System multi-degradation-stage residual life prediction method, system and equipment based on dynamic domain adaptive network, and storage medium

Abstract

The invention belongs to the technical field of residual life prediction, and particularly relates to a system multi-degradation-stage residual life prediction method and system based on a dynamic domain adaptive network, equipment and a storage medium. The method comprises the following steps: acquiring source domain monitoring data and target domain monitoring data of a system, and preprocessing the source domain monitoring data and the target domain monitoring data; extraction of multi-scale features and construction of health indexes; performing degradation stage division based on the IGTS; constructing a dynamic domain adaptive network containing the VBGRU and training the model; and inputting target domain monitoring data into the trained model, and outputting a residual life prediction value of the target domain system. In addition, the invention further provides a system, equipment and a storage medium for executing the method. The system not only can effectively extract the multi-scale time sequence dependence features in the system degradation process, but also can adaptively compare complete local and local features between domains, and effectively improves the generalization ability of residual life prediction in a variable working condition scene.

Description

Technical Field

[0001] This invention belongs to the field of remaining lifetime prediction technology, specifically relating to a method, system, device and storage medium for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, and particularly to a method and system for predicting the remaining lifetime under different operating conditions based on deep transfer learning technology. Background Technology

[0002] As industrial equipment becomes increasingly complex, integrated, and intelligent, prognostics and health management (PHM) technologies are becoming increasingly important. Remaining Useful Life (RUL) prediction is one of the key technologies in PHM. Accurate RUL prediction can effectively assess the health status of a system, providing valuable suggestions for maintenance decisions, thereby reducing downtime and economic losses, and improving machine reliability.

[0003] Existing RUL prediction methods are mainly divided into model-based methods and data-driven methods. Because complex systems exhibit diverse failure modes and intricate mechanisms, making it difficult to establish accurate physical models, data-driven methods that do not rely on prior knowledge, especially deep learning-based methods such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory Networks (LSTMs), have received widespread attention in the field of RUL prediction due to their powerful feature extraction and nonlinear mapping capabilities.

[0004] However, in real-world industrial scenarios, deep learning-based RUL prediction methods still face the following significant challenges: the performance of deep learning models typically relies on the assumption that training and test data follow the same distribution. But in practical applications, equipment often operates under different conditions, such as varying loads, speeds, and ambient temperatures, leading to significant differences in the feature distributions of the source and target domains. This distributional difference causes a substantial decrease in the model's generalization ability when predicting across different operating conditions.

[0005] To address the issue of distributional discrepancies, transfer learning, particularly domain adaptation (DA) techniques, has been widely applied. However, it also has the following drawbacks:

[0006] First, there are limitations to global alignment. Existing global domain adaptive methods typically attempt to align the overall distribution of the source and target domains. However, system performance degradation during operation exhibits multi-stage characteristics, such as a healthy stage, early degradation, and severe degradation. Using only global alignment can easily obscure fine-grained features between subdomains at different degradation stages, leading to insufficient prediction accuracy.

[0007] Second, the shortcomings of local alignment: Although some subdomain adaptive methods attempt to align similar subdomains, these methods often ignore the importance of global distribution alignment, resulting in the neglect of the overall distribution differences between domains.

[0008] Third, there is a lack of dynamic balance; existing methods struggle to find a dynamic balance between global and local alignment, and cannot adaptively adjust the importance weights of both according to the training process.

[0009] Furthermore, when processing multi-sensor monitoring data, existing feature extraction methods struggle to collaboratively extract multi-scale temporal features and spatial correlation features. In particular, existing network structures often fail to adequately capture local degradation features between adjacent time points, limiting the model's ability to perceive subtle degradation trends.

[0010] Therefore, how to effectively extract multi-scale local enhancement features, divide the degradation process into multiple stages, and then achieve high-precision RUL prediction across different working conditions by dynamically adjusting the alignment weights of the global and local systems is a technical problem that urgently needs to be solved. Summary of the Invention

[0011] Therefore, the first objective of this invention is to provide a method and system for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, in order to solve the problem that existing transfer learning global domain adaptive methods are prone to confusing fine-grained features between subdomains of different degradation stages in system RUL prediction; and to overcome the technical problem that it is difficult to collaboratively extract multi-scale temporal features and spatial correlation features in the local feature extraction process of multi-sensor data, resulting in insufficient RUL prediction accuracy across different operating conditions.

[0012] A second objective of this invention is to provide an electronic device and a computer-readable storage medium based on the above-described remaining lifetime prediction method and system.

[0013] In a first aspect, the present invention provides a method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, comprising the following steps:

[0014] S1. Acquire the source domain monitoring data and target domain monitoring data of the system, and perform preprocessing; the source domain monitoring data has a remaining lifetime label, and the target domain monitoring data does not have a label;

[0015] S2. Construct a multi-scale feature extraction module. Input the preprocessed data into the multi-scale feature extraction module. Extract multi-scale temporal degradation features through multiple parallel convolutional branches. Use the global channel attention mechanism to adaptively fuse the extracted multi-scale temporal degradation features to obtain health indicator features that reflect the degradation state of the system.

[0016] S3. Construct a degradation stage segmentation module, and use the information gain-based time segmentation method to segment the health indicator features of the source domain to obtain multiple degradation stages and their corresponding stage labels.

[0017] S4. Construct a dynamic domain adaptive network, which includes a feature extractor, a subdomain classifier, and a remaining lifetime predictor; the feature extractor adopts a variational autoencoder-bidirectional gated recurrent unit network to characterize the randomness and uncertainty of degradation data and to extract degradation feature data of long sequences.

[0018] S5. Train the dynamic domain adaptive network using source domain data and target domain data; during training, use a dynamic weight adjustment mechanism to adaptively adjust the weights of global domain distribution alignment and local subdomain distribution alignment, and update the network parameters by minimizing the total loss function; the total loss function includes variational autoencoder-bidirectional gated recurrent unit network loss, subdomain classification loss, global domain adaptation loss and local subdomain adaptation loss.

[0019] S6. Input the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network, and output the remaining lifetime prediction value of the target domain system.

[0020] In an optional embodiment of the present invention, the multi-scale feature extraction module in step S2 includes three parallel temporal convolution branches, each of which contains multiple serial TCN residual blocks; the TCN residual blocks include dilated causal convolutional layers, weight normalization layers, ReLU activation layers and Dropout layers, and use direct mapping of the input layer as residual connections.

[0021] The three temporal convolutional branches are configured with convolutional kernels of sizes 3, 5, and 7, respectively. The receptive field is then expanded by dilated causal convolutional layers to extract short, medium, and long-distance temporal dependencies.

[0022] The multi-scale feature extraction module introduces a multi-scale local feature enhancement mechanism in each temporal convolutional branch. It uses the output of the TCN residual block of the previous layer in each temporal convolutional branch as the input of the next layer, and accumulates the outputs of all TCN residual blocks in each temporal convolutional branch to capture the relevant local features between adjacent time points.

[0023] The global channel attention mechanism uses global average pooling and global max pooling to extract global degradation information of different sensor channels. Then, it generates the weights of attention for each channel through a Sigmoid activation layer, and performs weighted fusion on the extracted multi-scale temporal degradation features to finally form health indicator features that reflect the degradation state of the system.

[0024] In an optional embodiment of the present invention, step S3 employs a time segmentation method based on information gain, specifically including:

[0025] Define a cost function based on the information gain and segment length resulting from segmenting a given time series;

[0026] The objective of maximizing information gain is transformed into minimizing weighted entropy.

[0027] The optimal split point is solved by using dynamic programming algorithm, and the entire degradation process of the source domain is discretized into multiple degradation stages and their corresponding stage labels, which serve as the basis for subdomain partitioning.

[0028] In an optional embodiment of the present invention, the variational autoencoder-bidirectional gated cyclic unit network in step S4 includes:

[0029] A bidirectional gated recurrent unit layer is used to capture the bidirectional temporal degradation patterns of the input feature data, enabling the extraction of long-sequence degradation feature data;

[0030] The variational autoencoder layer is used to map the input feature data to the mean and standard deviation distribution of a low-dimensional latent space, and to generate latent features through reparameterization techniques to characterize the randomness and uncertainty of degraded data.

[0031] Both the subdomain classifier and the remaining lifetime predictor consist of fully connected layers; the subdomain classifier contains three fully connected layers with 128, 64 and 10 neurons respectively; the remaining lifetime predictor contains three fully connected layers with 128, 64 and 1 neurons respectively; and a Dropout layer is applied between every two fully connected layers with a Dropout rate of 0.2.

[0032] In one optional embodiment of the present invention, the training process in step S5 specifically includes:

[0033] The source domain data and target domain data are input into the multi-scale feature extraction module to obtain health indicator features, and the source domain health indicator features are divided into multiple degradation stage source domain health indicator features.

[0034] By inputting the source domain health index features and target domain health index features of multiple degradation stages into the variational autoencoder-bidirectional gated recurrent unit network, the potential feature distribution of the corresponding domain is obtained.

[0035] The latent features of the corresponding domain are input into the subdomain classifier; for source domain data, the cross-entropy loss between its predicted class and the true stage label is calculated as the subdomain classification loss; for target domain data, pseudo-labels are generated using the output of the subdomain classifier, and the pseudo-labels are used for subsequent local subdomain distribution alignment.

[0036] The latent features of the source domain data are input into the remaining lifetime predictor, and the mean square error between the predicted value and the actual remaining lifetime label is calculated as the remaining lifetime prediction loss.

[0037] Meanwhile, the dynamic weight adjustment mechanism in step S5 is specifically as follows:

[0038] Calculate global feature alignment metric Alignment metrics with local features ;

[0039] Calculate the adaptive coefficient according to the formula. ;

[0040]

[0041] Using the adaptive coefficients The dynamic domain adaptive loss is calculated by weighted summation of the global domain adaptation loss and the local subdomain adaptation loss. ;

[0042]

[0043] Specifically, when the global feature differences are small, the weight of local feature differences is increased; when the local feature differences are small, the weight of global feature differences is increased.

[0044] The global domain adaptation loss is measured using the maximum mean difference (MMD), and the local subdomain adaptation loss is measured using the local maximum mean difference (LMMD).

[0045] When calculating the Local Maximum Mean Difference (LMMD), for source domain samples, the stage labels obtained in step S3 are used to calculate the weights; for target domain samples, the pseudo labels generated by the subdomain classifier in step S5 are used to calculate the weights.

[0046] Furthermore, the total loss function Represented as:

[0047]

[0048] in, For the variational autoencoder-bidirectional gated cyclic unit network loss, For dynamic domain adaptive loss, Subdomain classification loss; and For the balance coefficient, the variational autoencoder-bidirectional gated recurrent unit network loss is expressed as the sum of the remaining lifetime prediction loss and the latent space regularization constraint term solved through variational inference; The value is 0.1. The value is 0.3.

[0049] In an optional embodiment of the present invention, the condition for the training to end in step S5 is:

[0050] Set a fixed number of training rounds of 50;

[0051] The Adam optimizer was used during training, with the initial learning rate set to 0.005 and the learning rate scheduler used to dynamically adjust the learning rate.

[0052] When the value of the total loss function stabilizes and stops decreasing, or when the set number of training rounds is reached, training is stopped, and the network parameters with the minimum loss value at this point are taken as the trained model parameters.

[0053] In an optional embodiment of the present invention, the specific process of inputting the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network, and outputting the remaining lifetime prediction value of the target domain system in step S6 includes:

[0054] The target domain monitoring data is input into the trained multi-scale feature extraction module, and the target domain features are extracted using the convolution kernels and attention weights with fixed parameters.

[0055] The target domain features are input into the feature extractor in the trained dynamic domain adaptive network to obtain latent features.

[0056] The latent features are input into the trained remaining lifetime predictor, and the weighted summation and activation function mapping of the fully connected layer directly outputs the remaining lifetime prediction value of the target domain system.

[0057] Secondly, this invention provides a system for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, comprising:

[0058] The data acquisition module is used to acquire source domain monitoring data and target domain monitoring data of the system;

[0059] The multi-scale feature extraction module is used to extract and fuse multi-scale local enhancement features through parallel temporal convolutional branches and global channel attention mechanisms to construct health indicators.

[0060] The degradation stage segmentation module is used to divide source domain data into multiple degradation stages based on the information gain time segmentation method;

[0061] The dynamic domain adaptive prediction module includes a variational autoencoder-bidirectional gated recurrent unit network, a subdomain classifier, and a remaining lifetime predictor; it is used to perform transfer learning training by dynamically adjusting global and local alignment weights and output the remaining lifetime prediction results.

[0062] Thirdly, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the aforementioned method for predicting the remaining lifetime of a system based on a dynamic domain adaptive network in multiple degradation stages.

[0063] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for predicting the remaining lifetime of a system based on a dynamic domain adaptive network in multiple degradation stages.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] (1) This invention addresses the problems of insufficient extraction of local degradation features and weak multi-scale information fusion capability in multi-sensor time-series data by proposing a method for constructing health indicators with multi-scale local feature enhancement. This method effectively extracts multi-scale temporal dependent features in the system degradation process through multi-scale temporal convolution structure and local feature enhancement mechanism, providing a feature representation basis for subsequent remaining lifetime prediction.

[0066] (2) The degradation features are further divided into multiple subdomains according to the degradation stage division module. A variational autoencoder-bidirectional gated recurrent unit network is proposed. It can capture the bidirectional degradation law in the working condition fluctuation through the bidirectional gated recurrent unit, and use the variational autoencoder to compress the degradation data to a low-dimensional potential space to enhance the prediction robustness under complex working conditions.

[0067] (3) The remaining lifetime prediction method based on dynamic domain adaptive network was evaluated under cross-domain conditions using an aircraft turbofan engine dataset. For the proposed dynamic domain adaptive network, the method flexibly adjusts the importance weights of global and local distributions to adaptively align global and local features between domains, achieving global alignment of degraded features and local alignment of subdomain features. Attached Figure Description

[0068] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0069] Figure 1 This is a schematic diagram of the overall process of the prediction method involved in this invention.

[0070] Figure 2 This is a schematic diagram of the TCN residual block involved in the present invention.

[0071] Figure 3 This is a schematic diagram of the multi-scale feature extraction module involved in the present invention.

[0072] Figure 4 This is a schematic diagram of the DDAN principle involved in this invention.

[0073] Figure 5 This invention relates to a health index map constructed using the FD001 No. 5 training set engine.

[0074] Figure 6 This invention relates to a comparative chart of the assessment capabilities of a comprehensive evaluation index model for health indicators constructed using the FD001 dataset.

[0075] Figure 7 This is a comparison chart of the RUL prediction results of two models using OC2-OC3 as an example task in this invention.

[0076] Figure 8 This is a comparison chart of the results of 12 cross-domain RUL predictions using a random engine in the target domain test set, as per the present invention.

[0077] Figure 9 This is a visualization feature distribution map of t-SNE involved in the present invention. Detailed Implementation

[0078] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0079] Deep transfer learning technology has been successfully applied to predict remaining lifetime under different operating conditions. However, due to the multi-stage degradation of system performance during operation, existing global domain adaptive transfer learning methods are prone to confounding fine-grained features between subdomains of different degradation stages in system RUL prediction; and in the process of local feature extraction of multi-sensor channel data, it is difficult to collaboratively extract multi-scale temporal features and spatial correlation features, resulting in insufficient RUL prediction accuracy across operating conditions.

[0080] Therefore, this invention proposes a method for predicting the remaining lifetime of a system across multiple degradation stages using a dynamic domain adaptive network. First, a multi-scale feature extraction module is constructed for feature extraction. While effectively extracting time-series information, a multi-scale local feature enhancement mechanism is further used to capture local degradation features of the system, enhancing the model's multi-scale feature extraction capability. Second, a degradation stage segmentation module is introduced to divide the features into stages. Third, the proposed dynamic domain adaptive network is used for migration prediction of the system under varying operating conditions, aligning data features at different degradation stages. This method achieves both global alignment of degradation features and local alignment of fine-grained features between subdomains.

[0081] Based on the technical approach provided above, such as Figure 1 As shown, the method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network provided by this invention specifically includes the following steps:

[0082] S1. Acquire the source domain monitoring data and target domain monitoring data of the system, and perform preprocessing; the source domain monitoring data has a remaining lifetime label, and the target domain monitoring data does not have a label;

[0083] S2. Construct a multi-scale feature extraction module. Input the preprocessed data into the multi-scale feature extraction module. Extract multi-scale temporal degradation features through multiple parallel convolutional branches. Use the global channel attention mechanism to adaptively fuse the extracted multi-scale temporal degradation features to obtain health indicator features that reflect the degradation state of the system.

[0084] S3. Construct a degradation stage segmentation module, and use the information gain-based time segmentation method to segment the health indicator features of the source domain to obtain multiple degradation stages and their corresponding stage labels.

[0085] S4. Construct a dynamic domain adaptive network, which includes a feature extractor, a subdomain classifier, and a remaining lifetime predictor; the feature extractor adopts a variational autoencoder-bidirectional gated recurrent unit network to characterize the randomness and uncertainty of degradation data and to extract degradation feature data of long sequences.

[0086] S5. Train the dynamic domain adaptive network using source domain data and target domain data; during training, use a dynamic weight adjustment mechanism to adaptively adjust the weights of global domain distribution alignment and local subdomain distribution alignment, and update the network parameters by minimizing the total loss function; the total loss function includes variational autoencoder-bidirectional gated recurrent unit network loss, subdomain classification loss, global domain adaptation loss and local subdomain adaptation loss.

[0087] S6. Input the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network, and output the remaining lifetime prediction value of the target domain system.

[0088] In step S1, it is necessary to acquire and preprocess the source domain monitoring data and target domain monitoring data of the system. The source domain monitoring data and target domain monitoring data can be acquired through multi-channel sensors. For example, the acquired monitoring data is the filtered operating data of an aircraft turbofan engine under different operating conditions and fault types. The data preprocessing involves normalizing the multi-channel sensor monitoring data and then using a sliding window method to extract time feature segments.

[0089] In step S2, the multi-scale feature extraction module synchronously extracts temporal degradation features through a parallel branch temporal convolutional network and adaptively fuses key degradation features from multi-sensor data using a global channel attention mechanism, thereby significantly improving the accuracy of remaining lifetime prediction. When constructing the multi-scale feature extraction module, a shallow feature extraction unit is first built, followed by a deep feature extraction unit. The shallow feature extraction unit mainly includes a dilated causal convolutional layer, a ReLU activation layer, and a Dropout layer. The deep feature extraction unit mainly includes three parallel temporal convolutional branches, each containing multiple serial TCN residual blocks. It should be noted that the TCN residual blocks include a dilated causal convolutional layer, a weight normalization layer, a ReLU activation layer, and a Dropout layer, and use a direct mapping of the input layer as the residual connection. For example, each temporal convolutional branch contains three serial TCN residual blocks.

[0090] The structure of the TCN residual block is as follows: Figure 2 As shown in the figure. The weight normalization layer and the Dropout layer are mainly used to suppress network noise and optimize the network training effect. In addition, a 1×1 convolutional block is introduced to ensure that the input and output dimensions are the same.

[0091] The multi-scale feature extraction module configures different layers using two key parameters: different convolutional kernel sizes and dilation factors. This expands the receptive field of the convolutional kernels without significantly increasing computation, thus more effectively capturing long-range dependencies in long sequence data. For this purpose, as... Figure 3 As shown, the multi-scale feature extraction module is designed with three parallel temporal convolution branches, using three convolution kernels of different sizes to extract features from the input data. The sizes of the three convolution kernels are set to 3, 5, and 7, respectively. Then, the receptive field is expanded through dilated causal convolutional layers to extract short, medium, and long-distance temporal dependencies.

[0092] Furthermore, the multi-scale feature extraction module introduces a multi-scale local feature enhancement mechanism in each temporal convolutional branch to capture relevant local features between adjacent time points. Specifically, the output of the TCN residual block from the previous layer in each temporal convolutional branch is used as the input to the next layer, and the outputs of all TCN residual blocks in each temporal convolutional branch are accumulated. Convolutional block layers at different times extract features at different time scales. By using a global channel attention mechanism to weight and fuse the features extracted at different scales, the entire RUL prediction model can obtain a more representative feature representation, thereby better distinguishing different operating states and degradation stages of the system and improving the accuracy of remaining lifetime prediction. Based on this, the accumulated outputs of all TCN residual blocks in each temporal convolutional branch, superimposed with the input layer data, become the final output of that branch.

[0093] like Figure 3 The diagram shows the structure of the multi-scale feature extraction module, where FC is a fully connected layer and k is the size of the convolutional kernel. Each temporal convolutional branch uses a temporal convolutional block (TCNB) to extract multi-scale local information layer by layer. After temporal feature extraction, the dual-channel global attention module AM ​​extracts degraded global information from different sensor channels using global average pooling and global max pooling, respectively. Then, a one-dimensional convolutional layer and a ReLU activation layer are used to adjust the feature dimension and increase the nonlinear representation of the model. Finally, a sigmoid activation layer is used to obtain the attention weights for each channel, and the weights are multiplied by the input data to obtain the corresponding output. The AM module can dynamically allocate weights for different sensor channels, strengthen feature information sensitive to degradation, suppress noise interference, and achieve adaptive fusion of key spatial information. This multi-scale feature extraction module superimposes multi-scale temporal features and attention-weighted spatial features layer by layer through residual connections, ultimately forming a composite health indicator. The dual-channel global attention module AM ​​is the global channel attention mechanism.

[0094] It should be noted that the dilation factors d=1, 2 and 4 are set in both the dilated causal convolutional layer in the shallow feature extraction unit and the dilated causal convolutional layer in the TCN residual block.

[0095] In step S3, considering the inherent similarity between different stages in the system degradation process and the infinity of RUL labels, this invention proposes a degradation stage segmentation module for RUL prediction, which discretizes the entire degradation process of the source domain into multiple degradation stages. For example, the IGTS method, i.e., the information gain time segmentation method, is used to achieve stage segmentation. This method finds key time points that reflect the health status of the system by quantifying the information entropy changes of features. This method takes the feature vector space that reflects the degradation trend of the system as input to obtain the system health status segmentation under unsupervised monitoring. For the feature set of the time series, under the constraint of the number of segmentations, a cost function is constructed by introducing the concept of information gain. Information gain is the expected reduction in entropy generated by segmenting a given time series. The higher the gain value, the more significant the difference in degradation mode before and after the point, and the more suitable it is as a stage boundary. Its cost function is defined as shown in formula (1):

[0096]

[0097] in, Indicates information gain. Indicates the number of segments. Indicates the first The length of the segment Indicates the length of the entire time series. This represents the entropy of the entire time series. Indicates the first The entropy of a segment.

[0098] For the feature set of time series This feature set Contains A feature sequence, Represented as the first ... ; This represents the length of each sequence. Therefore, in formula (1) Indicates the first Entropy of a segment Indicates the first The first segment The probability of a feature sequence. Indicates the number of feature sequences. For the first The time division point of the segment, Indicates the first The first feature sequence The value of each data point. Indicates the first The first feature sequence The value of each data point.

[0099] To find the optimal split point within the dynamic programming framework, the compatibility issue between information gain and traditional dynamic programming needs to be addressed, specifically the problem that the non-separability of the original cost function prevents direct recursive optimization. To address this, an adaptive transformation is performed on the cost function, converting the maximization of information gain into minimizing the weighted entropy objective, as shown in equation (2), thus satisfying the decomposability requirement of dynamic programming.

[0100]

[0101] Finally, by solving for the optimal split point using the IGTS algorithm with dynamic programming, the entire degradation time series of the source domain can be discretized into multiple degradation stages and their corresponding stage labels, which serve as the basis for subdomain partitioning. Each stage corresponds to a specific degradation state of the system from health to failure. This partitioning result not only solves the problem of infinite RUL labels but also provides multi-stage input data for subsequent staged prediction models, thereby improving the accuracy of RUL prediction.

[0102] In step S4, a dynamic domain adaptive network needs to be constructed. During this construction process, a feature extractor must first be built, followed by a subdomain classifier and a remaining lifetime predictor.

[0103] For multi-stage degradation time-series data that has been segmented, RUL prediction needs to consider how to better capture the long-term dependencies of degradation features within each stage and represent and interpret the randomness and uncertainty of the degradation data. To this end, this invention proposes a VBGRU network that integrates the interpretability of a variational autoencoder (VAE) with the temporal modeling advantages of a bidirectional gated recurrent unit (Bi-GRU). This network achieves the representation of staged degradation states and the interpretability of the data, providing reliable input for subsequent RUL prediction. Therefore, constructing a feature extractor is equivalent to constructing a VBGRU network. Specifically, in this embodiment, see [link to relevant documentation]. Figure 4 The VBGRU network comprises: a Bi-GRU layer with bidirectional gated recurrent units (Bi-GRU) to capture the bidirectional temporal degradation patterns of the input feature data, enabling the extraction of long-sequence degraded feature data; and a Variational Autoencoder (VAE) layer to map the input feature data to the mean and standard deviation distributions of a low-dimensional latent space, and to generate latent features through reparameterization techniques to characterize the randomness and uncertainty of the degraded data. This invention employs the VBGRU network to extract long-sequence features, overcoming the gradient vanishing and exploding problems inherent in traditional recurrent neural networks during feature extraction.

[0104] It should be noted that the reparameterization technique is implemented through the following formula.

[0105]

[0106] In the formula, As a potential feature, The mean of the low-dimensional latent space. The standard deviation of the low-dimensional latent space. For high-dimensional input features, To obey The standard normal distribution This is the set of weights and bias parameters for VBGRU.

[0107] The VBGRU network can use latent variables to describe the distribution of the input data. In this process, the loss function of the VBGRU network can be expressed as the true remaining lifetime value. Compared with the predicted value The sum of the mean square error term and the latent space regularization constraint term obtained through variational inference is shown in Equation (4):

[0108]

[0109] In the formula, This represents the number of samples used during the training process.

[0110] Next, a subdomain classifier and a remaining lifespan predictor are constructed. For example, both the subdomain classifier and the remaining lifespan predictor consist of fully connected layers; the subdomain classifier contains three fully connected layers with 128, 64, and 10 neurons respectively; the remaining lifespan predictor contains three fully connected layers with 128, 64, and 1 neurons respectively; and a Dropout layer is applied between every two fully connected layers with a Dropout rate of 0.2.

[0111] Furthermore, since complex systems exhibit different data representation distributions at different degradation stages, relying solely on global domain adaptation may obscure fine-grained features across subdomains of different degradation stages. Moreover, dynamically balancing the importance of each distribution type is crucial for feature alignment when aligning global and local distributions. Therefore, this invention also introduces a dynamic weight adjustment mechanism into the dynamic domain adaptive network that captures fine-grained information and flexibly adjusts the global and local weight distributions for RUL prediction of complex systems.

[0112] A framework for constructing the RUL prediction model is established using the VBGRU network and a dynamic weight adjustment mechanism. This framework is named Dynamic Domain Adaptive Network (DDAN), meaning DDAN is equivalent to the RUL prediction model. Its structure is as follows: Figure 4 As shown. First, the source domain data with stage labels generated by the degradation stage segmentation module and the target domain monitoring data are used as input to the model. Second, the mean values ​​of the source and target domains obtained from the output of the VBGRU network are used. and standard deviation Then, the feature distribution difference is minimized through a dynamic weight adjustment mechanism. Specifically, in this embodiment, the specific training process of the RUL prediction model is as follows: the source domain data and target domain data are input into the multi-scale feature extraction module to obtain health indicator features, and the source domain health indicator features are divided into multi-degradation stage source domain health indicator features; the multi-degradation stage source domain health indicator features and target domain health indicator features are input into the VBGRU network to obtain the latent feature distribution of the corresponding domain; the latent features of the corresponding domain are input into the subdomain classifier, and for the source domain data, the cross-entropy loss between its predicted category and the true stage label is calculated as the subdomain classification loss; for the target domain data, pseudo-labels are generated using the output of the subdomain classifier, and the pseudo-labels are used for subsequent local subdomain distribution alignment; the latent features of the source domain data are input into the remaining life predictor, and the mean square error between the predicted value and the true remaining life label is calculated as the remaining life prediction loss.

[0113] The dynamic weight adjustment mechanism will be described in detail below.

[0114] In cross-condition RUL prediction, the difference in feature distribution between domains is a core issue leading to the degradation of model generalization performance. The differences in data distribution across different domains make aligning extracted degenerate features using only global or local distributions challenging. Furthermore, dynamically balancing the importance of each distribution type is crucial for feature alignment when aligning global and local distributions. To address this issue, this invention proposes a dynamic weight adjustment mechanism. This mechanism adaptively aligns global and local features across domains by flexibly adjusting the importance weights of global and local distributions. Its design goal is to achieve the mean of the staged probability features output by the VBGRU network. and standard deviation As input, global alignment ensures overall consistency between domains, subdomain alignment captures fine-grained differences in each degradation stage, and dynamic weight coefficients are introduced to balance the importance of both, ultimately achieving feature alignment.

[0115] Next, we will first explain the method for global domain adaptive distribution alignment and then calculate the global feature alignment metric. .

[0116] Domain adaptation encourages feature extractors to map samples from different domains to the same hidden representation space. Given a source domain... Target domain ,in and They represent the first The feature variables and corresponding RUL labels of each source domain sample. Indicates the first Feature variables of a target domain sample and represents the total number of samples in the source domain and the target domain, respectively. Global domain adaptive distribution alignment methods typically use the maximum mean difference (MMD) to measure the difference between the data distributions in the source domain and the target domain, as shown in formula (5).

[0117]

[0118] In the formula, Represents the regenerating nucleus Hilbert space. and Representing the source domain respectively and target domain Characteristic variables of the sample This represents the transformation from the initial data to the regenerating kernel Hilbert space. This indicates that the random variable x follows a p-distribution; This indicates that the pair follows the source domain distribution. samples After feature mapping Calculate the expected value of the result. This indicates that the distribution follows the target domain distribution. samples After feature mapping Calculate the expected value of the result.

[0119] Secondly, the method for adaptive distribution alignment of local subdomains is explained, and the local feature alignment metric is calculated. .

[0120] While global alignment can reduce overall discrepancies, it may obscure fine-grained features between subdomains represented by different degradation stages. Local subdomain adaptive distribution alignment aims to bring samples belonging to corresponding subdomains closer together in the mapped latent space. and Divided into C subdomains and Where c represents the category label and Then, Local Maximum Mean Difference (LMMD) is introduced, which measures the Hilbert-Schmidt norm between the kernel-averaged embeddings of the empirical distributions of relevant subdomains in the source and target domains, taking into account the weights of different samples. To align the relevant subdomains, LMMD is expressed as follows:

[0121]

[0122] In the formula, and They are and The distribution, This represents the input with a characteristic kernel and the function mapping to the Hilbert space of the reproducing kernel. Indicates category label The expectation, that is, in all categories Weighted average on, This indicates that the random variable x follows a set pattern. distributed; This indicates that the pair follows the source domain distribution. samples After feature mapping Calculate the expected value of the result. This indicates that the distribution follows the target domain distribution. samples After feature mapping Calculate the expected value of the result.

[0123] LMMD can measure the expected value of local distributional variance. It assumes each sample is weighted... For each category, there exists an unbiased estimate:

[0124]

[0125] In the formula, and They represent the categories respectively. of and The weights. Among them, and The following conditions must be met:

[0126]

[0127] Based on this, for source domain samples / target domain samples Real UL Labels / Fake Labels Its weight The expression is shown in formula (8):

[0128]

[0129] in, It is a label vector The Each element.

[0130] For labeled samples in the source domain The weights are calculated using one-hot encoding of labels obtained by dividing the degradation data generated by the multi-scale feature extraction module through the IGTS method into degradation stages. For unlabeled samples in the target domain Using pseudo-labels to calculate .

[0131] Since changes in global and local distributions affect the alignment of degenerate feature distributions differently, and considering the evolving distributional differences between source and target features throughout iterative training, the ratio of these two distributional distances must be dynamically adjusted. To better balance the importance of global and local feature differences at each training stage, adaptive coefficients are used. The weights of global and local features are dynamically adjusted using the following formula:

[0132]

[0133] Using the adaptive coefficients The dynamic domain adaptive loss is calculated by weighted summation of the global domain adaptation loss and the local subdomain adaptation loss. ;

[0134]

[0135] when When the difference is close to 0, the global feature difference is minimal, and more emphasis should be placed on the difference in local features; when When the value is close to 1, the local feature difference is minimal, and more attention should be paid to the global feature difference. Therefore, the dynamic weight adjustment mechanism can dynamically adjust the weights of global and local feature differences according to the magnitude of the differences, ensuring the representation of cross-domain degenerate features.

[0136] Furthermore, considering the mean distribution obtained from the VBGRU network and standard deviation distribution It can well describe the different stages of the system degradation process, with special feature representation and good interpretability. Therefore, by using the mean and standard deviation as inputs to formula (10), domain adaptation for different degradation stages can be well achieved.

[0137] Finally, the RUL prediction model of this invention, during training, first pre-trains a subdomain classifier using source domain data. Then, the target domain data is passed through the subdomain classifier to obtain subdomain pseudo-labels. Through iterative training, the accuracy of the subdomain classifier gradually improves, thereby obtaining accurate subdomain labels. For all samples in the source domain... Subdomain classification loss Defined as:

[0138]

[0139] in, The labels obtained by classifying the HI health indicator data into stages of degradation using the IGTS method are used as the true labels. Data features extracted for the model For classifiers, It is the cross-entropy loss function. Therefore, according to formulas (4)-(11), The total loss function of the entire prediction model It can be defined as:

[0140]

[0141] in, For VBGRU network loss, For dynamic domain adaptive loss, Subdomain classification loss; and For the balance coefficient, the VBGRU network loss is expressed as the sum of the remaining lifetime prediction loss and the latent space regularization constraint term solved through variational inference; The value is 0.1. The value is 0.3.

[0142] In step S5, detailed parameters for the end of training are set for the RUL prediction model involved in this invention, as follows: the fixed number of training rounds is set to 50; the Adam optimizer is used during training, the initial learning rate is set to 0.005, and the learning rate is dynamically adjusted using the learning rate scheduler; when the value of the total loss function tends to stabilize and no longer decreases, or when the set number of training rounds is reached, training is stopped, and the network parameters with the minimum loss value at this time are taken as the trained model parameters.

[0143] In step S6, after the entire prediction model has been trained, the specific process of inputting the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network, and outputting the remaining lifetime prediction value of the target domain system includes: inputting the target domain monitoring data into the trained multi-scale feature extraction module, and extracting target domain features using convolutional kernels and attention weights with fixed parameters; inputting the target domain features into the feature extractor in the trained dynamic domain adaptive network to obtain latent features; inputting the latent features into the trained remaining lifetime predictor, and directly outputting the remaining lifetime prediction value of the target domain system through weighted summation and activation function mapping of fully connected layers.

[0144] Based on this, the present invention also provides a system for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, comprising: a data acquisition module for acquiring source domain monitoring data and target domain monitoring data of the system; a multi-scale feature extraction module for extracting and fusing multi-scale local enhancement features through parallel temporal convolutional branches and global channel attention mechanisms to construct health indicators; a degradation stage segmentation module for dividing the source domain data into multiple degradation stages based on the information gain temporal segmentation method; and a dynamic domain adaptive prediction module comprising a variational autoencoder-bidirectional gated recurrent unit network, a subdomain classifier, and a remaining lifetime predictor; used for transfer learning training by dynamically adjusting global and local alignment weights and outputting the remaining lifetime prediction results.

[0145] The RUL migration prediction system for systems under varying operating conditions proposed in this invention consists of four modules: a data acquisition module, a multi-scale feature extraction module, a degradation stage segmentation module, and a dynamic domain adaptive prediction module. The dynamic domain adaptive prediction module further comprises a VBGRU network, a subdomain classifier, and a remaining lifetime predictor. In this system, the multi-scale feature extraction module fuses multi-sensor features to construct a health index; the degradation stage segmentation module uses information gain time segmentation to segment degradation data into multiple stages; the dynamic domain adaptive prediction module aligns global and local data features for different degradation stages; and the subdomain classifier and remaining lifetime predictor perform RUL subdomain stage classification and remaining lifetime prediction tasks. First, preprocessed sensor data is fed into the multi-scale feature extraction module to construct a health index HI containing degradation features. The extracted source domain features are then transmitted to the VBGRU network, subdomain classifier, and remaining lifetime predictor in the degradation stage segmentation module and the dynamic domain adaptive prediction module. Then, the VBGRU network loss and the subdomain classification loss are calculated respectively. Simultaneously, the extracted source and target domain features are combined in the dynamic domain adaptive prediction module to calculate the dynamic domain adaptive loss and update the proposed model parameters. Finally, the trained multi-scale feature extraction module and the trained remaining lifetime predictor are combined to predict the test data in the target domain.

[0146] Furthermore, the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the aforementioned method for predicting the remaining lifetime of a system based on a dynamic domain adaptive network across multiple degradation stages. Simultaneously, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the aforementioned method for predicting the remaining lifetime of a system based on a dynamic domain adaptive network across multiple degradation stages.

[0147] To further verify the predictive performance of this method, this invention uses the C-MAPSS dataset released by NASA and performs RUL predictions under different operating conditions, comparing the prediction method proposed in this invention with several traditional and cutting-edge RUL prediction methods.

[0148] Experimental data description: The C-MAPSS dataset consists of four sub-datasets: FD001, FD002, FD003, and FD004. Simulations were conducted under different combinations of operating conditions and fault types, such as... As shown:

[0149] Table 1 Description of the aircraft turbofan engine dataset

[0150]

[0151] In Table 1, FD001 has relatively simple operating conditions and fault types, while FD002 and FD004 have more operating conditions, and FD003 and FD004 have more fault types. The training set and test set are from different turbofan engines. The training set records the sampled values ​​of each time series within the complete cycle from the start of engine operation to the occurrence of a fault, including engine degradation data; while the test set records the state parameter values ​​at a certain point in time before the fault, including engine failure data. The purpose of this experiment is to predict the remaining service life of the engines in the test set.

[0152] As shown in Table 1, the four subsets have different conditions. The research objective of this invention is to train the model using only source domain data labeled under certain operating conditions and target domain data unlabeled under another operating condition. Furthermore, the prediction tasks under various cross-domain combinations are summarized in Table 2. For the training and testing datasets, taking task OC1-OC2 as an example, degraded data under operating condition FD001 is used as the source data, and degraded data under operating condition FD002 is used as the target domain data.

[0153] In addition, operating condition FD001 contains 100 training engines and 100 test engines. Operating condition FD002 contains 260 training engines and 259 test engines. The model is trained using the 100 labeled engine data and 260 unlabeled engine data from FD001 as the training dataset, and the 259 test engine data from FD002 are used as the test dataset.

[0154] Table 2 Cross-domain prediction tasks

[0155]

[0156] This invention uses the root mean square error (RMSE) and relative accuracy (RA) to evaluate the predictive performance of a prediction task, and the formulas are as follows:

[0157]

[0158]

[0159] in, and These are the actual RUL value and the predicted RUL value. This represents the number of test samples. RMSE reflects the deviation between the predicted RUL and the true RUL; the smaller the RMSE and the larger the RA, the better the predictive performance of the method.

[0160] For the construction of the HI (Highly Indicative Hierarchical Information), a multi-scale feature extraction module is used to generate the HI from the aero-engine data. The constructed HI is as follows: Figure 5 As shown in the figure, the constructed HI accurately represents the engine's degradation. When the system is in the early stages of operation, the degradation rate is low due to the absence of defects. After a period of time, the degradation rate becomes significantly higher than in the early stages. This indicates that the system is operating under abnormal conditions and may have developed some defects.

[0161] To evaluate the overall capability of HI, taking into account its monotonicity Mon, correlation Corr, and robustness Rob, this invention defines a comprehensive index CI that includes these three indicators, as shown in the following formula:

[0162]

[0163] like Figure 6 As shown, to demonstrate the superiority of the multi-scale feature extraction module (Model 1), a multi-scale temporal convolutional network (Model 2) was constructed in this experiment. Model 2 only uses a multi-scale structure and does not use a multi-scale local feature enhancement mechanism. Model 3 uses only the single-branch local feature enhancement mechanism of the method in this invention. The constructed HI was evaluated using the comprehensive index CI. Figure 6 It can be seen that the HI constructed by the multi-scale feature extraction module can more effectively characterize the degradation of system performance, and the HI constructed by the method of this invention is more suitable for RUL prediction. In addition, it can be seen that the HI in Model 3 is better than that in Model 2, indicating that the local feature enhancement structure can better capture the relevant local features between adjacent time points.

[0164] To improve the domain adaptation and RUL prediction performance of the DDAN model in this invention, the window width, Dropout rate, initial learning rate, batch size, and loss weight coefficients are adjusted. and The parameters were determined through grid search. The hyperparameter details of the DDAN model were used for 12 cross-domain RUL prediction tasks, as shown in Table 3. Furthermore, a variable-learning-rate Adam optimizer was employed to optimize the training process of the DDAN model, resulting in rapid convergence in the early stages of training. Optimal model parameters were obtained when the loss value stabilized.

[0165] Table 3 Selection of Hyperparameters

[0166]

[0167] Based on the RMSE and RA evaluation metrics, 12 cross-domain RUL prediction experiments were conducted, as shown in Table 2, to compare the RUL prediction performance of the DDAN model with the Source-Only model and the VBGRU-MMD model. Furthermore, to ensure a fair comparison among the three models, the multi-scale feature extraction module and RUL prediction module from the DDAN model were also applied to the Source-Only model and the VBGRU-MMD model.

[0168] The experimental results are shown in Table 4. It can be seen that the method of the present invention can achieve better prediction performance compared with Source-Only without transfer learning and VBGRU-MMD with conventional global alignment.

[0169] Table 4. Results of ablation comparison experiments

[0170]

[0171] Among them, such as Figure 7 As shown, the OC2-OC3 task is used as an example to compare the RUL prediction results of DDAN and Source-Only. In most cases, the value of the proposed DDAN is closer to the actual situation, which can also represent that the method of the present invention has better superiority.

[0172] like Figure 8 As shown, to demonstrate the predictive performance of RUL throughout its entire lifecycle, this experiment used a random engine from the target domain test dataset to perform 12 cross-domain RUL predictions. Figure 8 (a)- Figure 8 As can be seen from (l), compared with Source-Only, the RUL prediction value of DDAN in this invention is closer to the actual situation, especially in the initial and final stages of the entire life cycle, that is, the proposed method has good predictive ability.

[0173] Finally, the cross-domain RUL prediction performance of the method of this invention is compared with that of several representative domain adaptation methods, including Deep Adaptive Network (DAN), Domain Adaptive Neural Network (DANN), CORAL, ADARUL, LSTM-DANN, CNN-DANN, and Unsupervised Subdomain Adversarial Network (USDAN). Among the compared methods, DAN is a domain adaptation method that uses the MK-MMD metric function to minimize the distribution difference between the source and target domains. DANN represents a deep domain adaptation architecture based on adversarial learning mechanisms. CORAL aligns distributions by minimizing the covariance shift between different domains. ADARUL is an adversarial domain adaptation prediction method based on LSTM networks. LSTM-DANN combines LSTM with adversarial domain adaptation techniques to address the domain shift problem. USDAN utilizes an unsupervised subdomain adversarial network for cross-domain RUL prediction. The comparative experimental results of RMSE are shown in Table 5, with bold results indicating the best performance and underlined results indicating the second-best performance.

[0174] As shown in Table 5, our proposed method outperforms other methods on almost all tasks. Some deep learning-based transfer learning methods, such as LSTM-DANN and CNN-DANN, can improve cross-domain RUL prediction performance. However, using only global DA methods may obfuscate fine-grained features between subdomains represented by different degradation stages, thus reducing RUL prediction performance. USDAN aligns local features between different domains through unsupervised subdomain adversarial networks and multilinear adjustment schemes, but ignores the influence of global features on feature alignment. In contrast, the proposed DDAN can capture fine-grained information from multiple degradation stages, and its dynamic weight adjustment mechanism reduces distribution differences between different domains by aligning global and local features, thereby achieving good prediction performance. In this sense, Table 5 demonstrates that DDAN has superior performance compared to traditional transfer learning methods.

[0175] Table 5. Comparison of RMSE results predicted by RUL

[0176]

[0177] Furthermore, in the domain adaptation tasks OC1-OC4 on the CMAPSS dataset, this invention uses t-SNE to visualize the feature distribution in different domains, and the visualization results are as follows: Figure 9 As shown. From Figure 9 As can be seen, the domain shift is small, and the domain features are effectively aligned. The degradation trajectories between different domains are consistent, which allows for a more intuitive understanding of the discriminative contribution of deep domain adaptation to domain-invariant features in the target domain. It is evident that this invention achieves varying degrees of improvement across different evaluation metrics, and the prediction accuracy also increases to varying degrees.

[0178] The above are merely embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for predicting the remaining lifetime of a system in multiple degradation stages based on dynamic domain adaptive networks, characterized in that, Includes the following steps: S1. Acquire the source domain monitoring data and target domain monitoring data of the system, and perform preprocessing; the source domain monitoring data has a remaining lifetime label, and the target domain monitoring data does not have a label; S2. Construct a multi-scale feature extraction module. Input the preprocessed data into the multi-scale feature extraction module. Extract multi-scale temporal degradation features through multiple parallel convolutional branches. Use the global channel attention mechanism to adaptively fuse the extracted multi-scale temporal degradation features to obtain health indicator features that reflect the degradation state of the system. S3. Construct a degradation stage segmentation module, and use the information gain-based time segmentation method to segment the health indicator features of the source domain to obtain multiple degradation stages and their corresponding stage labels. S4. Construct a dynamic domain adaptive network, which includes a feature extractor, a subdomain classifier, and a remaining lifetime predictor. The feature extractor employs a variational autoencoder-bidirectional gated recurrent unit network to characterize the randomness and uncertainty of degraded data and to extract degraded feature data from long sequences. S5. Train the dynamic domain adaptive network using source domain data and target domain data; During training, a dynamic weight adjustment mechanism is used to adaptively adjust the weights of global domain distribution alignment and local subdomain distribution alignment, and the network parameters are updated by minimizing the total loss function; the total loss function includes variational autoencoder-bidirectional gated recurrent unit network loss, subdomain classification loss, global domain adaptation loss and local subdomain adaptation loss; S6. Input the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network, and output the remaining lifetime prediction value of the target domain system.

2. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, The multi-scale feature extraction module in step S2 includes three parallel temporal convolution branches, each of which contains multiple serial TCN residual blocks. The TCN residual blocks include dilated causal convolutional layers, weight normalization layers, ReLU activation layers, and Dropout layers, and use direct mapping of the input layer as residual connections. The three temporal convolutional branches are configured with convolutional kernels of sizes 3, 5 and 7 respectively, and the receptive field is expanded by dilated causal convolutional layers to extract short, medium and long-distance temporal dependencies. The multi-scale feature extraction module introduces a multi-scale local feature enhancement mechanism in each temporal convolutional branch. It uses the output of the TCN residual block of the previous layer in each temporal convolutional branch as the input of the next layer, and accumulates the outputs of all TCN residual blocks in each temporal convolutional branch to capture the relevant local features between adjacent time points. The global channel attention mechanism uses global average pooling and global max pooling to extract global degradation information of different sensor channels. Then, it generates the weights of attention for each channel through a Sigmoid activation layer, and performs weighted fusion on the extracted multi-scale temporal degradation features to finally form health indicator features that reflect the degradation state of the system.

3. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, Step S3 employs a time segmentation method based on information gain, specifically including: Define a cost function based on the information gain and segment length resulting from segmenting a given time series; The objective of maximizing information gain is transformed into minimizing weighted entropy. The optimal split point is solved by using dynamic programming algorithm, and the entire degradation process of the source domain is discretized into multiple degradation stages and their corresponding stage labels, which serve as the basis for subdomain partitioning.

4. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, The variational autoencoder-bidirectional gated cyclic unit network in step S4 includes: A bidirectional gated recurrent unit layer is used to capture the bidirectional temporal degradation patterns of the input feature data, enabling the extraction of long-sequence degradation feature data; The variational autoencoder layer is used to map the input feature data to the mean and standard deviation distribution of the low-dimensional latent space, and generate latent features through reparameterization techniques to characterize the randomness and uncertainty of degraded data. Both the subdomain classifier and the remaining lifetime predictor consist of fully connected layers; the subdomain classifier contains three fully connected layers with 128, 64 and 10 neurons respectively; the remaining lifetime predictor contains three fully connected layers with 128, 64 and 1 neurons respectively; and a Dropout layer is applied between every two fully connected layers with a Dropout rate of 0.

2.

5. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, The training process in step S5 specifically includes: The source domain data and target domain data are input into the multi-scale feature extraction module to obtain health indicator features, and the source domain health indicator features are divided into multiple degradation stage source domain health indicator features. By inputting the source domain health index features and target domain health index features of multiple degradation stages into the variational autoencoder-bidirectional gated recurrent unit network, the potential feature distribution of the corresponding domain is obtained. The latent features of the corresponding domain are input into the subdomain classifier; for source domain data, the cross-entropy loss between its predicted class and the true stage label is calculated as the subdomain classification loss; for target domain data, pseudo-labels are generated using the output of the subdomain classifier, and the pseudo-labels are used for subsequent local subdomain distribution alignment. The latent features of the source domain data are input into the remaining lifetime predictor, and the mean square error between the predicted value and the actual remaining lifetime label is calculated as the remaining lifetime prediction loss. Meanwhile, the dynamic weight adjustment mechanism in step S5 is specifically as follows: Calculate global feature alignment metric Alignment metrics with local features ; Calculate the adaptive coefficient according to the formula. ; ， Using the adaptive coefficients The dynamic domain adaptive loss is calculated by weighted summation of the global domain adaptation loss and the local subdomain adaptation loss. ; ， Specifically, when the global feature differences are small, the weight of local feature differences is increased; when the local feature differences are small, the weight of global feature differences is increased. The global domain adaptation loss is measured using the maximum mean difference (MMD); the local subdomain adaptation loss is measured using the local maximum mean difference (LMMD). When calculating the Local Maximum Mean Difference (LMMD), for source domain samples, the stage labels obtained in step S3 are used to calculate the weights; for target domain samples, the pseudo labels generated by the subdomain classifier in step S5 are used to calculate the weights. Furthermore, the total loss function Represented as: ， in, For the variational autoencoder-bidirectional gated cyclic unit network loss, For dynamic domain adaptive loss, Subdomain classification loss; and For the balance coefficient, the variational autoencoder-bidirectional gated recurrent unit network loss is expressed as the sum of the remaining lifetime prediction loss and the latent space regularization constraint term solved through variational inference; The value is 0.

1. The value is 0.

3.

6. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, The condition for the training to end in step S5 is: Set a fixed number of training rounds of 50; The Adam optimizer was used during training, with the initial learning rate set to 0.005 and the learning rate scheduler used to dynamically adjust the learning rate. When the value of the total loss function stabilizes and stops decreasing, or when the set number of training rounds is reached, training is stopped, and the network parameters with the minimum loss value at this point are taken as the trained model parameters.

7. The method for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network according to claim 1, characterized in that, The specific process of inputting the target domain monitoring data into the trained multi-scale feature extraction module and dynamic domain adaptive network in step S6, and outputting the remaining lifetime prediction value of the target domain system, includes: The target domain monitoring data is input into the trained multi-scale feature extraction module, and the target domain features are extracted using the convolution kernels and attention weights with fixed parameters. The target domain features are input into the feature extractor in the trained dynamic domain adaptive network to obtain latent features. The latent features are input into the trained remaining lifetime predictor, and the weighted summation and activation function mapping of the fully connected layer directly outputs the remaining lifetime prediction value of the target domain system.

8. A system for predicting the remaining lifetime of a system in multiple degradation stages based on a dynamic domain adaptive network, characterized in that, include: The data acquisition module is used to acquire source domain monitoring data and target domain monitoring data of the system; The multi-scale feature extraction module is used to extract and fuse multi-scale local enhancement features through parallel temporal convolutional branches and global channel attention mechanisms to construct health indicators. The degradation stage segmentation module is used to divide source domain data into multiple degradation stages based on the information gain time segmentation method; The dynamic domain adaptive prediction module includes a variational autoencoder-bidirectional gated recurrent unit network, a subdomain classifier, and a remaining lifetime predictor; it is used to perform transfer learning training by dynamically adjusting global and local alignment weights and output the remaining lifetime prediction results.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the system multi-degradation stage remaining lifetime prediction method based on dynamic domain adaptive networks as described in any one of claims 1-7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the method for predicting the remaining lifetime of a system based on a dynamic domain adaptive network as described in any one of claims 1-7.

Images