Cross-system multi-source heterogeneous data fusion and trend prediction method

Abstract

The present application relates to the field of artificial intelligence and data processing technology, especially to a cross-system multi-source heterogeneous data fusion and trend prediction method, comprising: converting multi-source data into dimensionless data based on mean value and scale factor and mapping into a hidden state matrix to construct an initial full connection graph; calculating the norm change rate limit value of the perturbation vector to obtain the maximum Lyapunov index, and activating the physical operator accordingly; extracting the maximum singular value of the Jacobian matrix as the Lipschitz constant, and calculating the partial derivative two norm as the response weight; performing physical disconnection on the edges with out-of-bound response weight with the Lipschitz constant as the truncation threshold to generate a dynamic adjacency matrix as a topology mask; inputting the dynamic adjacency matrix and the hidden state matrix into a graph convolution network, extracting the connected features to calculate the local mask residual, mapping the maximum Lyapunov index into a penalty weight to construct a loss function and performing network parameter update, and outputting the prediction result.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and data processing technology, and in particular to a method for cross-system multi-source heterogeneous data fusion and trend prediction. Background Technology

[0002] In complex industrial scenarios such as health monitoring of deep-sea high-pressure oil and gas pipeline networks and operation and maintenance of large electromechanical equipment, systems typically need to integrate multi-source heterogeneous data such as fluid pressure, pipe wall vibration, and acoustic emission in real time to deduce the trend of state evolution. Under complex operating conditions, the object under test is highly susceptible to external excitation or sudden changes in the internal environment, thereby inducing the underlying physical phase transition.

[0003] Most existing trend prediction schemes employ data fusion models that combine graph neural networks with physical information constraints. However, existing graph network-based physical constraint fusion prediction techniques suffer from the following significant drawbacks in practical applications: Most existing fusion models use static physical partial differential operators for residual constraints. When a physical phase transition occurs in the system, the underlying manifold dynamics undergo fundamental changes, and existing static physical models cannot adaptively follow this abrupt change. This leads to the network still forcibly using invalid rigid physical equations to constrain graph topology derivation during the manifold evolution transition phase. This results in severe conflicts between real data and erroneous physical conservation boundaries, causing the prediction model to exhibit sharp divergence in local physical residuals and the loss function to fall into local convergence deadlock under extremely complex phase transition conditions. Ultimately, this leads to severe lag in trend prediction or even complete model failure. Summary of the Invention

[0004] To overcome the above shortcomings, this invention provides a cross-system multi-source heterogeneous data fusion and trend prediction method, which aims to improve the problem that traditional fusion prediction mostly adopts static physical models, which cannot adapt to complex working conditions and sudden changes, thus causing the model to easily diverge and the prediction to fail.

[0005] This invention provides the following technical solution: a method for cross-system multi-source heterogeneous data fusion and trend prediction, comprising the following steps:

[0006] S1. Based on the mean and scaling factor, the acquired multi-source data is transformed into dimensionless data, and the dimensionless data is mapped into a hidden state matrix to construct an initial fully connected graph of the hidden state matrix.

[0007] S2. Calculate the limit of the norm change rate of the perturbation vector of the hidden state matrix to obtain the maximum Lyapunov exponent; activate partial differential operators from the operator library as physical operators based on the maximum Lyapunov exponent;

[0008] S3. Extract the maximum singular value of the Jacobian matrix of the physical operator under the hidden state matrix as the Lipschitz constant; calculate the L2 norm of the partial derivatives between nodes in the initial fully connected graph as the response weights.

[0009] S4. Using the Lipschitz constant as the truncation threshold, physically disconnect the nodes whose response weights exceed the limit in the initial fully connected graph, and generate a dynamic adjacency matrix as a topological mask.

[0010] S5. Input the dynamic adjacency matrix and the hidden state matrix into the graph convolutional network to output the predicted features; use the dynamic adjacency matrix as a topological mask to filter and extract connectivity features from the predicted features; calculate the local mask residual based on the connectivity features and the physical operator; map the maximum Lyapunov exponent to a penalty weight; construct a loss function using the penalty weight weighted mean square error and the local mask residual; update the network parameters; and output the prediction result.

[0011] By adopting the above technical solution, the Lyapunov exponent activates the physical operator and generates a topological mask with the Lipschitz constant to guide the residual optimization, thereby realizing the dynamic cross-linking of graph topology and physical phase transition. This improves the problem that traditional fusion prediction mostly uses static physical models, which cannot adapt to sudden changes in complex working conditions, resulting in the model being prone to divergence and prediction failure.

[0012] The present invention has the following beneficial effects:

[0013] 1. In this invention, physical operators are activated by Lyapunov exponents and topological masks are generated by Lipschitz constants to guide residual optimization, thereby realizing the dynamic cross-linking of graph topology and physical phase transition. This improves the problem that traditional fusion prediction mostly uses static physical models, which cannot adapt to sudden changes in complex working conditions, resulting in easy model divergence and prediction failure.

[0014] 2. In this invention, by using empirical mean and scaling factor to transform dimensionless data and map it into a hidden state matrix, the physical dimension barrier of cross-system features is eliminated, thereby improving the problem that traditional data fusion mostly adopts direct cascading of original inputs, which causes network input feature distribution imbalance and difficulty in convergence due to huge numerical differences.

[0015] 3. In this invention, by using a dynamic adjacency matrix to perform masking operations to filter and extract physical connectivity features, pseudo-correlation distorted edges that lack causal relationships in the network are accurately filtered out. This improves the problem that traditional graph neural networks mostly use global topology aggregation, which forcibly retains physical noise paths, resulting in distortion in the extraction of local state transmission features.

[0016] 4. In this invention, by mapping the maximum Lyapunov exponent to the penalty weight fusion mask residual and mean square error, the physical regularization constraint strength is adaptively adjusted, thereby improving the problem that traditional neural networks mostly use fixed penalty coefficients, which cannot adjust the weights with the phase transition stage of the system, thus causing the model to get stuck in local optimization deadlock. Attached Figure Description

[0017] Figure 1 The flowchart is a cross-system multi-source heterogeneous data fusion and trend prediction method proposed in this invention;

[0018] Figure 2 This is a flowchart of the multi-source data dimensionlessization and initial fully connected graph construction process of the cross-system multi-source heterogeneous data fusion and trend prediction method proposed in this invention.

[0019] Figure 3 The flowchart shows the Lyapunov phase transition determination and operator activation process of the cross-system multi-source heterogeneous data fusion and trend prediction method proposed in this invention.

[0020] Figure 4 This is a flowchart of the dynamic boundary solution and topology mask hard truncation generation process of the cross-system multi-source heterogeneous data fusion and trend prediction method proposed in this invention.

[0021] Figure 5 This is a flowchart of the joint loss residual optimization and network feedback iteration process of the cross-system multi-source heterogeneous data fusion and trend prediction method proposed in this invention. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] Example 1:

[0024] In the first embodiment of the present invention, the present invention provides a method for cross-system multi-source heterogeneous data fusion and trend prediction, such as... Figures 1-2 As shown, it includes the following steps:

[0025] S1. Based on the mean and scaling factor, the acquired multi-source data is transformed into dimensionless data, and the dimensionless data is mapped into a hidden state matrix to construct an initial fully connected graph of the hidden state matrix.

[0026] Furthermore, in S1, the acquired multi-source data is transformed into dimensionless data based on the mean and scaling factor, including:

[0027] Establish historical time windows corresponding to different physical perception modalities, and extract multi-source data of each physical perception modality within the historical time windows to construct a historical sample set;

[0028] Extract the empirical distribution mean and empirical standard deviation scaling factor of the historical sample set for each physical perception modality;

[0029] The acquired real-time multi-source data is subtracted from the empirical distribution mean of the corresponding physical sensing mode, and the difference is divided by the empirical standard deviation scaling factor to perform data centralization and scaling processing, outputting dimensionless data.

[0030] In S1, the dimensionless data is mapped to the hidden state matrix, and the initial fully connected graph of the hidden state matrix is ​​constructed, including:

[0031] Construct a feature mapping network that includes fully connected layers and non-linear activation functions;

[0032] The dimensionless data extracted from the cross system are concatenated and then input into the feature mapping network to perform forward propagation feature encoding, thereby obtaining the hidden state matrix in a unified representation space.

[0033] Extract the feature vectors of each state dimension in the hidden state matrix as graph topology nodes, and establish initial bidirectional connection edges between all graph topology nodes to generate an initial fully connected graph.

[0034] Specifically, after acquiring the multi-source raw data set of the associated systems, the system establishes a historical time window with a fixed time span for each physical sensing mode, taking into account the objectively existing differences in absolute physical dimensions among cross-system data such as fluid pressure and mechanical vibration velocity contained in the multi-source raw data set. The system extracts multi-source data from each physical sensing mode within the historical time window to construct a historical sample set. Based on the historical sample set, the system extracts the corresponding empirical distribution mean and empirical standard deviation scaling factor for each physical sensing mode. The system subtracts the empirical distribution mean of the corresponding physical sensing mode from the acquired real-time multi-source data, and divides the difference result of the aforementioned subtraction operation by the empirical standard deviation scaling factor to perform data centralization and scaling processing, outputting dimensionless data stripped of absolute physical units. The corresponding data centralization and scaling processing mathematical equations are configured as follows:

[0035] ;

[0036] in Indicates the first Dimensionless data output from a physical sensing mode; Indicates the first The amount of multi-source data input for each physical sensing mode in real-time acquisition state; Indicates the first The empirical distribution mean of each physical sensing modality extracted from the historical sample set; Indicates the first The empirical standard deviation scaling factor extracted from the historical sample set of each physical perception modality.

[0037] By extracting the empirical distribution mean and empirical standard deviation scaling factor to perform dimensionless mapping, the technical defects of heterogeneous sensor data due to huge differences in absolute dimensions and numerical magnitudes are improved. This avoids the problem of underlying physical dimension imbalance and model training non-convergence caused by the polarization of input feature distribution when the prediction model is directly cascaded across system features.

[0038] The system constructs a feature mapping network containing fully connected layers and nonlinear activation functions. Dimensionless data extracted from across systems is concatenated along the data channel dimension. This concatenated dimensionless data is then input into the feature mapping network to perform forward propagation feature encoding, extracting the hidden state matrix within a unified representation space. The mathematical mapping equation for forward propagation feature encoding is defined as:

[0039] ;

[0040] in This represents the hidden state matrix in the unified representation space, which is the output of the forward propagation of the feature mapping network. This represents the global dimensionless data matrix after cross-physical sensing modal extraction and splicing operations; This represents the learnable encoder weight matrix of the fully connected layer in a feature mapping network. This represents the learnable bias vector of the fully connected layer in a feature mapping network. ; represents a nonlinear activation function that introduces the nonlinear mapping capability of the network.

[0041] After obtaining the hidden state matrix, the system extracts the feature vectors of each state dimension in the hidden state matrix as graph topology nodes, and establishes initial connection edges in both directions between all graph topology nodes to generate the initial fully connected graph required for graph convolution calculation.

[0042] By constructing a feature mapping network containing fully connected layers and nonlinear activation functions to complete the unified representation space mapping and initialize the fully interconnected topology, the technical defects of multi-source heterogeneous data in the original acquisition space where features are fragmented and cannot be directly used for graph topology calculation are improved. This avoids the problem of information silos in subsequent trend prediction caused by the difficulty of extracting cross-system related features, and constructs an underlying data architecture that supports cross-physical modality node graph inference.

[0043] like Figure 3As shown, S2, calculate the limit of the norm change rate of the hidden state matrix perturbation vector to obtain the maximum Lyapunov exponent; activate partial differential operators from the operator library as physical operators based on the maximum Lyapunov exponent;

[0044] Furthermore, in S2, the limit of the norm change rate of the hidden state matrix perturbation vector is calculated to obtain the maximum Lyapunov exponent, including:

[0045] Based on preset delay time parameters and embedding dimension parameters, a phase space reconstruction operation is performed on the time series of the hidden state matrix to obtain the reconstructed phase space trajectory;

[0046] In the reconstructed phase space trajectory, an initial state point is specified, and a small perturbation vector is introduced to locate the accompanying state point;

[0047] As time steps advance, the state evolution paths of the initial state point and the accompanying state point are tracked synchronously, and the ratio of the Euclidean distance of the two state evolution paths at the current time step to the initial Euclidean distance is extracted as the norm change rate.

[0048] Logarithmic operations are performed on the norm rate of change, and the time-averaged limit is calculated under the condition that the time step approaches infinity. The time-averaged limit is then used as the maximum Lyapunov exponent.

[0049] In S2, partial differential operators are activated from the operator library as physical operators based on the maximum Lyapunov exponent, including:

[0050] Multiple candidate partial differential operators representing different physical evolution phases are pre-loaded in the operator library, and the corresponding Lyapunov exponent decision interval is labeled for each candidate partial differential operator;

[0051] By comparing the boundary parameters of the maximum Lyapunov index with the decision intervals of each Lyapunov index, the target decision interval covering the maximum Lyapunov index is screened out.

[0052] The candidate partial differential operators that have a unique mapping relationship with the target determination interval are awakened by the conditional indicator function, and the awakened candidate partial differential operators are established as the physical operators that take effect in the current time step.

[0053] Specifically, the system uses the feature-mapped hidden state matrix as the input stream for the underlying computation. Based on preset delay time parameters and embedding dimension parameters, a phase space reconstruction operation is performed on the time series contained in the hidden state matrix to extract the reconstructed phase space trajectory. The computation unit specifies an initial state point in the reconstructed phase space trajectory and introduces a small perturbation vector to locate the corresponding companion state point. As the evolution time step progresses, the system synchronously tracks the state evolution paths of the initial state point and the companion state point in the phase space. The Euclidean distance between the two state evolution paths at the current time step is extracted, and its ratio with the Euclidean distance at the initial time step is calculated to obtain the norm change rate, which characterizes the degree of divergence of the state trajectory. The natural logarithm of the obtained norm change rate is performed to calculate the time average limit value under the condition that the evolution time step tends to infinity, and the maximum Lyapunov exponent is output as a quantitative indicator characterizing the current physical chaos level of the system. The mathematical equations for the above phase space divergence rate derivation are configured as follows:

[0054] ;

[0055] in This represents the maximum Lyapunov exponent of the time-step extrapolation output; Indicates the discrete evolution time step; This represents the initial Euclidean distance of the small perturbation vector introduced at the initial moment; Indicates elapsed time The current Euclidean distance between the initial state point and the accompanying state point after evolution; This represents a function measured in the Euclidean norm. This represents the function for the natural logarithm operation.

[0056] By calculating the maximum Lyapunov exponent of the reconstructed phase space trajectory, the technical deficiency of traditional multi-source fusion data models in lacking a means to quantitatively assess the degree of chaos in the underlying dynamic system is improved. This avoids the practical problem of trend inference lagging when the system faces sudden phase transition due to failure to identify the distortion of the underlying physical manifold in time.

[0057] After obtaining the maximum Lyapunov exponent, the processing logic and data flow of activating partial differential operators from the operator library as physical operators based on the maximum Lyapunov exponent are as follows.

[0058] An operator library is pre-built in the system's underlying memory. This library pre-loads multiple candidate partial differential operators representing different physical evolution phases. The system pre-calibrates the corresponding Lyapunov exponent decision interval for each candidate operator. The processor extracts the maximum Lyapunov exponent from the pre-processed output, compares the boundary parameters of the maximum Lyapunov exponent with each Lyapunov exponent decision interval, and filters out the target decision interval covering the current maximum Lyapunov exponent from the pre-set interval set. A conditional indicator function is used to wake up the candidate partial differential operators that have a unique mapping relationship with the target decision interval. The woken candidate partial differential operators are then established as the physical operators effective at the current time step for output configuration. The control equation for this dynamic operator switching mechanism is defined as follows:

[0059] ;

[0060] in This indicates the physical operator that is activated and effective by the system at the current time step; This represents the total number of candidate partial differential operators in the operator library, and its value is a positive integer greater than 1. Indicates the internal retrieval index number of the candidate partial differential operator; This indicates a conditional indicator function that outputs a value of 1 when the internal comparison logic is true and a value of 0 when the internal comparison logic is false. Indicates the first The lower boundary parameter of the Lyapunov exponential decision interval bound to each candidate partial differential operator; Indicates the first The upper boundary parameter of the Lyapunov exponential decision interval bound to each candidate partial differential operator; Indicates the first retrieval call Candidate partial differential operators.

[0061] By introducing a partial differential operator activation mechanism bound to the dynamic maximum Lyapunov exponent decision interval, the technical defect of static physical constraint model in the evolution of system wide envelope conditions, which causes the residual calculation to deviate from the real physical state due to the use of fixed rigid physical equations, is improved. This avoids the phenomenon of local feature deduction divergence caused by the network prediction features violating the underlying physical conservation boundary when the measured object undergoes a physical modal change.

[0062] like Figure 4 As shown, S3 extracts the maximum singular value of the Jacobian matrix of the physical operator under the hidden state matrix as the Lipschitz constant; calculates the L2 norm of the partial derivatives between nodes in the initial fully connected graph as the response weights;

[0063] Furthermore, in S3, the maximum singular value of the Jacobian matrix of the physical operator under the hidden state matrix is ​​extracted as the Lipschitz constant, including:

[0064] Substitute the state variables of each dimension in the hidden state matrix as independent variables into the physical operator;

[0065] The first-order partial derivatives of the output vector of the physical operator with respect to each independent variable are calculated, and the Jacobian matrix is ​​constructed by arranging all the elements of the first-order partial derivatives obtained from the calculation.

[0066] Perform singular value decomposition on the Jacobian matrix and extract all singular value elements from the resulting diagonal matrix.

[0067] The target singular value with the largest value is retrieved from all singular value elements and defined as the global convergence boundary constant of the system's nonlinear mapping, which is used as the Lipschitz constant.

[0068] In S3, the L2 norm of the partial derivatives between nodes in the initial fully connected graph is calculated as the response weights, including:

[0069] Traverse the initial fully connected graph to extract the first and last graph nodes that have connections;

[0070] Obtain the row vector features of the first and last graph nodes in the hidden state matrix, respectively.

[0071] Input the row vector features of the first graph node into the physical operator to obtain the output response, and solve the partial derivative matrix of the output response with respect to the row vector features of the last graph node.

[0072] Calculate the absolute L2 norm of the partial derivative matrix and define the absolute L2 norm as the response weight of the state evolution between nodes.

[0073] Specifically, the system extracts the hidden state matrix and activated physical operators from the previous steps. Each dimension of the hidden state matrix is ​​set as an independent variable and substituted into the governing equations of the physical operators. The first-order partial derivatives of the physical operator's output vector with respect to each independent variable are calculated, and the Jacobian matrix of the system's multidimensional state mapping is constructed based on the partial derivative results. The mathematical equation configuration for constructing the Jacobian matrix in this step is as follows:

[0074] ;

[0075] in This represents the Jacobian matrix of the physical operator at the current time step in the hidden state space; Indicates the currently active physical operator; This represents the hidden state matrix in the unified representation space; This represents the partial derivative operator.

[0076] The processor performs singular value decomposition (SVD) on the constructed Jacobian matrix, extracting all singular value elements from the resulting diagonal matrix. These singular value elements are then sorted by value, and the target singular value with the largest value is retrieved. This target singular value is designated as the global convergence boundary constant characterizing the system's nonlinear mapping features, thus establishing the Lipschitz constant for this time step in the output. The corresponding extraction equation is defined as:

[0077] ;

[0078] in This represents the Lipschitz constant of the output calculated at the current time step; Indicates the current discrete evolution time step; This represents the operation function for extracting the maximum singular value of a matrix.

[0079] The Lipschitz constant is established based on the maximum singular value of the Jacobian matrix, which solves the technical defect of neural networks lacking a strict mathematical and physical convergence boundary when dealing with nonlinear multi-source heterogeneous data dimensionality reduction mapping. It avoids the practical problems of non-convergence of physical loss function and model training collapse caused by disordered gradient propagation in the subsequent topology derivation process of graph networks. It provides an objective and rigid mathematical benchmark for disconnection and pruning of the underlying topology structure.

[0080] The system iterates through the topology of the initial fully connected graph, extracting the leading and trailing nodes with connectivity. It matches the hidden state matrix to obtain the row vector features mapped to the leading and trailing nodes. These row vector features are then substituted into the physical operator as input to obtain the corresponding output response. The partial derivative matrix of the output response with respect to the row vector features of the trailing nodes is then calculated. The absolute L2 norm of this partial derivative matrix is ​​calculated and defined as the response weight characterizing the true evolution strength between node states. The equation for calculating the L2 norm of the partial derivative is as follows:

[0081] ;

[0082] in This represents the calculated response weights between the nodes in the first and last graphs. This represents the row vector feature corresponding to the first graph node in the hidden state matrix; This represents the row vector feature corresponding to the node in the hidden state matrix of the terminal graph; This represents the measure function for calculating the absolute 2 norm of a matrix.

[0083] By extracting the L2 norm of the partial derivatives between nodes as response weights, the technical defect of traditional graph neural networks that rely solely on the statistical similarity of data surface to construct node connection weights is overcome. This removes the transmission interference caused by non-causal pseudo-correlation feature edges in multi-source sensing data, ensuring that the quantitative evaluation of the connection strength between graph topology nodes fully conforms to the underlying real physical partial differential transmission law.

[0084] like Figure 4 As shown, S4, using the Lipschitz constant as the truncation threshold, physically disconnects the edges of nodes whose response weights exceed the limit in the initial fully connected graph, generating a dynamic adjacency matrix as a topological mask.

[0085] Furthermore, in S4, using the Lipschitz constant as the truncation threshold, physical disconnections are performed on the nodes whose response weights exceed the limit in the initial fully connected graph, generating a dynamic adjacency matrix as a topological mask, including:

[0086] Construct an initial adjacency matrix with the same node size as the initial fully connected graph;

[0087] Compare the response weights of the edges connecting each node in the initial fully connected graph with the Lipschitz constant values;

[0088] For node connections with response weights not less than the Lipschitz constant, the matrix elements corresponding to the node connections in the initial adjacency matrix are set to zero to perform physical disconnection.

[0089] For node connections with response weights less than the Lipschitz constant, the corresponding matrix element in the initial adjacency matrix is ​​set to one to preserve the physical path, and a dynamic adjacency matrix is ​​output based on all updated matrix elements.

[0090] Specifically, the system extracts the initial fully connected graph topology data, including the response weights between each node, and the Lipschitz constants output from the pre-calculated values. An initial adjacency matrix with the same node size as the initial fully connected graph is constructed in the computation space. The processor extracts the response weights carried by the edges of each node in the initial fully connected graph, comparing the extracted response weight values ​​with the Lipschitz constant values. For node edges with response weights not less than the Lipschitz constant, the system determines that the path of this edge has undergone physical topology distortion, and forcibly sets the corresponding matrix element in the initial adjacency matrix to zero to perform a physical disconnection. For node edges with response weights less than the Lipschitz constant, it determines that the edge conforms to the physical law constraint boundary, and sets the corresponding matrix element in the initial adjacency matrix to one to preserve the actual physical path. After completing the traversal and comparison of all node edges in the graph, the system outputs a dynamic adjacency matrix as the subsequent network topology mask based on all updated matrix elements.

[0091] The update control equations for the above topological truncation and dynamic adjacency matrix generation are defined as follows:

[0092] ;

[0093] in This represents the first element in the output dynamic adjacency matrix. The graph node and the first Assigning scores to the matrix elements between graph nodes; In the initial fully connected graph, the first... The graph node and the first The actual response weights extracted between each graph node; This represents the Lipschitz constant extracted at the current time step as the global convergence boundary constant of the nonlinear mapping; a value of 1 indicates that the physical path between the corresponding nodes is preserved; a value of 0 indicates that the physical connection between the corresponding nodes is broken.

[0094] By introducing the Lipschitz constant as a rigid truncation threshold to filter the initial fully connected graph, this method overcomes the technical flaw of traditional graph neural networks that forcibly retain fully connected features when processing dimensionality reduction mapping of high-dimensional heterogeneous data, leading to distortion and disorder of the underlying physical transmission paths. This step cuts off pseudo-correlated edges that only have accidental correlations on the data surface but lack the support of underlying physical differential laws, removes physical noise paths in the input graph structure, and ensures that the topology passed to the spatiotemporal graph convolutional network absolutely obeys the underlying physical conservation laws. This lays an objective and realistic graph structure foundation for subsequent steps to accurately extract effective connected features and calculate local physical residuals.

[0095] like Figure 5 As shown in step S5, the dynamic adjacency matrix and hidden state matrix are input into the graph convolutional network to output the predicted features; the dynamic adjacency matrix is ​​used as a topological mask to filter the predicted features and extract connected features; the local mask residual is calculated based on the connected features and physical operators; the maximum Lyapunov exponent is mapped to the penalty weight; the weighted mean square error of the penalty weight and the local mask residual are used to construct the loss function and perform network parameter updates, and the prediction result is output.

[0096] Furthermore, in S5, the dynamic adjacency matrix and hidden state matrix are input into the graph convolutional network to output predicted features; the dynamic adjacency matrix is ​​used as a topological mask to perform a masking operation on the predicted features to extract connected features; and the mask residual is calculated based on the connected features and physical operators, including:

[0097] Obtain the degree matrix of the dynamic adjacency matrix, and combine the degree matrix, dynamic adjacency matrix, and graph convolution learnable weights to perform a neighborhood aggregation-based graph convolution operation on the hidden state matrix to output preliminary predicted features.

[0098] The dynamic adjacency matrix is ​​used as a binarization filter, and a matrix dot product-level masking operation is performed between the dynamic adjacency matrix and the predicted features.

[0099] The feature nodes in the physical disconnection state are filtered out by masking, while the feature nodes in the physical path are retained and aggregated to generate connected features.

[0100] The connectivity features are substituted as independent variables into the control equations of the physical operator to solve for the partial derivative residuals of the equations, and the mask residuals are output.

[0101] In S5, the maximum Lyapunov exponent is mapped to the penalty weights, the mean squared error and mask residual are fused to construct the loss function, and gradient updates are performed. The output prediction results include:

[0102] The maximum Lyapunov exponent is input into the exponential scaling function to calculate the penalty weight that limits the physical regularization intensity.

[0103] The mean square error is obtained by calculating the mean square error between the predicted and extrapolated values ​​of connectivity features and the actual value of the labels collected by the system.

[0104] The penalty weights are used to perform a penalty coefficient multiplication on the mask residuals, and the result of the multiplication is arithmetically added to the mean square error to generate a loss function;

[0105] The network connection weights are updated based on the gradient direction of the partial derivative of the loss function with respect to the network parameters, and the iteration is repeated until the loss function converges and the prediction result is output.

[0106] Specifically, the system extracts the dynamically generated adjacency matrix and the hidden state matrix in multi-dimensional space as the underlying input data stream of the graph convolutional network. The computation unit performs row element summation on the dynamically generated adjacency matrix to obtain the degree matrix representing the node connection density. Combining the degree matrix, the dynamically generated adjacency matrix, and the learnable graph convolution weights contained in the network itself, the system performs graph convolution operations on the hidden state matrix based on neighborhood aggregation feature fusion, outputting the predicted features of the initial evolution stage. The corresponding spatiotemporal graph convolution mathematical evolution equation is configured as follows:

[0107] ;

[0108] in This represents the predicted features output by the network after one graph convolution inference. This represents the hidden state matrix of the input at the current discrete time step; This represents the dynamic adjacency matrix after adding the self-loop attribute; This represents the degree matrix corresponding to the dynamic adjacency matrix; This indicates that the graph convolutions within a graph convolutional network can be learned weights; This represents the activation function for a nonlinear mapping.

[0109] After obtaining the initial prediction features, the system sets the dynamic adjacency matrix of the binarized attributes as a spatial structure mask filter. A matrix-level masking operation is performed between the dynamic adjacency matrix and the initial prediction features at the corresponding node dimensions. The masking operation automatically filters out feature nodes in the dynamic adjacency matrix whose element values ​​are zero (indicating physical disconnection), and forcibly retains node features whose element values ​​are one (indicating effective physical pathways). After feature aggregation, the final connectivity features are generated. The extracted connectivity features are then substituted as independent variables into the control equations of the currently active physical operator to perform a numerical solution operation for the partial derivative residuals of the equations, outputting a quantitative index of the degree of conflict between the underlying physical laws of the system: the mask residual.

[0110] This approach upgrades the dynamic adjacency matrix to a low-level mask filter and calculates local mask residuals, overcoming the technical shortcoming of traditional physical information neural networks where forced calculation of global residuals on the global topology graph leads to the smoothing of local high-frequency errors. This solution establishes the boundary for residual calculation, avoiding gradient divergence caused by nodes without physical causal relationships participating in regularization constraints.

[0111] The processor receives the maximum Lyapunov exponent from the preceding calculation, inputs this exponent into a built-in exponent scaling function for adaptive scaling transformation, and calculates the penalty weights specifically used to limit the strength of physical regularization. It extracts the connectivity features from the preceding output and decodes them to obtain their physical inference values. The mean squared error between these inference values ​​and the actual collected label values ​​is calculated to establish the mean squared error of the data fitting dimension.

[0112] The system uses penalty weights to perform a penalty coefficient multiplication on the mask residuals to scale the residual magnitude. The scaled mask residual values ​​are then arithmetically summed with the mean squared error value to generate the joint loss function that guides network evolution. The analytical equation for this loss function is defined as:

[0113] ;

[0114] in This represents the joint loss function generated by fusing the error branches; This represents the total number of data samples whose labels were actually collected by the system participating in model training. The predicted and extrapolated numerical value representing the connectivity features of a single sample channel output; This represents the actual tag value collected by the corresponding single sample channel system. This represents the penalty weight that limits the physical regularization strength generated by the maximum Lyapunov exponential mapping transformation; This represents the mask residual obtained by solving the output under the constraints of the currently active physical operator.

[0115] After calculating the joint loss function, which includes both data-driven and physics-driven errors, the computational unit performs partial derivative calculations with respect to the network parameters to obtain the gradient descent direction for backpropagation. The calculated partial derivative gradient directions are then used to continuously update the network connection weights until the joint loss function converges, at which point the network training process stops and a stable prediction result is output.

[0116] By directly mapping the maximum Lyapunov exponent, which characterizes the critical state of phase transition, to the penalty weight to construct the joint loss function, the technical shortcoming of traditional loss optimization methods, which is that the fixed penalty coefficient makes it impossible to adaptively and dynamically adjust the physical constraint strength as the system moves across dynamic operating conditions, is overcome. This prevents the prediction model from falling into local convergence deadlock due to excessively strong or weak physical constraints during the transition phase of manifold evolution, and thus establishes the global optimization capability of the prediction system.

[0117] Example 2:

[0118] In the health monitoring scenario of deep-sea high-pressure oil and gas pipeline networks, the system needs to integrate multi-source heterogeneous data such as fluid pressure, pipe wall vibration, and acoustic emission in real time to deduce pipeline leakage trends. Under complex seabed conditions, when pipelines encounter ocean current vibrations or internal scaling, they are prone to physical phase transitions, such as the transition from elastic micro-deformation to plastic yielding of the pipe material or the transition from laminar to turbulent flow of the internal fluid. Existing graph network-based physical constraint fusion prediction techniques have serious defects when facing such phase transition evolution: the static physical partial differential operators they use cannot adapt to the new manifold dynamics after the phase transition, and when processing cross-modal data dimensionality reduction mapping, they often generate false topological connections that completely violate the underlying physical causal transmission laws due to overfitting high-frequency noise. This rigid and erroneous graph topology, superimposed on the rigid physical equations of failure, directly leads to the sharp divergence of physical residuals of local nodes and the gradient explosion of the joint loss function when predicting the critical failure trend of the pipeline network, ultimately causing model deadlock and serious underreporting and prediction lag of catastrophic pipeline rupture accidents. To address the aforementioned problems, this invention provides a method for cross-system multi-source heterogeneous data fusion and trend prediction, the structure of which is as follows: Figure 1 As shown. The specific implementation process of this method is as follows:

[0119] By performing dimensionless processing on multi-source data based on mean and scale factor, the absolute physical dimension barrier between cross-modal sensing data is eliminated, laying a benchmark for the unified mapping of hidden state matrices. This method calculates the maximum Lyapunov exponent of the hidden state evolution trajectory, accurately detecting the physical phase transition critical point of the system, and adaptively activating a partial differential operator from the operator library that strictly matches the current manifold, overcoming the failure problem of static rigid equations under wide envelope conditions. Furthermore, the maximum singular value of the Jacobian matrix of this physical operator is extracted as the Lipschitz constant, establishing the global convergence boundary for the mapping of nonlinear systems. Using this boundary as a rigid threshold, physically distorted edges with excessive response weights in the initial fully connected graph are cut off to generate a dynamic adjacency matrix, filtering out pseudo-correlation noise at the graph structure level. Finally, the dynamic adjacency matrix is ​​subdivided into a topological mask to extract physical connectivity features, and the penalty weight of the mask residual is adaptively adjusted using the maximum Lyapunov exponent to construct a joint loss function. This technical solution deeply locks the physical topology, phase transition adaptive operator, and network error during backpropagation, avoiding gradient divergence of the model under extremely complex phase transition conditions and ensuring that the prediction results strictly obey the underlying physical causal and dynamic conservation laws.

[0120] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for cross-system multi-source heterogeneous data fusion and trend prediction, characterized in that, Includes the following steps: S1. Based on the mean and scaling factor, the acquired multi-source data is transformed into dimensionless data, and the dimensionless data is mapped into a hidden state matrix to construct an initial fully connected graph of the hidden state matrix. S2. Calculate the limit of the norm change rate of the perturbation vector of the hidden state matrix to obtain the maximum Lyapunov exponent; activate partial differential operators from the operator library as physical operators based on the maximum Lyapunov exponent; S3. Extract the maximum singular value of the Jacobian matrix of the physical operator under the hidden state matrix as the Lipschitz constant; The L2 norm of the partial derivatives between nodes in the initial fully connected graph is calculated as the response weights; S4. Using the Lipschitz constant as the truncation threshold, physically disconnect the nodes whose response weights exceed the limit in the initial fully connected graph, and generate a dynamic adjacency matrix as a topological mask. S5. Input the dynamic adjacency matrix and the hidden state matrix into the graph convolutional network to output the predicted features; use the dynamic adjacency matrix as a topological mask to filter and extract connectivity features from the predicted features; calculate the local mask residual based on the connectivity features and the physical operator; map the maximum Lyapunov exponent to a penalty weight; construct a loss function using the penalty weight weighted mean square error and the local mask residual; update the network parameters; and output the prediction result.

2. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S1, the process of converting the acquired multi-source data into dimensionless data based on the mean and scaling factor includes: Establish historical time windows corresponding to different physical perception modes, and extract the multi-source data of each physical perception mode within the historical time windows to construct a historical sample set; Extract the empirical distribution mean and empirical standard deviation scaling factor of the historical sample set for each physical perception modality; The multi-source data in the acquired real-time state is subtracted from the mean of the empirical distribution of the corresponding physical sensing mode, and the difference is divided by the empirical standard deviation scaling factor to perform data centralization and scaling processing, and the dimensionless data is output.

3. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S1, mapping the dimensionless data to a hidden state matrix and constructing an initial fully connected graph of the hidden state matrix includes: Construct a feature mapping network that includes fully connected layers and non-linear activation functions; The dimensionless data extracted from the cross-system is concatenated and input into the feature mapping network to perform forward propagation feature encoding, thereby obtaining the hidden state matrix in a unified representation space; Extract the feature vectors of each state dimension in the hidden state matrix as graph topology nodes, and establish initial bidirectional connection edges between all graph topology nodes to generate the initial fully connected graph.

4. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S2, calculating the limit of the norm change rate of the hidden state matrix perturbation vector to obtain the maximum Lyapunov exponent includes: Based on preset delay time parameters and embedding dimension parameters, a phase space reconstruction operation is performed on the time series of the hidden state matrix to obtain the reconstructed phase space trajectory; An initial state point is specified in the reconstructed phase space trajectory, and a small perturbation vector is introduced to locate the accompanying state point; The state evolution paths of the initial state point and the accompanying state point are tracked synchronously as the time step progresses, and the ratio of the Euclidean distance of the two state evolution paths at the current time step to the initial Euclidean distance is extracted as the norm change rate. Logarithmic operation is performed on the norm change rate, and the time average limit value is calculated under the condition that the time step tends to infinity. The time average limit value is used as the maximum Lyapunov exponent.

5. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S2, the step of activating a partial differential operator from the operator library as a physical operator based on the maximum Lyapunov exponent includes: The operator library is preloaded with multiple candidate partial differential operators that characterize different physical evolution phases, and the corresponding Lyapunov exponent determination interval is assigned to each candidate partial differential operator. By comparing the boundary parameters of the maximum Lyapunov index with the determination intervals of each Lyapunov index, a target determination interval covering the maximum Lyapunov index is identified. The candidate partial differential operators that have a unique mapping relationship with the target determination interval are awakened by the conditional indication function, and the awakened candidate partial differential operators are established as the physical operators that take effect at the current time step.

6. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S3, extracting the maximum singular value of the Jacobian matrix of the physical operator under the hidden state matrix as the Lipschitz constant includes: Substitute each dimension of the hidden state matrix as an independent variable into the physical operator; The first-order partial derivatives of the output vector of the physical operator with respect to each independent variable are calculated, and the Jacobian matrix is ​​constructed by arranging all the elements of the first-order partial derivatives obtained from the calculation. Perform singular value decomposition on the Jacobian matrix to extract all singular value elements in the resulting diagonal matrix. The target singular value with the largest value is retrieved from all singular value elements, and the target singular value is defined as the global convergence boundary constant of the system's nonlinear mapping, which is used as the Lipschitz constant.

7. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S3, calculating the L2 norm of the partial derivatives between nodes in the initial fully connected graph as response weights includes: Traverse the initial fully connected graph to extract the first and last graph nodes with connection relationships; Obtain the row vector features corresponding to the first and last graph nodes in the hidden state matrix, respectively. The row vector features of the first end graph node are input into the physical operator to obtain the output response, and the partial derivative matrix of the output response with respect to the row vector features of the last end graph node is solved. Calculate the absolute L2 value of the partial derivative matrix, and define the absolute L2 value as the response weight of the state evolution between nodes.

8. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S4, the step of physically disconnecting the nodes whose response weights exceed the limit in the initial fully connected graph using the Lipschitz constant as a truncation threshold, and generating a dynamic adjacency matrix as a topological mask, includes: Construct an initial adjacency matrix with the same node size as the initial fully connected graph; The response weights of the edges connecting each node in the initial fully connected graph are compared with the Lipschitz constant values. For node connections whose response weight is not less than the Lipschitz constant, the matrix elements corresponding to the node connections in the initial adjacency matrix are set to zero to perform physical disconnection. For node connections whose response weights are less than the Lipschitz constant, the corresponding matrix elements in the initial adjacency matrix are set to one to preserve the physical path, and the dynamic adjacency matrix is ​​output based on all updated matrix elements.

9. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S5, the step of inputting the dynamic adjacency matrix and the hidden state matrix into the graph convolutional network to output predicted features; using the dynamic adjacency matrix as a topological mask to perform a masking operation on the predicted features to extract connectivity features; and calculating the mask residual based on the connectivity features and the physical operator includes: Obtain the degree matrix of the dynamic adjacency matrix, and combine the degree matrix, the dynamic adjacency matrix, and the graph convolution learnable weights to perform a graph convolution operation based on neighborhood aggregation on the hidden state matrix, and output the initial predicted features. The dynamic adjacency matrix is ​​used as a binary filter, and the dynamic adjacency matrix and the predicted features are subjected to a matrix dot product-level masking operation. The masking operation filters out feature nodes that are physically disconnected, retains node features that are physically connected, and aggregates them to generate the connectivity feature. The connectivity feature is substituted as an independent variable into the control equation of the physical operator to solve the partial derivative residuals of the equation, and the mask residual is output.

10. The method for cross-system multi-source heterogeneous data fusion and trend prediction according to claim 1, characterized in that, In S5, the step of mapping the maximum Lyapunov exponent to penalty weights, fusing the mean squared error and the mask residual to construct a loss function and performing gradient updates, and outputting the prediction result includes: The maximum Lyapunov exponent is input into the exponential scaling function to calculate the penalty weight that limits the physical regularization intensity. The mean square error is obtained by calculating the mean square error between the predicted and extrapolated values ​​of the connectivity features and the actual values ​​of the labels collected by the system. The penalty weights are used to perform a penalty coefficient multiplication on the mask residuals, and the multiplication result is arithmetically summed with the mean square error to generate the loss function; The network connection weights are updated based on the gradient direction of the partial derivative of the loss function with respect to the network parameters, and the iteration is repeated until the loss function converges and the prediction result is output.

Images