Electrical design method and electrical design system

Abstract

The invention relates to the technical field of electrical engineering, and discloses an electrical design method and an electrical design system.The electrical design method comprises the following steps that design input parameters are collected and preprocessed, an initial state vector is generated, and a coupling modeling matrix of an electromagnetic field, a thermal field and a mechanical field is constructed; inputting the state vector into a reinforcement learning model to optimize design parameters, solving a coupling matrix in edge calculation hardware to obtain a physical field response, updating the response to the state vector, and feeding back the response to the reinforcement learning model for the next round of design optimization; an electrical design system comprises an input module, a modeling module, an optimization module, a solving module and a feedback module. According to the method, a reinforcement learning driven coupling modeling and action generation method is adopted, multi-physics field collaborative response modeling in the electrical design process is achieved, and the technical effect of dynamically and finely capturing complex interaction relations among electromagnetic, thermal, mechanical and other fields is achieved.

Description

Technical Field

[0001] The present invention relates to the field of electrical engineering technology, and in particular to an electrical design method and an electrical design system. Background Art

[0002] The design complexity of electrical systems is increasing in numerous application scenarios, including intelligent manufacturing, new energy equipment, and precision electrical control. Faced with increasingly compact space layouts, high-density integrated device configurations, and highly coupled working environments involving multiple physical fields (such as electricity, heat, and force), traditional static electrical design methods alone often struggle to meet the comprehensive performance, reliability, and adaptability requirements of modern industry.

[0003] Existing electrical design technologies already have relatively mature processes for modeling and solving problems. For example, engineers use rule-based topology modeling tools to build a preliminary system structure and utilize finite element analysis (FEA) or electromagnetic field simulation platforms to perform static field solutions, enabling basic electrical performance evaluations. Some commercial software also provides semi-automated wiring and structural optimization modules with certain simulation visualization capabilities and preliminary parametric tuning methods. These methods offer advantages such as ease of operation and stable calculations in the design of low- to medium-complexity electrical systems, and can also achieve a closed-loop design verification in relatively simple physical scenarios.

[0004] However, traditional methods mostly rely on empirical configuration. Once the parameter relationship becomes complex in the coupling scenario, the model is easily deformed and distorted, the feedback mechanism lags, and the adjustment effect is discounted. The existing process lacks technical means to dynamically adjust the design action according to the physical field response. The state vector update is often divorced from the actual feedback, resulting in a rigid design path and poor adaptability. In the simulation solution process, most of them rely on a centralized computing architecture. When faced with concurrent computing of multiple physical fields, resource scheduling is not sensitive, making it difficult to quickly complete design verification in an edge environment. In addition, the action generation in the optimization process is generally fixed strategy control, which lacks the ability to learn and self-iterate, resulting in the design results easily falling into local optimality and difficulty in breaking through the global optimization bottleneck under complex objective functions. To this end, those skilled in the art have proposed an electrical design method and electrical design system to solve the above problems. Summary of the Invention

[0005] In response to the shortcomings of the existing technology, the present invention provides an electrical design method and an electrical design system, which solve the problems in the existing technology such as insufficient multi-physical field interactive response modeling capabilities, lack of feedback mechanism, strong design path rigidity and lack of efficient and intelligent optimization strategies.

[0006] To achieve the above objectives, the present invention is implemented through the following technical solutions: an electrical design method and an electrical design system, comprising the following steps:

[0007] Step 1: Collect the input parameters required for electrical design and perform preprocessing to generate the initial state vector;

[0008] Step 2: constructing a coupling modeling matrix of the electromagnetic field, thermal field and mechanical field based on the initial state vector;

[0009] Step 3: Input the state vector into the reinforcement learning model and optimize the design parameters in combination with the coupling modeling matrix;

[0010] Step 4: Solve the coupling modeling matrix in the edge computing hardware according to the optimization results to obtain the responses of each physical field;

[0011] Step 5: Update the physical field response to the state vector and feed it back to the reinforcement learning model for the next round of design action generation.

[0012] Preferably, the step 1 includes:

[0013] Collect electrical characteristic parameters, including rated voltage, current, resistance, inductance, and capacitance;

[0014] Obtain three-dimensional environmental field information to form a tensor input of fixed dimension;

[0015] Set constraint parameters, including maximum allowable temperature, maximum allowable structural displacement, and maximum design cost;

[0016] The acquired parameters are normalized and structured to generate a state vector as the input for modeling and optimization.

[0017] Preferably, the step 2 includes:

[0018] Construct an electromagnetic field finite element model based on vector potential form;

[0019] Construct the thermal field finite element model corresponding to the heat conduction control equation;

[0020] Construct a finite element model of mechanical fields based on structural dynamics;

[0021] The above three physical models are combined into a multi-physics coupling matrix and unified modeling is adopted using a block diagonal structure.

[0022] Preferably, the step three includes:

[0023] Define state space and action space, where the state space includes electrical parameters, field response parameters and environmental information;

[0024] Build a reinforcement learning model, including a policy network and constraint mapping structure;

[0025] Set up a multi-objective reward function based on a combination of thermal constraints, electrical performance, and cost factors;

[0026] The output action variables are used to adjust structural parameters, electrical component configuration and topology design schemes.

[0027] Preferably, the step 4 includes:

[0028] dividing the coupling modeling matrix into dense blocks and sparse blocks;

[0029] Allocate dense blocks to high-performance solver cores, and schedule sparse blocks through on-chip cache;

[0030] Solve the modeling matrix in parallel on the edge computing platform;

[0031] Output physical response results, including node temperature, structural displacement, and current distribution.

[0032] Preferably, the step five includes:

[0033] Receive the solution result of step 4 and update the state vector;

[0034] Inputting the state vector into the reinforcement learning model;

[0035] Calculating the next action output using the model;

[0036] Continue the next round of physical modeling and solving until the design stop condition is reached.

[0037] Preferably, the electromagnetic field finite element model satisfies the following relationship:

[0038] The electromagnetic field adopts the form of vector potential, and the constructed electromagnetic finite element equation is:

[0039] Electromagnetic stiffness matrix K em The relationship between it and the vector potential a is:

[0040] K em a=f em ;

[0041] Where: a is the vector potential column vector at the node; K em Electromagnetic stiffness matrix; f em is the equivalent electromagnetic source term vector.

[0042] Preferably, the thermal field finite element modeling satisfies the following relationship:

[0043] The node temperature vector θ satisfies the heat balance equation:

[0044]

[0045] Where: C th is the heat capacity matrix; K this the thermal conductivity matrix; θ is the node temperature column vector; f th is the node heat source term.

[0046] Preferably, the reward function satisfies the following relationship:

[0047] At each optimization step t, the reward function R t Expressed as:

[0048]

[0049] Where: α t , β t , γ t is the dynamic weight of each target; P rated is the system rated power; P loss is the calculated power loss; C t is the current design cost; C max is the maximum allowable cost; T t is the maximum node temperature; T max The maximum temperature allowed.

[0050] An electrical design system, comprising:

[0051] Input module, used to collect electrical parameters, environmental parameters and design constraints, and generate state vectors;

[0052] a modeling module, configured to construct coupled electromagnetic, thermal, and mechanical field models based on the state vector and output a multi-physics coupling matrix;

[0053] an optimization module, configured to receive the state vector and the modeling matrix and output a designed action through a reinforcement learning model;

[0054] A solution module, configured to receive the modeling matrix and the design action, perform accelerated solution in the edge computing platform, and output a physical response;

[0055] A feedback module is configured to receive the physical response and update the state vector, and feed the updated state vector back to the optimization module.

[0056] The present invention provides an electrical design method and an electrical design system. They have the following beneficial effects:

[0057] 1. The present invention adopts a reinforcement learning-driven coupling modeling and action generation method to realize the collaborative response modeling of multiple physical fields in the electrical design process, achieving the technical effect of being able to dynamically and precisely capture the complex interactive relationships between electromagnetic, thermal, mechanical and other fields. Compared with the single-field or weakly coupled modeling methods in the existing technology, this solution solves the problems of insufficient simulation of multi-physical field interactions and large deviations in calculation results.

[0058] 2. The present invention introduces an edge computing platform in the solution stage, and allocates and schedules the coupling matrix according to dense blocks and sparse blocks, which improves the efficiency and response speed of numerical calculations, and achieves the advantages of high hardware resource utilization and strong computational concurrency. Different from the traditional cloud platform-based remote modeling and solution technology path, the present invention effectively makes up for its shortcomings such as high computing delay and inflexible on-site deployment.

[0059] 3. The present invention establishes a real-time feedback mechanism between the state vector and the physical response, so that the reinforcement learning model updates its actions based on the actual response data in each round of iteration. Therefore, the design optimization path is in line with reality and the response is more accurate. Compared with traditional design schemes based on empirical rules or static objective functions, it avoids the problems of large adjustment blind spots and low optimization efficiency.

[0060] 4. The present invention realizes intelligent adjustment and adaptive evolution of the electrical design process by constructing a closed-loop optimization process of coupled physical field-state vector-reinforcement learning-action feedback. It can automatically switch design strategies according to target and environmental changes. Compared with existing technical models that rely on manual parameter adjustment or fixed processes, it effectively solves the bottlenecks of low automation and weak adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the method flow of the present invention;

[0062] Figure 2 Schematic diagram of the system architecture of the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the present specification. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] Please see the attached Figure 1 , an embodiment of the present invention provides an electrical design method, comprising the following steps:

[0065] Step S1, collecting the input parameters required for electrical design and preprocessing them to generate an initial state vector;

[0066] Specifically, in this embodiment, step S1 is mainly responsible for collecting the input parameters required for electrical design and preprocessing them to generate an initial state vector. This step is the starting point of the entire electrical design method, ensuring that subsequent modeling, optimization, and solution can be based on accurate and standardized input data. Input parameters include electrical characteristic parameters, environmental field information, and design constraints. By normalizing these input parameters, parameters of different dimensions and dimensions can be uniformly represented, providing accurate basic data for subsequent modeling and optimization.

[0067] Generally, input parameter collection is determined based on actual application requirements and design goals. Their accuracy directly impacts the quality of the design results and optimization effectiveness. Alternatively, electrical characteristic parameters can be obtained from device manuals, user specifications, or measuring instruments. Environmental field information can be acquired through sensors, physical simulation, or 3D modeling tools. Proper design ensures that these input data are accurately collected and processed.

[0068] In the specific implementation, various electrical characteristic parameters, including rated voltage, current, resistance, inductance, and capacitance, are first collected using appropriate measuring equipment or tools. Electrical characteristic parameters are fundamental parameters of electrical equipment and systems, determining their performance and operating conditions. For example, resistance determines the resistance to current flow in a circuit, which in turn affects important characteristics such as power loss and temperature rise.

[0069] Next, environmental field information is acquired and formed into a fixed-dimensional tensor input, typically including factors such as temperature, humidity, and radiation. This environmental field information reflects the impact of the external environment on the performance of electrical equipment and is particularly important for optimizing aspects such as thermal management and electromagnetic interference. For example, in electrical design, the operating temperature of the equipment may affect the performance and reliability of electrical components. Therefore, accurately acquiring environmental information and considering its impact on system design is crucial.

[0070] In some embodiments, environmental field information can be collected via sensors or predicted using simulation models. In this case, the information collected by the sensors can be converted into three-dimensional environmental field data and input into subsequent modeling and optimization processes in the form of tensors. For certain complex environmental fields, simulation models can provide more accurate data, allowing design solutions to be verified under multiple external conditions.

[0071] To facilitate subsequent modeling and optimization, all collected parameters require preprocessing. The primary task of preprocessing is to normalize the collected electrical characteristic parameters and environmental field information. The goal of normalization is to convert data of varying dimensions and scales into a unified, standardized format, enabling seamless integration of input data from diverse sources and providing consistent input for subsequent steps. Through normalization, the system effectively eliminates dimensional differences between parameters, avoiding calculation errors or data biases caused by inconsistent dimensions.

[0072] During the normalization process, the following formula can be used:

[0073]

[0074] in: is the normalized value; X is the original input parameter; X min and X max In this way, all parameters are scaled to the range of [0, 1], ensuring the consistency and comparability of the data.

[0075] After normalization, all input parameters form a state vector. This state vector serves as the initial input for the entire electrical design and contains important data such as electrical characteristics, environmental information, and constraints. This state vector not only provides foundational data for subsequent modeling, optimization, and solution, but also ensures that reinforcement learning models can be trained and optimized to consistent standards.

[0076] In a specific embodiment, the state vector is constructed by mapping and normalizing each collected parameter one by one. All input data, including electrical parameters, environmental field information, and design constraints, are uniformly processed and converted into components of the state vector. This data not only reflects the specific characteristics required for the electrical design, but also reflects the design goals and constraints. By aggregating this data into a standardized state vector, the system can effectively integrate multi-dimensional information into an easy-to-process format, ensuring smooth data connection and accuracy between subsequent steps.

[0077] Specifically, the state vector from step S1 is passed as input to step S2 for subsequent physical modeling and optimization calculations. Therefore, step S1 is not only a preliminary data collection and processing step but also a key step in laying a solid foundation for the entire electrical design process. This comprehensive and meticulous processing of input data ensures data consistency and calculation accuracy in subsequent steps.

[0078] In one possible implementation, in addition to basic electrical characteristics and environmental information, design constraints can also be defined in this step. For example, constraints such as maximum temperature, maximum structural displacement, and maximum design cost can influence decisions during subsequent optimization. Therefore, ensuring these constraints are accurately collected and processed is an essential part of step S1.

[0079] The initial state vector generated by collecting, normalizing, and mapping the input parameters in step S1 provides accurate and standardized basic data for subsequent physical modeling, reinforcement learning optimization, and solution. This state vector is key to the entire electrical design method, carrying the initial input information in the system and serving as the starting point for multi-physics coupled modeling and optimization solutions.

[0080] Step S2, constructing a coupling modeling matrix of the electromagnetic field, thermal field and mechanical field based on the initial state vector;

[0081] Specifically, in this embodiment, the main task of step S2 is to construct a coupled modeling matrix for the electromagnetic, thermal, and mechanical fields based on the initial state vector generated in step S1. Through this process, the modeling and calculation of different physical fields are organically combined to form a unified multi-physics coupling model, providing the necessary mathematical foundation for subsequent reinforcement learning optimization and edge computing solutions.

[0082] Specifically, the coupled modeling matrix for the electromagnetic, thermal, and mechanical fields is established using the finite element method (FEM). This method numerically discretizes the equations for each physical field, enabling rapid and accurate quantification of the responses of each field during the subsequent solution process. The electromagnetic, thermal, and mechanical fields are interconnected via the coupling matrix, ensuring that each field can be solved and optimized within a unified framework.

[0083] The coupling between electromagnetic, thermal, and mechanical fields is typically highly complex, necessitating the merging of these physical field models into a unified model using a block diagonal structure. This allows the influence of these different physical fields to be considered simultaneously within the overall model, while avoiding excessive computational redundancy during the solution process.

[0084] Electromagnetic field modeling typically uses the finite element method based on vector potentials. Based on the fundamental principles of electromagnetic fields, electric and magnetic fields can be described using vector potentials, thus avoiding the need to directly solve the electric and magnetic field equations. The finite element equations for electromagnetic fields, assuming the electromagnetic field propagates in free space, can be expressed as:

[0085] K em a=f em ;

[0086] Where: a is the vector potential column vector at the node; K em Electromagnetic stiffness matrix; f em is the equivalent electromagnetic source term vector.

[0087] Through this electromagnetic field model, the interaction between electric and magnetic fields and their impact on electrical design can be effectively described.

[0088] Specifically, the electromagnetic stiffness matrix K em It is the electromagnetic field control equation discretized by the finite element method, which is usually determined by factors such as the material properties of the electromagnetic medium, current distribution, and external electric field. em It includes the electric or magnetic field sources applied by external power sources, which have a direct impact on the electromagnetic response of the system.

[0089] Thermal field modeling involves the finite element equations for heat conduction, which are usually described by the heat capacity matrix and the thermal conductivity matrix. Specifically, the node temperature vector θ satisfies the following thermal balance equation:

[0090]

[0091] Where: C th is the heat capacity matrix; K th is the thermal conductivity matrix; θ is the node temperature column vector; f th is the node heat source term.

[0092] The thermal field model is based on the principle of conservation of energy. It helps predict the temperature changes of equipment or systems under different working environments by calculating the heat conduction process of materials under different conditions.

[0093] In the coupled modeling of the thermal field, the heat capacity matrix C th Represents the heat capacity characteristics of the system, that is, the system's ability to store heat; thermal conductivity matrix K th The matrices reflect the thermal conductivity of each component in the system. These physical properties vary with factors such as material type, size, and structure. By calculating these matrices, the system can accurately predict temperature changes, thereby optimizing temperature control and thermal management in electrical design.

[0094] Mechanical field modeling is primarily based on structural dynamics theory, taking into account the mechanical properties of materials and structural deformation. In electrical design, mechanical field analysis typically involves analyzing the strength, stiffness, displacement, and stress of devices and components. Finite element equations for mechanical fields help evaluate the mechanical performance of systems by describing the deformation and response of objects under external forces. Specific modeling methods involve discretizing the structural elements, calculating the mechanical response of each element, and aggregating this response into the overall structural response.

[0095] In mechanical field modeling, system deformation is typically described by a structural stiffness matrix and external force vectors. These external force vectors include forces acting on the structure, such as electromagnetic forces and gravity. By constructing a structural dynamics model, we can effectively analyze the mechanical behavior of the device under electrical and external conditions, further optimizing the mechanical performance of the design.

[0096] In this embodiment, the coupling of electromagnetic, thermal, and mechanical fields is achieved by combining the finite element equations for each physical field into a unified multi-physics coupling matrix. This matrix adopts a block diagonal structure, with the model of each physical field occupying a diagonal block. This structure allows the mutual influence of multiple physical fields to be uniformly described within the same mathematical framework, avoiding duplicate calculations and information loss.

[0097] In the multiphysics coupling matrix, the coupling relationships between electromagnetic, thermal, and mechanical fields are reflected through corresponding cross terms. For example, changes in the electromagnetic field affect the temperature distribution, which in turn can cause structural deformation in the mechanical field; deformation in the mechanical field can further affect the distribution of the electromagnetic field. These complex coupling relationships are reflected through the coupling terms in the matrix, providing real-time feedback on the interactions between the various physical fields during subsequent optimization and solution.

[0098] Step S3: input the state vector into the reinforcement learning model and optimize the design parameters in combination with the coupling modeling matrix;

[0099] Specifically, in this embodiment, the main task of step S3 is to input the multi-physics field coupling modeling matrix constructed in step S2 and the initial state vector generated in step S1 into the reinforcement learning model, and then combine the current state information to optimize the design parameters. This step is a key part of the present invention, aiming to automatically adjust the design parameters through a reinforcement learning (RL) algorithm, so that the electrical design can be optimized under multiple objective constraints to meet the design requirements.

[0100] In the previous steps, the initial state vector and coupled modeling matrix have been prepared and contain all the necessary parameters and physical field responses of the electrical system. Step S3 optimizes the design solution in the current state to the optimal action through the training and inference of the reinforcement learning model. This process is crucial to the optimization effect of the design. The reinforcement learning model uses the information in the state vector to evaluate the effects of different design parameters and select the best action (such as adjusting the configuration of electrical components or changing the structural design) so that the final design maximizes the objective function while satisfying multiple constraints.

[0101] Typically, the input to a reinforcement learning model is the current system state vector, and the output is an optimized design solution. Through interaction with the environment, the reinforcement learning algorithm continuously evaluates the impact of its actions on the goal and optimizes itself based on the reward function. Specifically, the design of the state space and action space ensures rapid iteration and optimization at different design stages.

[0102] In a specific embodiment, the input state vector includes electrical parameters (such as voltage, current, and resistance), field response parameters (such as the distribution of electromagnetic, thermal, and mechanical fields), and environmental information (such as external temperature and humidity). The reinforcement learning model selects actions based on this information, thereby optimizing the design solution. The optimized design solution can achieve the optimal performance and cost balance while satisfying the constraints by controlling the configuration of electrical components and the structural topology.

[0103] In reinforcement learning, defining the state space and action space is a key step. The state space is usually composed of all important variables involved in the design process, including electrical parameters, physical field responses, and environmental factors. Specifically, the state vector can include the following:

[0104] Electrical parameters: including but not limited to rated voltage, current, resistance, inductance, capacitance, etc.;

[0105] Physical field response parameters: including temperature, displacement, current, etc. of electromagnetic field, thermal field, and mechanical field;

[0106] Environmental factors: including external environmental information, such as temperature, humidity, radiation, etc.

[0107] The action space defines actionable design parameters or decisions. For example, the action space might include adjusting electrical component configurations, changing the electrical system structure, selecting different materials, or optimizing the design topology. These actions serve as the decision outputs of the reinforcement learning model to guide the design optimization process.

[0108] In this embodiment, the core of the reinforcement learning model is a policy network, which is trained to select the optimal action based on the input state. Typically, the policy network's input is the current state vector, and its output is a probability distribution representing the action choices. Through repeated training, the model gradually learns which design actions lead to the optimal design results.

[0109] During the training process of the reinforcement learning model, a reward function is used to guide model learning. The reward function is usually defined based on a comprehensive consideration of design goals and constraints. Specifically, the reward function of the present invention combines multiple target factors, such as power loss, electrical performance, cost control, and constraints such as temperature and structural displacement. The form of the reward function can be expressed as:

[0110]

[0111] Where: α t , β t , γ t is the dynamic weight of each target; P rated is the system rated power; P loss is the calculated power loss; C t is the current design cost; C max is the maximum allowable cost; T t is the maximum node temperature; T max The maximum temperature allowed.

[0112] The reward function gives a numerical feedback based on the achievement of these goals to guide the model for optimization.

[0113] In one possible implementation, the model gradually improves the accuracy of its design decisions through continuous iterative optimization. As training progresses, the model learns how to balance various design objectives and find the optimal design solution while satisfying the constraints.

[0114] The output of the reinforcement learning model is an optimized design action, specifically an adjustment to design parameters. The optimization process uses the model to calculate the optimal design action for the current state and feeds this back into the system. In practice, these actions can include adjusting electrical component parameters, optimizing structural design, modifying topology, and so on. Through multiple rounds of optimization, the design gradually converges to the optimal solution.

[0115] In some embodiments, the optimized design can be verified using automated tools or simulation software to ensure its feasibility and effectiveness in real-world applications. In some cases, the optimization may involve complex design decisions, such as selecting specific materials or adjusting component layouts, based on the reinforcement learning model's predictions and evaluations of different actions.

[0116] Step S4: solving the coupling modeling matrix in the edge computing hardware according to the optimization results to obtain the responses of each physical field;

[0117] Specifically, in this embodiment, the main task of step S4 is to solve the coupling modeling matrix obtained through reinforcement learning optimization in step S3 on the edge computing platform to obtain the response of each physical field. This step is crucial in the entire electrical design process. It not only completes the numerical solution of the theoretical model but also accelerates the solution process through an efficient computing platform, ensuring the efficiency and feasibility of the design.

[0118] In the aforementioned steps, step S3 uses reinforcement learning to optimize the design action that suits the current design goal. Furthermore, the coupled models of multiple physical fields are solved. In this step, the solution process is primarily based on these coupled modeling matrices. The specific mathematical model and calculation method directly determine the accuracy and efficiency of the physical response.

[0119] The physical field coupling matrix involved in electrical design is typically highly complex, encompassing the cross-influences between multiple physical fields. Therefore, solving this matrix typically requires a high-performance computing platform. Edge computing is particularly crucial in practical applications, especially when faced with a large number of parameters and complex calculations.

[0120] In one possible implementation, the coupled modeling matrix is ​​divided into dense and sparse blocks to improve computational efficiency. Dense blocks typically contain information-intensive components, making them suitable for processing on high-performance computing cores; sparse blocks, on the other hand, are simpler and suitable for scheduling via on-chip cache. This approach reduces wasted computing resources and significantly improves solution speed.

[0121] In the coupled modeling matrix for electromagnetic, thermal, and mechanical fields, each physical field has its own independent matrix representation, but the cross terms between these matrices reflect the coupling relationship between the different physical fields. Specifically, changes in the electromagnetic field affect the temperature distribution in the thermal field, and changes in the thermal field can cause structural deformation in the mechanical field, which in turn affects the distribution of the electromagnetic field.

[0122] In order to effectively solve the coupling modeling matrix, numerical solution methods such as the finite element method (FEM) are usually used. Through the finite element method, the entire physical field is divided into several discrete units, and the physical response of each unit is solved by numerical integration. In the electromagnetic field model, the electromagnetic stiffness matrix (K em ) and the potential vector (a), the distribution of electric and magnetic fields is obtained; in the thermal field model, the heat capacity matrix (C th ) and thermal conductivity matrix (K th ) to obtain the distribution of temperature field; in the mechanical field model, the deformation of the structure is determined by the structural stiffness matrix (K mech ) and the external load matrix (F mThe coupling relationship of ech) is described.

[0123] The specific solution of the coupling matrix can be achieved through the following steps:

[0124] A coupled x = b;

[0125] Among them: A coupled is the multi-physics coupling matrix; x is a vector containing the responses of physical fields such as electromagnetic fields, thermal fields, and mechanical fields; b is the external excitation term or boundary condition.

[0126] To improve solution efficiency, this embodiment uses an edge computing platform. The advantage of edge computing is that it can transfer computing tasks from traditional cloud servers to the device side or the edge of the network, thereby reducing data transmission time and computing latency, which is particularly important in applications that require real-time processing. Through the edge computing platform, the solution of the coupling modeling matrix can be performed efficiently on the device, ensuring real-time feedback of data during the electrical design process.

[0127] In some embodiments, the edge computing platform can further improve computing speed by processing dense and sparse blocks in parallel using high-performance GPUs or dedicated hardware accelerators. For dense blocks, the powerful parallel computing capabilities of the GPU can be utilized; for sparse blocks, on-chip cache can be used for rapid scheduling, reducing pressure on memory bandwidth.

[0128] By solving the coupled modeling matrix in parallel on the edge computing platform, the calculation process can be effectively accelerated. During this process, physical responses such as current, voltage, temperature, displacement, etc. will be output as solution results. These physical responses are crucial for subsequent design optimization and verification. Specifically, physical responses can include:

[0129] The distribution of electromagnetic fields, such as the distribution of electric and magnetic field strengths;

[0130] Temperature distribution of thermal field, such as node temperature and heat flux density;

[0131] Structural deformations from mechanical fields such as nodal displacements, stresses, and strains.

[0132] These response results provide an important basis for verifying and improving electrical designs. By analyzing the physical response, we can further determine whether the design meets the constraints (such as temperature, displacement, current, etc.) and decide whether further optimization adjustments are needed.

[0133] Once the responses of the various physical fields are solved on the edge computing platform, the results are fed back to the reinforcement learning model in step S3. Based on these feedback results, the model evaluates whether the current design meets expectations and further adjusts design parameters based on changes in the physical responses, leading to the next round of optimization. In this closed-loop design process, design objectives and constraints are continuously optimized and adjusted, ultimately resulting in an optimal electrical design.

[0134] In step S5, the physical field response is updated to the state vector and fed back to the reinforcement learning model for the next round of design action generation.

[0135] Specifically, in this embodiment, the primary task of step S5 is to update the physical field response obtained in step S4 to the initial state vector and feed the updated state vector back to the reinforcement learning model, thereby providing a basis for generating the next round of design actions. This step is a key step in the present invention's process, enabling adaptive adjustment and continuous optimization of the optimization process through a feedback mechanism.

[0136] In the aforementioned steps, physical field response results (such as temperature, current, electric field, magnetic field, and mechanical deformation) are obtained through parallel calculations on the edge computing platform and used as feedback information. Based on this, the updated state vector serves as input to the reinforcement learning model for the next round of optimization. This mechanism not only enhances the dynamic adaptability of the design solution but also ensures continuous improvement throughout the design process, ultimately achieving the optimal design.

[0137] Generally speaking, the introduction of a physical response feedback mechanism can enhance the model's self-learning capabilities, allowing each optimized design to be further adjusted based on actual feedback. Alternatively, the physical response feedback can include not only electromagnetic, thermodynamic, and mechanical responses, but also other factors that may affect the design, such as material aging and environmental changes.

[0138] In a specific implementation, the implementation of step S5 can be divided into several key sub-steps, as follows:

[0139] In step S4, the physical field response results (such as the distribution of the electromagnetic field, changes in the temperature field, and stress and displacement in the mechanical field) obtained by the edge computing platform are fed back into the system's state vector. Specifically, the physical response data is integrated and mapped back to the relevant positions of the state vector to form an updated state vector. This updated state vector contains the new design information and can accurately reflect the current design's performance in the physical field.

[0140] In some embodiments, the physical response results are weighted and synthesized with the input parameters in the original state vector to generate a new state vector. This new state vector can be viewed as "feedback" to the design, combining the optimization results with the environmental response, allowing subsequent optimization to more precisely adjust the design strategy.

[0141] The state vector update mechanism is a key innovation of this invention. By updating the state vector based on feedback from the physical field, the reinforcement learning model can continuously adjust its decision-making process to optimize the design solution. The updated state vector includes electrical parameters, field response parameters (such as temperature, current, displacement, etc.), and environmental information (such as external temperature and humidity). This updated information is fed into the reinforcement learning algorithm to guide the model in selecting the optimal design action.

[0142] Specifically, the updated state vector can be generated as follows:

[0143] S new =S old +ΔS;

[0144] Where: S new is the updated state vector; S old is the old state vector; ΔS is the adjustment to the state vector caused by the change in physical response. This update formula indicates that the change in the state vector is based on the impact of the physical response on the design.

[0145] The updated state vector is fed back into the reinforcement learning model as input. Based on this new input, the reinforcement learning model recalculates and generates the next round of designed actions. Specifically, the reinforcement learning model evaluates the effects of different actions in the current state, calculates the corresponding rewards, and selects the optimal action for design adjustment.

[0146] In some embodiments, the reinforcement learning model's reward function adjusts based on each change in the physical response, enabling the model to continuously improve design performance with each optimization iteration. The reward function typically integrates design goals and constraints, such as power loss, electrical performance, and temperature control, to provide a comprehensive evaluation.

[0147] Through feedback from the reinforcement learning model, the system is able to generate new design actions during each round of optimization. Design actions include adjusting electrical component configurations, optimizing structural topology, selecting different materials, etc. These adjustments will be optimized based on the latest state vector and physical response.

[0148] Specifically, design actions are generated based on feedback data from the current state and the optimization objective. During each feedback cycle, the reinforcement learning model calculates a set of possible actions based on the updated state vector and selects the optimal action based on the reward function. This process not only improves design efficiency but also ensures the feasibility of the design results within multiple constraints.

[0149] The feedback mechanism in this embodiment forms a closed-loop optimization process by continuously updating the state vector, feeding back the physical response, and generating design actions. Through multiple rounds of optimization iterations, the design continuously converges towards the optimal solution, ensuring that the system achieves optimal performance while meeting the constraints.

[0150] In some embodiments, the design process may encounter environmental changes or system parameter adjustments. In this case, the feedback mechanism can automatically adjust the design parameters to ensure that the system can still maintain an efficient and stable working state under the new conditions.

[0151] An electrical design system described below and an electrical design method described above may refer to each other.

[0152] Please see the attached Figure 2 , an electrical design system, comprising:

[0153] Input module, used to collect electrical parameters, environmental parameters and design constraints, and generate state vectors;

[0154] Modeling module, used to build coupled electromagnetic, thermal and mechanical field models based on the state vector and output the multi-physics coupling matrix;

[0155] The optimization module receives the state vector and modeling matrix and outputs the designed action through the reinforcement learning model;

[0156] The solver module receives the modeling matrix and design actions, performs accelerated solving in the edge computing platform, and outputs the physical response;

[0157] Feedback module, which receives physical responses and updates the state vector, and feeds the updated state vector back to the optimization module

[0158] The system of this embodiment can be used to execute the above method embodiments, and its principles and technical effects are similar, so they will not be repeated here.

[0159] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. An electrical design method, characterized in that: The following steps are involved: Step 1: Collect the input parameters required for electrical design and perform preprocessing to generate the initial state vector; Step 2: constructing a coupling modeling matrix of the electromagnetic field, thermal field and mechanical field based on the initial state vector; Step 3: Input the state vector into the reinforcement learning model and optimize the design parameters in combination with the coupling modeling matrix; Step 4: Solve the coupling modeling matrix in the edge computing hardware according to the optimization results to obtain the responses of each physical field; Step 5: Update the physical field response to the state vector and feed it back to the reinforcement learning model for the next round of design action generation.

2. An electrical design method according to claim 1, characterized in that: The step one comprises: Collect electrical characteristic parameters, including rated voltage, current, resistance, inductance, and capacitance; Obtain three-dimensional environmental field information to form a tensor input of fixed dimension; Set constraint parameters, including maximum allowable temperature, maximum allowable structural displacement, and maximum design cost; The acquired parameters are normalized and structured to generate a state vector as the input for modeling and optimization.

3. An electrical design method according to claim 1, characterized in that: The second step includes: Construct an electromagnetic field finite element model based on vector potential form; Construct the thermal field finite element model corresponding to the heat conduction control equation; Construct a mechanical field finite element model based on structural dynamics; The above three physical models are combined into a multi-physics coupling matrix and unified modeling is adopted using a block diagonal structure.

4. An electrical design method according to claim 1, characterized in that: The step three includes: Define state space and action space, where the state space includes electrical parameters, field response parameters and environmental information; Build a reinforcement learning model, including a policy network and constraint mapping structure; Set up a multi-objective reward function based on a combination of thermal constraints, electrical performance, and cost factors; The output action variables are used to adjust structural parameters, electrical component configuration and topology design schemes.

5. An electrical design method according to claim 1, characterized in that: The fourth step includes: dividing the coupling modeling matrix into dense blocks and sparse blocks; Allocate dense blocks to high-performance solver cores, and schedule sparse blocks through on-chip cache; Solve the modeling matrix in parallel on the edge computing platform; Output physical response results, including node temperature, structural displacement, and current distribution.

6. An electrical design method according to claim 1, characterized in that: The step five includes: Receive the solution result of step 4 and update the state vector; Inputting the state vector into the reinforcement learning model; Calculating the next action output using the model; Continue the next round of physical modeling and solving until the design stop condition is reached.

7. An electrical design method according to claim 3, characterized in that: The electromagnetic field finite element model satisfies the following relationship: The electromagnetic field adopts the form of vector potential, and the constructed electromagnetic finite element equation is: Electromagnetic stiffness matrix K em The relationship between it and the vector potential a is: K em ·a=f em ; Where: a is the vector potential column vector at the node; K em Electromagnetic stiffness matrix; f em is the equivalent electromagnetic source term vector.

8. An electrical design method according to claim 3, characterized in that: The thermal field finite element modeling satisfies the following relationship: The node temperature vector θ satisfies the heat balance equation: Where: C th is the heat capacity matrix; K th is the thermal conductivity matrix; θ is the node temperature column vector; f th is the node heat source term.

9. An electrical design method according to claim 4, characterized in that: The reward function satisfies the following relationship: At each optimization step t, the reward function R t Expressed as: Where: α t , β t , γ t is the dynamic weight of each target; P rated is the system rated power; P loss is the calculated power loss; C t is the current design cost; C max is the maximum allowable cost; T t is the maximum node temperature; T max The maximum temperature allowed.

10. An electrical design system, applied to an electrical design method according to any one of claims 1 to 9, characterized in that: include: Input module, used to collect electrical parameters, environmental parameters and design constraints, and generate state vectors; a modeling module, configured to construct coupled electromagnetic, thermal, and mechanical field models based on the state vector and output a multi-physics coupling matrix; an optimization module, configured to receive the state vector and the modeling matrix and output a designed action through a reinforcement learning model; A solution module, configured to receive the modeling matrix and the design action, perform accelerated solution in the edge computing platform, and output a physical response; A feedback module is configured to receive the physical response and update the state vector, and feed the updated state vector back to the optimization module.

Images