An agent development and product optimization method and apparatus

Abstract

The application discloses an agent development and product optimization method and equipment. The method comprises the following steps: calling a product design agent encapsulating an industry knowledge graph; generating an initial design scheme through an inter-agent communication protocol; searching for optimization in a design parameter space by using a proxy model; constructing a physical simulation model; inputting physical sensor data as a boundary condition into the physical simulation model to simulate and predict data; calculating the error of the sensor and the simulation prediction data by using a neural network algorithm based on physical information; if the error exceeds a preset setting value, updating the material physical parameters by using a loss function containing data items and physical constraint items and transmitting the updated material physical parameters to the physical simulation model for re-simulation; if the scheme does not exceed the preset setting value, judging whether the performance index meets the design requirement, and if yes, outputting a final product design scheme. The application improves the prediction accuracy of the simulation model.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method, device, storage medium and program product for developing and optimizing intelligent agents. Background Technology

[0002] With the rapid development of artificial intelligence technology, intelligent agents, as software entities capable of autonomously perceiving their environment, making decisions, and executing tasks, have been widely used in product design and engineering simulation. To lower the development threshold, existing intelligent agent development platforms typically provide visual orchestration systems that allow users to build workflows by dragging and dropping components, and integrate intelligent agent repositories containing pre-built modules.

[0003] In some product design applications of related technologies, the working mode typically manifests as an open-loop processing flow based on preset rules. Specifically, the system invokes the design agent generation scheme according to user configuration and relies on integrated simulation modules for performance verification. The operation of these simulation modules is based on standard material parameters stored in a material library. These parameters are usually average values ​​measured under idealized laboratory conditions, possessing a high degree of determinism and ideality.

[0004] However, in the actual manufacturing process, the performance of product materials is affected by a variety of real-world factors such as processing technology, ambient temperature and humidity, and batch differences, which cause deviations between their actual physical properties and standard parameters. This results in a large error between the simulation prediction results of related technologies and the actual test results. Summary of the Invention

[0005] This application provides a method and apparatus for developing and optimizing intelligent agents, which can improve the accuracy of intelligent agent design results.

[0006] Firstly, this application provides a method for developing and optimizing intelligent agents, comprising: creating an intelligent agent workflow; calling a product design intelligent agent encapsulated with an industry knowledge graph from an intelligent agent repository; mapping the output results of a general intelligent agent in the software domain to the input parameters of the product design intelligent agent through an inter-agent communication protocol, wherein the message format defined by the inter-agent communication protocol includes a knowledge graph reference field; inputting the input parameters into the product design intelligent agent to generate an initial design scheme; using a proxy model to search in the design parameter space of the initial design scheme to obtain preliminary optimized product design parameters; constructing a corresponding physical simulation model based on the preliminary optimized product design parameters; and inputting physical sensor data as simulation boundary conditions into the physical simulation model. The simulation prediction data is obtained, and the physical sensor data is the data collected by physical sensors deployed in the product prototype or product operating environment. The physical sensor data includes at least one of temperature, pressure and vibration parameters. The error between the physical sensor data and the simulation prediction data is calculated using a neural network algorithm based on physical information. If the error exceeds a preset threshold, the material physical parameters are updated using a loss function that includes data fitting terms and physical constraint terms, and the updated material physical parameters are fed back to the physical simulation model for resimulation. If the error does not exceed the preset threshold, it is determined whether the physical performance indicators obtained from the simulation calculation meet the design requirements. If they do not meet the requirements, the product design parameters are readjusted and the physical simulation model is rebuilt. If they meet the requirements, the final product design scheme is output.

[0007] This embodiment uses temperature, pressure, and vibration parameters collected by physical sensors deployed in the product prototype or actual operating environment as simulation boundary conditions input into the physical simulation model. This allows the simulation calculation to obtain constraint information from the real physical system. A neural network algorithm based on physical information is used to calculate the error between the physical sensor data and the simulation prediction data. When this error exceeds a preset threshold, the material physical parameters are updated using a loss function that includes data fitting terms and physical constraint terms, and then fed back to the physical simulation model for recalculation. This constructs a closed-loop feedback mechanism from sensor measurement to simulation correction. This mechanism transforms material parameters from static standard values ​​to dynamically calibrated values ​​based on actual physical responses. While the data fitting term ensures that the simulation results closely approximate the measured data, the physical constraint term ensures that the updated material parameters still conform to the basic constraints of physical laws, improving the accuracy of the simulation model in predicting the physical performance of the real product.

[0008] In conjunction with some embodiments of the first aspect, in some embodiments, the industry knowledge graph is stored using a graph database, including material nodes, process nodes, and constraint relationship edges. The message format defined by the inter-agent communication protocol includes a message header, a message body, and a knowledge graph reference field. The knowledge graph reference field includes a knowledge entry identifier and a query condition expression. The output results of the general intelligent agent in the software domain are mapped to the input parameters of the product design intelligent agent through the inter-agent communication protocol. Specifically, this includes: constructing the message body of the general intelligent agent according to the inter-agent communication protocol; filling in the type identifier and attribute constraint conditions of the target knowledge entry in the knowledge graph reference field; and filling the output results into the message body. After receiving the message body, the product design intelligent agent locates the target knowledge entry in the industry knowledge graph according to the knowledge graph reference field; and associating the output results in the message body with the target knowledge entry to generate the input parameters of the product design intelligent agent.

[0009] This embodiment utilizes a knowledge graph reference field in the message format of the inter-agent communication protocol. This field includes a knowledge entry identifier and a query condition expression, allowing the general agent to fill in the type identifier and attribute constraints of the target knowledge entry when constructing the message body. After receiving the message, the product design agent directly locates the target knowledge entry in the industry knowledge graph based on the reference field and associates it with the output result in the message body. This transforms the cross-domain data mapping process from implicit inference at the receiving end to explicit declaration at the sending end. This pre-reference mechanism ensures that the output of the general agent carries precise location information pointing to the product design domain knowledge entity during the transmission stage, reducing the semantic alignment cost in cross-domain agent collaboration and improving the accuracy of the conversion from general requirements to domain design input.

[0010] In conjunction with some embodiments of the first aspect, in some embodiments, input parameters are input into the product design agent to generate an initial design scheme. Specifically, this includes: sending the input parameters transmitted via the inter-agent communication protocol to the product design agent; the input parameters including product functional requirements and performance constraints; retrieving candidate material nodes that meet the performance constraints from the industry knowledge graph based on the input parameters, wherein the yield strength and elastic modulus of the candidate material nodes both meet the lower limit requirements of the performance constraints; traversing the constraint relationship edges of the candidate material nodes, filtering process nodes compatible with the candidate materials, and forming a set of material-process combination schemes; sorting the set of material-process combination schemes according to a preset optimization objective function, wherein the optimization objective function comprehensively considers material cost parameters and process complexity, and selecting a preset number of optimal combination schemes of objective function values ​​as candidate design schemes; determining the geometric topology of the product structure based on the product functional requirements, assigning corresponding material properties to each candidate design scheme, and generating an initial design scheme including geometric dimensions, material number, and process type identifier.

[0011] This embodiment stores material-process compatibility knowledge in an industry knowledge graph by connecting material nodes and process nodes with constraint edges. After retrieving candidate material nodes based on performance constraints, compatible process nodes are directly selected by traversing the constraint edges of these candidate material nodes. Material performance verification and process compatibility judgment are integrated into a single traversal of the knowledge graph. Each combination in the resulting material-process combination scheme set has been verified by the constraints implicit in the graph structure. This parallel verification mechanism based on graph traversal enables the system to simultaneously meet performance constraints and process compatibility requirements when generating candidate design schemes, reducing the number of iterations in the initial design scheme generation process and improving the efficiency of generating a set of feasible design schemes from input parameters.

[0012] In conjunction with some embodiments of the first aspect, in some embodiments, a surrogate model is used to search in the design parameter space to obtain preliminary optimized product design parameters. Specifically, this includes: using a Gaussian process regression model as a surrogate model, generating an initial sample point set in the design parameter space composed of the geometric dimensions of the initial design scheme using the Latin hypercube sampling method; for each sample point in the initial sample point set, calling a simplified physical model for rapid performance evaluation to obtain the corresponding performance response value; training the Gaussian process regression model based on the initial sample point set and the corresponding performance response value; determining the next point to be evaluated in the design parameter space using the expected improvement criterion, which is used to calculate the expected improvement of the unsampled point in the current design parameter space relative to the known optimal performance response value; performing performance evaluation on the point to be evaluated, adding the point to be evaluated and its corresponding performance response value to the sample point set and updating the Gaussian process regression model; repeating the steps of determining the point to be evaluated and updating the model until the improvement in the performance response value is less than the convergence threshold or the number of iterations reaches the preset maximum number of iterations in a consecutive preset number of iterations; and outputting the design parameters corresponding to the optimal performance response value obtained during the iteration process as the preliminary optimized product design parameters.

[0013] This embodiment constructs a probabilistic proxy mapping from design parameters to performance response using a Gaussian process regression model. This model can not only predict the mean performance response of unsampled locations based on the evaluated sample point set, but also quantify the uncertainty of the prediction. Furthermore, it employs an expected improvement criterion to calculate the expected improvement of unsampled points relative to the known optimal performance response value in the design parameter space. This criterion comprehensively considers the potential for improvement in the predicted response value and the exploratory value of the prediction uncertainty, positioning the next point to be evaluated at a location that balances local depth optimization and global region exploration. This optimization mechanism based on a probabilistic model and adaptive sampling criterion allows the system to dynamically adjust the sampling strategy according to existing evaluation results, concentrating computational resources in regions with high expected improvement while avoiding repeated sampling in low-value regions. This reduces the number of performance evaluations required to reach the convergence threshold and improves the computational resource utilization efficiency of the design parameter optimization process.

[0014] In conjunction with some embodiments of the first aspect, in some embodiments, physical sensor data is used as simulation boundary conditions and input into a physical simulation model to obtain simulation prediction data. Specifically, this includes: receiving temperature, pressure, and vibration parameters collected by physical sensors and preprocessing the physical sensor data; mapping the preprocessed physical sensor data to corresponding nodes or units in the physical simulation model based on the deployment location of the physical sensors in the product prototype or product operating environment; setting the mapped physical sensor data as boundary conditions of the physical simulation model, where the boundary conditions include at least one of temperature boundaries, pressure boundaries, and displacement boundaries; and solving the physical simulation model based on the boundary conditions. The simulation model's control equations are used to obtain the distribution results of temperature, stress, and flow fields. It is then determined whether coupling effects exist between the physical fields in the distribution results. If coupling effects exist, multiphysics coupling calculations are performed. Coupling effects include at least one of thermal-stress coupling, fluid-structure interaction, and electrothermal coupling. In the multiphysics coupling calculations, coupling transfer relationships are established between the control equations of adjacent physical fields according to the coupling type. The control equations of each physical field are solved alternately using an iterative solver until the calculation results of each physical field converge. The physical quantity values ​​of nodes or units corresponding to the physical sensor deployment locations in the converged physical field distribution results are extracted as simulation prediction data.

[0015] This embodiment maps sensor measurement data to the spatial distribution boundary conditions of the simulation model, enabling the simulation calculation to obtain driving inputs that reflect real operating conditions. Temperature field calculation can capture non-uniform thermal strain in local hot spots based on measured temperature distribution. Stress field solution can identify the real load transfer path at the contact interface based on measured pressure boundary. Flow field analysis can reflect the dynamic response characteristics of the structure under fluid excitation based on measured vibration parameters. The coupling transmission relationship between the control equations established after determining the existence of coupling effects allows the thermal expansion of materials caused by the temperature field to change the geometric configuration of the stress field in real time. The material stiffness degradation caused by the stress field can feed back and affect the boundary deformation of the flow field. This bidirectional transmission mechanism between physical fields ensures that the coupling effect continues to propagate in the control equations of each field through the iterative solution process until a self-consistent equilibrium state is reached, improving the accuracy of the physical simulation model in reproducing the multi-physics coupling behavior of the product under actual operating conditions.

[0016] In conjunction with some embodiments of the first aspect, in some embodiments, after outputting the final product design scheme, the method further includes: storing agent development data, product design data, physical simulation data, and product operation data in a data platform to construct a data association graph, which includes agent version nodes, design parameter nodes, simulation result nodes, test data nodes, and fault record nodes; using a graph attention network algorithm to perform node embedding learning on the data association graph, generating feature vector representations of each node, calculating the association strength between nodes through the cosine similarity between feature vector representations, and using the association strength as the weight of each edge in the data association graph; searching for directed paths from fault record nodes to design parameter nodes in the data association graph, calculating path scores according to the cumulative weights of each edge on the path, and selecting a preset number of paths with the highest path scores as key causal paths; identifying the design parameter nodes that have the greatest impact on the occurrence of faults based on the key causal paths, generating parameter adjustment suggestions, which include target design parameters; and automatically passing the target design parameters in the parameter adjustment suggestions as constraints to the product design agent when creating a new agent workflow through a visual orchestration engine.

[0017] This embodiment employs a graph attention network algorithm to learn the feature vector representations of each node in the data association graph and quantify the association strength between nodes. This allows heterogeneous data that was originally stored in a scattered manner to establish semantic similarity-based connections within the graph structure. This deep learning-based association strength calculation can capture nonlinear implicit associations between data without relying on predefined association rules. When searching for directed paths from fault record nodes to design parameter nodes, the path score quantifies the causal transmission strength from fault phenomena to design decisions by accumulating the weights of each edge. The key causal path with the highest path score reveals which design parameters have a dominant influence on the occurrence of faults through simulation model settings, optimization target selection, or agent version differences. The identified target design parameters are automatically passed as constraints to the product design agent in subsequent workflows, enabling the system to avoid verified high-risk parameter configurations in advance during the new product design phase. This improves the automation and accuracy of extracting design knowledge from historical fault data and applying it to product iteration.

[0018] In conjunction with some embodiments of the first aspect, in some embodiments, before outputting the final product design scheme, the method further includes: calling a cost assessment agent from the agent repository, and passing the material number and process type identifier in the product design scheme that meets the performance requirements to the cost assessment agent; calculating the total product cost through the cost assessment agent based on the material cost parameters and process complexity in the industry knowledge graph; when the total product cost exceeds the preset cost limit, readjusting the product design parameters, and adding cost constraints when adjusting the design parameters.

[0019] This embodiment calculates the total product cost by calling a cost assessment agent before outputting the final design scheme. When the total cost exceeds a preset upper limit, the design parameters are readjusted, and cost constraints are added during the adjustment process. This ensures that subsequent parameter optimization is carried out under the dual constraints of performance objectives and cost constraints. This cost feedback-driven parameter adjustment mechanism can guide the optimization algorithm to search for low-cost regions in the design parameter space, such as by selecting cheaper alternative materials or simplifying process complexity to reduce the total cost. At the same time, it ensures that the adjusted design scheme still meets performance requirements, reduces the number of invalid optimizations caused by cost-performance mismatch during design iteration, and improves the efficiency of obtaining feasible product design schemes under budget constraints.

[0020] Secondly, embodiments of this application provide an intelligent agent development and product optimization device, which includes: one or more processors and a memory; the memory is coupled to the one or more processors, and the memory is used to store computer program code, the computer program code including computer instructions, and the one or more processors call the computer instructions to cause the intelligent agent development and product optimization device to perform the method as described in the first aspect and any possible implementation thereof.

[0021] Thirdly, embodiments of this application provide a computer program product containing instructions that, when the computer program product is run on an intelligent agent development and product optimization device, cause the intelligent agent development and product optimization device to perform the method described in the first aspect and any possible implementation thereof.

[0022] Fourthly, embodiments of this application provide a computer-readable storage medium including instructions that, when executed on an agent development and product optimization device, cause the agent development and product optimization device to perform the method described in the first aspect and any possible implementation thereof.

[0023] Understandably, the intelligent agent development and product optimization equipment provided in the second aspect, the computer program product provided in the third aspect, and the computer storage medium provided in the fourth aspect are all used to execute the methods provided in the embodiments of this application. Therefore, the beneficial effects they can achieve can be referred to the beneficial effects in the corresponding methods, and will not be repeated here.

[0024] One or more technical solutions provided in the embodiments of this application have at least the following technical effects or advantages:

[0025] 1. This application uses physical sensor data as simulation boundary conditions and employs a neural network algorithm based on physical information to calculate the error. When the error exceeds a threshold, the material parameters are updated through a loss function that includes data fitting terms and physical constraint terms and fed back to the simulation model. This constructs a closed-loop feedback mechanism that transforms the material parameters from static standard values ​​to dynamically calibrated values ​​based on the actual response, thereby improving the prediction accuracy of the simulation model.

[0026] 2. This application uses a Gaussian process regression model to predict performance response and quantify uncertainty. It uses the expected improvement criterion to locate the evaluation point in a position that takes into account both local optimization and global exploration. This allows the system to dynamically adjust the sampling strategy based on the evaluation results and concentrate computing resources in the region with high expected improvement, thereby reducing the number of evaluations required to reach convergence and improving the efficiency of computing resource utilization.

[0027] 3. This application establishes semantic connections between heterogeneous data by quantifying the correlation strength between nodes using a graph attention network algorithm. It reveals the dominant influencing parameters by searching the directed path from the fault node to the design parameter node and accumulating edge weights to calculate the path score. The identified target parameters are automatically passed to subsequent workflows as constraints, thereby improving the automation and accuracy of extracting design knowledge from historical fault data. Attached Figure Description

[0028] Figure 1 This is a flowchart illustrating the intelligent agent development and product optimization method in an embodiment of this application;

[0029] Figure 2 This is another flowchart illustrating the intelligent agent development and product optimization method in the embodiments of this application;

[0030] Figure 3 This is a schematic diagram of the physical device structure of an intelligent agent development and product optimization equipment in the embodiments of this application. Detailed Implementation

[0031] The terminology used in the following embodiments of this application is for the purpose of describing particular embodiments only and is not intended to be limiting of this application. As used in the specification of this application, the singular expressions “a,” “an,” “the,” “the,” and “this” are intended to include the plural expressions as well, unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this application refers to any or all possible combinations including one or more of the listed items.

[0032] Hereinafter, the terms "first" and "second" are used for descriptive purposes only and should not be construed as implying or suggesting relative importance or implicitly indicating the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature, and in the description of the embodiments of this application, unless otherwise stated, "multiple" means two or more.

[0033] The following describes the process of the method provided in this implementation. Please refer to [link / reference]. Figure 1 This is a flowchart illustrating the intelligent agent development and product optimization method in an embodiment of this application.

[0034] S101. Create an intelligent agent workflow and call the product design intelligent agent encapsulated with industry knowledge graph from the intelligent agent repository.

[0035] Among them, the intelligent agent repository refers to a centralized resource management platform that stores pre-built intelligent agent modules, providing intelligent agent invocation services based on function classification, version management, and access control. An industry knowledge graph refers to a domain-specific professional knowledge network structure stored in a graph database, organizing industry knowledge such as material properties, process parameters, and design constraints by representing knowledge entities through nodes and relationships between entities through edges. A product design intelligent agent refers to an autonomous decision-making software entity that encapsulates domain-specific professional knowledge and reasoning capabilities in product design, capable of automatically retrieving knowledge graphs, generating candidate solutions, and outputting design parameters based on input functional requirements and performance constraints.

[0036] This step is triggered when a user initiates a product development task or the system receives a product design request. Specifically, the user drags and drops agent nodes onto a canvas using the graphical interface of the visual orchestration engine, defining the data flow and execution order between nodes to form a directed acyclic graph (DAG) structure. The system sends a call request to the agent repository, which locates the corresponding product design agent based on the identifier and loads the agent's execution code, configuration parameters, and associated industry knowledge graph access interface into the workflow runtime environment. During the loading process, the product design agent initializes its connection with the industry knowledge graph, establishing a communication channel from the agent's internal inference engine to the graph database query interface.

[0037] S102. The output results of general intelligent agents in the software domain are mapped to the input parameters of product design intelligent agents through the inter-agent communication protocol. The message format defined by the inter-agent communication protocol includes knowledge graph reference fields.

[0038] Among them, the inter-agent communication protocol refers to the standardized interface specification that defines the message exchange format, transmission rules, and semantic mapping mechanism between different intelligent agents, including message headers, message bodies, and knowledge graph reference fields. A general-purpose intelligent agent in the software domain refers to an intelligent agent module that is not bound to specific industry knowledge and can handle general software development tasks such as requirements analysis and architecture design. Knowledge graph reference fields are data fields in the message format specifically used to point to specific knowledge entities in the industry knowledge graph, including knowledge entry identifiers and query condition expressions.

[0039] When the general-purpose agent completes the software requirements analysis task and generates output results, a data mapping process between cross-domain agents is triggered. Specifically, the general-purpose agent constructs a message data packet according to the inter-agent communication protocol specification, fills in the agent identifiers of the sender and receiver in the message header, encapsulates the output result data in the message body, and fills in the type identifier and attribute constraints of the knowledge entries required by the product design domain in the knowledge graph reference field. After receiving the message, the product design agent parses the query condition expression in the knowledge graph reference field, performs a graph traversal query in the internally encapsulated industry knowledge graph, and locates the target knowledge entries that meet the constraints. The system associates and binds the output results in the message body with the queried knowledge entries, converts the software requirement description into product functional requirement parameters through semantic mapping rules, maps performance constraints into product performance indicator constraints, and generates structured input parameters that the product design agent can directly use.

[0040] S103. Input the input parameters into the product design intelligent agent to generate an initial design scheme.

[0041] The input parameters refer to the set of structured data received by the product design agent to initiate the design reasoning process, including the product's functional goals, performance constraints, and domain knowledge associations.

[0042] After receiving the mapped input parameters, the product design agent triggers the initial design scheme generation process. Specifically, the product design agent parses the performance constraints in the input parameters, searches for candidate material nodes in the internal industry knowledge graph, selects a set of materials whose physical properties meet the lower limit of the constraints, and identifies compatible process nodes by traversing the constraint relationship edges of the material nodes, forming multiple material-process combination candidate schemes. The agent scores and ranks the candidate schemes according to a preset optimization objective function, which comprehensively considers material cost and process complexity, and selects a preset number of combinations with the best scores as the design basis. The agent determines the geometric topology of the product based on the product's functional requirements, assigns corresponding material attribute parameters to each candidate material-process combination, and generates an initial design scheme set containing complete geometric information, material numbers, and process identifiers, providing a feasible design space starting point for subsequent parameter optimization.

[0043] S104. Using a proxy model, search the design parameter space of the initial design scheme to obtain the preliminary optimized product design parameters.

[0044] Among them, the surrogate model refers to an approximate mapping model from design parameters to performance response constructed using machine learning algorithms such as Gaussian process regression, used to replace computationally expensive precise physical simulation. The design parameter space refers to a multidimensional continuous space composed of geometric parameters in the initial design scheme, where each point represents a specific set of design parameter values.

[0045] After the initial design scheme is generated but before the complete physical simulation model is built, a parameter optimization process based on a surrogate model is triggered. Specifically, the system uses surrogate modeling methods such as Gaussian process regression to generate initial sample points in the design parameter space through sampling methods. A simplified physical model is called to evaluate the performance of the sample points, and the surrogate model is trained based on the sample points and their performance response values. The surrogate model establishes a probabilistic mapping relationship between design parameters and performance response, and can predict the mean performance and uncertainty of unsampled points. The system uses optimization criteria to determine the next point to be evaluated in the parameter space. After evaluating the performance of this point, it is added to the sample set and the surrogate model is updated. Through an iterative process, the system gradually approaches the optimal parameter region. The search terminates when the performance improvement converges or reaches the preset number of iterations, and the design parameters corresponding to the optimal performance obtained during the iteration process are output as the preliminary optimization results.

[0046] S105. Construct the corresponding physical simulation model based on the preliminary optimized product design parameters.

[0047] Among them, the physical simulation model refers to a numerical calculation model that can predict the physical performance of a product, based on the finite element method or computational fluid dynamics method, and includes geometric mesh, material properties, boundary conditions and governing equations.

[0048] After the proxy model completes parameter optimization and outputs preliminary optimized parameters, the physical simulation model construction process is triggered. Specifically, based on the geometric dimensions in the preliminary optimized product design parameters, the system generates a 3D geometric model of the product in a computer-aided design environment. The geometric model is then meshed to generate a finite element mesh, with mesh nodes and elements forming the discretized basis for the simulation calculation. Based on the material number in the design parameters, the system extracts the corresponding physical property parameters of the material from the material database, including elastic modulus, Poisson's ratio, thermal conductivity, and density, and assigns these material properties to the corresponding regions or parts in the geometric model. The system sets the simulation boundary conditions and loads based on the product's expected operating conditions, including fixed constraints, applied loads, and initial temperature fields. It selects an appropriate physics solver type, such as a structural mechanics solver, a heat conduction solver, or a fluid dynamics solver, completing the conversion from optimized parameters to an executable simulation model.

[0049] S106. Use physical sensor data as simulation boundary conditions and input it into the physical simulation model to obtain simulation prediction data. The physical sensor data is the data collected by physical sensors deployed in the product prototype or product operating environment. The physical sensor data includes at least one of temperature, pressure and vibration parameters.

[0050] Simulation boundary conditions refer to the constraints or loads applied to the model boundaries or specific locations when the physical simulation model solves the governing equations. Simulation prediction data refers to the numerical values ​​of the physical field distribution output by the physical simulation model after solving the governing equations at the corresponding sensor locations.

[0051] After the physical simulation model is built, when the system receives real-time or historical measurement data from physical sensors, it triggers a simulation calculation process based on the actual measurement data. Specifically, the system receives temperature, pressure, and vibration parameters collected by the sensors, performs preprocessing such as filtering, noise reduction, and outlier detection on the raw data, and maps the preprocessed measurement data to nodes or elements at corresponding locations in the simulation model mesh according to the spatial deployment location of the sensors in the product prototype. The system sets the mapped temperature data as the temperature boundary condition for thermal simulation, converts the pressure data into the load boundary condition for structural simulation, and processes the vibration data as the excitation input for dynamic simulation. This drives the physical simulation model to perform solution calculations based on the actual boundary conditions, obtaining the distribution results of physical fields such as temperature and stress fields, and extracting the node physical quantity values ​​corresponding to the sensor locations in the simulation model as simulation prediction data.

[0052] S107. The error between the physical sensor data and the simulation prediction data is calculated using a neural network algorithm based on physical information.

[0053] Among them, the neural network algorithm based on physical information refers to a deep learning algorithm that embeds physical law constraints into the loss function of the neural network, which can simultaneously fit the data and satisfy the physical equations.

[0054] After the physical simulation model completes its calculations and outputs simulation prediction data, the system triggers an error calculation and evaluation process. Specifically, the system pairs the physical quantity values ​​measured by physical sensors at specific locations with the predicted values ​​output by the simulation model at the corresponding location nodes, and uses a neural network algorithm based on physical information to construct an error evaluation model. This neural network includes not only data fitting terms such as the mean square error between sensor data and simulation data in its loss function, but also embeds physical control equations as regularization constraints, ensuring that the error calculation process considers both data consistency and physical consistency. The neural network learns the systematic deviation patterns between sensor data and simulation data through training, outputting error values ​​for each measurement point. The system performs statistical analysis on the errors at multiple measurement points, calculating indicators such as average error and maximum error to quantify the current prediction accuracy of the simulation model.

[0055] S108. If the error exceeds the preset threshold, update the material physical parameters using a loss function that includes data fitting terms and physical constraint terms, and then send the updated material physical parameters back to the physical simulation model for resimulation.

[0056] Here, the preset threshold refers to the upper limit of error used to determine whether the accuracy of the simulation model's predictions is acceptable. The data fitting term refers to the term in the loss function used to minimize the difference between the simulation results and the sensor measurements. The physical constraint term refers to the term in the loss function used to ensure that the updated material parameters satisfy physical laws.

[0057] When the error calculation result exceeds a preset threshold, the system determines that the material parameters of the current simulation model deviate significantly from the actual material properties, triggering a model parameter correction process. Specifically, the system constructs a loss function that includes data fitting terms and physical constraint terms. The data fitting term calculates the sum of squared residuals between the simulation prediction data and the sensor measurement data, while the physical constraint term embeds physical equations such as the material constitutive relation equation and conservation laws to ensure that the material behavior still conforms to basic physical laws during parameter updates. The system uses optimization algorithms such as gradient descent to minimize this loss function, calculates the gradient of the loss function with respect to the material parameters through backpropagation, and iteratively updates the values ​​of material physical parameters such as elastic modulus and Poisson's ratio along the gradient direction. The updated material parameters are fed back to the physical simulation model, replacing the original standard material parameters. The system re-executes the simulation calculation to obtain new simulation prediction data based on the corrected parameters, repeating the error calculation and parameter update process until the error is reduced below the threshold.

[0058] S109. If the error does not exceed the preset threshold, determine whether the physical performance indicators obtained from the simulation calculation meet the design requirements. If not, readjust the product design parameters and rebuild the physical simulation model.

[0059] Physical performance indicators refer to quantitative metrics that characterize the physical properties of a product, obtained through simulation calculations, including maximum stress, maximum deformation, and natural frequency. Design requirements refer to the performance objectives and constraints that the product needs to meet, typically specified when inputting design tasks.

[0060] When the error calculation result does not exceed the preset threshold, the system determines that the simulation model has reliable predictive capabilities and triggers the performance index verification process. Specifically, the system extracts key performance index values ​​from the physical field distribution results obtained from the simulation calculation, such as the maximum equivalent stress from the stress field, the maximum deformation from the displacement field, and the highest temperature value from the temperature field. The extracted performance indexes are compared with the allowable ranges specified in the design requirements. If the performance indexes exceed the allowable ranges, it indicates that the current design parameter configuration cannot meet the product performance requirements. The system triggers the parameter adjustment process, determines the parameter adjustment strategy based on the direction of the performance index exceeding the limit, such as increasing the wall thickness or selecting a higher strength material if the stress is too high. The system modifies the design parameter values ​​according to the adjustment strategy, reconstructs the physical simulation model based on the new design parameters, and returns to execute the sensor data input and simulation calculation steps, forming an iterative closed loop of design-simulation-verification.

[0061] S110. If satisfied, output the final product design scheme.

[0062] The final product design scheme refers to the product design result that has undergone iterative optimization through design, simulation, and verification and has passed physical performance verification, including geometric dimensions, material selection, process planning, and predicted performance indicators.

[0063] When all physical performance indicators obtained from simulation calculations meet the design requirements, the system determines that the current design parameter configuration has reached an acceptable state and triggers the final solution output process. Specifically, the system summarizes complete information on the current design parameters, including the optimized and adjusted geometric dimensions, material number, process type identifier, and material physical parameters corrected by sensor data. It extracts the key performance indicator values ​​calculated by the simulation model as performance prediction results. The system organizes the above design information and performance data into a structured product design document. This document includes a 3D geometric model file, material and process specifications, performance verification reports, and simulation analysis results. Through the output interface of the visual orchestration engine, the final product design solution is transmitted to the downstream manufacturing stage or user review stage, completing the entire design process from requirement input to solution output.

[0064] The above embodiments describe the overall process framework of the agent development and product optimization method. In order to more clearly demonstrate the specific implementation mechanism and technical details of each step, the following describes the core technical links such as agent-to-agent communication protocol mapping, knowledge graph retrieval, agent model optimization, and multi-physics coupling simulation in a more detailed manner, taking the specific application scenario of turbine blade design as an example.

[0065] The following provides a more detailed description of the process of the method provided in this implementation. Please refer to [link / reference]. Figure 2 This is another flowchart illustrating the intelligent agent development and product optimization method in this application embodiment.

[0066] S201. Construct the message body of a general intelligent agent according to the inter-agent communication protocol, fill in the type identifier and attribute constraints of the target knowledge entry in the knowledge graph reference field, and fill the output results into the message body.

[0067] The message body refers to the structured data fields in the communication protocol used to carry actual business data, including task type, performance index values, and business logic parameters. The type identifier of the target knowledge boundary is the label of the knowledge entity category in the knowledge graph, such as ontology system classifications like materials, processes, and equipment. Attribute constraints refer to the filtering of conditions applied to the attribute values ​​of the target knowledge type.

[0068] This step achieves semantic interoperability across agents through standardized message bodies. After the general agent completes task processing, it needs to pass the output to downstream agents. Knowledge graph references are used as a semantic bridging mechanism to map general performance to specialized domain knowledge.

[0069] For example, after the demand analysis agent processes the design of the cooling blade channel, it outputs that it needs to withstand a maximum operating temperature of 1200℃ and an oxidizing environmental pressure of 2.5MPa. The knowledge graph reference field then fills in the type identifier material: high-temperature alloy, and the attribute constraint "WHERE m.melting_point>=1400 AND m.oxidation_resistance_temp>=1200". This design decouples the responsibilities of semantic understanding: the general agent is responsible for identifying the task type and providing type identifier prompts, while the specialized agent is responsible for parsing specific knowledge entities in the knowledge graph.

[0070] The message body construction includes three key steps: first, the output results are formed into performance contract parameters; second, the required knowledge type is inferred based on domain knowledge; and third, the performance contract is transformed into knowledge graph query syntax.

[0071] S202. After receiving the message body, the product design intelligent agent locates the target knowledge entry in the industry knowledge graph based on the knowledge graph reference fields.

[0072] This step leverages the ontology hierarchy of the knowledge graph to achieve efficient retrieval. The type identifier provided by the knowledge graph's reference fields narrows the search scope from the entire graph to specific category subgraphs, while attribute constraints further refine the location within those subgraphs. This two-level filtering strategy fully utilizes the indexing mechanism of the graph database.

[0073] After receiving the message body, the product design agent parses the knowledge graph reference field {entity_type: Material: HighTemperatureAlloy, constraint: WHERE m.melting_point>=1400 AND m.oxidation_resistance_temp>=1200}. The first-level filtering performs MATCH(m: Material: HighTemperatureAlloy) to locate a subset of 12 nodes labeled as high-temperature alloys in the ontology level. This step utilizes a hash index built on the node labels in the graph database, eliminating the need to traverse all 3820 material nodes. The second-level filtering performs attribute filtering on the 12 candidate nodes, reading the melting_point attribute value of each node and quickly comparing values ​​using a pre-built B-tree index on the attribute field. For example, Inconel718, with a melting point of 1336℃, was filtered out because it did not meet the requirement of >=1400℃. Ultimately, it was located to two target knowledge items: CMSX-4 (melting point 1477℃, oxidation resistance temperature 1250℃) and PWA1484 (melting point 1485℃, oxidation resistance temperature 1280℃).

[0074] The advantage of this hierarchical retrieval is that it reduces the query complexity from O(N) to O(N_subset). When the size of the subgraph is much smaller than the full graph, the performance improvement is significant, reducing the query time from 450ms for a full table scan to 85ms for type filtering plus 35ms for attribute filtering.

[0075] S203. Associate the output results in the message body with the target knowledge items to generate the input parameters of the product design intelligent agent.

[0076] This step binds requirements and knowledge through an explicit association structure, making the source of constraints traceable in subsequent design optimization processes. When performance constraints change, the system directly determines material suitability through the association structure without re-executing the knowledge graph query. The agent establishes a correspondence between performance constraints in the message body and the located target knowledge entries. For example, the maximum operating temperature of 1200℃ is associated with the attribute "oxidation_resistance_temp: 1250℃" of the CMSX-4 node, generating {{Design value: 1200, unit: ℃, corresponding material attribute: oxidation_resistance_temp, material node reference: CMSX-4, attribute value: 1250}} in the input parameters. This explicit association structure records three layers of information: design requirement value, corresponding material attribute name, and actual attribute value of the material node.

[0077] The core value of this design lies in its rapid response to constraint changes. If the operating temperature needs to be adjusted to 1250℃, the system directly reads the "attribute value: 1250" in the associated structure to determine that CMSX-4 has reached its upper limit for oxidation resistance temperature, and needs to switch to PWA1484 (oxidation resistance temperature 1280℃), avoiding re-execution of the knowledge graph query and reducing the material suitability judgment from 120ms to 5ms. Simultaneously, the agent associates the functional requirement of blade cooling with the internal cooling channel design feature nodes in the knowledge graph, providing knowledge support for subsequent geometric topology generation. The final aggregated input parameters include four performance constraint fields, two candidate material node references, and one functional requirement field.

[0078] S204. Send the input parameters transmitted by the inter-agent communication protocol to the product design agent. The input parameters include product functional requirements and performance constraints.

[0079] This step uses message queues to achieve asynchronous decoupling between agents, improving the workflow's concurrent processing capabilities. In synchronous calls, the requirements analysis agent must wait for a response from the product design agent; if the latter is temporarily busy performing complex simulation calculations, the former will be blocked.

[0080] The system serializes input parameters into JSON message packets according to the inter-agent communication protocol and sends them to the input queue "queue.product_design.input" of the product design agent via a RabbitMQ message queue. The message packet contains metadata such as sender identifier, receiver identifier, message type, and timestamp, as well as all business data related to the input parameters. Message sending employs a persistent model, storing a copy on the RabbitMQ server's disk to prevent message loss.

[0081] After the product design agent's message listener detects a new message, it pulls it from the queue and deserializes it, verifying the completeness of required fields, including checking if `product_function` exists, if the `performance_constraints` array contains at least one constraint, and if the `material_references` array is not empty. Once verification is successful, the input parameters are loaded into the design inference engine, which extracts product functional requirements to implement the blade cooling function. The design pattern is identified as internal flow channel cooling, and performance constraints are extracted and parsed into a standardized format such as `{parameter name: max_operating_temp, value: 1200, unit: ℃, constraint type: upper limit}`.

[0082] The advantage of asynchronous mode is that the demand analysis agent returns immediately after sending a message to continue processing the next task, and the product design agent pulls messages from the queue for processing when its computing resources are available. Messages can be cached in the queue for up to 24 hours, which increases the concurrent processing capability of the entire workflow from 1 task / minute in serial mode to 15 tasks / minute.

[0083] S205. Based on the input parameters, retrieve candidate material nodes that meet the performance index constraints in the industry knowledge graph. The yield strength and elastic modulus of the candidate material nodes both meet the lower limit requirements of the performance index constraints.

[0084] In this context, candidate material nodes refer to material entity nodes in the knowledge graph that satisfy all performance index constraints. The lower limit requirement refers to the minimum allowable value that the material properties specified in the performance index constraints must reach.

[0085] This step employs a multi-level filtering strategy to progressively reduce the size of the candidate set, leveraging the indexing mechanism of the graph database to improve query performance. Different performance metrics exhibit significant differences in selectivity; therefore, executing high-selectivity constraints first minimizes the amount of data required for subsequent filtering.

[0086] For example, a product design agent sends a query request to the aerospace materials knowledge graph, executing the Cypher statement "MATCH(m: Material)WHERE m.yield_strength>=900ANDm.elastic_modulus>=180ANDm.density<=9.0ANDm.oxidation_resistance_temp>=1200RETURNm". The graph database uses a four-level filtering: the first level, yield strength filtering, reduces the number of nodes from 3820 to 526, filtering out 86.2% of the nodes that do not meet the requirements, because the yield strength of most aluminum and titanium alloys in aerospace materials is below 900 MPa. The second level, elastic modulus filtering, reduces the number of nodes from 526 to 358; the third level, density filtering, reduces the number of nodes from 358 to 156; and the fourth level, oxidation resistance temperature filtering, reduces the number of nodes from 156 to 3.

[0087] For example, CMSX-4 (yield strength 980 MPa, elastic modulus 130 GPa, density 8.7 g / cm³, oxidation resistance temperature 1250℃) met all constraints and was retained. Inconel718, while meeting the yield strength (1034 MPa), elastic modulus (214 GPa), and density (8.19 g / cm³), was filtered out because its oxidation resistance temperature of 980℃ did not meet the requirement of >=1200℃. Ultimately, CMSX-4, PWA1484, and René N5 were selected as the three candidate material nodes.

[0088] S206. Traverse the constraint edges of candidate material nodes, filter out process nodes that are compatible with the candidate materials, and form a set of material-process combination schemes.

[0089] Here, constraint edges refer to the edges in the knowledge graph that connect material nodes and process nodes, representing the compatibility or exclusion constraints between materials and processes. The material-process combination set refers to the set of paired solutions consisting of candidate material nodes and their compatible process nodes.

[0090] This step improves retrieval efficiency by pre-storing the constraint relationships between materials and processes in the knowledge graph, transforming compatibility judgments from rule engine calls to graph traversal operations. Material and process compatibility involves complex physicochemical constraints; explicitly encoding this knowledge into a graph structure enables rapid access.

[0091] The agent performs a graph traversal for each candidate material node, iterating along all relational edges from the node along the outgoing edges, filtering for edges of type COMPATIBLE_WITH and extracting the target process node. For example, the CMSX-4 node is connected to 5 outgoing edges, of which 3 COMPATIBLE_WITH edges point to directional solidification casting, single crystal casting, and precision casting process nodes respectively, 1 INCOMPATIBLE_WITH edge points to a forging process node, and 1 REQUIRES_SPECIAL_TREATMENT edge points to a vacuum heat treatment process node. The agent reads the attributes of the compatible process nodes, including process type, pouring temperature, cooling rate, equipment cost, processing cycle, etc., forming three combination schemes: {CMSX-4, directional solidification casting}, {CMSX-4, single crystal casting}, and {CMSX-4, precision casting}.

[0092] After performing the same operation on nodes PWA1484 and RenéN5, the aggregation yields a material-process combination scheme set containing 8 elements. Detailed compatibility constraints can be annotated on the design edges. For example, the COMPATIBLE_WITH edge between CMSX-4 and directional solidification casting stores {temperature range: 1480-1520℃, lower limit of cooling rate: 8℃ / min}. The agent can obtain the effective range of process parameters by reading the edge attributes, reducing the material-process compatibility judgment from 80ms of calling an external rule engine to 12ms of graph traversal.

[0093] S207. Sort the material-process combination scheme set according to the preset optimization objective function. The optimization objective function takes into account material cost parameters and process complexity, and selects the optimal combination scheme of the preset number of objective function values ​​as candidate design schemes.

[0094] Among them, the optimization objective function refers to a mathematical function that quantitatively evaluates the material-process combination scheme, taking into account multiple dimensions such as cost, performance, and manufacturability.

[0095] This step balances the trade-offs between different evaluation dimensions using a multi-objective optimization framework. The agent constructs an indicator system comprising five evaluation dimensions: material cost (weight 0.25), process cost (weight 0.20), performance margin (weight 0.30), manufacturability (weight 0.15), and delivery cycle (weight 0.10). This weighting reflects the aerospace industry's emphasis on performance reliability. The highest weight for performance margin reflects the safety-first design philosophy, while the relatively low weight for delivery cycle is because aerospace components prioritize quality over speed.

[0096] When calculating the scores for each combination scheme, normalization is used to eliminate the dimensional differences between different indicators. For example, for the {CMSX-4, single crystal casting} scheme: the material cost score is calculated using the formula score_cost=1-(cost_actual-cost_min) / (cost_max-cost_min), where the CMSX-4 price of 850 yuan / kg corresponds to a score of 0.35 relative to the price range [420, 1280] yuan / kg; the process cost score, with an investment of 12 million yuan in single crystal casting equipment, corresponds to a score of 0.42; and the performance margin score, with a 4.2% margin between the CMSX-4 oxidation resistance temperature of 1250℃ and the required 1200℃, corresponds to a score of 0.68. The overall score = 0.25×0.35+0.20×0.42+0.30×0.68+0.15×0.78+0.10×0.62=0.548.

[0097] After scoring and ranking the eight combined schemes, the agent did not simply select the highest-scoring scheme, but further verified the performance margin. The top-ranked scheme, {René N5, directional solidification casting}, scored 0.612, but its material density of 8.95 g / cm³ was close to the constraint limit of 9.0 g / cm³, with a margin of only 0.56%, posing a risk of exceeding the limit. Therefore, the second-ranked scheme, {PWA1484, single-crystal casting}, with a score of 0.587, was selected. This scheme achieved a performance margin of 12.3% and balanced all indicators, reflecting the principle of prioritizing robustness in engineering practice.

[0098] S208. Determine the geometric topology of the product structure based on the product functional requirements, assign corresponding material properties to each candidate design scheme, and generate an initial design scheme including geometric dimensions, material number and process type identifier.

[0099] Among them, Gaussian process regression model refers to a nonparametric regression method based on Gaussian process, which uses a coordination function to characterize the correlation of the input space and outputs predicted values ​​and their uncertainties. Latin hypercube sampling is a sampling method that generates uniformly distributed sample points in a multidimensional parameter space, ensuring uniform marginal distribution at each level.

[0100] This step employs parametric modeling technology to enable rapid response to design changes, and generates initial sample points covering the design space through Latin hypercube sampling, providing a data foundation for subsequent surrogate model training.

[0101] The agent invokes the CAD kernel API to initiate the modeling process and create the basic geometric framework of the blade: the blade body uses NURBS curves to define the aerodynamic shape, the blade root uses a tenon-tooth connection structure, and the blade crown uses a peripheral band structure. Crucially, these dimensions are all defined as variable parameters rather than fixed values.

[0102] Based on product functional requirements, the blade cooling function is implemented, and the intelligent agent considers the geometric features of the cooling channels inside the blade. The internal cooling channel design pattern node is searched from the knowledge graph, and recommended parameters are read: number of channels 5, diameter dε[2,3] mm, wall thickness t≥1.5 mm, and channel width≥4 mm. The intelligent agent creates a 5-channel heat sink, built along the blade height direction.

[0103] Geometric constraints are established to ensure design effectiveness: the constraint between the cooling channel diameter and the thinnest part of the blade thickness is d≤0.4×t_min, the constraint between the channel thickness and the diameter is s≥1.5×d, and the constraint between the channel wall thickness and pressure is t_wall≥P×d / (2×σ_allow). These constraint parameters are defined using constraint definitions, and the CAD system automatically verifies the constraints and updates the geometry when the parameters are adjusted.

[0104] The design parameter space is defined as 4-dimensional: {channel diameter dε[2.0, 3.0] mm, channel distance ε[4.0, 5.5] mm, leaf root tooth height hε[6, 10] mm, leaf crown thickness tε[2.5, 3.5] mm}. Latin hypercube sampling is used to generate 50 initial sample points in this space, ensuring that each partition and each interval is sampled only once, thus making the sample points more evenly distributed in space.

[0105] S209. Using a Gaussian process regression model as a surrogate model, an initial sample point set is generated in the design parameter space composed of the geometric dimensions of the initial design scheme through the Latin hypercube sampling method.

[0106] Among them, Gaussian process regression model refers to a nonparametric regression method based on Gaussian process, which uses a coordination function to characterize the correlation of the input space and outputs predicted values ​​and their uncertainties. Latin hypercube sampling is a sampling method that generates uniformly distributed sample points in a multidimensional parameter space, ensuring uniform marginal distribution at each level.

[0107] This step employs parametric modeling technology to enable rapid response to design changes, and generates initial sample points covering the design space through Latin hypercube sampling, providing a data foundation for subsequent surrogate model training.

[0108] The agent invokes the CAD kernel API to initiate the modeling process and create the basic geometric framework of the blade: the blade body uses NURBS curves to define the aerodynamic shape, the blade root uses a tenon-tooth connection structure, and the blade crown uses a peripheral band structure. Crucially, these dimensions are all defined as variable parameters rather than fixed values.

[0109] Based on product functional requirements, the blade cooling function is implemented, and the intelligent agent considers the geometric features of the cooling channels inside the blade. The internal cooling channel design pattern node is searched from the knowledge graph, and recommended parameters are read: number of channels 5, diameter dε[2,3] mm, wall thickness t≥1.5 mm, and channel width≥4 mm. The intelligent agent creates a 5-channel heat sink, built along the blade height direction.

[0110] Geometric constraints are established to ensure design effectiveness: the constraint between the cooling channel diameter and the thinnest part of the blade thickness is d≤0.4×t_min, the constraint between the channel thickness and the diameter is s≥1.5×d, and the constraint between the channel wall thickness and pressure is t_wall≥P×d / (2×σ_allow). These constraint parameters are defined using constraint definitions, and the CAD system automatically verifies the constraints and updates the geometry when the parameters are adjusted.

[0111] The design parameter space is defined as 4-dimensional: {channel diameter dε[2.0, 3.0] mm, channel distance ε[4.0, 5.5] mm, leaf root tooth height hε[6, 10] mm, leaf crown thickness tε[2.5, 3.5] mm}. Latin hypercube sampling is used to generate 50 initial sample points in this space, ensuring that each partition and each interval is sampled only once, thus making the sample points more evenly distributed in space.

[0112] S210. For each sample point in the initial sample point set, call the simplified physical model to perform a fast performance evaluation and obtain the corresponding performance response value.

[0113] Simplified physical model refers to a computational model that has been reasonably simplified from the complete model, thereby rapidly reducing computational costs while ensuring accuracy. Performance response value refers to the calculated performance index of the design scheme under specific operating conditions, such as maximum selection, highest physical temperature, and maximum priority.

[0114] This step verifies the performance of the design scheme under real-world conditions through multiphysics coupling simulation. The turbine blade's operating environment involves high temperature, high pressure, and high failure rate complexity, requiring refined simulation analysis to consider coupling effects and temperature-related material nonlinearities.

[0115] The intelligent agent imports the 3D geometric model into the finite element preprocessor to perform mesh generation. The blade body adopts a hexahedral dominant mesh with an element size of 0.8mm, totaling 350,000 elements; the cooling channel wall adopts a boundary layer mesh with a first layer thickness of 0.05mm and a middle layer thickness of 1.2mm to capture the instantaneous temperature resonance change area; the blade root tenon contact surface is refined to 0.3mm.

[0116] When defining material properties, the PWA1484 parameters extracted from S208 were assigned to the leaf body mesh: elastic modulus 130 GPa, Poisson's ratio 0.31, and density 8.7 g / cm3. Yield strength and strength coefficient were represented by linear interpolation using a support structure to account for temperature dependence, making the simulation more closely resemble real material behavior.

[0117] Boundary conditions and contaminants are defined as follows: fixed constraints are applied to the blade root tenon teeth, a gas pressure barrier (distributed pressure field imported from CFD analysis) is applied to the blade surface, the gas temperature is 1450℃ as a heat inhibitor, a convective heat transfer coolant temperature of 800℃ and a convection rate of 5000W / (m²·K) is applied to the inner surface of the cooling channel, and the mesh is rotated at 12000rpm to generate a centrifugal mesh.

[0118] Activation simulation was performed. First, short-circuit thermal analysis revealed the temperature field distribution, with the highest temperature at the blade leading edge reaching 1238℃, and approaching 950℃ near the cooling channel. The temperature field was then preloaded into the structural analysis. The activated field showed a maximum Von Mises strength of 879 MPa at the blade root tenon root, both lower than the bending strength of PWA1484 at the temperature, meeting synchronous strength requirements and with a safety factor of approximately 1.05. The initial seismic field showed a maximum northeastward deviation of 2.87 mm at the blade tip, within the design margin.

[0119] S211. Train a Gaussian process regression model based on the initial sample point set and the corresponding performance response values.

[0120] This step utilizes Gaussian process regression to construct a probabilistic mapping model from design parameters to performance response. Compared to ordinary regression models that only output predicted values, Gaussian process regression also outputs prediction uncertainty, providing confidence information for subsequent optimization strategies. In sparse sample regions, the prediction uncertainty is high, guiding the algorithm to increase sampling density in those regions.

[0121] The agent organizes the input (4-dimensional design parameters) and output (maximum stress value) of 50 initial sample points into training datasets X_train and y_train. The squared exponential covariance function k(x, x') = σ²exp(-||x-x'||² / (2l²)) is chosen, where σ² is the signal variance and l is the length scale parameter. This function assumes that similar design parameters correspond to similar performance responses.

[0122] Hyperparameter optimization was performed, and σ² = 125 MPa² and l = [0.3, 0.4, 0.5, 0.35] mm were determined by maximizing the marginal likelihood function (each design parameter dimension corresponds to a different length scale). The trained model can predict the mean μ(x) and standard deviation σ(x*) of the performance response for any design parameter x. For example, for the parameter point {d = 2.3 mm, s = 4.5 mm, h = 7 mm, t = 3 mm}, the predicted maximum stress mean is 815 MPa and the standard deviation is 38 MPa, indicating that the actual stress at this point has a 95% probability of falling within the interval [739, 891] MPa.

[0123] Model validation employed leave-one-out cross-validation, where one sample point was left for testing each time, with the remaining 49 samples used for training. The mean absolute error (MAE) after 50 tests was calculated to be 22 MPa, and the coefficient of determination (R²) was 0.93, indicating that the model possesses good generalization ability. Compared to directly calling finite element analysis, which takes 45 minutes each time, the surrogate model prediction takes only 0.02 seconds, significantly improving performance evaluation speed and enabling large-scale optimization searches.

[0124] S212. The expected improvement criterion is used to determine the next point to be evaluated in the design parameter space. The expected improvement criterion is used to calculate the expected improvement of the unsampled points in the current design parameter space relative to the known optimal performance response value.

[0125] Among them, the expected improvement criterion (EI) refers to the expected improvement of candidate points relative to known optimal values ​​based on the mean and uncertainty predicted by the current surrogate model, taking into account the balance between development and exploration.

[0126] This step aims to strike a balance between developing near known optimal solutions and exploring regions of high uncertainty by improving the criteria.

[0127] The agent reads the optimal performance response value y_best = 682 MPa from the current sample point set. For unsampled points in the design parameter space, the mean μ(x) and standard deviation σ(x) are predicted using a trained Gaussian process model, and the expected improvement EI(x) = (y_best - μ(x))Φ(Z) + σ(x)φ(Z) is calculated, where Z = (y_best - μ(x)) / σ(x).

[0128] For example, candidate point A has a predicted mean μ = 695 MPa and a standard deviation σ = 45 MPa, resulting in an EI of 18.3 MPa; candidate point B has a predicted mean μ = 720 MPa and a standard deviation σ = 62 MPa, resulting in an EI of 21.5 MPa. Although the predicted mean of candidate point A is closer to the current optimum, candidate point B has higher uncertainty, thus a greater improvement is desired.

[0129] The agent uses optimization algorithms to improve EI(x) across the entire design parameter space, finding the point with the greatest expected improvement as the evaluation point. In this iteration, the evaluation point EI=23.7MPa is located in a sparse region of the current samples, ensuring the discovery of new, excellent designs.

[0130] S213. Perform performance evaluation on the points to be evaluated, add the points to be evaluated and their corresponding performance response values ​​to the sample point set, and update the Gaussian process regression model.

[0131] This step achieves continuous improvement in the accuracy of the surrogate model through iterative updates. The initial surrogate model is trained based on 50 samples, and the prediction error is relatively large in some areas. Each time a new sample point is added, it is equivalent to supplementing information in the weak parts of the model and gradually correcting the prediction bias.

[0132] The agent performs precise finite element analysis on the evaluation point {d=2.4mm, s=5.1mm, h=7.8mm, t=2.9mm}. First, the CAD model geometry is updated based on the parameter values: the cooling channel diameter is adjusted to 2.4mm, the spacing to 5.1mm, the blade root tooth height to 7.8mm, and the blade crown thickness to 2.9mm. The mesh is then re-generated with 342,000 elements (the number of elements changes slightly due to geometric fine-tuning). The same boundary conditions and loads are applied, and the solver outputs simulation results after 42 minutes: the maximum stress is 673MPa, lower than the current optimal value of 682MPa, indicating a new optimal design has been found.

[0133] The agent adds new sample points {x=[2.4, 5.1, 7.8, 2.9] mm, y=673 MPa} to the sample point set, increasing the sample size from 50 to 51. Gaussian process regression training is re-executed, and the hyperparameters are optimized to obtain updated σ²=118 MPa² and l=[0.28, 0.39, 0.52, 0.33] mm. Comparing the model predictions before and after the update, the standard deviation of the prediction in the region near the evaluation point decreases from 62 MPa to 35 MPa, indicating a significant reduction in uncertainty in this region and local model refinement.

[0134] The updated model achieved a cross-validation metric of MAE = 19 MPa (down from 22 MPa) and R² = 0.95 (up from 0.93) on 51 sample points, indicating that the model accuracy continues to improve with the increase in samples. This iterative optimization strategy will be repeated subsequently, selecting the point with the greatest expected improvement each time to evaluate and update the model, gradually approaching the globally optimal design. Compared to traditional grid search, which requires evaluating 160,000 points, Bayesian optimization typically converges within 100-200 iterations, reducing the total computational cost by three orders of magnitude.

[0135] S214. Repeat the steps of determining the point to be evaluated and updating the model until the performance response value improvement is less than the convergence threshold or the number of iterations reaches the preset maximum number of iterations in a consecutive preset number of iterations.

[0136] This step utilizes an iterative optimization framework to automatically optimize design parameters. Bayesian optimization uses a surrogate model to guide the search direction, sampling at the location where the expected improvement is greatest in each iteration.

[0137] The agent enters an iterative loop. During the second iteration of training, the sample point set contains 51 points, and the current optimal performance response value y_best = 673 MPa. Execute S212 to find the point to be evaluated with the largest EI, and execute S213 to perform finite element analysis to obtain a performance response value of 658 MPa, with a performance improvement of Δy = 15 MPa. Add the new sample to the set to update the model.

[0138] The improvement amount Δy = 7 MPa in the 3rd iteration, 3 MPa in the 4th iteration, 1 MPa in the 5th iteration, and 0.5 MPa in the 6th iteration is less than the convergence value of 2 MPa. The agent detects that the improvement amount in three consecutive iterations is less than the convergence threshold, satisfying the convergence criterion. At this point, the total number of iterations is 6, which is much less than the initial maximum number of iterations of 100.

[0139] The entire optimization process started with 50 initial samples. Adding only 6 sample points reduced the maximum stress from 682 MPa to 646.5 MPa, improving performance by 5.2%. After terminating the iteration, the agent recorded the optimized stretching curve. The first three iterations showed significant improvement, and subsequent iterations showed convergence, validating the desired improvement based on rapid approximation of the optimal region and advanced fine-grained search.

[0140] S215. The design parameters corresponding to the optimal performance response value obtained during the output iteration process are used as the product design parameters for preliminary optimization.

[0141] This step extracts the global optimal solution from all sample points accumulated during the optimization process, completing the parameter optimization stage based on the surrogate model.

[0142] Specifically, the agent traverses a set of 56 sample points (50 initial + 6 iterations), each sample point recording the design parameter vector x and the corresponding accurate simulation performance response value y. The sample point with the smallest performance response value is found: x_opt = [2.3 mm, 5.0 mm, 8.0 mm, 3.0 mm], y_opt = 646.5 MPa. This point is evaluated in the 5th iteration.

[0143] The intelligent agent outputs the following preliminary optimized design parameters: cooling channel diameter d = 2.3 mm, which is 8% smaller than the initial design of 2.5 mm. The smaller diameter reduces the weakening of the blade structure; channel spacing s = 5.0 mm, which is 19% larger than the initial 4.2 mm. Increasing the spacing increases the wall thickness between channels and improves the structural strength; blade root tooth height h = 8.0 mm, which is the same as the initial 8 mm, indicating that this parameter has little impact on stress; blade crown thickness t = 3.0 mm, which is the same as the initial 3 mm, and the impact is also insignificant.

[0144] These optimized parameters reduced the maximum stress from 879 MPa (strain of the blade root tenon in S210) in the initial design to 646.5 MPa, a reduction of 26.4%, thus improving the structural safety margin. The stress reduction mechanism, analyzed by the intelligent agent, is as follows: increasing the channel spacing increases the channel wall thickness from 3.5 mm to 4.2 mm, reducing the stress in this area from 765 MPa to 645 MPa; decreasing the channel diameter improves the overall blade stiffness and reduces bending deformation by 12%.

[0145] The agent writes the optimization results into a design report, including: optimal parameter values, optimal performance response values, percentage improvement relative to the initial design, convergence curve of the optimization process, and parameter-performance scatter plots for all 56 sample points. The scatter plots show the distribution of sample points in the parameter space, verifying the uniformity of Latin hypercube sampling and the effectiveness of Bayesian optimization, with sample points gradually clustering towards the optimal region.

[0146] S216. Receive temperature, pressure and vibration parameters collected by physical sensors and preprocess the physical sensor data.

[0147] This step acquires real-world operating data through physical sensors, providing boundary condition inputs for the digital twin model and achieving virtual-real mapping.

[0148] Specifically, the intelligent agent reads real-time data streams from the sensors via the data acquisition system. Temperature sensors are deployed at five measuring points on the turbine blade surface, with a sampling frequency of 10Hz. The data format is timestamp-temperature pairs, for example, {t=1678345678.123s, T=1235.8℃}. Pressure sensors are deployed at the leading and trailing edges of the blade, with a sampling frequency of 50Hz. The data format is {t=1678345678.125s, P=1.73MPa}. Vibration sensors (accelerometers) are deployed at the blade root, with a sampling frequency of 1000Hz. The data format is {t=1678345678.126s, a_x=12.3m / s², a_y=8.7m / s², a_z=15.2m / s²}.

[0149] The data preprocessing process begins with time synchronization. Since different sensors have different sampling frequencies, the data needs to be aligned to a unified time grid. Linear interpolation is used to resample all sensor data to 10Hz (the lowest sampling frequency). Next, outlier detection is performed. Three outliers were identified in the temperature sensor data {T=2150℃, T=-50℃, T=1850℃}, far exceeding the physically reasonable range. These were determined to be sensor malfunctions or electromagnetic interference. The 3σ criterion was used to remove the outliers, and they were filled in using linear interpolation of the preceding and following values.

[0150] After preprocessing, the data is stored as a standardized time-series data table with fields including {timestamp, sensor_id, sensor_type, location, value, unit, quality_flag}. The quality_flag indicates the data quality (0=normal, 1=interpolation, 2=filtering, 3=abnormal), providing traceability for subsequent analysis. Finally, 60 seconds of steady-state operating condition data were obtained, with 5 temperature measurement points × 600 time steps = 3000 data points, 2 pressure measurement points × 600 = 1200 data points, and 1 vibration measurement point × 600 = 600 data points (triaxial components).

[0151] S217. Based on the deployment location of the physical sensors in the product prototype or product operating environment, map the preprocessed physical sensor data to the corresponding nodes or units in the physical simulation model.

[0152] This step establishes the geometric mapping relationship between physical space and digital space, injecting sensor measurement data into the correct position of the simulation model.

[0153] The agent reads the sensor configuration file, which records the physical coordinates of each sensor. For example, temperature sensor T1 is located at coordinates (12.3, 45.2, 8.5) mm at the leading edge of the blade root, pressure sensor P1 is located at coordinates (11.8, 45.2, 40.2) mm at the outer surface of the leading edge, and vibration sensor A1 is located at coordinates (20.0, 45.2, 0) mm at the bottom of the blade root.

[0154] The agent loads a finite element mesh model containing 350,000 nodes and executes a nearest neighbor search algorithm to find the nearest node for each coordinate of the sensor as the mapping target. Temperature sensor T1 is mapped to Node_15823, and the Euclidean distance of 0.06mm is less than the mesh size of 0.8mm, so the mapping is effective.

[0155] For sensors with excessively large distances between nearest nodes, multi-node weighted interpolation mapping is used. Pressure sensor P1 finds the tetrahedral elements of the surrounding four node structures, calculates the centroid coordinates (λ1=0.42, λ2=0.31, λ3=0.18, λ4=0.09), and distributes the sensor data to the four nodes in a weighted manner according to the centroid coordinates.

[0156] The velocity measured by the vibration sensor needs to be mapped to the nodal degrees of freedom. The agent converts the velocity data into an equivalent inertial force F = m × a applied to the leaf root node, where m is the element mass, and simulates the bearing response through dynamic simulation.

[0157] After mapping, the sensor data at each time step is associated with the nodes of the simulation model: temperature data T1(t) → Node_15823.Temperature(t), pressure data P1(t) is weighted and distributed to 4 nodes, and vibration data A1(t) → Node_1.Force(t).

[0158] S218. Set the mapped physical sensor data as the boundary conditions of the physical simulation model. The boundary conditions include at least one of temperature boundary, pressure boundary and displacement boundary.

[0159] This step uses graph neural networks to automatically discover the root causes of failures and solidify the experience into system constraints. The system writes agent development data, product implementation design data, physical simulation data, and product operation data into a data platform to build a data relationship graph, creates nodes for each type of data, and creates related edges through business logic, forming a complete traceability from design parameters to failure records.

[0160] A graph theory network algorithm is used to learn node embeddings in the data association graph. Node features are aggregated and the cosine similarity between nodes is calculated as edge weights. The data includes 1580 design schemes, 4200 simulations, and 860 failure cases from two years of historical data. The model's AUC for predicting associations reaches 0.89.

[0161] The system searches backward from the fault node to the design parameter node to locate the root cause. Starting from the fault, it traverses along the incoming edges and calculates the path score by multiplying the edge weights. After identifying 453 paths, the top 10 with the highest scores are identified as key causal paths, including blade wall thickness (0.38), cooling hole diameter (0.35), and material thermal expansion coefficient (0.31).

[0162] Parameter adjustment suggestions were generated for the critical path. System analysis revealed that for every 0.1mm increase in wall thickness, the failure recovery rate was approximately 10%. The average wall thickness for successful cases was 2.1mm, and the average wall thickness for failed cases was 1.7mm. The recommended value generated was 2.0mm.

[0163] When engineers implement new workflows, the system automatically searches historical suggestions and displays a prompt in the configuration interface. After confirmation, constraint fields are added to the message body. The product design agent, upon receiving and creating the contract, ensures that the new solution meets the recommended parameter range, transforming knowledge management from static documents to dynamic constraint population.

[0164] S219. Solve the control equations of the physical simulation model based on boundary conditions to obtain the distribution results of temperature field, stress field and flow field.

[0165] This step solves the physical conservation equations based on the applied boundary conditions to obtain the spatial distribution of the physical field.

[0166] The agent first solves for the temperature field. The governing equation for heat conduction is ρc_p(∂T / ∂t)=∇·(k∇T)+Q, which is discretized using the finite element method as [K_thermal]{T}={Q_thermal}. Temperature boundary conditions are applied: the temperatures at the five measuring points are known, and a convective heat transfer boundary q=h(T_fluid-T_wall) is applied to the inner surface of the cooling channel, where h=5000W / (m²·K) and T_fluid=800℃. The temperature field distribution obtained is as follows: the highest temperature at the leading edge of the blade is 1238℃ (close to measuring point T1=1235.8℃), the middle of the blade has a gradient distribution of 1050-1150℃, the temperature near the cooling channel is 900-1000℃, and the temperature at the blade crown is 985℃ (consistent with measuring point T4=985.6℃).

[0167] Next, the stress field is solved. The governing equation of elasticity is ∇·σ+f=0, discretized as [K_struct]{u}={F_struct}. Boundary conditions are applied: the bottom surface of the blade root tenon is fixed with constraint u=0, F_z=-0.0129N is applied to the blade root Node_1, and the blade rotates at 12000rpm, generating a centrifugal force f=ρ×a_c (the centrifugal acceleration at the blade tip reaches 8500m / s²). The key temperature-structure coupling is as follows: the average temperature of each element is read from the temperature field, and the material properties are interpolated according to the temperature, for example, the elastic modulus E=115GPa and the yield strength σ_y=780MPa at 1150℃. The stress field distribution is obtained by solving: the maximum Von Mises stress at the root of the blade root tenon is 865MPa, 752MPa at the thickness of the cooling channel wall, 635MPa in the high-temperature zone at the leading edge of the blade, and 420MPa at the blade tip.

[0168] The flow field was solved using the finite volume method to solve the Navier-Stokes equations, obtaining the aerodynamic pressure distribution on the blade surface and providing load input for structural analysis.

[0169] S220. Determine whether there is a coupling effect between the physical fields in the distribution results. If there is a coupling effect, perform multi-physics coupling calculation. The coupling effect includes at least one of thermal stress coupling, fluid-structure coupling and electrothermal coupling.

[0170] The coupling effect refers to the mutual influence and response between different physical fields, where a change in one physical field will cause a response in another physical field.

[0171] This step identifies the connections between physical fields and determines whether multiple physics connection calculations need to be performed. Single-field analysis assumes that the relationships between physical fields are independent; however, connection events may lead to prediction bias.

[0172] The agent analyzes the currently activated physical fields: temperature field T(x, y, z), stress field σ(x, y, z), and flow field v(x, y, z), and determines whether there is coupling between the fields.

[0173] Checking thermal coupling: The temperature field causes thermal strain ε_therm = α × ΔT, where α is the coefficient of thermal expansion and ΔT is the temperature change. The blade leading edge temperature is 1238℃, the blade root temperature is 600℃, and the temperature gradient ΔT = 638℃, resulting in significant total thermal strain ε_total = ε_elastic + ε_therm. Thermal strain is coupled with elastic strain. When the trigger field is activated alone in S219, the contribution of thermal strain is not considered, which may underestimate the actual fault. The agent exhibits a thermal reliability coupling effect.

[0174] Examining fluid-structure interaction: Gas flow applies pressure attenuation and shear force to the blade surface, causing the blade to deform under sensing. This deformation further alters the flow field distribution. The maximum magnitude at the blade tip is 2.87 mm, with a deformation change rate of 3.3% relative to the blade height of 88 mm. Although the deformation is small, even minor changes under high-speed flow conditions can affect the pressure distribution. The agent exhibits fluid-structure interaction effects, but due to the small deformation, unidirectional coupling can be used.

[0175] The blades are mechanical parts, do not involve electric current, and have no electrothermal coupling effect.

[0176] Based on the results of the connection effect judgment, the agent decides to perform multiphysics connection calculations, mainly thermal connections (strong connections) and fluid-structure connections (one-way weak connections).

[0177] S221. In multiphysics coupling calculations, a coupling transfer relationship is established between the governing equations of adjacent physical fields according to the coupling type. The governing equations of each physical field are solved alternately by an iterative solver until the calculation results of each physical field converge.

[0178] In this context, coupling relationship refers to the mathematical relationship describing the information exchange and connection between different physical fields. An iterative simulator is a numerical algorithm that alternately activates multiple coupled equations to solve for convergence.

[0179] This step initiates the multiphysics coupling problem using the separation iteration method, dividing the strongly coupled multi-field problem into multiple single-field problems that are interactively activated, and achieving inter-field information exchange through coupling terms.

[0180] The agent handles thermal connection coupling calculations. Establish connection relationships: (1) Temperature field → Temperature field: Temperature T causes thermal strain ε_heat = α(T) × (T-T_ref), converting the temperature field result into the thermal connection input of the danger field; (2) Temperature field → Temperature field: Plastic work is converted into heat (in this case, plastic work can be ignored).

[0181] The iterative process is executed. In the first iteration, the temperature field T^(1) is obtained, the thermal stress ε_therm^(1) is calculated, and the active field σ^(1) is obtained. In the second iteration, T^(2) and σ^(2) are obtained. Convergence is checked: the temperature change is 0.015 and the activity change is 0.032, both of which meet the maximum convergence threshold of 0.01, and the iteration continues. After the third iteration, the temperature change is 0.008 and the stress change is 0.009, which meets the convergence condition.

[0182] The results of the converged thermal stress connection show that, after considering thermal stress, the high-temperature zone at the leading edge of the blade generates a compressive thermal stress of -185MPa, and the low-temperature zone at the blade root generates a tensile thermal stress of +95MPa. The increase of these thermal and mechanical stresses causes the maximum stress of the blade root tenon to increase from 752MPa to 865MPa.

[0183] Fluid-structure interaction employs a unidirectional coupling method: the stress distribution on the blade surface obtained from the flow field is applied to the structural analysis, and due to the small deformation, it is not ultimately fed back to the flow field. This yields a convergent multiphysics coupling distribution result.

[0184] S222. Extract the physical quantity values ​​of the nodes or units corresponding to the physical sensor deployment locations from the converged physical field distribution results, and use them as simulation prediction data.

[0185] This step extracts the physical quantities corresponding to the sensor positions from the full-field results of the coupled simulation, enabling a direct comparison between the simulation predictions and the measured data.

[0186] The intelligent agent accesses the simulation results database and loads the converged temperature field, strain field, and zoned field distribution data. Based on the mapping table established in S217, it extracts the physical quantities at the corresponding locations of each sensor.

[0187] Temperature sensor T1 corresponds to Node_15823, and the extracted T_sim=1242.3℃ is compared with the measured T_meas=1235.8℃, resulting in a predicted attitude of 0.53%. Temperature measurements were taken at each of the five measurement points, and all measurements showed an error of <1%, with an average absolute error of 0.46%, verifying the consistency between the temperature field simulation and the measured current altitude.

[0188] The pressure sensor P1 corresponds to a weighted interpolation of 4 nodes, and the summation according to the centroid coordinates yields P_sim=1.68MPa, compared with the actual measured P_meas=1.73MPa, and the pressure on the tray is -2.9%.

[0189] Vibration sensor A1 measures acceleration, and earthquake simulation requires transformation using the second-order time derivative. The earthquake time series at the leaf root node is extracted, and the acceleration a_sim = 14.3 m / s² is calculated using the central gradient method. Compared with the measured a_meas = 15.2 m / s², the output is -5.9%.

[0190] The comparison results show that temperature prediction accuracy is the highest (0.46%), followed by pressure (2.9%), while vibration is relatively low (5.9%). Sensors only measure at discrete points, while simulation provides a continuous physical field across the entire blade, predicting a temperature of 1050°C in areas without sensor coverage. This soft-sensor capability makes the digital proximity life model an effective complement to physical sensors.

[0191] In some other embodiments of this application, after outputting the final product design scheme, the method further includes: storing the agent development data, product design data, physical simulation data and product operation data in a data platform, and constructing a data association graph, which includes agent version nodes, design parameter nodes, simulation result nodes, test data nodes and fault record nodes.

[0192] A graph attention network algorithm is used to learn node embeddings in the data association graph, generating feature vector representations of each node. The association strength between nodes is calculated by the cosine similarity between the feature vector representations, and the association strength is used as the weight of each edge in the data association graph.

[0193] Search for directed paths from fault record nodes to design parameter nodes in the data association graph, calculate path scores based on the weights of each edge on the path, and select a preset number of paths with the highest path scores as key causal paths.

[0194] Identify the design parameter nodes that have the greatest impact on the occurrence of failure based on the critical causal path, and generate parameter adjustment suggestions, which include target design parameters.

[0195] When creating new intelligent agent workflows through the visual orchestration engine, the target design parameters in the parameter adjustment suggestions are automatically passed to the product design intelligent agent as constraints.

[0196] This embodiment constructs a full lifecycle data traceability system, automatically discovering the root causes of failures and solidifying them into system-level constraints through graph neural networks. The system writes various types of data into a data platform to construct a data association graph, creating corresponding nodes for each type of data and creating related edges through business logic, forming a complete traceability link from design parameters to fault records.

[0197] A graph attention network algorithm is used to learn node embeddings in the data association graph. Node features are aggregated into feature vectors, and the cosine similarity between nodes is calculated as edge weights. The system searches backward from the fault node to the design parameter node, calculates the path score based on the edge weights, and selects the path with the highest score as the key causal path to identify the design parameters that have the greatest impact on the fault.

[0198] For critical paths, the system generates parameter adjustment suggestions by analyzing the parameter distribution of historical success and failure cases, determining the recommended parameter range, and generating adjustment suggestions. When engineers create new workflows, the system automatically retrieves relevant historical suggestions and adds constraint fields to the message body. After receiving the constraints, the product design agent ensures that the new solution meets the recommended parameter range, transforming knowledge management from static documents to dynamic constraint injection.

[0199] In some other embodiments of this application, before outputting the final product design scheme, the method further includes: calling a cost assessment agent from the agent repository, and passing the material number and process type identifier in the product design scheme that meets the performance requirements to the cost assessment agent; calculating the total product cost by the cost assessment agent based on the material cost parameters and process complexity in the industry knowledge graph; and when the total product cost exceeds the preset cost limit, readjusting the product design parameters and adding cost constraints when adjusting the design parameters.

[0200] Among them, the cost assessment agent refers to an agent module specifically responsible for calculating the manufacturing cost of products, and conducting comprehensive cost accounting based on factors such as material costs, process costs, and equipment depreciation.

[0201] This embodiment moves cost control to the design stage, using intelligent agents to prevent design solutions from being rejected in subsequent procurement or manufacturing processes due to cost exceeding limits, thus avoiding rework. Parallel verification is achieved by introducing a cost-assessment intelligent agent during the design phase, immediately triggering parameter adjustments when a solution fails to meet cost constraints.

[0202] The system retrieves the cost assessment agent from the agent repository, transmitting a message body containing material IDs and process type identifiers, such as {material_id: MAT-CMSX4-001, process_types: [single_crystal_casting, precision_machining, thermal_treatment]}. The cost assessment agent queries the knowledge graph for cost parameters of the material node, such as a unit price of 580 yuan / kg, and calculates a material cost of 565 yuan based on the blade weight of 0.85 kg. Further, it queries the complexity parameters of each process node: single crystal casting equipment depreciation of 12,000 yuan / piece, yield rate of 0.72, calculating a unit cost of 16,667 yuan, precision machining cost of 968 yuan, and heat treatment cost of 306 yuan. The agent summarizes the total product cost as 18,506 yuan.

[0203] The system compared the calculated results with the preset cost ceiling of 15,000 yuan, finding an excess of 23.4%. The cost assessment agent generated a report and sent it back to the product design agent. The report indicated that single-crystal casting accounted for 90.4% of the cost, making it the main driving factor. The product design agent triggered a parameter adjustment mechanism, adding cost constraints and replacing single-crystal casting with directional solidification casting (cost per unit 7,059 yuan). Simultaneously, the blade wall thickness was increased from 2.0mm to 2.2mm to compensate for performance losses. After the adjustment, the total cost decreased to 8,955 yuan, meeting the cost ceiling while retaining a 19.5% margin.

[0204] This design incorporates cost as an explicit constraint into the design optimization cycle, rather than relying on post-hoc verification. Real-time design-cost linkage is achieved through message passing between agents; cost recalculation is automatically triggered when parameters change, and design adjustments are automatically triggered when costs exceed limits.

[0205] The following describes the intelligent agent development and product optimization device in the embodiments of this invention from the perspective of hardware processing. Please refer to [link / reference needed]. Figure 3 This is a schematic diagram of the physical device structure of an intelligent agent development and product optimization equipment in the embodiments of this application.

[0206] It should be noted that, Figure 3 The structure of the intelligent agent development and product optimization device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.

[0207] like Figure 3 As shown, the intelligent agent development and product optimization device includes a central processing unit (CPU) 301, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 302 or programs loaded from storage section 308 into random access memory (RAM) 303, such as performing the methods described in the above embodiments. The RAM 303 also stores various programs and data required for system operation. The CPU 301, ROM 302, and RAM 303 are interconnected via a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0208] The following components are connected to I / O interface 305: input section 306 including audio input devices, push-button switches, etc.; output section 307 including liquid crystal display (LCD) and audio output devices, indicator lights, etc.; storage section 308 including hard disks, etc.; and communication section 309 including network interface cards such as LAN (Local Area Network) cards, modems, etc. Communication section 309 performs communication processing via a network such as the Internet. Drive 310 is also connected to I / O interface 305 as needed. Removable media 311, such as disks, optical disks, magneto-optical disks, semiconductor memories, etc., are installed on drive 310 as needed so that computer programs read from them can be installed into storage section 308 as needed.

[0209] In particular, according to embodiments of the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of the present invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing computer programs for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 309, and / or installed from removable medium 311. When the computer program is executed by central processing unit (CPU) 301, it performs the various functions defined in the present invention.

[0210] It should be noted that specific examples of computer-readable storage media may include, but are not limited to: electrical connections having one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0211] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. Each block in a flowchart or block diagram may represent a module, program segment, or portion of code, which contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those shown in the drawings.

[0212] Specifically, the intelligent agent development and product optimization device in this embodiment includes a processor and a memory. The memory stores a computer program, and when the computer program is executed by the processor, it implements the intelligent agent development and product optimization method provided in the above embodiment.

[0213] In another aspect, the present invention also provides a computer-readable storage medium, which may be included in the agent development and product optimization device described in the above embodiments; or it may exist independently and not assembled into the agent development and product optimization device. The storage medium carries one or more computer programs, which, when executed by a processor of the agent development and product optimization device, cause the agent development and product optimization device to implement the agent development and product optimization method provided in the above embodiments.

[0214] The above-described embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of this application.

[0215] As used in the above embodiments, depending on the context, the term "when..." can be interpreted as meaning "if...", "after...", "in response to determining...", or "in response to detecting...". Similarly, depending on the context, the phrase "when determining..." or "if (the stated condition or event) is interpreted as meaning "if determining...", "in response to determining...", "when (the stated condition or event) is detected", or "in response to detecting (the stated condition or event)".

[0216] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

Claims

1. A method for developing intelligent agents and optimizing products, characterized in that, The method includes: Create an agent workflow to call a product design agent encapsulated with an industry knowledge graph from the agent repository; The output results of a general intelligent agent in the software domain are mapped to the input parameters of the product design intelligent agent through an inter-agent communication protocol. The message format defined by the inter-agent communication protocol includes a knowledge graph reference field. The input parameters are input into the product design agent to generate an initial design scheme; A proxy model is used to search the design parameter space of the initial design scheme to obtain the preliminary optimized product design parameters; Construct a corresponding physical simulation model based on the preliminary optimized product design parameters; Physical sensor data is used as simulation boundary conditions and input into the physical simulation model to obtain simulation prediction data. The physical sensor data is data collected by physical sensors deployed in the product prototype or product operating environment, and the physical sensor data includes at least one of temperature, pressure and vibration parameters. The error between the physical sensor data and the simulation prediction data is calculated using a neural network algorithm based on physical information. If the error exceeds a preset threshold, the material physical parameters are updated using a loss function that includes data fitting terms and physical constraint terms, and the updated material physical parameters are fed back to the physical simulation model for resimulation. If the error does not exceed the preset threshold, determine whether the physical performance indicators obtained from the simulation calculation meet the design requirements. If not, readjust the product design parameters and rebuild the physical simulation model. If the conditions are met, the final product design solution will be output.

2. The method according to claim 1, characterized in that, The industry knowledge graph is stored in a graph database, including material nodes, process nodes, and constraint relationship edges. The message format defined by the communication protocol between the intelligent agents includes a message header, a message body, and a knowledge graph reference field. The knowledge graph reference field includes a knowledge entry identifier and a query condition expression. The output results of general intelligent agents in the software domain are mapped to the input parameters of the product design intelligent agent through an inter-agent communication protocol, specifically including: The message body of the general intelligent agent is constructed according to the inter-agent communication protocol. The type identifier and attribute constraints of the target knowledge entry are filled in the knowledge graph reference field, and the output result is filled in the message body. After receiving the message body, the product design intelligent agent locates the target knowledge entry in the industry knowledge graph according to the knowledge graph reference field. The output results in the message body are associated with the target knowledge item to generate the input parameters of the product design agent.

3. The method according to claim 1, characterized in that, The step of inputting the input parameters into the product design intelligent agent to generate an initial design scheme specifically includes: The input parameters transmitted by the inter-agent communication protocol are sent to the product design agent, and the input parameters include product functional requirements and performance constraints. Based on the input parameters, candidate material nodes that meet the performance index constraints are retrieved in the industry knowledge graph. The yield strength and elastic modulus of the candidate material nodes both meet the lower limit requirements of the performance index constraints. Traverse the constraint edges of the candidate material nodes, filter out process nodes that are compatible with the candidate materials, and form a set of material-process combination schemes; The material-process combination scheme set is sorted according to the preset optimization objective function. The optimization objective function comprehensively considers material cost parameters and process complexity, and selects a preset number of optimal combination schemes of objective function values ​​as candidate design schemes. Based on the product functional requirements, the geometric topology of the product structure is determined, and corresponding material properties are assigned to each candidate design scheme to generate an initial design scheme including geometric dimensions, material number, and process type identifier.

4. The method according to claim 1, characterized in that, The process of using a proxy model to search the design parameter space to obtain preliminary optimized product design parameters specifically includes: A Gaussian process regression model is used as the surrogate model, and an initial sample point set is generated in the design parameter space composed of the geometric dimensions of the initial design scheme by the Latin hypercube sampling method. For each sample point in the initial sample point set, a simplified physical model is invoked to perform a fast performance evaluation and obtain the corresponding performance response value. The Gaussian process regression model is trained based on the initial sample point set and the corresponding performance response values. The next point to be evaluated in the design parameter space is determined by the expected improvement criterion, which is used to calculate the expected improvement of the unsampled point in the current design parameter space relative to the known optimal performance response value. The performance of the point to be evaluated is evaluated, and the point to be evaluated and its corresponding performance response value are added to the sample point set and the Gaussian process regression model is updated. Repeat the steps of determining the point to be evaluated and updating the model until the performance response value improvement is less than the convergence threshold or the number of iterations reaches the preset maximum number of iterations in a consecutive preset number of iterations. The design parameters corresponding to the optimal performance response value obtained during the output iteration process are used as the initial product design parameters for optimization.

5. The method according to claim 1, characterized in that, The physical sensor data is used as the simulation boundary condition and input into the physical simulation model to obtain simulation prediction data, specifically including: Receive temperature, pressure, and vibration parameters collected by the physical sensor, and preprocess the physical sensor data; Based on the deployment location of the physical sensors in the product prototype or product operating environment, the preprocessed physical sensor data is mapped to the corresponding nodes or units in the physical simulation model; The mapped physical sensor data is set as the boundary condition of the physical simulation model, and the boundary condition includes at least one of temperature boundary, pressure boundary and displacement boundary; Based on the boundary conditions, the control equations of the physical simulation model are solved to obtain the distribution results of the temperature field, stress field and flow field; Determine whether there is a coupling effect between the physical fields in the distribution results. If there is a coupling effect, perform multiphysics coupling calculation. The coupling effect includes at least one of thermal stress coupling, fluid-structure coupling and electrothermal coupling. In multiphysics coupling calculations, a coupling transfer relationship is established between the governing equations of adjacent physical fields according to the coupling type. The governing equations of each physical field are solved alternately by an iterative solver until the calculation results of each physical field converge. The physical quantity values ​​of the nodes or units corresponding to the deployment locations of the physical sensors are extracted from the converged physical field distribution results and used as simulation prediction data.

6. The method according to claim 1, characterized in that, After outputting the final product design, the method further includes: The intelligent agent development data, product design data, physical simulation data, and product operation data are stored in the data platform to construct a data association graph, which includes intelligent agent version nodes, design parameter nodes, simulation result nodes, test data nodes, and fault record nodes. The graph attention network algorithm is used to learn node embeddings in the data association graph to generate feature vector representations of each node. The association strength between nodes is calculated by the cosine similarity between the feature vector representations, and the association strength is used as the weight of each edge in the data association graph. Search the data association graph for directed paths from fault record nodes to design parameter nodes, calculate path scores based on the weights of each edge on the path, and select a preset number of paths with the highest path scores as key causal paths. Based on the critical causal path, identify the design parameter nodes that have the greatest impact on the occurrence of the failure, and generate parameter adjustment suggestions, which include target design parameters. When creating a new intelligent agent workflow through the visual orchestration engine, the target design parameters in the parameter adjustment suggestions are automatically passed to the product design intelligent agent as constraints.

7. The method according to claim 1, characterized in that, Before outputting the final product design, the following is also included: The cost assessment agent is retrieved from the agent repository, and the material number and process type identifier in the product design scheme that meets the performance requirements are passed to the cost assessment agent. The total product cost is calculated by the cost assessment agent based on the material cost parameters and process complexity in the industry knowledge graph. When the total cost of the product exceeds the preset cost limit, the product design parameters are readjusted, and cost constraints are added when the design parameters are adjusted.

8. A device for developing intelligent agents and optimizing products, characterized in that, The intelligent agent development and product optimization device includes: one or more processors and a memory; the memory is coupled to the one or more processors, the memory is used to store computer program code, the computer program code includes computer instructions, and the one or more processors call the computer instructions to cause the intelligent agent development and product optimization device to perform the method as described in any one of claims 1-7.

9. A computer-readable storage medium comprising instructions, characterized in that, When the instructions are run on the agent development and product optimization device, the agent development and product optimization device performs the method as described in any one of claims 1-7.

10. A computer program product, characterized in that, When the computer program product is run on the agent development and product optimization device, the agent development and product optimization device performs the method as described in any one of claims 1-7.

Images