Wind turbine generator state monitoring method based on single-source domain self-adaption

Abstract

The invention discloses a wind turbine generator state monitoring method based on single-source domain self-adaption. The method comprises the following steps: S1, collecting running state historical data of generator systems of multiple different wind turbines at different geographic positions; s2, selecting a wind turbine generator needing to be monitored and analyzed as a target field unit, and taking other wind turbine generators as source field units; s3, the plurality of source field unit training sets comprise a source field 1, a source field 2... a source field N; s4, respectively inputting each different source domain unit training set and target domain unit training set into the convolutional neural network single-source domain adaptive model for training and learning; s5, preprocessing the online data of the operation state of the generator system of the unit in the target domain; s6, selecting a source domain unit more suitable for the target domain; and S7, monitoring the operation state of the generator of the wind turbine generator. According to the method, a new thought and a new method are provided for establishing an effective state monitoring model in a scene that data in the target domain is missing and the optimal source domain needs to be selected from a plurality of source domains.

Description

Technical Field

[0001] The present invention belongs to the technical field of wind turbine condition monitoring, and particularly relates to a method for monitoring the state of a wind turbine generator based on single-source domain adaptation. Background Art

[0002] Wind energy, as a renewable energy source, has been widely utilized. In order to effectively convert wind energy into electrical energy, wind turbines play a crucial role as key equipment. Wind turbines are usually deployed in remote onshore areas or nearshore and offshore regions. With the progress of technology, the diameter of wind turbine blades has been continuously increasing, and the nacelle position has become higher and higher, resulting in high daily operation and maintenance costs of wind turbines. There is data indicating that the operation and maintenance costs of large-scale wind farms account for about 15% - 35% of the wind farm revenue. The generator is one of the key components of a wind turbine, which directly affects the efficiency of wind energy conversion and the power generation quality. During the dynamic operation process, the generator bears complex loads and is prone to failures. Once a failure occurs in the generator, it is difficult to repair, and it will cause a long downtime of the unit. Therefore, it is necessary to monitor the state of the wind turbine generator.

[0003] Common methods for monitoring the state of wind turbine generators can be divided into physical model methods and data-driven methods. Due to the complex structure of wind turbines, physical model methods often face problems such as high modeling complexity and low model accuracy. Given that the supervisory control and data acquisition (SCADA) system has obtained the operation state data of wind turbines through sensors integrated on key components such as generators and gearboxes, without the need to install additional sensors, analyzing the SCADA data through data-driven methods has become a potential low-cost solution for unit condition monitoring. Therefore, in recent years, many studies have monitored the state of wind turbines by analyzing SCADA data.

[0004] Active power is a key feature of wind turbines and can reflect the health state of the generator. When a failure occurs in a wind turbine, the power output will change accordingly. By analyzing the power prediction residuals, the operation state of the wind turbine generator can be effectively analyzed. In recent years, the research on deep learning algorithms has been increasingly in-depth, and it has demonstrated powerful capabilities in processing high-dimensional complex data. Deep learning can efficiently model and analyze the SCADA data of wind turbines, thereby achieving more accurate power prediction and condition monitoring.

[0005] Although deep learning has made great progress in practical applications, it still has certain limitations in some application scenarios. For example, deep learning requires a large amount of labeled and identically distributed training data. However, newly installed wind turbines often face the situation of insufficient training data, resulting in a decline in model accuracy. Transfer learning focuses on cross-domain knowledge transfer and has the potential to solve the above problems. Summary of the Invention

[0006] The purpose of the present invention is to provide a method for monitoring the generator status of wind turbines based on single-source domain adaptation, which effectively solves the problem of insufficient training data for newly installed wind turbines by using the data of source domain turbines.

[0007] To achieve the above purpose, the present invention adopts the following technical solutions:

[0008] A method for monitoring the generator status of wind turbines based on single-source domain adaptation includes the following steps:

[0009] S1. Collect the historical operation status data of the generators of multiple wind turbines with different geographical locations;

[0010] S2. Select the wind turbine to be monitored and analyzed as the target domain turbine, and other wind turbines as the source domain turbines. Respectively preprocess the historical operation status data sets of the generators of multiple source domain turbines with sufficient data and one target domain turbine with less data to obtain multiple source domain turbine training sets and target domain turbine training sets;

[0011] S3. The multiple source domain turbine training sets include source domain 1, source domain 2... source domain N. Source domain 1 contains the data of two wind turbines in the same wind farm. Similarly, source domain N also contains the data of two wind turbines in the same wind farm;

[0012] S4. Input the multiple different source domain turbine training sets and target domain turbine training sets into the convolutional neural network single-source domain adaptation model for training and learning, and save the trained model structure and parameters;

[0013] S5. Preprocess the online operation data of the generator of the target domain turbine to obtain the target domain turbine test set, and input the target domain turbine test set into the model saved in step S4 to predict the generator power;

[0014] S6. Each source domain turbine and the target domain turbine training set perform domain adaptation to obtain a single-source domain adaptation model. Through the mean square error (MSE) between the predicted value and the actual value of the generator power of the multiple different source domain single-source domain adaptation models on the target domain turbine test set, select the source domain turbine that is more suitable for the data distribution of the target domain turbine;

[0015] S7. Monitor the operating state of the generator of the wind turbine by combining the selected source domain units and the domain adaptation analysis of the residual between the predicted power and the actual power of the generator in the combined domain.

[0016] As a preferred solution, the data distributions of the source domain and the target domain are different, and the data distributions between the source domains are also different.

[0017] As a preferred solution, in step S2, the wind turbine characteristic parameters included in the source domain data and the target domain data are: grid current, grid voltage, 30s average wind direction, 30s average wind speed, generator speed, generator winding temperature 1, generator drive end bearing temperature, generator non-drive end bearing temperature. Calculate the characteristics with high correlation with the generator through the Pearson correlation coefficient, and delete the characteristics with high redundancy. Finally, 6 characteristics are selected: grid current, 30s average wind direction, 30s average wind speed, generator speed, generator winding temperature 1, and generator drive end bearing temperature.

[0018] As a preferred solution, in S2, after deleting the abnormal data, the data needs to be normalized during data preprocessing, and the formula is as follows:

[0019]

[0020] Where x' represents the normalized value, and x max 、x min represent the maximum and minimum values of the feature respectively.

[0021] As a preferred solution, in S3, source domain 1 contains the data of two wind turbines in the same wind farm. These two wind turbines are from the same manufacturer and have the same rated power level; due to the high similarity of the external environment, operating conditions and data distribution, the differences between the source domains are small. Therefore, integrating the data of the two wind turbines not only effectively expands the number of source domain samples, but also provides more sufficient source domain information for model training. Similarly, source domain N also contains the data of two wind turbines in the same wind farm.

[0022] As a preferred solution, in S4, the specific steps of the convolutional neural network domain adaptation model are as follows:

[0023] Step 4.1: Construct a convolutional neural network domain adaptation model, which mainly consists of a feature extraction layer, a domain adaptation layer, and a regression layer;

[0024] Step 4.2: Input the source domain and target domain training sets into the feature extraction layer, and set the learning rate, the number of neurons in the hidden layer, the training cycle, and the training batch parameters of each layer;

[0025] Step 4.3: Set the error loss, which consists of two parts. One part is the regression loss, and the other part is the feature difference loss. Calculate it according to the following formula;

[0026]

[0027] Step 4.4: Calculate the training error loss of the network model and update the model parameters using the Adam optimization algorithm;

[0028] Step 4.5: Save the model when the maximum number of iterations is reached.

[0029] As a preferred solution, in Step 4.2, the Adam optimization algorithm is used for each layer, and the learning rate is set using a dynamic learning rate decay strategy. The formula is as follows:

[0030]

[0031] Where LR and lr represent the learning rate of the current epoch and the initial learning rate respectively, and e and E represent the current epoch and the total number of epochs respectively. lr is set to 1e-3 and E is set to 50.

[0032] As a preferred solution, in Step 4.2, one-dimensional CNN is used to extract the features of the wind turbine. The output of the convolutional layer is:

[0033]

[0034] Where, represents the output of the j-th channel of the l-th convolutional layer; represents the output of the (l-1)-th layer; N j represents the combination of the input feature maps; represents the convolutional kernel matrix; is the bias; * represents the convolutional operation; f(×) represents the ReLU activation function.

[0035] As a preferred solution, in Step 4.2, a max pooling layer is added after the convolutional layer, which is expressed as follows:

[0036]

[0037] Where represents the value of the t-th neuron in the i-th feature vector of the l-th layer, where t ∈ [(j-1)W+1, jW]; W represents the width of the pooling area; represents the corresponding value of the neuron in the (l+1)-th layer.

[0038] As a preferred solution, in Step 4.3, the mean square error MSE calculation method is as follows:

[0039]

[0040] where y i is the actual power value, is the power value predicted by the model, and n is the number of samples.

[0041] As a preferred solution, in step 4.3, the MMD loss calculation method is as follows:

[0042]

[0043] where is an unbiased estimator of, represents the source domain dataset and the target domain dataset the squared MMD distance of, and The formulas for are as follows:

[0044]

[0045] where, E p represents the expectation of the function on the source domain dataset, E q represents the expectation of the function on the target domain dataset, x s represents the time series data of the source domain, x t represents the time series data of the target domain, and φ(×) represents the mapping.

[0046] The design concept of the present invention is:

[0047] In view of the problem that the training data of newly installed wind turbine generators is insufficient and the condition monitoring work cannot be effectively carried out, the present invention proposes a method for monitoring the condition of wind turbine generators based on single-source domain adaptation: collect the historical operation data of the generators of multiple wind turbine generators with different geographical locations; select the wind turbine generator to be monitored as the target domain unit and other wind turbine generators as the source domain units, and preprocess the historical operation data sets of the generators of multiple source domain units with sufficient data and one target domain unit with less data respectively to obtain multiple source domain unit training sets and target domain unit training sets; the multiple source domain unit training sets include source domain 1, source domain 2... source domain N, and source domain 1 contains the data of two wind turbine generators in the same wind farm. Similarly, source domain N also contains the data of two wind turbine generators in the same wind farm; input the multiple different source domain unit training sets and target domain unit training sets into the convolutional neural network single-source domain adaptation model for training and learning, and save the trained model structure and parameters; preprocess the online operation data of the target domain unit generator to obtain the target domain unit test set, input the target domain unit test set into the saved model to predict the generator power; through the mean square error (MSE) between the predicted value and the actual value of the generator power of the multiple different source domain single-source domain adaptation models on the target domain unit test set, select the source domain unit that is more suitable for the data distribution of the target domain unit, run the model 30 times, count the mean square error (MSE) between the predicted value and the actual value of the generator power of each single-source domain adaptation model after 30 runs, draw a violin plot and select the best source domain; refer to the selected source domain unit and combine domain adaptation analysis of the residual between the predicted power and the actual power of the generator to monitor the condition of the wind turbine generator.

[0048] The beneficial effects of the present invention are as follows:

[0049] In view of the insufficient data volume of newly installed wind turbines, which makes it impossible to train an accurate wind turbine status monitoring model, the present invention proposes a method for monitoring the generator status of wind turbines based on single-source domain adaptation. This method inputs the training data sets of multiple different source domain turbines and target domain turbines into a convolutional neural network single-source domain adaptation model for training and learning, and saves the trained model structure and parameters; each source domain turbine and target domain turbine training data set undergoes domain adaptation to obtain a single-source domain adaptation model. By calculating the mean square error MSE between the predicted generator power value and the actual value of multiple different single-source domain adaptation models on the target domain turbine test data set, the source domain turbine that is suitable for the data distribution of the target domain turbine (with the smallest MSE) is selected; the operation status of the generator of the target domain wind turbine is monitored with reference to the selected source domain turbine and the domain adaptation model. The present invention provides a new idea and method for establishing an effective status monitoring model in the scenario where target domain data is missing and the best source domain needs to be selected from multiple source domains. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 is a flowchart of a method for monitoring the generator status of wind turbines based on single-source domain adaptation;

[0051] Figure 2 is a comparison diagram of the actual power value and the predicted power value of two different single-source domain adaptation models based on the target domain online monitoring data;

[0052] Figure 3 is a power residual diagram of two different single-source domain adaptation models based on the target domain online monitoring data;

[0053] Figure 4 is an enlarged power residual diagram of two different single-source domain adaptation models based on the target domain online monitoring data;

[0054] Figure 5 is a power residual diagram of a non-transfer model based on the target domain online monitoring data;

[0055] Figure 6 is a violin diagram of the power prediction regression loss after two different single-source domain adaptation models based on the target domain online monitoring data run 30 times respectively. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0056] The following describes the implementation modes of the present invention through specific embodiments. Those skilled in the art can easily understand other advantages and effects of the present invention from the content of this specification. The present invention can also be implemented or applied through other different specific implementation modes. Various details of this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that, without conflict, the following embodiments and the features in the embodiments can be combined with each other.

[0057] Referring to the attached Figure 1 , a method for monitoring the generator state of a wind turbine based on single-source domain adaptation, comprising the following steps:

[0058] S1. Collect historical operation state data of generators of multiple wind turbines with different geographical locations;

[0059] In step S1, the specific information of the wind turbines is shown in the following table:

[0060]

[0061] S2. Select the wind turbine to be monitored and analyzed as the target domain unit, and other wind turbines as the source domain units. Respectively preprocess the historical operation state data sets of the generators of multiple source domain units with sufficient data and one target domain unit with less data to obtain multiple source domain unit training sets and a target domain unit training set;

[0062] The data distributions of the source domain and the target domain are different, and the data distributions among the source domains are also different.

[0063] In step S2, the target domain includes one wind turbine (Unit 1#). The wind turbine characteristic parameters included in the source domain data and the target domain data are: grid current, grid voltage, 30s average wind direction, 30s average wind speed, generator speed, generator winding temperature 1, generator drive end bearing temperature, generator non-drive end bearing temperature, real-time lateral vibration value, real-time longitudinal vibration value. Calculate the characteristics with high correlation with the generator through the Pearson correlation coefficient, and delete the characteristics with high redundancy. Finally, select 6 characteristics: grid current, 30s average wind direction, 30s average wind speed, generator speed, generator winding temperature 1, generator drive end bearing temperature. The wind turbine characteristic parameters are shown in the following table:

[0064]

[0065]

[0066] In S2, after deleting abnormal data, the data needs to be normalized during data preprocessing, and the formula is as follows:

[0067]

[0068] Among them, x' represents the normalized value, and x max 、x min represent the maximum and minimum values of the feature respectively.

[0069] In the specific implementation case of the present invention, only two source domain unit training sets are included.

[0070] S3. The two source domain unit training sets include Source Domain 1 and Source Domain 2. Source Domain 1 contains data of two wind turbines in the same wind farm. Similarly, Source Domain 2 also contains data of two wind turbines in the same wind farm.

[0071] In step S3, Source Domain 1 contains data of two wind turbines (Unit 2 and Unit 3) in the same wind farm. These two wind turbines are from the same manufacturer and have the same rated power level. Due to the high similarity of the external environment, operating conditions and data distribution, the differences between source domains are small. Therefore, integrating the data of the two wind turbines not only effectively expands the number of source domain samples, but also provides more sufficient source domain information for model training. Similarly, Source Domain 2 also contains data of two wind turbines (Unit 4 and Unit 5) in the same wind farm. The power levels of Source Domain 1 and Source Domain 2 are different.

[0072] S4. Input the two different source domain unit training sets and the target domain unit training set into the convolutional neural network single source domain adaptive model for training and learning, and save the trained model structure and parameters;

[0073] In step S4, the specific steps of the domain adaptive model of the convolutional neural network are as follows:

[0074] Step 4.1: Construct a convolutional neural network adaptive model, which mainly consists of a feature extraction layer, a domain adaptive layer, and a regression layer;

[0075] Step 4.2: Input the source domain and target domain training sets into the feature extraction layer, and set the learning rate, the number of neurons in the hidden layer, the training cycle, and the training batch parameters of each layer;

[0076] Step 4.3: Set the error loss, which consists of two parts. One part is the regression loss, and the other part is the feature difference loss. Calculate according to the following formula;

[0077]

[0078] Step 4.4: Calculate the training error loss of the network model, and update the model parameters using the Adam optimization algorithm;

[0079] Step 4.5: Save the model when the maximum number of iterations is reached.

[0080] In Step 4.2, the Adam optimization algorithm is adopted for each layer, the training batch is set to 32, and the learning rate is set using a dynamic learning rate decay strategy. The formula is as follows:

[0081]

[0082] where LR and lr represent the learning rate of the current epoch and the initial learning rate respectively, and e and E represent the current epoch and the total number of epochs respectively. lr is set to 1e-3 and E is set to 50.

[0083] In Step 4.2, one-dimensional CNN is used to extract the features of the wind turbine. The output of the convolutional layer is:

[0084]

[0085] where represents the output of the j-th channel of the l-th convolutional layer; represents the output of the (l - 1)-th layer; N j represents the combination of the input feature maps; represents the convolutional kernel matrix; is the bias; * represents the convolution operation; f(×) represents the ReLU activation function.

[0086] In Step 4.2, a max pooling layer is added after the convolutional layer, which is expressed as follows:

[0087]

[0088] where represents the value of the t-th neuron in the i-th feature vector of the l-th layer, where t ∈ [(j - 1)W + 1, jW]; W represents the width of the pooling area; represents the corresponding value of the neuron in the (l + 1)-th layer.

[0089] In Step 4.3, the mean squared error MSE is calculated as follows:

[0090]

[0091] where y i is the actual power value, is the power value predicted by the model, and n is the number of samples.

[0092] In step 4.3, after extracting the features of the source domain and the target domain, in order to better utilize the source domain information and solve the problem of data distribution differences, it is necessary to align the distributions between the source domain and the target domain. The present invention uses the Maximum Mean Discrepancy (MMD) method to reduce the distribution differences between the source domain and the target domain. MMD is a metric method that is frequently used in transfer learning. It is used for two-sample tests, assuming that two probability distributions p and q are equal, and deciding whether to accept this hypothesis according to different sample test methods. Assuming that the source domain dataset and the target domain dataset follow probability distributions p and q respectively, the squared MMD distance for the source domain dataset and the target domain dataset is:

[0093]

[0094] where E p denotes the expectation of the function on the source domain dataset, and E q denotes the expectation of the function on the target domain dataset. denotes the Reproducing Kernel Hilbert Space (RKHS). φ(×) is a mapping used to map the source domain data and the target domain data to the RKHS. The kernel function is defined as the inner product of the mappings, expressed as follows:

[0095] k(x i , x j ) = <φ(x i ), φ(x j )> (8)

[0096] where <,> represents the inner product operation.

[0097] When the maximum mean difference between two probability distributions p and q is 0, p = q, and the opposite result is the same. In practice, the squared distance between the significant kernel mean embeddings estimated by MMD is:

[0098]

[0099] where is an unbiased estimator of . The distribution differences between the source domain and the target domain are measured by the above formula. represent the source domain and the target domain respectively.

[0100] x s represents the time series data of the source domain, and x t represents the time series data of the target domain.

[0101] The loss formula of MMD is expressed as:

[0102]

[0103] In step 4.3, As the main loss function, a larger weight should be given. As the secondary loss function, a smaller weight should be given, and β is set to 0.05.

[0104] S5. Preprocess the online operation data of the generator sets in the target domain to obtain a test set of the generator sets in the target domain, and input the test set of the generator sets in the target domain into the model saved in step S4 to predict the generator power.

[0105] S6. Select the source domain generator sets that are more suitable for the data distribution of the target domain generator sets through the mean square error (MSE) between the predicted values and the actual values of the generator power on the test set of the target domain generator sets by two different single-source domain adaptation models in different source domains.

[0106] S7. Refer to the selected source domain generator sets and combine domain adaptation to analyze the residuals between the predicted power and the actual power of the generator to monitor the operating status of the wind turbine generator.

[0107] In step S7, when monitoring the status of the wind turbine generator, it is necessary to set a threshold. The quantile method is used to set the residual threshold, and the 99th percentile of the data is defined as the threshold. When the residual exceeds this threshold, it indicates that the unit may have an abnormality or a fault. To reduce the phenomenon of false alarms caused by the impact of short-term extreme weather on the monitoring model, the principle of warning is considered when the model has three consecutive over-threshold phenomena within 24 hours. Figure 2 It is a comparison chart of the actual power value and the predicted power value of two different single-source domain adaptation models in the target domain test set. It can be seen from the chart that the predicted value and the actual value of the power basically coincide. Figure 3 It is a power residual chart of two different single-source domain adaptation models in the target domain test set. Figure 4 It is an enlarged power residual chart of two different single-source domain adaptation models in the target domain test set. Figure 5 It is a power residual chart of the non-transfer learning model in the target domain test set. We can see that the non-transfer learning model has frequent false alarm times and a late warning time. There are few points exceeding the threshold at the alarm, and the status monitoring effect is not good. The single-source domain adaptation model has very few false alarm times and an early warning time. The number of points exceeding the threshold at the alarm is large, and the effect is obvious. This is because the single-source domain adaptation model has learned some common characteristics or fault-related patterns of wind turbine generators during the adaptation process. When a fault occurs in the target domain generator set, the adaptation model can more easily identify these characteristics, so the state monitoring of single-source domain adaptation is more effective than non-transfer learning. Figure 6It is a violin plot of the power prediction regression loss after the single-source domain adaptation models in two different source domains are run 30 times on the target domain test set respectively. We can find that the regression loss of the model in source domain 2 is smaller than that of the model in source domain 1. Therefore, the data of the wind field in source domain 2 is selected.

[0108] The embodiments described above are only used to describe the preferred embodiments of the present invention, rather than to limit the scope of the present invention. Without departing from the design spirit of the present invention, various deformations and improvements made by those of ordinary skill in the art to the technical solutions of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for monitoring the status of a wind turbine generator based on single source domain adaptation, characterized in that: The following steps are involved: S1. Collect historical data on the operating status of multiple wind turbine generators in different geographical locations; S2. Select the wind turbines that need to be monitored and analyzed as the target domain units, and the other wind turbines as the source domain units, and pre-process the collected historical data sets of the generator operation status of the source domain units and one target domain unit respectively to obtain multiple source domain unit training sets and target domain unit training sets; S3, multiple source domain unit training sets include source domain 1, source domain 2...source domain N, where each source domain contains data of two wind turbine units in the same wind farm; S4, inputting a plurality of different source domain unit training sets and target domain unit training sets into the convolutional neural network single source domain adaptive model for training and learning, and saving the trained model structure and parameters; S5, preprocessing the online operation data of the target field unit generator to obtain the target field unit test set, inputting the target field unit test set into the model saved in step S4, and predicting the generator power; S6. A single-source domain adaptive model is obtained by performing domain adaptation on each source domain unit and target domain unit training set. The source domain unit suitable for the target domain unit data distribution is selected through the mean square error (MSE) between the generator power prediction value and the actual value of multiple different source domain single-source domain adaptive models on the target domain unit test set; S7. Monitor the operating status of the wind turbine generator by referring to the residual between the predicted power and the actual power of the generator in combination with the domain adaptive analysis of the selected source domain unit.

2. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 1, characterized in that: The data distribution of the source domain unit and the target domain unit is different, and the data distribution of each source domain unit is also different.

3. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 1, characterized in that: In S2, the characteristic parameters of wind turbines included in the source domain unit data and the target domain unit data are: grid current, grid voltage, 30s average wind direction, 30s average wind speed, generator speed, generator winding temperature, generator drive end bearing temperature, and generator non-drive end bearing temperature.

4. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 1, characterized in that: In S2, preprocessing requires normalizing the data after deleting abnormal data. The formula is as follows: In the formula, x' represents the normalized value, x max 、x min They represent the maximum and minimum values ​​of the characteristic variables respectively.

5. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 1, characterized in that: In S3, source domain 1 contains data of two wind turbines in the same wind farm, where the two wind turbines are from the same manufacturer and have the same rated power level; the same is true for source domain N.

6. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 1, characterized in that: The specific process steps of S4 are as follows: Step 4.1: Construct a convolutional neural network single-source domain adaptation model, which includes a feature extraction layer, a domain adaptation layer, and a regression layer; Step 4.2: Input the training set data of the source domain unit and the target domain unit into the feature extraction layer, and set the learning rate of each layer, the number of neurons in the hidden layer, the training cycle, and the training batch parameters; Step 4.3: Set the error loss, which consists of two parts, one is the regression loss and the other is the feature difference loss, calculated as follows; Step 4.4: Calculate the training error loss of the network model and use the Adam optimization algorithm to update the model parameters; Step 4.5: Save the model after reaching the maximum number of iterations.

7. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 6 is characterized in that: In step 4.2, a one-dimensional CNN is used to extract the features of the wind turbine, where the output of the convolutional layer is: in, Represents the output of the jth channel of convolutional layer l; represents the output of the l-1th layer; N j Represents the combination of input feature maps; Represents the convolution kernel matrix; is the bias; * represents the convolution operation; f(×) represents the ReLU activation function.

8. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 6, characterized in that: In step 4.2, a maximum pooling layer is added after the convolutional layer, which is expressed as follows: in, represents the value of the tth neuron in the ith feature vector of the lth layer, t∈[(j-1)W+1,jW]; W represents the width of the pooling area; P i l+1 (j) represents the value corresponding to the neurons in layer l+1.

9. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 6, characterized in that: The mean square error MSE calculation method in step 4.3 is as follows: Among them, y i is the actual power value, is the model prediction power value, and n is the number of samples.

10. The method for monitoring the status of a wind turbine generator based on single source domain adaptation according to claim 6, characterized in that: The MMD loss calculation method in step 4.3 is as follows: in, yes The unbiased estimator of Represents the source domain dataset and target domain datasets The MMD square distance, and The formula is as follows: Among them, E p represents the expectation of the function on the source domain dataset, E q represents the expectation of the function on the target domain dataset, x s represents the time series data of the source domain, x t represents the time series data of the target domain, and φ(×) represents the mapping.

Images