A kind of vehicle-mounted traction fan bearing abnormality evaluation method based on digital-analog drive

Abstract

The present application relates to the technical field of fan motor, solves the technical problem that the existing technology is difficult to collect abnormal data, the cost is high, the abnormal data is scarce, and the model is under-fitted due to insufficient training data, especially relates to a kind of abnormal evaluation method of vehicle-mounted traction fan bearing based on digital model driving, including obtaining the limited prior abnormal measured data of fan bearing, based on the real geometry and working condition parameters of fan bearing, construct four-degree-of-freedom dynamics model to obtain mass simulation dynamic response signal, construct deep branch migration network model.The present application makes up the short board that fan bearing abnormal measured data is scarce under vehicle-mounted scene, does not need to rely on a large number of difficult to obtain measured abnormal data, greatly reduces the time cost and economic cost of data collection, ensures the working condition adaptability of data, provides high-quality source domain data support for model training, avoids the under-fitting problem of model due to insufficient data.

Description

Technical Field

[0001] This invention relates to the field of wind turbine motor technology, and in particular to a method for evaluating bearing anomalies in vehicle-mounted traction wind turbines based on digital model drive. Background Technology

[0002] As a key component of the locomotive's power source ventilation and cooling system, the onboard traction fan plays a crucial role in ensuring the safe and stable operation of the locomotive. Due to the harsh operating environment, the bearings of onboard traction fans are highly susceptible to abnormal failures. Failure to detect and address these failures in a timely manner can lead to bearing jamming, fan shutdown, and even systemic power system malfunctions, causing serious safety hazards and economic losses. Therefore, current anomaly assessments of onboard traction fan bearings typically rely on deep transfer learning algorithms. These algorithms collect massive amounts of measured vibration signals from similar bearings in both normal and abnormal states to train a classification model for anomaly identification. However, the low incidence of bearing anomaly failures in onboard scenarios, the difficulty and high cost of collecting measured anomaly data, and the resulting scarcity of anomaly data lead to underfitting of the model due to insufficient training data. This results in low anomaly identification accuracy, poor generalization ability, and an inability to meet the anomaly assessment needs of small sample scenarios. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a method for evaluating the anomalies of vehicle-mounted traction fan bearings based on digital model-driven methods. This method solves the technical problem that the scarcity of anomaly data due to the difficulty and high cost of collecting measured anomaly data in existing technologies leads to underfitting of the model due to insufficient training data. This method aims to improve the accuracy and generalization ability of anomaly identification.

[0004] To solve the above-mentioned technical problems, the present invention provides the following technical solution: a method for evaluating the abnormality of bearings in vehicle-mounted traction wind turbines based on digital model driving, the method comprising the following steps:

[0005] S1. Obtain finite prior anomaly measurement data of wind turbine bearings, and construct a four-degree-of-freedom dynamic model based on the actual geometric and operating parameters of the wind turbine bearings to obtain massive simulation dynamic response signals.

[0006] S2. Construct a deep branch transfer network model for extracting abnormal features of wind turbine bearing vibration signals. It includes two deep branch networks with consistent structures. The deep branch network includes a feature extraction module and a feature fitting and classification module that extract features layer by layer.

[0007] S3. Using limited prior abnormal measured data as target domain data and massive simulated dynamic response signals as source domain data, and performing preprocessing, the data is then substituted into the deep branch transfer network model. After that, the cross-entropy loss function is used to calculate the model loss value, and the model parameters are iteratively updated to complete the training and optimization of the deep branch transfer network model.

[0008] S4. Deploy the trained and optimized deep branch transfer network model at the vehicle-mounted wind turbine site, perform anomaly assessment on the wind turbine bearing status, execute corresponding operation and maintenance operations based on the preset assessment results, and return to step S1.

[0009] Furthermore, step S1 specifically includes the following steps:

[0010] S11. Obtain finite prior anomaly measurement data of the wind turbine bearing under preset operating conditions, wherein the preset operating conditions include rotational frequency, sampling frequency and sampling time.

[0011] S12. Based on the geometric characteristics of the wind turbine bearing and Hertz contact theory, calculate the comprehensive contact stiffness between the rolling elements of the wind turbine bearing and the raceway surfaces of the inner and outer rings, and establish the relationship between contact deformation and load. The expressions are as follows:

[0012] ;

[0013] ;

[0014] In the formula, This refers to the overall contact stiffness between the rolling elements of the wind turbine bearing and the raceways of the inner and outer rings. The load-deformation coefficient, This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the inner ring. This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the outer ring. For Hertzian contact force, This represents the radial contact deformation of the contact pair. The radial contact velocity of the contact pair, This is the damping coefficient between the inner and outer raceways;

[0015] S13. Based on the concentrated spring-mass-damping model of the wind turbine bearing, construct the four-degree-of-freedom dynamic equilibrium equations of the wind turbine bearing under healthy conditions, as shown in the following expression:

[0016] ;

[0017] In the formula, This refers to the total mass of the bearing inner ring and the support shaft. This refers to the total mass of the bearing outer ring and bearing housing. These refer to the stiffness of the support shaft and the bearing housing, respectively. These are the damping coefficients of the support shaft and the bearing housing, respectively. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively. These represent the vibration displacements of the bearing outer ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing inner ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing outer ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing inner ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing outer ring in the X and Y directions, respectively. For the external load applied in the Y direction to the inner ring of the bearing, For the first Parameters of each rolling element located in the load zone For the first The total contact deformation of a rolling element at any angular position For the first The contact velocity of a rolling element at any angular position For the first The angular position of the center of each rolling element during its motion;

[0018] S14. Simplify the inner and outer raceway defects of the bearing as rectangular pits that run through the entire width of the raceway. Assuming that the rolling element does not contact the bottom of the defect, the movement trajectory of the rolling element through the raceway defect is characterized by a half-sine function. Calculate the displacement excitation representing the depth of the rolling element falling into the raceway defect, as shown in the following expression:

[0019] ;

[0020] In the formula, For the displacement excitation of the rolling element, To fall into the maximum depth of the defect, The angle of the rolling center corresponds to the defect. This represents the angular position of the defect relative to the Y-axis.

[0021] S15. Substituting the displacement excitation of the rolling elements into the four-degree-of-freedom dynamic equilibrium equations, we obtain the modified dynamic model of the wind turbine bearing, expressed as:

[0022] ;

[0023] In the formula, This refers to the radial clearance of the bearing. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively.

[0024] Furthermore, based on the actual geometry and operating parameters of the wind turbine bearing, as well as the principles of classical mechanics and kinematics, the parameters of each rolling element in the load zone, the total contact deformation of each rolling element at any angular position, the contact velocity of each rolling element at any angular position, and the angular position of the center of each rolling element during motion are calculated. The expressions for these calculations are as follows:

[0025] ;

[0026] ;

[0027] ;

[0028] ;

[0029] In the formula, For the radial clearance of the bearing, This refers to the number of rolling elements in the bearing. The initial position angle of the rolling element. To maintain the frame angular velocity, For time.

[0030] Furthermore, the feature extraction module includes a shallow feature extraction module, a mid-level feature extraction module, and a multi-scale deep feature extraction module, and the multi-scale deep feature extraction module is provided with a DepthConcatenation layer for splicing deep features.

[0031] Furthermore, the convolutional layers in the feature extraction module are used for data feature extraction. The feature maps output by the convolutional layers of each feature extraction module are calculated, as shown in the following expression:

[0032] ;

[0033] In the formula, For the first The output of the first convolutional layer Each feature map For activation function, The number of input features, For the first The first convolutional layer Each input feature For convolution kernel, For the first The convolutional layer corresponds to the first Bias terms of the output feature maps;

[0034] The feature maps output by the convolutional layers are compressed, and the feature maps output by the pooling layers of each feature extraction module are calculated, as shown in the following expression:

[0035] ;

[0036] In the formula, For the first The output of the first convolutional layer Each feature map This is the pooling function;

[0037] Based on convolutional kernels of different sizes in different branches, deep features at different scales are extracted. The deep features output from each branch are then concatenated using a DepthConcatenation layer. The expression for the concatenated deep features is as follows:

[0038] ;

[0039] In the formula, For deep features, The feature extraction results for each branch, This represents the number of branches within the module.

[0040] Furthermore, the feature fitting and classification module includes multiple fully connected layers. Image features obtained through convolution and pooling processes are integrated in the fully connected layers. Except for the last layer, each neuron in the remaining fully connected layers is... The output value after activation function, the output value of each neuron in the current fully connected layer and The activation functions and expressions are as follows:

[0041] ;

[0042] ;

[0043] In the formula, For the current fully connected layer The output value of each neuron For activation function, For the next level The number of neurons and the current layer Connection weights between neurons For the next fully connected layer The output value of each neuron For the current fully connected layer Bias terms for each neuron.

[0044] Furthermore, the statement based on the last fully connected layer The function classifies different samples, transforms the output data into predicted probabilities for each category, and calculates the predicted probabilities for each category of the bearing. The expression is as follows:

[0045] ;

[0046] In the formula, For bearing number The predicted probabilities of each category, For bearing number The original output values ​​for each category, This represents the total number of bearing categories. For the first Perform natural exponentiation on the original output value of the class.

[0047] Furthermore, in step S3, the specific steps include the following:

[0048] S31. Substitute the preprocessed target domain data and source domain data into the deep branch-transfer network model to perform feature extraction and classification prediction, and obtain the prediction probabilities of the source domain data and target domain data respectively.

[0049] S32. Based on the predicted probabilities of the source domain data and the target domain data, the cross-entropy loss function is used to calculate the cross-entropy loss of the source domain data and the target domain data, as well as the total model loss, respectively. The expressions are as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] In the formula, The cross-entropy loss is the loss of the source domain data. The true labels for the source domain data samples. The predicted probability of the source domain data samples. The cross-entropy loss is for the target domain data. The true labels for the target domain data samples. The predicted probability of the data samples in the target domain. This represents the total loss value of the model. The source domain loss weight coefficients are... For sample categories;

[0054] S33. Based on gradient descent, all trainable parameters of the source domain pre-training branch and the target domain fine-tuning branch are updated synchronously. The update expression is as follows:

[0055] ;

[0056] In the formula, For the first Network parameters after the next iteration For the first Network parameters after the next iteration For learning rate, The total loss function of the model is relative to the network parameters. The gradient;

[0057] S34. For the parameters of the mid-level feature extraction module in the multi-scale deep feature extraction module, calculate and update the gradient using the multi-branch gradient aggregation method. The gradient calculation expression is:

[0058] ;

[0059] In the formula, The parameters of the mid-layer feature extraction module are used as the overall loss function of the model. gradient, For the first The output features of each parallel branch affect the parameters. gradient, For the model's total loss function, the first... The gradient of the output feature of each parallel branch;

[0060] S35. Input the preprocessed target domain data and source domain data into the deep branch-transfer network model after the current iteration, and calculate the classification accuracy of the deep branch-transfer network model. The calculation expression is:

[0061] ;

[0062] In the formula, To measure the classification accuracy of the deep branch transfer network model, The number of validation set samples whose predictions match the true labels. This represents the total number of samples in the source domain validation set and the target domain validation set.

[0063] S36. Determine whether the deep branch transfer network model has been trained based on whether the classification accuracy is greater than a preset threshold or whether the number of iterations has reached a preset upper limit.

[0064] If so, then model training is complete;

[0065] If not, return to step S31, adjust the learning rate, and continue training.

[0066] Furthermore, the evaluation results include damage, lack of lubrication, and normal condition of raceway or rolling element parts;

[0067] If the damage is determined to be to raceway or rolling element parts, a stop command is triggered.

[0068] If an oil shortage is detected, a shutdown command will be triggered.

[0069] If the condition is determined to be normal, the bearing will continue to operate.

[0070] By employing the above technical solution, the present invention provides a method for evaluating the abnormality of bearings in vehicle-mounted traction wind turbines based on digital model driving, which has at least the following beneficial effects:

[0071] 1. This invention generates a massive amount of simulated dynamic response signals by constructing a four-degree-of-freedom dynamic model based on real geometry and operating parameters. This makes up for the lack of measured data on abnormal wind turbine bearings in vehicle-mounted scenarios. It eliminates the need to rely on a large amount of hard-to-obtain measured abnormal data, significantly reducing the time and economic costs of data acquisition. At the same time, the simulation data is constructed based on real parameters and considers time-varying displacement excitation, ensuring the adaptability of the data to operating conditions. This provides high-quality source domain data support for model training and avoids the problem of model underfitting due to insufficient data.

[0072] 2. This invention, through the construction of a deep branch transfer network model and the combination of a cross-domain training strategy, achieves efficient adaptation between simulation data and measured data, ensuring the synchronous learning of general features and real-world scene features. The layer-by-layer progressive feature extraction module can fully capture the differentiated features of different anomaly types. The combination of the cross-entropy loss function and the gradient descent method achieves precise optimization of model parameters, effectively eliminating the domain differences between simulation data and measured data. This enables the trained model to possess both the general feature learning ability given by simulation data and the real-world scene recognition ability adapted to measured data, significantly improving the accuracy and generalization performance of anomaly assessment.

[0073] 3. The deep branch transfer network model of this invention achieves gradual abstraction and refinement from shallow and medium-level basic features to deep multi-scale features through a layer-by-layer progressive feature extraction module. Compared with existing single-scale feature extraction models, it can more accurately capture the core feature differences of different anomaly types. The feature fitting and classification module further strengthens the mapping relationship between features and anomaly categories through the synergistic effect of fully connected layers and the Softmax function, ensuring that the model can accurately distinguish the three types of states being evaluated, avoiding misjudgment of anomaly types, and providing a reliable decision-making basis for subsequent operation and maintenance.

[0074] 4. This invention, by deploying the deep branch migration network model on the vehicle-mounted wind turbine site, can continuously return the steps of collecting measured data based on on-site operation and maintenance feedback, supplement the measured data, update the simulation model parameters, and then iteratively optimize the network model. It can dynamically adapt to changes such as drift and equipment aging in vehicle-mounted operating conditions, effectively solving the problem that existing models cannot be continuously updated after deployment and the long-term evaluation accuracy decreases, thus ensuring the long-term stability and reliability of wind turbine bearing anomaly evaluation. Attached Figure Description

[0075] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings:

[0076] Figure 1 This is a flowchart of the anomaly assessment method of the present invention;

[0077] Figure 2 This is a schematic diagram of the displacement function of the rolling element of the present invention as it passes through the raceway defect;

[0078] Figure 3 This is a structural diagram of the deep branch migration network model of the present invention. Detailed Implementation

[0079] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0080] Current bearing anomaly assessment methods based on measured data rely on a large amount of labeled measured anomaly data. However, abnormal operating conditions of vehicle-mounted traction wind turbine bearings are not numerous, resulting in a scarcity of measured anomaly data. This leads to insufficient model training data, resulting in low anomaly recognition accuracy and poor generalization ability. To compensate for the lack of measured data on wind turbine bearing anomalies in vehicle-mounted scenarios, ensure the adaptability of data operating conditions, provide high-quality measured data support for model training, and avoid model underfitting due to insufficient data, this invention proposes a model-driven method for assessing anomalies in vehicle-mounted traction wind turbine bearings. Figures 1-3 As shown, the method includes the following steps:

[0081] S1. Obtain limited prior anomaly measurement data of the wind turbine bearing. This is achieved by deploying vibration acceleration sensors on the end cover of the motor shaft extension of the vehicle-mounted wind turbine bearing. Based on the actual geometric and operating parameters of the wind turbine bearing, a four-degree-of-freedom dynamic model is constructed to compensate for the scarcity of measured anomaly data. This model acquires a large amount of simulated dynamic response signals that are highly compatible with real operating conditions. Displacement excitation is incorporated to ensure that the generated simulation signals can accurately reproduce the dynamic response characteristics of the bearing under different states. The displacement excitation under fault conditions is represented by a half-sine function, which eliminates the need for a large amount of difficult-to-obtain measured anomaly data, significantly reducing the time and economic costs of data acquisition while ensuring the data's adaptability to operating conditions. The specific steps include:

[0082] S11. Obtain finite prior anomaly measurement data of the wind turbine bearing under preset operating conditions. The preset operating conditions include rotational frequency (usually set to 3000 r / min), sampling frequency (usually set to 20 kHz), and sampling time (usually set to 30 s). The wind turbine bearing is usually a deep groove ball bearing.

[0083] S12. Based on the geometric characteristics of the wind turbine bearing and Hertz contact theory, calculate the comprehensive contact stiffness between the rolling elements of the wind turbine bearing and the raceway surfaces of the inner and outer rings, and establish the relationship between contact deformation and load, as shown in the following expressions:

[0084] ;

[0085] ;

[0086] In the formula, This refers to the overall contact stiffness between the rolling elements of the wind turbine bearing and the raceways of the inner and outer rings. The load-deformation coefficient, This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the inner ring. This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the outer ring. For Hertzian contact force, This represents the radial contact deformation of the contact pair. The radial contact velocity of the contact pair, This is the damping coefficient between the inner and outer raceways.

[0087] S13. Based on the concentrated spring-mass-damping model of the wind turbine bearing, construct the four-degree-of-freedom dynamic equilibrium equations of the wind turbine bearing under healthy conditions, as shown in the following expression:

[0088] ;

[0089] In the formula, This refers to the total mass of the bearing inner ring and the support shaft. This refers to the total mass of the bearing outer ring and bearing housing. These refer to the stiffness of the support shaft and the bearing housing, respectively. These are the damping coefficients of the support shaft and the bearing housing, respectively. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively. These represent the vibration displacements of the bearing outer ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing inner ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing outer ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing inner ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing outer ring in the X and Y directions, respectively. This refers to the external load applied in the Y direction to the inner ring of the bearing.

[0090] in, For the first Parameters of each rolling element located in the load zone For the first The total contact deformation of a rolling element at any angular position For the first The contact velocity of a rolling element at any angular position For the first The angular positions of the centers of the rolling elements during the motion are calculated using the following expressions:

[0091] ;

[0092] ;

[0093] ;

[0094] ;

[0095] In the formula, For the radial clearance of the bearing, This refers to the number of rolling elements in the bearing. The initial position angle of the rolling element. To maintain the frame angular velocity, For time.

[0096] S14. Simplify the defects in the inner and outer raceways of the bearing as rectangular pits that run through the entire width of the raceway. Assuming that the rolling element does not contact the bottom of the defect, the movement trajectory of the rolling element through the defect is characterized by a half-sine function. The variation law of the displacement excitation of the rolling element is determined by the depth of the defect into which the rolling element falls. The displacement excitation of the rolling element is calculated, and the expression is as follows:

[0097] ;

[0098] In the formula, For the displacement excitation of the rolling element, To fall into the maximum depth of the defect, The angle of the rolling center corresponds to the defect. This represents the angular position of the defect relative to the Y-axis.

[0099] S15. Considering the displacement excitation of the rolling elements, substituting the displacement excitation of the rolling elements into the four-degree-of-freedom dynamic equilibrium equations, we obtain the modified dynamic model of the wind turbine bearing, expressed as:

[0100] ;

[0101] In the formula, This refers to the radial clearance of the bearing. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively.

[0102] S2. Construct a deep branch transfer network model for extracting abnormal features of wind turbine bearing vibration signals. It includes two deep branch networks with consistent structures. Each deep branch network includes a feature extraction module and a feature fitting and classification module with progressive extraction at each level. This can ensure the synchronous learning of general features and real scene features. The progressive feature extraction module can fully capture the differentiated features of different anomaly types and achieve efficient adaptation between simulation data and measured data.

[0103] The feature extraction module includes a shallow feature extraction module, a mid-level feature extraction module, and a multi-scale deep feature extraction module, as well as three convolutional layers. The multi-scale deep feature extraction module has a DepthConcatenation layer for concatenating deep features. The convolutional layers in each feature extraction module are used for data feature extraction. The feature maps output by the convolutional layers of each feature extraction module are calculated using the following expression:

[0104] ;

[0105] In the formula, For the first The output of the first convolutional layer Each feature map For activation function, The number of input features, For the first The first convolutional layer Each input feature For convolution kernel, For the first The convolutional layer corresponds to the first The bias term of the output feature map.

[0106] The feature maps output by the convolutional layers are compressed, and the feature maps output by the pooling layers of each feature extraction module are calculated, as shown in the following expression:

[0107] ;

[0108] In the formula, For the first The output of the first convolutional layer Each feature map This is the pooling function.

[0109] Based on convolutional kernels of different sizes in different branches, deep features at different scales are extracted. The deep features output from each branch are then concatenated using a DepthConcatenation layer. The expression for the concatenated deep features is as follows:

[0110] ;

[0111] In the formula, For deep features, The feature extraction results for each branch, This represents the number of branches within the module.

[0112] The shallow feature extraction module extracts basic local features through convolution and pooling, while the mid-level feature extraction module extracts abstract local features through convolution and pooling. The multi-scale deep feature extraction module extracts multi-scale deep features through convolution and pooling using branches of convolution kernels of different sizes. Through the convolution and pooling process of shallow → mid → multi-scale deep layers, the module gradually extracts differential features from the details of the original signal to the abnormal types. Finally, the multi-scale features are concatenated through the DepthConcatenation layer, thereby achieving accurate differentiation of the three states of bearing evaluation.

[0113] The feature fitting and classification module includes multiple fully connected layers. Image features obtained from convolution and pooling processes are integrated in the fully connected layers. Except for the last layer, the neurons in the remaining fully connected layers are... The output value after activation function, the output value of each neuron in the current fully connected layer and The activation functions and expressions are as follows:

[0114] ;

[0115] ;

[0116] In the formula, For the current fully connected layer The output value of each neuron For activation function, For the next level The number of neurons and the current layer Connection weights between neurons For the next fully connected layer The output value of each neuron For the current fully connected layer Bias terms for each neuron.

[0117] Based on the last fully connected layer The function classifies different samples by converting the output data into probabilities in the range [0,1]. The function expression is:

[0118] ;

[0119] In the formula, For bearing number The predicted probabilities of each category, For bearing number The original output values ​​for each category, This represents the total number of bearing categories. For the first Perform natural exponentiation on the original output value of the class; through a fully connected layer and The synergistic effect of the functions further strengthens the mapping relationship between features and anomaly categories, ensuring that the model can accurately distinguish the three types of states being evaluated and avoiding misjudgment of anomaly types.

[0120] A feature extraction module is constructed, comprising a shallow feature extraction module, a mid-level feature extraction module, and a multi-scale deep feature extraction module. The feature extraction module and the feature fitting and classification module serve as the model pre-training branch and fine-tuning branch, respectively, forming a complete deep branch transfer network model. Through the collaborative design of general feature learning and real-world scenario adaptation, the complementary value of massive simulation data in the source domain and limited measured data in the target domain is fully utilized, effectively eliminating cross-domain differences and achieving accurate extraction and classification of abnormal features of wind turbine bearings in small-sample scenarios, ensuring the model's generalization ability and evaluation reliability.

[0121] S3. Using limited prior anomaly-measured data as the target domain data and massive simulated dynamic response signals as the source domain data, the target domain data and source domain data are preprocessed and then substituted into the deep branch-transfer network model. The model loss value is then calculated using the cross-entropy loss function, and the model parameters are iteratively updated using gradient descent to complete the training and optimization of the deep branch-transfer network model. By combining the cross-entropy loss function and gradient descent, accurate optimization of the model parameters is achieved, effectively eliminating the domain differences between simulated and measured data, thereby improving the accuracy and generalization performance of anomaly assessment. The specific steps include the following:

[0122] S31. Substitute the preprocessed target domain data and source domain data into the deep branch transfer network model to perform feature extraction and classification prediction, and obtain the prediction probabilities of the source domain data and target domain data respectively. This can provide core input for the calculation of the total loss of the model, support cross-domain collaborative learning and parameter optimization of the two branches, and at the same time quantify the model's feature learning and classification adaptation effect on the two types of data.

[0123] S32. Based on the predicted probabilities of the source domain data and the target domain data, the cross-entropy loss function is used to calculate the cross-entropy loss of the source domain data and the target domain data, as well as the total model loss, respectively. The expressions are as follows:

[0124] ;

[0125] ;

[0126] ;

[0127] In the formula, The cross-entropy loss is the loss of the source domain data. The true labels for the source domain data samples. The predicted probability of the source domain data samples. The cross-entropy loss is for the target domain data. The true labels for the target domain data samples. The predicted probability of the data samples in the target domain. This represents the total loss value of the model. The source domain loss weight coefficients are... For sample categories.

[0128] S33. Based on gradient descent, all trainable parameters of the source domain pre-training branch and the target domain fine-tuning branch are updated synchronously. The update expression is as follows:

[0129] ;

[0130] In the formula, For the first Network parameters after the next iteration For the first Network parameters after the next iteration For learning rate, The total loss function of the model is relative to the network parameters. The gradient.

[0131] S34. For the parameters of the mid-level feature extraction module in the multi-scale deep feature extraction module, calculate and update the gradient using the multi-branch gradient aggregation method. The gradient calculation expression is:

[0132] ;

[0133] In the formula, The parameters of the mid-layer feature extraction module are used as the overall loss function of the model. gradient, For the first The output features of each parallel branch affect the parameters. gradient, For the model's total loss function, the first... The gradient of the output feature of each parallel branch.

[0134] S35. Input the preprocessed target domain data and source domain data into the deep branch-transfer network model after the current iteration, and calculate the classification accuracy of the deep branch-transfer network model. The calculation expression is:

[0135] ;

[0136] In the formula, To measure the classification accuracy of the deep branch transfer network model, The number of validation set samples whose predictions match the true labels. This represents the total number of samples in the source domain validation set and the target domain validation set.

[0137] S36. There are two conditions for determining that the deep branch transfer network model training is complete: the classification accuracy is greater than a preset threshold and the number of iterations reaches a preset upper limit. If either of the above conditions is met, the model training is complete; otherwise, return to step S31, adjust the learning rate, and continue training.

[0138] S4. Deploy the trained and optimized deep branch transfer network model at the vehicle-mounted wind turbine site, perform anomaly assessment on the wind turbine bearing status, execute corresponding maintenance operations based on the preset assessment results, and return to step S1; if it is determined that the raceway or rolling element parts are damaged, trigger a shutdown command to stop and replace the wind turbine bearing; if it is determined that there is a lack of oil, trigger a shutdown command to stop and replenish the lubricating oil; if it is determined to be in a normal state, keep the bearing running.

[0139] This anomaly assessment method acquires finite prior anomaly measurement data of wind turbine bearings. Based on the actual geometric and operating parameters of the wind turbine bearings, a four-degree-of-freedom dynamic model is constructed to obtain massive simulated dynamic response signals. A deep branch transfer network model, including two structurally consistent deep branch networks, is constructed. The finite prior anomaly measurement data (target domain data) and massive simulated dynamic response signals (source domain data) are preprocessed and then substituted into the deep branch transfer network model. The model loss value is then calculated using the cross-entropy loss function, and the model parameters are iteratively updated using the gradient descent method to complete the training and optimization of the deep branch transfer network model. The trained and optimized deep branch transfer network model is deployed at the wind turbine site to assess the anomaly status of the wind turbine bearings. Based on the preset assessment results, corresponding operation and maintenance operations are executed, and the process of obtaining wind turbine bearing data is returned, forming a logical closed loop.

[0140] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0141] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0142] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for evaluating bearing anomalies in a vehicle-mounted traction fan based on digital model driving, characterized in that, The method includes the following steps: S1. Obtain finite prior anomaly measurement data of wind turbine bearings, and construct a four-degree-of-freedom dynamic model based on the actual geometric and operating parameters of the wind turbine bearings to obtain massive simulation dynamic response signals. S2. Construct a deep branch transfer network model for extracting abnormal features of wind turbine bearing vibration signals. It includes two structurally consistent deep branch networks. The deep branch networks include a feature extraction module and a feature fitting and classification module that extract features layer by layer. The feature extraction module includes a shallow feature extraction module, a mid-level feature extraction module, and a multi-scale deep feature extraction module, as well as three convolutional layers. The multi-scale deep feature extraction module has a DepthConcatenation layer for splicing deep features. The convolutional layers in each layer of the feature extraction module are used for extracting data features. Calculate the feature maps output by the convolutional layers of each feature extraction module. S3. Using finite prior abnormal measured data as the target domain data and massive simulated dynamic response signals as the source domain data, preprocessing is performed, and then the data is substituted into the deep branch-transfer network model. The cross-entropy loss function is then used to calculate the model loss value, and the model parameters are iteratively updated to complete the training and optimization of the deep branch-transfer network model. The specific steps in step S3 include the following: S31. Substitute the preprocessed target domain data and source domain data into the deep branch-transfer network model to perform feature extraction and classification prediction, and obtain the prediction probabilities of the source domain data and target domain data respectively. S32. Based on the predicted probabilities of the source domain data and the target domain data, the cross-entropy loss function is used to calculate the cross-entropy loss of the source domain data and the target domain data, as well as the total model loss, respectively. The expressions are as follows: ； ； ； In the formula, The cross-entropy loss is the loss of the source domain data. The true labels for the source domain data samples. The predicted probability of the source domain data samples. The cross-entropy loss is for the target domain data. The true labels for the target domain data samples. The predicted probability of the data samples in the target domain. This represents the total loss value of the model. The source domain loss weight coefficients are... For sample categories; S33. Based on gradient descent, all trainable parameters of the source domain pre-training branch and the target domain fine-tuning branch are updated synchronously. The update expression is as follows: ； In the formula, For the first Network parameters after the next iteration For the first Network parameters after the next iteration For learning rate, The total loss function of the model is relative to the network parameters. The gradient; S34. For the parameters of the mid-level feature extraction module in the multi-scale deep feature extraction module, calculate and update the gradient using the multi-branch gradient aggregation method. The gradient calculation expression is: ； In the formula, The parameters of the mid-layer feature extraction module are used as the overall loss function of the model. gradient, For the first The output features of each parallel branch affect the parameters. gradient, For the model's total loss function, the first... The gradient of the output feature of each parallel branch; S35. Input the preprocessed target domain data and source domain data into the deep branch-transfer network model after the current iteration, and calculate the classification accuracy of the deep branch-transfer network model. The calculation expression is: ； In the formula, To measure the classification accuracy of the deep branch transfer network model, The number of validation set samples whose predictions match the true labels. This represents the total number of samples in the source domain validation set and the target domain validation set. S36. Determine whether the deep branch transfer network model has been trained based on whether the classification accuracy is greater than a preset threshold or whether the number of iterations has reached a preset upper limit. If so, then model training is complete; If not, return to step S31, adjust the learning rate, and continue training; S4. Deploy the trained and optimized deep branch transfer network model at the vehicle-mounted wind turbine site, perform anomaly assessment on the wind turbine bearing status, execute corresponding operation and maintenance operations based on the preset assessment results, and return to step S1.

2. The anomaly assessment method according to claim 1, characterized in that, Step S1 specifically includes the following steps: S11. Obtain finite prior anomaly measurement data of the wind turbine bearing under preset operating conditions, wherein the preset operating conditions include rotational frequency, sampling frequency and sampling time. S12. Based on the geometric characteristics of the wind turbine bearing and Hertz contact theory, calculate the comprehensive contact stiffness between the rolling elements of the wind turbine bearing and the raceway surfaces of the inner and outer rings, and establish the relationship between contact deformation and load. The expressions are as follows: ； ； In the formula, This refers to the overall contact stiffness between the rolling elements of the wind turbine bearing and the raceways of the inner and outer rings. The load-deformation coefficient, This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the inner ring. This refers to the contact stiffness between the rolling elements of the wind turbine bearing and the raceway of the outer ring. For Hertzian contact force, This represents the radial contact deformation of the contact pair. The radial contact velocity of the contact pair, This is the damping coefficient between the inner and outer raceways; S13. Based on the concentrated spring-mass-damping model of the wind turbine bearing, construct the four-degree-of-freedom dynamic equilibrium equations of the wind turbine bearing under healthy conditions, as shown in the following expression: ； In the formula, This refers to the total mass of the bearing inner ring and the support shaft. This refers to the total mass of the bearing outer ring and bearing housing. These refer to the stiffness of the support shaft and the bearing housing, respectively. These are the damping coefficients of the support shaft and the bearing housing, respectively. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively. These represent the vibration displacements of the bearing outer ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing inner ring in the X and Y directions, respectively. These represent the vibration velocities of the bearing outer ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing inner ring in the X and Y directions, respectively. These represent the vibration accelerations of the bearing outer ring in the X and Y directions, respectively. For the external load applied in the Y direction to the inner ring of the bearing, For the first Parameters of each rolling element located in the load zone For the first The total contact deformation of a rolling element at any angular position For the first The contact velocity of a rolling element at any angular position For the first The angular position of the center of each rolling element during its motion; S14. Simplify the inner and outer raceway defects of the bearing as rectangular pits that run through the entire width of the raceway. Assuming that the rolling element does not contact the bottom of the defect, the movement trajectory of the rolling element through the raceway defect is characterized by a half-sine function. Calculate the displacement excitation representing the depth of the rolling element falling into the raceway defect, as shown in the following expression: ； In the formula, For the displacement excitation of the rolling element, To fall into the maximum depth of the defect, The angle of the rolling center corresponds to the defect. This represents the angular position of the defect relative to the Y-axis. S15. Substituting the displacement excitation of the rolling elements into the four-degree-of-freedom dynamic equilibrium equations, we obtain the modified dynamic model of the wind turbine bearing, expressed as: ； In the formula, This refers to the radial clearance of the bearing. These represent the vibration displacements of the bearing inner ring in the X and Y directions, respectively.

3. The anomaly assessment method according to claim 2, characterized in that, Based on the actual geometry and operating parameters of the wind turbine bearing, as well as the principles of classical mechanics and kinematics, the parameters of each rolling element in the load zone, the total contact deformation of each rolling element at any angular position, the contact velocity of each rolling element at any angular position, and the angular position of the center of each rolling element during motion are calculated. The expressions for these calculations are as follows: ； ； ； ； In the formula, For the radial clearance of the bearing, This refers to the number of rolling elements in the bearing. The initial position angle of the rolling element. To maintain the frame angular velocity, For time.

4. The anomaly assessment method according to claim 1, characterized in that, The convolutional layers in the feature extraction module are used to extract data features. The feature maps output by the convolutional layers of each feature extraction module are calculated using the following expression: ； In the formula, For the first The output of the first convolutional layer Each feature map For activation function, The number of input features, For the first The first convolutional layer Each input feature For convolution kernel, For the first The convolutional layer corresponds to the first Bias terms of the output feature maps; The feature maps output by the convolutional layers are compressed, and the feature maps output by the pooling layers of each feature extraction module are calculated, as shown in the following expression: ； In the formula, For the first The output of the first convolutional layer Each feature map This is the pooling function; Based on convolutional kernels of different sizes in different branches, deep features at different scales are extracted. The deep features output from each branch are then concatenated using a DepthConcatenation layer. The expression for the concatenated deep features is as follows: ； In the formula, For deep features, The feature extraction results for each branch, This represents the number of branches within the module.

5. The anomaly assessment method according to claim 4, characterized in that, The feature fitting and classification module includes multiple fully connected layers. Image features obtained from convolution and pooling processes are integrated in the fully connected layers. Except for the last layer, each neuron in the remaining fully connected layers is... The output value after activation function, the output value of each neuron in the current fully connected layer and The activation functions and expressions are as follows: ； ； In the formula, For the current fully connected layer The output value of each neuron For activation function, For the next level The number of neurons and the current layer Connection weights between neurons For the next fully connected layer The output value of each neuron For the current fully connected layer Bias terms for each neuron.

6. The anomaly assessment method according to claim 5, characterized in that, Based on the last fully connected layer The function classifies different samples, transforms the output data into predicted probabilities for each category, and calculates the predicted probabilities for each category of bearings. The expression is as follows: ； In the formula, For bearing number The predicted probabilities of each category, For bearing number The original output values ​​for each category, This represents the total number of bearing categories. For the first Perform natural exponentiation on the original output value of the class.

7. The anomaly assessment method according to claim 1, characterized in that, The evaluation results include damage, lubrication deficiency, and normal status of raceway or rolling element parts; If the damage is determined to be to raceway or rolling element parts, a stop command is triggered. If an oil shortage is detected, a shutdown command will be triggered. If the condition is determined to be normal, the bearing will continue to operate.

Images