Unit health assessment method based on deep learning

Abstract

The invention discloses a unit health assessment method based on deep learning. The method comprises the following steps: (1) acquiring and preprocessing original data of a unit; (2) key variables are selected based on a physical model and expert knowledge, and feature extraction is carried out through an automatic encoder and other technologies; (3) optimizing and improving a cabin temperature model, a gearbox temperature model and a generator bearing temperature model on the basis of GRU time sequence prediction, performing model training by using historical data, and optimizing model parameters; and finally obtaining a cabin temperature model, a gearbox temperature model and a generator bearing temperature model. (4) performing health assessment on the unit by using the three models; according to the invention, more comprehensive and reliable health condition evaluation is provided.

Description

Technical Field

[0001] The present invention relates to the technical field of wind power generation equipment maintenance and management, and specifically to a unit health assessment method based on deep learning. Background Art

[0002] Traditional assessment methods for the health of wind turbines mainly include vibration monitoring, temperature monitoring, and performance parameter analysis. The vibration monitoring method relies on the vibration sensors of key components to collect data to determine faults. The disadvantages are that the data is difficult to interpret, is greatly affected by the environment, has high collection and processing costs, is limited in dimension, and has poor real-time performance. The temperature monitoring method uses temperature sensors to track the temperature of heating parts. The problem is that it indirectly reflects the situation, is insensitive to potential faults, and the sensor accuracy and ambient temperature will interfere with the judgment. The performance parameter analysis method is to find operating anomalies by analyzing the performance indicators of the unit. The disadvantages are that the diagnosis of complex faults is weak, it is easily affected by external factors, and the perception of early faults is poor. Summary of the invention

[0003] Purpose of the invention: The purpose of the present invention is to provide a unit health assessment method based on deep learning, which solves the problems of insufficient processing of high-dimensional noise data, weak dynamic feature capture ability, poor correlation of related factors, etc. of traditional assessment methods by integrating multi-dimensional data denoising, time series feature mining and multi-factor collaborative analysis, so as to achieve comprehensive and accurate assessment and forward-looking prediction of the unit health status.

[0004] Technical solution: The unit health assessment method based on deep learning described in the present invention comprises the following steps: (1) Obtain the original data of the unit and perform preprocessing; (2) Feature selection: Select key variables based on physical models and expert knowledge, and extract features through techniques such as autoencoders; (3) Model selection and construction: Based on the optimization and improvement of GRU time series prediction, the cabin temperature model, gearbox temperature model, and generator bearing temperature model are selected, and historical data are used to train the models and optimize the model parameters. Finally, the three models of "cabin temperature model", "gearbox temperature model", and "generator bearing temperature model" are obtained; (4) Use three models to conduct health assessment of the unit.

[0005] Furthermore, in step (1), the original data of the unit includes: generator-related temperature data: average generator winding temperature, average generator nacelle side bearing temperature, average generator speed, average generator inlet water temperature, average generator stator side outlet water temperature 1, average generator rotor side outlet water temperature 2, average generator gearbox side bearing temperature; grid-related data: average grid voltage L1N, L2N, L3N, average grid current L1, L2, L3, average grid active power, average grid reactive power; blade angle data: average blade angle 3A, average blade angle 3B, average blade angle 1A , average blade angle 1B, average blade angle 2A, average blade angle 2B; cabin related data: average cabin temperature, average cabin humidity, average cabin air-to-air cooling inner circulation outlet temperature, average cabin air-to-air cooling inner circulation inlet temperature, average cabin air-to-air cooling outer circulation inlet temperature, average cabin air-to-air cooling outer circulation outlet temperature; other temperature data: average hub temperature, average outdoor temperature, average gearbox main bearing temperature, average gearbox oil temperature, average gearbox coolant temperature, average gearbox generator side bearing temperature; gearbox related data: average gearbox main pump oil temperature, average gearbox cooling pump outlet pressure.

[0006] Furthermore, in step (1), the preprocessing is as follows: cleaning the data, removing noise and outliers; and dividing the data into training set and test set.

[0007] Furthermore, in step (3), the GRU network is used to establish three modules, namely, "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model", including: setting the number of GRU units to 128, optimizing the model parameters using the Adam optimizer, and adjusting the model by the root mean square error and the average error; wherein, the output variable of the "cabin temperature model" is the average cabin temperature; the output variable of the "gearbox temperature model" is the average gearbox hub side bearing temperature; and the output variable of the "generator bearing temperature model" is the average generator gearbox side bearing temperature.

[0008] Furthermore, in step (4), real-time wind turbine data is imported, and the model output data of step 3 is used, combined with characteristic data such as "average generator speed", "average grid active power", "average grid current L1", and "average grid voltage L1N", and the unit health is evaluated using the quartile method.

[0009] The present invention provides a unit health assessment system based on deep learning, comprising: Acquisition module: used to acquire the original data of the unit and perform preprocessing; Selection module: used to select key variables based on physical models and expert knowledge, and extract features through technologies such as autoencoders; Model building module: used to optimize and improve the cabin temperature model, gearbox temperature model, and generator bearing temperature model based on GRU time series prediction, use historical data to train the model and optimize the model parameters; finally, the three models of "cabin temperature model", "gearbox temperature model", and "generator bearing temperature model" are obtained; Assessment module: used to conduct health assessment of the unit.

[0010] Furthermore, in the acquisition module, the generator-related temperature data: average generator winding temperature, average generator nacelle side bearing temperature, average generator speed, average generator water inlet temperature, average generator stator side water outlet temperature 1, average generator rotor side water outlet temperature 2, average generator gearbox side bearing temperature; grid-related data: average grid voltage L1N, L2N, L3N, average grid current L1, L2, L3, average grid active power, average grid reactive power; blade angle data: average blade angle 3A, average blade angle 3B, average blade angle 1A, average blade Angle 1B, average blade angle 2A, average blade angle 2B; cabin related data: average cabin temperature, average cabin humidity, average cabin air-to-air cooling inner circulation outlet temperature, average cabin air-to-air cooling inner circulation inlet temperature, average cabin air-to-air cooling outer circulation inlet temperature, average cabin air-to-air cooling outer circulation outlet temperature; other temperature data: average hub temperature, average outdoor temperature, average gearbox main bearing temperature, average gearbox oil temperature, average gearbox coolant temperature, average gearbox generator side bearing temperature; gearbox related data: average gearbox main pump oil temperature, average gearbox cooling pump outlet pressure.

[0011] Furthermore, in the step acquisition module, the preprocessing is as follows: cleaning the data, removing noise and outliers; and dividing the data into training sets and test sets.

[0012] Furthermore, in the model building module, the GRU network is used to establish three modules, namely "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model", including: setting the number of GRU units to 128, optimizing the model parameters using the Adam optimizer, and adjusting the model by the root mean square error and the average error; among which, the output variable of the "cabin temperature model" is the average cabin temperature; the output variable of the "gearbox temperature model" is the average gearbox hub side bearing temperature; and the output variable of the "generator bearing temperature model" is the average generator gearbox side bearing temperature.

[0013] Furthermore, in the evaluation module, real-time wind turbine data is imported, and the model output data of step 3 is used, combined with characteristic data such as "average generator speed", "average grid active power", "average grid current L1", and "average grid voltage L1N", and the unit health is evaluated using the quartile method.

[0014] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: applying deep learning algorithms to the health assessment of wind turbines can automatically extract features from complex time series data and capture nonlinear relationships, which can provide more accurate fault predictions compared to traditional statistical methods or shallow machine learning algorithms. By providing early warning of potential problems, maintenance can be arranged at the best time to minimize unplanned downtime and ensure the stability of power supply. Combining multi-dimensional data, a comprehensive assessment of the status of the entire machine can be achieved, rather than relying solely on a single indicator, thereby providing a more comprehensive and reliable health status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 is a flow chart of the present invention; Figure 2 is the cabin temperature model creation process of the present invention; Figure 3 This is the cabin temperature prediction model effect of the present invention; Figure 4 The gearbox hub side bearing temperature model creation process of the present invention; Figure 5 This is the gearbox hub side bearing temperature prediction model effect of the present invention; Figure 6 The generator bearing temperature model creation process of the present invention; Figure 7 This is the effect of the generator bearing temperature prediction model of the present invention; Figure 8 is the unit health assessment process of the present invention; Fig. 9 This is the unit health scoring effect of the present invention. DETAILED DESCRIPTION

[0016] The technical solution of the present invention is further described below in conjunction with the accompanying drawings.

[0017] like Figure 1 As shown, an embodiment of the present invention provides a distributed wind and solar access distribution network voltage control method based on reinforcement learning, comprising the following steps: S1. Data preprocessing: Clean the raw operating data from the system, detect, correct or delete damaged, inaccurate or unsuitable data from the data, remove noise and outliers, and make the data adapt to the model and match the model's requirements.

[0018] S2. Feature selection: Select key variables based on physical models and expert knowledge, and further refine features through technologies such as autoencoders.

[0019] S3, Model selection and construction: Based on the optimization and improvement model of GRU (Gated Recurrent Unit) time series prediction, historical data is used to train the model and optimize the model parameters. Finally, three models are obtained: "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model".

[0020] (31) Cabin temperature modeling like Figure 2 As shown in the figure, the cabin temperature model is established using the GRU optimization and improved version (where the optimal model parameters Adam model parameters are: 0x000001AC7BEF63D0, number of GRU units: 128, mean square error (MSE): 0.0635, mean absolute error (MAE): 0.2069, root mean square error (RMSE): 0.252), and the model is trained using the training set data. The model obtained in the model folder is the model file of the "cabin temperature model". Use the generated "cabin temperature model" to import the test set data and predict the cabin temperature.

[0021] Among them, the output variable is the average cabin temperature, and the input variables are shown in Table 1.

[0022] Table 1 Description of input variables for cabin temperature modeling ; The model is built under Python 3.9, and the required libraries and their versions are shown in Table 2.

[0023] Table 2 Libraries and versions required for model building ; like Figure 3 As shown in Figure 2, the overall trend of the predicted value curve and the actual value curve is highly similar, and the fit is high. In the process of changing from 0 to 2000 time steps, the fluctuation trends of the two are basically the same.

[0024] The error indicators of the model also show that it has high accuracy. The mean absolute error (MAE) is 0.20689453488484848, the mean square error (MSE) is 0.06342892797911112, and the root mean square error (RMSE) is 0.25195892611090625. The values ​​are relatively small, indicating that the deviation between the predicted value and the actual value of the model is small, and the average cabin temperature can be predicted more accurately.

[0025] (32) Gearbox temperature modeling like Figure 4As shown in the figure, the gearbox temperature model is established by using the GRU optimization and improved version (where the optimal model parameters Adam model parameters are: 0x000002100A88AF40, number of GRU units: 128, mean square error (MSE): 0.0107, mean absolute error (MAE): 0.0826, root mean square error (RMSE): 0.1032), and the model is trained using the training set data. The model obtained in the model folder is the model file of the "gearbox temperature model". Import the test set data and use the trained "gearbox temperature model" to predict the gearbox temperature.

[0026] Among them, the output variable is the average gearbox hub side bearing temperature, and the input variables are shown in Table 3.

[0027] Table 3 Description of input variables of gearbox temperature prediction model ; The model is built under Python 3.9, and the required libraries and their versions are shown in Table 4.

[0028] Table 4 Libraries and versions required for model building ; like Figure 5 As shown in the figure, the fluctuation trends of the actual value curve and the predicted value curve are relatively similar. In the range of time steps from 0 to 2000, the two closely follow each other.

[0029] From the error indicators, the mean absolute error (MAE) is 0.08263912380776989, the mean square error (MSE) is 0.010654813326130275, and the root mean square error (RMSE) is 0.10322215520967519. The values ​​are relatively small, indicating that the deviation between the model prediction value and the actual value is low. The model can predict the average gearbox temperature more accurately.

[0030] (33) Generator bearing temperature modeling like Figure 6 As shown in the figure, the generator bearing temperature model is established by using the GRU optimization and improved version (where the optimal model parameters Adam model parameters are: 0x0000026205587D00, number of GRU units: 128, mean square error (MSE): 0.0029, mean absolute error (MAE): 0.0444, root mean square error (RMSE): 0.0535), and the training set data is used to train the model. The model obtained in the model folder is the model file of the "generator gearbox side bearing temperature model". Import the test set data and use the "generator gearbox side bearing temperature model" to predict the generator bearing temperature.

[0031] The output variable is the average generator gearbox side bearing temperature, and the input variables are shown in Table 5.

[0032] Table 5 Description of input variables for generator bearing temperature modeling ; The model is built under Python 3.9, and the required libraries and their versions are shown in Table 6.

[0033] Table 6 Libraries and versions required for model building ; like Figure 7 As shown in the figure, comparing the actual value curve and the predicted value curve, the overall trend of the two is highly consistent. In the time step range from 0 to 2000, the fluctuation trend over time is basically the same.

[0034] From the error indicators, the mean absolute error (MAE) is 0.04439538816541983, the mean square error (MSE) is 0.0028634280521347318, and the root mean square error (RMSE) is 0.05351100870040418. These values ​​are small, indicating that the error between the model prediction value and the actual value is small, and the model can accurately predict the average generator gearbox side bearing temperature.

[0035] S4, such as Figure 8 As shown, real-time wind turbine data is imported, and the model output data of step 3 is used, combined with characteristic data such as "average generator speed", "average grid active power", "average grid current L1", and "average grid voltage L1N", and the quartile method is used to evaluate the unit health. The details are as follows: The unit health score uses real-time wind turbine data to call the "generator bearing temperature model", "gearbox temperature model" and "nacelle temperature model" to obtain the "average gearbox hub side bearing temperature", "average generator gearbox side bearing temperature" and "average nacelle temperature" respectively. Combined with real-time wind turbine data such as "average generator speed", the quartile method is used to evaluate the health of the wind turbine.

[0036] The output variable is the unit health score, and the input variables are shown in Table 7.

[0037] Table 7 Selected measurement points for unit health score ; The relative degradation matrix is ​​calculated according to the following formula: (1); Where: x is the actual monitored value of the variable; [xmin , x max ] is the normal operating range of the monitored quantity.

[0038] Then the computer group health score value is calculated according to the following formula: (2); The quartile method is used to determine whether the value of each variable is normal, and three types of judgments are given: [exceeding the upper limit, normal, exceeding the lower limit]. The quartile calculation method is as follows: Calculate the second quartile, that is, the median Q2: (3); Calculate the first and third quartiles Q1 and Q3. When n=2k (k=0,1,2...), divide the data sample X into two parts from Q2. Q2 is not included in the two parts of data. Calculate the median of the two parts according to formula (3) and ,but .when When: (4); When n=4k+3, we have: (5); By interquartile range: (6); The inner limit of outliers in the data sample X can be determined as: (7); Health score calculation: Based on the model output results, combined with the quartile method and other statistical indicators, the health score of each component of the computer group is calculated.

[0039] Specific test: Environment configuration: Use Pandas, Numpy, Matplotlib, sklearn, Keras-TensorFlow and other libraries to build the development platform in Python 3.9 environment.

[0040] Data import: Read pre-processed wind turbine operation data from CSV files.

[0041] Model loading and application: Three independent GRU network models are established for the cabin temperature, gearbox temperature, and generator bearing temperature, and saved as .h5 format model files. By loading these models, the temperature change trend in the future can be predicted, and then the maximum and minimum values ​​of the "average gearbox hub side bearing temperature", "average generator gearbox side bearing temperature", "average cabin temperature" calculated by the model and the original features "average generator speed", "average grid active power", "average grid current L1", "average grid voltage L1N" are selected. The upper and lower quartiles of each variable in the historical data are calculated. According to the formula in the modeling process, it is judged whether each variable in the actual machine operation data exceeds the upper and lower thresholds and the results are given. The health degree calculates the degradation degree according to the value of each parameter, and the health score is calculated by the formula.

[0042] like Fig. 9 As shown in the unit health evaluation effect diagram, the curve trend is relatively smooth and coherent, without drastic fluctuations. This shows that the model's evaluation at different data points is relatively stable, without excessive fluctuations, ensuring the consistency of the evaluation results.

[0043] By processing the data using the quartile method, we can effectively eliminate interference from extreme values, making the health assessment more representative and reliable. Judging from the trend presented by the curve, the model can more reasonably reflect the changing trend of the unit's health.

Claims

1. A unit health assessment method based on deep learning, characterized in that: The following steps are involved: (1) Obtain the original data of the unit and perform preprocessing; (2) Feature selection: Select key variables based on physical models and expert knowledge, and extract features through autoencoders; (3) Model selection and construction: Based on the optimization and improvement of GRU time series prediction, the "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model" models are selected, and historical data are used to train the models and optimize the model parameters; finally, the "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model" are obtained; (4) Use three models to conduct health assessment of the unit.

2. The unit health assessment method based on deep learning according to claim 1, characterized in that: In step (1), the original data of the unit include: generator-related temperature data: average generator winding temperature, average generator nacelle side bearing temperature, average generator speed, average generator inlet water temperature, average generator stator side outlet water temperature 1, average generator rotor side outlet water temperature 2, average generator gearbox side bearing temperature; grid-related data: average grid voltage L1N, L2N, L3N, average grid current L1, L2, L3, average grid active power, average grid reactive power; blade angle data: average blade angle 3A, average blade angle 3B, average blade angle 1A, average Blade angle 1B, average blade angle 2A, average blade angle 2B; cabin related data: average cabin temperature, average cabin humidity, average cabin air-to-air cooling inner circulation outlet temperature, average cabin air-to-air cooling inner circulation inlet temperature, average cabin air-to-air cooling outer circulation inlet temperature, average cabin air-to-air cooling outer circulation outlet temperature; other temperature data: average hub temperature, average outdoor temperature, average gearbox main bearing temperature, average gearbox oil temperature, average gearbox coolant temperature, average gearbox generator side bearing temperature; gearbox related data: average gearbox main pump oil temperature, average gearbox cooling pump outlet pressure.

3. The unit health assessment method based on deep learning according to claim 2, characterized in that: In step (1), the preprocessing is as follows: cleaning the data, removing noise and outliers; and dividing the data into training set and test set.

4. The unit health assessment method based on deep learning according to claim 1, characterized in that: In step (3), the GRU network is used to establish three modules, namely, "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model", including: setting the number of GRU units to 128, optimizing the model parameters using the Adam optimizer, and adjusting the model using the root mean square error and the average error; wherein, the output variable of the "cabin temperature model" is the average cabin temperature; the output variable of the "gearbox temperature model" is the average gearbox hub side bearing temperature; and the output variable of the "generator bearing temperature model" is the average generator gearbox side bearing temperature.

5. The unit health assessment method based on deep learning according to claim 4 is characterized in that: In step (4), real-time wind turbine data is imported, and the model output data of step (3) is used, combined with the characteristic data of "average generator speed", "average grid active power", "average grid current L1", and "average grid voltage L1N", and the unit health is evaluated using the quartile method.

6. A unit health assessment system based on deep learning, characterized in that: include: Acquisition module: used to obtain the original data of the unit and perform preprocessing; Selection module: used to select key variables based on physical models and expert knowledge, and perform feature extraction through autoencoders; Model building module: used to optimize and improve the "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model" based on GRU time series prediction, use historical data to train the model and optimize the model parameters; finally, the three models of "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model" are obtained; Assessment module: used to conduct health assessment of the unit.

7. The unit health assessment system based on deep learning according to claim 6, characterized in that: In the acquisition module, the generator-related temperature data: average generator winding temperature, average generator nacelle side bearing temperature, average generator speed, average generator inlet water temperature, average generator stator side outlet water temperature 1, average generator rotor side outlet water temperature 2, average generator gearbox side bearing temperature; grid-related data: average grid voltage L1N, L2N, L3N, average grid current L1, L2, L3, average grid active power, average grid reactive power; blade angle data: average blade angle 3A, average blade angle 3B, average blade angle 1A, average blade angle 1 B. Average blade angle 2A, average blade angle 2B; Cabin related data: average cabin temperature, average cabin humidity, average cabin air-to-air cooling inner circulation outlet temperature, average cabin air-to-air cooling inner circulation inlet temperature, average cabin air-to-air cooling outer circulation inlet temperature, average cabin air-to-air cooling outer circulation outlet temperature; Other temperature data: average hub temperature, average outdoor temperature, average gearbox main bearing temperature, average gearbox oil temperature, average gearbox coolant temperature, average gearbox generator side bearing temperature; Gearbox related data: average gearbox main pump oil temperature, average gearbox cooling pump outlet pressure.

8. The unit health assessment system based on deep learning according to claim 7, characterized in that: In the step acquisition module, the preprocessing is as follows: cleaning the data, removing noise and outliers; and dividing the data into training sets and test sets.

9. The unit health assessment system based on deep learning according to claim 6, characterized in that: In the model building module, the GRU network is used to establish three modules, namely "cabin temperature model", "gearbox temperature model" and "generator bearing temperature model", including: setting the number of GRU units to 128, using Adam optimizer to optimize model parameters, and adjusting the model by root mean square error and average error; among them, the output variable of the "cabin temperature model" is the average cabin temperature; the output variable of the "gearbox temperature model" is the average gearbox hub side bearing temperature; the output variable of the "generator bearing temperature model" is the average generator gearbox side bearing temperature.

10. The unit health assessment system based on deep learning according to claim 9, characterized in that: In the evaluation module, real-time wind turbine data is imported, and the model output data of the model building module is used. Combined with the characteristic data of "average generator speed", "average grid active power", "average grid current L1", and "average grid voltage L1N", the unit health is evaluated using the quartile method.

Images