Multi-agent cooperation method and device for electromagnetic spectrum monitoring and analysis

Abstract

The invention provides a multi-agent cooperation method and device for electromagnetic spectrum monitoring and analysis, and belongs to the technical field of artificial intelligence. The method comprises the following steps: receiving a task demand, and obtaining interface information of an electromagnetic detection tool model; analyzing the task demand through the first agent to generate a preliminary task intention; generating a task planning book containing a plurality of subtasks based on the preliminary task intention and the interface information; calling an electromagnetic detection tool model to execute a subtask through a task scheduling module of the second agent, and obtaining intermediate result information and final label information; and based on the intermediate result information and the final label information, performing feature fusion and intention inference through a second agent, and generating an intention inference result of the electromagnetic signal and an evidence chain supporting the intention inference result. Based on the method, the invention also provides a multi-agent cooperation device for electromagnetic spectrum monitoring and analysis. According to the invention, the recall rate and robustness of spectral analysis are improved, and the detection rate of anomaly detection is improved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to a multi-agent collaborative method and device for electromagnetic spectrum monitoring and analysis. Background Technology

[0002] With breakthroughs achieved by large-scale language models in natural language understanding and generalization of complex tasks, LLM-based autonomous intelligent agent (Agentic AI) technology has become an emerging direction in the field of artificial intelligence. The core characteristic of this type of intelligent agent is its ability to independently complete complex multi-step tasks through autonomous perception, reasoning, planning, and execution. Its core technical challenges focus on how to accurately parse and structure high-level user instructions into executable sub-task sequences, and how to perform reliable final intent reasoning and decision-making based on heterogeneous, multimodal intermediate results. Electromagnetic spectrum monitoring technology has wide-ranging applications in the public safety field. Typical uses include, but are not limited to, monitoring unauthorized drone flights, spectrum compliance supervision, security monitoring of public events and large venues, electromagnetic safety protection of critical infrastructure, and emergency communication support. With the widespread adoption of wireless communication, video transmission, the Internet of Things, and unmanned systems in urban and industrial environments, the electromagnetic environment exhibits characteristics such as multi-source, multi-protocol, multi-bandwidth, and non-steady-state characteristics, posing significant challenges to traditional monitoring and identification methods.

[0003] In general domains, LLM-based autonomous agents demonstrate strong planning capabilities in tasks such as dialogue and programming. However, when applied to highly specialized industrial and technical fields, such as electromagnetic spectrum monitoring, they face the following technical bottlenecks: First, lack of domain knowledge: General-purpose LLMs lack expertise in the electromagnetic signal domain (such as modulation classification and protocol identification), making it difficult to understand technical terminology and generate task plans that conform to domain logic. Second, insufficient interpretability of reasoning: Traditional end-to-end detection models are "black boxes," lacking transparency in their decision-making process and failing to provide a convincing chain of reasoning evidence, which is a fatal flaw in safety-critical public safety monitoring scenarios. Third, static task flow: Existing electromagnetic monitoring systems mostly adopt preset, static analysis processes, unable to adjust and adaptively plan in real time according to dynamically changing electromagnetic environments and diverse user needs, resulting in poor flexibility.

[0004] Therefore, there is an urgent need in this field for an intelligent processing method that can deeply integrate domain knowledge, possess dynamic task planning, interpretable intention reasoning, and autonomous reflection and optimization capabilities. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention proposes a multi-agent collaborative method and device for electromagnetic spectrum monitoring and analysis, which significantly improves the recall and robustness of spectrum analysis, enhances the anomaly detection rate, and reduces the false alarm rate.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: A multi-agent cooperative method for electromagnetic spectrum monitoring and analysis includes the following steps: Receive task requirements for electromagnetic spectrum monitoring and analysis, and obtain interface information for electromagnetic detection tool models; The task requirements are analyzed by a pre-defined first intelligent agent to generate a preliminary task intent; and a task plan containing multiple sub-tasks is generated based on the preliminary task intent and the interface information. The task scheduling module of the second intelligent agent calls the electromagnetic detection tool model to execute the sub-tasks of the task plan and obtains the intermediate result information and final label information generated by the electromagnetic detection tool model. Based on the intermediate result information and the final label information, feature fusion and intent reasoning are performed by a preset second intelligent agent to generate the intent inference result of the electromagnetic signal and the evidence chain supporting the intent inference result.

[0007] The present invention also proposes a multi-agent cooperative device for electromagnetic spectrum monitoring and analysis, comprising at least one processor and a memory, wherein the memory stores a computer program, and the computer program, when executed by the at least one processor, implements the aforementioned multi-agent cooperative method for electromagnetic spectrum monitoring and analysis.

[0008] The effects described in the invention are merely those of the embodiments, and not all the effects of the invention. One of the above technical solutions has the following advantages or beneficial effects: This invention proposes a multi-agent collaborative method and device for electromagnetic spectrum monitoring and analysis, belonging to the field of artificial intelligence technology. The method includes the following steps: receiving a task requirement for electromagnetic spectrum monitoring and analysis, and obtaining interface information of an electromagnetic detection tool model; parsing the task requirement through a pre-set first agent to generate a preliminary task intent; and generating a task plan containing multiple sub-tasks based on the preliminary task intent and the interface information; calling the electromagnetic detection tool model to execute the sub-tasks of the task plan through a task scheduling module of a second agent, and obtaining intermediate result information and final label information generated by the electromagnetic detection tool model; and performing feature fusion and intent reasoning through a pre-set second agent based on the intermediate result information and final label information to generate an intent inference result for the electromagnetic signal and an evidence chain supporting the intent inference result. Based on this multi-agent collaborative method for electromagnetic spectrum monitoring and analysis, a multi-agent collaborative device for electromagnetic spectrum monitoring and analysis is also proposed. This invention overcomes the performance degradation problem of a single end-to-end model when facing multi-source, multi-format, and non-stationary signals by decomposing the overall task into multiple sub-tasks such as perception, classification, feature fusion, and anomaly detection, and processing them in parallel or serial by an encapsulated second intelligent agent. This significantly improves the recall and robustness of spectrum analysis, increases the anomaly detection rate, and reduces the false alarm rate.

[0009] This invention utilizes a second intelligent agent to perform semantic fusion and causal reasoning on the results of multi-tool models, enabling the output of high-level intent inference conclusions and evidence chains. This elevates the analysis level from the physical and signaling layers to the cognitive and decision-making layers, providing regulatory personnel with high-value intelligence that can be directly used for decision-making.

[0010] This invention, through a reflective mechanism, learns from the results of each execution and automatically optimizes subsequent task planning strategies. This allows the system to continuously evolve, becoming increasingly intelligent with use, and significantly improving its robustness in unfamiliar and complex electromagnetic environments.

[0011] In this invention, the two intelligent agents are functionally decoupled and interact through a standardized interface. The new electromagnetic detection tool model only needs to register its interface information to be recognized and invoked by the system. This "plug-and-play" architecture enables the system to quickly integrate the latest analysis algorithms, easily expand to address new threats, and significantly reduce system maintenance and upgrade costs. Attached Figure Description

[0012] Figure 1 This is a flowchart of a multi-agent cooperative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of a multi-agent cooperative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of the present invention; Figure 3 This is the user interface of the electromagnetic signal detection platform during the detection of unauthorized drone flights in Embodiment 1 of the present invention. Figure 4 This is a schematic diagram of a multi-agent collaborative device for electromagnetic spectrum monitoring and analysis proposed in Embodiment 2 of the present invention. Detailed Implementation

[0013] To clearly illustrate the technical features of this solution, the invention will be described in detail below through specific embodiments and in conjunction with the accompanying drawings. The following disclosure provides many different embodiments or examples for implementing different structures of the invention. To simplify the disclosure of the invention, components and arrangements of specific examples are described below. Furthermore, reference numerals and / or letters may be repeated in different examples. This repetition is for simplification and clarity and does not in itself indicate a relationship between the various embodiments and / or arrangements discussed. It should be noted that the components illustrated in the drawings are not necessarily drawn to scale. Descriptions of well-known components, processing techniques, and processes are omitted in this invention to avoid unnecessarily limiting the invention.

[0014] Example 1 Embodiment 1 of this invention proposes a multi-agent collaborative method for electromagnetic spectrum monitoring and analysis, which aims to solve the technical problems existing in the prior art of electromagnetic spectrum monitoring and analysis.

[0015] Figure 1 This is a flowchart of a multi-agent cooperative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of the present invention; Figure 2 This is an architecture diagram of a multi-agent cooperative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of the present invention; combined with Figure 1 and Figure 2 The execution process is described.

[0016] In step S100, a user interface is set up at the front end, and an intelligent agent system is set up at the back end. The user interface includes: 1. A user input dialog box, where the user inputs the electromagnetic spectrum signal task background, the set of available tool models, and analysis requirements; 2. A tool model set display box, where users can upload or replace specific tool models at the system back end, allowing users and the first intelligent agent to understand the types and number of available models; 3. A task sheet display box, displaying the task sheet generated by the first intelligent agent, the operation execution flow and steps of the second intelligent agent, and providing a visual representation of the electromagnetic spectrum signal analysis process; 4. An intent inference result display box, displaying the intent inference result and the reference task sheet for executing the task. This display box requires a user interaction dialog box, allowing users to evaluate the correctness and error of the intent inference result and the correctness and error of the task sheet process referenced by the result. This information is used by the intent inference reflection module within the system to provide feedback and update the internal parameters of the first intelligent agent, thereby better completing subsequent tasks.

[0017] The intelligent agent system in this application includes a first intelligent agent and a second intelligent agent; The method for obtaining the first intelligent agent is as follows: obtain the first basic large language model; use the first training dataset to perform supervised fine-tuning on the first basic large language model to obtain the first intelligent agent; wherein, the first training dataset contains multiple task requirement texts in the field of electromagnetic monitoring, and the structured task planning document corresponding to each task requirement text; The method for obtaining the second intelligent agent is as follows: obtain a second basic large language model; use a second training dataset to perform supervised fine-tuning on the second basic large language model to obtain the second intelligent agent; wherein, the second training dataset contains multiple electromagnetic signal feature data, as well as the intent inference conclusion and evidence chain description corresponding to each feature data.

[0018] In this application, both the first and second agents are essentially Large Language Models (LLMs), which possess powerful capabilities for handling open-ended language tasks. However, for the fields of autonomous analysis and cognitive adversarial processing in electromagnetic spectrum analysis, LLMs still have shortcomings in areas such as domain terminology and abbreviation recognition. Therefore, before deploying the first and second agents, it is necessary to fine-tune them using specific datasets.

[0019] In step S110, the overall task requirements in text form are received through the user interface, and the interface information of the electromagnetic detection tool model is obtained.

[0020] First, electromagnetic measurement equipment is deployed in the area to be monitored to collect raw data streams, which are then normalized. The raw data streams include I / Q streams, power spectrum (PSD), short-time Fourier transform (STFT) frames, radar echoes, synchronous video frames, and acoustic data. Then, the raw data streams are normalized to tensor input types usable by the numerous electromagnetic detection tool models deployed within the system.

[0021] The user provides a textual description of their needs for inferring the intent of the captured signal, tailored to different scenarios and time periods. The interface information includes the functionality of each tool model and the required input / output data types.

[0022] In step S120, the task requirements are parsed by a preset first intelligent agent to generate a preliminary task intent; and based on the preliminary task intent and the interface information, a task plan containing multiple sub-tasks is generated; the specific steps include: The task requirements are semantically understood to generate a preliminary task intent; the preliminary task intent includes the task output target and execution steps. Based on the initial task intent and in conjunction with the functions described in the interface information, task decomposition and tool matching are performed. The matching results are iteratively optimized to generate a structured task plan, which defines the task type, input data source, electromagnetic detection tool model to be called, execution priority, and dependencies between subtasks for each subtask.

[0023] In step S130, the task scheduling module of the second intelligent agent calls the electromagnetic detection tool model to execute the sub-tasks of the task plan, and obtains the intermediate result information and final label information generated by the electromagnetic detection tool model; the specific steps include: According to the dynamic task scheduling sequence, the normalized electromagnetic data is bound and invoked with the electromagnetic detection tool model required by the sub-task; the electromagnetic detection tool model includes an automatic modulation classification model, a protocol identification model, a tracking and positioning model, a noise filtering model, and a signal enhancement model; the electromagnetic detection tool model is also the set of sub-intelligent agents in this application.

[0024] Monitor the execution status of each electromagnetic detection tool model and collect its output intermediate result information and final label information; wherein, the intermediate result information includes text descriptions or intermediate features in the reasoning process of the electromagnetic detection tool model, and the final label information includes the classification, identification or location conclusions drawn by the electromagnetic detection tool model after reasoning on the input data.

[0025] In step S140, based on the intermediate result information and the final label information, feature fusion and intent reasoning are performed by a preset second intelligent agent to generate an intent inference result of the electromagnetic signal and an evidence chain supporting the intent inference result; the specific steps include: In the computational graph of the second agent performing intent reasoning, the intent reasoning result is used as the starting point to backpropagate and calculate the gradient of its output layer with respect to the input layer. Extract the gradient value on the input node corresponding to the final label information, and use the absolute value of the gradient value as a quantitative indicator of the contribution of the corresponding final label information to the intent inference result; Based on the quantitative indicators, all final label information is sorted, and a preset number of evidence with the highest contribution is selected.

[0026] In step S150, feedback information regarding the intent inference result is received; based on the intent inference result and / or feedback information, the first intelligent agent reflects on the process of generating the task plan; based on the reflection result, the internal parameters of the first intelligent agent are updated.

[0027] The process of generating the task plan is reflected by the first intelligent agent, specifically as follows: based on the intent inference result, the contribution of each final tag information to the intent inference result is determined; evidence is screened based on the contribution; the causal relationship between the evidence and the intent inference result is analyzed, and the task planning logic is adjusted based on the analysis result.

[0028] Based on the intent inference result, the contribution of each final label information to the intent inference result is determined. Specifically, in the computation graph of the second agent performing intent reasoning, the intent inference result is used as the starting point, and the gradient of its output layer to the input layer is calculated by backpropagation; the gradient value on the input node corresponding to the final label information is extracted, and the absolute value of the gradient value is used as a quantitative indicator of the contribution of the corresponding final label information to the intent inference result; based on the quantitative indicator, all final label information is sorted, and a preset number of evidence with the highest contribution is selected.

[0029] In this application, the collaboration between the first and second intelligent agents is achieved through the evolution of the planning strategy of the first intelligent agent, the evolution of the intent knowledge of the second intelligent agent, and the evolution of the toolset of the task scheduling module of the second intelligent agent.

[0030] The multi-agent cooperative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of this invention can be applied to different scenarios. The following describes the situational awareness process of this invention in a field scenario: Broadband field monitoring refers to short-term or continuous scanning of a wide bandwidth spectrum in open or semi-open field environments. The purpose is to quickly detect and identify unconventional emission events, infer their intentions for timely intervention, and analyze the underlying causes.

[0031] Firstly, in selecting tool models, models more suitable for situational awareness tasks in outdoor scenarios should be chosen. For example, automatic modulation classification models can quickly identify modulation types such as ASK / FSK / QAM and are noise-resistant. For protocol identification models, a dedicated outdoor wireless protocol identification model should be selected to quickly match unconventional protocols and exclude legitimate civilian signals. For tracking and positioning models, a portable TDOA positioning model and a signal strength triangulation model should be selected for low-power, fast source location, adaptable to outdoor terrain. Additionally, noise filtering and signal enhancement models should be included to handle interference from complex electromagnetic environments in the field and improve signal quality.

[0032] In the fine-tuning step of the first intelligent agent, the dataset used for fine-tuning should contain more than 100 texts of task requirements for field scenarios and corresponding task book labels, as shown in the following example: Requirement text: "Unknown high-frequency signals were detected during scanning in a forest reserve, suspected to be from an illegal mapping drone. A rapid analysis of its intent is required." Task label: {"plan": [{"step":1, "description":"Filters ambient noise and enhances target signals", "tool":"Noise Filtering Model", "args":{"input_data":"Raw I / Q streams and PSD data from the field"}}, {"step":2, "description":"Identify signal modulation type", "tool":"Lightweight CNN modulation classification model", "args":{"items":"Signals processed in step 1"}}, {"step":3, "description":"Matches drone-specific protocols", "tool":"Field Drone Protocol Recognition Model", "args":{"modulation_type":"Result of Step 2"}}, {"step":4, "description":"Location of signal transmitter", "tool":"TDOA positioning model", "args":{"signal_features":"Results of Step 3"}}]} In the fine-tuning step for the second agent, the dataset required for fine-tuning contains over 500 field scenarios, namely paired data of "tool model output labels and intent results". Example: Input (tool model labels): {"Modulation type": "QAM-16", "Protocol type": "a certain radar's proprietary protocol", "Location result": "core area of ​​a forest reserve", "Signal strength": "medium persistence"}. Output (intent result): "The transmission intent is illegal mapping, with a medium-high risk level. It is recommended to immediately dispatch a patrol team to the location area for intervention." The fine-tuning process adopts a LoRA lightweight fine-tuning scheme, freezing the core parameters of the LLM, training only the electromagnetic spectrum domain-related layers, and controlling the training time to within 8 hours to adapt to rapid deployment requirements.

[0033] This application involves deploying portable electromagnetic measurement equipment in the field area to be monitored during data acquisition and demand deployment, specifically including: Core equipment: broadband spectrum analyzer (bandwidth ≥ 200MHz, supports continuous scanning), portable I / Q data acquisition unit, solar power module (for outdoor scenarios without mains power).

[0034] Auxiliary equipment: Environmental sensors (monitoring temperature, humidity, and wind speed for subsequent signal interference analysis). 4G communication module, satellite communication module (remote data transmission backup).

[0035] The data acquisition includes I / Q streams, power spectrum (PSD), short-time Fourier transform (STFT) frames, radar echoes (for UAV / aircraft signals), synchronized video frames (monitoring footage of the deployment area to assist in intent verification), and acoustic data.

[0036] The raw data is normalized to a tensor type with a unified dimension of [32, 32, 1] to adapt to the input requirements of various electromagnetic detection tool models. The raw data and normalized data are synchronously stored on an edge server and retained for 72 hours to meet the needs of field mission retrospective analysis. Users submit their overall task requirements in text format through the system, which must include three core elements (example): "Monitoring area: Section A of the western mountainous area; Task objective: Quickly detect unconventional UAV signals; Intervention requirements: Output the intended results and intervention suggestions within 1 hour." Users upload the interface information of currently available field electromagnetic detection tool models through the system, with an example format: {"tool_name":"Lightweight CNN Modulation Classification Model", "function":"Identify the modulation type of anti-interference signals in the field", "input_type":"tensor[32,32,1]", "output_type":"Text (Modulation Type + Confidence Level)", "response_time":"≤0.5 seconds"} The first agent captures user request text and tool model information, and generates preliminary task intent through rapid semantic understanding using a fine-tuned LLM: Task Output Objective: Complete the signal identification, localization, and intent inference of unconventional UAVs in Section A of the western mountainous area within 1 hour. Brief Execution Steps: Signal Filtering → Modulation Classification → Protocol Identification → Localization → Intent Inference.

[0037] The first intelligent agent iteratively optimizes the initial intent, generating a sequence of subtasks (task list) in JSON format, highlighting the priority, timeout threshold, and dependencies of the field scenario; Example: {"task_book_id":"Field Monitoring", "overall_objective":"Inference of unconventional drone signal intent in Section A of the western mountainous area", "sub_tasks": [{"task_id":"T1", "task_type":"signal preprocessing", "input_data_index":"Data Acquisition Device 1-20240520-0900", "preprocessing": "Noise filtering + signal enhancement", "tool":"Noise Filtering Model", "priority":"highest", "timeout_threshold":"10 minutes", "dependencies":[]}, {"task_id":"T2", "task_type":"modulation classification", "input_data_index":"Output result of T1", "preprocessing": "None", "tool":"Lightweight CNN modulation classification model", "priority":"high", "timeout_threshold":"15 minutes", "dependencies": ["T1"]}, {"task_id": "T3", "task_type": "Protocol recognition", "input_data_index": "Output result of T2", "preprocessing": "Protocol feature extraction", "tool": "Field UAV protocol recognition model", "priority": "High", "timeout_threshold": "15 minutes", "dependencies": ["T2"]}, {"task_id": "T4", "task_type": "Positioning", "input_data_index": "Output result of T3 + GPS information of acquisition device", "preprocessing": "Positioning data calibration", "tool": "TDOA positioning model", "priority": "Medium", "timeout_threshold": "20 minutes", "dependencies": ["T3"]}, {"task_id": "T5", "task_type": "Intent inference", "input_data_index": "Output results of T2 + T3 + T4", "preprocessing": "Feature fusion", "tool": "The second intelligent agent", "priority": "Highest", "timeout_threshold": "10 minutes", "dependencies": ["T4"]}]}。

[0038] Then perform task scheduling and intent inference, read the subtask sequence of the task book, parse the priority and dependencies, and adopt the strategy of "executing high-priority tasks first and parallelizing tasks without dependencies".

[0039] The next step for the second intelligent agent is to combine the input data to collect intermediate result information from each tool model (such as "T2 intermediate result: signal modulation type confidence 92%) and final label information (such as "T3 final label: a certain illegal drone private protocol").

[0040] The reasoning process is as follows: Feature fusion: Weighted fusion of features such as modulation type, protocol type, positioning result, and signal strength.

[0041] Intent matching: Based on the fine-tuned LLM, match the intent library of the field scene (including typical intents such as illegal mapping, signal interference and illegal communication).

[0042] The final result example is: "The signal was transmitted with the intent of illegal mapping, and the risk level is medium to high."

[0043] Evidence chain generation: Output the reasoning process and core evidence. Example: "1. The modulation type is QAM-16 (confidence level 92%), which is consistent with the characteristics of high-definition mapping signals of UAVs; 2. The protocol matches a certain illegal UAV private protocol (confidence level 88%); 3. The positioning result is located in the core protected area of ​​the mountainous area (prohibited mapping area); 4. The signal duration is 30 minutes, and there is no legal registration mark."

[0044] The scope of protection of this invention is not limited to specific numerical values ​​in the process of situational awareness in the field. Those skilled in the art can make reasonable selections based on the actual situation.

[0045] The execution logic of the intent inference reflection module is as follows: First, it receives the intent inference result and user feedback. Then, it reflects on the task book generation process. If the intent inference is accurate, it strengthens the logic of the corresponding sub-task sequence (e.g., retaining the "signal filtering → modulation classification → protocol identification" process). If there is a deviation (e.g., failing to identify anti-interference signals), it optimizes the task book (e.g., adding the "anti-interference signal detection" sub-task). Finally, it updates the parameters, fine-tuning the parameters of the first agent through gradient descent. The update cycle is once a day (for iterative optimization of batch tasks in the field). For the evidence chain storage part, it first performs reverse gradient calculation, that is, the reflection module, based on the intent inference result, reverse-calculates each input label pair. The contribution gradient of the results (e.g., "protocol type label contribution gradient 0.7, which is TOP-1") is then used to preserve the evidence chain. First, the top 3 labels with the highest contribution (e.g., protocol type, location result, modulation type) are automatically selected and saved to the edge server in the format of "task book ID + label information + contribution gradient + causal relationship". Then, manual review is performed. Evidence chain data is exported weekly, and domain experts analyze the correctness of the causal relationship. If it is correct, the reasoning path is strengthened; if it is incorrect, the task book execution steps are corrected (e.g., adjusting the sub-task priority). At the same time, the preservation period is specified. The core evidence chain is permanently preserved, and the original data is preserved for 30 days.

[0046] The multi-agent collaborative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of this invention can be applied to different scenarios. The following describes the process of detecting unauthorized drone flights: The core objective is to achieve real-time capture, rapid identification, accurate positioning, and intent inference of unauthorized drone signals within the airport's airspace and a surrounding area of ​​5-10 kilometers through multi-agent collaboration, providing timely intervention basis for airport air traffic control departments and preventing unauthorized drones from interfering with flight takeoffs and landings and threatening aviation safety.

[0047] First, an electromagnetic spectrum intent reasoning multi-agent system is built. The system architecture must be adapted to the high-density signal environment and security control requirements of airports. The front-end user interface adopts a design combining a dedicated terminal for the airport air traffic control command center and a mobile monitoring APP. The terminal supports multi-screen split-screen display, while the APP allows patrol personnel to receive real-time warning information. Both have offline caching capabilities to cope with temporary airport network interruptions. Among the core modules that the interface should include, the electromagnetic detection tool model display window should be labeled with "High Sensitivity," "Low False Alarm Rate," and "Airport Dedicated," highlighting models for UAV signal acquisition, protocol parsing, and precise positioning. The window will also simultaneously display the currently available airport electromagnetic detection tool model interface information, as shown in the following example format: {"tool_name":"Enhanced CNN Modulation Classification Model", "function":"Identify the modulation type of drone signals in the 2.4GHz / 5.8GHz bands at airports and eliminate interference from civil aviation communications", "input_type":"tensor[64,64,1]", "output_type":"Text (Modulation type + Confidence level + Interference exclusion description)", "response_time":"≤0.3 seconds", "applicable_scenario":"Signal detection around airport runways and over the apron", "maintenance_status":"Normal and available"}, Figure 3 This is the user interface of the electromagnetic signal detection platform for detecting unauthorized drone flights in Embodiment 1 of the present invention. The interactive dialog box window supports text input and quick selection of standard air traffic control commands, such as "suspected unauthorized flight signal near runway entrance 1". It also has an emergency warning pop-up function, with a red pop-up alert triggered by high-risk signals. The task list display window is presented in a concise text format, highlighting "task timeliness requirements" and "intervention priority". It also fully displays the airport unauthorized drone flight detection task list generated by the system, as shown in the following format example: {"tool_name":"Enhanced CNN Modulation Classification Model", "function":"Identify the modulation type of drone signals in the 2.4GHz / 5.8GHz bands at airports and eliminate interference from civil aviation communications", "input_type":"tensor[64,64,1]", "output_type":"Text (Modulation type + Confidence level + Interference exclusion description)", "response_time":"≤0.3 seconds", "applicable_scenario":"Signal detection around airport runways and over the apron", "maintenance_status":"Normal and available"}, The interactive dialog box supports text input and quick selection of standard air traffic control commands, such as "suspected unauthorized flight signal near runway threshold 1." It also features an emergency warning pop-up function, with a red pop-up alert triggered by high-risk signals. The task list display window presents a concise text format, highlighting "task timeliness requirements" and "intervention priority." It also fully displays the airport drone unauthorized flight detection task list generated by the system, as shown in the following example format: {"task_book_id":"airport-drone-detection-20240610-001", "overall_objective":"Identify, locate, and infer the intent of unauthorized drones flying over the north side of runway 3 at the airport to ensure the safety of flights departing at 10:30". "task_initiator":"Airport Air Traffic Control Center", "initiate_time":"2024-06-1009:50:00", "required_completion_time":"2024-06-1010:00:00", "sub_tasks": [{"task_id":"T1", "task_type":"signal preprocessing", "input_data_index":"Airport Base Station 3-20240610-0948 (Data Collection Point on the North Side of Runway 3)" "preprocessing": "Airport radar interference filtering + drone weak signal enhancement", "tool":"Airport Signal Noise Reduction Model", "priority":"highest", "timeout_threshold":"3 minutes", "dependencies":[], "responsible_module":"Sub-smart agent set - Signal preprocessing module"}, {"task_id":"T2", "task_type":"modulation classification", "input_data_index":"T1 output result (preprocessed I / Q stream)", "preprocessing":"Signal frequency band focus (2.4GHz / 5.8GHz)", "tool":"Enhanced CNN Modulation Classification Model", "priority":"highest", "timeout_threshold":"4 minutes", "dependencies":["T1"], "responsible_module":"Sub-agent set - Modulation classification module", "output_requirement":"Specify modulation type (FSK / QAM, etc.), confidence level ≥90%, interference exclusion description"} {"task_id":"T3", "task_type":"protocol identification", "input_data_index":"T2 Output Result (Modulation Type + Signal Characteristics)", "preprocessing": "Protocol feature extraction + comparison of legitimate civil aviation agreements", "tool":"Airport Unmanned Aerial Vehicle Protocol Identification Model", "priority":"highest", "timeout_threshold":"4 minutes", "dependencies":["T2"], "responsible_module":"Sub-agent set - Protocol identification module", "output_requirement":"Protocol type, brand, legal filing status, description of suspicious characteristics"}, {"task_id":"T4", "task_type":"Precise positioning", "input_data_index":"T3 output results (protocol information) + arrival time of base station 3 / 4 / 5 signals + coordinates of surveillance camera (number C12)", "preprocessing":"Location data calibration (excluding building obstruction errors)", "tool":"TDOA+AOA Fusion Localization Model", "priority":"high", "timeout_threshold":"5 minutes", "dependencies":["T3"], "responsible_module":"Sub-agent set - localization module", "output_requirement":"Location coordinates (latitude and longitude), error range ≤ 10 meters, surrounding geographical landmarks (e.g., "1.2 kilometers north of runway end 3, near the shelterbelt")"}, {"task_id":"T5", "task_type":"intent inference", "input_data_index":"T2 (modulation type) + T3 (protocol information) + T4 (location result) + flight schedule (flight HU7801 departing at 10:30)", "preprocessing":"Feature weighted fusion (location and flight correlation have the highest weight)", "tool":"Secondary Agent", "priority":"highest", "timeout_threshold":"4 minutes", "dependencies":["T4"], "responsible_module":"Second Agent", "output_requirement":"Type of unauthorized flight, threat level, scope of impact, intervention recommendations, and chain of inference evidence"}, "emergency_contact":"Engineer Li from the Airport Security Department", "backup_plan":"If a tool model fails, it will automatically switch to the backup model (e.g., if T2 fails, it will switch to the airport-specific SVM modulation classification model)"}.

[0048] The reflection process display window focuses on the discrepancies between the task description and the actual test results, such as "the protocol of a certain new type of black-flying drone is not covered, and model training needs to be supplemented"; the intention inference result display window should highlight "black flight type", "threat level" and "suggested intervention measures", such as "consumer-grade black-flying drone - high threat - immediately start directional jamming", and at the same time associate it with geographical information such as airport runways and aprons to help air traffic controllers quickly determine the scope of impact. Both the first and second intelligent agents are deployed based on high-performance LLM, selecting model versions that support millisecond-level response times. They are deployed on a local airport server cluster to avoid cloud transmission delays, and a backup server is configured to ensure uninterrupted system operation. The sub-agent set needs to select electromagnetic detection tool models suitable for the airport scenario. Among these, the automatic modulation classification model prioritizes enhanced CNN models that support multi-band UAV signal recognition, accurately identifying the 2.4GHz and 5.8GHz frequency band signal modulation types commonly used in consumer and industrial UAVs. The protocol recognition model must include an airport-specific UAV protocol library, covering the proprietary protocols of mainstream unauthorized UAV brands and custom protocols of illegally modified equipment, quickly eliminating interference signals from legitimate civil aviation communications and airport ground equipment. The tracking and positioning model uses multi-base station collaborative TDOA+AOA fusion positioning technology, combined with the coordinates of surveillance cameras deployed around the airport, to achieve accurate positioning of unauthorized UAVs within 10 meters. Supplementary tool models must include a UAV signal noise reduction model and a multi-signal separation model. The former is used to filter electromagnetic interference generated by airport runway radar and ground navigation equipment, while the latter is used for signal separation when simultaneously capturing multiple unauthorized UAV signals, avoiding mutual interference that could affect the detection results.

[0049] The first intelligent agent fine-tuning dataset needs to contain more than 1,500 actual task requirement texts from airport air traffic controllers and corresponding task book tags. An example of the requirement text is "An unidentified flying object was detected over runway 3 at the airport, suspected to be a drone flying illegally. The signal type and location need to be confirmed within 5 minutes, and it needs to be determined whether it will affect flight landing." The corresponding task book tag needs to include the complete process of signal acquisition, modulation classification, protocol identification, localization, and intent inference, and clearly define the time requirements for each step. The format can refer to the above example of an airport drone illegal flight detection task book.

[0050] The second intelligent agent's fine-tuning dataset needs to contain more than 800 paired data entries of "tool model output label + intent result". An example of an input label is: "Modulation type: FSK, Protocol type: a certain consumer-grade drone's proprietary protocol; Location result: 1 km north of runway 3; Signal strength: strong; Flight altitude: 300 meters; Associated flight: HU7801 departing at 10:30". The output intent result must clearly specify "unauthorized flight type, threat level, impact range, and intervention suggestion". Fine-tuning execution adopts a scheme combining incremental training and LoRA, retaining the LLM's basic language understanding capabilities while strengthening the recognition of airport electromagnetic signal domain terminology and the task logic for unauthorized flight scenarios.

[0051] Field data collection needs to be adapted to the high-density signal environment and all-weather monitoring requirements of airports. Regarding equipment deployment, fixed electromagnetic measurement base stations should be deployed at 2-kilometer intervals within a 5-10 kilometer radius of the airport's airspace. Each base station should be equipped with an ultra-wideband spectrum analyzer (bandwidth coverage of 100MHz-6GHz), a high-gain directional antenna, and a 24-hour uninterrupted power supply system. Portable mobile monitoring equipment, carried by patrol personnel, should also be deployed at both ends of the airport runway and around the apron to address the need for detecting signal blind spots in areas covered by fixed base stations. Auxiliary equipment should include meteorological sensors to monitor the impact of weather factors such as wind speed and precipitation on the UAV's flight path, as well as a 4G / 5G dual-link communication module to ensure real-time transmission of collected data to the airport's local server. A satellite communication backup link should also be configured to address ground communication interruptions caused by extreme weather.

[0052] The data collection content should focus on covering the common signal types of unauthorized drone flights. Core data includes I / Q streams (sampling rate increased to 5MSps to ensure high-speed drone signal capture), power spectrum (PSD), and short-time Fourier transform (STFT) frames (frame length shortened to 5ms to improve real-time signal acquisition efficiency). The collection time, coordinates of the collection equipment, and corresponding airport flight take-off and landing schedules should be recorded simultaneously to facilitate subsequent correlation analysis of the impact of unauthorized drone signals on flights. Supplementary data should include synchronous video frames from airport surveillance cameras (to verify the matching of visual and electromagnetic signal characteristics of unauthorized drones), airport air traffic control radar data (excluding signal interference from civil airliners and legal general aviation aircraft), and voice feedback from ground personnel (such as the location and appearance description of unauthorized drones discovered by patrol personnel).

[0053] Data normalization processing must meet the airport's requirements for data consistency across multiple devices. The preprocessing module built into the acquisition device must normalize the raw data collected from different base stations and different types of devices into tensor types (with dimensions uniformly set to [64, 64, 1]), and simultaneously label the signal strength, signal-to-noise ratio, and interference level during data acquisition to facilitate subsequent tool models adjusting detection parameters based on data quality. Data storage adopts a three-tiered storage scheme of "real-time transmission + local backup + cloud archiving." Real-time data is transmitted to the airport command center server for immediate analysis, local backup retains 7 days of data for retrospective querying, and cloud archiving retains 3 months of data for model optimization and incident tracing. All data must be stored encrypted in accordance with airport information security management regulations.

[0054] User requests are submitted through a dedicated terminal at the airport air traffic control command center. Air traffic controllers can choose preset task templates (such as "detecting unauthorized flights around the runway", "investigating signals over the apron", and "focusing on monitoring during flight take-off and landing periods") or manually enter custom requests. The request text must include the monitoring area (accurate to the specific location of the airport), task timeframe (such as "outputting preliminary results within 3 minutes"), and key focus (such as "prioritizing the identification of modified drone signals").

[0055] The input of tool model interface information is the responsibility of airport technical maintenance personnel. Each time a tool model is added or updated, the model interface information must be uploaded to the system. In addition to the core tool model interface example mentioned above, the format of the newly added model interface information is as follows: {"tool_name":"Multi-signal separation model", "function":"Simultaneously separates signals from multiple unauthorized flights at airports to avoid signal superposition interfering with detection results", "input_type":"Mixed I / Q streams + signal quantity estimation", "output_type":"Single-target I / Q stream (sorted by signal strength) + separation confidence", "response_time":"≤0.8 seconds", "applicable_scenario": "Scenario where multiple unauthorized flying targets appear simultaneously around the airport", "maintenance_status":"Normal and available", "version":"V2.1", "update_time":"2024-05-20"} The system automatically verifies the compatibility of the model interface with the existing system, generates case status identifiers as available, pending testing, and unavailable, and records the model's update time and maintenance responsibility to facilitate subsequent troubleshooting.

[0056] Task analysis should prioritize timeliness and safety in airport scenarios. The first agent captures the task requirement text submitted by air traffic controllers and the information of currently available tool models in real time. Through a fine-tuned LLM, it quickly completes semantic understanding, focusing on extracting key information from the task, such as "whether the monitoring area involves the runway or apron," "whether the task is urgent," and "whether it is related to flight take-off and landing plans," to generate a preliminary task intent. The initial task intent should clearly define the task output objectives and brief execution steps. The output objectives should quantify the timeliness and accuracy requirements, such as "complete the modulation classification and protocol identification of signals over runway 1 within 2 minutes, with a positioning error of no more than 20 meters." The brief execution steps should be ordered according to the urgency of the task. For urgent tasks (such as detecting unauthorized flight signals before flight takeoff), the signal acquisition and identification stage should be prioritized. For non-urgent tasks (such as routine inspections), signal feature storage and model training data accumulation stages can be added. For example, the initial execution steps for routine inspection tasks are "signal preprocessing → modulation classification → protocol identification → feature storage → abnormal signal marking," while the initial execution steps for urgent tasks are "signal preprocessing (priority) → modulation classification (parallel) → protocol identification (parallel) → positioning → intent inference." After reading the sub-task sequence of the airport drone unauthorized flight detection task book, the task scheduling module of the second intelligent agent first analyzes the priority and dependency of each sub-task. Combining the principles of safety first and timeliness first in the airport scenario, it adopts a strategy of preemptive execution of high-priority tasks and parallel collaboration of dependent tasks. For example, in the "emergency unauthorized flight detection at runway 3" task, when T1 (signal preprocessing, highest priority) is executed, the system will simultaneously load the enhanced CNN modulation classification model parameters and airport-specific protocol library required by T2 (modulation classification), and start the signal synchronization reception channels of base stations 3 and 4. After T1 completes signal denoising and filtering, it immediately triggers the parallel execution of T2 (modulation classification) and T3 (protocol identification). At this time, the task scheduling module will monitor the computing power usage of the two sub-tasks in real time. If T2 has insufficient computing power, it will temporarily allocate the resources of the allocated server to ensure that the timeout rate of a single sub-task is less than 1%. For routine inspection tasks, independent subtasks can be executed flexibly at off-peak times. For example, T4 (signal feature storage) can start preprocessing when T3 (protocol identification) is 70% complete, which avoids wasting resources and ensures that the overall task is completed within the required time. When the second intelligent agent infers the intent of unauthorized drone flights at the airport, it will focus on collecting intermediate result information and final label information from various tool models. The intermediate results must include key data specific to the airport scenario: such as the intermediate result of T2 (multi-band modulation classification): "The modulation type of the 2.4GHz band is FSK, with a confidence level of 95%; there is no effective signal in the 5.8GHz band, and the interference signal accounts for 12% (mainly airport ground support communication)"; the final label information of T3 (protocol identification and legality verification): "The protocol type matches the private protocol of a certain consumer-grade unauthorized drone (without civil aviation registration mark), the owner is unknown, the signal duration has reached 18 minutes, and it is moving towards runway 1."

[0057] The reasoning process is optimized around airport security requirements. First, feature fusion is performed, weighting and fusing features such as modulation type, protocol legality, location results, signal movement trajectory, and related flight dynamics. Among these, "distance between the location result and the runway or flight path," "legality of protocol registration," and "signal movement direction" have the highest weights, aligning with the airport's core judgment needs regarding whether a signal threatens flight safety. For example, in a runway-side unauthorized flight detection task, if the location result shows an unauthorized flight signal only 200 meters from the runway centerline and moving along the flight landing path, this feature's weight will temporarily increase to 35%, prioritizing the triggering of a high-risk warning logic. Next, intent matching is performed. Based on the LLM (Local Limit Model) fine-tuned from airport unauthorized flight data, it matches the airport's exclusive intent library, which includes typical intents such as "malicious interference with flight takeoffs and landings," "illegal photography of airport facilities," "mistakenly entering the airspace," and "illegal transportation." It also considers airport security levels (e.g., during peak travel periods like Spring Festival or important flight operations, the matching threshold for "malicious interference" intent is reduced by 15%). Finally, the final result and evidence chain are generated. The output must clearly specify "unauthorized flight intent type, threat level, scope of impact, and intervention measures," and the evidence chain must include the aircraft... The following are examples of compliance and relevance data specific to this scenario: The intent inference result is "The signal transmission intent is to maliciously interfere with flight takeoffs and landings, with an extremely high threat level"; the chain of reasoning evidence is: "1. The modulation type is FSK, which is consistent with the characteristics of drone remote control signals, and the signal strength is continuously increasing, indicating that the drone is approaching the airport; 2. The protocol matches a private protocol of a modified black-flying drone, without a Civil Aviation Administration filing mark, and the protocol contains "anti-interference frequency hopping" instructions, which has the characteristics of deliberately avoiding monitoring; 3. The positioning result is located 300 meters east of runway 1, and the movement trajectory prediction shows that the next 2 minutes..." 4. Related flight status: Flight HU7801 (landing at 10:30) has entered the fifth approach phase, currently at an altitude of 1000 meters. The unauthorized flight signal may interfere with the flight navigation system; 5. The signal lasted for 22 minutes, during which there was no legitimate declaration signal exchange; The intervention recommendation is to "immediately activate the directional jamming equipment at runway 1 to cut off the drone's remote control link; notify the airport security patrol team to go to the location coordinates, carrying an anti-drone capture net; simultaneously notify the air traffic control department, and suggest that Flight HU7801 temporarily go around and reapproach after the interference is cleared."

[0058] The intent inference and reflection module is embedded in the first intelligent agent. Its execution logic in airport scenarios needs to be optimized by combining feedback from air traffic controllers and the results of airport security incident handling. First, it receives intent inference results and multi-dimensional feedback, including effectiveness evaluations of intervention measures submitted by air traffic controllers through the command terminal, on-site handling reports from security departments, and flight operation records. Then, it reflects on the task generation process. If the intent inference is accurate and the intervention measures are effective (e.g., successfully avoiding flight interference), it strengthens the corresponding sub-task sequence logic, for example, retaining the emergency task of "signal preprocessing → parallel modulation classification and protocol identification → positioning → trajectory prediction → intent inference". The process involves adding the case's feature data (such as "anti-interference frequency hopping protocol features" and "black flight trajectory on the approach route") to the first intelligent agent's fine-tuning dataset. If there are deviations (such as mistakenly identifying a legal agricultural drone as an illegal flight, leading to the temporary closure of the apron), the reasons for the deviations are analyzed in depth. If the task book does not include the "List of Legally Operated Drones Around the Airport," a "T3-1 (Legal Reporting Comparison)" step is added to the sub-task sequence to call the airport air traffic control reporting database interface for secondary verification. If the tool model has a low accuracy rate in recognizing the "Special Protocol for Agricultural Drones," the model is marked as needing supplementary training, and the backup model calling rules are updated. The parameter update process is adapted to the 24-hour operation characteristics of the airport and is carried out during the intervals between flight takeoffs and landings. The parameters of the first agent are fine-tuned through gradient descent, with a focus on optimizing modules such as "priority allocation of sub-tasks in different areas of the airport", "weight of flight correlation features", and "logic for identifying new black flight protocols". After each parameter update, 10 sets of historical typical cases need to be tested to ensure that the accuracy of task book generation is not less than 95% before it can be used the next day. The evidence chain preservation strategy must meet the requirements of airport security incident tracing and civil aviation supervision. First, reverse gradient calculation is performed. The reflection module, based on the difference between the intent inference result and the actual handling result, reverse-calculates the contribution gradient of each input label to the intent inference. For example, in the case of malicious flight interference, the label "Location Result and Runway Distance" contributes a gradient of 0.85 (TOP-1), the label "Protocol Anti-interference Characteristics" contributes a gradient of 0.78 (TOP-2), and the label "Flight Approach Status" contributes a gradient of 0.72 (TOP-3). Subsequently, evidence chain preservation is carried out, automatically selecting the labels with the highest TOP-3 contributions. Combined with sub-task execution logs, air traffic control instruction records, and on-site handling video clip indexes, the data is organized into structured data according to the "Airport Security Incident Evidence Specification," with an example format: "Task ID: airport-dron e-detection-20240610-001; Tag information 1: Location result (300 meters east of runway 1, within the approach channel), contribution gradient 0.85, causal relationship: This location directly threatens the safety of landing flights and is the core basis for determining the intent of "malicious interference"; Tag information 2: Protocol type (a modified private protocol for illegal flights, containing anti-interference frequency hopping commands), contribution gradient 0.78, causal relationship: The protocol has the characteristics of deliberately avoiding monitoring, ruling out the possibility of "mistaken intrusion"; Tag information 3: Flight status (HU7801 flight five-sided approach, altitude 1000 meters), contribution gradient 0.72, causal relationship: The flight is in a critical flight phase, and the risk of illegal flight signal interference is extremely high; Retention time: 2024-06-10 11:30:00; Storage location: Airport security database - illegal flight event partition. Manual review is conducted by the Airport Safety Management Department in conjunction with the air traffic control technical team. Every Friday from 16:00 to 18:00, evidence chain data for all unauthorized flight detection tasks of the week is exported. The focus is on verifying the compliance of the intent inference logic with civil aviation safety regulations, whether the intervention measures comply with the airport's emergency response procedures, and whether the data records are complete and traceable. If the evidence chain is verified to be correct, the reasoning path is included in the airport's best practice library for unauthorized flight intent inference, and it will be prioritized for subsequent similar tasks. If errors are found, the execution steps of the task book are corrected (such as adding a suspicious feature marking sub-task after T4), and the person responsible for the data collection process is held accountable to ensure the integrity of the evidence chain. The retention period strictly follows the regulatory requirements of the Civil Aviation Administration: core evidence chains are permanently retained, evidence chains for ordinary unauthorized flight incidents are retained for 5 years, and original collected data is retained for 90 days. After the expiration, the data can be destroyed according to the prescribed procedures only after being reviewed and approved by the civil aviation regulatory department. The destruction record must be kept for 10 years for future reference.

[0059] The scope of protection of this invention is not limited to specific values ​​in the process of detecting unauthorized drone flights; those skilled in the art can make reasonable selections based on the actual situation.

[0060] The scope of protection of this invention is not limited to the two examples of situational awareness in the field and detection of unauthorized drone flights listed in Example 1. Those skilled in the art can make reasonable choices based on the actual situation.

[0061] The multi-agent collaborative method for electromagnetic spectrum monitoring and analysis proposed in Embodiment 1 of this invention decomposes the overall task into multiple sub-tasks such as perception, classification, feature fusion, and anomaly detection, which are then processed in parallel or serially by an encapsulated second agent. This overcomes the performance degradation problem of a single end-to-end model when facing multi-source, multi-format, and non-steady-state signals, thereby significantly improving the recall and robustness of spectrum analysis, increasing the anomaly detection rate, and reducing the false alarm rate.

[0062] Example 2 The present invention also proposes a device, Figure 4 This is a schematic diagram of a multi-agent collaborative device for electromagnetic spectrum monitoring and analysis proposed in Embodiment 2 of the present invention.

[0063] At the hardware level, the electronic device 400 includes a processor 410, and optionally, an internal bus 420, a network interface 430, and memory. The memory may include main memory, such as high-speed random-access memory (RAM), or it may also include non-volatile memory, such as at least one disk drive. Of course, the electronic device may also include other hardware required for other business operations. The processor 410, network interface 430, and memory can be interconnected via an internal bus 420. This internal bus 420 can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be categorized as an address bus, data bus, control bus, etc. For ease of illustration, only a single bidirectional arrow is used in this diagram, but this does not imply that there is only one bus or one type of bus. The memory is used to store programs. Specifically, the program can include program code, which includes computer operation instructions. The memory can include main memory 440 and non-volatile memory 450, and provides instructions and data to the processor 410. Processor 410 reads the corresponding computer program from non-volatile memory 450 into memory 440 and then runs it, forming a device for locating the target user at the logical level. Processor 410 executes the program stored in memory and specifically performs the following: In step S100, a user interaction interface is set up at the front end, and an intelligent agent system is set up at the back end.

[0064] In step S110, the overall task requirements in text form are received through the user interface, and the interface information of the electromagnetic detection tool model is obtained.

[0065] In step S120, the task requirements are parsed by a preset first intelligent agent to generate a preliminary task intent; and based on the preliminary task intent and the interface information, a task plan containing multiple sub-tasks is generated. In step S130, the task scheduling module of the second intelligent agent calls the electromagnetic detection tool model to execute the sub-tasks of the task plan, and obtains the intermediate result information and final label information generated by the electromagnetic detection tool model. In step S140, based on the intermediate result information and the final label information, feature fusion and intent reasoning are performed by a preset second intelligent agent to generate the intent inference result of the electromagnetic signal and the evidence chain supporting the intent inference result; In step S150, feedback information regarding the intent inference result is received; based on the intent inference result and / or the feedback information, the process of generating the task plan is reflected upon by the first intelligent agent; based on the reflection result, the internal parameters of the first intelligent agent are updated.

[0066] Figure 1It can be applied to processor 410, or implemented by processor 410. The processor may be an integrated circuit chip with signal processing capabilities. In the implementation process, each step of the above method can be completed by the integrated logic circuit in the hardware of the processor or by instructions in the form of software. The processor mentioned above can be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc.; it can also be a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the various methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of this application can be directly embodied as being executed by a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software module can reside in a mature storage medium in the field, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in memory, and the processor reads information from the memory and, in conjunction with its hardware, completes the steps of the above method. Of course, in addition to software implementation, the electronic device of this application does not exclude other implementation methods, such as logic devices or a combination of hardware and software, etc. In other words, the execution subject of the following processing flow is not limited to each logic unit, but can also be hardware or logic devices.

[0067] The description of the relevant parts of the multi-agent collaborative device for electromagnetic spectrum monitoring and analysis provided in Embodiment 2 of this application can be found in the detailed description of the corresponding parts of the multi-agent collaborative method for electromagnetic spectrum monitoring and analysis provided in Embodiment 1 of this application, and will not be repeated here.

[0068] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that the elements inherent in a process, method, article, or apparatus that includes a list of elements are included. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element. Additionally, portions of the technical solutions provided in the embodiments of this application that are consistent with the implementation principles of corresponding technical solutions in the prior art have not been described in detail to avoid excessive elaboration.

[0069] While specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art can make other modifications or variations based on the above description. It is neither necessary nor possible to exhaustively describe all embodiments here. Various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A multi-agent collaborative method for electromagnetic spectrum monitoring and analysis, characterized in that, The method comprises the following steps: Receiving a task requirement for electromagnetic spectrum monitoring and analysis, and obtaining interface information of an electromagnetic detection tool model; Analyzing the task requirement by a preset first intelligent agent to generate a preliminary task intention; Based on the preliminary task intention and the interface information, a task planning document containing multiple sub-tasks is generated; Through a task scheduling module of a second intelligent agent, the electromagnetic detection tool model is called to execute the sub-tasks of the task planning document, and intermediate result information and final label information generated by the electromagnetic detection tool model are obtained; Based on the intermediate result information and the final label information, feature fusion and intention reasoning are performed by a preset second intelligent agent to generate an intention inference result of the electromagnetic signal and an evidence chain supporting the intention inference result.

2. The method of claim 1, wherein, The method comprises the following steps: Analyzing the task requirement by a preset first intelligent agent to generate a preliminary task intention; and based on the preliminary task intention and the interface information, a task planning document containing multiple sub-tasks is generated; Specifically: Performing semantic understanding on the task requirement to generate a preliminary task intention; the preliminary task intention includes a task output target and an execution step; Based on the preliminary task intention and in combination with the functions described in the interface information, task decomposition and tool matching are performed; 3. The method of claim 1, wherein, The matching result is iteratively optimized to generate a structured task planning document, wherein the task planning document is used to define the task type, input data source, required electromagnetic detection tool model, execution priority, and dependency relationship between sub-tasks of each sub-task. Through a task scheduling module of a second intelligent agent, the electromagnetic detection tool model is called to execute the sub-tasks of the task planning document, and intermediate result information and final label information generated by the electromagnetic detection tool model are obtained, specifically: According to a dynamic task scheduling sequence, the normalized electromagnetic data are bound and called with the electromagnetic detection tool model required by the sub-tasks; the electromagnetic detection tool model includes an automatic modulation classification model, a protocol identification model, a tracking and positioning model, a noise filtering model, and a signal enhancement model; 4. The method of claim 1, wherein, The execution state of each electromagnetic detection tool model is monitored, and the intermediate result information and final label information output by the electromagnetic detection tool model are collected; wherein the intermediate result information includes text description or intermediate features in the reasoning process of the electromagnetic detection tool model, and the final label information includes classification, identification, or positioning conclusions obtained by the electromagnetic detection tool model after reasoning on the input data. Based on the intermediate result information and the final label information, feature fusion and intention reasoning are performed by a preset second intelligent agent to generate an intention inference result of the electromagnetic signal and an evidence chain supporting the intention inference result, specifically: Based on a pre-defined domain knowledge rule base, corresponding weight coefficients are assigned to different feature categories in the feature vector to generate a weighted comprehensive feature vector; The weighted comprehensive feature vector is matched and reasoned with a preset scene intention knowledge base to generate an intention inference result of the electromagnetic signal emission source behavior; Based on the weighted comprehensive feature vector and the matching and reasoning process, an evidence chain supporting the intention inference result is generated, and the evidence chain at least includes feature sources, reasoning logic paths, and contribution degrees of each evidence to the final conclusion.

5. The method of claim 1, wherein, The method further comprises: receiving feedback information for the intention inference result; based on the intention inference result and / or the feedback information, reflecting on the process of generating the task planning document by the first agent; updating the internal parameters of the first agent based on the reflection result.

6. The method of claim 5, wherein, The reflection on the process of generating the task planning document by the first agent comprises: based on the intention inference result, determining the contribution degree of each final label information to the intention inference result; based on the contribution degree, screening out evidence; analyzing the causal relationship between the evidence and the intention inference result, and adjusting the task planning logic based on the analysis result.

7. The method of claim 6, wherein, Based on the intention inference result, the contribution degree of each final label information to the intention inference result is determined, which comprises: in the calculation graph of the intention reasoning of the second agent, the intention inference result is taken as the starting point, and the gradient of the output layer to the input layer is calculated by back propagation; extracting the gradient value on the input node corresponding to the final label information, and taking the absolute value of the gradient value as a quantitative index of the contribution degree of the corresponding final label information to the intention inference result; based on the quantitative index, all final label information is sorted, and a preset number of evidence with the highest contribution degree is screened out.

8. The method according to any one of claims 1 to 7, characterized in that, The method obtained by the first agent is: obtaining a first basic large language model; using a first training data set to supervise the fine-tuning of the first basic large language model to obtain a first agent; wherein the first training data set contains multiple task demand texts in the electromagnetic monitoring field, and the structured task planning document corresponding to each task demand text. The method obtained by the second agent is: obtaining a second basic large language model; using a second training data set to supervise the fine-tuning of the second basic large language model to obtain a second agent; wherein the second training data set contains multiple electromagnetic signal feature data, and the intention inference conclusion and evidence chain description corresponding to each feature data.

9. The method of claim 8, wherein, The method further comprises: through the planning strategy evolution of the first agent, the intention knowledge evolution of the second agent, and the tool set evolution of the task scheduling module of the second agent, the cooperation of the first agent and the second agent is realized, which comprises: The first agent updates its parameters according to the reflection result, optimizes its understanding and planning ability for electromagnetic monitoring tasks; In the reasoning process of the second agent, the scene intention knowledge base called by the second agent is incrementally updated based on the electromagnetic signal intention mode after reflection; The task scheduling module of the second agent can dynamically integrate the electromagnetic detection tool model introduced after updating according to the updated task planning document. 10.A multi-agent collaborative device for electromagnetic spectrum monitoring and analysis, comprising at least one processor and a memory, the memory storing a computer program, characterized in that, The computer program is executed by the at least one processor to realize a multi-agent cooperative method for electromagnetic spectrum monitoring and analysis according to any one of claims 1-9.

Images