A method for predicting the remaining life of a methanol fuel supply system

Abstract

The application discloses a kind of methanol fuel supply system remaining useful life prediction method.There is comprising: obtaining methanol fuel supply system multivariate monitoring data and constructing sliding window observation sequence;Physical directed graph and structure mask are constructed based on system physical topology and control logic;Space interaction modeling is carried out under the constraint of structure mask, and spatial coupling feature sequence is extracted;Time series dependence modeling is carried out using Transform encoder;Channel level gate weight is generated based on spatial representation sequence, and structural condition modulation is carried out to time series hidden representation;Time aggregation is carried out to the spatiotemporal representation after fusion, and the remaining useful life prediction result of multiple key components is output by regression layer.The application can improve the accuracy and stability of remaining useful life prediction under the complex coupling degradation condition of methanol fuel supply system.

Description

Technical Field

[0001] This invention belongs to the field of methanol fuel supply system fault prediction and health management technology, specifically relating to a method for predicting the remaining life of a methanol fuel supply system. Background Technology

[0002] Methanol, as a marine fuel with good storage and transportation conditions and application prospects, is being used more and more widely in marine power systems. The methanol fuel supply system, as a key subsystem connecting fuel storage, transportation, pressurization, filtration, heat exchange, and supply processes, directly affects the stable operation of the marine power system due to its operational safety, reliability, and the health status of its key components.

[0003] A methanol fuel supply system typically comprises several key components, including storage tanks, pumps, filters, heat exchangers, valves, and pipelines. These components form a complex coupled system through physical connections, fluid transfer relationships, and control logic. During long-term operation, key components such as booster pumps, filters, heat exchangers, and sealing structures are affected by factors such as fluid impact, operating condition switching, corrosion, wear, blockage, and heat load fluctuations, gradually leading to performance degradation. This can result in problems such as abnormal supply pressure, flow fluctuations, reduced heat exchange capacity, and even leaks. Therefore, predicting the remaining life of key components in a methanol fuel supply system is crucial for identifying degradation trends in advance, optimizing maintenance schedules, and ensuring the safe operation of the system.

[0004] Existing monitoring methods for methanol fuel supply systems typically rely on sensor data acquisition, threshold alarms, or univariate trend analysis. Most methods focus on independent monitoring of measurement points such as pressure, temperature, flow rate, and vibration, emphasizing alarms and responses after abnormal conditions occur. While these methods can reflect the current operating status of the system to some extent, they struggle to fully explore the coupling relationships between multiple measurement points and variables, and they also fail to characterize the spatiotemporal features of degradation and evolution of key components.

[0005] Furthermore, existing remaining lifetime prediction methods are mostly based on traditional data-driven models or general time series modeling methods, typically inputting multivariate monitoring data directly into models such as recurrent neural networks, convolutional neural networks, or Transformers for learning. While these methods can extract certain time series patterns, they generally suffer from the following shortcomings: First, they do not fully utilize the physical topological relationships between pumps, filters, heat exchangers, valves, and pipelines within the methanol fuel supply system, leading to the introduction of non-physical correlations that do not conform to actual physical connections during spatial interaction modeling. Second, under complex operating conditions and multi-component coupled degradation scenarios, the spatiotemporal dependencies between different measurement points in the system are quite complex, and relying solely on general time series models makes it difficult to effectively distinguish between true degradation information and noise disturbances caused by operating condition fluctuations. Third, existing methods often focus on single-component lifetime prediction, making it difficult to achieve joint remaining lifetime prediction for multiple key components within the methanol fuel supply system.

[0006] Furthermore, methanol fuel supply systems exhibit significant operating condition switching characteristics during operation, such as startup, steady-state operation, load changes, shutdown, and abnormal operating condition switching. The influence paths and response intensities between variables at various measurement points may change under different operating conditions. If the prior knowledge of the system's physical structure and the coupling characteristics of operating conditions are ignored in the prediction model, it can easily lead to biases in the extraction of spatiotemporal features, thereby reducing the accuracy and stability of the remaining life prediction results.

[0007] Therefore, there is an urgent need to propose a method for predicting the remaining lifetime of methanol fuel supply systems by combining the physical topology, control logic, and multivariate time-series data. By introducing prior constraints on physical structure during the spatiotemporal feature extraction process, the model's ability to characterize the coupling degradation behavior of key components can be enhanced, thereby improving the accuracy, stability, and engineering applicability of remaining lifetime prediction for multiple key components under complex operating conditions. Summary of the Invention

[0008] To address the problem that existing methods for predicting the remaining life of methanol fuel supply systems struggle to fully utilize the system's physical topology, control logic relationships, and multivariate temporal coupling information, resulting in insufficient prediction accuracy and stability under complex operating conditions, this invention provides a method for predicting the remaining life of methanol fuel supply systems, comprising the following steps: S1. Obtain multivariate monitoring data of the methanol fuel supply system and construct a sliding window observation sequence to form the original input features for prediction; S2. Based on the physical topology and control logic of the methanol fuel supply system, construct a physical directed graph representing the connection relationship of components and its corresponding structural mask; S3. Under the structural mask constraint, perform spatial interaction modeling on the observation sequence and extract spatial coupling feature sequences that conform to the system's physical priors; S4. Use a Transformer encoder to perform long-range dependency modeling on the spatially coupled feature sequence to obtain a temporal hidden representation that integrates the global temporal context; S5. Generate channel-level gating weights based on the spatially coupled feature sequence, and perform structural condition modulation on the temporal hidden representation to achieve spatial prior-guided spatiotemporal feature fusion. S6. Perform time aggregation on the fused spatiotemporal features and output the remaining life prediction results of multiple key components in the methanol fuel supply system through the regression layer.

[0009] Furthermore, step S1 specifically includes: S11. Obtain the methanol fuel supply system at any point in its life cycle. The multivariate observation data below combines external operating condition input variables with system response variables to form an observation vector. : S12. Using a length of The sliding time window truncates the observation vector and uses the current time as the threshold. Construct a window-level input tensor as the window alignment end. .

[0010] Furthermore, step S2 specifically includes: S21. Extract the pipeline connection relationships, medium flow paths, pressure propagation paths, heat exchange coupling paths, and valve control links within the methanol fuel supply system. Abstract sensor measurement points and key components into nodes, and construct the following three types of prior adjacency matrices: in Represents the prior adjacency matrix of physical connections; Indicates the total number of nodes in the graph; in, The prior adjacency matrix represents the mechanism of propagation. in, Represents the prior adjacency matrix of the control logic; S22. Weighted fusion of the three types of prior adjacency matrices to construct a comprehensive structural prior matrix for the methanol fuel supply system. : in, , and These are the weight coefficients for physical connection priors, mechanism propagation priors, and control logic priors, respectively. , , , ; By adding a self-connect to the comprehensive structure prior matrix, an extended structure matrix is ​​obtained. : in, It is the identity matrix; S23. Normalize the extended structure matrix to obtain the normalized structure prior matrix. : in, for The degree matrix; To suppress information propagation between physically reachable nodes, a structural shielding matrix is ​​constructed. : in, The shielding penalty constant is greater than zero. To extend the structure matrix The Middle line, number The elements of the column.

[0011] Furthermore, step S3 specifically includes: S31. At time , observation vector Mapping nodes to the embedding space yields the node embedding matrix. : in, For node-wise linear mapping functions, Embed items for node locations or component types. , Where F is the spatial feature dimension, and F is the total number of graph nodes; S32. Construct the query matrix, key matrix, and value matrix: in, , and Representing time respectively The query matrix, key matrix, and value matrix; , and These are the learnable parameter matrices for the query matrix, key matrix, and value matrix, respectively. S33. To ensure that both the structural priors and the current degenerate state participate in the spatial correlation calculation, the construction time... Degenerate bias matrix , its first Line number Column elements Represented as: in, , , They are nodes The deviation of pressure, flow rate, and temperature from the healthy baseline. Represents a node To the node Distance along the physical path of the methanol fuel supply system , , This is the bias coefficient; Construct a correlation matrix constrained by both structural priors and degenerate biases. : in, Represents the normalized structure prior matrix; Represents the degenerate bias matrix; Represents the structure mask matrix; and These represent the adjustment coefficients for the structural prior term and the degenerate bias term, respectively; The hidden layer feature dimension of the attention mechanism; S34. Regarding the correlation matrix After normalization, the spatial interaction weight matrix is ​​obtained. : Obtain spatial aggregation output : The node-level space is obtained through residual connection and normalization. : S35. Node-level space Expand and map into time step spatial representation vectors : in, This represents the matrix flattening operation; This is a fusion mapping function used to compress the node-level spatial representation into a fixed-dimensional vector; This leads to the formation of a window-level spatial representation sequence. : ; Each represents a continuous segment within the sliding window. The spatial representation vector corresponding to each time step.

[0012] Furthermore, step S4 specifically includes: S41. Spatial representation sequence Injection timing position encoding To form the initial timing input : in, This is the position encoding adjustment coefficient; S42. Input the initial timing sequence Send in The layer-time encoder yields the final temporal hide. : in, Indicates by A time-series coding map formed by cascading Transformer encoders.

[0013] Furthermore, step S5 specifically includes: S51. To enable multi-source structural priors to play a role again in the temporal fusion stage, the node-level space... and normalized structure prior matrix Mapping is performed to obtain the structural prior summary vector. : in, This is a compression mapping function; S52. Spatial representation sequence Compression results and structural prior summary vector The inputs are concatenated to construct a dual-source gated input, and structural conditional gated weights are generated. : in, This represents vector concatenation. For the Sigmoid function, and These are learnable parameters; Expand the gating weights to: in To gate the structural conditions weights Replication and expansion along the time dimension The extended gating weight matrix obtained after this step; Simultaneously, the structural prior summary vector is expanded to: in, This indicates a copy-and-expand operation along the time dimension; Indicates the length of the sliding window; To extract the structural prior summary vector Replication and expansion along the time dimension The extended structure summary matrix obtained after this step; S53. Under the combined constraints of the dual-source gating and the structural prior summary, the temporal hiding... Conditional modulation is performed to obtain the fused spatiotemporal representation. : in, and These are the adjustment coefficients for the gated modulation term and the structural prior summary injection term, respectively. Represents element-wise product. The representation layer normalization operation is used to mitigate scale differences between different feature channels.

[0014] Furthermore, step S6 specifically includes: S61. Representation of the merged spacetime Perform weighted time aggregation to obtain a window-level global representation. : in, For the first The aggregate weights corresponding to each time step; Indicates the first [number] within the sliding window The fused spatiotemporal representation corresponding to each time step; satisfying: S62. For key component sets Each key component Each component has a dedicated lifespan. Perform initial remaining lifetime prediction to obtain initial lifetime estimate. : S63. Based on the coupling degradation relationship among multiple key components in the methanol fuel supply system, a system coupling correction term is introduced to correct the initial lifetime estimate, resulting in the final remaining lifetime prediction. : in, Indicates the first Initial remaining life estimates for key components; Indicates key components For key components The coupling influence coefficient; Indicates the first in the methanol fuel supply system Key components at time The initial remaining lifetime estimate; Indicates time Key components The degradation intensity index; This represents the coupling correction mapping function, used to characterize the correction of the remaining life prediction results of the current critical component by the degradation state of related critical components.

[0015] Furthermore, step S6 also includes: S64, Based on Real Remaining Life Labels Define the joint loss function Through multi-objective collaborative optimization, a deep integration of data-driven prediction and physical prior constraints is achieved; in, Represents the set of key components The number of key components; Indicates the first Loss weights corresponding to each key component; This represents the error loss function between the predicted remaining lifetime and the actual remaining lifetime. Indicates the first The predicted final remaining life of each key component; Indicates the first Key components at time The true remaining life label; and These represent the weight coefficients of the structural consistency constraint term and the coupling constraint term, respectively. This represents the loss of structural consistency. This represents the coupling constraint loss.

[0016] Structural consistency loss The consistency between the constraint space representation and the structural prior is expressed as: in, Denotes the Frobenius norm; Coupling constraint loss The physical rationality used to constrain the lifespan relationship of coupled components is expressed as: in, Indicates key components With key components The allowable lifetime difference threshold between them; This represents a hinge-type penalty function. Indicates the first The predicted final remaining life of each key component.

[0017] Compared with the prior art, the significant advantages of this invention are as follows: 1) This invention constructs a physical directed graph and structural mask based on the physical topology and control logic of the methanol fuel supply system, thereby realizing the physical constraint modeling of the spatial coupling relationship of multivariable monitoring data. It introduces the actual connection relationship, state propagation path and control link of the system into the feature extraction process, reduces the propagation of information that does not conform to the actual mechanism, and improves the accuracy of the characterization of the degradation state of key components and the physical credibility of the prediction results.

[0018] 2) By constructing a sliding window observation sequence and combining spatial interaction modeling and Transformer temporal dependency modeling, this invention achieves the joint extraction of local degradation evolution features and global temporal context information. This can more effectively distinguish between short-term disturbances caused by operating condition switching and operational fluctuations and the actual degradation trend of key components, thereby improving the accuracy, stability and robustness of remaining life prediction under complex operating conditions.

[0019] 3) This invention improves the systematicness and engineering applicability of methanol fuel supply system life assessment by establishing a spatial prior-guided spatiotemporal feature fusion mechanism and a remaining life prediction method based on the joint output of multiple key components. At the same time, it can comprehensively consider the coupling degradation relationship between multiple components within the system and make unified predictions on the remaining life of multiple key components, providing a basis for predictive maintenance, maintenance timing optimization and operation and maintenance decisions, thereby reducing the risk of unplanned downtime and improving system operation safety and operation and maintenance efficiency. Attached Figure Description

[0020] To more clearly illustrate the technical solutions of the embodiments of this application, 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.

[0021] Figure 1 This is a schematic diagram of the overall process of a method for predicting the remaining life of a methanol fuel supply system according to the present invention. Figure 2 This is a schematic diagram illustrating the physical topology, mechanism propagation relationship, control logic relationship, and structural mask construction process in this invention; Figure 3 This is a schematic diagram of the spatial interaction and spatiotemporal fusion network structure constrained by structural priors in this invention. Detailed Implementation

[0022] The following is in conjunction with the instruction manual appendix. Figure 1-3 The present invention provides a method for predicting the remaining life of a methanol fuel supply system, which includes the following steps: S1. Obtain multivariate monitoring data of the methanol fuel supply system and construct a sliding window observation sequence to form the original input features for prediction; S1 specifically includes: S11. Obtain the methanol fuel supply system at any point in its life cycle. The multivariate observation data below combines external operating condition input variables with system response variables to form an observation vector. : in, Input the set of variables for the operating conditions. For the set of system response variables, The total dimension of the measurement point variables; S12. Using a length of The sliding time window truncates the observation vector and uses the current time as the threshold. Construct a window-level input tensor as the window alignment end. : in, The length of the sliding window. Indicates continuous within the window The observation vector at each time step; S2. Based on the physical topology and control logic of the methanol fuel supply system, construct a physical directed graph representing the connection relationship of components and its corresponding structural mask; S2 specifically includes the following steps: S21. Extract the pipeline connection relationships, medium flow paths, pressure propagation paths, heat exchange coupling paths, and valve control links within the methanol fuel supply system. Abstract sensor measurement points and key components into nodes, and construct the following three types of prior adjacency matrices: in Represents the prior adjacency matrix of physical connections; Indicates the total number of nodes in the graph; elements in Represents a node With nodes Does the physical topology of the methanol fuel supply system contain direct connections or physical interactions? in, The prior adjacency matrix represents the mechanism of propagation. elements in Represents a node For nodes Does it involve mechanisms such as pressure propagation, flow coupling, temperature transfer, or heat transfer effects? in, Represents the prior adjacency matrix of the control logic; elements in Represents a node With nodes Are there control links, valve interlocking relationships, or logical connections for switching operating conditions between them?

[0023] S22. Weighted fusion of the three types of prior adjacency matrices to construct a comprehensive structural prior matrix for the methanol fuel supply system. : in, , and These are the weight coefficients for physical connection priors, mechanism propagation priors, and control logic priors, respectively. , , , ; To preserve the influence of the node's own state on subsequent representations, a self-connection is added to the comprehensive structure prior matrix to obtain the extended structure matrix. : in, It is the identity matrix; S23. Normalize the extended structure matrix to obtain the normalized structure prior matrix. : in, for The degree matrix; To suppress information propagation between physically reachable nodes, a structural shielding matrix is ​​constructed. : in, The shielding penalty constant is greater than zero. To extend the structure matrix The Middle line, number Elements of the column. Synthetic prior matrix. and structural shielding matrix Together, they form the structural constraint basis for the spatial interaction stage.

[0024] S3. Under the structural mask constraint, perform spatial interaction modeling on the observation sequence and extract spatial coupling feature sequences that conform to the system's physical priors; Step S3 specifically includes: S31. At time , observation vector Mapping nodes to the embedding space yields the node embedding matrix. : in, For node-wise linear mapping functions, Embed items for node locations or component types. , Where F is the spatial feature dimension, and F is the total number of graph nodes; S32. Construct the query matrix, key matrix, and value matrix: in, , and Representing time respectively The query matrix, key matrix, and value matrix; , and These are the learnable parameter matrices for the query matrix, key matrix, and value matrix, respectively. S33. To ensure that both the structural priors and the current degenerate state participate in the spatial correlation calculation, the construction time... Degenerate bias matrix , its first Line number Column elements Represented as: in, , , They are nodes The deviation of pressure, flow rate, and temperature from the healthy baseline. Represents a node To the node Distance along the physical path of the methanol fuel supply system , , This is the bias coefficient; Construct a correlation matrix constrained by both structural priors and degenerate biases. : in, Represents the normalized structure prior matrix; Represents the degenerate bias matrix; Represents the structure mask matrix; and These represent the adjustment coefficients for the structural prior term and the degenerate bias term, respectively; The hidden layer feature dimension of the attention mechanism; Through the above formula, the multi-source structural prior does not exist only as a static topological constraint, but participates in the formation of spatial interaction weights together with the degradation deviation at the current moment, thereby improving the spatial modeling's responsiveness to the actual degradation state of the methanol fuel supply system.

[0025] S34. Regarding the correlation matrix After normalization, the spatial interaction weight matrix is ​​obtained. : Obtain spatial aggregation output : The node-level space is obtained through residual connection and normalization. : S35. Node-level space Expand and map into time step spatial representation vectors : in, This represents the matrix flattening operation; This is a fusion mapping function used to compress the node-level spatial representation into a fixed-dimensional vector; This leads to the formation of a window-level spatial representation sequence. : ; Each represents a continuous segment within the sliding window. The spatial representation vector corresponding to each time step.

[0026] S4. Use a Transformer encoder to perform long-range dependency modeling on the spatially coupled feature sequence to obtain a temporal hidden representation that integrates the global temporal context; Step S4 specifically includes: S41. Spatial representation sequence Injection timing position encoding To form the initial timing input : in, This is the position encoding adjustment coefficient; S42. Input the initial timing sequence Send in The layer-time encoder yields the final temporal hide. : in, Indicates by A time-series coding map formed by cascading Transformer encoders.

[0027] S5. Generate channel-level gating weights based on the spatially coupled feature sequence, and perform structural condition modulation on the temporal hidden representation to achieve spatial prior-guided spatiotemporal feature fusion. Step S5 specifically includes: S51. To enable multi-source structural priors to play a role again in the temporal fusion stage, the node-level space... and normalized structure prior matrix Mapping is performed to obtain the structural prior summary vector. : in, This is a compression mapping function used to compress the spatial representation under structural constraints into a structural prior summary; S52. Spatial representation sequence Compression results and structural prior summary vector The inputs are concatenated to construct a dual-source gated input, and structural conditional gated weights are generated. : in, This represents vector concatenation. For the Sigmoid function, and Learnable parameters; gate weights It originates from both spatial representations obtained after spatial interaction and structural prior summaries, thus essentially reflecting the combined effect of structural information and the current degradation state of the methanol fuel supply system.

[0028] Expand the gating weights to: in To gate the structural conditions weights Replication and expansion along the time dimension The extended gating weight matrix obtained after this step; Simultaneously, the structural prior summary vector is expanded to: in, This indicates a copy-and-expand operation along the time dimension; Indicates the length of the sliding window; To extract the structural prior summary vector Replication and expansion along the time dimension The extended structure summary matrix obtained after this step; S53. Under the combined constraints of the dual-source gating and the structural prior summary, the temporal hiding... Conditional modulation is performed to obtain the fused spatiotemporal representation. : in, and These are the adjustment coefficients for the gated modulation term and the structural prior summary injection term, respectively. Represents element-wise product. The representation layer normalization operation is used to mitigate scale differences between different feature channels.

[0029] Through the above steps, the second injection of multi-source structural priors into the spatiotemporal fusion process is achieved: that is, during the spatial interaction phase, through... , and After directly constraining spatial correlation, structural prior summarization is then applied in the temporal fusion stage. and gating weights The temporal hidden representation is remodulated so that structural priors are carried through the entire spatiotemporal fusion process.

[0030] S6. Perform time aggregation on the fused spatiotemporal features and output the remaining life prediction results of multiple key components in the methanol fuel supply system through the regression layer.

[0031] Step S6 specifically includes: S61. Representation of the merged spacetime Perform weighted time aggregation to obtain a window-level global representation. : in, For the first The aggregate weights corresponding to each time step; Indicates the first [number] within the sliding window The fused spatiotemporal representation corresponding to each time step; satisfying: S62. For key component sets Each key component Each component has a dedicated lifespan. Perform initial remaining lifetime prediction to obtain initial lifetime estimate. : S63. Based on the coupling degradation relationship among multiple key components in the methanol fuel supply system, a system coupling correction term is introduced to correct the initial lifetime estimate, resulting in the final remaining lifetime prediction. : in, Indicates the first Initial remaining life estimates for key components; Indicates key components For key components The coupling influence coefficient; Indicates the first in the methanol fuel supply system Key components at time The initial remaining lifetime estimate; Indicates time Key components The degradation intensity index; This represents the coupling correction mapping function, used to characterize the correction of the remaining life prediction results of the current critical component by the degradation state of related critical components.

[0032] S64, Based on Real Remaining Life Labels Define the joint loss function Through multi-objective collaborative optimization, a deep integration of data-driven prediction and physical prior constraints is achieved; in, Represents the set of key components The number of key components; Indicates the first Loss weights corresponding to each key component; This represents the error loss function between the predicted remaining lifetime and the actual remaining lifetime. Indicates the first The predicted final remaining life of each key component; Indicates the first Key components at time The true remaining life label; and These represent the weight coefficients of the structural consistency constraint term and the coupling constraint term, respectively. This represents the loss of structural consistency. This represents the coupling constraint loss.

[0033] Structural consistency loss The consistency between the constraint space representation and the structural prior is expressed as: in, Denotes the Frobenius norm; Coupling constraint loss The physical rationality used to constrain the lifespan relationship of coupled components is expressed as: in, Indicates key components With key components The allowable lifetime difference threshold between them; This represents a hinge-type penalty function, used to impose a penalty when the lifespan of critical components exceeds a preset physical reasonable range; Indicates the first The predicted final remaining life of each key component.

[0034] In this specification, the same or similar parts between the various embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the descriptions of the embodiments described later are relatively simple, and relevant parts can be referred to the descriptions of the foregoing embodiments.

[0035] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining life of a methanol fuel supply system, characterized in that, Includes the following steps: S1. Obtain multivariate monitoring data of the methanol fuel supply system and construct a sliding window observation sequence to form the original input features for prediction; S2. Based on the physical topology and control logic of the methanol fuel supply system, construct a physical directed graph representing the connection relationship of components and its corresponding structural mask; S3. Under the structural mask constraint, perform spatial interaction modeling on the observation sequence and extract spatial coupling feature sequences that conform to the system's physical priors; S4. Use a Transformer encoder to perform long-range dependency modeling on the spatially coupled feature sequence to obtain a temporal hidden representation that integrates the global temporal context; S5. Generate channel-level gating weights based on the spatially coupled feature sequence, and perform structural condition modulation on the temporal hidden representation to achieve spatial prior-guided spatiotemporal feature fusion. S6. Perform time aggregation on the fused spatiotemporal features and output the remaining life prediction results of multiple key components in the methanol fuel supply system through the regression layer.

2. The method according to claim 1, characterized in that, Step S1 specifically includes: S11. Obtain the methanol fuel supply system at any point in its life cycle. The multivariate observation data below combines external operating condition input variables with system response variables to form an observation vector. : S12. Using a length of The sliding time window truncates the observation vector and uses the current time as the threshold. Construct a window-level input tensor as the window alignment end. .

3. The method according to claim 2, characterized in that, Step S2 specifically includes: S21. Extract the pipeline connection relationships, medium flow paths, pressure propagation paths, heat exchange coupling paths, and valve control links within the methanol fuel supply system. Abstract sensor measurement points and key components into nodes, and construct the following three types of prior adjacency matrices: ； in Represents the prior adjacency matrix of physical connections; Indicates the total number of nodes in the graph; ； in, The prior adjacency matrix represents the mechanism of propagation. ； in, Represents the prior adjacency matrix of the control logic; S22. Weighted fusion of the three types of prior adjacency matrices to construct a comprehensive structural prior matrix for the methanol fuel supply system. : ； in, , and These are the weight coefficients for physical connection priors, mechanism propagation priors, and control logic priors, respectively. , , , ; By adding a self-connect to the comprehensive structure prior matrix, an extended structure matrix is ​​obtained. : ； in, It is the identity matrix; S23. Normalize the extended structure matrix to obtain the normalized structure prior matrix. : ； in, for The degree matrix; To suppress information propagation between physically reachable nodes, a structural shielding matrix is ​​constructed. : ； in, The shielding penalty constant is greater than zero. To extend the structure matrix The Middle line, number The elements of the column.

4. The method according to claim 3, characterized in that, Step S3 specifically includes: S31. At time , observation vector Mapping nodes to the embedding space yields the node embedding matrix. : ； in, For node-wise linear mapping functions, Embed items for node locations or component types. , Where F is the spatial feature dimension, and F is the total number of graph nodes; S32. Construct the query matrix, key matrix, and value matrix: ； in, , and Representing time respectively The query matrix, key matrix, and value matrix; , and These are the learnable parameter matrices for the query matrix, key matrix, and value matrix, respectively. S33. To ensure that both the structural prior and the current degenerate state participate in the spatial correlation calculation simultaneously, the construction time... Degenerate bias matrix , its first Line number Column elements Represented as: ； in, , , They are nodes The deviation of pressure, flow rate, and temperature from the healthy baseline. Represents a node To the node Distance along the physical path of the methanol fuel supply system , , This is the bias coefficient; Construct a correlation matrix constrained by both structural priors and degenerate biases. : ； in, Represents the normalized structure prior matrix; Represents the degenerate bias matrix; Represents the structure mask matrix; and These represent the adjustment coefficients for the structural prior term and the degenerate bias term, respectively; The hidden layer feature dimension of the attention mechanism; S34. Regarding the correlation matrix After normalization, the spatial interaction weight matrix is ​​obtained. : ； Obtain spatial aggregation output : ； The node-level space is obtained through residual connection and normalization. : ； S35. Node-level space Expand and map into time step spatial representation vectors : ； in, This represents the matrix flattening operation; This is a fusion mapping function used to compress the node-level space into a fixed-dimensional vector; This leads to the formation of a window-level spatial representation sequence. : ； Each represents a continuous segment within the sliding window. The spatial representation vector corresponding to each time step.

5. The method according to claim 4, characterized in that, Step S4 specifically includes: S41. Spatial representation sequence Injection timing position encoding To form the initial timing input : ； in, This is the position encoding adjustment coefficient; S42. Input the initial timing sequence Send in The layer-time encoder yields the final temporal hide. : ； in, Indicates by A time-series coding map formed by cascading Transformer encoders.

6. The method according to claim 5, characterized in that, Step S5 specifically includes: S51. To enable multi-source structural priors to play a role again in the temporal fusion stage, the node-level space... and normalized structure prior matrix Mapping is performed to obtain the structural prior summary vector. : ； in, This is a compression mapping function; S52. Spatial representation sequence Compression results and structural prior summary vector The inputs are concatenated to construct a dual-source gated input, and structural conditional gated weights are generated. : ； in, This represents vector concatenation. For the Sigmoid function, and These are learnable parameters; Expand the gating weights to: ； in To gate the structural conditions weights Replication and expansion along the time dimension The extended gating weight matrix obtained after this step; Simultaneously, the structural prior summary vector is expanded to: ； in, This indicates a copy-and-expand operation along the time dimension; Indicates the length of the sliding window. To extract the structural prior summary vector Replication and expansion along the time dimension The extended structure summary matrix obtained after this step; S53. Under the combined constraints of the dual-source gating and the structural prior summary, the temporal hiding... Conditional modulation is performed to obtain the fused spatiotemporal representation. : ； in, and These are the adjustment coefficients for the gated modulation term and the structural prior summary injection term, respectively. Represents element-wise product. The representation layer normalization operation is used to mitigate scale differences between different feature channels.

7. The method according to claim 6, characterized in that, Step S6 specifically includes: S61. Representation of the merged spacetime Perform weighted time aggregation to obtain a window-level global representation. : ； in, For the first The aggregate weights corresponding to each time step; Indicates the first [number] within the sliding window The fused spatiotemporal representation corresponding to each time step; satisfying: ； S62. For key component sets Each key component Each component has a dedicated lifespan. Perform initial remaining lifetime prediction to obtain initial lifetime estimate. : ； S63. Based on the coupling degradation relationship among multiple key components in the methanol fuel supply system, a system coupling correction term is introduced to correct the initial lifetime estimate, resulting in the final remaining lifetime prediction. : ； in, Indicates the first Initial remaining life estimates for key components; Indicates key components For key components The coupling influence coefficient; Indicates the first in the methanol fuel supply system Key components at time The initial remaining lifetime estimate; Indicates time Key components The degradation intensity index; This represents the coupling correction mapping function, used to characterize the correction of the remaining life prediction results of the current critical component by the degradation state of related critical components.

8. The method according to claim 7, characterized in that, Step S6 also includes: S64, Based on Real Remaining Life Labels Define the joint loss function Through multi-objective collaborative optimization, a deep integration of data-driven prediction and physical prior constraints is achieved; ； in, Represents the set of key components The number of key components; Indicates the first Loss weights corresponding to each key component; This represents the error loss function between the predicted remaining lifetime and the actual remaining lifetime. Indicates the first The predicted final remaining life of each key component; Indicates the first Key components at time The true remaining life label; and These represent the weight coefficients of the structural consistency constraint term and the coupling constraint term, respectively. This represents the loss of structural consistency. This represents the coupling constraint loss.

9. The method according to claim 8, characterized in that, Structural consistency loss The consistency between the constraint space representation and the structural prior is expressed as: ； in, Denotes the Frobenius norm; Coupling constraint loss The physical rationality used to constrain the lifespan relationship of coupled components is expressed as: ； in, Indicates key components With key components The allowable lifetime difference threshold between them; This represents a hinge-type penalty function. Indicates the first The predicted final remaining life of each key component.

Images