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

Abstract

The invention discloses a wind power converter open-circuit fault robustness diagnosis method considering wind speed fluctuation. The method comprises the following steps: establishing three groups of typical equivalent wind speed data at the height of a hub, and randomly selecting one group of typical equivalent wind speed data for modeling; obtaining a three-phase current signal through simulation and experiment, and carrying out real-time amplitude normalization on the three-phase current signal; continuously sampling the preprocessed three-phase current data by using a sliding window, converting time sequence data into image data by using a Markov transfer field, and finally combining to form a whole data set; cleaning and disrupting a data set, dividing the data set into a training set and a verification set according to a proportion, carrying out synchronous training based on an improved GRU-ResCNN-Atnn combined neural network, and storing a trained model; and selecting the remaining two groups of equivalent wind speed data to generate a test data set, evaluating model performance and finely adjusting parameters, and determining an optimal diagnosis model. In the actual operation of the unit, the abnormal state of the converter power tube is monitored and diagnosed in real time through the model.

Description

Technical Field

[0001] The present invention relates to the field of fault maintenance of wind power systems, and specifically to a robust diagnostic method for open - circuit faults of converters in wind power systems considering wind speed fluctuations. Background Art

[0002] In recent years, driven by the "dual - carbon" goal, the new energy industry has developed vigorously, especially the wind power market, whose scale has expanded rapidly. The main battlefield of wind power generation is gradually shifting from land to sea, extending from near - shore waters to deep - sea areas. The permanent - magnet direct - drive technology is widely used in offshore wind turbines due to its high efficiency, stability, and low maintenance cost. In the electrical composition of wind turbines, power electronics technology plays a crucial role. It is the bridge connecting the generator and the power grid, and its reliability issues have attracted wide attention in the academic and industrial circles. The reliability of the converter is of great significance for the stable operation of the wind power system. Therefore, in - depth research on the fault diagnosis method of power devices in power converters and timely handling of related problems are crucial for improving the reliability of permanent - magnet direct - drive wind power systems.

[0003] Although there are numerous methods to improve the reliability of power electronic converters, permanent - magnet direct - drive wind turbines are often in coastal or even deep - sea environments. Coupled with the fact that the nacelle of the unit is tens of meters high in the air, the converter faces complex and changeable climate challenges such as high temperature, high humidity, salt - fog corrosion, lightning, and typhoons, making faults difficult to completely avoid.

[0004] The core component of a wind power converter is the power switch tube, which is also the most vulnerable link in the conversion device. The faults of power switch tubes mainly appear in two forms: short - circuit and open - circuit. Short - circuit faults are extremely destructive. It is neither reasonable nor feasible to achieve their diagnostic protection through software algorithms, and usually, a standard hardware short - circuit protection circuit needs to be used to solve this problem. In contrast, the open - circuit fault of the switch tube does not cause serious over - current or over - voltage phenomena in the short term. Therefore, it may persist without triggering the system protection mechanism and causing immediate shutdown. However, under harsh working conditions or when operating with faults for a long time, such faults will lead to an increase in the heat generation and loss of the switching device, which may further cause cascading faults of other components in the wind power system, even resulting in catastrophic consequences and high maintenance costs. Summary of the Invention

[0005] Object of the Invention: The object of the present invention is to provide a robust diagnostic method for open - circuit faults of converters in wind power systems considering wind speed fluctuations, so as to improve the operation reliability of offshore wind farms. In order to effectively improve the operation reliability of wind power systems, this paper proposes an innovative fault diagnosis method - a robust diagnostic technology for open - circuit faults of converters in wind power systems considering wind speed fluctuations. This technology aims to achieve precise monitoring and fault diagnosis of the converter status of wind turbines.

[0006] Technical solution. To achieve the above invention purpose, the present invention proposes a robust diagnosis method for open-circuit faults of a wind power system converter considering wind speed fluctuations, and the method includes the following steps:

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

[0008] S2. Based on the selected wind speed data, obtain the three-phase current signals output by the converter through simulation and experiments, and preprocess the three-phase current using the real-time amplitude normalization method;

[0009] S3. Continuously sample the preprocessed three-phase current data through a sliding window considering wind speed fluctuations to obtain a time series data set, use the Markov transfer field to transform the time series data set to obtain an image data set, and combine the time series data set and the image data set into an overall data set;

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

[0011] S5. Select the remaining two groups of equivalent wind speed data, obtain a test data set according to steps S2 - S3, clean this test set and apply it to the saved model, and adjust the model parameters accordingly to establish an optimal diagnosis model;

[0012] S6. During the actual operation of the unit, use the optimal diagnosis model to monitor and diagnose the abnormal state of the power tubes of the unit converter in real time.

[0013] Further, the specific method of step S1 is as follows:

[0014] Step 1.1) Measure the actual wind speed at the hub height in the operating environment of the wind turbine, and screen out three groups of wind speed data that can reflect the typical change patterns of the wind speed in this area, including stable wind speed, fluctuating wind speed, and gust wind speed Among them, 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 the wind shear effect and the tower effect, establish an equivalent wind speed model for the wind speed data V w The expression is:

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

[0017] Among them, V eq is the equivalent wind speed, V ws represents the wind shear component, V ts represents the tower shadow effect component, and the units are all m / s, V ws and V ts are calculated by equations (2) and (3) respectively:

[0018]

[0019]

[0020] In the formula, R is the blade radius (m); α is the wind shear index, and its value ranges from 0.1 to 0.7; H is the hub height (m); ω m represents the impeller angular frequency (rad / s), which is obtained through wind speed data V w and wind turbine simulation; β i is the blade azimuth angle (rad), and the subscript i is the blade number; D t is the tower barrel radius (m); d represents the distance from the impeller plane to the center line of the tower barrel (m); the equivalent wind speed data is obtained through V w and formulas (1)-(3):

[0021]

[0022] Step 1.3) Randomly select a set of wind speed data {V eq = v′ r , v′ i1 , ··· v′ i2 , ··· v′ inum} from V

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

[0024] Step 2.1) Use Simulink to establish a wind turbine simulation model, and use a permanent magnet synchronous motor to simulate the wind power system on a back-to-back test platform. Conduct simulations and experiments on the health of the converter, single-tube faults, and double-tube faults. Input the selected wind speed data into the simulation model and the test platform, and measure the three-phase current signal i abc of the converter; when modeling the wind speed data, it is V r , and when adjusting the parameters, it is the remaining two groups;

[0025] Step 2.2) Perform real-time normalization preprocessing on the three-phase current to eliminate the influence of current fluctuations caused by wind speed fluctuations on fault diagnosis. When the converter is operating normally, the three-phase current is in a sinusoidal balanced state, and the expression is:

[0026]

[0027] wherein, i a , i b , i c are three-phase currents, I m is the real-time amplitude of the current, ω is the angular frequency, θ is the initial phase angle. Formula (4) indicates that using the Park transformation to obtain the real-time amplitude I m can perform a normalization preprocessing operation on the current. The expression of I m is as follows:

[0028]

[0029] wherein, i d , i q are respectively the direct-axis and quadrature-axis components obtained by the Park transformation of the three-phase currents. Then the normalized three-phase currents

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

[0031] Step 3.1) Establish a sliding window considering wind speed fluctuations using the generator speed autoregressive model to eliminate the influence of speed fluctuations caused by wind speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment. The formula is as follows:

[0032]

[0033] wherein, n rt is the speed value at time t, c is a constant, is the model parameter, q is the model order, ε is the error term; the autoregressive model parameters are obtained by fitting the historical speed data using the least squares method. According to the speeds n r1 , n r2 …, n rq at the previous q moments, predict the speed n rq+1 at the next moment, and thus obtain the average speed:

[0034]

[0035] Assume that the reference size of the sliding window is W b , the reference moving step is S b , and the reference speed is n rb . Then the current sliding window size and moving step are respectively:

[0036]

[0037] wherein, rand is the random integer function;

[0038] Step 3.2) Use the sliding window considering wind speed fluctuations for the normalized three-phase currents Continuous sampling is performed, and interpolation processing is used to make the lengths of all time-series samples consistent, obtaining a time-series dataset {TimeSeries = (T 1 , y 1 ), (T 2 , y 2 ), ··· (T x , y x ), ··· (T n , y n )}, where n is the number of samples, y x is the x-th sample label, x ∈ [1, n]; T x is the x-th sample data, and {T x = x 1 , x 2 , ··· x t , ··· x N}, where N is the sequence length, t is the time point, t ∈ [1, N], and x t is the current value at time t;

[0039] Step 3.3) Convert the three-phase current time-series dataset TimeSeries into an image dataset {Image = (M 1 , y 1 ), (M 2 , y 2 ), ··· (M x , y x ), ··· (M n , y n )} through the Markov transition field MTF, where M x is the Markov transition field obtained by MTF transformation of T x . The process of performing MTF transformation on any time-series sample T x into M x is as follows:

[0040] 1) Quantile sequence partitioning of sample T x

[0041] Divide the sample T x into D discrete distribution quantile units according to the value distribution, marked with the quantile d j , d j ∈ [1, D]. According to the division of each unit, map each time-domain value x t to the corresponding quantile unit d j , and convert the time-series sample into a quantile sequence {d j1 , d j2 ,..., d jD} represented by quantiles;

[0042] ​2) Construct the sample T x 's transition probability matrix P x

[0043] With the help of the single-step transition probability and multi-step transition probability defined in the Markov chain, that is, Equation (10), construct the sample T x 's Markov transition probability matrix P x ;

[0044]

[0045] In the formula, P i,i-1 is the single-step transition probability, indicating the probability that an element in the quantile region d i-1 at time t - 1 transfers to the quantile region d i at the next time t. P i,j is the multi-step transition probability, indicating the probability that an element in the quantile region d j at time t - 1 transfers to the quantile region d i at the next time t, where the gap between i and j is greater than 1; by statistically analyzing the transition situation of the quantile sequence at each time, the specific transition probability is obtained, and all the transition probabilities of the Markov chain are arranged along the transition rule to construct the sample T x 's D×D Markov transition probability matrix P x , as shown in Equation (11):

[0046]

[0047] 3) Construct the Markov transition field M x of the sample T x

[0048] Based on P x , arrange the samples according to the time series rule to obtain the Markov transition field of the sample T x as follows:

[0049]

[0050] Among them, p i,j represents the transition probability of the corresponding quantile relationship of d i , d j on the matrix P x , 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 an overall dataset {Data = (TimeSeries, Image) = [(T 1 , M 1 ), y1 , [(T 2 , M 2 ), y 2 , ··· [(T x , M x ), y x , ··· [(T n , M n ), y n}, where n is the number of samples.

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

[0053] Step 4.1) First, clean and shuffle the dataset Data, remove abnormal samples by calculating the standard deviation, randomly select 80% as the training set, and 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 unit GRU as branch 1 to process the three-phase current time series data, introduce skip connections at the middle position of the normal gated recurrent neural network chain structure, map the intermediate hidden state to the same dimension as the final hidden state. Suppose there are L gated recurrent units in total, then the hidden state of the current sequence at the middle position is denoted as h [L / 2] , and at the same time, introduce a fully connected layer in the skip connection, 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 recurrent unit is denoted as h L , then the final hidden state output by the improved recurrent neural network is h gru = h L + s(wh [L / 2] + b), and h gru contains the non-linear transformation of h [L / 2] ;

[0057] Use the residual convolutional neural network ResCNN as branch 2 to process the three-phase current MTF transformed image. For the input MTF image M and convolutional kernel K, the output MTF feature map Y is expressed as:

[0058]

[0059] In the formula, i and j are the positions of the output feature map, m and l are the sizes of the convolutional kernel, ω ij and b ijThey are the corresponding weight and offset coefficients respectively. After the convolutional layer, the activation function ReLU(x) = max(0, x) is connected. The batch normalization (BN) layer is placed after the convolutional layer or fully connected layer and before the activation function to normalize the activation values in each batch; the convolutional branch of the diagnostic model passes through the adaptive average pooling layer to adjust the final feature map Y' to the specified output size. Assume the size of Y' is H in ×W in , and the expected output size is H out ×W out , then the expected MTF feature map Y” is calculated by 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' respectively, i', j' are the positions of input_region, and N Y is the number of elements in input_region, and i, j are the positions of the expected feature map Y”;

[0062] Use the attention mechanism encoder to fuse the output features of branch 1 and branch 2, and extract the final feature structure for fault classification;

[0063] Step 4.3) Input the training set and the validation set into the established combined 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 other groups of equivalent wind speed data, and repeat steps S2 - S3 to obtain the test data set;

[0066] Step 5.2) Evaluate the diagnostic accuracy of the model saved in S4 through the test set. If the evaluation result shows that the model has not reached the preset best effect, adjust the hyperparameters of the neural network, and return to step S4 to retrain the model. Repeat this process until the diagnostic accuracy of the model on the test set reaches the best effect;

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

[0068] Further, in step S6, when the unit is actually operating, the three-phase current signals are subjected to real-time amplitude normalization and sliding window sampling according to the methods in steps S2 - S3 to obtain sequence samples. In the background analysis system, an MTF image is generated using the MTF transformation module. The sequence samples and the MTF image are input into the above-trained optimal diagnosis model to perform real-time monitoring and diagnosis on the abnormal state of the unit's converter power tubes.

[0069] Beneficial effects: The present invention discloses a robust diagnosis method for open-circuit faults of a converter in a wind power system considering wind speed fluctuations. Using the three-phase output current of the converter as the diagnosis signal, firstly, Park transformation is innovatively used to estimate and correct the real-time amplitude of the three-phase current for normalizing the three-phase current. Then, the normalized three-phase current signals are used as the original signals, and a dataset is made through a sliding window sampling method considering wind speed fluctuations. One path transforms it into a three-channel two-dimensional matrix through Markov transfer field (MTF) and sends it into the residual convolutional neural network branch, and the other path passes through the improved recurrent neural network branch. Finally, the outputs of the two branches are extracted into two-dimensional features and fed into the attention mechanism encoder for feature fusion, and finally, fault recognition is performed through the softmax layer. By normalizing the current signal through the real-time current amplitude and using a sliding window sampling method considering wind speed fluctuations to make the dataset, the influence of wind speed fluctuations on fault diagnosis is effectively solved. At the same time, this combined structure can utilize the super strong two-dimensional feature learning ability of the convolutional neural network, the time series processing ability that the recurrent neural network is good at, and the feature fusion ability of the attention mechanism to fully extract the feature information contained in the three-phase current signals of the converter, with high detection accuracy, and can effectively improve the operation reliability of the wind turbine generator set. Description of the Drawings

[0070] Figure 1 is the schematic diagram of a permanent magnet direct drive wind power system;

[0071] Figure 2 is the framework diagram of the abnormal state detection method for the converter of the wind turbine generator set;

[0072] Figure 3 is the typical wind speed curve and the equivalent wind speed curve diagram.

[0073] Figure 4 is the comparison diagram of the normalization effect between the real-time current amplitude normalization method and other methods;

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

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

[0076] Figure 7It is the accuracy and training loss graph of the combined neural network model. Detailed implementation manners

[0077] The present invention will be described in detail below in combination with specific embodiments and the accompanying drawings.

[0078] This embodiment is based on a robust diagnosis method for open - circuit faults of a converter in a wind power system considering wind speed fluctuations: The schematic diagram of a permanent - magnet direct - drive wind power system is as Figure 1 shown, which is composed of a wind turbine, a permanent - magnet synchronous generator, a machine - side rectifier, a grid - side inverter, etc. Among them, the machine - side rectifier and the grid - side inverter form a back - to - back topology structure, which can realize bidirectional energy flow and convert the electric energy with variable frequency and amplitude generated by the generator into alternating current that meets the grid - connection requirements of frequency and amplitude and send it into the grid. The framework diagram of the abnormal state detection method for the converter of a wind turbine is as Figure 2 shown. First, three groups of equivalent wind speed data at the hub height are established. Then, according to these data, three - phase current signals output by the converter are obtained through simulation and experiments, and real - time amplitude normalization processing is performed on the current signals. Subsequently, the normalized three - phase current signals are used as the original signals, and a data set is made through a sliding - window sampling method considering wind speed fluctuations. Then, the made data set is used to construct and train a model. After the model training is completed, the performance of the model is further optimized by adjusting the model to improve the accuracy of diagnosis. Finally, the trained model is applied to the actual scenario for real - time state detection and fault diagnosis.

[0079] The present invention proposes a robust diagnosis method for open - circuit faults of a converter in a wind power system considering wind speed fluctuations, and the method includes the following steps:

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

[0081] S2. Based on the selected wind speed data, obtain three - phase current signals output by the converter through simulation and experiments, and perform pre - processing on the three - phase current using the real - time amplitude normalization method;

[0082] S3. Continuously sample the pre - processed three - phase current data through a sliding window considering wind speed fluctuations to obtain a time - series data set, convert the time - series data set using a Markov transfer field to obtain an image data set, and combine the time - series data set and the image data set into an overall data set;

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

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

[0085] S6. When the unit is actually operating, use the optimal diagnostic model to monitor and diagnose the abnormal state of the unit's converter power tubes in real time.

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

[0087] Step 1.1) Measure the actual wind speed at the hub height in the operating environment of the wind turbine, and screen out three groups of wind speed data that can reflect the typical change pattern of the wind speed in this area, such as Figure 3 the three groups of wind speed data shown in (a) 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 the wind shear effect and tower shadow effect, establish an equivalent wind speed model for the wind speed data V w The expression is:

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

[0090] where V eq is the equivalent wind speed, V ws represents the wind shear component, V ts represents the tower shadow effect component, and the units are all m / s. V ws and V ts are calculated by equations (2) and (3) respectively:

[0091]

[0092] In the formula, R is the blade radius (m); α is the wind shear exponent, with a value between 0.1 and 0.7; H is the hub height (m); ω m represents the impeller angular frequency (rad / s), which is obtained through the wind speed data V w and wind turbine simulation; β i is the blade azimuth angle (rad), and the subscript i is the blade number; D t is the tower radius (m); d represents the distance from the impeller plane to the tower center line (m); the equivalent wind speed data is obtained through V w and formulas (1) - (3):

[0093]

[0094] Step 1.3) Randomly select the first group of wind speed data {V eq from V r = v′ 11 , v′ 12 , ··· v′ 1num}, Figure 3 (b) is the specific waveform of V r . It can be seen that it contains obvious fluctuations compared with the wind speed wind1 before equivalence.

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

[0096] Step 2.1) Use Simulink to establish a wind turbine simulation model, and use a permanent magnet synchronous motor to simulate the wind power system on a back-to-back test platform. Conduct simulations and experiments on the converter health, single-tube fault, and double-tube fault. Input the selected wind speed data into the simulation model and the test platform, and measure the three-phase current signal i abc as shown in Figure 4 (a); when modeling the wind speed data, it is V r , and when adjusting the parameters, it is the remaining two groups;

[0097] Step 2.2) Perform real-time normalization preprocessing on the three-phase current to eliminate the influence of current fluctuations caused by wind speed fluctuations on fault diagnosis. When the converter is operating normally, the three-phase current is in a sinusoidal balanced state, and the expression is:

[0098]

[0099] In the formula, i a , i b , i c are the three-phase currents, I m is the real-time amplitude of the current, ω is the angular frequency, θ is the initial phase angle. Formula (4) shows that using the Park transformation to obtain the real-time amplitude I m can perform normalization preprocessing on the current. The expression of I m is as follows:

[0100]

[0101] In the formula, i d , i q are the direct-axis and quadrature-axis components obtained by the Park transformation of the three-phase currents respectively. Then the normalized three-phase currents are as shown in Figure 4 (b) is the waveform of the three-phase current signal i abc after normalization by the real-time current amplitude, Figure 4 (c) is the three-phase current signal i abcThe waveform normalized by the MinMaxScaler method shows that real-time amplitude normalization can effectively eliminate the influence of wind speed fluctuations on the change of current amplitude.

[0102] Further, the specific method of step S3 is as follows:

[0103] Step 3.1) Use the generator speed autoregressive model to establish a sliding window considering wind speed fluctuations, eliminate the influence of speed fluctuations caused by wind speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment. The formula is as follows:

[0104]

[0105] In the formula, n rt is the speed value at time t, c is a constant, is the model parameter, q is the model order, and ε is the error term; the autoregressive model parameters are obtained by fitting the historical speed data combined with the least squares method. According to the speeds n r1 , n r2 …, n rq at the previous q moments, predict the speed n rq+1 at the next moment, and thus obtain the average speed:

[0106]

[0107] Let the reference size of the sliding window be W b , the reference moving step be S b , and the reference speed be n rb . Then the current sliding window size and moving step are respectively:

[0108]

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

[0110] Step 3.2) Use the sliding window considering wind speed fluctuations to continuously sample the normalized three-phase current and make all time series sample lengths consistent through interpolation processing to obtain the time series dataset {TimeSeries = (T 1 , y 1 ), (T 2 , y 2 ), ·(T x , y x ), ·(T n , y n )}, where n is the number of samples, y x is the label of the x-th sample, x ∈ [1, n]; T x is the data of the x-th sample, and there is {T x = x1 , x 2 , ·x t , ·x N}, where N is the sequence length, t is the time point, t ∈ [1, N], and x t is the current value at time t;

[0111] Step 3.3) Convert the three-phase current time series dataset TimeSeries into an image dataset {Image = (M 1 , y 1 ), (M 2 , y 2 ), ··· (M x , y x ), ··· (M n , y n )} through the Markov transfer field MTF, where M x is the Markov transfer field obtained by MTF transformation. The process of performing MTF transformation on any time series sample T x into M x is as follows: x

[0112] 1) Quantile sequence division of sample T x

[0113] Divide the sample T x into D discrete distribution quantile units according to the value distribution, and use the quantile d j to mark, d j ∈ [1, D]. According to the division of each unit, map each time domain value x t to the corresponding quantile unit d j to convert the time series sample into a quantile sequence {d j1 , d j2 ,..., d jD} represented by quantiles;

[0114]

[0115] 2) Construct the transition probability matrix P x of sample T x

[0116] With the help of the single-step transition probability and multi-step transition probability defined in the Markov chain, that is, Equation (10), construct the Markov transition probability matrix P x of sample T x ;

[0116]

[0117] In the formula, P i,i-1 is the single-step transition probability, indicating that at time t - 1, it is in the quantile region d i-1The element at the next moment t transfers to the quantile region d i with probability P i,j which is the multi-step transition probability, indicating that the element in the quantile region d j at time t-1 transfers to the quantile region d i at the next moment t, where the gap between i and j is greater than 1; by statistically analyzing the transitions of the quantile sequence at each moment, the specific transition probability is obtained, and all the transition probabilities of the Markov chain are arranged along the transition rule to construct the sample T x The D×D Markov transition probability matrix P x is as shown in Equation (11):

[0118]

[0119] 3) Construct the sample T x The Markov transition field M x

[0120] Based on P x the samples are arranged according to the time sequence rule to obtain the sample T x The Markov transition field is as follows:

[0121]

[0122] where p i,j represents the transition probability of the d i , d j corresponding quantile relationship in the matrix P x , and the diagonal elements are the corresponding self-transition probabilities. For the time series sample T x of the current signal, the MTF transformation process is as Figure 5 shown, and the time series sample is transformed into the Markov transition field M x and then displayed as an RGB image.

[0123] Step 3.4) Combine the time series dataset TimeSeries and the image dataset Image into an overall dataset {Data=(TimeSeries,Image)=[(T 1 , M 1 ), y 1 , [(T 2 , M 2 ), y 2 , ···[(T x , M x ), y x , ···[(T n , M n ), y n}, where n is the number of samples.

[0124] Further, the specific method of step S4 is as follows:

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

[0126] Step 4.2) Establish a combined neural network model based on the improved GRU-ResCNN-Atnn. The specific structure is as Figure 6 shown. After the improved GRU and ResCNN process samples in parallel, information is fused through the 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 to map the middle hidden state to the same dimension as the final hidden state. Suppose there are L gated recurrent units in total, then the hidden state of the current sequence at the middle position is denoted as h [L / 2] At the same time, a fully connected layer is introduced in the skip connection, 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 recurrent unit is denoted as h L Then the final hidden state output by the improved recurrent neural network is h gru = h L + s(wh [L / 2] + b), and h gru contains the non-linear transformation of h [L / 2] ;

[0129] The residual convolutional neural network ResCNN is used as branch 2 to process the three-phase current MTF transformation image. For the input MTF image M and convolutional kernel K, the output MTF feature map Y is expressed as:

[0130]

[0131] In the formula, i and j are the positions of the output feature map, m and l are the sizes of the convolutional kernel, ω ij and b ijThey are the corresponding weight and offset coefficient respectively. After the convolutional layer, the activation function ReLU(x) = max(0, x) is connected. The batch normalization (BN) layer is placed after the convolutional layer or fully connected layer and before the activation function to normalize the activation values in each batch; the convolutional branch of the diagnostic model passes through the adaptive average pooling layer to adjust the final feature map Y' to the specified output size. Assume the size of Y' is H in ×W in and the expected output size is H out ×W out , then the expected MTF feature map Y” is calculated by 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' respectively, i', j' are the positions of input_region, N Y is the number of elements in input_region, and i, j are the positions of the expected feature map Y”;

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

[0135] Step 4.3) Input the training set and the validation set into the established combined 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 other groups of equivalent wind speed data, and repeat steps S2 - S3 to obtain the test data set;

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

[0139] Step 5.3) Save the best diagnostic model for online diagnosis during actual operation. Its accuracy curves and loss curves on the training set and test set are respectively as Figure 7 (a) and 7(b) shown.

[0140] Further, in the step S6, when the unit is actually operating, the three-phase current signals are subjected to real-time amplitude normalization and sliding window sampling according to the method in steps S2 - S3 to obtain sequence samples. In the background analysis system, an MTF image is generated by using the MTF transformation module. The sequence samples and the MTF image are input into the above-mentioned trained optimal diagnosis model to perform real-time monitoring and diagnosis on the abnormal state of the unit's converter power tubes.

[0141] Table 1 Main hyperparameters of the model

[0142]

Claims

1. A robust diagnosis method for open-circuit faults of wind power system converters considering wind speed fluctuations, characterized in that: The method comprises the following steps: S1, establish three sets of equivalent wind speed data at hub height, and randomly select one set for modeling; S2, obtaining the three-phase current signal output by the converter through simulation and experiment based on the selected wind speed data, and preprocessing the three-phase current using a real-time amplitude normalization method; S3, continuously sampling the preprocessed three-phase current data through a sliding window considering wind speed fluctuations to obtain a time series data set, transforming the time series data set using a Markov transition field to obtain an image data set, and combining the time series data set with the image data set into an overall data set; S4, after cleaning and shuffling the data set, divide it into a training set and a validation set according to a predetermined ratio, input the training set and the validation set into a combined neural network model based on the improved GRU-ResCNN-Atnn for synchronous training, and save the trained model; S5, selecting the remaining two sets of equivalent wind speed data, obtaining a test data set according to steps S2-S3, applying the test data set to the saved model after cleaning, and adjusting the model parameters accordingly to establish the optimal diagnostic model; S6: When the unit is actually running, the abnormal state of the unit's converter power tube is monitored and diagnosed in real time through the optimal diagnostic model.

2. A robust diagnosis method for open-circuit fault of a wind power system converter considering wind speed fluctuation according to claim 1, characterized in that: The specific method of step S1 is as follows: 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 pattern in the area, including stable wind speed, fluctuating wind speed and gusts. Among them, 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; Step 1.2) Considering wind shear effect and tower effect, establish wind speed data V w The equivalent wind speed model is expressed as: V eq =V w +V ws +V ts (1) Among them, V eq is the equivalent wind speed, V ws represents the wind shear component, V ts Represents the tower shadow effect component, the unit is m / s, V ws and V ts Calculated by formula (2) and formula (3) respectively: Where R is the blade radius (m); α is the wind shear index, which ranges from 0.1 to 0.7; H is the hub height (m); ω m represents the impeller angular frequency (rad / s), through the wind speed data V w and wind turbine simulation acquisition; β i is the blade azimuth (rad), subscript i is the blade number; D t is the tower radius (m); d is the distance from the impeller plane to the tower centerline (m); V w And formula (1)-(3) to get the equivalent wind speed data: Step 1.3) From V eq A set of wind speed data {V r =v i '1,v i '2,…v i ' num }, where i represents a random integer value between 1 and 3.

3. A robust diagnosis method for open-circuit fault of a wind power system converter considering wind speed fluctuation according to claim 2, characterized in that: The specific method of step S2 is as follows: Step 2.1) Use Simulink to build a wind turbine simulation model, and use the permanent magnet synchronous motor tow test platform to simulate the wind power system, simulate and experiment on the converter health, single-tube fault and double-tube fault, input the selected wind speed data into the simulation model and the experimental platform, and measure the converter three-phase current signal i abc ; V for wind speed data modeling r , when adjusting the parameters, there are two groups left; Step 2.2) Perform real-time normalization preprocessing on the three-phase current to eliminate the influence of current fluctuation caused by wind speed fluctuation on fault diagnosis. When the converter is operating normally, the three-phase current is in a sinusoidal balance state, and the expression is: In the formula, i a 、i b 、i c is the three-phase current, I m is the real-time current amplitude, ω is the angular frequency, θ is the initial phase angle, and formula (4) shows that the real-time amplitude I can be obtained by using Park transform. m The current can be normalized and preprocessed, I m The expression is as follows: In the formula, i d 、i q They are the direct-axis and quadrature-axis components of the three-phase current obtained by Park transformation, so the normalized three-phase current is 4. A robust diagnosis method for open-circuit fault of a wind power system converter considering wind speed fluctuation according to claim 3, characterized in that: The specific method of step S3 is as follows: Step 3.1) Use the generator speed autoregressive model to establish a sliding window that takes into account wind speed fluctuations to eliminate the impact of speed fluctuations caused by wind speed fluctuations on fault diagnosis. First, use the autoregressive model to predict the speed value at the next moment. The formula is as follows: Where n rt is the speed value at time t, c is a constant, is the model parameter, q is the model order, and ε is the error term. The autoregressive model parameters are obtained by fitting the historical speed data with the least square method, and the speed n at the previous q moments is r1 ,n r2 …,n rq , predict the speed n at the next moment rq+1 , thus the average speed is: Assume the sliding window base size is W b , the moving step length is S b , the speed reference is n rb , then the current sliding window size and moving step size are: In the formula, rand is a random integer function; Step 3.2) Use a sliding window that takes into account wind speed fluctuations to normalize the three-phase currents Continuous sampling is performed, and the length of all time series samples is made consistent through interpolation processing, and the time series data set {TimeSeries=(T1,y1),(T2,y2),·(T x ,y x ),·(T n ,y n )}, where n is the number of samples, y x is the x-th sample label, x∈[1,n]; T x is the xth sample data, and {T x =x1,x2,·x t ,·x N }, where N is the sequence length, t is the time point, t∈[1,N], x t is the current value at time t; Step 3.3) The three-phase current time series data set TimeSeries is converted into an image data set {Image=(M1,y1),(M2,y2),…(M x ,y x ),…(M n ,y n )}, where M x T x The Markov transition field obtained by MTF transformation is x Transform MTF to M x The process is as follows: 1) Sample T x Quantile sequence partitioning The sample T x Divide the distribution of values ​​into D discrete distribution quantile units, using quantile d j Mark, d j ∈[1,D], according to the division of each unit, each time domain value x t Mapped to the corresponding quantile unit d j On the other hand, the time series samples are converted into quantile sequences represented by quantiles {d j1 ,d j2 ,...,d jD }; 2) Construct sample T x The transition probability matrix P x With the help of the single-step transition probability and multi-step transition probability defined in the Markov chain, that is, formula (10), we construct the sample T x The Markov transition probability matrix P x ; Where P i,i-1 is the single-step transition probability, indicating that it is in the quantile region d at time t-1 i-1 The elements of are transferred to the quantile area d at the next moment t i The probability, P i,j is the multi-step transition probability, indicating that it is in the quantile region d at time t-1 j The elements of are transferred to the quantile area d at the next time t i The probability that the difference between i and j is greater than 1; by statistically analyzing the changes in the quantile sequence at each moment, the specific transition probability is obtained, and all the transition probabilities of the Markov chain are arranged along the transition law to construct the sample T x The D×D Markov transition probability matrix P x , as shown in formula (11): 3) Construct sample T x The Markov transition field M x In P x Based on the time sequence rule, the samples are arranged to obtain sample T x The Markov transition field of is as follows: Among them, p i,j Indicates d i , d j The corresponding quantile relationship is in the matrix P x The transition probability on the diagonal is the corresponding self-transition probability; Step 3.4) Combine the time series dataset TimeSeries and the image dataset Image into an overall dataset {Data = (TimeSeries, Image) = [(T1, M1), y1], [(T2, M2), y2], ... [(T x , M x ),y x ],…[(T n , M n ), y n ]}, n is the number of samples.

5. A robust diagnosis method for open-circuit fault of a wind power system converter considering wind speed fluctuation according to claim 4, characterized in that: The specific method of step S4 is as follows: Step 4.1) First, clean and shuffle the data set Data, remove abnormal samples by calculating the standard deviation, randomly select 80% as the training set, and 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 in 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 that there are L gated recurrent units in total, the hidden state of the current sequence in the middle position is recorded as h [L / 2] At the same time, a fully connected layer is introduced in the skip connection, and its formula is as follows: y=s(wx+b) (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, s is the ReLU function, and the hidden state of the output current sequence of the last gated recurrent unit is recorded 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 of The residual convolutional neural network ResCNN is used as branch 2 to process the three-phase current MTF transformation image. For the input MTF image M and convolution kernel K, the output MTF feature map Y is expressed as: Where i and j are the positions of the output feature maps, m and l are the sizes of the convolution kernels, and ω ij and b ij are the corresponding weights and offset coefficients respectively. The activation function ReLU(x)=max(0,x) is connected after the convolution layer. The batch normalization (BN) layer is placed after the convolution layer or the fully connected layer and before the activation function to normalize the activation values ​​in each batch. The convolution branch of the diagnostic model adjusts the final feature map Y' to the specified output size through the adaptive average pooling layer. Assume that the size of Y' is H in ×W in , the desired output size is H out ×W out , then the expected MTF characteristic graph Y" is calculated by the following formula: Where 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 input_region, and N Y is the number of elements in input_region, i and j are the positions of the expected feature map Y"; Use the attention mechanism encoder to fuse the output features of branch 1 and branch 2 to extract the final feature structure for fault classification; Step 4.3) Input the training set and the validation set into the established combined neural network model for synchronous training, and save the trained model.

6. The robustness diagnosis method for open-circuit fault of a wind power system converter considering wind speed fluctuation according to claim 5 is characterized in that: The specific method in step S5 is as follows: Step 5.1) Select another two sets of equivalent wind speed data and repeat steps S2-S3 to obtain a test data set; Step 5.2) Evaluate the diagnostic accuracy of the model saved in S4 through the test set. If the evaluation result shows that the model has not yet achieved the preset optimal effect, adjust the hyperparameters of the neural network, return to step S4 to retrain the model, and repeat this process until the diagnostic accuracy of the model on the test set reaches the optimal effect; Step 5.3) Save the best diagnosis model for online diagnosis in actual operation.

7. The method for robust diagnosis of open-circuit faults of converters in wind power systems considering wind speed fluctuations according to claim 6 is characterized in that: In step S6, when the unit is actually running, the three-phase current signal is amplitude normalized and sampled with a sliding window in real time according to the method in steps S2-S3 to obtain a sequence sample, and in the background analysis system, an MTF transformation module is used to generate an MTF image, and the sequence sample and the MTF image are input into the above-mentioned trained optimal diagnosis model to perform real-time monitoring and diagnosis on the abnormal state of the power tube of the unit converter.

Images