A knowledge graph relationship consistency verification method and system based on physical simulation feedback

Abstract

The application relates to a knowledge graph relationship consistency verification method and system based on physical simulation feedback. Triples in which entities in a knowledge graph have material attributes and relationship predicates involve physical or spatial relationships are extracted; through mapping rules, parameters and constraint conditions that can be recognized by a physical simulation engine are obtained; a three-dimensional physical scene is constructed in a virtual simulation space; the physical simulation engine is driven to obtain a dynamic physical interaction state in the three-dimensional physical scene; physical state data is extracted from the dynamic physical interaction state; and knowledge graph relationship consistency verification is performed according to whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description. The application realizes simulation and verification of laws of the physical world, can identify error knowledge that implies physical conflicts, and solves the problem that pure text models cannot understand physical dynamic laws. The application can be widely applied in the fields of artificial intelligence and natural language processing technologies.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and natural language processing technology, and more specifically, to a method, system, computer device, and storage medium for verifying the consistency of knowledge graph relationships based on physical simulation feedback. Background Technology

[0002] As a crucial infrastructure for artificial intelligence, the quality of knowledge graphs directly impacts their application effectiveness; therefore, quality control is a critical step in the knowledge graph construction process. Knowledge graph relationship consistency refers to the consistency between entities within a knowledge graph in terms of logic, semantics, structure, temporal sequence, and multi-source fusion, ensuring that the relationships conform to ontology constraints and objective facts. It is a core indicator for evaluating the quality of knowledge graphs.

[0003] Existing methods for verifying the consistency of knowledge graph relationships mainly include: I. Knowledge Graph Validation Methods Based on Pre-trained Language Models These methods utilize the commonsense reasoning capabilities of large-scale pre-trained models such as GPT-3 and BERT for knowledge verification. However, the knowledge of pre-trained language models originates from textual data, essentially modeling statistical patterns in text rather than a true understanding of the physical world. The models may inherit incorrect knowledge due to biases or errors in the training data, failing to provide reliable physical judgments. For scenarios requiring quantitative analysis (such as determining whether a container with a specific pore size can store a liquid of a specific viscosity), language models struggle to provide accurate answers. Furthermore, the model's reasoning process lacks interpretability, making it difficult to trace the root cause of errors.

[0004] II. Knowledge Completion and Validation Methods Based on Knowledge Graph Embedding These methods learn low-dimensional vector representations of entities and relations and use vector operations to evaluate the plausibility of triples. Knowledge graph embedding only captures the statistical regularities of knowledge and cannot understand the dynamic evolution of the physical world; there is no necessary connection between distance or matching degree in the vector space and physical plausibility; the generalization ability is limited for entity combinations that do not appear in the training data; and it cannot handle the influence of continuous physical parameters (such as density, viscosity, and size).

[0005] III. Rule-based Knowledge Graph Cleaning Method These methods verify the legitimacy of knowledge through rules defined by domain experts. However, rule formulation relies on domain expert knowledge, which is costly and difficult to cover all scenarios; it is also difficult to encode complex physical laws (such as fluid dynamics, rigid body collisions, and flexible body deformation); the rules lack flexibility and cannot handle fuzzy or uncertain situations; and rule-based methods are difficult to apply to cross-domain and cross-modal knowledge verification.

[0006] IV. Common Sense Reasoning Evaluation Method Based on Embodied Intelligent Simulation Environment These methods utilize 3D physical simulation environments (such as AI2-THOR and NVIDIA Isaac Sim) to evaluate the commonsense execution capabilities of artificial intelligence agents or large language models. These methods are primarily used to evaluate and test the visual or action planning capabilities of agents, and are not intended for the systematic cleaning of massive symbolic "knowledge graph triples," thus they cannot be directly applied to the consistency verification and proactive construction of large-scale knowledge graphs.

[0007] In general, existing knowledge graph relationship consistency verification methods have the following main drawbacks: First, semantic understanding is limited to the text level and cannot capture the dynamic laws of the physical world; Second, the lack of explainable physical evidence makes it difficult to trace the source of the verification results; Third, it consumes a lot of computing resources and has low verification efficiency; IV. Inability to handle the effects of continuous physical parameters; Fifth, the test results cannot be fed back into the knowledge graph, and there is a lack of a closed-loop correction mechanism. Summary of the Invention

[0008] To address the shortcomings of the existing technologies, this invention aims to provide a method for verifying the consistency of knowledge graph relationships based on physical simulation feedback.

[0009] The method for verifying the consistency of knowledge graph relationships based on physical simulation feedback of the present invention includes the following steps: Analyze the entity relation triples in the knowledge graph and extract triples where the entity has material attributes and the relation predicate involves physical or spatial relations as triples to be verified. By using preset mapping rules, the physical simulation engine can identify the parameters and constraints of the triplet to be verified. Based on the parameters and constraints, a three-dimensional physical scene is constructed in the virtual simulation space, and initial boundary conditions are set. Drive the physics simulation engine to obtain the dynamic physical interaction state of each entity in the three-dimensional physical scene; Monitor physical state data and extract physical state data (such as retention rate, stress, etc.) from the dynamic physical interaction state. The consistency of knowledge graph relationships is verified based on whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description. If the physical state data does not conform to the constraints of the corresponding knowledge graph semantic description, the corresponding knowledge graph relationship is determined to have failed the consistency verification, and the knowledge graph is corrected.

[0010] In a practical application, if querying the material knowledge base fails to map entity names to physical material parameters, i.e., the knowledge graph is missing parameters, this invention determines the parameter range based on the general statistical values ​​of entity categories; Bayesian optimization or Monte Carlo sampling is performed within the parameter range to generate multiple sets of simulation parameter instances for verification; and only when all parameter instances pass the verification is the corresponding knowledge graph relationship determined to have passed the consistency verification.

[0011] In a practical application, the present invention first inputs the parameters and constraints into a pre-trained physical proxy model to obtain physical state data and prediction confidence. If the prediction confidence is higher than a preset threshold, the consistency of knowledge graph relationships is verified based on whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description.

[0012] In a practical application, this invention calls a smooth particle fluid dynamics algorithm for fluid-like entities and monitors in real time whether the set of fluid particle coordinates is completely within the bounding box of the container. If the particle retention rate is lower than a preset threshold after a preset simulation time, it is determined that it is physically invalid. For soft-body entities, a position-based dynamics algorithm is called. For rigid-body entities, a rigid-body dynamics solver is called.

[0013] In a practical application, the modified knowledge graph described in this invention reduces the confidence weight of triples that fail the consistency verification or deletes them; or modifies the relation description to a dynamic process description that matches the abnormal physical phenomenon, automatically generates and suggests new triples to the knowledge graph.

[0014] In a practical application, the present invention uses existing methods to verify the consistency of relationships for triples where entities do not have material properties or relational predicates do not involve physical or spatial relationships.

[0015] This invention also provides a knowledge graph relationship consistency verification system based on physical simulation feedback, including... The semantic extraction and filtering module parses entity relation triples in the knowledge graph and extracts triples where the entity has material attributes and the relation predicate involves physical or spatial relations as triples to be verified. The semantic parameterization module obtains the physical simulation engine's recognizable parameters and constraints for the triples to be verified through preset mapping rules. The scene reconstruction module, based on the parameters and constraints, constructs a three-dimensional physical scene in the virtual simulation space and sets initial boundary conditions; The physics simulation module drives the physics simulation engine to obtain the dynamic physical interaction state of each entity in the three-dimensional physical scene. The data monitoring module extracts key physical state data from the dynamic physical interaction state; The consistency verification module verifies the consistency of knowledge graph relationships based on whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description.

[0016] The present invention also provides an electronic device, including a processor and a memory storing computer program instructions; the processor executes the computer program instructions to implement the steps in the method of the present invention described above.

[0017] The present invention also provides a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the steps of the method of the present invention described above.

[0018] This invention transforms symbolic knowledge triples into a three-dimensional physical scene and drives a physics simulation engine to perform dynamic evolution calculations, thereby simulating and verifying the laws of the physical world. It solves the core problem that pure text models cannot understand the dynamic laws of physics; it can identify erroneous knowledge with implicit physical conflicts, such as common-sense illusions like "water is stored in a basket" or "a feather crushes a stone"; the verification results are based on real physical calculations, possessing high credibility and reliability; compared to existing methods based on text or vector spaces, physics simulation can capture complex physical characteristics such as temporal evolution, spatial interaction, and energy conversion.

[0019] This invention avoids expensive physical simulation of all relationships by using a pre-defined semantic filtering mechanism and physical proxy model, significantly reducing computational costs and making physical verification of large-scale knowledge graphs feasible.

[0020] This invention proposes a counterfactual relation generation mechanism that can automatically generate new triples that conform to physical laws based on anomalies observed during physical simulation. It not only eliminates erroneous knowledge but also proactively enriches the semantic hierarchy of the knowledge graph; forming a closed-loop optimization mechanism of "verification-correction-enrichment" to continuously improve the quality of the knowledge graph; and the generated counterfactual relations have high physical credibility and can serve as candidates for knowledge graph expansion.

[0021] This invention addresses the common problem of missing parameters in knowledge graphs by employing probability intervals and Monte Carlo sampling methods. It enables effective verification even in the absence of specific physical parameters; verification using multiple parameter instances covers the uncertainties of the parameter space, avoiding misjudgments; and it infers reasonable parameter value ranges from the verification results, supplementing the knowledge graph with parameter information.

[0022] This invention generates a causal analysis report at the physical level by monitoring state data and events during the physical simulation process. The verification results have clear physical basis, allowing users to understand the physical causes of erroneous knowledge; the causal analysis report provides guidance for correcting the knowledge graph; and it improves the credibility and user acceptance of the verification results. This invention provides a heterogeneous simulation architecture with multi-physics engine collaboration. A collaborative simulation architecture is designed for different types of physical phenomena, and it can automatically select appropriate simulation algorithms based on entity attributes. For fluid entities, the SPH algorithm is used; for flexible bodies, the PBD algorithm is used; and for rigid bodies, the rigid body dynamics solver is used, all coordinated and managed through a unified interface layer. This enables efficient handling of complex physical interaction scenarios, improving simulation accuracy and efficiency while maintaining system scalability. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 This is a system structure diagram of the present invention.

[0025] Figure 3 This is a flowchart illustrating a method according to an embodiment of the present invention.

[0026] Figure 4 This is a system structure diagram according to one embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and specific embodiments, but this does not constitute any limitation on the present invention.

[0028] The knowledge graph relationship consistency verification method based on physical simulation feedback provided by this invention has the following process: Figure 1 As shown, the system structure is as follows Figure 2 As shown.

[0029] In a practical application, the overall architecture of the system shown in this invention is as follows: Figure 4 As shown, it mainly includes the following core modules: (1) Semantic extraction and filtering module: used to parse the entity attributes and relation predicates of the entity relation triples to be verified in the knowledge graph. When it is identified that the entity has material attributes and the relation predicate involves physical or spatial relations, the physical consistency verification process is triggered. (2) Semantic parameterization module: It is used to convert the triples to be verified into physical material parameters, geometric models and dynamic constraints that can be recognized by the physical simulation engine through a pre-set mapping rule library; (3) Scene reconstruction module: used to automatically construct the corresponding three-dimensional physical scene in the virtual simulation space based on the parameters and constraints, and set the initial boundary conditions; (4) Physics simulation module: used to drive the physics simulation engine to perform time step iterations and calculate the dynamic physical interaction state of each entity in the three-dimensional physical scene; (5) Data monitoring module: used to extract physical state data from the dynamic physical interaction state of the simulation process; (6) Consistency verification module: compare the physical state data with the constraints of the semantic description of the knowledge graph. If the two do not match, the corresponding knowledge graph relationship is determined to have failed the consistency verification and the knowledge graph is corrected.

[0030] In a practical application, the overall process of the knowledge graph relationship consistency verification method based on physical simulation feedback provided by this invention is as follows: Figure 3 As shown.

[0031] The specific steps are as follows: Step S1: Semantic Extraction and Filtering S1.1: Obtain the triple T = <subject entity s, relational predicate p, object entity o> from the knowledge graph.

[0032] S1.2: Parse entity attributes: Query the ontology library to obtain the type tags of entities s and o; query the entity knowledge base to obtain the attribute list and attribute values ​​of the entities; determine whether the entities have material attributes (Matter type or related attributes).

[0033] S1.3: Analyze relational predicates: Perform semantic decomposition on relational predicate p to identify its core semantics and implicit semantics; determine whether the predicate implies physical interaction relationships (such as collision, support, crushing, etc.); determine whether the predicate implies spatial relationships (such as inclusion, contact, attachment, etc.); determine whether the predicate implies causal relationships (such as causing, inducing, influencing, etc.).

[0034] S1.4: Trigger Decision: If the entity has material attributes and the relational predicate implies physical or spatial relations, then the physical verification process is triggered and proceed to step S2; otherwise, physical verification is skipped and marked as "no physical verification required" or other verification methods are used.

[0035] Step S2: Semantic parameterization S2.1: Material Parameter Mapping: Query the material knowledge base to map entity names to physical material parameters; for fluid entities, obtain parameters such as density, viscosity, and surface tension; for rigid entities, obtain parameters such as density, elastic modulus, friction coefficient, and fracture threshold; for flexible entities, obtain parameters such as mass, stiffness, damping, and deformation threshold.

[0036] S2.2: Geometric Model Acquisition: Query the geometric model library to obtain the 3D geometric model of the entity; if the model does not exist in the library, call the parametric modeling program to generate an approximate model; set the model's size, position, orientation and other parameters.

[0037] S2.3: Constraint transformation: Transform relational predicates into physical constraints. For example, "stored in" is transformed into the fluid particle coordinate set being completely within the container bounding box, "crush" is transformed into the target entity undergoing fracture deformation or structural collapse, and "support" is transformed into contact force balance with no penetration.

[0038] S2.4: Fuzzy parameter imputation (when parameters are missing): Determine the parameter value range based on the general statistical values ​​of entity categories; perform Monte Carlo sampling within the range to generate N sets of parameter instances (N is recommended to be 100-1000); perform subsequent verification on each set of parameter instances, and only when all instances pass the verification is the consistency of the knowledge graph relationship deemed to have passed.

[0039] Step S3: Scene Reconstruction S3.1: Scene Initialization: Create a scene container in the virtual simulation space; set global physical parameters (gravitational acceleration 9.8 m / s²). 2 Set simulation time parameters (total duration, time step, etc.); set air resistance coefficient, etc.

[0040] S3.2: Entity Instantiation: Based on the geometric model and material parameters obtained in step S2, instantiate each entity in the scene; set the initial position, velocity, attitude and other states of the entity; for fluid entities, set the parameters of the fluid emitter (emission position, emission rate, particle size, etc.).

[0041] S3.3: Constraint Configuration: Based on the constraints transformed in step S2, configure the corresponding physical constraints in the scene; set boundary conditions (such as fixed boundaries, periodic boundaries, etc.); configure collision detection and response parameters.

[0042] S3.4: Sensor Deployment: Deploy virtual sensors at key locations to monitor physical state data, such as deploying a particle counter at the bottom of a fluid container to monitor retention rate; or deploying a stress sensor on a rigid body surface to monitor impact impulse.

[0043] Step S4: Optimized physical simulation and data monitoring S4.1: Proxy Model Prediction (Optimization Step): Input the scene parameters into the pre-trained physical proxy model (such as PINNs network); the proxy model outputs the predicted physical state and confidence level; if the prediction confidence level is ≥ the threshold (such as 0.9), the prediction result is used directly, and step S4.2 is skipped; if the prediction confidence level is < the threshold, step S4.2 is entered to start full-precision simulation. S4.2: Full-Precision Physical Simulation: First, select a suitable physics engine and algorithm: for fluid-like entities, use the SPH (Smooth Particle Hydrodynamics) algorithm; for soft-body-like entities, use the PBD (Position-Based Dynamics) algorithm; for rigid-body-like entities, use the rigid-body dynamics solver. Second, execute time-step iterations, including detecting collisions and contacts, calculating internal and external forces, solving motion equations, updating entity states (position, velocity, deformation, etc.), recording sensor data, etc., and monitoring key physical indicators (such as particle retention rate, peak stress, deformation, etc.) in real time.

[0044] S4.3: Anomaly Detection: Real-time detection of abnormal physical events during simulation. Common anomalies include: Leakage: Fluid particles penetrate the container; Clipping: Rigid bodies penetrate each other; Collapse: Structural instability and collapse; Sliding: Sliding of contact surfaces; Bounce: Bounce after a collision; Deformation: Unexpected deformation.

[0045] Step S5: Consistency Verification and Feedback S5.1: Consistency Judgment: Compare the simulation results with the semantic description of the original triplet; extract key indicators (such as retention rate, stress, deformation, etc.) in the simulation process; judge whether it violates physical laws based on the preset threshold: if the key indicators meet the expectations, it is determined that the original knowledge is physically valid, that is, it is consistent; if the key indicators violate the expectations, it is determined that the original knowledge is physically invalid, that is, it is inconsistent.

[0046] S5.2: Causal Analysis: When a knowledge graph relationship fails consistency verification, analyze the physical causes; identify key factors leading to violations of physical laws; generate causal chains: for example, "Leakage → Aperture larger than particle diameter → Net cannot stop gasoline". S5.3: Knowledge Graph Correction: Reduce the confidence weight of triples that fail consistency verification (e.g., reduce to 0.1 or delete directly); generate counterfactual relation triples based on the type of abnormal event: Leakage event → generate "<subject, leaked at, object>", Sliding event → generate "<subject, slipped at, object>", Bouncing event → generate "<subject, impact, object>" or "<subject, landed on, object>", Deformation event → generate "<subject, deformed at, object>" or "<subject, adapted, object>"; insert the newly generated triples into the knowledge graph or push them to the review queue.

[0047] S5.4: Validation Report Generation: Generate a detailed validation report, including the original triplet content, validation results (valid / invalid), key physical index data, descriptions and cause analyses of abnormal events, suggested corrective actions, confidence scores, etc.

[0048] The following are several specific embodiments to illustrate the method and system of the present invention.

[0049] Example 1: Verification of the physical consistency of fluid-container relationship This embodiment details how to verify the consistency of the relationship of the triple T = <gasoline, stored in, net>.

[0050] (1) Problem Background The knowledge graph contains the relation triple <gasoline, stored in, net>, which is syntactically valid, but is it feasible in the physical world? A net is a porous, flexible container, and gasoline is a low-viscosity fluid. The validity of the "store" relation needs to be verified through physical simulation.

[0051] (2) Execution process Step S1: Semantic Extraction and Scene Filtering Parsing the entity "gasoline": Querying the ontology, the type tag for "gasoline" is "Matter", and it has the "fluid" attribute; Parsing the entity "Net": Querying the ontology database, the type tag for "Net" is "Artifact", and it has the attributes "container" and "soft body". The relational predicate "stored in" is analyzed: semantic analysis identifies it as a spatial containment relationship, implying physical interaction; Triggering decision: The subject is matter, and the predicate implies spatial relationships, triggering the physical verification process.

[0052] Step S2: Semantic parameterization Material mapping: "Gasoline" is mapped to fluid parameters, with a density ρ = 740 kg / m³. 3 The kinematic viscosity ν = 0.6 cSt and the surface tension σ = 0.02 N / m are given. The "net" is mapped to flexible body parameters: mesh aperture d = 5 mm (known), mesh wire diameter δ = 0.5 mm, and the material is nylon (elastic modulus E = 3 GPa). Geometric model: "Gasoline", no fixed geometry, represented by a particle system, with particle diameter set to 1 mm; "Net", a cylindrical mesh container, 15 cm in diameter and 20 cm in height; Constraint transformation: The constraint condition "stored at" is transformed into the evolution end time, and the gasoline particle retention rate R ≥ 80% (threshold setting); Fuzzy parameter processing: In this example, the mesh aperture is known, so sampling is unnecessary. If the aperture is unknown, sampling will be performed within the range of [2mm, 10mm].

[0053] Step S3: Scene Reconstruction Scene initialization: Create a virtual space of 10m × 10m × 10m with gravity g = 9.8 m / s². 2 down; Entity instantiation: Generate a net model in the center of the scene with fixed edges (constraining 6 degrees of freedom); place a fluid emitter 10 cm directly above the net with an emission rate of 10 ml / s and an emission duration of 1 second; the number of gasoline particles is approximately 10,000 (based on volume calculation). Constraint configuration: Set particle-mesh collision detection, friction coefficient μ = 0.3; Sensor deployment: Particle counters are deployed at the bottom and sides of the net to count the number of particles in real time.

[0054] Step S4: Simulation Evolution Calculation Proxy model prediction: Assuming this is the first time this instance is validated and there is no historical data, skip the proxy model; Full-precision simulation: The SPH algorithm using the NVIDIA Flex engine (or Bullet Physics engine, MuJoCo engine) is employed; time step Δt = 0.001 s, total simulation time T = 10 s (10,000 steps); each step calculates: inter-particle interaction forces (pressure, viscosity), particle-mesh collision forces, and gravity; real-time monitoring reveals that in the first second, the emitter continuously emits particles, increasing the number of particles in the net; in the second second, emission stops, and particles begin to pass through the mesh holes; from the third to the tenth second, particles continuously leak, decreasing the number of particles in the net; at the tenth second: the particle retention rate R = 4.2%; Anomaly Detection: A "leak" event has been detected, with a large number of particles passing through the mesh holes.

[0055] Step S5: Consistency Verification and Feedback Consistency judgment: If the retention rate R = 4.2% < the threshold of 80%, the original triplet is determined to be physically invalid; Causal analysis: The physical reason is that the particle diameter (1 mm) is less than the mesh aperture (5 mm), allowing the particle to pass through the mesh under the influence of gravity. The viscous force is insufficient to resist gravity, preventing the particle from adhering to the mesh. Knowledge graph correction: Delete the original triple <gasoline, stored in, net>, confidence level reduced to 0; generate counterfactual relation <gasoline, leaked in, net>, confidence level 0.9; generate new knowledge <net, not applicable to, fluid storage>; Verification Report: Verification Result: Not valid; Original Triad: <Gasoline, Stored in, Mesh Bag>; Key Indicator: Particle Retention Rate = 4.2% (Threshold 80%); Abnormal Event: Leakage; Physical Cause: Mesh Aperture (5mm) > Particle Diameter (1mm), Viscous Force Insufficient to Resist Gravity; Suggested Correction: <Gasoline, Leaked in, Mesh Bag>.

[0056] (3) Technical effects Through this embodiment, the system successfully identified and eliminated erroneous knowledge that violated the laws of physics, and generated new knowledge that conformed to physical reality, thus verifying the effectiveness of the present invention.

[0057] Example 2: Verification of the physical consistency of the interaction relationship between a flexible body and a rigid body This embodiment details how to verify the consistency of the relationship of the triple T = <feather, crushed, stone>, and solve the problem of "common sense illusion".

[0058] (1) Problem Background Generative AI models may output relational triples <feather, crush, stone>, which are syntactically valid but severely violate common sense in physics. The validity of this relation needs to be verified through physical simulation.

[0059] (2) Execution process Step S1: Semantic Extraction and Scene Filtering Analyzing the entity "feather": the type tag is "Matter", and it has the attributes of "soft body" and "lightweight"; The entity "Stone" is analyzed: its type label is "Matter", and it has the attributes of "rigid body" and "high hardness". The relational predicate "crushing" was analyzed and identified as a physical stress relationship, implying structural damage. Triggering decision: Both the subject and object are matter, and the predicate implicitly contains a physical stress relationship, triggering verification.

[0060] Step S2: Semantic parameterization Material mapping: The "feather" is mapped to soft body parameters, with mass m = 0.005 kg and density ρ = 500 kg / m³. 3 The elastic modulus E = 0.1 GPa, stiffness k = 10 N / m; the "stone" is mapped to rigid body parameters, density ρ = 2600 kg / m³. 3 Elastic modulus E = 50 GPa, fracture threshold σ_c = 10 MPa; Geometric models: "Feather", a thin, flexible body model, 10 cm long, 2 cm wide, and 0.5 mm thick; "Stone", a rigid sphere model, 5 cm in radius; Constraint transformation: Transforming the constraint condition "crushing" into the stone producing fracture deformation or structural collapse, i.e., stress σ > fracture threshold σ_c; Fuzzy parameter processing: Feather weight is estimated using intervals: [0.001 kg, 0.01 kg], with 100 parameter instances sampled within the interval.

[0061] Step S3: Scene Reconstruction Scene initialization: Create a virtual space with gravity g = 9.8 m / s². 2 down; Entity instantiation: Generate a stone model at the bottom of the scene and fix its position; generate a feather model 50 cm directly above the stone and give it an initial velocity v = 0 (free fall), or give the feather an initial velocity v = 5 m / s downward (simulate throwing). Constraint configuration: Set collision detection, restitution coefficient e = 0.3; Sensor deployment: Deploy stress sensors on the stone surface to monitor impact stress, feather deformation, and bounce height.

[0062] Step S4: Simulation Evolution Calculation Proxy model (PINNs) prediction: Input parameters, the Proxy model predicts collision stress σ ≈ 0.001 MPa, confidence level 0.85. Since the confidence level < threshold 0.9, full-precision simulation is started. Gaussian process regression (GPR) and decision tree ensemble methods (such as Random Forest, XGBoost) can also be used to replace the PINNs physical Proxy model.

[0063] Full-precision simulation: The PBD algorithm is used to handle the soft body (feather), and the rigid body solver is used to handle the stone. The time step Δt = 0.0001 s, and the total simulation time T = 2 s. The simulation process is as follows: from 0 to 0.32 seconds, the feather is in free fall, and its velocity increases. At 0.32 seconds, the feather contacts the stone surface, and a collision occurs. The collision calculation is: feather kinetic energy E_k = 0.5 × m × v 2 =0.5 × 0.005 × (0.32 × 9.8) 2 The impact impulse is approximately 0.0008 J, I = √(2 × m × E_k) ≈0.0028 N·s, and the impact stress is σ = I / A ≈ 0.0014 MPa (assuming the contact area A = 2 cm²). 2 Analysis showed that the impact stress σ = 0.0014 MPa << the fracture threshold σ_c = 10 MPa. Between 0.32 and 0.5 seconds, the feather underwent flexible deformation and then rebounded. After 0.5 seconds, the feather either adhered to the stone surface or rebounded slightly before falling. Anomaly detection: No stone breakage or collapse was detected, but "bounce" and "deformation" events were detected.

[0064] Step S5: Consistency Verification and Feedback Consistency judgment: The collision stress σ = 0.0014 MPa << the fracture threshold σ_c, and no "crushing" phenomenon occurred. Therefore, the original triplet is determined to be physically invalid. Causal analysis: Physical reasons: feathers are extremely light (0.005 kg) and have very little kinetic energy (0.0008 J), resulting in impact stress far below the fracture threshold of the stone. Stones are extremely hard, while feathers are extremely soft; the impact energy is primarily absorbed by the flexible deformation of the feathers. Knowledge graph correction: The original triple <feather, crush, stone> was deleted, and the confidence level was reduced to 0. Based on actual observed phenomena (collision, bounce, attachment), the counterfactual relations <feather, land on, stone> were generated with a confidence level of 0.95, <feather, cover, stone> with a confidence level of 0.85, and <feather, impact, stone> with a confidence level of 0.80. New knowledge was generated: <feather, cannot be crushed, stone> with a confidence level of 0.98. Verification Report: Verification Result: Not valid; Original Triad: <Feather, Crushed, Stone>; Key Indicator: Impact Stress = 0.0014 MPa (Fracturing Threshold 10 MPa); Abnormal Events: Rebound, Deformation; Physical Cause: Feathers are extremely light, and their kinetic energy is insufficient to generate destructive stress; Suggested Correction: <Feather, Falling, Stone>.

[0065] (3) Technical effects This embodiment successfully identified and eliminated "common sense illusions," verifying the effectiveness of the invention in addressing erroneous knowledge output by generative AI models. Simultaneously, the system generated new knowledge consistent with common sense based on physical simulation results, achieving proactive enrichment of the knowledge graph.

[0066] Example 3: Verification under multi-parameter uncertainty scenarios This embodiment illustrates how to handle the situation of missing parameters in a knowledge graph, and verifies it through fuzzy parameter imputation and Monte Carlo sampling.

[0067] (1) Problem Background In a knowledge graph, there exists a relation triple T = <water, stored in, container X>, where the type of "container X" is known to be "mesh container," but the specific pore size parameter is unknown. The consistency of this relation needs to be verified.

[0068] (2) Execution process Step S2: Semantic parameterization (fuzzy parameter imputation) Material mapping: "Water" is mapped to fluid parameters, with density ρ = 1000 kg / m³. 3 Viscosity ν = 1.0 cSt; "Container X" is mapped to flexible body parameters, pore size d is unknown, but the type is known to be "mesh container". Parameter range determination: Query the statistical database of the "mesh container" category. The aperture range is [d_min, d_max] = [0.5 mm, 20 mm]; the particle diameter is set to d_p = 2 mm (based on the surface tension of water).

[0069] Monte Carlo sampling (or Bayesian optimization): Randomly sample N = 500 aperture values ​​within the interval [0.5 mm, 20 mm]; construct a scene for each aperture value and perform simulation.

[0070] Validation strategy: If all sampled instances pass the validation (retention rate R ≥ 80%), the knowledge graph relationship is considered to have passed the consistency validation; if some instances pass, the pass rate and parameter sensitivity analysis are output; if all instances fail, the knowledge graph relationship is considered to have failed the consistency validation.

[0071] Steps S3-S5: Perform scenario reconstruction, evolutionary calculation, and consistency verification for each set of parameter instances. (3) Verification results After 500 sets of simulation verification: For examples with aperture d < 2 mm (approximately 120 sets): particle retention rate R ≥ 95%, which passed the test; For examples with aperture d ∈ [2 mm, 5 mm] (approximately 200 sets): particle retention rate R ∈ [30%, 95%], which partially passed the test; For examples with aperture d > 5 mm (approximately 180 sets): particle retention rate R < 30%, which failed the test.

[0072] Conclusion: The validity of the original triplet depends on the container pore size, and it holds true when the pore size is smaller than the particle diameter. It is recommended to supplement the knowledge graph with parameter information, or label it as "condition holds". Generate new knowledge: <Water, can be stored in, small-pore mesh containers>, condition: pore size < 2 mm As described in the embodiments, this invention solves the core problem that pure text models cannot understand the dynamic laws of physics; it can identify erroneous knowledge with implicit physical conflicts, such as common-sense illusions like "water is stored in a basket" or "a feather crushes a stone"; the verification results are based on real physical calculations and have high credibility and reliability. Through a pre-implemented semantic filtering mechanism, the scale of knowledge to be verified is compressed to 5%-15% of the original, significantly reducing computational costs; through a physical proxy model, it can quickly predict about 60%-80% of scenarios without starting full-precision simulation. Overall computational efficiency is improved by 3-5 times, making physical verification of large-scale knowledge graphs feasible.

[0073] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Those skilled in the art can modify or make equivalent substitutions to the technical solutions of the present invention. The scope of protection of the present invention is defined by the claims.

Claims

1. A method for verifying the consistency of knowledge graph relationships based on physical simulation feedback, comprising the following steps: Analyze the entity relation triples in the knowledge graph and extract triples where the entity has material attributes and the relation predicate involves physical or spatial relations as triples to be verified. By using preset mapping rules, the physical simulation engine can identify the parameters and constraints of the triplet to be verified. Based on the parameters and constraints, a three-dimensional physical scene is constructed in the virtual simulation space, and initial boundary conditions are set. Drive the physics simulation engine to obtain the dynamic physical interaction state of each entity in the three-dimensional physical scene; Monitor physical state data and extract physical state data from the dynamic physical interaction state; The consistency of knowledge graph relationships is verified by checking whether the physical state data conforms to the constraints of the corresponding semantic description of the knowledge graph.

2. The method as described in claim 1, characterized in that, If the physical state data does not conform to the constraints of the corresponding knowledge graph semantic description, the corresponding knowledge relationship is determined to have failed the consistency verification and the knowledge graph is corrected.

3. The method as described in claim 1, characterized in that, If the knowledge graph is missing parameters, the parameter range is determined based on the general statistical values ​​of the entity categories; Bayesian optimization or Monte Carlo sampling is performed within the parameter range to generate multiple sets of simulation parameter instances for verification; and the corresponding knowledge relationship is determined to have passed the consistency verification only when all parameter instances pass the verification.

4. The method as described in claim 1, characterized in that, First, the parameters and constraints are input into the pre-trained physical proxy model to obtain physical state data and prediction confidence. If the prediction confidence is higher than a preset threshold, the consistency of knowledge graph relationships is verified based on whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description.

5. The method as described in claim 1, characterized in that, For fluid-type entities, the smooth particle fluid dynamics algorithm is invoked, and the coordinate set of fluid particles is monitored in real time to see if it is completely within the bounding box of the container. If the particle retention rate is lower than a preset threshold after a preset simulation time, it is determined that it is physically invalid. For soft-body-type entities, the position-based dynamics algorithm is invoked. For rigid-body-type entities, the rigid-body dynamics solver is invoked.

6. The method as described in claim 1, characterized in that, The modified knowledge graph involves reducing the confidence weight of triples that fail the consistency verification, or deleting them; or modifying the relation description to a dynamic process description that matches abnormal physical phenomena, automatically generating and suggesting new triples to the knowledge graph.

7. The method as described in claim 1, characterized in that, For triples where the entity does not have material properties or the relational predicate does not involve physical or spatial relationships, existing methods are used to verify relational consistency.

8. A knowledge graph relationship consistency verification system based on physical simulation feedback, including... The semantic extraction and filtering module parses entity relation triples in the knowledge graph and extracts triples where the entity has material attributes and the relation predicate involves physical or spatial relations as triples to be verified. The semantic parameterization module obtains the physical simulation engine's recognizable parameters and constraints for the triples to be verified through preset mapping rules. The scene reconstruction module, based on the parameters and constraints, constructs a three-dimensional physical scene in the virtual simulation space and sets initial boundary conditions; The physics simulation module drives the physics simulation engine to obtain the dynamic physical interaction state of each entity in the three-dimensional physical scene. The data monitoring module extracts physical state data from the dynamic physical interaction state; The consistency verification module verifies the consistency of knowledge graph relationships based on whether the physical state data conforms to the constraints of the corresponding knowledge graph semantic description.

9. An electronic device comprising a processor and a memory storing computer program instructions; wherein the processor executes the computer program instructions to implement the steps of the methods described in claims 1-7.

10. A computer-readable storage medium storing computer program instructions that, when executed by a processor, implement the steps of the methods described in claims 1-7.

Images