient and standardized training sample set through blade acoustic simulation algorithm and multi-dimensional data enhancement strategy, which fundamentally solves the problems of scarcity of early crack acoustic samples of wind turbine blades, inconsistent data distribution and easy drowning of weak whistling signals by environmental noise. Combined with Mel spectrum mapping to retain high-frequency whistling characteristics, the model can obtain stable time-frequency feature expression under high noise and small sample conditions.

[0030] The present application introduces a double-scale time-frequency attention module (DSTFA) in the MobileNetV3 lightweight network, which enhances the feature response of the high-frequency whistling sensitive area in the time domain and the frequency domain at the same time, significantly improves the recognition ability of micro cracks and early damage, and realizes the effective alignment of simulation data and real wind field data by combining MMD domain adaptation and entropy minimization regularization term. High generalization performance can be obtained with only a small amount of measured samples. This design has high precision and low computational complexity, and can be directly deployed on mobile terminals or portable devices to realize real-time offline detection.

[0031] The present application further proposes an inter-blade dynamic comparison and historical health baseline self-evolution collaborative diagnosis mechanism, which suppresses common mode interference through transverse difference spectrum to realize precise positioning of single-blade faults; Through longitudinal baseline self-evolution, early warning of gradual degradation and effective identification of three-blade synchronous aging are realized, so that the system has long-term self-adaptation ability. This mechanism can adapt to the natural aging of equipment for a long time and continuously maintain the diagnosis accuracy, thereby constructing an offline, deployable and self-adapting end-to-end wind turbine blade acoustic fault detection system, which greatly reduces the operation and maintenance cost and improves the safety and reliability of long-term operation of the wind farm.

[0032] The application avoids the redundant steps of traditional manual feature extraction by integrating end-to-end acoustic feature extraction and classification, significantly improving the detection efficiency and robustness; at the same time, the proposed model structure has the characteristics of small parameter quantity and low computational overhead, and can run in real time on portable or embedded terminals, suppresses common mode interference through horizontal difference spectrum, and adapts to natural aging of equipment through longitudinal baseline self-evolution, meeting the needs of on-site rapid diagnosis and intelligent operation and maintenance of wind turbine generators. BRIEF DESCRIPTION OF DRAWINGS

[0033] Figure 1 is a flowchart of the wind turbine blade damage detection method of the application;

[0034] Figure 2 is a data processing flowchart of the application;

[0035] Figure 3 is a schematic diagram of the deep learning diagnosis model structure of the application;

[0036] Figure 4 is a relative position diagram of the rotpsgen algorithm microphone and the wind turbine;

[0037] Figure 5 is an application execution flowchart of the system of the application.

[0038] In the figure, 1. Blade, 2. Hub, 3. Nacelle, 4. Tower, 5. Microphone, 6. Simulated cracking noise source. DETAILED DESCRIPTION

[0039] The application will be further described below in conjunction with the drawings and examples.

[0040] The structures, proportions, sizes, etc. shown in the drawings of the present specification are only used to cooperate with the content disclosed in the specification, to be understood and read by those skilled in the art, and do not define the limiting conditions under which the application can be implemented, so they do not have technical significance. Any modification of the structure, change of the proportion relationship or adjustment of the size, without affecting the effect and purpose that the application can produce, should still fall within the scope of the technical content disclosed by the application. At the same time, the terms such as "upper", "lower", "left", "right", "middle" and "one" in the present specification are only for the convenience of clear understanding and description, and are not used to limit the scope in which the application can be implemented, and the change or adjustment of the relative relationship, without substantially changing the technical content, is also regarded as the scope in which the application can be implemented.

[0041] As shown in Figure 1 , the application proposes a wind turbine blade health intelligent detection method based on sound signals, and the flowchart is as follows.

[0042] AsFigure 2 As shown, the acoustic data of normal operation and abnormal operation of the simulation generation part of the application is simulated, the data is augmented, the acoustic signals during the operation of the simulation equipment are simulated, and the data set is expanded, then Mel spectrum feature extraction is performed, the acoustic characteristics of blade cracking are highlighted, and a Mel spectrogram is generated.

[0043] As shown Figure 3 As shown, the deep learning recognition module adopts an improved MobileNetV3 convolutional neural network structure for feature learning and classification, and realizes binary classification recognition output of the equipment operation state.

[0044] As shown Figure 4 As shown, the application simulates the relative position of the microphone and the wind turbine when the wind turbine blade rotates the sound data by the rotpsgen algorithm, as shown in the figure, wherein 1. Blade, 2. Hub, 3. Nacelle, 4. Tower, 5. Microphone, 6. Simulated cracking noise source.

[0045] As shown Figure 5 As shown, the implementation mode of the system on the mobile terminal includes a mobile device, a mobile terminal model deployment module, and a result display module. The mobile device is responsible for sound collection and preprocessing; the mobile terminal model deployment module performs spectrum conversion and model prediction process; the result display module realizes visual analysis and abnormal early warning of the detection result, so as to realize real-time monitoring and intelligent diagnosis of the equipment operation state.

[0046] A wind turbine blade acoustic fault detection method based on transfer learning, comprising:

[0047] 1. Sound data simulation

[0048] In order to solve the problem of extremely scarce early blade crack acoustic samples in actual wind fields, the application adopts the rotpsgen algorithm, generates parameter-controllable and physically-consistent standardized fault howling samples through physical modeling of rotating monopole sound source + Doppler effect + propagation delay, and effectively supplements the training set when the measured data is insufficient. The algorithm function is:

[0049] (1)

[0050] Wherein: qm is the monopole source intensity amplitude (4-12), freq is the point sound source signal frequency (Hz, 500-6000), x0, y0, z0 are the three-dimensional coordinates of the microphone (m), R is the blade radius (m), n is the blade speed (turns / min, 10-30), and nbld is the number of blades.

[0051] First, calculate the blade angular velocity:

[0052] (2)

[0053] where ω is the angular velocity in rad / s and n is the blade speed.

[0054] The angular position of the sound source at the emission time T is:

[0055] θ(T) = ω · T (3)

[0056] where, θ(T) is the angular position, ω is the angular velocity.

[0057] The function q(T) of the sound source intensity variation over time is defined as:

[0058] (4)

[0059] where qm is the monopole source intensity amplitude, f is the signal frequency (input freq), ω(T) is a random number sequence ∈ [0, 1] used to simulate high-frequency random disturbances, T is the time.

[0060] The instantaneous spatial coordinates of the rotating sound source at time T are:

[0061] (5)

[0062] where R is the blade radius, θ is the angular position of the blade at time T.

[0063] The instantaneous distance from the sound source to the microphone is:

[0064] (6)

[0065] where x0, y0, z0 are the three-dimensional coordinates of the microphone, R is the blade radius, θ is the angular position of the blade at time T.

[0066] Considering the sound wave propagation delay, the microphone receiving time t is:

[0067] (7)

[0068] where a is the sound speed (take 348 m / s).

[0069] The instantaneous sound pressure at the microphone (including geometric attenuation and Doppler correction) is:

[0070] (8)

[0071] where, p 0 (t)is the instantaneous sound pressure, q(T) is the function of sound source intensity changing with time, T is the time, r(T) is the instantaneous distance from the sound source to the microphone, a is the sound speed, R is the radius of the blade, ω is the angular velocity, x0, y0, z0 are the three-dimensional coordinates of the microphone, θ is the angular position of the blade at time T, the second term in the denominator is the Doppler correction generated by the projection of the sound source speed in the observation direction.

[0072] Since the received time sequence t(T) is non-uniform, linear interpolation is used to obtain a uniform sampling sound pressure sequence:

[0073] (9)

[0074] where, p u (t u ) is the final uniform sampling sound pressure, p 0 (t) is the instantaneous sound pressure, t u is the uniform time grid (sampling rate 44100 Hz), Interp is the linear interpolation operation.

[0075] The algorithm output is a standard.wav audio file (sampling rate 44100 Hz), simulating the acoustic characteristics of healthy and crack fault states; combined with the field measured signal (sampling rate 16-48 kHz, resampled and aligned), a hybrid acoustic data set is formed, providing sufficient and standardized sample basis for subsequent data enhancement and model training.

[0076] 2. Data enhancement processing

[0077] Data augmentation is performed on both simulated and measured audio data to improve the robustness of the model to environmental noise, wind speed changes, and other factors, and to address small sample and class imbalance issues. Augmentation techniques include: time stretching, adjusting the playback speed of the audio by resampling, with a speed variation range of 0.8 to 1.2 times; volume perturbation, randomly adjusting the gain with an amplitude range of -9 dB to +9 dB; random noise addition, adding Gaussian noise with a noise factor range of 0.001 to 0.02; frequency shift, adjusting the pitch by spectral transformation with a shift range of -3 to +3 semitones; high-pass filtering, using a 4th-order Butterworth filter with a cutoff frequency range of 2 kHz to 8 kHz; time axis shift, randomly shifting the audio by no more than 20% of its length; signal clipping, limiting the signal amplitude within a threshold range of 0.5 to 0.9. Each augmentation randomly combines 2 to 3 techniques to generate augmented audio data, with the goal of reaching a pre-set number of samples for each class (normal and abnormal). The augmented data is stored in a separate directory with the file names appended with identifiers to distinguish between original and augmented samples. This step enhances the generalization ability of the model through diversified data augmentation.

[0078] 3. Mel-spectrogram feature extraction

[0079] The augmented audio data is subjected to time-frequency analysis, and the short-time Fourier transform (STFT) is used to extract frequency domain features to capture the acoustic differences in blade operating conditions. The calculation formula is:

[0080] (10)

[0081] where x[m] is the input signal, w[m-n] is the window function, the window function length M=2048 points, m is the summation index, indicating the time sample position within the window, n is the time frame index, k is the frequency index, the hop length is 512 points, and the sampling rate remains the original value, j is the imaginary unit.

[0082] Subsequently, the linear frequency is mapped to the Mel frequency domain through 128 groups of Mel filter banks, and the mapping formula is:

[0083] (11)

[0084] where, f lin is the linear frequency (unit: Hz), f m is the mapped Mel frequency (unit: Mel).

[0085] The logarithmic amplitude of the filtered power spectrum is taken to obtain the Mel spectrogram, and the calculation formula is:

[0086] (12)

[0087] where X Mel (n, m) is the magnitude of the mel-filtered spectrum, ε is a small constant to prevent the log from being singular, n is the time frame index, and c is the mel-frequency channel index (total of 128).

[0088] The final two-dimensional mel-spectrogram is drawn as a fixed-size image (224x224 pixels, magma mapping) and converted to three-channel RGB format to adapt to the input of the convolutional neural network. This process realizes the mapping of the time-domain acoustic signal to the mel-frequency domain feature matrix, effectively preserving the blade high-frequency whistling energy distribution, providing robust acoustic features for subsequent model classification. The converted spectrogram is stored according to the normal and abnormal categories, preserving the main frequency structure of the acoustic signal, which helps to distinguish between normal and abnormal operating states.

[0089] The generated mel-spectrogram is used to build the training dataset and validation dataset. The training data is randomly cropped (to 224x224 pixels), randomly horizontally flipped, and normalized (mean [0.485, 0.456, 0.406], standard deviation [0.229, 0.224, 0.225]); the validation data is adjusted to 256x256 pixels, then cropped to 224x224 pixels in the center and normalized.

[0090] In the training and inference stage, the system first performs sound source separation on the enhanced single-path mixed audio, using the DOA azimuth angle estimation algorithm combined with beamforming technology to generate three virtual sound channels for the blades; then the mel-spectrogram feature extraction described in step 3 is performed on each of the three virtual sound channels, obtaining three 224x224 pixel RGB images; the dynamic health baseline (pixel-by-pixel median) of the three spectrograms is calculated in real time, and the difference spectrum is obtained by subtracting the baseline from each spectrogram; At the same time, each blade maintains an individual historical health baseline, and the deviation degree of the current spectrogram from the baseline is calculated to obtain three longitudinal evolution scalars; the three transverse difference spectrograms are averaged channel by channel to obtain a comprehensive difference spectrogram as the main input image, and the three longitudinal evolution scalars are concatenated to the global feature vector before the classifier, forming a single-sample input highly consistent with the traditional training distribution, ensuring that the model maintains its generalization performance while increasing the single-blade positioning and early warning capabilities.

[0091] 4. Model construction and weight initialization

[0092] MobileNetV3-Large model is used as the basic architecture, which contains convolutional layers, reverse residual modules (containing depth separable convolution and squeeze-excitation modules) and classifiers. A double-scale time-frequency attention module (DSTFA) is inserted at the output end of the last reverse residual module (feature dimension 960). The module calculates the average pooling and maximum pooling features in the time and frequency dimensions respectively, and generates attention weights through two layers of 1x1 convolution A t and A f where A t The dimension of is:

[0093] (13)

[0094] where B is the batch size and T is the number of frames in the time dimension.

[0095] A f The dimension of is:

[0096] (14)

[0097] where B is the batch size and F is the number of mel frequency bands in the frequency dimension.

[0098] The final output is:

[0099] (15)

[0100] where X is the input feature map, A t is the time attention weight map, A f is the frequency attention weight map.

[0101] The module increases the parameter amount by about 0.23M and the calculation overhead by about 0.05 GFLOPs. The output dimension of the classifier is modified to 2 (normal and abnormal), and the original classification head is replaced with a new fully connected layer structure (input dimension 1280, output dimension 2). The model weights are initialized using the Kaiming initialization method for convolutional layers, and the standard normal distribution (mean 0, standard deviation 0.01) is used for fully connected layers. The batch normalization layer weights are set to 1 and the bias is set to 0. The pre-trained weights on the ImageNet large image data are loaded, and the classification head weights that do not match the binary classification task are removed. All feature extraction layer parameters except the DSTFA module, the last reverse residual module (1x1 convolutional layer) and the classifier are frozen to preserve the general feature extraction capability. An efficient feature extraction model is constructed to provide a foundation for transfer learning.

[0102] 5. Transfer learning and domain adaptation fine-tuning

[0103] To adapt to the actual wind farm environment, the pre-trained model in the experimental environment is applied to the actual scene using transfer learning. Audio data during the actual wind farm operation is collected, processed according to the steps of acoustic data simulation and data enhancement, a mel-spectrogram dataset is generated, and random cropping and flipping are used to enhance the samples. The pre-trained model weights are loaded, all feature extraction layer parameters except the DSTFA module, the last residual module (1x1 convolution layer) and the classifier are fixed, and only the trainable parameters are updated. Joint training of simulated data and measured data is adopted, and the total loss is defined as: total loss is defined as:

[0104] (16)

[0105] where, is the classification cross-entropy loss; is the maximum mean difference (MMD) regular term between the simulated domain and the measured domain features, which acts on the output features of the DSTFA module to reduce the distribution difference between the domains; is the predictive entropy minimization regular term, which acts on the classifier output probability to encourage the model to generate high confidence for the measured samples, and λ 1 and λ 2 is a hyperparameter used to balance the weights of each loss term.

[0106] Verify different freezing strategies with a small amount of measured samples and a large number of simulated samples (normal and abnormal each several), compare the verification accuracy and feature convergence, and determine the optimal freezing layer range (only unfreeze the DSTFA module, the last 1x1 convolution layer and the classifier). Use the cross-entropy loss function and the Adam optimizer (learning rate 0.0001, batch size 4) to optimize the model parameters through 10 iterations, calculate the prediction results through forward propagation, and adjust the parameters through back propagation. Monitor performance indicators (such as accuracy, recall rate, F1 value) during training, save the model weights with the highest verification accuracy, and generate visual curves of training loss and verification accuracy to analyze model performance. Until the verification accuracy reaches a high level, save the fine-tuned model weights. This step adjusts the model through domain adaptive regularization strategy and a small amount of actual data to ensure that it adapts to the acoustic distribution characteristics of the real wind farm and improves the model's generalization performance.

[0107] 6. Offline detection and result determination

[0108] The fine-tuned model is deployed on a portable computing platform (such as a smartphone equipped with a processor, a storage unit, and a display module, supporting Python runtime environment and deep learning framework PyTorch). The system composition includes: an audio input module (microphone, sampling rate 16 kHz to 48 kHz, collecting 10-15 second audio segments); an original audio storage module (automatically saving original.wav audio files to local encrypted storage, file name containing timestamp and device ID, supporting subsequent model update and data backtracking); sound source separation and dual-flow feature extraction module (using DOA azimuth angle estimation algorithm combined with beamforming technology, virtually separating single audio into three-blade sound channels, generating three 224x224 mel spectrum graphs respectively; real-time calculation of dynamic health baseline (pixel-by-pixel median) for each spectrum graph, subtracting the baseline from each spectrum graph to obtain a horizontal difference spectrum graph; at the same time, each blade independently maintains an individual historical health baseline, and the deviation degree of the current spectrum graph from the baseline is calculated to obtain a longitudinal evolution feature; three horizontal difference spectrum graphs are averaged channel by channel to form a comprehensive difference spectrum graph, and three longitudinal deviation degree scalars are concatenated to the global feature vector before the classifier to form a single comprehensive input consistent with the distribution in the training stage); data processing module (performing dual-flow feature extraction and fusion as described above, outputting a single 224x224 pixel RGB comprehensive image); model inference module (running MobileNetV3 model, loading fine-tuned weights); result output module (displaying single-blade diagnosis results (blade A / B / C normal / abnormal), early degradation warning level, or three-blade synchronous attenuation recognition and probability).

[0109] The operator records short-time operation audio near the wind farm, and the system automatically performs sound source separation, dual-flow feature extraction and fusion, then performs model inference, outputs single-blade fault positioning, early performance degradation warning or three-blade synchronous attenuation recognition results. When all blades are diagnosed as healthy for more than 5 consecutive times, the current comprehensive difference spectrum graph is used to slowly update (update rate 0.001-0.01) the historical health baseline of each blade; as soon as an abnormality or warning occurs, the baseline update is immediately frozen to prevent baseline pollution. At the same time, the original audio, three-blade virtual sound channels and corresponding inference results (including labels, probabilities, timestamps) are structured into dataset entries for subsequent model incremental training and performance iteration. The detection process does not require real-time sensor data or network connection, and is suitable for rapid offline detection in wild or offshore wind farms. Through single-microphone sound source separation, horizontal-longitudinal dual-flow collaborative diagnosis and safe self-evolution mechanism, the lightweight model not only realizes efficient fault diagnosis and single-blade precise positioning, but also provides a basis for device life cycle adaptive early warning and continuous learning.

[0110] 7. Visualization and auxiliary analysis

[0111] The detection system outputs the blade health state results, generates a single comprehensive mel spectrum, three lateral difference spectrum graphs, a frequency response curve and a single blade longitudinal deviation trend curve, and provides an artificial review for the operation and maintenance personnel. The lateral difference spectrum graph highlights the high-frequency whistling abnormal area of the fault blade, and the longitudinal trend curve displays the deviation of each blade over time and marks the warning threshold. The system supports storing multiple detection results, realizing long-term tracking and statistical analysis of blade state evolution through historical comparison, and generating a trend report. The operation and maintenance personnel can view the spectrum graph and curve through the display module of the portable platform, and make a comprehensive judgment combined with the classification results. This step enhances the transparency and reliability of diagnosis through visualization tools, and assists operation and maintenance decision-making.

[0112] The end-to-end acoustic diagnosis process integrates acoustic signal simulation, data enhancement, feature extraction, model training, offline detection and visualization analysis, single microphone sound source separation, inter-blade dynamic comparison and historical baseline self-evolution collaborative diagnosis, forms an independent end-to-end diagnosis system, and supports rapid fault identification, single blade fault accurate positioning, early warning of progressive damage, and three-blade synchronous attenuation identification and operation and maintenance decision-making under non-real-time monitoring conditions.

[0113] The present application solves the problem of insufficient samples by combining acoustic signal simulation with measured data; data enhancement and mel spectrum extraction make full use of the frequency domain characteristics of acoustic signals; the lightweight design of MobileNetV3 ensures efficient inference; transfer learning adapts to complex wind field environment through a small amount of actual data, reducing training cost; portable offline detection meets the needs of field operation; single microphone sound source separation combined with lateral difference spectrum and longitudinal historical baseline self-evolution mechanism realizes single blade fault accurate positioning, early warning of progressive performance degradation and accurate identification of three-blade synchronous aging, upgrading traditional "post-diagnosis" to real "predictive maintenance". Compared with traditional methods, the present application significantly improves the accuracy and adaptability of fault detection, is suitable for various wind turbine blade fault detection scenarios, and has high universality and practicality.

[0114] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A wind turbine blade acoustic fault detection method based on transfer learning, characterized in that, Comprise: (1) acoustic data simulation, using rotpsgen algorithm, input noise amplitude, frequency, blade coordinates, radius, speed, number of blades, generate signal, simulate crack fault; combined with field recording, form a mixed data set; (2) data enhancement processing, data enhancement processing is carried out on the simulated and measured audio data; (3) Mel spectrum feature extraction, short-time Fourier transform and Mel filter are used to generate frequency spectrum, which is converted into RGB image, Mel spectrum is generated, acoustic characteristics are retained, and model input is generated; Short-time Fourier transform is used to obtain audio frequency spectrum, the window function length is set to 2048 points, the hop length is set to 512 points, and the sampling rate is kept as the original value; the frequency is mapped to the Mel frequency domain through the Mel filter bank, and the two-dimensional Mel spectrum is generated by logarithmic amplitude transformation; the spectrum is drawn as a fixed size image, and the magma color mapping is used without border; the single channel spectrum is converted into three channel RGB format to adapt to the input requirements of convolutional neural network; Short-time Fourier transform is used to extract frequency domain features to capture the acoustic differences of blade running state, and the calculation formula is: (10) Wherein, wherein, x[m] is the input signal, w[m-n] is the window function, the window function length M=2048 points, m is the summation index, represents the time sample position in the window, n is the frame index, k is the frequency index, the hop length is 512 points, the sampling rate is kept as the original value, and j is the imaginary unit; The linear frequency is mapped to the Mel frequency domain through the Mel filter bank, and the mapping formula is: (11) wherein, f lin is the linear frequency, f m is the mapped mel frequency; The logarithmic amplitude of the filtered power spectrum is taken to obtain the Mel spectrum, and the calculation formula is: (12) wherein, X mel (n,c) is the mel-filtered spectral magnitude, Epsilon is a small constant to prevent log singularity, n is the time frame index, c is the mel-frequency channel index; (4) model construction and weight initialization, using MobileNetV3-Large, loading ImageNet pre-trained weight, adding double scale time-frequency attention module after the output of the last residual module, and changing the classification head to full connection layer; freeze the feature layer except DSTFA, the last residual module 1x1 convolution layer and the classifier; (5) transfer learning and domain adaptation fine tuning, joint training is carried out by using Mel spectrum and measured samples, fine tuning the weights of the last convolution layer and the classifier of MobileNetV3, using mixed data set, cross entropy loss and Adam optimizer iteration, adapting to wind field environment, and improving generalization performance; (6) offline detection and result determination, deploying the model on a portable platform, recording audio, generating frequency spectrum, and outputting normal / abnormal result and probability; (7) visualization and auxiliary analysis, generating Mel spectrum and frequency response curve, supporting manual review, realizing blade state trend analysis through history comparison, and assisting operation and decision making; (8) end-to-end acoustic diagnosis process, integrating acoustic signal simulation, data enhancement, feature extraction, model training, offline detection and visualization analysis, forming an independent end-to-end diagnosis system, supporting rapid fault identification and operation and decision making under non real-time monitoring conditions.

2. The wind turbine blade acoustic fault detection method based on transfer learning of claim 1, wherein, In step (1), the rotpsgen algorithm function is: Rotpsgen(qm,freq,x0,y0,z0,R,n,nbld) (1), Wherein: qm is the monopole source strength amplitude, freq is the point sound source signal frequency, x0, y0, z0 are the three-dimensional coordinates of the microphone, R is the blade radius, n is the blade speed, and nbld is the number of blades.

3. The wind turbine blade acoustic fault detection method based on transfer learning of claim 1, wherein, In the step (2), the data enhancement processing includes: time stretching, adjusting the audio playback speed by resampling, the speed change range being 0.8 to 1.2 times; volume disturbance, randomly adjusting the gain, the amplitude range being -9dB to +9dB; random noise superposition, adding Gaussian noise, the noise factor range being 0.001 to 0.02; frequency offset, adjusting the pitch by spectrum transformation, the shift range being -3 to +3 semitones; high-pass filtering, using a 4th order Butterworth filter, the cutoff frequency range being 2kHz to 8kHz; time axis shift, the random shift amount being not more than 20% of the audio length; signal amplitude limiting, limiting the signal amplitude in the threshold range of 0.5 to 0.9; randomly combining 2 to 3 kinds of technologies each time for enhancement, and generating the enhanced audio data, the target being to reach a preset number of each type of sample.

4. The wind turbine blade acoustic fault detection method based on transfer learning of claim 1, wherein, In the step (4), the MobileNetV3-Large model is used as the basic architecture, including a convolution layer, an inverse residual module and a classifier; the output dimension of the classifier is modified to 2, and the original classification head is replaced by a new fully connected layer structure; the model weight is initialized by the Kaiming initialization method for the convolution layer, and the standard normal distribution is used for the initialization of the fully connected layer, the batch normalization layer weight is set to 1, and the bias is set to 0; the pre-trained weight on a large image data set is loaded, and the classification head weight that does not match the binary classification task is removed.

5. The wind turbine blade acoustic fault detection method based on transfer learning of claim 1, wherein In the step (4), the double-scale time-frequency attention module calculates the average pooling and the maximum pooling features in the time dimension and the frequency dimension respectively in parallel, generates the attention weight through two layers of 1×1 convolution A t and A f wherein A t The dimension of is: B ×1× T ×1 （13） wherein, B is the batch size, T is the number of frames in the time dimension; A f The dimensions are: B ×1×1× F （14） wherein, B is the batch size, F is the number of mel-frequency bands in the frequency dimension; The final output is: Y = X ⊙ A t +X ⊙ A f （15） wherein, X is an input feature map, A t is a temporal attention weight map, A f is a frequency attention weight map.

6. The wind turbine blade acoustic fault detection method based on transfer learning of claim 1, wherein In the step (5), the simulation data and the measured data are jointly trained, and the total loss L total is defined as: L total =L ce + lambda 1 L MMD + lambda 2 L entropy （16） where, L ce is the classification cross-entropy loss; L MMD is the maximum mean discrepancy regularizer between the simulated and measured domains; L entropy is the predictive entropy minimization regularizer, acting on the classifier output probabilities, encouraging the model to generate high confidence predictions lambda 1 and lambda 2 is a hyperparameter used to balance the weights of the loss terms.

Images