Large model component recommendation method based on EMC test frequency spectrum curve and rectification suggestion

Abstract

The invention relates to a large model component recommendation method based on an EMC (Electro Magnetic Compatibility) test spectrum curve and a corresponding rectification suggestion, which comprises the following steps: acquiring the EMC test spectrum curve and the corresponding rectification suggestion, taking the rectification suggestion as a cue project, generating and using function description, a parameter list and an existing case of an existing device according to a text form, and constructing a device function knowledge graph, and after construction, generating a GraphRAG knowledge base. Taking the rectification suggestion as input, and combining with the device function knowledge graph to generate a device type selection recommendation suggestion; aiming at the reliability of recommendation suggestions, a DeepSeek solution is adopted, four types of experts of EMC and a GRPO technology are designed based on an MoE technology, a reinforcement learning network of a reward mechanism and a reference strategy is improved, whether a result is automatically output or manual support needs to be provided is judged according to a result given by evaluation and the requirement of iteration frequency limitation, and the reliability of the recommendation suggestions is improved. Through recommendation of reinforcement learning and human help, device recommendation suggestions are given, and landing application of a large model assistant effect is realized.

Description

Technical Field

[0001] This invention relates to artificial intelligence recommendation suggestions for components used in EMC testing and rectification, and particularly to RAG technology for generating reliability recommendations using large models and inference reinforcement learning technology. Background Technology

[0002] EMC (Electromagnetic Compatibility) testing essentially covers all electronic and electrical equipment. However, the EMC characteristics, especially EMI (Electromagnetic Emission) characteristics, of different devices vary due to their different internal electromagnetic noise sources. Different components are used to address noise suppression in different noise modules. Since this task requires experienced technicians for component matching, it is largely a repetitive task based on experience. To reduce costs and increase efficiency, companies aim to implement this using artificial intelligence technology. This invention utilizes large-scale modeling and reinforcement learning techniques to generate accurate component recommendations for relevant rectification.

[0003] Employing large-scale modeling and reinforcement learning techniques requires building an existing enterprise knowledge database and related professional knowledge base. Based on this, improvements to component recommendations require inference technology within relevant vertical domains to ensure accurate recommendations. To address these issues, the private database can utilize existing, commonly used enterprise private database technologies, such as general RAG technology, while the inference technology employs existing PPO, MoE, and GRPO technologies. However, current methods for recommending specific components based on improvements to EMC test spectrum curve generation still face the following pressing problems: (1) Problem of efficient use of private databases. Enterprises collect a lot of case libraries. These case libraries are not only text information, but also include various parameters and selection information of components. The usual RAG cannot solve the problem of providing comprehensive information from case libraries.

[0004] (2) Hybrid Reasoning Problem. In the field of EMC testing and rectification, the location, shape, and pixel distribution of sensors often vary greatly due to differences in the type of equipment and the role of the circuit itself. Having a general expert review these issues can lead to a lack of specialized knowledge from the expert, resulting in inaccurate recommendations that fail to reflect the actual rectification needs and severely impact the model's recommendations.

[0005] (3) Inference Focus Issues. Because the inference provided by general-purpose models does not consider the risks associated with professional EMC testing and rectification recommendations, or inaccurate component recommendations, the general DeepSeek GRPO reinforcement learning technique cannot provide better generation and inference in this specific domain. Inaccurate generated recommendations will also severely impact the model's recommendations. Summary of the Invention

[0006] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions. This method includes the following steps: 1. Obtain the EMC test spectrum curve and corresponding rectification suggestions. These rectification suggestions are generated as prompt words in text format.

[0007] 2. Using existing device functional descriptions, parameter lists, and existing cases, construct a device functional construction and generation graph retrieval database (GraphRAG knowledge base).

[0008] 3. Using DeepSeek's solution, four types of EMC experts (power supply, clock, motor, and synthesis) were designed based on MoE technology.

[0009] 4. Based on GRPO technology, the reinforcement learning network with improved reward mechanism and reference strategy is implemented. Through reinforcement learning training, rectification schemes including device recommendations are generated, achieving a fully automated and more accurate recommendation process from text suggestions to device recommendations.

[0010] Compared with existing technologies, the advantages of this invention lie in a large-model component recommendation method based on EMC test spectrum curves and corresponding rectification suggestions. This method combines a GraphRAG database generated from existing cases and improves the efficiency of existing database utilization, addresses the issues of weak expert expertise in rectification models, and the lack of consideration for EMC specificities in the inference process through the design of a multi-expert channel specializing in EMC and a professional reward mechanism in reinforcement learning. Compared with traditional large-model techniques, this invention, in integrating the given case database, employs GraphRAG technology to increase the accuracy of database utilization, introduces DeepSeek's multi-expert technology (MoE) to solve the problem of weak model expert expertise, and designs an EMC inference reward mechanism, enabling DeepSeek's GRPO technology to perform inference more vertically within this field.

[0011] Other features and advantages of the invention will become clear from the following detailed description of exemplary embodiments of the invention with reference to the accompanying drawings. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments of the invention and, together with their description, serve to explain the principles of the invention.

[0013] Figure 1 Overall project implementation plan; Figure 2This is a diagram illustrating the framework for constructing a graph vector database (GraphRAG) based on existing data examples according to an embodiment of the present invention. Figure 3 For example, a diagram showing the relationship between the rectification plan and the corresponding components; Figure 4 Flowcharts for the four expert channels; Figure 5 A schematic diagram of the GRPO algorithm; its patented innovation lies in its reward mechanism. r i . Detailed Implementation

[0014] Various exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings. It should be noted that, unless otherwise specifically stated, the relative arrangement, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0015] The following description of at least one exemplary embodiment is merely illustrative and is in no way intended to limit the invention or its application or use.

[0016] Techniques, methods, and equipment known to those skilled in the art may not be discussed in detail, but where appropriate, such techniques, methods, and equipment should be considered part of the specification.

[0017] In all the examples shown and discussed herein, any specific values ​​should be interpreted as merely exemplary and not as limitations. Therefore, other examples of exemplary embodiments may have different values.

[0018] It should be noted that similar labels and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be discussed further in subsequent figures.

[0019] See Figure 1 As shown, a novel data framework combining artificial intelligence big model technology aims to provide recommendations for the final selection of components based on the rectification plan given after EMC inspection and the existing rectification and matching component database, using graph retrieval technology and big model technology.

[0020] Combination Figure 1 It can be seen that the provided generation methods mainly include: Step S110: The obtained rectification cases are initially provided in PDF format, containing flowcharts of the corresponding problem analysis, logical algorithms for rectification measures, and detailed explanations. These documents also contain a lot of information presented in visual forms such as flowcharts and tables. For these pages, data AI preprocessing technology is used to convert this page information into JSON format. Each JSON object captures the information from each page of the document providing the case. The JSON data is then converted into text using LLM. At this point, the basic language for the prompt word project is constructed.

[0021] Step S120 involves constructing the GraphRAG database, which translates text blocks into rectification suggestions, component entities, and the relationships between them. The specific steps are as follows: 1. Convert the extracted entities and relations into a graph element summary.

[0022] 2. Graph element summaries to graph communities: Organize graph element summaries into graph communities.

[0023] 3. Graph Community to Community Summary: Generate a summary for each graph community.

[0024] Step S130 involves querying the database through steps S110 to S120 to generate a query-related graph summary and component recommendation suggestions. This suggestion uses a general large model to generate more complete component recommendations. The system evaluates these suggestions to determine if they meet the user's requirements; this evaluation employs SFT technology. If the evaluation meets the requirements, proceed to step S150 to directly generate component recommendation suggestions; otherwise, proceed to steps S1401 to S1402.

[0025] Step S1401, based on DeepSeek's large model solution and MoE technology, this patent designs four expert modules—power supply, clock, motor, and synthesis—depending on function and device. (Among them...) Figure 4 ): Power Supply: This section primarily addresses issues related to power supply modules and provides recommendations for the selection of corresponding components. Clock: This section primarily addresses issues related to clock-related modules and provides recommendations for the selection of corresponding components. Motors: This section primarily addresses issues related to motor-related modules and provides recommendations for the selection of corresponding components. In summary: similar to a general practitioner, it provides general solutions and component selection recommendations; This step provides new recommendations; Step S1402 employs the reinforcement learning technique GRPO proposed by DeepSeek. The innovation lies in our proprietary EMC establishment mechanism: the reward for multiple suggestion outputs is a weighted reward mechanism based on the suggestion's highest match with GraphRAG.

[0026] GRPO based on deepSeek Figure 5 ), the r of this patent i The method is: The reward function obtained in the i-th iteration is: r i =cos(GraphRAG, S1401's suggestion) If the suggestion is not satisfactory, the system will perform iterative optimization, repeating the process from S140 to S1401 to S1402.

[0027] The above loop count is set to an upper limit of K. If the requirement is met within K loops, the system will provide accurate component recommendations. If the K iterations are still unsatisfactory, the system requests human experts to supplement and improve the system.

Claims

1. A method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions, characterized in that, include: Obtain the EMC test spectrum curve and corresponding rectification suggestions. These rectification suggestions are generated as prompt words in text format. Using the functional specifications, parameter lists, and existing cases of existing devices, construct a device functional knowledge graph, and then generate the GraphRAG knowledge base; The rectification suggestions are used as input, and combined with the device functional knowledge graph, device selection recommendations are generated. To ensure the reliability of the recommendations, DeepSeek's solution is adopted, which designs four types of EMC experts based on MoE technology. These four types of experts include power supply, clock, motor, and synthesis. Based on GRPO technology, the reinforcement learning network with improved reward mechanism and reference strategy is used to generate rectification schemes including devices through reinforcement learning training, realizing a fully automatic and more accurate recommendation process from text suggestions to device recommendations.

2. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 1, characterized in that, By analyzing existing problems and drawing on professional knowledge, a set of standard templates for prompt word engineering was constructed, thus forming a questioning paradigm for large-scale model rectification suggestions.

3. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 2, characterized in that, The rectification suggestions were initially provided as a PDF document containing a flowchart of the problem analysis and location, the logic algorithm of the rectification measures, and detailed explanations. These documents also contained a lot of information presented in visual forms such as flowcharts and tables. For these pages, data AI preprocessing technology was used to convert these pages into JSON format. Each JSON object captured the information of each page of the document providing the case, and the JSON data was converted into text through LLM.

4. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 2, characterized in that, The rectification case analysis document uses a GraphRAG database to implement the transformation of text blocks into rectification suggestions, component entities, and the relationships between them. The specific steps are as follows: ① Convert the extracted entities and relations into graph element summaries; ② Graph element summaries to graph communities: Organize graph element summaries into graph communities; ③ Graph community to community summary: Generate a summary for each graph community; ④ Automatic recommendation generation: Component recommendation suggestions are generated by querying relevant graph summaries.

5. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 3, characterized in that, The system generates detailed recommendations through LLM calls.

6. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 3, characterized in that, The system evaluates the adequacy and reliability of the generated recommendations, and performs large-scale model reasoning checks and recommendations improvement.

7. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 3, characterized in that, The patent suggests the existence of insufficient reasoning, and based on DeepSeek's large model solution, according to MoE technology, this patent designs four expert modules including power supply, clock, motor and synthesis, depending on the function and device.

8. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 3, characterized in that, To determine if there is insufficient reasoning, we adopted the reinforcement learning technique GRPO proposed by DeepSeek. The innovation lies in our design of a recommendation similarity mechanism based on existing EMC case data. In this mechanism, the reward for multiple suggestion outputs is a weighted reward based on the suggestion with the highest similarity to GraphRAG.

9. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 8, characterized in that, Based on the sufficiency assessment of the recommendations, if the recommendations are deemed insufficient, the steps of claims 5-8 are repeated for iterative optimization.

10. The method for recommending large-scale components based on EMC test spectrum curves and corresponding rectification suggestions as described in claim 9, characterized in that, When the number of iterations reaches a preset threshold K, the optimization task is transferred to human experts for processing, ultimately outputting comprehensive and accurate recommendations.

Images