A topological perception space-time graph fusion automobile energy system life prediction method

Abstract

The present application belongs to the technical field of automobile active system health management and residual life prediction, and particularly relates to a kind of topological perception space-time graph fusion automobile active system life prediction method.The method obtains the multi-sensor time sequence monitoring signal of the predicted component in the automobile active system and slices into input sequence, introduces a learnable global node to obtain a space-time enhanced input representation, dynamically generates a time-dependent graph and a space-dependent graph;In the time module, bidirectional state space scanning and graph propagation are performed in parallel, and in the space module, bidirectional state space scanning and graph propagation are performed in parallel after sorting the sensor node spectrum, the two modules are executed in turn and repeated for a predetermined number of space-time graph state space blocks, and a space-time degradation representation is obtained;Input the prediction head and the projection head, and output the residual life prediction value and the health factor sequence constrained by physical consistency.The present application combines time sequence dependence and spatial coupling relationship for life prediction, and outputs an interpretable health factor to assist in analyzing the degradation trend.

Description

Technical Field

[0001] This invention belongs to the field of automotive power system health management and remaining life prediction technology, specifically involving a method for predicting the life of automotive power systems based on topological sensing spatiotemporal graph fusion. Background Technology

[0002] Key components of automotive power systems, such as automotive batteries, engines, drive motors, and transmissions, are typically subjected to complex operating conditions such as high temperature, high pressure, variable load, vibration, shock, and periodic charging and discharging during actual service. As operating time increases, internal materials, structural components, and functional units gradually age, wear down, and degrade, adversely affecting system reliability. Therefore, predicting the remaining service life of equipment provides crucial information for maintenance planning, mitigating the risk of sudden downtime, and ensuring operational safety.

[0003] Existing methods for predicting remaining lifetime still have certain limitations. Traditional data-driven methods typically rely on manually constructed degradation features or pre-defined statistical indicators, making it difficult to fully characterize the long-range temporal correlations in multi-sensor signals and the dynamic changes throughout the degradation process. Prediction results are easily affected when equipment is under complex operating conditions or when signal noise is strong. While some deep learning methods can automatically extract features, they usually treat multi-sensor data as ordinary time-series matrices, failing to adequately model the correlations between sensors, and particularly struggling to reflect the dynamically changing sensor coupling structures during degradation.

[0004] When processing long window sequences, attention-based temporal modeling methods often involve high computational complexity. Directly modeling fully connected relationships across all time steps or sensor nodes may weaken the representation of local physical correlations and introduce order bias due to fixed node arrangements. For equipment degradation processes driven by the coupling of multiple components and physical quantities, these shortcomings limit the model's ability to characterize the true degradation mechanism. Equipment degradation is usually not a completely disordered process; its health status declines over time. However, existing end-to-end models often directly provide remaining lifetime predictions, lacking auxiliary health factors corresponding to the degradation process, making it difficult to explain the degradation state changes upon which the prediction results are based. Based on these issues, it is necessary to propose a remaining lifetime prediction method that can simultaneously handle long-range temporal dependencies, dynamic sensor topology relationships, and the evolution characteristics of health factors. Summary of the Invention

[0005] The purpose of this invention is to provide a method for predicting the lifespan of automotive active systems based on topological sensing spatiotemporal graph fusion, in order to solve the technical problems of existing methods for predicting the remaining lifespan of key components of automotive active systems, which are unable to simultaneously characterize long-range temporal dependencies, dynamic sensor coupling structures, and insufficient interpretability of degradation processes in multi-sensor degradation data.

[0006] The present invention achieves the above objectives through the following technical solutions: A method for predicting the lifespan of a vehicle's active power system based on topology-aware spatiotemporal graph fusion includes: The multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system are acquired. The multi-sensor time-series monitoring signals include at least one of temperature, pressure, speed, voltage, current, capacity, internal resistance or vibration signals that characterize the operating state and degradation state of the component to be predicted. The signals are normalized and sliced ​​by a sliding window to obtain a multivariable time-series input sequence. By introducing learnable global time nodes and global spatial nodes into the multivariate temporal input sequence, a spatiotemporally enhanced input representation is obtained; Based on the spatiotemporal enhancement input representation, the original time step and global time node are used together as time graph nodes, and the original sensor node and global spatial node are used together as spatial graph nodes. Pearson correlation coefficients are calculated respectively, and time-related adjacency matrices and spatial-related adjacency matrices are dynamically generated. In each spatiotemporal graph state space block, a time module is executed according to the time-related adjacency matrix to obtain the time module output; after transposing the time module output by dimension, a spatial module is executed according to the spatial-related adjacency matrix to obtain the spatial module output; The temporal and spatial modules respectively perform gated fusion of bidirectional state space scanning features and correlation-weighted graph propagation features within their respective modules; the temporal and spatial modules execute and repeat a preset number of spatiotemporal graph state space blocks in sequence to obtain spatiotemporal degradation representations. The spatiotemporal degradation characterization is input into the remaining lifespan prediction head and the health factor projection head, respectively, and the remaining lifespan prediction value and auxiliary health factor sequence of the component to be predicted are output. Physical consistency constraints are applied to the auxiliary health factor sequence.

[0007] As a further optimization, the component to be predicted is selected from at least one of the following: automotive power battery, automotive engine, drive motor, transmission, and other components in the automotive power system that have a monitorable degradation process; when the component to be predicted is an automotive power battery, the multi-sensor time-series monitoring signal includes one or more of voltage, current, capacity, internal resistance, or temperature; when the component to be predicted is an automotive engine, drive motor, or transmission, the multi-sensor time-series monitoring signal includes one or more of temperature, pressure, speed, or vibration.

[0008] As a further optimization scheme, the time-related adjacency matrix and the spatially related adjacency matrix are obtained through the following steps: Using the original time step and the global time node together as time graph nodes, the Pearson correlation coefficient between the feature vectors of different time graph nodes is calculated to obtain the time-related adjacency matrix; Using both the original sensor nodes and global spatial nodes as spatial graph nodes, the Pearson correlation coefficient between feature sequences of different spatial graph nodes is calculated to obtain the spatial correlation adjacency matrix. The time-dependent adjacency matrix and the spatially dependent adjacency matrix are both dynamically calculated and generated based on the current input window, and are used to characterize the window-level time-varying degradation relationship.

[0009] As a further optimization, the time-series memory features are obtained through the following steps: Selective state recursive scanning is performed along the forward and reverse directions of the time series to obtain bidirectional time memory features; Based on the aforementioned temporal correlation adjacency matrix, correlation-weighted graph propagation is performed on the enhanced temporal location features to obtain temporal graph topological features; The time series memory feature is obtained by fusing the bidirectional time memory feature with the time map topology feature.

[0010] As a further optimization, the spatial sequence memory feature is obtained through the following steps: Top-k sparsification is performed on the spatially related adjacency matrix to retain connections with strong absolute correlation for each node; Inject weak connections into the sparsed spatial correlation adjacency matrix to enhance connectivity; A Laplacian matrix is ​​constructed based on the absolute weights of the spatially correlated adjacency matrix after connectivity enhancement. The Fiedler vector of the Laplacian matrix is ​​solved, and the nodes are sorted according to the magnitude of their components in the Fiedler vector to determine the arrangement order of the spatial nodes. The node with the highest degree centrality is selected as the anchor point, and the direction of the Fiedler vector is normalized by sign. A bidirectional state space scan is performed according to the sorted node sequence, and the original spatial node index order is restored by inverse permutation after the scan is completed to obtain the spatial sequence memory feature.

[0011] As a further optimization scheme, the spatiotemporal degradation characterization is obtained through the following steps: In the time module, the learnable time gating parameters are mapped to the (0,1) interval using the Sigmoid function, and the obtained time gating coefficients are used to perform weighted fusion of bidirectional time memory features and time graph topological features. The output of the time module is transposed and then input into the spatial module. The learnable spatial gating parameters are mapped to the (0,1) interval by the Sigmoid function, and the obtained spatial gating coefficients are used to perform weighted fusion of bidirectional spatial memory features and spatial graph topological features. The time module and the space module are executed sequentially, and the spatiotemporal graph state space block is repeated a preset number of times. The output of the last spatiotemporal graph state space block is used as the spatiotemporal degradation representation.

[0012] As a further optimization scheme, the correlation-weighted graph propagation nonnegates and rows normalizes the temporal or spatial correlation adjacency matrix constructed based on the Pearson correlation coefficient, and uses the normalized graph propagation weight matrix to perform weighted aggregation and linear mapping on the input features of the current module to obtain the corresponding graph topology features.

[0013] As a further optimization, the step of mapping the spatiotemporal degradation characterization to an auxiliary health factor sequence includes: The high-dimensional hidden states corresponding to each enhanced time position in the spatiotemporal degradation representation are mapped to one-dimensional scalar values ​​by the health factor projection head, and the scalar values ​​are constrained to a preset range by the Sigmoid function to obtain the health factor sequence within the window.

[0014] As a further optimization scheme, the physical consistency constraints include: End-of-life anchoring constraints are used to bring the health factor values ​​at the end of the enhanced time series closer to the normalized remaining life label. Monotonic degradation constraint is used to penalize non-physical increases in health factor values ​​at adjacent augmentation time locations; Amplitude constraint is used to constrain the preset degradation amplitude between the initial health factor value of the window and the health factor value at the end of the enhanced time series. The monotonic degradation constraint is an asymmetric monotonic penalty with a tolerance boundary, used to generate a penalty when the growth rate of the health factor value at adjacent enhancement time positions exceeds the preset tolerance boundary. The health factor value is a one-dimensional scalar output by the health factor projection head at the corresponding enhancement time position.

[0015] A topology-aware spatiotemporal graph fusion-based vehicle active system life prediction system is applied to implement the above-mentioned vehicle active system life prediction method. The system includes: The preprocessing module is used to acquire multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system, and obtain a multivariable time-series input sequence through normalization processing and sliding window slicing; wherein, the multi-sensor time-series monitoring signals include physical quantity signals characterizing the operating state and degradation state of the component to be predicted; The node enhancement module is used to perform feature mapping on the multivariate temporal input sequence and introduce learnable global time nodes and global spatial nodes to obtain a spatiotemporally enhanced input representation. The graph construction module is used to calculate the Pearson correlation coefficient based on the spatiotemporal enhancement input representation, using the original time step and global time node as time graph nodes, and the original sensor node and global spatial node as spatial graph nodes, and dynamically generate the time-related adjacency matrix and the spatial-related adjacency matrix respectively. The time modeling module is used to perform bidirectional state space scanning and time graph propagation in parallel based on the time-related adjacency matrix, and to obtain the output of the time module by fusing bidirectional time memory features and time graph topology features through learnable time gating parameters; The spatial modeling module is used to perform topology-aware bidirectional state space scanning and spatial graph propagation in parallel based on the spatially related adjacency matrix, and to obtain the spatial module output by fusing bidirectional spatial memory features and spatial graph topological features through learnable spatial gating parameters; The spatiotemporal block iteration module is used to execute the time modeling module and the spatial modeling module in sequence, and use the output of the spatial module after dimensional recovery as the input of the next spatiotemporal graph state space block, and repeat the execution of a preset number of spatiotemporal graph state space blocks, and use the output of the last spatiotemporal graph state space block as the spatiotemporal degradation representation. The lifetime prediction module is used to output the remaining lifetime prediction value of the component to be predicted based on the spatiotemporal degradation characterization. The health factor module is used to perform gradient isolation processing on the spatiotemporal degradation representation, map the hidden states at each enhanced time location to a health factor sequence, apply physical consistency constraints to the health factor sequence, and obtain auxiliary health representation.

[0016] The beneficial effects of this invention are as follows: 1. This invention dynamically constructs temporal and spatial correlation graphs, enabling the correlations between time steps and between sensor nodes to change dynamically with the input window, providing dynamic graph structure information for characterizing degradation states. Compared to a pre-defined fixed topology graph, this invention can depict the changes in temporal dependencies and sensor coupling relationships at different degradation stages, making the model's representation of the degradation process more closely match the actual data distribution.

[0017] 2. This invention models multi-sensor degradation sequences in the temporal dimension using a bidirectional state-space scanning structure. This effectively characterizes long-range dependencies and staged evolutionary features during device degradation, while reducing the computational cost of long sequence modeling. Simultaneously, through graph sparsification, connectivity enhancement, spectral ordering, and anchor point normalization, the spatial correlation graph is topologically reorganized, making the spatial state-space modeling process more closely match the actual sensor correlation structure. Furthermore, by setting learnable gating parameters, the model can adaptively adjust the fusion ratio between state-space scanning features and graph propagation features during training.

[0018] 3. This invention maps latent states to a sequence of health factors by setting up an independent health factor representation branch, and introduces end-point anchoring constraints, monotonicity constraints, and amplitude constraints, enabling the output health factor sequence to reflect the degradation trend of equipment health status over service time. This auxiliary health factor is decoupled from the main remaining life prediction path and is used only as auxiliary information for degradation trend analysis and prediction result interpretation, without changing the remaining life prediction value itself, thus providing an interpretable degradation basis for the model prediction results. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall process of the method for predicting the lifespan of an automotive power system based on topology-sensing spatiotemporal graph fusion provided in an embodiment of the present invention; Figure 2 This is a schematic diagram of the signal acquisition and preprocessing process provided in an embodiment of the present invention; Figure 3 These are the static average spatial diagram, the later window-level dynamic spatial diagram, and their difference comparison diagram provided in the embodiments of the present invention; Figure 4 This is a Top-k topology visualization of the spatial dynamic PCC graph under different degradation stages provided in the embodiments of the present invention; Figure 5 This is an example of the auxiliary health factor degradation curve and terminal anchoring relationship diagram provided in the embodiments of the present invention; Figure 6 This is a comparison diagram of auxiliary health factor sequences within a window at different degradation stages provided in the embodiments of the present invention; Figure 7 This is a graph showing the change in the effective contribution rate of the time-map propagation branch and the spatial-map propagation branch as a function of the degradation process, provided in an embodiment of the present invention. Detailed Implementation

[0020] The term "automotive power system" as used in this specification refers to a system for storing, generating, converting, or transferring driving energy in a vehicle, including automotive power batteries, vehicle engines, drive motors, transmissions, and related functional components. The following description provides specific application scenarios and requirements for this specification, intended to enable those skilled in the art to manufacture and use the contents of this specification. Various partial modifications to the disclosed embodiments will be apparent to those skilled in the art, and the general principles defined herein can be applied to other embodiments and applications without departing from the spirit and scope of this specification. Therefore, this specification is not limited to the embodiments shown, but rather to the widest scope consistent with the claims.

[0021] The terminology used herein is for the purpose of describing particular exemplary embodiments only and is not restrictive. For example, unless the context clearly indicates otherwise, the singular forms “a,” “an,” and “the” used herein may also include the plural forms. When used in this specification, the terms “comprising,” “including,” and / or “containing” mean that the associated integers, steps, operations, elements, and / or components are present, but do not exclude the presence of one or more other features, integers, steps, operations, elements, components, and / or groups, or that other features, integers, steps, operations, elements, components, and / or groups may be added to the system / method.

[0022] Considering the following description, these and other features of this specification, as well as the operation and function of the related components of the structure, and the economy of assembly and manufacture of the parts, can be significantly improved. All of these form part of this specification with reference to the accompanying drawings. However, it should be clearly understood that the drawings are for illustrative and descriptive purposes only and are not intended to limit the scope of this specification. It should also be understood that the drawings are not drawn to scale.

[0023] The flowcharts used in this specification illustrate operations of a system implementation according to some embodiments of this specification. It should be clearly understood that the operations shown in the flowcharts are executed in the order indicated; operations without data dependencies can be executed in parallel. Furthermore, one or more additional operations can be added to the flowcharts. One or more operations can be removed from the flowcharts.

[0024] First Embodiment like Figure 1 and Figure 2 As shown, this embodiment of the invention provides a method for predicting the lifespan of a vehicle's active power system based on topology-aware spatiotemporal graph fusion. The method includes the following steps: Step S1: Obtain multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system. The multi-sensor time-series monitoring signals include at least one of temperature, pressure, speed, voltage, current, capacity, internal resistance or vibration signals that characterize the operating state and degradation state of the component to be predicted. After normalization and sliding window slicing, a multivariable time-series input sequence is obtained.

[0025] Specifically, the component to be predicted is an automotive power battery, automotive engine, drive motor, transmission, or other components in the automotive power system that have a monitorable degradation process. The multi-sensor time-series monitoring signal includes physical parameters and their derived characteristics reflecting the operating and degradation states of the component to be predicted.

[0026] Taking automotive power batteries as an example, the collected multi-sensor time-series monitoring signals include one or more of the following: single cell voltage, battery pack voltage, charging and discharging current, capacity, internal resistance, and temperature; taking automotive engines as an example, the collected multi-sensor time-series monitoring signals include one or more of the following: temperature, pressure, speed, and vibration; taking drive motors or transmissions as examples, the collected multi-sensor time-series monitoring signals include one or more of the following: temperature, speed, and vibration.

[0027] The acquired raw time-series monitoring signals are used to form a raw time-series signal matrix according to a preset sampling frequency or cycle period. Since different sensors have different dimensions and numerical ranges, directly using the raw data for modeling may lead to unstable model training. Therefore, the acquired signals need to be normalized. Normalization methods can include min-max normalization to keep the data distribution of each sensor channel within a similar range; alternatively, standardization or normalization methods based on training set statistical parameters can be used. This embodiment does not limit the specific method used.

[0028] For the normalized multi-channel time-series signal, a sliding window slice is performed according to a preset window length L and sliding step size r to obtain multiple locally degenerate window samples. For the i-th window, its input sample is represented as... Where N represents the number of sensor channels or feature nodes, and L represents the length of the time window. Each window sample corresponds to a remaining lifetime prediction label, which is the remaining lifetime value at the end of the window and is normalized according to the upper limit of the lifespan or failure cycle of the component to be predicted, to obtain a normalized remaining lifetime label. .

[0029] Step S2: Introduce learnable global time nodes and global spatial nodes into the multivariate temporal input sequence to obtain a spatiotemporally enhanced input representation.

[0030] For the window samples First, a linear mapping is performed along the time dimension, then the dimension is transposed, and a linear mapping is performed along the sensor node dimension to obtain the mapped window representation: In the formula, This represents a time-dimensional linear mapping for each sensor sequence. This represents a node-dimensional linear mapping of sensor features at each time location, where represents the _th_... The feature matrix of a window after two linear mappings; Concatenate global time nodes to the time dimension: Then, the global spatial nodes are stitched together to the sensor node dimension: In the formula, This represents a learnable global time point. Represents a learnable global space node. Indicates spatiotemporal augmentation input; This represents the intermediate feature representation obtained by concatenating global time nodes into the time dimension.

[0031] The global time node and global spatial node participate in the Pearson correlation coefficient calculation in step S3. Specifically, the original time step and the global time node together constitute the time graph node, and the original sensor node and the global spatial node together constitute the spatial graph node. The global spatial node and the original sensor node jointly participate in the Top-k sparsification of the spatial correlation adjacency matrix, the construction of the Laplacian matrix, and spectral sorting.

[0032] The global time node is used to aggregate the overall time context information between different time locations within the window; the global spatial node is used to aggregate the global spatial state information between different sensor nodes at the same time location.

[0033] Step S3: Based on the spatiotemporal enhancement input representation, calculate the Pearson correlation coefficient with time location and spatial node as nodes, and dynamically generate the temporal correlation adjacency matrix and the spatial correlation adjacency matrix.

[0034] For the current window sample, the spatiotemporal augmentation input representation includes One original time step and one global time node, and One raw sensor node and one global space node. This indicates the number of original time steps contained in the sliding window. This indicates the number of original sensor nodes or monitoring feature channels.

[0035] This step uses the Pearson correlation coefficient to measure the degree of linear correlation between the feature vectors of two nodes. For two feature vectors whose correlation is to be calculated... and Its Pearson correlation coefficient is expressed as: In the formula, This represents the number of feature components contained in the feature vector of the node whose correlation is to be calculated; in the calculation of the time correlation graph. In spatial correlation map calculation . and They represent the eigenvectors respectively. and Middle Each feature component Indicates the feature component index; and They represent the eigenvectors respectively. and The mean; Represents the L2 norm; This represents a very small positive number used to prevent the denominator from being zero.

[0036] For the time correlation graph, represented by the spatiotemporal enhancement input... Each original time step and a global time node together serve as nodes in the time map. For any two augmented time locations, their feature values ​​are extracted from all original sensor nodes and the global spatial node, forming corresponding node feature vectors. The correlation between these two vectors is then calculated using the Pearson correlation coefficient, yielding the time-related adjacency matrix. The size of the time-dependent adjacency matrix is Its matrix elements represent the correlation strength between two enhanced time positions, used to characterize the dependencies between different time states within the window and between the original time step and the global time node.

[0037] For the spatial correlation map, represented by the spatiotemporal enhancement input... One original sensor node and one global spatial node together form the spatial graph nodes. For any two augmented spatial nodes, their feature sequences at all original time steps and global time steps are extracted to form corresponding node feature vectors. The correlation between the two is then calculated using the Pearson correlation coefficient mentioned above to obtain the spatial correlation adjacency matrix. The size of the spatially related adjacency matrix is Its matrix elements represent the correlation strength between two enhanced space nodes, which are used to characterize the dynamic coupling relationship between the original sensor nodes within the window and between the original sensor nodes and the global space nodes.

[0038] Both the temporally relevant adjacency matrix and the spatially relevant adjacency matrix are dynamically calculated and generated based on the current window samples, rather than using a pre-defined fixed topology. As the operating conditions and degradation stages of the component to be predicted change, the node characteristics and related adjacency matrices corresponding to different window samples change accordingly, thereby characterizing the dynamic evolution of time-dependent relationships and sensor coupling relationships during the device degradation process.

[0039] After calculating the temporally and spatially correlated adjacency matrices, the positional encoding is superimposed onto the spatiotemporally enhanced input representation to obtain the input of the first spatiotemporal graph state space block, and this is denoted as... The positional encoding is not involved in the calculation of the aforementioned Pearson correlation coefficient and related adjacency matrix.

[0040] To verify the dynamic characteristics of the spatial correlation map, this embodiment performs statistical analysis on the spatial correlation under different degradation stages. Specifically, for each window sample, a corresponding spatial correlation adjacency matrix is ​​constructed based on the Pearson correlation coefficient between each sensor channel, and the mean and standard deviation of the absolute values ​​of the off-diagonal elements of the matrix are further calculated. The mean is used to characterize the overall correlation strength between sensors, and the standard deviation is used to reflect the dispersion of the coupling relationships between different sensors.

[0041] As shown in Table 1, as the normalized remaining lifetime label decreased from 0.944 to 0.160, the mean absolute value of the spatial correlation coefficient increased from 0.126 to 0.378, and the standard deviation increased from 0.154 to 0.405. These changes indicate that as equipment degradation deepens, the correlation between sensors strengthens, and the differences between different sensor pairs become more pronounced. Therefore, the window-level spatial correlation diagram can reflect changes in the degradation stage and can serve as input for subsequent spatial diagram state-space modeling.

[0042] Table 1. Statistical results of spatial correlation at different degradation stages Figure 3 The static average spatial graph, the later window-level dynamic spatial graph, and their absolute differences are further illustrated. The static average spatial graph is obtained by averaging the spatial correlation adjacency matrices of multiple window samples, while the later window-level dynamic spatial graph is calculated from a single later window sample. Figure 3 It can be seen that the static average spatial map has a relatively weak overall correlation, while the later window-level dynamic spatial map shows more obvious positive correlation, negative correlation and local block structure. The two differ greatly across multiple sensor pairs, indicating that the static average map weakens the sensor coupling characteristics under specific degradation stages.

[0043] Based on the above results, this invention uses window-level Pearson correlation coefficients to construct spatial correlation maps, so that the spatial topology is updated as the input window changes. Compared with fixed topology maps or global average maps, this is more suitable for expressing sensor coupling relationships under different degradation stages.

[0044] Step S4: Based on the time-related adjacency matrix, perform parallel time-dimensional bidirectional state space scanning and time graph propagation on the input features of the current spatiotemporal graph state space block, and fuse the two types of features through learnable time gating parameters to obtain the time module output.

[0045] According to the above embodiments, the input features of the current spatiotemporal graph state space block enter the bidirectional state space scanning branch and the graph propagation branch, respectively.

[0046] In the bidirectional state-space scanning branch, selective state recursion scanning is performed along both the forward and reverse directions of the time series. Specifically, the forward scan recursively scans from the beginning to the end of the window, capturing the degenerative evolution path from the past to the present; the reverse scan recursively scans from the end of the window back to the beginning, capturing information feedback from the future to the present. The hidden states obtained from the forward and reverse scans are combined to obtain bidirectional temporal memory features.

[0047] In the graph propagation branch, the time-step features are propagated using a correlation-weighted graph based on the temporal correlation adjacency matrix to obtain the temporal graph topological features. For the th The nth window sample, in the nth In each time module, the time-related adjacency matrix is ​​first nonnegated and row-normalized. Then, the resulting time graph propagation weight matrix is ​​used to perform weighted aggregation and linear mapping on the input features of the current time module to obtain the time graph topological features. In the formula, Indicates the window sample number; Indicates the state space block number of the spacetime diagram; Indicates the first Bidirectional selective state space scan operation in each time module; Indicates the first Bidirectional temporal memory features obtained by scanning the forward and reverse state space of a window sample in the time dimension. Indicates the first The time-related adjacency matrix corresponding to each window sample; This indicates the number of enhanced time locations after being added to the global time node; This indicates that the time-dependent adjacency matrix is ​​nonnegated element-wise. This indicates that Softmax normalization is performed along the row direction of the adjacency matrix; This represents the normalized time-plot propagation weight matrix; Indicates the first The first window sample input Features of a time module of a spatiotemporal graph state space block, wherein This indicates the number of space nodes after being added to the global space node; Indicates the first The learnable linear mapping matrix of the graph propagation branches in each time module; This represents the topological features of the time-dependent graph obtained through propagation.

[0048] The bidirectional time memory features and the time graph topology features are weighted and fused using learnable time gating parameters, and then processed by layer normalization, multilayer perceptron and residual connection to obtain the time module output.

[0049] The first The learnable time-gated parameters of the time modules in the state space blocks of the spatiotemporal diagram are mapped to the Sigmoid function. Interval, obtain the time gating coefficient: In the formula, Indicates the first Learnable gating parameters of time modules in a spatiotemporal graph state space block; Represents the Sigmoid function; This represents the corresponding time gating coefficient.

[0050] Based on the aforementioned time gating coefficient, the bidirectional time memory features and time graph topological features are weighted and fused: In the formula, This indicates the characteristics of bidirectional temporal memory; Representing the topological features of the time graph; This represents the gated fusion feature of the time module. When When the value approaches 1, the gating ratio of temporal map topological features in the fusion result is relatively high; when When the value approaches 0, the gating ratio of bidirectional temporal memory features in the fusion result is relatively high.

[0051] The time-gated fusion features are sequentially subjected to layer normalization, multilayer perceptron processing, and random deactivation, and then residually concatenated with the input features of the time module to obtain the output of the time module: In the formula, Indicates the first Input features of a spatiotemporal graph state space block; This indicates the layer normalization operation in the time module; This represents a multilayer perceptron in the time module; Indicates a random deactivation operation; Indicates the first The window sample passes through the first The output after each time module.

[0052] Step S5: Transpose the time module output, perform topology-aware bidirectional state space scanning and spatial graph propagation in parallel according to the spatial correlation adjacency matrix, and fuse the two types of features through learnable spatial gating parameters to obtain the spatial module output.

[0053] First, the spatial correlation adjacency matrix is ​​subjected to Top-k sparsity, that is, for each spatial node, only the node with the strongest absolute correlation is retained. For each of the following connections, set the remaining weakly correlated connections to zero. This represents the number of neighbors retained for each spatial node. The advantage of Top-k sparsity is that it can preserve the main sensor coupling relationships related to the degradation state within the current window and filter out some weakly correlated or redundant connections, thereby reducing subsequent computation and improving modeling efficiency.

[0054] Figure 4 The Top-k topology corresponding to the spatial dynamic PCC graph at different degradation stages is shown. Nodes S1 to S14 represent different sensor nodes, and the lines between nodes represent the highly correlated connections retained after Top-k sparsification of the current window-level spatial correlation graph. The line color distinguishes between positive and negative correlations, and the line thickness characterizes the strength of the Pearson correlation. In the early degradation stage, the Top-k spatial topology connections are relatively dispersed, and the overall edge weight strength is weak. In the middle degradation stage, the edge weight strength increases, and more obvious correlations begin to form between local sensor nodes. In the late degradation stage, highly correlated connections are more pronounced, and some sensor nodes exhibit high connectivity. These results indicate that Top-k sparsification based on the spatial dynamic PCC graph can retain the main sensor coupling relationships related to the degradation state within the current window and filter out some weakly correlated or redundant connections, thus forming a sparse topology more suitable for subsequent spectral ordering and state space scanning. For ease of demonstration, Figure 4 Only the Top-k connections between the original sensor nodes are shown. The actual calculation of the model also includes global spatial nodes.

[0055] Then, weak connections are injected into the sparsified spatially correlated adjacency matrix to enhance connectivity. It's understandable that if the sparsified graph has breaks, subsequent spectral sorting operations cannot obtain effective Fiedler vectors. Connectivity is ensured by adding edges with minimal weights to the sparsified adjacency matrix.

[0056] Next, a degree matrix and a Laplacian matrix are constructed based on the absolute weights of the spatially correlated adjacency matrix after connectivity enhancement. A pre-defined diagonal regularization term is added to the Laplacian matrix to improve the numerical stability of the eigenvalue decomposition. The Fiedler vector (i.e., the eigenvector corresponding to the second smallest eigenvalue) of this Laplacian matrix is ​​then solved. The nodes are sorted according to the magnitude of their components in the Fiedler vector to determine the spatial node arrangement order.

[0057] It is important to emphasize that, due to the sign uncertainty of the Fiedler vector, the node with the highest degree centrality is selected as the anchor point. When the component of the anchor point in the Fiedler vector is less than the mean of all components of the Fiedler vector, the sign of the Fiedler vector is flipped, so that the spectral sorting direction of samples from different windows has a unified reference. This ensures that the sorting results of samples from different windows have a consistent reference benchmark.

[0058] Based on the sorted node sequence, the temporal module output (i.e., the spatial module input) after dimensional transpose is rearranged according to the new spatial node order, and a bidirectional state space scan is performed along this node sequence. Selective state recursive scans are performed along both the forward and reverse directions of the sorted node sequence to obtain bidirectional spatial memory features. After the scan is completed, the node features are restored to the original sensor index order through inverse permutation to obtain the spatial sequence memory features.

[0059] After spectral sorting, bidirectional state-space scanning, and inverse permutation, the bidirectional spatial memory features are obtained: In the formula, Indicates the first The window sample at the th The spatial module input is the space module after dimensional transpose in the spatiotemporal graph state space block; Indicates according to the first The node permutation matrix is ​​determined by the Fiedler vectors of each window sample; Represents the corresponding inverse permutation matrix; Indicates the first Bidirectional selective state space scan operation in each spatial module; This represents the bidirectional spatial memory feature after restoring the original node order.

[0060] In parallel with the topology-aware bidirectional state-space scanning branch, graph propagation is performed on the spatial node features based on the spatially correlated adjacency matrix to obtain the spatial graph topological features: In the formula, Indicates the first The original spatial correlation adjacency matrix corresponding to each window sample; This represents the spatial graph propagation weight matrix after ReLU nonnegativity and row Softmax normalization; Indicates the first Learnable linear mapping matrices in each spatial graph propagation branch; This represents the topological features of the spatial graph obtained through spatial graph propagation.

[0061] The first The learnable spatial gating parameters of each spatial module are mapped to a sigmoid function. Interval, obtain spatial gating coefficient: Based on the aforementioned spatial gating coefficient, the bidirectional spatial memory features and spatial graph topological features are weighted and fused: In the formula, Indicates the first Learnable scalar gating parameters in each spatial module; This represents the corresponding spatial graph propagation branch gating coefficient; This represents the bidirectional spatial memory feature obtained after spectral sorting, bidirectional state-space scanning, and inverse permutation processing. This represents the topological features of the spatial graph obtained by graph propagation based on the spatial relevance adjacency matrix; This indicates the gated fusion characteristics of the spatial module; Indicates the first Layer normalization operations in each spatial module; This represents a multilayer perceptron in the space module; Indicates a random deactivation operation; This indicates the output of the space module. When... When the value approaches 1, the gating ratio of spatial graph topological features in the fusion result is relatively high; when When the value approaches 0, the gating ratio of bidirectional spatial memory features in the fusion result is relatively high.

[0062] The spatially gated fusion features are sequentially processed through layer normalization, multilayer perceptron, and random deactivation, and then residually connected with the input features of the spatial module to obtain the spatial module output. Dimension recovery is then performed on the spatial module output to obtain the first... Output of each spatiotemporal graph state space block: In the formula, This indicates a dimension transpose operation; Indicates the first The window sample passes through the first The output after a spacetime state space block of the spatiotemporal graph.

[0063] In this way, the spatial graph state space modeling process can extract spatial degradation features along a topological structure that better conforms to the dynamic coupling relationship of sensors, thereby alleviating the order bias problem caused by directly serializing according to sensor numbers.

[0064] Step S6: Execute the time module and the space module in sequence, and repeat the spatiotemporal graph state space block a preset number of times to obtain the spatiotemporal degradation characterization.

[0065] In the formula, Indicates the total number of preset spatiotemporal graph state space blocks; Indicates the first The window sample passes through the first The output after a state space block of a spatiotemporal graph; Indicates the first The spatiotemporal degradation characterization is ultimately obtained from the window samples.

[0066] In each spatiotemporal diagram state space block, the output of the previous spatiotemporal diagram state space block is first input into the time module to obtain the time module output; the time module output is then transposed in dimensions and input into the space module; the space module output is then transposed back to the time dimension first arrangement and used as the input of the next spatiotemporal diagram state space block.

[0067] The process involves repeatedly executing a preset number of spatiotemporal graph state space blocks, with the output of the last block serving as the spatiotemporal degradation representation. This representation simultaneously includes long-range time-dependent information during the device degradation process and dynamic spatial coupling information between sensor nodes.

[0068] Step S7: Input the spatiotemporal degradation characterization into the remaining lifespan prediction head and the health factor projection head respectively, output the remaining lifespan prediction value and the auxiliary health factor sequence, and apply physical consistency constraints to the auxiliary health factor sequence.

[0069] Specifically, the spatiotemporal degradation representation is expanded into a one-dimensional feature vector and then input into the remaining lifetime prediction head. The remaining lifetime prediction value of the component to be predicted is output through nonlinear mapping. At the same time, gradient isolation processing is performed on the spatiotemporal degradation representation, and the hidden states of each enhanced time position after gradient isolation are input into the health factor projection head to generate an auxiliary health factor sequence.

[0070] For the i-th window sample, its spatiotemporal degradation characterization is denoted as... The remaining lifetime prediction branch first flattens the spatiotemporal degradation representation, then inputs it into the remaining lifetime prediction head, and outputs a normalized remaining lifetime prediction value: In the formula, This represents the normalized remaining lifetime prediction value for the i-th window sample. Indicates the feature flattening operation; Indicates remaining lifetime prediction head; This represents the trainable parameters of the remaining lifetime prediction head.

[0071] The health factor representation branch performs gradient isolation processing on the hidden states at each enhanced temporal location in the spatiotemporal degradation representation, and generates a health factor sequence within a window through the health factor projection head: In the formula, HI i,t H represents the health factor value at the t-th enhancement time position in the i-th window sample; i,t The W represents the spatiotemporal degradation hidden state at the corresponding augmented time position; L represents the number of time steps contained in the original sliding window, T=L+1 represents the number of augmented time positions after adding the global time node, t=1,…,T; HI and b HI These represent the weight matrix and bias of the health factor projection head, respectively; Represents the Sigmoid function; This represents the gradient isolation operator, which keeps the input values ​​unchanged during forward computation and blocks the gradient propagation of the physical consistency loss to the spatiotemporal feature extraction backbone during backpropagation.

[0072] In the formula, This indicates the number of samples in a training batch. Indicates the window sample number; Indicates the first Normalized remaining lifetime labels corresponding to each window sample; This represents the normalized remaining lifetime prediction value output by the model; This represents a preset tolerance boundary that allows for localized growth of adjacent health factors; This represents the minimum degradation spread required between the initial and final health factors in an enhanced time series. and These represent the weighting coefficients for the monotonic degradation loss and the amplitude loss, respectively. Indicates the current training round; This represents the number of preheating cycles required for the physical consistency loss weight to increase linearly to its maximum value; The maximum weight representing the physical consistency loss; Indicates the first The physical consistency loss weights used during each round of training. In this embodiment, , , , Although the model training uses a joint loss mechanism, under the gradient isolation mechanism, the physical consistency loss is only applied to the auxiliary health factor representation module, while the remaining lifespan main prediction path is primarily optimized by the remaining lifespan prediction loss. Since the auxiliary health factor representation module uses gradient isolation, the gradient relationship between the physical consistency loss and the main prediction path can be expressed as: In the formula, This represents the mean square error loss in the remaining lifetime prediction. This indicates anchorage loss in the lowest health factor; This represents the monotonic degradation loss with a tolerance boundary; This indicates the initial and final amplitude loss of health factors; This represents the physical consistency loss consisting of the above three items; This represents the total loss used during model training; This represents the spatiotemporal feature extraction backbone parameters used to extract time-dependent and spatially coupled information. This represents the trainable parameters of the health factor projector. Since the input to the health factor projector is gradient isolated, the physical consistency loss does not propagate gradients to the spatiotemporal feature extraction backbone.

[0073] After training, the model outputs the predicted remaining lifespan of the component to be predicted, and may selectively output an auxiliary health factor sequence. This sequence can be used to observe the relative degradation trend within the observation window and provide auxiliary interpretation of the prediction results. The physical consistency loss, through end anchoring, approximate monotonicity, and first-to-last-amplitude constraints, makes the auxiliary health factors tend to form a characterization with degradation significance, but does not directly participate in the calculation of the remaining lifespan prediction.

[0074] The health factor projection head generates a health factor sequence based on the hidden states after gradient isolation. Physical consistency constraints only apply to the health factor representation branch and do not participate in the parameter updates of the remaining lifespan prediction branch. The gradient isolation mechanism ensures that the main remaining lifespan prediction path is not directly affected by the loss of physical consistency. The remaining lifespan prediction value is output independently by the main prediction branch, while the health factor sequence serves as auxiliary information for degradation trend analysis and interpretation of prediction results.

[0075] To further verify the degradation characterization effect of auxiliary health factor sequences, this embodiment performs a visual analysis of health factor sequences at different degradation stages. Figure 5 The relationship between normalized remaining life expectancy labels, window-starting health factors, enhanced health factors at the end of the time series, and the amplitude of health factors within the window is illustrated. Figure 5 It can be seen that as the degradation process progresses, both the normalized remaining lifetime label and the health factor at the end of the enhanced time series show a downward trend, and the two change in the same direction. The health factor at the beginning of the window is generally higher than the health factor at the end of the enhanced time series, indicating that the health factor within the window can reflect the change process of health status from high to low, which verifies the effectiveness of the end anchoring constraint and the spread constraint.

[0076] Figure 6 The data shows health factor sequences at different stages of degradation. In the early stages, health factors are generally high with small fluctuations within the time window; in the middle stages, local fluctuations begin to appear, and the difference between the start and end values ​​increases; in the late stages, the amplitude of health factors further increases, enhancing the health factor at the end of the time series to approach the lower remaining lifespan label. This demonstrates that the auxiliary health factor sequence can reflect changes in health status at different stages of degradation. The monotonic degradation constraint ensures that the sequence maintains a non-increasing trend overall, while the amplitude constraint guarantees that the sequence has a sufficient degradation amplitude, avoiding degradation into an approximately constant sequence.

[0077] Figure 7 The diagram illustrates how the effective contribution rates of the time-map propagation branch and the spatial-map propagation branch change with the equipment degradation process. It should be noted that the... The gating coefficients in each time module and spatial module are obtained by mapping the corresponding learnable gating parameters through the Sigmoid function. The gating coefficients are fixed parameters after the model is trained and do not change dynamically with a single window of samples.

[0078] In the formula, Indicates a time module or a space module; Indicates the first The window sample at the th Effective contribution rate of graph propagation branches in each corresponding module; This represents the fixed gating coefficients obtained from training the corresponding module; This represents the output of the graph propagation branch; This indicates the output of a bidirectional state-space scan branch; This represents the Frobenius norm calculated for the branch output feature tensor.

[0079] For different window samples, the feature amplitudes output by the bidirectional state-space scan branch and the graph propagation branch will change with the device operating conditions and degradation stages. Therefore, this embodiment calculates the effective contribution rate of the graph propagation branch based on the ratio of the gated weighted graph propagation feature norm to the sum of the feature norms of the two gated weighted branches. This effective contribution rate is affected by both the fixed gating coefficient and the input related branch feature amplitude, and is used to reflect the relative actual contribution of the graph propagation branch in the current window sample.

[0080] Figure 7 The temporal graph curves represent the effective contribution of temporal graph propagation features relative to bidirectional temporal memory features within the temporal module, while the spatial graph curves represent the effective contribution of spatial graph propagation features relative to bidirectional spatial memory features within the spatial module. An increase in the effective contribution rate indicates that the gated weighted output of the corresponding graph propagation branch is enhanced relative to the state space scan branch; a decrease in the effective contribution rate indicates that the model in the current stage relies more on bidirectional state space scan features in the corresponding module.

[0081] from Figure 7 It can be seen that the effective contribution rates of the temporal and spatial propagation branches show different trends with the degradation process, indicating that the temporal topological relationship and the sensor's spatial topological relationship have different relative effects at different degradation stages.

[0082] During the inference phase, no real remaining lifespan label is required. Only the multi-sensor monitoring data of the component to be predicted at the current stage needs to be input. The monitoring data is processed according to the same normalization parameters, window length and sliding step size as in the training phase and then input into the model. At the same time, the remaining lifespan prediction value and the auxiliary health factor sequence can be obtained.

[0083] According to the above embodiments, this invention dynamically constructs window-level temporal and spatial correlation graphs, introduces learnable global nodes for information enhancement, extracts sequence memory features from sensor node sequences sorted along the time dimension and spectrum using bidirectional state-space scanning, and performs gating fusion of the bidirectional state-space scanning features with the corresponding graph propagation features within the time and spatial modules respectively, ultimately outputting both the remaining lifetime prediction value and the auxiliary health factor sequence. These techniques achieve a joint representation of long-range temporal dependencies and sensor spatial coupling relationships in multi-sensor degradation data, solving the technical problem that existing remaining lifetime prediction methods struggle to simultaneously process both types of information and lack interpretable output. The auxiliary health factor branch, through end-anchoring constraints, monotonic degradation constraints, and amplitude constraints, provides the trend of health status changes within the window without interfering with the main prediction path, thus aiding in the interpretation of the prediction results.

[0084] Second Embodiment A topology-aware spatiotemporal graph fusion-based vehicle active system life prediction system is applied to implement the vehicle active system life prediction method as described in the first embodiment. The system includes: The preprocessing module is used to acquire multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system. These signals include at least one of the following: temperature, pressure, rotational speed, voltage, current, capacity, internal resistance, or vibration signals characterizing the operating and degradation states of the component. The signals are normalized and sliced ​​using a sliding window to obtain a multivariable time-series input sequence. The specific processing flow of this module is consistent with step S1 in Example 1, and will not be repeated here.

[0085] The node enhancement module is used to perform feature mapping on the multivariate temporal input sequence and introduce learnable global time nodes and global spatial nodes to obtain a spatiotemporally enhanced input representation; the specific processing flow of this module is consistent with step S2 of the first embodiment.

[0086] The graph construction module is used to calculate the Pearson correlation coefficient based on the spatiotemporal enhancement input representation, using the original time step and global time node as time graph nodes, and the original sensor node and global spatial node as spatial graph nodes, and dynamically generate the temporal correlation adjacency matrix and the spatial correlation adjacency matrix respectively; the specific processing flow of this module is the same as step S3 of the first embodiment.

[0087] The time modeling module is used to perform parallel time-dimensional bidirectional state-space scanning and time graph propagation based on the time-related adjacency matrix, and to obtain the time module output by fusing bidirectional time memory features and time graph topological features through learnable time gating parameters. The specific processing flow of this module is consistent with step S4 of Example 1.

[0088] The spatial modeling module is used to perform topology-aware bidirectional state space scanning and spatial graph propagation in parallel based on the spatially related adjacency matrix, and to obtain the spatial module output by fusing bidirectional spatial memory features and spatial graph topological features through learnable spatial gating parameters. The specific processing flow of this module is consistent with step S5 of Example 1.

[0089] The spatiotemporal block iteration module is used to execute the time modeling module and the spatial modeling module in sequence. The output of the spatial module is restored to its dimension and used as the input of the next spatiotemporal graph state space block. The module is executed repeatedly for a preset number of spatiotemporal graph state space blocks. The output of the last spatiotemporal graph state space block is used as the spatiotemporal degradation representation. The specific processing flow of this module is the same as step S6 of the first embodiment.

[0090] The lifetime prediction module is used to output the predicted remaining lifetime value of the component to be predicted based on the spatiotemporal degradation characterization. The specific processing flow of this module is consistent with the remaining lifetime prediction part in step S7 of Embodiment 1.

[0091] The health factor module performs gradient isolation processing on the spatiotemporal degradation representation, maps the hidden states at each enhanced time location to a health factor sequence, and applies physical consistency constraints to the health factor sequence to obtain an auxiliary health representation. The specific processing flow of this module is consistent with the auxiliary health factor representation part in step S7 of Embodiment 1.

[0092] In practical applications, this system can be deployed in vehicle controllers, vehicle remote monitoring platforms, or automotive operation and maintenance management platforms. Taking a vehicle engine as an example, during vehicle operation or engine bench testing, temperature, pressure, speed, and vibration sensors installed on the engine collect operational data at a preset sampling frequency. The collected multi-sensor time-series data is cleaned and formatted before being input into the preprocessing module to obtain the current window sample. The node enhancement module introduces global time nodes and global spatial nodes into the current window sample, and the graph construction module dynamically calculates the time-related adjacency matrix and the spatial-related adjacency matrix. The time modeling module and the spatial modeling module extract long-range time-dependent features and sensor spatial coupling features, respectively. The spatiotemporal block iteration module repeatedly executes a preset number of spatiotemporal graph state space blocks to obtain the spatiotemporal degradation characterization of the current window sample. Based on this, the life prediction module outputs the remaining life prediction value of the vehicle engine.

[0093] Taking automotive power batteries as an example, this system connects to monitoring data such as voltage, current, capacity, internal resistance, and temperature collected by the battery management system. During vehicle charging, discharging, and driving, it continuously outputs the predicted value of the remaining cycle life of the automotive power battery, providing a basis for power battery maintenance or replacement.

[0094] It should be noted that the above application scenarios are only used to illustrate the deployment and usage of the system. This invention is not limited to specific models of automotive power batteries, automotive engines, drive motors or transmissions. Other components in automotive power systems with monitorable degradation processes can also use this invention to predict their remaining lifespan.

[0095] During the inference phase, the system only needs to input the multi-sensor monitoring data collected at the current moment, process the data according to the same normalization parameters, window length and sliding step size as in the training phase, and input it into the model to obtain the remaining lifespan prediction value at the current moment, while outputting the auxiliary health factor sequence.

[0096] Maintenance personnel or vehicle controllers can make maintenance decisions for the components to be predicted based on the remaining life prediction value, and observe the changing trend of the health status of the components to be predicted based on the auxiliary health factor sequence. Without relying on manual feature extraction or experience threshold judgment, degradation assessment and life prediction can be completed online during equipment operation.

[0097] The above modules can be integrated into a single processor, or they can be implemented separately by multiple processors or dedicated hardware modules. The modules interact with each other via a data bus or communication interface, thereby enabling end-to-end prediction of the remaining life and degradation trend analysis of the components to be predicted in the automotive dynamic system.

[0098] Experimental verification To verify the modeling capabilities of this invention for degradation signals related to engines and power batteries in automotive power systems, this embodiment uses publicly available engine degradation datasets and lithium-ion battery degradation datasets for experimental verification. These publicly available datasets are used to verify the invention's ability to process multi-sensor time-series degradation data and its remaining lifetime prediction performance, and are not intended to limit the specific model and structure of the components to be predicted in the automotive power system.

[0099] (a) Verification of engine degradation data The aero-engine experiments used the NASA-released C-MAPSS dataset, which comprises four sub-datasets, FD001 to FD004, simulating the degradation process of turbofan engines under different operating conditions and failure modes. Each sub-dataset recorded time-series monitoring data from 21 sensor channels. During the experiments, model training and evaluation were performed according to the training and testing sets provided in each sub-dataset, maintaining consistent normalization methods, window lengths, and sliding step sizes throughout the training and testing phases. For comparison, this experiment selected nine existing methods—HierCorrPool, LOGO, STFA, HAGCN, STAGNN, GRU_CM, GraphSAGE, STGNN, and GWNet—as baseline models, using root mean square error (RMSE), mean absolute error (MAE), and the remaining life prediction scoring function Sv1 as evaluation metrics. RMSE and MAE measure the numerical deviation between predicted values ​​and actual remaining life; smaller values ​​indicate lower prediction errors. The scoring function Sv1 comprehensively evaluates the differentiated impact of early and delayed predictions on operational decisions.

[0100] As shown in Table 2 (Table 2 shows the comparative experimental results on the C-MAPSS dataset), the method of this invention (denoted as STGMamba in the table) achieved an RMSE of 12.2258, a MAE of 8.8510, and a score Sv1 of 2.6411 on the FD001 subset, all three metrics being superior to the comparative methods such as HierCorrPool, LOGO, and STFA. On the FD002 subset, the method of this invention achieved an RMSE of 12.7766, a MAE of 9.2565, and a score of 2.7248, again achieving the best prediction performance. On the FD003 subset, the RMSE was 11.3761, the MAE was 8.1396, and the score was 2.3516. On the FD004 subset, the RMSE was 14.0040, the MAE was 9.9152, and the score was 4.1566. Experimental results on four subset datasets show that the method of the present invention can achieve low prediction errors and scores under different operating conditions and fault modes. The experiment verifies the ability of the present invention to model long-range time dependence and sensor coupling relationship in the degradation signal of engine-type multi-sensor sensors.

[0101] Table 2 Comparison Experiment Results of CMAPSS Dataset To further verify the performance of the method of this invention on large-scale datasets, this experiment also used the N-CMAPSS dataset for validation. The N-CMAPSS dataset has more sensor channels and a longer lifecycle of data compared to the C-MAPSS dataset. As shown in Table 3 (Table 3 shows the experimental results of the N-CMAPSS dataset comparison), in the three test units of Unit 11, Unit 14, and Unit 15, the RMSE of the method of this invention were 6.3432, 7.2809, and 3.7731, respectively, and the MAE were 5.1376, 5.8737, and 2.7066, respectively, all of which are superior to the comparative methods such as GraphSAGE and GWNet. These results further verify the prediction performance of the method of this invention on large-scale datasets.

[0102] Table 3 Comparison Experiment Results of N-CMAPSS Dataset (II) Verification of degradation data for power batteries The lithium-ion battery experiments utilized the NASA Battery dataset and the MIT Battery dataset. The NASA Battery dataset contains charge-discharge cycle monitoring data for battery samples such as B0005, B0006, and B0018, recording multi-dimensional features including voltage, current, capacity, and temperature. The MIT Battery dataset contains capacity degradation data from multiple commercial lithium-ion batteries during cycle aging. The experiments employed the same normalization method and windowing strategy as the validation of engine-type degradation data, using the normalized remaining battery life as the prediction target. Evaluation metrics included RMSE, MAE, and the scoring function Sv2. Sv2 was used to evaluate the consistency between the life prediction results and the degradation trend; a higher Sv2 value indicates better overall prediction performance.

[0103] As shown in Table 4 (Table 4 presents the comparative experimental results on the NASA Battery dataset), on battery sample B0005, the RMSE of the method of this invention is 6.8106, the MAE is 5.6500, and the score is 0.2597; on battery sample B0006, the RMSE is 7.7328, the MAE is 6.3400, and the score is 0.2387; on battery sample B0018, the RMSE is 3.6020, the MAE is 3.2307, and the score is 0.2423. Compared with comparative methods such as HierCorrPool, LOGO, STFA, and HAGCN, the method of this invention achieves lower prediction errors on most battery samples.

[0104] Table 4 Comparison Experiment Results of NASA Battery Dataset As shown in Table 5 (Table 5 presents the comparative experimental results on the MIT Battery dataset), the RMSE of the method of this invention is 67.8977, the MAE is 51.0356, and the score is 0.4541, with the score being superior to all the comparative methods. The experimental results combining engine-type degradation data and power battery-type degradation data demonstrate that the method of this invention can effectively model the multi-sensor degradation signals related to two types of key components of automotive power systems.

[0105] Table 5. Comparison Experiment Results of the MIT Battery Dataset Based on experimental results combining engine and power battery degradation data, the method of this invention achieved low prediction errors across all test datasets and maintained stable prediction performance under different operating conditions and different types of degraded equipment. These experimental results verify that the spatiotemporal degradation features extracted by this invention through dynamic graph construction, bidirectional state space scanning, topology-aware spatial sorting, and learnable gating fusion have excellent characterization capabilities. They can effectively support the remaining life prediction task of key components in automotive power systems such as vehicle engines and automotive power batteries, providing a basis for component inspection and replacement decisions in automotive operation and maintenance scenarios.

[0106] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features with similar functions disclosed in this disclosure.

[0107] Furthermore, except for the data dependency order explicitly defined in this specification, for operations without data dependencies, the execution order can be adjusted or they can be executed in parallel as needed. In certain environments, multitasking and parallel processing may be advantageous. Similarly, although several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0108] The embodiments described above are merely examples of several implementations of the present invention, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention.

Claims

1. A method for predicting the lifespan of a vehicle's active power system based on topological sensing spatiotemporal graph fusion, characterized in that, include: The multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system are acquired. The multi-sensor time-series monitoring signals include at least one of temperature, pressure, speed, voltage, current, capacity, internal resistance or vibration signals that characterize the operating state and degradation state of the component to be predicted. The signals are normalized and sliced ​​by a sliding window to obtain a multivariable time-series input sequence. By introducing learnable global time nodes and global spatial nodes into the multivariate temporal input sequence, a spatiotemporally enhanced input representation is obtained; Based on the spatiotemporal enhancement input representation, the original time step and global time node are used together as time graph nodes, and the original sensor node and global spatial node are used together as spatial graph nodes. Pearson correlation coefficients are calculated respectively, and time-related adjacency matrices and spatial-related adjacency matrices are dynamically generated. In each spatiotemporal graph state space block, a time module is executed according to the time-related adjacency matrix to obtain the time module output; After transposing the time module output, the spatial module is executed according to the spatial correlation adjacency matrix to obtain the spatial module output; The time module and the space module are executed sequentially, and a preset number of spatiotemporal graph state space blocks are repeated. The output of the last spatiotemporal graph state space block is used as the spatiotemporal degradation characterization of the component to be predicted. The spatiotemporal degradation characterization is input into the remaining lifespan prediction head and the health factor projection head, respectively, and the remaining lifespan prediction value and auxiliary health factor sequence of the component to be predicted are output. Physical consistency constraints are applied to the auxiliary health factor sequence.

2. The method for predicting the lifespan of a vehicle's active power system based on topology-sensing spatiotemporal graph fusion according to claim 1, characterized in that, The component to be predicted is selected from at least one of the following: automotive power battery, automotive engine, drive motor, transmission, and other components in the automotive power system that have a monitorable degradation process; When the component to be predicted is an automotive power battery, the multi-sensor timing monitoring signal includes one or more of voltage, current, capacity, internal resistance, or temperature; When the component to be predicted is a vehicle engine, drive motor, or transmission, the multi-sensor timing monitoring signal includes one or more of temperature, pressure, rotational speed, or vibration.

3. The method for predicting the lifespan of a vehicle's active power system based on topological sensing spatiotemporal graph fusion according to claim 1, characterized in that, The time-dependent adjacency matrix and the spatially dependent adjacency matrix are obtained through the following steps: Using the original time step and the global time node together as time graph nodes, calculate the Pearson correlation coefficient between the feature vectors of any two time graph nodes to obtain the time-related adjacency matrix; Using the original sensor nodes and global spatial nodes as spatial graph nodes, the Pearson correlation coefficient between the feature sequences of any two spatial graph nodes is calculated to obtain the spatial correlation adjacency matrix. The time-related adjacency matrix and the spatially related adjacency matrix are both dynamically calculated and generated based on the current input window.

4. The method for predicting the lifespan of a vehicle's active power system based on topology-sensing spatiotemporal graph fusion according to claim 1, characterized in that, The time module output is obtained through the following steps: Selective state recursive scanning is performed along the forward and reverse directions of the time series to obtain bidirectional time memory features; Based on the aforementioned temporal correlation adjacency matrix, correlation-weighted graph propagation is performed on the enhanced temporal location features to obtain temporal graph topological features; The bidirectional time memory feature is fused with the time map topology feature to obtain the time module output.

5. The method for predicting the lifespan of a vehicle's active power system based on topology-sensing spatiotemporal graph fusion according to claim 1, characterized in that, The bidirectional spatial memory feature in the spatial module is obtained through the following steps: Top-k sparsification is performed on the spatially related adjacency matrix to retain connections with strong absolute correlation for each node; Inject weak connections into the sparsed spatial correlation adjacency matrix to enhance connectivity; A Laplacian matrix is ​​constructed based on the absolute weights of the spatially correlated adjacency matrix after connectivity enhancement. The Fiedler vector of the Laplacian matrix is ​​solved, and the nodes are sorted according to the magnitude of their components in the Fiedler vector to determine the arrangement order of the spatial nodes. The node with the highest degree centrality is selected as the anchor point, and the direction of the Fiedler vector is normalized by sign. Perform a bidirectional state space scan according to the sorted node sequence, and restore the original spatial node index order by inverse permutation after the scan is completed to obtain the bidirectional spatial memory feature.

6. The method for predicting the lifespan of a vehicle's active power system based on topology-sensing spatiotemporal graph fusion according to claim 1, characterized in that, The spatiotemporal degradation characterization is obtained through the following steps: In the time module of each spatiotemporal graph state space block, the input of the spatiotemporal graph state space block is input into the time dimension bidirectional state space scanning branch and the time graph propagation branch respectively to obtain bidirectional time memory features and time graph topological features; The learnable time-gated parameters are mapped to time-gated coefficients using the Sigmoid function, and the bidirectional time memory features and the time graph topology features are weighted and fused according to the time-gated coefficients. After layer normalization, multilayer perceptron and residual connection processing, the time module output is obtained. The output of the time module is transposed and then input into the spatial module. The spatial nodes are spectrally sorted according to the spatial correlation adjacency matrix. A bidirectional state space scan is performed along the sorted node sequence and the original node order is restored by inverse permutation to obtain bidirectional spatial memory features. At the same time, spatial graph propagation is performed according to the spatial correlation adjacency matrix to obtain spatial graph topological features. The learnable spatial gating parameters are mapped to spatial gating coefficients using the Sigmoid function, and the bidirectional spatial memory features and the spatial graph topological features are weighted and fused according to the spatial gating coefficients. After layer normalization, multilayer perceptron and residual connection processing, the spatial module output is obtained. The output of the spatial module is used to restore its dimension and is used as the input of the next spatiotemporal graph state space block; the spatiotemporal graph state space block is executed repeatedly for a preset number of times, and the output of the last spatiotemporal graph state space block is used as the spatiotemporal degradation representation.

7. The method for predicting the lifespan of a vehicle's active power system based on topology-sensing spatiotemporal graph fusion according to claim 6, characterized in that, Correlation-weighted graph propagation is based on the temporal or spatial correlation adjacency matrix, the input features X of the current temporal or spatial module, and the learnable linear mapping matrix W. g The calculated graph topological feature H graph Represented as: ； ； In the formula, This represents a time-related adjacency matrix or a spatial-related adjacency matrix constructed based on the Pearson correlation coefficient; Represents the function for nonnegating the adjacency matrix; This indicates normalization along the row direction of the adjacency matrix; This represents the graph propagation weight matrix after nonnegation and row normalization; Indicates the input characteristics of the current time module or space module; This represents the learnable linear mapping matrix in the graph propagation branch; This represents the graph topological features obtained from correlation-weighted graph propagation.

8. The method for predicting the lifespan of a vehicle's active power system based on topological sensing spatiotemporal graph fusion according to claim 1, characterized in that, The auxiliary health factor sequence is obtained through the following steps: The high-dimensional hidden states corresponding to each enhanced time position in the spatiotemporal degradation representation are mapped to one-dimensional scalar values ​​by the health factor projection head, and the scalar values ​​are constrained to a preset range by the Sigmoid function to obtain the health factor sequence within the window.

9. The method for predicting the lifespan of a vehicle's active power system based on topological sensing spatiotemporal graph fusion according to claim 8, characterized in that, The physical consistency constraints include: End-of-life anchoring constraints are used to bring the health factor values ​​at the end of the enhanced time series closer to the normalized remaining life label. Monotonic degradation constraint is used to penalize non-physical increases in health factor values ​​at adjacent augmentation time locations; Amplitude constraint is used to constrain the preset degradation amplitude between the initial health factor value of the window and the health factor value at the end of the enhanced time series. The monotonic degradation constraint is an asymmetric monotonic penalty with a tolerance boundary, used to generate a penalty when the growth rate of the health factor value at adjacent enhancement time positions exceeds the preset tolerance boundary. The health factor value is a one-dimensional scalar output by the health factor projection head at the corresponding enhancement time position.

10. A topology-sensing spatiotemporal graph fusion-based vehicle active system life prediction system, applied to implement the vehicle active system life prediction method according to any one of claims 1 to 9, characterized in that, The system includes: The preprocessing module is used to acquire multi-sensor time-series monitoring signals of the component to be predicted in the automotive power system, and obtain a multivariable time-series input sequence through normalization processing and sliding window slicing; wherein, the multi-sensor time-series monitoring signals include physical quantity signals characterizing the operating state and degradation state of the component to be predicted; The node enhancement module is used to introduce learnable global time nodes and global spatial nodes into the multivariate temporal input sequence to obtain a spatiotemporally enhanced input representation; The graph construction module is used to calculate the Pearson correlation coefficient based on the spatiotemporal enhancement input representation, using the original time step and global time node as time graph nodes and the original sensor node and global spatial node as spatial graph nodes, and dynamically generate the time-related adjacency matrix and the spatial-related adjacency matrix. The time modeling module is used to perform bidirectional state space scanning, time graph propagation, and time-gated fusion along the time dimension based on the time-related adjacency matrix to obtain the time module output; The spatial modeling module is used to perform spectral sorting of spatial nodes according to the spatial correlation adjacency matrix, and to perform bidirectional state space scanning, inverse permutation, spatial graph propagation and spatial gating fusion along the sorted node sequence to obtain the spatial module output; The spatiotemporal block iteration module is used to execute the time modeling module and the space modeling module in sequence, and repeat a preset number of spatiotemporal graph state space blocks, and use the output of the last spatiotemporal graph state space block as the spatiotemporal degradation representation. The lifetime prediction module is used to output the remaining lifetime prediction value of the component to be predicted based on the spatiotemporal degradation characterization. The health factor module is used to perform gradient isolation processing on the spatiotemporal degradation representation, map the hidden states at each enhanced time location to a health factor sequence, apply physical consistency constraints to the health factor sequence, and obtain auxiliary health representation.

Images