MMC converter bridge arm inductance parameter type fault diagnosis method and system based on Newton-Raphson optimization CNN-SVM hybrid model

Abstract

The invention provides an MMC converter bridge arm inductance parameter type fault diagnosis method and system based on a Newton-Raphson optimization CNN-SVM hybrid model. The method comprises the following steps: S1, collecting current signals of a corresponding bridge arm in a modular multilevel converter under normal and fault conditions; s2, time-frequency domain feature extraction is carried out on the collected current signals, and feature vectors with high fault correlation are screened out from time domain and frequency domain features; s3, establishing a CNN-SVM model for outputting a fault diagnosis result according to the input feature vector; s4, a Newton-Raphson optimization algorithm is adopted to optimize hyper-parameters of the CNN-SVM model; and S5, taking the screened feature vectors with strong correlation with the fault as input data, training the CNN-SVM model after hyper-parameter optimization, and applying the trained CNN-SVM model to MMC converter bridge arm inductance parameter type fault diagnosis. According to the invention, through combination of deep learning and an optimization algorithm, the accuracy, real-time performance and reliability of fault diagnosis are significantly improved.

Description

Technical Field

[0001] This invention relates to the field of fault diagnosis technology, and has a method and system for diagnosing inductance parameter-based faults in MMC converter bridge arms based on a Newton-Raphson optimized CNN-SVM hybrid model. Background Technology

[0002] Modular multilevel converters (MMCs), as a key component of flexible DC transmission technology, have been widely used in high-voltage DC transmission and renewable energy grid integration. With their advantages such as high voltage output quality and modular design, MMCs have become an important technical means to achieve efficient and stable power transmission. However, during the long-term operation of MMCs, parametric failures of key components such as bridge arm inductors have gradually become prominent. Inductor parametric failures typically refer to deviations in inductance values; for example, when the inductance value is lower than 80% of the rated value, the inductor is considered to have failed. This type of failure leads to decreased system stability, affects power quality, and may even trigger more serious system failures, causing system paralysis.

[0003] Traditional fault diagnosis methods rely primarily on human experience or simple signal processing techniques, such as monitoring changes in the amplitude of the bridge arm current to determine if an inductor is faulty. However, this method is too coarse and cannot accurately capture the aging trends and subtle changes in inductor parameters. In actual operation, inductor aging is a slow and gradual process, and its parameter changes may be very small. Simple amplitude monitoring alone can easily miss early signs of faults, resulting in insufficient sensitivity and accuracy in fault diagnosis.

[0004] While model predictive control (MMC)-based inductor identification methods can provide some parameter estimation, these methods are highly dependent on model accuracy and system operating conditions. If the model is flawed or the system operates under non-ideal conditions, the accuracy and reliability of the diagnostic results will be significantly reduced. Furthermore, these methods also suffer from limitations in computational complexity and real-time performance, making it difficult to meet the demands of MMC systems for rapid and accurate fault diagnosis.

[0005] While inductor identification based on digital twin technology theoretically boasts high accuracy, practical applications face significant challenges. The creation and maintenance of digital twin models are costly, and the system demands high real-time performance. For systems with complex topologies and numerous components, such as MMC, the application of digital twin technology still faces many challenges, and its current application in bridge arm inductor fault diagnosis is relatively limited.

[0006] Although the least squares-based inductance identification method is relatively simple to calculate, it is easily affected by outliers and may get stuck in local optima in nonlinear cases, leading to inaccurate identification results.

[0007] While data-driven inductor identification methods reduce model dependence, they require extensive training with large amounts of historical data and are highly demanding in terms of both data quality and quantity. Their identification performance is also limited when data is insufficient or noisy. Furthermore, the model's generalization ability is limited when faced with new fault modes or changes in system operating conditions, necessitating extensive retraining and data collection.

[0008] These problems limit the effectiveness of existing methods in practical engineering, making it difficult to meet the requirements of high precision, fast response, and high reliability. Therefore, developing a new fault diagnosis method to improve the accuracy and reliability of MMC bridge arm inductor fault diagnosis is of significant practical importance. Summary of the Invention

[0009] The purpose of this invention is to propose a method and system for fault diagnosis of inductor parameters in MMC converter bridge arms based on a Newton-Raphson optimized CNN-SVM hybrid model. By combining deep learning and optimization algorithms, the accuracy, real-time performance, and reliability of fault diagnosis are significantly improved. This method can not only automatically extract complex time-frequency domain features and capture early signs of inductor aging, but also enhance the adaptability and robustness of the model through optimization algorithms, effectively coping with noise interference and parameter variations under different operating conditions. Its efficient data processing capabilities and rapid diagnostic output give it significant advantages in practical engineering applications, providing strong support for the stable operation of power systems.

[0010] To achieve the above objectives, the technical solution of the present invention is as follows:

[0011] This invention proposes a fault diagnosis method for inductance parameters of MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model, which specifically includes the following steps:

[0012] Step 1: Collect current signals of the corresponding bridge arm in the modular multilevel converter under normal and fault conditions;

[0013] Step 2: Extract time-frequency domain features from the acquired current signal, and select feature vectors with strong correlation to the fault from the time-domain and frequency-domain features;

[0014] Step 3: Establish a CNN-SVM model to output fault diagnosis results based on the input feature vector. The CNN-SVM model includes an input layer, a convolution-pooling block, an SVM layer, and an output layer. The convolution-pooling block includes several alternating convolutional layers and pooling layers.

[0015] Step 4: Optimize the hyperparameters of the CNN-SVM model using the Newton-Raphson optimization algorithm;

[0016] Step 5: Use the selected feature vectors that are highly correlated with the fault as input data to train the hyperparameter-optimized CNN-SVM model, and then use the trained CNN-SVM model for MMC converter arm inductance parameter fault diagnosis.

[0017] Preferably, the time-domain feature extraction of the acquired bridge arm current signal specifically includes:

[0018] Calculate the quantitative characteristics of the current signal, including average value, maximum value, minimum value, peak-to-peak value, root mean square, variance, and standard deviation;

[0019] Calculate the infinitesimal rigidity characteristics of the bridge arm current signal, including skewness, kurtosis, peak factor, waveform factor, impulse factor, margin factor, and energy.

[0020] Preferably, the collected bridge arm current signal is subjected to frequency domain feature extraction based on fast Fourier transform. Specifically, after performing fast Fourier transform on the time domain signal, the amplitude spectrum, power spectral density, and total frequency domain energy are calculated from the frequency domain signal.

[0021] The Euclidean norm is used to process the total energy in the frequency domain. The formula for the total energy in the frequency domain is as follows:

[0022]

[0023] Among them, G i (f) represents the power spectral density; k represents the frequency index; k1 and k n These represent the upper and lower limits of the frequency range, respectively.

[0024] Preferably, the step of selecting features with strong correlation to the fault from the time domain and frequency domain features specifically involves: calculating the Spearman correlation coefficient between the feature and the fault through Spearman correlation analysis, and using the features with a correlation coefficient greater than or equal to the threshold as the final feature vector.

[0025] Preferably, the correlation coefficient threshold is set to 0.75, and features with a correlation coefficient greater than or equal to 0.75 are used as the final feature vector.

[0026] Preferably, the SVM layer uses the features after the last pooling as input and employs a basis function as the kernel function for classification.

[0027] Preferably, the optimization of the hyperparameters of the CNN-SVM model using the Newton-Raphson optimization algorithm specifically includes the following steps:

[0028] Step 4.1, Parameter Initialization: This includes initializing the population size N. pThe determinant DF is used to control population diversity, the adaptive coefficient δ is used to dynamically adjust the search step size, the maximum number of iterations M, and the iteration counter i, where DF∈(0,1), i=0;

[0029] Step 4.2: Population Initialization: Randomly generate an initial population within the search space. The position vector of each individual in the population represents a candidate solution. The dimension of the individual's position vector is consistent with the dimension of the variables in the problem to be optimized; IT represents the current iteration number obtained according to the iteration counter i, and n represents the nth individual;

[0030] Step 4.3, Fitness Assessment: Calculate the fitness function value for each individual. The fitness function is the objective function of the problem to be optimized, and the parameter X is specified. b and X w ; where X b X represents the individual with the best fitness in the current population, i.e., a candidate for the global optimum; w This represents the individual with the worst fitness in the current population;

[0031] Step 4.4, Iteration Termination Judgment: Check if the iteration counter i has reached the maximum number of iterations M: if i ≥ M, then jump to step 4.9; otherwise, execute step 4.5;

[0032] Step 4.5, Adaptive Coefficient Update: Dynamically update the adaptive coefficient δ based on the current iteration state.

[0033]

[0034] Where Max_IT represents the maximum number of iterations;

[0035] Step 4.6, Constructing the difference vector: Let x n For N in this iteration p For a given individual in the population, Δx is the Newton-Raphson search correction factor, which is x n +Δx is converted to X w , will x n -Δx is converted to X b The function value f(x) n Convert ) to x n Then the offset ε is converted to NRSR:

[0036]

[0037] Where randn is a normally distributed random number with a mean of 0 and a variance of 1;

[0038] Introducing another parameter ρ to guide the population to move in the correct direction: in [1, N]p Randomly select r1 and r2 within the interval, and satisfy r1≠r2, and calculate ρ:

[0039]

[0040] where a and b are random numbers between (0,1);

[0041] Calculate the new position vector for the next iteration The formula is as follows:

[0042]

[0043] where r′2 represents a random number between (0,1), The three types of parameters are calculated by the following formula:

[0044]

[0045] Step 4.7, Newton-Raphson correction and selection: Use the trap avoidance operator to avoid local optima; generate a random number rand∈[0,1], if rand<DF, then use the trap avoidance operator to determine the parameter The calculation formula is as follows:

[0046]

[0047]

[0048] where θ1 and θ2 are uniformly distributed random numbers between (-1,1) and (-0.5,0.5) respectively; μ1 and μ2 are random numbers, generated according to the following formula; the judgment criterion for the parameter β is as follows:

[0049]

[0050] Δ is a random number between 0-1;

[0051] If rand≥DF, then retain the calculation result of step 4.6;

[0052] Step 4.8, Iteration count update: Update X at the current position b and X w , let i = i + 1, and return to step 4.4 to continue the iteration;

[0053] Step 4.9, Output the optimal initial value: When the number of iterations reaches the preset maximum number of iterations M, the iteration terminates, and the individual with the best fitness value is selected as the optimal initial value solution of the Newton-Raphson algorithm for subsequent model training.

[0054] This invention proposes a fault diagnosis system for MMC converter arm inductance parameters based on a Newton-Raphson optimized CNN-SVM hybrid model, comprising a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the aforementioned MMC converter arm inductance parameter fault diagnosis method.

[0055] Compared with the prior art, the present invention has the following beneficial effects:

[0056] This invention addresses the shortcomings of existing modular multilevel converter (MMC) arm inductor fault diagnosis technologies, such as insufficient accuracy, poor real-time performance, weak adaptability, and low reliability. It proposes an arm inductor fault diagnosis method based on an improved convolutional neural network (CNN), which effectively solves the deficiencies of the existing technology.

[0057] Insufficient accuracy: Traditional methods rely on human experience or simple signal processing, making it difficult to capture subtle changes in the early stages of inductor aging. This invention, through an improved CNN model, can automatically extract complex features of the bridge arm current, including time-domain statistical features and frequency-domain energy distribution features, significantly improving the accuracy of fault diagnosis and enabling earlier detection of signs of inductor aging.

[0058] Poor real-time performance: Model-based predictive control methods are computationally complex and difficult to meet the needs of real-time online monitoring. The improved CNN model used in this invention, combined with feature extraction methods such as Fast Fourier Transform (FFT), has high computational efficiency, can quickly process data and output diagnostic results in real time, meeting the real-time requirements of actual operation.

[0059] Weak adaptability: Existing methods suffer a significant drop in parameter estimation accuracy when faced with complex changes in operating conditions. This invention introduces the Newton-Raphson optimization algorithm (NRBO) to optimize the hyperparameters of the CNN model, making the model more adaptable and able to maintain stable diagnostic performance under different operating conditions.

[0060] Low reliability: Least squares-based and data-driven methods are susceptible to outliers and data noise, leading to inaccurate identification results. This invention uses a time-frequency domain co-analysis method to simultaneously capture time-domain and frequency-domain features. Combined with the powerful feature learning capabilities of CNNs, this effectively reduces noise interference and improves the reliability of diagnostic results. Attached Figure Description

[0061] Figure 1 This is a topology diagram of an MMC inverter circuit;

[0062] Figure 2 This is a flowchart of the MMC bridge arm inductor aging fault diagnosis process of the present invention;

[0063] Figure 3This is the CNN-SVM network structure of the present invention;

[0064] Figure 4 This is a flowchart of the Newton-Raphson optimization algorithm of the present invention. Detailed Implementation

[0065] The following is in conjunction with the appendix Figure 1-3 The technical solution of the present invention will be described in detail below.

[0066] This invention proposes a fault diagnosis method for the inductance parameters of MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model. This invention is applicable to scenarios requiring high-precision, fast-response, and high-reliability fault diagnosis of inductance parameters in Modular Multilevel Converters (MMC), such as high-voltage direct current transmission and renewable energy grid connection. This method can effectively monitor changes in the parameters of the arm inductance, promptly detect potential faults, and ensure the stable operation of the power system. Figure 1 The topology of the MMC inverter circuit is shown.

[0067] This invention proposes a fault diagnosis method for inductance parameters of MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model, which specifically includes the following steps:

[0068] Step 1: Acquire the current signal of the corresponding bridge arm in the modular multilevel converter (MMC) under normal and fault conditions.

[0069] Step 2: Time-frequency domain feature extraction;

[0070] Step 2.1: Extract time-domain features from the bridge arm current signal. Calculate the quantitative features of the current signal, including average, maximum, minimum, peak-to-peak value, root mean square, variance, and standard deviation. The specific calculation formula is as follows:

[0071] Average value of the current data curves of the upper / lower bridge arms on the same side:

[0072] Where i ava i represents the mean value of the current in the corresponding side arm. n is the sampled data value of the signal, and N represents the total number of samples of the current of the corresponding side arm.

[0073] Extract the extreme value characteristics (maximum value i) of the current waveform of the upper / lower arm on the same side. max and minimum value i min And based on the range calculation, the peak-to-peak current value within the power frequency cycle is obtained: i f =i max -i min

[0074] Root mean square (RMS):

[0075] variance:

[0076] Standard deviation:

[0077] Step 2.2: Calculate the characteristic components of the rigidity of the bridge arm current signal, including skewness, kurtosis, peak factor, waveform factor, impulse factor, margin factor, and energy. The specific calculation formula is as follows:

[0078] Skewness:

[0079] kurtosis:

[0080] Peak factor: i peak / i RMS

[0081] Waveform factor:

[0082] Pulse factor:

[0083] Margin factor:

[0084] Total energy of the signal after processing with the Euclidean norm:

[0085] Among them, i peak Defined as the peak value of the arm current signal, which is the maximum value of the arm current during that time period; E i Defined as the total energy of the signal, this energy value can effectively characterize the overall intensity distribution characteristics of the signal, E i 'This is the result after processing with the Euclidean norm.'

[0086] Step 2.3: Extract frequency domain features from the bridge arm current signal using Fast Fourier Transform. Perform amplitude analysis on the time-domain signal to obtain the features at the corresponding frequency; the corresponding amplitude spectrum is shown in the following formula:

[0087]

[0088] Where N is the signal length (i.e., the number of sample points), i[n] represents the time-domain signal, i[k] represents the frequency-domain signal, k is the frequency index, and j is the imaginary unit. Since the above formula includes both real and imaginary components, Re(i[k]) represents the real component of the bridge arm current after FFT decomposition, while Im(i[k]) represents the imaginary component of the bridge arm current after FFT decomposition. The amplitude is the combination of the real and imaginary components, and its expression is:

[0089] For frequency domain signals, the power spectral density function is used to quantitatively characterize the energy distribution characteristics. The formula for the power spectral density function is as follows:

[0090]

[0091] The Euclidean norm is used to process the total energy in the frequency domain, and the improved formula for the total energy in the frequency domain is as follows:

[0092]

[0093] Where k represents the index in the frequency domain; k1 and k n These represent the upper and lower limits of the frequency range, respectively.

[0094] Step 2.3: Through Spearman correlation analysis, select features with strong correlation to the fault from the time domain and frequency domain features, that is, features with a correlation coefficient greater than or equal to 0.75 as the final feature vector.

[0095] Step 3: Build the neural network model. The network model structure mainly includes an input layer, convolutional layers, pooling layers, SVM layers, and an output layer, with the following structural relationships: Figure 3 As shown.

[0096] Convolutional Layer: A convolutional layer consists of multiple feature planes, each composed of multiple neurons. Each neuron is connected to a local region of the feature plane in the previous layer through a convolution kernel. The convolution kernel is a weight matrix. The convolutional layers of a CNN extract different features from the input through convolution operations. The entire convolution process can be described by the following formula:

[0097]

[0098] In the formula: This represents the output value of the v-th neuron in the output feature surface u of the convolutional layer. ω represents the output value of the nth neuron on the input feature surface m; a(i)b(j) f represents the weight parameters from the i-th neuron in the input feature surface a to the j-th neuron in the output feature surface b; b represents the bias term corresponding to the output feature surface; cov The activation function is represented by the ReLU function.

[0099] f ReLU (x) = max(0,x)

[0100] Pooling Layers: The input to pooling layers comes from multiple feature maps in the convolutional layers. Each feature map in the pooling layer corresponds to a feature map in the convolutional layer, and the number of feature maps remains constant. Neurons in the pooling layer are connected to local receptive regions in the input layer, allowing each neuron to process only a small portion of the input data, and preventing overlap between the input regions of different neurons. Pooling typically includes max pooling and average pooling, with max pooling offering the best performance and being the most widely used. Its formula is shown below:

[0101] m i(a,b) =max (n-1)ω＜t＜nω {o i(x,y)}n∈1,2,...,p

[0102] Among them o i(x，y) This represents the y-th neuron in the x-th mapping of the i-th layer. ω represents the width of the convolution kernel, and n represents the n-th pooling kernel.

[0103] Support Vector Machine (SVM) Layer: After several convolution-pooling operations in the CNN structure, the features from the final pooling are used as input features for multi-class classification. The SVM finds the optimal hyperplane to maximize the margin between classes, thus separating the support vectors of different classes as much as possible. This invention uses a radial basis function (RBF) as the kernel function for classification. The expression for the RBF is as follows:

[0104]

[0105] Here, g represents the kernel function parameters. Furthermore, by introducing a penalty factor and the principle of maximizing the margin, classification accuracy is ensured while controlling model complexity, thus theoretically avoiding the risk of overfitting.

[0106] Step 4: Optimize the hyperparameters of the CNN-SVM model using the Newton-Raphson optimization algorithm (NRBO).

[0107] Step 4.1: Parameter Initialization. Set the control parameters for the Newton-Raphson Optimization (NRBO) algorithm, including: Population size: N p ; Determinant: DF∈(0,1), used to control population diversity; Adaptive coefficient: δ, used to dynamically adjust the search step size; Maximum number of iterations: M; Initial iteration counter i=0.

[0108] Step 4.2: Population Initialization. An initial population is randomly generated within the search space, consisting of N... p It consists of 10 individuals, and the position vector of each individual represents a candidate solution. Its position vector dimension is consistent with the variable dimension of the problem to be optimized. IT represents the current iteration number obtained according to the iteration counter i;

[0109] Step 4.3: Fitness Assessment. Calculate the fitness function value for each individual. The fitness function is the objective function of the problem to be optimized or a variant thereof, and specifies the parameter X. b and X w , where: X b X represents the individual with the best fitness in the current population, i.e., a candidate for the global optimum; w It represents the individual with the worst fitness in the current population and is used to measure population diversity or as a difference basis vector.

[0110] Step 4.4: Iteration Termination Judgment. Check if the iteration counter i has reached the maximum number of iterations M: if i ≥ M, then jump to step 4.9; otherwise, execute step 4.5.

[0111] Step 4.5: Adaptive Coefficient Update. The adaptive coefficient δ is dynamically updated based on the current iteration state. The update rule is as follows:

[0112]

[0113] Where IT represents the current iteration number, Max_IT represents the maximum iteration number, and the parameter δ is adjusted automatically during each iteration.

[0114] Step 4.6: Construct the difference vector. x n For N in this iteration p For a given individual in the population, Δx is the Newton-Raphson search correction factor, which is x n +Δx is converted to X w (its position in its neighborhood is greater than x) n (better location), x n -Δx is converted to X b (its position in its neighborhood is greater than x) n (Poorer location), function value f(x) n Convert ) to x n Then the offset ε is converted into NRSR. Calculate NRSR:

[0115]

[0116] Where randn is a normally distributed random number with a mean of 0 and a variance of 1; X b For the optimal position, X w This is the worst-case position. The function value f(x) is... n Convert ) to x nThis is to save the computational time of the algorithm. In addition, the proposed NRSR improves NRBO by considering random actions, increasing diversity, and avoiding local optima during the optimization process, while significantly reducing the number of iterations. The proposed NRBO is improved by introducing another parameter ρ, which can guide the population to move in the correct direction. Randomly select r1 and r2 in the interval [1, N p , and satisfy r1≠r2, and calculate ρ:

[0117]

[0118] where represents the current position vector, and a and b are random numbers between (0, 1).

[0119] Calculate the new position vector for the next iteration The formula is as follows:

[0120]

[0121] where r2′ represents a random number between (0, 1), The three types of parameters are calculated by the following formula:

[0122]

[0123] Step 4.7: Newton-Raphson correction and selection, using the trap avoidance operator (TAO) to avoid local optima. Generate a random number rand∈[0, 1]. If rand < DF, then use the trap avoidance operator to determine the parameter The calculation formula is as follows:

[0124]

[0125] where θ1 and θ2 are uniformly random numbers between (-1, 1) and (-0.5, 0.5) respectively; μ1 and μ2 are random numbers, generated according to the following formula; the judgment criterion for the parameter β is as follows:

[0126]

[0127] Δ and rand are random numbers between 0 - 1. Since the selection of the parameters μ1 and μ2 is random, the population becomes more diverse and escapes from the local optimal solution, which helps to improve its diversity.

[0128] If rand≥DF, then retain the calculation result of step 4.6.

[0129] Step 4.8: Update the iteration count. Update X at the current position b and X w , let i = i + 1, and return to step 4.4 to continue the iteration.

[0130] Step 4.9: Output the optimal initial value. The iteration terminates when the number of iterations reaches the preset maximum number of iterations M. The individual with the best fitness value is selected as the optimal initial value solution for the Newton-Raphson algorithm and used for subsequent model training.

[0131] Step 5: Use the extracted feature vectors as input data to train the CNN-SVM model optimized with NRBO. By adjusting the model parameters and optimizing the model structure, the accuracy and generalization ability of the model can be improved.

[0132] In summary, this invention employs a comprehensive technical solution to achieve efficient diagnosis of inductor faults in modular multilevel converters (MMC) bridge arms. The technical solution includes:

[0133] First, time-domain and frequency-domain features of the bridge arm inductor current are extracted. By calculating various statistical features and frequency characteristics, the characteristics of the current signal are comprehensively captured. Then, correlation analysis is used to screen out features that are strongly correlated with faults, and an accurate feature vector is constructed.

[0134] In terms of model construction, this invention builds a fault diagnosis model based on an improved convolutional neural network (CNN). The CNN model includes convolutional layers and pooling layers; convolutional operations extract local features, while pooling operations reduce feature dimensionality. Furthermore, a support vector machine (SVM) is introduced as a classifier on top of the CNN, forming a CNN-SVM model. SVM, with its superior classification ability and generalization performance, significantly improves the accuracy and reliability of diagnosis, enabling the model to maintain stable diagnostic performance even when facing complex and changing working conditions.

[0135] To further optimize model performance, this invention employs the Newton-Raphson optimization algorithm (NRBO) to optimize hyperparameters. This algorithm, leveraging the Newton-Raphson search rule (NRSR) and trap avoidance operator (TAO), quickly and accurately searches for the optimal combination of hyperparameters, significantly improving model performance and stability, and enabling it to maintain efficient diagnostic capabilities under various complex conditions.

[0136] This invention also proposes a fault diagnosis system for MMC converter arm inductance parameters based on a Newton-Raphson optimized CNN-SVM hybrid model. The system is characterized by comprising a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs any of the steps in the aforementioned MMC converter arm inductance parameter fault diagnosis method.

[0137] The above are preferred embodiments of the present invention. Any changes made to the technical solution of the present invention that do not exceed the scope of the technical solution of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model, characterized in that, Specifically, the following steps are included: Step 1: Collect current signals of the corresponding bridge arm in the modular multilevel converter under normal and fault conditions; Step 2: Extract time-frequency domain features from the acquired current signal, and select feature vectors with strong correlation to the fault from the time-domain and frequency-domain features; Step 3: Establish a CNN-SVM model to output fault diagnosis results based on the input feature vector. The CNN-SVM model includes an input layer, a convolution-pooling block, an SVM layer, and an output layer. The convolution-pooling block includes several alternating convolutional layers and pooling layers. Step 4: Optimize the hyperparameters of the CNN-SVM model using the Newton-Raphson optimization algorithm; Step 5: Use the selected feature vectors that are highly correlated with the fault as input data to train the hyperparameter-optimized CNN-SVM model, and then use the trained CNN-SVM model for MMC converter arm inductance parameter fault diagnosis.

2. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 1, characterized in that, The time-domain feature extraction of the acquired bridge arm current signal specifically includes: Calculate the quantitative characteristics of the current signal, including average value, maximum value, minimum value, peak-to-peak value, root mean square, variance, and standard deviation; Calculate the infinitesimal rigidity characteristics of the bridge arm current signal, including skewness, kurtosis, peak factor, waveform factor, impulse factor, margin factor, and energy.

3. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 1, characterized in that, Frequency domain feature extraction based on fast Fourier transform is performed on the collected bridge arm current signal. Specifically, after performing fast Fourier transform on the time domain signal, the amplitude spectrum, power spectral density, and total frequency domain energy are calculated from the frequency domain signal. The Euclidean norm is used to process the total energy in the frequency domain. The formula for the total energy in the frequency domain is as follows: Among them, G i (f) represents the power spectral density; k represents the frequency index; k1 and k n These represent the upper and lower limits of the frequency range, respectively.

4. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 1, characterized in that, The specific steps for selecting features with strong correlation to faults from time-domain and frequency-domain features are as follows: Spearman correlation analysis is performed to calculate the Spearman correlation coefficient between the features and the faults, and features with correlation coefficients greater than or equal to the threshold are used as the final feature vectors.

5. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 4, characterized in that, The correlation coefficient threshold is set to 0.75, and features with a correlation coefficient greater than or equal to 0.75 are used as the final feature vector.

6. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 1, characterized in that, The SVM layer takes the features after the last pooling as input and uses the radial basis function as the kernel function for classification.

7. The method for diagnosing inductance parameter-based faults in MMC converter arms based on a Newton-Raphson optimized CNN-SVM hybrid model according to claim 1, characterized in that, The optimization of the hyperparameters of the CNN-SVM model using the Newton-Raphson optimization algorithm specifically includes the following steps: Step 4.1, Parameter Initialization: This includes initializing the population size N. p The determinant DF is used to control population diversity, the adaptive coefficient δ is used to dynamically adjust the search step size, the maximum number of iterations M, and the iteration counter i, where DF∈(0,1), i=0; Step 4.2: Population Initialization: Randomly generate an initial population within the search space. The position vector of each individual in the population represents a candidate solution. The dimension of the individual's position vector is consistent with the dimension of the variables in the problem to be optimized; IT represents the current iteration number obtained according to the iteration counter i, and n represents the nth individual; Step 4.3, Fitness Assessment: Calculate the fitness function value for each individual. The fitness function is the objective function of the problem to be optimized, and the parameter X is specified. b and X w ; where X b X represents the individual with the best fitness in the current population, i.e., a candidate for the global optimum; w This represents the individual with the worst fitness in the current population; Step 4.4, Iteration Termination Judgment: Check if the iteration counter i has reached the maximum number of iterations M: if i ≥ M, then jump to step 4.9; otherwise, execute step 4.5; Step 4.5, Adaptive Coefficient Update: Dynamically update the adaptive coefficient δ based on the current iteration state. Where Max_IT represents the maximum number of iterations; Step 4.6, Constructing the difference vector: Let x n For N in this iteration p For a given individual in the population, Δx is the Newton-Raphson search correction factor, which is x n +Δx is converted to X w , will x n -Δx is converted to X b The function value f(x) n Convert ) to x n Then the offset ε is converted to NRSR: Where randn is a normally distributed random number with a mean of 0 and a variance of 1; Introducing another parameter ρ to guide the population to move in the correct direction: in [1, N] p Randomly select r1 and r2 within the interval, such that r1≠r2, and calculate ρ: Where a and b are random numbers between (0,1); Calculate the new position vector for the next iteration The formula is as follows: Where r2′ represents a random number between (0,1), The three types of parameters are calculated using the following formula: Step 4.7, Newton-Raphson correction and selection: Use the trap avoidance operator to avoid local optima; generate a random number rand ∈ [0, 1], if rand < DF, then use the trap avoidance operator to determine the parameter The calculation formula is as follows: Where θ1 and θ2 are uniformly random numbers between (-1, 1) and (-0.5, 0.5); μ1 and μ2 are random numbers generated by the following formula; the criteria for determining the parameter β are as follows: Δ is a random number between 0 and 1; If rand≥DF, then retain the calculation result of step 4.6; Step 4.8, Iteration Count Update: Update the X value at the current position. b and X w Let i = i + 1, and return to step 4.4 to continue iterating; Step 4.9: Output the optimal initial value: When the number of iterations reaches the preset maximum number of iterations M, the iteration terminates. The individual with the best fitness value is selected as the optimal initial value solution of the Newton-Raphson algorithm for subsequent model training.

8. A fault diagnosis system for the inductance parameters of an MMC converter arm based on a Newton-Raphson optimized CNN-SVM hybrid model, characterized in that, It includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the MMC converter arm inductance parameter type fault diagnosis method as described in any one of claims 1-7.

Images