Wind turbine generator gearbox state monitoring method based on passive domain field self-adaption

Abstract

The invention discloses a wind turbine generator gearbox state monitoring method based on passive domain field self-adaption, which comprises the following steps of: respectively selecting a source domain unit and a target domain unit, and performing data preprocessing on historical operation data of the source domain unit and the target domain unit to obtain a data set I and a data set II; training by using the data set I to obtain a source domain pre-training model; inputting the data set 2 into the source domain pre-training model to obtain a corresponding gear box oil pool temperature prediction value; performing domain self-adaption on the source domain pre-training model in combination with the data set II and the corresponding oil pool temperature predicted value, so as to obtain a passive domain domain self-adaption model; based on the data set II, determining a temperature health threshold value of the gearbox oil pool through a passive domain self-adaptive model; and determining the gearbox operation state of the current target domain unit based on the oil pool temperature health threshold value and the passive domain self-adaptive model. Under the condition that source domain unit data is not used and target domain unit data is deficient, a new thought is provided for constructing a wind turbine gearbox state monitoring model.

Description

Technical Field

[0001] This invention relates to the field of wind turbine condition monitoring technology, specifically to a wind turbine gearbox condition monitoring method based on passive domain adaptive design. Background Technology

[0002] With increasing environmental pollution and a deepening energy crisis, the global demand for clean energy is growing. Wind energy, as an important renewable energy source, is widely utilized. Wind turbines, as key equipment for converting wind energy into electricity, are experiencing a continuous increase in installed capacity. The gearbox, as its core transmission component, uses a multi-stage precision gear structure to accelerate the low-speed rotation of the main shaft to the high-speed rotation required by the generator, making it a crucial carrier for converting mechanical energy into electrical energy. Due to long-term exposure to high loads and vibrations, gearboxes are prone to failures such as tooth surface wear, bearing damage, and lubrication deterioration. Statistics show that gearbox-related failures account for more than 30% of mechanical failures in wind turbines, making it one of the components with the highest failure rate. Once a gearbox fails, it not only causes turbine downtime but also leads to substantial economic losses due to prolonged maintenance cycles. Therefore, effectively monitoring the operating status of wind turbine gearboxes and reducing maintenance costs is one of the core issues in intelligent operation and maintenance of wind turbines.

[0003] Existing methods for wind turbine gearbox condition monitoring include signal analysis, physical modeling, and data-driven methods. Among these, the data-driven method, relying on supervisory control and data acquisition (SCADA) systems for modeling and analysis, requires no additional sensors and has low deployment costs, thus becoming a research hotspot in wind turbine gearbox condition monitoring. This method typically uses a large amount of labeled data to build a high-accuracy condition identification model. However, in reality, wind turbine gearbox fault samples are relatively scarce, and the accumulation of operating data from newly commissioned units is limited, leading to significant sample imbalance and data scarcity problems in the model training process. Furthermore, differences in data distribution among different wind turbines further limit the model's generalization ability. To address these challenges, researchers have introduced domain adaptation (DA) methods. However, traditional domain adaptation methods require source domain turbine data, but in the wind power industry, wind farm data often involves core operational information and sensitive equipment parameters. Due to data privacy regulations and corporate data confidentiality requirements, cross-field sharing of source domain turbine data is often difficult to achieve, resulting in the inability to meet this migration condition. Summary of the Invention

[0004] The purpose of this invention is to provide a passive domain adaptive method for wind turbine gearbox condition monitoring. In the source domain pre-training stage, a source domain pre-training model is trained based on source domain turbine data, including a shared feature extractor and two predictors with identical structures but different initialization parameters. In the target domain migration stage, the corresponding predicted gearbox oil sump temperature is obtained using the source domain pre-training model and target domain turbine data. The prediction consistency difference between the target domain and source domain turbine data is minimized by fine-tuning the shared feature extractor, achieving distribution alignment between the target and source domain turbine data, and obtaining a passive domain adaptive model. Finally, residual analysis of the gearbox oil sump temperature is combined to monitor the operating status of the gearbox in the target domain.

[0005] In an embodiment of the present invention, a method for monitoring the condition of a wind turbine gearbox based on passive domain adaptive domain is provided, comprising:

[0006] Step 1. Determine the source domain units based on the historical operating data of the wind turbine units; perform data preprocessing on the historical operating data of the source domain units to obtain the source domain unit dataset;

[0007] Step 2. Determine the target domain units; perform data preprocessing on the historical operating data of the target domain units to obtain Data Set 2 for the target domain units;

[0008] Step 3. Train the GRFC model based on the source domain group's dataset to obtain the source domain pre-trained model;

[0009] Step 4. Based on the input features in dataset 2 of the target domain unit, obtain the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2 through the source domain pre-trained model;

[0010] Step 5. Based on the input features in dataset 2 of the target domain unit and the gearbox oil sump temperature prediction values ​​that correspond one-to-one with the input features in dataset 2, perform domain adaptation on the source domain pre-trained model to obtain the passive domain adaptive model.

[0011] Step 6. Based on the target domain unit's dataset 2, determine the gearbox oil sump temperature health threshold using a passive domain adaptive model;

[0012] Step 7. Determine the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model.

[0013] As a preferred embodiment of the present invention, step 1, which involves preprocessing the historical operating data of the source domain units to obtain the source domain unit dataset, specifically includes:

[0014] Step 11. Perform data cleaning on the historical operating data of the source domain units;

[0015] Step 12. Perform feature filtering on the historical operational data after data cleaning;

[0016] Step 13. Normalize the data after feature filtering to obtain dataset one.

[0017] As a preferred embodiment of the present invention, step 2, which involves preprocessing the historical operating data of the target domain units to obtain the second dataset of the source domain units, specifically includes:

[0018] Step 21. Perform data cleaning on the historical operating data of the target domain units;

[0019] Step 22. Perform feature filtering on the historical operational data after data cleaning;

[0020] Step 23. Normalize the data after feature filtering to obtain dataset two.

[0021] As a preferred embodiment of the present invention, step 3, training the GRFC model to obtain the source domain pre-trained model, specifically includes:

[0022] Step 31. Construct the GRFC (gated recurrent unit - fully connected neural network) model: The GRFC model includes a shared feature extractor and a first predictor and a second predictor with the same structure but different initialization parameters. The shared feature extractor is composed of a gated recurrent unit (GRU), and both the first and second predictors are composed of a fully connected neural network (FCNN).

[0023] Step 32. Load the features from Dataset 1 of the source domain units into the GRFC model and set the hyperparameters;

[0024] Step 33. Set the training error loss;

[0025] Step 34. Calculate the training error loss of the GRFC model and update the model parameters using the Adam optimization algorithm;

[0026] Step 35. Determine whether the training error loss has reached the preset value 1. If the training error loss has reached the preset value 1, save the model parameters to obtain the source domain pre-trained model; otherwise, return to step 34.

[0027] As a preferred embodiment of the present invention, in step 4, the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset two is obtained through the source domain pre-trained model using the following formula: , in, For a set of input features in dataset two of the target domain units, For shared feature extractors, As the first predictor, For the second predictor, This is the predicted temperature of the gearbox oil sump corresponding to a certain set of input features.

[0028] As a preferred embodiment of the present invention, step 5, in which the source domain pre-trained model undergoes domain adaptation to obtain a source-free domain adaptive model, specifically includes:

[0029] Step 51. Load the input features from dataset 2 of the target domain unit and the gearbox oil sump temperature prediction data corresponding one-to-one with the input features in dataset 2 into the source domain pre-trained model and set the hyperparameters;

[0030] Step 52. Freeze the first predictor and the second predictor;

[0031] Step 53. Set the combined error loss, which includes regression loss and prediction consistency loss; , in, For the combination error loss, To regress the loss, To predict consistency loss, Weights for predicting consistency loss;

[0032] Step 54. Calculate the combined error loss of the source domain pre-trained model and update the model parameters using the Adam optimization algorithm;

[0033] Step 55. Determine whether the combined error loss has reached the preset value two. If the combined error loss has reached the preset value two, save the model parameters to obtain the passive domain adaptive model; otherwise, return to step 54.

[0034] As a preferred embodiment of the present invention, in step 53, the regression loss is calculated using the following formula: ,

[0035] in, To regress the loss, This is the predicted gearbox oil sump temperature corresponding to a certain set of input features. The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

[0036] In a preferred embodiment of the present invention, in step 53, the predicted consistency loss is calculated using the following formula: , in, To predict consistency loss, The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

[0037] As a preferred embodiment of the present invention, step 6, determining the gearbox oil sump temperature health threshold based on the target domain unit's dataset two using a passive domain adaptive model specifically includes:

[0038] Step 61. Based on the input features in dataset 2 of the target domain unit, obtain the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset 2 through the passive domain adaptive model;

[0039] Step 62. Calculate the gearbox oil sump temperature residual based on the actual gearbox oil sump temperature value in dataset 2 and the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2;

[0040] Step 63. Determine the maximum gearbox oil sump temperature residual under normal operating conditions of the gearbox of the target domain unit based on the temperature residual of all gearbox oil sumps, and set the maximum gearbox oil sump temperature residual as the healthy threshold of gearbox oil sump temperature.

[0041] As a preferred embodiment of the present invention, step 7, determining the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model, specifically includes:

[0042] Step 71. Obtain the current actual operating data of the target domain units;

[0043] Step 72. Based on the input features of the current actual operating data, obtain the predicted gearbox oil sump temperature corresponding to the input features of the current actual operating data through a passive domain adaptive model;

[0044] Step 73. Calculate the current gearbox oil sump temperature residual based on the actual value of the gearbox oil sump temperature in the current actual operating data and the predicted value of the gearbox oil sump temperature corresponding to the input characteristics of the current actual operating data;

[0045] Step 74. Determine whether the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold. If the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold, then the gearbox status of the current target domain unit is determined to be abnormal; otherwise, the gearbox status of the current target domain unit is normal.

[0046] In summary, the present invention has the following beneficial effects:

[0047] In the source domain pre-training stage, a source domain pre-training model is trained based on source domain unit data. This model includes a shared feature extractor and two predictors with identical structures but different initialization parameters. In the target domain migration stage, the corresponding gearbox oil sump temperature prediction value is obtained using the source domain pre-training model and target domain unit data. By fine-tuning the shared feature extractor, the prediction consistency difference of the target domain unit data is minimized to achieve distribution alignment between the target domain and source domain unit data, resulting in a source domain-free adaptive model. Finally, residual analysis of the gearbox oil sump temperature is combined to monitor the gearbox operating status of the target domain units. The proposed method achieves domain adaptation using only the source domain pre-training model and target domain unit data, even when source domain unit data is not used and target domain unit data is scarce. This method fundamentally avoids the privacy risks of sharing source domain data across wind farms, meets the data confidentiality needs of wind power companies, and has significant practical value and broad application prospects.

[0048] Further or more detailed beneficial effects will be described in conjunction with specific embodiments in the detailed implementation. Attached Figure Description

[0049] Figure 1 A flowchart of a wind turbine gearbox condition monitoring method based on passive domain adaptive method according to an embodiment of the present invention is shown;

[0050] Figure 2 A schematic diagram of a wind turbine gearbox condition monitoring method based on passive domain adaptive method according to an embodiment of the present invention is shown;

[0051] Figure 3 The data visualization diagrams of the source domain and target domain units in Experiment Group 1 are shown;

[0052] Figure 4 This shows a data visualization of the source and target domain units in Experiment Group 2;

[0053] Figure 5 The status monitoring results of the C gearbox of the wind turbine are shown;

[0054] Figure 6 The status monitoring results of the Z gearbox of the wind turbine are shown. Detailed Implementation

[0055] Embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While some embodiments of the invention are shown in the drawings, it should be understood that the invention can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of the invention. It should be understood that the accompanying drawings and embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the invention.

[0056] In the description of embodiments of the present invention, the term "comprising" and similar terms should be understood as open-ended inclusion, i.e., "including but not limited to". The term "based on" should be understood as "at least partially based on". The term "one embodiment" or "the embodiment" should be understood as "at least one embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other explicit and implicit definitions may also be included below.

[0057] like Figure 1 and Figure 2 As shown in the figure, this embodiment proposes a wind turbine gearbox condition monitoring method based on passive domain adaptive design, including:

[0058] Step 1. Determine the source domain units based on the historical operating data of the wind turbine units; perform data preprocessing on the historical operating data of the source domain units to obtain a unified dataset of the source domain units.

[0059] Assume there is a wind turbine A in region A, a wind turbine B in region B, a wind turbine C in region C, and so on, with a wind turbine Z in region Z. This step requires first obtaining the historical operating data for wind turbine A, wind turbine B, wind turbine C, and so on, for wind turbine Z. Then, by reviewing the historical operating data, a wind turbine with "sufficient healthy training data" is identified. This embodiment assumes that wind turbine A has "sufficient healthy training data," so wind turbine A can be identified as the source region turbine.

[0060] Once the source domain wind turbine (i.e., wind turbine A) is determined, data preprocessing of its historical operating data is necessary. Wind turbines are typically deployed in harsh environments, making SCADA systems susceptible to interference during data acquisition and transmission. This can result in some acquired data containing information irrelevant to the turbine's operating status. Furthermore, situations such as wind curtailment or unplanned shutdowns can generate significant amounts of noisy data. Therefore, data preprocessing, as the first step in condition monitoring, aims to provide high-quality data and prevent noisy data from affecting model learning. The data preprocessing process mainly includes cleaning, feature selection, and normalization operations.

[0061] In step 1 of this embodiment, data preprocessing of the historical operating data of the source domain units to obtain the source domain unit dataset specifically includes:

[0062] Step 11. Perform data cleaning on the historical operating data of the source domain units. Specifically, this may involve removing noisy data (such as missing value data, shutdown data, power limit data, data outside the normal operating range of the wind turbine units, and outlier data).

[0063] Step 12. Feature filtering is performed on the historical operating data after data cleaning. The source domain unit data contains numerous wind turbine features. Given that gearbox oil sump temperature is an important indicator of gearbox operating status, features with high correlation to gearbox oil sump temperature are calculated using the mutual information method, and features with high redundancy are removed using the Pearson correlation coefficient. The final filtered features include generator active power, generator drive end bearing temperature, generator non-drive end bearing temperature, main bearing temperature, nacelle control cabinet temperature, #3 propeller motor temperature, hub temperature, gearbox inlet oil pressure, gearbox oil circuit filter pre-pressure, generator stator U-phase coil temperature, gearbox high-speed shaft non-drive end bearing temperature, pitch battery cabinet 3 temperature, pitch drive 3 temperature, tower base control cabinet temperature, and gearbox oil sump temperature.

[0064] Step 13. Normalize the data after feature filtering to obtain dataset one. This step can be performed using the following formula for normalization: , in, For a certain data value of a certain feature, The minimum data value for a certain feature. The maximum data value of a certain feature. This is the normalized value of a specific data point for a given feature. Normalization eliminates the influence of dimensions on the data.

[0065] Step 2. Determine the target domain unit; perform data preprocessing on the historical operating data of the target domain unit to obtain Data Set 2 for the target domain unit.

[0066] This step designates the wind turbines requiring condition monitoring as the target domain units. There can be one, two, or more target domain units. For ease of understanding, this embodiment assumes there is only one target domain unit, which can be wind turbine C.

[0067] Once the target domain turbine unit (i.e., wind turbine unit C) is determined, data preprocessing of the historical operating data of wind turbine unit C is required. In step 2 of this embodiment, the data preprocessing of the historical operating data of the target domain turbine unit to obtain the target domain turbine unit's dataset two specifically includes:

[0068] Step 21. Perform data cleaning on the historical operating data of the target domain units. Specifically, this may involve removing noisy data (such as missing value data, shutdown data, power limit data, data outside the normal operating range of wind turbine units, and outlier data).

[0069] Step 22. Perform feature filtering on the historical operating data after data cleaning. The filtering features of the target domain units need to be the same as those of the source domain units. Therefore, the features filtered in this step also include generator active power, generator drive end bearing temperature, generator non-drive end bearing temperature, main bearing temperature, nacelle control cabinet temperature, #3 propeller motor temperature, hub temperature, gearbox inlet oil pressure, gearbox oil circuit filter pre-oil pressure, generator stator U-phase coil temperature, gearbox high-speed shaft non-drive end bearing temperature, pitch battery cabinet 3 temperature, pitch drive 3 temperature, tower bottom control cabinet temperature, and gearbox oil sump temperature.

[0070] Step 23. Normalize the data after feature filtering to obtain dataset two. This step can be performed using the following formula for normalization: , in, For a certain data value of a certain feature, The minimum data value for a certain feature. The maximum data value of a certain feature. This is the normalized value of a specific data point for a given feature. Normalization eliminates the influence of dimensions on the data.

[0071] Step 3. Train the GRFC model based on the dataset of the source domain group to obtain the source domain pre-trained model.

[0072] Step 3, training the GRFC model to obtain the source domain pre-trained model, specifically includes:

[0073] Step 31. Construct the GRFC model: The GRFC model includes a shared feature extractor and a first predictor and a second predictor with the same structure but different initialization parameters. The shared feature extractor is composed of GRU, and both the first and second predictors are composed of FCNN.

[0074] The gated loop unit of the shared feature extractor extracts features from wind turbine data using the following formula: , , , , in, To update the door, It is the sigmoid activation function. To update the gate weight matrix, This is the hidden state from the previous moment. For the current moment's data, To reset the door, To reset the gate weight matrix, In the candidate hidden state, For element-wise multiplication, This is the final hidden state.

[0075] The fully connected neural network with a first predictor and a second predictor maps the features extracted by the feature extractor to the output space using the following formula: , in, To output the feature vector, For activation function, This is the weight matrix. For the extracted feature vector, This is the bias vector.

[0076] Step 32. Load the features from dataset 1 of the source domain units into the GRFC model and set the hyperparameters. Set the learning rate, number of hidden layer neurons, training cycle, and training batch parameters for each layer.

[0077] Step 33. Set the training error loss.

[0078] Step 34. Calculate the training error loss of the GRFC model and update the model parameters using the Adam optimization algorithm.

[0079] In this embodiment, the learning rate setting adopts a dynamic learning rate decrease strategy, as shown in the following formula: ,

[0080] in, The learning rate for the current loop. The initial learning rate, This represents the current loop count. This represents the total number of iterations. In this embodiment, It is 1e-4, It is 100.

[0081] Step 35. Determine whether the training error loss has reached a preset value of 1. The preset value of 1 is set to 0.001. When the training error loss reaches the preset value of 1, save the model parameters to obtain the source domain pre-trained model; otherwise, return to step 34.

[0082] The source domain generator set dataset 1 includes input and output features. The input features are: generator active power, generator drive-end bearing temperature, generator non-drive-end bearing temperature, main bearing temperature, nacelle control cabinet temperature, #3 propeller motor temperature, hub temperature, gearbox inlet oil pressure, gearbox oil pressure before the filter, generator stator U-phase coil temperature, gearbox high-speed shaft non-drive-end bearing temperature, pitch battery cabinet 3 temperature, pitch actuator 3 temperature, and tower base control cabinet temperature. The output feature is the gearbox oil sump temperature. In this step, the input and output features of dataset 1 are used together to train the GRFC model, thereby obtaining the source domain pre-trained model.

[0083] Step 4. Based on the input features in dataset 2 of the target domain unit, obtain the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2 through the source domain pre-trained model.

[0084] Dataset 2 for the target domain unit also includes input and output features. The input features are generator active power, generator drive-end bearing temperature, generator non-drive-end bearing temperature, main bearing temperature, nacelle control cabinet temperature, #3 propeller motor temperature, hub temperature, gearbox inlet oil pressure, gearbox oil pressure before the filter, generator stator U-phase coil temperature, gearbox high-speed shaft non-drive-end bearing temperature, pitch battery cabinet 3 temperature, pitch drive 3 temperature, and tower bottom control cabinet temperature. The output feature is the gearbox oil sump temperature. However, in this step, only the input features of dataset 2 are used; the output features are not. This step inputs the input features of dataset 2 into the source domain pre-trained model (which has already been trained in step 3), causing the source domain pre-trained model to output a predicted value for the gearbox oil sump temperature. After each set of input features is input into the source domain pre-trained model, the source domain pre-trained model will output a corresponding predicted value for the gearbox oil sump temperature.

[0085] In step 4, the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset 2 is obtained through the source domain pre-trained model using the following formula: ,

[0086] in, For a set of input features in dataset two of the target domain units, For shared feature extractors, As the first predictor, For the second predictor, This is the predicted gearbox oil sump temperature corresponding to a certain set of input features. The input to the shared feature extractor is the input features in dataset two. The output of the shared feature extractor serves as the input to both the first and second predictors. The outputs of both the first and second predictors are initial predicted gearbox oil sump temperatures. In this embodiment, the final predicted gearbox oil sump temperature is obtained by averaging the two initial predicted gearbox oil sump temperatures. This method yields more accurate predicted gearbox oil sump temperatures.

[0087] Step 5. Based on the input features in dataset 2 of the target domain unit and the gearbox oil sump temperature prediction values ​​that correspond one-to-one with the input features in dataset 2, perform domain adaptation on the source domain pre-trained model to obtain the passive domain adaptive model.

[0088] In step 5 of this embodiment, the process of performing domain adaptation on the source domain pre-trained model to obtain a source-free domain adaptive model specifically includes:

[0089] Step 51. Load the input features from dataset 2 of the target domain unit and the gearbox oil sump temperature prediction data corresponding one-to-one with the input features in dataset 2 into the source domain pre-trained model and set the hyperparameters;

[0090] Step 52. Freeze the first predictor and the second predictor;

[0091] Step 53. Set the combined error loss, which includes regression loss and prediction consistency loss; ,

[0092] in, For the combination error loss, To regress the loss, To predict consistency loss, The weight for predicting consistency loss is set to 0.2;

[0093] Step 54. Calculate the combined error loss of the source domain pre-trained model and update the model parameters using the Adam optimization algorithm;

[0094] In this embodiment, the learning rate setting adopts a dynamic learning rate decrease strategy, as shown in the following formula: , in, The learning rate for the current loop. The initial learning rate, This represents the current loop count. This represents the total number of iterations. In this embodiment, It is 1e-5. It is 50.

[0095] Step 55. Determine whether the combined error loss has reached the preset value two. The preset value two is set to 0.001. When the combined error loss reaches the preset value two, save the model parameters to obtain the passive domain adaptive model; otherwise, return to step 54.

[0096] Specifically, in step 53, the regression loss is calculated using the following formula: , in, To regress the loss, This is the predicted gearbox oil sump temperature corresponding to a certain set of input features. The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

[0097] In step 53, the prediction consistency loss is calculated using the following formula: , in, To predict consistency loss, The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

[0098] The second dataset for the target domain unit includes input and output features. The input features are generator active power, generator drive-end bearing temperature, generator non-drive-end bearing temperature, main bearing temperature, nacelle control cabinet temperature, #3 propeller motor temperature, hub temperature, gearbox inlet oil pressure, gearbox oil pressure before the filter, generator stator U-phase coil temperature, gearbox high-speed shaft non-drive-end bearing temperature, pitch battery cabinet 3 temperature, pitch drive 3 temperature, and tower bottom control cabinet temperature. The output feature is the gearbox oil sump temperature. In this step, only the input features of dataset two are used (excluding the output features) along with the predicted gearbox oil sump temperature (obtained through the source domain pre-trained model in step 5) corresponding to each set of input features to adjust the source domain pre-trained model, transforming it into a passive domain adaptive model.

[0099] Step 6. Based on the target domain unit's dataset 2, determine the gearbox oil sump temperature health threshold using a passive domain adaptive model.

[0100] In step 6 of this embodiment, determining the gearbox oil sump temperature health threshold based on dataset 2 of the target domain unit using a passive domain adaptive model specifically includes:

[0101] Step 61. Based on the input features in dataset 2 of the target domain unit, obtain the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset 2 through the passive domain adaptive model.

[0102] After inputting the input features from dataset 2 into the passive domain adaptive model (obtained through step 5), the predicted gearbox oil sump temperature corresponding to the input features from dataset 2 can be obtained directly.

[0103] Step 62. Calculate the gearbox oil sump temperature residual based on the actual gearbox oil sump temperature value in dataset 2 and the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2.

[0104] The output feature in dataset two is the gearbox oil sump temperature, which is the actual value of the gearbox oil sump temperature. Subtracting the predicted value of the gearbox oil sump temperature from the actual value of the gearbox oil sump temperature yields the gearbox oil sump temperature residual.

[0105] Step 63. Determine the maximum gearbox oil sump temperature residual under normal operating conditions of the gearbox of the target domain unit based on the temperature residual of all gearbox oil sumps, and set the maximum gearbox oil sump temperature residual as the healthy threshold of gearbox oil sump temperature.

[0106] Step 7. Determine the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model.

[0107] In step 7 of this embodiment, determining the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model specifically includes:

[0108] Step 71. Obtain the current actual operating data of the target domain units. Subsequently, it is only necessary to obtain the actual operating data of the target domain units (i.e., wind turbine C). The latest set of actual operating data is the current actual operating data.

[0109] Step 72. Based on the input features of the current actual operating data, obtain the predicted gearbox oil sump temperature corresponding to the input features of the current actual operating data through the passive domain adaptive model. After inputting the input features from the current actual operating data into the passive domain adaptive model (obtained through step 5), the passive domain adaptive model will output a predicted value of the current gearbox oil sump temperature.

[0110] Step 73. Calculate the current gearbox oil sump temperature residual based on the actual value of the gearbox oil sump temperature in the current actual operating data and the predicted value of the gearbox oil sump temperature corresponding to the input characteristics of the current actual operating data.

[0111] Step 74. Determine whether the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold. If the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold, then the gearbox status of the current target domain unit (i.e., wind turbine C) is determined to be abnormal, and an early warning is triggered; otherwise, the gearbox status of the current target domain unit is normal.

[0112] This invention selects wind turbine A as the source domain wind turbine, and wind turbine C and wind turbine Z as the target domain wind turbines. Two experimental groups are set up. Experimental group one consists of the source domain wind turbine (wind turbine A) and the target domain wind turbine (wind turbine C), and experimental group two consists of the source domain wind turbine (wind turbine A) and the target domain wind turbine (wind turbine Z). Figures 3-6 Figures showing the relevant experimental results: Figure 3 This section presents the data visualization results of the source domain units and target domain units in Experiment Group 1. Figure 3 (a) Specifically presents the original data distribution characteristics of wind turbine A and wind turbine C; Figure 3 (b) The data distribution characteristics of the two generator sets after processing by the passive domain adaptive method proposed in this invention are presented. Figure 3 (b) It can be seen that the data distribution of the two units after processing has a high degree of consistency. Figure 4 This section presents the data visualization results for the source domain units and target domain units in Experiment Group 2. Figure 4 (a) Specifically presents the original data distribution characteristics of wind turbine A and wind turbine Z; Figure 4 (b) The data distribution characteristics of the two generator sets after processing by the passive domain adaptive method proposed in this invention are presented. Figure 4 (b) shows that the data distribution difference between the two units is significantly reduced after processing. Combined with... Figure 3 and Figure 4 The experimental results clearly demonstrate that the passive domain adaptive method proposed in this invention can effectively optimize the data distribution consistency between the source domain units and the target domain units, fully proving the effectiveness and practicality of the method. Figure 5 The results of condition monitoring of the oil sump temperature in the C gearbox of the wind turbine are presented. Figure 5 (a) Presents the gearbox oil sump temperature residual obtained using the source domain pre-training method; Figure 5 (b) Presents the gearbox oil sump temperature residual obtained using the passive domain adaptive method proposed in this invention. Figure 5 (a) and Figure 5(b) shows that the gearbox oil sump temperature residual obtained by the passive domain adaptive method has better smoothness under the healthy operation state of the unit and more significant abnormal characteristics under the fault state of the unit; and the alarm time triggered by this method is October 28, which is 20 days earlier than the alarm time of the existing SCADA system, and can realize early warning of faults. Figure 6 The results of condition monitoring of the oil sump temperature in the Z gearbox of the wind turbine are displayed. Figure 6 (a) Presents the gearbox oil sump temperature residual obtained using the source domain pre-training method; Figure 6 (b) Presents the gearbox oil sump temperature residual obtained using the passive domain adaptive method proposed in this invention. Figure 6 (b) It can be seen that the gearbox oil sump temperature residual obtained by the passive domain adaptive method has better smoothness and no false alarms occurred during the entire monitoring period, which further demonstrates the stability and reliability of the method in condition monitoring.

[0113] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for monitoring the condition of a wind turbine gearbox based on passive domain adaptive characteristics, characterized in that, include: Step 1. Determine the source domain units based on historical operating data of wind turbine units; Data preprocessing is performed on the historical operating data of the source domain units to obtain the first dataset of the source domain units; Step 2. Determine the target domain units; perform data preprocessing on the historical operating data of the target domain units to obtain Data Set 2 for the target domain units; Step 3. Train the GRFC model based on the source domain group's dataset to obtain the source domain pre-trained model; Step 4. Based on the input features in dataset 2 of the target domain unit, obtain the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2 through the source domain pre-trained model; Step 5. Based on the input features in dataset 2 of the target domain unit and the gearbox oil sump temperature prediction values ​​that correspond one-to-one with the input features in dataset 2, perform domain adaptation on the source domain pre-trained model to obtain the passive domain adaptive model. Step 6. Determine the gearbox oil sump temperature health threshold based on dataset 2 of the target domain unit and through a passive domain adaptive model; Step 7. Determine the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model.

2. The method according to claim 1, characterized in that, Step 1, specifically the data preprocessing of the historical operating data of the source domain units to obtain the source domain unit dataset, includes: Step 11. Perform data cleaning on the historical operating data of the source domain units; Step 12. Perform feature filtering on the historical operational data after data cleaning; Step 13. Normalize the data after feature filtering to obtain dataset one.

3. The method according to claim 1, characterized in that, Step 2, specifically the data preprocessing of the historical operating data of the target domain units to obtain the second dataset of the target domain units, includes: Step 21. Perform data cleaning on the historical operating data of the target domain units; Step 22. Perform feature filtering on the historical operational data after data cleaning; Step 23. Normalize the data after feature filtering to obtain dataset two.

4. The method according to claim 1, characterized in that, Step 3, training the GRFC model to obtain the source domain pre-trained model, specifically includes: Step 31. Construct the GRFC (gated recurrent unit - fully connected neural network) model: The GRFC model includes a shared feature extractor and a first predictor and a second predictor with the same structure but different initialization parameters. The shared feature extractor is composed of a gated recurrent unit (GRU), and both the first and second predictors are composed of a fully connected neural network (FCNN). Step 32. Load the features from Dataset 1 of the source domain units into the GRFC model and set the hyperparameters; Step 33. Set the training error loss; Step 34. Calculate the training error loss of the GRFC model and update the model parameters using the Adam optimization algorithm; Step 35. Determine whether the training error loss has reached the preset value 1. If the training error loss has reached the preset value 1, save the model parameters to obtain the source domain pre-trained model; otherwise, return to step 34.

5. The method according to claim 1, characterized in that, In step 4, the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset 2 is obtained through the source domain pre-trained model using the following formula: ， in, For a set of input features in dataset two of the target domain units, For shared feature extractors, As the first predictor, For the second predictor, This is the predicted temperature of the gearbox oil sump corresponding to a certain set of input features.

6. The method according to claim 1, characterized in that, In step 5, the process of performing domain adaptation on the source domain pre-trained model to obtain a source-free domain adaptive model specifically includes: Step 51. Load the input features from dataset 2 of the target domain unit and the gearbox oil sump temperature prediction data corresponding one-to-one with the input features in dataset 2 into the source domain pre-trained model and set the hyperparameters; Step 52. Freeze the first predictor and the second predictor; Step 53. Set the combined error loss, which includes regression loss and prediction consistency loss; ， in, For the combination error loss, To regress the loss, To predict consistency loss, Weights for predicting consistency loss; Step 54. Calculate the combined error loss of the source domain pre-trained model and update the model parameters using the Adam optimization algorithm; Step 55. Determine whether the combined error loss has reached the preset value two. If the combined error loss has reached the preset value two, save the model parameters to obtain the passive domain adaptive model; otherwise, return to step 54.

7. The method according to claim 6, characterized in that, In step 53, the regression loss is calculated using the following formula: ， in, To regress the loss, This is the predicted gearbox oil sump temperature corresponding to a certain set of input features. The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

8. The method according to claim 6, characterized in that, In step 53, the predicted consistency loss is calculated using the following formula: ， in, To predict consistency loss, The first predictor is the gearbox oil sump temperature prediction value obtained by the first predictor based on a certain set of input features. The second predictor is the predicted temperature of the second gearbox oil sump obtained by a certain set of input features. This represents the sample size for the predicted gearbox oil sump temperature.

9. The method according to claim 1, characterized in that, In step 6, determining the gearbox oil sump temperature health threshold based on dataset 2 of the target domain unit and using a passive domain adaptive model specifically includes: Step 61. Based on the input features in dataset 2 of the target domain unit, obtain the predicted gearbox oil sump temperature corresponding one-to-one with the input features in dataset 2 through the passive domain adaptive model; Step 62. Calculate the gearbox oil sump temperature residual based on the actual gearbox oil sump temperature value in dataset 2 and the gearbox oil sump temperature prediction value corresponding one-to-one with the input features in dataset 2; Step 63. Determine the maximum gearbox oil sump temperature residual under normal operating conditions of the gearbox of the target domain unit based on the temperature residual of all gearbox oil sumps, and set the maximum gearbox oil sump temperature residual as the healthy threshold of gearbox oil sump temperature.

10. The method according to claim 1, characterized in that, In step 7, determining the gearbox operating status of the current target domain unit based on the gearbox oil sump temperature health threshold and the passive domain adaptive model specifically includes: Step 71. Obtain the current actual operating data of the target domain units; Step 72. Based on the input features of the current actual operating data, obtain the predicted gearbox oil sump temperature corresponding to the input features of the current actual operating data through a passive domain adaptive model; Step 73. Calculate the current gearbox oil sump temperature residual based on the actual value of the gearbox oil sump temperature in the current actual operating data and the predicted value of the gearbox oil sump temperature corresponding to the input characteristics of the current actual operating data; Step 74. Determine whether the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold. If the current gearbox oil sump temperature residual is greater than the gearbox oil sump temperature health threshold, then the gearbox status of the current target domain unit is determined to be abnormal; otherwise, the gearbox status of the current target domain unit is normal.

Images