System, device and storage medium for predicting residual life of multiple degradation stages of system based on dynamic domain adaptive network

Abstract

The present application belongs to the technical field of residual life prediction, and in particular to a system multi-degradation phase residual life prediction method, system, device and storage medium based on a dynamic domain self-adaptive network. The method comprises the following steps: obtaining source domain monitoring data and target domain monitoring data of the system and performing preprocessing; multi-scale feature extraction and health index construction; IGTS-based degradation phase division; dynamic domain self-adaptive network containing VBGRU is constructed, and the model is trained; the target domain monitoring data is input into the trained model, and the residual life prediction value of the target domain system is output. In addition, the present application also provides a system, device and storage medium for executing the method. The system can not only effectively extract multi-scale time sequence dependent features in the system degradation process, but also adaptively align global and local features between domains, effectively improving the generalization ability of residual life prediction under variable working condition scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of remaining useful life prediction, and particularly relates to a system multi-degradation phase remaining useful life prediction method, system, device and storage medium based on a dynamic domain self-adaptive network, and especially relates to a remaining useful life prediction method and system based on deep migration learning technology under cross-working conditions. BACKGROUND

[0002] With the development of industrial equipment towards complexity, integration and intelligence, the Prognostics and Health Management (PHM) technology of the system is particularly important. Among them, the Remaining Useful Life (RUL) prediction is one of the key technologies of PHM. Accurate RUL prediction can effectively evaluate the health condition of the system and provide valuable suggestions for maintenance decision-making, thereby reducing downtime and economic loss and improving the reliability of machine operation.

[0003] The existing RUL prediction methods are mainly divided into model-based methods and data-driven methods. Due to the diversity of failure modes and the complexity of mechanisms of complex systems, it is difficult to establish an accurate physical model, therefore, the data-driven method which does not rely on prior knowledge, especially the deep learning-based method such as convolutional neural network CNN, long short-term memory network LSTM, etc., has attracted widespread attention in the field of RUL prediction due to its powerful feature extraction and nonlinear mapping ability.

[0004] However, in actual industrial scenarios, the deep learning-based RUL prediction method still faces the following severe challenges: the performance of the deep learning model usually depends on the assumption that the training data and the test data conform to the same distribution. However, in actual applications, the equipment usually runs under different working conditions, such as different loads, speeds and environmental temperatures, resulting in significant differences in feature distribution between the source domain and the target domain. This distribution difference will cause the model to have a significant decline in generalization ability when predicting across working conditions.

[0005] In order to solve the problem of distribution difference, transfer learning (Transfer Learning) and domain adaptation (Domain Adaptation, DA) technology are widely used, but there are also the following defects:

[0006] First, the defect of global alignment; the existing global domain adaptation method usually tries to align the overall distribution of the source domain and the target domain. However, the performance degradation of the system in the running process presents a multi-stage feature, such as the health stage, the early degradation stage and the serious degradation stage. Only using global alignment can easily confuse the fine-grained features between different degradation stage sub-domains, resulting in insufficient prediction accuracy.

[0007] Second, the defect of local alignment; although there are some sub-domain adaptive methods trying to align similar sub-domains, these methods often ignore the importance of global distribution alignment, resulting in the overall distribution difference between domains being ignored.

[0008] Third, the lack of dynamic balance; existing methods are difficult to find a dynamic balance between global alignment and local alignment, and cannot adaptively adjust the importance weight of the two according to the training process.

[0009] In addition, when processing multi-sensor monitoring data, the existing feature extraction method is difficult to cooperatively extract multi-scale time sequence features and spatial correlation features. In particular, for the local degradation features between adjacent time points, the existing network structure often captures insufficiently, limiting the model's ability to perceive subtle degradation trends.

[0010] Therefore, how to effectively extract multi-scale local enhanced features, and divide the degradation process into multiple stages, and then realize high-precision RUL prediction across working conditions by dynamically adjusting the alignment weight of global and local, is a technical problem to be solved at present. SUMMARY

[0011] In view of this, the first purpose of the present application is to provide a system multi-degradation phase remaining useful life prediction method and system based on a dynamic domain adaptive network, to solve the problem that the existing global domain adaptive method of transfer learning easily confuses the fine-grained features between sub-domains of different degradation stages in system RUL prediction; and to overcome the technical problem that it is difficult to cooperatively extract multi-scale time sequence features and spatial correlation features in the local feature extraction process of multi-sensor data, thereby leading to insufficient RUL prediction accuracy across working conditions.

[0012] The second purpose of the present application is to provide an electronic device and a computer readable storage medium based on the above remaining useful life prediction method and system.

[0013] In the first aspect, the present application provides a system multi-degradation phase remaining useful life prediction method based on a dynamic domain adaptive network, comprising the following steps:

[0014] S1, obtaining source domain monitoring data and target domain monitoring data of a system and preprocessing; the source domain monitoring data has a remaining useful life label, and the target domain monitoring data does not have a label;

[0015] S2, constructing a multi-scale feature extraction module, inputting the preprocessed data into the multi-scale feature extraction module, extracting multi-scale time sequence degradation features through multiple parallel convolution branches, and adaptively fusing the extracted multi-scale time sequence degradation features by using a global channel attention mechanism to obtain health index features reflecting the degradation state of the system;

[0016] S3, a degradation stage division module is constructed, and a time segmentation method based on information gain is used to divide the health index features of the source domain to obtain multiple degradation stages and corresponding stage labels;

[0017] S4, a dynamic domain adaptation network is constructed, the dynamic domain adaptation network includes a feature extractor, a sub-domain classifier and a remaining life predictor; the feature extractor adopts a variational autoencoder-bidirectional gated recurrent unit network to represent the randomness and uncertainty of the degradation data and extract long sequence degradation feature data;

[0018] S5, the dynamic domain adaptation network is trained using source domain data and target domain data; in the training process, a dynamic weight adjustment mechanism is used to adaptively adjust the weights of global domain distribution alignment and local sub-domain 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, sub-domain classification loss, global domain adaptation loss and local sub-domain adaptation loss;

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

[0020] As an optional embodiment of the application, the multi-scale feature extraction module in step S2 includes three parallel time convolution branches, each time convolution branch includes a plurality of serial TCN residual blocks; the TCN residual block includes an inflation causal convolution layer, a weight normalization layer, a ReLU activation layer and a Dropout layer, and uses the direct mapping of the input layer as the residual connection.

[0021] The three time convolution branches are respectively provided with convolution kernels with sizes of 3, 5 and 7, and then the receptive field is expanded through the inflation causal convolution layer to extract short, medium and long distance time dependence.

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

[0023] The global channel attention mechanism uses global average pooling and global maximum pooling to extract degradation global information of different sensor channels, respectively, and then generates the weight of each channel attention through a Sigmoid activation layer, weights and fuses the extracted multi-scale time series degradation features, and finally forms health index features reflecting the system degradation state.

[0024] As an optional embodiment of the present application, the time segmentation method based on information gain is used in step S3, which specifically comprises:

[0025] A cost function is defined, which is based on the information gain and the length of the line segment generated by segmenting the given time series;

[0026] The maximization of information gain is converted into the minimization of a weighted entropy objective;

[0027] The dynamic programming algorithm is used to solve the optimal segmentation point, and the entire degradation process of the source domain is discretized into multiple degradation stages and their corresponding stage labels as the basis for sub-domain division.

[0028] As an optional embodiment of the present application, the variational autoencoder-bidirectional gated recurrent unit network in step S4 comprises:

[0029] A bidirectional gated recurrent unit layer is used to capture the bidirectional time series degradation law of the input feature data, and to realize the extraction of long sequence degradation feature data;

[0030] A 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 to generate latent features through the reparameterization technique, to represent the randomness and uncertainty of the degradation data.

[0031] The sub-domain classifier and the remaining life predictor are both composed of fully connected layers; wherein the sub-domain classifier contains three fully connected layers with neuron numbers of 128, 64 and 10 respectively; the remaining life predictor contains three fully connected layers with neuron numbers of 128, 64 and 1 respectively; and a Dropout layer is applied between each two fully connected layers, with a Dropout rate of 0.2.

[0032] As an optional embodiment of the present application, the training process in step S5 specifically comprises:

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

[0034] The multi-degradation stage source domain health indicator features and the target domain health indicator features are input into the variational autoencoder-bidirectional gated recurrent unit network to obtain the latent feature distribution of the corresponding domain;

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

[0036] The potential 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.

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

[0038] calculating a global feature alignment measure and a local feature alignment measure ;

[0039] calculating an adaptive coefficient according to the formula ;

[0040]

[0041] using the adaptive coefficient to perform weighted summation on the global domain adaptive loss and the local subdomain adaptive loss to calculate a dynamic domain adaptive loss ;

[0042]

[0043] wherein when the global feature difference is small, the weight of the local feature difference is increased; and when the local feature difference is small, the weight of the global feature difference is increased.

[0044] wherein the global domain adaptive loss is measured by maximum mean difference MMD; and the local subdomain adaptive loss is measured by local maximum mean difference LMMD.

[0045] In calculating the local maximum mean difference LMMD, for the source domain samples, the stage label obtained in step S3 is used to calculate the weight; and for the target domain samples, the pseudo label generated by the subdomain classifier in step S5 is used to calculate the weight.

[0046] In addition, the total loss function is expressed as:

[0047]

[0048] wherein, is the variational autoencoder-bidirectional gated recurrent unit network loss, is the dynamic domain adaptive loss, is the subdomain classification loss; and is a balance coefficient, wherein the variational autoencoder-bidirectional gated recurrent unit network loss is expressed as the sum of the remaining life prediction loss and the latent space regularization constraint item solved by variational inference; takes a value of 0.1, takes a value of 0.3.

[0049] As an optional embodiment of the present application, the condition for ending the training in step S5 is:

[0050] The fixed number of training rounds is set to 50;

[0051] The Adam optimizer is used in the training process, the initial learning rate is set to 0.005, and the learning rate scheduler is used to dynamically adjust the learning rate.

[0052] When the value of the total loss function tends to be stable and no longer decreases, or the set number of training rounds is reached, the training is stopped, and the network parameters with the minimum loss value at this time are taken as the trained model parameters.

[0053] As an optional embodiment of the present application, the specific process of inputting the target domain monitoring data into the trained multi-scale feature extraction module and the dynamic domain adaptive network to output the remaining life 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 to extract the target domain features using the fixed parameters of the convolution kernel and the attention weight.

[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 life predictor, and the remaining life prediction value of the target domain system is directly output through the weighted sum of the full connection layer and the activation function mapping.

[0057] In a second aspect, the present application provides a system multi-degradation phase remaining life prediction system based on a dynamic domain adaptive network, comprising:

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

[0059] A multi-scale feature extraction module is used to extract and fuse multi-scale local enhanced features through parallel time convolution branches and global channel attention mechanisms to construct health indicators.

[0060] A degradation phase division module is used to divide the source domain data into multiple degradation phases based on the information gain time segmentation method.

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

[0062] In a third aspect, the present application provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the above-mentioned system multi-degradation phase residual life prediction method based on a dynamic domain adaptive network when executing the program.

[0063] In a fourth aspect, the present application provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by a processor to implement the above-mentioned system multi-degradation phase residual life prediction method based on a dynamic domain adaptive network.

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

[0065] (1) The present application proposes a health index construction method with multi-scale local feature enhancement to solve the problem of insufficient local degradation feature extraction and weak multi-scale information fusion capability in multi-sensor time series data. The method effectively extracts multi-scale time series dependent features in the system degradation process through a multi-scale time convolution structure and a local feature enhancement mechanism, providing a feature representation basis for subsequent residual life prediction.

[0066] (2) According to the phase division module, the degradation features are further divided into multiple sub-domains, and 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 latent space, enhancing the prediction robustness under complex working conditions.

[0067] (3) The residual life prediction method based on the dynamic domain adaptive network is evaluated on the aircraft turbofan engine dataset under cross-domain conditions. For the proposed dynamic domain adaptive network, the method adjusts the importance weight of the global and local distribution flexibly, and adaptively aligns the global and local features between domains, realizing the global alignment of degradation features and the local alignment of sub-domain features. BRIEF DESCRIPTION OF DRAWINGS

[0068] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0069] Figure 1 The present application relates to a prediction method overall flowchart.

[0070] Figure 2 The present application relates to a TCN residual block structure diagram.

[0071] Figure 3 Structure diagram of the multiscale feature extraction module involved in the present application.

[0072] Figure 4 Principle diagram of the DDAN involved in the present application.

[0073] Figure 5 Health index diagram constructed by the engine of the No. 5 training set FD001 involved in the present application.

[0074] Figure 6 Comparison diagram of the evaluation ability of the health index constructed by the comprehensive evaluation index evaluation model of the FD001 data set involved in the present application.

[0075] Figure 7 Comparison diagram of the RUL prediction results of two models with OC2-OC3 as an example task involved in the present application.

[0076] Figure 8 Comparison diagram of the 12 times of cross-domain RUL prediction results with random engines in the target domain test set involved in the present application.

[0077] Figure 9 t-SNE visualization feature distribution diagram involved in the present application. DETAILED DESCRIPTION

[0078] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0079] Deep transfer learning technology has been successfully applied to the remaining useful life prediction under different working conditions. However, due to the multi-stage degradation characteristics of the system performance during operation, the existing global domain adaptive method of transfer learning is prone to confuse the fine-grained features between different degradation stage sub-domains in the system RUL prediction; and in the process of local feature extraction of multi-sensor channel data, it is difficult to cooperatively extract multi-scale time sequence features and spatial correlation features, thereby resulting in insufficient RUL prediction accuracy under cross-working conditions.

[0080] Therefore, the application provides a system multi-degradation phase remaining useful life prediction method based on a dynamic domain adaptive network.

[0081] Based on the above provided technical route, as shown in Figure 1 The system multi-degradation phase remaining useful life prediction method based on the dynamic domain adaptive network provided by the application specifically comprises the following steps:

[0082] S1, source domain monitoring data and target domain monitoring data of a system are acquired and preprocessed; the source domain monitoring data has a remaining useful life label, and the target domain monitoring data does not have a label;

[0083] S2, a multi-scale feature extraction module is constructed, the preprocessed data is input into the multi-scale feature extraction module, multi-scale time sequence degradation features are extracted through a plurality of parallel convolution branches, and global channel attention mechanisms are used to adaptively fuse the extracted multi-scale time sequence degradation features, so as to obtain health index features reflecting a degradation state of the system;

[0084] S3, a degradation phase division module is constructed, an information gain-based time segmentation method is used to divide the health index features of the source domain, and a plurality of degradation phases and corresponding phase labels are obtained;

[0085] S4, a dynamic domain adaptive network is constructed, the dynamic domain adaptive network comprises a feature extractor, a subdomain classifier and a remaining useful life predictor; the feature extractor adopts a variational autoencoder-bidirectional gated recurrent unit network to represent randomness and uncertainty of degradation data and extract long sequence degradation feature data;

[0086] S5, the dynamic domain adaptive network is trained by using source domain data and target domain data; in the training process, a dynamic weight adjustment mechanism is used to adaptively adjust weights of global domain distribution alignment and local subdomain distribution alignment, and network parameters are updated by minimizing a total loss function; the total loss function comprises a variational autoencoder-bidirectional gated recurrent unit network loss, a subdomain classification loss, a global domain adaptive loss and a local subdomain adaptive 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 life prediction value of the target domain system.

[0088] In step S1, source domain monitoring data and target domain monitoring data of the system need to be acquired and preprocessed. The source domain monitoring data and the target domain monitoring data can be acquired by a multi-channel sensor. For example, the acquired monitoring data is the running data of an aircraft turbofan engine under different working conditions and fault categories after screening. The data preprocessing is to normalize the multi-channel sensor monitoring data, and then use a sliding window method to intercept time characteristic segments.

[0089] In step S2, the multi-scale feature extraction module synchronously extracts time sequence degradation features through parallel time convolution networks, and adaptively fuses key degradation features in multi-sensor data through a global channel attention mechanism, thereby significantly improving the remaining life prediction accuracy. When constructing the multi-scale feature extraction module, a shallow feature extraction unit needs to be constructed first, and then a deep feature extraction unit. The shallow feature extraction unit mainly includes an inflation causal convolution layer, a ReLU activation layer and a Dropout layer. The deep feature extraction unit mainly includes three parallel time convolution branches, and each time convolution branch contains multiple serial TCN residual blocks. It should be noted that the TCN residual block includes an inflation causal convolution layer, a weight normalization layer, a ReLU activation layer and a Dropout layer, and uses the direct mapping of the input layer as the residual connection. For example, each time convolution branch contains 3 serial TCN residual blocks.

[0090] The structure of the TCN residual block is shown in Figure 2 . Among them, the weight normalization layer and the Dropout layer are mainly used to suppress network noise and optimize network training effect. In addition, in order to ensure that the input and output dimensions are the same, a 1x1 convolution block is introduced.

[0091] The multi-scale feature extraction module configures different layers by means of two key parameters of different convolution kernel sizes and expansion factors, expands the receptive field of the convolution kernel without increasing too much calculation amount, and more effectively captures the long-distance dependence relationship in the long sequence data. Therefore, as shown in Figure 3 , the multi-scale feature extraction module designs three parallel time convolution branches, and uses three different size convolution kernels to extract features from the input data. The sizes of the three convolution kernels are set to 3, 5 and 7 respectively, and then the inflation causal convolution layer is used to expand the receptive field for extracting short, medium and long distance time dependence.

[0092] In addition, the multi-scale feature extraction module introduces a multi-scale local feature enhancement mechanism in each time convolution branch to capture the relevant local features between adjacent time points, i.e., the output of the TCN residual block of the previous layer in each time convolution branch is taken as the input of the next layer, and the outputs of all TCN residual blocks in each time convolution branch are accumulated. Different convolution block layers at different times extract different time scale features, and the entire RUL prediction model can obtain more representative feature representations by weighting and fusing the features extracted at different scales using the global channel attention mechanism, thereby better distinguishing different operating states and degradation stages of the system and improving the accuracy of the remaining useful life prediction. On this basis, the accumulation of the outputs of all TCN residual blocks on each time convolution branch and the data of the input layer will be taken as the final output of the branch.

[0093] As shown in Figure 3 The structure diagram of the multi-scale feature extraction module is shown in the figure, wherein FC is a fully connected layer, and k is the size of the convolution kernel. Each time convolution branch uses a time convolution block TCNB to extract multi-scale local information layer by layer. After time feature extraction, the degradation global information of different sensor channels is extracted by using global average pooling and global maximum pooling through a dual-channel global attention module AM, respectively. Then, a one-dimensional convolution layer and a ReLU activation layer are used to adjust the feature dimension and increase the non-linear representation of the model. Finally, the weight of each channel attention is obtained through a Sigmoid activation layer, and the weight is multiplied by the input data to obtain the corresponding output. The AM module can dynamically allocate the weights of different sensor channels, strengthen the feature information sensitive to degradation, suppress noise interference, and realize adaptive fusion of spatial dimension key information. The multi-scale feature extraction module stacks the multi-scale time features and the attention-weighted spatial features layer by layer through residual connection, and finally forms a composite health index. Among them: the dual-channel global attention module AM is a global channel attention mechanism.

[0094] It should be noted that, in the dilated causal convolution layer in the shallow feature extraction unit, or in the dilated causal convolution layer in the TCN residual block, the expansion factor d = 1, 2 and 4 are set.

[0095] In step S3, considering the inherent similarity of different stages in the system degradation process and the infinity of the RUL label, the application proposes a degradation stage division 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 realize stage segmentation. This method finds the key time points that can reflect the system health state by quantifying the information entropy change of the features. This method takes the feature vector space that can reflect the system degradation trend as input to obtain the system health state division under unsupervised monitoring. For a time series feature set, under the constraint of a given segmentation number, a cost function is constructed by introducing the concept of information gain. The information gain is the reduction of the entropy expectation generated by segmenting the given time series. The higher the gain value, the more significant the difference between the degradation patterns before and after the point, and the more suitable it is as a stage boundary. The definition of the cost function is shown in formula (1):

[0096]

[0097] Wherein, represents the information gain, represents the number of segments, represents the length of the th segment, represents the length of the entire time series, represents the entropy of the entire time series, represents the entropy of the th segment.

[0098] For a time series feature set , the feature set contains feature sequences, represents the th feature sequence, wherein ; represents the length of each sequence. Therefore, in formula (1), represents the entropy of the th segment, represents the probability of the th feature sequence in the th segment, represents the number of feature sequences, is the time segmentation point of the th segment, represents the value of the th data of the th feature sequence, represents the value of the th data of the th feature sequence.

[0099] In order to solve the optimal segmentation point in the dynamic programming framework, the compatibility problem of information gain and traditional dynamic programming needs to be solved, that is, the problem of non-separable original cost function leading to the problem of not being able to directly optimize recursion. In view of this, the cost function is adaptively converted, and the maximum information gain is converted into the minimum weighted entropy target as shown in formula (2), so as to meet the decomposability requirement of dynamic programming.

[0100]

[0101] Finally, the optimal segmentation point is solved by the IGTS algorithm of dynamic programming, and the whole degradation time sequence of the source domain can be discretized into multiple degradation stages and corresponding stage labels, which are used as the basis for sub-domain division. Each stage corresponds to a specific degradation state of the system from health to failure. The division result not only solves the problem of infinite RUL label, but also provides multi-stage input data for subsequent phased prediction model, thereby improving the RUL prediction accuracy.

[0102] In step S4, a dynamic domain adaptive network needs to be constructed. In the construction process, first, a feature extractor needs to be constructed, and then a sub-domain classifier and a remaining useful life predictor need to be constructed.

[0103] For the divided multi-stage degradation time sequence data, the RUL prediction needs to consider how to better capture the long-time dependence relationship of the degradation characteristics in each stage and represent and explain the randomness and uncertainty of the degradation data. Therefore, the present application proposes a VBGRU network that combines the interpretability of the variational autoencoder VAE with the time sequence modeling advantage of the bidirectional gated recurrent unit Bi-GRU, to realize the representation of the phased degradation state and the data interpretability, and to provide reliable input for subsequent RUL prediction. Therefore, constructing the feature extractor is to construct the VBGRU network. In the embodiment, referring to Figure 4 , the VBGRU network includes: a bidirectional gated recurrent unit Bi-GRU layer for capturing the bidirectional time sequence degradation law of the input feature data and realizing the extraction of long sequence degradation feature data; and a variational autoencoder VAE layer for mapping the input feature data to the mean and standard deviation distribution of the low-dimensional latent space, and generating latent features through the reparameterization technique to represent the randomness and uncertainty of the degradation data. The present application adopts the VBGRU network to realize the extraction of long sequence features, and overcomes the gradient disappearance and explosion problem in the feature extraction process of the traditional recurrent neural network.

[0104] It should be noted that the implementation of the reparameterization technique is realized by the following formula.

[0105]

[0106] In the formula, is the latent feature, is the mean of the low-dimensional latent space, is the standard deviation of the low-dimensional latent space, is the high-dimensional input feature, is the standard normal distribution, is the set of weight and bias parameters of the VBGRU.

[0107] The VBGRU network can be used to describe the distribution of the input data with the hidden variable. In this process, the loss function of the VBGRU network can be expressed as the sum of the mean square error term between the remaining life true value and the predicted value and the latent space regularization constraint term solved by variational inference, as shown in equation (4):

[0108]

[0109] In the formula, is the number of samples in the training process.

[0110] Next, the sub-domain classifier and the remaining life predictor are constructed. For example, the sub-domain classifier and the remaining life predictor are both composed of fully connected layers; wherein the sub-domain classifier contains three fully connected layers with neuron numbers of 128, 64 and 10 respectively; the remaining life predictor contains three fully connected layers with neuron numbers of 128, 64 and 1 respectively; and a Dropout layer is applied between each two fully connected layers, and the Dropout rate is 0.2.

[0111] In addition, since the complex system has different data representation distributions at different degradation stages, if only global domain adaptation is used, it may confuse the fine-grained features between different degradation stage sub-domains, and in addition, in aligning global and local distributions, dynamically balancing the importance of each distribution type is crucial for feature alignment. Therefore, the application also introduces a dynamic weight adjustment mechanism that can capture fine-grained information and flexibly adjust the global and local weight distribution in the dynamic domain adaptation network, which is used for RUL prediction of complex systems.

[0112] The construction framework of the RUL prediction model is established by the VBGRU network and the dynamic weight adjustment mechanism, and the framework is named as dynamic domain adaptation network DDAN, that is, DDAN is equivalent to the RUL prediction model. Its structure is shown in Figure 4 First, the source domain data with stage labels generated by the degradation stage division module and the target domain monitoring data are used as the input of the model. Secondly, the mean and the 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 respectively represent the total number of samples in the source domain and the target domain. The global domain adaptive distribution alignment method usually uses the maximum mean difference MMD to measure the difference between the source domain data distribution and the target domain data distribution, as shown in equation (5).

[0117]

[0118] In the formula, represents the reproducing kernel Hilbert space, and respectively represent the feature variables of the samples in the source domain and the target domain , represents the transformation from the initial data to the reproducing kernel Hilbert space, represents that the random variable x obeys the p distribution; represents the expectation of the result after the feature mapping of the sample obeying the source domain distribution , represents the expectation of the result after the feature mapping of the sample obeying the target domain distribution .

[0119] Secondly, the local sub-domain adaptive distribution alignment method is described, and the local feature alignment measure is calculated.

[0120] Although the global alignment can reduce the overall difference, it may confuse the fine-grained features between the sub-domains represented by different degradation stages. The local sub-domain adaptive distribution alignment method is to promote the samples belonging to the corresponding sub-domain to be closer in the mapped hidden space, and divide and into C sub-domains and , wherein c represents the class label and . Then, the local maximum mean difference LMMD is introduced, which measures the Hilbert-Schmidt norm between the kernel mean embedding of the relevant sub-domain empirical distribution in the source domain and the target domain while considering the weights of different samples, and considers the weights of different samples. In order to align the relevant sub-domains, LMMD is represented as the following formula:

[0121]

[0122] In the formula, and are the distributions of and , represents the function mapping of the input with the feature kernel and the reproducing kernel Hilbert space, denotes the expectation over the class labels , i.e., the weighted average over all classes , denotes that the random variable x follows a distribution; denotes the expectation over the results of the feature mapping for samples following the source domain distribution ; denotes the expectation over the results of the feature mapping for samples following the target domain distribution .

[0123] LMMD can measure the expectation of local distribution difference. Assuming that each sample belongs to each class with weight , there is an unbiased estimate:

[0124]

[0125] In the formula, and respectively represent the weight of and belonging to the class . Among them, and should satisfy the following conditions:

[0126]

[0127] On this basis, for source domain samples / target domain samples , the real RUL label / pseudo label , the expression of the weight is shown in formula (8):

[0128]

[0129] Among them, is the th element of the label vector .

[0130] For the labeled samples in the source domain , the weight is calculated using the one-hot encoding of the label obtained by dividing the degradation stage of the degradation data generated by the multi-scale feature extraction module through the IGTS method. For the unlabeled samples in the target domain , the pseudo label is used to calculate .

[0131] Since the changes of global distribution and local distribution affect the alignment of degradation feature distribution in different ways, and considering that the distribution difference between source features and target features evolves constantly during the whole iterative training process, the proportion of the two distribution distances must be dynamically adjusted. In order to better adjust the importance of global and local feature differences at each training stage, an adaptive coefficient is used to dynamically adjust the weights of global and local features, and the calculation formula is as follows:

[0132]

[0133] The adaptive coefficient is used to dynamically adjust the weights of global and local features, and the calculation formula is as follows:

[0134]

[0135] When is close to 0, the global feature difference is the smallest, and the local feature difference should be emphasized more; when is close to 1, the local feature difference is the smallest, and the global feature difference should be paid more attention to. Therefore, the dynamic weight adjustment mechanism can dynamically adjust the weights of global and local feature differences according to the amplitude of global and local feature differences, and ensure the representation of cross-domain degradation features.

[0136] In addition, considering that the mean distribution and the standard deviation distribution obtained by the VBGRU network can well describe the different stages of the system in the degradation process, and have special feature representation and good interpretability. Therefore, taking the mean and standard deviation as the input of formula (10) can well realize the domain adaptation of different degradation stages.

[0137] Finally, the RUL prediction model involved in the present application, in the training process, first uses the source domain data to pre-train the sub-domain classifier. Subsequently, the target domain data is passed through the sub-domain classifier to obtain the sub-domain pseudo label. Through iterative training, the accuracy of the sub-domain classifier gradually improves, so as to obtain accurate sub-domain labels. For all samples on the source domain, the sub-domain classification loss is defined as:

[0138]

[0139] Wherein, is the label obtained by the IGTS method for degradation stage division of health index HI data, which is taken as the true label, is the data feature extracted by the model, is the classifier, ​is a cross-entropy loss function. Therefore, according to formulas (4)-(11), the total loss function of the entire prediction model can be defined as:

[0140]

[0141] wherein, is the VBGRU network loss, is the dynamic domain adaptation loss, is the sub-domain classification loss; and is a balance coefficient, wherein the VBGRU network loss is represented as the sum of the remaining life prediction loss and the latent space regularization constraint term solved by variational inference; takes a value of 0.1, takes a value of 0.3.

[0142] In step S5, detailed parameters for ending the training of the RUL prediction model involved in the application are set, and specifically as follows: the fixed training round number is set to 50 times; the Adam optimizer is used in the training process, the initial learning rate is set to 0.005, and the learning rate scheduler is used to dynamically adjust the learning rate; when the value of the total loss function tends to be stable and no longer decreases, or reaches the set training round number, the 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, when the entire prediction model is trained, the target domain monitoring data is input into the trained multi-scale feature extraction module and the dynamic domain adaptation network, and the specific process of outputting the remaining life prediction value of the target domain system includes: inputting the target domain monitoring data into the trained multi-scale feature extraction module to extract the target domain features by using the fixed parameters of the convolution kernel and the attention weight; inputting the target domain features into the feature extractor in the trained dynamic domain adaptation network to obtain the latent features; inputting the latent features into the trained remaining life predictor to directly output the remaining life prediction value of the target domain system through the weighted summation and activation function mapping of the full connection layer.

[0144] On this basis, the application further provides a system multi-degradation stage remaining useful life prediction system based on a dynamic domain adaptive network, comprising: a data acquisition module, used for acquiring source domain monitoring data and target domain monitoring data of a system; a multi-scale feature extraction module, used for extracting and fusing multi-scale local enhanced features through a parallel time convolution branch and a global channel attention mechanism to construct a health index; a degradation stage division module, used for dividing the source domain data into multiple degradation stages based on an information gain time segmentation method; and a dynamic domain adaptive prediction module, comprising a variational autoencoder-bidirectional gated recurrent unit network, a subdomain classifier and a remaining useful life predictor; and used for performing transfer learning training by dynamically adjusting global and local alignment weights and outputting a remaining useful life prediction result.

[0145] The RUL migration prediction system of the system under variable working conditions comprises four modules: a data acquisition module, a multi-scale feature extraction module, a degradation stage division 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 useful life predictor. In the system, the multi-scale feature extraction module fuses multi-sensor features to construct a health index; the degradation stage division module divides the degradation data into multiple stages by the information gain time segmentation method; the dynamic domain adaptive prediction module aligns global data features and local data features of different degradation stages; and the subdomain classifier and the remaining useful life predictor perform RUL subdomain stage classification and remaining useful life prediction tasks. First, the sensor data after data preprocessing is transmitted to the multi-scale feature extraction module to construct a health index HI containing degradation features, and the extracted source domain features are then transmitted to the degradation stage division module, the VBGRU network in the dynamic domain adaptive prediction module, the subdomain classifier and the remaining useful life predictor. Then, the VBGRU network loss and the subdomain classification loss are calculated. At the same time, the extracted source domain and target domain features are combined in the dynamic domain adaptive prediction module, the dynamic domain adaptive loss is calculated, and the proposed model parameters are updated. Finally, the trained multi-scale feature extraction module and the trained remaining useful life predictor are combined to predict the test data of the target domain.

[0146] In addition, the application further provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the above-mentioned system multi-degradation stage remaining useful life prediction method based on a dynamic domain adaptive network when executing the program. At the same time, the application further provides a computer readable storage medium having a computer program stored thereon, wherein the computer program is executed by the processor to implement the above-mentioned system multi-degradation stage remaining useful life prediction method based on a dynamic domain adaptive network.

[0147] To further verify the prediction performance of the method, the C-MAPSS dataset published by NASA is used to perform RUL prediction under different working conditions, and several traditional and cutting-edge RUL prediction methods are compared with the prediction method proposed in the application.

[0148] Experimental data description: The C-MAPSS dataset is composed of four sub-datasets FD001, FD002, FD003 and FD004, simulated under different working conditions and fault type combinations, as shown in

[0149] Table 1 Description of the aero turbine engine dataset

[0150]

[0151] In Table 1, the operating conditions and fault types of FD001 are relatively simple, the working conditions of FD002 and FD004 are more, and the fault types of FD003 and FD004 are more. The training set and the test set come from different turbine engines, the training set records the sampling values of each time series during the complete period from the start of the engine to the occurrence of the fault, which contains the degradation data of the engine; while the test set records the state parameter values at a certain time point before the fault, which contains the failure data of the engine, and the purpose of this experiment is to predict the remaining useful life of the engine in the test set.

[0152] As can be seen from Table 1, the four sub-datasets have different conditions. The research goal of the application is to use only the source domain data labeled under a certain operating condition and the target domain data unlabeled under another operating condition to train the model. In addition, various cross-domain prediction tasks are summarized in Table 2, for the training dataset and the test dataset, taking the task OC1-OC2 as an example, using the degradation data under the working condition FD001 as the source data, and using the degradation data under the working condition FD002 as the target domain data.

[0153] In addition, the working condition FD001 contains 100 training engines and 100 test engines. The working condition FD002 contains 260 training engines and 259 test engines. The 100 labeled engine data in FD001 and the 260 unlabeled engine data are used as the training dataset to train the model, and the 259 test engine data in FD002 are used as the test dataset.

[0154] Table 2 Cross-domain prediction tasks

[0155]

[0156] ​The present application uses root mean square error RMSE and relative accuracy RA to evaluate the prediction performance of the prediction task, and the formula is as follows:

[0157]

[0158]

[0159] wherein, and are the RUL true value and the RUL predicted value respectively, is the number of test samples. RMSE reflects the deviation between the predicted RUL and the true RUL, and the smaller the RMSE, the greater the RA, indicating that the prediction effect of the method is better.

[0160] For the construction of HI, the multi-scale feature extraction module is used to generate HI for aero-engine data. The constructed HI is as shown in Figure 5 The chart shows that the constructed HI accurately represents the degradation of the engine. When the system is in the early stage of operation, the degradation rate is low because there is no defect. After a period of time, the degradation rate has a larger value than the early stage. This indicates that the system is running in an abnormal state and may produce some defects.

[0161] In order to evaluate the overall ability of HI, considering its monotonicity Mon, correlation Corr and robustness Rob, the present application defines a comprehensive index CI containing the three indexes, and the formula is as follows:

[0162]

[0163] As shown in Figure 6 , in order to prove the superiority of the multi-scale feature extraction module, i.e. model 1, a multi-scale time convolution network, i.e. model 2, is constructed in this experiment, which only uses a multi-scale structure and does not use a multi-scale local feature enhancement mechanism, and a local feature enhancement mechanism using only a single branch in the method of the present application, i.e. model 3, and the comprehensive index CI is used to evaluate the constructed HI. From Figure 6 It can be seen that the HI constructed by the multi-scale feature extraction module can more effectively represent the degradation of system performance, and the HI constructed by the method of the present application 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] In order to improve the field adaptation and RUL prediction performance of the DDAN model of the present application, the window width, Dropout rate, initial learning rate, batch size and loss weight coefficient and The hyperparameters of the DDAN model are determined by grid search. The details of the hyperparameters are shown in Table 3 for the 12 cross-domain RUL prediction tasks. In addition, an Adam optimizer with a variable learning rate is used to optimize the training process of the DDAN model, so that the model has a fast convergence speed at the beginning of training. When the loss value tends to be stable, the optimal model parameters can be obtained.

[0165] Table 3 Selection of hyperparameters

[0166]

[0167] Based on the evaluation indicators of RMSE and RA, 12 cross-domain RUL prediction experiments as shown in Table 2 are carried out to compare the RUL prediction performance of the DDAN model with the Source-Only model and the VBGRU-MMD model. In addition, in order to ensure a fair comparison between the three models, the multi-scale feature extraction module and the RUL prediction module in the DDAN model are also applied to the Source-Only model and the VBGRU-MMD model.

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

[0169] Table 4 Ablation comparison experiment results

[0170]

[0171] Among them, as shown in Figure 7 , taking the OC2-OC3 task 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 also represents that the method of the present application has better superiority.

[0172] As shown in Figure 8 , in order to prove the prediction performance of RUL in the whole life cycle, 12 cross-domain RUL predictions are carried out in this experiment using random engines in the target domain test data set. From Figure 8 (a)- Figure 8 , it can be seen that compared with Source-Only, the RUL prediction value of the DDAN of the present application is closer to the true situation, especially in the initial and end stages of the whole life cycle, that is, the proposed method has good predictability.

[0173] Finally, the cross-domain RUL prediction performance of the method of the present application is compared with several representative domain adaptation methods, including deep adaptation network DAN, domain adaptation neural network DANN, CORAL, ADARUL, LSTM-DANN, CNN-DANN and unsupervised sub-domain 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 domain and the target domain. DANN represents a deep domain adaptation architecture based on an adversarial learning mechanism. CORAL aligns the distribution by minimizing the covariance shift between different domains. ADARUL is an adversarial domain adaptation prediction method based on the LSTM network. LSTM-DANN combines LSTM with adversarial domain adaptation technology to solve the domain transfer problem. USDAN uses an unsupervised sub-domain adversarial network for cross-domain RUL prediction. The comparison results of RMSE are shown in Table 5, and the bold results represent the best performance and the underlined results represent the second best results.

[0174] As can be seen from Table 5, the present method has achieved better results than other methods on almost all tasks. Some deep learning-based transfer learning methods, such as LSTM-DANN, CNN-DANN, etc., can improve the cross-domain RUL prediction performance. However, the use of global DA methods alone can confuse the fine-grained features between sub-domains represented by different degradation stages, thereby reducing the RUL prediction performance. USDAN aligns the local features between different domains through an unsupervised sub-domain adversarial network and a multi-linear adjustment scheme, but ignores the influence of global features on feature alignment. In contrast, the DDAN proposed in the present application can capture the fine-grained information of multiple degradation stages, and the dynamic weight adjustment mechanism reduces the distribution difference between different domains by aligning global and local features, thereby achieving good prediction performance. In this sense, it can be proved from Table 5 that DDAN has better superiority compared with traditional transfer learning methods.

[0175] Table 5 RMSE results of comparative RUL prediction

[0176]

[0177] In addition, in the domain adaptation tasks OC1-OC4 on the CMAPSS data set, the present application uses t-SNE to visualize the feature distribution in different domains, and the visualization results are shown in Figure 9 As can be seen from Figure 9 , the domain shift is small, and the domain features are effectively aligned, and the degradation trajectories between different domains are consistent, which can more intuitively understand the discriminative contribution of domain-invariant features in the target domain by deep domain adaptation. It can be seen that the present application has different degrees of improvement in different evaluation indicators, and the prediction accuracy also has different degrees of rise.

[0178] The above merely provides the application examples, and does not limit the patent scope of the application. Any equivalent structure or equivalent process transformation based on the content of the application specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the 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, and adaptively fuse the extracted multi-scale temporal degradation features using a global channel attention mechanism to obtain health indicator features reflecting the system degradation state. The multi-scale feature extraction module includes three parallel temporal convolutional branches, each containing multiple serial TCN residual blocks; the TCN residual blocks include dilated causal convolutional layers, weight normalization layers, ReLU activation layers, and Drop layers. The out layer uses the direct mapping of the input layer as the residual connection; the three temporal convolutional branches are set 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 temporal dependencies; the multi-scale feature extraction module introduces a multi-scale local feature enhancement mechanism in each temporal convolutional branch, takes 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, and then generates the weights of each channel attention through a Sigmoid activation layer. The extracted multi-scale temporal degradation features are weighted and fused to finally form 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, wherein: The variational autoencoder-bidirectional gated recurrent unit network includes: a bidirectional gated recurrent unit layer, used to capture the bidirectional temporal degradation patterns of the input feature data and realize the extraction of long-sequence degradation feature data; and a variational autoencoder layer, 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 the degradation data. Both the subdomain classifier and the remaining lifetime predictor are composed 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. 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. 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, 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.

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, 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.

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 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.

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 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.

6. 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.

7. 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 1-6.

8. 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 1-6.

Images