A non-contact method for monitoring the icing status of wind turbine blades

Abstract

A non-contact method for monitoring the icing state of wind turbine blades includes: acquiring wind turbine images; conducting noise tests on the wind turbine rotor composed of iced blades to obtain the wind turbine noise signal under icing conditions; processing the wind turbine noise signal using an eigenvalue-optimized beamforming algorithm to obtain sound pressure level spectral characteristics; acquiring noise sources based on the wind turbine images and identifying the coordinates of the maximum noise source, establishing the relationship between the noise source and different icing states in non-icing locations; constructing a positive relationship between icing states at different locations, with different masses, and with different incoming flow angles of attack, and the increase in the total sound pressure level of the wind turbine; inputting the positive relationship into a BP neural network model and a local linear wavelet neural network model respectively to obtain monitoring results. This invention utilizes acoustic array technology to achieve non-contact measurement of horizontal axis iced wind turbine noise and thus monitor its icing state, realizing the monitoring of wind turbine blade noise under different icing states.

Description

Technical Field

[0001] This invention belongs to the field of wind turbine fault diagnosis, specifically relating to a non-contact method for monitoring the icing status of wind turbine blades. Background Technology

[0002] Icing has become a significant factor affecting wind power generation losses in cold regions. In extremely low temperatures, rain and snow easily lead to icing on wind turbine blades. Icing alters the blade's geometry, affects its aerodynamic and mechanical characteristics, and impacts the aerodynamic noise performance of the wind turbine, causing significant changes in noise levels. It also reduces the power generation efficiency of the wind turbine generator set, affects the expected lifespan of wind turbine components, and in severe cases, can even lead to blade breakage and turbine collapse. Therefore, detecting icing on wind turbine blades is crucial to ensuring the safe operation of wind turbine generator sets and the power grid system.

[0003] Icing monitoring technology primarily employs contact-based measurement, where sensors are installed on or inside the blades. This method not only struggles to achieve real-time monitoring of the icing status of existing wind turbines but is also very costly. Since icing affects the evolution mechanism of wind turbine noise, the stable characteristics of the acoustic spectrum after icing can be extracted, and the correlation between factors such as icing morphology and location and these stable characteristics can be established. This allows for monitoring of the icing status by observing changes in the wind turbine noise spectrum. Summary of the Invention

[0004] To address the problems of the prior art, this invention proposes a non-contact method for monitoring the icing status of wind turbine blades. The method includes: acquiring noise signals and image data of wind turbine blades in real time; inputting the noise signals and image data into a trained wind turbine blade icing status monitoring model to obtain monitoring results.

[0005] Training the wind turbine blade icing status monitoring model includes:

[0006] S1. Acquire wind turbine images; use an acoustic array to test the noise of the wind turbine composed of icing blades to obtain the wind turbine noise signal under icing conditions;

[0007] S2. The eigenvalue optimization beamforming algorithm is used to process the wind turbine noise signal to obtain the sound pressure level spectrum characteristics; the noise source is obtained from the wind turbine image, and the coordinates of the maximum noise source are identified. The relationship between the noise source and different icing states is established in the non-icing position.

[0008] S3. Based on the spectral characteristics of sound pressure level, construct a positive relationship between icing at different locations, with different masses, and with different incoming flow angles of attack, and the increase in the total sound pressure level of the wind turbine. Collect all the positive relationships to obtain a positive relationship dataset.

[0009] S4. Input the positive relations in the positive relation dataset into the BP neural network model to obtain the first icing state monitoring results;

[0010] S5. Input the positive relations in the positive relation dataset into the local linear wavelet neural network model to obtain the second icing state monitoring results;

[0011] S6. Determine the optimal icing state based on the relationship between the noise source and different icing states, the monitoring results of the first icing state, and the monitoring results of the second icing state.

[0012] S7. Calculate the loss function of the model based on the optimal icing state, continuously adjust the model, and complete the model training when the loss function converges.

[0013] The beneficial effects of this invention are:

[0014] This invention utilizes acoustic array technology to achieve non-contact measurement of horizontal axis icing wind turbine noise and thus monitor its icing state. It enables the monitoring of wind turbine blade noise under different icing conditions, extracts stable characteristics of the acoustic spectrum after icing, and establishes the correlation between factors such as icing morphology and location and their stable acoustic spectrum characteristics. This invention establishes a neural network model for predicting blade icing, and monitors the stable spectral characteristics of wind turbine noise sound pressure level by subsequently monitoring changes in the spectral characteristics, thereby monitoring the icing state of wind turbine blades. This provides a new method for non-contact and rapid monitoring of wind turbine blade icing. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the beamforming principle of the present invention;

[0016] Figure 2 This is a diagram of the BP neural network structure of the present invention;

[0017] Figure 3 This is a diagram of the LLWNN neural network structure of the present invention;

[0018] Figure 4 This is an overall flowchart of the present invention;

[0019] Figure 5 This is a diagram of the noise monitoring system of the present invention;

[0020] Figure 6 This is a graph showing the change in noise source location as a function of icing mass according to the present invention.

[0021] Figure 7This is a schematic diagram of the quadratic fitting equation for icing on the leading edge and windward side of the blade in this invention.

[0022] Figure 8 This is a comparison chart of the predicted icing mass and the actual icing mass of the blade's windward side in this invention.

[0023] Figure 9 This is a graph showing the relative error of BP and LLWNN in predicting the icing mass of the windward side of the blade according to the present invention.

[0024] Figure 10 This is a comparison chart of the predicted icing quality and the actual icing quality of the blade leading edge according to the present invention.

[0025] Figure 11 The error in predicting the leading edge icing quality of blades using BP and LLWNN in this invention. Detailed Implementation

[0026] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0027] A non-contact method for monitoring the icing status of wind turbine blades includes: acquiring noise signals and image data of wind turbine blades in real time; inputting the noise signals and image data into a trained wind turbine blade icing status monitoring model to obtain monitoring results.

[0028] like Figure 4 As shown, training the wind turbine blade icing status monitoring model includes:

[0029] S1. Acquire wind turbine images; use an acoustic array to test the noise of the wind turbine composed of icing blades to obtain the wind turbine noise signal under icing conditions;

[0030] S2. The eigenvalue optimization beamforming algorithm is used to process the wind turbine noise signal to obtain the sound pressure level spectrum characteristics; the noise source is obtained from the wind turbine image, and the coordinates of the maximum noise source are identified. The relationship between the noise source and different icing states is established in the non-icing position.

[0031] S3. Based on the spectral characteristics of sound pressure level, construct a positive relationship between icing at different locations, with different masses, and with different incoming flow angles of attack, and the increase in the total sound pressure level of the wind turbine. Collect all the positive relationships to obtain a positive relationship dataset.

[0032] S4. Input the positive relations in the positive relation dataset into the BP neural network model to obtain the first icing state monitoring results;

[0033] S5. Input the positive relations in the positive relation dataset into the local linear wavelet neural network model to obtain the second icing state monitoring results;

[0034] S6. Determine the optimal icing state based on the relationship between the noise source and different icing states, the monitoring results of the first icing state, and the monitoring results of the second icing state.

[0035] S7. Calculate the loss function of the model based on the optimal icing state, continuously adjust the model, and complete the model training when the loss function converges.

[0036] In this embodiment, acquiring wind turbine noise signals and images under icing conditions includes using a 112-microphone acoustic array to conduct noise tests on the wind turbine composed of iced blades, monitoring wind turbine noise under icing conditions at different locations, with different masses of icing, and with different incoming flow angles of attack. The noise signals of the wind turbine are collected by the acoustic array and recorded by a signal acquisition instrument, and images of the operating wind turbine are captured by an embedded camera.

[0037] The eigenvalue-optimized beamforming algorithm is used to process wind turbine noise signals. This includes: processing the noise signal using a time-domain weighted summation formula to obtain the noise source and its spectral characteristics; and further processing the noise source and its spectral characteristics using the eigenvalue-optimized beamforming algorithm to obtain the sound pressure level spectral characteristics. Beamforming technology determines the sound source location based on the "delay and summation" processing of the sound signal acquired by the microphone. By establishing a focusing grid model, acoustic information is collected, focused backward onto the grid nodes, and the acoustic signals at the focal point are phase-compensated and summed for output. The output at the sound source location is amplified by forward superposition while attenuating the output at other focal points, thus effectively identifying the sound source.

[0038] Beamforming principle as follows Figure 1 As shown, O1 and O2 are the origins of the reference acoustic array and the evaluation plane, respectively, and r l Let be the coordinate vector of the l-th microphone (l = 1, 2, ..., M), M be the number of microphones, H be the focusing noise source point, and r be the coordinate vector. Assuming the noise source is a monopole point source, the radiated wave is an expression for the sound pressure signal received by the microphones in the acoustic array:

[0039] p l (t)=p(t-Δ l (r))

[0040] In the formula, r is the coordinate vector of the m-th microphone (l = 1, 2, ..., M), l is the number of microphones, r is the coordinate vector of the point where we measure the sound, and p l (t) is the sound pressure signal received by microphone l, Δ l (r) is the time difference between the noise source arriving at microphone m and the reference microphone.

[0041] The formula for delay compensation is:

[0042]

[0043] In the formula, c is the speed of sound, r is the position of grid point i, and r l (r i )=|r i -r l | is the distance from microphone l to grid point i.

[0044] After compensating for the time delay of each microphone signal, each sound pressure level from the focal point is summed and normalized in the time domain to produce the output. The formula is as follows:

[0045]

[0046] In the formula, It is the normalized output of the time-domain delay-weighted summation, where t is time. The focus vector is the wavenumber, M is the number of microphones, and ω is the wavenumber. l It is a weight function in the generalized space.

[0047] The sound signal collected by the 112-channel microphone array is input into the formula above to obtain the noise source and noise spectrum characteristics, and the noise source of the rotating blade is located.

[0048] The EigenValue Optimized Beamforming (EVOB) algorithm uses cross-power eigenvalue optimization based on beamforming to significantly improve the resolution and accuracy of acoustic imaging and signal processing, and is suitable for a variety of different test scenarios and needs.

[0049] Consider a general L-dimensional data model x = s + v + n, where α is an L-dimensional data model consisting of a signal s ∈ α plus interference v ∈ α and noise n ∈ α. Assume s, v, and n are uncorrelated and have zero mean. The measurement covariance matrix is ​​obtained from the formula as follows:

[0050] R = E[xx] H ] = R ss +R vv +R nn

[0051] In the formula, Rss =E[ss H ] is the signal covariance, R vv =E[vv H ] is the interference covariance, R nn =E[nn H ] is the noise covariance.

[0052] In beamforming of matching direction and matching subspace, the lth column W of multi-rank beamforming o =[W o，1 , ..., W o，r The eigenvalue-optimized beamforming is obtained from the following formula:

[0053] W o，l =R -1 Ψ(Ψ H R -1 Ψ) -1 q l

[0054] In the formula, Ψ is a known matrix L×p (p<L) with orthogonal columns. l Q = [q1, q2, ..., q] l The lth column of []. Direction-matched and subspace-matched beamforming is a multi-rank combination technique where the output power of each mode (each subspace dimension) extracted by the eigenvalue beamformer is summed to obtain an estimate of the signal power. When the signal power is distributed over the signal subspace Ψ, the signal covariance is multi-rank; either the signal with a covariance rank of 1 is interfering, each eigenvalue beamformer captures a portion of the signal power, and the sum of the output power is a multi-rank combination used to capture all the signal power. Angular resolution decreases as the signal power acquired by the beamformer increases. Each eigenvalue-optimized beamformer extracts signal information from orthogonal subspace modes at different resolutions.

[0055] In this embodiment, two ANN models are trained using a positive relation dataset to accurately predict the relationship between leaf icing mass and the increase in sound pressure level. The two ANN models include a Back Propagation (BP) neural network and a Local Linear Wavelet Neural Network (LLWNN). Each neuron is a perceptron; the received signal weights are summed as the input to its activation function f(x), and the output is fed to the neuron in the first hidden layer below. The hidden layer can contain one or more neurons. The output layer reads the output value of the last neuron in the hidden layer and provides the prediction result. Its BP neural network structure diagram is shown below. Figure 2As shown.

[0056] Wavelet Neural Networks (WNNs) replace the activation function of a backpropagation (BP) neural network with a wavelet function obtained by translation and scaling of the wavelet basis function Ψ(x). This overcomes the influence of local extrema and extracts the characteristics of local signals. However, when dealing with problems with many parameters, multiple hidden layers are required, resulting in high computational complexity. To realize the training and prediction process of a real-time wind field detection system using neural networks, computational resource constraints need to be considered. Therefore, an improved version of the WNN, the LLWNN, is adopted, which requires only one hidden layer to achieve efficient computation. Its LLWNN neural network structure diagram is shown below. Figure 3 As shown.

[0057] The hidden layers of the LLWNN neural network consist of perceptrons and a new type of neuron (wavelet neurons), with each wavelet neuron corresponding to a perceptron. Each wavelet neuron performs a wavelet transform on each input element and multiplies the transform results to obtain the multivariate wavelet function Ψ. j (x1,x2,…x m Output:

[0058]

[0059]

[0060] In the formula, A = {a} j,i}, D={d j,i} are the scaling and translation factors of the wavelet function, respectively. Each wavelet neuron has a corresponding perceptron, the activation function of which is f(x) = x, and their outputs are v. j These are used as weights for the corresponding wavelet neuron output values. Let L be the number of wavelet neurons in the neural network, then the final predicted output value is:

[0061]

[0062] Where, x m Let v be the m-th data point, L be the number of neurons in the neural network, and v be the number of neurons in the network. j The weights corresponding to the output values ​​of the wavelet neural network,

[0063] In this embodiment, a monitoring system based on non-contact measurement of the icing status of a horizontal axis wind turbine is constructed. The system's equipment and instruments mainly include: a VC6023P photoelectric tachometer, a 112-microphone acoustic array, an anemometer, a temperature and humidity meter, CAE Noise Inspector analysis software, an HS6020 sound calibrator, a 600W small wind turbine, etc. A noise monitoring system is also included. Figure 5 As shown.

[0064] To ensure the accuracy of the monitoring data, a reflective sheet was attached to the back of the wind turbine blade root during monitoring. A VC6023P photoelectric tachometer was used to monitor the actual rotational speed of the wind turbine, and the speed was stabilized after 10 minutes of operation. To ensure positioning accuracy and minimize the generation of false noise sources, the microphone array was fixed on the constructed platform, ensuring that the plane of the microphone array was parallel to the plane of the wind turbine impeller.

[0065] Noise signals from the wind turbines were acquired by an acoustic array and recorded by a signal acquisition instrument, with images of the operating wind turbines captured by an embedded camera. All signals were ultimately input to a computer via a network cable. The sampling time was 10 seconds. A-weighted sound pressure level and one-third octave band spectrum analysis were used, and an eigenvalue-optimized beamforming algorithm was selected to identify noise source distribution and analyze sound pressure level spectral characteristics. The coordinates of the largest noise source in the noise source identification were obtained using NumPy's stack and the WHERE function in Python.

[0066] This study establishes the relationship between noise source location and icing mass under different icing conditions. Noise around wind turbines is generated due to the interaction between the operating blades and the surrounding fluid. Specifically, the surrounding fluid experiences lift from the individual blade elements, resulting in periodic changes in particle velocity. Icing alters the original airfoil of each blade, causing instability in the flow field around the blades, and this airfoil change leads to fluid velocity variations. The rotating blades are also subject to Coriolis and centrifugal forces. Higher icing mass results in greater Coriolis and centrifugal forces, leading to a greater fluid movement along the blade spanwise towards the blade tip, thus shifting the noise source's spanwise position towards the blade tip. Therefore, monitoring the noise source location allows for analysis of the airfoil changes caused by icing, thereby monitoring the icing status of wind turbine blades.

[0067] When the windward side of the blade is covered with ice, the correlation coefficient R 2 The value is 0.997. The relationship between the noise source location and the icing quality is as follows:

[0068] f1(x) = 0.0054x 2 +0.0214x+0.6241

[0069] Under the condition of icing at the leading edge of the blade, the correlation coefficient R 2 The value is equal to 0.990. The relationship between the noise source location and the icing quality is also as follows:

[0070] f2(x)=0.0142x 2 +0.0029x+0.6248

[0071] The curve showing the change in noise source location with icing mass is shown below. Figure 6 As shown.

[0072] The total sound pressure level (SPL) is defined as the noise increase ΔOASPL. The larger the absolute value of the increase in SPL, the greater the impact of icing on this increase. ΔOASPL is defined as:

[0073] ΔOASPL=OASPL IcedTE -OASPL NoicedTE

[0074] Among them, OASPL IcedTE It is the total sound pressure level (OASPL) of the icing blades in dB. NoicedTE This represents the total sound pressure level (SPL) of the un-iced blade in dB. A quadratic polynomial was used to fit the relationship between the icing mass and the increase in total SPL in three different frequency bands: 0.63-1.25kHz, 1.25-3.15kHz, and 3.15-8kHz. The fitted quadratic polynomial equation is shown below. Figure 7 .according to Figure 7 The data in the figure shows the judgment coefficient R of the quadratic fitting equation for the increase in total noise and the ice mass at different ice locations. 2 The values ​​are all greater than 0.93, indicating a good fit between the two. Therefore, under appropriate conditions, the fitting equation can be used to calculate the increase in sound pressure level of iced blades corresponding to any icing mass between 0 and the maximum icing mass.

[0075] In this embodiment, a deep neural network prediction model is built, and two ANN models are trained: a BP neural network and a local linear wavelet neural network, in order to accurately predict the relationship between the icing mass of the blades and the increase in sound pressure level.

[0076] First, a 3-layer backpropagation (BP) neural network is constructed, with n input layer nodes. i The value is 3, representing the increase in total sound pressure level across three different frequency bands: 0.63kHz-1.25kHz, 1.25kHz-3.15kHz, and 3.15kHz-8kHz. The number of output layer nodes, n... o A value of 1 corresponds to the icing mass. To select a suitable number of hidden layer nodes n... neu Typically, n is selected according to the formula. neu ∈[4,13], the value corresponding to the optimal training effect of the network is determined by enumeration. Meanwhile, to avoid overfitting during training, n should be reduced as much as possible while still meeting accuracy requirements. neu This makes the structure more compact.

[0077]

[0078] Where a is a constant, a∈(1,10).

[0079] Number of training samples n s At least the number n of weights and biases used in the network to be trained. var 5-10 times, n var The following formula can be used to calculate and select the network to ensure that it has sufficiently high performance and generalization ability.

[0080] n var =(n i +1)n neu +n0(n neu +1)

[0081] A backpropagation (BP) neural network model was used for prediction. Logsig and purelin activation functions were used between the hidden and output layers of the training samples, respectively. The learning function (learngdm) and the Levenberg-Marquardt training function were used for training. The target root mean square error was 0.001, the maximum number of training iterations was 1000, and the learning rate was 0.01. Multiple neural networks were trained using different combinations of hidden layer nodes and training samples, and predictions were made for different icing locations on the test set. The coefficient of determination (R²) was used to calculate the predictions. 2 The network prediction performance was evaluated, and it was found that when the number of hidden layer nodes was 5 and the number of training samples was 349, the R value in the test set was [missing information]. 2 The values ​​are all above 0.93, indicating that the network has good prediction performance.

[0082] An LLWNN neural network structure was built in Matlab using the Morlet function, with n input layer nodes. i The output layer has 3 nodes, n. o The target root mean square error (RMSE) is 0.001, with a value of 1. Increasing the number of wavelet neurons in the hidden layer leads to increased computational complexity, thus affecting the prediction error. Therefore, by varying the maximum number of neurons in the hidden layer to achieve the target RMSE, the coefficient of determination R in neural network performance evaluation can be observed. 2 The value is used to judge the quality of the trained network. The LLWNN neural network achieves good prediction results even when the number of wavelet neurons in the hidden layer is less than 12. During programming and testing, it was found that the prediction error was minimized when the number of hidden layer neurons was 10, and the coefficient of determination R in the test set was the highest. 2 All values ​​were above 0.93, indicating good training results. The wavelet basis functions used in the LLWNN neural network in this paper are as follows:

[0083]

[0084] To test the predictive ability of artificial neural networks for leaf icing conditions, a relative error rate E is defined between the predicted icing mass and the actual icing mass of the leaf.rel and average error rate E ave .

[0085]

[0086]

[0087] In the formula, M i For the actual icing mass, M j Let k = 1, 2, 3, ..., 10, representing the icing mass predicted by the artificial neural network. Ten different icing mass samples were randomly assigned to the windward side and leading edge of the wind turbine blade, respectively. Based on the relationships described in the previous chapter, the increase in total sound pressure level (SPL) within three different frequency bands—0.63kHz-1.25kHz, 1.25kHz-3.15kHz, and 3.15kHz-8kHz—was calculated. Using trained BP and LLWNN neural networks, the increase in SPL was input to predict the icing mass of the wind turbine blade. A comparison chart of the predicted icing mass and the actual icing mass at the windward side and leading edge, along with a table showing the relative error between the BP and LLWNN predictions of the icing mass at the windward side of the blade, are provided. Figures 8-11 As shown.

[0088] The average relative error rates for the icing mass predicted by the LLWNN neural network on the windward side and leading edge of the blade were 3.14% and 7.35%, respectively, while the average relative error rates for the icing mass predicted by the BP neural network were 5.60% and 9.09%, respectively. Therefore, the LLWNN has improved prediction capability compared to the BP neural network. The LLWNN neural network can effectively solve the problem of local extrema, thus enabling the gradient descent algorithm to find the globally optimal parameters, making it more suitable for this embodiment. By subsequently monitoring the stable spectral characteristics of the wind turbine noise sound pressure level, and then monitoring the icing status of the wind turbine blades, a solution is provided for the non-contact rapid monitoring and de-icing of wind turbine blade icing.

[0089] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A non-contact method for monitoring the icing status of wind turbine blades, characterized in that, include: Real-time acquisition of noise signals and image data of wind turbine blades; The noise signal and image data are input into the trained wind turbine blade icing status monitoring model to obtain the monitoring results; Training the wind turbine blade icing status monitoring model includes: S1. Acquire wind turbine images; use an acoustic array to perform noise testing on the wind turbine composed of icing blades to obtain the wind turbine noise signal under icing conditions; S2. The eigenvalue optimization beamforming algorithm is used to process the wind turbine noise signal to obtain the sound pressure level spectrum characteristics; the noise source is obtained from the wind turbine image, and the coordinates of the maximum noise source are identified. The relationship between the noise source and different icing states under different icing locations is established. S3. Based on the spectral characteristics of sound pressure level, construct a positive relationship between icing at different locations, with different masses, and with different incoming flow angles of attack, and the increase in the total sound pressure level of the wind turbine. Collect all the positive relationships to obtain a positive relationship dataset. S4. Input the positive relations in the positive relation dataset into the BP neural network model to obtain the first icing state monitoring results; S5. Input the positive relations in the positive relation dataset into the local linear wavelet neural network model to obtain the second icing state monitoring results; S6. Determine the optimal icing state based on the relationship between the noise source and different icing states, the monitoring results of the first icing state, and the monitoring results of the second icing state. S7. Calculate the loss function of the model based on the optimal icing state, continuously adjust the model, and complete the model training when the loss function converges.

2. The non-contact method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that, The eigenvalue-optimized beamforming algorithm is used to process wind turbine noise signals, including: processing the wind turbine noise signals using a time-domain weighted summation formula to obtain the noise source and noise spectrum characteristics; and processing the noise source and noise spectrum characteristics using the eigenvalue-optimized beamforming algorithm to obtain the sound pressure level spectrum characteristics.

3. The non-contact method for monitoring the icing status of wind turbine blades according to claim 1, characterized in that, The local linear wavelet neural network model processes the positive relationship by: establishing an LLWNN neural network structure using the Morlet function, with 3 nodes in the input layer, 1 node in the output layer, and a target root mean square error of 0.001; and using the trained LLWNN neural network to process the positive relationship to obtain the second icing state monitoring result.

Images