Method for detecting abnormal state of converter of wind turbine generator based on GAF-CNN-Transform

Abstract

The invention discloses a GAF-CNN-Transform-based wind turbine generator converter abnormal state detection method. The method comprises the following steps: S1, continuously sampling three-phase current signals of a wind power system machine-side converter and a grid-side converter through a sliding window to obtain a time sequence data set; s2, preprocessing the time sequence data set, and converting one-dimensional time domain information into two-dimensional graphic information by using a GAF (Gray Angle Field); and S3, establishing a CNN-Transform neural network model, and carrying out abnormal state detection and classification on the wind power system converter through a two-dimensional image data training model. According to the method, the current signals are converted into the image data, dependency information between the data can be fully mined through the CNN-Transform model, the detection precision is high, and the operation reliability of the wind turbine generator can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind power system fault inspection and repair, and in particular relates to a method for detecting abnormal state of a wind turbine converter based on GAF-CNN-Transformer. Background Art

[0002] In recent years, the main field of wind power generation is shifting from onshore to offshore, and from nearshore to offshore. Thanks to the advantages of permanent magnet direct drive technology, high efficiency, good reliability and low maintenance cost, most offshore wind turbines are permanent magnet direct drive. In the electrical components of wind turbines, power electronics technology is a key technology for energy conversion and plays an important role in connecting generators and power grids. Its reliability issues have become a research hotspot in academia and industry. The reliability of converters is crucial in wind power systems. Therefore, studying the fault diagnosis method of power devices in power converters and handling them in a timely manner is of great significance to improving the reliability of permanent magnet direct drive wind power systems.

[0003] Although there are many ways to improve the reliability of power electronic converters, permanent magnet direct-drive wind turbines are usually located in coastal or even offshore areas, and the nacelles are usually located at an altitude of tens of meters. The converters of the units face complex environmental conditions, such as high temperature, high humidity, salt spray corrosion, lightning, typhoons and other severe weather conditions, and failures are inevitable. In 2007, the Egmond aan Zee offshore wind farm in the Netherlands had a high outage rate due to electrical system failures (converters, control systems, etc.), resulting in huge economic losses.

[0004] Wind power converters are mainly composed of power switch tubes, which are the most vulnerable link in the converter. Usually, the main forms of power switch tube failure are short circuit failure and open circuit failure. Short circuit failure is highly destructive. The method of implementing switch tube short circuit failure diagnosis and protection based on software algorithm is not reasonable and difficult to implement. It is generally completed by standard hardware short circuit protection circuit. The open circuit failure of the switch tube will not cause serious overcurrent or overvoltage in a short time. It can last for a period of time and will not trigger system protection. Therefore, it will not cause the system to shut down immediately in a short time. However, under harsh working conditions or long-term operation with faults, the heat and loss of the switch device will increase. The open circuit failure will cause secondary failures of other components in the wind power system, and even cause catastrophic failures, resulting in higher maintenance costs.

[0005] Therefore, in order to improve the operational reliability of wind power systems, this paper proposes a wind turbine converter abnormal state detection method based on GAF-CNN-Transformer to perform state detection and fault diagnosis on wind turbine converters. Summary of the invention

[0006] In order to solve the above problems, a wind turbine converter abnormal state detection method based on GAF-CNN-Transformer is proposed to improve the operation reliability of offshore wind farms.

[0007] The present invention is specifically a method for detecting abnormal state of a wind turbine converter based on GAF-CNN-Transformer. The power system state assessment method comprises the following steps:

[0008] S1: Continuously sample the three-phase current signals of the wind power system generator-side and grid-side converters through a sliding window to obtain a time series data set;

[0009] S2: Preprocess the time series data set and use the Gram Angle Field (GAF) to convert the one-dimensional time domain information into two-dimensional graphic information.

[0010] S3: Establish a CNN-Transformer neural network model and train the model through two-dimensional graphic data to detect and classify abnormal states of wind power system converters.

[0011] Furthermore, in step S1, the sliding window size N needs to satisfy the following formula:

[0012]

[0013] Where: T is the three-phase current period, T s is the current sensor sampling period. Since the operating frequencies of the generator-side converter and the grid-side converter are different, the sliding window sizes on the generator side and the grid side are also different.

[0014] The sliding window moving step size adopts random step size to improve the diversity of data, thereby enhancing the generalization ability of the model and reducing overfitting. The random step size formula is:

[0015]

[0016] Where: rand is a random integer function, and the return value is a random integer in the interval.

[0017] Furthermore, in step S2, the time series data set extracted by the sliding window is first cleaned and normalized. Data cleaning is performed by first-order low-pass filtering, and the calculation formula is:

[0018] Y(n)=αX(n)+(1-α)Y(n-1) (3)

[0019] Where: ɑ is the filter coefficient; X(n) is the current sampling value; Y(n-1) is the previous filter output value; Y(n) is the current filter output value.

[0020] The cleaned data is normalized by Z-Score, which can convert data of different magnitudes into the same magnitude and measure them uniformly with the calculated Z-score value to ensure comparability between data. n Perform Z-Score standardization, and the calculation formula is:

[0021]

[0022] Where: is the sample mean, S is the sample standard deviation, The newly generated sequence value.

[0023] First, the cleaned and normalized data is mapped to polar coordinates, and the calculation formula is shown in Equation 5.

[0024]

[0025] Where: t i is the timestamp; m means dividing the unit length of polar coordinates into m equal parts. This method of converting time series into polar coordinates has two advantages. First, the transformation is a bijection. There is a one-to-one correspondence with φ. Second, the transformation retains the time information, and the time value can be determined through the radius coordinate.

[0026] After obtaining the data information in polar coordinate form, the Gram angular field (GAF) matrix can be calculated using Equation 6.

[0027]

[0028] The three-phase current data calculated by GAF is three sets of two-dimensional matrices with a shape of (3, 64, 64), and the value range of each element is between (-1, 1). In order to facilitate the visualization of image data, the data is mapped to the interval (0, 1), corresponding to the RGB color system. The mapping function is shown in Equation 7.

[0029]

[0030] Furthermore, in step S3, the neural network model adopts a CNN-Transformer serial structure, and the working state of the wind power converter is detected by training the data set. The model loss function adopts the cross-entropy loss function (Cross-entropy Loss), the optimization algorithm adopts Adam, and uses the Relu activation function.

[0031] Different blocks of the CNN network are connected through a residual structure to build a deep model to fully extract feature information, and the generalization ability of the model is improved by adding a batch normalization layer. The output shape of the CNN network is (1, 512), which is transformed into a (16, 32) matrix in groups of 32 data as the input of the Transformer network.

[0032] The Transformer model uses a self-attention mechanism for information interaction and transmission. This mechanism allows the model to focus on information at different positions when processing sequence data, and can effectively capture the global information of the input data. Since the calculation of this attention mechanism can be performed in parallel, the Transformer model has extremely high computational efficiency and parallel processing capabilities, and the model can learn rich data features in a short time.

[0033] The output of the Transformer first passes through the dropout layer to avoid overfitting of the model, and then passes through the softmax layer to output the final result of the converter state detection.

[0034] Compared with the prior art, the beneficial effects are as follows: the present invention discloses a method for detecting abnormal state of wind turbine converter based on GAF-CNN-Transformer. Taking the three-phase output current of the converter as the diagnostic signal, the diagnostic signal is firstly sampled continuously through a sliding window to obtain a time series data set, followed by data cleaning and normalization operations, and a two-dimensional image data set is constructed using GAF, and finally a CNN-Transformer model is built to train, identify and classify the image data set, and realize open circuit fault detection of the power tube module of the converter of the wind power system. The current signal is converted into image data, and the dependency information between the data can be fully mined through the CNN-Transformer model, which has high detection accuracy and can effectively improve the operation reliability of the wind turbine. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the schematic diagram of the permanent magnet direct drive wind power system;

[0036] Figure 2 The framework diagram of the abnormal state detection method of the wind turbine converter is shown in FIG.

[0037] Figure 3 This is the waveform diagram of the partial open circuit fault of the wind power system converter;

[0038] Figure 4 This is the structure diagram of the CNN-Transformer neural network model;

[0039] Figure 5 Figure 2 is the accuracy and training loss graph of the CNN-Transformer neural network model. DETAILED DESCRIPTION

[0040] The present invention is described in detail below with reference to specific embodiments and accompanying drawings.

[0041] This embodiment is based on the GAF-CNN-Transformer wind turbine converter abnormal state detection method: The permanent magnet direct drive wind power system schematic diagram is as follows Figure 1 As shown in the figure, the three-phase output current of the converter is used as the diagnostic signal. First, the diagnostic signal is continuously sampled through a sliding window to obtain a time series data set, then data cleaning and normalization operations are performed, and a two-dimensional image data set is constructed using GAF. Finally, a CNN-Transformer model is built to train, identify and classify the image data set, and the open circuit fault detection of the power tube module of the wind power system converter is realized. The detection method framework is shown in Figure 2 shown.

[0042] The specific steps are as follows.

[0043] S1: The three-phase current signals of the wind power system machine-side and grid-side converters are continuously sampled and annotated through a sliding window to obtain a time series data set.

[0044] S2: Preprocess the time series data set and use the Gram Angle Field (GAF) to convert the one-dimensional time domain information into two-dimensional graphic information.

[0045] S3: Establish a CNN-Transformer neural network model and train the model through two-dimensional image data to detect and classify abnormal states of wind power system converters.

[0046] In step S1, the sliding window size N must satisfy the following formula:

[0047]

[0048] Where: T is the three-phase current period, T s is the current sensor sampling period. Since the operating frequencies of the generator-side converter and the grid-side converter are different, the sliding window sizes on the generator side and the grid side are also different.

[0049] The sliding window moving step size adopts random step size to improve the diversity of data, thereby enhancing the generalization ability of the model and reducing overfitting. The random step size formula is:

[0050]

[0051] Where: rand is a random integer function, and the return value is a random integer in the interval.

[0052] In the embodiment, the operating conditions of the converter are divided into 22 states, including normal state, single open circuit fault and double open circuit fault, and the specific division is shown in Table 1. Each operating condition is simulated in Simulink, the simulation time is set to 50s, the sampling frequency of the three-phase current signal is set to 1000Hz, and 1000 samples under each operating condition are obtained by continuous sampling through the sliding window. Each sample contains 3*1000 points of data, and the 22 operating conditions have a total of 22000 samples. The waveforms of the three-phase current samples under some operating conditions are as follows: Figure 3 shown.

[0053] Table 1 22 operating conditions of converter

[0054]

[0055] In step S2, firstly, the time series data set extracted by the sliding window is cleaned and normalized. Data cleaning is performed by first-order low-pass filtering, and the calculation formula is:

[0056] Y(n)=αX(n)+(1-α)Y(n-1) (3)

[0057] Where: ɑ is the filter coefficient; X(n) is the current sampling value; Y(n-1) is the previous filter output value; Y(n) is the current filter output value.

[0058] The cleaned data is normalized by Z-Score, which can convert data of different magnitudes into the same magnitude and measure them uniformly with the calculated Z-score value to ensure comparability between data. n Perform Z-Score standardization, and the calculation formula is:

[0059]

[0060] Where: is the sample mean, S is the sample standard deviation, The newly generated sequence value.

[0061] The cleaned and normalized dataset is then converted into a two-dimensional image dataset through the Gram Angular Field. The Gram Angular Field (GAF) uses trigonometric functions and angular formulas to define the correlation between each time point.

[0062] First, the sequence data obtained after cleaning and normalization Mapped to polar coordinates, the calculation formula is as follows:

[0063]

[0064] Where: t iis the timestamp; m means dividing the unit length of polar coordinates into m equal parts. This method of converting time series into polar coordinates has two advantages. First, the transformation is a bijection. There is a one-to-one correspondence with the angle φ. Second, the transformation retains the time information, and the time value can be determined through the radius coordinate.

[0065] After obtaining the data information in polar coordinate form, the Gram angular field (GAF) matrix can be calculated by the following formula.

[0066]

[0067] Where: φ i (i=1, 2, ..., n) is the angle of the sequence data point on the polar coordinates obtained after the i-th cleaning and normalization; n is the total number of sequence data points.

[0068] The three-phase current data is calculated by GAF to obtain three sets of two-dimensional matrices with a shape of (3, 64, 64), and the value range of each element is between (-1, 1). In order to facilitate the visualization of image data, the data is mapped to the interval (0, 1), corresponding to the RGB color system, and the final image data set is obtained. The mapping function is as follows:

[0069]

[0070] In step S3, the neural network model adopts a CNN-Transformer serial structure, and detects the working state of the wind power converter by training the image data set. The model loss function adopts the cross-entropy loss function (Cross-entropyLoss), the optimization algorithm adopts Adam, and uses the Relu activation function.

[0071] Different blocks of the CNN network are connected through a residual structure to build a deep model to fully extract feature information, and the generalization ability of the model is improved by adding a batch normalization layer. The output shape of the CNN network is (1, 512), which is transformed into a (16, 32) matrix in groups of 32 data as the input of the Transformer network.

[0072] The Transformer model uses a self-attention mechanism for information interaction and transmission. This mechanism allows the model to focus on information at different positions when processing sequence data, and can effectively capture the global information of the input data. Since the calculation of this attention mechanism can be performed in parallel, the Transformer model has extremely high computational efficiency and parallel processing capabilities, and the model can learn rich data features in a short time.

[0073] The output of the Transformer model first passes through a dropout layer to avoid overfitting of the model, and then passes through a softmax layer to output the final result of the converter state detection.

[0074] The CNN-Transformer model structure is as follows Figure 4 As shown, the accuracy and training loss of the model are Figure 5 shown.

[0075] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not intended to limit the present invention. A person skilled in the art should understand that the specific implementation of the present invention can be modified or replaced by equivalents, but these modifications or changes are within the scope of protection of the pending claims.

Claims

1. A wind turbine converter abnormal state detection method based on GAF-CNN-Transformer, characterized in that: The following steps are involved: S1: Continuously sample the three-phase current signals of the wind power system generator-side and grid-side converters through a sliding window to obtain a time series data set; S2: Preprocess the time series data set and use the Gram Angular Field (GAF) to convert the one-dimensional time domain information into two-dimensional image information. S3: Establish a CNN-Transformer neural network model and train the model through two-dimensional image data to detect and classify abnormal states of wind power system converters.

2. According to claim 1, a method for detecting abnormal state of wind turbine converter based on GAF-CNN-Transformer is characterized in that: In step S1, the sliding window size N needs to satisfy the following formula: Where: T is the three-phase current period, T s is the current sensor sampling period. Since the operating frequencies of the generator-side converter and the grid-side converter are different, the sliding window sizes on the generator side and the grid side are also different. The sliding window moving step size adopts a random step size to improve the diversity of data, thereby enhancing the generalization ability of the model and reducing overfitting; the random step size formula is: Where: rand is a random integer function, and the return value is a random integer in the interval.

3. A wind turbine converter abnormal state detection method based on GAF-CNN-Transformer according to claim 2, characterized in that: In step S2, firstly, data cleaning and normalization operations are performed on the time series data set extracted by the sliding window; data cleaning is performed by first-order low-pass filtering, and the calculation formula is: Y(n)=αX(n)+(1-α)Y(n-1) (3) Where: ɑ is the filter coefficient; X(n) is the current sampling value; Y(n-1) is the last filter output value; Y(n) is the current filter output value; The cleaned data is normalized by Z-Score, which converts data of different magnitudes into the same magnitude and uses the calculated Z-score value to ensure comparability between data. n Perform Z-Score standardization, and the calculation formula is: Where: is the sample mean, S is the sample standard deviation, is the newly generated sequence value; The cleaned and normalized dataset is then converted into a two-dimensional image dataset through the Gram Angular Field. The Gram Angular Field (GAF) uses trigonometric functions and angular formulas to define the correlation between each time point; First, the sequence data obtained after cleaning and normalization Mapped into polar coordinates, the calculation formula is shown in formula (5). Where: t i is the timestamp; m means dividing the unit length of the polar coordinate into m equal parts; After obtaining the data information in polar coordinate form, the Gram angular field (GAF) matrix is ​​calculated by equation 6: Where: φ i (i=1, 2, ..., n) is the angle of the i-th sequence data point obtained after cleaning and normalization in polar coordinates; n is the total number of sequence data points; The three-phase current data calculated by GAF is three sets of two-dimensional matrices with a shape of (3, 64, 64), and the value range of each element is between (-1, 1). In order to facilitate the visualization of image data, the data is mapped to the (0, 1) interval, corresponding to the RGB color system; the mapping function is shown in Formula 7:

4. A method for detecting abnormal state of wind turbine converter based on GAF-CNN-Transformer according to claim 3, characterized in that: In the step S3, the neural network model adopts a CNN-Transformer serial structure, and the working state of the wind power converter is detected by training the data set; the model loss function adopts a cross-entropy loss function (Cross-entropy Loss), the optimization algorithm adopts Adam, and uses a Relu activation function; Different blocks of the CNN network are connected through a residual structure to build a deep model to fully extract feature information, and the generalization ability of the model is improved by adding a batch normalization layer; the output shape of the CNN network is (1, 512), which is transformed into a (16, 32) matrix as a group of 32 data as the input of the Transformer network; The Transformer model uses a self-attention mechanism for information interaction and transmission. This mechanism allows the model to focus on information at different positions when processing sequence data, and can effectively capture the global information of the input data. Since the calculation of this attention mechanism can be performed in parallel, the Transformer model has extremely high computational efficiency and parallel processing capabilities, and the model can learn rich data features in a short time.

5. A method for detecting abnormal state of wind turbine converter based on GAF-CNN-Transformer according to claim 4, characterized in that: The output of the Transformer model first passes through a dropout layer to avoid overfitting of the model, and then passes through a softmax layer to output the final result of the converter state detection.

Images