Cross-domain fan bearing remaining service life prediction method based on adversarial domain adaptation
Through the anti-domain domain adaptation cross-domain fan bearing residual service life prediction method, the problem of scarce fan bearing failure samples is solved, and the accurate prediction of experimental bench bearings to real fan bearings is achieved, which improves the prediction accuracy and cross-domain adaptability of the model.
Patent Information
- Application Number
- CN202510657077.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-21
- Publication Date
- 2025-08-08
AI Technical Summary
In the prediction of residual service life of fan bearings, due to the scarcity of fault samples and the small available training data, the model prediction effect is poor, making it difficult to achieve accurate migration and multiplexing from the experimental bench bearing to the real fan bearing.
The residual service life prediction method of cross-domain fan bearings based on adversarial domain adaptation is adopted. By constructing a weight-sharing feature extraction module, domain classification module and domain adaptation residual service life prediction module, combining minimum-maximum normalization, fusion attention mechanism and multi-core maximum mean difference loss function, feature migration and distribution alignment are achieved, and prediction accuracy is improved.
The accurate construction and migration and reuse of the residual service life prediction model from the experimental bench bearing to the real fan bearing is realized, and the accuracy and cross-domain adaptability of the residual service life prediction of the fan bearing are improved.
Smart Images
Figure CN120449706A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the technical field of remaining service life prediction of wind turbine bearings, and specifically relates to a cross-domain remaining service life prediction method for wind turbine bearings based on adversarial domain adaptation. Background Art
[0002] As global dependence on fossil fuels gradually decreases, the use of wind energy not only aligns with the global pursuit of sustainable development but also becomes a key pathway to achieving carbon neutrality, playing an increasingly important role in the global energy system. While the wind power industry is booming, it also faces new challenges. Wind turbines operate under complex conditions for long periods of time, leading to a high incidence of transmission system failures. Bearings, as core transmission components, can cause serious accidents such as downtime if they fail. Therefore, accurately predicting the remaining service life of wind turbine bearings is crucial for avoiding catastrophic consequences and protecting personnel and equipment.
[0003] The advantages of deep learning-based methods in the field of remaining useful life prediction lie primarily in their powerful data processing capabilities. They can improve the accuracy and robustness of remaining useful life prediction by automatically extracting features and capturing nonlinearities and global dependencies. However, their "end-to-end" approach relies heavily on data quality and quantity, and only performs well under big data conditions. However, in real-world applications, wind turbine bearing monitoring data often exhibits a scarcity of fault samples and limited available training data, which can severely impact the model's predictive effectiveness.
[0004] Therefore, there is a need for a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation that can accurately construct and migrate the remaining service life prediction model from the experimental bench bearing to the real wind turbine bearing, and improve the accuracy of wind turbine bearing remaining service life prediction. Summary of the Invention
[0005] The purpose of the present invention is to provide a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation. By improving the transferable prediction model, it can realize the accurate construction and migration reuse of the remaining service life prediction model from the experimental bench bearing to the real wind turbine bearing, thereby improving the remaining service life prediction accuracy of the wind turbine bearing.
[0006] To achieve the above object, the technical solution adopted by the present invention is:
[0007] A cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation includes the following steps:
[0008] Step S1: performing normalization preprocessing on the vibration data of the experimental platform bearing and the fan bearing respectively;
[0009] Step S2: Construct a remaining useful life prediction model, which includes a weight-sharing feature extraction module, a domain classification module, and a domain-adaptive remaining useful life prediction module, wherein:
[0010] The feature extraction module is a convolutional neural network-gated recurrent unit architecture integrated with an attention mechanism, including a data input layer, a convolutional neural network convolution operation layer, a gated recurrent unit timing modeling layer, an attention weight allocation layer, and a feature output layer. The convolutional neural network convolution operation layer of the feature extraction module captures local features of the vibration signal through convolution operations, the gated recurrent unit timing modeling layer is used to model timing dependencies, and the attention weight allocation layer uses the attention mechanism to enhance the model's ability to focus on key features, thereby improving the model's performance when processing long time series data.
[0011] The domain classification module consists of two fully connected layers and one domain classification layer. Its input is the high-level features of the source and target domains extracted by the feature extraction module. It uses an adversarial training mechanism to reduce the difference in feature distribution between the source and target domains, and uses adversarial game theory to obfuscate domain differences.
[0012] The domain-adaptive remaining useful life prediction module includes a fully connected layer and a linear regression layer, with the output of the feature extraction module as input. The fully connected layer maps the source domain features and the target domain features to a shared latent space, using a multi-kernel maximum mean difference loss function to quantify the distribution difference between the test bench bearing data and the wind turbine bearing data. By minimizing this difference, the model adapts between the different data sources and generates shared features with domain invariance. The linear regression layer maps the features to the remaining useful life value, establishing a mapping relationship from the feature space to the remaining useful life value.
[0013] Step S3: training the remaining useful life prediction model according to the total loss function to achieve feature migration;
[0014] Step S4: pre-process the target wind turbine bearing data set and input it into the trained remaining service life prediction model to verify the prediction effect.
[0015] A further improvement of the technical solution of the present invention is that in step S1, the vibration signal is preprocessed using the minimum-maximum normalization method to constrain the data to a fixed interval [0, 1], eliminate dimensional differences, and provide a stable input benchmark for the remaining service life prediction model to accelerate the convergence process of the prediction model and improve training efficiency. The process is shown in the following formula:
[0016]
[0017] Where, represents the preprocessing result of the i-th data point of the j-th bearing, and accordingly, They represent the minimum and maximum values of the original data of the j-th bearing respectively.
[0018] A further improvement of the technical solution of the present invention is that: in step S2, the feature extraction module forms an end-to-end processing flow from the original vibration signal to the timing feature; the original vibration signal is enhanced by multi-layer convolution to achieve local sensitive feature enhancement, and after the feature dimension is reduced by the pooling operation, it is fused into a feature vector by the fully connected layer; the gated recurrent unit timing modeling layer of the joint attention mechanism models the key timing dependencies of the feature vector, and finally outputs the extracted features.
[0019] A further improvement of the technical solution of the present invention is that: in step S2, the error between the domain category and the domain label output by the domain classification module is quantified using the binary cross entropy function to maximize the discrimination ability of the classifier and ensure the controllability of the feature confusion effect. The domain classification loss function L D It can be expressed as follows:
[0020]
[0021] Where N represents the number of batch samples, d i ∈{0,1} represents the sample domain label.
[0022] A further improvement of the technical solution of the present invention is that: in step S2, the domain-adaptive remaining useful life prediction module introduces the multi-core maximum mean difference as a regularization constraint, and with the help of the kernel function, the features of the source domain and the target domain are mapped to the reproducing kernel Hilbert space, and the distribution difference of the features of the two domains is quantitatively calculated to align the cross-domain feature distributions; the multi-core maximum mean difference is based on the traditional maximum mean difference method, and by combining multiple kernel functions, it adaptively selects the kernel suitable for the current data to comprehensively measure the distribution difference.
[0023] A further improvement of the technical solution of the present invention is that: the source domain feature X S and target domain features X T The multi-core maximum mean difference metric formula is as follows:
[0024]
[0025] Where l represents the number of layers with the maximum mean difference among multiple cores, H K denotes the reproducing kernel Hilbert space of a particular kernel, and E[·] denotes the expectation of a given distribution;
[0026] The characteristic kernel function k is obtained as follows:
[0027]
[0028] Where, β u represents the weight assigned to each kernel function, ku (·) represents u basic kernels, and r represents the total number of kernel functions.
[0029] A further improvement of the technical solution of the present invention is that the domain-adaptive remaining useful life prediction module uses a regression network to construct a predictor, maps the degradation features output by the feature extraction module to the remaining useful life to achieve prediction, and simultaneously reduces the loss function between the remaining useful life prediction value and the label value, and uses a mean square error function to improve accuracy. Its expression is shown in the following formula:
[0030]
[0031] In the formula, rul i Indicates the actual remaining service life tag value, rul pi Represents the prediction result, and N represents the batch size.
[0032] A further improvement of the technical solution of the present invention is that: in step S3, the loss function is first optimized and trained to optimize the remaining useful life prediction model parameters and improve the cross-domain generalization capability, and then the Adam optimizer is used to train the model to find a suitable parameter solution.
[0033] A further improvement of the technical solution of the present invention is that the total loss L of the remaining useful life prediction model in the back propagation process includes the regression loss L R , MK-MMD loss L M and domain classification loss L D , the expression is as follows:
[0034] L=L R +λ1L M +λ2L D
[0035] Where λ1 and λ2 represent the trade-off coefficients, which control the strength of distribution alignment and domain adversarial optimization respectively.
[0036] Due to the adoption of the above technical solution, the technical advancements achieved by the present invention are:
[0037] The present invention is based on a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation. By improving the transferable prediction model, it can achieve accurate construction and migration reuse of the remaining service life prediction model from the experimental bench bearing to the real wind turbine bearing, thereby improving the remaining service life prediction accuracy of the wind turbine bearing.
[0038] The present invention adopts a convolutional neural network-gated recurrent unit architecture that integrates an attention mechanism to capture global long-term dependencies while extracting local key features, thereby enhancing feature expression capabilities.
[0039] This paper proposes a dual-constraint feature distribution alignment strategy. It embeds a multi-core maximum average difference metric function in the feature space to quantify the difference in feature distribution between the source and target domains. It also combines the adversarial training strategy of the domain classification module to achieve dual constraint feature distribution alignment. BRIEF DESCRIPTION OF THE DRAWINGS
[0040] Figure 1 This is a flowchart of a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to the present invention;
[0041] Figure 2 This is a diagram of the remaining useful life prediction model architecture of the present invention;
[0042] Figure 3 This is the architecture diagram of the convolutional neural network-gated recurrent unit based on the fusion attention mechanism in the present invention;
[0043] Figure 4 is the prediction result graph of the cross-domain prediction task a in the present invention;
[0044] Figure 5 It is the prediction result diagram of the cross-domain prediction task b in the present invention. DETAILED DESCRIPTION
[0045] The present invention is described in further detail below in conjunction with the embodiments:
[0046] like Figure 1 As shown, the present invention provides a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation, comprising the following steps:
[0047] Step S1: performing normalization preprocessing on the vibration data of the experimental platform bearing and the fan bearing respectively;
[0048] Specifically, the vibration signal is preprocessed using the minimum-maximum normalization method to constrain the data to a fixed interval [0, 1], eliminate dimensional differences, and provide a stable input benchmark for the remaining service life prediction model to accelerate the convergence process of the prediction model and improve training efficiency. The process is shown in the following formula:
[0049]
[0050] Where, represents the preprocessing result of the i-th data point of the j-th bearing, and accordingly, Respectively represent the minimum and maximum values of the original data of the j-th bearing;
[0051] Step S2: Figure 2As shown in Figure 1, a remaining useful life prediction model is constructed, which includes a weight-sharing feature extraction module, a domain classification module, and a domain-adaptive remaining useful life prediction module; θ f ,θ p ,θ d denote the parameters of the feature extraction module, the domain adaptation remaining useful life prediction module, and the domain classification module, respectively;
[0052] Among them, such as Figure 3 As shown in the figure, the feature extraction module is a convolutional neural network-gated recurrent unit architecture integrated with an attention mechanism, including a data input layer, a convolutional neural network convolution operation layer, a gated recurrent unit timing modeling layer, an attention weight allocation layer, and a feature output layer. The convolutional neural network convolution operation layer of the feature extraction module captures the local features of the vibration signal through convolution operations, the gated recurrent unit timing modeling layer is used to model timing dependencies, and the attention weight allocation layer uses the attention mechanism to enhance the model's ability to focus on key features, thereby improving the model's performance when processing long time series data.
[0053] Specifically, the feature extraction module forms an end-to-end processing flow from raw vibration signals to temporal features. The raw vibration signals are enhanced with locally sensitive features through multi-layer convolution, reduced in dimension through pooling, and then fused into feature vectors by a fully connected layer. The gated recurrent unit temporal modeling layer, combined with the attention mechanism, models the key temporal dependencies of the feature vectors and ultimately outputs the extracted features.
[0054] The domain classification module consists of two fully connected layers and one domain classification layer. Its input is the high-level features of the source and target domains extracted by the feature extraction module. Its core goal is to reduce the feature distribution differences between the source and target domains through adversarial training and further obfuscate domain differences through adversarial game.
[0055] Specifically, the domain classification module forms an adversarial game mechanism with the feature extraction module by distinguishing whether the extracted features belong to the source domain or the target domain: when the feature classification accuracy improves, the feature extractor is forced to generate shared features that are difficult to identify the domain. This dynamic optimization process can effectively improve the cross-domain adaptability of the prediction model. During the adversarial training process, the gradient reversal layer between the domain classification module and the feature extraction module can pass the feature identity to the domain classification module during the forward propagation process, and invert the gradient of the domain classification module during the backpropagation process before passing it to the feature extraction module.
[0056] Furthermore, the binary cross entropy function is used to quantify the error between the domain category and the domain label output by the domain classification module to maximize the discriminative ability of the classifier and ensure the controllability of the feature confusion effect. The domain classification loss function L D It can be expressed as follows:
[0057]
[0058] Where N represents the number of batch samples, d i ∈{0,1} represents the sample domain label;
[0059] The domain-adaptive RLS prediction module consists of a fully connected layer and a linear regression layer, with the output of the feature extraction module as input. The fully connected layer maps the source and target domain features into a shared latent space, using a multi-kernel maximum mean difference loss function to quantify the distribution differences between the bench bearing data and the wind turbine bearing data. By minimizing this difference, the model can better adapt between different data sources and generate shared features with domain invariance. The linear regression layer maps the features to the RLS value, establishing a mapping relationship from the feature space to the RLS value.
[0060] Specifically, the domain-adaptive remaining useful life prediction module introduces multi-kernel maximum mean difference as a regularization constraint. Its essence is to use the kernel function to map the features of the source and target domains into the reproducing kernel Hilbert space, quantitatively calculating the distribution difference of the features of the two domains to align the cross-domain feature distributions. Multi-kernel maximum mean difference is based on the traditional maximum mean difference method. By combining multiple kernel functions, it adaptively selects the kernel that is suitable for the current data to comprehensively measure the distribution difference.
[0061] More specifically, the source domain feature X S and target domain features X T The multi-core maximum mean difference metric formula is as follows:
[0062]
[0063] Where l represents the number of layers with the maximum mean difference among multiple cores, H K denotes the reproducing kernel Hilbert space of a particular kernel, and E[·] denotes the expectation of a given distribution;
[0064] The characteristic kernel function k is obtained as follows:
[0065]
[0066] Where, β u represents the weight assigned to each kernel function, k u (·) represents u basic kernels, and r represents the total number of kernel functions;
[0067] The domain-adaptive remaining useful life prediction module uses a regression network to construct a predictor, maps the degradation features output by the feature extraction module to the remaining useful life to achieve prediction, and at the same time reduces the loss function between the remaining useful life prediction value and the label value. The mean square error function is used to improve the accuracy. Its expression is shown in the following formula:
[0068]
[0069] In the formula, rul i Indicates the actual remaining service life tag value, rul pi Represents the prediction result, N represents the batch size;
[0070] Step S3: training the remaining useful life prediction model according to the total loss function to achieve feature migration;
[0071] The structural parameters of the prediction model are as follows: the input of the feature extraction module is a one-dimensional vibration signal with a shape of 1×2560. During the convolution operation, the convolution kernel size in the first convolution layer is 3, the stride and padding are both 1, and the output size is 2560. After passing through the first pooling layer with the window size and stride set to 2, the output is 1280. After that, the output size of the second convolution layer and pooling layer is 640. After entering the gated recurrent unit layer and the attention layer operation, the output is 128. The output of the feature extraction module enters the remaining service life prediction module and the domain classification module, and the final output size after passing through the fully connected layer is 1.
[0072] Specifically, we first optimize the loss function to optimize the remaining useful life prediction model parameters and improve cross-domain generalization capabilities. Then, we use the Adam optimizer to train the model and find the appropriate parameter solution.
[0073] Among them, the total loss L of the remaining service life prediction model in the back propagation process includes the regression loss L R , MK-MMD loss L M and domain classification loss L D , the expression is as follows:
[0074] L=L R +λ1L M +λ2L D
[0075] Where λ1 and λ2 represent the trade-off coefficients, which control the strength of distribution alignment and domain adversarial optimization respectively;
[0076] Step S4: pre-processing the target wind turbine bearing data set and inputting it into the trained remaining service life prediction model to verify the prediction effect;
[0077] Both migration tasks a and b use five bearings (1-1, 1-2, 2-1, 2-2, and 3-2) from the PRONOSTIA test bench bearing dataset as source domain training data. The target domain training data comes from bearing monitoring data from a wind turbine. After data preprocessing, these data are fed into the trained remaining useful life prediction model. The test set, which evaluates the prediction performance of the proposed method, is based on bearing operating data from another wind turbine.
[0078] The prediction results of cross-domain prediction tasks a and b are as follows Figure 4 and Figure 5 As shown in the two cross-domain migration prediction experiments, the predicted values of the present invention fluctuate slightly around the true values. Although there are local deviations, they show a highly synchronized degradation trend and a high degree of fit, which verifies the cross-domain migration capability of the remaining useful life prediction model for global degradation knowledge.
[0079] In summary, the present invention adopts a cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation and improves the transferable prediction model. It can realize the accurate construction and migration reuse of the remaining service life prediction model from the experimental bench bearing to the real wind turbine bearing, thereby improving the remaining service life prediction accuracy of the wind turbine bearing.
[0080] 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. A cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation, characterized by The following steps are involved: Step S1: performing normalization preprocessing on the vibration data of the experimental platform bearing and the fan bearing respectively; Step S2: Constructing a remaining useful life prediction model, which includes a weight-sharing feature extraction module, a domain classification module, and a domain-adaptive remaining useful life prediction module, wherein: The feature extraction module is a convolutional neural network-gated recurrent unit architecture integrated with an attention mechanism, including a data input layer, a convolutional neural network convolution operation layer, a gated recurrent unit timing modeling layer, an attention weight allocation layer, and a feature output layer. The convolutional neural network convolution operation layer of the feature extraction module captures local features of the vibration signal through convolution operations, the gated recurrent unit timing modeling layer is used to model timing dependencies, and the attention weight allocation layer uses the attention mechanism to enhance the model's ability to focus on key features, thereby improving the model's performance when processing long time series data. The domain classification module consists of two fully connected layers and one domain classification layer. Its input is the high-level features of the source and target domains extracted by the feature extraction module. It uses an adversarial training mechanism to reduce the difference in feature distribution between the source and target domains, and uses adversarial game theory to obfuscate domain differences. The domain-adaptive remaining useful life prediction module includes a fully connected layer and a linear regression layer, with the output of the feature extraction module as input. The fully connected layer maps the source domain features and the target domain features to a shared latent space, using a multi-kernel maximum mean difference loss function to quantify the distribution difference between the test bench bearing data and the wind turbine bearing data. By minimizing this difference, the model adapts between the different data sources and generates shared features with domain invariance. The linear regression layer maps the features to the remaining useful life value, establishing a mapping relationship from the feature space to the remaining useful life value. Step S3: training the remaining useful life prediction model according to the total loss function to achieve feature migration; Step S4: pre-process the target wind turbine bearing data set and input it into the trained remaining service life prediction model to verify the prediction effect.
2. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 1 is characterized by: In step S1, the vibration signal is preprocessed using the minimum-maximum normalization method to constrain the data to a fixed interval [0, 1], eliminate dimensional differences, and provide a stable input benchmark for the remaining service life prediction model to accelerate the convergence process of the prediction model and improve training efficiency. The process is shown in the following formula: Where, represents the preprocessing result of the i-th data point of the j-th bearing, and accordingly, They represent the minimum and maximum values of the original data of the j-th bearing respectively.
3. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 1 is characterized by: In step S2, the feature extraction module forms an end-to-end processing flow from the original vibration signal to the timing feature; the original vibration signal is enhanced by multiple layers of convolution to achieve local sensitive feature enhancement, and after the feature dimension is reduced by the pooling operation, it is fused into a feature vector by the fully connected layer; the gated recurrent unit timing modeling layer of the joint attention mechanism models the key timing dependencies of the feature vector, and finally outputs the extracted features.
4. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 3 is characterized by: In step S2, the binary cross entropy function is used to quantify the error between the domain category and the domain label output by the domain classification module to maximize the discriminative ability of the classifier and ensure the controllability of the feature confusion effect. The domain classification loss function L D It can be expressed as follows: Where N represents the number of batch samples, d i ∈{0,1} represents the sample domain label.
5. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 4 is characterized by: In step S2, the domain-adaptive RSU prediction module introduces the multi-kernel maximum mean difference as a regularization constraint. With the help of the kernel function, the features of the source and target domains are mapped into the reproducing kernel Hilbert space, and the distribution difference of the features of the two domains is quantified to align the cross-domain feature distributions. Multi-kernel maximum mean difference is based on the traditional maximum mean difference method. By combining multiple kernel functions, it adaptively selects the kernel suitable for the current data to comprehensively measure the distribution difference.
6. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 5 is characterized by: Source domain feature X S and target domain features X T The multi-core maximum mean difference metric formula is as follows: Where l represents the number of layers with the maximum mean difference among multiple cores, H K denotes the reproducing kernel Hilbert space of a particular kernel, and E[·] denotes the expectation of a given distribution; The characteristic kernel function k is obtained as follows: Where, β u represents the weight assigned to each kernel function, k u (·) represents u basic kernels, and r represents the total number of kernel functions.
7. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 6 is characterized by: The domain-adaptive remaining useful life prediction module uses a regression network to construct a predictor, maps the degradation features output by the feature extraction module to the remaining useful life to achieve prediction, and at the same time reduces the loss function between the remaining useful life prediction value and the label value. The mean square error function is used to improve the accuracy. Its expression is shown in the following formula: In the formula, rul i Indicates the actual remaining service life tag value, rul pi Represents the prediction result, and N represents the batch size.
8. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 1 is characterized by: In step S3, the loss function is first optimized and trained to optimize the remaining useful life prediction model parameters and improve the cross-domain generalization capability. Then, the Adam optimizer is used to train the model to find a suitable parameter solution.
9. The cross-domain wind turbine bearing remaining service life prediction method based on adversarial domain adaptation according to claim 8 is characterized by: The total loss L of the remaining useful life prediction model during back propagation includes the regression loss L R , MK-MMD loss L M and domain classification loss L D , the expression is as follows: L=L R +λ1L M +λ2L D Where λ1 and λ2 represent the trade-off coefficients, which control the strength of distribution alignment and domain adversarial optimization respectively.