Method for intelligently determining on and off of a lightning diverter wire of a wind turbine blade lightning receptor

Abstract

The application discloses a method for intelligently determining the conduction and disconnection of a lightning conductor of a wind turbine blade lightning receptor, and relates to the technical field of wind driven generators.The application comprises: a sine wave harmonic superposition excitation source for generating a composite excitation signal; a reflection coefficient of a Y-shaped lightning conductor structure is calculated based on a characteristic impedance; an architecture of a KAN model is designed, an approximate equivalent function of a harmonic signal is selected as a base function of the KAN model, the KAN model is trained using response data of the Y-shaped lightning conductor system, and network weights and biases are optimized; a ResNet network structure comprising a plurality of residual blocks is constructed, parameters of the residual blocks are configured, and the ResNet model is trained using signal data with autocorrelation and cross-correlation characteristics; and the fault probability is evaluated by comparing the current signal with the ResNet model.The application can more meticulously analyze the characteristics and state of the lightning conductor by synthesizing sine waves of different frequencies to generate rich frequency components.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of wind power generators, and particularly relates to a method for intelligently determining the conduction and disconnection of a lightning receptor lightning conductor of a wind turbine blade. BACKGROUND

[0002] According to the Global Wind Report 2024, the global growth rate of wind energy is rapid. However, with the rapid development of the wind power industry, the safe operation of wind turbine generators has become increasingly prominent. Lightning strikes can cause serious damage to wind turbine generators, bringing huge economic burdens and adverse effects to wind farms. Therefore, ensuring the safe operation of wind turbine generators, especially in terms of lightning protection, is crucial for the stable development of wind farms.

[0003] The wind turbine blade is one of the key components of a wind turbine generator, and the lightning receptor lightning conductor inside the blade is responsible for protecting the blade from lightning damage in lightning weather. The lightning conductor safely guides the lightning current to the ground, thereby avoiding damage to the blade due to lightning strikes. However, the lightning conductor may experience broken lines or unstable on-off problems during long-term operation due to factors such as material aging, mechanical stress, and environmental erosion, which will greatly reduce the lightning protection effect of the blade and increase the risk of lightning strikes on the wind turbine generator.

[0004] Traditional lightning conductor detection methods mainly rely on manual inspection and simple electrical measurements, which have problems such as low efficiency, insufficient accuracy, and inability to monitor in real time. With the development of technology, people have begun to explore more advanced detection techniques to improve the efficiency and accuracy of detection. For example, detection techniques based on acoustic, optical, and electromagnetic principles have been proposed and applied in actual detection work. These techniques capture specific signals generated by lightning conductors when they are damaged to assess the integrity of the lightning conductor. Existing fault diagnosis methods for wind turbine blade lightning conductors include:

[0005] 1. Visual inspection: manually inspect the external condition of the lightning conductor to find obvious breaks or damage.

[0006] Disadvantages: unable to detect internal damage or damage in hidden locations, low efficiency, and prone to missed detection.

[0007] 2. Resistance test: measure the resistance of the lightning conductor to assess its continuity.

[0008] Disadvantages: can only determine the on-off state, cannot locate the specific break point, and is greatly affected by environmental factors.

[0009] 3. Time Domain Reflectometry (TDR): uses the propagation time of electromagnetic pulses in the lightning conductor to locate the fault position.

[0010] Disadvantages: high precision requirements for equipment, high cost, and may be disturbed in complex environments.

[0011] 4. Acoustic detection: Identify faults by capturing the sound signals generated when the lightning conductor breaks.

[0012] Disadvantages: Susceptible to environmental noise, limited detection range.

[0013] 5. Infrared thermal imaging: Detect thermal anomalies in lightning conductors using thermal imaging technology.

[0014] Disadvantages: High cost, high technical requirements for operators, and limited detection of fault types.

[0015] 6. Vibration analysis: Indirectly determine the integrity of lightning conductors by analyzing blade vibration characteristics.

[0016] Disadvantages: Influenced by various factors such as blade mechanical state, detection results are not direct and accurate.

[0017] Existing detection methods, although to some extent can locate faults, but they are usually limited by detection accuracy, cost, equipment complexity and requirements for operating environment. Fault diagnosis of complex systems usually relies on single or limited signal types and traditional signal processing methods, which have limitations in dealing with highly nonlinear and dynamic systems. Therefore, the method for determining the continuity of lightning conductors proposed in this paper aims to solve the above problems. SUMMARY

[0018] The purpose of the present application is to provide a method for intelligently determining the continuity and disconnection of the lightning conductor of the fan blade lightning arrester. By synthesizing sine waves of different frequencies, a rich frequency component is generated, which can more carefully analyze the characteristics and state of the lightning conductor, helping to identify and locate small changes and potential faults on the lightning conductor. By constructing a complex neural network model, it can learn and extract features from a large amount of data, and then realize advanced analysis of signals and fault diagnosis, solving the existing problems.

[0019] To solve the above technical problems, the present application is realized by the following technical scheme:

[0020] The present application is a method for intelligently determining the continuity and disconnection of the lightning conductor of the fan blade lightning arrester. To enhance the fault detection capability of the lightning arrester of the wind turbine blade, realize the rapid and accurate evaluation of the conductivity performance of the lightning conductor, and ensure the continuous effectiveness of its lightning protection function, including.

[0021] Sine wave harmonic superposition excitation source:

[0022] The invention discards the traditional square wave or pulse excitation mode, and instead adopts a sine wave as the excitation source. By superimposing 1 to 29 prime harmonic waves, the complex reflection signal of the Y-shaped lightning arrester wire can be simulated. Since there is no simple integer multiple relationship between the prime harmonic waves, the interference between different harmonic components is small. Ensure that each harmonic component independently represents a specific system characteristic, reducing confusion between harmonic components. In the invention, the excitation source does not use square wave or pulse mode, but uses a signal excitation method based on sine wave harmonic superposition, which can more accurately simulate and analyze the characteristics and state of the lightning arrester wire.

[0023] Further, the superposition mode of the invention not only enriches the frequency components of the signal, but also improves the characterization ability of the signal to the lightning arrester wire. A sine wave can represent a complex waveform by superimposing the fundamental frequency and its harmonics. A periodic signal can be represented as:

[0024] x(t)=A1sin(2πf0t+φ1)+A2sin(2π3f0t+φ2)+...+A M sin(2π29f0t+φ M )

[0025] Formula (1)

[0026] Where A is the amplitude of the fundamental frequency to the 29th harmonic, φ is the phase of the corresponding harmonic, f0 is the fundamental frequency, and t is the time. In the sine wave superposition, the amplitude of the fundamental frequency and each harmonic should be adjusted according to the frequency response characteristics of the system. Generally, the amplitude of the low-frequency component (such as the fundamental frequency) is larger, while the amplitude of the high-frequency component is smaller. This is because high-frequency signals may be attenuated in complex systems, but still need to maintain a certain intensity to capture the high-frequency characteristics of the system. For prime harmonic waves, the amplitude is set according to the decreasing rule:

[0027] A n =A0 / n α ;

[0028] Where n is the number of prime harmonic waves, A0 is the amplitude of the fundamental frequency, and α is a decay factor. This ensures that the amplitude of the high-frequency component is not too large, while retaining sufficient high-frequency information.

[0029] Take a Y-shaped network with characteristic impedance Z c for each branch as an example. For a Y-shaped lightning arrester structure, the reflection coefficient Γ can be determined by the ratio of the fault point reflection signal voltage to the incident voltage at that point:

[0030]

[0031] Where Z L is the load impedance, and Z C is the characteristic impedance of the transmission line.

[0032] The transmission coefficient is defined as:

[0033]

[0034] From equation 2, equation 3 can also be written as:

[0035] γ L = 1 + Γ L ;

[0036] If k'2, k'3 are the reflection coefficients on the branch, and the equivalent reflection coefficients are k2, k3, which satisfy:

[0037]

[0038] k1, k2, k3 satisfy:

[0039]

[0040] Combining the frequency domain representation of the sinusoidal wave superposition signal and the reflection coefficient Γ, the corresponding relationship between x(t) and k value is established by training the KAN model.

[0041] Further, the Kolmogorov-Arnold Networks (KAN) model application: KAN is a neural network architecture inspired by the Kolmogorov superposition theorem, and the goal of this model is to approximate complex function mappings with limited resources. Mathematically, Kolmogorov's theorem states that any multivariate continuous function can be approximated by a combination of a finite number of simpler functions. The mathematical expression is as follows:

[0042]

[0043] Each neuron in the KAN model can be seen as a basis function, and the linear combination of these basis functions can approximate the target function. For the Y-type fan blade lightning wire system, the basis function can be the approximate equivalent function of the harmonic signal, such as the harmonic components of the sine wave or cosine wave. With 29 harmonics as the minimum term of the basis function, it ensures that the model can capture the high-frequency characteristics of the signal. A system parameter calculation method using the KAN model is adopted, which can be continuously optimized and adjusted according to new data to adapt to changes in different environments and conditions.

[0044] Further, the KAN network structure includes an input layer, a hidden layer, and an output layer. The input layer receives signals represented by basis functions, and the hidden layer processes the signals through a nonlinear activation function. The KAN model is trained using response data of a Y-type lightning protection line system, and after training is completed, the KAN model can output parameters of the system, such as k1, k0, k2, and k3. These parameters describe the response characteristics of the system to different harmonic components. The local function expression of each harmonic signal is Then the basis function can be represented as:

[0045]

[0046] Here q corresponds to different harmonics (such as 1st, 3rd, 5th harmonics, etc.). From the above, the following relationship can be established:

[0047]

[0048] Further, a ResNet fault detection model is constructed:

[0049] A deep neural network ResNet is constructed, which solves the gradient vanishing problem in deep network training by introducing residual learning. The network contains multiple residual blocks, and the residual block is represented as:

[0050] H(x) = F(x) + x

[0051] where F(x) is the transformation inside the residual block, and H(x) is the output of the residual block.

[0052] Each residual block is composed of multiple convolutional layers and activation functions. The input features include the autocorrelation and cross-correlation of the reflection signals of multiple branches of the Y-type lightning arrester. The autocorrelation module captures the feature changes of the signal by calculating the correlation of the input signal itself at different time points. For each input signal x i , the autocorrelation function R(x i ) is calculated as:

[0053]

[0054] The cross-correlation module is used to calculate the correlation between different input signals, similar to the cross-attention mechanism. This module captures the interaction information between different branches of the signal. A deep neural network fault diagnosis model based on residual network is proposed to automatically extract the features of the signal and learn the relationship between the signal and the fault, improving the accuracy and reliability of fault diagnosis.

[0055] For two input signals x i and x j , the cross-correlation function C(x i , x j ) is calculated as:

[0056]

[0057] In ResNet, the autocorrelation and cross-correlation functions automatically learn the characteristics of the signals during the training process. In the three parts of the Y-shaped structure (at the main input, branch 1, and branch 2), the signal is excited respectively, and when the test signal encounters an impedance change at a certain part of the Y-shaped structure, part of the signal will be reflected back. At the same time, part of the signal will continue to propagate along the structure. By exciting at different points, multiple sets of reflected and transmitted signals can be collected, which contain detailed information about the Y-shaped structure. Using the above-mentioned signal data to form the input of the training set to train a deep learning model, the binary classification result gives the probability of whether the system has a fault, and the output predicts the health status of the system. This model can learn the normal state characteristics of the Y-shaped structure under normal conditions. When the system is in a fault state, by comparing the current acquired signal with the trained fault-free model, it can be evaluated whether the system has a fault.

[0058] The present application has the following beneficial effects:

[0059] 1. By superimposing multiple harmonics of the sine wave, the characteristics and state of the lightning conductor can be more accurately simulated and analyzed, and high-precision detection of the on-off of the lightning conductor can be realized.

[0060] 2. Using a deep learning model, especially an advanced architecture similar to ResNet, the features of the signal can be automatically extracted, and the relationship between the signal and the fault can be learned, improving the accuracy and reliability of fault diagnosis.

[0061] 3. The deep learning model has the ability of self-adaptive learning, which can continuously optimize and adjust according to new data to adapt to changes in different environments and conditions.

[0062] 4. By analyzing the autocorrelation and cross-correlation of multiple reflected signals of the Y-shaped branch structure, the signal data can be more comprehensively utilized, and the comprehensiveness and systematicness of detection can be improved.

[0063] 5. The signal excitation method based on harmonic superposition can work in a noisy environment, and the deep learning model can learn the signal characteristics in different environments, improving the robustness of detection.

[0064] 6. When the fan blade is in motion, the signal acquisition system can respond to the change of the blade position in real time and quickly collect reflected and transmitted signals. Due to the movement of the blade, the characteristics of the signal may change over time, and the fault detection algorithm can adapt to dynamic changes and accurately identify fault characteristics.

[0065] Of course, any product implementing the present application does not necessarily need to achieve all the advantages described above at the same time. Attached Figure Description

[0066] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0067] Figure 1 This is a schematic diagram of the system model of the detection method;

[0068] Figure 2 This is a schematic diagram of a Y-type lightning protection wire unit structure;

[0069] Figure 3 To introduce residual units with autocorrelation and cross-correlation;

[0070] Figure 4 This is the KAN-ResNet network structure. Detailed Implementation

[0071] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0072] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0073] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0074] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0075] In addition, in the description of the present application and the appended claims, the terms "first", "second", "third", etc. are used only to distinguish descriptions and cannot be understood as indicating or implying relative importance.

[0076] In the present application, the reference to "one embodiment" or "some embodiments" and the like means that the particular feature, structure or characteristic described in connection with this embodiment is included in at least one embodiment of the application. Thus, the appearance of the phrases "in one embodiment", "in some embodiments", "in other embodiments", "in yet other embodiments" and the like in various places throughout this specification are not necessarily all referring to the same embodiment, unless otherwise specifically stated. The terms "comprise", "include", "have" and their conjugates mean "including but not limited to", unless otherwise specifically stated.

[0077] Embodiment I:

[0078] Referring to Figures 1-4 As shown in the drawings, the present application provides an improved signal excitation method and artificial intelligence fault diagnosis model to enhance the fault detection capability of the lightning receptor of the wind turbine blade, to realize rapid and accurate evaluation of the conductive performance of the lightning conductor, to ensure the continuous effectiveness of its lightning protection function; the present application is a method for intelligently determining the on and off of the lightning conductor of the wind turbine blade lightning receptor, comprising:

[0079] First, prepare the hardware equipment required for wind turbine blade lightning conductor detection, including signal generator, sensor, data acquisition system, computing device, etc., install necessary software tools, such as software environment for signal processing and neural network training, use the signal generator to generate a fundamental frequency sine wave, and generate a composite excitation signal according to the 1 to 29 prime number times of harmonic superposition mode; in the present application, the excitation source does not use square wave or pulse mode, but uses a signal excitation method based on sine wave harmonic superposition, which can more accurately simulate and analyze the characteristics and state of the lightning conductor;

[0080] Adjust the amplitude and phase of each harmonic to adapt to the characteristics of the specific Y-type lightning conductor, inject the composite excitation signal into the main input point of the Y-type lightning conductor structure, install sensors on branch 1 and branch 2 of the Y-type structure, and collect reflected and transmitted signals. Perform autocorrelation and cross-correlation analysis on the collected signals, extract the time delay and intensity characteristics of the signals, and use the extracted characteristic data as the input of the deep learning model.

[0081] The architecture of the KAN model is designed, including the input layer, the hidden layer, and the output layer. The approximate equivalent function of the harmonic signal is selected as the basis function of the KAN model. The response data of the Y-type lightning conductor system is used to train the KAN model, and the network weights and biases are optimized. A ResNet network structure containing multiple residual blocks is constructed, each residual block is composed of multiple convolutional layers and activation functions, the parameters of the residual block are configured, and the signal data with autocorrelation and cross-correlation features are used to train the ResNet model. Under the condition of no fault, the ResNet model is trained to learn the normal state characteristics of the Y-type structure; under the condition of fault, the signal data is collected, and the trained model is used for fault detection. By comparing the current signal with the fault-free model, the fault probability is evaluated. The detection results of the model are tested and analyzed to verify the accuracy of fault detection. According to the test results, the model parameters and structure are adjusted and optimized.

[0082] Embodiment two:

[0083] As shown in Figures 1-4 , as an embodiment provided by the present application, the present application is a method for intelligently determining the conduction and disconnection of the lightning conductor of the fan blade lightning receptor. The traditional square wave or pulse excitation method is abandoned, and a sine wave is used as the excitation source instead. By superimposing 1 to 29 prime number harmonics, the complex reflection signal of the Y-type structure lightning conductor can be simulated. Since there is no simple integer multiple relationship between the prime number harmonics, the interference between different harmonic components is small, ensuring that each harmonic component independently represents a specific system characteristic and reducing the confusion between harmonic components; including the following steps:

[0084] Step Y001: Sine wave harmonic superposition excitation source: use a signal generator to generate a fundamental sine wave, and generate a composite excitation signal according to a 1 to 29 prime number harmonic superposition mode; as an embodiment provided by the present application, preferably, the superposition mode not only enriches the frequency components of the signal, but also improves the characterization ability of the signal to the characteristics of the lightning conductor. A sine wave can represent a complex waveform through the superposition of the fundamental frequency and its harmonics. A sine wave one cycle signal is represented as:

[0085] x(t)=A1sin(2πf0t+φ1)+A2sin(2π3f0t+φ2)+...+A M sin(2π29f0t+φ M ) (1)

[0086] Where A is the amplitude of the fundamental frequency to the 29th harmonic, φ is the phase of the corresponding harmonic, f0 is the fundamental frequency, and t is the time;

[0087] Step Y002: Calculate the reflection coefficient of the Y-type lightning conductor structure based on the characteristic impedance;

[0088] Step Y003: design the architecture of the KAN model, select the approximate equivalent function of the harmonic signal as the basis function of the KAN model, train the KAN model using the response data of the Y-type lightning rod system, and optimize the network weight and bias;

[0089] Step Y004: construct a ResNet network structure containing multiple residual blocks, configure the parameters of the residual blocks, and train the ResNet model using signal data with autocorrelation and cross-correlation characteristics;

[0090] Step Y005: evaluate the fault probability by comparing the current signal with the ResNet model.

[0091] As an embodiment provided by the present application, preferably, the generated composite excitation signal is injected into the main input point of the Y-type lightning rod structure, sensors are installed on branch 1 and branch 2 of the Y-type lightning rod structure to collect reflected and transmitted signals.

[0092] As an embodiment provided by the present application, preferably, in step Y001, when the sine waves are superimposed, the amplitudes of the fundamental frequency and each harmonic in the sine wave superposition should be adjusted according to the frequency response characteristics of the system. Generally, the amplitude of the low-frequency component (such as the fundamental frequency) is larger, and the amplitude of the high-frequency component is smaller. This is because the high-frequency signal may be greatly attenuated in a complex system, but still needs to maintain a certain strength to capture the high-frequency characteristics of the system. For prime harmonic waves, the amplitudes are set according to the decreasing rule; the amplitudes of the fundamental frequency and each harmonic are adjusted according to the frequency response characteristics of the system, and the adjustment method is:

[0093] For prime harmonic waves, the amplitudes are set according to the decreasing rule;

[0094] The amplitude is set as: A n = A0 / n α ;

[0095] Where n is the order of the prime harmonic, A0 is the amplitude of the fundamental frequency, and a is a decay factor, which can ensure that the amplitude of the high-frequency component is not too large, while retaining sufficient high-frequency information.

[0096] As an embodiment provided by the present application, preferably, the architecture of the KAN model includes an input layer, a hidden layer and an output layer; the input layer receives signals represented by basis functions, and the hidden layer processes signals through a nonlinear activation function.

[0097] As an embodiment provided by the present application, preferably, taking a Y-type network with characteristic impedance of each branch as an example, for a Y-type lightning rod structure, the reflection coefficient Γ of the Y-type lightning rod structure is determined by the ratio of the fault point reflection signal voltage to the incident voltage at that point, and the reflection coefficient Γ is:

[0098]

[0099] wherein Z L is the load impedance, Z C is the characteristic impedance of the transmission line;

[0100] The transmission coefficient of the Y-shaped lightning conductor structure is:

[0101]

[0102] From equation 2, equation 3 can also be written as, γ L = 1 + Γ L (4) ;

[0103] If k'2, k'3 are the reflection coefficients on branch 1 and branch 2 of the Y-shaped lightning conductor structure respectively, and the equivalent reflection coefficients are k2, k3, which satisfy:

[0104]

[0105] Then k1, k2, k3 satisfy:

[0106]

[0107] Combining the frequency domain representation of the sinusoidal wave superposition signal and the reflection coefficient Γ, the corresponding relationship between x(t) and k value is established by training the KAN model.

[0108] Embodiment three:

[0109] As an embodiment provided by the present application, preferably, the Kolmogorov-Arnold Networks (KAN) model is applied: KAN is a neural network architecture inspired by Kolmogorov's superposition theorem, and the goal of this model is to approximate complex function mapping with limited resources. In mathematics, Kolmogorov's theorem shows that any multivariate continuous function can be approximated by a combination of a finite number of simpler functions. The mathematical expression is as shown in the formula:

[0110]

[0111] As an embodiment provided by the present application, preferably, each neuron in the KAN model can be regarded as a basis function, and the linear combination of these basis functions can approximate the target function. For the Y-shaped lightning conductor system of the fan blade, the basis function can be the approximate equivalent function of the harmonic signal, such as the harmonic component of the sine wave or cosine wave. Taking the minimum term of 29 harmonics as the basis function ensures that the model can capture the high-frequency characteristics of the signal.

[0112] As an embodiment provided by the present application, preferably, the KAN network structure includes an input layer, a hidden layer and an output layer. The input layer receives signals represented by basis functions, and the hidden layer processes the signals through a nonlinear activation function. The KAN model is trained using response data of a Y-type lightning rod system, and after the training is completed, the KAN model can output parameters of the system, such as k1, k0, k2 and k3. These parameters describe the response characteristics of the system to different harmonic components; the local function expression of each harmonic signal is Therefore, the basis function can be represented as:

[0113]

[0114] Based on the KAN model, the corresponding relationship between the periodic signal and the reflection coefficient is established based on the frequency domain representation of the sine wave superposition signal and the reflection coefficient, where q corresponds to different harmonics (such as 1st, 3rd, 5th harmonics, etc.). The following relationship can be established according to formula (4):

[0115]

[0116] where q corresponds to different harmonics, k is the reflection coefficient, x is a periodic signal, and n is the number of prime harmonics.

[0117] As an embodiment provided by the present application, preferably, a deep neural network ResNet is constructed, and the ResNet solves the gradient disappearance problem in the training of a deep network by introducing residual learning. The network includes multiple residual blocks; in the ResNet network structure, each residual block is composed of multiple convolutional layers and activation functions, and the residual block is represented as:

[0118] H(x)=F(x)+x (10);

[0119] where F(x) is the transformation inside the residual block, and H(x) is the output of the residual block.

[0120] As an embodiment provided by the present application, preferably, a cross-correlation module is used to calculate the correlation between different input signals, similar to the cross-attention mechanism. The module captures the interaction information between different branches; in the ResNet network structure, the input features include the autocorrelation and cross-correlation of the reflection signals of multiple branches of the lightning arrester Y-type lightning rod; the autocorrelation module captures the feature changes of the signal by calculating the correlation of the input signal itself at different time points; for each input signal x i , the autocorrelation function R(x i ) is calculated as:

[0121]

[0122] The cross-correlation module captures the interaction information between signals of different branches by calculating the correlation between different input signals. i and x j The cross-correlation function C(x i , x j ) is calculated as:

[0123]

[0124] In the ResNet, the autocorrelation and cross-correlation functions automatically learn the features of the signals during the training process. The signals are excited at the three parts of the Y-shaped structure (the main input, branch 1, and branch 2). When the test signal encounters an impedance change at a certain part of the Y-shaped structure, part of the signal will be reflected back. At the same time, part of the signal will continue to propagate along the structure. By exciting at different points, multiple sets of reflected and transmitted signals can be collected, which contain detailed information of the Y-shaped structure. The acquired signal data is used to form the input of the training set to train a deep learning model. The binary classification result gives the probability of whether the system has a fault, and the output predicts the health state of the system. This model can learn the normal state features of the Y-shaped structure under the condition of no fault. When the system is in a fault state, by comparing the currently acquired signal with the trained fault-free model, it can be evaluated whether the system has a fault.

[0125] As an embodiment provided by the present application, preferably, the autocorrelation and cross-correlation analysis of the reflected and transmitted signals collected by the sensors on branch 1 and branch 2 of the Y-shaped lightning conductor structure is performed, and the time delay and intensity features of the signals are extracted, and the extracted feature data is used as the input of the deep learning model.

[0126] The signals are excited at the main input, branch 1, and branch 2 of the Y-shaped lightning conductor structure. When the test signal encounters an impedance change at a certain part of the Y-shaped lightning conductor structure, part of the signal will be reflected back, and part of the signal will continue to propagate along the structure.

[0127] By exciting at different points, multiple sets of reflected and transmitted signals are collected.

[0128] The acquired reflected and transmitted signal data is used to form the input of the training set to train a deep learning model. The binary classification result gives the probability of whether the system has a fault, and the output predicts the health state of the system.

[0129] As an embodiment provided by the present application, preferably, the ResNet network structure can learn the normal state features of the Y-shaped lightning conductor structure under the condition of no fault. When the system is in a fault state, by comparing the currently acquired signal with the trained fault-free model, it can be evaluated whether the system has a fault.

[0130] The method for intelligently determining the conduction and disconnection of the lightning rod wire of the fan blade lightning receptor can more accurately simulate and analyze the characteristics and state of the lightning rod wire by superimposing multiple harmonics of a sine wave, thereby realizing high-precision detection of the on-off of the lightning rod wire. By using a deep learning model, especially an advanced architecture similar to ResNet, the features of the signal can be automatically extracted, and the relationship between the signal and the fault can be learned, thereby improving the accuracy and reliability of fault diagnosis. The deep learning model has the ability of adaptive learning, which can continuously optimize and adjust according to new data to adapt to changes in different environments and conditions. By analyzing the autocorrelation and cross-correlation of multiple reflected signals of the Y-shaped branch structure, the signal data can be more comprehensively utilized, and the comprehensiveness and systematicness of detection can be improved. The signal excitation method based on harmonic superposition can work in a noisy environment, and the deep learning model can learn the signal features in different environments to improve the robustness of detection. When the fan blade is in motion, the signal acquisition system can respond to the change of the blade position in real time and quickly collect the reflected and transmitted signals. Due to the movement of the blade, the characteristics of the signal may change over time, and the fault detection algorithm can adapt to dynamic changes and accurately identify fault features.

[0131] In the description of the present specification, the description of the terms "one embodiment", "example", "specific example", and the like means that the specific features, structures, materials, or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative description of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0132] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and specifically describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their entire scope and equivalents.

Claims

1. A method of intelligently determining the on and off of a lightning diverter conductor of a wind turbine blade lightning receptor, characterized by, The method comprises the following steps: Step Y001: a sine wave harmonic superposition excitation source is used to generate a composite excitation signal by using a signal generator to generate a fundamental frequency sine wave and in a harmonic superposition mode of 1 to 29 prime numbers; Step Y002: a reflection coefficient of the Y-shaped lightning conductor structure is calculated based on a characteristic impedance; Step Y003: an architecture of a KAN model is designed, an equivalent function of a harmonic signal is selected as a base function of the KAN model, the KAN model is trained by using response data of the Y-shaped lightning conductor system, and network weights and biases are optimized; Step Y004: a ResNet network structure comprising a plurality of residual blocks is constructed, parameters of the residual blocks are configured, and the ResNet model is trained by using signal data with autocorrelation and cross-correlation characteristics; Step Y005: a fault probability is evaluated by comparing a current signal with the ResNet model; The composite excitation signal generated is injected into a main input point of the Y-shaped lightning conductor structure, sensors are installed on branch 1 and branch 2 of the Y-shaped lightning conductor structure, and reflection and transmission signals are collected; In the step Y001, a one-period signal of the sine wave is represented as: x(t) = A1sin(2πf0t + φ1) + A2sin(2π3f0t + φ2) +... + A M sin(2π29f0t + φ M ) wherein A is an amplitude of the fundamental frequency to the 29th harmonic, φ is a phase of the corresponding harmonic, f0 is the fundamental frequency, and t is time; In the ResNet network structure, each residual block is composed of a plurality of convolutional layers and an activation function, and the residual block is represented as: H(x) = F(x) + x; wherein F(x) is a transformation inside the residual block, H(x) is a residual block output, and x is a one-period signal; In the ResNet network structure, the input features include autocorrelation and cross-correlation of the reflection signals of the multiple branches of the Y-shaped lightning rod; the autocorrelation module captures the characteristic change of the signal by calculating the correlation of the input signal itself at different time points; for each input signal x i , the autocorrelation function R(x i ) is calculated as: The cross-correlation module captures the interaction information between signals of different branches by calculating the correlation between different input signals. For two input signals x i and x j , the cross-correlation function C(x i , x j ) is calculated as:

2. The method of claim 1, wherein, In the step Y001, when the sine wave is superimposed, the amplitudes of the fundamental frequency and each harmonic are adjusted according to the frequency response characteristics of the system, and the adjustment method is as follows: For prime number harmonics, the amplitudes are set according to a decreasing rule; The amplitude is set to: A n = A0 / n α ; wherein n is the number of prime number harmonics, A0 is the amplitude of the fundamental frequency, and α is a decay factor.

3. The method of claim 1, wherein the method is performed by the intelligent lightning protection system of claim 1. The architecture of the KAN model comprises an input layer, a hidden layer and an output layer; the input layer receives a signal represented by a base function, and the hidden layer processes the signal through a nonlinear activation function.

4. The method of claim 1, wherein the method is performed by the intelligent lightning protection system of claim 1. The reflection coefficient Γ of the Y-shaped lightning conductor structure is determined by a ratio of a fault point reflection signal voltage to an incident voltage at the fault point, and the reflection coefficient Γ is: where Z L is the load impedance, Z C is the characteristic impedance of the transmission line; The transmission coefficient of the Y-shaped lightning conductor structure is: If k′2 and k′3 are reflection coefficients of branch 1 and branch 2 of the Y-shaped lightning conductor structure respectively, and the equivalent reflection coefficients k2 and k3 satisfy: Then k1, k2 and k3 satisfy:

5. The method of claim 4, wherein the method further comprises: The base function of the harmonic signal is represented as: Based on the KAN model, the corresponding relationship between a periodic signal and a reflection coefficient is established by combining a frequency domain representation of a sine wave superposition signal and a reflection coefficient, and is as follows: wherein q corresponds to different harmonics, k is a reflection coefficient, x is a one-period signal, and n is the number of prime number harmonics.

6. The method of claim 1, wherein, The reflection and transmission signals collected by the sensors on branch 1 and branch 2 of the Y-shaped lightning conductor structure are subjected to autocorrelation and cross-correlation analysis, time delay and intensity characteristics of the signals are extracted, and the extracted characteristic data are used as inputs of a deep learning model; Signal excitation is respectively conducted at the main input, branch 1 and branch 2 of the Y-shaped lightning wire structure, when the test signal meets impedance change at a certain part of the Y-shaped lightning wire structure, part of the signal is reflected back, and part of the signal continues to propagate along the structure through the part; By exciting at different points, multiple sets of reflected and transmitted signals are collected; The reflected and transmitted signal data is used as the input of the training set to train a deep learning model, and the binary classification result gives the probability of whether the system has a fault, and the output is the prediction of the health state of the system.

7. The method of claim 1, wherein the method further comprises: determining whether the lightning rod is in a conductive state or a non-conductive state. The ResNet network structure can learn the normal state characteristics of the Y-shaped lightning wire structure under the condition of no fault; when the system is in a fault state, the current obtained signal is compared with the trained fault-free model to evaluate whether the system has a fault.

Images