Integrated simulation evaluation method for complex equipment system

Abstract

This invention proposes an integrated simulation evaluation method for complex equipment systems, relating to the field of industrial simulation technology. It constructs a directed coupled evaluation network comprising simulation model nodes, actual equipment data interface nodes, and environmental model nodes. During the operation of this evaluation network, dynamic field information reflecting the confidence status of each node is calculated in parallel. Based on this dynamic field information, adaptive fusion of virtual simulation data and actual equipment data is performed in a distributed manner at each node, outputting an integrated evaluation result. Online parameter correction is triggered based on the uncertainty propagation revealed by the dynamic field information. Based on the historical records of the average steady-state accuracy potential difference and average uncertainty flow of each link in the evaluation network, high-confidence simulation logic substructures are automatically identified and compiled into reconfigurable hardware logic. This improves the real-time performance, accuracy, and reusability of the evaluation.

Description

Technical Field

[0001] This application relates to the field of industrial simulation technology, and in particular to an integrated simulation evaluation method for complex equipment systems. Background Technology

[0002] In the development and performance verification of complex equipment systems (such as aviation, aerospace, ships, missiles, high-speed rail, etc.), a segmented iterative approach combining virtual simulation and physical testing is commonly used. However, with the continuous increase in system functional density and coupling, this traditional approach has revealed the following shortcomings: Virtual simulation models and physical equipment data are often compared offline, making it difficult for simulation outputs to promptly incorporate real-world operating conditions, resulting in insufficient model credibility verification. Uncertainties within the system (such as sensor noise, model simplification errors, and interface delays) propagate and amplify along information chains, and existing evaluation systems struggle to dynamically describe and quantify this uncertainty propagation process at the network level. Evaluation processes are often conducted offline after the fact, failing to identify key error sources and automatically adjust model parameters in real-time on-site, leading to a disconnect between evaluation and optimization. High-confidence simulation logic lacks automatic identification and solidification mechanisms, hindering rapid reuse and hardware-level acceleration in subsequent similar tasks, resulting in significant repetitive computational overhead. Complex equipment system simulation involves large-scale interactive computations across multiple domains and scales. Existing methods lack targeted optimization for high-confidence simulation logic, still employing a full-process software simulation model, leading to high computational loads and long execution times, failing to meet the demands of rapid iterative equipment evaluation. Furthermore, indiscriminate resource allocation wastes hardware resources, further restricting simulation efficiency improvements. Summary of the Invention

[0003] To address the aforementioned issues, this application provides an integrated simulation evaluation method for complex equipment systems. Through a dynamic evaluation network, it achieves reliable fusion and real-time correction of simulation data and measured data, and solidifies high-confidence simulation logic into reconfigurable hardware modules, thereby improving the real-time performance, accuracy, and reusability of the evaluation.

[0004] This invention provides an integrated simulation and evaluation method for complex equipment systems, including: Construct a directed coupled evaluation network that includes simulation model nodes, actual equipment data interface nodes, and environment model nodes; During the operation of the evaluation network, dynamic field information reflecting the credibility status of each node is computed in parallel, and based on the dynamic field information, the virtual simulation data and actual equipment data are adaptively fused in a distributed manner at each node to output an integrated evaluation result. Based on the uncertainty propagation revealed by the dynamic field information, online parameter correction of relevant simulation model nodes is triggered and executed; Based on the historical records of the average steady-state accuracy difference and average uncertainty flow of each link in the evaluation network, high-confidence simulation logic substructures are automatically identified, and the substructures are compiled into reconfigurable hardware logic.

[0005] Preferably, constructing the directed coupled evaluation network includes: Based on the interaction relationship between the simulation model, the actual equipment data interface, and the environment model, the directed coupled evaluation network is constructed. The actual equipment data interface node is defined as the reference node of the network, and it is given a constant highest confidence benchmark. Edge weight coefficients are configured on the directed connection edges between nodes, and the edge weight coefficients are used to quantify the uncertainty propagation characteristics between nodes.

[0006] Preferably, the dynamic field information includes the local uncertainty intensity of each node and the precision potential field of the entire network; The accuracy potential value of the reference node is a constant high value; the accuracy potential value of non-reference nodes is iteratively updated according to the state of their upstream nodes; at each update time, the update rule is as follows: first, select the highest value from the accuracy potential values ​​of all upstream nodes at the previous time as the basis, and then perform information decay on the highest value; then subtract the joint loss caused by the uncertainty flow transmitted from all upstream nodes and weighted by their local uncertainty intensity; finally, compare the calculation result with a preset lower limit value of accuracy potential, and take the larger one as the accuracy potential value of the node at the current time.

[0007] Precision potential of non-reference nodes at the next time step Obtained through the following formula: in, The inherent attenuation coefficient of information transmission; 0 < <1; It is the maximum value of the precision potential among all upstream nodes of node i; Let be the intensity of the local uncertainty of upstream node j at time t; Let be the uncertainty propagation coefficient for the edge from node j to node i; The uncertainty loss coefficient; This is the lower limit of the accuracy potential common to all non-reference nodes; >0.

[0008] Preferably, the step of performing adaptive data fusion in a distributed manner across nodes based on dynamic field information includes: For any node to be evaluated, gather all source paths that can provide data to it; For each source path, its path contribution weight is calculated. This weight is positively correlated with the precision potential of the path terminal node and negatively correlated with the intensity of the accumulated uncertain flow on the path. Based on the contribution weight of each path, the data from different paths are weighted and fused to obtain the final output of the node.

[0009] Preferably, the aggregation of all source paths that can provide data to it includes: Determine the constraint rules for reverse path search based on at least one preset path filtering condition; Starting from the current node to be evaluated, a reverse graph traversal is performed along the directed edges of the directed coupled evaluation network according to the constraint rules. Nodes that are reached and meet the preset filtering conditions are identified as valid source nodes; the directed path from the valid source node to the current node to be evaluated is identified as a valid source path. After performing reverse tracing and obtaining the initial set of valid source paths, multiple paths pointing to the same valid source node are filtered, and only the path with the minimum cumulative uncertainty flow intensity is retained. The path selection criteria include maximum search depth, node type restrictions, precision potential threshold, and key node set. The cumulative uncertain flow intensity is determined by summation or maximum value attenuation.

[0010] Preferably, let R be the set of all valid source paths, and for each path r∈R in the set R, its contribution weight to the current node to be evaluated is... Obtained using the following formula: R is the set of all valid source paths; src(r) represents the starting node of path r; γ is the uncertainty penalty coefficient, 0 < γ < 2, and is a constant greater than 0; Let src(r) be the precision potential value of node src(r) at the current time. The accumulated uncertain flow intensity on path r; the summation in the denominator traverses all paths p in set R; It is a constant, 0 < <0.01.

[0011] Preferably, the step of triggering online parameter correction based on dynamic field information includes: Real-time monitoring of the intensity of uncertain flow from the target simulation model node to its key downstream node or reference node, wherein the intensity of uncertain flow is obtained based on the product of the local uncertainty intensity of the target simulation model node and the uncertainty propagation coefficient of the connecting edge; When the intensity of the uncertain flow exceeds a preset threshold, online correction of the target simulation model nodes is triggered; During the calibration process, the model parameters are adjusted based on the gradient information of the node's output error with respect to its own parameters, and in combination with the direction of the precision potential gradient between the node and downstream nodes.

[0012] Preferably, the adjustment of model parameters includes: A continuously differentiable correction guidance factor is determined based on the precision potential difference from the current node to its critical downstream node. Based on the gradient of the correction guidance factor and the loss function constructed based on the node output error, the parameter update amount is calculated and the node parameters are updated. After the parameter update is completed, a correction flow associated with the correction behavior is generated, which is used to propagate along the directed coupled evaluation network and affect the accuracy potential update of downstream nodes.

[0013] Preferably, the step of automatically identifying high-confidence simulation logic substructures and compiling the substructures into reconfigurable hardware logic includes: Based on the operational data of the evaluation network after reaching a statistically stable state, the steady-state precision potential difference and average uncertainty flow of each link are calculated; wherein, the steady-state precision potential difference is defined as the difference in precision potential values ​​of the nodes at both ends of the link in the steady state, and the average uncertainty flow is defined as the average value of the uncertainty intensity transmitted through the link during the statistical period. Based on the calculated steady-state accuracy potential difference and average uncertainty flow of each link, the maximum connected subnetwork that meets the preset condition is identified from the evaluation network as the high-confidence causal subnetwork. The simulation logic corresponding to the high-confidence causal subnet, including the topological connection relationship between its nodes and the fixed parameters of each node, is compiled into hardware description language code, and a hardware acceleration module that can be configured and run on a programmable logic device is generated based on the code.

[0014] Preferably, the preset conditions include: the average steady-state precision potential difference of all links within the sub-network is higher than a first preset threshold, and the average uncertainty flow of all boundary links of the sub-network is lower than a second preset threshold; the boundary link is a link connecting the internal nodes of the sub-network with the external nodes.

[0015] A directed coupled network is constructed, linking the simulation model, actual equipment data interface, and environment model. Actual equipment data is used as the baseline node, assigned a constant highest confidence level, providing a reliable reference for data fusion and correction, thus reducing system bias at the source. Dynamic field information reflects the node confidence status in real time. During data fusion, path weights are positively correlated with accuracy potential and negatively correlated with uncertainty flow, ensuring that highly reliable data dominates the fusion results and improving the accuracy of the evaluation results. Online parameter correction is triggered based on the uncertainty propagation revealed by the dynamic field. Model parameters are adjusted by combining the accuracy potential gradient direction and output error gradient, while the correction flow propagates the positive effect throughout the network, continuously optimizing the model output accuracy and suppressing uncertainty diffusion. Data fusion employs a reverse tracing and path filtering mechanism. Constraints such as maximum search depth and accuracy potential threshold, combined with redundant path deduplication, reduce invalid computations and improve fusion efficiency. High-confidence causal subnets are automatically identified and compiled into reconfigurable hardware logic, significantly reducing the runtime of similar simulation tasks and solving the large-scale computational efficiency bottleneck of complex equipment simulation. The hardware acceleration module can be reused in similar evaluation scenarios without repeated compilation and parameter configuration, shortening the simulation iteration cycle and improving the overall process efficiency of complex equipment development and performance evaluation. The accuracy potential value of non-reference nodes is dynamically updated iteratively. Parameter correction and data fusion are based on real-time dynamic field information, enabling the system to adapt to changes in network state and fluctuations in data quality, and to adapt to the dynamic evolution characteristics of complex equipment simulation. The multi-layer fault-tolerant design (accuracy potential lower limit, path filtering and deduplication, uncertainty threshold triggered correction) effectively addresses issues such as data deviation and link fluctuations, improving the system's stability and anti-interference capability under complex working conditions. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the integrated simulation evaluation method for complex equipment systems according to an embodiment of this application. Detailed Implementation

[0017] The present application will now be further described in conjunction with the accompanying drawings and specific embodiments. It should be noted that, without conflict, the various embodiments or technical features described below can be arbitrarily combined to form new embodiments.

[0018] Example 1, see Figure 1 This embodiment provides an integrated simulation and evaluation method for complex equipment systems, including: Construct a directed coupled evaluation network that includes simulation model nodes, actual equipment data interface nodes, and environment model nodes; During the operation of the evaluation network, dynamic field information reflecting the credibility status of each node is computed in parallel, and based on the dynamic field information, the virtual simulation data and actual equipment data are adaptively fused in a distributed manner at each node to output an integrated evaluation result. Based on the uncertainty propagation revealed by the dynamic field information, online parameter correction of relevant simulation model nodes is triggered and executed; Based on the historical records of the average steady-state accuracy difference and average uncertain flow of each link in the evaluation network, high-confidence simulation logic substructures are automatically identified, and the substructures are compiled into reconfigurable hardware logic to accelerate the execution of subsequent similar evaluation tasks.

[0019] The simulation model, the real equipment data interface, and the environment model are abstracted into nodes and connected according to real data / logical relationships to form a computing network that can process both virtual and real data simultaneously.

[0020] During network operation, the precision potential and uncertainty flow of each node are calculated in parallel. Using these two dynamic fields, the network dynamically and adaptively determines which data—simulated or measured—is more reliable and what weight it should have at each node, thereby fusing them to arrive at an optimal local evaluation result.

[0021] The system automatically monitors the propagation path and intensity of uncertainties in the network. When a simulation model is found to be the main source of error, it immediately triggers online correction of its internal parameters to make its output closer to real data, thereby improving the overall reliability of the assessment from the root.

[0022] After the network is running stably, it automatically identifies those local computational logics that have been repeatedly verified and have a high degree of confidence, and solidifies them into dedicated hardware modules for direct use by subsequent similar tasks, thereby achieving a leap in evaluation speed.

[0023] In one possible implementation, constructing the directed coupled evaluation network includes: The simulation model unit, the actual equipment data acquisition interface, and the environment model unit are defined as nodes, and a directed coupled network is constructed according to the data / logic dependency relationship. The actual equipment data acquisition interface is defined as the baseline node only, and it is given a constant highest confidence baseline. Edge weight coefficients are configured on the directed connection edges between nodes, and the edge weight coefficients are used to quantify the uncertainty propagation characteristics between nodes.

[0024] The specific implementation steps include: Step 1: Identify and Define Nodes: Define each simulation sub-model in a complex equipment system that has clear inputs and outputs and can perform independent calculations (such as propulsion system model, flight control model, radar cross section calculation model) as a node, with its input and output variables serving as the interface for interaction between the node and the outside world; define each hardware interface or data stream that directly collects data from physical equipment (such as sensors, buses) as a baseline node. For example, GPS positioning data interface, engine speed sensor interface; define models that simulate the external environment (such as atmospheric model, terrain elevation model, electromagnetic interference model) as a node.

[0025] Step 2: Construct directed edges based on interaction relationships: Iterate through the input and output variables of all nodes. If the output variable of node A is the input variable of node B, then create a directed edge from A to B. If there is bidirectional data exchange (such as in iterative solving), then create two edges in opposite directions. Finally, all nodes and edges form a directed graph, which is the directed coupled evaluation network.

[0026] The edge weight coefficient (i.e., the uncertainty propagation coefficient) is used to quantitatively characterize the uncertainty propagation strength from node j to node i. It can be determined in at least one of the following ways: Method 1: Sampling is performed within the typical input range of node j to obtain a set of input samples X.

[0027] Record the output Yj corresponding to node j and the output Yi corresponding to node i.

[0028] Calculate the partial derivative or regression coefficient of the output Yi of node i with respect to the output Yj of node j, take its absolute value and normalize it, and use it as an estimate of the edge weight coefficient. This reflects the sensitivity of node i to changes in the output of node j.

[0029] Method 1: Collect the output error sequence of node j and the output error sequence of node i during historical operation or testing of the system.

[0030] Calculate the correlation coefficient (such as the Pearson correlation coefficient) between two error sequences.

[0031] The absolute value of this correlation coefficient is used as the edge weight coefficient. This reflects the statistical strength of the correlation between the errors of the two nodes as they change together throughout history.

[0032] Method 3: For known connections with well-defined physical relationships (such as a linear amplifier model), their gain coefficients can be directly set as edge weights.

[0033] For connections that are complex or difficult to quantify, an empirical value can be preset within the interval [0.5, 1] ​​based on information such as the simplification level of the model and known accuracy.

[0034] During operation, the edge weights can be recalculated periodically using either Method 1 or Method 2 based on newly collected data, thus achieving online adaptive updates.

[0035] In this embodiment, the simulation logic substructure refers to a set of local computing units in the evaluation network that have been dynamically verified and have high confidence. The core consists of nodes (simulation model / data interface / environment model), directed edges between nodes (data / logic dependencies), and fixed parameters (model parameters, edge weight coefficients). It is the smallest functional module that can independently complete a specific simulation function and output stable credibility.

[0036] High confidence level is a core criterion for judgment, for example: The link's average steady-state accuracy potential difference must be higher than the first preset threshold (with low internal node reliability transmission loss). At the same time, the average uncertainty traffic of the boundary link is lower than the second preset threshold (low error risk in external propagation). Essentially, it is a local logic that has been dynamically verified through network operation, and whose output results are highly consistent with the real physical scene with negligible uncertainty.

[0037] The structure is complete and independent: it consists of multiple nodes connected by directed edges to form a connected sub-network, with clear data / logic dependencies between nodes (such as the link combination of sensor data interface → flight control simulation model → rudder system simulation model); it can independently complete a specific simulation function (such as flight attitude calculation, power system performance evaluation) and output valid results without relying on external nodes of the network.

[0038] Parameters and logic can be fixed: Includes fixed parameters within the nodes (such as the corrected parameters of the simulation model and the steady-state parameters of the environmental model). Includes fixed edge weight coefficients between nodes (the uncertainty propagation coefficients tend to stabilize after online adaptive updates); These parameters and logic combinations are reusable and can be directly called in similar simulation tasks without the need for iterative optimization.

[0039] Relationship with the overall evaluation network It is a local core subset of the overall evaluation network: automatically identified and screened from the whole network, retaining the high-confidence logic that plays a key role in the simulation results, and eliminating low-confidence, redundant nodes or links; It serves as a carrier for hardware acceleration reuse: by being compiled into reconfigurable hardware logic (such as FPGA configuration code), it becomes an independent hardware acceleration module that can be directly called in subsequent similar evaluation tasks, avoiding redundant calculations and improving simulation real-time performance. Collaboration with the remaining network components: The hardware-fixed substructures can be connected to the overall network as trusted computing units, working in conjunction with other dynamically optimized nodes / links to balance simulation efficiency and global accuracy.

[0040] Taking aerospace equipment simulation as an example, the evaluation network may contain connected subnetworks of GPS data interface (baseline node) → positioning solution model (simulation node) → flight attitude fusion model (simulation node) → environmental interference model (environment node): After the network has been running in a steady state, if the precision potential difference of each link within the sub-network is higher than 0.8 (first threshold) and the uncertainty flow of the boundary link connected to the external node is lower than 0.2 (second threshold), then the sub-network is identified as a high-confidence simulation logic substructure. The positioning solution model parameters, attitude fusion algorithm, and uncertainty transmission coefficient between nodes contained therein are all fixed and compiled into a hardware acceleration module. Subsequent attitude evaluation tasks for similar aerospace equipment can directly call this module without re-calibrating parameters and fusing data, which greatly improves the simulation speed.

[0041] By abstracting heterogeneous simulation models, measured data, and environmental models into nodes with clear interfaces, and constructing directed edges based on their real data / logical dependencies, the complex equipment system is mapped into a computational graph with a clear structure and well-defined relationships. This provides a standardized and computable object model for subsequent unified dynamic field calculations, data fusion, and uncertainty analysis, solving the technical problem of the difficulty in unified management and interaction of virtual and real models in traditional methods.

[0042] The actual equipment data acquisition interface is clearly and uniquely defined as the benchmark node and given a constant highest credibility. The credibility of all virtual simulation data will be relative to this benchmark for measurement and calibration, thereby ensuring that the entire evaluation system always closely follows physical reality, avoiding the systematic bias that may exist in pure virtual simulation, and significantly improving the physical credibility and engineering practical value of the evaluation results.

[0043] The edge weight coefficients (uncertainty propagation coefficients) configured on the connection edges explicitly quantify the propagation intensity of errors or uncertainties between models in the evaluation network. This enables the system to not only evaluate the accuracy of individual nodes, but also to trace and predict the diffusion path and amplification effect of uncertainties throughout the entire equipment system. This provides a direct basis for locating key bottlenecks affecting the overall accuracy of the system and identifying weak links, making subsequent online parameter correction and resource optimization more targeted and improving the efficiency and pertinence of evaluation and optimization.

[0044] The network-based modeling approach, based on nodes and edges, allows the framework to flexibly adapt to changes in the composition of different equipment systems. Adding or replacing models only requires adding or deleting nodes and edges accordingly, without needing to reconstruct the entire evaluation process. The edge weight configuration mechanism also supports initial setting and dynamic updates through various methods such as sensitivity analysis and historical data learning, enabling the network to self-improve as understanding deepens and data accumulates. This provides a feasible technical architecture for building a simulation evaluation system that adapts to the entire lifecycle and sustainable evolution of complex equipment.

[0045] In one possible implementation, the dynamic field information includes the local uncertainty intensity of each node and the precision potential field of the entire network; The accuracy potential value of the reference node is a constant high value; the accuracy potential value of the non-reference node is iteratively updated according to the state of its upstream node; at each update time, the update rule is as follows: first, select the highest value from the accuracy potential values ​​of all upstream nodes at the previous time as the basis, and then perform information decay on the highest value; then subtract the joint loss caused by the uncertainty flow transmitted from all upstream nodes and weighted by their local uncertainty intensity; finally, compare the calculation result with a preset lower limit value of accuracy potential, and take the larger one as the accuracy potential value of the node at the current time.

[0046] In one possible implementation, the precision potential of the non-reference node at the next time step... Obtained through the following formula: in, The inherent attenuation coefficient of information transmission; 0 < <1; It is the maximum value of the precision potential among all upstream nodes of node i; Let be the intensity of the local uncertainty of upstream node j at time t; Let be the uncertainty propagation coefficient of the edge from node j to node i, i.e., the edge weight coefficient. The uncertainty loss coefficient; This is the lower limit of the accuracy potential common to all non-reference nodes; >0; The accuracy can be determined offline before actual deployment, taking into account the specific simulation scenario, equipment type, and system accuracy requirements. For example, it can be set to 0.1, 0.2, etc.

[0047] The node precision potential describes the reliability or precision level of the output data of a single node i at the current moment, and is a scalar value.

[0048] The overall network precision potential field refers to the set or distribution state composed of the precision potential values ​​of all nodes; it is a snapshot of the overall precision potential values ​​of all nodes. The dynamic update of node precision potential values ​​drives the evolution of the overall network precision potential field. Simultaneously, the distribution state of the overall network precision potential field (e.g., where potential values ​​are high, where gradients are large) reflects the propagation path and quality of information and uncertainty in the network.

[0049] The intensity of local uncertainty at a node is a measure of the output instability or error range caused by model simplification, parameter errors, measurement noise, etc. within the node itself. It is an endogenous property.

[0050] In one possible implementation, the intensity of the local uncertainty is obtained based on the difference between the probability distribution of the node's output data and a preset baseline probability distribution.

[0051] In one possible implementation, the intensity of the local uncertainty is expressed by the following equation: in, The cumulative probability distribution of the output data of node j at time t can be estimated through real-time statistics. The cumulative probability distribution of the baseline data for node j can be obtained from historical calibration data; this integral value directly reflects the overall area of ​​difference between the two distributions.

[0052] In one possible implementation, the local uncertainty intensity of node j The model residual method can be used for real-time calculation, specifically: For simulation model nodes This can be calculated as the root mean square error (RMSE) of the residual sequence between the node's output value and the synchronous measured reference value from the baseline node within the most recent time window [tT,t], and then normalized. For nodes without direct measured reference values, its It can be recursively estimated based on the uncertainty intensity of its upstream nodes and the network topology, or assigned a conservative nominal value determined based on prior knowledge of the model.

[0053] To ensure that all parameters work collaboratively on a uniform scale, an offline calibration phase is required after the system's initial run or a major update. During this phase, by injecting standard test data or using historical datasets, the dynamic field update formula is run, and the coefficients (such as the uncertainty loss coefficient and the baseline scale of each edge weight coefficient) are adjusted to ensure that the overall network accuracy potential is distributed within a reasonable and stable range (e.g., […]). The set of coefficients determined after this calibration will serve as the baseline parameters for subsequent online operation.

[0054] Initial values ​​of the precision potential of all non-reference nodes at system startup. The lower limit of the accuracy potential can be uniformly set. After the system is running, when the change in the accuracy potential value of all nodes is less than a preset small threshold in multiple consecutive time steps, the system is determined to have entered a statistically stable state. At this time, historical data for identifying high-confidence substructures can be started.

[0055] The attenuation coefficient can be preset based on the physical characteristics or logical association strength of the connection edge, or dynamically calculated in the following ways: Based on historical consistency: The average consistency (such as correlation coefficient) of the output data of the two nodes connected by the edge in the historical operation is statistically analyzed. The higher the consistency, the closer the decay coefficient is to 1 (the smaller the decay).

[0056] Based on model complexity: If the data transmitted by the edge originates from a complex simulation model, an attenuation coefficient less than 1 can be set according to the inherent complexity of the model or the known uncertainty.

[0057] Preset empirical values: For known, high-fidelity data transmission links (such as rigorously verified physical interfaces), the attenuation coefficient can be preset to 0.95-1.0; for links with lower fidelity or containing simplified models, it can be preset to 0.7-0.9.

[0058] By defining the precision potential and giving its full-network iterative update formula, a collaborative computing model is established in which the credibility of all nodes is interconnected and dynamically influenced. The credibility of any node is no longer determined solely by its own performance, but rather by the credibility status of all its upstream nodes and the uncertainty transmission effect. This allows the assessment of the overall credibility of a complex system to be established for the first time on a rigorous mathematical basis that reflects the inherent information flow dependencies.

[0059] This method achieves a paradigm shift from static point evaluation to dynamic field evaluation, revealing the spatiotemporal propagation law of uncertainty. Traditional methods evaluate individual models in isolation, while this method explicitly models how uncertainty is transmitted from one model and affects downstream models through the iterative formula of the accuracy potential. This makes the evaluation no longer a static scoring of the results, but a real-time analysis of the error propagation dynamics, which can predict systemic risks and locate error amplification nodes in advance.

[0060] The precision potential field provides a dynamic reliability measure that varies over time and across nodes. This provides a direct and optimal basis for weight allocation in data fusion based on path contribution weights (nodes with high potential values ​​contribute more). Simultaneously, the precision potential difference in steady state directly identifies the critical links in the network that cause the greatest loss of information fidelity, indicating the highest priority for model optimization and hardware acceleration resource allocation.

[0061] The dynamic field (precision potential and uncertainty flow) serves as the judgment benchmark for the sensing signal and hardware logic identification in subsequent online parameter correction.

[0062] In one possible implementation, the data adaptive fusion based on dynamic field information distributed across nodes includes: For any node to be evaluated, gather all source paths that can provide data to it; For each source path, its path contribution weight is calculated. This weight is positively correlated with the precision potential of the path terminal node and negatively correlated with the intensity of the accumulated uncertain flow on the path. Based on the contribution weight of each path, the data from different paths are weighted and fused to obtain the final output of the node.

[0063] In one possible implementation, the aggregation of all source paths that can provide data to the current node to be evaluated specifically includes: Determine the constraint rules for reverse path search based on at least one preset path filtering condition.

[0064] Perform a constraint-based reverse traversal, starting from the current node to be evaluated, and traverse the graph in reverse (i.e. against the direction of data flow) along the directed edges of the directed coupled evaluation network according to the constraint rules.

[0065] Nodes that are reached and meet the preset filtering conditions are identified as valid source nodes; the directed path from the valid source node to the current node to be evaluated is identified as a valid source path.

[0066] The path filtering conditions include, but are not limited to: Maximum search depth: The maximum number of directed edges that can be traced. The default global maximum search depth is 3~5, which can be adjusted according to the network complexity. For example, in a complex equipment simulation network, after data transmission has passed through about 5 nodes, uncertainty has accumulated significantly, and continuing to trace is meaningless. Node type restrictions: For example, only trace back to simulation model nodes or baseline nodes.

[0067] Precision potential threshold: Only trace node branches whose precision potential value is higher than the preset threshold.

[0068] Key node set: A predefined set of key upstream nodes; the search is limited to this set.

[0069] After performing reverse tracing and obtaining the initial set of valid source paths, multiple paths pointing to the same valid source node are filtered, and only the path with the minimum cumulative uncertainty flow intensity is retained.

[0070] The final output of the node refers to the evaluation value of the current status or performance index of the equipment subsystem or functional module represented by the node after multi-source data fusion.

[0071] Specifically, this may include (but is not limited to) the following types of data: State parameters: such as engine thrust estimate, radar detection range estimate, and aircraft attitude angle estimate.

[0072] Performance metrics: such as real-time evaluation values ​​of system response time, tracking accuracy, and energy efficiency.

[0073] Health or credibility: A quantitative score for whether the unit is working properly and whether the data is reliable.

[0074] Each node, based on its pre-defined role in the simulation evaluation system, has one or more target variables that need to be output. The final output is the optimal fusion estimate of the values ​​of these target variables.

[0075] When a node has multiple evaluation target variables that need to be output, the adaptive data fusion process can be performed independently for each target variable. For each target variable, the system aggregates the source paths that can provide the data for that variable, calculates the path contribution weights based on the data related to that variable, and finally obtains an independent fusion estimate for that target variable.

[0076] The set of key nodes can be determined through network topology analysis (such as out-degree and betweenness centrality) or by the importance of information flow in historical data.

[0077] In one possible implementation, let R be the set of all valid source paths, and for each path r∈R in the set R, let its contribution weight to the current node to be evaluated be... Obtained using the following formula: R is the set of all valid source paths; src(r) represents the starting node of path r, i.e., the valid source node; γ is the uncertainty penalty coefficient, 0 < γ < 2, and is a constant greater than 0; Let src(r) be the precision potential value of node src(r) at the current time. The accumulated uncertain flow intensity on path r; the summation in the denominator traverses all paths p in set R; It is a constant, 0 < <0.01.

[0078] γ is usually initially set to 1. If the fusion result is found to be insensitive to uncertainty, it is increased; if it is too sensitive and causes weight concentration, it is decreased. ε is used to prevent the denominator from being zero, and is usually set to a small value related to machine precision (such as 1e-10).

[0079] The data from the source path must be aligned to the same time or the same operating condition.

[0080] The cumulative uncertain flow intensity is calculated using either a summation method or a maximum value attenuation method.

[0081] In one implementation, the accumulated uncertainty flow intensity along the entire directed path from the upstream source node to the current downstream node to be evaluated is calculated as follows: for each directed connection edge on the path, the local uncertainty intensity of its starting node at the current moment is multiplied by the uncertainty propagation coefficient pre-configured for that edge to obtain the instantaneous uncertainty flow intensity of that edge; subsequently, the instantaneous uncertainty flow intensities of all directed edges on the path are summed, and the sum is the accumulated uncertainty flow intensity.

[0082] In another implementation, the accumulated uncertainty flow intensity along the entire directed path from the upstream source node to the current downstream node to be evaluated is calculated as follows: First, find the node with the maximum local uncertainty intensity at the current moment among the starting nodes of all directed edges on the path, and use this maximum intensity value as the basis; then, multiply this maximum value by a decay coefficient that is negatively correlated with the path length (i.e., the number of directed edges contained in the path), and the resulting product is the accumulated uncertainty flow intensity.

[0083] By taking the data transmission path in the network as the core evaluation object and calculating the accumulated uncertainty of the path, a quantitative evaluation of the information transmission channel quality is achieved. This enables the fusion system not only to determine whether the data is accurate, but also to diagnose whether the inaccurate data is a source problem or contaminated during transmission. This allows for tracing the root cause of evaluation errors and provides unprecedented insight for precise system optimization.

[0084] Each node's fusion decision relies solely on the dynamic field of global broadcast and local network topology information, forming a distributed collaborative intelligence: without a central decision-maker, each node can make the optimal local fusion decision for itself based on the overall network situation. This architecture is naturally resistant to local failures and easy to expand, perfectly matching the distributed and modular physical characteristics of complex equipment systems, and significantly enhancing the overall robustness of the system.

[0085] The core algorithm of this scheme reduces computational complexity from exponential to linear by introducing a strategy of maximum search depth and unique optimal path for a single source node, without sacrificing core performance (due to the minimal contribution of excessively long or suboptimal paths). This successfully addresses the critical challenge of advanced algorithms failing to perform computationally in engineering applications, enabling a fusion framework with profound theoretical implications to run stably in complex equipment simulation and evaluation systems with extremely high real-time requirements.

[0086] The fusion weights depend on the evaluation results of the dynamic field, and the better fusion output will be fed back to improve the data quality of the relevant nodes and reduce their uncertainty, thereby generating a better dynamic field and fusion weights in the next cycle.

[0087] In one possible implementation, triggering online parameter correction based on dynamic field information includes: Real-time monitoring of the intensity of uncertain flow from the target simulation model node to its key downstream node or reference node, wherein the intensity of uncertain flow is obtained based on the product of the local uncertainty intensity of the target simulation model node and the uncertainty propagation coefficient of the connecting edge; When the intensity of the uncertain flow exceeds a preset threshold, online correction of the target simulation model nodes is triggered; During the calibration process, the model parameters are adjusted based on the gradient information of the node's output error with respect to its own parameters, and in combination with the direction of the precision potential gradient between the node and downstream nodes; where the model refers to the mathematical model or surrogate model corresponding to the simulation model node.

[0088] For example: aerodynamic coefficients (such as lift coefficient and drag coefficient) in aerodynamic calculation models; PID parameters in control system models; heat transfer coefficients or specific heat capacity in thermodynamic models; weights and biases in neural network surrogate models.

[0089] Among them, the key downstream nodes are determined sequentially according to the following rules: Rule 1 (Static topology takes precedence): The node with the largest out-degree among the direct downstream nodes of the source node is selected as the primary key downstream node. If multiple nodes have the same out-degree, all of them are included in the monitoring list. This rule, based on the network structure, identifies the downstream hub with the strongest information dissemination capability.

[0090] Rule 2 (Dynamic State Correction): The accuracy potential of all directly downstream nodes of the source node is monitored in real time. The node with the lowest accuracy potential is designated as the dynamically critical downstream node. This rule aims to protect the weakest link in the network, the one most vulnerable to uncertainties.

[0091] Rule 3 (Goal-Oriented Anchoring): Identify all directed paths from the source node to any baseline node (the actual equipment data interface node). The most frequently occurring direct downstream nodes on these paths are identified as critical downstream nodes. This rule ensures the fidelity of the most direct and critical paths that have the greatest impact on the final evaluation results.

[0092] Rule 4 (Comprehensive Weighted Judgment): Weights are assigned to the above rules, and a comprehensive score is calculated for each direct downstream node. The node with the highest score is identified as the key downstream node. The weights can be dynamically adjusted according to the simulation stage or system objectives; for example, rule three is given more weight in the early verification phase, while rule two is given more weight during stable operation.

[0093] Default and succession rules: If a unique downstream node cannot be determined according to the above rules, all eligible downstream nodes are considered critical downstream nodes, and the sum of their uncertain flow intensities is monitored. Furthermore, the system allows for manual pre-setting of downstream nodes related to specific functions or security as critical downstream nodes.

[0094] The adjusted model parameters include: A loss function is constructed based on the node output error, and the gradient of the loss function with respect to the adjustable parameters of the node is calculated. The loss function L can be defined as the sum of squares (i.e., mean square error, MSE) between the output data of the simulation model node at the current moment and the measured reference data collected through the actual equipment data interface node corresponding to the same moment or the same working condition. Based on the precision potential difference from the current node to its critical downstream node Determine a continuously differentiable correction guiding factor. ; Based on the aforementioned correction guiding factor The gradient of the loss function constructed based on the node output error is used to calculate the parameter update amount and update the node parameters; And after completing the parameter update, generate a correction flow associated with this correction behavior. The correction stream is used to propagate along the directed coupled evaluation network and affects the accuracy potential update of downstream nodes.

[0095] The correction guiding factor The precision potential difference value A monotonically increasing function that satisfies: when When >0, >0; when When <0, <0.

[0096] The correction guiding factor It is a hyperbolic tangent function or a sigmoid function.

[0097] The parameter update amount This can be achieved using the following formula: in, L represents the node adjustable parameters, which are the set of numerically adjustable mathematical variables within the nodes of the simulation model used to describe their input-output mapping relationship, and whose adjustment is aimed at making the model output closer to the behavior of the real physical system; L is the loss function defined by the node output error. The learning rate; The correction flow The intensity and magnitude of this parameter update and the aforementioned correction guiding factor The absolute values ​​are positively correlated.

[0098] For example, the correction stream The strength can be obtained by the following formula: in, The correction gain coefficient is a preset constant (e.g., 0.1, 0.5); it can be obtained by calibrating the positive impact on downstream nodes using historical correction data. The norm of the parameter update vector (such as the L2 norm) represents the correction magnitude; Dimensionless, with a value range of [0,1], used to unify the scale differences of parameters at different nodes.

[0099] When a downstream node updates its precision potential value, the dynamic field information it relies on also includes the correction flow transmitted from the upstream node. The update rule for its precision potential value is as follows: on the basic precision potential value calculated based on the precision potential information and uncertainty flow information of the upstream node, a compensation term formed by weighting the upstream correction flow is added to obtain the final precision potential value.

[0100] The compensation term is obtained by weighted summation of the correction flow intensity transmitted from all upstream nodes to the current node.

[0101] The final precision potential of downstream node k at the next time step Obtained by the following formula; The basic value of the precision potential of node k is calculated based on the precision potential value of the upstream node and the uncertain flow information. The corrected flow intensity is transmitted from upstream node j; This is the corresponding correction flow influence coefficient, which is the weight in the weighted summation of the correction flow intensity transmitted from all upstream nodes to the current node; For the compensation weighting coefficient, 0 < <1; This is the lower limit of the precision potential; Represents the set of all upstream nodes of node k; Let j be the intensity of the local uncertainty of the upstream node. Let λ be the uncertainty propagation coefficient on the edge from node j to node k; λ is the uncertainty loss coefficient.

[0102] It can be obtained in any of the following ways: Preset empirical values: For high-fidelity, strongly connected links, the preset value is close to 1.0 (e.g., 0.9); for simplified models or weak connections, the preset value is smaller (e.g., 0.5).

[0103] Related to the uncertainty transmission coefficient: can be defined as =1-β* Where β is the adjustment coefficient, 0 < β < 1 / This indicates that the stronger the uncertainty propagation of an edge, the lower the efficiency of its correction effect propagation (because the structure itself is prone to amplifying errors).

[0104] Based on historical data analysis: The actual improvement rate of downstream node accuracy potential after upstream node correction is statistically analyzed and used as... The estimated value.

[0105] Compensation weight coefficient This is used to balance the proportion of losses caused by uncertain flow and compensation brought by correction flow in the accuracy potential update.

[0106] It can be determined in any of the following ways: μ is preset based on the system's level of trust in positive feedback, such as =0.2.

[0107] μ can be designed to be negatively correlated with the average uncertainty level of the entire network. When the overall network uncertainty is high, the value should be lowered. A cautiously optimistic attitude is maintained regarding the effect of a single correction. Once the network as a whole is stable, the adjustment should be increased. The overall network average uncertainty level is the weighted average of the local uncertainty intensity of all nodes.

[0108] As a hyperparameter, it is optimized and determined during the simulation testing phase before system deployment.

[0109] At the same time, the uncertainty flow triggers the correction first, then the correction flow is generated and transmitted to the downstream. The downstream node synchronously incorporates the uncertainty flow loss and the correction flow compensation in the accuracy potential update at the next time.

[0110] The integrated simulation evaluation method is a dynamic process that runs cyclically over discrete time steps. At each simulation time t, the dynamic field information of each node in the network is updated according to the state of the previous time step. The online parameter correction process is tightly coupled with this dynamic field update process, forming a closed loop in time. The coordination mechanism is as follows: First, the dynamic field state at the current time t is used to make a judgment (especially the intensity of the uncertain flow from the simulation model node to the key downstream node); if it exceeds the threshold, the online correction of the node is triggered immediately.

[0111] The calibration process uses the current precision potential difference as a guide to calculate the parameter update amount and update the model parameters of the nodes accordingly. Simultaneously, a calibration flow is generated based on the magnitude and direction of this calibration.

[0112] After completing all possible corrections at time t, the dynamic field update phase begins. At this point, the uncertain flow and correction flow generated from each node will simultaneously propagate along the directed edges to their respective downstream nodes k.

[0113] When downstream node k calculates its precision potential at the next time step t+1, it will simultaneously and comprehensively consider two types of influences from upstream: Negative impact (loss): The decrease in precision potential caused by the uncertainty intensity of upstream node j after being weighted by the transfer coefficient.

[0114] Positive impact (compensation): The accuracy potential compensation brought about by the correction flow generated by the correction behavior of upstream node j after being weighted by the influence coefficient.

[0115] The above scheme integrates four types of rules: static topology, dynamic state, goal orientation, and comprehensive weighting. It can identify hub nodes with strong information dissemination capabilities, protect weak links with low accuracy potential, and anchor the core path that affects the final assessment, ensuring that no key downstream nodes are missed or deviated from.

[0116] It supports dynamic adjustment and manual preset of rule weights, adapting to the needs of different stages such as the initial stage of simulation verification and the stable operation period. It can prioritize the fidelity of core scenarios or security-related nodes, improving the flexibility and engineering adaptability of correction triggering. The supplementary rules solve the problem of judgment of multiple nodes in parallel, and avoid correction omissions by monitoring the sum of uncertain flow intensity, further strengthening the integrity of the triggering logic.

[0117] The MSE loss function is constructed based on the measured data. The parameters are adjusted by combining the parameter gradient and the precision potential difference value as a guide factor. This ensures that the model output is close to the real physical system, and the guide factor fits the dynamic field state of the entire network, avoiding parameter oscillation caused by blind correction.

[0118] The calibration guide factor employs a monotonically increasing hyperbolic tangent or sigmoid function, strictly matching the positive and negative directions of the accuracy potential difference value to ensure that the calibration direction is consistent with the overall network credibility optimization requirements, thereby improving the accuracy of parameter adjustment. Parameters such as the learning rate and guide factor are flexibly configurable and can be tailored to different model types (e.g., aerodynamic models, neural network surrogate models) to adapt to the parameter calibration needs of simulations of various complex equipment.

[0119] The correction flow intensity is positively correlated with the parameter update amplitude and the absolute value of the guiding factor, enabling the positive effect of effective correction to propagate along the network, achieving single-point correction and benefiting the entire network, thus making up for the lack of global linkage in traditional local correction.

[0120] Downstream node accuracy potential updates synchronously integrate uncertainty loss and correction flow compensation. By balancing the two types of influence through the correction flow influence coefficient and the compensation weight coefficient, uncertainty diffusion is suppressed and positive correction effect is fully absorbed, thereby promoting the continuous improvement of the network's overall credibility.

[0121] The correction flow impact coefficient supports multiple configuration methods and can adapt to different link characteristics such as high-fidelity / simplified models and strong / weak connections. The compensation weight coefficient can dynamically adapt to the uncertainty level of the entire network, enhancing the system's adaptability under complex operating conditions.

[0122] The timing logic of calibration triggering → parameter adjustment → calibration flow generation → dynamic field update is clearly defined to avoid conflicts between the propagation of uncertain flows and calibration flows, ensuring orderly connection of each link and improving the stability of system operation. Online parameter calibration and dynamic field update form a closed-loop collaboration. The calibration decision at each moment is based on the current dynamic field state, and the calibration effect feeds back into the dynamic field update at the next moment, continuously optimizing the simulation evaluation accuracy.

[0123] In one possible implementation, the dynamic identification of high-confidence simulation logic substructures and the compilation of said substructures into reconfigurable hardware logic includes: Based on the operational data of the evaluation network after reaching a statistically stable state, the steady-state precision potential difference and average uncertainty flow of each link are calculated; wherein, the steady-state precision potential difference is defined as the difference in precision potential values ​​of the nodes at both ends of the link in the steady state, and the average uncertainty flow is defined as the average value of the uncertainty intensity transmitted through the link during the statistical period. Based on the calculated steady-state accuracy potential difference and average uncertainty flow of each link, the maximum connected subnetwork that meets the preset condition is identified from the evaluation network as the high-confidence causal subnetwork. The simulation logic corresponding to the high-confidence causal subnet, including the topological connection relationship between its nodes and the fixed parameters of each node, is compiled into hardware description language code, and a hardware acceleration module that can be configured and run on a programmable logic device is generated based on the code.

[0124] The average steady-state accuracy potential difference of all links within the sub-network is higher than the first preset threshold, and the average uncertainty flow of all boundary links of the sub-network is lower than the second preset threshold; the boundary link is the link connecting the internal node and the external node of the sub-network; the first threshold is, for example, 0.7~0.9, and the second threshold is, for example, 0.1~0.3, which can be optimized according to the simulation accuracy requirements.

[0125] The statistically stable state refers to the state in which, within M consecutive time steps (M is a preset positive integer, such as M=100), the change in the precision potential value of more than 90% of the nodes in the entire network is less than the preset change threshold. The calculation of the average uncertain flow is based on a continuous time window after the evaluation network enters a statistically stable state. The length of this window can be preset (e.g., containing data for N=1000 time steps) or can continue until the end of the system task phase. Within this window, for each directed edge, the time average of its instantaneous uncertain flow intensity is calculated as the average uncertain flow of that link.

[0126] The preset conditions are used to filter links; identifying high-confidence causal subnets includes: first, filtering all links that meet the preset conditions from the evaluation network; then, finding all connected components formed by these filtered links; finally, determining the connected component containing the most nodes as the high-confidence causal subnet.

[0127] The fixed parameters of each node refer to the time average of the adjustable parameters within each simulation model node or the instantaneous value at the end of the statistical steady state during the statistical period. Preferably, the time average is used to smooth out random fluctuations and obtain more representative steady-state parameters. These fixed parameters, along with the topological relationships between nodes, are translated into corresponding computational units (such as arithmetic logic units, lookup tables) and their configuration constants in the hardware description language.

[0128] The node topology, node fixed parameters (corrected model parameters), and data interaction logic of the high-confidence causal subnet are analyzed. Generate Hardware Description Language (HDL): Convert simulation logic into Verilog or VHDL code (e.g., node models correspond to hardware computing units, and links correspond to data transmission buses). HDL code is synthesized into a netlist file using FPGA synthesis tools (such as Xilinx Vivado and Intel Quartus), and then configured into a programmable logic device (FPGA / CPLD) to form a reusable hardware acceleration module.

[0129] Subnetworks are selected based on a dual metric of steady-state accuracy potential difference and average uncertainty flow. This ensures that the selected high-confidence causal subnetworks have high internal link reliability and low external uncertainty leakage, resulting in simulation logic outputs that better reflect real-world scenarios and improving the reliability and credibility of simulation evaluation results. Clearly defined preset conditions and boundary link definitions for subnetwork selection prevent low-confidence logic from being mixed into the hardware solidification process, avoiding simulation deviations caused by unreliable logic at the source and providing high-quality core logic support for the performance evaluation of complex equipment. The high-confidence simulation logic is compiled into reconfigurable hardware logic, generating acceleration modules adapted to programmable devices such as FPGAs. The parallelism and high speed of hardware computation significantly reduce the runtime of similar simulation evaluation tasks, resolving efficiency bottlenecks caused by multi-node interaction and large-scale computation in complex equipment simulation. The solidified hardware acceleration modules can be reused in similar simulation scenarios without requiring recompilation and parameter configuration each time, shortening the iteration cycle of simulation tasks and improving the overall efficiency of complex equipment development and performance evaluation.

[0130] By selecting sub-networks based on steady-state network operation data, the system can dynamically adapt to changes in network operation (such as improved reliability after node parameter correction), ensuring that the hardware-based solidified sub-network always represents the most reliable core logic in the current network, thus avoiding the adaptability issues caused by static solidification. The first and second preset thresholds can be flexibly adjusted according to simulation accuracy requirements and equipment scenario characteristics, adapting to both the stringent requirements of high-fidelity simulation and the efficiency demands of rapid iteration scenarios, enhancing the engineering adaptability of the solution. Hardware solidification is performed only on high-confidence simulation logic, avoiding resource waste caused by indiscriminate hardware implementation, allowing limited hardware resources to be concentrated on serving the core simulation logic, improving hardware resource utilization. The independence and stability of the high-confidence sub-network can reduce interaction conflicts between hardware modules and software simulation parts, lowering the overall system maintenance and debugging costs, and improving the operational stability of complex equipment simulation systems.

[0131] Example 2: This example provides an integrated simulation and evaluation system for complex equipment systems, used to implement the steps of the integrated simulation and evaluation method for complex equipment systems described in Example 1. The system includes: The evaluation network construction module is used to build a directed coupled evaluation network that includes simulation model nodes, actual equipment data interface nodes, and environment model nodes. The evaluation module is used to perform parallel calculations of dynamic field information reflecting the credibility status of each node during the operation of the evaluation network, and based on the dynamic field information, to perform adaptive fusion of virtual simulation data and actual equipment data in a distributed manner at each node, and output an integrated evaluation result. The online calibration module is used to trigger and perform online parameter calibration of relevant simulation model nodes based on the uncertainty propagation revealed by the dynamic field information. The compilation module is used to automatically identify high-confidence simulation logic substructures based on the historical records of the average steady-state accuracy potential difference and average uncertainty flow of each link in the evaluation network, and compile the substructures into reconfigurable hardware logic for accelerated execution of subsequent similar evaluation tasks.

[0132] The working principle and effect of the above technical solution are the same as those in Example 1, and will not be repeated here.

[0133] This invention also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any embodiment of this invention.

[0134] This invention also provides a computer-readable storage medium for storing a computer program. When the computer program is executed, it implements the steps of the method described in Embodiment 1 of this invention. The specific implementation method is consistent with the implementation method and the technical effects achieved in the above method embodiments, and some contents will not be repeated.

[0135] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the invention without departing from the principles and spirit of the invention, and all such changes should fall within the protection scope of the claims of the present invention.

Claims

1. An integrated simulation and evaluation method for complex equipment systems, characterized in that, include: Construct a directed coupled evaluation network that includes simulation model nodes, actual equipment data interface nodes, and environment model nodes; During the operation of the evaluation network, dynamic field information reflecting the credibility status of each node is computed in parallel, and based on the dynamic field information, the virtual simulation data and actual equipment data are adaptively fused in a distributed manner at each node to output an integrated evaluation result. Based on the uncertainty propagation revealed by the dynamic field information, online parameter correction of relevant simulation model nodes is triggered and executed; Based on the historical records of the average steady-state accuracy difference and average uncertainty flow of each link in the evaluation network, high-confidence simulation logic substructures are automatically identified, and the substructures are compiled into reconfigurable hardware logic.

2. The simulation evaluation method according to claim 1, characterized in that, The construction of the directed coupled evaluation network includes: Based on the interaction relationship between the simulation model, the actual equipment data interface, and the environment model, the directed coupled evaluation network is constructed. The actual equipment data interface node is defined as the reference node of the network, and it is given a constant highest confidence benchmark. Edge weight coefficients are configured on the directed connection edges between nodes, and the edge weight coefficients are used to quantify the uncertainty propagation characteristics between nodes.

3. The simulation evaluation method according to claim 2, characterized in that, The dynamic field information includes the local uncertainty intensity of each node and the precision potential field of the entire network; The precision potential value of the reference node is a constant high value; the precision potential value of non-reference nodes is iteratively updated according to the state of their upstream nodes; at each update time, the update rule is as follows: first, select the highest value from the precision potential values ​​of all upstream nodes at the previous time as the basis, and then perform information decay on the highest value; then subtract the joint loss caused by the uncertainty flow transmitted from all upstream nodes and weighted by their local uncertainty intensity; finally, compare the calculation result with a preset precision potential lower limit value, and take the larger one as the precision potential value of the node at the current time. Precision potential of non-reference nodes at the next time step Obtained through the following formula: in, The inherent attenuation coefficient of information transmission; 0 < <1; It is the maximum value of the precision potential among all upstream nodes of node i; Let be the intensity of the local uncertainty of upstream node j at time t; Let be the uncertainty propagation coefficient for the edge from node j to node i; The uncertainty loss coefficient; This is the lower limit of the accuracy potential common to all non-reference nodes; >

0.

4. The simulation evaluation method according to claim 1, characterized in that, The process of achieving adaptive data fusion distributed across nodes based on dynamic field information includes: For any node to be evaluated, gather all source paths that can provide data to it; For each source path, its path contribution weight is calculated. This weight is positively correlated with the precision potential of the path terminal node and negatively correlated with the intensity of the accumulated uncertain flow on the path. Based on the contribution weight of each path, the data from different paths are weighted and fused to obtain the final output of the node.

5. The simulation evaluation method according to claim 4, characterized in that, The aggregation of all source paths that can provide data to it includes: Determine the constraint rules for reverse path search based on at least one preset path filtering condition; Starting from the current node to be evaluated, a reverse graph traversal is performed along the directed edges of the directed coupled evaluation network according to the constraint rules. Nodes that are reached and meet the preset filtering conditions are identified as valid source nodes; the directed path from the valid source node to the current node to be evaluated is identified as a valid source path. After performing reverse tracing and obtaining the initial set of valid source paths, multiple paths pointing to the same valid source node are filtered, and only the path with the minimum cumulative uncertainty flow intensity is retained. The path selection criteria include maximum search depth, node type restrictions, precision potential threshold, and key node set. The cumulative uncertain flow intensity is determined by summation or maximum value attenuation.

6. The simulation evaluation method according to claim 4, characterized in that, Let R be the set of all valid source paths. For each path r∈R in the set R, its contribution weight to the current node to be evaluated is... Obtained using the following formula: R is the set of all valid source paths; src(r) represents the starting node of path r; γ is the uncertainty penalty coefficient, 0 < γ < 2, and is a constant greater than 0; Let src(r) be the precision potential value of node src(r) at the current time. The accumulated uncertain flow intensity on path r; the summation in the denominator traverses all paths p in set R; It is a constant, 0 < <0.

01.

7. The simulation evaluation method according to claim 1, characterized in that, The step of triggering online parameter correction based on dynamic field information includes: Real-time monitoring of the intensity of uncertain flow from the target simulation model node to its key downstream node or reference node, wherein the intensity of uncertain flow is obtained based on the product of the local uncertainty intensity of the target simulation model node and the uncertainty propagation coefficient of the connecting edge; When the intensity of the uncertain flow exceeds a preset threshold, online correction of the target simulation model nodes is triggered; During the calibration process, the model parameters are adjusted based on the gradient information of the node's output error with respect to its own parameters, and in combination with the direction of the precision potential gradient between the node and downstream nodes.

8. The simulation evaluation method according to claim 7, characterized in that, The adjusted model parameters include: A continuously differentiable correction guidance factor is determined based on the precision potential difference from the current node to its critical downstream node. Based on the gradient of the correction guidance factor and the loss function constructed based on the node output error, the parameter update amount is calculated and the node parameters are updated. After the parameter update is completed, a correction flow associated with the correction behavior is generated, which is used to propagate along the directed coupled evaluation network and affect the accuracy potential update of downstream nodes.

9. The simulation evaluation method according to claim 1, characterized in that, The automatic identification of high-confidence simulation logic substructures and the compilation of these substructures into reconfigurable hardware logic includes: Based on the operational data of the evaluation network after reaching a statistically stable state, the steady-state precision potential difference and average uncertainty flow of each link are calculated; wherein, the steady-state precision potential difference is defined as the difference in precision potential values ​​of the nodes at both ends of the link in the steady state, and the average uncertainty flow is defined as the average value of the uncertainty intensity transmitted through the link during the statistical period. Based on the calculated steady-state accuracy potential difference and average uncertainty flow of each link, the maximum connected subnetwork that meets the preset condition is identified from the evaluation network as the high-confidence causal subnetwork. The simulation logic corresponding to the high-confidence causal subnet, including the topological connection relationship between its nodes and the fixed parameters of each node, is compiled into hardware description language code, and a hardware acceleration module that can be configured and run on a programmable logic device is generated based on the code.

10. The simulation evaluation method according to claim 9, characterized in that, The preset conditions include: the average steady-state accuracy potential difference of all links within the sub-network is higher than a first preset threshold, and the average uncertainty flow of all boundary links of the sub-network is lower than a second preset threshold; the boundary link is a link connecting the internal nodes of the sub-network with the external nodes.

Images