A device residual life prediction method and device based on hierarchical semantic graph reasoning

Abstract

The application discloses a device residual life prediction method and device based on hierarchical semantic graph reasoning, and relates to the technical field of data processing. The method comprises the following steps: acquiring multi-source domain labeled samples and target domain unlabeled samples; extracting sample space-time features and decoupling the sample space-time features into domain-invariant features and domain-specific features; constructing a basic concept graph composed of domain nodes and degradation stage nodes and propagating the basic concept graph; combining and mapping into domain-stage composite nodes, constructing a context interaction graph for reasoning; performing hard retrieval of semantic context in a source domain, and obtaining semantic context based on similarity weighting aggregation in a target domain; modulating the domain-invariant features by using the semantic context to obtain fused features; and jointly training by using multiple losses, and outputting residual life prediction values from the fused features. The application realizes cross-condition degradation semantic transfer, reduces stage semantic mismatch and negative transfer, and improves prediction accuracy.

Description

Technical Field

[0001] This application relates to the field of data processing technology, and in particular to a method and apparatus for predicting the remaining life of equipment based on hierarchical semantic graph reasoning. Background Technology

[0002] Remaining lifetime, characterizing the time an equipment can remain operational from its current state until it reaches a failure threshold, is a key indicator in prediction and health management. Deep learning-based data-driven methods can automatically extract degradation features from sensor time series, but they often implicitly assume that training and test data are identically distributed. In practical applications, different speeds, loads, equipment models, and operating conditions can cause significant domain shifts, leading to performance degradation of models trained on source conditions on target conditions. Unsupervised domain adaptation, utilizing joint training with labeled source and unlabeled target domains, can mitigate domain shifts without labeling the remaining lifetime of the target domain. However, equipment degradation is continuous, irreversible, and stage-dependent. Existing methods merely align statistical distributions in a flat feature space, easily mismatching samples at different degradation stages with similar numerical distributions, resulting in semantic mismatch and negative transfer. Therefore, how to achieve effective transfer of degradation semantics across operating conditions without using the remaining lifetime label of the target domain is a pressing technical problem to be solved in this field. Summary of the Invention

[0003] The purpose of this application is to provide a method and apparatus for predicting the remaining life of a device based on hierarchical semantic graph reasoning, which can solve the problems of degradation stage mismatch and negative transfer caused by existing multi-source unsupervised domain adaptation methods ignoring the structured semantic relationship between the domain and the degradation stage.

[0004] To achieve the above objectives, this application provides the following solution: In a first aspect, this application provides a method for predicting the remaining lifetime of a device based on hierarchical semantic graph reasoning, including: Acquire samples from multiple source and target domains, wherein the samples from the source domains are labeled with remaining lifetime. Spatiotemporal feature extraction and decoupling are performed on all the samples to obtain domain-invariant features and domain-specific features; The remaining lifetime labels of the samples in the source domain are discretized to obtain degradation stage information of the samples in each source domain; based on the domain information and the degradation stage information of each sample, a basic concept graph containing domain nodes and degradation stage nodes is constructed; the domain nodes are updated based on the domain-specific features, the degradation stage nodes are updated based on the domain-invariant features, and all updated nodes are propagated on the basic concept graph; The propagated domain nodes and the propagated degenerate stage nodes are combined and mapped to generate composite nodes representing the joint state of the domain and the stage; a context interaction graph is constructed based on the composite nodes, and graph attention reasoning is performed on the context interaction graph to obtain the context node representation; Based on the domain information and degradation stage information of the source domain sample, the corresponding semantic context is retrieved from the context node representation for the source domain sample; based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, the corresponding semantic context is aggregated for the target domain sample. Modulation parameters are generated based on the semantic contexts, and the domain-invariant features are modulated using the modulation parameters to obtain fused features; wherein, the modulation parameters include a scale vector and a bias vector; The spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module are jointly optimized and trained using multiple preset loss functions. The trained remaining lifetime prediction module outputs the remaining lifetime prediction value based on the fusion features of the target domain samples. The preset loss functions include a regression loss constructed based on the fusion features of the source domain samples.

[0005] Optionally, the step of extracting and decoupling spatiotemporal features from all the samples specifically includes: The sensor channels at each sampling time are treated as graph nodes. The association weights between different sensor channels are calculated and neighborhood information is aggregated through the spatial graph attention mechanism to obtain the spatially enhanced sequence. The spatially enhanced sequence is fed into a temporal coding network after adding positional encoding. Global average pooling is then performed on the output of the temporal coding network to obtain the encoded features. The temporal coding network includes a multi-head self-attention sub-layer and a feedforward network sub-layer. The encoded features are mapped to a domain-invariant feature space and a domain-specific feature space through two independent mapping networks, respectively.

[0006] Optionally, the basic concept diagram includes: Domain nodes that correspond one-to-one with each source domain and target domain; Each discrete degradation stage corresponds to a degradation stage node, and adjacent degradation stage nodes are connected in the order of the equipment degradation process to form a stage manifold chain. Fully connected edges between any two domain nodes, and cross-type connected edges between each domain node and each degeneration stage node.

[0007] Optionally, updating the domain nodes based on domain-specific features, updating the degradation stage nodes based on domain-invariant features, and propagating the updated nodes across the basic concept graph specifically includes: The mean of domain-specific features of samples belonging to the same domain in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the domain node according to a preset momentum coefficient to complete the update of the domain node; The mean of the domain-invariant features of samples belonging to the same degradation stage in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the degradation stage node according to the preset momentum coefficient to complete the update of the degradation stage node. Based on the edge connections of the basic concept graph, neighborhood information is aggregated for all updated nodes to complete the information propagation.

[0008] Optionally, the step of combining and mapping the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; constructing a context interaction graph based on the composite nodes, and performing graph attention reasoning on the context interaction graph to obtain context node representations, specifically includes: The representation of each propagated domain node is concatenated with the representation of each propagated degenerate stage node, and the concatenation result is projected to a unified dimension through a nonlinear mapping layer to generate a domain-stage composite node. Establish fully connected edges between all the composite nodes to form the context interaction graph; A multi-head graph attention network is used to calculate the attention weight between any two composite nodes. The information of adjacent nodes is aggregated according to the attention weight, and the context node representation is obtained through residual connection.

[0009] Optionally, the step of retrieving the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and aggregating the corresponding semantic context for the target domain sample based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, specifically includes: For a source domain sample, based on the source domain to which the source domain sample belongs and the degradation stage of the source domain sample, the corresponding context node representation is directly indexed from the context node representation, and the indexed context node representation is used as the semantic context of the source domain sample. For a target domain sample, the cosine similarity between the domain-invariant features of the target domain sample and the node representations of each degradation stage is calculated. The cosine similarity after temperature coefficient scaling is normalized exponentially to obtain the soft matching weights of each degradation stage. The soft matching weights are then used to perform a weighted summation on all context node representations corresponding to the target domain, and the weighted summation result is used as the semantic context of the target domain sample.

[0010] Optionally, the step of generating modulation parameters based on each of the semantic contexts and using the modulation parameters to modulate the domain-invariant features to obtain fused features specifically includes: The semantic context is input into the scale generation network and the bias generation network respectively to obtain the scale vector and bias vector with the same dimension as the domain invariant feature. The neighborhood-invariant features are scaled element-wise using the sum of the unit vector and the scale vector as a scaling factor, and then superimposed with the bias vector to obtain the fused features.

[0011] Optionally, the multiple preset loss functions include remaining lifetime regression loss, domain adversarial loss, domain reconstruction loss, feature orthogonality loss, and stage consistency comparison loss; the joint optimization training of the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module using multiple preset loss functions specifically includes: The remaining lifetime regression loss is constructed using the mean squared error between the predicted remaining lifetime and the actual remaining lifetime of the source domain samples. The domain-invariant features and the domain discriminator are connected by a gradient inversion layer to construct the domain adversarial loss; The domain reconstruction loss is constructed based on the cross-entropy of the real domain number of the sample reconstructed from the domain-specific features by the domain classifier. The feature orthogonal loss is constructed based on the normalized cross-correlation between the domain-invariant feature matrix and the domain-specific feature matrix in the current training batch; Samples with the same degradation stage are constructed as positive sample pairs, and samples with different degradation stages are constructed as negative sample pairs, thus constructing the stage consistency comparison loss. The remaining lifetime regression loss, the domain adversarial loss, the domain reconstruction loss, the feature orthogonality loss, and the stage consistency comparison loss are weighted and summed to obtain the total loss; The total loss is used to jointly optimize and train the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module.

[0012] Optionally, the degradation stage of the target domain sample is determined by the degradation stage corresponding to the maximum weight in the soft matching weight; the stage consistency comparison loss simultaneously constrains samples belonging to the same degradation stage in different domains to move closer to each other in the domain-invariant feature space, and constrains samples belonging to different degradation stages to move further apart in the domain-invariant feature space.

[0013] Secondly, this application provides a device for predicting remaining lifetime based on hierarchical semantic graph reasoning, comprising: The data acquisition module is used to acquire samples from multiple source domains and a target domain, wherein the samples from the source domains are labeled with remaining lifetime. The spatiotemporal feature extraction and decoupling module is used to extract and decouple spatiotemporal features from all the samples to obtain domain-invariant features and domain-specific features. The graph construction and update module is used to discretize the remaining lifetime labels of the samples in the source domain to obtain degradation stage information of the samples in each source domain; based on the domain information and the degradation stage information of each sample, a basic concept graph containing domain nodes and degradation stage nodes is constructed; the domain nodes are updated based on the domain-specific features, the degradation stage nodes are updated based on the domain-invariant features, and all updated nodes are propagated on the basic concept graph; The graph reasoning module is used to combine and map the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; construct a context interaction graph based on the composite nodes, and perform graph attention reasoning on the context interaction graph to obtain the context node representation; The context retrieval module is used to retrieve the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and to aggregate the corresponding semantic context for the target domain sample based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node. The feature modulation module is used to generate modulation parameters based on each semantic context, and to modulate the domain-invariant features using the modulation parameters to obtain fused features; The joint optimization module is used to perform joint optimization training on the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module using multiple preset loss functions. The remaining lifetime prediction module is used to output the remaining lifetime prediction value based on the fusion features of the target domain samples after training.

[0014] Thirdly, this application provides a computer device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of the device remaining lifetime prediction method based on hierarchical semantic graph reasoning as described above.

[0015] Fourthly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the device remaining lifetime prediction method based on hierarchical semantic graph reasoning as described above.

[0016] Fifthly, this application provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the device remaining lifetime prediction method based on hierarchical semantic graph reasoning as described above.

[0017] According to the specific embodiments provided in this application, the following technical effects are disclosed: This application provides a method and apparatus for predicting the remaining lifetime of a device based on hierarchical semantic graph reasoning. By acquiring samples from multiple source and target domains and labeling the source domain samples with remaining lifetime information, a data foundation is provided for subsequent cross-domain transfer learning, making it possible to predict remaining lifetime even when the target domain is unlabeled. By extracting and decoupling spatiotemporal features from all samples to obtain domain-invariant and domain-specific features, it can effectively capture the spatial coupling relationships between sensors and long-range degradation dependencies in the temporal dimension, while suppressing the interference of operating noise on cross-domain degradation knowledge transfer. By discretizing the remaining lifetime labels of the source domain samples, degradation stage information of each source domain sample is obtained, and based on… A basic concept graph containing domain nodes and degradation stage nodes is constructed based on domain information and degradation stage information. Then, two types of nodes are updated based on domain-specific and domain-invariant features, and information is propagated on the basic concept graph. This allows the model to explicitly organize the structured semantic relationships between the domain and degradation stages, enabling stage nodes to absorb continuous degradation information from adjacent stages and shared information from various domains, providing stable and referential anchors for the unlabeled target domain. Composite nodes are generated by combining and mapping the propagated domain nodes with degradation stage nodes. A context interaction graph is then constructed based on these composite nodes, and graph attention inference is performed to obtain context node representations. This expands domain adaptation from flat statistical alignment to having a degradation order. Structured semantic reasoning of domain relationships can reduce misalignment between different degradation stages. By retrieving the corresponding semantic context for source domain samples from context node representations based on domain and degradation stage information, and aggregating the corresponding semantic context for target domain samples based on the similarity between the domain-invariant features of target domain samples and nodes at each degradation stage, labeled source domains can perform deterministic context retrieval using real-world supervised information. Simultaneously, it provides soft matching criteria for unlabeled target samples using degradation stage prototypes, avoiding single-point hard matching based solely on unstable pseudo-labels. Furthermore, it generates modulation parameters based on each semantic context and uses these parameters to modulate domain-invariant features. By obtaining fused features, the semantic context obtained from graph reasoning is fed back to the domain-invariant features through feature-level modulation, forming a closed loop of "feature extraction - graph reasoning - context retrieval - feature modulation - lifetime prediction", allowing structured priors to directly participate in sample-level prediction. By using multiple preset loss functions for joint optimization training, and using the trained lifetime prediction module to output the lifetime prediction value based on the fused features of the target domain samples, and the preset loss functions include a regression loss constructed based on the fused features of the source domain samples, the model can maintain the separability of the degradation stage while pursuing domain alignment, avoiding negative transfer, thereby improving the accuracy, robustness and interpretability of lifetime prediction across operating conditions as a whole. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 A flowchart illustrating a method for predicting the remaining life of a device based on hierarchical semantic graph reasoning, provided in an embodiment of this application; Figure 2 A schematic diagram of the overall architecture of the device remaining lifetime prediction method based on hierarchical semantic graph reasoning provided in this application; Figure 3 This is a schematic diagram illustrating the source domain hard context retrieval and target domain soft context aggregation process provided in this application. Figure 4 A schematic diagram illustrating the hierarchical semantic graph construction process, which evolves from a basic concept graph to a context interaction graph, provided for this application. Figure 5 The figures show partial remaining lifetime prediction results of this application on two rolling bearing datasets, where (a) to (f) correspond to transfer tasks A, B, C, D, E and F respectively; Figure 6 The diagram shows the remaining lifetime prediction results of this application under three hard disk drive migration tasks, where (a) to (f) correspond to migration tasks G1, G2, G3, H1, H2 and I1, respectively; Figure 7 The image shows the 3D t-SNE feature distribution of this application on the migration task E, where (a) and (b) represent the feature distribution of the baseline method and the method of this application when coloring by the domain, respectively, and (c) and (d) represent the feature distribution of the baseline method and the method of this application when coloring by the degradation stage, respectively. Figure 8 This is a graph showing the relative performance improvement of this application compared to the suboptimal method on twelve bearing migration tasks; Figure 9 A graph showing the average performance comparison of different multi-source domain adaptation methods across the entire dataset; Figure 10 Figure (a) shows the overall importance ranking of SMART attributes on the hard disk learned in this application on the transfer task G; Figure (b) shows the attention curves of the top six SMART attributes as the degradation stage changes; and Figure (c) shows the normalized attention heatmap between the five degradation stages and the twelve SMART attributes. Figure 11 The figure shows the ablation experiment results of this application on the XJTU-SY rolling bearing dataset; Figure 12 A schematic diagram of the functional modules of a device remaining life prediction device based on hierarchical semantic graph reasoning provided in an embodiment of this application; Figure 13 This is a schematic diagram of the structure of a computer device provided in an embodiment of this application. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] To make the above-mentioned objectives, features and advantages of this application more apparent and understandable, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0022] In one exemplary embodiment, such as Figure 1 and Figure 2 As shown, a method for predicting the remaining lifespan of a device based on hierarchical semantic graph reasoning is provided. This method is executed by a computer device, specifically by a terminal or server alone, or by both a terminal and a server, and includes the following steps 101 to 107. Wherein: Step 101: Obtain samples from multiple source domains and target domains, wherein the samples from the source domains are labeled with remaining lifetime.

[0023] In this embodiment of the application, the acquisition There is one source domain and one target domain, each domain corresponding to a certain operating condition, equipment model, or data distribution. The samples in the source domain and the samples in the target domain are represented as follows: ; ; in, Representing the One source domain dataset, , Represents the total number of source domains; Representing the The first in the source domain One sample, For multivariate time window samples, This corresponds to the actual remaining lifespan value; Represents the target domain dataset; Represents the first in the target domain A multivariate time window sample; Represents the length of the time window; This represents the number of sensor channels.

[0024] In this embodiment, according to the preset time window length and number of sensor channels A sliding time window process is applied to the monitoring sequences of equipment in various fields to form multivariate time window samples. The true remaining lifetime labels of the source domain samples are retained, while the true remaining lifetime labels of the target domain samples are not provided to the model. The training objective is to jointly train a prediction function using all labeled samples from the source domain and unlabeled samples from the target domain, without using the true remaining lifetime labels of the target domain, so that this prediction function maps the target domain time window samples to the predicted remaining lifetime values ​​of the target domain.

[0025] Step 102: Perform spatiotemporal feature extraction and decoupling on all the samples to obtain domain-invariant features and domain-specific features.

[0026] As an optional implementation, the spatiotemporal feature extraction and decoupling of all the samples specifically includes: The sensor channels at each sampling time are treated as graph nodes. The association weights between different sensor channels are calculated and neighborhood information is aggregated through the spatial graph attention mechanism to obtain the spatially enhanced sequence. The spatially enhanced sequence is fed into a temporal coding network after adding positional encoding. Global average pooling is then performed on the output of the temporal coding network to obtain the encoded features. The temporal coding network includes a multi-head self-attention sub-layer and a feedforward network sub-layer. The encoded features are mapped to a domain-invariant feature space and a domain-specific feature space through two independent mapping networks, respectively.

[0027] In this embodiment of the application, for any time window sample , take time steps Below Each sensor channel serves as a node in a fully connected spatial graph, and the spatial graph attention mechanism is used to extract cross-channel spatial correlations. Spatial coupling often exists between device monitoring channels due to physical structure, control relationships, or shared loads. Processing each channel separately makes it difficult to express these coupling relationships that change with operating states. Therefore, this embodiment does not pre-fix channel importance but instead calculates the correlation score between any two channels based on the node characteristics of the current time step.

[0028] The formula for calculating the correlation score is: ; in, Represents time step Downstream from channel node To channel node Unnormalized association scores during information transmission; and Representing time steps Lower channel node and channel nodes eigenvectors; and These are the trainable parameters for the spatial graph attention layer.

[0029] The formula for calculating the normalized spatial attention weights is: ; in, Indicates time step Lower channel node For channel nodes Normalized spatial attention weights; Indicates channel node The set of neighbors.

[0030] Channel Node The node representation after aggregating neighbor information is as follows: ; in, Indicates spatially enhanced sequences; This indicates a unified hidden dimension. This represents the nonlinear activation function of the spatial graph attention layer.

[0031] After the weighted aggregation described above, each channel node incorporates information from other channels related to the current state. The node representations at the same time step are then merged into a spatially augmented representation, which is arranged chronologically to form a spatially augmented sequence. This sequence retains the original time window length and transforms isolated sensor readings into features containing cross-channel structural information, serving as input to the temporal coding network.

[0032] In this embodiment, the spatial augmentation sequence is input into a temporal coding network after adding positional encoding. In this embodiment, the temporal coding network is a Transformer encoder, including a multi-head self-attention sublayer and a feedforward network sublayer. Positional encoding is used to preserve the sequential relationship between each time step in the degradation sequence—specifically, it can be implemented using sine / cosine positional encoding, with its dimension matching the feature dimension of the spatial augmentation sequence, and is superimposed onto the temporal dimension of the spatial augmentation sequence through addition—enabling the multi-head self-attention to establish long-distance dependencies across the entire time window. Compared to using only local computations of adjacent time steps, this structure can simultaneously focus on degradation trends, short-term fluctuations, and anomalous shocks.

[0033] The scaled dot product attention representation of a single attention head is: ; in, Represents the query matrix; Represents the bond matrix; Representative value matrix; The key vector dimension represents a single attention head.

[0034] No. The output calculation formula for each self-attention head is as follows: ; in, Representing the The output of a self-attention head; , , These represent the query, key, and value projection matrices of the dir attention head, respectively.

[0035] The multi-head self-attention operation is as follows: ; in, This represents the output of the multi-head self-attention operation. Represents the total number of self-attention heads; This represents vector concatenation operations; This represents the multi-head output projection matrix. Different attention heads learn the correlations between time steps in different projection spaces, and the concatenated result is projected onto the output to form a unified temporal representation. Temporal coding networks also include feedforward networks, residual connections, and layer normalization.

[0036] Then, output the encoder. Perform global average pooling over time: ; in, This represents the complete feature sequence output by the Transformer encoder. Indicates that the Transformer is at the 1st epoch. Output at each time step; This represents the sample-level spatiotemporal encoding features after global average pooling; Indicates the length of the time window.

[0037] Global average pooling compresses the encoding results at each time step into a fixed-dimensional sample-level representation. This makes subsequent graph node updates and remaining lifetime predictions no longer dependent on variable time locations. This representation simultaneously contains actual equipment degradation information and information from domains such as operating conditions and model. If directly used for cross-domain alignment, domain-specific noise may be incorrectly transferred as degradation knowledge.

[0038] In this embodiment of the application, the spatiotemporal coding features are explicitly decoupled using two independent multilayer perceptrons: ; ; in, and Representing domain-invariant features and domain-specific features; and Represents domain-invariant branches and domain-specific branches; and Dimensions representing domain-invariant features and domain-specific features.

[0039] The domain-invariant branch preserves degradation patterns that can be reused across operating conditions. Its output is used for degradation stage prototype updates, target domain stage matching, context modulation, and remaining lifetime regression. The domain-specific branch preserves domain cues such as rotational speed, load, hard drive model, or workload. Its output is used to update domain nodes and reconstruct domain categories. The two branches are not arbitrarily split, but are jointly trained through subsequent domain adversarial processes, domain reconstruction, and feature orthogonality constraints, so that degradation information and operating condition information are concentrated in their respective subspaces.

[0040] This implementation method effectively captures the coupling relationship between monitoring channels of the equipment as the operating state changes by adaptively calculating the correlation weights between different sensor channels through a spatial graph attention mechanism; it establishes long-range dependencies in the time dimension through a temporal coding network, which can simultaneously focus on degradation trends, short-term fluctuations, and abnormal shocks; and it decouples the encoded features into domain-invariant features and domain-specific features through two independent mapping networks, so that degradation information and operating condition information are concentrated in their respective subspaces, thereby suppressing the interference of operating condition noise on cross-domain degradation knowledge transfer.

[0041] Step 103: Discretize the remaining lifetime labels of the samples in the source domain to obtain degradation stage information of the samples in each source domain; construct a basic concept graph containing domain nodes and degradation stage nodes based on the domain-specific features of each sample and the degradation stage information; update the domain nodes based on the domain-specific features of the domain, update the degradation stage nodes based on the domain-invariant features of the domain, and propagate the updated nodes on the basic concept graph.

[0042] In this embodiment, the continuous remaining lifetime of source domain samples is mapped to discrete degradation stages. Remaining lifetime prediction is essentially a continuous value regression, while the degradation stage nodes in the graph need to have clear and reusable discrete semantics. Therefore, the lifecycle is proportionally divided according to the ratio of the sample's remaining lifetime to the corresponding device's complete lifecycle, ensuring that devices with different absolute lifecycle lengths fall into common healthy, early degradation, mid-term degradation, late degradation, or failure stages. The degradation stage number is determined according to the following formula: ; in, Indicates the discrete degradation stage number; and These represent the actual remaining lifetime and the corresponding maximum lifetime of the sample, respectively. This indicates the total number of degradation stages.

[0043] This mapping assigns samples near the maximum remaining lifetime to the initial healthy stage and samples nearing failure to the final degradation stage, while ensuring that the stage number monotonically changes with the degradation process. The resulting source domain stage labels are not only used to organize the stage nodes of the basic concept graph, but also serve as supervision information for subsequent source domain hard context retrieval and stage consistency learning.

[0044] As an optional implementation, the basic concept diagram includes: Domain nodes that correspond one-to-one with each source domain and target domain; Each discrete degradation stage corresponds to a degradation stage node, and adjacent degradation stage nodes are connected in the order of the equipment degradation process to form a stage manifold chain. Fully connected edges between any two domain nodes, and cross-type connected edges between each domain node and each degeneration stage node.

[0045] In this embodiment of the application, a basic concept diagram is constructed. Node set Includes domain node set and the set of nodes in the degradation stage The domain node set contains One source domain node and one target domain node; the degenerate stage node set contains Each stage node is defined. The edge set is constructed according to three types of physical and semantic relationships. First, adjacent degradation stage nodes are sequentially connected to form a stage manifold chain, used to maintain the continuous and irreversible sequential relationship of equipment degradation and avoid directly treating distant stages as equivalent. Second, full connections are established between all domain nodes, enabling the model to learn the potential similarities and differences between different operating conditions or equipment models. Third, cross-type connections are established between domain nodes and degradation stage nodes, allowing operating condition information to interact with degradation stage information. The basic concept graph thus forms a stable atomic semantic dictionary, providing referable domain and stage anchors for the unlabeled target domain.

[0046] In step 103, "updating the domain nodes based on the domain-specific features, updating the degenerate stage nodes based on the domain-invariant features, and propagating the information of all updated nodes on the basic concept graph" is a higher-level summary of the node update and propagation operations. Its specific implementation methods include weighted fusion update based on momentum coefficient and neighborhood information aggregation based on graph edge relationships.

[0047] As an optional implementation, the step of updating the domain nodes based on domain-specific features, updating the degradation stage nodes based on domain-invariant features, and propagating information from all updated nodes onto the basic concept graph specifically includes: The mean of domain-specific features of samples belonging to the same domain in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the domain node according to a preset momentum coefficient to complete the update of the domain node; The mean of the domain-invariant features of samples belonging to the same degradation stage in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the degradation stage node according to the preset momentum coefficient to complete the update of the degradation stage node. Based on the edge connections of the basic concept graph, neighborhood information is aggregated for all updated nodes to complete the information propagation.

[0048] In this embodiment, in each mini-batch training iteration, domain nodes are updated with domain-specific features, and degenerate stage nodes are updated with domain-invariant features. If nodes are set to fixed parameters independent of the samples, the graph semantics cannot reflect changes in the current feature space in a timely manner; if the node representation is completely replaced by the mean of a single mini-batch, it is easily affected by random sampling and pseudo-label fluctuations. Therefore, this embodiment uses momentum updating to smoothly integrate historical node representations with the current mini-batch statistics. The domain node embedding is updated according to the following formula: ; in, Indicates the first Embedding of nodes in various domains; Indicates the momentum coefficient; This represents the set of samples belonging to this domain in the current batch; It represents domain-specific characteristics.

[0049] Update the embedding of degenerate stage nodes using domain-invariant features: ; in, Indicates the first Embedded nodes and stage prototypes for each degradation stage; This represents the set of samples in the current batch that belong to this degradation stage; Represents domain-invariant features.

[0050] Domain nodes are calibrated using the mean of domain-specific features from samples within the same domain, enabling them to aggregate relevant operating conditions or model information. Stage nodes are calibrated using the mean of domain-invariant features from samples within the same stage, allowing them to gradually form a degenerate stage prototype independent of any specific domain. Momentum coefficients are used to balance historical knowledge and current batch information, thereby obtaining stable and updatable node representations during training.

[0051] In this embodiment of the application, graph information propagation is performed on the dynamically updated basic concept nodes: ; in, This indicates the node embedding after propagation; Represents a node The set of neighbors; Indicates the weight of the neighbor's contribution; This represents the propagation layer weight matrix. This represents the nonlinear activation function of the propagation layer of the basic concept graph.

[0052] Information propagation selects neighbors and performs weighted aggregation based on edge connections. Stage nodes can absorb continuous degradation information from adjacent stages and shared information from various domains, while domain nodes can also perceive the states of different stages and other domains. The propagated stage nodes are no longer simply the average of the current batch, but stable prototypes that take into account the degradation order and cross-domain relationships; the propagated domain nodes summarize the domain characteristics of the corresponding working condition throughout the entire degradation process.

[0053] This implementation method involves constructing a basic concept graph containing domain nodes and degradation stage nodes, and establishing stage manifold chains, domain fully connected edges, and cross-type connected edges. This allows the model to explicitly organize the hierarchical relationship between the domain and degradation stages. By smoothly integrating historical node representations with current mini-batch statistics through momentum updates, nodes remain stable and updatable during training. Furthermore, information propagation on the basic concept graph enables stage nodes to absorb continuous degradation information from adjacent stages and shared information from various domains, providing a structured atomic semantic foundation for subsequent context interaction graph reasoning.

[0054] Step 104: Combine and map the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; construct a context interaction graph based on the composite nodes, and perform graph attention reasoning on the context interaction graph to obtain context node representations.

[0055] As an optional implementation, the step of combining and mapping the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; constructing a context interaction graph based on the composite nodes, and performing graph attention reasoning on the context interaction graph to obtain context node representations, specifically includes: The representation of each propagated domain node is concatenated with the representation of each propagated degenerate stage node, and the concatenation result is projected to a unified dimension through a nonlinear mapping layer to generate a domain-stage composite node. Establish fully connected edges between all the composite nodes to form the context interaction graph; A multi-head graph attention network is used to calculate the attention weight between any two composite nodes. The information of adjacent nodes is aggregated according to the attention weight, and the context node representation is obtained through residual connection.

[0056] In this embodiment, the propagated domain nodes and degradation stage nodes are combined and mapped to generate domain-stage composite nodes. The domain nodes and stage nodes in the basic concept graph represent two types of atomic semantics, but remaining lifetime prediction requires further answering the question of "what state should a domain exhibit when it is in a certain degradation stage?" Directly establishing this joint state based on the target domain label is not feasible because the target domain does not have a true degradation stage label. This embodiment combines each domain node with each general stage prototype, thereby generating complete candidate joint states for all source and target domains.

[0057] The combined mapping is represented as: ; in, Representation of domain and stage The initial composite node; and These represent the domain embedding and stage embedding after propagation, respectively; Indicates a composite mapping layer; and These represent the weight matrix and bias vector of the combined mapping layer, respectively.

[0058] The composite mapping layer projects heterogeneous domain and stage information onto a unified dimension and forms a composite representation through nonlinear transformation. Specifically, even if the target domain lacks stage labels, hypothetical states can be generated using target domain nodes and each general stage prototype, thus completing the candidate semantics for all degenerate stages of the target domain. These candidate states are not the final matching results but are used for subsequent graph reasoning and soft retrieval.

[0059] In this embodiment of the application, all The fully connected composite nodes form a context interaction graph. Furthermore, it utilizes a multi-head graph attention network for high-order semantic reasoning. The purpose of employing a fully connected topology is to avoid pre-defining which domains or stages can transfer information between each other; instead, graph attention automatically determines the importance of information sources based on node content. For example... Figure 4 As shown, the basic concept graph evolves into a context interaction graph through combination mapping, realizing a hierarchical semantic organization from atomic semantics to joint states. For any two nodes in the context interaction graph... and , No. Calculate unnormalized correlation for each graph attention head: ; in, Indicates the first Attention to nodes in the graph For nodes The correlation score; and These are the trainable parameters for the attention head.

[0060] The attention weights for the normalized graph are: ; in, Indicates the first Attention to nodes in the graph For nodes Normalized attention weights.

[0061] The context nodes after graph reasoning are represented as follows: ; in, The context node representation after graph reasoning; This indicates the total number of attention heads in the graph; This represents the aggregate weight of node information.

[0062] After normalization, the attention weights reflect the node The cross-domain and cross-stage experience included is relevant to nodes. The contribution is that multiple graph attention heads aggregate all node information in parallel from different relation subspaces, and residual connections preserve the current joint state itself. The resulting context node representation simultaneously includes information about the current stage in this domain, transferable degradation experience from other source domains, and global semantics of adjacent or related stages, rather than being a simple concatenation of domain nodes and stage nodes.

[0063] In this implementation, the representation of each propagated domain node is concatenated with the representation of each propagated degenerate stage node and nonlinearly combined and mapped to generate domain-stage composite nodes, thus generating complete candidate joint states for all source and target domains. By establishing a fully connected context interaction graph between composite nodes and using a multi-head graph attention network for high-order semantic reasoning, the model can automatically determine the importance of information sources based on node content, and aggregate all node information in parallel from different relational subspaces to obtain context node representations containing cross-domain degenerate knowledge.

[0064] Step 105: Based on the domain information and degradation stage information of the source domain sample, retrieve the corresponding semantic context for the source domain sample from the context node representation; based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, aggregate the corresponding semantic context for the target domain sample.

[0065] In this embodiment of the application, "retrieval" and "aggregation" are generalizations of two different processing methods in this step: the source domain uses a deterministic index, and the target domain uses a probability-weighted approach based on similarity.

[0066] As an optional implementation, the step of retrieving the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and aggregating the corresponding semantic context for the target domain sample based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, specifically includes: For a source domain sample, based on the source domain to which the source domain sample belongs and the degradation stage of the source domain sample, the corresponding context node representation is directly indexed from the context node representation, and the indexed context node representation is used as the semantic context of the source domain sample. For a target domain sample, the cosine similarity between the domain-invariant features of the target domain sample and the node representations of each degradation stage is calculated. The cosine similarity after temperature coefficient scaling is normalized exponentially to obtain the soft matching weights of each degradation stage. The soft matching weights are then used to perform a weighted summation on all context node representations corresponding to the target domain, and the weighted summation result is used as the semantic context of the target domain sample.

[0067] In this embodiment of the application, hard context retrieval is performed on the source domain samples. For example... Figure 3 As shown, the source domain sample has a real domain ID, and the real remaining lifetime of the sample can be determined by the degradation stage ID through step 103. Therefore, the domain-stage node corresponding to the sample can be uniquely located in the context interaction graph. Directly using this node as the source domain semantic context can avoid introducing unnecessary matching errors when reliable supervision is available, and establish a definite training objective for the context modulation network. ; in, Represents the semantic context of the source domain sample; Represents the source domain With the real stage The corresponding context node.

[0068] The context obtained from source domain hard retrieval not only describes the source domain number and stage number, but also contains higher-order information aggregated from other domains and stages. During training, this context participates in feature modulation and remaining lifetime prediction, enabling the model to learn how to transform structured degradation knowledge into sample-level prediction corrections.

[0069] In this embodiment, soft context aggregation based on degradation stage prototypes is performed on the target domain samples. Since the target domain samples do not have true degradation stage labels, directly selecting a single pseudo-stage would cause early prediction errors, leading to context retrieval falling into incorrect nodes and further amplifying the bias. Therefore, this embodiment first calculates the cosine similarity between the target domain's domain-invariant features and all degradation stage prototypes, and then obtains the stage probability distribution after temperature scaling and normalization. ; in, Indicates the target domain sample pair of the first Soft matching weights for each stage; Indicates the domain-invariant features of the target domain; Indicates a stage prototype; This indicates the matching temperature coefficient.

[0070] The semantic context of soft aggregation of the target domain is: ; in, This represents the semantic context of soft aggregation of the target domain; Indicates stage weight; Indicates the target domain is in the th order. Context nodes under each candidate stage.

[0071] A larger soft-match weight indicates that the target domain sample is closer to the corresponding degradation stage. Using this weight to perform a weighted summation of the context node representations of the target domain across all candidate stages can simultaneously preserve reasonable information from multiple stages at stage boundaries or for uncertain samples, reducing semantic jumps caused by a single hard pseudo-label. The temperature coefficient is used to control the concentration of the probability distribution.

[0072] In one possible implementation, the degradation stage of the target domain sample is determined by the degradation stage corresponding to the maximum weight in the soft matching weights; the stage consistency contrast loss simultaneously constrains samples belonging to the same degradation stage in different domains to move closer to each other in the domain-invariant feature space, and constrains samples belonging to different degradation stages to move further apart in the domain-invariant feature space.

[0073] Specifically, to construct the positive and negative sample relationship in phase consistency contrastive learning, the pseudo-degenerate stage of the target domain samples is determined based on the maximum soft-match weight: ; in, This indicates the pseudo-degeneration stage of the target domain samples; This represents the soft matching weight for each candidate stage.

[0074] The pseudo-degenerate stage is only used to establish stage consistency constraints; the target domain context itself still uses the complete probability distribution for soft aggregation. This provides discrete class relationships for contrastive learning without losing the uncertainty of the target sample in adjacent stages.

[0075] This implementation method avoids unnecessary matching errors when reliable supervision is available by performing hard context retrieval on source domain samples according to the source domain and degradation stage to which the source domain samples belong. By calculating the cosine similarity between the domain-invariant features of the target domain samples and the prototypes of each degradation stage and performing soft matching, reasonable information of multiple stages is preserved at the stage boundaries or uncertain samples, thereby reducing the semantic jumps and error amplification caused by a single hard pseudo-label.

[0076] Step 106: Generate modulation parameters based on each semantic context, and use the modulation parameters to modulate the domain-invariant features to obtain fused features.

[0077] In step 106, “modulation parameters” is a higher-level summary of the control signal generated in this step, and its specific form is a scale vector and a bias vector; the specific implementation of “modulation” is a feature-level linear transformation (scaling and translation).

[0078] As an optional implementation, the step of generating modulation parameters based on each semantic context and modulating the domain-invariant features using the modulation parameters to obtain fused features specifically includes: The semantic context is input into the scale generation network and the bias generation network respectively to obtain the scale vector and bias vector with the same dimension as the domain invariant feature. The neighborhood-invariant features are scaled element-wise using the sum of the unit vector and the scale vector as a scaling factor, and then superimposed with the bias vector to obtain the fused features.

[0079] In this embodiment, modulation parameters are generated based on the sample semantic context vector. These modulation parameters include a scale vector and a bias vector. Feature-level modulation is performed on the neighborhood-invariant features using these modulation parameters, and a predicted remaining lifetime value is output. The source or target domain semantic context is uniformly denoted as... This context cannot be simply added to domain-invariant features because the feature dimensions that need to be enhanced or suppressed are different at different degradation stages. This embodiment uses two independent multilayer perceptrons to generate scale and bias vectors respectively, enabling graph reasoning knowledge to change the amplitude and baseline of the original representation according to the feature dimensions. The modulation parameters are expressed as: ; in, and These represent the scale vector and bias vector of the feature modulation, respectively; Represents the semantic context of the sample; and This indicates that a network is generated using two parameters.

[0080] The fusion characteristics after modulation are: ; in, Indicates the fusion characteristics after modulation; Represents domain-invariant features; and Control each dimension's scaling and offset separately; This indicates element-wise multiplication.

[0081] The remaining life prediction value is: ; in, This indicates the predicted remaining lifespan of the equipment. This indicates the remaining lifetime prediction module; This indicates the fusion feature.

[0082] The residual scaling design preserves the original domain-invariant features and selectively amplifies or weakens dimensions related to the current domain-stage state using the scaling vector, while the bias vector supplements context-related information. The fused features are then fed into the remaining lifetime prediction module to obtain the predicted value. This forms a closed loop of "feature extraction - graph reasoning - context retrieval - feature modulation - lifetime prediction," where the graph structure is no longer just supplementary supervision but directly participates in the prediction process for each sample.

[0083] In this implementation, the sample semantic context is input into the scale generation network and the bias generation network respectively to obtain the scale vector and bias vector with the same dimension as the domain-invariant feature. This allows graph reasoning knowledge to change the magnitude and baseline of the original representation according to the feature dimension. The original domain-invariant features are preserved through residual scale design, and the dimension related to the current domain-stage state is selectively amplified or weakened by the scale vector, so that the structured prior directly participates in the sample-level prediction.

[0084] Step 107: The spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module are jointly optimized and trained using multiple preset loss functions. The trained remaining lifetime prediction module outputs the remaining lifetime prediction value based on the fusion features of the target domain samples. The preset loss functions include a regression loss constructed based on the fusion features of the source domain samples.

[0085] As an optional implementation, the multiple preset loss functions include remaining lifetime regression loss, domain adversarial loss, domain reconstruction loss, feature orthogonality loss, and stage consistency comparison loss; the joint optimization training of the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module using multiple preset loss functions specifically includes: The remaining lifetime regression loss is constructed using the mean squared error between the predicted remaining lifetime and the actual remaining lifetime of the source domain samples. The domain-invariant features and the domain discriminator are connected by a gradient inversion layer to construct the domain adversarial loss; The domain reconstruction loss is constructed based on the cross-entropy of the real domain number of the sample reconstructed from the domain-specific features by the domain classifier. The feature orthogonal loss is constructed based on the normalized cross-correlation between the domain-invariant feature matrix and the domain-specific feature matrix in the current training batch; Samples with the same degradation stage are constructed as positive sample pairs, and samples with different degradation stages are constructed as negative sample pairs, thus constructing the stage consistency comparison loss. The remaining lifetime regression loss, the domain adversarial loss, the domain reconstruction loss, the feature orthogonality loss, and the stage consistency comparison loss are weighted and summed to obtain the total loss; The total loss is used to jointly optimize and train the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module.

[0086] In this embodiment of the application, the remaining lifetime regression loss is calculated for the source domain samples: ; in, Represents the remaining lifetime regression loss; The number of source domain samples participating in regression supervision; Represents the remaining lifespan return head; Representing the Fusion characteristics of samples from each source domain; This represents the corresponding actual remaining lifespan.

[0087] This loss function uses only source domain samples with true remaining lifetime labels, constraining the remaining lifetime prediction module's output to closely approximate the true lifetime, and serves as the task anchor for the entire joint training process. The purpose of this loss is to ensure that subsequent domain alignment and graph inference always serve remaining lifetime prediction, rather than simply acquiring features that are indistinguishable within the domain but lack degradation information.

[0088] In this embodiment, a domain adversarial loss is constructed using a gradient inversion layer and a domain discriminator: ; in, Representing the field of combating losses; This represents the number of samples in the current small batch; Represents the number of source domains; Representing the The real domain number of each sample; Domain discriminant; Represents a gradient inversion layer; Representing the Domain-invariant features of each sample.

[0089] The neighborhood discriminator attempts to identify the sample source using neighborhood-invariant features, while the gradient inversion layer causes the feature extractor to update in the opposite direction, making the source and target domains difficult to distinguish within the neighborhood-invariant subspace. This adversarial process is used to remove distributional differences caused by rotational speed, load, hard drive model, or workload, but does not directly act on neighborhood-specific branches.

[0090] In this embodiment, the real domain number is reconstructed using domain-specific features: ; in, Representative domain restructuring losses; This represents the number of samples in the current small batch; Representing the The real domain number of each sample; Reconstruct the classifier for the representative domain; Representing the Domain-specific features of each sample.

[0091] Domain reconstruction classifiers require that domain-specific features can correctly recover the domain to which a sample belongs, giving a clear direction to the condition information excluded by the domain-invariant branch. This prevents the feature extractor from deceiving the domain discriminator by discarding the entire input information and provides a stable and interpretable basis for updating domain nodes.

[0092] In this embodiment of the application, in order to reduce the information overlap between domain-invariant features and domain-specific features, a feature orthogonal loss is constructed: ; in, Orthogonal loss representing representative features; This represents the matrix composed of the neighborhood-invariant features in the current mini-batch; A matrix representing the domain-specific features in the current mini-batch; This represents the Frobenius norm.

[0093] Adversarial loss and reconstruction loss constrain the two branches from the perspective of classifiability, while orthogonal loss further reduces the correlation between the two feature matrices from a geometric perspective. The three work together to ensure that the domain-invariant branch focuses on expressing the degradation process, while the domain-specific branch focuses on expressing the differences in working conditions, reducing the repetition of the same information in the two subspaces.

[0094] In this embodiment of the application, the consistency comparison loss for the sample construction stage in the same degradation stage is as follows: ; in, Representative stage consistency comparison loss; This represents the set of anchor point samples participating in the comparative learning; Representative and anchor sample A set of positive samples at the same stage of degradation; Representative anchor point sample The set of candidate comparison samples; Represents invariant characteristics of the domain; Represents the cosine similarity function; Represents the temperature coefficient.

[0095] The formula for calculating cosine similarity is: ; in, Representative vector with vector Cosine similarity; and This represents two eigenvectors to be compared.

[0096] The source domain samples are used to construct a positive sample set based on the true degradation stage, while the target domain samples are used to construct the set based on the pseudo-degradation stage obtained in step 105. Cross-domain samples at the same stage are brought closer together, while samples at different stages are pushed further apart, ensuring that the features after domain alignment still maintain clear and orderly stage boundaries along the healthy to failed direction. The stage consistency contrast loss simultaneously constrains samples belonging to the same degradation stage in different domains to move closer to each other in the domain-invariant feature space, and constrains samples belonging to different degradation stages to move further apart in the same domain-invariant feature space.

[0097] In this embodiment, the remaining lifetime regression loss, domain adversarial loss, domain reconstruction loss, feature orthogonality loss, and stage consistency comparison loss are weighted and summed, and the model parameters are updated; the overall optimization objective is: ; in, Represents the total loss; , , , , These represent the remaining lifetime regression loss, domain adversarial loss, domain reconstruction loss, feature orthogonality loss, and stage consistency comparison loss, respectively.

[0098] Update all trainable parameters using a gradient-based optimizer: ; in, Represents all trainable parameters of the model; Represents the learning rate; This represents the gradient of the total loss with respect to the model parameters.

[0099] Each weight is used to balance prediction accuracy, domain invariance, information integrity, and stage separability. During backpropagation, the spatiotemporal feature extraction and decoupling, basic concept graph propagation, combinatorial mapping, context interaction graph inference, feature modulation, and remaining lifetime prediction modules are updated end-to-end; the basic concept nodes are calibrated according to the momentum method in step 103, in addition to gradient updates. After multiple iterations, the model simultaneously obtains cross-domain transferable features and structured degradation context.

[0100] In this implementation, the remaining lifetime regression loss provides a lifetime prediction benchmark, the domain adversarial loss removes condition clues from domain-invariant features, the domain reconstruction loss ensures that separated domain information is preserved, the feature orthogonality loss reduces redundant information between the two subspaces, and the stage consistency loss maintains cross-domain degenerate semantics. The five losses used together avoid the loss of fine-grained information related to lifetime prediction by pursuing domain alignment alone.

[0101] The method developed in this application will be tested numerically on bearing performance degradation datasets and hard disk performance degradation datasets.

[0102] In one embodiment, a unified hidden dimension is achieved. Domain-invariant feature dimensions Domain-specific feature dimensions All values ​​are set to 64; batch size is set to 64; maximum number of training rounds is set to 100; momentum coefficient is set to 64. Set to 0.9; soft-matching temperature coefficient and comparative learning temperature coefficient All can be set to 0.5; initial learning rate Set to 0.001; Domain adversarial loss weight Set to 0.1; Domain Reconstruction Loss Weight Feature orthogonal loss weights Compared with stage consistency, the loss weight All values ​​are set to 0.05. These values ​​are intended to illustrate a feasible configuration and are not intended to limit the scope of protection.

[0103] In the bearing implementation, the IEEE PHM2012 and XJTU-SY datasets were used. The IEEE PHM2012 dataset includes three combinations of speed and radial load; the XJTU-SY dataset includes three different combinations of speed and load. Four bearing instances from two of these load conditions were used as multiple source domains, and bearing instances from the remaining load condition were used as unlabeled target domains, constituting transfer tasks A to F. The number of bearing input feature channels... Set to 8, time window length Set to 32, the total number of degradation stages. Set it to 8.

[0104] The prediction curve of the method in this application is closer to the actual remaining life curve over the entire degradation lifespan, as shown in the example. Figure 5 As shown in (a)-(f), the prediction fluctuations and overestimations in the later degradation stage are suppressed. Taking the migration task D1 as an example, the root mean square error of the method in this application is 0.0781, which is about 12.7% lower than the second-best comparative method with a root mean square error of 0.0895.

[0105] like Figure 7 As shown in (a)-(d) in the paper, when coloring by domain, the separation degree between source domain samples and target domain samples in the feature space is reduced after adopting the method of this application; when coloring by degradation stage, samples of the same degradation stage form a more compact cluster and form a continuous semantic trajectory from the healthy stage to the failure stage, indicating that hierarchical semantic graph reasoning and feature-level modulation can simultaneously promote domain alignment and stage semantic preservation.

[0106] In this implementation, experimental verification on a bearing dataset demonstrates that the proposed method can maintain a low prediction error throughout the entire degradation lifecycle, and the prediction fluctuations in the later degradation stage are effectively suppressed. Hierarchical semantic graph reasoning and feature-level modulation can simultaneously promote domain alignment and stage semantic preservation.

[0107] In one hard disk drive embodiment, Backblaze hard disk monitoring data from 2023 to 2024 was used. Twelve SMART monitoring attributes were retained, including read error rate, number of reallocated sectors, seek error rate, power-on time, number of reported uncorrectable errors, power-down rollback count, load / unload loop count, temperature, number of currently pending sectors, number of uncorrectable sectors, cumulative number of written logical block addresses, and cumulative number of read logical block addresses. Migration tasks G, H, and I were formed using ST4000DM000 and ST8000DM002 models as two source domains and ST6000DX000, ST4000DM005, and ST8000DM005 models as three target domains. The number of hard disk input feature channels was also considered. Set to 12, time window length Set to 30, the total number of degradation stages. Set it to 5.

[0108] like Figure 6 As shown in (a)-(f), the method of this application can generate a smooth prediction curve that closely approximates the actual degradation trend under different hard drive models, capacities, rotation speeds, and workloads. In migration task G, the root mean square error of the method of this application is 0.0270, which is about 29.1% lower than the second-best comparative method with a root mean square error of 0.0381. This indicates that the spatiotemporal feature decoupling and hierarchical semantic graph reasoning can effectively handle the correlation between heterogeneous SMART attributes and the domain offset between hard drive models.

[0109] like Figure 10As shown in (a)-(c), during the healthy phase, the cumulative number of written logical block addresses receives high attention; as degradation progresses, the importance of the current number of pending sectors, the number of uncorrectable sectors, and the read error rate increases; during the failure phase, power-on time receives high attention. Therefore, the method of this application can change the degree of attention paid to different monitoring attributes according to the degradation stage, and the resulting stage-perceived attention is consistent with the physical degradation pattern of the hard disk.

[0110] In particular, by implementing this method and conducting experiments on hard disk drive datasets, it was demonstrated that the method of this application can effectively handle the correlation between heterogeneous SMART attributes and the domain offset between hard disk models under different hard disk models, capacities, speeds and workloads. Furthermore, the obtained stage-aware attention is consistent with the physical degradation law of hard disks, indicating that the method of this application has good interpretability.

[0111] Figure 8 Figures (a) and (b) show the relative performance improvement of this application compared to the suboptimal method (DWDAN method (Domain Weighted Distribution Adaptation Network)) on twelve bearing migration tasks. Figure 8 The paper demonstrates the performance improvement of the proposed method compared to the suboptimal methods across all twelve bearing migration tasks. Figure 8 As can be seen, the method of this application achieves better prediction accuracy than the suboptimal method on all migration tasks, with significant performance improvements on some tasks, indicating that the method of this application has good applicability and stability in different bearing operating condition migration scenarios.

[0112] Figure 9 Figures (a) and (b) show the average performance comparison of different multi-source domain adaptation methods on the entire dataset. Figure 9 Lieutenant General's method of this application (i.e. Figure 9 The CIGDA method in this paper is compared with various existing multi-source domain adaptation methods, including distribution alignment-based methods, adversarial training-based methods, and graph neural network-based methods. Figure 9 As can be seen, the average performance of the proposed method is better than that of the comparative methods on all evaluation datasets, indicating that the proposed method can effectively utilize transferable degradation knowledge in multi-source domains through hierarchical semantic graph reasoning and structured domain adaptation mechanisms, and obtain more accurate remaining lifetime prediction results under the condition of unlabeled target domain.

[0113] By implementing this method and comparing its comprehensive performance across multiple migration tasks, the superiority of the proposed method over existing multi-source domain adaptation methods was verified, demonstrating that hierarchical semantic graph reasoning and structured domain adaptation mechanisms can effectively improve the accuracy and stability of cross-condition remaining lifetime prediction.

[0114] To verify the effectiveness of each component of the model, the following model variants are constructed: Model A removes feature modulation guided by the context interaction graph; Model B replaces feature-level modulation with linear addition; Model C removes stage consistency contrastive loss; Model D removes domain adversarial loss, domain reconstruction loss, and feature orthogonality loss. The complete method in this application simultaneously retains context interaction graph inference, feature-level modulation, stage consistency constraints, and feature decoupling constraints. Figure 11 As shown in (a)-(c), in Task D, the complete method reduces the root mean square error (RMSE) from 0.1847 to 0.0781 compared to Model A, which removes graph-guided context modulation; Model B, which replaces feature-level modulation with linear addition, has an RMSE of 0.0919, higher than the complete method; in Task E, Model D, which removes all feature decoupling losses, achieves an RMSE of 0.1938, while the complete method is 0.0857. These results demonstrate that context-interactive graph inference, feature-level modulation, stage consistency, and feature decoupling work together to predict cross-domain remaining lifetime.

[0115] In this implementation, the role of each component of the model was verified through ablation experiments, demonstrating that context interaction graph reasoning, feature-level modulation, stage consistency constraints, and feature decoupling constraints all contribute to the predictive performance of the method in this application. The combination of various technical features can synergistically improve the accuracy of cross-domain remaining lifetime prediction.

[0116] It should be noted that the number of layers and specific neurons in the graph attention network, temporal coding network, multilayer perceptron, domain discriminator, and remaining lifetime prediction module described in this specification can be adjusted according to the device type and computing resources; the total number of degradation stages, time window length, feature dimensions, and weights of each loss can also be determined through validation data. Any approach that uses domain nodes and degradation stage nodes to construct a basic concept graph, forms a domain-stage context interaction graph through combination mapping, retrieves or aggregates semantic context based on stage prototypes, and modulates domain-invariant features using the semantic context falls within the scope of the technical concept of this application.

[0117] By implementing steps 101 to 107 above, this application provides a method and apparatus for predicting the remaining lifetime of a device based on hierarchical semantic graph reasoning. By acquiring samples from multiple source and target domains and labeling the source domain samples with remaining lifetime tags, a data foundation is provided for subsequent cross-domain transfer learning, making it possible to predict remaining lifetime even when the target domain is unlabeled. By extracting and decoupling spatiotemporal features from all samples to obtain domain-invariant and domain-specific features, it is possible to effectively capture the spatial coupling relationship between sensors and the long-range degradation dependency in the temporal dimension, while suppressing the interference of operating noise on cross-domain degradation knowledge transfer. The remaining lifetime tags of the source domain samples are discretized to obtain the remaining lifetime tags of each source domain. The model extracts degradation stage information from samples and constructs a basic concept graph containing domain nodes and degradation stage nodes based on domain information and degradation stage information. Then, it updates both types of nodes based on domain-specific and domain-invariant features and propagates information on the basic concept graph. This allows the model to explicitly organize the structured semantic relationships between the domain and degradation stages, enabling stage nodes to absorb continuous degradation information from adjacent stages and shared information from various domains, providing stable and referential anchors for the unlabeled target domain. By combining and mapping the propagated domain nodes with degradation stage nodes to generate composite nodes, a context interaction graph is constructed based on these composite nodes, and graph attention inference is performed to obtain context node representations. This allows the domain to adapt from flat statistical alignment expansion. The method expands to structured semantic reasoning with degradation order and domain relationships, reducing misalignment between different degradation stages. It retrieves the corresponding semantic context for source domain samples from context node representations based on domain and degradation stage information, and aggregates the corresponding semantic context for target domain samples based on the similarity between domain-invariant features and nodes at each degradation stage. This enables labeled source domains to perform deterministic context retrieval using real-world supervised information, while providing soft matching criteria for unlabeled target samples using degradation stage prototypes, avoiding single-point hard matching based solely on unstable pseudo-labels. Furthermore, it generates modulation parameters based on each semantic context and utilizes these parameters to modify domain-invariant features. Modulation is performed to obtain fused features, and the semantic context obtained by graph reasoning is fed back to the domain-invariant features through feature-level modulation, forming a closed loop of "feature extraction - graph reasoning - context retrieval - feature modulation - lifetime prediction", which allows structured priors to directly participate in sample-level prediction. By using multiple preset loss functions for joint optimization training, and using the trained lifetime prediction module to output the lifetime prediction value based on the fused features of the target domain samples, the preset loss functions include a regression loss constructed based on the fused features of the source domain samples. This allows the model to maintain the separability of the degradation stage while pursuing domain alignment, avoiding negative transfer, thereby improving the accuracy, robustness and interpretability of lifetime prediction across operating conditions as a whole.

[0118] Based on the same inventive concept, this application also provides a device for predicting the remaining lifetime of a device based on hierarchical semantic graph reasoning, which is used to implement the device remaining lifetime prediction method based on hierarchical semantic graph reasoning described above. The solution provided by this device is similar to the solution described in the above method. Therefore, the specific limitations of one or more embodiments of the device for predicting the remaining lifetime of a device based on hierarchical semantic graph reasoning provided below can be found in the limitations of the device remaining lifetime prediction method based on hierarchical semantic graph reasoning described above, and will not be repeated here.

[0119] In one exemplary embodiment, such as Figure 12 As shown, a device for predicting remaining lifetime based on hierarchical semantic graph reasoning is provided, comprising: The data acquisition module 201 is used to acquire samples from multiple source domains and target domains, wherein the samples from the source domains are labeled with remaining lifetime. The spatiotemporal feature extraction and decoupling module 202 is used to extract and decouple spatiotemporal features from all the samples to obtain domain-invariant features and domain-specific features. The graph construction and update module 203 is used to discretize the remaining lifetime labels of the samples in the source domain to obtain degradation stage information of the samples in each source domain; based on the domain information and the degradation stage information of each sample, a basic concept graph containing domain nodes and degradation stage nodes is constructed; the domain nodes are updated based on the domain-specific features, the degradation stage nodes are updated based on the domain-invariant features, and all updated nodes are propagated on the basic concept graph; Graph reasoning module 204 is used to combine and map the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; construct a context interaction graph based on the composite nodes, and perform graph attention reasoning on the context interaction graph to obtain context node representations; The context retrieval module 205 is used to retrieve the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and to aggregate the corresponding semantic context for the target domain sample based on the similarity between the domain invariant features of the target domain sample and each degradation stage node. The feature modulation module 206 is used to generate modulation parameters according to each of the semantic contexts, and use the modulation parameters to modulate the domain-invariant features to obtain fused features; The joint optimization module 207 is used to perform joint optimization training on the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph and the remaining lifetime prediction module using multiple preset loss functions. The remaining lifetime prediction module 208 is used to output the remaining lifetime prediction value based on the fusion features of the target domain samples using the trained remaining lifetime prediction module.

[0120] In one exemplary embodiment, a computer device is provided, which may be a server or a terminal, and its internal structure diagram may be as follows. Figure 13 As shown, the computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operating system and computer programs in the non-volatile storage media to run. The database stores device remaining lifetime prediction data based on hierarchical semantic graph reasoning. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a device remaining lifetime prediction method based on hierarchical semantic graph reasoning.

[0121] Those skilled in the art will understand that Figure 13 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0122] In one exemplary embodiment, a computer device is also provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0123] In one exemplary embodiment, a computer-readable storage medium is provided storing a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0124] In one exemplary embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above-described method embodiments.

[0125] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data must comply with relevant regulations.

[0126] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM).

[0127] The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0128] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0129] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the methods and core ideas of this application. Furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the remaining life of equipment based on hierarchical semantic graph reasoning, characterized in that, The device remaining lifetime prediction method based on hierarchical semantic graph reasoning includes: Acquire samples from multiple source and target domains, wherein the samples from the source domains are labeled with remaining lifetime. Spatiotemporal feature extraction and decoupling are performed on all the samples to obtain domain-invariant features and domain-specific features; The remaining lifetime labels of the samples in the source domain are discretized to obtain degradation stage information of the samples in each source domain; based on the domain information and the degradation stage information of each sample, a basic concept graph containing domain nodes and degradation stage nodes is constructed; the domain nodes are updated based on the domain-specific features, the degradation stage nodes are updated based on the domain-invariant features, and all updated nodes are propagated on the basic concept graph; The propagated domain nodes and the propagated degenerate stage nodes are combined and mapped to generate composite nodes representing the joint state of the domain and the stage; a context interaction graph is constructed based on the composite nodes, and graph attention reasoning is performed on the context interaction graph to obtain the context node representation; Based on the domain information and degradation stage information of the source domain sample, the corresponding semantic context is retrieved from the context node representation for the source domain sample; based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, the corresponding semantic context is aggregated for the target domain sample. Modulation parameters are generated based on the semantic contexts, and the domain-invariant features are modulated using the modulation parameters to obtain fused features; wherein, the modulation parameters include a scale vector and a bias vector; The spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module are jointly optimized and trained using multiple preset loss functions. The trained remaining lifetime prediction module outputs the remaining lifetime prediction value based on the fusion features of the target domain samples. The preset loss functions include a regression loss constructed based on the fusion features of the source domain samples.

2. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The process of extracting and decoupling spatiotemporal features from all the samples specifically includes: The sensor channels at each sampling time are treated as graph nodes. The association weights between different sensor channels are calculated and neighborhood information is aggregated through the spatial graph attention mechanism to obtain the spatially enhanced sequence. The spatially enhanced sequence is fed into a temporal coding network after adding positional encoding. Global average pooling is then performed on the output of the temporal coding network to obtain the encoded features. The temporal coding network includes a multi-head self-attention sub-layer and a feedforward network sub-layer. The encoded features are mapped to a domain-invariant feature space and a domain-specific feature space through two independent mapping networks, respectively.

3. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The basic concept diagram includes: Domain nodes that correspond one-to-one with each source domain and target domain; Each discrete degradation stage corresponds to a degradation stage node, and adjacent degradation stage nodes are connected in the order of the equipment degradation process to form a stage manifold chain. Fully connected edges between any two domain nodes, and cross-type connected edges between each domain node and each degeneration stage node.

4. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 3, characterized in that, The process of updating the domain nodes based on domain-specific features, updating the degradation stage nodes based on domain-invariant features, and propagating the updated nodes across the basic concept graph specifically includes: The mean of domain-specific features of samples belonging to the same domain in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the domain node according to a preset momentum coefficient to complete the update of the domain node; The mean of the domain-invariant features of samples belonging to the same degradation stage in the current training batch is calculated, and the calculated mean is weighted and fused with the historical representation of the degradation stage node according to the preset momentum coefficient to complete the update of the degradation stage node. Based on the edge connections of the basic concept graph, neighborhood information is aggregated for all updated nodes to complete the information propagation.

5. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The propagated domain node and the propagated degradation stage node are combined and mapped to generate a composite node representing the joint state of the domain and the stage. A context interaction graph is constructed based on the composite nodes, and graph attention reasoning is performed on the context interaction graph to obtain the context node representation, specifically including: The representation of each propagated domain node is concatenated with the representation of each propagated degenerate stage node, and the concatenation result is projected to a unified dimension through a nonlinear mapping layer to generate a domain-stage composite node. Establish fully connected edges between all the composite nodes to form the context interaction graph; A multi-head graph attention network is used to calculate the attention weight between any two composite nodes. The information of adjacent nodes is aggregated according to the attention weight, and the context node representation is obtained through residual connection.

6. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The step of retrieving the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and aggregating the corresponding semantic context for the target domain sample based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node, specifically includes: For a source domain sample, based on the source domain to which the source domain sample belongs and the degradation stage of the source domain sample, the corresponding context node representation is directly indexed from the context node representation, and the indexed context node representation is used as the semantic context of the source domain sample. For a target domain sample, the cosine similarity between the domain-invariant features of the target domain sample and the node representations of each degradation stage is calculated. The cosine similarity after temperature coefficient scaling is normalized exponentially to obtain the soft matching weights of each degradation stage. The soft matching weights are then used to perform a weighted summation on all context node representations corresponding to the target domain, and the weighted summation result is used as the semantic context of the target domain sample.

7. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The step of generating modulation parameters based on each semantic context and using the modulation parameters to modulate the domain-invariant features to obtain fused features specifically includes: The semantic context is input into the scale generation network and the bias generation network respectively to obtain the scale vector and bias vector with the same dimension as the domain invariant feature. The neighborhood-invariant features are scaled element-wise using the sum of the unit vector and the scale vector as a scaling factor, and then superimposed with the bias vector to obtain the fused features.

8. The equipment remaining life prediction method based on hierarchical semantic graph reasoning according to claim 1, characterized in that, The multiple preset loss functions include remaining lifetime regression loss, domain adversarial loss, domain reconstruction loss, feature orthogonality loss, and stage consistency comparison loss; the joint optimization training of the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module using multiple preset loss functions specifically includes: The remaining lifetime regression loss is constructed using the mean squared error between the predicted remaining lifetime and the actual remaining lifetime of the source domain samples. The domain-invariant features and the domain discriminator are connected by a gradient inversion layer to construct the domain adversarial loss; The domain reconstruction loss is constructed based on the cross-entropy of the real domain number of the sample reconstructed from the domain-specific features by the domain classifier. The feature orthogonal loss is constructed based on the normalized cross-correlation between the domain-invariant feature matrix and the domain-specific feature matrix in the current training batch; Samples with the same degradation stage are constructed as positive sample pairs, and samples with different degradation stages are constructed as negative sample pairs, thus constructing the stage consistency comparison loss. The remaining lifetime regression loss, the domain adversarial loss, the domain reconstruction loss, the feature orthogonality loss, and the stage consistency comparison loss are weighted and summed to obtain the total loss; The total loss is used to jointly optimize and train the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module.

9. The method for predicting remaining equipment lifetime based on hierarchical semantic graph reasoning according to claim 8, characterized in that, The degradation stage of the target domain sample is determined by the degradation stage corresponding to the maximum weight in the soft matching weight; the stage consistency comparison loss simultaneously constrains samples belonging to the same degradation stage in different domains to move closer to each other in the domain-invariant feature space, and constrains samples belonging to different degradation stages to move further apart in the domain-invariant feature space.

10. A device for predicting remaining lifespan of an equipment based on hierarchical semantic graph reasoning, characterized in that, The device remaining lifetime prediction device based on hierarchical semantic graph reasoning includes: The data acquisition module is used to acquire samples from multiple source domains and a target domain, wherein the samples from the source domains are labeled with remaining lifetime. The spatiotemporal feature extraction and decoupling module is used to extract and decouple spatiotemporal features from all the samples to obtain domain-invariant features and domain-specific features. The graph construction and update module is used to discretize the remaining lifetime labels of the samples in the source domain to obtain degradation stage information of the samples in each source domain; based on the domain information and the degradation stage information of each sample, a basic concept graph containing domain nodes and degradation stage nodes is constructed; the domain nodes are updated based on the domain-specific features, the degradation stage nodes are updated based on the domain-invariant features, and all updated nodes are propagated on the basic concept graph; The graph reasoning module is used to combine and map the propagated domain nodes with the propagated degenerate stage nodes to generate composite nodes representing the joint state of the domain and stage; construct a context interaction graph based on the composite nodes, and perform graph attention reasoning on the context interaction graph to obtain the context node representation; The context retrieval module is used to retrieve the corresponding semantic context for the source domain sample from the context node representation based on the domain information and degradation stage information of the source domain sample; and to aggregate the corresponding semantic context for the target domain sample based on the similarity between the domain-invariant features of the target domain sample and each degradation stage node. The feature modulation module is used to generate modulation parameters based on each semantic context, and to modulate the domain-invariant features using the modulation parameters to obtain fused features; The joint optimization module is used to perform joint optimization training on the spatiotemporal feature extraction and decoupling process, the basic concept graph, the context interaction graph, and the remaining lifetime prediction module using multiple preset loss functions. The remaining lifetime prediction module is used to output the remaining lifetime prediction value based on the fusion features of the target domain samples after training.

Images