Electromagnetic signal autonomous analysis method and product

Abstract

The invention discloses an electromagnetic signal autonomous analysis method and product, relates to the field of radio, and is used for improving the efficiency, accuracy, specialty and intelligent level of electromagnetic signal analysis. The method comprises the following steps of: performing semantic analysis on a natural language request of a user by utilizing a large language model (LLM), identifying question and answer type or analysis type user intentions, and entering a corresponding question and answer task or analysis task process according to the user intentions; searching key information corresponding to the analysis result from a pre-constructed multi-modal knowledge base by utilizing LLM according to the analysis result; replying the question and answer task according to the key information, or generating a program control command for the analysis task, and verifying the program control command; and translating the program control command passing the verification into a control instruction supported by the electromagnetic signal analyzer, and controlling the electromagnetic signal analyzer to execute an analysis task on the electromagnetic signal to obtain an analysis result. The electromagnetic signal analysis operation threshold is reduced, and the engineering efficiency is improved.

Description

Technical Field

[0001] This invention relates to the field of radio technology, and in particular to an autonomous electromagnetic signal analysis method and product. Background Technology

[0002] With the rapid development of fifth-generation mobile communication (5G) technology, the effective utilization of signal resources and the accurate analysis of wireless signals have become indispensable and crucial aspects of research and development and production. Electromagnetic signal analyzers, as high-precision electronic measuring instruments, are widely used in fields such as electromagnetic signal transmission characteristic analysis, modulation quality analysis, and consistency verification. However, the current mainstream electromagnetic signal analysis mode still heavily relies on engineers manually configuring core parameters such as center frequency, analysis bandwidth, and subcarrier spacing through the graphical user interface or programmable commands of the electromagnetic signal analyzer. This process requires a high level of professional knowledge from the operators, has low measurement efficiency, and is prone to measurement failures, false alarms, or measurement result deviations due to configuration errors or omissions when dealing with multiple parameter configuration combinations such as frequency, bandwidth, and subcarrier spacing.

[0003] In recent years, Large Language Model (LLM) technology has developed rapidly, possessing powerful semantic understanding and natural language generation capabilities, providing a new technical approach to solving the aforementioned problems. By combining LLM with Retrieval Augmented Generation (RAG) mechanisms, semantic understanding of electromagnetic signal-related technical documents, signal analysis procedures, and program control commands can be achieved. On the one hand, this supports natural language-driven intelligent question answering, quickly responding to user inquiries about analysis methods and parameter configurations; on the other hand, it can automatically parse user natural language requirements, intelligently completing core parameters and automatically generating instrument control commands, fundamentally improving human-computer interaction efficiency and the automation level of the analysis process. Other solutions attempt to achieve natural language control of instruments through large models, but these general-purpose solutions lack both a professional knowledge base and compliance verification and error correction mechanisms.

[0004] Chinese patent document CN118584195A discloses a low-cost spectrum analysis method for 5G communication systems. This method achieves high-speed data transmission through a DMA + frequency sweep module architecture and a Xilinx IP core, reducing the hardware cost of electromagnetic signal analysis. However, this solution focuses on hardware cost optimization and uses MATLAB for data analysis. It does not introduce an automated parameter configuration and command generation mechanism, still requiring manual writing of test scripts and adjustment of analysis parameters. When faced with the diverse signal types of 5G NR, the operation efficiency is low and analysis deviations are easily caused by human error.

[0005] Chinese patent document CN118227537A discloses an intelligent general-purpose measurement and control instrument based on a large model and its construction method. This method, through modular hardware design and the construction of an instrument-specific knowledge base, achieves natural language control and test workflow generation driven by a large model, and is adaptable to various types of measurement and control instruments. However, this method lacks parameter compliance verification and error correction mechanisms for measurement and control scenarios, and can only generate basic control logic. If the SCPI commands output by the large model have parameter combination conflicts or exceed the instrument's hardware capabilities, it can easily lead to equipment malfunctions or test failures, making it difficult to directly adapt to high-precision, highly constrained professional scenarios such as electromagnetic signal analysis. Summary of the Invention

[0006] The purpose of this invention is to provide an autonomous electromagnetic signal analysis method and product that addresses all or part of the problems mentioned above, thereby lowering the threshold for electromagnetic signal analysis and improving its efficiency, accuracy, professionalism, and intelligence.

[0007] The technical solution adopted in this invention is as follows: An autonomous electromagnetic signal analysis method, characterized by comprising: S1. Use a Large Language Model (LLM) to perform semantic parsing on the user's natural language request, identify the user's intent to answer questions or perform analysis, and enter the corresponding question-answering task or analysis task process according to the user's intent; wherein, the analysis task extracts and verifies the analysis parameters of the request. S2. Using LLM, retrieve key information corresponding to the analysis results from a pre-built multimodal knowledge base; answer question-and-answer tasks based on the key information, or generate program control commands for the analysis task; the program control commands include a complete chain of instrument initialization instructions → parameter configuration instructions → analysis execution instructions → result acquisition instructions; the multimodal knowledge base is built based on the electromagnetic signal analyzer operation manual; S3. For analysis tasks, verify the program control commands generated by LLM; S4. Translate the verified programmable commands into control instructions supported by the electromagnetic signal analyzer, control the electromagnetic signal analyzer to perform the analysis task of electromagnetic signals, and obtain the analysis results.

[0008] In addition, this application also provides a computer program product, including a computer program that, when executed by a processor, performs an autonomous electromagnetic signal analysis method.

[0009] In summary, due to the adoption of the above technical solution, the beneficial effects of the present invention are: The proposed autonomous electromagnetic signal analysis method leverages the powerful learning and reasoning capabilities of the LLM (Limited Language Analyzer) to generate autonomous question-and-answer or analysis commands for electromagnetic signals through natural language interaction. This eliminates the reliance on traditional methods requiring specialized personnel and lowers the entry barrier. Furthermore, a professional and comprehensive multimodal knowledge base is constructed based on the electromagnetic signal analyzer's operation manual to support LLM retrieval. Through LLM's powerful natural language understanding and retrieval capabilities, it can efficiently, accurately, and rigorously organize question-and-answer knowledge or generate programmable commands based on the multimodal knowledge base, exhibiting a high degree of professionalism. Simultaneously, the multimodal knowledge base, constructed through a two-layer structure of vector database and knowledge graph, allows LLM to further uncover potential related knowledge, improving the accuracy of responses or commands. By automatically generating programmable commands strictly according to the command formats, parameter ranges, usage conditions, and combinations supported by the electromagnetic signal analyzer, LLM demonstrates a high level of intelligence and compliance. By using the LLM's verification and reflection mechanism, the compliance and legality of program control commands are automatically verified. When verification fails, parameters are automatically corrected based on professional knowledge. This solves the pain points of incomplete or inaccurate external input information and inability to troubleshoot, thus improving the reliability of the solution. Attached Figure Description

[0010] The present invention will be described by way of example and with reference to the accompanying drawings, wherein: Figure 1 This is a flowchart of an autonomous electromagnetic signal analysis method.

[0011] Figure 2 This is a flowchart of user intent recognition and task mapping.

[0012] Figure 3 This is a flowchart of the multimodal knowledge base construction process.

[0013] Figure 4 This is a flowchart of the key information retrieval process.

[0014] Figure 5 This is a flowchart for verifying program control commands.

[0015] Figure 6 , Figure 7 This is a request-response interaction diagram for two examples of question-and-answer requests.

[0016] Figure 8 This is a request-response interaction diagram for analyzing a request example.

[0017] Figure 9 This is a flowchart of an analysis request embodiment of the autonomous electromagnetic signal analysis method. Detailed Implementation

[0018] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0019] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0020] Current mainstream electromagnetic signal analysis methods still heavily rely on engineers manually configuring core parameters such as center frequency, analysis bandwidth, and subcarrier spacing through the graphical user interface or programmable commands of the electromagnetic signal analyzer. This process demands a high level of professional knowledge from the operators, has low measurement efficiency, and is prone to measurement failures, false alarms, or measurement result deviations due to errors or omissions when dealing with multiple parameter configurations such as frequency, bandwidth, and subcarrier spacing. Although some solutions attempt to achieve natural language control of the instrument through large models, these general-purpose solutions lack both a professional knowledge base and compliance verification and error correction mechanisms, making them difficult to directly adapt to high-precision, highly constrained professional scenarios such as electromagnetic signal analysis.

[0021] To address the problems existing in the prior art, this application proposes an autonomous electromagnetic signal analysis method and product based on a large model. This method not only supports natural language interaction, intelligent parameter extraction, and automatic generation of signal analysis commands, but also has a complete knowledge base and compliance verification and error correction mechanisms, thereby comprehensively improving the efficiency, accuracy, and intelligence level of electromagnetic signal analysis.

[0022] The electromagnetic signal autonomous analysis method proposed in this application is based on LLM implementation. Leveraging the powerful understanding and reasoning capabilities of LLM, it is supported by a professional, multi-dimensional, multimodal knowledge base. Through sophisticated guidance, it achieves autonomous analysis of electromagnetic signals (including question-and-answer or analysis process).

[0023] like Figure 1 As shown, the autonomous electromagnetic signal analysis method includes the following steps: S1. The Large Language Model (LLM) is used to semantically parse the user's natural language requests, identifying question-and-answer or analysis-based user intents. Based on the user intent, the request proceeds to the corresponding question-and-answer or analysis task flow. Question-and-answer requests enter the question-and-answer task flow, while analysis-based requests enter the analysis task flow. In the analysis task flow, analysis parameters are extracted and verified from the user's natural language request. These analysis parameters are the user-defined parameters for electromagnetic signal analysis, such as center frequency and bandwidth. These parameters instruct the electromagnetic signal analyzer to ultimately perform the analysis according to these parameters. Before this, the LLM needs to generate corresponding programmable commands to instruct the electromagnetic signal analyzer to operate based on these set analysis parameters.

[0024] As an optional implementation, step S1 includes the following sub-steps: S11. Perform intent recognition and task mapping on the natural language requests input by the user to achieve question-and-answer functions or analyze automated process parsing.

[0025] Intent recognition leverages the semantic understanding capabilities of LLM (Limited Language Management) to determine whether a user's request is for knowledge-based question answering or signal analysis. The LLM's decision (or response) regarding intent recognition is thus defined as either "question answering" or "analysis." Specifically, for example... Figure 2 As shown, after the user request is given to the LLM, the LLM performs semantic modeling and context parsing on the content of the user request, and learns to identify the user's intent type (question-answering or analysis) through few-shot prompts.

[0026] For example, prompts can be designed as follows: "You are an intelligent intent recognition assistant in the field of communications. You need to identify the most relevant operational intents from the user's current input and historical conversations." Optional intents include: Q&A: When users ask questions, such as "What are the program control commands for xxx?" or "What are the steps for xxx?" Analysis: When the user wants to configure the signal and provides the parameters.

[0027] Example 1: User: What are the steps for customizing the signal configuration in the 5G NR downlink? Output: Questions and answers.

[0028] Example 2: User: I want to customize the 5G NR downlink signal configuration, specifically the parameters are... Output: Analysis.

[0029] Based on the user's input request, determine whether the user's intent is "question and answer" or "analysis".

[0030] For example, when a user enters "I want to analyze the 5G NR downlink custom configuration signal," the system can automatically determine that the request belongs to the analysis category and further map it to the corresponding analysis task flow, achieving accurate mapping from natural language to analysis task type. This intent recognition process supports the automatic differentiation between question-and-answer type (such as querying the meaning of a command function) and analysis type (such as initiating an analysis process) requests, and can call the appropriate module for subsequent processing according to the actual task.

[0031] S12. Extract the analysis parameters requested in the analysis task.

[0032] After successfully identifying the user's analytical intent, key analytical parameters are further extracted from the user's natural language request. These analytical parameters may involve multiple dimensions, such as center frequency, channel bandwidth, subcarrier spacing, modulation scheme, etc.

[0033] S13. Verify the extracted analytical parameters.

[0034] As an optional implementation method, the verification of analytical parameters includes integrity verification and compliance verification.

[0035] S131. Based on the pre-built standard parameter template, perform integrity checks on the analysis parameters and indicate any missing parameters.

[0036] The extracted analysis parameters are structured and populated into a standard parameter template. The standard parameter template is a structured template that contains the parameter items required to generate the program control command. This allows for comparison of the completeness of the analysis parameters, ensuring that all the parameter items required for command generation are complete and preventing problems such as command execution errors or parameter call failures.

[0037] For example, the parameter template for the 5G NR downlink Test Model configuration signal is as follows: "Center Frequency": Format "XGHz", "XMHz", "XHz", where X is at least 0 and at most the maximum frequency limit of the device, with the default being "2.6GHz"; The "Test Model configuration file" is in the format "NR_FR1_TM3_1a_TDD_100MHz_30kHz".

[0038] If a parameter is detected to be missing, the user will be prompted to fill in the missing parameter information in natural language, or a default value suggestion or fill-in option will be provided.

[0039] S132. Based on the preset parameter constraints, perform compliance checks on each analysis parameter and indicate non-compliant parameters.

[0040] After confirming the completeness of the analysis parameters, a validity check is performed on each parameter. This check is conducted with reference to the parameter constraints, such as the parameter value ranges and combined constraints specified in the electromagnetic signal analyzer's operation manual. Technically, this is achieved by constructing a rule template description document to describe the parameter constraints, and then verifying each analysis parameter item by item.

[0041] For example, some rule templates for custom configuration signals in the 5G NR downlink are as follows: "Channel bandwidth": format "XMHz"; If the "frequency band deployment range" is "FR1≤3GHz" or "FR1>3GHz", the channel bandwidth value range is: "XMHz, where X can be 5, 10, 15, 20, 25, 30, 40, 50, 60, 80, 90, 100". If the "Frequency Band Deployment Range" is "FR2", the channel bandwidth range is: "XMHz, where X can be 50, 100, 200, or 400".

[0042] If an invalid analysis parameter is found, the specific reason for the error will be reported in natural language, and alternative effective parameter combinations can be recommended.

[0043] The aforementioned verification mechanism enhances the robustness of automated analysis, ensuring that the final generated programmable commands can be correctly executed within the range supported by the analyzer.

[0044] S2. Using LLM, retrieve key information corresponding to the parsing results from a pre-built multimodal knowledge base; answer the question-and-answer task based on the key information, or generate program control commands for the analysis task.

[0045] In step S1, the LLM performs semantic parsing on the natural language request input by the user, thereby parsing out relevant semantics and fields, such as "5G NR", "downlink", and "configuration". Key information is then retrieved from the multimodal knowledge base based on the semantics and fields.

[0046] The multimodal knowledge base is built based on the electromagnetic signal analyzer operation manual, thus ensuring the professionalism and traceability of the knowledge information.

[0047] As an optional implementation, step S2 includes the following sub-steps: S21. Construct a multimodal knowledge base. If the multimodal knowledge base has already been constructed in advance, this sub-step can be omitted.

[0048] Multimodal knowledge bases include vector databases and knowledge graphs. As an optional implementation, such as... Figure 3 As shown, the method for constructing the multimodal knowledge base includes: S211. Preprocess the raw text data in the electromagnetic signal analyzer operation manual, including data cleaning, structure conversion and format standardization, to enhance data consistency and parsability.

[0049] The electromagnetic signal analyzer operation manual is text data. During the text data cleaning phase, firstly, to address the common non-core information noise in the manual, a combination of regular expression matching and rule engines is used to accurately remove copyright notices, duplicate headers and footers, page numbers, and meaningless placeholder text (such as "[See Appendix XX]" or "General Instructions for XXX Series Instruments," etc., which are general expressions unrelated to electromagnetic signal analysis). Simultaneously, the text character encoding is standardized to UTF-8 format, and garbled symbols, mixed full-width and half-width characters, and redundant spaces and line breaks are removed. Secondly, to address information redundancy and conflicts, text similarity algorithms (such as cosine similarity) are used to identify and merge duplicate content (such as repeated descriptions of the same SCPI command operation steps in different chapters, explanations of the same parameters, etc.). For contradictory information (such as inconsistent value ranges for the same analysis parameter in different chapters), cross-validation is performed using protocol specifications and the electromagnetic signal analyzer hardware parameter manual. Accurate information that conforms to protocol standards and matches the actual functions of the equipment is prioritized for retention, while erroneous or ambiguous statements are removed. As for structured transformation and format standardization, these are conventional methods of text preprocessing, designed to facilitate computer processing of input data.

[0050] S212. Perform text block division and vectorization on the preprocessed text data, and store the divided text blocks, the corresponding metadata of the text blocks, and the standardized semantic vectors in the vector database, and build a semantic similarity index for each standardized semantic vector.

[0051] During the text segmentation stage, there are two segmentation methods. Based on the content characteristics of the electromagnetic signal analyzer's operation manual, the text is split according to semantic relevance: For short texts with independent semantics, such as program control command descriptions and analysis parameter definitions, a single-entry independent block approach is adopted to ensure that each text block corresponds to a complete semantic unit; for long texts such as operation procedures and troubleshooting, they are split into coherent sub-text blocks based on logical connectors and thematic relevance.

[0052] In addition, metadata tags such as type, topic, and chapter are added to all text blocks to facilitate filtering and tracing during subsequent searches.

[0053] In the text vectorization stage, semantic embedding models, such as the Beijing Academy of Artificial Intelligence Generative Embeddings (BGE) model and the Moka Massive Mixed Embedding (M3E) model, are used to embed text blocks. Semantic embedding models can accurately capture the semantic relationships between technical terms and construct a stable and effective semantic vector space. Text blocks are concatenated with metadata tags and input into the model, transforming them into fixed-dimensional semantic vectors.

[0054] Finally, the standardized semantic vectors, after vectorization, are associated with their corresponding text blocks and metadata tags and stored in a vector database. A semantic similarity index is constructed for each standardized semantic vector to facilitate retrieval and matching. The vector database supports metadata filtering and Top-K semantic retrieval, forming an efficiently accessible vector database that provides data support for LLM to acquire operational knowledge through the RAG mechanism.

[0055] S213. Entity extraction, relation extraction, entity disambiguation, and pronoun resolution are performed on the preprocessed text data to construct a multi-level knowledge graph covering instrument control instructions, parameter items, and usage condition entity types.

[0056] First, LLM leverages its domain semantic understanding capabilities to perform entity extraction, automatically identifying core entities such as instrument control commands, parameter items, and usage conditions, and classifying and labeling them by type. In the relationship extraction stage, LLM analyzes the semantic connections and logical structure in the text to automatically uncover relationships such as "containment," "constraint," and "fit" between entities, generating standardized triples. For entity disambiguation and referential resolution, LLM uses contextual semantic reasoning to automatically identify different expressions of the same entity (e.g., "DL" and "downlink"), clarify the boundaries of synonymous entities (e.g., "analysis bandwidth" and "signal bandwidth"), and establish a mapping relationship between pronouns and the entities they refer to (e.g., "this parameter" corresponds to "center frequency").

[0057] The entire process requires no human intervention. Through the semantic understanding capabilities of LLM, it achieves fully automated construction from unstructured text to multi-level knowledge graphs, ensuring that the multi-level knowledge graphs can accurately cover the core operational elements of the instrument while maintaining the logical consistency and semantic coherence of entity relationships.

[0058] S22. Construct a prompt template for the question-answering or analysis task. This prompt template defines the constraints of the task.

[0059] Here is an example of a Q&A task prompt template: Based on the knowledge base of the electromagnetic signal analyzer operation manual, please answer the following user questions regarding independent electromagnetic signal analysis: {User Request}; The answer must meet the following (i.e., the task constraints): 1. Prioritize the use of technical content explicitly recorded in the electromagnetic signal analyzer's operation manual, including but not limited to instrument operation procedures, parameter value ranges, and programmable command formats; 2. Avoid using vague expressions; for expressions such as "maybe" or "probably," specify the underlying conditions. 3. If the question is outside the scope of the knowledge base, the user should be clearly informed that "the current question does not match the relevant content in the manual, and it is recommended to supplement the analysis of the scenario details".

[0060] Example reference: If the user's question is "What is the minimum center frequency value when analyzing 5G NR downlink signals", the answer should include "According to Chapter X of the manual, the minimum center frequency value for 5G NR downlink signal analysis is 100kHz, which is applicable to low-frequency signal acquisition scenarios".

[0061] This prompt template guides LLM to accurately call structured information from vector databases and multi-level knowledge graphs by clearly defining "knowledge base citation requirements, expression rigor constraints, and boundary handling rules." This ensures that the question-and-answer results are both technically accurate and scenario-appropriate. At the same time, it reduces model comprehension bias and avoids generic answers by providing example references.

[0062] The template for the analysis task prompt is as follows: The user-input electromagnetic signal autonomous analysis requirement is as follows: {User Request}, which requires the completion of the following tasks (i.e., task constraints): 1. First, identify the core analysis task type (such as "5G NR downlink custom signal modulation quality analysis" or "uplink signal power spectrum detection"), and extract key analysis parameters (including but not limited to center frequency, analysis bandwidth, subcarrier spacing, modulation method, etc.). If analysis parameters are missing, mark "Parameters to be supplemented: XXX (referring to the missing analysis parameters)". 2. By combining vector databases and multi-level knowledge graphs, generate programmable commands for the corresponding analysis process. The commands must include a complete chain of "instrument initialization command → parameter configuration command → analysis execution command → result acquisition command".

[0063] S23. Input the request and prompt template into the LLM so that the LLM can retrieve the corresponding key information for the request from the multimodal knowledge base according to the constraints indicated by the prompt template.

[0064] As can be seen from the prompt template defined above, its "{User Request}" field is a reserved location for filling in user requests. By filling the request into this field and inputting it into the LLM, the LLM can retrieve the corresponding key information for the request from the multimodal knowledge base.

[0065] LLM will retrieve corresponding key information from vector databases and multi-level knowledge graphs, respectively. As an optional implementation method, such as... Figure 4 As shown, the relevant key information is retrieved from the multimodal knowledge base for the request, including: Key information corresponding to the request is retrieved from both a vector database and a multi-level knowledge graph. Specifically, in the vector database retrieval stage, the text block associated with the standardized semantic vector that has the highest semantic similarity to the request is retrieved from the vector database. In the multi-level knowledge graph retrieval stage, at least the instrument control instructions corresponding to the requested entity are retrieved from the multi-level knowledge graph.

[0066] Specifically, the process of retrieving key information includes: S231. Retrieve key information from the vector database.

[0067] In the vector retrieval stage, the user input question is first converted into a semantic vector, and then the text block with the highest semantic similarity is matched in the vector database. For example, when the user inputs "analysis command to generate 5G NR downlink 3.5GHz signal", the vector retrieval stage will prioritize locating text blocks containing semantics such as "3.5GHz band configuration" and "downlink SCPI command", and quickly return basic information directly related to the analysis scenario, such as instrument operation instructions and parameter range constraints, to ensure that the knowledge acquired by the model is scenario-adaptable.

[0068] S232. Retrieve key information in a multi-level knowledge graph.

[0069] In the multi-level knowledge graph retrieval process, a strategy combining local retrieval mode and global retrieval mode is adopted.

[0070] For requests in a single dimension (such as specific parameter configuration, single command generation, etc.), the system retrieves parameter items, usage conditions, and instrument control instructions corresponding to the requested entity from a multi-level knowledge graph through a local retrieval mode. Based on its characteristics of "clear target length and format, avoidance of absolute expression, and support for data reference", the system generates accurate control command fragments or parameter descriptions.

[0071] For example, when a user inputs "Generate the center frequency configuration command for the 5G NR downlink 3.5GHz signal," the local retrieval mode first specifies the required response format (including the SCPI command, function description, and parameter constraints) and length range. Then, based on data obtained from a multi-level knowledge graph local retrieval, such as "SCPI command format corresponding to the 3.5GHz center frequency" and "parameter value range 100kHz-6GHz," it generates a response: "SCPI command: :CONFigure:NR:DL:FREQ 3.5GHz; Function description: Configure the center frequency for 5G NR downlink analysis; Parameter constraints: This frequency should be within the instrument's supported range of 100kHz-6GHz and must match the analysis bandwidth (a bandwidth of at least 20MHz is recommended to cover the entire signal frequency band)." Simultaneously, all command formats and parameter ranges reference explicit data from the electromagnetic signal analyzer's operation manual in the multimodal knowledge base, ensuring traceability. This mode can quickly respond to single command generation needs, providing a standardized foundational module for the subsequent construction of a complete command chain.

[0072] For requests from multiple dimensions, a global retrieval mode is used to extract multiple key nodes from the requests and sort them by importance. The parameter items, usage conditions and instrument control instructions corresponding to the entities of each key node are retrieved from the multi-level knowledge graph. The instrument control instructions are then connected according to the importance of the key nodes to obtain a complete instrument control instruction chain, and the relationships between the instrument control instructions are marked.

[0073] Specifically, for users' complex and multi-dimensional analysis needs (such as "completing the modulation quality analysis of a custom 5G NR downlink signal, including parameter configuration, command execution, and result acquisition"), the global retrieval mode integrates knowledge from multiple modules through the Map-Reduce framework to generate a complete programmable command solution covering the entire process. In the Map phase, based on the "modulation quality analysis process relationships" and "key parameters and commands at each stage" obtained through global retrieval of a multi-level knowledge graph, multiple key nodes are extracted from the request and marked with importance scores: for example, "instrument initialization (90% importance)," "center frequency configuration (95% importance)," "subcarrier spacing setting (85% importance)," and "modulation method selection (80% importance)," etc. Each key node corresponds to a specific knowledge source. In the Reduce phase, LLM simulates the perspectives of multiple "domain analysts" to integrate the knowledge. From a "parameter adaptation perspective," the compliance of the combination of "3.5GHz center frequency + 30kHz subcarrier spacing + 16QAM modulation" is verified. From a "process logic perspective," the step sequence of "initialization → parameter configuration → execution analysis → result acquisition" is confirmed. From a "scenario optimization perspective," the suggestion to "enable signal amplification in weak signal environments" is added. Finally, the instrument control commands are retrieved based on the importance of key nodes, generating a logically coherent chain of instrument control commands. The relationships and optimization suggestions between each instrument control command are also marked, ensuring that the response covers both the core process and the specificities of the scenario. This model effectively handles complex tasks with multiple stages and constraints, upgrading from single commands to a complete process solution, and providing comprehensive command support for automated analysis of electromagnetic signal analyzers.

[0074] S3. For analysis tasks, verify the programmable commands generated by LLM.

[0075] In step S3, compliance and validity checks are performed on the program control commands: S31. Perform static syntax validation on the program control commands using the preset program control command syntax rules.

[0076] Based on a pre-defined set of syntax rules for electromagnetic signal analyzer programmable commands, static syntax validation is performed on the programmable commands generated by the LLM. This set of rules is constructed according to the standard instruction protocols supported by the electromagnetic signal analyzer and is used to verify whether the structure, parameter order, command spelling, and other elements of the programmable commands conform to the analyzer specifications, ensuring that command execution will not fail or the device will respond abnormally due to syntax errors.

[0077] S32. Based on the instruction set and parameter range supported by the electromagnetic signal analyzer, verify the validity of the analysis parameters in the program control command.

[0078] Specifically, based on the valid value range or enumeration set of analysis parameters such as analysis frequency, bandwidth, and modulation method recorded in the operation manual of the electromagnetic signal analyzer, it is determined whether the analysis parameters in the program control command exceed the corresponding valid value range or whether there is a situation where they do not match the enumeration set. If so, the corresponding analysis parameters are determined to be invalid.

[0079] In addition, as a preferred implementation, such as Figure 5 As shown, when the analysis parameters of the program control command are determined to be invalid, the LLM's reflection and correction mechanism or user confirmation mechanism is automatically triggered.

[0080] If the analysis parameters are determined to be invalid, the LLM's reflection mechanism is activated. This involves backtracking and analyzing previously generated programmable commands, adjusting the analysis parameters within the commands, or reconstructing the command content, and then re-verifying the commands. Specifically, the LLM first identifies the core reason for the invalid analysis parameters. Then, combining the parameter constraints supported by the analyzer in the multimodal knowledge base with the scenario logic of electromagnetic signal analysis, it generates multiple correction schemes. If the adjustments made according to the correction schemes still do not meet the scenario requirements, the command is reconstructed. During the correction process, the LLM records the logical basis for each adjustment step to ensure the compliance and scenario adaptability of the corrected programmable commands.

[0081] If the re-verification still fails, display the failed items and the reason for the failure (in natural language). The explanation should include the specific illegal parameter, the failure type, the associated analyzer parameter standard, and a brief explanation of the technical logic behind the failure to avoid users submitting invalid requests due to unclear information.

[0082] The system provides users with optional modification suggestions for analysis parameters and receives user feedback on analysis parameters, or requests users to modify analysis parameters. This forms a closed-loop optimization process involving human-machine collaboration. The optional modification suggestions must be designed based on the user's original needs and the analyzer's capabilities to ensure operability. If the suggestions cannot fully match the user's request, the system explicitly requests the user to supplement key information, such as key analysis parameters or a clear usage scenario. After receiving user feedback, LLM re-executes the generation of program control commands and verification processes based on the updated analysis parameters.

[0083] This process continues until the programmable command passes verification, or until a specific request cannot be fulfilled, such as when the analyzer supports the designed analysis parameters or the specified use case.

[0084] S4. Translate the verified programmable commands into control instructions supported by the electromagnetic signal analyzer, control the electromagnetic signal analyzer to perform the analysis task of electromagnetic signals, and obtain the analysis results.

[0085] Taking the LLM-generated program control command as an example of an SCPI command, in one optional implementation, step S4 includes the following sub-steps: S41. Embed the SCPI command into an automated analysis script written in Python. This analysis script calls the interface of the electromagnetic signal analyzer through the PyVISA library. This automated analysis script can also be automatically generated by the LLM using a prompt template. By introducing a few examples of typical analysis procedures, signal analyzer interface structure, and command call syntax into the prompt template, the LLM can guide a large model to generate a structurally sound, logically complete, and executable analysis script based on a few-sample prompt mechanism.

[0086] S42. By executing the analysis script, the electromagnetic signal analyzer is controlled to complete the entire process of initialization configuration, parameter setting, analysis execution, and result acquisition. Measurement results may include EVM (Error Vector Magnitude), frequency offset, constellation diagram, power spectrum, etc.

[0087] S43. Generate and output measurement results such as EVM, frequency offset, constellation diagram, and power spectrum in natural language format. For graphical results, LLM will first extract key features from the graph and then associate the feature descriptions with the numerical results.

[0088] Based on the ideas of this application, embodiments of this application also provide a computer program product, including a computer program that, when executed by a processor, performs an autonomous electromagnetic signal analysis method.

[0089] like Figures 6-9 This is a schematic diagram illustrating the interaction of the method proposed in the embodiments of this application in an application with two question-answering requests and one analysis request, wherein... Figure 9 The detailed implementation process of the electromagnetic signal analysis request is illustrated. These examples demonstrate the feasibility, effectiveness, accuracy, and autonomy of the method proposed in this application.

[0090] Compared with existing technologies, this application achieves significant advantages in terms of interactivity, knowledge support accuracy, command compliance, result value transformation, and overall benefits through the synergy of four core modules: natural language parsing module, knowledge base construction and retrieval module, program control command generation and verification module, and result collection and semantic feedback module. All of these advantages address the core pain points of existing technologies.

[0091] At the interaction level, the natural language parsing module of this application relies on the dedicated intent template for electromagnetic signal analysis and few-shot learning to accurately identify professional needs such as "5G NR downlink signal analysis", automatically extract core analysis parameters and complete missing completion and anomaly screening. Non-professional users do not need to master SCPI commands or protocol terminology to operate it, which solves the problem of "heavy reliance on professional knowledge and difficulty for non-professional users to get started" in traditional solutions.

[0092] At the knowledge support level, the multimodal knowledge base construction and retrieval module uses RAG technology to automatically extract and construct structured maps of protocol and instrument manual knowledge, breaking through the limitations of existing solutions such as "knowledge lag and insufficient association mining".

[0093] At the command generation and verification level, the programmable command generation and verification module can generate a complete SCPI command chain covering "initialization-configuration-execution-collection". It forms a double layer of protection through static syntax verification and parameter validity verification. When verification fails, it can also start a large model reflection mechanism to automatically correct parameters, which solves the pain points of existing technologies such as "incomplete commands, high error rate and difficulty in troubleshooting". At the result feedback level, the result collection and semantic feedback module automatically completes the format conversion, feature extraction and professional interpretation of the raw data, and generates a report, avoiding the process fragmentation of the existing solution that "only outputs raw data and requires secondary manual analysis".

[0094] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A method for autonomous analysis of electromagnetic signals, characterized in that, include: S1. Use a Large Language Model (LLM) to perform semantic parsing on the user's natural language request, identify the user's intent to answer questions or perform analysis, and enter the corresponding question-answering task or analysis task process according to the user's intent; wherein, the analysis task extracts and verifies the analysis parameters of the request. S2. Using LLM, retrieve key information corresponding to the analysis results from a pre-built multimodal knowledge base; answer question-and-answer tasks based on the key information, or generate program control commands for the analysis task; the program control commands include a complete chain of instrument initialization instructions → parameter configuration instructions → analysis execution instructions → result acquisition instructions; the multimodal knowledge base is built based on the electromagnetic signal analyzer operation manual; S3. For analysis tasks, verify the program control commands generated by LLM; S4. Translate the verified programmable commands into control instructions supported by the electromagnetic signal analyzer, control the electromagnetic signal analyzer to perform the analysis task of electromagnetic signals, and obtain the analysis results.

2. The autonomous electromagnetic signal analysis method as described in claim 1, characterized in that, The methods for verifying the analytical parameters include: Based on a pre-built standard parameter template, the completeness of the analysis parameters is checked, and missing parameters are indicated. Based on preset parameter constraints, each analysis parameter is checked for compliance, and non-compliant parameters are flagged.

3. The autonomous electromagnetic signal analysis method as described in claim 1, characterized in that, The method for constructing the multimodal knowledge base includes: Preprocess the raw text data in the electromagnetic signal analyzer operation manual, including data cleaning, structure conversion and format standardization; The preprocessed text data is divided into text blocks and vectorized. The divided text blocks, the corresponding metadata of the text blocks, and the standardized semantic vectors are associated and stored in the vector database. A semantic similarity index is built for each standardized semantic vector. Entity extraction, relation extraction, entity disambiguation, and pronoun resolution are performed on the preprocessed text data to construct a multi-level knowledge graph covering instrument control commands, parameter items, and usage condition entity types.

4. The autonomous electromagnetic signal analysis method as described in claim 3, characterized in that, Using LLM, based on the parsing results, key information corresponding to the parsing results is retrieved from a pre-built multimodal knowledge base, including: Construct a prompt template for a question-answering task or an analysis task, wherein the prompt template defines the constraints of the task; The request and the prompt template are input into the LLM so that the LLM can retrieve the corresponding key information for the request from the multimodal knowledge base according to the constraints indicated by the prompt template.

5. The autonomous electromagnetic signal analysis method as described in claim 4, characterized in that, Retrieving relevant key information from the multimodal knowledge base for the request, including: Key information corresponding to the request is retrieved from the vector database and the multi-level knowledge graph, respectively. In the vector database retrieval stage, the text block associated with the standardized semantic vector with the highest semantic similarity to the request is retrieved from the vector database. In the multi-level knowledge graph retrieval stage, at least the instrument control instructions corresponding to the entity of the request are retrieved from the multi-level knowledge graph.

6. The autonomous electromagnetic signal analysis method as described in claim 5, characterized in that, In the multi-level knowledge graph retrieval process, a strategy combining local retrieval mode and global retrieval mode is adopted for retrieval; For a single-dimensional request, the local retrieval mode is used to retrieve the parameter items, usage conditions, and instrument control instructions corresponding to the entity in the request from the multi-level knowledge graph. For requests from multiple dimensions, the global retrieval mode is used to extract multiple key nodes from the request and sort them by importance. The parameter items, usage conditions and instrument control instructions corresponding to the entity of each key node are retrieved from the multi-level knowledge graph. The instrument control instructions are connected according to the importance sort of the key nodes to obtain a complete instrument control instruction chain, and the relationship between each instrument control instruction is marked.

7. The autonomous electromagnetic signal analysis method as described in claim 1, characterized in that, The programmable commands generated by the LLM are validated, including: Static syntax validation is performed on the program control commands using preset program control command syntax rules; Based on the instruction set and parameter range supported by the electromagnetic signal analyzer, the validity of the analysis parameters in the programmable commands is verified.

8. The autonomous electromagnetic signal analysis method as described in claim 7, characterized in that, Based on the instruction set and parameter range supported by the electromagnetic signal analyzer, the validity of the analysis parameters in the programmable commands is verified, including: Based on the legal value range or enumeration set of the analysis parameters recorded in the operation manual of the electromagnetic signal analyzer, it is determined whether the analysis parameter in the program control command exceeds the corresponding legal value range or whether there is a situation where it does not match the enumeration set. If so, the corresponding analysis parameter is determined to be invalid.

9. The autonomous electromagnetic signal analysis method as described in claim 8, characterized in that, Validating the analysis parameters in the program control command also includes: If the analysis parameters are determined to be invalid, the LLM reflection mechanism is activated to adjust the analysis parameters in the program control command, or to reconstruct the content of the program control command and then re-verify the program control command. If the re-verification still fails, the system will display the items that failed verification and the reasons for the failure. Provide users with optional modification suggestions for analysis parameters and receive user feedback on analysis parameters, or request users to modify analysis parameters; The generated program control commands and verification process are re-executed based on the updated analysis parameters; This process continues until the program control command passes verification, or until a clear request cannot be fulfilled.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it performs the electromagnetic signal autonomous analysis method as described in any one of claims 1-9.

Images