Wind turbine gearbox bearing temperature status detection method based on deep learning model
Through the wind turbine gearbox bearing temperature state detection method based on deep learning model, the weighted sum and clustering technology is used to solve the problem of low prediction accuracy caused by bearing parameter data fluctuation, and the accuracy and accuracy of the detection results are improved.
Patent Information
- Application Number
- CN202411132301.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-08-19
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-08-19
AI Technical Summary
In the prior art, the large fluctuation in bearing parameters data leads to low model prediction accuracy, large error in detection results, and low accuracy.
The temperature state detection method of wind turbine gearbox bearing based on deep learning model is adopted. By training the classification network model and the timing network model respectively, the accuracy and prediction accuracy of the trained model are obtained, and the bearing is calculated using the weighted sum method. The probability of temperature abnormality further improves the accuracy of the detection results through the fluctuation degree clustering of the difference sequence.
The accuracy of bearing temperature state detection is improved, the error of detection results is reduced, and the prediction accuracy of the model is improved.
Smart Images

Figure CN118656665B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bearing detection, and in particular to a method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model. Background Art
[0002] The gearbox bearing in a wind turbine is one of the key components, and its health directly affects the operating safety and performance of the entire wind turbine. The traditional regular maintenance method is usually for operators to check and maintain the bearings one by one, which increases maintenance costs and prolongs downtime, resulting in low bearing maintenance efficiency.
[0003] The Chinese patent application document with publication number CN116256170A discloses a bearing status detection method, device and computer-readable storage medium. The bearing status detection method includes: obtaining a vibration signal of the bearing; performing frequency domain conversion on the vibration signal to obtain a frequency signal in the frequency domain; determining the alarm characteristic index corresponding to the frequency band to which the frequency signal in the frequency domain belongs in multiple frequency bands; and reflecting the health status of the bearing based on the alarm characteristic index corresponding to the frequency band to which it belongs. The above patent provides a technical solution for judging the health status of the bearing using the bearing vibration signal, and in the prior art, the health status of the bearing can be predicted by setting a neural network model.
[0004] However, when the above-mentioned related technical scheme uses the neural network model to predict the bearing, the vibration signal of the bearing is obtained and the vibration signal is processed to obtain the state of the bearing. However, when collecting the parameter data of the bearing, due to the influence of the external environment or the influence of the sensor itself, the collected parameter data of the bearing has a certain volatility, that is, the parameter data of the bearing is not in a stable state. In the prediction through the neural network model, if the parameter data involved in the training tends to a stable line, the accuracy of the model prediction is higher. If the amplitude of the trained parameter data on the curve changes greatly, the difficulty of the model to make predictions is higher and the accuracy is lower. Therefore, data fluctuations have a certain impact on the results of obtaining the bearing state, resulting in large errors in the detection results and low accuracy. Summary of the invention
[0005] In order to solve the above-mentioned problem of low model prediction accuracy caused by large fluctuations in bearing parameters, the purpose of the present invention is to provide a method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model.
[0006] The method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model provided by the present invention comprises the following steps:
[0007] The pre-built classification network model and time series network model are trained respectively to obtain the accuracy of the trained classification network model and the prediction accuracy of the time series network model;
[0008] Inputting the real-time collected bearing parameters into the trained classification network model and the time series network model respectively to obtain the first abnormal probability and the second abnormal probability; the bearing parameters include bearing temperature data;
[0009] The accuracy rate and the prediction accuracy rate are used to perform weighted summation on the first abnormal probability and the second abnormal probability respectively, and the final bearing temperature abnormal probability is obtained;
[0010] The prediction accuracy is as follows: obtaining difference sequences corresponding to multiple groups of bearing temperature data sequences during the training process of the time series network model, and obtaining the fluctuation degree of each difference sequence; clustering the fluctuation degree to obtain multiple cluster clusters; the bearing temperature data sequence corresponds to the fluctuation degree one by one;
[0011] According to the prediction results of the bearing temperature data series corresponding to each fluctuation degree in the cluster, the accuracy of each cluster is obtained; the accuracy of the cluster corresponding to the fluctuation degree of the bearing temperature data collected in real time is used as the prediction accuracy.
[0012] The beneficial effects of the present invention are as follows: the corresponding difference sequence is obtained through the bearing temperature sequence, the fluctuation degree of the difference sequence is further obtained, the fluctuation degree is clustered, the prediction accuracy of the timing network model is obtained by using the clustering clusters, and the prediction results are weighted and summed using the accuracy of the classification network model and the prediction accuracy of the timing network model to obtain the final bearing temperature abnormality probability, thereby improving the accuracy of the detection results.
[0013] In one embodiment, the specific process of training the temporal network model is as follows:
[0014] Obtain a bearing temperature data sequence from the bearing temperature data, and set a corresponding label for the sequence;
[0015] Input the bearing temperature data sequence of preset length and the corresponding label into the pre-built time series network model, and the loss function of the time series network model is the cross entropy loss;
[0016] Update the model parameters using a gradient descent algorithm, and terminate the training in response to the time series network model reaching a set maximum number of training times or the time series network model loss being less than a set threshold;
[0017] Get the trained temporal network model.
[0018] In one embodiment, before training the time series network model, a step of determining the optimal length of the bearing temperature data sequence is also included, specifically:
[0019] a. Use a pre-built sliding window to slide along the bearing temperature data sequence to obtain multiple groups of bearing temperature data sequences of preset lengths;
[0020] b. Obtaining the unsteady probability of each group of bearing temperature data sequences of the preset length according to the unit root test method, and calculating the variance of multiple unsteady probabilities; calculating the unsteadiness of the corresponding bearing temperature data sequence using the unsteady probability and the variance of the unsteady probability;
[0021] c. Change the length of the sliding window and repeat steps ab to obtain multiple degrees of unsteadiness. The sliding window length corresponding to the minimum degree of unsteadiness is taken as the optimal length of the bearing temperature data sequence.
[0022] The beneficial effects of the present invention are as follows: by calculating the probability of non-stationarity of the bearing temperature data sequence and the variance of the non-stationarity probability, the non-stationarity of the bearing temperature sequence is finally calculated, and the optimal length of the bearing temperature data sequence is obtained according to the non-stationarity, so that the stationarity of the input bearing temperature data sequence can be made constant, so that the timing network model can predict the correct result.
[0023] In one embodiment, the unsteadiness of the corresponding bearing temperature data sequence is calculated by using the unsteady probability and the variance of the unsteady probability, and the calculation method is:
[0024] Where M is the unsteadiness of the bearing temperature data sequence, represents the probability of non-stationarity of the bearing temperature data sequence in the sliding window when the sliding window slides for the i-th time, represents the variance of all non-stationary probabilities, is the summation formula.
[0025] The beneficial effects of the present invention are as follows: the formula can be used to accurately quantify the unsteadiness of the bearing temperature data sequence, thereby improving the calculation accuracy of the unsteadiness of the bearing temperature data sequence, so as to facilitate determination of the optimal length of the bearing temperature data sequence.
[0026] In one embodiment, the training method of the classification network model is:
[0027] Obtain a data set, which includes bearing parameters and labels, wherein the bearing parameters also include ambient temperature data and bearing speed data;
[0028] The classification network model is trained using the data set to obtain a trained classification network model.
[0029] In one embodiment, the fluctuation degree calculation formula of the difference sequence is:
[0030]
[0031] Among them, ent represents the volatility of the difference sequence, Represents the probability of the jth fluctuation value occurring.
[0032] The beneficial effects of the present invention are as follows: the fluctuation degree of the difference sequence can be accurately quantified by using the formula, thereby improving the calculation accuracy of the difference degree, so as to facilitate clustering of the fluctuation degree and improve the clustering accuracy.
[0033] In one embodiment, the bearing temperature abnormality probability is calculated by the following formula:
[0034]
[0035] in, Indicates the probability of abnormal bearing temperature during real-time detection, Represents the accuracy of the classification network model, represents the prediction accuracy of the temporal network model, represents the first abnormal probability of bearing temperature output by the classification network model, Represents the second abnormal probability of bearing temperature output by the time series network model.
[0036] The beneficial effect of the present invention is that by weighted summing the first abnormal probability and the second abnormal probability respectively, the abnormal probability of the bearing temperature can be accurately calculated, thereby improving the accuracy of the calculation result of the bearing abnormal probability.
[0037] In one embodiment, the classification network model is a support vector machine model.
[0038] Based on the above scheme, the scheme of the present invention determines the optimal length of the bearing temperature data sequence so that the bearing temperature data sequence input to the model has a higher stability, further enabling the timing network model to output accurate prediction results, thereby improving the prediction accuracy of the timing network model; calculating the difference sequence of the bearing temperature data sequence, calculating the degree of fluctuation of the difference sequence, clustering the degree of fluctuation to obtain clustering clusters, and obtaining the prediction accuracy of the timing network model through the clustering clusters, and at the same time combining the accuracy of the classification network model, weightedly summing the first abnormal probability and the second abnormal probability respectively, thereby improving the accuracy of the detection results. BRIEF DESCRIPTION OF THE DRAWINGS
[0039] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0040] Figure 1 The flowchart schematically shows the steps of the method for detecting the temperature status of the gearbox bearing of a wind turbine generator set based on the deep learning model of the present invention. DETAILED DESCRIPTION
[0041] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the specific implementation methods, structures, features and effects of the technical solutions proposed by the present invention are described in detail below in conjunction with the accompanying drawings and preferred embodiments. In the following description, different one embodiment or another embodiment does not necessarily refer to the same embodiment.
[0042] Figure 1 The flowchart schematically shows the steps of the method for detecting the temperature status of the gearbox bearing of a wind turbine generator set based on the deep learning model of the present invention.
[0043] A method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model comprises the following steps:
[0044] S1. Train the pre-built classification network model and timing network model respectively to obtain the accuracy of the classification network model and the prediction accuracy of the timing network model.
[0045] S11. Train the classification network model.
[0046] Collect bearing parameters at the same time, including: ambient temperature data, bearing speed data, bearing temperature data, and set labels for the bearing parameters at the same time.
[0047] The bearing temperature and ambient temperature are measured by installing a temperature probe at the bearing of the gearbox of the wind turbine. The bearing temperature is the temperature of the outer ring of the bearing, and the ambient temperature is the temperature of the outer surface of the gearbox. The speed of the shaft installed on the inner ring of the bearing is measured by a photoelectric sensor. It should be noted that the speed of the shaft is the speed of the inner ring of the bearing. The collected historical ambient temperature data, bearing speed data, and bearing temperature data constitute a data set. The data set contains multiple groups of bearing parameter data, each group of bearing parameter data includes ambient temperature data, bearing speed data, and bearing temperature data. Each group of data corresponds to a label, and its label data is 0 or 1, where 0 represents a normal bearing and 1 represents an abnormal bearing. The data set and the corresponding labels are input into the classification network model, and the classification network model is trained to obtain a trained classification network model.
[0048] The bearing parameter data is input into the trained classification network model, and the output is the probability that the bearing state is 1. In other words, the output is the probability that the bearing is abnormal, and this probability is used as the first abnormality probability.
[0049] The reason for constructing a classification network model using ambient temperature data, bearing speed data, and bearing temperature data is that in a high temperature environment, the wind turbine gearbox cannot fully dissipate heat, resulting in a high temperature of the bearing, which in turn affects the working efficiency and service life of the bearing; in a low temperature environment, the bearing temperature is low, and the lubricating oil cannot be fully melted, resulting in poor lubrication of the bearing, increasing the friction when the bearing rotates, and making the bearing temperature high; the rotation of the shaft during the operation of the wind turbine increases the friction of the bearing, resulting in an increase in the bearing temperature, affecting the working efficiency and service life of the bearing. Therefore, ambient temperature and bearing speed can affect the bearing temperature state. The classification network model is one of a perceptron, a support vector machine, a decision tree, and a random forest. In this embodiment, the classification network model is a support vector machine.
[0050] S12. Calculate the accuracy of the trained classification network model.
[0051] The input of the trained classification network model is ambient temperature data, bearing speed data, and bearing temperature data, and the output is the bearing abnormality probability; during the training process, the output result of the classification network model is compared and calculated with the actual result to calculate the accuracy of the classification network model.
[0052] For example, if the probability of bearing abnormality is greater than the preset threshold value 0.6, the bearing temperature is abnormal, and if it is less than the preset threshold value 0.6, the bearing temperature is normal. The number of data with normal bearing temperature output by the statistical classification network model is 8, and the number of data with abnormal bearing temperature is 2. There are 7 normal bearing temperature labels and 3 abnormal bearing temperature labels in the data set. The accuracy of the classification network model is 1-|8-7| / (7+3)=90%.
[0053] S13, the step of determining the optimal length of the bearing temperature data sequence is specifically:
[0054] a. Use a pre-built sliding window to slide along the bearing temperature data sequence to obtain multiple groups of bearing temperature data sequences of preset lengths;
[0055] b. Obtain the unsteady probability of each group of bearing temperature data sequences of the preset length according to the unit root test method, and calculate the variance of multiple unsteady probabilities; calculate the unsteadiness of the corresponding bearing temperature data sequence using the unsteady probability and the variance of the unsteady probability, and the unsteadiness calculation method is:
[0056] Where M is the unsteadiness of the bearing temperature data sequence, represents the probability of non-stationarity of the bearing temperature data sequence in the sliding window when the sliding window slides for the i-th time, represents the variance of all non-stationary probabilities, ∑ is the summation formula;
[0057] c. Change the length of the sliding window and repeat steps ab to obtain multiple degrees of unsteadiness. The sliding window length corresponding to the minimum degree of unsteadiness is taken as the optimal length of the bearing temperature data sequence.
[0058] Among them, the unit root test method is a statistical test method, which is mainly used to detect whether there is a unit root in time series data. The existence of a unit root means that the time series is non-stationary, that is, its mean and variance change over time. Obtaining the probability of non-stationarity based on the unit root test method is an existing technology and will not be elaborated here.
[0059] Exemplarily, the bearing temperature data sequence is ( , , , , … ), the length of the sliding window is 3, the sliding window slides along the bearing temperature data sequence, and multiple groups of bearing temperature data sequences with a length of 3 are obtained, which are ( , , )、( , , )、( , , )……( , , ), according to the unit root test method, the corresponding non-stationary probability of multiple groups of length sequences is obtained , , … , calculate the probability of non-stationarity , , … Variance Finally, the unstationary degree of the sequence with a length of 3 is calculated according to the unstationary degree formula .
[0060] The length of the sliding window is adjusted to 4, and the sliding window slides along the bearing temperature data sequence to obtain multiple groups of bearing temperature data sequences with a length of 4, which are respectively ( , , , )、( , , , )、( , , , )……( , , , ), according to the unit root test method, the corresponding non-stationary probability of multiple groups of length sequences is obtained , , … , calculate the probability of non-stationarity , , … Variance Finally, the unstationary degree of the sequence with a length of 4 is calculated according to the unstationary degree formula .
[0061] Finally, we get multiple groups of unsteadiness , , ..., take the sequence length corresponding to the minimum value as the optimal length of the bearing temperature data sequence. For example, the optimal length of the bearing temperature data sequence is 3. It should be noted that the smaller the value and the sum of the variance of the non-stationary probability are, the more stable the bearing temperature data sequence is.
[0062] When training a time series network model, the stationarity of the input data affects the accuracy of the model. When other conditions are the same, stationary data is more likely to produce a model with better detection effects. Stationary sequences are more likely to be captured by the model with regularities and trends, and can usually produce more accurate prediction results.
[0063] S14. Train the temporal network model.
[0064] Construct a time series network model; obtain a bearing temperature data sequence of optimal length from a data set, wherein the bearing temperature data sequence is a plurality of groups; input the bearing temperature data sequence of optimal length and the label corresponding to the sequence into the time series network model, wherein the label is the label corresponding to the bearing temperature data at the last moment of the sequence, and the label is the abnormal probability label, and the label data is 0 or 1, wherein 0 represents a normal bearing and 1 represents an abnormal bearing; the loss function of the model is a cross entropy loss; the model parameters are updated using a gradient descent algorithm, and the training is terminated in response to the time series network model reaching a set maximum number of training times or the loss of the time series network model being less than a set threshold; and a trained time series network model is obtained.
[0065] The bearing temperature data sequence is input into the trained time series network model, and the output is the probability that the bearing state is 1. In other words, the output is the probability that the bearing is abnormal, and this probability is used as the second abnormality probability.
[0066] The time series network model is one of an autoregressive model, a moving average model, an exponential smoothing model, and a neural network model. In this embodiment, the time series network model is a long short-term memory network model in the neural network model.
[0067] S15, calculating the prediction accuracy of the time series prediction network model, including the following steps:
[0068] In one embodiment, the difference sequences corresponding to the multiple groups of bearing parameter sequences in the training process of the time series network model are obtained to obtain the fluctuation degree of each difference sequence. The function for calculating the fluctuation degree is:
[0069]
[0070] Among them, ent represents the volatility of the difference sequence, It represents the probability of the j-th fluctuation value occurring. The logarithmic function in the function expression of the fluctuation degree takes 2 as the base, or other numbers as the base. The fluctuation value is the difference between two adjacent values in the difference sequence.
[0071] For example, for the difference sequence ( , , , , ), the fluctuation values are ( , , , ), the probability of fluctuation value 1 is 3 / 4=75%, and the probability of fluctuation value 2 is 1 / 4=25%.
[0072] The fluctuation degree is clustered to obtain multiple clusters; the accuracy of each cluster is obtained based on the prediction result of the bearing parameter sequence corresponding to the difference sequence corresponding to each fluctuation degree in the cluster; the accuracy of the cluster corresponding to the fluctuation degree of the bearing parameter data collected in real time is used as the prediction accuracy.
[0073] For example, the length of the bearing temperature data sequence is 3, and the multiple groups of bearing temperature data sequences are respectively ( , , )、( , , )、( , , )……( , , ), the difference sequence corresponding to the multiple groups of bearing temperature data sequences obtained by the difference method is ( - , - )、( - , - )、( - , - ),……,( - , - ), calculate the volatility of the difference sequence based on the function, and get the volatility of the difference sequence , , ,…… , cluster the results to get multiple clusters. For example, the number of clusters is three, namely , , It should be noted that, for the sake of convenience, it is assumed that the number of clusters is three. In actual operation, the specific result can be four or more. Then, the accuracy corresponding to the clusters is calculated. Assuming that the clusters The number of samples in is 10, of which 8 are correctly predicted by the temporal network model. Then the cluster The corresponding accuracy is 80%.
[0074] In other embodiments, a validation set is used to validate the trained time series prediction network model. The validation set is multiple groups of bearing temperature data sequences and labels. The validation set is input into the time series prediction network model to obtain prediction results. The ratio of the number of samples with correct prediction results to the total number of samples is used as the accuracy rate. For example, there are 20 groups of bearing temperature data sequences, and 15 groups have correct prediction results. Then the accuracy rate of the prediction of the time series network model is 75%.
[0075] S2. Input the bearing parameters collected in real time into the trained classification network model and the time series network model respectively to obtain the first abnormal probability and the second abnormal probability.
[0076] The real-time collected ambient temperature data, bearing speed data, and bearing temperature data are input into the classification network model, and the classification network model outputs the first abnormality probability; the real-time collected bearing temperature data sequence is input into the timing network model, and the timing network model outputs the second abnormality probability. The first abnormality probability and the second abnormality probability are decimals between 0 and 1. The larger the value, the closer the bearing temperature is to abnormality.
[0077] S3. Use the accuracy rate and the prediction accuracy rate to perform weighted summation on the first abnormal probability and the second abnormal probability respectively to obtain a final bearing temperature abnormal probability.
[0078] Calculate the difference sequence of the bearing temperature data sequence collected in real time and the fluctuation degree of the difference sequence; calculate the cluster corresponding to the fluctuation degree; and record the accuracy corresponding to the cluster as the prediction accuracy of the time series network model at this time.
[0079] For example, the difference sequence corresponding to the collected bearing temperature data sequence is calculated, and then the fluctuation degree corresponding to the difference sequence is calculated. The difference between the fluctuation degree and the center value of each cluster is calculated, and the cluster with the smallest difference is taken as the cluster corresponding to the bearing temperature data sequence. The accuracy of this cluster is used as the prediction accuracy of the time series network model at this time. Assume that the cluster is , then the cluster The accuracy rate is the prediction accuracy rate corresponding to the temporal network model at this time.
[0080] The final calculation formula for the bearing temperature abnormality probability is:
[0081] ;
[0082] in, Indicates the probability of abnormal bearing temperature during real-time detection, Represents the accuracy of the classification network model, represents the prediction accuracy of the temporal network model, represents the first abnormal probability of bearing temperature output by the classification network model, Indicates the second abnormal probability of the bearing temperature output by the time series network model. When the abnormal probability of the bearing temperature detected in real time is greater than the set threshold, an alarm is issued in time. For example, the set threshold is 0.7. When the abnormal probability of the bearing temperature detected in real time is greater than 0.7, it means that the bearing temperature is in an abnormal state at this time, and an alarm is issued in time.
[0083] Although this specification has shown and described a number of embodiments of the present invention, it will be apparent to those skilled in the art that such embodiments are provided by way of example only. Those skilled in the art will conceive of many modifications, changes and alternatives without departing from the ideas and spirit of the present invention. It should be understood that in the practice of the present invention, various alternatives to the embodiments of the present invention described herein may be employed.
Claims
1. A method for detecting the temperature status of a wind turbine gearbox bearing based on a deep learning model, characterized in that: The following steps are involved: The pre-built classification network model and time series network model are trained respectively to obtain the accuracy of the classification network model and the prediction accuracy of the time series network model; Inputting the real-time collected bearing parameters into the classification network model and the timing network model respectively to obtain the first abnormal probability and the second abnormal probability; the bearing parameters include bearing temperature data; The accuracy rate and the prediction accuracy rate are used to perform weighted summation on the first abnormal probability and the second abnormal probability respectively, and the final bearing temperature abnormal probability is obtained; The prediction accuracy is as follows: obtaining difference sequences corresponding to multiple groups of bearing temperature data sequences during the training process of the time series network model, calculating the fluctuation degree of each difference sequence; clustering the fluctuation degree to obtain multiple cluster clusters; According to the prediction results of the bearing temperature data series corresponding to each fluctuation degree in the cluster, the accuracy of each cluster is obtained; the accuracy of the cluster corresponding to the fluctuation degree of the bearing temperature data collected in real time is used as the prediction accuracy; Before training the time series network model, the step of determining the optimal length of the bearing temperature data sequence is also included, specifically: a. Use a pre-built sliding window to slide along the bearing temperature data sequence to obtain multiple groups of bearing temperature data sequences of preset lengths; b. Obtaining the unsteady probability of each group of bearing temperature data sequences of the preset length according to the unit root test method, and calculating the variance of multiple unsteady probabilities; calculating the unsteadiness of the corresponding bearing temperature data sequence using the unsteady probability and the variance of the unsteady probability; c. Change the length of the sliding window, repeat steps ab to obtain multiple degrees of unsteadiness, and take the sliding window length corresponding to the minimum degree of unsteadiness as the optimal length of the bearing temperature data sequence; The unsteadiness of the corresponding bearing temperature data sequence is calculated by using the unsteady probability and the variance of the unsteady probability, and the calculation method is: in, is the unsteadiness of the bearing temperature data series, represents the probability of non-stationarity of the bearing temperature data sequence in the sliding window when the sliding window slides for the i-th time, represents the variance of all non-stationary probabilities; The calculation formula for the volatility of the difference sequence is: Among them, ent represents the volatility of the difference sequence, Represents the probability of the jth fluctuation value occurring.
2. The method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model according to claim 1, characterized in that: The training method of the temporal network model is: Obtain a bearing temperature data sequence from the bearing temperature data, and set a corresponding label for the sequence; Input the bearing temperature data sequence of preset length and the corresponding label into the pre-built time series network model, and the loss function of the time series network model is the cross entropy loss; Update the model parameters using a gradient descent algorithm, and terminate the training in response to the time series network model reaching a set maximum number of training times or the time series network model loss being less than a set threshold; Get the trained temporal network model.
3. The method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model according to claim 1, characterized in that: The training method of the classification network model is: Obtain a data set, which includes bearing parameters and labels, wherein the bearing parameters also include ambient temperature data and bearing speed data; The classification network model is trained using the data set to obtain a trained classification network model.
4. The method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model according to claim 1, characterized in that: The bearing temperature abnormality probability is calculated by the following formula: in, Indicates the probability of abnormal bearing temperature during real-time detection, Represents the accuracy of the classification network model, represents the prediction accuracy of the temporal network model, represents the first abnormal probability of bearing temperature output by the classification network model, Represents the second abnormal probability of bearing temperature output by the time series network model.
5. The method for detecting the temperature state of a wind turbine gearbox bearing based on a deep learning model according to claim 1, characterized in that: The classification network model is a support vector machine model.
