Wind power bearing life prediction method, system and device and storage medium
By introducing a convolutional length short-time memory network model with self-attention module and overlapping sampling strategy in the life prediction of wind power bearings, the problems of low training efficiency and low prediction accuracy in traditional LSTMs when predicting wind power bearing life are solved, and more efficient and accurate prediction results are achieved.
Patent Information
- Application Number
- CN202411488060.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-23
- Publication Date
- 2025-03-14
AI Technical Summary
When traditional LSTM predicts the remaining service life of wind power bearings, there are common problems of low training efficiency and low prediction accuracy.
A convolutional length and short-time memory network model with self-attention module is adopted, combined with overlapping sampling strategy and autoencoder algorithm, the bearing life cycle vibration data is converted into a feature matrix, and prediction is made through the attention convolutional length and short-time memory network.
It improves the accuracy and training efficiency of prediction results, reduces data complexity, enhances the utilization of important information, and improves prediction performance.
Smart Images
Figure CN119623235A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of simulation analysis technology, and in particular to a method, system, device and storage medium for predicting the life of a wind power bearing. Background Art
[0002] Wind turbine bearings are widely used in wind turbine equipment because of their strong load-bearing capacity and good starting performance. Under complex working conditions and harsh environments, bearings are prone to failure, which seriously affects the performance of the entire machine. Therefore, in order to prevent mechanical system safety accidents caused by bearing failure and improve the operating efficiency of wind turbines, predicting the remaining useful life (RUL) of bearings has considerable engineering application significance.
[0003] At present, physics-based and data-driven methods have been identified as the main RUL prediction methods. Physics-based methods mainly describe the degradation process of mechanical components by establishing mathematical models based on failure mechanisms. However, for some complex mechanical systems, the physical principles of damage are difficult to understand, which makes it impossible to establish accurate mathematical models, limiting the application of these methods. In recent years, due to the rise of artificial intelligence theory, data-driven methods have gradually become the mainstream RUL prediction solution. They rely on methods such as statistics and machine learning, rather than physical knowledge of mechanical components, to adaptively model the collected degradation data to describe the evolution of their degradation process. At present, some researchers are studying RUL prediction methods based on convolutional neural networks.
[0004] Due to its recursive structure, recurrent neural network (RNN) can learn the correlation of sequence data in the time dimension, which is very suitable for time series data. However, RNN has problems with gradient explosion and gradient vanishing during training, and long-range dependencies are difficult to capture. Long short-term memory (LSTM) and gated recurrent unit (GRU) solve this problem by introducing multiple gating mechanisms. LSTM and GRU are widely used in RUL prediction of mechanical equipment due to their advantages in long-term memory. In summary, the classic RNN variant is considered to be a more appropriate prediction method and has been widely used in time series prediction.
[0005] Although classic RNN variants such as LSTM have been widely used in many fields, they rely on fully connected layers based on dense matrix multiplication operations and require a large amount of data to train the network. In addition, as the bearing degrades throughout its life cycle, the closer the vibration signal is to the fault threshold, the more degradation information it contains. LSTM does not highlight important degradation information in the prediction process and assumes that the input data makes the same contribution to the output. Therefore, in actual industrial applications, traditional LSTM generally has problems with low training efficiency and low prediction accuracy when predicting the RUL of bearings. Summary of the invention
[0006] In view of the deficiencies in the prior art, the present invention provides a method for predicting the life of a wind turbine bearing, which solves the problems of low training efficiency and low prediction accuracy in the conventional LSTM in the prior art when predicting the RUL of bearings.
[0007] According to an embodiment of the present invention, a method for predicting the life of a wind turbine bearing includes:
[0008] S1: Acquire the bearing life cycle vibration data, and convert the bearing life cycle vibration data into a feature matrix;
[0009] S2: Input the last P rows in the feature matrix into the attention convolution long short-term memory network model to obtain the prediction matrix for the next moment, concatenate the prediction matrix with the feature matrix and normalize it to obtain the prediction value, and then compare the prediction value with the fault threshold;
[0010] S3: Repeat step S2 until the predicted value is greater than the fault threshold, and then estimate the bearing life according to the number of repetitions.
[0011] Preferably, before step S2, the first sample is selected from the feature matrix as a training matrix, and the remaining samples are normalized, and an overlapping sampling strategy is used to expand the number of samples to obtain a new feature matrix.
[0012] Preferably, after obtaining the new feature matrix, the first p rows and the last row of the feature matrix are used as training data and labels to train the attention convolutional long short-term memory network model.
[0013] Preferably, the method for obtaining bearing life cycle vibration data includes:
[0014] Sampling multiple degradation features of the bearing during degradation at fixed sampling intervals, and fitting the corresponding variation curves;
[0015] According to the monotonicity and trend of the change curve, multiple bearing life cycle vibration data are screened out.
[0016] Preferably, the AE algorithm is used to process the bearing life cycle vibration data to obtain a feature matrix.
[0017] Preferably, the AE algorithm is used to process the bearing life cycle vibration data to obtain a feature matrix.
[0018] Preferably, the bearing life estimation formula is as follows
[0019] RUL=τ×Δt
[0020] Among them, τ is the number of iterations, and Δt is the sampling interval of vibration data during the bearing life cycle.
[0021] On the other hand, according to an embodiment of the present invention, a wind turbine bearing life prediction system is also provided. The system adopts the above-mentioned wind turbine bearing life prediction method, including:
[0022] The sampling module is used to sample multiple degradation characteristics during the bearing degradation process, fit the corresponding change curve, and obtain the bearing life cycle vibration data after screening.
[0023] A prediction module, which is used to convert the bearing life cycle vibration data into a feature matrix and obtain a prediction matrix for the bearing at the next moment according to the feature matrix, and concatenate the prediction matrix and the feature matrix;
[0024] An estimation module is used to compare the predicted value with the fault threshold and estimate the bearing life.
[0025] On the other hand, according to an embodiment of the present invention, there is also provided a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the above-mentioned method for predicting the life of a wind turbine bearing.
[0026] On the other hand, according to an embodiment of the present invention, there is further provided a computer storage medium storing a computer program, wherein when the computer program is executed by a processor, the processor executes the above-mentioned method for predicting the life of a wind turbine bearing.
[0027] Compared with the prior art, the present invention has the following beneficial effects:
[0028] 1) A convolutional long short-term memory network model with a self-attention module is introduced to focus on the degraded data, which is beneficial to the prediction results and adaptively utilizes the corresponding important information to improve the prediction performance.
[0029] 2) The overlapping sampling strategy is used to expand the number of samples, and after obtaining the predicted data, it is reconnected with the original data to greatly increase the data volume. At the same time, the prediction process no longer relies on the fully connected layer, the data complexity of the training and prediction process is reduced, the training efficiency is greatly improved, and the prediction accuracy can also be guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS
[0030] Figure 1 This is a flow chart of wind turbine bearing life prediction according to an embodiment of the present invention.
[0031] Figure 2 The training and prediction curves of the attention convolutional long short-term memory network of the embodiment of the present invention.
[0032] Figure 3This is a comparison chart of the prediction capabilities of different methods for a data set according to an embodiment of the present invention.
[0033] Figure 4 This is a comparison chart of the convergence speed of different methods in the data set of the embodiment of the present invention. DETAILED DESCRIPTION
[0034] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.
[0035] like Figure 1 As shown, the embodiment of the present invention proposes a method for predicting the life of a wind turbine bearing, comprising:
[0036] S1: Acquire the bearing life cycle vibration data, and convert the bearing life cycle vibration data into a feature matrix;
[0037] In order to obtain more comprehensive bearing degradation information, it is necessary to comprehensively adopt multi-domain features. In the bearing degradation process, there are significant differences in the features that reflect the bearing degradation trend. Some features are insensitive to the degradation state of the bearing, so these features with poor performance need to be eliminated. Therefore, we consider monotonicity (Mon) and trend (Tre) as screening features, and a single metric cannot fully evaluate the applicability of degradation features. Therefore, this invention uses a linear combination of Mon and Tre as a comprehensive evaluation:
[0038]
[0039] Among them, Cem is a comprehensive evaluation index.
[0040] In this invention, we use several features listed in Table 1:
[0041] Table 1
[0042]
[0043] After that, let n and Δt represent the sampling times and sampling interval of the same feature respectively, to sample the above filtered features and obtain the matrix X∈R n×m , where m is the m features with higher comprehensive evaluation selected, and the selected m features are the bearing life cycle vibration data.
[0044] After removing the features with low evaluation values, different time domain and frequency domain features still exist in the high-dimensional space, and there are redundancy and conflict problems. Therefore, the dimensionality reduction algorithm is used to fuse multiple features into VHI. Dimensionality reduction algorithms such as linear discriminant analysis cannot reflect the characteristics of nonlinear degradation of wind turbine bearings. The manifold learning algorithm based on nonlinear dimensionality reduction such as local linear embedding requires that the data is closed and locally linear in the high-dimensional space, and is very sensitive to parameter selection. The AE algorithm is an unsupervised neural network that can automatically learn the distribution information of the original data and realize nonlinear dimensionality reduction from the compressed space representation without losing important information.
[0045] Therefore, AE is used to process the matrix X to generate VHI. When the loss function of the AE algorithm is minimized, its output is close to the input. At this time, the feature matrix V∈R n×1 .
[0046] Many studies assume that the root mean square of acceleration is fixed, and it has an obvious step signal which is considered as the end of the life cycle, while the actual trend at this time shows a very rapid growth. This is because the confidence in predicting the slow-moving trend is high, so it is not suitable for sudden changes. Therefore, in the present invention, the root mean square value of acceleration has an obvious step signal as the VHI failure time, and it is used as the fault threshold.
[0047] S2: Input the last P rows in the feature matrix into the attention convolution long short-term memory network model to obtain the prediction matrix for the next moment, concatenate the prediction matrix with the feature matrix and normalize it to obtain the prediction value, and then compare the prediction value with the fault threshold;
[0048] For the same bearing, it is usually necessary to use it for a long time and collect enough sample data to accurately predict its service life. However, long-term use itself will cause wear on the bearing. Even if its service life is predicted, the bearing cannot be sold. Therefore, it is only possible to use it for a short time and collect a small amount of sample data to predict the bearing. Due to the small amount of sample data, it is easy to make the prediction result inaccurate. Therefore, it is necessary to expand the number of samples and use the overlapping sampling strategy to expand the number of samples. After obtaining the predicted data, it is re-joined with the original data to greatly increase the data volume. At the same time, the prediction process no longer relies on the fully connected layer, the data complexity of the training and prediction process is reduced, the training efficiency is greatly improved, and the prediction accuracy can also be guaranteed.
[0049] The first sample n1 is selected from the feature matrix V as the training matrix. After normalization, the normalized training matrix U = (u1, u2, ..., u n1 ), U∈R n×1, and then the overlapping sampling strategy is used to expand the number of training samples and obtain the matrix W:
[0050]
[0051] Where p is the dimension of the input segment data, w i =[u i u i+1 …u n1-p+i-1 ].
[0052] In addition, before using the attention convolutional long short-term memory network model, the model needs to be trained. The first p rows and the last row in W are taken as training data and labels to train the attention convolutional long short-term memory network. The training process is described as follows:
[0053] minLoss(w p+1 ,f(w1,w2,…,w p ))=MSE(w p+1 ,f(w1,w2,…,w p ))
[0054] Where MSE and f represent the mean square error function and the attention convolutional long short-term memory network, respectively.
[0055] After training is completed, Figure 2 As shown, the number of iterations τ is introduced, the initial value is 0, and then the last p rows in W are input into the attention convolution long short-term memory network to obtain the prediction data for the next moment:
[0056]
[0057] In the formula, is the prediction data for the next moment. At the same time, the number of iterations is τ+1. At this time, the prediction matrix and the feature matrix are concatenated. The overall prediction data at this time is as follows:
[0058]
[0059] Update the matrix W to
[0060] By introducing a convolutional long short-term memory network model with a self-attention module, focusing on the degraded data is beneficial to the prediction results, and the corresponding important information is adaptively utilized to improve the prediction performance.
[0061] S3: Repeat step S2 until the predicted value is greater than the fault threshold, and then estimate the bearing life according to the number of repetitions.
[0062] Calculate the normalized predicted values: If the predicted value is less than the fault threshold, S2 is repeated. If the predicted value is greater than the fault threshold, the bearing is considered to have failed. At this time, the remaining life of the bearing is:
[0063] RUL=τ×Δt.
[0065] In order to verify the effectiveness of the method disclosed in the present invention, wind turbine gearbox bearing data is used for experimental verification:
[0066] 1) Data collection and description: Wind energy is a renewable energy source that has developed quite rapidly due to its clean and abundant resource reserves. However, due to the harsh operating environment of wind turbines, the failure rate of bearings is often very high, resulting in the interruption of the entire unit and even serious accidents. Accurate and efficient RUL prediction of wind turbine gearbox bearings helps to understand the future health status of bearings and reduce the operation and maintenance costs of wind farms.
[0067] The attention convolutional long short-term memory network is applied to the RUL evaluation of wind turbine gearbox bearings. The test bearing used in the wind turbine gearbox is TIMKEN LM770949, with a theoretical speed of 85.71r / min. The accelerometer (sampling frequency is 25.6kHz) installed on the low-speed shaft collects the bearing vibration signal for one year, 5.12s / day, and 4 times at equal intervals every day. Therefore, the wind farm collects 131072 vibration amplitude sampling points every day. The model was verified using a set of wind turbine gearbox bearing data. In addition to the wind farm shutdown and maintenance, the dataset contains 760 and 876 samples.
[0068] 2) Results and analysis: Similarly, AE uses several features with the largest comprehensive evaluation value to fuse into VHI. The AE hyperparameters are the same as those in Section V-A3. The vibration data collected by the wind turbine bearing is greatly affected by external interference. Smoothing the fused feature points can effectively weaken the interference of external factors such as impact. Without considering the static phase data, the RMS value of the acceleration of the bearing shows an obvious step signal at 4140h, at which time the bearing fails. The VHI amplitude corresponding to this time in the data set is 1.441 as the corresponding threshold. Therefore, the data set uses 390 (300-690) samples for research. The input segment data size, the channel size of the cyclic state, and the convolution kernel size of the data set are set to 30, 10, and 1×3, respectively. Other hyperparameters are the same as those in Section V-A3. In order to avoid contingency, five comparative experiments were conducted. Due to the complex and changeable working environment of the wind farm, the training data samples are smaller than the samples of the test bench, so only 30 feature points in the future can be predicted.
[0069] from Figure 2 It can be seen that this method can still predict the trend of future service life. Figure 3As shown in the figure, the RMSE, MAE, and MAPE of the attention convolution long short-term memory network in the four quantitative evaluation indicators are 0.068, 0.054, and 4.113, which are significantly lower than the other five methods when predicting the next 30 feature points. When the actual RUL is 180h, the RUL predicted by the dataset is 159h, and the prediction error is 11.67%; therefore, the attention convolution long short-term memory network in the dataset has the highest score of 0.667. At the same time, the convergence speed of the attention convolution long short-term memory network is significantly faster than other models, such as Figure 4 Therefore, in practical engineering applications, the attention convolutional long short-term memory network is suitable for RUL prediction of wind turbine gearbox bearings.
[0070] On the other hand, an embodiment of the present invention further provides a wind turbine bearing life prediction system, which adopts the above-mentioned wind turbine bearing life prediction method, including:
[0071] The sampling module is used to sample multiple degradation characteristics during the bearing degradation process, fit the corresponding change curve, and obtain the bearing life cycle vibration data after screening.
[0072] A prediction module, which is used to convert the bearing life cycle vibration data into a feature matrix and obtain a prediction matrix for the bearing at the next moment according to the feature matrix, and concatenate the prediction matrix and the feature matrix;
[0073] An estimation module is used to compare the predicted value with the fault threshold and estimate the bearing life.
[0074] On the other hand, an embodiment of the present invention further provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the above-mentioned method for predicting the life of a wind turbine bearing.
[0075] On the other hand, an embodiment of the present invention further provides a computer storage medium storing a computer program. When the computer program is executed by a processor, the processor executes the above-mentioned method for predicting the life of a wind turbine bearing.
[0076] Finally, it should be noted that the above embodiments are only used to illustrate the technical solution of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solution of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.
Claims
1. A method for predicting the life of a wind turbine bearing, characterized in that: include: S1: Acquire the bearing life cycle vibration data, and convert the bearing life cycle vibration data into a feature matrix; S2: Input the last P rows in the feature matrix into the attention convolution long short-term memory network model to obtain the prediction matrix for the next moment, concatenate the prediction matrix with the feature matrix and normalize it to obtain the prediction value, and then compare the prediction value with the fault threshold; S3: Repeat step S2 until the predicted value is greater than the fault threshold, and then estimate the bearing life according to the number of repetitions.
2. A wind turbine bearing life prediction method according to claim 1, characterized in that: Before step S2, the first sample is selected from the feature matrix as the training matrix, and the remaining samples are normalized. The overlapping sampling strategy is used to expand the number of samples to obtain a new feature matrix.
3. A wind turbine bearing life prediction method as claimed in claim 2, characterized in that: After obtaining the new feature matrix, the first p rows and the last row of the feature matrix are used as training data and labels to train the attention convolutional long short-term memory network model.
4. A wind turbine bearing life prediction method according to claim 1, characterized in that: Methods for obtaining bearing life cycle vibration data include: Sampling multiple degradation features of the bearing during degradation at fixed sampling intervals, and fitting the corresponding variation curves; According to the monotonicity and trend of the change curve, multiple bearing life cycle vibration data are screened out.
5. A wind turbine bearing life prediction method according to claim 1, characterized in that: The AE algorithm is used to process the bearing life cycle vibration data to obtain the characteristic matrix.
6. A wind turbine bearing life prediction method according to claim 1, characterized in that: The acceleration variation curve during the bearing degradation process is extracted, and the root mean square of the acceleration is calculated, and the root mean square is used as the fault threshold.
7. A wind turbine bearing life prediction method according to claim 1, characterized in that: The bearing life estimation formula is as follows RUL=τ×Δt Among them, τ is the number of iterations, and Δt is the sampling interval of vibration data during the bearing life cycle.
8. A wind turbine bearing life prediction system, characterized by: The system adopts a wind turbine bearing life prediction method as described in any one of claims 1 to 7, comprising: The sampling module is used to sample multiple degradation characteristics during the bearing degradation process, fit the corresponding change curve, and obtain the bearing life cycle vibration data after screening. A prediction module, which is used to convert the bearing life cycle vibration data into a feature matrix and obtain a prediction matrix for the bearing at the next moment according to the feature matrix, and concatenate the prediction matrix and the feature matrix; An estimation module is used to compare the predicted value with the fault threshold and estimate the bearing life.
9. A computer device, characterized in that: It comprises a memory and a processor, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes a wind power bearing life prediction method according to any one of claims 1 to 7.
10. A computer storage medium, characterized in that: A computer program is stored, and when the computer program is executed by a processor, the processor executes a wind power bearing life prediction method according to any one of claims 1 to 7.