Simulation signal combat training system and training method thereof

Abstract

The invention relates to the technical field of military simulation and training, discloses a training system for analog signal combat and a training method thereof, and aims to solve the problems of insufficient confrontation authenticity, strategy solidification, cognitive depth deficiency and weak multi-domain cooperative capability of a traditional training system. According to the method, a dynamic confrontation deduction architecture with linkage of a physical domain, an electromagnetic domain and a cognitive domain is constructed, a distributed radio frequency array, a programmable attenuator and a multipath reflection simulator are deployed to realize high-fidelity signal propagation simulation, and full-band signal perception and feature extraction are completed in combination with a broadband digital receiver cluster and a central data fusion platform. An intelligent confrontation agent cluster driven by deep reinforcement learning and a cross-domain collaborative decision center are introduced; according to the technical scheme, the authenticity of the training environment, the intelligent adaptability of confrontation, the objectivity of evaluation and the sustainable evolution ability of the system are remarkably improved, and the signal game, dynamic response and cooperative command ability of trainees are effectively hammered.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of military simulation and training, and particularly relates to a training system for simulating signal combat and a training method thereof. BACKGROUND

[0002] With the accelerated evolution of modern war patterns towards informatization and intelligentization, simulating signal combat training has become a core means to improve the electronic countermeasure capability and battlefield situation awareness level of troops. Traditional training systems mostly rely on pre-set scripts and fixed signal parameters to construct the countermeasure scene, and the underlying logic is based on a static electromagnetic environment model and linear deduction rules, which is difficult to truly restore the dynamic countermeasure behavior of signal jamming, spectrum hopping and multi-source deception in complex battlefield. In a high-intensity electromagnetic countermeasure environment, the signal features of the enemy evolve in real time with the tactical intent and equipment state, while the existing systems lack the ability to adaptively model the signal behavior patterns, often leading to distorted training scenes, insufficient countermeasure intensity, and delayed tactical feedback, making it difficult for trainees to establish a deep understanding of real battlefield signal game. In addition, commanders have increasing needs for post-mortem analysis of the training process, tactical effectiveness evaluation and personalized subject customization. Traditional systems cannot support multi-level, multi-role and multi-round deep countermeasure deduction due to single data dimension, difficulty in behavior tracing and rigid scene reconstruction.

[0003] Among them, the training method for simulating signal combat focuses on constructing countermeasure signal bodies with dynamic evolution ability and interactive tactical decision chains, and the core goal is to realize immersive countermeasure training in the whole process of "detection-judgment-disturbance-evaluation" through high-fidelity signal behavior simulation and closed-loop tactical feedback mechanism. This direction requires the system not only to simulate the emission characteristics and modulation strategies of enemy radar, communication and navigation multi-system signals, but also to adjust the countermeasure intensity and deception mode in real time according to the trainee's operation, forming an intelligent countermeasure entity with tactical intent.

[0004] The existing technology generally has three major defects of single signal behavior modeling dimension, lack of adaptive evolution of countermeasure strategy, and disconnection between training process and tactical evaluation. Most systems can only simulate static signal sources with fixed frequency bands and fixed modulation modes, and cannot dynamically adjust signal parameters and emission strategies according to the battlefield situation; the countermeasure engine mostly adopts rule-driven mode, lacks behavior prediction and intent deduction ability based on machine learning, leading to mechanical repetition of countermeasure behavior and lack of tactical depth; the training data collection is limited to signal transceiver logs, and the closed-loop feedback chain of "operation-response-effectiveness-evaluation" cannot be established, resulting in superficial post-mortem analysis and inability to accurately locate tactical shortfalls. Under the background of joint operations and multi-domain cooperation, the above defects seriously restrict the ability of troops to quickly identify threats, flexibly develop countermeasures and cooperatively implement electronic attack and defense in complex electromagnetic environments, and there is an urgent need to build a simulated signal combat training system with intelligent signal evolution, dynamic countermeasure deduction and deep tactical evaluation capabilities. SUMMARY

[0005] To solve the above technical problems, the present application is proposed. Embodiments of the present application provide a training system for simulating signal operations and a training method thereof, which realizes high-fidelity simulation and adaptive evolution of signal game behavior in a complex battlefield environment by constructing a multi-level dynamic counter-deployment architecture and integrating real-time interaction mechanisms in physical, electromagnetic, and cognitive domains. The system takes a distributed agent cluster as the core carrier, relies on a time-space synchronization engine and a multi-source heterogeneous data fusion hub, establishes a closed-loop training environment covering the whole chain of signal generation, propagation, interference, identification, decision-making, and feedback, enables trainees to hone their signal countermeasures formulation ability, dynamic response ability, and collaborative command ability in a highly realistic virtual battlefield, and thus breaks through the fundamental limitations of traditional static scripted training modes in terms of confrontation reality, strategy diversity, and cognitive depth.

[0006] According to an aspect of the present application, a control method mentioned in the training system for simulating signal operations and the training method thereof is provided, which includes: By deploying a distributed radio frequency signal transceiver array in a physical space, a three-dimensional electromagnetic field simulation environment covering a specified geographic area is constructed, wherein each radio frequency unit is configured with an independent frequency modulation module, a power control module, and a directional pattern reconfigurable antenna structure, supporting dynamic generation of simulated radiation source signals with specific modulation patterns, time domain envelopes, and spatial directivity in a specified frequency band.

[0007] By embedding a programmable attenuator array and a multipath reflection simulator in the electromagnetic propagation path, the path loss, multipath time delay spread, Doppler shift, and polarization rotation effects are calculated and applied in real time according to the preset terrain data and atmospheric refraction model, so that the simulated signals exhibit consistent fading characteristics and spatial dispersion behavior with the real battlefield environment during propagation.

[0008] By deploying a wideband digital receiver cluster at the receiving end, each receiver integrates a high-speed analog-to-digital converter, a digital down-conversion module, and a real-time spectrum analysis engine, which are used for full-band continuous sampling, channelization processing, and instantaneous spectral feature extraction of the superimposed multi-source signals in space, and synchronously upload the original time domain waveform data and frequency domain feature vectors to the central data fusion platform.

[0009] By constructing an intelligent adversarial agent cluster based on a deep reinforcement learning framework, each agent is configured with an independent state perception module, a policy generation module, and an action execution module. The state perception module receives multi-dimensional battlefield situation information from the central data fusion platform, including enemy and friendly signal source positions, spectrum occupation maps, interference intensity distribution, and communication link quality indicators. The policy generation module outputs optimal signal transmission parameter adjustment instructions, interference strategy selection instructions, or evasion maneuver instructions based on the current situation and historical interaction experience. The action execution module converts the instructions into specific radio frequency control signals and distributes them to the corresponding physical devices for execution.

[0010] By establishing a cross-domain collaborative decision-making hub, the hub receives local decision proposals from all intelligent adversarial agents, resolves conflicts, allocates resources, and optimizes global strategies based on pre-set command hierarchy and collaboration rule library. Finally, it generates a unified sequence of operational instructions and synchronously distributes them to each execution unit, ensuring that the multi-agent system maintains tactical consistency and strategic synergy in complex adversarial environments.

[0011] By constructing a cognitive domain behavior modeling engine, the engine collects speech instructions, operation logs, decision delays, and situation annotation data of participating commanders during training. Using natural language processing and behavior sequence mining algorithms, it constructs a commander cognitive load model, a decision preference model, and a tactical style portrait. The modeling results are input as feedback signals to the policy generation module of the intelligent adversarial agent, allowing it to dynamically adjust the intensity and complexity of the strategy, achieving adaptive matching of training difficulty.

[0012] By deploying a real-time performance evaluation module, the module quantifies the comprehensive performance of participating parties and adversarial agents based on pre-set multi-dimensional evaluation index system, including signal suppression success rate, communication link maintenance duration, spectrum resource utilization rate, tactical target achievement degree, and cognitive decision efficiency, after each adversarial round ends. It generates a visual evaluation report, providing data support for subsequent training plan optimization.

[0013] By constructing a historical adversarial case library and a strategy evolution engine, the engine continuously records signal interaction sequences, decision paths, and environmental parameters in each training adversarial, uses graph neural networks to structure model historical cases, extracts high-frequency successful strategy patterns and typical failure causes, and iteratively optimizes the strategy space of intelligent adversarial agents through genetic algorithms. This makes the system have the ability of continuous self-evolution, avoiding the training content from falling into pattern fixation.

[0014] According to another aspect of the present application, a control system for a simulated signal warfare training system and a training method thereof is provided, which includes: The distributed radio frequency signal transceiving array is used for arranging a plurality of independently controllable radio frequency transmitting and receiving nodes in a preset topological structure in a three-dimensional space. Each node comprises a radio frequency front-end module, a baseband processing module, and a time-space synchronization module. The radio frequency front-end module supports continuous tuning in a frequency band of 300 MHz to 6 GHz, and the output power dynamic range covers -30 to +30 dBm. The baseband processing module supports generating twelve standard modulation patterns including frequency modulation, phase modulation, and quadrature amplitude modulation. The time-space synchronization module realizes nanosecond-level time alignment of the whole system through a Beidou timing signal and a fiber time distribution network.

[0015] The programmable attenuator array is used for deploying radio frequency devices with dynamic attenuation characteristics at key nodes of an analog signal propagation path. Each attenuator unit supports program-controlled adjustment in a range of 0 to 120 dBm with a step of 0.5 dBm. The control instructions are generated by a propagation model calculation module in real time according to current analog terrain elevation data, vegetation coverage density, and atmospheric humidity parameters, ensuring that the signal attenuation characteristics are strictly coupled with the environmental physical parameters.

[0016] The multipath reflection simulator is used for deploying a plurality of movable reflecting surface units in a specified spatial region. Each reflecting surface unit is composed of a metal mesh structure and a servo drive mechanism. The servo drive mechanism adjusts the spatial azimuth and elevation angles of the reflecting surface in real time according to the reflection path calculated by a ray tracing algorithm, so that the incident signal produces specular reflection and diffuse reflection components in accordance with the principles of geometric optics. The reflection coefficient and polarization characteristics can be dynamically configured according to the material electromagnetic parameter library.

[0017] The wideband digital receiver cluster is used for deployment at key listening locations in a simulated battlefield. Each receiver comprises a 12-bit analog-to-digital converter with a sampling rate of 5 GHz, a digital down-conversion channel with a real-time bandwidth of 200 MHz, and a frequency spectrum analysis unit based on fast Fourier transform. The output data is uploaded in real time to a central data fusion platform through a gigabit Ethernet interface, and the data format complies with the unified timestamp marking and channel identification protocol.

[0018] The intelligent counter-agent cluster is used for deploying a plurality of software entities with autonomous decision-making capabilities in a virtual command hierarchy. Each agent entity runs in an independent computing container. The state perception module subscribes to the battlefield situation updates published by the central data fusion platform through a message queue. The strategy generation module constructs an action value function using the deep deterministic policy gradient algorithm. The action execution module sends parameter adjustment instructions to the physical device control layer through a remote procedure call interface, and the instruction execution status is confirmed in real time through a heartbeat packet mechanism.

[0019] The cross-domain cooperative decision-making hub is configured to deploy a global strategy coordinator at the top layer of the system architecture, the coordinator including a conflict detection module, a resource arbitration module and an instruction synthesis module, the conflict detection module performing time-space overlap analysis and target consistency verification according to decision proposals submitted by each agent, the resource arbitration module allocating spectrum, power and space direction resources according to preset priority rules and resource occupation state tables, and the instruction synthesis module encapsulating arbitration results into standardized instruction packets and distributing the instruction packets to each execution unit through a reliable multicast protocol.

[0020] The cognitive domain behavior modeling engine is configured to be deployed on a training monitoring server, the engine including a speech recognition subsystem, an operation log analysis subsystem and a decision trajectory mining subsystem, the speech recognition subsystem converting commander voice instructions into structured semantic labels by using an end-to-end neural network model, the operation log analysis subsystem extracting key operation events and response delays according to a time series alignment algorithm, and the decision trajectory mining subsystem performing clustering analysis on behavior patterns of commanders at different battlefield stages by using a hidden Markov model, and outputting a cognitive load index and a tactical tendency vector.

[0021] The real-time performance evaluation module is configured to automatically trigger an evaluation process after each confrontation round ends, the module including an index calculation submodule, a weight configuration submodule and a report generation submodule, the index calculation submodule calculating twenty-eight preset quantitative indexes according to original sensor data and system logs, the weight configuration submodule supporting a trainer to dynamically adjust weight coefficients of each index according to a training target, and the report generation submodule outputting evaluation results to a visualization terminal in the form of a radar chart, a heat map and a trend curve.

[0022] The historical confrontation case library is configured to persistently store complete data records of each training confrontation, including original radio frequency waveforms, spectrum feature vectors, agent decision sequences, environment parameter snapshots and evaluation result summaries, and a columnar database structure is adopted for data storage, supporting multi-dimensional retrieval and correlation analysis according to time ranges, confrontation types, participating units and tactical targets.

[0023] The strategy evolution engine is configured to periodically extract strategy patterns from successful confrontation samples in the historical case library, the engine including a graph structure construction module, a strategy pattern mining module and a genetic optimization module, the graph structure construction module modeling each confrontation process as a state transition graph, with nodes representing battlefield situation snapshots and edges representing agent decision actions, the strategy pattern mining module identifying high-frequency subgraph structures by using a graph convolution network, and the genetic optimization module encoding the identified strategy patterns into chromosome individuals, generating a new generation of strategy populations by crossover, mutation and fitness selection operations and injecting the new generation of strategy populations into the intelligent confrontation agent cluster.

[0024] Compared with the prior art, the application has the advantages and positive effects that: First, the system first realizes the deep coupling and closed-loop interaction of the physical domain radio frequency signal generation and propagation process, the electromagnetic domain interference and anti-interference game, and the cognitive domain command decision behavior, breaking through the architectural limitations of traditional training systems that separate signal simulation, electronic countermeasures, and command and control, making the training environment have the multi-domain linkage characteristics and dynamic evolution ability of real battlefield.

[0025] Second, by deploying intelligent countermeasure agent clusters with independent decision-making capabilities and introducing a deep reinforcement learning framework, the system can dynamically adjust the countermeasure strategy according to the real-time performance of the participants, avoiding the problem of fixed countermeasure behavior and lack of adaptability in traditional scripted training, significantly improving the challenge and unpredictability of training, and effectively honing the participants' on-the-spot response ability.

[0026] Third, the introduction of the cognitive domain behavior modeling engine enables the system to quantitatively evaluate the cognitive load and decision-making efficiency of commanders and adaptively adjust the countermeasure intensity and complexity accordingly, achieving precise matching of training difficulty and participant ability level, avoiding frustration caused by excessive difficulty or training inefficiency caused by low difficulty, and improving the utilization efficiency of training resources and the smoothness of individual ability growth curve.

[0027] Fourth, the collaborative work of the historical countermeasure case library and the strategy evolution engine enables the system to have continuous self-optimization capability, automatically extracting high-frequency successful strategies and typical failure modes from massive training data, and continuously iterating and optimizing the strategy space of intelligent countermeasure agents through genetic algorithms, ensuring that the training content is always at the forefront of tactics and avoiding the decline of training effectiveness due to tactical solidification.

[0028] Fifth, the collaborative control of distributed radio frequency arrays, programmable attenuators, and multipath reflection simulators enables the system to accurately reproduce the electromagnetic propagation characteristics of complex terrain, urban canyons, and jungle sheltered typical battlefield environments in a laboratory environment, without relying on external field equipment for high-fidelity signal countermeasure training, significantly reducing training costs and organizational implementation difficulty.

[0029] Sixth, the multi-dimensional quantitative index system provided by the real-time performance evaluation module enables the training effect evaluation to be driven by objective data instead of subjective experience, and the instructor can accurately identify the weaknesses of the participants in signal recognition, interference avoidance, resource scheduling, and collaborative command through radar charts and heat map visualization tools, providing a clear direction for subsequent targeted reinforcement training.

[0030] Seventh, the introduction of the cross-domain collaborative decision-making hub enables the multi-agent system to maintain tactical consistency and strategic synergy in complex countermeasure environments, avoiding local optimal decisions that conflict with global goals, ensuring that both sides in the training process follow reasonable command levels and coordination rules, and improving the tactical rationality and command system integrity of the training scene.

[0031] Eighth, the system architecture adopts modular design and standardized interface protocol, supports flexible increase and decrease of the number of radio frequency nodes, adjustment of the size of the adversarial agent or replacement of specific functional modules according to training needs, has good scalability and customizability, and can adapt to the diversified needs of different military branches, different combat scenarios and different training stages.

[0032] Ninth, all physical device control instructions are uniformly scheduled and state-monitored through the central data fusion platform, ensuring the stability and safety of system operation, avoiding training interruption or device damage due to device out of control or instruction conflict, while supporting remote fault diagnosis and parameter rollback, improving system operation efficiency and reliability.

[0033] Tenth, by incorporating the operation behavior, voice instruction and decision trajectory of the trainees into the system feedback loop, the training process not only exercises the technical operation ability, but also strengthens the command cognitive ability and tactical thinking ability, realizing the paradigm upgrade from "skill training" to "cognitive training", and providing technical support for cultivating compound command talents who can adapt to future intelligent warfare. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is the overall technical scheme architecture schematic diagram of the training system and the training method of the analog signal combat proposed by the present application; Figure 2 is the core principle framework schematic diagram of the cooperative evolution of the intelligent adversarial agent cluster and the cognitive domain behavior modeling engine in the present application. DETAILED DESCRIPTION

[0035] To realize high-fidelity simulation and adaptive evolution of signal game behavior in complex battlefield environment, the present embodiment constructs a training system and a training method for simulating signal combat. The method realizes the three-domain closed-loop interaction mechanism of physical domain, electromagnetic domain and cognitive domain, relies on distributed intelligent agent cluster and multi-source heterogeneous data fusion hub, and establishes a dynamic adversarial deduction architecture covering the whole chain of signal generation, propagation, interference, identification, decision and feedback. The system level is deployed with modular hardware units and standardized software entities, which supports the reproduction of real battlefield electromagnetic environment in laboratory environment, and realizes the systematic training of the trainees' adversarial strategy formulation ability, dynamic response ability and cooperative command ability.

[0036] Please refer to Figure 1 and Figure 2 , the method comprises the following steps: S1, by deploying a distributed radio frequency signal transceiver array in a physical space, a three-dimensional electromagnetic field simulation environment covering a specified geographic area is constructed, wherein each radio frequency unit is configured with an independent frequency modulation module, a power control module and a directional pattern reconfigurable antenna structure, supporting dynamically generating simulated radiation source signals with specific modulation patterns, time domain envelopes and spatial directivity in a specified frequency band.

[0037] In the S1 step, first, according to the training scene requirements, a plurality of radio frequency transmitting and receiving nodes are laid out in a three-dimensional space according to a preset topological structure. Each node includes a radio frequency front-end module, a baseband processing module and a time-space synchronization module. The radio frequency front-end module supports continuous tuning in the three hundred megahertz to six gigahertz frequency band, and the output power dynamic range covers negative thirty to positive thirty dBm, the frequency tuning step precision is ten kilohertz, and the power adjustment resolution is zero point one dBm. The baseband processing module has twelve standard modulation pattern generators built-in, including frequency modulation, phase modulation, and quadrature amplitude modulation, each modulation pattern supports independent configuration of symbol rate, roll-off coefficient and constellation mapping rule. The directional pattern reconfigurable antenna structure is composed of eight element phased arrays, supporting dynamic adjustment of beam pointing in the range of zero to three hundred sixty degrees in azimuth angle and negative ninety to ninety degrees in elevation angle, and the beam main lobe width can be continuously adjusted between five to sixty degrees according to task requirements.

[0038] In the signal generation process, the system allocates independent carrier frequency, transmission power, modulation parameter and spatial pointing angle for each radio frequency unit according to the instructions issued by the central data fusion platform. The baseband processing module generates corresponding baseband I / Q data stream according to the instructions, which is sent to the radio frequency front-end module for up-conversion and power amplification after digital-to-analog conversion. The time-space synchronization module realizes nanosecond-level time alignment of the whole system through Beidou timing signal and fiber time distribution network, ensuring that all radio frequency units transmit signals synchronously under the unified time reference. Each radio frequency unit performs self-checking process before transmission, including power calibration, frequency locking state detection and antenna standing wave ratio measurement, if any parameter exceeds the preset threshold, local alarm is triggered and transmission is suspended, at the same time, the abnormal state code is reported to the central data fusion platform.

[0039] S2, by embedding programmable attenuator array and multipath reflection simulator in the electromagnetic propagation path, according to the preset terrain data and atmospheric refraction model, the atmospheric refraction model includes the following two core parts: 1. Refractive Index Gradient Model This is the key to realize the "refraction effect". The system estimates the refractive index of the atmosphere through real-time collected atmospheric temperature and humidity data; using these data to calculate the gradient of the refractive index with height or space (i.e. ); This gradient is used to correct the propagation path (bending) of electromagnetic waves, thereby affecting: Doppler shift; signal arrival time; reception angle.

[0040] Here is a model of refractive index based on the spatial distribution of temperature, humidity, and air pressure, usually in the form of:

[0041] Where: Absolute temperature (K), Total air pressure (hPa), Water vapor pressure (hhPa).

[0042] This formula is the standard model recommended by the International Telecommunication Union (ITU), also known as the ITU-R P834 model; 2, Atmospheric Absorption Loss Model, the absorption of specific frequency electromagnetic waves by air molecules (such as oxygen O2, water vapor H2O). The absorption intensity is different at different frequencies (for example: 22GHz and 60GHz are strong absorption bands of water vapor) and is related to path length, humidity, and temperature.

[0043] Common models include: ITU-R P.676: Standard Model of Atmospheric Absorption Attenuation; Mie scattering model: suitable for rain and fog particle scattering; Hagmann model: for millimeter wave communication.

[0044] The atmospheric refraction model is actually a comprehensive propagation environment model that not only includes refraction but also absorption.

[0045] Real-time calculation and application of path loss, multipath time delay spread, Doppler shift, and polarization rotation effects make the simulated signal exhibit consistent fading characteristics and spatial dispersion behavior with real battlefield environment during propagation.

[0046] In the S2 step, the system first loads the digital elevation model and land cover type map corresponding to the current training scene, combines real-time collected atmospheric temperature and humidity data to construct a three-dimensional spatial propagation environment parameter field. The propagation model calculation module uses ray tracing algorithm to geometrically model all possible propagation paths of each radio frequency signal from the transmitting point to the receiving point, and calculates the free space loss, diffraction loss, reflection loss, and atmospheric absorption loss of each path. For each path, the system calculates the Doppler shift based on the path length and atmospheric refractive index gradient, and calculates the polarization rotation matrix based on the incident angle and reflection surface material properties.

[0047] A programmable attenuator array is deployed at key nodes of the propagation path, each attenuator unit supports program-controlled adjustment in the range of zero to one hundred and twenty dBm with zero point five dBm steps. The propagation model calculation module decomposes the total loss value of each path into several segments and assigns them to the attenuator units at the corresponding locations, generating a sequence of attenuation control instructions. The instruction sequence contains target attenuation values, adjustment rates, and effective time stamps, which are sent to each attenuator unit through a dedicated control bus. After receiving the instructions, the attenuator unit adjusts the attenuation value at the specified time point and at the specified rate, and confirms the execution status to the propagation model calculation module through the feedback channel.

[0048] The multipath reflection simulator deploys multiple movable reflective surface units in a specified spatial region, each reflective surface unit is composed of a metal mesh structure and a servo drive mechanism. The servo drive mechanism adjusts the spatial azimuth and elevation angles of the reflective surface in real time according to the reflection path calculated by the ray tracing algorithm, so that the incident signal produces specular reflection and diffuse reflection components consistent with the principles of geometric optics. The reflection coefficient and polarization characteristics are dynamically configured based on the material electromagnetic parameter library, supporting electromagnetic response simulation of multiple typical battlefield materials such as copper, aluminum, concrete, and vegetation. The servo drive mechanism has a built-in position closed-loop controller to ensure that the angle adjustment error of the reflective surface is less than zero point one degree, and the angle adjustment response time is less than fifty milliseconds. The system performs collision detection before each angle adjustment, and if the target angle will cause mechanical interference, it automatically selects a suboptimal angle and reports a conflict log.

[0049] S3, by deploying a wideband digital receiver cluster at the receiving end, each receiver integrates a high-speed analog-to-digital converter, a digital down-conversion module, and a real-time spectrum analysis engine, for full-band continuous sampling, channelization processing, and instantaneous spectral feature extraction of the superimposed multi-source signals in space, and synchronously uploading the original time-domain waveform data and frequency-domain feature vectors to the central data fusion platform.

[0050] In the S3 step, the wideband digital receiver cluster is deployed at key listening locations in the simulated battlefield, each receiver contains a twelve-bit analog-to-digital converter with a sampling rate of five gigahertz, a digital down-conversion channel supporting a real-time bandwidth of two hundred megahertz, and a spectrum analysis unit based on fast Fourier transform. The receiver antenna receives the superimposed multi-source radio frequency signals in space, which are sent to the analog-to-digital converter for digital sampling after low-noise amplification and anti-aliasing filtering. The sampled data stream enters the digital down-conversion module, which performs digital mixing and decimation filtering according to the pre-set center frequency and bandwidth parameters, and outputs the baseband I / Q data stream.

[0051] The real-time spectrum analysis engine processes the baseband I / Q data stream in segments, each segment being 8,192 points long, and performs a fast Fourier transform after windowing with a Hanning window function to obtain the frequency domain amplitude spectrum and phase spectrum. The spectrum analysis engine further extracts the instantaneous spectral feature vector, including the spectral centroid, spectral entropy, peak frequency, bandwidth occupancy rate, and modulation recognition features. The original time domain waveform data and frequency domain feature vector are uploaded in real time to the central data fusion platform through a gigabit Ethernet interface, and the data format complies with a unified timestamp marking and channel identification protocol. Each data packet contains a sampling start timestamp, a receiver identifier, a center frequency, a bandwidth parameter, an original I / Q data block, and a feature vector array. The receiver has a built-in data buffering mechanism that can temporarily store up to ten seconds of data when the network is congested, and when the network recovers, the data is transmitted in time sequence to ensure data integrity.

[0052] S4, by constructing an intelligent adversarial agent cluster based on a deep reinforcement learning framework, each agent is configured with an independent state perception module, a policy generation module, and an action execution module. The state perception module receives multi-dimensional battlefield situation information from the central data fusion platform, including enemy and friendly signal source positions, spectrum occupation maps, interference intensity distribution, and communication link quality indicators. The policy generation module outputs the optimal signal transmission parameter adjustment instructions, interference strategy selection instructions, or evasion maneuver instructions based on the current situation and historical interaction experience. The action execution module converts the instructions into specific radio frequency control signals and issues them to the corresponding physical devices for execution.

[0053] In the S4 step, the intelligent adversarial agent cluster deploys multiple software entities with autonomous decision-making capabilities in the virtual command hierarchy, and each agent entity runs in an independent computing container. The state perception module subscribes to battlefield situation updates published by the central data fusion platform through a message queue. The situation information includes enemy signal source three-dimensional coordinates, spectrum occupation heat maps, interference signal intensity spatial distribution maps, own communication link signal-to-noise ratio time series curves, and resource occupation state tables. The state perception module preprocesses the original situation data, including coordinate system conversion, data normalization, and feature splicing, to generate a uniform dimension state vector input to the policy generation module.

[0054] The policy generation module adopts a deep deterministic policy gradient algorithm to construct an action value function, and a neural network structure thereof includes three fully connected layers, each layer having 512, 256 and 128 neurons respectively, and a rectified linear unit as an activation function. The network input is a current state vector, and the output is a continuous action vector including a carrier frequency adjustment amount, a transmission power adjustment amount, a beam pointing angle adjustment amount, a modulation pattern index and an interference strategy identifier. The policy generation module performs forward inference in each decision cycle to generate an optimal action vector, and adds random disturbance according to an exploration noise mechanism to maintain policy diversity. The training of the action value function adopts an experience replay mechanism, randomly samples batches from historical interaction data for gradient descent update, and the initial learning rate is 0.0001, which is attenuated by 10% every 100,000 steps.

[0055] The action execution module converts the action vector into specific radio frequency control signals, and sends parameter adjustment instructions to the physical device control layer through a remote procedure call interface. The instruction content includes a target device identifier, a parameter type, a target value and an effective timestamp. The action execution module starts a heartbeat packet mechanism after sending the instruction, sends a state query request to the target device every 10 milliseconds, and if no confirmation response is received for three consecutive times, it is determined that the instruction execution fails, triggering a local retry mechanism and reporting an abnormal event. The policy generation module updates the internal state according to the instruction execution result, records a positive reward if the instruction execution is successful, and records a negative reward and adjusts the subsequent policy generation weight if the execution fails.

[0056] S5, by establishing a cross-domain collaborative decision-making hub, the hub receives local decision proposals from all intelligent counter-agent, according to the preset command level structure and collaboration rule base, conflict resolution, resource allocation and global strategy optimization, finally generate unified combat instruction sequence and synchronous distribution to each execution unit, ensure that the multi-agent system in complex counter-environment maintain tactical consistency and strategic synergy.

[0057] In the S5 step, the cross-domain collaborative decision-making hub deploys a global strategy coordinator at the top layer of the system architecture, which includes a conflict detection module, a resource arbitration module and an instruction synthesis module. The conflict detection module performs spatiotemporal overlap analysis and target consistency verification based on the decision proposals submitted by each agent. The spatiotemporal overlap analysis calculates the intersection of the spectrum resources, spatial regions and time windows involved in each proposal, and if there is a resource conflict, it is marked as a conflict proposal. The target consistency verification evaluates the contribution of each proposal to the global target according to the preset tactical target priority table, and if the proposal target deviates from the global target by more than a threshold, it is marked as an invalid proposal.

[0058] The resource arbitration module allocates spectrum, power and spatial direction resources to the agents according to a preset priority rule and a resource occupation state table. The priority rule includes three dimensions of task urgency, agent command level and historical contribution, and each dimension is assigned a different weight coefficient. The resource occupation state table records the allocation state and release time of all resources in real time. The resource arbitration module uses a greedy algorithm to allocate the optimal available resource to each conflicting proposal, ensuring maximum resource utilization. For proposals that cannot be allocated resources, the resource arbitration module generates a resource request queue, and reallocates resources in priority order after the resources are released.

[0059] The instruction synthesis module encapsulates the arbitration results into standardized instruction messages. The message format includes instruction type, target agent identifier, parameter list, effective timestamp and check code. The instruction synthesis module distributes the instruction messages to each execution unit through a reliable multicast protocol. The multicast protocol uses an acknowledgement retransmission mechanism to ensure that all target units correctly receive the instructions. The execution unit returns an acknowledgement message after receiving the instruction, and the instruction synthesis module updates the global state after receiving all the acknowledgement messages. If an acknowledgement message is not received within a timeout period, the retransmission mechanism is triggered. The cross-domain collaborative decision-making hub executes a decision-making cycle every second to ensure that the system response delay is less than one second.

[0060] S6, by constructing a cognitive domain behavior modeling engine, the engine collects the voice instructions, operation logs, decision delay and situation annotation data of the participating commanders during the training process, uses natural language processing and behavior sequence mining algorithms to construct the commander's cognitive load model, decision preference model and tactical style portrait, and inputs the modeling results as feedback signals to the strategy generation module of the intelligent counter-agent, so that it can dynamically adjust the counter-strength and strategy complexity, and realize adaptive matching of training difficulty.

[0061] In the S6 step, the cognitive domain behavior modeling engine is deployed on the training monitoring server. The engine includes a speech recognition subsystem, an operation log analysis subsystem and a decision trajectory mining subsystem. The speech recognition subsystem uses an end-to-end neural network model to convert the commander's voice instructions into structured semantic labels. The neural network model is based on a convolutional recurrent architecture, with speech spectrum features as input and probability distribution of predefined instruction set as output. The speech recognition subsystem performs acoustic model adaptation during recognition, dynamically adjusting model parameters based on the commander's voice features to ensure an accuracy rate of no less than 95%.

[0062] The operation log analysis subsystem extracts key operation events and response delays based on a time series alignment algorithm. The operation log includes interface click events, parameter adjustment events, and instruction sending events, each log record containing event type, timestamp, and operation object. The time series alignment algorithm dynamically time warps the operation log and the battlefield situation change sequence, calculating operation response delays and operation effectiveness indicators. The decision trajectory mining subsystem uses a hidden Markov model to cluster the behavior patterns of commanders at different battlefield stages. The state space of the hidden Markov model corresponds to tactical decision types, and the observation sequence corresponds to the operation event sequence. The model parameters are trained using maximum likelihood estimation.

[0063] The cognitive domain behavior modeling engine outputs a cognitive load index and a tactical tendency vector. The cognitive load index is calculated based on speech pause frequency, operation error rate, and decision delay, and the tactical tendency vector is generated based on decision trajectory clustering results. The modeling results are pushed in real time to the strategy generation module of the intelligent counteragent through a message queue. The strategy generation module dynamically adjusts the exploration noise amplitude based on the cognitive load index. The higher the cognitive load index, the smaller the exploration noise amplitude, reducing the complexity of the counterstrategy. The strategy generation module adjusts the output weights of the action value function based on the tactical tendency vector, strengthening the strategy branches that match the commander's tactical style and improving the training relevance.

[0064] S7, by deploying a real-time performance evaluation module, which, after each counter-round, quantifies the comprehensive performance of the participating parties and the counteragent based on a pre-set multi-dimensional evaluation index system, including signal suppression success rate, communication link maintenance time, spectrum resource utilization rate, tactical target achievement degree, and cognitive decision efficiency, and generates a visual evaluation report, providing data support for subsequent training plan optimization.

[0065] In step S7, the real-time performance evaluation module automatically triggers the evaluation process after each counter-round. The module includes an index calculation submodule, a weight configuration submodule, and a report generation submodule. The index calculation submodule calculates the pre-set twenty-eight quantitative indicators based on the original sensor data and system logs, including signal suppression success rate, communication link maintenance time, spectrum resource utilization rate, tactical target achievement degree, and cognitive decision efficiency. The signal suppression success rate is calculated as the proportion of time when the interference signal power exceeds the target signal power threshold. The communication link maintenance time is calculated as the length of the continuous time period when the signal-to-noise ratio is above the demodulation threshold. The spectrum resource utilization rate is calculated as the proportion of the allocated spectrum bandwidth to the total available bandwidth. The tactical target achievement degree is calculated based on a pre-set target completion degree scoring table. The cognitive decision efficiency is calculated based on decision delay and operation accuracy.

[0066] The weight configuration submodule supports the instructor to dynamically adjust the weight coefficients of each index according to the training target, and the weight coefficients are stored in the configuration file, supporting multi-dimensional configuration according to the training subject, the level of the trainees and the tactical stage. The report generation submodule outputs the evaluation results to the visualization terminal in the form of radar chart, heat map and trend curve. The radar chart shows the relative scores of each index, the heat map shows the spatial distribution of spectrum resource utilization, and the trend curve shows the change process of key indicators over time. The evaluation report includes comprehensive score, single index score, improvement suggestion and historical comparison data, and supports export to standard format file for subsequent analysis.

[0067] S8, by constructing a historical confrontation case library and a strategy evolution engine, the engine continuously records the signal interaction sequence, decision path and environment parameter in each training confrontation, uses a graph neural network to structure the historical cases, extracts high-frequency successful strategy patterns and typical failure causes, and iteratively optimizes the strategy space of the intelligent confrontation agent through a genetic algorithm, so that the system has the ability of continuous self-evolution, avoiding the pattern fixation of training content.

[0068] In the S8 step, the historical confrontation case library persistently stores the complete data records of each training confrontation, including the original radio frequency waveform, the spectrum feature vector, the agent decision sequence, the environment parameter snapshot and the evaluation result summary. The data storage format adopts a columnar database structure, supporting multi-dimensional retrieval and correlation analysis according to time range, confrontation type, participating unit and tactical target. Each record contains a unique identifier, a timestamp, a scene identifier, a trainee identifier, a confrontation agent identifier, an original data block and a metadata index. The database adopts a distributed architecture, supporting horizontal expansion to cope with data growth, with a data write delay of less than fifty milliseconds and a query response time of less than two hundred milliseconds.

[0069] The strategy evolution engine extracts strategy patterns from successful confrontation samples in the historical case library regularly, and the engine includes a graph structure construction module, a strategy pattern mining module and a genetic optimization module. The graph structure construction module models each confrontation process as a state transition graph, with nodes representing battlefield situation snapshots and edges representing agent decision actions, and node attributes including situation feature vectors and edge attributes including action parameters and reward values. The strategy pattern mining module uses a graph convolution network to identify high-frequency subgraph structures, which includes three layers of graph convolution layers, each with an output dimension of one hundred and twenty-eight and a rectified linear unit activation function, using adjacency matrix normalization and residual connection structure. The graph convolution network outputs node embedding vectors, which are identified as high-frequency strategy pattern subgraphs through clustering algorithms.

[0070] The genetic optimization module encodes the identified policy patterns into chromosome individuals, with chromosome encoding containing policy network weights, exploration parameters, and resource allocation rules. The genetic optimization module generates a new generation of policy population through crossover, mutation, and fitness selection operations. The crossover operation adopts a uniform crossover strategy, the mutation operation adopts Gaussian noise disturbance, and the fitness function is calculated based on the average reward value of the strategy in the historical case. The new generation of policy population is injected into the intelligent adversarial agent cluster through the strategy injection interface, replacing low fitness individuals, and completing the iterative optimization of the policy space. The strategy evolution engine executes an optimization cycle every twenty-four hours to ensure continuous evolution of the system strategy.

[0071] The system includes a distributed radio frequency signal transceiving array, a programmable attenuator array, a multipath reflection simulator, a wideband digital receiver cluster, an intelligent adversarial agent cluster, a cross-domain collaborative decision hub, a cognitive domain behavior modeling engine, a real-time performance evaluation module, a historical adversarial case library, and a strategy evolution engine.

[0072] The distributed radio frequency signal transceiving array is arranged with multiple independently controllable radio frequency transmitting and receiving nodes in a preset topological structure in three-dimensional space. Each node includes a radio frequency front-end module, a baseband processing module, and a time-space synchronization module. The radio frequency front-end module supports continuous tuning within a frequency band of 300 MHz to 6 GHz, with an output power dynamic range of -30 to +30 dBm, a frequency tuning step precision of 10 kHz, and a power regulation resolution of 0.1 dBm. The baseband processing module supports generating twelve standard modulation patterns, each supporting independent configuration of symbol rate, roll-off factor, and constellation mapping rules. The time-space synchronization module achieves nanosecond-level time alignment of the entire system through Beidou timing signals and fiber time distribution network, with a time synchronization accuracy better than 10 ns.

[0073] The programmable attenuator array is deployed with radio frequency devices with dynamic attenuation characteristics at key nodes of the simulated signal propagation path. Each attenuator unit supports program-controlled adjustment within a range of 0 to 120 dBm with a step of 0.5 dBm. The attenuator unit has a built-in digital control interface that supports remote command reception and state feedback, with a command response time of less than 5 ms and an attenuation accuracy better than 0.2 dBm. The attenuator unit adopts temperature compensation design to ensure stable attenuation characteristics in an environmental temperature range of -20°C to +70°C.

[0074] The multi-path reflection simulator deploys multiple movable reflecting surface units within a specified spatial region, each reflecting surface unit composed of a metal mesh structure and a servo drive mechanism. The servo drive mechanism adjusts the azimuth and elevation angles of the reflecting surface in real time according to the reflection path calculated by the ray tracing algorithm, with an angle adjustment range covering zero to three hundred and sixty degrees in azimuth and negative ninety to ninety degrees in elevation, and an angle adjustment accuracy better than one degree. The reflecting surface unit supports dynamic configuration of the reflection coefficient and polarization characteristics, with a reflection coefficient adjustment range of zero point one to zero point nine and polarization characteristics supporting linear polarization, circular polarization, and elliptical polarization mode switching.

[0075] The wideband digital receiver cluster is deployed at key listening locations in the simulated battlefield, with each receiver containing a twelve-bit analog-to-digital converter with a sampling rate of five gigahertz, a digital down-conversion channel supporting a real-time bandwidth of two hundred megahertz, and a fast Fourier transform-based spectrum analysis unit. The analog-to-digital converter has a spurious-free dynamic range of better than seventy dBm, the digital down-conversion channel supports arbitrary center frequency configuration, and the spectrum analysis unit performs an eight thousand one hundred and ninety-two point fast Fourier transform with a spectral resolution of better than zero point six one hertz. The receiver output data is uploaded in real time through a gigabit Ethernet interface with a data transmission delay of less than ten milliseconds.

[0076] The intelligent countermeasure agent cluster deploys multiple software entities with autonomous decision-making capabilities in the virtual command hierarchy, with each agent entity running in an independent computing container. The state-aware module subscribes to battlefield situation updates published by the central data fusion platform through a message queue, which uses a publish-subscribe mode and supports processing one hundred thousand messages per second. The strategy generation module uses the deep deterministic policy gradient algorithm, with a neural network inference delay of less than five milliseconds. The action execution module sends parameter adjustment instructions to the physical device control layer through a remote procedure call interface, with an instruction transmission delay of less than one millisecond, and a heartbeat packet mechanism ensures real-time monitoring of instruction execution status.

[0077] The cross-domain collaborative decision-making hub deploys a global strategy coordinator at the top of the system architecture, with a conflict detection module performing spatio-temporal overlap analysis and target consistency verification with an analysis delay of less than fifty milliseconds. The resource arbitration module allocates resources based on preset priority rules and resource occupation state tables, with a resource allocation algorithm execution time of less than one hundred milliseconds. The instruction synthesis module encapsulates the arbitration results into standardized instruction packets and distributes them to each execution unit through a reliable multicast protocol, with a multicast protocol ensuring an instruction delivery rate of no less than ninety-nine point nine nine percent.

[0078] The cognitive domain behavior modeling engine is deployed on a training monitoring server, the speech recognition subsystem uses an end-to-end neural network model, the speech recognition accuracy is not less than 95%, and the recognition delay is less than 200 milliseconds. The operation log analysis subsystem extracts key operation events according to a time series alignment algorithm, and the alignment accuracy is better than 10 milliseconds. The decision trajectory mining subsystem uses a hidden Markov model to cluster behavior patterns, the model training convergence time is less than 10 minutes, and the clustering accuracy is not less than 90%.

[0079] The real-time performance evaluation module includes an index calculation submodule, a weight configuration submodule, and a report generation submodule. The index calculation submodule calculates twenty-eight quantitative indicators based on raw sensor data and system logs, with a calculation delay of less than one second. The weight configuration submodule supports dynamic adjustment of the weight coefficients of each indicator, and the weight update takes effect in less than 50 milliseconds. The report generation submodule outputs the evaluation results in the form of radar charts, heat maps, and trend curves, with a graph rendering delay of less than 200 milliseconds.

[0080] The historical confrontation case library stores the complete data records of each training confrontation, and the data storage format uses a columnar database structure, supporting writing 10,000 records per second, with a query response time of less than 200 milliseconds. The data storage capacity supports expansion to petabyte level, with a data compression ratio of not less than 5:1, ensuring that the storage cost is controllable.

[0081] The strategy evolution engine includes a graph structure construction module, a strategy pattern mining module, and a genetic optimization module. The graph structure construction module models the confrontation process as a state transition graph, with a modeling delay of less than 10 seconds. The strategy pattern mining module uses a graph convolution network to identify high-frequency subgraph structures, with a network inference delay of less than 5 seconds. The genetic optimization module generates a new generation of strategy population through crossover, mutation, and fitness selection operations, with a single optimization period execution time of less than one hour, and a strategy injection interface ensures seamless replacement of old strategies with new strategies.

[0082] The above method and system realize deep coupling of the physical domain, electromagnetic domain, and cognitive domain, construct a high-fidelity, adaptive, and evolvable simulated signal combat training environment, and effectively improve the confrontation ability and command decision-making level of trainees in complex electromagnetic environments.

Claims

1. A training method for analog signal operations, characterized by, Comprise: By deploying a distributed radio frequency signal transceiver array in the physical space, a three-dimensional electromagnetic field simulation environment covering a specified geographic area is constructed; By embedding a programmable attenuator array and a multipath reflection simulator in the electromagnetic propagation path, the path loss, multipath time delay spread, Doppler shift and polarization rotation effects are calculated and applied in real time according to the pre-set terrain data and atmospheric refraction model in the three-dimensional electromagnetic field simulation environment; By deploying a wideband digital receiver cluster at the receiving end, each receiver in the wideband digital receiver cluster is integrated with a high-speed analog-to-digital converter, a digital down-conversion module and a real-time spectrum analysis engine; By constructing an intelligent adversarial agent cluster based on a deep reinforcement learning framework, each agent in the intelligent adversarial agent cluster is configured with an independent state perception module, a policy generation module and an action execution module; By establishing a cross-domain collaborative decision-making hub, the cross-domain collaborative decision-making hub receives local decision proposals from all intelligent adversarial agents; By constructing a cognitive domain behavior modeling engine, the cognitive domain behavior modeling engine collects voice instructions, operation logs, decision delays and situation annotation data of participating commanders during the training process; By deploying a real-time performance evaluation module, the real-time performance evaluation module evaluates the performance of each agent according to a pre-set multi-dimensional evaluation index system after each adversarial round; By constructing a historical adversarial case library and a strategy evolution engine, the strategy evolution engine continuously records the signal interaction sequence, decision path and environment parameters in each training adversarial.

2. The training method of analog signal campaign according to claim 1, characterized in that, By deploying a distributed radio frequency signal transceiver array in the physical space, a three-dimensional electromagnetic field simulation environment covering a specified geographic area is constructed, comprising: Multiple radio frequency transmitting and receiving nodes are arranged in a pre-set topological structure in a three-dimensional space, each node including a radio frequency front-end module, a baseband processing module and a time-space synchronization module; The radio frequency front-end module supports continuous tuning within a frequency range of 300 MHz to 6 GHz, with an output power dynamic range of -30 to +30 dBm, a frequency tuning step precision of 10 kHz and a power adjustment resolution of 0.1 dBm; The baseband processing module supports generating twelve standard modulation patterns including frequency modulation, phase modulation and quadrature amplitude modulation, each modulation pattern supporting independent configuration of symbol rate, roll-off coefficient and constellation mapping rules; The reconfigurable antenna structure is composed of an eight-element phased array, supporting dynamic adjustment of beam pointing within a range of 0 to 360 degrees in azimuth and -90 to +90 degrees in elevation, with a beam main lobe width continuously adjustable between 5 and 60 degrees; The time-space synchronization module achieves nanosecond-level time alignment of the entire system through Beidou timing signals and fiber time distribution network, with a time synchronization accuracy better than 10 ns.

3. The training method of analog signal campaign according to claim 2, characterized in that, By embedding a programmable attenuator array and a multipath reflection simulator in the electromagnetic propagation path, the path loss, multipath time delay spread, Doppler shift and polarization rotation effects are calculated and applied in real time according to the pre-set terrain data and atmospheric refraction model, comprising: Load the digital elevation model and land cover type map corresponding to the current training scene, combine the real-time collected atmospheric temperature and humidity data to construct the three-dimensional space propagation environment parameter field; The propagation model calculation module performs geometric modeling on all possible propagation paths of each radio frequency signal from the transmitting point to the receiving point according to the ray tracing algorithm, and calculates the free space loss, diffraction loss, reflection loss and atmospheric absorption loss of each path; The programmable attenuator array is deployed at the key nodes of the propagation path, and each attenuator unit supports programmed adjustment in the range of zero to one hundred and twenty dBm with a step of zero point five dBm, the instruction response time is less than five milliseconds, and the attenuation accuracy is better than zero point two dBm; The multipath reflection simulator deploys multiple movable reflecting surface units, each reflecting surface unit is composed of a metal mesh structure and a servo drive mechanism, and the servo drive mechanism adjusts the spatial azimuth and elevation angle of the reflecting surface in real time according to the reflection path calculated by the ray tracing algorithm, and the angle adjustment accuracy is better than zero point one degree, and the angle adjustment response time is less than fifty milliseconds.

4. The training method of analog signal campaign according to claim 3, characterized in that, By deploying a wideband digital receiver cluster at the receiving end, each receiver integrates a high-speed analog-to-digital converter, a digital down-conversion module, and a real-time spectrum analysis engine, including: The wideband digital receiver is deployed at the key listening position of the simulated battlefield, and each receiver includes a twelve-bit analog-to-digital converter with a sampling rate of five gigahertz, a digital down-conversion channel supporting a real-time bandwidth of two hundred megahertz, and a spectrum analysis unit based on fast Fourier transform; The receiver antenna receives the superimposed multi-source radio frequency signal in space, which is sent to the analog-to-digital converter for digital sampling after low-noise amplification and anti-aliasing filtering; The digital down-conversion module performs digital mixing and decimation filtering according to the preset center frequency and bandwidth parameters, and outputs baseband I / Q data stream; The real-time spectrum analysis engine processes the baseband I / Q data stream in segments, each segment has a length of eighty-one thousand nine hundred and ninety-two points, and performs fast Fourier transform after windowing with a Hanning window function to obtain the frequency domain amplitude spectrum and phase spectrum, and extracts the spectral centroid, spectral entropy, peak frequency, bandwidth occupancy rate and modulation recognition features; The original time domain waveform data and frequency domain feature vector are uploaded in real time to the central data fusion platform through a gigabit Ethernet interface, and the data format follows a unified timestamp marking and channel identification protocol.

5. The training method of analog signal campaign according to claim 4, characterized in that, By constructing an intelligent adversarial agent cluster based on a deep reinforcement learning framework, each agent is configured with an independent state perception module, a policy generation module, and an action execution module, including: The intelligent adversarial agent deploys multiple software entities with autonomous decision-making capabilities in the virtual command hierarchy, and each agent entity runs in an independent computing container; The state perception module subscribes to the battlefield situation updates published by the central data fusion platform through a message queue, and performs coordinate system conversion, data normalization, and feature splicing on the original situation data to generate a state vector of uniform dimension; The policy generation module uses the deep deterministic policy gradient algorithm to construct an action value function, and its neural network structure includes three fully connected layers with five hundred and twelve, two hundred and fifty-six, and one hundred and twenty-eight neurons respectively, and the activation function is a rectified linear unit. The output is a continuous action vector, including carrier frequency adjustment, transmission power adjustment, beam pointing angle adjustment, modulation pattern index and interference strategy identifier; The action execution module converts the action vector into specific radio frequency control signals, sends parameter adjustment instructions to the physical device control layer through the remote procedure call interface, and monitors the instruction execution status through the heartbeat packet mechanism. If no confirmation response is received for three consecutive times, the local retry mechanism is triggered and an abnormal event is reported.

6. The training method of analog signal campaign according to claim 5, characterized in that, By establishing a cross-domain collaborative decision-making hub, the hub receives local decision proposals from all intelligent adversarial agents, resolves conflicts, allocates resources, and optimizes global strategies based on pre-set command hierarchy and collaboration rule library, including: The conflict detection module analyzes the spatial and temporal overlap and target consistency of the decision proposals submitted by each agent. If there is a resource conflict or the target deviation exceeds the threshold, it is marked as a conflict or invalid proposal. The resource arbitration module allocates frequency, power, and spatial direction resources based on pre-set priority rules and resource occupation status table. The priority rules include task urgency, agent command level, and historical contribution. The instruction synthesis module encapsulates the arbitration results into standardized instruction packets and distributes them to each execution unit through a reliable multicast protocol. The multicast protocol uses an acknowledgement and retransmission mechanism to ensure that all target units correctly receive the instructions. The cross-domain collaborative decision-making hub executes a decision cycle every second to ensure that the system response delay is less than one second.

7. The training method of analog signal campaign according to claim 6, characterized in that, By building a cognitive domain behavior modeling engine, the engine collects voice instructions, operation logs, decision delays, and situation annotation data of participating commanders during training, including: The voice recognition subsystem uses an end-to-end neural network model to convert commander voice instructions into structured semantic tags. The neural network model is based on a convolutional recurrent architecture, with an accuracy of not less than 95% and a recognition delay of less than 200 milliseconds. The operation log analysis subsystem extracts key operation events and response delays based on time series alignment algorithms with an alignment accuracy of better than 10 milliseconds. The decision trajectory mining subsystem uses a hidden Markov model to cluster and analyze the behavior patterns of commanders at different battlefield stages. The model training convergence time is less than 10 minutes, and the clustering accuracy is not less than 90%. The cognitive domain behavior modeling engine outputs a cognitive load index and a tactical tendency vector as feedback signals to the strategy generation module of the intelligent adversarial agent, dynamically adjusting the exploration noise amplitude and the action value function output weight.

8. The training method of analog signal campaign according to claim 7, characterized in that, By deploying a real-time performance evaluation module, the module quantitatively scores based on a pre-set multi-dimensional evaluation index system after each adversarial round, including: The index calculation submodule calculates twenty-eight quantitative indicators based on raw sensor data and system logs, including signal suppression success rate, communication link maintenance time, spectrum resource utilization rate, tactical target achievement degree, and cognitive decision efficiency, with a calculation delay of less than one second. The weight configuration submodule allows instructors to dynamically adjust the weight coefficients of each indicator based on training objectives, with a weight update effective time of less than 50 milliseconds. The report generation submodule outputs the evaluation results in the form of radar charts, heat maps, and trend curves to the visualization terminal. The graph rendering delay is less than 200 milliseconds, and the evaluation report includes comprehensive scores, individual indicator scores, improvement suggestions, and historical comparison data.

9. A training system for training a method of combat using the analog signal according to any one of claims 1 to 8, characterized in that, including: A distributed RF transceiver array is used to deploy multiple independently controllable RF transmitting and receiving nodes in a preset topology in three-dimensional space, each node including an RF front-end module, a baseband processing module, and a space-time synchronization module. The RF front-end module supports continuous tuning within a frequency band of 300 MHz to 6 GHz, with an output power dynamic range of -30 to +30 dBm. The baseband processing module supports the generation of twelve standard modulation patterns, including frequency modulation, phase modulation, and quadrature amplitude modulation. The space-time synchronization module achieves nanosecond-level time alignment of the entire system through Beidou timing signals and a fiber-optic time distribution network. A programmable attenuator array is used to deploy RF devices with dynamic attenuation characteristics at key nodes in the analog signal propagation path. Each attenuator unit supports programmable adjustment within a range of 0 to 120 dBm with a step size of 0.5 dBm. The control instructions are generated in real time by a propagation model calculation module based on current analog terrain elevation data, vegetation coverage density, and atmospheric humidity parameters. A multipath reflection simulator is used to deploy multiple movable reflective surface units in a specified spatial region. Each reflective surface unit is composed of a metal mesh structure and a servo drive mechanism. The servo drive mechanism adjusts the spatial azimuth and elevation angles of the reflective surface in real time based on the reflection path calculated by the ray tracing algorithm, so that the incident signal produces mirror reflection and diffuse reflection components that conform to the principles of geometric optics. A wideband digital receiver cluster is used to deploy at key listening locations in the simulated battlefield. Each receiver includes a 12-bit analog-to-digital converter with a sampling rate of 5 GHz, a digital down-conversion channel with a real-time bandwidth of 200 MHz, and a frequency spectrum analysis unit based on the Fast Fourier Transform. The output data is uploaded in real time to the central data fusion platform through a Gigabit Ethernet interface. An intelligent adversarial agent cluster is used to deploy multiple software entities with autonomous decision-making capabilities in the virtual command hierarchy. Each agent entity runs in an independent computing container. The state perception module subscribes to battlefield situation updates published by the central data fusion platform through a message queue. The strategy generation module constructs an action value function using the Deep Deterministic Policy Gradient algorithm. The action execution module sends parameter adjustment instructions to the physical device control layer through a remote procedure call interface. A cross-domain collaborative decision-making hub is used to deploy a global strategy coordinator at the top of the system architecture. The coordinator includes a conflict detection module, a resource arbitration module, and an instruction synthesis module. The conflict detection module performs temporal overlap analysis and target consistency verification based on the decision proposals submitted by each agent. The resource arbitration module allocates frequency, power, and spatial direction resources based on preset priority rules and resource occupation status tables. The instruction synthesis module encapsulates the arbitration results into standardized instruction packets and distributes them to each execution unit through a reliable multicast protocol. The cognitive domain behavior modeling engine is deployed on the training monitoring server, and the engine includes a speech recognition subsystem, an operation log analysis subsystem, and a decision trajectory mining subsystem. The speech recognition subsystem converts the commander's voice command into a structured semantic label using an end-to-end neural network model. The operation log analysis subsystem extracts key operation events and response delays based on a time sequence alignment algorithm. The decision trajectory mining subsystem uses a hidden Markov model to cluster and analyze the behavior patterns of commanders at different battlefield stages. The real-time performance evaluation module automatically triggers the evaluation process after each confrontation round ends. The module includes an index calculation submodule, a weight configuration submodule, and a report generation submodule. The index calculation submodule calculates twenty-eight preset quantitative indicators based on raw sensor data and system logs. The weight configuration submodule allows instructors to dynamically adjust the weight coefficients of each indicator based on training objectives. The report generation submodule outputs the evaluation results in the form of radar charts, heat maps, and trend curves to a visualization terminal. The historical confrontation case library is used to persistently store the complete data records of each training confrontation, including original radio frequency waveforms, spectral feature vectors, agent decision sequences, environment parameter snapshots, and evaluation result summaries. The data storage format uses a columnar database structure, supporting multi-dimensional retrieval and correlation analysis based on time range, confrontation type, participating unit, and tactical target. The strategy evolution engine extracts strategy patterns from successful confrontation samples in the historical case library on a regular basis. The engine includes a graph structure construction module, a strategy pattern mining module, and a genetic optimization module. The graph structure construction module models each confrontation process as a state transition graph. The strategy pattern mining module uses a graph convolution network to identify high-frequency subgraph structures. The genetic optimization module encodes the identified strategy patterns into chromosome individuals, generates a new generation of strategy population through crossover, mutation, and fitness selection operations, and injects the new strategy population into the intelligent confrontation agent cluster.

10. The analog signal-of-the-operation training system of claim 9, wherein, The distributed radio frequency signal transceiver array is used for: Laying out multiple radio frequency transmitting and receiving nodes in a preset topological structure in three-dimensional space, each node including a radio frequency front-end module, a baseband processing module, and a space-time synchronization module. The radio frequency front-end module supports continuous tuning in the three hundred megahertz to six gigahertz frequency band, with an output power dynamic range of negative thirty to positive thirty dBm, a frequency tuning step precision of ten kilohertz, and a power regulation resolution of zero point one dBm. The baseband processing module supports generating twelve standard modulation patterns, each supporting independent configuration of symbol rate, roll-off coefficient, and constellation mapping rules. The directional pattern reconfigurable antenna structure is composed of eight-element phased array, supporting dynamic adjustment of beam pointing in the range of zero to three hundred sixty degrees in azimuth angle and negative ninety to ninety degrees in elevation angle, with a beam main lobe width continuously adjustable between five and sixty degrees. The space-time synchronization module achieves nanosecond-level time alignment of the entire system through Beidou timing signals and fiber-optic time distribution networks, with a time synchronization accuracy better than ten nanoseconds.

Images