Degradation equipment online domain migration residual life prediction method
By combining comparative learning and adversarial learning, the problems of data requirements and model fixation in online domain adaptation of equipment are solved, enabling efficient prediction and accurate migration of equipment remaining lifespan to adapt to changes in operating conditions.
Patent Information
- Application Number
- CN202511010082.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-22
- Publication Date
- 2025-11-21
AI Technical Summary
Existing methods for predicting the remaining useful life of equipment require target domain data or labels, the model is fixed and cannot be updated online, and the generalization performance degrades when the domain differences are too large.
A prediction model based on contrastive learning is pre-trained and an online cross-condition training based on adversarial learning is conducted. The model is updated using source domain data and online-collected target domain data. The distribution differences are aligned by momentum encoding and Gaussian kernel function, and loss functions at the individual level and distribution level are introduced to achieve online domain adaptation.
It achieves online model updates and domain adaptation without requiring additional target domain data and labels, significantly improving the generalization ability and transfer prediction accuracy of the prediction model.
Smart Images
Figure CN120995833A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of predictive maintenance and health management of equipment, and more particularly to a method for predicting the remaining life of degraded equipment in the field of online migration. Background Technology
[0002] Remaining useful life (RUL) prediction technology is a technique that predicts equipment failure time using historical data or mechanistic models. This technology can avoid unexpected downtime and optimize maintenance decisions. Prediction techniques mainly include data-driven and model-driven methods. Data-driven methods use sensor data for modeling, requiring no specialized knowledge and having strong applicability; model-driven methods rely on physical mechanisms, offering high accuracy but poor generalization.
[0003] In real-world applications, equipment data distribution dynamically changes due to shifts in operating conditions or evolving faults, making it difficult for traditional offline models to adapt. To address this issue, researchers have proposed domain adaptation methods, which achieve knowledge transfer by reducing the distributional differences between the source and target domains. Typical domain adaptation methods include: domain adversarial networks based on adversarial training, which learn domain-invariant features through a domain discriminator; and distribution alignment strategies, which reduce the source / target domain distributional differences using metrics such as maximum mean square distance (MMD) and correlation matrix alignment. However, these existing methods still have certain limitations: they require target domain data or labels, while industrial data is usually online and unlabeled; most models are fixed after training and cannot be updated online to adapt to dynamic degradation; and generalization performance degrades when the domain differences are too large.
[0004] Therefore, in order to solve the above problems, this invention proposes a method for predicting the online domain adaptive lifespan of degraded equipment under changing operating conditions, based on the ideas of contrastive learning and adversarial learning. Summary of the Invention
[0005] The present invention aims to provide an online domain migration remaining lifetime prediction method for degraded equipment, in order to solve the problems of existing prediction methods requiring target domain data or labels, fixed models that cannot be updated online, and declining generalization performance when domain differences are too large.
[0006] To achieve the above objectives, the present invention provides the following technical solution:
[0007] A method for predicting the remaining lifetime of degraded equipment via online domain migration includes a prediction model pre-training part based on contrastive learning and an online cross-condition training part based on adversarial learning.
[0008] The pre-training part of the prediction model based on contrastive learning includes the following steps:
[0009] A1. In the source domain data, select two sets of data with similar remaining useful life labels and construct similar sample pairs:
[0010]
[0011] Where k represents the number of sample pairs in the same training batch, and Shared remaining useful life label y i ;
[0012] A2. Using a momentum coding strategy, the network parameters of the main branch are updated in real time through backpropagation during training, while the network parameters of the other branch do not participate in gradient updates but slowly follow the changes in the main branch parameters through exponential moving average. The feature extraction and prediction process of the main branch and sub-branch of the Siamese neural network is represented as follows:
[0013]
[0014] in, and These represent the prediction processes for the main branch and the sub-branch, respectively. and These represent the main branch extraction process;
[0015] The supervised prediction loss of the main branch is expressed in the following form:
[0016]
[0017] Among them, y i This represents the true remaining lifetime of the i-th sample;
[0018] A3. Introduce individual-level contrastive loss to make each pair of similar samples closer in the output representation space. The individual-level contrastive loss is expressed as follows:
[0019]
[0020] in, sim(u,v)=u T v represents the dot product similarity, and the denominator iterates through the division... and In addition to the remaining 2k-2 implicit negative samples;
[0021] A4. To ensure a high degree of overlap in the overall distribution of the entire batch of data, a Gaussian kernel function is used for MMD alignment, and the following loss function at the distribution level is defined:
[0022]
[0023] Where χ denotes the reproducing kernel Hilbert space mapping, and the kernel function is chosen as γ(u,v) = exp(-||uv|| 2 / 2σ 2 Then, the total loss function compared to pre-training is expressed as:
[0024] L COM =L RUL +α·L IND +β·L DIS
[0025] By minimizing the total loss function, the network parameters of the backbone feature extraction module and the predictor can be updated in reverse, while the network parameters of the sub-branches can be updated using EMA momentum.
[0026] The online cross-condition training component based on adversarial learning includes the following steps:
[0027] B1. Data x collected from equipment operating online under target conditions t Accordingly, monitoring data for the corresponding time period is selected from the source domain devices, and the average is taken to obtain x. s The corresponding average remaining life label is y. s The pre-trained main branch feature extraction network and predictor network are respectively represented as F. θ and F φ After inputting the corresponding source domain data into the feature extractor and predictor, the predicted remaining lifetime is obtained. The online supervised prediction loss is defined as follows:
[0028]
[0029] B2. After inputting the online target domain data into the feature extractor and predictor, the estimated remaining lifetime under the current operating state can be obtained. The two RUL prediction values are then sent to the domain discriminator, resulting in the following form:
[0030]
[0031] Where, d t and d s D represents the output of the domain discriminator predicting the RUL values for the source and target domains, respectively. ψ The domain discriminator consists of hidden layers and fully connected layers with sigmoid activation functions; GRL stands for gradient inversion layer.
[0032] B3. Define the following domain discriminator labels:
[0033]
[0034] The online adversarial loss of the output layer is defined as follows:
[0035]
[0036] B4. Because the number of samples in each online training session is small, the model is more prone to overfitting to noise. Therefore, a regularization term is added, and the total loss function expression is as follows:
[0037]
[0038] Since the goal of the feature extractor and predictor is to maximize the classification error, while the goal of the domain discriminator is to minimize the classification error, during forward propagation, let GRL(x) = x, and during backward propagation, let:
[0039]
[0040] Where λ adv This represents the gradient coefficient, which is always a positive number.
[0041] Through the above process, whenever a new batch of online monitoring data is collected, the parameters of the feature extraction network, predictor, and domain discriminator can be updated in reverse by minimizing the loss function, thereby achieving continuous optimization and domain adaptation of the online prediction model.
[0042] Furthermore, in step A2, the predictor generally uses a fully connected layer, and feature extraction can be selected as needed.
[0043] The principle and beneficial effects of this technical solution:
[0044] (1) The method of this invention is based on the idea of output layer adversarial, and uses online collected target domain data and corresponding source domain data to update the existing prediction model. This process does not require any additional target domain data and target domain remaining lifetime labels, and realizes the migration from the source domain to the target domain online, which is more in line with the actual operating conditions of degraded equipment.
[0045] (2) The present invention is based on the contrastive learning method, which uses source domain data with similar remaining lifetime labels to train the model and utilizes the invariant features of the learning domain to significantly improve the generalization ability of the prediction model. The advantages of the present invention have been verified on actual aerospace engine datasets. Attached Figure Description
[0046] Figure 1 This is a flowchart of the method of the present invention;
[0047] Figure 2The present invention provides the RUL curves for a typical engine under six migration tasks; a) FD001→FD002; b) FD001→FD003; c) FD001→FD004; d) FD002→FD001; e) FD002→FD003; f) FD002→FD004.
[0048] Figure 3 The present invention provides the RUL curves for a typical engine under six other migration tasks; g) is FD003→FD001; h) is FD003→FD002; i) is FD003→FD004; j) is FD004→FD001; k) is FD004→FD002; l) is FD004→FD003. Detailed Implementation
[0049] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments:
[0050] like Figure 1 The diagram shown is a complete flowchart of the comparative pre-training and online adversarial domain adaptation process in the online domain migration remaining lifetime prediction method for degraded equipment according to the present invention.
[0051] The following is a detailed description of the method of the present invention:
[0052] 1. Pre-training of prediction models based on contrastive learning
[0053] To improve the generalization ability of the prediction model, the model is pre-trained offline based on source domain data and the idea of contrastive learning before online domain adaptation.
[0054] In the source domain data, select two sets of data with similar remaining useful life labels to construct "similar sample pairs":
[0055]
[0056] Where k represents the number of sample pairs in the same batch, and Shared remaining useful life label y i .
[0057] Since only positive sample pairs with similar lifetime labels participate in training, and there are no negative sample pairs, contrastive learning is prone to going to extremes: mapping all inputs to the same constant vector, resulting in insufficient discriminative power. To avoid this problem, this invention adopts a momentum encoding strategy. During training, the network parameters of the main branch are updated in real time through backpropagation, while the network parameters of the other branch do not participate in gradient updates and slowly follow the changes in the parameters of the main branch through exponential moving average (EMA).
[0058] The feature extraction and prediction process of the main branch and sub-branch of the contrastive learning Siamese neural network can be represented as follows:
[0059]
[0060] in, and These represent the prediction processes for the main branch and the sub-branch, respectively. and These represent the main branch extraction process; the predictor generally uses a fully connected layer, and feature extraction can be selected as needed.
[0061] The supervised prediction loss of the main branch can be expressed in the following form:
[0062]
[0063] Among them, y i This represents the true remaining lifetime of the i-th sample;
[0064] While supervised regression loss ensures accuracy in the source domain, it cannot actively eliminate distributional differences between the source and target domains. To enable the model to quickly adapt to the unknown target domain while maintaining high accuracy in the source domain, contrastive loss needs to be introduced.
[0065] To make each pair of similar samples closer in the output representation space, we introduce an individual-level contrastive loss:
[0066]
[0067] in, sim(u,v)=u T v represents the dot product similarity, and the denominator iterates through the division... and In addition to the remaining 2k-2 implicit negative samples;
[0068] In addition to alignment at the individual level, it is also desirable for the overall distribution of the entire batch of data to be highly consistent. Therefore, a Gaussian kernel function is used for MMD alignment, and the following distribution-level loss function is defined:
[0069]
[0070] Where χ denotes the reproducing kernel Hilbert space mapping, and the kernel function is chosen as γ(u,v) = exp(-||uv|| 2 / 2σ 2 Then, the total loss function compared to pre-training is expressed as:
[0071] L COM =L RUL +α·L IND +β·LDIS
[0072] By minimizing the total loss function, the network parameters of the backbone feature extraction module and the predictor can be updated in reverse, while the network parameters of the sub-branches can be updated using EMA momentum.
[0073] Compared to ordinary pre-training methods, this method can learn domain-invariant features more effectively, thereby further improving the generalization ability of the prediction model.
[0074] 2. Online cross-condition training based on adversarial learning
[0075] After offline pre-training is completed, in order to enable the prediction model to better adapt to the target working conditions, this invention updates the model online based on the idea of adversarial learning.
[0076] Assume that data x is currently collected from the equipment operating online under the target condition. t Accordingly, monitoring data for the corresponding time period is selected from the source domain devices, and the average is taken to obtain x. s The corresponding average remaining life label is y. s .
[0077] The pre-trained main branch feature extraction network and predictor network are respectively represented as F θ and F φ After inputting the corresponding source domain data into the feature extractor and predictor, the predicted remaining lifetime is obtained. The online supervised prediction loss is defined as follows:
[0078]
[0079] By inputting online target domain data into the feature extractor and predictor, an estimate of the remaining lifetime under the current operating condition can be obtained. The two RUL prediction values are then sent to the domain discriminator, resulting in the following form:
[0080]
[0081] Where, d t and d s D represents the output of the domain discriminator predicting the RUL values for the source and target domains, respectively. ψ The domain discriminator consists of hidden layers and fully connected layers with sigmoid activation functions; GRL stands for gradient inversion layer.
[0082] We define the following domain discriminator labels:
[0083]
[0084] The online adversarial loss of the output layer is defined as follows:
[0085]
[0086] Because the number of samples in each online training session is small, the model is more prone to overfitting to noise. Therefore, a regularization term is added, and the total loss function expression is as follows:
[0087]
[0088] Since the goal of the feature extractor and predictor is to maximize the classification error (making the source and target domains indistinguishable), while the goal of the domain discriminator is to minimize the classification error (making the source and target domains easily distinguishable), therefore, during forward propagation, let GRL(x) = x, and during backward propagation, let:
[0089]
[0090] Where λ adv This represents the gradient coefficient, which is always a positive number.
[0091] Through the above process, whenever a new batch of online monitoring data is collected, the parameters of the feature extraction network, predictor, and domain discriminator can be updated in reverse by minimizing the loss function, thereby achieving continuous optimization and domain adaptation of the online prediction model. This method enables the model to better transfer to the target operating conditions, significantly improving the accuracy of transfer prediction. Furthermore, this process requires no additional target domain data or labels, only source domain data and online-collected target domain data, which better reflects the actual operating conditions of degraded equipment.
[0092] Specific implementation process:
[0093] To verify the feasibility and effectiveness of the method of this invention, data from actual engines were used to predict the RUL curves of typical engines under 12 migration tasks, such as... Figure 2 and Figure 3 The figure shows the RUL curve prediction results of a typical engine under 12 migration tasks according to the present invention. Figure 2 a) is FD001→FD002; b) is FD001→FD003; c) is FD001→FD004; d) is FD002→FD001; e) is FD002→FD003; f) is FD002→FD004; Figure 3In the figure, g) represents FD003→FD001; h) represents FD003→FD002; i) represents FD003→FD004; j) represents FD004→FD001; k) represents FD004→FD002; and l) represents FD004→FD003. As can be seen from the figure, the accuracy is high in the early stages, indicating that contrastive learning effectively extracted shared features between samples with similar lifetimes during pre-training. The transition at the inflection point is smooth, with no significant lag or advance in the prediction curve. Even in the high-speed degradation stage near the end, the predicted line can still track the ground truth within a relatively small range of fluctuation, achieving smooth convergence. Even in transfer tasks with significantly different operating conditions, the error between the predicted lifetime and the actual lifetime is small, further demonstrating the advantages of the online domain adaptation strategy.
[0094] The above descriptions are merely embodiments of the present invention, and common technical solutions or characteristics known in the schemes are not described in detail here. For those skilled in the art, various modifications and improvements can be made without departing from the technical solutions of the present invention, and these should also be considered within the scope of protection of the present invention. These modifications and improvements will not affect the effectiveness of the implementation of the present invention or the practicality of the patent. The scope of protection claimed in this application shall be determined by the content of its claims, and the specific embodiments described in the specification can be used to interpret the content of the claims.
Claims
1. A method for predicting the remaining life of a degradation device migrating online field, characterized by, The prediction model comprises a pre-training part based on contrastive learning and an online cross-condition training part based on adversarial learning. The pre-training part based on contrastive learning comprises the following steps: A1. In the source domain data, select two groups of data similar in remaining useful life label to construct similar sample pairs; wherein k represents the number of sample pairs in the same training batch, and sharing the remaining useful life label y i ; A2. Use a momentum coding strategy, and the network parameters of the main branch are updated in real time through back propagation during training, and the network parameters of the other branch are not involved in gradient update, and the main branch and the sub-branch of the contrastive learning twin neural network are slowly followed by the main branch parameter change through exponential moving average, and the feature extraction and prediction process of the main branch and the sub-branch of the contrastive learning twin neural network is represented as follows: wherein, and respectively denote the prediction process of the main branch and the sub-branch, and respectively denote the extraction process of the main branch; The supervised prediction loss of the main branch is represented as follows: wherein y i represents the true residual lifetime of the ith sample; A3. Introducing individual-level contrastive loss makes each pair of similar samples closer in the output representation space, and the individual-level contrastive loss is represented as follows: wherein sim(u,v) = u T v denotes the dot product similarity, the denominator is traversed by and the remaining 2k-2 implicit negative samples; A4. In order to make the overall distribution of the whole batch of data highly coincide, a Gaussian kernel function is used for MMD alignment, and a distribution-level loss function is defined as follows: Where χ denotes the reproducing kernel Hilbert space mapping, and the kernel function is chosen as γ(u,v) = exp(-||uv|| 2 / 2σ 2 Then, the total loss function compared to pre-training is expressed as: L COM = L RUL + a · L IND + β · L DIS By minimizing the total loss function, the network parameters of the backbone feature extraction module and the predictor can be updated in reverse, and the network parameters of the sub-branch are updated using EMA momentum; The online cross-condition training part based on adversarial learning comprises the following steps: B1. Collecting data x from the target working condition of the online running equipment t , and selecting the monitoring data of the corresponding time period from the source domain equipment, and obtaining x after averaging s , the corresponding average remaining life label is y s ; the pre-trained main branch feature extraction network and the predictor network are represented as F θ and F φ , respectively. After inputting the corresponding source domain data into the feature extractor and the predictor, the predicted remaining useful life is obtained The online supervised prediction loss is defined as follows: B2, after the online target domain data is input into the feature extractor and the predictor, a remaining useful life estimation value under a current operating state can be obtained The two RUL prediction values are then sent to the domain discriminator to obtain the following form: wherein d t and d s respectively represent the output results of the domain discriminator predicting the RUL values of the source domain and the target domain, D ψ represents a domain discriminator composed of a hidden layer and a fully connected layer with a sigmoid activation function, and GRL is a gradient reversal layer. B3. Define the domain discriminator label as follows: The output layer online adversarial loss is defined as follows: B4. Since the number of samples for each online training is small, the model is more likely to overfit to noise, so a regularization term is added, and the total loss function expression is as follows: Since the goal of the feature extractor and the predictor is to maximize the classification error, and the goal of the domain discriminator is to minimize the classification error, during forward propagation, set GRL(x)=x, and during back propagation, set: where λ adv represents the gradient coefficient, which is always positive; Through the above process, whenever a new batch of online monitoring data is collected, the parameters of the feature extraction network, the predictor and the domain discriminator can be updated in reverse by minimizing the loss function, so as to realize the continuous optimization and field adaptation of the online prediction model.
2. The method of claim 1, wherein: In step A2, the predictor generally adopts a fully connected layer, and the feature extraction can be selected as needed.