Wind turbine generator blade acoustic fault detection method based on transfer learning
By combining transfer learning and lightweight networks with dynamic comparison between blades, the problems of high noise interference and data scarcity in early fault detection of wind turbine blades are solved, achieving high-precision fault identification and location, which is suitable for rapid diagnosis and intelligent operation and maintenance of wind turbines.
Patent Information
- Application Number
- CN202511793419.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-02
- Publication Date
- 2025-12-30
- Estimated Expiration
- 2045-12-02
AI Technical Summary
Existing methods for detecting faults in wind turbine blades are difficult to identify early micro-cracks and high-frequency whistling in high-noise environments. Furthermore, the large scale of model training data and high computational resource consumption make it difficult to deploy in real time on resource-constrained end-side devices. There is also a lack of standardized training samples and transferable end-to-end modeling frameworks.
Employing a transfer learning-based approach, a training sample set is constructed through acoustic simulation and multidimensional data augmentation. Using a lightweight MobileNetV3 network and a dual-scale time-frequency attention module, combined with dynamic comparison between blades and a self-evolution mechanism of historical health baselines, the system achieves accurate identification of early cracks and high-frequency whistling features and single-blade fault location, supporting offline detection.
It achieves high-precision fault identification and location under small sample conditions, reduces computational load, is suitable for deployment in portable devices, has long-term adaptive capability, and meets the needs of rapid diagnosis and intelligent operation and maintenance of wind turbine units.
Smart Images

Figure CN121234273A_ABST
Abstract
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 signals into a deep neural network through beamforming, high-pass filtering and multiple frequency spectrum features, while providing a client-server interaction architecture. Although this method improves the signal-to-noise ratio and automation level, it has high system complexity and large model calculation, which makes it difficult to be deployed in real time on the resource-limited side. Moreover, it still has limitations in recognizing early cracks and high-frequency whistling features 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 recognize acoustic features of micro cracks and early damage, which makes it 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 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 recognition of early cracks and high-frequency whistling features, accurate positioning of single-blade faults, three-blade synchronous attenuation recognition 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: A wind turbine blade acoustic fault detection method based on transfer learning, comprising: 1) Acoustic data simulation 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. 2) Data enhancement processing The simulated and a small amount of measured audio data are processed for data enhancement 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 problems. 3) Mel spectrogram feature extraction The short-time Fourier transform and 128 sets of Mel filters are used to generate a spectrogram, which is converted into a 224x224 pixel RGB image, to generate a Mel spectrogram that retains acoustic features for model input. Model construction and weight initialization MobileNetV3-Large is used as the backbone network, and ImageNet pre-trained weights are loaded. After the output of the last reverse residual module, the DSTFA (Dual-Scale Time-Frequency Attention Module) module is inserted, and the classification head is replaced with a fully connected layer. All feature layers except the DSTFA module, the last 1x1 convolution layer, and the classifier are frozen; 5) Transfer learning and domain adaptation fine-tuning Joint training is performed using enhanced simulated samples and measured samples. Based on the above model, only the unfrozen module parameters are fine-tuned. The MMD (Maximum Mean Discrepancy) and Entropy regular terms are introduced to construct a joint loss function (including cross-entropy), which is iterated 10 times by the Adam optimizer to align the features of the simulation domain and the measured domain, adapt to the wind field environment, and improve the model generalization performance; 6) Offline detection and result determination The model is deployed on a portable platform, and about 10-15 seconds of three-blade running audio is recorded to generate Mel spectrograms, infer normal / abnormal results and probabilities, single-blade fault positioning results or early warning levels, and automatically save the original audio and inference results to the local. It is suitable for offline detection in the field and does not require network connection. Cumulative data supports subsequent model updates and baseline self-evolution; 7) Visualization and auxiliary analysis 224x224 Mel spectrograms and frequency response curves, three-blade transverse difference spectrograms, and single-blade longitudinal trend curves are generated to support manual review, historical comparison for blade state trend analysis, and auxiliary 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, as well as blade dynamic comparison and historical baseline self-evolution collaborative diagnosis, an independent end-to-end diagnostic system is formed to support rapid fault identification and operation and decision-making under non-real-time monitoring conditions; 9) Inter-blade dynamic comparison and historical baseline self-evolution collaborative diagnosis and early warning After generating the three-blade Mel spectrogram in step 6), the dynamic health benchmark of the three-blade spectrogram is calculated in real time, and the transverse difference spectrogram is obtained by subtracting each blade from the benchmark; At the same time, each blade maintains an independent personal historical health baseline, and the deviation degree of the current spectrogram from the baseline is calculated to obtain the longitudinal evolution characteristics; The transverse difference spectrogram and the longitudinal evolution characteristics 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 diagnosing all health for more than 5 times in a row, update each blade baseline with current data slowly, and freeze update immediately when abnormal to prevent baseline pollution.
[0011] The beneficial effects of the present application are: The present application firstly 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 early crack acoustic sample scarcity, inconsistent data distribution and weak whistling signal easily submerged by environmental noise of wind turbine blades. 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.
[0012] 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 frequency domain at the same time, significantly improves the recognition ability of micro cracks and early damage, and combines MMD domain adaptation and entropy minimization regularization term to realize effective alignment of simulated data and real wind field data, which can obtain high generalization performance 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.
[0013] The present application further proposes an inter-blade dynamic comparison and historical health baseline self-evolution collaborative diagnosis mechanism, which realizes precise positioning of single-blade faults by suppressing common mode interference through transverse difference spectrograms; Through longitudinal baseline self-evolution, it realizes early warning of gradual degradation and effective identification of three-blade synchronous aging, 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.
[0014] This invention integrates end-to-end acoustic feature extraction and classification, avoiding redundant steps in traditional manual feature extraction and significantly improving detection efficiency and robustness. Meanwhile, the proposed model structure features small parameter count and low computational overhead, enabling it to run in real time on portable or embedded terminals. It suppresses common-mode interference through lateral difference spectrum and adapts to the natural aging of equipment through longitudinal baseline self-evolution, meeting the needs of rapid on-site diagnosis and intelligent operation and maintenance of wind turbine units. Attached Figure Description
[0015] Figure 1 This is a flowchart of the wind turbine blade damage detection method of the present invention; Figure 2 This is a data processing flowchart of the present invention; Figure 3 This is a schematic diagram of the deep learning diagnostic model structure of the present invention; Figure 4 This is a diagram showing the relative positions of the microphone and the wind turbine in the Rotpsgen algorithm. Figure 5 This is a flowchart illustrating the application execution process of the system of the present invention.
[0016] In the diagram, 1. blade, 2. hub, 3. nacelle, 4. tower, 5. microphone, 6. simulated cracking noise source. Detailed Implementation
[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0018] The structures, proportions, and sizes illustrated in the accompanying drawings are merely for illustrative purposes and to aid those skilled in the art in understanding and reading the invention. They are not intended to limit the scope of the invention and therefore have no substantial technical significance. Any modifications to the structure, changes in proportions, or adjustments to size, provided they do not affect the effectiveness or purpose of the invention, should still fall within the scope of the technical content disclosed herein. Furthermore, the terms "upper," "lower," "left," "right," "middle," and "one" used in this specification are merely for clarity and not intended to limit the scope of the invention. Changes or adjustments to their relative relationships, without substantially altering the technical content, should also be considered within the scope of the invention's implementation.
[0019] like Figure 1 As shown in the figure, this invention proposes an intelligent health detection method for wind turbine blades based on sound signals, and the process is illustrated in the figure.
[0020] like Figure 2As shown, this invention simulates and generates acoustic data of both normal and abnormal operation, performs data enhancement to simulate the acoustic signals during equipment operation and expands the dataset, then extracts Mel spectrum features to highlight the acoustic features of blade cracking, and generates a Mel spectrum.
[0021] like Figure 3 As shown, the deep learning recognition module uses an improved MobileNetV3 convolutional neural network structure for feature learning and classification, and achieves binary classification recognition output of device operating status.
[0022] like Figure 4 As shown in the figure, the relative positions of the microphone and the wind turbine are as follows when the Rotpsgen algorithm is used to simulate the sound data of the rotating blade of the wind turbine. In the figure, 1. Blade, 2. Hub, 3. Nacelle, 4. Tower, 5. Microphone, 6. Simulated cracking noise source.
[0023] like Figure 5 As shown, the system of the present invention, implemented on a mobile device, includes a mobile device, a mobile model deployment module, and a result display module. The mobile device is responsible for sound acquisition and preprocessing; the mobile model deployment module performs spectrogram conversion and model prediction; and the result display module enables visual analysis of the detection results and anomaly warnings, thereby achieving real-time monitoring and intelligent diagnosis of the device's operating status.
[0024] A method for detecting acoustic faults in wind turbine blades based on transfer learning, comprising: 1. Sound data simulation To address the extreme scarcity of acoustic samples of early blade cracks in actual wind farms, this invention employs the Rotpsgen algorithm. Through physical modeling using a rotating monopole sound source, Doppler effect, and propagation delay, it generates standardized fault whistling samples with controllable parameters and physical consistency, effectively supplementing the training set when measured data is insufficient. The algorithm function is as follows: (1) Where: qm is the intensity amplitude of the monopole source (4~12), freq is the frequency of the point source signal (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 rotation speed (rpm, 10~30), and nbld is the number of blades.
[0025] First, calculate the blade's rotational angular velocity: (2) Where ω is the angular velocity in rad / s and n is the blade rotational speed.
[0026] The angular position of the sound source at emission time T is: θ(T) = ω · T(3) in, θ(T) For the angular position, ω ω is the angular velocity.
[0027] The function q()T, which defines the change of sound source intensity over time, is defined as: (4) Where qm is the intensity amplitude of the monopole source, and f is the signal frequency (input freq). ω(T) The random number sequence ∈ [0,1] is used to simulate high-frequency random disturbances. T For a moment.
[0028] The instantaneous spatial coordinates of the rotating sound source at time T are: (5) Where R is the blade radius. θ Let be the rotation angle position of the blade at time T.
[0029] The instantaneous distance from the sound source to the microphone is: (6) Where x0, y0, z0 are the three-dimensional coordinates of the microphone, and R is the blade radius. θ Let be the rotation angle position of the blade at time T.
[0030] Considering the sound wave propagation delay, the microphone reception time t is: (7) Where a is the speed of sound (taken as 348 m / s).
[0031] The instantaneous sound pressure level at the microphone (including geometric attenuation and Doppler correction) is: (8) in, p 0 (t) Let be the instantaneous sound pressure, q(T) be the function of sound source intensity as a function of time, T be the time interval, r(T) be the instantaneous distance from the sound source to the microphone, a be the speed of sound, and R be the blade radius. ω Let x0, y0, z0 be the angular velocity, and x0, y0, z0 be the microphone's three-dimensional coordinates. θ The second term in the denominator is the Doppler correction generated by the projection of the sound source velocity onto the observation direction, representing the rotation angle position of the blade at time T.
[0032] Since the received time series t(T) is non-uniform, linear interpolation is used to obtain a uniformly sampled sound pressure sequence: (9) in, p u(t u ) To achieve uniform sampling of sound pressure, p 0 (t) For instantaneous sound pressure, t u The time grid is uniform (sampling rate 44100 Hz). Interp This is a linear interpolation operation.
[0033] The algorithm outputs a standard .wav audio file (sampling rate 44100 Hz) to simulate the acoustic characteristics of healthy and cracked fault states. Combined with field-measured signals (sampling rate 16-48 kHz, resampled and aligned), a hybrid acoustic dataset is formed, providing a sufficient and standardized sample basis for subsequent data augmentation and model training.
[0034] 2. Data Augmentation Processing Data augmentation was performed on simulated and measured audio data to improve the model's robustness to environmental noise, wind speed changes, and other factors, and to address small sample sizes and class imbalance issues. Augmentation techniques included: time stretching (adjusting audio playback speed by resampling, with speed variations ranging from 0.8 to 1.2 times); volume perturbation (randomly adjusting gain, with amplitude ranging from -9 dB to +9 dB); random noise superposition (adding Gaussian noise, with a noise factor ranging from 0.001 to 0.02); frequency shifting (adjusting pitch through spectral transformation, shifting by -3 to +3 semitones); high-pass filtering (using a 4th-order Butterworth filter, with a cutoff frequency ranging from 2 kHz to 8 kHz); time axis shifting (random shift not exceeding 20% of the audio length); and signal limiting (limiting signal amplitude within a threshold range of 0.5 to 0.9). Two to three techniques were randomly combined for each augmentation, generating enhanced audio data with the goal of reaching a preset number of samples per class (normal and abnormal). The enhanced data was stored in a classification directory, with filenames appended with identifiers to distinguish it from the original and enhanced samples. This step enhances the model's generalization ability by diversifying the data.
[0035] 3. Mel spectrogram feature extraction Time-frequency analysis was performed on the enhanced audio data, and short-time Fourier transform (STFT) was used to extract frequency domain features to capture the acoustic differences in blade operating states. The calculation formula is as follows: (10) Where x[m] is the input signal, w[mn] is the window function with a length M = 2048 points, m is the summation index representing 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 unchanged. j It is the imaginary unit.
[0036] The linear frequency is then mapped to the Mel frequency domain using 128 Mel filter banks, and the mapping formula is as follows: (11) in, f lin Linear frequency (unit: Hz) f m The mapped Mel frequency (unit: Mel).
[0037] Taking the logarithmic magnitude of the filtered power spectrum yields the Mel spectrum, calculated using the following formula: (12) Among them, X Mel (n, m) represents the amplitude of the spectrum after Mel filtering, ε is a small constant to prevent logarithmic singularity, n is the time frame index, and c is the Mel frequency channel index (128 in total).
[0038] The final result is a two-dimensional Mel-Navigate spectrum, which is plotted as a fixed-size image (224×224 pixels, Magma mapping) and converted to a three-channel RGB format to fit the input of the convolutional neural network. This process maps the time-domain acoustic signal to the Mel-Navigate frequency domain feature matrix, effectively preserving the high-frequency whistling energy distribution of the blades and providing robust acoustic features for subsequent model classification. The converted spectrum is stored according to normal and abnormal categories, retaining the main frequency structure of the acoustic signal, which helps to distinguish between normal and abnormal operating states.
[0039] The generated Mel spectrograms were used to construct the training and validation datasets. The training data was 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 was adjusted to 256x256 pixels, then centered and cropped to 224x224 pixels before being normalized.
[0040] During the training and inference phases, the system first performs source separation on the enhanced single-channel mixed audio, and uses the DOA azimuth estimation algorithm combined with beamforming technology to generate a three-blade virtual channel. Then, the Mel spectrogram feature extraction described in step 3 is performed on the three virtual channels to obtain three 224×224 pixel RGB images. The dynamic health baseline (median per pixel) of the three spectrograms is calculated in real time, and the horizontal difference spectrum is obtained by subtracting the baseline from each spectrogram. At the same time, each blade independently maintains its own historical health baseline, and the deviation of the current spectrogram from the baseline is calculated to obtain three longitudinal evolution scalars. The three horizontal difference spectrograms are averaged channel by channel to form a comprehensive difference spectrum as the main input image, and the three longitudinal evolution scalars are concatenated to the global feature vector in front of the classifier, thereby forming a single-sample input that is highly consistent with the traditional training distribution, ensuring that the model maintains generalization performance while increasing the single-blade localization and early warning capabilities.
[0041] 4. Model Building and Weight Initialization The MobileNetV3-Large model is used as the basic architecture, which includes convolutional layers, inverse residual modules (containing depthwise separable convolutions and squeeze-activation modules), and a classifier. A dual-scale time-frequency attention module (DSTFA) is inserted at the output of the last inverse residual module (feature dimension 960). This module computes average pooling and max pooling features in parallel across the time and frequency dimensions, respectively, and generates attention weights through two 1×1 convolutional layers. A t and A f ,in A t The dimensions are: (13) Where B is the batch size and T is the number of frames in the time dimension; A f The dimensions are: (14) Where B is the batch size and F is the number of Mel bands in the frequency dimension; The final output is: (15) Where X is the input feature map. A t For time-based attention weights, A f This is a frequency attention weight map.
[0042] This module adds approximately 0.23M parameters and incurs a computational cost of approximately 0.05 GFLOPs. The classifier output dimension is modified to 2 (normal and abnormal), and the original classifier head is replaced with a new fully connected layer structure (input dimension 1280, output dimension 2). Model weights are initialized using the Kaiming initialization method for convolutional layers, and fully connected layers are initialized with a standard normal distribution (mean 0, standard deviation 0.01). Batch normalization layer weights are set to 1, and biases are set to 0. Weights pre-trained on large ImageNet image datasets are loaded, and classifier head weights that are incompatible with the binary classification task are removed. All feature extraction layer parameters are frozen except for the DSTFA module, the last inverse residual module (1×1 convolutional layer), and the classifier to retain general feature extraction capabilities. An efficient feature extraction model is constructed, providing a foundation for transfer learning.
[0043] 5. Transfer learning and domain adaptation fine-tuning To adapt to real-world wind field environments, transfer learning was used to apply the pre-trained model from the experimental environment to the actual scenario. Audio data from actual wind field operation was collected and processed according to acoustic data simulation and data augmentation steps to generate a Mel spectrogram dataset. Random pruning and flipping were then applied to enhance the samples. The pre-trained model weights were loaded, and all feature extraction layer parameters except for the DSTFA module, the last inverted residual module (1×1 convolutional layer), and the classifier were fixed; only the trainable parameters were updated. Joint training with simulated and measured data was used, and the total loss was defined as: Total Loss Defined as: (16) in, For classification cross-entropy loss; The maximum mean difference (MMD) regularization term between the simulated and measured domain features is applied to the output features of the DSTFA module to reduce the distribution differences between the domains. To minimize the predictive entropy, a regularization term is applied to the classifier's output probability, encouraging the model to generate high-confidence predictions for the tested samples. λ 1 and λ 2 This is a hyperparameter used to balance the weights of the various loss terms.
[0044] Different freezing strategies were validated using a small number of real-world samples and a large number of simulated samples (some normal and some abnormal). Validation accuracy and feature convergence were compared to determine the optimal range of frozen layers (only the DSTFA module, the last 1×1 convolutional layer, and the classifier were unfrozen). The model parameters were optimized through 10 iterations using the cross-entropy loss function and the Adam optimizer (learning rate 0.0001, batch size 4). Prediction results were calculated via forward propagation, and parameters were adjusted via backpropagation. Performance metrics (such as accuracy, recall, and F1 score) were monitored during training. The model weights with the highest validation accuracy were saved, and a visualization curve of training loss versus validation accuracy was generated to analyze model performance. The model weights were saved after fine-tuning until a high level of validation accuracy was achieved. This step adjusted the model using a domain-adaptive regularization strategy and a small amount of real-world data to ensure it adapted to the acoustic distribution characteristics of real wind fields and improve model generalization performance.
[0045] 6. Offline detection and result determination The fine-tuned model is deployed on a portable computing platform (such as a smartphone, equipped with a processor, storage unit, and display module, supporting the Python runtime environment and the deep learning framework PyTorch). The system consists of: an audio input module (microphone, sampling rate 16 kHz to 48 kHz, capturing 10-15 second audio segments); a raw audio storage module (automatically saving raw .wav audio files to local encrypted storage, with filenames containing timestamps and device IDs, supporting subsequent model updates and data backtracking); and a sound source separation and dual-stream feature extraction module (using DOA azimuth estimation algorithm combined with beamforming technology to virtually separate a single audio stream into three blade channels, generating three 224×224 Mel spectrograms; calculating the dynamic health baseline (pixel-by-pixel median) of the three spectrograms in real time, subtracting each spectrogram from the baseline to obtain a lateral difference spectrum; simultaneously, each blade independently maintains its own historical health baseline and calculates the current spectrum). The deviation of the image from the baseline yields the longitudinal evolution features; the three lateral difference spectra are averaged channel by channel into a single comprehensive difference spectra, and the three longitudinal deviation scalars are concatenated to the global feature vector before the classifier to form a single comprehensive input that is completely consistent with the distribution during the training phase; the data processing module (performs the above dual-stream feature extraction and fusion, and outputs a single 224×224 pixel RGB comprehensive image); the model inference module (runs the MobileNetV3 model and loads fine-tuned weights); the result output module (displays the single-leaf diagnostic results (leaf A / B / C normal / abnormal), early degradation warning level, or identification and probability of simultaneous decay of three leaves).
[0046] The operator records short-duration audio near the wind farm. The system automatically performs sound source separation, dual-stream feature extraction and fusion, and then performs model inference, outputting single-blade fault location, early performance degradation warning, or three-blade synchronous attenuation identification results. When all blades are diagnosed as healthy for more than five consecutive times, the historical health baseline of each blade is slowly updated (update rate 0.001-0.01) using the current comprehensive difference spectrum. Once an anomaly or warning occurs, the baseline update is immediately frozen to prevent baseline contamination. Simultaneously, the original audio, the three-blade virtual channels, and the corresponding inference results (including labels, probabilities, and timestamps) are structured and stored as dataset entries for subsequent incremental model training and performance iteration. The detection process does not require real-time sensor data or network connection, making it suitable for rapid offline detection in field or offshore wind farms. Through single-microphone sound source separation, lateral-longitudinal dual-stream collaborative diagnosis, and a safety self-evolution mechanism, the lightweight model not only achieves efficient fault diagnosis and accurate single-blade location but also provides a foundation for adaptive early warning and continuous learning throughout the equipment's lifecycle.
[0047] 7. Visualization and Auxiliary Analysis While outputting the blade health status results, the detection system generates a single comprehensive Mel-ray spectrogram, three lateral difference spectrograms, a frequency response curve, and a longitudinal deviation trend curve for individual blades, for manual review by maintenance personnel. The lateral difference spectrogram highlights the high-frequency whistling abnormal area of the faulty blade, and the longitudinal trend curve shows the change in deviation of each blade over time and marks the warning threshold. The system supports storing multiple detection results, enabling long-term tracking and statistical analysis of blade status evolution through historical comparison, and generating trend reports. Maintenance personnel can view the spectrograms and curves through the display module of the portable platform and make comprehensive judgments based on the classification results. This step enhances the transparency and reliability of diagnosis through visualization tools, assisting maintenance decision-making.
[0048] The end-to-end acoustic diagnostic process integrates acoustic signal simulation, data augmentation, feature extraction, model training, offline detection and visualization analysis, single-microphone sound source separation, dynamic comparison between blades and historical baseline self-evolution collaborative diagnosis, forming an independently operating end-to-end diagnostic system that supports rapid fault identification, precise location of single-blade faults, early warning of progressive damage, and synchronous attenuation identification and maintenance decision-making for three blades under conditions without real-time monitoring.
[0049] This invention addresses the problem of insufficient samples by combining acoustic signal simulation with measured data; data augmentation and Mel-spectrum extraction fully utilize the frequency domain features of acoustic signals; the lightweight design of MobileNetV3 ensures efficient inference; transfer learning adapts to complex wind field environments with a small amount of actual data, reducing training costs; portable offline detection meets the needs of field operations; single-microphone sound source separation combined with a self-evolution mechanism of lateral difference spectrograms and longitudinal historical baselines enables precise location of single-blade faults, advance warning of progressive performance degradation months in advance, and accurate identification of simultaneous aging of three blades, upgrading traditional "post-diagnosis" to true "predictive maintenance." Compared with traditional methods, this invention significantly improves the accuracy and adaptability of fault detection, is applicable to various wind turbine blade fault detection scenarios, and has high versatility and practicality.
[0050] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.
Claims
1. A wind turbine blade acoustic fault detection method based on transfer learning, characterized in that, Comprise: (1) Acoustic data simulation, using the 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 spectrogram feature extraction, short-time Fourier transform and Mel filter are used to generate spectrogram, which is converted into RGB image, Mel spectrogram is generated, acoustic features are retained, and model input is generated; (4) Model construction and weight initialization, MobileNetV3-Large is used, ImageNet pre-trained weight is loaded, double scale time-frequency attention module is added after the output of the last residual module, and full connection layer is changed to the classification head; freeze the feature layer except DSTFA, the last residual module 1x1 convolution layer and the classifier; (5) Transfer learning and domain adaptation fine tuning, use enhanced Mel spectrogram and measured samples for joint training, fine tune MobileNetV3 last convolution layer and classifier weight, use mixed data set, cross entropy loss and Adam optimizer iteration, adapt to wind field environment, improve generalization performance; (6) Offline detection and result determination, deploy the model on the portable platform, record the audio, generate the spectrogram, infer the normal / abnormal result and probability; (7) Visualization and auxiliary analysis, generate Mel spectrogram and frequency response curve, support manual review, realize blade state trend analysis through history comparison, assist operation and decision; (8) End-to-end acoustic diagnosis process, integrate acoustic signal simulation, data enhancement, feature extraction, model training, offline detection and visualization analysis, form an independent end-to-end diagnosis system, support fast fault identification and operation decision under no real-time monitoring condition.
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: (1), Wherein: qm is the monopole source intensity 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 step (2), the data enhancement processing includes: time stretching, adjusting the audio playback speed by resampling, the speed change range is 0.8 to 1.2 times; volume disturbance, randomly adjusting the gain, the amplitude range is-9 dB to +9 dB; random noise superposition, adding Gaussian noise, the noise factor range is 0.001 to 0.02; frequency shift, adjusting the pitch by spectrum transform, the shift range is-3 to +3 semitones; high-pass filter, using 4th order Butterworth filter, cutoff frequency range is 2 kHz to 8 kHz; time axis shift, the random shift amount is not more than 20% of the audio length; signal limiting, limit the signal amplitude in the threshold range of 0.5 to 0.9; randomly combine 2 to 3 kinds of technology each time, generate enhanced audio data, the target is to reach the 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 (3), the short-time Fourier transform is used to obtain the audio 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 a mel filter bank, and a two-dimensional mel spectrum graph is generated through logarithmic amplitude transformation; the spectrum graph is drawn as a fixed size image, the magma color mapping is used, and there is no border; the single-channel spectrum graph is converted into a three-channel RGB format to adapt to the input requirements of the convolutional neural network.
5. The wind turbine blade acoustic fault detection method based on transfer learning of claim 4, wherein, The frequency domain features are extracted by using the short-time Fourier transform to capture the acoustic differences of the blade running state, and the calculation formula is: (10) 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, which represents the time sample position in the window, n is the time 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.
6. The wind turbine blade acoustic fault detection method based on transfer learning as claimed in claim 4, wherein, The linear frequency is mapped to the mel frequency domain through a mel filter bank, and the mapping formula is: (11) wherein, f lin is the linear frequency, f m is the mapped mel frequency.
7. The wind turbine blade acoustic fault detection method based on transfer learning as claimed in claim 4, wherein, The logarithmic amplitude of the filtered power spectrum is taken to obtain the mel spectrum graph, and the calculation formula is: (12) where X Mel (n, m) is the magnitude of the mel-filtered spectrum, ε is a small constant to prevent log singularity, n is the time frame index, and c is the mel-frequency channel index.
8. The wind turbine blade acoustic fault detection method based on transfer learning as claimed in claim 1, wherein, In the step (4), the MobileNetV3-Large model is used as the basic architecture, which includes convolutional layers, reverse residual modules and classifiers; the output dimension of the classifier is modified to 2, and the original classification head is replaced with a new fully connected layer structure; the model weight is initialized by using the Kaiming initialization method for the convolutional layer, the standard normal distribution for the fully connected layer, the batch normalization layer weight set to 1, and the bias set to 0; the pre-trained weight on the large image dataset is loaded, and the classification head weight that does not match the binary classification task is removed.
9. The wind turbine blade acoustic fault detection method based on transfer learning as claimed in claim 1, characterized in that 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 is: (13) Wherein, B is the batch size, and T is the number of frames in the time dimension; A f The dimensions are: (14) Wherein, B is the batch size, and F is the number of mel bands in the frequency dimension; The final output is: (15) wherein X is an input feature map, A t is a time attention weight map, A f is a frequency attention weight map.
10. The wind turbine blade acoustic fault detection method based on transfer learning as claimed in claim 1, characterized in that In step (5), the simulation data and the measured data are jointly trained, and the total loss is defined as: (16) where, is the categorical cross-entropy loss; is the maximum mean discrepancy regularizer between the simulated and measured domains; is the predictive entropy minimization regularizer, acting on the classifier output probabilities, encouraging the model to generate high confidence predictions λ 1 and λ 2 is a hyperparameter balancing the weights of the loss terms.
Citation Information
Patent Citations
A method and system for detecting damage to wind turbine blades based on deep learning
CN114565006B
Draught fan blade fault acoustic detection method and system based on deep learning
CN116631442A
Wind turbine blade crack detection system based on acoustic signature analysis
CN119491796B
Equipment defect analysis method and system based on voiceprint system and transfer learning
CN120089158A
Motor fault detection method and system based on voiceprint recognition
CN120431966A
Cited By
Edge calculation-oriented heterogeneous electromagnetic signal multi-mode lightweight identification method and device
CN121615094A
Wind turbine generator blade efficiency diagnosis method and device based on field operation data
CN121760895A
Wind turbine blade efficiency diagnosis method and device based on field operation data
CN121760895B
