A method and system for detecting damage to wind turbine blades based on deep learning

Abstract

The present invention discloses a method and system for detecting damage to wind turbine blades based on deep learning, including collecting blade data and performing preprocessing; updating model parameters using an optimization algorithm; inputting the data into a deep learning prediction model to adaptively extract data features; calculating the true value and predicted value of the health index; calculating the root mean square error and determining the condition of the blade; calculating the entropy value and energy value of the data and determining blade anomalies; inputting the data into a deep learning diagnosis model to adaptively extract fault features and labels; matching the fault labels with a fault database to obtain the blade diagnosis result and outputting it to the wind farm control center, and simultaneously updating the fault database. By adaptively extracting data signal features, the present invention accurately and timely monitors early damage to the blades according to the health index, and detects the type, location and degree of blade damage, without relying on artificial prior experience. By updating the model parameters using an optimization algorithm, the method has higher detection accuracy and wider adaptability.

Description

Technical Field

[0001] The present invention relates to the field of blade damage detection, and in particular to a method and system for detecting wind turbine blade damage based on deep learning. Background Art

[0002] The blade is one of the key components of a wind turbine. Blade damage directly affects the life of the blade and the reliability of the stable operation of the wind turbine. At present, blade defect detection usually involves setting acceleration sensors on the blade and judging whether there is damage to the blade according to the stress test analysis results. However, this method can only judge whether there is damage on the blade and cannot accurately obtain the damage location, resulting in low detection accuracy of this type of method; Fiber Bragg grating detection requires embedding Fiber Bragg grating sensors inside the material during blade manufacturing, which is difficult to manufacture, and the sensors may fail or be damaged after long-term operation, making it difficult to maintain; Ultrasonic testing is a non-destructive testing method that uses the acoustic performance differences of materials and their defects to detect internal defects of the blade through the reflection of ultrasonic propagation waveforms and the energy change of the penetration time. Affected by the subjective judgment of the tester, its detection cycle is long and it is more suitable for static monitoring before leaving the factory; Infrared imaging detection has low detection sensitivity for positions with deeper pre-damage and is greatly affected by environmental factors, making it difficult to perform real-time detection of the blade. In the prior art, there are also blade damage detection methods based on data-driven, which extract the time domain, frequency domain, and time-frequency domain features of the blade vibration signal, and judge whether there is damage to the blade according to the features. This method relies on manual inspection experience, has low diagnostic efficiency and accuracy, and has poor adaptability for detecting wind turbines in different wind farms.

[0003] For example, a "Device and Method for Synchronously Detecting Wind Turbine Blade Damage" disclosed in a Chinese patent document, with the publication number CN108386323A, the device includes a hoisting mechanism, a propulsion mechanism, a synchronization mechanism, a moving mechanism, a circular ring acquisition mechanism, and an information processing unit. The synchronization mechanism includes a synchronization mechanism base body and an intelligent locking device. The circular ring acquisition mechanism includes an infrared imager and an infrared flash excitation source adjustment system. The method includes: the hoisting mechanism lifts and positions the device, the pushing mechanism pushes the detection device to be synchronized with the wind turbine, the moving mechanism transports the circular ring acquisition mechanism to the detection position, the adjustment system adjusts the height, angle, and frequency of the infrared flash, the infrared imager collects image data and sends it to the information processing unit, the information processing unit constructs an infrared three-dimensional thermal image of the blade for analysis and judgment, the device resets, and the detection ends. Although this invention realizes online synchronous detection of wind turbine blade damage and analyzes and judges the damage through the infrared three-dimensional thermal image of the wind turbine blade, shortening the detection time, there are problems such as difficult installation and maintenance of the detection equipment, poor stability, low detection accuracy and sensitivity, and being greatly affected by environmental factors. Summary of the Invention

[0004] The present invention aims to overcome the problems of poor stability, low accuracy, and dependence on manual prior experience in the damage detection of wind turbine blades in the prior art, and provides a method and system for wind turbine blade damage detection based on deep learning.

[0005] To achieve the above object, the present invention adopts the following technical solutions:

[0006] A method for wind turbine blade damage detection based on deep learning includes the following steps: S1: Real-time collect the original audio and vibration data of the blade and perform preprocessing, and divide the preprocessed data into a training sample set X, a validation sample set Y, and a test sample set Z; S2: Update the parameters of the deep learning model and the machine learning prediction model using an optimization algorithm. The deep learning model includes a deep learning prediction model and a deep learning diagnosis model; S3: Input the data of the training sample set X in S1 into the deep learning prediction model to adaptively extract data features; S4: Calculate the true value of the health index according to the data features extracted in S3, and input the true value data of the health index into the machine learning prediction model to obtain the predicted value of the health index; S5: Calculate the root mean square error according to the true value and the predicted value of the health index, and compare the root mean square error with a preset error threshold. When the root mean square error is less than the preset error threshold, it is determined that the blade is normal and this detection ends; when the root mean square error is greater than the preset error threshold, go to S6; S6: Calculate the entropy value and energy value of the data, and compare the entropy value and energy value with a preset entropy threshold and energy threshold. When the entropy value is greater than the entropy threshold and the energy value is greater than the energy threshold, it is determined that the blade is abnormal and go to S7; otherwise, it is determined that the blade is normal and this detection ends; S7: Input the preprocessed data in S1 into the deep learning diagnosis model to adaptively extract fault features and fault labels; S8: Match the fault labels with the blade fault database to obtain the damage location, damage type, and damage degree of the blade, and output the diagnosis result to the wind farm control center and synchronously update the blade fault database. Compared with the prior art, the present invention can accurately and timely monitor the early damage of the blade according to the health index by adaptively extracting the features of audio and vibration signals, detect the damage type, location, and degree of the blade, without relying on manual prior experience, and update the parameters of the deep learning model by using an optimization algorithm, making the detection accuracy of this method higher and the adaptability wider.

[0007] As a preferred solution of the present invention, the output of the machine learning prediction model in S4 is: β = H T (I / C + HH T ) -1 y, define the kernel matrix Ω ELM = HH T , the matrix element Ω ELM (i, j) = h(x i)h(x j ) = K(x i , x j ), then the output of the machine learning prediction model is where the kernel function K(x i , y i ) = exp(-γ||x i , x j || 2 ), where h(x) is the output matrix of the hidden layer, β represents the connection weight between the hidden layer and the output layer, C is the penalty coefficient, and γ is the kernel parameter.

[0008] As a preferred solution of the present invention, the calculation formula for the true value of the S4 health index is:

[0009]

[0010] In the formula, is the real-time data feature, y t is the damage data feature, and N is the sequence length.

[0011] As a preferred solution of the present invention, the calculation formula for the root mean square error in S5 is: where N is the number of test samples, is the predicted value of the health index, and y i is the true value of the health index.

[0012] As a preferred solution of the present invention, the calculation formula for the entropy value in S6 is:

[0013]

[0014] where m is the embedding dimension; r is the similarity tolerance; E SE (y (s) , m, r) is the sample entropy; is the number of m and m+1 dimensional space vectors of the coarse-grained sequence.

[0015] As a preferred solution of the present invention, the calculation formula for the energy value in S6 is:

[0016]

[0017] where E(j,i) represents the energy value of the i-th node on the j-th layer; p s (n,j,k) is the wavelet packet coefficient.

[0018] As a preferred solution of the present invention, the role of the optimization algorithm in S2 is stage-based and is implemented according to the SCADA wind speed parameters and noise levels.

[0019] A wind turbine blade damage detection system based on deep learning, comprising a signal acquisition module for acquiring blade audio and vibration signal data; a working condition division module for setting reference thresholds based on SCADA wind speed parameters and noise levels; a preprocessing module for processing irrelevant information such as noise and environmental interference in audio and vibration signals; a deep learning feature extraction module for adaptively extracting high-dimensional features of audio and vibration signals; a deep learning prediction module for outputting real-time monitoring results of the blade; a deep learning diagnosis module for adaptively diagnosing the damage location, type, and degree of the blade; and a display and transmission module for displaying the diagnosis results and transmitting them to the wind farm control center.

[0020] As a preferred embodiment of the present invention, the system further includes a central processor for realizing the operation and control functions of the entire system; and a memory for storing program instructions for the processor to execute the method for early damage detection of wind turbine generator blades and related data generated during the execution of the program instructions.

[0021] Therefore, the present invention has the following beneficial effects: By adaptively extracting the features of audio and vibration signals, the present invention accurately and timely monitors the early damage of the blade according to health indicators, and detects the damage type, location, and degree of the blade, without relying on artificial prior experience. By optimizing the algorithm to update the parameters of the deep learning model, the detection accuracy of this method is higher and the adaptability is wider. Description of the Drawings

[0022] Figure 1 is a flowchart of the method for detecting damage to wind turbine blades of the present invention;

[0023] Figure 2 is a flowchart of blade damage monitoring in an embodiment of the present invention;

[0024] Figure 3 is a flowchart of blade damage diagnosis in an embodiment of the present invention;

[0025] Figure 4 is a schematic diagram of the structure of the deep learning prediction model of the present invention;

[0026] Figure 5 is a schematic diagram of the structure of the deep learning diagnosis model of the present invention;

[0027] Figure 6 is a waveform diagram of the original audio and vibration data in an embodiment of the present invention;

[0028] Figure 7 is a waveform diagram of the data after band-pass filtering to remove noise interference in an embodiment of the present invention;

[0029] Figure 8 is a distribution diagram of part of the data with abnormal health indicators in the test sample set in an embodiment of the present invention;

[0030] Figure 9 is the multi-scale sample entropy data graph of the embodiment of the present invention;

[0031] Figure 10 is the wavelet energy data graph of the embodiment of the present invention;

[0032] Figure 11 is the schematic structural diagram of the system of the present invention;

[0033] Figure 12 is the schematic diagram of the installation position of the system device of the present invention on the wind turbine;

[0034] In the figure: 1. First acceleration sensor; 2. Second acceleration sensor; 3. Third acceleration sensor; 4. Pickup; 5. Central processing unit. Specific embodiments

[0035] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments.

[0036] As Figure 1 shown, a method for detecting blade damage of a wind turbine based on deep learning includes blade damage monitoring and blade damage diagnosis. As Figure 2 the blade damage monitoring flowchart of the embodiment of the present invention, Figure 3 is the blade damage diagnosis flowchart of the embodiment of the present invention, including:

[0037] S1: Real-time obtain the original audio and vibration signal data of the wind turbine blade, and preprocess the data. Divide the preprocessed data into a training sample set X, a verification sample set Y, and a test sample set Z. Specifically, the preprocessing process includes filter noise reduction and signal decomposition and reconstruction. Use a Butterworth band-pass filter to filter the original audio and vibration signal data of the wind turbine blade to remove environmental noise, bird calls, and mechanical vibration noise. The start and stop frequency bands of the Butterworth band-pass filter are set to 800 Hz to 10 KHz. Then, perform wavelet packet decomposition on the filtered audio and vibration data. The wavelet basis function is selected as the Morlet wavelet, and its formula is:

[0038]

[0039] where C is the normalization constant during reconstruction, e is the irrational number, and t is the time series.

[0040] S2: Optimize the deep learning model and the machine learning prediction model using an optimization algorithm, and update the parameters of the deep learning model and the machine learning prediction model. Among them, the deep learning model includes a deep learning prediction model and a deep learning diagnosis model. Specifically, the optimization algorithm uses the simulated annealing particle swarm algorithm, sets the acceleration factors c1 = c2 = 1.5, the number of particle swarms N = 20, the inertia weight ω start = 0.9, ω end = 0.4, the parameter optimization interval (C, γ) = [0.01, 200], the maximum number of iterations is 100, the simulated annealing parameter λ = 0.7, the initialization of the deep learning prediction model parameters is ω = rand(H, L), b = zeros(H, 1), where ω is the weight matrix, b is the bias vector, H represents the number of neurons in the hidden layer, and L represents the number of neurons in the input layer of the DBN. The role of the optimization algorithm is phased and is implemented according to the SCADA wind speed parameters and the noise level. Among them, the training set data input to the deep learning prediction model is more than the test set data.

[0041] The specific process of optimizing by the optimization algorithm is as follows: Initialize the parameters of the optimization algorithm, the deep learning model, and the machine learning prediction model. Normalize the training sample set X using the matlab function mapminmax, and the normalization interval is [0, 1]. According to the validation sample set Y and the fitness function formula

[0042]

[0043] for verification, where N is the number of samples in the validation sample set Y, is the predicted value of the health index, y i is the true value of the health index. If the RMSE reaches the optimal, import the optimized parameters into the deep learning model and the machine learning prediction model for optimization; otherwise, no optimization will be performed.

[0044] S3: Use the training sample set X data as the input of the deep learning prediction model, and adaptively extract data features. Among them, the activation function of the deep learning prediction model uses sigmoid, and its expression is: The deep learning prediction model is only used to unsupervised extract the features of audio and vibration data and does not include supervised backpropagation fine-tuning.

[0045] The structure of the deep learning prediction model is as Figure 4As shown, the training sample set X is used as the input for deep learning prediction to adaptively and unsupervised extract high-dimensional features. The obtained high-dimensional features are input into the machine learning prediction model. The root mean square error is calculated between the true value and the predicted value of the health index, and the number of neurons, the number of network layers, the random weight values, and the bias parameters of each layer of the deep learning prediction model are determined. The accuracy and reliability of the model are verified through the validation sample set Y.

[0046] S4: Calculate the true value of the health index according to the adaptively extracted data features and formulas, and use the true value of the health index as the input of the machine learning prediction model to obtain the predicted value of the health index; specifically, the output of the machine learning prediction model is: β = H T (I / C + HH T ) -1 y, define the kernel matrix Ω ELM = HH T , the matrix element Ω ELM (i,j) = h(x i )h(x j ) = K(x i ,x j ), then the output of the machine learning prediction model is

[0047]

[0048] where the kernel function K(x i ,y i ) = exp(-γ||x i ,x j || 2 ), h(x) is the output matrix of the hidden layer, β represents the connection weight between the hidden layer and the output layer, C is the penalty coefficient, and γ is the kernel parameter.

[0049] The formula for calculating the true value of the health index is:

[0050]

[0051] where is the real-time data feature, y t is the damage data feature, and N is the sequence length.

[0052] S5: Obtain the true value and the predicted value of the health index, calculate the root mean square error according to the root mean square error (RMSE residual) calculation formula, and compare the root mean square error with the error threshold. If the root mean square error is less than the error threshold, it is determined that the blade is normal and this detection ends; if the root mean square error is greater than the preset error threshold, proceed to the next step; specifically, the preset error threshold R is 0.8, and the root mean square error calculation formula is:

[0053]

[0054] Among them, N is the number of test samples, is the predicted value of the health index, y i is the true value of the health index.

[0055] S6: Calculate the entropy value and energy value of the training sample set X data, and compare the entropy value and energy value with the pre-set entropy threshold and energy threshold. If the entropy value is greater than the entropy threshold and the energy value is greater than the energy threshold, it is determined that the blade is abnormal and proceed to the next step; otherwise, it is determined that the blade is normal and this detection ends; specifically, the output of the multi-scale sample entropy, that is, the entropy value calculation formula is:

[0056]

[0057] Among them, m is the embedding dimension, r is the similarity tolerance, E SE (y (s) , m, r) is the sample entropy, is the number of m and m+1 dimensional space vectors of the coarse-grained sequence.

[0058] The wavelet packet energy formula is:

[0059]

[0060] Among them, E(j,i) represents the energy value of the i-th node at the j-th level; p s (n,j,k) is the wavelet packet coefficient.

[0061] In this embodiment, the detection result is as Figures 6 - 10 shown, and the blade abnormality can be detected in time through the multi-scale sample entropy and energy ratio. Figure 6 is the original audio and vibration data collected; Figure 7 is the data after removing noise interference through band-pass filtering, and the preprocessed data is divided into a training set, a validation set, and a test set, and normalized processing is performed; Figure 8 is part of the data with abnormal health indicators in the test sample set. At this time, it is found that there is data with an RMSE value greater than the threshold of 0.8; Figure 9 and Figure 10 Extract the multi-scale entropy and wavelet energy from the currently abnormal data respectively, and it is found that their multi-scale entropy and wavelet energy increase abnormally in the subsequent time period, and it is determined that the blade is damaged.

[0062] S7: Use the preprocessed data after normalization (including the training sample set X, the validation sample set Y, and the test sample set Z) as the input of the deep learning diagnosis model, and adaptively extract fault features and fault labels. Specifically, for the visible layer units v = {v1, v2, v3, …… v i} and the hidden layer units h = {h1, h2, h3, …… h i} under the conditions of the weight matrix ω, the threshold α of the visible layer units, and the threshold b of the hidden layer units, let θ = {ω, α, b}. For all j, v i ∈(0, 1) and h i ∈(0, 1). The energy function of the original deep learning diagnosis model, the restricted Boltzmann machine (RBM), is:

[0063]

[0064] Adopt the continuous restricted Boltzmann machine (CRBM) with Gaussian distribution. The CRBM uses the CD algorithm to adjust its parameters. At this time, the energy function is:

[0065]

[0066] where σ i is the standard deviation vector of the visible nodes with Gaussian noise added.

[0067] Use the sparse regularization unsupervised pre-training model, defined as f = f C + λf S , where f c represents the maximum likelihood function of the CRBM, λ is the regularization coefficient, and f s represents the sparse regularization function;

[0068] The output function of the improved deep learning diagnosis model is:

[0069]

[0070] where m is the number of neurons, P(v i ) is the independent distribution function, λ is the regularization coefficient, and f s represents the sparse regularization function.

[0071] By using the improved CRBM, that is, the deep learning diagnosis model, and introducing the regularization coefficient, the training / diagnosis speed of the model can be accelerated, the loss during the training process can be reduced, and the diagnosis accuracy and robustness of blade damage faults can be improved.

[0072] Specifically, the training process of its diagnosis model is as follows Figure 5As shown in the figure, the normalized preprocessed data is imported into the deep learning diagnosis model. The deep learning diagnosis model adopts four hidden layers, and the top classifier uses KELM to supervise the extraction of high-dimensional features of audio and vibration signals. The forward training uses the CD algorithm, and the reverse fine-tuning uses the BPNN algorithm. Its fault label is composed of the extracted health indicators. The extracted high-dimensional features are input into the KELM classification model, and the feature label of the fault sample is output through the classifier and compared with the blade damage database to obtain the blade damage recognition result.

[0073] S8: Match the fault label with the blade fault database to obtain the damage location, damage type, and damage degree of the blade, and output the diagnosis result to the wind farm control center and synchronously update the blade fault database.

[0074] As Figure 11 shown, the present invention preferably includes a wind turbine blade damage detection system based on deep learning, including a signal acquisition module. The signal acquisition module includes a microphone and an acceleration sensor for collecting blade audio and vibration signal data; a working condition division module, which includes a speed division module and a noise level division module, and sets reference thresholds according to the SCADA wind speed parameters and noise levels; a preprocessing module for processing irrelevant information such as noise and environmental interference of audio and vibration signals, determining the input parameters of the deep learning prediction module and the deep learning diagnosis module, and the output parameter health indicator is the prediction variable; a deep learning feature extraction module for adaptively extracting high-dimensional features of audio and vibration signals; a deep learning prediction module for outputting the real-time monitoring result of the blade; further including a machine learning prediction module for obtaining the predicted value of the health indicator; a deep learning diagnosis module for adaptively diagnosing the damage location, type, and degree of the blade; a display module for displaying the diagnosis result; a transmission module for transmitting the diagnosis result to the wind farm control center; further including a central processor to realize the operation and control functions of the whole system; a memory for storing the program instructions for the processor to execute the method for early damage detection of wind turbine generator blades and the relevant data generated during the execution of the program instructions; the memory also contains a fault sample database for matching with the fault label.

[0075] Among them, the establishment of the deep learning prediction module includes: determining the training sample set X, validation sample set Y, and test sample set Z after data preprocessing; using the training sample set X as the input of the deep learning prediction model, adaptively and unsupervised extracting high-dimensional features, inputting the obtained features into the machine learning prediction model, and determining the number of neurons, network layers, random weights, and bias parameters of each layer of the deep learning prediction model according to the error of predicting health indicators by the machine learning prediction. The establishment of the deep learning diagnosis module includes: migrating the number of neurons, learning rate, random weights, and bias parameters of the deep learning prediction model to the deep learning diagnosis model, using the processed data (including the training sample set X, validation sample set Y, and test sample set Z) as the input of the deep learning diagnosis model, using the fault label as the output data, adaptively extracting high-dimensional features, and comparing the output fault label of the deep learning diagnosis model with the blade fault database to obtain the blade damage type, location, and degree.

[0076] As Figure 12 shown, according to the installation positions and functions realized by the main devices of the wind turbine blade early damage detection system according to the exemplary embodiments of the present invention, the central processor 5 is installed inside the nacelle and connected to the wind farm computing center, the microphone 4 is installed at a position 2 m from the root of the tower, and the first acceleration sensor 1, the second acceleration sensor 2, and the third acceleration sensor 3 are respectively installed at a position 1 / 3 from the root of the blade to collect audio and vibration data in real time.

[0077] Turn on the system control switch, collect the blade audio and vibration data in real time, adjust the RMSE threshold according to the wind speed division module and the noise level division module. When it is detected that the data RMSE is greater than the set threshold, record the operating data of the blade in the next 10 minutes as the test sample set Z, calculate the multi-scale sample entropy and wavelet packet energy through the abnormal index detection module. If no abnormality is found, release the memory and continue to monitor the blade operating state; when it exceeds the set threshold, start the deep learning diagnosis module of the central processor, display the diagnosis result through the display module, and send the diagnosis result to the wind farm control center and synchronously update the fault sample database.

[0078] According to the exemplary embodiments of the present invention, a computing device is also provided. The computing device includes a processor and a memory. The memory is used to store a computer program. When the computer program is executed by the processor, the processor executes the computer program of the above-mentioned wind turbine blade early damage detection method.

[0079] An exemplary embodiment according to the present invention further provides a readable storage medium storing a computer program. The computer-readable storage medium stores a computer program that, when executed by a processor, causes the processor to execute the above-described method for early damage detection of a wind turbine blade. Examples of the computer-readable storage medium include: read-only memory, random access memory, compact disc read-only memory, magnetic tape, floppy disk, optical data storage device, and carrier wave (such as data transmission via the Internet through a wired or wireless transmission path).

[0080] By using the method and system for early damage detection of a wind turbine blade according to the exemplary embodiment of the present invention, it is possible to detect early abnormalities of the blade in real time and accurately, and detect the damage position, damage type, and damage degree of the blade, improving the accuracy and stability of early damage detection of the blade, thereby effectively reducing the deployment management and operation and maintenance costs.

[0081] As described above, the above are only specific embodiments of the present invention, but the protection scope of the present invention is not limited thereto. Any changes or substitutions that can be thought of without creative work shall be covered within the protection scope of the present invention.

Claims

1. A method for detecting damage to wind turbine blades based on deep learning, characterized in that, It includes the following steps: S1: Collect the original audio and vibration data of the blade in real time, preprocess them, and divide the preprocessed data into a training sample set X, a validation sample set Y, and a test sample set Z; S2: Update the parameters of the deep learning model and the machine learning prediction model using an optimization algorithm. The deep learning model includes a deep learning prediction model and a deep learning diagnosis model; S3: Input the data of the training sample set X in S1 into the deep learning prediction model to adaptively extract data features; S4: Calculate the true value of the health index based on the data features extracted in S3, input the true value data of the health index into the machine learning prediction model, and obtain the predicted value of the health index; S5: Calculate the root mean square error based on the true value and the predicted value of the health index, compare the root mean square error with a preset error threshold. When the root mean square error is less than the preset error threshold, it is determined that the blade is normal and this detection ends; when the root mean square error is greater than the preset error threshold, go to S6; S6: Calculate the entropy value and energy value of the data, and compare the entropy value and energy value with a preset entropy threshold and energy threshold. When the entropy value is greater than the entropy threshold and the energy value is greater than the energy threshold, it is determined that the blade is abnormal and go to S7; Otherwise, it is determined that the blade is normal and this detection ends; S7: Input the preprocessed data in S1 into the deep learning diagnosis model to adaptively extract fault features and fault labels; S8: Match the fault labels with the blade fault database to obtain the damage location, damage type, and damage degree of the blade, and output the diagnosis result to the wind farm control center and synchronously update the blade fault database.

2. The method for detecting damage to the blades of a wind turbine based on deep learning according to claim 1, characterized in that, The output of the machine learning prediction model in S4 is as follows: β = H T (I / C + HH T ) -1 y, define the kernel matrix Ω ELM = HH T , the matrix element Ω ELM (i, j) = h(x i )h(x j ) = K(x i , x j ), then the output of the machine learning prediction model is where the kernel function K(x i , y i ) = exp(-γ||x i , x j || 2 ), where h(x) is the output matrix of the hidden layer, β represents the connection weight between the hidden layer and the output layer, C is the penalty coefficient, and γ is the kernel parameter.

3. The method for detecting damage to the blades of a wind turbine based on deep learning according to claim 1, wherein The calculation formula for the true value of the health index in S4 is: In the formula, is the real-time data feature, y t is the damage data feature, and N is the sequence length.

4. A method for detecting damage to a wind turbine blade based on deep learning according to claim 1 or 3, characterized in that, The root mean square error calculation formula in S5 is as follows: where N is the number of test samples, is the predicted value of the health indicator, and y i is the true value of the health indicator.

5. A method for detecting damage to a wind turbine blade based on deep learning according to claim 1, characterized in that The calculation formula for the entropy value in S6 is: where m is the embedding dimension, r is the similarity tolerance, and E SE (y (s) , m, r) is the sample entropy, is the number of m- and (m+1)-dimensional space vectors of the coarse-grained sequence.

6. A method for detecting damage to a wind turbine blade based on deep learning according to claim 1, characterized in that, The calculation formula for the energy value in S6 is: Among them, E(j, i) represents the energy value of the i-th node at the j-th hierarchy; p s (n, j, k) are wavelet packet coefficients.

7. A method for detecting damage to a wind turbine blade based on deep learning according to claim 1, characterized in that The role of the optimization algorithm in S2 is phased and is implemented based on the SCADA wind speed parameter and noise level.

8. A wind turbine blade damage detection system applicable to the wind turbine blade damage detection method based on deep learning described in claim 1, characterized in that, It includes: A signal acquisition module for collecting blade audio and vibration signal data; A working condition division module for setting reference thresholds based on the SCADA wind speed parameter and noise level; A preprocessing module for processing irrelevant information such as noise and environmental interference of audio and vibration signals; A deep learning feature extraction module for adaptively extracting high-dimensional features of audio and vibration signals; A machine learning prediction module for outputting real-time monitoring results of the blade; A deep learning diagnosis module for adaptively diagnosing the damage location, type, and degree of the blade; A display and transmission module for displaying the diagnosis result and transmitting it to the wind farm control center.

9. The a damage detection system for a wind turbine blade based on deep learning according to claim 8, characterized in that it further It includes: A central processing unit to realize the operation and control functions of the entire system; A memory for storing program instructions for the processor to execute the method for early damage detection of wind turbine blades and related data generated during the execution of the program instructions.

Images