A wind power system converter open-circuit fault robustness diagnosis method considering wind speed fluctuation

Abstract

This invention discloses a robust diagnostic method for open-circuit faults in wind power converters considering wind speed fluctuations: Three sets of typical equivalent wind speed data at the hub height are established, and one set is randomly selected for modeling. Three-phase current signals are acquired through simulation and experiments, and their amplitudes are normalized in real time. A sliding window is used to continuously sample the preprocessed three-phase current data, and a Markov transfer field is used to convert the time-series data into image data, which are then combined to form a unified dataset. The dataset is cleaned and shuffled, proportionally divided into training and validation sets, and synchronously trained based on an improved GRU-ResCNN-Atnn combined neural network. The trained model is saved. The remaining two sets of equivalent wind speed data are selected to generate a test dataset to evaluate model performance and fine-tune parameters, establishing the optimal diagnostic model. In actual unit operation, this model is used for real-time monitoring and diagnosis of abnormal states of the converter power transistors.

Description

Technical Field

[0001] This invention relates to the field of wind power system fault diagnosis and maintenance, and specifically to a robust diagnostic method for open-circuit faults in wind power system converters that takes into account wind speed fluctuations. Background Technology

[0002] In recent years, driven by the "dual carbon" target, the new energy industry has flourished, especially the wind power market, which has expanded rapidly. The main battlefield for wind power generation is gradually shifting from land to sea, extending from nearshore waters to the deep sea. Permanent magnet direct drive technology, due to its high efficiency, stability, and low maintenance costs, is widely used in offshore wind turbines. In the electrical structure of wind turbines, power electronics technology plays a crucial role, serving as the bridge connecting the generator and the power grid. Its reliability has attracted widespread attention from academia and industry. The reliability of the converter is of paramount importance to the stable operation of the wind power system. Therefore, in-depth research into fault diagnosis methods for power devices in power converters and timely handling of related problems are essential for improving the reliability of permanent magnet direct drive wind power systems.

[0003] Despite numerous methods for improving the reliability of power electronic converters, permanent magnet direct-drive wind turbines are often located in coastal or even offshore environments. In addition, the turbine nacelles are located tens of meters high, which exposes the converters to complex and variable climate challenges such as high temperature, high humidity, salt spray corrosion, lightning, and typhoons, making it difficult to completely avoid failures.

[0004] The core component of a wind power converter is the power switching transistor, which is also the most vulnerable part of the converter. Power switching transistor failures mainly manifest as short circuits and open circuits. Short circuit faults are extremely destructive; diagnosing and protecting them through software algorithms is both unreasonable and difficult to implement, typically requiring standard hardware short-circuit protection circuits. In contrast, open-circuit faults in switching transistors do not cause severe overcurrent or overvoltage in the short term, and therefore may persist without triggering system protection mechanisms for immediate shutdown. However, under harsh operating conditions or prolonged operation with a fault, such faults lead to increased heat generation and losses in the switching devices, potentially triggering a cascading failure of other components in the wind power system, even causing catastrophic consequences and exorbitant repair costs. Summary of the Invention

[0005] Purpose of the Invention: The purpose of this invention is to provide a robust diagnostic method for open-circuit faults in wind power system converters that considers wind speed fluctuations, thereby improving the operational reliability of offshore wind farms. To effectively enhance the operational reliability of wind power systems, this paper proposes an innovative fault diagnosis method—a robust diagnostic technique for open-circuit faults in wind power system converters that considers wind speed fluctuations. This technique aims to achieve accurate monitoring and fault diagnosis of the state of wind turbine converters.

[0006] Technical Solution: To achieve the above-mentioned objectives, this invention proposes a robust diagnostic method for open-circuit faults in wind power system converters, taking into account wind speed fluctuations. This method includes the following steps:

[0007] S1. Establish three sets of equivalent wind speed data at the hub height, and randomly select one set for modeling;

[0008] S2, based on the selected wind speed data, the three-phase current signal output by the converter is obtained through simulation and experiment, and the three-phase current is preprocessed using the real-time amplitude normalization method;

[0009] S3. By considering the wind speed fluctuation, the preprocessed three-phase current data is continuously sampled to obtain a time series dataset. The time series dataset is then transformed using a Markov transfer field to obtain an image dataset. The time series dataset and the image dataset are then combined into a single dataset.

[0010] S4. After cleaning and shuffling the dataset, divide it into training set and validation set according to a predetermined ratio. Input the training set and validation set into the combined neural network model based on the improved GRU-ResCNN-Atnn for synchronous training and save the model after training.

[0011] S5. Select the remaining two sets of equivalent wind speed data, obtain the test dataset according to steps S2-S3, clean this test dataset and apply it to the saved model, and adjust the model parameters accordingly to establish the optimal diagnostic model.

[0012] S6, during actual operation of the unit, uses the optimal diagnostic model to monitor and diagnose abnormal states of the unit's converter power transistors in real time.

[0013] Furthermore, the specific method for step S1 is as follows:

[0014] Step 1.1) Measure the actual wind speed at the hub height in the wind turbine operating environment, and select three sets of wind speed data that can reflect the typical wind speed variation patterns in the region, including steady wind speed, fluctuating wind speed, and gusts. Where num represents the number of wind speed data, the first row of the matrix represents the stable wind speed, the second row represents the fluctuating wind speed, and the third row represents the gust wind speed.

[0015] Step 1.2) Considering wind shear effect and tower effect, establish wind speed data V w The equivalent wind speed model is expressed as follows:

[0016] V eq =V w +V ws +V ts (1)

[0017] Among them, V eq For the equivalent wind speed, V ws V represents the wind shear component. ts The tower shadow effect component is represented by units of m / s and V. ws and V ts The results are obtained from equations (2) and (3) respectively:

[0018]

[0019]

[0020] In the formula, R is the blade radius (m); α is the wind shear index, ranging from 0.1 to 0.7; H is the hub height (m); ω m Representing the impeller angular frequency (rad / s), obtained from wind speed data V w And wind turbine simulation acquisition; β i D represents the blade azimuth angle (rad), where the subscript i is the blade number; t d represents the tower radius (m); d represents the distance from the impeller plane to the tower centerline (m); through V w The equivalent wind speed data are obtained by combining formulas (1)-(3):

[0021]

[0022] Step 1.3) From V eq A set of wind speed data {V} is randomly selected from the data. r =v′ i1 ,v′ i2 ,···v′ inum}, where i represents a random integer value between 1 and 3.

[0023] Furthermore, the specific method for step S2 is as follows:

[0024] Step 2.1) Use Simulink to build a wind turbine simulation model, and use a permanent magnet synchronous motor to drive an experimental platform to simulate the wind power system. Simulate and experiment on converter health, single-tube fault, and dual-tube fault. Input the selected wind speed data into the simulation model and experimental platform, and measure the three-phase current signal i of the converter. abc Wind speed data is modeled as V r When adjusting the parameters, there are two remaining sets;

[0025] Step 2.2) Perform real-time normalization preprocessing on the three-phase currents to eliminate the impact of wind speed fluctuations on fault diagnosis. When the converter is operating normally, the three-phase currents are in a sinusoidal equilibrium state, expressed as:

[0026]

[0027] In the formula, i a i b i c For three-phase current, I m Let I be the real-time amplitude of the current, ω be the angular frequency, and θ be the initial phase angle. Equation (4) shows that the real-time amplitude I can be obtained using the Park transform. m The current can be normalized preprocessed, I m The expression is as follows:

[0028]

[0029] In the formula, i d i q These are the direct-axis and quadrature-axis components of the three-phase current obtained through the Park transform, respectively, thus yielding the normalized three-phase current.

[0030] Furthermore, the specific method for step S3 is as follows:

[0031] Step 3.1) Establish a sliding window that considers wind speed fluctuations using an autoregressive model of generator speed to eliminate the impact of wind speed fluctuations on speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment, as shown in the following formula:

[0032]

[0033] In the formula, n rt Let be the rotational speed at time t, and c be a constant. Here, q represents the model parameters, ε represents the model order, and ε represents the error term. The autoregressive model parameters are obtained by fitting historical rotational speed data using the least squares method, based on the rotational speed n at the previous q time points. r1 ,n r2 …,n rq Predict the rotational speed n at the next moment rq+1 Thus, the average rotational speed is obtained:

[0034]

[0035] Let the reference size of the sliding window be W. b The reference step size is S. b The speed reference is n rb The current sliding window size and movement step size are as follows:

[0036]

[0037] In the formula, rand is a random integer function;

[0038] Step 3.2) Apply a sliding window that takes into account wind speed fluctuations to the normalized three-phase current. Continuous sampling is performed, and interpolation is used to make all time series samples have the same length, resulting in a time series dataset {TimeSeries=(T1,y1),(T2,y2),·(T x ,y x ),·(T n ,y n )}, where n is the number of samples, y x Let x be the label of the x-th sample, x∈[1,n]; T x Let x be the x-th sample data, and let {T} x =x1,x2,·x t ,·x N}, where N is the sequence length, t is the time point, t∈[1,N], x t The value of the current at time t;

[0039] Step 3.3) Convert the three-phase current time series dataset TimeSeries into an image dataset {Image = (M1,y1), (M2,y2), ..., (M...} using Markov Transfer Field MTF. x ,y x ),···(M n ,y n )}, where M x For T x The Markov transfer field obtained by the MTF transform is applied to any time series sample T. x Perform MTF transformation to M x The process is as follows:

[0040] 1) Sample T x Quantile sequence partitioning

[0041] Sample T x The values ​​are divided into D discrete quantile units, using quantile d. j Marker, d j ∈[1,D], based on the division of each unit, each time-domain value x t Mapped to the corresponding quantile unit d j The above transforms the time-series samples into a quantile sequence {d} expressed in quantiles. j1 ,d j2 ,...,d jD};

[0042] 2) Construct sample T x The transition probability matrix P x

[0043] Using the single-step transition probability and multi-step transition probability defined in Markov chains, i.e., equation (10), we construct sample T. xMarkov transition probability matrix P x ;

[0044]

[0045] In the formula, P i,i-1 The single-step transition probability indicates that the device is in the quantile region d at time t-1. i-1 The element is transferred to the quantile region d at the next time t. i The probability, P i,j The multi-step transition probability represents the probability that the value is in the quantile region d at time t-1. j The element is transferred to the quantile region d at the next time t. i The probability of transition is given by the formula, where the difference between i and j is greater than 1. The specific transition probabilities are obtained by statistically analyzing the changes in the quantile sequence at each time step. All transition probabilities of the Markov chain are then arranged along the transition rules to construct the sample T. x The D×D Markov transition probability matrix P x As shown in equation (11):

[0046]

[0047] 3) Construct sample T x Markov transfer field M x

[0048] In P x Based on this, the samples are arranged according to the time sequence to obtain sample T. x The Markov transition field is shown below:

[0049]

[0050] Where, p i,j d i d j The corresponding quantile relationship in matrix P x The transition probabilities on the top and the elements on the diagonal are the corresponding self-transition probabilities;

[0051] Step 3.4) Combine the time series dataset TimeSeries and the image dataset Image into a single dataset {Data=(TimeSeries,Image)=[(T1,M1),y1],[(T2,M2),y2],···[(T x M x ),y x ],···[(T n M n ), y n ]}, where n is the number of samples.

[0052] Furthermore, the specific method for step S4 is as follows:

[0053] Step 4.1) First, clean and shuffle the dataset Data, remove outlier samples by calculating the standard deviation, randomly select 80% as the training set, and use the remaining 20% ​​as the test set.

[0054] Step 4.2) Establish a combined neural network model based on the improved GRU-ResCNN-Atnn; use the gated recurrent neural network GRU as branch 1 to process the three-phase current time series data, introduce a jump connection at the middle position of the normal gated recurrent neural network chain structure, and map the intermediate hidden state to the same dimension as the final hidden state. Assuming there are L gated recurrent units in total, the hidden state of the current sequence at the middle position is denoted as h. [L / 2] Meanwhile, a fully connected layer is introduced into the skip connections, and its formula is as follows:

[0055] y = s(wx + b) (13)

[0056] In the formula, x and y are the input and output vectors of the fully connected layer, w and b are the weight matrix and bias vector of the fully connected layer, respectively, s is the ReLU function, and the hidden state of the output current sequence of the last gated recursive unit is denoted as h. L Then the final hidden state of the improved recurrent neural network output is h. gru =h L +s(wh [L / 2] +b), h gru Contains h [L / 2] Nonlinear transformation;

[0057] Using a residual convolutional neural network (ResCNN) as branch 2 to process the three-phase current MTF transform image, for the input MTF image M and convolutional kernel K, the output MTF feature map Y is represented as:

[0058]

[0059] In the formula, i and j are the positions of the output feature maps, m and l are the sizes of the convolution kernel, and ω ij and b ij These are the corresponding weights and offset coefficients. The convolutional layer is followed by the activation function ReLU(x) = max(0,x). A batch normalization (BN) layer is placed after the convolutional or fully connected layer and before the activation function to normalize the activation values ​​in each batch. The convolutional branch of the diagnostic model uses an adaptive average pooling layer to adjust the final feature map Y' to the specified output size, assuming the size of Y' is H. in ×W in The desired output size is H out ×Wout The desired MTF feature map Y” is then calculated using the following formula:

[0060]

[0061] In the formula, input_height and input_width are the length and width of each calculation region input_region in Y', i' and j' are the positions of the input_region, and N is the length of the input_region. Y The number of elements in the input_region is denoted by , and i and j are the positions of the desired feature map Y".

[0062] An attention mechanism encoder is used to fuse the output features of branch 1 and branch 2 to extract the final feature structure used for fault classification.

[0063] Step 4.3) Input the training set and validation set into the established combinatorial neural network model for synchronous training, and save the trained model.

[0064] Furthermore, the specific method in step S5 is as follows:

[0065] Step 5.1) Select two more sets of equivalent wind speed data and repeat steps S2-S3 to obtain the test dataset;

[0066] Step 5.2) Evaluate the diagnostic accuracy of the model saved in S4 using the test set. If the evaluation results show that the model has not yet reached the preset best performance, adjust the hyperparameters of the neural network and return to step S4 to retrain the model. Repeat this process until the model achieves the best diagnostic accuracy on the test set.

[0067] Step 5.3) Save the best diagnostic model for online diagnosis in actual operation.

[0068] Furthermore, in step S6, during actual operation of the unit, the three-phase current signal is normalized in real time and sampled by a sliding window according to the method in steps S2-S3 to obtain a sequence sample. In the background analysis system, an MTF image is generated using the MTF transformation module. The sequence sample and the MTF image are input into the above-trained optimal diagnostic model to monitor and diagnose the abnormal state of the unit's converter power tube in real time.

[0069] Beneficial Effects: This invention discloses a robust diagnostic method for open-circuit faults in wind power system converters that considers wind speed fluctuations. Using the three-phase output current of the converter as the diagnostic signal, it innovatively utilizes Park transform to estimate and correct the real-time amplitude of the three-phase current for normalization. Then, the normalized three-phase current signal is used as the original signal, and a dataset is created using a sliding window sampling method considering wind speed fluctuations. One path transforms the signal into a three-channel two-dimensional matrix using a Markov Transfer Field (MTF) and feeds it into a residual convolutional network branch, while the other path uses an improved recurrent neural network branch. Finally, the outputs of the two branches are extracted into two-dimensional features and fed into an attention mechanism encoder for feature fusion. Finally, a softmax layer is used for fault identification. By normalizing the current signal using real-time current amplitude and employing a sliding window sampling method considering wind speed fluctuations to create the dataset, the influence of wind speed fluctuations on fault diagnosis is effectively addressed. Meanwhile, this combined structure can fully extract the feature information hidden in the three-phase current signal of the converter by utilizing the super two-dimensional feature learning ability of convolutional neural networks, the time series processing ability of recurrent neural networks, and the feature fusion ability of attention mechanism. It has high detection accuracy and can effectively improve the operational reliability of wind turbine units. Attached Figure Description

[0070] Figure 1 This is a schematic diagram of a permanent magnet direct-drive wind power system.

[0071] Figure 2 A framework diagram of an abnormal state detection method for wind turbine converters;

[0072] Figure 3 The diagram shows a typical wind speed curve and an equivalent wind speed curve.

[0073] Figure 4 This is a comparison chart showing the normalization effects of the real-time current amplitude normalization method with other methods.

[0074] Figure 5 This is the MTF transformation diagram of the three-phase current of the converter.

[0075] Figure 6 This is a structural diagram of the combinatorial neural network model established in this paper.

[0076] Figure 7 This is a graph showing the accuracy and training loss of a combined neural network model. Detailed Implementation

[0077] The present invention will now be described in detail with reference to specific embodiments and accompanying drawings.

[0078] This embodiment is based on a robust diagnostic method for open-circuit faults in wind power system converters that takes into account wind speed fluctuations: The schematic diagram of a permanent magnet direct-drive wind power system is shown below. Figure 1As shown, it consists of a wind turbine, a permanent magnet synchronous generator, a turbine-side rectifier, and a grid-side inverter. The turbine-side rectifier and the grid-side inverter form a back-to-back topology, enabling bidirectional energy flow and converting the variable frequency and amplitude electrical energy generated by the generator into alternating current with frequency and amplitude meeting grid connection requirements before being fed into the grid. The framework diagram of the wind turbine converter abnormal state detection method is shown below. Figure 2 As shown, firstly, three sets of equivalent wind speed data for hub height are established. Then, based on these data, the three-phase current signals output by the converter are obtained through simulation and experiments, and the current signals are normalized in real time. Subsequently, the normalized three-phase current signals are used as the original signals, and a dataset is created using a sliding window sampling method that considers wind speed fluctuations. Then, a model is built and trained using the created dataset. After the model is trained, its performance is further optimized by adjusting the model to improve diagnostic accuracy. Finally, the trained model is applied to a real-world scenario for real-time status detection and fault diagnosis.

[0079] This invention proposes a robust diagnostic method for open-circuit faults in wind power system converters that takes into account wind speed fluctuations. The method includes the following steps:

[0080] S1. Establish three sets of equivalent wind speed data at the hub height, and randomly select one set for modeling;

[0081] S2, based on the selected wind speed data, the three-phase current signal output by the converter is obtained through simulation and experiment, and the three-phase current is preprocessed using the real-time amplitude normalization method;

[0082] S3. By considering the wind speed fluctuation, the preprocessed three-phase current data is continuously sampled to obtain a time series dataset. The time series dataset is then transformed using a Markov transfer field to obtain an image dataset. The time series dataset and the image dataset are then combined into a single dataset.

[0083] S4. After cleaning and shuffling the dataset, divide it into training set and validation set according to a predetermined ratio. Input the training set and validation set into the combined neural network model based on the improved GRU-ResCNN-Atnn for synchronous training and save the model after training.

[0084] S5. Select the remaining two sets of equivalent wind speed data, obtain the test dataset according to steps S2-S3, clean this test dataset and apply it to the saved model, and adjust the model parameters accordingly to establish the optimal diagnostic model.

[0085] S6, during actual operation of the unit, uses the optimal diagnostic model to monitor and diagnose abnormal states of the unit's converter power transistors in real time.

[0086] Furthermore, the specific method for step S1 is as follows:

[0087] Step 1.1) Measure the actual wind speed at the hub height in the wind turbine operating environment and select patterns that reflect typical wind speed changes in the region, such as... Figure 3 (a) shows the three sets of wind speed data. Where num represents the number of wind speed data, the first row of the matrix represents wind1, the second row represents wind2, and the third row represents wind3;

[0088] Step 1.2) Considering wind shear effect and tower effect, establish wind speed data V w The equivalent wind speed model is expressed as follows:

[0089] V eq =V w +V ws +V ts (1)

[0090] Among them, V eq For the equivalent wind speed, V ws V represents the wind shear component. ts The tower shadow effect component is represented by units of m / s and V. ws and V ts The results are obtained from equations (2) and (3) respectively:

[0091]

[0092] In the formula, R is the blade radius (m); α is the wind shear index, ranging from 0.1 to 0.7; H is the hub height (m); ω m Representing the impeller angular frequency (rad / s), obtained from wind speed data V w And wind turbine simulation acquisition; β i D represents the blade azimuth angle (rad), where the subscript i is the blade number; t d represents the tower radius (m); d represents the distance from the impeller plane to the tower centerline (m); through V w The equivalent wind speed data are obtained by combining formulas (1)-(3):

[0093]

[0094] Step 1.3) From V eq The first set of wind speed data was randomly selected from the data. r =v′ 11 ,v′ 12 ,···v′ 1num}, Figure 3 (b) is V r The specific waveform shows that it contains significant fluctuations compared to the equivalent forward wind speed wind1.

[0095] Furthermore, the specific method for step S2 is as follows:

[0096] Step 2.1) Use Simulink to build a wind turbine simulation model, and use a permanent magnet synchronous motor to drive an experimental platform to simulate the wind power system. Simulate and experiment on converter health, single-tube fault, and dual-tube fault. Input the selected wind speed data into the simulation model and experimental platform, and measure the three-phase current signal i of the converter. abc like Figure 4 As shown in (a); the wind speed data is V when modeling. r When adjusting the parameters, there are two remaining sets;

[0097] Step 2.2) Perform real-time normalization preprocessing on the three-phase currents to eliminate the impact of wind speed fluctuations on fault diagnosis. When the converter is operating normally, the three-phase currents are in a sinusoidal equilibrium state, expressed as:

[0098]

[0099] In the formula, i a i b i c For three-phase current, I m Let I be the real-time amplitude of the current, ω be the angular frequency, and θ be the initial phase angle. Equation (4) shows that the real-time amplitude I can be obtained using the Park transform. m The current can be normalized preprocessed, I m The expression is as follows:

[0100]

[0101] In the formula, i d i q These are the direct-axis and quadrature-axis components of the three-phase current obtained through the Park transform, respectively, thus yielding the normalized three-phase current. like Figure 4 (b) shows the three-phase current signal i abc The waveform after real-time current amplitude normalization. Figure 4 (c) is the three-phase current signal i abc The waveform after normalization using the MinMaxScaler method shows that real-time amplitude normalization can effectively eliminate the influence of wind speed fluctuations on current amplitude changes.

[0102] Furthermore, the specific method for step S3 is as follows:

[0103] Step 3.1) Establish a sliding window that considers wind speed fluctuations using an autoregressive model of generator speed to eliminate the impact of wind speed fluctuations on speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment, as shown in the following formula:

[0104]

[0105] In the formula, n rt Let be the rotational speed at time t, and c be a constant. Here, q represents the model parameters, ε represents the model order, and ε represents the error term. The autoregressive model parameters are obtained by fitting historical rotational speed data using the least squares method, based on the rotational speed n at the previous q time points. r1 ,n r2 …,n rq Predict the rotational speed n at the next moment rq+1 Thus, the average rotational speed is obtained:

[0106]

[0107] Let the reference size of the sliding window be W. b The reference step size is S. b The speed reference is n rb The current sliding window size and movement step size are as follows:

[0108]

[0109] In the formula, rand is a random integer function;

[0110] Step 3.2) Apply a sliding window that takes into account wind speed fluctuations to the normalized three-phase current. Continuous sampling is performed, and interpolation is used to make all time series samples have the same length, resulting in a time series dataset {TimeSeries=(T1,y1),(T2,y2),·(T x ,y x ),·(T n ,y n )}, where n is the number of samples, y x Let x be the label of the x-th sample, x∈[1,n]; T x Let x be the x-th sample data, and let {T} x =x1,x2,·x t ,·x N}, where N is the sequence length, t is the time point, t∈[1,N], x t The value of the current at time t;

[0111] Step 3.3) Convert the three-phase current time series dataset TimeSeries into an image dataset {Image = (M1,y1), (M2,y2), ..., (M...} using Markov Transfer Field MTF. x ,y x ),···(M n ,y n )}, where M xFor T x The Markov transfer field obtained by the MTF transform is applied to any time series sample T. x Perform MTF transformation to M x The process is as follows:

[0112] 1) Sample T x Quantile sequence partitioning

[0113] Sample T x The values ​​are divided into D discrete quantile units, using quantile d. j Marker, d j ∈[1,D], based on the division of each unit, each time-domain value x t Mapped to the corresponding quantile unit d j The above transforms the time-series samples into a quantile sequence {d} expressed in quantiles. j1 ,d j2 ,...,d jD};

[0114] 2) Construct sample T x The transition probability matrix P x

[0115] Using the single-step transition probability and multi-step transition probability defined in Markov chains, i.e., equation (10), we construct sample T. x Markov transition probability matrix P x ;

[0116]

[0117] In the formula, P i,i-1 The single-step transition probability indicates that the device is in the quantile region d at time t-1. i-1 The element is transferred to the quantile region d at the next time t. i The probability, P i,j The multi-step transition probability represents the probability that the value is in the quantile region d at time t-1. j The element is transferred to the quantile region d at the next time t. i The probability of transition is given by the formula, where the difference between i and j is greater than 1. The specific transition probabilities are obtained by statistically analyzing the changes in the quantile sequence at each time step. All transition probabilities of the Markov chain are then arranged along the transition rules to construct the sample T. x The D×D Markov transition probability matrix P x As shown in equation (11):

[0118]

[0119] 3) Construct sample T x Markov transfer field Mx

[0120] In P x Based on this, the samples are arranged according to the time sequence to obtain sample T. x The Markov transition field is shown below:

[0121]

[0122] Where, p i,j d i d j The corresponding quantile relationship in matrix P x The transition probabilities on the left and right sides are represented by the elements on the diagonal, which are the corresponding self-transition probabilities. For a time-series sample T of a current signal... x The MTF conversion process is as follows Figure 5 As shown, the time-series sample is transformed into a Markov transition field M. x It is then displayed as an RGB image.

[0123] Step 3.4) Combine the time series dataset TimeSeries and the image dataset Image into a single dataset {Data=(TimeSeries,Image)=[(T1,M1),y1],[(T2,M2),y2],···[(T x M x ),y x ],···[(T n M n ), y n ]}, where n is the number of samples.

[0124] Furthermore, the specific method for step S4 is as follows:

[0125] Step 4.1) First, clean and shuffle the dataset Data, remove outlier samples by calculating the standard deviation, randomly select 80% as the training set, and use the remaining 20% ​​as the test set.

[0126] Step 4.2) Establish a combined neural network model based on the improved GRU-ResCNN-Atnn, with the specific structure as follows: Figure 6 As shown, after improving GRU and ResCNN and processing samples in parallel, information is fused through an attention mechanism, and finally the diagnostic result is output through softmax. The gated recurrent neural network GRU is used as branch 1 to process the three-phase current time series data. A skip connection is introduced at the middle position of the normal gated recurrent neural network chain structure, mapping the intermediate hidden state to the same dimension as the final hidden state. Assuming there are L gated recurrent units, the hidden state of the current sequence at the middle position is denoted as h. [L / 2] Meanwhile, a fully connected layer is introduced into the skip connections, and its formula is as follows:

[0127] y = s(wx + b) (13)

[0128] In the formula, x and y are the input and output vectors of the fully connected layer, w and b are the weight matrix and bias vector of the fully connected layer, respectively, s is the ReLU function, and the hidden state of the output current sequence of the last gated recursive unit is denoted as h. L Then the final hidden state of the improved recurrent neural network output is h. gru =h L +s(wh [L / 2] +b), h gru Contains h [L / 2] Nonlinear transformation;

[0129] Using a residual convolutional neural network (ResCNN) as branch 2 to process the three-phase current MTF transform image, for the input MTF image M and convolutional kernel K, the output MTF feature map Y is represented as:

[0130]

[0131] In the formula, i and j are the positions of the output feature maps, m and l are the sizes of the convolution kernel, and ω ij and b ij These are the corresponding weights and offset coefficients. The convolutional layer is followed by the activation function ReLU(x) = max(0,x). A batch normalization (BN) layer is placed after the convolutional or fully connected layer and before the activation function to normalize the activation values ​​in each batch. The convolutional branch of the diagnostic model uses an adaptive average pooling layer to adjust the final feature map Y' to the specified output size, assuming the size of Y' is H. in ×W in The desired output size is H out ×W out The desired MTF feature map Y” is then calculated using the following formula:

[0132]

[0133] In the formula, input_height and input_width are the length and width of each calculation region input_region in Y', i' and j' are the positions of the input_region, and N is the length of the input_region. Y The number of elements in the input_region is denoted by , and i and j are the positions of the desired feature map Y".

[0134] An attention mechanism encoder is used to fuse the output features of branch 1 and branch 2 to extract the final feature structure used for fault classification.

[0135] Step 4.3) Input the training set and validation set into the established combinatorial neural network model for synchronous training, and save the trained model.

[0136] Furthermore, the specific method in step S5 is as follows:

[0137] Step 5.1) Select two more sets of equivalent wind speed data and repeat steps S2-S3 to obtain the test dataset;

[0138] Step 5.2) Evaluate the diagnostic accuracy of the model saved in S4 using the test set. If the evaluation results show that the model has not yet reached the preset optimal effect, adjust the hyperparameters of the neural network and return to step S4 to retrain the model. Repeat this process until the model achieves the optimal diagnostic accuracy on the test set. The adjusted learning rate is 0.0005, and the other main hyperparameters are shown in Table 1.

[0139] Step 5.3) Save the optimal diagnostic model for online diagnosis in actual operation. Its accuracy and loss curves on the training and test sets are shown below. Figure 7 As shown in (a) and 7(b).

[0140] Furthermore, in step S6, during actual operation of the unit, the three-phase current signal is normalized in real time and sampled by a sliding window according to the method in steps S2-S3 to obtain a sequence sample. In the background analysis system, an MTF image is generated using the MTF transformation module. The sequence sample and the MTF image are input into the above-trained optimal diagnostic model to monitor and diagnose the abnormal state of the unit's converter power tube in real time.

[0141] Table 1. Main hyperparameters of the model

[0142]

Claims

1. A robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations, characterized in that, The method includes the following steps: S1. Establish three sets of equivalent wind speed data at the hub height, and randomly select one set for modeling, including the following steps: Step 1.1) Measure the actual wind speed at the hub height in the wind turbine operating environment, and select three sets of wind speed data that can reflect the typical wind speed change patterns in the target area, including steady wind speed, fluctuating wind speed, and gusts. Where num represents the number of wind speed data, the first row of the matrix represents the stable wind speed, the second row represents the fluctuating wind speed, and the third row represents the gust wind speed. S2, based on the selected wind speed data, the three-phase current signal output by the converter is obtained through simulation and experiment, and the three-phase current is preprocessed using the real-time amplitude normalization method; S3, by continuously sampling the preprocessed three-phase current data using a sliding window that considers wind speed fluctuations, a time-series dataset is obtained. This time-series dataset is then transformed using a Markov transfer field to obtain an image dataset. Finally, the time-series dataset and the image dataset are combined into a single dataset; this includes the following steps: Step 3.1) Establish a sliding window that considers wind speed fluctuations using an autoregressive model of generator speed to eliminate the impact of wind speed fluctuations on speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment, as shown in the following formula: (6) In the formula, n rt Let φ be the rotational speed at time t, c be a constant, and φ be the rotational speed at time t. i Here, q represents the model parameters, ε represents the model order, and ε represents the error term. The autoregressive model parameters are obtained by fitting historical rotational speed data using the least squares method, based on the rotational speed n at the previous q time points. r1 , n r2 …,n rq Predict the rotational speed n at the next moment rq+1 Thus, the average rotational speed is obtained: (7) Let the reference size of the sliding window be W. b The reference step size is S. b The speed reference is n rb The current sliding window size and movement step size are as follows: (8) (9) In the formula, rand is a random floor function; S4. After cleaning and shuffling the dataset, divide it into training and validation sets according to a predetermined ratio. Input the training and validation sets into a combined neural network model based on the improved GRU-ResCNN-Atnn for simultaneous training, and save the trained model. This includes the following steps: Step 4.1) First, clean and shuffle the dataset Data, remove outlier samples by calculating the standard deviation, randomly select 80% as the training set, and use the remaining 20% ​​as the test set; Step 4.2) Establish a combined neural network model based on the improved GRU-ResCNN-Atnn; use the gated recurrent neural network GRU as branch 1 to process the three-phase current time series data, introduce a jump connection at the middle position of the normal gated recurrent neural network chain structure, and map the intermediate hidden state to the same dimension as the final hidden state. Assuming there are L gated recurrent units in total, the hidden state of the current sequence at the middle position is denoted as h. [L / 2] Meanwhile, a fully connected layer is introduced into the skip connections, and its formula is as follows: (13) In the formula, x and y are the input and output vectors of the fully connected layer, w and b are the weight matrix and bias vector of the fully connected layer, respectively, s is the ReLU function, and the hidden state of the output current sequence of the last gated recursive unit is denoted as h. L Then the final hidden state of the improved recurrent neural network output is h gru Includes h [L / 2] Nonlinear transformation; Using a residual convolutional neural network (ResCNN) as branch 2 to process the three-phase current MTF transform image, for the input MTF image M and convolutional kernel K, the output MTF feature map Y is represented as: (14) In the formula, i and j are the positions of the output feature maps, and m and l are the sizes of the convolution kernels. and b ij These are the corresponding weights and offset coefficients, connected to the activation function after the convolutional layer. Batch normalization (BN) layers are placed after convolutional or fully connected layers and before activation functions to normalize the activation values ​​in each batch. The convolutional branches of the diagnostic model use adaptive average pooling layers to adjust the final feature map Y' to a specified output size, assuming Y' is of size H. in ×W in The desired output size is H out ×W out The desired MTF feature map Y'' is then calculated using the following formula: (15) In the formula, input_height and input_width are the length and width of each calculation region input_region in Y', i' and j' are the positions of the input_region, and N is the length of the input_region. Y The number of elements in the input_region is denoted by , and i and j are the positions of the desired feature map Y''. An attention mechanism encoder is used to fuse the output features of branch 1 and branch 2 to extract the final feature structure used for fault classification. S5. Select the remaining two sets of equivalent wind speed data, obtain the test dataset according to steps S2-S3, clean this test dataset and apply it to the saved model, and adjust the model parameters accordingly to establish the optimal diagnostic model. S6, during actual operation of the unit, uses the optimal diagnostic model to monitor and diagnose abnormal states of the unit's converter power transistors in real time.

2. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 1, characterized in that, Step S1 also includes the following steps: Step 1.2) Considering wind shear effect and tower effect, establish wind speed data V w The equivalent wind speed model is expressed as follows: (1) Among them, V eq For the equivalent wind speed, V ws V represents the wind shear component. ts The tower shadow effect component is represented by units of m / s and V. ws and V ts The results are obtained from equations (2) and (3) respectively: (2) (3) In the formula, R is the blade radius (m); α is the wind shear index, which ranges from 0.1 to 0.7; and H is the hub height (m). Representing the impeller angular frequency (rad / s), obtained from wind speed data V w Simulation of wind turbine generator sets; β i D represents the blade azimuth angle (rad), where the subscript i is the blade number; t d represents the tower radius (m); d represents the impeller. Distance from the plane to the centerline of the tower (m); through V w The equivalent wind speed data are obtained by combining formulas (1)-(3): ； Step 1.3) From V eq A set of wind speed data was randomly selected. , where i represents a random integer value between 1 and 3.

3. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 2, characterized in that, The specific method for step S2 is as follows: Step 2.1) Use Simulink to build a wind turbine simulation model, and use a permanent magnet synchronous motor to drive an experimental platform to simulate the wind power system. Simulate and experiment on converter health, single-tube fault, and dual-tube fault. Input the selected wind speed data into the simulation model and experimental platform, and measure the three-phase current signal i of the converter. abc Wind speed data is modeled as V r When adjusting the parameters, there are two remaining sets; Step 2.2) Perform real-time normalization preprocessing on the three-phase currents to eliminate the impact of wind speed fluctuations on fault diagnosis. When the converter is operating normally, the three-phase currents are in a sinusoidal equilibrium state, expressed as: (4) In the formula, i a i b i c For three-phase current, I m This represents the real-time amplitude of the current. Let θ be the angular frequency and θ be the initial phase angle. Equation (4) shows that the real-time amplitude I can be obtained using the Park transform. m The current can be normalized preprocessed, I m The expression is as follows: (5) In the formula, i d i q These are the direct-axis and quadrature-axis components of the three-phase current obtained through the Park transform, respectively, thus yielding the normalized three-phase current. .

4. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 3, characterized in that, Step S3 also includes the following steps: Step 3.2) Apply a sliding window that takes into account wind speed fluctuations to the normalized three-phase current. Continuous sampling is performed, and interpolation is used to ensure that all time-series samples have the same length, resulting in a time-series dataset. Where n is the number of samples, y x For the x-th sample label, ;T x Let x be the x-th sample data, and have Where N is the sequence length and t is the time point. x t The value of the current at time t; Step 3.3) Convert the three-phase current time series dataset TimeSeries into an image dataset using Markov Transfer Field MTF. , of which M x For T x The Markov transfer field obtained by the MTF transform is applied to any time series sample T. x Perform MTF transformation to M x The process is as follows: 1) Sample T x Quantile sequence partitioning Sample T x The values ​​are divided into D discrete quantile units, using quantile d. j Marker, d j ∈[1,D], based on the division of each unit, each time-domain value x t Mapped to the corresponding quantile unit d j The above transforms the time-series samples into a quantile sequence {d} expressed in quantiles. j1 ,d j2 ,...,d jD }; 2) Construct sample T x The transition probability matrix P x Using the single-step transition probability and multi-step transition probability defined in Markov chains, i.e., equation (10), we construct sample T. x Markov transition probability matrix P x ; (10) In the formula, P i,i-1 The single-step transition probability indicates that the device is in the quantile region d at time t-1. i-1 The element is transferred to the quantile region d at the next time t. i The probability, P i,j The multi-step transition probability represents the probability that the value is in the quantile region d at time t-1. j The element is transferred to the quantile region d at the next time t. i The probability of transition is given by the quantile sequence at each time step, where the difference between i and j is greater than 1. The specific transition probabilities are obtained by statistically analyzing the changes in the quantile sequence at each time step. All transition probabilities of the Markov chain are then arranged along the transition rules to construct the sample T. x The D×D Markov transition probability matrix P x As shown in equation (11): (11) 3) Construct sample T x Markov transfer field M x In P x Based on this, the samples are arranged according to the time sequence to obtain sample T. x The Markov transition field is shown below: (12) Where, p i,j d i d j The corresponding quantile relationship in matrix P x The transition probabilities on the top and the elements on the diagonal are the corresponding self-transition probabilities; Step 3.4) Combine the TimeSeries dataset and the Image dataset into a single dataset. , where n is the number of samples.

5. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 4, characterized in that, Step S4 also includes the following steps: Step 4.3) Input the training set and validation set into the established combinatorial neural network model for synchronous training, and save the trained model.

6. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 5, characterized in that, The specific method in step S5 is as follows: Step 5.1) Select two more sets of equivalent wind speed data and repeat steps S2-S3 to obtain the test dataset; Step 5.2) Evaluate the diagnostic accuracy of the model saved in S4 using the test set. If the evaluation results show that the model has not yet reached the preset optimal effect, adjust the hyperparameters of the neural network and return to step S4 to retrain the model. Repeat this process until the model achieves the optimal diagnostic accuracy on the test set. Step 5.3) Save the optimal diagnostic model for online diagnosis in actual operation.

7. The robust diagnostic method for open-circuit faults in wind power system converters considering wind speed fluctuations according to claim 6, characterized in that: In step S6, during actual operation of the unit, the three-phase current signal is normalized in real time and sampled by a sliding window according to the methods in steps S2-S3 to obtain a sequence sample. In the background analysis system, an MTF image is generated using the MTF transformation module. The sequence sample and the MTF image are input into the above-trained optimal diagnostic model to monitor and diagnose the abnormal state of the unit's converter power tube in real time.

Images