Fan blade abnormal sound labeling sample generation method, abnormal sound detection method and system

Abstract

This application discloses a method for generating abnormal noise annotation samples, an abnormal noise detection method, and a system for wind turbine blades, relating to the field of wind power equipment condition monitoring technology. The method includes configuring the motion parameters of the sound source as the blade rotates, simulating the Doppler frequency shift effect to generate an instantaneous frequency change sequence and an abnormal noise simulation signal, superimposing these signals with background noise and annotating their time-frequency domain location information to generate training samples for training a target detection model. In the detection scenario, the wind turbine operating audio is acquired and a spectrogram is generated, which is input into the target detection model to detect the abnormal noise trajectory. Based on the detection results, the existence of the abnormal noise and the location parameters of the sound source are determined. This application can solve the problem of poor model generalization ability caused by the scarcity of real abnormal noise samples and achieve accurate localization of abnormal noise sources.

Description

Technical Field

[0001] This application relates to the field of wind power equipment condition monitoring technology, and more specifically, to a method for generating abnormal noise annotation samples for wind turbine blades, an abnormal noise detection method and system. Background Technology

[0002] As the core component for capturing wind energy, wind turbine blades operate under complex conditions such as variable loads, wind erosion, and alternating temperatures for extended periods, making them prone to surface cracks or edge damage. In existing wind power operation and maintenance scenarios, acoustic monitoring technology for blade damage has been applied to some extent. The mainstream approach typically involves collecting the turbine's operating audio through microphones deployed on the nacelle or tower. Subsequently, deep learning models such as convolutional neural networks are used to train the audio features, learning the differences between normal and abnormal signals, and ultimately outputting a binary classification result indicating whether abnormal noise exists or is normal.

[0003] However, in existing technologies, abnormal noises caused by early damage to real wind turbine blades are sporadic events, resulting in an extreme scarcity of high-quality labeled samples suitable for model training. This leads to low robustness of existing detection models. In particular, it is often difficult to maintain stable recognition accuracy when facing low signal-to-noise ratio environments or short-term abnormal noise signals. Summary of the Invention

[0004] In view of the above problems, this application proposes a method for generating abnormal noise annotation samples, an abnormal noise detection method, and a system for wind turbine blades. This method generates simulated abnormal noise samples carrying abnormal noise annotation information, overcoming the problem of poor model robustness caused by the scarcity of real wind turbine blade abnormal noise samples, and improving the detection model's accuracy in recognizing abnormal noise signals. The specific solution is as follows:

[0005] Firstly, a method for generating annotation samples of abnormal noise from wind turbine blades is provided, including:

[0006] Configure the motion parameters of the sound source corresponding to the abnormal noise as it rotates with the fan blades;

[0007] Based on the motion parameters, the Doppler frequency shift effect generated by the sound source relative to the audio acquisition device during the rotation process is simulated to obtain the instantaneous frequency change sequence of the sound source during the rotation process;

[0008] Based on the instantaneous frequency change sequence, generate an abnormal sound simulation signal with frequency shift characteristics;

[0009] The abnormal noise simulation signal is superimposed with the background noise during the operation of the wind turbine to obtain an abnormal noise simulation sample, and the abnormal noise information is labeled.

[0010] In one possible design, in another implementation of the first aspect of the embodiments of this application, the motion parameters include the angular velocity or rotation period of the wind turbine blades, and the distance of the sound source from the center of the wind turbine hub.

[0011] In one possible design, in another implementation of the first aspect of this application, the process of simulating the Doppler frequency shift effect of the sound source relative to the audio acquisition device during rotation, based on the motion parameters, to obtain the instantaneous frequency change sequence of the sound source during rotation includes:

[0012] Based on the motion parameters, calculate the instantaneous relative velocity of the sound source with respect to the audio acquisition device during the rotation process;

[0013] Based on the instantaneous relative computing speed and the configured heterophonic fundamental frequency, the instantaneous frequency at each moment is calculated using the Doppler frequency shift formula to obtain the instantaneous frequency change sequence.

[0014] In one possible design, in another implementation of the first aspect of the embodiments of this application, the process of calculating the instantaneous relative velocity of the sound source with respect to the audio acquisition device during rotation, based on the motion parameters, includes:

[0015] The linear velocity of the sound source as the blade rotates is determined based on the motion parameters.

[0016] Based on the relative position between the current rotational position of the sound source and the audio acquisition device, the velocity component of the linear velocity in the direction of the line connecting the sound source and the audio acquisition device is determined as the instantaneous relative motion velocity.

[0017] In one possible design, in another implementation of the first aspect of the embodiments of this application, the motion parameters include the angular velocity or rotation period of the wind turbine blades, and the distance of the sound source from the center of the wind turbine hub.

[0018] The process of determining the linear velocity of the sound source as it rotates with the blade based on the motion parameters includes:

[0019] The linear velocity of the sound source as it rotates with the blades is determined by multiplying the distance of the sound source from the center of the wind turbine hub with the rotational angular velocity of the wind turbine blades.

[0020] or,

[0021] The rotational angular velocity of the wind turbine blades is calculated based on the rotation period, and the linear velocity of the sound source as the blades rotate is determined by multiplying the distance of the sound source from the center of the wind turbine hub with the calculated rotational angular velocity of the wind turbine blades.

[0022] In one possible design, in another implementation of the first aspect of the embodiments of this application, the abnormal sound information labeled on the abnormal sound simulation sample includes:

[0023] The location information of the abnormal sound simulation sample in the time-frequency domain.

[0024] In one possible design, in another implementation of the first aspect of the embodiments of this application, the abnormal sound information labeled on the abnormal sound simulation sample further includes:

[0025] The position parameters of the sound source corresponding to the simulated abnormal sound signal on the wind turbine blade.

[0026] In one possible design, in another implementation of the first aspect of the embodiments of this application, the location information in the time-frequency domain includes:

[0027] The start and end positions of the simulated noise sample on the time axis, and the lowest and highest frequencies on the frequency axis.

[0028] In one possible design, another implementation of the first aspect of the embodiments of this application further includes:

[0029] The simulated samples of abnormal sounds labeled with abnormal sound information are used to train a target detection model, which is used to automatically detect abnormal sound sources from the spectrogram of the wind turbine operating audio.

[0030] Secondly, a method for detecting abnormal noise in wind turbine blades is provided, including:

[0031] Acquire the audio signal to be detected generated during the operation of the wind turbine;

[0032] The audio signal to be detected is subjected to time-frequency transformation to generate a spectrogram;

[0033] The spectrogram is input into a trained target detection model, which is used to detect the trajectory shape of the abnormal sound signal in the spectrogram caused by the Doppler frequency shift effect due to the blade rotation.

[0034] Based on the detection results output by the target detection model, the final abnormal sound detection result is determined.

[0035] In one possible design, in another implementation of the second aspect of the embodiments of this application, the detection result output by the target detection model includes the detection box corresponding to the abnormal sound signal in the spectrogram and the confidence level of the detection box, and the detection box covers the time-frequency trajectory of the abnormal sound signal;

[0036] The process of determining the final abnormal sound detection result includes:

[0037] If a target detection box with a confidence level higher than a preset threshold exists, it is determined that there is an abnormal sound signal in the audio signal to be detected;

[0038] If there is no detection box with a confidence level higher than the preset threshold, it is determined that there is no abnormal sound signal in the audio signal to be detected.

[0039] In one possible design, in another implementation of the second aspect of the embodiments of this application, the process of determining the final abnormal sound detection result further includes:

[0040] When it is determined that there is an abnormal sound signal in the audio signal to be detected, the abnormal sound signal information is output;

[0041] The abnormal sound signal information includes the start and end positions of the abnormal sound signal on the time axis, and the lowest and highest frequencies on the frequency axis.

[0042] In one possible design, in another implementation of the second aspect of the embodiments of this application, the process of determining the final abnormal sound detection result further includes:

[0043] Extract the highest and lowest frequencies of the abnormal sound signal from the target detection box;

[0044] Based on the highest and lowest frequencies, and combined with the Doppler frequency shift relationship, the linear velocity of the sound source corresponding to the abnormal sound signal as the fan blades rotate is determined.

[0045] Based on the linear velocity of the sound source corresponding to the abnormal noise signal as it rotates with the wind turbine blades, and the rotation period or angular velocity of the wind turbine blades, the distance from the sound source corresponding to the abnormal noise signal to the center of the wind turbine hub is calculated and used as the location parameter of the sound source corresponding to the abnormal noise signal.

[0046] Thirdly, a wind turbine blade abnormal noise detection system is provided, including: audio acquisition equipment and data processing equipment;

[0047] The audio acquisition device is used to acquire the audio signal to be detected generated during the operation of the fan;

[0048] The data processing device is used to implement each step of the wind turbine blade noise detection method described in any of the second aspects of this application.

[0049] Fourthly, a readable storage medium is provided, on which a computer program is stored, wherein when the computer program is executed by a processor, it implements the steps of the method for generating wind turbine blade noise annotation samples as described in any of the preceding first aspects of this application, or implements the steps of the method for detecting wind turbine blade noise as described in any of the preceding second aspects of this application.

[0050] Fifthly, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps of the method for generating wind turbine blade noise annotation samples as described in any of the first aspects of this application, or implements the steps of the method for detecting wind turbine blade noise as described in any of the second aspects of this application.

[0051] By employing the aforementioned technical solution, this application obtains simulated noise samples labeled with noise information by configuring motion parameters, simulating the Doppler frequency shift effect, generating simulated noise signals with frequency shift characteristics, and superimposing them with background noise for annotation. Through the combination of motion parameters and the Doppler frequency shift effect, the dynamic frequency changes of the noise signal caused by blade rotation can be reproduced mechanistically, solving the problem that traditional methods cannot generate samples with realistic dynamic characteristics. By superimposing the simulated signal with background noise, complex field environments are effectively simulated, improving the generalization ability of the data. This application can generate high-quality simulated noise samples, significantly improving the robustness of subsequent noise detection models in recognizing low signal-to-noise ratio, short-duration noise signals.

[0052] Furthermore, in the detection scenario, this application converts the audio signal acquired through time-frequency transformation into a two-dimensional spectrogram, highlighting the Doppler frequency shift trajectory characteristics of the abnormal sound signal caused by the blade rotation, thus solving the problem that traditional time-domain analysis is difficult to capture dynamic frequency changes. Secondly, a specially trained target detection model is introduced. Utilizing the trained model's keen ability to capture specific trajectory patterns in the image (the trajectory pattern corresponding to the abnormal sound signal caused by the Doppler frequency shift effect due to blade rotation), automatic identification and localization of the abnormal sound signal is achieved, overcoming the problem of low recognition accuracy of traditional binary classification models when facing low signal-to-noise ratio environments or short-duration abnormal sound signals. In addition, the target detection model of this application can locate the time-frequency boundary data of the abnormal sound signal in the spectrogram, providing basic data for subsequent calculation of the specific location of the sound source.

[0053] It should be noted that the method for generating wind turbine blade noise annotation samples and the method for detecting wind turbine blade noise provided in the above embodiments of this application belong to the same general inventive concept.

[0054] In an embodiment of the labeled sample generation method, by configuring the motion parameters of the sound source corresponding to the abnormal noise as it rotates with the wind turbine blades, the Doppler frequency shift effect generated by the sound source relative to the audio acquisition device during rotation is simulated. This obtains the instantaneous frequency change sequence of the sound source during rotation, thereby generating an abnormal noise simulation signal with frequency shift characteristics. This signal is then superimposed with background noise to form labeled training samples. The core of this method lies in using the Doppler frequency shift effect to physically simulate the blade abnormal noise signal, generating training data carrying dynamic frequency shift characteristics.

[0055] In an embodiment of the abnormal noise detection method, a spectrogram is generated by performing time-frequency transformation on the audio signal to be detected during wind turbine operation. A target detection model is then used to detect the trajectory shape corresponding to the abnormal noise signal in the spectrogram caused by the Doppler frequency shift effect due to blade rotation, thereby determining whether abnormal noise exists and outputting the corresponding detection result. The core of this method lies in using a target detection model to identify and locate abnormal noise signals exhibiting Doppler frequency shift characteristics in the spectrogram.

[0056] Both methods involve the Doppler frequency shift effect caused by the rotation of wind turbine blades: the sample generation method simulates the Doppler frequency shift effect to construct training samples, while the detection method identifies the Doppler frequency shift effect in actual signals to detect abnormal noise. They are technically related (solving the reliable identification and location of abnormal noise signals from wind turbine blades) and their technical solutions complement each other (both focus on processing the Doppler frequency shift effect), belonging to the same inventive concept. Attached Figure Description

[0057] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0058] Figure 1 This application provides a schematic flowchart of a method for generating noise annotation samples for wind turbine blades.

[0059] Figure 2 A schematic diagram illustrating the Doppler effect principle of abnormal noise in wind turbine blades, provided in an embodiment of this application;

[0060] Figure 3 A schematic diagram of a detection box for heterophonic signals on a spectrogram provided in this application embodiment;

[0061] Figure 4 A schematic diagram illustrating the positional relationship between the sound source during the rotation of wind turbine blades, the center of the wind turbine hub, and the audio acquisition device, provided in an embodiment of this application;

[0062] Figure 5 This is a schematic flowchart of a method for detecting abnormal noise in wind turbine blades, provided in an embodiment of this application. Detailed Implementation

[0063] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0064] In the field of wind power equipment condition monitoring and fault diagnosis, wind turbine blades are subject to long-term effects such as variable loads, wind erosion, and alternating temperatures, making them prone to surface cracks or edge damage. During rotation, blade damage generates abnormal friction with the air, emitting short-duration, non-stationary noise signals. Furthermore, influenced by the blade's rotational motion, these noise signals exhibit a typical Doppler frequency shift effect. Existing acoustic monitoring solutions largely rely on manual inspections or conventional classification models.

[0065] However, in the current wind power operation and maintenance scenario, real blade noise faults are sporadic events, and labeled samples that can be used for model training are extremely scarce; this results in insufficient robustness of the model in recognizing low signal-to-noise ratio and short-term noise signals, leading to a high false alarm and false negative rate.

[0066] Furthermore, traditional classification models have not been optimized for training to address the Doppler frequency shift characteristics caused by blade rotation. Some solutions even aim to eliminate Doppler frequency shift characteristics by distributing multiple microphones, making it difficult for traditional classification models to capture the dynamic frequency shift trajectory of abnormal sound signals, further reducing the accuracy of abnormal sound signal recognition.

[0067] Furthermore, traditional solutions only achieve binary classification detection of the presence or absence of abnormal noise through classification models, which cannot accurately locate the source of the abnormal noise on the wind turbine blades, making it difficult to support subsequent targeted operation and maintenance work.

[0068] This application provides a method for generating abnormal noise annotation samples for wind turbine blades, as well as a method for detecting abnormal noise in wind turbine blades, which can at least solve some of the problems existing in the current acoustic monitoring schemes.

[0069] First participate Figure 1 This application provides a method for generating noise annotation samples for wind turbine blades, which may include the following steps:

[0070] Step S100: Configure the motion parameters of the sound source corresponding to the abnormal noise as it rotates with the fan blades.

[0071] Among them, motion parameters can refer to the set of physical quantities used to describe the changes of the sound source on the wind turbine blade as it rotates. Their function is to define the motion state of the sound source in the simulation environment and provide basic input data for subsequent simulation of the Doppler frequency shift effect.

[0072] In some examples, motion parameters may include the angular velocity or rotation period τ of the wind turbine blades, and the distance L from the sound source to the center of the wind turbine hub. These parameters can be configured using preset typical operating condition values ​​or set based on actual wind turbine operation data. By configuring diverse combinations of motion parameters, the motion characteristics of sound sources under different wind turbine models and operating conditions can be covered, thereby generating simulation data with broad representativeness. This step aims to construct the physical boundary conditions of the simulation, ensuring that the generated abnormal sound signals conform to the actual motion laws of wind turbine blade rotation.

[0073] Step S110: Based on the motion parameters, simulate the Doppler frequency shift effect generated by the sound source relative to the audio acquisition device during the rotation process to obtain the instantaneous frequency change sequence of the sound source during the rotation process.

[0074] Among them, the simulated Doppler frequency shift effect can refer to the process of calculating the dynamic change of the received frequency caused by the relative motion between the sound source and the audio acquisition device based on the acoustic Doppler principle.

[0075] Audio acquisition equipment can be installed on the wind turbine tower or other fixed locations to collect audio signals during the operation of the wind turbine.

[0076] The instantaneous frequency change sequence is a set of frequency values ​​calculated at each instant by discretizing time and using the Doppler frequency shift formula combined with the current instantaneous relative velocity. By traversing a complete rotation cycle, a periodic sinusoidal instantaneous frequency change sequence can be obtained. This sequence accurately reproduces the dynamic drift characteristics of heterophonic frequencies caused by the rotation of wind turbine blades, providing core data support for generating realistic simulation signals.

[0077] Reference Figure 2 This is a schematic diagram illustrating the Doppler effect principle of wind turbine blade noise provided in an embodiment of this application. It illustrates the instantaneous frequencies corresponding to different positions of the wind turbine blade.

[0078] Step S120: Generate an abnormal sound simulation signal with frequency shift characteristics based on the instantaneous frequency change sequence.

[0079] The generation of the simulated sound signal can refer to synthesizing an audio waveform whose frequency dynamically changes over time in the time domain, based on the instantaneous frequency change sequence obtained in the previous step. This process can be achieved using phase accumulation or direct frequency modulation, that is, controlling the output frequency of the oscillator according to the instantaneous frequency at each moment, thereby synthesizing a continuously changing sound wave signal.

[0080] As a pure fault characteristic signal, the abnormal noise simulation signal exhibits waveform characteristics highly consistent with those produced by real blade cracks. For example, a 10-second sinusoidal frequency-modulated signal can be generated, with its frequency fluctuating between 1800Hz and 2200Hz according to a sequence of instantaneous frequency changes, simulating the whistling sound of an abnormal noise during one rotation of the blade. The signal generated in this way not only contains the fundamental frequency information of the abnormal noise but also completely preserves the Doppler frequency shift trajectory characteristics caused by rotational motion, solving the problem of capturing dynamic features in real fault samples.

[0081] Step S130: Superimpose the abnormal noise simulation signal with the background noise during the operation of the wind turbine to obtain an abnormal noise simulation sample, and label the abnormal noise information.

[0082] The background noise during wind turbine operation can refer to the environmental sound generated by the wind turbine under normal operating conditions, specifically including one or more of wind noise, gearbox noise, and generator noise. Optionally, the background noise can be collected from multiple different wind turbine operating conditions.

[0083] The superposition process can involve linearly superimposing the generated simulated noise signal with the collected real background noise signal at a preset signal-to-noise ratio to simulate a scenario where abnormal noise is masked by strong background noise under real-world conditions. For example, multiple signal-to-noise ratios such as -5dB, 0dB, and 5dB can be selected to generate simulated noise samples with different noise levels, thereby enhancing the diversity of the training data.

[0084] Labeling abnormal sound information refers to annotating the generated abnormal sound simulation samples. The annotation methods can be varied, such as: labeling the abnormal sound simulation samples with a binary classification tag indicating whether they carry an abnormal sound; or, labeling the abnormal sound simulation samples with time-frequency domain feature markers, indicating the location information of the abnormal sound simulation signal in the time-frequency domain, specifically including the start and end positions of the abnormal sound on the time axis, and the lowest and highest frequencies on the frequency axis. In practice, a short-time Fourier transform can be performed on the abnormal sound simulation samples to generate a spectrogram, and then a rectangular area containing the abnormal sound signal can be selected on the spectrogram, recording the spatiotemporal coordinates of that area. (Refer to...) Figure 3 As shown, a rectangular detection box is selected on the spectrogram to indicate the location of the abnormal sound signal, which covers the start and end time and frequency range of the abnormal sound signal.

[0085] This automated annotation method generates a large number of training samples with accurate labels, eliminating the need for manual listening and annotation of each sample, and greatly reducing data preparation costs.

[0086] The abnormal sound simulation samples labeled with abnormal sound information obtained in this step can be used as training data for the subsequent target detection model. This target detection model is used to automatically detect abnormal sound sources from the spectrogram of the wind turbine operating audio.

[0087] The method described in this application generates simulated noise samples with noise annotations by configuring motion parameters, simulating the Doppler frequency shift effect, generating frequency shift feature signals, and superimposing them with background noise. By combining motion parameters with the Doppler frequency shift formula, the dynamic frequency changes caused by blade rotation can be reproduced mechanistically, solving the problem that traditional methods cannot generate samples with realistic dynamic features. By superimposing the simulated signal with background noise under various operating conditions at different signal-to-noise ratios, complex field environments are effectively simulated, improving the generalization ability of the data. This method can generate a large number of high-quality noise simulation samples, enriching the number of training samples and improving the robustness and generalization of the trained target detection model.

[0088] In some embodiments of this application, the motion parameters configured in the aforementioned step S100 may include the angular velocity ω or rotation period τ of the wind turbine blades, and the distance L from the sound source to the center of the wind turbine hub.

[0089] The angular velocity ω of a wind turbine blade can refer to the angle through which the blade rotates per unit time; the rotation period τ can refer to the time required for the blade to complete one rotation. The two are reciprocals of each other, i.e., ω = 2π / τ.

[0090] The distance L from the sound source to the center of the wind turbine hub can refer to the radial length of the damaged point producing the abnormal sound on the blade relative to the center of rotation. This distance determines the radius of the circular motion of the sound source. By clarifying these two core geometric and kinematic parameters, the circular motion state of the sound source in space can be accurately described, providing the necessary physical quantity input for subsequent calculations of instantaneous relative velocity.

[0091] Furthermore, the volume v can also be configured. s It can be taken as 340m / s under standard operating conditions, or it can be corrected according to the ambient temperature.

[0092] It is also possible to configure the fundamental frequency f0 of the sound source, which refers to the inherent frequency of the abnormal sound signal continuously emitted by the crack (or damaged part) on the blade during rotation due to friction, resonance or airflow disturbance.

[0093] This application concretizes abstract motion parameters into measurable engineering parameters, giving the Doppler frequency shift simulation process a clear physical meaning. The angular velocity or rotation period of the wind turbine blades, in conjunction with the distance between the sound source and the center of the wind turbine hub, jointly determine the magnitude of the linear velocity of the sound source as the blades rotate. Linear velocity is the direct driving factor for the Doppler frequency shift effect. Different combinations of angular velocity and radius will generate simulation signals with different frequency variation amplitudes, thus flexibly adapting to different wind turbine models or different damage locations, ensuring that the generated abnormal sound simulation samples cover diverse actual working conditions.

[0094] Furthermore, in some embodiments of this application, the aforementioned step S110 is described as an optional implementation process for simulating the Doppler frequency shift effect generated by the sound source relative to the audio acquisition device during rotation based on motion parameters, thereby obtaining the instantaneous frequency change sequence of the sound source during rotation.

[0095] In one alternative implementation, the instantaneous relative velocity of the sound source with respect to the audio acquisition device during rotation can be calculated based on motion parameters.

[0096] Instantaneous relative velocity can refer to the projection component of the linear velocity of the sound source along the line connecting the sound source and the audio acquisition device when the sound source moves in a circle with the blade. Its function is to quantify the dynamic rate at which the sound source approaches or moves away from the audio acquisition device, and directly determine the amplitude and positive or negative direction of the Doppler frequency shift.

[0097] In this embodiment, the linear velocity of the sound source rotating with the blade can be determined based on the motion parameters. Based on the current rotation position of the sound source and the relative orientation between the audio acquisition device and the sound source, the velocity component of the linear velocity in the direction of the line connecting the sound source and the audio acquisition device can be determined as the instantaneous relative motion velocity.

[0098] Combination Figure 4 As shown, the audio acquisition device can be installed at the bottom of the wind turbine tower. Since the tower is wider at the bottom than at the top, the wind turbine blades, the center O of the wind turbine hub, and the position M of the audio acquisition device can be approximated as being in the same vertical plane.

[0099] When the motion parameters include the angular velocity ω of the wind turbine blades and the distance L from the sound source to the center of the hub, the product of the two can be taken as the linear velocity V=ωL of the sound source as it rotates with the blades.

[0100] When the motion parameters include the rotation period τ of the wind turbine blades and the distance L from the sound source to the center of the hub, the rotational angular velocity of the wind turbine blades can be calculated based on the rotation period τ. The angular velocity can then be multiplied by the distance L from the sound source to the center of the hub to obtain the linear velocity V = 2πL / τ of the sound source as it rotates with the blades.

[0101] Subsequently, the angle θ(t) is determined based on the geometric relationship between the rotational position of the sound source at time t and the microphone position. θ(t) is the angle between the ray pointing from the sound source position to the center O of the wind turbine hub at time t and the ray pointing from the sound source position to the position M of the audio acquisition device. Then, the instantaneous relative velocity V... D (t)=V⋅sin(θ(t)).

[0102] When the sound source is located at point A, θ(t) = π, then V D (t)=0; when the sound source is located at point B, θ(t)=π / 2, then V D θ(t) = V; when the sound source is located at point C, θ(t) = 0, then V D (t)=0; when the sound source is located at point D, θ(t)=3π / 2, then V D (t) = -V. Figure 4 The example only uses the angle θ when the sound source is located at point D.

[0103] Furthermore, based on the instantaneous relative computation speed V calculated above... D (t) and the configured heterophonic fundamental frequency The instantaneous frequency at each moment is calculated using the Doppler frequency shift formula. This yields the instantaneous frequency change sequence.

[0104] The Doppler frequency shift formula can be expressed as:

[0105]

[0106] in, V represents the speed of sound. D (t) represents the relative velocity between the sound source and the audio acquisition device at different moments during the rotation process. When the sound source moves towards the audio acquisition device, V... D (t) is a negative number, V when the sound source moves away from the audio acquisition device. D (t) is a positive number.

[0107] When the sound source moves toward the audio acquisition device to point D, the instantaneous relative velocity is at its minimum, and the calculated instantaneous frequency reaches its peak value; when the sound source moves away from the audio acquisition device to point B, the instantaneous relative velocity is at its maximum, and the calculated instantaneous frequency reaches its trough value.

[0108] The method in this embodiment first calculates the instantaneous relative velocity based on motion parameters, clarifying the dynamic geometric relationship between the sound source and the audio acquisition device. Then, this velocity variable is coupled with the fundamental frequency of the abnormal sound and substituted into the Doppler frequency shift formula to complete the calculation from motion parameters to acoustic frequency characteristics. Based on this, the obtained instantaneous frequency change sequence is directly used as the driving basis for generating the abnormal sound simulation signal, ensuring a high degree of consistency between the simulation sample and the real fault signal in the time-frequency domain characteristics. This effectively solves the technical problems of insufficient model training data and poor generalization ability caused by the scarcity of real fault samples.

[0109] In some embodiments of this application, after superimposing the abnormal noise simulation signal with the background noise during wind turbine operation to obtain abnormal noise simulation samples, abnormal noise information can be labeled on the abnormal noise simulation samples to obtain training data. The labeling process will be described below.

[0110] In one optional annotation method, the simulated abnormal sound sample can be labeled with a binary label indicating whether it carries an abnormal sound, that is, the labeled abnormal sound information is a binary label.

[0111] The training data obtained using this annotation method, after training the object detection model, can yield a binary classification detection model capable of distinguishing whether an audio signal carries abnormal sounds. Furthermore, since this application can simulate and generate a large amount of training data, it can train an object detection model with better robustness and stronger generalization.

[0112] In another optional annotation method, the abnormal sound information annotated for the abnormal sound simulation sample may include the location information of the abnormal sound simulation sample in the time-frequency domain.

[0113] In this context, the location information in the time-frequency domain refers to the geometric coordinates of the simulated noise sample on a two-dimensional time-frequency image (i.e., spectrogram). This location information is obtained by mapping the generated simulated noise sample to the time-frequency plane using a Short-Time Fourier Transform (STFT) and then extracting the boundary parameters of its energy concentration region. Its function is to serve as a true label in the supervised learning process, guiding the target detection model to learn the distribution pattern and Doppler frequency shift trajectory characteristics of the noise signal in the time-frequency domain.

[0114] In one optional implementation, the position information of the aforementioned abnormal sound simulation sample in the time-frequency domain may include the start and end positions of the abnormal sound simulation sample on the time axis, and the lowest and highest frequencies on the frequency axis.

[0115] In another optional implementation, the location information of the simulated abnormal sound sample in the time-frequency domain can be set according to the duration of the simulated abnormal sound sample on the time axis and the frequency shift range on the frequency axis. A short-time Fourier transform can be performed on the simulated abnormal sound sample to generate a spectrogram, and then a rectangular detection area containing the abnormal sound signal can be selected on the spectrogram as the location information for annotation, or the spatiotemporal coordinates of the detection area can be recorded as the location information for annotation.

[0116] like Figure 3 As shown, this is typically represented by the coordinate parameters of a rectangular detection box that completely covers the energy distribution area of ​​the abnormal sound signal. For example, when the simulated abnormal sound sample starts at 2.5 seconds and ends at 3.8 seconds on the time axis, and fluctuates between 1200Hz and 1850Hz on the frequency axis due to the Doppler effect, the positional information in the time-frequency domain can be labeled as a quadruplet of data containing the start time of 2.5s, the end time of 3.8s, the lowest frequency of 1200Hz, and the highest frequency of 1850Hz, or the corresponding bounding box coordinates.

[0117] After generating simulated abnormal noise samples, this application further uses the positional information of the sample signal in the time-frequency domain as the core annotation content. This ensures that the generated training samples not only contain realistic acoustic features but also have precise spatiotemporal structure labels. Using these annotated samples with precise positional information for model training effectively solves the problem of poor model generalization ability caused by the scarcity of real fault samples. This enables the trained target detection model to accurately capture the sinusoidal frequency shift trajectory generated by blade rotation, thus improving the target detection model's accuracy in recognizing abnormal noise signals.

[0118] In another optional annotation method, the abnormal sound information annotated for the abnormal sound simulation sample may also include the position parameters of the sound source corresponding to the abnormal sound simulation signal on the wind turbine blade.

[0119] The positional parameter of the sound source on the wind turbine blade refers to quantitative data characterizing the specific distribution location of the abnormal noise source on the blade's geometry. This positional parameter originates from the physical model input configured when generating the abnormal noise simulation signal, and may specifically include the radial distance L from the sound source to the center of the wind turbine hub. This positional parameter serves as a ground truth label for training samples, used to establish the mapping relationship between time-frequency domain signal features and the physical spatial location of the sound source during model training, or to evaluate the computational accuracy of the sound source localization algorithm during model validation.

[0120] This embodiment expands the abnormal noise information labeled on the simulated noise samples to include the location parameters of the sound source on the wind turbine blades, forming a synergistic effect with the time-frequency domain location information defined in the previous embodiment. The time-frequency domain location information clarifies the trajectory of the abnormal noise signal in the spectrogram, solving the problem of detecting when and where the abnormal noise occurs; while the location parameters of the sound source on the wind turbine blades clarify the physical root cause of the abnormal noise, solving the problem of locating the source of the abnormal noise from the blades. The combination of the two allows a single simulated sample to simultaneously possess detection and recognition features and a ground truth label for localization. This not only enriches the information dimension of the training samples and improves the target detection model's ability to understand Doppler frequency shift trajectories, but also provides a direct verification benchmark for the sound source localization algorithm based on Doppler frequency shift relationships described later.

[0121] In some embodiments of this application, after obtaining simulated samples of abnormal sounds labeled with abnormal sound information, these samples can be used to train a target detection model. The target detection model is used to automatically detect abnormal sound sources from the spectrogram of the wind turbine operating audio.

[0122] Object detection models can refer to neural network architectures based on deep learning. Depending on the annotation content of the simulated noise samples, object detection models can employ various structures. For example, for binary classification object detection models, neural network models such as CNNs can be used. For object detection models that need to identify the trajectory corresponding to the noise signal caused by the Doppler frequency shift effect due to blade rotation in the spectrogram, YOLO models or other types of neural network models can be used.

[0123] Taking the latter model as an example, this model is configured to receive a two-dimensional time-frequency image (spectral graph) as input and output the detection box and confidence score of a specific target in the image. The target detection model is trained through end-to-end supervised learning using labeled abnormal sound simulation samples generated in the aforementioned steps. During training, the abnormal sound simulation samples are converted into spectrograms, and the labeled abnormal sound information is mapped to rectangular detection box labels in the spectrogram. The model continuously adjusts its internal weight parameters through a backpropagation algorithm to minimize the loss function between the predicted detection box and the true labeled box, thereby learning the specific periodic frequency shift trajectory characteristics of the abnormal sound signal in the time-frequency domain due to the Doppler effect.

[0124] This training method, which utilizes large-scale simulation data, effectively solves the problem of model overfitting caused by the scarcity of real fault samples, and significantly improves the generalization ability and recognition robustness of the target detection model for short-term abnormal sound signals in low signal-to-noise ratio environments.

[0125] Considering that in the field of wind turbine blade condition monitoring, abnormal noise signals generated by blade damage have short-term, non-stationary characteristics, and are subject to a typical Doppler frequency shift effect due to blade rotational motion, existing acoustic monitoring schemes mostly rely on convolutional neural networks for sound classification. These schemes can only achieve binary classification to determine the presence or absence of abnormal noise, making it difficult to capture the dynamic frequency shift trajectory of the noise signal. This results in insufficient robustness in identifying low signal-to-noise ratio, short-term abnormal noise signals, and the inability to provide quantitative references for the sound source location, hindering targeted maintenance and repair. Therefore, some embodiments of this application further provide a method for detecting abnormal noise in wind turbine blades, referring to... Figure 5 As shown, the method for detecting abnormal noise in wind turbine blades includes the following steps:

[0126] Step S200: Acquire the audio signal to be detected generated during the operation of the fan.

[0127] The audio signal to be detected can refer to the continuous sound wave signal captured by audio acquisition equipment deployed around the wind turbine during normal operation or fault condition. This signal originates from aerodynamic noise generated by the friction between the wind turbine blades and the air during rotation, mechanical noise generated by components such as gearboxes and generators, and abnormal friction noise that may be caused by cracks or edge damage on the blade surface.

[0128] Audio acquisition equipment is typically set to a high-sensitivity microphone, which can be installed at the bottom of the wind turbine tower or in a specific part of the nacelle to ensure that it can effectively receive sound source signals in the area where the blades rotate and sweep.

[0129] The audio signal to be detected obtained in this step can be acquired in real time by an audio acquisition device, or it can be the audio signal to be detected pre-stored in a database.

[0130] Step S210: Perform time-frequency transformation on the audio signal to be detected to generate a spectrogram.

[0131] Time-frequency transformation is the process of converting a one-dimensional time-domain audio signal into a two-dimensional time-frequency domain image. The generated spectrogram can intuitively show the distribution of signal energy as time and frequency change. This transformation process specifically includes framing the audio signal to be detected, windowing it, and performing a short-time Fourier transform (STFT) or wavelet transform on each frame to calculate the frequency components and their intensities at each time point.

[0132] In a spectrogram, the horizontal axis represents time, and the vertical axis represents frequency. The grayscale value or color intensity of a pixel represents the signal energy intensity at the corresponding time and frequency. Due to the periodic Doppler frequency shift characteristic of abnormal noise signals caused by the rotation of wind turbine blades, these signals appear as specific sinusoidal or quasi-sinusoidal bright band trajectories in the spectrogram, which significantly differs from the random distribution of background noise. This transformation converts the difficult-to-identify dynamic frequency shift patterns hidden in the time-domain waveform into geometric features in the image, providing a suitable input data structure for feature extraction in deep learning models.

[0133] Step S220: Input the spectrogram into the trained target detection model. The target detection model is used to detect the trajectory shape corresponding to the abnormal sound signal in the spectrogram caused by the Doppler frequency shift effect due to the blade rotation.

[0134] Among them, the trained object detection model is an image processing model built on a deep neural network. It can adopt the YOLO series model architecture and has end-to-end real-time detection capabilities.

[0135] During the training phase, the target detection model used a large number of spectrogram samples with annotated abnormal sound trajectories. These samples included both precisely annotated simulation data generated by the Doppler effect simulation in the aforementioned embodiments and a small amount of real fault data that had been manually verified. This enabled the model to learn the unique time-frequency trajectory shape of abnormal sound signals, rather than a simple spectral energy distribution.

[0136] The object detection model automatically scans the input spectrogram, identifies and locates abnormal sound regions that conform to Doppler frequency shift patterns, and outputs bounding boxes containing the locations of the abnormal sound signals. Specifically, the model can extract local texture and shape features from the spectrogram through convolutional layers and use a regression head to predict the coordinates of the bounding boxes and the class confidence. For example, when a bright band appears in the spectrogram whose frequency periodically increases and then decreases over time, the model identifies it as an abnormal sound trajectory caused by blade rotation and generates a rectangular box that tightly wraps around the trajectory.

[0137] This step uses target detection technology to replace traditional classification technology. It can not only identify whether there are abnormal sounds in the audio signal under test, but also initially solve the problem of the location of abnormal sounds in the time and frequency domain, significantly improving the detection rate of short-term and weak abnormal sound signals.

[0138] Step S230: Determine the final abnormal sound detection result based on the detection results output by the target detection model.

[0139] The detection results typically include the detection bounding box predicted by the model (such as the coordinates of the detection bounding box) and its corresponding confidence score. The detection bounding box covers the time-frequency trajectory of the abnormal sound signal.

[0140] This application can determine the presence or absence of abnormal noise based on a preset confidence threshold. If there is a target with a confidence level higher than the preset threshold (such as 0.6 or 0.7) in the detection box output by the model, it is determined that there is abnormal noise from the wind turbine blades in the current audio signal to be detected; conversely, if the confidence level of all detection boxes is lower than the threshold or no target box is detected, it is determined that there is no abnormal noise.

[0141] Furthermore, this step can be combined with the Non-Maximum Suppression (NMS) algorithm to remove redundant detection boxes with high overlap and low confidence, retaining the best detection result as the final conclusion.

[0142] This application's method transforms a one-dimensional audio signal into a two-dimensional spectrogram through time-frequency transformation, highlighting the Doppler frequency shift trajectory characteristics of the abnormal sound signal caused by blade rotation, thus solving the problem that traditional time-domain analysis struggles to capture dynamic frequency changes. Secondly, a specially trained target detection model is introduced, utilizing its keen ability to capture specific trajectory patterns in images to achieve automatic identification and localization of abnormal sound signals, overcoming the shortcomings of traditional classification models such as poor generalization ability and inability to distinguish between background noise and transient abnormal sounds. This scheme, which deeply integrates the Doppler effect with artificial intelligence algorithms, not only significantly improves the accuracy and robustness of abnormal sound detection, especially for short-duration abnormal sounds in low signal-to-noise ratio environments, but also provides accurate time-frequency boundary data for subsequent calculation of the specific location of the sound source based on frequency shift extrema.

[0143] In some embodiments of this application, the process of determining the final abnormal sound detection result in step S230 may further include:

[0144] When it is determined that there is an abnormal sound signal in the audio signal to be detected, the abnormal sound signal information is output as part of the abnormal sound detection result.

[0145] The abnormal sound signal information includes the start and end positions of the abnormal sound signal on the time axis, as well as the lowest and highest frequencies on the frequency axis.

[0146] This application achieves refined output of abnormal noise detection results through the above steps. Based on this, the system transforms the geometric features of the detection frame into specific time-frequency physical quantities, making the abstract model output interpretable and measurable. By extracting the start and end positions on the time axis, the system can accurately pinpoint the time period in which the abnormal noise occurred, facilitating maintenance personnel to quickly trace back the corresponding wind turbine operating conditions. Simultaneously, by combining the lowest and highest frequencies on the frequency axis, the spectral characteristics of the abnormal noise are fully described.

[0147] Furthermore, based on the highest and lowest output frequencies, the linear velocity of the sound source as the blades rotate can be calculated, and the specific distance between the sound source and the hub center can be deduced by combining the fan speed, thus accurately locating the abnormal sound source.

[0148] In some embodiments of this application, the process of determining the final abnormal noise detection result in step S230 may also include the process of calculating the position of the abnormal noise source on the wind turbine blades. That is, the final abnormal noise detection result may also include the sound source position parameters corresponding to the abnormal noise signal.

[0149] The process of calculating the sound source location parameters corresponding to the abnormal sound signal may include:

[0150] S1. Extract the highest and lowest frequencies of the abnormal sound signal from the target detection box.

[0151] The highest frequency can refer to the peak frequency of the abnormal sound signal on the frequency trajectory in the spectrogram due to the Doppler shift effect, and the lowest frequency can refer to the valley frequency on that trajectory. These two frequency characteristics directly reflect the extreme values ​​of the sound source's velocity relative to the audio acquisition device. Specifically, the received frequency reaches its highest when the sound source rotates with the fan blades to be closest to the audio acquisition device and its relative velocity is at its maximum towards the acquisition device; conversely, the received frequency reaches its lowest when the sound source rotates away from the audio acquisition device and its relative velocity is at its maximum away from the acquisition device.

[0152] By analyzing the detection box coordinate information output by the target detection model, the coverage range of the detection box on the frequency axis is read, and the maximum and minimum values ​​within this range are determined as the highest and lowest frequencies, respectively.

[0153] S2. Based on the highest and lowest frequencies, and combined with the Doppler frequency shift relationship, determine the linear velocity of the sound source corresponding to the abnormal sound signal as it rotates with the fan blades.

[0154] This step aims to use the extracted frequency extrema to inversely calculate the circular velocity of the sound source. Specifically, based on the Doppler effect principle, the highest frequency... With the lowest frequency These correspond to the extreme cases of the sound source moving towards and away from the audio acquisition device, respectively, and are related to the fundamental frequency of the sound source. Speed ​​of sound Furthermore, there is a definite mathematical relationship between the linear velocity V of the sound source rotating with the wind turbine blades. Specifically:

[0155] In conjunction with the preceding text Figure 4 As can be seen from the corresponding text description, when the sound source is located at point D, the linear velocity V of the sound source as the blade rotates is... D (t) is minimized, V D (t) = -V, corresponding to the highest instantaneous frequency, with a value of .

[0156] When the sound source is located at point B, the linear velocity V of the sound source as it rotates with the blade is... D(t) is at its maximum, V D (t)=V, corresponding to the lowest instantaneous frequency, taking a value of .

[0157] According to the Doppler frequency shift formula:

[0158]

[0159]

[0160] Based on the above formula, the following is derived:

[0161]

[0162] The linear velocity V of the sound source rotating with the fan blades can be calculated using the above formula.

[0163] S3. Based on the linear velocity of the sound source corresponding to the abnormal sound signal as it rotates with the fan blades, and the rotation period or angular velocity of the fan blades, calculate the distance from the center of the fan hub to the sound source corresponding to the abnormal sound signal, and use this distance as the position parameter of the sound source corresponding to the abnormal sound signal.

[0164] Combining the formula for the linear velocity of the sound source as it rotates with the blade:

[0165]

[0166] The derivation leads to the following conclusion:

[0167]

[0168] This embodiment eliminates the uncertainty of dependence on the original fundamental frequency of abnormal noise by co-operating the highest and lowest frequencies. It directly quantifies the linear velocity of the sound source as the blade rotates using the frequency shift amplitude, and further combines it with the linear velocity calculation formula to derive the distance between the sound source and the center of the wind turbine hub, thereby achieving precise localization of the sound source. In other words, this application not only solves the model generalization problem caused by the scarcity of real abnormal noise samples, but also overcomes the limitation of traditional acoustic monitoring schemes in accurately locating the source, providing a precise spatial coordinate reference for the early damage repair of wind turbine blades.

[0169] In some embodiments of this application, a wind turbine blade abnormal noise detection system is further provided, which includes an audio acquisition device and a data processing device.

[0170] The audio acquisition device is used to collect the audio signals to be detected generated during the operation of the fan.

[0171] The data processing device is used to execute each step of the wind turbine blade abnormal noise detection method described in any of the foregoing embodiments, thereby obtaining the final abnormal noise detection result.

[0172] The data processing device can refer to an electronic device with data computing and logic processing capabilities, such as an industrial control computer, embedded processor, server, or cloud computing node. The data processing device communicates with the audio acquisition device and receives the audio signal to be detected transmitted by the audio acquisition device. Upon receiving the audio signal, the program instructions running inside the data processing device are triggered, executing the various steps of the wind turbine blade abnormal noise detection method.

[0173] Specifically, the data processing equipment first performs a short-time Fourier transform on the input audio signal to be detected to generate a spectrogram. Then, the spectrogram is input into a pre-trained target detection model. This model can identify the trajectory pattern corresponding to the abnormal noise signal in the spectrogram caused by the Doppler frequency shift effect due to blade rotation, and outputs detection results including detection boxes and confidence scores. Finally, based on the detection results, the data processing equipment determines whether abnormal noise exists and can further calculate the distance from the sound source to the center of the wheel hub by combining the Doppler frequency shift relationship, thus completing the entire process from signal acquisition to fault location.

[0174] The hardware architecture of the data processing equipment can be flexibly configured according to system deployment requirements. It can be integrated into the local control cabinet of the wind turbine to realize edge computing, or it can transmit data to a remote data center for centralized processing via the network.

[0175] This application also provides a computer program product, including computer-readable instructions. When the computer-readable instructions are run on an electronic device, the electronic device enables the electronic device to implement any of the wind turbine blade abnormal noise annotation sample generation methods or any of the wind turbine blade abnormal noise detection methods provided in this application.

[0176] This application also provides a computer-readable storage medium carrying one or more computer programs. When the one or more computer programs are executed by an electronic device, the electronic device can implement any of the wind turbine blade abnormal noise annotation sample generation methods or any of the wind turbine blade abnormal noise detection methods provided in this application.

[0177] It should also be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. In addition, in the device embodiment drawings provided in this application, the connection relationship between modules indicates that they have a communication connection, which can be implemented as one or more communication buses or signal lines.

[0178] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented by means of software plus necessary general-purpose hardware, or it can be implemented by special-purpose hardware including application-specific integrated circuits, special-purpose CPUs, special-purpose memory, special-purpose components, etc. Generally, any function performed by a computer program can be easily implemented by corresponding hardware, and the specific hardware structure used to implement the same function can also be diverse, such as analog circuits, digital circuits, or special-purpose circuits. However, for this application, software program implementation is more often the preferred implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a readable storage medium, such as a computer floppy disk, USB flash drive, mobile hard disk, ROM, RAM, magnetic disk, or optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, training equipment, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0179] In the above embodiments, implementation can be achieved, in whole or in part, through software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product.

[0180] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions may be transmitted from one website, computer, training device, or data center to another website, computer, training device, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that a computer can store or a data storage device such as a training device or data center that integrates one or more available media. The available media may be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., DVDs), or semiconductor media (e.g., solid-state drives (SSDs)).

[0181] The various embodiments in this specification are described in a progressive manner. Each embodiment focuses on the differences from other embodiments. The various embodiments can be combined as needed, and the same or similar parts can be referred to each other.

Claims

1. A method for generating a sample of abnormal noise annotations for wind turbine blades, characterized in that, include: Configure the motion parameters of the sound source corresponding to the abnormal noise as it rotates with the fan blades; Based on the motion parameters, the Doppler frequency shift effect generated by the sound source relative to the audio acquisition device during the rotation process is simulated to obtain the instantaneous frequency change sequence of the sound source during the rotation process; Based on the instantaneous frequency change sequence, generate an abnormal sound simulation signal with frequency shift characteristics; The abnormal noise simulation signal is superimposed with the background noise during the operation of the wind turbine to obtain an abnormal noise simulation sample, and the abnormal noise information is labeled.

2. The method according to claim 1, characterized in that, The motion parameters include the angular velocity or rotation period of the wind turbine blades, and the distance between the sound source and the center of the wind turbine hub.

3. The method according to claim 1, characterized in that, Based on the motion parameters, the process of simulating the Doppler frequency shift effect of the sound source relative to the audio acquisition device during rotation, and obtaining the instantaneous frequency change sequence of the sound source during rotation, includes: Based on the motion parameters, calculate the instantaneous relative velocity of the sound source with respect to the audio acquisition device during the rotation process; Based on the instantaneous relative computing speed and the configured heterophonic fundamental frequency, the instantaneous frequency at each moment is calculated using the Doppler frequency shift formula to obtain the instantaneous frequency change sequence.

4. The method according to claim 3, characterized in that, The process of calculating the instantaneous relative velocity of the sound source with respect to the audio acquisition device during rotation, based on the motion parameters, includes: The linear velocity of the sound source as the blade rotates is determined based on the motion parameters. Based on the relative position between the current rotational position of the sound source and the audio acquisition device, the velocity component of the linear velocity in the direction of the line connecting the sound source and the audio acquisition device is determined as the instantaneous relative motion velocity.

5. The method according to claim 4, characterized in that, The motion parameters include the angular velocity or rotation period of the wind turbine blades, and the distance between the sound source and the center of the wind turbine hub. The process of determining the linear velocity of the sound source as it rotates with the blade based on the motion parameters includes: The linear velocity of the sound source as it rotates with the blades is determined by multiplying the distance of the sound source from the center of the wind turbine hub with the rotational angular velocity of the wind turbine blades. or, The rotational angular velocity of the wind turbine blades is calculated based on the rotation period, and the linear velocity of the sound source as the blades rotate is determined by multiplying the distance of the sound source from the center of the wind turbine hub with the calculated rotational angular velocity of the wind turbine blades.

6. The method according to claim 1, characterized in that, The abnormal sound information labeled for the abnormal sound simulation samples includes: The location information of the abnormal sound simulation sample in the time-frequency domain.

7. The method according to claim 6, characterized in that, The abnormal sound information annotated for the abnormal sound simulation samples also includes: The position parameters of the sound source corresponding to the simulated abnormal sound signal on the wind turbine blade.

8. The method according to claim 6, characterized in that, The location information in the time-frequency domain includes: The start and end positions of the simulated noise sample on the time axis, and the lowest and highest frequencies on the frequency axis.

9. The method according to any one of claims 1-8, characterized in that, Also includes: The simulated samples of abnormal sounds labeled with abnormal sound information are used to train a target detection model, which is used to automatically detect abnormal sound sources from the spectrogram of the wind turbine operating audio.

10. A method for detecting abnormal noise in wind turbine blades, characterized in that, include: Acquire the audio signal to be detected generated during the operation of the wind turbine; The audio signal to be detected is subjected to time-frequency transformation to generate a spectrogram; The spectrogram is input into a trained target detection model, which is used to detect the trajectory shape of the abnormal sound signal in the spectrogram caused by the Doppler frequency shift effect due to the blade rotation. Based on the detection results output by the target detection model, the final abnormal sound detection result is determined.

11. The method according to claim 10, characterized in that, The detection results output by the target detection model include the detection boxes corresponding to the heterophonic signals in the spectrogram and the confidence scores of the detection boxes, wherein the detection boxes cover the time-frequency trajectory of the heterophonic signals; The process of determining the final abnormal sound detection result includes: If a target detection box with a confidence level higher than a preset threshold exists, it is determined that there is an abnormal sound signal in the audio signal to be detected; If there is no detection box with a confidence level higher than the preset threshold, it is determined that there is no abnormal sound signal in the audio signal to be detected.

12. The method according to claim 11, characterized in that, The process of determining the final abnormal sound detection result also includes: When it is determined that there is an abnormal sound signal in the audio signal to be detected, the abnormal sound signal information is output; The abnormal sound signal information includes the start and end positions of the abnormal sound signal on the time axis, and the lowest and highest frequencies on the frequency axis.

13. The method according to claim 11, characterized in that, The process of determining the final abnormal sound detection result also includes: Extract the highest and lowest frequencies of the abnormal sound signal from the target detection box; Based on the highest and lowest frequencies, and combined with the Doppler frequency shift relationship, the linear velocity of the sound source corresponding to the abnormal sound signal as the fan blades rotate is determined. Based on the linear velocity of the sound source corresponding to the abnormal noise signal as it rotates with the wind turbine blades, and the rotation period or angular velocity of the wind turbine blades, the distance from the sound source corresponding to the abnormal noise signal to the center of the wind turbine hub is calculated and used as the location parameter of the sound source corresponding to the abnormal noise signal.

14. A wind turbine blade abnormal noise detection system, characterized in that, include: Audio acquisition equipment and data processing equipment; The audio acquisition device is used to acquire the audio signal to be detected generated during the operation of the fan; The data processing device is used to execute each step of the wind turbine blade abnormal noise detection method according to any one of claims 10 to 13.

Images