CN-122019620-A - Large model component recommendation method based on EMC test spectrum curve and rectification suggestion

CN122019620ACN 122019620 ACN122019620 ACN 122019620ACN-122019620-A

Abstract

The invention relates to a large model component recommendation method based on EMC test spectrum curves and corresponding correction suggestions, which comprises the steps of obtaining EMC test spectrum curves and corresponding correction suggestions, generating the correction suggestions as prompt word engineering according to text forms, constructing a device function knowledge graph by using function descriptions, parameter lists and existing cases of the existing devices, and generating a GraphRAG knowledge base after construction. Aiming at the reliability of the recommended suggestion, adopting a DeepSeek solution, designing four types of experts of EMC and GRPO technology based on MoE technology, improving a reinforcement learning network of a rewarding mechanism and a reference strategy, judging whether an automatic output result or a person is required to provide support according to the result given by evaluation and the requirement of iteration number limitation, and giving out the recommended suggestion of the device by reinforcement learning and the help of the person so as to realize the landing application of the large model assistant function.

Inventors

WU WENMING
HE DONGMING
CHENG JINLI
QIN TINGYU
CAO YINGJING
WU JIAFEN

Assignees

深圳市比创达电子科技有限公司

Dates

Publication Date: 20260512
Application Date: 20251231

Claims (10)

1. A large model component recommendation method based on EMC test spectrum curves and corresponding rectification suggestions is characterized by comprising the following steps: Acquiring an EMC test spectrum curve and a corresponding correction proposal, wherein the correction proposal is used as a prompt word project and is generated in a text form; Constructing a device function knowledge graph by using the function description, the parameter list and the existing cases of the existing device, and generating a GraphRAG knowledge base after constructing; Taking the correction suggestion as input, and generating a device type selection recommendation suggestion by combining the device function knowledge graph; Aiming at the reliability of the recommended suggestion, adopting a DeepSeek solution, and designing four types of experts of EMC based on MoE technology, wherein the four types of experts comprise a power supply, a clock, a motor and synthesis; Based on GRPO technology, the reinforcement learning network of the rewarding mechanism and the reference strategy is improved, and through the training of reinforcement learning, a rectification scheme comprising devices is generated, so that the process from text suggestion to full-automatic and accurate recommendation of device recommendation is realized.
2. The large model component recommendation method based on EMC test spectrum curves and corresponding correction suggestions as claimed in claim 1, wherein a set of standard templates of corresponding prompt word engineering are constructed by analyzing existing problems and expertise, so that a questioning paradigm of the large model correction suggestions is formed.
3. The large model component recommendation method based on EMC test spectral curves and corresponding reform suggestions according to claim 2, wherein the reform suggestions are reform case-resolved documents originally provided in PDF format, including a reform corresponding problem analysis positioning flowchart, a reform measure logic algorithm, and detailed descriptions, the documents also include a lot of information presented in visual forms such as flowcharts and tables, for the pages, data AI preprocessing technology is used to convert the pages into JSON format, each JSON object captures information of each page of the document providing a case, and JSON data is converted into text through LLM.
4. The large model component recommendation method based on EMC test spectral curves and corresponding modification suggestions according to claim 2, wherein the modification case analysis document constructs GraphRAG a database, and realizes text block-to-modification suggestions, component entities and relationships among them. The method comprises the following specific steps: ① Converting the extracted entities and relations into a diagram element abstract; ② Organizing the graphic element abstracts into graphic communities; ③ Generating abstracts for each graph community; ④ And generating an automatic recommendation suggestion, namely generating a component recommendation suggestion through inquiring the related graph abstract.
5. The large model component recommendation method based on EMC test spectral curves and corresponding rectification suggestions according to claim 3, wherein the system generates detailed recommendation suggestions through LLM calls.
6. The large model component recommendation method based on EMC test spectral curves and corresponding rectification suggestions according to claim 3, wherein the recommendation suggestion is subjected to systematic evaluation to generate a suggestion that is sufficient and reliable, and the large model reasoning checks and the suggestion is perfect.
7. The large model component recommendation method based on EMC test spectral curves and corresponding rectification suggestions according to claim 3, wherein whether insufficient reasoning exists or not is suggested, the large model solution based on DeepSeek, according to the MoE technology, the patent designs four expert modules including power supply, clock, motor and synthesis according to different functions and equipment.
8. The large model component recommendation method based on EMC test spectral curves and corresponding rectification suggestions according to claim 3, wherein whether insufficient reasoning exists in the suggestions or not is characterized in that a reinforcement learning technology of GRPO proposed by DeepSeek is adopted, and the innovation is that a recommendation similarity mechanism based on existing EMC case data is designed, wherein rewards output by a plurality of suggestions are a weighted rewards mechanism with the highest similarity matching of the suggestions and GraphRAG.
9. The large model component recommendation method based on EMC test spectral curves and corresponding rectification suggestions according to claim 8, wherein the iterative optimization is performed by repeating the steps of claims 5-8 if the suggestions are insufficient based on the sufficiency determination result of the suggestions.
10. The large model component recommendation method based on the EMC test spectrum curve and the corresponding rectification suggestion according to claim 9, wherein when the number of iterative optimization reaches a preset threshold value K times, the optimization task is transferred to a human expert for processing, and finally, a comprehensive and accurate recommendation suggestion is output.

Description

Large model component recommendation method based on EMC test spectrum curve and rectification suggestion Technical Field The invention relates to an artificial intelligence recommendation suggestion for components used for EMC test rectification, in particular to a RAG technology and an reasoning reinforcement learning technology for generating reliability recommendation by a large model. Background The EMC (electromagnetic compatibility) test object covers basically all electronic and electric equipment, and the EMC characteristics of different equipment, especially the EMI (electromagnetic disturbance emission) characteristics, can be different due to different electromagnetic noise sources inside the equipment, and noise suppression treatment measures aiming at different noise modules can be modified by different components. Because this task requires experienced technicians to perform component matching, these tasks are largely repeated empirically, and for enterprise cost reduction and efficiency improvement, the enterprise goal is to implement with artificial intelligence technology. The invention uses large model technique and reinforcement learning technique to recommend suggestion for generating relevant accurate correction components. By using the large model technology and the reinforcement learning technology, an enterprise existing knowledge database and a related professional knowledge base are required to be constructed, and on the basis, an improved scheme is adopted to recommend components, and an inference technology in the related vertical field is required, so that accurate recommendation suggestion can be ensured to be obtained. To address these issues, the private database may employ existing general-purpose enterprise private database technology, such as general-purpose RAG technology, while the inference technology employs existing PPO, moE, and GRPO technologies. However, the current rectification scheme for EMC test spectrum curve generation technology recommends specific components, and the following problems to be solved still exist: (1) The private database is efficiently utilized. The enterprise collects many case libraries, the case libraries are not only text information, but also various parameters and types of information corresponding to components and parts, and the information of the case libraries cannot be obtained in an omnibearing way by using a common RAG. (2) Hybrid reasoning problems. In the field of EMC test rectification, the locations, shapes and pixel distributions involved in sensing are often quite different due to the kind of the device itself, the role differences of the circuit itself, etc. The problems are audited by a general expert, so that the problem of lack of knowledge in the special professional field of the expert can be generated, the real correction requirement can not be well reflected, and the recommendation proposal given has certain inaccuracy, so that the recommendation proposal of the model is seriously influenced. (3) Reasoning focuses on the problem. The reasoning given by the general large model is that the risks caused by inaccurate recommended components are not considered in the professional EMC test correction proposal. Therefore, with the common DeepSeek's GRPO reinforcement learning technique, no better generation and reasoning can be done in this particular field. If the generated recommendation is not authentic, the recommendation of the model is also seriously affected. Disclosure of Invention The invention aims to overcome the defects of the prior art and provide a large model component recommendation method based on EMC test spectrum curves and corresponding correction suggestions. The method comprises the following steps: 1. and acquiring an EMC test spectrum curve and a corresponding correction proposal, wherein the correction proposal is used as a prompt word project and is generated in a text form. 2. The device function build and generate map retrieval database (GraphRAG knowledge base) is built using the function description, parameter list and existing cases of the existing device. 3. Four classes of experts (power, clock, motor and synthesis) for EMC were designed based on MoE technology with DeepSeek solutions. 4. Based on GRPO's technique, the reinforcement learning network of rewarding mechanism and reference strategy is improved. Through training of reinforcement learning, a rectification scheme comprising devices is generated, and a process from text suggestion to full-automatic and accurate recommendation of device recommendation is realized. Compared with the prior art, the method has the advantages that the large model component recommending method is based on EMC test spectrum curves and corresponding rectifying and modifying suggestions. The method combines GraphRAG databases generated by the existing cases, and improves the problems that the existing databases of a recommendation system are efficiently ut