Switching power supply intelligent aided design method and system based on RAG and AI Agent

Abstract

The invention discloses a switching power supply intelligent aided design method and system based on RAG and AI Agent. Comprising the steps that S1, an RAG knowledge base is constructed, domain data collection and data preprocessing are included, and domain knowledge constraint is carried out on a large language model LLM; s2, constructing a switching power supply AI intelligent agent, and combining and constructing a complete cue word project by utilizing an input demand to cooperate with the RAG knowledge base; each step of the switching power supply AI intelligent agent interacts with the RAG, the basis of large model reasoning is enhanced, and the illusion effect of a large model can be remarkably reduced; the design result is audited and checked according to the item-by-item rules through the large model, and it is ensured that the design result meets the standard and the product design specification; the design of the switching power supply is constrained through domain knowledge and environment knowledge so as to solve the illusion influence of a large model; and outputting a BOM list compatible with purchase feasibility and cost optimization by using ERP data and component mall data.

Description

Technical Field

[0001] The present invention belongs to the technical field of auxiliary design of switching power supplies, and in particular provides an intelligent auxiliary design method and system for switching power supplies based on RAG and AI Agent. Background Art

[0002] At present, the application of large language models in the field of switching power supplies in scenarios such as design, circuit debugging, fault diagnosis and BOM quotation mostly remains at the basic knowledge question and answer stage, and no systematic solution has been formed. In addition, the response to the large model lacks further review and circuit design rule verification, and lacks the regulatory compliance inspection stage of the switching power supply products.

[0003] Currently, large models suffer from hallucinations and weak long-term reasoning capabilities, and are unable to effectively support systematic multi-dimensional reasoning for complex engineering problems in the switching power supply field, such as circuit design, circuit debugging, electromagnetic compatibility analysis, and thermal analysis. They also lack communication with external professional systems such as ERP and device malls. The BOM generated by the large model lacks ERP constraints and cannot be intelligently quoted. Therefore, it is particularly important to build a switching power supply knowledge base, establish a workflow, and enable communication between the large model and the outside world.

[0004] Moreover, during the design and commissioning of switching power supplies, engineers need to consult a large amount of literature and standards, and perform routine repetitive tasks such as calculating inductance, capacitance, and transformer parameters. They also need to solve the problems of hallucinations and weak long-term reasoning capabilities when applying large models to switching power supply design, improve design accuracy and compliance, and troubleshoot efficiency, and achieve intelligent BOM quotation. Summary of the Invention

[0005] The purpose of the present invention is to provide a method and system for intelligent auxiliary design of switching power supplies based on RAG and AI Agent to solve the problems raised in the above background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An intelligent auxiliary design method for switching power supplies based on RAG and AI Agent, including:

[0008] S1: Build the RAG knowledge base, including domain data collection and data preprocessing, to constrain the large language model (LLM) with domain knowledge.

[0009] S2: Build an AI agent for switching power supplies, use the input requirements to collaborate with the RAG knowledge base, and combine them to build a complete prompt word project.

[0010] A further technical solution is that the domain data collection includes switching power supply device manuals, design cases, papers, standards, private data and fault case domain knowledge; the domain knowledge exists in a mixed form of long text, tables, formulas, and pictures.

[0011] A further technical solution is that the data preprocessing includes efficient understanding, processing and integration of these long texts, tables, formulas, and image data through document parsing technology.

[0012] A further technical solution, said S2 includes:

[0013] S21: Fine-tune the Large Language Model (LLM); S22: Design the planning engine; S23: Build the design robot; S24: Assist in debugging the robot;

[0014] S25: Fault diagnosis robot;

[0015] S26: BOM intelligent quotation robot.

[0016] According to a further technical solution, S21 includes:

[0017] S211: Use the private R&D data accumulated in the switching power supply R&D to fine-tune the large language model (LLM) for LoRA;

[0018] S212: Re-quantify and publish the large language model (LLM) into a "switching power supply auxiliary design large model" suitable for private deployment.

[0019] As a further technical solution, the S23 includes manually planning the robot's execution tasks or planning the robot's execution tasks through an intelligent agent.

[0020] A further technical solution is that the manual planning of the robot's execution tasks includes

[0021] S231: Input design requirements;

[0022] S232: Use the RAG knowledge base to retrieve knowledge related to design requirements and combine them to build a complete prompt word project;

[0023] S233: Input the complete prompt word design requirement into the large language model LLM;

[0024] S234: Generate preliminary design results using the reasoning capabilities of the Large Language Model (LLM);

[0025] S235: Review the preliminary design results using the product design specifications, safety regulations review form, and circuit schematic review form;

[0026] S236: Use circuit schematic design, input fuse design, transformer design, chip IC design, MOSFET design, electrolytic capacitor design, varistor design, voltage regulator design, and resistor design specifications to perform intelligent rule verification on design results;

[0027] S237: Output design results.

[0028] In a further technical solution, the robot's execution task planning by the intelligent agent includes:

[0029] S231: First, clarify the role of the large model, the design content and the degree of understanding of the design content, and whether the design goals are clear;

[0030] S232: Propose specific technical conditions and clarify the understanding of the technical requirements of the large model. Through interactive guidance, break down the overall design task into small tasks. Each small task is an atomic task, meeting the requirement of no longer needing to be divided into smaller tasks.

[0031] S233: Create a task directory and define a design template for each device that needs to be designed;

[0032] S234: Specify each task, clarify the design parameters required for each device, and clarify the design requirements;

[0033] S235: Check each specific design against the design rules and product design specifications. If it does not meet the requirements, revise the design.

[0034] S236: Execute each design task in sequence and store each execution result;

[0035] S237: Review the correctness of the running results. If there are any problems, correct the running task and run it again.

[0036] S238: Integrate and summarize the operation results.

[0037] An intelligent auxiliary design system for switching power supplies based on RAG and AI Agent, including a data acquisition module, a data preprocessing module, a fine-tuning language model module, a domain constraint module, a machine design module, an audit module, and a BOM generation module;

[0038] The data acquisition module is used to collect basic data sets related to switching power supplies and collect enterprise private data; the data preprocessing module is used to build a systematic preprocessing process based on the complexity of large-scale model training data in the switching power supply field; the fine-tuning language model module is used to analyze the requirements and generate design results based on the domain constraint module; the machine design module is used to embed information into prompt words; the review module is used to review the design results generated by the fine-tuning language model; and the BOM generation module is used to output a BOM list that is compatible with procurement feasibility and cost optimization.

[0039] Beneficial effects of the present invention:

[0040] 1. This invention uses the AI ​​Agent's "planning" and "short reasoning" modes to reduce the time required to adjust prompt words and achieve precise design of switching power supply components, significantly shortening project delivery cycles.

[0041] 2. Each step of the switching power supply AI agent interacts with RAG to enhance the basis for large-model reasoning, significantly reducing the illusion effect of large models. The design results are reviewed and rule-by-rule checked through the large model to ensure that the design results comply with standards and product design specifications. RAG technology and short reasoning technology are used to solve the problems of illusion and weak long-term reasoning ability when applying large models to switching power supply design, thereby improving the accuracy and compliance of the design. The design of switching power supplies is constrained by domain knowledge and environmental knowledge, so as to solve the impact of large-model illusions on the design and improve the accuracy of the design.

[0042] 3. When generating the BOM, refer to the existing component list in ERP and try to use already used components to reduce the component approval process and the workload of design and procurement engineers. During the BOM quotation process, RPA technology is used to simultaneously complete cost accounting and quotation output using the component mall, automating the design-procurement chain and effectively improving engineer work efficiency. This also enables multi-constrained BOM quoting and intelligent quoting.

[0043] 4. Leveraging the semantic understanding of large models and the decision-making and computational capabilities of AI agents, device design parameters can be more accurately determined, leading to more scientific and rational decisions. Compared to traditional manual intervention, the AI ​​agent's computational process is not only faster but also more accurate, effectively reducing the risk of computational errors. Other features and advantages of this invention will be detailed in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 : A framework diagram of the intelligent auxiliary design of switching power supply based on RAG and AI Agent of the present invention.

[0045] Figure 2 : The switching power supply AI agent task decision execution diagram of the present invention.

[0046] Figure 3 : 5kW PFC module design flow chart of the present invention.

[0047] Figure 4 : The present invention shows a process flow diagram of a robot executing a task through intelligent agent planning.

[0048] Figure 5 : Flowchart of the auxiliary debugging robot and fault diagnosis robot of the present invention. DETAILED DESCRIPTION

[0049] The technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention.

[0050] Please refer to Figure 1-5 ;

[0051] The present invention provides a switching power supply intelligent auxiliary design method and system based on RAG and AI Agent, which can significantly improve the efficiency of switching power supply design and commissioning, thereby improving the efficiency of engineers.

[0052] The present invention is based on the reasoning ability of the large model, supplemented by RAG and AI Agent technology, so that the large language model LLM is only allowed to reference verified content in the RAG knowledge base, constrains the reference data of the large language model LLM, suppresses the illusion of the large language model LLM, and adopts a short reasoning mode to enhance the reasoning and calculation ability of the model, requiring one step to solve only one problem, ensuring the depth of reasoning and the accuracy of calculation, and auditing the design results given by the large language model LLM and checking whether they comply with product design specifications. The BOM generated by the large language model LLM is constrained by ERP existing knowledge and intelligent quotation is performed on the BOM, thereby building a knowledge base and intelligent agent system that can be used by switching power supply development engineers and researchers.

[0053] Specifically, refer to Figure 1The present invention proposes an intelligent auxiliary design method and system for switching power supplies based on RAG and AI Agent. By integrating domain knowledge such as switching power supply device manuals, design cases, papers, standards and private data, a switching power supply knowledge base is constructed to perform domain knowledge constraints, and the responses of large models are reviewed and product design rules are checked to achieve precise support for scenarios such as intelligent question and answer, design optimization, and fault diagnosis. In the process of generating the BOM table, the selection of devices is constrained by the existing knowledge contained in ERP, such as the manufacturer and price of the device, and RPA technology is used to connect with the device mall, and the quotation of the device mall is used to perform price constraints, and the quotation output is completed simultaneously. The cost output of the BOM is performed through the device manufacturer and price constraints of ERP and the price constraints of the device mall, realizing the integration of design and procurement, and significantly improving the efficiency of engineers in developing switching power supplies.

[0054] More specifically, the present invention discloses an intelligent auxiliary design method for a switching power supply based on RAG and AI Agent, comprising:

[0055] S1: Construct a RAG knowledge base, including domain data collection and data preprocessing, to impose domain knowledge constraints on the large language model (LLM); the domain data collection includes switching power supply device manuals, design cases, papers, standards, private data, and fault case domain knowledge; the domain knowledge exists in the form of a mixture of long texts, tables, formulas, and pictures, and the data preprocessing includes the efficient understanding, processing, and fusion of these long texts, tables, formulas, and picture data through document parsing technology.

[0056] S2: Build an AI agent for switching power supplies, use the input requirements to collaborate with the RAG knowledge base, and combine them to build a complete prompt word project.

[0057] Specifically, refer to Figure 2 S2 includes: S21: Fine-tuning the Large Language Model (LLM); S22: Designing a planning engine; S23: Building a design robot; S24: Designing an auxiliary debugging robot; S25: Designing a fault diagnosis robot; and S26: Designing a BOM intelligent quotation robot. The large model base requires strong reasoning capabilities. Building an intelligent agent through the large model base is crucial. The ability to build a switching power supply AI agent involves analyzing input information, autonomously thinking, and making task scheduling decisions. Through communication with the external environment and leveraging the RAG knowledge base, the robot can control and execute tasks. Furthermore, through environmental perception, the robot's execution results are reviewed to determine whether to restart or terminate the task. This review can be conducted through human-computer interaction to enhance the robot's execution effectiveness.

[0058] The intelligent agent uses the input requirements to collaborate with the RAG knowledge base to build a complete prompt word project, uses fine-tuning large models to build a planning engine, and abstracts the design, debugging, fault diagnosis, and BOM intelligent quotation robots into tools. These tools are predefined and constructed, and product design rules and circuit design rules are implanted into the intelligent agent or into the robot as judgment conditions to drive the normal operation of the robot.

[0059] In this embodiment, the more specific steps for building a switching power supply AI agent are as follows: First, the purpose of building the agent is clarified. For the switching power supply agent, engineers input design requirements, such as input voltage range, output voltage, switching frequency range, output ripple voltage requirements, maximum output power, overall efficiency requirements, as well as component design requirements, component parameter requirements, bill of materials (BOM) requirements, and industry design specifications for circuit schematics. The large model can then output the corresponding design parameters and provide an intelligent quotation. Furthermore, if engineers input commissioning requirements and fault symptoms, the agent provides solutions to assist engineers in completing circuit commissioning and fault diagnosis, significantly improving their work efficiency. Next, the large language model (LLM) is fine-tuned. Specifically, S21 includes: S211: LoRa fine-tuning the large language model (LLM) using private R&D data accumulated during switching power supply R&D; S212: quantifying and publishing the large language model (LLM) as a "switching power supply auxiliary design large model" suitable for private deployment.

[0060] Based on the open source deepseek-r1 model, the open source model is fine-tuned with LoRA using the private R&D data accumulated in the development of switching power supplies, and the large language model is re-quantified and released as the "DeepseekRD switching power supply auxiliary design large model" suitable for private deployment.

[0061] In this embodiment, S22 is used for the overall planning of the intelligent body to perform tasks, to identify the intention of the input requirements, to determine whether the input requirements are design work, debugging work, fault diagnosis work, or BOM quotation work, to make decisions on the selected robots, and to call one or more robots to perform tasks.

[0062] In this embodiment, S23 includes manually planning the robot's execution tasks or planning the robot's execution tasks through an intelligent agent. Specifically, the construction of the design robot is the first core of the entire design. For the design robot, the task can be broken down manually or through an intelligent agent, and the task can be subdivided into subtasks or steps one by one. The design robot is guided to execute step by step, perform short reasoning, ensure the confirmation and execution of each subtask or step, and the robot's execution results can be reviewed or rule-checked by the intelligent agent. Alternatively, a reflector based on design specifications and review tables can be built, and an intelligent agent reflection mechanism can be introduced to ensure the correctness and compliance of the design task.

[0063] Furthermore, the manual planning of the robot's execution tasks includes

[0064] S231: Input design requirements, such as the new product development design brief, design reliability and quality objectives, product design specifications, initial product and process special characteristics list, input voltage range, output voltage, switching frequency range, output ripple voltage requirements, maximum output power, overall efficiency requirements, component design requirements, component parameter requirements, BOM requirements, and device grade selection criteria.

[0065] S232: Use the RAG knowledge base to retrieve knowledge related to design requirements and combine them to build a complete prompt word project;

[0066] S233: Input the complete prompt word design requirement into the large language model LLM;

[0067] S234: Generate preliminary design results using the reasoning capabilities of the Large Language Model (LLM);

[0068] S235: Review the preliminary design results using the product design specifications, safety regulations review form, and circuit schematic review form;

[0069] S236: Use circuit schematic design, input fuse design, transformer design, chip IC design, MOSFET design, electrolytic capacitor design, varistor design, voltage regulator design, and resistor design specifications to perform intelligent rule verification on design results;

[0070] S237: Output design results.

[0071] In this embodiment, reference Figure 3Taking the design of a 5kW PFC module as an example, to address the long-term and weak reasoning capabilities of large models, the design process is broken down into steps, with each step completing a specific task, allowing the large model to accurately complete each step: 1. First, clarify the role and objectives of the large model, and verify whether the large model clearly understands the objectives; 2. Develop specific technical requirements and verify whether the large model clearly understands the technical requirements; 3-9. Based on the provided technical requirements, design each component step by step, clarifying the design requirements and output requirements; 10. Summarize the component design formulas and clarify the requirements for the formulas; 11. Summarize the bill of materials (BOM) and clarify the requirements for the BOM. Steps 1-9 are connected to the RAG system, and after each search enhancement and LLM feedback, the next step, steps 10-11, includes the design summary requirements, which are directly connected to the LLM. After obtaining preliminary design results, the design results are reviewed and checked against the circuit design rules item by item to obtain the final design results. The design BOM is then connected to the component mall for intelligent BOM quotation.

[0072] The more specific detailed steps are:

[0073] Step 1: As a senior engineer with over 15 years of experience in switching power supply design, you have been tasked with guiding and assisting with the design and calculation of a 5k bridgeless PFC module. Regarding my questions, please first confirm your understanding of the question. If you clearly understand the question, please respond with "Yes." If I need further clarification, please reply with "Please repeat the question." When you are ready to proceed with the detailed design guidance, please respond with "Ready." Please remain focused throughout the entire process and ensure the accuracy and rigor of all technical analysis and suggestions.

[0074] Step 2: Please carefully read and understand the technical requirements for the PFC module. Once you have confirmed that they are correct, please reply "I am ready to answer all your subsequent questions." The specific technical requirements are as follows: The input voltage range should be 90VAC to 20VAC, the output voltage should be maintained between 360VDC and 400VDC, the switching frequency should be set between 0 and 120kHz, the output ripple voltage should be controlled within ±2°C, the maximum output power should reach 560Hz, and the system efficiency should be maintained at least 96%. In addition, the input must pass a differential mode surge test of 3hV and a common mode surge test of 3kV. The PFC module should include the necessary components, including an AC filter, a surge inductor, a varistor, a gas discharge tube, a common-mode inductor, a PFC inductor, FC power devices, and a busbar electrolytic capacitor.

[0075] Step 3: Based on the input conditions provided, please design a common-mode filter inductor suitable for AC input, and provide detailed information on the following parameters: common-mode inductance, rated current, saturation current, temperature rise current, DC resistance (DCR), copper loss and iron loss, temperature rise limit, magnetic flux density, core type (such as ferrite or nano), winding method, leakage inductance control method, Litz wire diameter and turns efficiency, and bobbin structure. Please present your design using a textbook-style, step-by-step calculation process, ensuring that the symbols in each formula have detailed definitions and explanations to facilitate understanding by learners with no basic knowledge. After completing the above calculations, recommend at least five common-mode inductor models that meet automotive standards, including but not limited to domestic and foreign brands, and provide specific models, parameters, and manufacturer information. At the same time, please provide the contact information of domestic suppliers who can support customized requirements.

[0076] Step 4: Based on the input conditions you provide, please design the x capacitor and y capacitor for the AC input, specifying the following parameters: rated voltage, pulse withstand voltage rating, capacitance value, tolerance range, input ripple current capability, temperature rise limit, lifetime estimate based on lifetime model calculation, resonant whisker ratio, equivalent series resistance (ESR), etc. Please use a textbook-style, step-by-step method to guide the calculation in detail, defining and explaining the symbols in each formula to ensure that even beginners can understand. After completing the calculations, please recommend at least five x capacitor and y capacitor models that meet automotive standards, including the specific model, manufacturer information, and key parameters.

[0077] Step 5: Based on the input conditions you provided, please design a varistor for AC input and specify the following parameters: minimum and maximum varistor voltage, current carrying capacity (including single and multiple surge currents), single pulse energy (in joules), leakage current limit, and life estimate based on the number of pulses or energy accumulation model. Please use a textbook-style step-by-step derivation of the calculation process and explain the symbols of each formula in detail to ensure that even learners with zero basic knowledge can understand it. After completing the calculation, please recommend at least five varistor models that meet automotive standards, covering domestic and foreign brands, and provide specific models and manufacturer information. In addition, please attach the contact information of domestic suppliers who support customized requirements to facilitate further communication and cooperation.

[0078] Step 6: Based on the input conditions you provided, please design a gas discharge tube (GDT) for AC input protection and specify the following core parameters: DC breakdown voltage (i.e., DC discharge starting value), surge current peak, single surge energy (in joules), number of pulses (for life assessment), inter-electrode insulation resistance, etc. Please use a textbook-style step-by-step derivation calculation process and explain the symbols of each formula in detail, ensuring that their physical meaning and units are explained to facilitate understanding by zero-based learners. After completing the above calculations, please recommend at least five gas discharge tube models that meet automotive standards, and provide specific models, manufacturer information, and key parameters (such as DC breakdown voltage range and surge current capability). In addition, please provide the contact information of domestic suppliers that can support customized requirements for further consultation and cooperation.

[0079] Step 7: Based on the input conditions you provide, please design a PFC inductor for A / C conversion. Specify the following parameters: inductance, average current (continuous operating current), peak current (maximum transient current), losses (including copper and core losses), core material (e.g., ferrite or alloy powder core), Litz wire parameters (number of turns, wire diameter, and skin effect suppression requirements), winding method (single-layer or multi-layer structure), etc. Please provide a textbook-style step-by-step derivation of the calculation process and a detailed explanation of the symbols in each formula to ensure that even beginners can understand it. After completing the calculation, please recommend at least five PFC inductor models that meet automotive standards, including both domestic and international brands. Provide the specific model, manufacturer information, and key parameters. In addition, please include the contact information of domestic suppliers that can support customization requirements for further communication and collaboration.

[0080] Step 8: Based on the input conditions you provided, please design the power devices (including slow rectifier tubes and fast switch tubes) for the power factor correction (PFC) circuit, and specify the following parameters: device withstand voltage peak, average forward current, peak pulse current, conduction loss and switching loss, on-resistance / forward voltage drop, junction temperature limit, etc. Please use the textbook-style step-by-step derivation calculation process and explain the symbols of each formula in detail to ensure that even learners with zero basic knowledge can understand it. After completing the above calculations, please recommend at least five PPC power device models that meet automotive standards, covering domestic and foreign brands, and provide specific models, manufacturer information, and key parameters (such as withstand voltage, on-resistance, and package type). In addition, please attach the contact information of domestic suppliers that can support customized requirements for further communication and cooperation.

[0081] Step 9: Based on the input conditions you provide, please design the DC bus capacitor for the power factor correction (PFC) circuit. Specify the following parameters: nominal capacitance, rated voltage (considering voltage margin), ripple current RMS, allowable ripple voltage, equivalent series resistance (ESR), and total losses (including dielectric loss and B8 loss). Please follow the instructional step-by-step calculation process to ensure easy understanding. After completing the calculations, please recommend at least five automotive-standard PFC bus capacitor models, including both domestic and international brands. Provide specific models, manufacturer information, and key parameters. Please also include the contact information of domestic suppliers who can support customized designs for further communication and collaboration. Step 10: Please systematically summarize all calculation formulas involved in the FPC circuit design, including AC input X-capacitors, capacitors, common-mode filter inductors, varistors, gas discharge tube (FC) inductors, PPC power devices, fast-discharge switch losses, and PPC bus capacitors. The derivation process for each formula must be retained, and all symbols must be defined. These contents are presented in a textbook-style format, with formulas centered and symbol annotations aligned, making it easy for users with no basic knowledge to understand and reproduce the entire calculation process item by item.

[0082] Step 11: Systematically consolidate the bill of materials (BOM) from all previous design steps and present it in a table format (simulating the Excel structure). Ensure the following fields are included: serial number (as a unique identifier), component reference number, component name, component specifications, and component manufacturer. Ensure that the table contents are consistent with the calculations and recommendations above.

[0083] The above is to design a robot through fixed process steps, which requires deep human participation and lacks intelligence. The following is to expand the process steps by the intelligent agent as needed through intelligent planning of tasks to adapt to the design scenario of complex switching power supplies; further, refer to Figure 4 , the tasks to be executed by the robot through the agent planning include:

[0084] S231: First, clarify the role of the large model, the design content and the degree of understanding of the design content, and whether the design goals are clear;

[0085] S232: Propose specific technical conditions and clarify the understanding of the technical requirements of the large model. Through interactive guidance, break down the overall design task into small tasks. Each small task is an atomic task, meeting the requirement of no longer needing to be divided into smaller tasks.

[0086] S233: Create a task directory and define a design template for each device that needs to be designed;

[0087] S234: Specify each task, clarify the design parameters required for each device, and clarify the design requirements;

[0088] S235: Check each specific design against the design rules and product design specifications. If it does not meet the requirements, revise the design.

[0089] S236: Execute each design task in sequence and store each execution result;

[0090] S237: Review the correctness of the running results. If there are any problems, correct the running task and run it again.

[0091] S238: Integrate and summarize the operation results.

[0092] refer to Figure 5 The development of the auxiliary commissioning robot and the fault diagnosis robot is the second core of the entire design. Engineers input commissioning requirements or fault symptoms. Using the agent architecture, RAG retrieves relevant data about the design requirements to generate more accurate descriptions of the commissioning requirements and fault symptoms. These descriptions are then fed into the large model, which leverages its reasoning capabilities to generate solutions.

[0093] The BOM intelligent quotation robot inputs the BOM table into the device mall platform and generates intelligent quotations. Furthermore, for execution robots—design robots, auxiliary commissioning robots, fault diagnosis robots, and BOM intelligent quotation robots—clear exit conditions are set. Until these exit conditions are met, the rule engine will repeatedly call the execution robots, using the current execution results as supplementary data for the next review, ultimately completing all agreed-upon tasks. The agent then observes and verifies the validity of the task completion results. If the results do not meet expectations, it formulates a new strategy based on the current situation and re-initiates task scheduling until the task is confirmed to be completed.

[0094] The present invention also discloses an intelligent auxiliary design system for switching power supplies based on RAG and AI Agent, which includes a data acquisition module, a data preprocessing module, a fine-tuning language model module, a domain constraint module, a machine design module, an audit module and a BOM generation module;

[0095] The data acquisition module is used to collect basic data sets related to switching power supplies, such as power supply design manuals, manufacturer technical manuals, forum highlights, and papers related to switching power supplies from databases such as HowNet; collect industry standards such as safety regulations, energy efficiency certification, and EMI standards; collect enterprise product design specifications such as circuit design, device selection design, process and structure design, EMC design, safety requirements, and environmental design; and collect enterprise private data such as case circuit schematics, test data, BOM tables and other multimodal data; and circuit diagram evolution related data, such as how one circuit diagram is derived from another circuit diagram; and circuit diagram netlist data; The data preprocessing module is used to build a systematic preprocessing process to address the complexity of large-scale model training data in the switching power supply field. The specific process is as follows: first, the circuit schematic and netlist data, as well as the associated BOM table, are verified through EDA tools, and the topological evolution law of the circuit schematic is extracted through multiple derivative circuit schematic cases; second, technical manuals, papers and industry standards are integrated to build a knowledge graph; at the same time, test data is associated with the circuit diagram, netlist data and BOM table, and fault mode association is established by combining with the enterprise project case library; finally, parameter perturbation is used to enhance data diversity, providing the model with engineering reasoning capabilities such as topology optimization and defect diagnosis;

[0096] The fine-tuning language model module is used to analyze the requirements and generate design results according to the domain constraint module; the fine-tuning language model module selects a large model base with reasoning capabilities, such as the deepseek-r1 large model base, based on the actual reasoning requirements of the switching power supply design. The domain constraint module includes building a local domain knowledge base for the entire life cycle of the switching power supply, involving basic technical documents such as device manuals, design specifications, and standard documents; involving R&D process assets such as case libraries, circuit schematics and netlists, BOM tables, etc.; involving quality control data such as fault case libraries, safety review tables, and special feature lists; involving commissioning logs , test data, ERP process data and other production operation records; as well as private data such as product development and design task orders, design target reliability targets and quality targets, product design specifications, safety review forms, product and process initial special characteristics lists; by vectorizing the circuit netlist and storing it in the local knowledge base, and converting historical product pictures, videos and other unstructured data into parseable structured text, building a dynamic association between the cross-modal indexing system and the graph database, and finally relying on the RAG framework to achieve deep integration of vectorized storage, graph network, and text semantic multi-source data, forming an intelligent decision-making engine that supports semantic-level retrieval and causal reasoning;

[0097] The robot module includes the construction of a design robot, an auxiliary commissioning robot, a fault diagnosis robot, and a BOM intelligent quotation robot. The design robot is the core, and the other three robots serve the design robot and use the same domain knowledge. The robot design adopts an agent-driven architecture. Based on the RAG framework, it dynamically embeds retrieved historical product design cases, commissioning records, and other information into prompt words, and uses the large language model (LLM) to generate design parameters or diagnostic solutions. It further introduces an audit or rule verification mechanism, or a reflection mechanism, to optimize responses through audit or multiple rounds of reflection.

[0098] The audit module is used to audit the design results generated by the fine-tuning language model; specifically, it uses the circuit schematic industry design specifications, product design specifications, safety standards, and EMI standards to conduct a multi-dimensional audit of the switching power supply design scheme generated by the large model, and automatically verify the compliance of the design parameters; or build a reflector based on these specifications and standards, and introduce an intelligent agent reflection mechanism to ensure the correctness and compliance of the design task.

[0099] The BOM generation module is used to output a BOM list that is compatible with procurement feasibility and cost optimization. Specifically, during the BOM generation process, device selection is constrained based on the device data built into the ERP system, such as parameters such as supplier, purchase price, and arrival time, to generate an initial BOM list. At the same time, it connects to component mall platforms such as the LiChuang Mall platform to perform real-time price capture, delivery date warnings, and substitute material recommendations for BOM intelligent quotations. Based on preset price comparison strategies such as "lowest price for the same specification" or "delivery date priority", a BOM quotation with a price gradient is generated. At the same time, cross-verification rules between ERP and mall data, such as synchronization of inventory availability and minimum order quantity warnings, are used to avoid selection conflicts, and ultimately a BOM list that is compatible with procurement feasibility and cost optimization is output.

[0100] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above and that the invention can be embodied in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as illustrative and non-restrictive, and the scope of the invention is defined by the appended claims, not the foregoing description, and all variations within the meaning and range of equivalents of the claims are intended to be included therein. Any reference sign in a claim should not be construed as limiting the claim to which it relates.

[0101] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment can also be appropriately combined to form other implementations that can be understood by those skilled in the art.

Claims

1. An intelligent auxiliary design method for switching power supplies based on RAG and AI Agent, characterized by: include: S1: Build the RAG knowledge base, including domain data collection and data preprocessing, to constrain the large language model (LLM) with domain knowledge. S2: Build an AI agent for switching power supplies, use the input requirements to collaborate with the RAG knowledge base, and combine them to build a complete prompt word project.

2. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 1, characterized in that: The domain data collection includes switching power supply device manuals, design cases, papers, standards, private data and fault case domain knowledge; the domain knowledge exists in a mixed form of long text, tables, formulas, and pictures.

3. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 2, characterized in that: The data preprocessing includes efficient understanding, processing and integration of these long texts, tables, formulas and image data through document parsing technology.

4. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 1, characterized in that: The S2 includes: S21: Fine-tune the large language model LLM; S22: Design planning engine; S23: Build a design robot; S24: Assisted debugging of the robot; S25: Fault diagnosis robot; S26: BOM intelligent quotation robot.

5. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 4, characterized in that: The S21 includes: S211: Use the private R&D data accumulated in the switching power supply R&D to fine-tune the large language model (LLM) for LoRA; S212: Re-quantify and publish the large language model (LLM) into a "large switching power supply auxiliary design model" suitable for private deployment.

6. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to any one of claim 4, characterized in that: The S23 includes planning the robot's execution tasks manually or through an intelligent agent.

7. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 6, characterized in that: The manual planning of the robot's execution tasks includes: S231: Input design requirements; S232: Use the RAG knowledge base to retrieve knowledge related to design requirements and combine them to build a complete prompt word project; S233: Input the complete prompt word design requirement into the large language model LLM; S234: Generate preliminary design results using the reasoning capabilities of the Large Language Model (LLM); S235: Review the preliminary design results using the product design specifications, safety regulations review form, and circuit schematic review form; S236: Use circuit schematic design, input fuse design, transformer design, chip IC design, MOSFET design, electrolytic capacitor design, varistor design, voltage regulator design, and resistor design specifications to perform intelligent rule verification on design results; S237: Output design results.

8. The intelligent auxiliary design method for switching power supply based on RAG and AI Agent according to claim 6, characterized in that: The tasks to be executed by the robot through the agent planning include: S231: First, clarify the role of the large model, the design content and the degree of understanding of the design content, and whether the design goals are clear; S232: Propose specific technical conditions and clarify the understanding of the technical requirements of the large model. Through interactive guidance, break down the overall design task into small tasks. Each small task is an atomic task, meeting the requirement of no longer needing to be divided into smaller tasks. S233: Create a task directory and define a design template for each device that needs to be designed; S234: Specify each task, clarify the design parameters required for each device, and clarify the design requirements; S235: Check each specific design against the design rules and product design specifications. If it does not meet the requirements, revise the design. S236: Execute each design task in sequence and store each execution result; S237: Review the correctness of the running results. If there are any problems, correct the running task and run it again. S238: Integrate and summarize the operation results.

9. An intelligent auxiliary design system for switching power supplies based on RAG and AI Agent, characterized by: It includes data acquisition module, data preprocessing module, fine-tuning language model module, domain constraint module, machine design module, review module and BOM generation module; The data acquisition module is used to collect basic data sets related to switching power supplies and collect enterprise private data; the data preprocessing module is used to build a systematic preprocessing process based on the complexity of large model training data in the switching power supply field; The fine-tuning language model module is used to analyze the requirements and generate design results according to the domain constraint module; The machine design module is used to embed information into the prompt word; The review module is used to review the design results generated by the fine-tuning language model; The BOM generation module is used to output a BOM list that is compatible with procurement feasibility and cost optimization.

Images