Wind turbine bearing life prediction method and device based on a hybrid model

Abstract

The present application relates to a method and device for predicting the bearing life of a wind turbine based on a hybrid model. The method includes: performing data conversion on the relevant operation data set of the wind turbine, dividing the data set in the form of amplitude probability density into a training set and a test set, and training a pre-constructed random forest regression model according to the training set; training a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indicators of the extracted training set, predicting the health indicators at the next moment according to the health indicator values of each sample before the prediction starting point in the test set and the health indicator prediction model, completing the prediction of the health indicator values at multiple future moments by cyclically carrying out the single-step prediction step, and using the health indicator threshold to judge the predicted health indicator values at multiple future moments, determining the bearing failure moment, and calculating the bearing life prediction value according to the time difference between the current moment and the failure moment. Using this method can improve the accuracy of bearing life prediction.

Description

Technical Field

[0001] The present application relates to the technical field of data processing, and particularly to a method and device for predicting the bearing life of a wind turbine based on a hybrid model. Background Art

[0002] With the intensification of environmental pollution and energy crisis, the development of clean, green and renewable new energy has become a global consensus. Among them, wind energy, with its advantages of wide distribution, huge reserves and small ecological impact, has become one of the main forces in the development of new energy. To maximize production efficiency, wind turbines are mostly deployed in areas with rich wind energy resources such as high altitudes and the sea, and thus face more arduous operation and maintenance challenges. Aging is an inevitable life process of mechanical components. As the usage time of the wind turbine increases, the fatigue damage, wear and other states of the bearing components will further deepen, affecting their performance and ultimately leading to scrapping (i.e., reaching the end of life). The components within each subsystem of the wind turbine are closely coordinated. The aging of the bearing components not only affects a single subsystem, but also spreads to the whole wind turbine. The unknown and sudden scrapping phenomenon will seriously endanger the operation safety and stability of the wind turbine. Therefore, accurate life prediction of bearings is an important support for ensuring the operation safety of wind turbines.

[0003] However, current bearing life prediction methods are mainly divided into methods based on physical models and methods based on data-driven. The monitoring signals generated during the use of bearings are an important information source for current prediction work. Therefore, the data-driven bearing life prediction method explores the aging trend starting from the monitoring signals, and is the mainstream method for current wind turbine bearing life prediction work. Among them, the data-driven methods can be roughly divided into methods based on reliability, time series, and machine learning. Their forms and processes are diverse and have been proven to have a certain prediction effect. However, the existing work generally has problems such as fixed model forms, weak generalization ability, and insufficient aging trend representation, resulting in insufficient reflection of trend information in the prediction results and limited prediction accuracy. Summary of the Invention

[0004] Based on this, it is necessary to provide a method and device for predicting the bearing life of a wind turbine based on a hybrid model that can improve the accuracy of bearing life prediction for the above technical problems.

[0005] A method for predicting the bearing life of a wind turbine based on a hybrid model, the method comprising:

[0006] Obtain a relevant operation data set of the wind turbine; perform data conversion on the relevant operation data set of the wind turbine to obtain a data set in the form of amplitude probability density;

[0007] Divide the amplitude probability density form dataset into a training set and a test set, and train a pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indexes;

[0008] Set a health index threshold for judging bearing failure according to the development law of the extracted health indexes, and train a prediction model of the 2D-CNN-BiLSTM hybrid structure according to the health indexes of the extracted training set to obtain a health index prediction model;

[0009] Use the health index construction model to construct health indexes for the test set data to obtain the health index values of each sample before the prediction starting point in the test set;

[0010] Predict the health index of the next moment according to the health index values of each sample before the prediction starting point in the test set and the health index prediction model, complete the prediction of health index values at multiple future moments through a loop of single-step prediction steps, and use the health index threshold to judge the predicted health index values at multiple future moments to determine the bearing failure moment, and calculate the bearing life prediction value according to the time difference between the current moment and the failure moment.

[0011] In one embodiment, learn the change trend of the bearing historical health indexes on the test set and update the health index prediction model, and predict the wind turbine bearing life in real time according to the updated health index prediction model.

[0012] In one embodiment, use the health index construction model to extract the historical health index values on the test set; use the health index prediction model trained with the training set as the basic model, and fine-tune the basic model with the historical health index values on the test set to complete the learning of the historical index value trend and obtain the updated health index prediction model; use the updated health index prediction model to perform single-step prediction of index values in a loop until the health index prediction value reaches the health index threshold to calculate the life prediction result of the model updated in real time.

[0013] In one embodiment, perform data conversion on the wind turbine related operation dataset to obtain an amplitude probability density form dataset, including:

[0014] Use the kernel density estimation method to calculate the probability density distribution of the wind turbine related operation dataset, sample within the amplitude range, and select the probability density values at several amplitudes as the samples after data conversion; the samples are the amplitude probability density form dataset.

[0015] In one embodiment, train a pre-constructed random forest regression model according to the training set to obtain a health index construction model, including:

[0016] Use a quartic function model to label the samples in the training set to obtain sample labels;

[0017] Take a number of probability density values obtained from data conversion as input and the sample labels as the target output to train a random forest regression model, and use the model output value as the constructed health index to obtain a health index construction model.

[0018] In one embodiment, train a prediction model of a 2D-CNN-BiLSTM hybrid structure according to the health index of the extracted training set to obtain a health index prediction model, including:

[0019] Sample the extracted health index with a fixed time window into a number of one-dimensional health index sequences; use a data reconstruction method to convert the one-dimensional health index sequences into a two-dimensional form to obtain two-dimensional time series samples;

[0020] Use the two-dimensional time series samples to train a prediction model of a 2D-CNN-BiLSTM hybrid structure, and complete the model training with the reconstructed two-dimensional data as the input and the health index value at the future moment as the target output to obtain a health index prediction model.

[0021] In one embodiment, according to the health index values of each sample before the prediction starting point in the test set and the health index prediction model, predict the health index at the next moment, complete the prediction of the health index values at multiple future moments through a loop of single-step prediction steps, and use the health index threshold to judge the predicted health index values at multiple future moments to determine the bearing failure moment, and calculate the bearing life prediction value according to the time difference between the current moment and the failure moment, including:

[0022] After determining the prediction starting point on the test set, extract the health index values at each moment within the previous time window of the starting moment, reconstruct the original one-dimensional health index sequence within the time window into two-dimensional data, and use the health index prediction model to receive the two-dimensional data and calculate and output the health index prediction value at the next moment, judge whether the prediction value reaches the health index threshold, if it does not reach the health index threshold, incorporate the prediction value into the original sequence, and at the same time use the sliding window method to update the sequence sample and repeat the data reconstruction and prediction steps until the health index prediction value reaches the health index threshold. When the health index prediction value reaches the health index threshold, perform the life prediction work, and calculate the duration experienced from the prediction starting point to the prediction end point as the bearing life prediction value.

[0023] A wind turbine bearing life prediction device based on a hybrid model, the device includes:

[0024] A data conversion module, configured to obtain a wind turbine-related operation data set; perform data conversion on the wind turbine-related operation data set to obtain an amplitude probability density form data set;

[0025] A health index construction model training module, which is used to divide the amplitude probability density form data set into a training set and a test set, and train a pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indexes;

[0026] A health index prediction model training module, which is used to set a health index threshold for judging bearing failure according to the development law of the extracted health indexes, and train a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indexes of the extracted training set to obtain a health index prediction model;

[0027] A health index construction module, which is used to construct health indexes for the test set data by using the health index construction model to obtain the health index values of each sample before the prediction starting point in the test set;

[0028] A bearing life prediction module, which is used to predict the health indexes at the next moment according to the health index values of each sample before the prediction starting point in the test set and the health index prediction model, complete the prediction of the health index values at multiple future moments through a loop of single-step prediction steps, and use the health index threshold to judge the predicted health index values at multiple future moments to determine the bearing failure moment, and calculate the bearing life prediction value according to the time difference between the current moment and the failure moment.

[0029] A computer device, including a memory and a processor, the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0030] Obtain the relevant operation data set of the wind turbine; perform data conversion on the relevant operation data set of the wind turbine to obtain a data set in the form of amplitude probability density;

[0031] Divide the data set in the form of amplitude probability density into a training set and a test set, and train a pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indexes;

[0032] Set a health index threshold for judging bearing failure according to the development law of the extracted health indexes, and train a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indexes of the extracted training set to obtain a health index prediction model;

[0033] Use the health index construction model to construct health indexes for the test set data to obtain the health index values of each sample before the prediction starting point in the test set;

[0034] Predict the health indicator at the next moment based on the health indicator values of each sample before the predicted starting point in the test set and the health indicator prediction model. Complete the prediction of health indicator values at multiple future moments by performing single-step prediction steps in a loop, and use the health indicator threshold to judge the predicted health indicator values at multiple future moments to determine the bearing failure moment. Calculate the predicted bearing life value based on the time difference between the current moment and the failure moment.

[0035] A computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the following steps are implemented:

[0036] Obtain the operation data set related to the wind turbine; perform data conversion on the operation data set related to the wind turbine to obtain a data set in the form of amplitude probability density;

[0037] Divide the data set in the form of amplitude probability density into a training set and a test set, and train the pre-constructed random forest regression model according to the training set to obtain a health indicator construction model; the health indicator construction model is used to extract health indicators;

[0038] Set the health indicator threshold for judging bearing failure according to the development law of the extracted health indicators, and train the prediction model of the 2D-CNN-BiLSTM hybrid structure according to the health indicators of the extracted training set to obtain a health indicator prediction model;

[0039] Use the health indicator construction model to construct health indicators for the test set data to obtain the health indicator values of each sample before the predicted starting point in the test set;

[0040] Predict the health indicator at the next moment based on the health indicator values of each sample before the predicted starting point in the test set and the health indicator prediction model. Complete the prediction of health indicator values at multiple future moments by performing single-step prediction steps in a loop, and use the health indicator threshold to judge the predicted health indicator values at multiple future moments to determine the bearing failure moment. Calculate the predicted bearing life value based on the time difference between the current moment and the failure moment.

[0041] The above-mentioned wind turbine bearing life prediction method and device based on a hybrid model. In this application, the bearing life prediction work is completed by adopting a health index construction model based on random forest and a health index single-step prediction model based on a 2D-CNN-BiLSTM hybrid structure. Among them, the health index construction model based on random forest can directly extract the health index of the device at the current moment from the probability density information of vibration data. The convolutional neural network part in the 2D-CNN-BiLSTM hybrid structure can refine key information from the original complex high-dimensional data and weaken the influence of noise. Introducing the convolutional part to process the original time series data in the health index single-step prediction model can effectively suppress the influence of the noise contained in the original time series, refine the time series information to help improve the prediction effect. The bidirectional long short-term memory network part in the 2D-CNN-BiLSTM hybrid structure completes the two-way transmission of data by stacking two layers of LSTM, comprehensively considering the information of the previous and subsequent moments in the process of processing time series data, showing better time series prediction performance. The combination of the convolutional neural network and the bidirectional long short-term memory network can effectively dig deep into the development trend of the health index, realize the accurate prediction of the health index, and help complete the life prediction. Using the trained random forest model to carry out the health index construction work on the test set data, the model calculates and outputs the health index values of each sample before the prediction starting point in the test set. Using the calculated health index values of the test set and the trained 2D-CNN-BiLSTM model to predict the health index at the next moment, the prediction of the health index values at multiple future moments is completed by repeatedly carrying out the single-step prediction steps. And using the set failure threshold to judge the predicted future health index values, determining the bearing failure moment, and calculating the bearing life prediction value according to the time difference between the current moment and the failure moment. By designing the idea of cyclic single-step prediction of the health index to achieve the prediction task, it is possible to provide a judgment on the future aging trend of the device on the basis of completing the life prediction, making the life prediction value more interpretable, and having great practical value for avoiding serious shutdown accidents caused by the operation failure of fan components. At the same time, using the historical health index data of the target device before the prediction starting point to retrain and fine-tune the basic model, helping the basic model learn the unique aging law of the target device to complete model update, thereby achieving further improvement of the prediction performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 FIG. is a schematic flow chart of a wind turbine bearing life prediction method based on a hybrid model in an embodiment;

[0043] Figure 2 FIG. is a flow block diagram of a wind turbine bearing life prediction method based on a hybrid model in an embodiment;

[0044] Figure 3 FIG. is a schematic framework diagram of a health index prediction model in an embodiment;

[0045] Figure 4 Schematic diagram of the life prediction results of the bearing for testing at different prediction starting points in another embodiment; (a) is the cyclic prediction result diagram of the health index with 30% of the full cycle as the prediction starting point, (b) is the cyclic prediction result diagram of the health index with 50% of the full cycle as the prediction starting point, (c) is the cyclic prediction result diagram of the health index with 70% of the full cycle as the prediction starting point, (d) is the cyclic prediction result diagram of the health index with 90% of the full cycle as the prediction starting point;

[0046] Figure 5 Block diagram of the real-time updated wind turbine bearing life prediction method in one embodiment;

[0047] Figure 6 Schematic diagram of the bearing life prediction effect in the method of real-time updating the model for the bearing for testing in one embodiment; (a) is the cyclic prediction effect diagram of the health index with 30% of the full cycle as the prediction starting point, (b) is the cyclic prediction effect diagram of the health index with 50% of the full cycle as the prediction starting point, (c) is the cyclic prediction effect diagram of the health index with 70% of the full cycle as the prediction starting point, (d) is the cyclic prediction effect diagram of the health index with 90% of the full cycle as the prediction starting point;

[0048] Figure 7 Comparison diagram of the life prediction effects between the method of real-time updating the model and the method of not updating the model in one embodiment;

[0049] Figure 8 Block diagram of the structure of a wind turbine bearing life prediction device based on a hybrid model in one embodiment;

[0050] Figure 9 Internal structure diagram of a computer device in one embodiment. Detailed implementation manners

[0051] In order to make the objectives, technical solutions and advantages of the present application clearer, the present application will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application, but not to limit the present application.

[0052] In one embodiment, as Figure 1 and Figure 2 shown, a wind turbine bearing life prediction method based on a hybrid model is provided, including the following steps:

[0053] Step 102, obtain the relevant operation data set of the wind turbine; perform data conversion on the relevant operation data set of the wind turbine to obtain a data set in the form of amplitude probability density.

[0054] The operation dataset related to the wind turbine is the vibration data of the wind turbine bearing throughout its life cycle. In this step, the full-cycle vibration data of 7 groups of bearings operating from the normal operation state to failure under the working conditions of a rotational speed of 1800 rpm and a load of 4000 N are used for method verification and testing. Each sampling sample has 2560 data points. Using the Gaussian kernel as the kernel function and 0.1 as the bandwidth, the kernel density estimation method is used to process the original vibration signals of the seven groups of bearings, and the vibration data is converted into a probability density form.

[0055] Step 104: Divide the amplitude probability density form dataset into a training set and a test set, and train the pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indexes.

[0056] Select 100 sampling points in the amplitude range [-50, 50] in an equal-width manner, and use the probability density at the corresponding amplitude as the sample feature after sampling for prediction work; use six groups of bearings out of the seven groups of bearings as the training set, and the data of one group of bearings as the test set.

[0057] Step 106: Set the health index threshold for judging bearing failure according to the development law of the extracted health indexes, and train the prediction model of the 2D-CNN-BiLSTM hybrid structure according to the health indexes of the extracted training set to obtain a health index prediction model.

[0058] Combining the characteristics of bearing aging and the monotonicity and trend of health indexes, select the quartic function as the source of the training set sample labels; mark the training set sample labels based on the quartic function to help the regression model construct health indexes with stronger trend and monotonicity; at the same time, relying on the interval characteristics of the quartic function values, the failure threshold can be set to 0, that is, the health index value is 0 when the bearing runs to the end point. After completing the sample label marking, the probability density values obtained after processing the original vibration signal form 100-dimensional input data, and the marked training set labels are used as the model target output to train the model based on random forest regression for health index value construction.

[0059] Sample the health indexes of the training set with a fixed time window into several one-dimensional health index sequences; use the data reconstruction method to convert the one-dimensional health index sequences into a two-dimensional form, and use the two-dimensional health index data as the input of the 2D-CNN-BiLSTM model, and use the health index value at the next moment of the taken time window as the model prediction target, thereby training the single-step prediction model of the health index based on the 2D-CNN-BiLSTM model. The model structure is as Figure 3 shown, which includes two parts: convolution processing and time series prediction.

[0060] The health index data of the training set is divided into several one-dimensional time series samples by combining a time window with a width of 200 and the sliding window method. Then, a small time window with a window width of 20 is used to further divide the one-dimensional time series, and the original one-dimensional health index sequence of 200×1 is reconstructed into a data form of 10×20.

[0061] After the reconstruction step, the one-dimensional sequence of the training set is reconstructed into two-dimensional data and input into the prediction model based on 2D-CNN-BiLSTM. The health index value at the next moment of the sequence is used as the prediction target to train the model, and single-step prediction of the health index is realized. The specific settings for model training are as follows: the batch size is 256, the loss function is MAE, the Adam algorithm is used to optimize the model parameters, training is carried out for 100 rounds, and the model with the smallest loss on the test set after more than 20 rounds of training is selected as the final single-step prediction model for the health index.

[0062] Step 108, use the model constructed by the health index to perform health index construction on the test set data to obtain the health index values of each sample before the prediction start point in the test set.

[0063] The test set data also undergoes kernel density estimation and sampling steps, and is processed from the original vibration data into a probability density form and input into the trained random forest regression model. After model calculation, the health index values corresponding to each sample at the corresponding moment before the prediction start point on the test set are output; through continuous calculation, a one-dimensional sequence of the health index on the test set is obtained.

[0064] Step 110, according to the health index values of each sample before the prediction start point in the test set and the health index prediction model to predict the health index at the next moment, complete the prediction of the health index values at multiple future moments through cyclic single-step prediction steps, and use the health index threshold to judge the predicted health index values at multiple future moments to determine the bearing failure moment, and calculate the bearing life prediction value according to the time difference between the current moment and the failure moment.

[0065] After determining the prediction start point, the one-dimensional health index sequence within a time window (window width of 200) before the start point is reconstructed into two-dimensional data and input into the prediction model. The single-step prediction value of the health index is calculated and output by the model. Judge whether the prediction value reaches the failure threshold. If it does not reach, the prediction value is incorporated into the known sequence of the health index, and the sequence is updated by the sliding window method once (that is, slide one step backward), and the data reconstruction and prediction steps are repeated until the prediction value reaches the failure threshold.

[0066] Perform cyclic single-step prediction on the health index value until the failure threshold is reached, and count the number of cycles experienced from the prediction start point to the end of the cycle as the life prediction value at the prediction start point. The prediction effect on the test set is as Figure 4As shown, (a), (b), (c), and (d) show the cyclic prediction effects of health indicators with the starting points of prediction at 30%, 50%, 70%, and 90% of the full cycle respectively. As the cyclic prediction work progresses, the final health indicator values all reach the failure threshold. It can be seen from the figure that as the bearing operation goes deeper, the aging trend reflected by the vibration signal gradually accumulates, bringing more aging information to the prediction model, and the life prediction effect also improves as the prediction starting point advances.

[0067] In the above-mentioned wind turbine bearing life prediction method based on a hybrid model, this application completes the bearing life prediction work by adopting a health indicator construction model based on random forest and a single-step prediction model of health indicators based on a 2D-CNN-BiLSTM hybrid structure. Among them, the health indicator construction model based on random forest can directly extract the health indicator of the device at the current moment from the probability density information of vibration data. The convolutional neural network part in the 2D-CNN-BiLSTM hybrid structure can extract key information from the original complex high-dimensional data and weaken the influence of noise. Introducing the convolutional part to process the original time series data in the single-step prediction model of health indicators can effectively suppress the influence of the noise contained in the original time series and extract time series information to help improve the prediction effect. The bidirectional long short-term memory network part in the 2D-CNN-BiLSTM hybrid structure completes the two-way transmission of data by stacking two layers of LSTM, comprehensively considers the information of the previous and subsequent moments in the process of processing time series data, and shows better time series prediction performance. The combination of the convolutional neural network and the bidirectional long short-term memory network can effectively dig deep into the development trend of health indicators, realize the accurate prediction of health indicators, and help complete the life prediction; use the trained random forest model to carry out the health indicator construction work on the test set data, and the model calculates and outputs the health indicator values of each sample before the prediction starting point in the test set; use the calculated health indicator values of the test set and the trained 2D-CNN-BiLSTM model to predict the health indicator at the next moment, and complete the prediction of health indicator values at multiple future moments by cyclically carrying out the single-step prediction steps; and use the set failure threshold to judge the predicted future health indicator values, determine the bearing failure moment, calculate the bearing life prediction value according to the time difference between the current moment and the failure moment, and realize the prediction task by designing the idea of cyclic single-step prediction of health indicators. On the basis of completing the life prediction, it can provide the judgment of the future aging trend of the device, making the life prediction value more interpretable and having great practical value for avoiding serious shutdown accidents caused by the operation failure of fan components. At the same time, use the historical health indicator data of the target device before the prediction starting point to retrain and fine-tune the basic model, help the basic model learn the unique aging law of the target device to complete model update, and thus realize the further improvement of prediction performance.

[0068] In one embodiment, after learning the change trend of the bearing historical health indicators on the test set, the health indicator prediction model is updated, and the bearing life of the wind turbine is predicted in real time according to the updated health indicator prediction model.

[0069] In one embodiment, a model is constructed using health indicators to extract the historical health indicator values on the test set; the health indicator prediction model trained with the training set is used as the basic model, and the historical health indicator values on the test set are used to fine-tune the basic model to complete the learning of the historical indicator value trend, and an updated health indicator prediction model is obtained; the updated health indicator prediction model is used to perform single-step prediction of the indicator values in a loop until the health indicator prediction value reaches the health indicator threshold, and the life prediction result updated in real time by the model is calculated.

[0070] In a specific embodiment, as Figure 5 shown, the single-step health indicator prediction model based on the 2D-CNN-BiLSTM model trained with the training set data is used as the basic model;

[0071] According to the historical health indicator data situation before the prediction starting point on the test set, the basic model is retrained and model fine-tuning is performed to obtain a single-step health indicator prediction model updated in real time; according to the single-step health indicator prediction result, it is judged whether the loop termination condition of the single-step prediction is reached; after the prediction is terminated, the duration from the prediction starting point to the end of the loop is calculated as the life prediction value; the test set health indicator prediction effect and life prediction effect in the case of model update and non-update are as Figure 6 shown, the loop health indicator prediction effect of the model updated in real time at different prediction starting points is better than that of the non-updated basic model; the test set life prediction errors in the case of model update and non-update are as Figure 7 shown, it can be seen that the model updated in real time, due to learning the historical aging data of the target device, has better prediction effects at the ten segmented points tested than the non-updated basic model, and as the running time continues, the bearing aging degree deepens and the aging information in the running data gradually accumulates, providing higher-quality data information for the prediction model, making the prediction error show an overall downward trend, but there are inevitably disturbances caused by factors such as noise at individual points. Therefore, the prediction errors of the model updated in real time and the basic model both show a fluctuation at 60%.

[0072] In one embodiment, data conversion is performed on the relevant operation data set of the wind turbine to obtain a data set in the form of amplitude probability density, including:

[0073] The probability density distribution of the relevant operation data set of the wind turbine is calculated by using the kernel density estimation method, sampling is performed within the amplitude range, and the probability density values at several amplitudes are selected as the samples after data conversion; the samples are data sets in the form of amplitude probability density.

[0074] In one embodiment, a pre-constructed random forest regression model is trained according to a training set to obtain a health index construction model, including:

[0075] Using a quartic function model to label the samples in the training set to obtain sample labels;

[0076] Taking a number of probability density values obtained in data transformation as input and the sample labels as the target output to train the random forest regression model, and using the model output value as the constructed health index to obtain the health index construction model.

[0077] In one embodiment, a prediction model with a 2D-CNN-BiLSTM hybrid structure is trained according to the health indices of the extracted training set to obtain a health index prediction model, including:

[0078] Sampling the extracted health indices with a fixed time window into a number of one-dimensional health index sequences; using a data reconstruction method to convert the one-dimensional health index sequences into a two-dimensional form to obtain two-dimensional time-series samples;

[0079] Using the two-dimensional time-series samples to train the prediction model with a 2D-CNN-BiLSTM hybrid structure, and completing the model training with the reconstructed two-dimensional data as the input and the health index value at the future moment as the target output to obtain the health index prediction model.

[0080] In a specific embodiment, the health indices of the training set are sampled with a fixed time window into a number of one-dimensional health index sequences by using the repeated segmentation and sliding window method, and a data reconstruction method is designed to convert the one-dimensional sequences into a two-dimensional data form; a single-step prediction model of the health index based on the 2D-CNN-BiLSTM model is trained by using the two-dimensional data of the health indices of the training set samples, with the two-dimensional health index data as the model input and the health index value at the next moment of the selected time window as the prediction target, and the training of the single-step prediction model of the health index is completed.

[0081] In one embodiment, according to the health index values of each sample before the prediction starting point in the test set and the health index prediction model to predict the health index at the next moment, the prediction of the health index values at multiple future moments is completed by repeatedly performing the single-step prediction step, and the predicted health index values at multiple future moments are judged by using the health index threshold to determine the bearing failure moment, and the bearing life prediction value is calculated according to the time difference between the current moment and the failure moment, including:

[0082] After determining the prediction starting point on the test set, extract the health index values at each moment within a time window before the starting moment, reconstruct the original one-dimensional health index sequence within the time window into two-dimensional data, and use the health index prediction model to receive the two-dimensional data and calculate the predicted value of the health index for the next moment. Determine whether the predicted value reaches the health index threshold. If the predicted value does not reach the health index threshold, include the predicted value in the original sequence. At the same time, use the sliding window method to update the sequence samples and repeat the data reconstruction and prediction steps until the predicted value of the health index reaches the health index threshold. When the predicted value of the health index reaches the health index threshold, perform the bearing life prediction work, and calculate the duration experienced from the prediction starting point to the prediction end point as the bearing life prediction value.

[0083] It should be understood that although Figure 1 the steps in the flowchart of Figure 1 are shown in sequence according to the indication of the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless there is a clear indication in this article, the execution of these steps has no strict order limit, and these steps can be executed in other orders. Moreover,

[0084] In one embodiment, as Figure 8 shown, a wind turbine bearing life prediction device based on a hybrid model is provided, including: a data conversion module 802, a health index construction model training module 804, a health index prediction model training module 806, a health index construction module 808, and a bearing life prediction module 810, where:

[0085] The data conversion module 802 is used to obtain the relevant operation data set of the wind turbine; perform data conversion on the relevant operation data set of the wind turbine to obtain a data set in the form of amplitude probability density;

[0086] The health index construction model training module 804 is used to divide the data set in the form of amplitude probability density into a training set and a test set, train a pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indexes;

[0087] The health index prediction model training module 806 is used to set the health index threshold for judging bearing failure according to the development law of the extracted health indexes, and train the prediction model of the 2D-CNN-BiLSTM hybrid structure according to the health indexes of the extracted training set to obtain a health index prediction model;

[0088] A health index construction module 808, configured to use a health index construction model to perform health index construction on the test set data, and obtain the health index values of each sample before the prediction start point in the test set;

[0089] A bearing life prediction module 810, configured to predict the health index at the next moment according to the health index values of each sample before the prediction start point in the test set and the health index prediction model, complete the prediction of the health index values at multiple future moments through a loop of single-step prediction steps, and use the health index threshold to judge the predicted health index values at multiple future moments, determine the bearing failure moment, and calculate the bearing life prediction value according to the time difference between the current moment and the failure moment.

[0090] For the specific limitations of the wind turbine bearing life prediction device based on the hybrid model, reference can be made to the limitations of the wind turbine bearing life prediction method based on the hybrid model in the above text, which will not be elaborated here. Each module in the above wind turbine bearing life prediction device based on the hybrid model can be implemented in whole or in part by software, hardware, and their combinations. The above-mentioned modules can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory of the computer device in the form of software, so as to facilitate the processor to call and execute the operations corresponding to the above modules.

[0091] In one embodiment, a computer device is provided. The computer device can be a terminal, and its internal structure diagram can be as Figure 9 shown. The computer device includes a processor, a memory, a network interface, a display screen, and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, it implements a wind turbine bearing life prediction method based on a hybrid model. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen. The input device of the computer device can be a touch layer covering the display screen, or a button, a trackball, or a touchpad provided on the housing of the computer device, or an external keyboard, touchpad, or mouse, etc.

[0092] Those skilled in the art can understand, Figure 9The structure shown is only a block diagram of some structures related to the solution of this application, and does not constitute a limitation on the computer device to which the solution of this application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine some components, or have different component arrangements.

[0093] Those of ordinary skill in the art can understand that all or part of the processes in the methods of the above embodiments can be completed by instructing relevant hardware through a computer program. The computer program can be stored in a non-volatile computer-readable storage medium. When the computer program is executed, it can include the processes of the embodiments of the above methods. Among them, any reference to a memory, storage, database, or other medium used in the various embodiments provided in this application can include non-volatile and / or volatile memories. Non-volatile memories can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memories can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and Rambus dynamic RAM (RDRAM), etc.

[0094] The technical features of the above embodiments can be combined arbitrarily. For the sake of brevity of description, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, it should be considered as the scope described in this specification.

[0095] The above-described embodiments merely represent several implementation manners of this application. Their descriptions are relatively specific and detailed, but they should not be construed as limiting the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the concept of this application, several modifications and improvements can still be made, and these all belong to the protection scope of this application. Therefore, the protection scope of the patent of this application should be subject to the appended claims.

Claims

1. A method for predicting the bearing life of a wind turbine based on a hybrid model, characterized in that, The method includes: Obtaining an operation dataset related to a wind turbine; performing data conversion on the operation dataset related to the wind turbine to obtain a dataset in the form of amplitude probability density; Dividing the dataset in the form of amplitude probability density into a training set and a test set, training a pre-constructed random forest regression model according to the training set to obtain a health index construction model; the health index construction model is used to extract health indices; Setting a health index threshold for judging bearing failure according to the development law of the extracted health indices, and training a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indices of the training set to obtain a health index prediction model; Using the health index construction model to construct health indices for the test set data to obtain the health index values of each sample before the prediction starting point in the test set; Predicting the health indices at the next moment according to the health index values of each sample before the prediction starting point in the test set and the health index prediction model, completing the prediction of the health index values at multiple future moments through cyclic single-step prediction steps, and judging the predicted health index values at multiple future moments by using the health index threshold to determine the bearing failure moment, and calculating the bearing life prediction value according to the time difference between the current moment and the failure moment.

2. The method according to claim 1, wherein The method further includes: Learning the change trend of the historical health indices of the bearing on the test set and updating the health index prediction model, and predicting the bearing life of the wind turbine in real time according to the updated health index prediction model.

3. The method according to claim 2, wherein The method further includes: Using the health index construction model to extract the historical health index values on the test set; using the health index prediction model trained with the training set as the basic model, and fine-tuning the basic model with the historical health index values on the test set to complete the learning of the historical index value trend to obtain an updated health index prediction model; using the updated health index prediction model to perform cyclic single-step prediction of the index values until the health index prediction value reaches the health index threshold to calculate the life prediction result of the model updated in real time.

4. The method according to any one of claims 1 to 3, characterized in that Performing data conversion on the operation dataset related to the wind turbine to obtain a dataset in the form of amplitude probability density, including: Calculating the probability density distribution of the operation dataset related to the wind turbine by using the kernel density estimation method, sampling within the amplitude range, and selecting the probability density values at several amplitudes as the samples after data conversion; the samples are the dataset in the form of amplitude probability density.

5. The method according to claim 1, wherein Training a pre-constructed random forest regression model according to the training set to obtain a health index construction model, including: Using a quartic function model to label the samples in the training set to obtain sample labels; Training a random forest regression model with several probability density values obtained in the data conversion as the input and the sample labels as the target output, and using the model output value as the constructed health index to obtain a health index construction model.

6. The method according to claim 1, characterized in that, Training a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indices of the training set to obtain a health index prediction model, including: Sample the extracted health indicators at fixed time windows into a number of one-dimensional health indicator sequences; use the data reconstruction method to transform the one-dimensional health indicator sequences into a two-dimensional form to obtain two-dimensional time-series samples; Use the two-dimensional time-series samples to train a prediction model with a 2D-CNN-BiLSTM hybrid structure. Complete the model training with the reconstructed two-dimensional data as the input and the health indicator values at future moments as the target output to obtain a health indicator prediction model.

7. The method according to claim 6, wherein According to the health indicator values of each sample before the prediction start point in the test set and the health indicator prediction model, predict the health indicators at the next moment. Complete the prediction of health indicator values at multiple future moments through a loop of single-step prediction steps, and use the health indicator threshold to judge the predicted health indicator values at multiple future moments to determine the bearing failure moment. Calculate the bearing life prediction value based on the time difference between the current moment and the failure moment, including: After determining the prediction start point on the test set, extract the health indicator values at each moment within the previous time window of the start point moment. Reconstruct the original one-dimensional health indicator sequence within the time window into two-dimensional data, and use the health indicator prediction model to receive the two-dimensional data and calculate and output the health indicator prediction value at the next moment. Judge whether the prediction value reaches the health indicator threshold. If it does not reach the health indicator threshold, incorporate the prediction value into the original sequence. At the same time, use the sliding window method to update the sequence samples and repeat the data reconstruction and prediction steps until the health indicator prediction value reaches the health indicator threshold. When the health indicator prediction value reaches the health indicator threshold, perform the life prediction work, and calculate the duration experienced from the prediction start point to the prediction end point as the bearing life prediction value.

8. A wind turbine bearing life prediction device based on a hybrid model, characterized in that, The device includes: A data conversion module, configured to obtain an operation dataset related to a wind turbine; perform data conversion on the operation dataset related to the wind turbine to obtain a dataset in the form of amplitude probability density; A health indicator construction model training module, configured to divide the dataset in the form of amplitude probability density into a training set and a test set, and train a pre-constructed random forest regression model according to the training set to obtain a health indicator construction model; the health indicator construction model is used to extract health indicators; A health indicator prediction model training module, configured to set a health indicator threshold for judging bearing failure according to the development law of the extracted health indicators, and train a prediction model with a 2D-CNN-BiLSTM hybrid structure according to the health indicators of the extracted training set to obtain a health indicator prediction model; A health indicator construction module, configured to use the health indicator construction model to construct health indicators for the test set data to obtain the health indicator values of each sample before the prediction start point in the test set; A bearing life prediction module, configured to predict the health indicators at the next moment according to the health indicator values of each sample before the prediction start point in the test set and the health indicator prediction model, complete the prediction of health indicator values at multiple future moments through a loop of single-step prediction steps, and use the health indicator threshold to judge the predicted health indicator values at multiple future moments to determine the bearing failure moment, and calculate the bearing life prediction value based on the time difference between the current moment and the failure moment.

Images