Online prediction method for residual service life of wind turbine generator bearing under limited sample

Abstract

The invention discloses an online prediction method for the residual service life of a wind turbine generator bearing under a limited sample, and belongs to the technical field of bearing life prediction, and the method comprises the following steps: S1, collecting the vibration data of a wind turbine generator bearing, segmenting the data according to a preset sampling interval, and obtaining a segmented vibration signal sequence; s2, extracting the health index HI in the step S1, designing a linear reliability index to correct the bearing degradation trend, forming an index system reflecting the bearing degradation state, then obtaining a mapping health index MHI through a linear interpolation method, and carrying out normalization processing on the mapped health index; s3, constructing an Informer network model integrated with an adaptive meta-learning (AML) strategy; s4, performing fine adjustment on the pre-training model in the step S3 by using historical data of a target domain; s5, target domain real-time data are input into the prediction model subjected to fine adjustment in the step S4, online real-time prediction of the remaining service life of the wind power bearing is achieved according to a set fault threshold value, the problem that samples are limited in the online RUL prediction process of the wind power generator bearing can be solved, and prediction efficiency and precision can be effectively improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing life prediction, and in particular relates to an online prediction method for the remaining service life of a wind turbine bearing under a finite sample. Background Art

[0002] Wind power, as an important renewable energy source, has rapidly grown globally, effectively alleviating the imbalance between electricity supply and demand. Bearings in wind turbines are critical components in the transmission system, primarily responsible for ensuring constrained relative rotation. Because wind turbines typically operate in harsh environments, bearing failures are prone to occur, leading to extended unplanned downtime and increased maintenance costs. Therefore, effectively predicting the remaining useful life (RUL) of wind turbine bearings can help plan maintenance decisions in advance, enhancing wind turbine reliability and safety while reducing maintenance costs.

[0003] Deep learning is widely used in the field of RUL prediction due to its powerful feature extraction and big data processing capabilities. Traditional RNN and its variants (such as LSTM, GRU) perform well in time series prediction, but their cyclic structure leads to low computational efficiency and makes it difficult to effectively capture the global dependencies of long sequences. Transformer breaks through the limitations of recursive computing through the encoder-decoder architecture and pure attention mechanism, supports parallel processing and explicitly models long-term dependencies, but it has problems such as large cumulative errors and low prediction efficiency. To optimize the above defects, the Informer network came into being. In the Informer network, the ProbSparse self-attention mechanism reduces the time complexity by sparse query screening, the generative decoder directly outputs multi-step prediction results to avoid error accumulation, and the distillation module compresses the sequence dimension layer by layer to improve computational efficiency. Since Informer has significant advantages in long-term sequence prediction, it can be used for RUL prediction of wind turbine bearings.

[0004] As an improved version of the Transformer, the Informer reduces time complexity and memory usage through the ProbSparse self-attention mechanism and avoids cumulative errors in multi-step predictions through a generative decoder, significantly improving the speed of long sequence prediction. However, due to the Informer's deep network structure, long training time, and insufficient degradation data, the efficiency and accuracy of RUL prediction need to be further improved.

[0005] Most existing data-driven methods achieve RUL prediction by learning the mapping relationship between degradation features and RUL values through offline training. These methods require specific RUL corresponding features, but labeled lifecycle datasets are difficult to obtain, limiting their practical application.

[0006] Therefore, there is a need for an online prediction method for the remaining service life of wind turbine bearings under limited samples that can solve the problem of limited samples in the online RUL prediction process of wind turbine bearings and effectively improve the prediction efficiency. Summary of the Invention

[0007] The purpose of the present invention is to provide an online prediction method for the remaining useful life of a wind turbine bearing under a limited sample, which can solve the problem of limited samples in the online RUL prediction process of wind turbine bearings and effectively improve the prediction efficiency and accuracy.

[0008] To achieve the above object, the technical solution adopted by the present invention is:

[0009] An online prediction method for the remaining service life of a wind turbine bearing under finite samples includes the following steps:

[0010] Step S1: collecting wind turbine bearing vibration data, segmenting the data according to a preset sampling interval, and obtaining a segmented vibration signal sequence;

[0011] Step S2: Extract the health index HI from step S1 and design a linear reliability index to correct the bearing degradation trend, forming an index system that reflects the bearing degradation status. Then, obtain the mapped health index MHI through linear interpolation and normalize the mapped health index.

[0012] Step S3: Construct an Informer network model that integrates an adaptive meta-learning AML strategy. The Informer network constitutes the basic model framework for remaining useful life prediction, which includes a position encoding module, a ProbSparse self-attention mechanism, an encoder distillation module, and a generative decoder. An adaptive meta-learning strategy is introduced, and offline training is performed using source domain data to obtain model initialization parameters. An adaptive learning rate update mechanism is used to improve the model convergence speed.

[0013] Step S4: Fine-tune the pre-trained model in step S3 using historical data from the target domain to achieve knowledge transfer and adapt the model to the characteristics of the target device;

[0014] Step S5: The target domain real-time data is input into the prediction model fine-tuned in step S4, and the online real-time prediction of the remaining service life of the wind turbine bearing is realized according to the set fault threshold.

[0015] A further improvement to the technical solution of the present invention is that in step S2, the health index HI of step S1 is extracted. For the test bench data, the maximum amplitude is used as its HI. The HI value gradually increases with the deterioration of the bearing and rises sharply at the end of its life. The rank-preserving regression method is used to determine the fault start time, and a linear reliability index is designed to correct the bearing degradation trend to form an indicator system reflecting the bearing degradation state. Then, the mapped health index MHI is obtained through the linear interpolation method, and the mapped health index is normalized.

[0016] A further improvement of the technical solution of the present invention is that the calculation formula of the linear reliability index is expressed as:

[0017]

[0018] Where, Represents the health indicator after mapping, t s Represents the time series corresponding to the mapped health indicator, k s and b s The function parameter representing the linear reliability index, k s and b s Respectively expressed as:

[0019]

[0020] b s =min(X)

[0021] Where, X={x1,x2,…,x m} represents the original health index, t o represents the time series of the original health indicator, max(·) represents the maximum value of the sequence, min(·) represents the minimum value of the sequence, and n is an integer;

[0022] The calculation formula of the linear interpolation method is expressed as:

[0023]

[0024] Where, X={x1,x2,…,x m} represents the original health index, Y={y1,y2,…,y m} represents the mapping health indicator, y i Located at x j and x j+1 between;

[0025] The normalization formula is expressed as:

[0026]

[0027] Where, X max and Xmin Represent the maximum and minimum values of the mapping health indicator respectively.

[0028] A further improvement of the technical solution of the present invention is that in step S3, the calculation formula of the adaptive learning rate update mechanism is expressed as:

[0029]

[0030] Where p1=0.9 and p2=0.999 represent the attenuation rate, ε=10 -8 is a constant used to avoid the denominator being zero, α t =10 -4 represents the learning rate at time t;

[0031] The calculation formula of the position encoding module of the Informer network is expressed as:

[0032]

[0033] Where, is the input of Informer, t represents the position number of the input sequence, i∈[0,d model -1], d model Represents the dimension of the Informer model, Represents the position-encoded input sequence.

[0034] A further improvement of the technical solution of the present invention is that the calculation formula of the ProbSparse self-attention mechanism is expressed as:

[0035]

[0036] Where, K and V represent the query matrix, key matrix and value matrix respectively, d model Indicates the dimension of the input, l K and Represent K and length.

[0037] A further improvement of the technical solution of the present invention is that the calculation formula of the encoder distillation module is expressed as:

[0038]

[0039] Where, represents the input sequence of the encoder layer i, [·] Att Represents a multi-head attention module, Convld(·) represents a one-dimensional convolution filter, RELU(·) represents an activation function, and MaxPool(·) represents a maximum pooling layer with a step size of 2, which can convert the input sequence Halved.

[0040] A further improvement of the technical solution of the present invention is that the calculation formula for updating the model parameters of the adaptive meta-learning strategy is expressed as:

[0041]

[0042]

[0043] In the formula, η represents the learning rate of task training, λ represents the learning rate of task verification, represents the loss calculated in task i over the support set, represents the loss calculated in task i through the query set, k represents the number of tasks, and f θ represents the model function with parameter θ.

[0044] A further improvement of the technical solution of the present invention is that the Informer network uses the mean square error (MSE) as the loss function for optimization, and the calculation formula is expressed as:

[0045]

[0046] Where b represents the batch size, x i represents the actual mapping health indicator value at time i, represents the predicted mapping health indicator value at time i.

[0047] A further improvement of the technical solution of the present invention is that: in step S4, the source domain data D s Support Set and queryset The initial model parameters θ are used Perform training to obtain θ′ i : A multi-step forecasting method is used to predict the values of multiple time steps based on the previous observations. After each forecast, the query set is used Update the model parameters via gradient descent to minimize the meta-objective function:

[0048] A further improvement of the technical solution of the present invention is that: in step S5, the pre-training model parameters θ′ are loaded from the meta-training, and the target domain training data Input the prediction model fine-tuned in step S4 to update the model parameters to θ * , according to the set fault threshold, use the target domain test data Realize online real-time prediction of the remaining service life of wind turbine bearings;

[0049] The remaining useful life RUL is calculated as the duration from the current moment to the predicted failure time. The calculation formula is expressed as:

[0050] RUL(t)=inf{r:f(t+r)>μ}

[0051] Where f(t+r) represents the predicted MHI value at time t+r, and μ represents the fault threshold of the bearing.

[0052] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention are:

[0053] The online prediction method for the remaining useful life of wind turbine bearings under limited samples in the present invention combines the advantages of adaptive meta-learning AML and Informer network, which can solve the problem of limited samples in the process of online RUL prediction of wind turbine bearings and effectively improve the prediction efficiency and accuracy.

[0054] The online prediction method for rolling load of wind turbine bearings based on the AML-Informer integrated network proposed in this paper realizes multi-step prediction and long-term prediction based on vibration data. The online prediction results generated by this method can provide timely technical support for maintenance plans.

[0055] The present invention introduces the linear reliability index (LRI) to realize the linear mapping of multiple hierarchical structures of health indicators, improves the monotonicity and trend characteristics of the hierarchical structure, can express sufficient degradation information, and contributes to accurate prediction.

[0056] The Informer used in this paper is applied for the first time as a long time series prediction method to the RUL prediction of wind turbine bearings, and is integrated with the AML strategy, which effectively improves the low precision and small sample problems in long-term prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 This is a flowchart of the online prediction method for the remaining service life of a wind turbine bearing under a limited sample according to the present invention;

[0058] Figure 2 It is a structural diagram of the Informer network in the present invention;

[0059] Figure 3 It is a structural diagram of the multi-step long-term prediction model in the present invention;

[0060] Figure 4 is the MHI prediction result of bearing 1-1 in task A of the present invention;

[0061] Figure 5 It is the RUL prediction result of bearing 1-1 in task A of the present invention, where a is the RUL prediction result with AML added, and b is the RUL prediction result without AML added. DETAILED DESCRIPTION

[0062] The present invention is described in further detail below in conjunction with the embodiments:

[0063] like Figure 1 As shown, the present invention provides an online prediction method for the remaining service life of a wind turbine bearing under a finite sample, comprising the following steps:

[0064] Step S1: collecting wind turbine bearing vibration data, segmenting the data according to a preset sampling interval, and obtaining a segmented vibration signal sequence;

[0065] Specifically, the process of collecting wind turbine bearing vibration data can be described as follows: First, the vibration data set of the bearing provided by the PRONOSTIA platform is used. This includes horizontal and vertical vibration data recorded from run to failure under three different operating conditions. The data is recorded at a sampling frequency of 25.6kHz, with each sampling data segment having an interval of 10 seconds, and the fault threshold is set to 20g;

[0066] Step S2: Extract the health index HI from step S1 and design a linear reliability index to correct the bearing degradation trend, forming an index system that reflects the bearing degradation status. Then, obtain the mapped health index MHI through linear interpolation and normalize the mapped health index.

[0067] Specifically, the health index (HI) of step S1 is extracted. For the test bench data, the maximum amplitude is used as its HI. The HI value gradually increases with the deterioration of the bearing and rises sharply at the end of its life. The rank-preserving regression method is used to determine the failure onset time. A linear reliability index is designed to correct the bearing degradation trend, forming an indicator system that reflects the bearing degradation state. Then, the mapped health index (MHI) is obtained through linear interpolation, and the mapped health index is normalized.

[0068] Among them, the calculation formula of the linear reliability index is expressed as:

[0069]

[0070] Where, Represents the health indicator after mapping, t s Represents the time series corresponding to the mapped health indicator, k s and b s The function parameter representing the linear reliability index, k s and v s Respectively expressed as:

[0071]

[0072] b s =min(X)

[0073] Where, X={x1,x2,…,x m} represents the original health index, t o represents the time series of the original health indicator, max(·) represents the maximum value of the sequence, min(·) represents the minimum value of the sequence, and n is an integer;

[0074] The calculation formula of the linear interpolation method is expressed as:

[0075]

[0076] Where, X={x1,x2,…,x m} represents the original health index, Y={y1,y2,…,y m} represents the mapping health indicator, y i Located at x j and x j+1 between;

[0077] The normalization formula is expressed as:

[0078]

[0079] Where, X max and X min Represent the maximum and minimum values of the mapping health indicator respectively;

[0080] Step S3: Construct an Informer network model that integrates the adaptive meta-learning AML strategy. The Informer network constitutes the basic model framework for remaining life prediction, such as Figure 2 As shown in the figure, it includes a position encoding module, a ProbSparse self-attention mechanism, an encoder distillation module, and a generative decoder. It introduces an adaptive meta-learning strategy, uses source domain data for offline training to obtain model initialization parameters, and uses an adaptive learning rate update mechanism to improve the model convergence speed.

[0081] The key hyperparameter settings for Informer include: prediction sequence length: 10, window size: 96, model size: 512, number of attention heads: 8, number of encoder layers: 2, number of decoder layers: 1, activation function: RELU for distillation layers, GELU for encoder and decoder, dropout rate: 0.25, training epochs: 30, batch size: 32, optimizer: Adam (learning rate 0.0001);

[0082] Adaptive learning rate update is implemented to achieve fast convergence. The calculation formula of the adaptive learning rate update mechanism is expressed as:

[0083]

[0084] Where p1=0.9 and p2=0.999 represent the attenuation rate, ε=10 -8 is a constant used to avoid the denominator being zero, α t =10 -4 represents the learning rate at time t;

[0085] The calculation formula of the position encoding module of the Informer network is expressed as:

[0086]

[0087] Where, is the input of Informer, t represents the position number of the input sequence, i∈[0,d model -1], d model Represents the dimension of the Informer model, Represents the position-encoded input sequence;

[0088] The calculation formula of ProbSparse self-attention mechanism is expressed as:

[0089]

[0090] Where, K and V represent the query matrix, key matrix and value matrix respectively, d model Indicates the dimension of the input, l K and Represent K and length;

[0091] The calculation formula of the encoder distillation module is expressed as:

[0092]

[0093] Where, represents the input sequence of the encoder layer i, [·] Att Represents a multi-head attention module, Convld(·) represents a one-dimensional convolution filter, RELU(·) represents an activation function, and MaxPool(·) represents a maximum pooling layer with a step size of 2, which can convert the input sequence halved;

[0094] The calculation formula for updating the model parameters of the adaptive meta-learning strategy is expressed as:

[0095]

[0096]

[0097] In the formula, η represents the learning rate of task training, λ represents the learning rate of task verification, represents the loss calculated in task i over the support set, represents the loss calculated in task i through the query set, k represents the number of tasks, and f θ represents the model function with parameter θ;

[0098] The Informer network uses mean square error (MSE) as the loss function for optimization. The calculation formula is expressed as:

[0099]

[0100] Where b represents the batch size, x i represents the actual mapping health indicator value at time i, represents the predicted mapping health indicator value at time i;

[0101] Step S4: Fine-tune the pre-trained model in step S3 using historical data from the target domain to achieve knowledge transfer and adapt the model to the characteristics of the target device;

[0102] Source domain data D s Support Set and queryset The initial model parameters θ are used Perform training to obtain θ′ i : Using multi-step forecasting methods, we can predict the values of multiple time steps based on previous observations, such as Figure 3 As shown, after each prediction, the query set Update the model parameters via gradient descent to minimize the meta-objective function:

[0103] Step S5: The target domain real-time data is input into the prediction model fine-tuned in step S4, and the remaining service life of the wind turbine bearing is predicted online in real time according to the set fault threshold;

[0104] Specifically, load the pre-trained model parameters θ′ from meta-training, and the target domain training data Input the prediction model fine-tuned in step S4 to update the model parameters to θ * , according to the set fault threshold, use the target domain test data Realize online real-time prediction of the remaining service life of wind turbine bearings;

[0105] The remaining useful life RUL is calculated as the duration from the current moment to the predicted failure time. The calculation formula is expressed as:

[0106] RUL(t)=inf{r:f(t+r)>μ}

[0107] Where f(t+r) represents the predicted MHI value at time t+r, and μ represents the fault threshold of the bearing.

[0108] Take bearing 1-1 in Task A as an example. Figure 4 The MHI prediction results are shown. Figure 5 The RUL prediction results with and without AML were compared. The experiment showed that after adding AML, the prediction error was significantly reduced (e.g. Figure 5 a), and the prediction stability is higher (as shown in Figure 5 (b) Through verification on experimental bearings on the PRONOSTIA platform and actual wind turbine generator bearings, the method of the present invention achieves high-precision online RUL prediction under limited samples, and its prediction efficiency is better than traditional models such as GRU and standard Transformer.

[0109] In summary, this invention effectively addresses issues such as insufficient sample size, low prediction accuracy, and large cumulative error through the integration of physical information fusion for linear degradation correction, adaptive meta-learning for small-sample generalization optimization, and an efficient long-sequence prediction architecture. This provides reliable technical support for predicting the remaining useful life of wind turbines. For practical wind turbine implementations, this method can be integrated with existing condition monitoring systems to provide continuous online prediction of bearing RUL, enabling proactive maintenance planning and reducing unplanned downtime costs.

[0110] It will be understood that the present invention is described by way of some embodiments, and it will be appreciated by those skilled in the art that various changes or equivalent substitutions may be made to these features and embodiments without departing from the spirit and scope of the present invention. In addition, under the teachings of the present invention, these features and embodiments may be modified to adapt to specific circumstances and materials without departing from the spirit and scope of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed herein, and all embodiments falling within the scope of the claims of this application are intended to be protected by the present invention.

Claims

1. An online prediction method for the remaining service life of wind turbine bearings under finite samples, characterized by The following steps are involved: Step S1: collecting wind turbine bearing vibration data, segmenting the data according to a preset sampling interval, and obtaining a segmented vibration signal sequence; Step S2: Extract the health index HI from step S1 and design a linear reliability index to correct the bearing degradation trend, forming an index system that reflects the bearing degradation status. Then, obtain the mapped health index MHI through linear interpolation and normalize the mapped health index. Step S3: Construct an Informer network model that integrates an adaptive meta-learning AML strategy. The Informer network constitutes the basic model framework for remaining useful life prediction, which includes a position encoding module, a ProbSparse self-attention mechanism, an encoder distillation module, and a generative decoder. An adaptive meta-learning strategy is introduced, and offline training is performed using source domain data to obtain model initialization parameters. An adaptive learning rate update mechanism is used to improve the model convergence speed. Step S4: Fine-tune the pre-trained model in step S3 using historical data from the target domain to achieve knowledge transfer and adapt the model to the characteristics of the target device; Step S5: The target domain real-time data is input into the prediction model fine-tuned in step S4, and the online real-time prediction of the remaining service life of the wind turbine bearing is realized according to the set fault threshold.

2. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 1 is characterized by: In step S2, the health index HI of step S1 is extracted. For the test bench data, the maximum amplitude is used as its HI. The HI value gradually increases with the deterioration of the bearing and rises sharply at the end of its life. The rank-preserving regression method is used to determine the fault start time, and a linear reliability index is designed to correct the bearing degradation trend to form an indicator system reflecting the bearing degradation status. Then, the mapped health index MHI is obtained through the linear interpolation method, and the mapped health index is normalized.

3. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 2 is characterized by: The calculation formula of the linear reliability index is expressed as: Where, Represents the health indicator after mapping, t s Represents the time series corresponding to the mapped health indicator, k s and b s The function parameter representing the linear reliability index, k s and b s Respectively expressed as: b s =min(X) Where, X={x1,x2,…,x m } represents the original health index, t o represents the time series of the original health indicator, max(·) represents the maximum value of the sequence, min(·) represents the minimum value of the sequence, and n is an integer; The calculation formula of the linear interpolation method is expressed as: Where, X={x1,x2,…,x m } represents the original health index, Y={y1,y2,…,y m } represents the mapping health indicator, y i Located at x j and x j+1 between; The normalization processing formula is expressed as: Where, X max and X min Represent the maximum and minimum values of the mapping health indicator respectively.

4. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 1 is characterized by: In step S3, the calculation formula of the adaptive learning rate update mechanism is expressed as: Where p1=0.9 and p2=0.999 represent the attenuation rate, ε=10 -8 is a constant used to avoid the denominator being zero, α t =10 -4 represents the learning rate at time t; The calculation formula of the position encoding module of the Informer network is expressed as: Where, is the input of Informer, t represents the position number of the input sequence, i∈[0,d model -1], d model Represents the dimension of the Informer model, Represents the position-encoded input sequence.

5. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 4 is characterized by: The calculation formula of the ProbSparse self-attention mechanism is expressed as: Where, K and V represent the query matrix, key matrix and value matrix respectively, d model Indicates the dimension of the input, l K and Represent K and length.

6. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 5 is characterized by: The calculation formula of the encoder distillation module is expressed as: Where, represents the input sequence of the encoder layer i, [·] Att Represents a multi-head attention module, Convld(·) represents a one-dimensional convolution filter, RELU(·) represents an activation function, and MaxPool(·) represents a maximum pooling layer with a step size of 2, which can convert the input sequence Halved.

7. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 6 is characterized by: The calculation formula for updating the model parameters of the adaptive meta-learning strategy is expressed as: In the formula, η represents the learning rate of task training, λ represents the learning rate of task verification, represents the loss calculated in task i over the support set, represents the loss calculated in task I through the query set, k represents the number of tasks, and f θ represents the model function with parameter θ.

8. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 7 is characterized by: The Informer network uses mean square error (MSE) as the loss function for optimization, and the calculation formula is expressed as: Where b represents the batch size, x i represents the actual mapping health indicator value at time i, represents the predicted mapping health indicator value at time i.

9. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 1 is characterized by: In step S4, the source domain data D s Support Set and queryset The initial model parameters θ are used Perform training to obtain θ i ′ : A multi-step forecasting method is used to predict the values of multiple time steps based on the previous observations. After each forecast, the query set is used Update the model parameters via gradient descent to minimize the meta-objective function:

10. The online prediction method for the remaining service life of a wind turbine bearing under a finite sample according to claim 1, characterized in that: In step S5, the pre-trained model parameters θ are loaded from meta-training ′ , target domain training data Input the prediction model fine-tuned in step S4 to update the model parameters to θ * , according to the set fault threshold, use the target domain test data Realize online real-time prediction of the remaining service life of wind turbine bearings; The remaining useful life RUL is calculated as the duration from the current moment to the predicted failure time. The calculation formula is expressed as: RUL(t)=inf{r:f(t+r)>μ} Where f(t+r) represents the predicted MHI value at time t+r, and μ represents the fault threshold of the bearing.

Images