A radio frequency signal demodulation analysis method and system fusing LLM and agent

Abstract

The application discloses a radio frequency signal demodulation analysis method and system fusing LLM and Agent, and relates to the field of wireless communication and automation testing. The method embeds LLM and Agent in a test instrument locally; an Agent controls the instrument to collect original IQ signals based on a test priority queue mechanism, and adopts a multi-constraint joint loss function to train a light-weight AI demodulator; the Agent relies on a 7-state finite state machine to schedule the demodulator to perform blind demodulation and output quantitative indicators; LLM performs semantic analysis and EVM early warning determination on the indicators, and deduces hardware tuning strategies; finally, the Agent parses the strategies into bottom-layer writing instructions, dynamically sets front-end hardware parameters, and triggers closed-loop retesting and incremental training. The application breaks through the control fault of high-level semantics to bottom-layer radio frequency hardware, realizes joint compensation of multiple physical damages and multi-task non-conflict scheduling, and greatly improves the demodulation accuracy and test automation level under complex working conditions.

Description

Technical Field

[0001] This invention relates to the field of wireless communication and automated testing, specifically to a radio frequency signal demodulation and analysis method and system that integrates LLM and Agent. Background Technology

[0002] Signal demodulation is a core function of RF test instruments and wireless communication receiving links, directly determining the reliability of communication transmission and the mass production test yield of RF devices. Traditional RF demodulation architectures typically employ multi-stage serial processing flows, including symbol synchronization, frequency synchronization, sampling clock synchronization, and explicit channel estimation. This traditional architecture relies heavily on time-domain searches, frequency-domain iterations, and complex matrix operations, resulting in high computational complexity, large hardware overhead, and high demodulation latency. Furthermore, traditional architectures struggle to fully and effectively compensate for underlying RF link defects such as local oscillator leakage, phase imbalance of in-phase and quadrature signals (IQ), and power amplifier nonlinearity. Performance deteriorates severely under extreme conditions such as multipath fading, strong interference, and low signal-to-noise ratio, and it only outputs cold, numerical results, lacking intelligent analysis and human-computer interaction capabilities.

[0003] In recent years, deep learning technology has been introduced into the field of radio frequency (RF) demodulation. For example, Chinese patent document CN113630130B discloses an end-to-end digital communication demodulation method, and Chinese patent document CN107147600A discloses a digital modulation signal demodulator based on neural networks. However, these existing technologies are limited to using shallow neural networks to achieve a direct mapping of signal waveforms to 0 and 1 symbols, essentially belonging to a purely single demodulation algorithm model. Such solutions lack multimodal human-computer interaction capabilities, lack depth of business semantic understanding, lack natural language parsing functions for fault root cause analysis, and cannot achieve intelligent programmable control of RF instruments, thus failing to meet the demands of modern RF testing for unmanned, automated, and intelligent operation and maintenance.

[0004] On the other hand, with the development of Large Language Models (LLM) and intelligent agent technologies, the industry has begun to explore the application of general-purpose large models in vertical fields. However, general-purpose LLMs have a large number of parameters and extremely high inference latency, making them unsuitable for direct embedded deployment in the edge hardware of RF instruments with limited computing power. More critically, general-purpose large models lack specialized semantic logic specific to the RF communication domain (such as EVM diagnostics and IQ imbalance mechanisms), and can only achieve general natural language casual conversation, failing to handle the demanding professional RF fault chain inference. At the same time, general-purpose agents lack dedicated business process scheduling mechanisms for RF signal acquisition, demodulation inference, parameter tuning, and retest verification, making them prone to task conflicts and edge computing resource contention when facing multimodal command concurrency in RF testing environments. Furthermore, the semantic conclusions of general-purpose large models cannot directly drive the underlying hardware of RF instruments, resulting in an insurmountable technical gap between high-level semantic analysis and low-level hardware program control. Summary of the Invention

[0005] To address the shortcomings of existing technologies, the technical problem this invention aims to solve is that existing deep learning demodulation methods are only single algorithm models, lacking semantic interaction and instrument hardware control capabilities. When applying general large models to the field of radio frequency testing, there are a series of core technical pain points, such as limited computing power for edge deployment, lack of RF professional semantics, inability of semantic conclusions to directly drive RF hardware, difficulty in joint compensation of multi-damage coupling, computing power scheduling conflicts caused by multi-modal instruction concurrency, and lack of implementation logic for underlying algorithms.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a method and system for demodulating and analyzing radio frequency signals by integrating an LLM (Large Language Model) and an Agent, wherein a Large Language Model (LLM) and an intelligent agent are embedded and deployed locally on a radio frequency test instrument; the LLM incorporates a radio frequency domain knowledge base and a fault chain reasoning template; the intelligent agent incorporates a radio frequency test finite state machine and multimodal instruction mapping rules; the method includes the following steps: S1. The intelligent agent controls the radio frequency test instrument to collect the original in-phase quadrature IQ signal to construct a training dataset. The neural network demodulation basic model is trained offline based on the multi-constraint joint loss function. The model after training convergence is then deployed to the edge inference unit after lightweight processing to obtain the AI ​​demodulator. S2. The radio frequency test instrument acquires the original IQ time domain signal of the test object in real time; S3. The intelligent agent schedules the AI ​​demodulator based on the RF test finite state machine; the AI ​​demodulator performs end-to-end blind demodulation on the original IQ time domain signal, completes time-frequency synchronization and RF link impairment compensation through implicit feature learning of the model, and outputs baseband demodulation symbols and RF demodulation quantization indicators. The RF demodulation quantization indicators include error vector amplitude EVM, local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, and IQ phase imbalance. S4. Input the RF demodulation quantization index into the Large Language Model (LLM); The LLM performs semantic analysis on the RF demodulation quantization index based on the fault chain reasoning template, determines the root cause of the RF link fault and the error vector amplitude (EVM) warning level, and outputs multimodal alarm information. S5. The Large Language Model (LLM) autonomously deduces and generates an RF hardware tuning strategy based on the root cause of the fault and the preset three-level mapping rules. The intelligent agent parses the RF hardware tuning strategy into hardware programmable instructions and sends them to the front-end RF hardware to dynamically adjust the RF bias, channel gain and phase compensation parameters. After the parameter adjustment is completed, the agent automatically controls the RF test instrument to perform retest demodulation, forming an adaptive tuning closed loop. S6. The large language model LLM intelligently filters out abnormal samples during the retesting process and supplements them with semantic labels for feedback. The intelligent agent schedules incremental iterative training to optimize the AI ​​demodulator.

[0007] According to the above technical solution, in step S1, the AI ​​demodulator is built based on a hybrid network structure of convolutional neural network (CNN) and long short-term memory network (LSTM). During the model training phase, a multi-constraint joint weighted constraint is adopted, consisting of temporal loss, constellation distortion loss, symbol error correction loss, and EVM error constraint loss. The weights of the total loss function are fixed as follows: temporal loss weight α=0.25, constellation distortion loss weight β=0.35, symbol error correction loss weight γ=0.25, and EVM error constraint loss weight δ=0.15. The constellation distortion loss is obtained by calculating the average Euclidean distance between the constellation point coordinates predicted by the model and the true coordinates of the ideal constellation; the symbol error correction loss is obtained by performing cross-entropy loss calculation on the number of modulation symbol categories, one-hot encoding of symbol truth, and model probability output.

[0008] According to the above technical solution, in step S1, when controlling the radio frequency test instrument to collect the original IQ signal to construct the training dataset, the intelligent agent performs multi-task computing resource scheduling based on the preset test priority queue mechanism. The scheduling rules of the test priority queue are as follows: the fault retest task has the highest priority, the real-time demodulation inference task has the second highest priority, the model incremental training task has the next highest priority, and the sample collection task in the idle state has the lowest priority. The priority queue exclusively or suspends the computing core of the edge inference hardware to avoid the conflict of computing resources when multiple tasks are concurrent.

[0009] According to the above technical solution, in steps S1 and S3, the AI ​​demodulator after training convergence performs structured pruning, low-bit quantization and layer fusion lightweight processing before deployment. The total number of model parameters of the AI ​​demodulator after lightweight processing is controlled within 1 million, and the inference delay of a single alternating signal is controlled within 10ms, so that it can be directly embedded into the local edge inference board of the RF test instrument for operation.

[0010] According to the above technical solution, in step S3, the radio frequency test finite state machine integrated in the intelligent agent kernel contains 7 standard core operating modes; The seven standard core operating modes are defined sequentially as idle state, sample acquisition state, model training state, demodulation inference state, LLM semantic analysis state, parameter tuning and retesting state, and incremental learning state. The intelligent agent drives and controls the RF test instrument to switch between the seven standard core operating modes based on the semantic mapping results of the received multimodal human-computer interaction commands.

[0011] According to the above technical solution, in step S3, when the AI ​​demodulator performs end-to-end blind demodulation on the original IQ time domain signal and outputs the RF demodulation quantization index, it simultaneously outputs the local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance and IQ phase imbalance measured by quantization at the network bottom layer in parallel. In step S4, the local oscillator leakage amplitude, IQ amplitude imbalance, and IQ phase imbalance are uniformly fed into the large language model LLM to match the built-in radio frequency communication professional knowledge base to complete the step-by-step reasoning and tracing of the root cause of the fault.

[0012] According to the above technical solution, in step S4, the specific threshold division rule for the Large Language Model (LLM) to perform the EVM early warning level determination is as follows: When the quantitative indicator meets the condition that EVM < 3.5%, it is judged to be at the normal level; When the quantitative indicator meets the condition of 3.5%≤EVM<8%, it is judged as a warning level; When the quantitative indicator meets the requirement of EVM≥8%, it is judged as an abnormal level and an immediate hardware-level interrupt alarm is triggered, which in turn triggers the local broadcasting unit to perform offline speech synthesis and broadcasting of the root cause of the fault.

[0013] According to the above technical solution, in step S5, the specific implementation action of the Large Language Model (LLM) generating hardware programmable instructions based on the preset semantic-RF parameter-hardware instruction three-level mapping rules is as follows: The generated natural language tuning strategy is converted into standard target parameter physical quantities, which include DC bias, channel gain increment and equalization filter coefficients. The target parameter physical quantity is converted into a write instruction to the register of the microcontroller of the attenuator and phase shifter at the front end of the RF test instrument by the intelligent agent, and the front end hardware parameters are dynamically tuned with an adjustment step accuracy of ±0.01dB.

[0014] This invention also provides a radio frequency signal demodulation and analysis system that integrates LLM and Agent, which is deployed inside a radio frequency test instrument to implement the radio frequency signal demodulation and analysis method that integrates LLM and Agent as described above. The system includes: a large language model LLM embedded decision module; an intelligent agent global task scheduling module; an AI intelligent demodulation inference module; and an instrument closed-loop control optimization module.

[0015] The present invention also provides a computer device, comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory is used to store computer program instructions; the at least one processor is configured to execute the computer program instructions to invoke its built-in arithmetic logic unit, floating-point arithmetic unit, and cache queue, and to perform large language model inference decision-making, agent task scheduling and flow, low-level AI demodulator execution, and hardware programmable instruction issuance actions by scheduling corresponding hardware boards, bus interfaces, and input / output interfaces, thereby realizing the various steps of the above-mentioned radio frequency signal demodulation and analysis method integrating LLM and Agent.

[0016] Compared with the prior art, the beneficial effects achieved by the present invention are: (1) It breaks through the technical gap between "high-level semantics" and "low-level hardware" and realizes closed-loop automated program control of RF instruments. This invention innovatively introduces a three-level mapping rule of "semantics-RF parameters-hardware instructions". After using domain distillation LLM for fault chain reasoning, the natural language strategy is directly converted into the low-level register writing instructions of the front-end attenuator and phase shifter microcontroller by the agent, thereby realizing the physical adaptive tuning of RF bias and channel gain, which greatly reduces the dependence of traditional RF testing on manual experience tuning.

[0017] (2) The underlying operator architecture solves the problems of "implicit time-frequency synchronization" and "joint compensation for multiple physical defects". The lightweight AI demodulator of this invention innovatively proposes a "quaternary multi-constraint joint loss function" that combines timing, constellation distortion, symbol error correction and EVM error, and designs a fixed weight allocation ratio. This enables the CNN-LSTM network to adaptively complete the deep feature extraction and joint compensation of local oscillator leakage, IQ amplitude-phase imbalance and power amplifier nonlinearity in the end-to-end blind demodulation process. The demodulation accuracy is significantly better than the traditional serial architecture in extreme scenarios such as low signal-to-noise ratio and multipath fading.

[0018] (3) The computing power allocation and scheduling based on multimodal and 7-state finite state machines effectively solves the preemption conflict of multiple tasks on edge devices. Facing the computing power bottleneck of large language model inference, neural network demodulation, and concurrent instrument data acquisition, this invention achieves the orderly arrangement and flow of multimodal instructions with extremely low system overhead by using a 7-state finite state machine built into the Agent for RF testing, combined with a priority queue mechanism. This meets the requirements of limited computing power of edge hardware for RF instruments, and enables complex multimodal interaction of RF signals to have a high degree of engineering and industrialization capabilities. Attached Figure Description

[0019] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a schematic diagram of the method flow according to an embodiment of the present invention; Figure 2 This invention provides an overall architecture block diagram of a radio frequency signal demodulation and analysis system that integrates LLM and Agent. Figure 3 This is a schematic diagram illustrating the principle of the CNN-LSTM hybrid network structure of the AI ​​demodulator in this embodiment of the invention. Figure 4 This is a schematic diagram of the system module composition according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the software interface and analysis results of a radio frequency signal demodulation and analysis system that integrates LLM and Agent, provided for an embodiment of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] It should be noted that the RF signal demodulation and analysis method integrating LLM and Agent provided in this embodiment of the invention is executed by an RF integrated test instrument, a programmable logic controller, or a dedicated embedded microprocessor system with high real-time computing capabilities. The device is physically configured with a pickup and playback unit for multimodal voice interaction, a hardware board for edge inference, and an RF acquisition front-end.

[0022] Example 1 Please refer to the following: Figure 1-2This embodiment takes 5G RF mass production comprehensive test instrument as a typical application scenario, and elaborates in detail the complete execution logic of the large language model LLM, intelligent agent and lightweight AI demodulator when the underlying algorithm is implemented.

[0023] Step S1: The intelligent agent controls the radio frequency test instrument to collect the original in-phase quadrature IQ signals to construct a training dataset. Based on the multi-constraint joint loss function, the neural network demodulation basic model is trained offline. After the training converges, the model is lightweighted and deployed to the edge inference unit to obtain the AI ​​demodulator.

[0024] When the RF test instrument acquires raw IQ signals to construct a training dataset, the intelligent agent automatically plans stringent RF test conditions with a signal-to-noise ratio ranging from -10dB to 30dB, constructing a labeled training set containing 12,000 data points. The intelligent agent performs multi-task computing resource scheduling based on a preset test priority queue mechanism. The scheduling rules of the test priority queue are as follows: fault retesting tasks have the highest priority, followed by real-time demodulation inference tasks, then incremental model training tasks, and finally, sample acquisition tasks in idle states have the lowest priority. By exclusively using or suspending the computing cores of the edge inference hardware through the priority queue, conflicts over computing resources during multi-task concurrency are avoided.

[0025] During the model training phase, the multi-constraint joint loss function defined in this invention is used for optimization, and the total loss function is... The execution code logic satisfies: ; The weight parameters are fixed as follows: .

[0026] The specific calculation formulas for each sub-loss term are as follows: Timing loss ; Constellation distortion loss The constellation distortion loss is obtained by calculating the average Euclidean distance between the constellation point coordinates predicted by the model and the true coordinates of the ideal constellation.

[0027] Code element error correction loss The symbol error correction loss is obtained by performing cross-entropy loss calculation on the number of modulation symbol categories, one-hot encoding of symbol truth values, and model probability output.

[0028] EVM error constraint loss ; After training convergence, the model undergoes structured pruning, 8-bit quantization, and layer fusion lightweighting. The total number of model parameters for the AI ​​demodulator after lightweighting is controlled to within 1 million (876,000 in actual measurements), and the inference latency for a single alternating signal is controlled to within 10ms (9.2ms in actual measurements), allowing it to be directly embedded into the local edge inference board of the RF test instrument for operation.

[0029] Step S2: The radio frequency test instrument acquires the original IQ time domain signal of the device under test in real time.

[0030] The RF test instrument acquires the signal under test in real time, and its sampling rate adjustment range supports 1MHz to 20GHz. It converts high-frequency analog waveforms in the physical space into raw IQ time-domain digital signals and pushes them into the local buffer queue.

[0031] Step S3: The intelligent agent schedules the AI ​​demodulator based on the RF test finite state machine; the AI ​​demodulator performs end-to-end blind demodulation on the original IQ time domain signal, completes time-frequency synchronization and RF link impairment compensation through implicit feature learning of the model, and outputs baseband demodulation symbols and RF demodulation quantization indicators. The RF demodulation quantization indicators include error vector amplitude (EVM), local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, and IQ phase imbalance.

[0032] Specifically, the RF test finite state machine integrated in the intelligent agent kernel contains 7 standard core operating modes; the 7 standard core operating modes are defined in sequence as idle state, sample acquisition state, model training state, demodulation inference state, LLM semantic analysis state, parameter tuning and retesting state, and incremental learning state; the intelligent agent drives and controls the RF test instrument to switch between the 7 standard core operating modes according to the semantic mapping results of the received multimodal human-computer interaction instructions.

[0033] Please see Figure 3 When the state machine transitions to the demodulation inference state, the AI ​​demodulator consistently employs 3 convolutional layers (3×3 kernels), 2 pooling layers, 2 bidirectional LSTM layers (64 hidden layer dimensions), and a fully connected output layer. The original IQ time-domain signal is directly input into the network to perform end-to-end blind demodulation, implicitly achieving time-frequency synchronization and outputting baseband demodulation symbols. Simultaneously, the local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, and IQ phase imbalance, measured in parallel at the bottom layer of the network, are output as the RF demodulation quantization indicators.

[0034] Step S4: Input the RF demodulation quantization index into the Large Language Model (LLM); The LLM performs semantic analysis on the RF demodulation quantization index based on the fault chain reasoning template, determines the root cause of the RF link fault and the error vector amplitude (EVM) warning level, and outputs multimodal alarm information.

[0035] In practice, the local oscillator leakage amplitude, IQ amplitude imbalance, and IQ phase imbalance are uniformly fed into the large language model LLM to match the built-in radio frequency communication professional knowledge base to complete the step-by-step reasoning and tracing of the root cause of the fault.

[0036] The specific threshold division rule for the Large Language Model (LLM) to determine the EVM warning level is as follows: when the quantitative indicator meets EVM < 3.5%, it is determined to be at the normal level; when the quantitative indicator meets 3.5% ≤ EVM < 8%, it is determined to be at the alert level; when the quantitative indicator meets EVM ≥ 8%, it is determined to be at the abnormal level, and a hardware-level interrupt alarm is triggered immediately, linking the local broadcasting unit to perform offline broadcasting of the root cause of the fault through speech synthesis.

[0037] Step S5: The Large Language Model (LLM) autonomously deduces and generates an RF hardware tuning strategy based on the root cause of the fault and the preset three-level mapping rules. The intelligent agent parses the RF hardware tuning strategy into hardware programmable instructions and sends them to the front-end RF hardware to dynamically adjust the RF bias, channel gain and phase compensation parameters. After the parameter adjustment is completed, the agent automatically controls the RF test instrument to perform retest demodulation, forming an adaptive tuning closed loop.

[0038] The Large Language Model (LLM) generates hardware programmable instructions based on a preset three-level mapping rule of semantic-RF parameter-hardware instruction. The specific implementation steps are as follows: the generated natural language tuning strategy is converted into standard target parameter physical quantities, including DC bias, channel gain increment, and equalization filter coefficients; an intelligent agent converts these target parameter physical quantities into register write instructions for the microcontrollers of the attenuator and phase shifter at the front end of the RF test instrument, dynamically tuning the front-end hardware parameters with an adjustment step accuracy of ±0.01dB. After parameter adjustment, the instrument is automatically triggered to perform retest demodulation.

[0039] Step S6: The Large Language Model (LLM) intelligently filters out abnormal samples during the retesting process and supplements them with semantic labels for feedback. The intelligent agent schedules incremental iterative training to optimize the AI ​​demodulator.

[0040] In actual operation, the Large Language Model (LLM) intelligently filters out abnormal samples during the closed-loop retesting process and supplements them with specific semantic tags, which are then sent back to the database. When the intelligent agent transitions to the incremental learning state in the state machine, it schedules edge computing power to perform incremental iterative training to continuously optimize the AI ​​demodulator's ability to resist device aging and temperature drift.

[0041] Example 2 To verify the rationality and technical superiority of the fixed weight combination of the multi-constraint joint loss function of the present invention, this embodiment sets up multiple sets of weight ratio ablation experiments and conducts comparative tests under the same dataset, the same network structure and the same training hyperparameters.

[0042] The specific group settings for the comparative experiment are as follows: Control group 1: Using quaternary loss with equal weights, i.e., setting... .

[0043] Control group 2: A configuration that prioritizes timing loss was adopted, i.e., setting... .

[0044] This invention group: adopts the fixed weights defined in this invention, that is, sets... .

[0045] Experimental results show that, under the fixed weight combination defined in this invention, the error vector amplitude (EVM) test error is reduced by 1.8% compared to the control group in scenarios with low signal-to-noise ratio and strong interference. Simultaneously, the constellation distortion correction rate is significantly improved by 23.5%. Furthermore, the bit error rate of the demodulated symbols is reduced by an order of magnitude compared to the control group. These quantitative data demonstrate that the overall demodulation performance and fault index fit are optimal, confirming that the weight ratio defined in this invention is the optimal configuration specifically for RF IQ signal demodulation, possessing outstanding substantive features and significant progress.

[0046] Example 3 In high-concurrency scenarios such as mass production testing on RF production lines, RF test instruments typically need to simultaneously access and process multiple interactive requests, including local touch commands, voice query commands, and remote retest commands from the host computer.

[0047] The intelligent agent plays a global task scheduling role. First, the agent uses its built-in multimodal instruction semantic mapping rules to uniformly parse requests from diverse sources. Then, relying on the 7-state finite state machine and test priority queue mechanism in the kernel, the agent prioritizes fault retesting and real-time demodulation tasks, and proactively postpones or suspends incremental learning of the scheduling model and idle sample collection tasks. Dedicated or suspended computing power follows the aforementioned priority queue rules. The system successfully achieves zero conflict under multimodal instruction concurrency and reasonable allocation of edge computing resources, meeting the hard real-time requirement of ≤10ms latency for edge hardware.

[0048] Example 4 To further illustrate the semantic decision-making actions of the Large Language Model (LLM) and the underlying control process of the intelligent agent in this invention, this embodiment objectively verifies the parsing of natural language instructions, contextual association capabilities, and operational performance in a specific operating environment. Please refer to [the relevant documentation / references]. Figure 5 The diagram shows the system software interface and analysis results.

[0049] (1) Multimodal instruction and context-related resolution execution mechanism: In the system's interactive interface, natural language commands (e.g., ...) are received from the user. Figure 5 Example commands in the code include: "Acquire and analyze the current signal" or "Train CNN-LSTM to recognize 64QAM distortion". After receiving the command, the Large Language Model (LLM) performs contextual reasoning by combining historical diagnostic records with current device parameters. For example... Figure 5 As shown, the chained inference built into LLM performs the following steps: Step 1: Perform intent recognition to confirm that it is a data collection and analysis task; Step 2: Schedule the corresponding AI end-to-end demodulator for rapid feature extraction and index analysis; Step 3: Generate a comprehensive result feedback that includes time-domain IQ, the sampled spectrum, and constellation diagrams before and after correction.

[0050] (2) Agent device access and scheduling performance test: To address compatibility issues with RF test equipment, the Agent in this system is configured with IQ acquisition control command specifications through a skill library to adapt to different models of test equipment. Operation logs extracted under continuous concurrent testing conditions show that the Agent's success rate in controlling the underlying hardware to perform signal acquisition and retesting actions is 99.2%. The average latency from receiving the command to issuing the hardware parameter tuning action is 12ms. This latency, combined with the 9.2ms inference latency of the underlying AI demodulator, meets the real-time requirements of the system's closed-loop control.

[0051] (3) Verification of the accuracy of intent recognition and diagnosis: The quantitative test data of system performance obtained from test samples in a real mass production test environment are as follows: For non-standardized natural language input, LLM has a statistically low error rate of 2.3% in recognizing test intent and key parameters. In terms of end-to-end joint diagnosis, the overall accuracy of the system's reasoning results is 98.7%. Among the core performance indicators, the accuracy rate of EVM detection under 64QAM modulation was 98.7%, and the accuracy rate of IQ amplitude-phase imbalance diagnosis was 96.4%. The success rate of automatic calibration using the Agent based on the LLM strategy to perform hardware parameter compensation is 99.1%.

[0052] The above quantitative data verifies that this system can achieve closed-loop programming of RF test parameters through semantic analysis.

[0053] Example 5 Please see Figure 4 Based on the same inventive concept, this invention also provides a radio frequency signal demodulation and analysis system integrating LLM and Agent. This system is integrated and deployed within the radio frequency test instrument to implement the steps in any of the above method embodiments. Specifically, the system includes a large language model (LLM) embedded decision module, an intelligent agent global task scheduling module, an AI intelligent demodulation inference module, and an instrument closed-loop control optimization module; wherein: Large Language Model (LLM) Embedded Decision Module: Deployed locally on the RF test instrument, it is used to store the RF domain knowledge base and the fault chain reasoning template, receive the RF demodulation quantization indicators output by the AI ​​intelligent demodulation reasoning module, and perform semantic analysis and error vector amplitude (EVM) early warning level determination based on the fault chain reasoning template, output multimodal alarm information, and deduce and generate RF hardware tuning strategies.

[0054] The intelligent agent global task scheduling module interacts and communicates with the large language model LLM embedded decision module. It is used to store the mapping rules between the RF test finite state machine and the multimodal instruction, receive and parse multimodal human-computer interaction instructions, and control the RF test instrument to perform jump scheduling and edge inference hardware computing resource allocation between the 7 standard core operating modes of the finite state machine based on the preset test priority queue mechanism.

[0055] AI Intelligent Demodulation Inference Module: Deployed on the local edge inference board of the RF test instrument, it loads the AI ​​demodulator after pruning, quantization, and lightweighting. The AI ​​Intelligent Demodulation Inference Module receives the scheduling instructions from the global task scheduling module of the intelligent agent, accesses the input raw IQ time domain signal, performs end-to-end blind demodulation through model implicit feature learning, and outputs baseband demodulation symbols and a complete set of RF demodulation quantization indicators, including local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, IQ phase imbalance, and EVM, and pushes them to the large language model LLM embedded decision module.

[0056] The instrument closed-loop control optimization module receives hardware programmable commands forwarded by the intelligent agent global task scheduling module. These commands are obtained by mapping and transforming the RF hardware tuning strategy generated by the large language model LLM embedded decision module. The instrument closed-loop control optimization module connects to the front-end RF hardware and dynamically tunes the DC bias, channel gain increment, and equalization filter coefficients of the front-end RF hardware with a preset adjustment step precision. After the parameter adjustment is completed, it automatically triggers the RF test instrument to perform closed-loop retesting.

[0057] Example 6 This invention also provides a computer device, including: at least one processor; and a memory communicatively connected to the at least one processor. The memory stores program instructions; the at least one processor is configured to execute the computer program instructions to invoke its built-in arithmetic logic unit, floating-point arithmetic unit, and cache queue, and to perform large language model inference decision-making, agent task scheduling and flow, low-level AI demodulator execution, and hardware programmable instruction issuance actions by scheduling corresponding hardware boards, bus interfaces, and input / output interfaces, thereby realizing the various steps of the aforementioned radio frequency signal demodulation and analysis method integrating LLM and Agent.

[0058] Example 7 This invention also provides a non-volatile computer-readable storage medium storing computer binary executable code. When the computer binary executable code is read and executed by the central processing unit, the instruction pipeline of the driving instrument sequentially completes operations such as raw signal time-domain acquisition, agent multimodal scheduling, low-level blind demodulation and impairment compensation, and large language model closed-loop hardware instruction mapping and distribution, so as to realize the various steps of the above-mentioned RF signal demodulation and analysis method integrating LLM and Agent.

[0059] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0060] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0061] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0062] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.

Claims

1. A radio frequency signal demodulation and analysis method integrating LLM and Agent, characterized in that, The Large Language Model (LLM) and the Intelligent Agent are embedded and deployed locally on the RF test instrument; the LLM has a built-in RF domain knowledge base and a fault chain reasoning template; the Intelligent Agent has a built-in RF test finite state machine and multimodal instruction mapping rules; the method includes the following steps: S1. The intelligent agent controls the radio frequency test instrument to collect the original in-phase quadrature IQ signal to construct a training dataset. The neural network demodulation basic model is trained offline based on the multi-constraint joint loss function. The model after training convergence is then deployed to the edge inference unit after lightweight processing to obtain the AI ​​demodulator. S2. The radio frequency test instrument acquires the original IQ time domain signal of the test object in real time; S3. The intelligent agent schedules the AI ​​demodulator based on the RF test finite state machine; the AI ​​demodulator performs end-to-end blind demodulation on the original IQ time domain signal, completes time-frequency synchronization and RF link impairment compensation through implicit feature learning of the model, and outputs baseband demodulation symbols and RF demodulation quantization indicators. The RF demodulation quantization indicators include error vector amplitude EVM, local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, and IQ phase imbalance. S4. Input the RF demodulation quantization index into the Large Language Model (LLM); The LLM performs semantic analysis on the RF demodulation quantization index based on the fault chain reasoning template, determines the root cause of the RF link fault and the error vector amplitude (EVM) warning level, and outputs multimodal alarm information. S5. The Large Language Model (LLM) autonomously deduces and generates an RF hardware tuning strategy based on the root cause of the fault and the preset three-level mapping rules. The intelligent agent parses the RF hardware tuning strategy into hardware programmable instructions and sends them to the front-end RF hardware to dynamically adjust the RF bias, channel gain and phase compensation parameters. After the parameter adjustment is completed, the agent automatically controls the RF test instrument to perform retest demodulation, forming an adaptive tuning closed loop. S6. The large language model LLM intelligently filters out abnormal samples during the retesting process and supplements them with semantic labels for feedback. The intelligent agent schedules incremental iterative training to optimize the AI ​​demodulator.

2. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In step S1, the AI ​​demodulator is built based on a hybrid network structure of convolutional neural network (CNN) and long short-term memory network (LSTM). During the model training phase, a multi-constraint joint weighted constraint is adopted, consisting of temporal loss, constellation distortion loss, symbol error correction loss, and EVM error constraint loss. The weights of the total loss function are fixed as follows: temporal loss weight α=0.25, constellation distortion loss weight β=0.35, symbol error correction loss weight γ=0.25, and EVM error constraint loss weight δ=0.

15. The constellation distortion loss is obtained by calculating the average Euclidean distance between the constellation point coordinates predicted by the model and the true coordinates of the ideal constellation; the symbol error correction loss is obtained by performing cross-entropy loss calculation on the number of modulation symbol categories, one-hot encoding of symbol truth, and model probability output.

3. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 2, characterized in that: In step S1, when controlling the radio frequency test instrument to collect raw IQ signals to construct a training dataset, the intelligent agent performs multi-task computing resource scheduling based on a preset test priority queue mechanism. The scheduling rules for the test priority queue are as follows: the fault retest task has the highest priority, followed by the real-time demodulation inference task, then the incremental training task of the model, and the sample collection task in the idle state has the lowest priority. By prioritizing or suspending the computing cores of edge inference hardware, conflicts over computing resources can be avoided during multi-task concurrency.

4. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In steps S1 and S3, the AI ​​demodulator after training convergence undergoes structured pruning, low-bit quantization, and layer fusion lightweight processing before deployment. The total number of model parameters of the AI ​​demodulator after lightweight processing is controlled within 1 million, and the inference delay of a single alternating signal is controlled within 10ms, so that it can be directly embedded into the local edge inference board of the RF test instrument for operation.

5. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In step S3, the radio frequency test finite state machine integrated in the intelligent agent kernel contains 7 standard core operating modes; The seven standard core operating modes are defined sequentially as idle state, sample acquisition state, model training state, demodulation inference state, LLM semantic analysis state, parameter tuning and retesting state, and incremental learning state. The intelligent agent drives and controls the RF test instrument to switch between the seven standard core operating modes based on the semantic mapping results of the received multimodal human-computer interaction commands.

6. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In step S3, when the AI ​​demodulator performs end-to-end blind demodulation on the original IQ time domain signal and outputs the RF demodulation quantization index, it simultaneously outputs the local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance and IQ phase imbalance measured by quantization at the network bottom layer in parallel. In step S4, the local oscillator leakage amplitude, IQ amplitude imbalance, and IQ phase imbalance are uniformly fed into the large language model LLM to match the built-in radio frequency communication professional knowledge base to complete the step-by-step reasoning and tracing of the root cause of the fault.

7. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In step S4, the specific threshold division rule for the Large Language Model (LLM) to perform the EVM early warning level determination is as follows: When the quantitative indicator meets the condition that EVM < 3.5%, it is judged to be at the normal level; When the quantitative indicator meets the condition of 3.5%≤EVM<8%, it is judged as a warning level; When the quantitative indicator meets the requirement of EVM≥8%, it is judged as an abnormal level and an immediate hardware-level interrupt alarm is triggered, which in turn triggers the local broadcasting unit to perform offline speech synthesis and broadcasting of the root cause of the fault.

8. The radio frequency signal demodulation and analysis method integrating LLM and Agent according to claim 1, characterized in that: In step S5, the specific implementation actions of the Large Language Model (LLM) generating hardware programmable instructions based on the preset semantic-RF parameter-hardware instruction three-level mapping rules are as follows: The generated natural language tuning strategy is converted into standard target parameter physical quantities, which include DC bias, channel gain increment and equalization filter coefficients. The target parameter physical quantity is converted into a write instruction to the register of the microcontroller of the attenuator and phase shifter at the front end of the RF test instrument by the intelligent agent, and the front end hardware parameters are dynamically tuned with an adjustment step accuracy of ±0.01dB.

9. A radio frequency signal demodulation and analysis system integrating LLM and Agent, characterized in that, Integrated and deployed within an RF test instrument, this method implements the RF signal demodulation and analysis method merging LLM and Agent as described in any one of claims 1 to 8, comprising a large language model (LLM) embedded decision module, an intelligent agent global task scheduling module, an AI intelligent demodulation inference module, and an instrument closed-loop control optimization module; wherein: The Large Language Model (LLM) embedded decision module is deployed locally on the RF test instrument. It is used to store the RF domain knowledge base and the fault chain reasoning template, receive the RF demodulation quantization indicators output by the AI ​​intelligent demodulation reasoning module, and perform semantic analysis and error vector amplitude (EVM) early warning level determination based on the fault chain reasoning template. It outputs multimodal alarm information and infers and generates RF hardware tuning strategies. The intelligent agent global task scheduling module interacts and communicates with the large language model LLM embedded decision module to solidify and store the mapping rules between the RF test finite state machine and the multimodal instruction, receive and parse multimodal human-computer interaction instructions, and control the RF test instrument to perform jump scheduling and edge inference hardware computing resource allocation between the 7 standard core operating modes of the finite state machine based on the preset test priority queue mechanism. The AI ​​intelligent demodulation inference module is deployed on the local edge inference board of the RF test instrument and loads the AI ​​demodulator after pruning, quantization and lightweight processing. The AI ​​intelligent demodulation inference module receives the scheduling instructions of the global task scheduling module of the intelligent agent, accesses the input raw IQ time domain signal, performs end-to-end blind demodulation through model implicit feature learning, and outputs baseband demodulation symbols and a complete set of RF demodulation quantization indicators including local oscillator leakage amplitude, in-phase quadrature IQ amplitude imbalance, IQ phase imbalance and EVM in parallel, and pushes them to the large language model LLM embedded decision module. The instrument closed-loop control optimization module receives hardware programmable instructions forwarded by the intelligent agent global task scheduling module. These hardware programmable instructions are obtained by mapping and transforming the RF hardware tuning strategy generated by the large language model LLM embedded decision module. The instrument closed-loop control optimization module is connected to the front-end RF hardware and dynamically tunes the DC bias, channel gain increment, and equalization filter coefficient of the front-end RF hardware with a preset adjustment step precision. After the parameter adjustment is completed, it automatically triggers the RF test instrument to perform closed-loop retesting.

10. A computer device, characterized in that, include: At least one processor; as well as A memory that is communicatively connected to the at least one processor; The memory is used to store computer program instructions; The at least one processor is configured to execute the computer program instructions to schedule corresponding hardware boards, bus interfaces, and arithmetic logic units, thereby implementing the various steps of the radio frequency signal demodulation and analysis method integrating LLM and Agent as described in any one of claims 1 to 8.

Images