A pre-warning method, system, product and device

Abstract

The application discloses a kind of early warning method, system, product and equipment, by acquiring the key situation characteristics of current battlefield;The key situation characteristics are input to large language model, generate multiple hypotheses about future enemy and me contact point;For each hypothesis, a lightweight combat simulation model is used to carry out multi-branch parallel deduction, each deduction branch is endowed with different random disturbance, and each branch outputs the deduction result representing enemy and me contact situation;Based on the deduction results of all deduction branches under the same hypothesis, the corresponding contact occurrence probability is determined;Based on the contact occurrence probability corresponding to each hypothesis, generate early warning information.The application quickly generates hypotheses by large language model and verifies by combining lightweight multi-branch parallel deduction, realizes active prediction and high-confidence early warning of future enemy and me contact point, with the advantages of fast response speed, high accuracy and strong interpretability.

Description

Technical Field

[0001] This application relates to the field of computer technology, and in particular to an early warning method, system, product, and device. Background Technology

[0002] Currently, situational awareness in the field of combat command mainly employs passive rule-triggered approaches, single-simulation simulations, or large-scale model inference. Passive rule-triggered approaches rely on manually preset rules, triggering alarms when real-time situational data meets the rule conditions; this is the most widely used approach. Single-simulation simulations use complex combat simulation engines to predict conflict points by simulating future situational evolution; they offer high reliability but are computationally complex. Large-scale model reasoning leverages the reasoning capabilities of large-scale language models for rapid situational analysis; they are fast but prone to illusions and unreasonable inferences.

[0003] However, the existing technical solutions mentioned above have the following technical problems: the passive rule-triggered scheme lags behind the situational changes by 5-15 minutes in terms of warning time, and the rules need to be manually defined and maintained, resulting in poor adaptability; the large model inference scheme, although fast, lacks reliability, with approximately 15%-25% of the inference results containing logical contradictions; the traditional simulation inference scheme has high reliability, but a single inference takes 5-30 minutes, which cannot support online real-time warnings. In summary, the existing technologies cannot simultaneously satisfy the proactive predictability, rapid response, and reliability of the warnings, resulting in delayed warnings, low accuracy, and poor adaptability.

[0004] Therefore, the inability of existing technologies to simultaneously address the timeliness, accuracy, and reliability of early warnings is a technical problem that urgently needs to be solved by those in the field. Summary of the Invention

[0005] In view of the above problems, this application provides an early warning method, system, product and equipment.

[0006] The embodiments of this application disclose the following technical solutions: The first aspect of this application provides an early warning method, including: Obtain key situational characteristics of the current battlefield; The key situational features are input into a large language model, which then performs reasoning based on a preset thinking framework to generate multiple hypotheses about the point of contact between the enemy and ourselves. For each of the aforementioned assumptions, a lightweight combat simulation model is used to perform multi-branch parallel simulations, where each simulation branch is given different random perturbations, and each simulation branch outputs simulation results representing the contact situation between the enemy and our side. Based on the deduction results of all deduction branches under the same assumption, the corresponding contact probability is determined; Early warning information is generated based on the probability of contact corresponding to each hypothesis.

[0007] A second aspect of this application provides an early warning system, including: The acquisition unit is used to acquire key situational characteristics of the current battlefield. The reasoning unit is used to input the key situational features into the large language model, so that the large language model can reason based on a preset thinking framework and generate multiple hypotheses about the enemy-friend contact point; The simulation unit is used to perform multi-branch parallel simulations for each of the above assumptions using a lightweight combat simulation model, wherein each simulation branch is given different random perturbations, and each simulation branch outputs simulation results characterizing the contact situation between the enemy and our side. The determination unit is used to determine the corresponding contact probability based on the deduction results of all deduction branches under the same assumption; The generation unit is used to generate early warning information based on the contact probability corresponding to each hypothesis.

[0008] A third aspect of this application provides an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the early warning method as described in the first aspect above.

[0009] A fourth aspect of this application provides a computer program product that, when run on a computer, executes the early warning method as described in the first aspect above.

[0010] A fifth aspect of this application provides a computer-readable storage medium storing instructions that, when executed on a terminal device, cause the terminal device to perform the early warning method as described in the first aspect above.

[0011] Compared with the prior art, this application has the following beneficial effects: By acquiring key situational characteristics of the current battlefield and inputting them into a large language model, the model's reasoning capabilities are leveraged to proactively generate multiple hypotheses about future enemy-friendly contact points. This shifts the approach from passively waiting for rule triggers to proactively predicting scenarios, resolving the issue of delayed early warning. Secondly, a lightweight combat simulation model is used for multi-branch parallel inference for each hypothesis. Each branch is assigned different random perturbations to simulate battlefield uncertainties. The lightweight design significantly reduces computational complexity, and the parallel architecture reduces inference time to milliseconds, meeting the requirements for real-time rapid response. Finally, the probability of contact is statistically calculated based on the inference results of all branches under the same hypothesis. Multi-branch statistical averaging effectively suppresses the illusions that may arise from single inferences of the large model and the random errors of single simulations. The output probability results have statistical reliability, and early warning information is generated accordingly, thus providing reliable early warning information while ensuring speed. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 A flowchart illustrating an early warning method provided in this application embodiment; Figure 2 This is a schematic diagram illustrating the process by which the large language model in this application reasones and generates hypotheses according to a three-layer thinking framework. Figure 3 This is a flowchart illustrating the structure and operation of the lightweight combat simulation module in this application. Figure 4 An overall architecture diagram provided for embodiments of this application; Figure 5 This is a structural diagram of an early warning system provided in an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0015] To facilitate understanding of the technical solutions provided in the embodiments of this application, the technical terms involved in the embodiments of this application will be explained below.

[0016] Large language models refer to deep learning models pre-trained on massive amounts of text data. They possess natural language understanding and generation capabilities, and can infer and output prediction results based on the input situational features.

[0017] Lightweight combat simulation models refer to simplified simulation models that retain only key combat factors such as troop movement, fire coverage, casualty calculation, logistical constraints, and communication delays, which significantly reduces computational complexity compared to traditional complex simulation models.

[0018] Multi-branch parallel simulation refers to running multiple independent simulation branches simultaneously for the same hypothesis, with each branch using the same lightweight model but given different random perturbations.

[0019] To facilitate understanding of the technical solutions provided in the embodiments of this application, the background technology involved in the embodiments of this application will be described below.

[0020] As mentioned earlier, existing situational awareness technologies in the field of combat command suffer from the following main problems: First, passive rule-triggered schemes require pre-defined conditions to be met before issuing an alert, with the alert time typically lagging behind situational changes by 5 to 15 minutes, making proactive prediction impossible. Second, all alert rules require manual definition and maintenance, resulting in rule update cycles that can last for days to weeks when enemy tactics or the battlefield environment change, leading to poor adaptability. Third, independent judgment schemes cannot uncover the implicit relationships between situational elements, resulting in an alert accuracy rate of only 45% to 60%, with a false alarm rate as high as 25% to 40%. Fourth, traditional simulation and deduction schemes require the establishment of detailed combat models, with each simulation taking 5 to 30 minutes, making it impossible to support online real-time alerts. Fifth, using large models for situational reasoning can easily lead to illusions, with approximately 15% to 25% of the reasoning results containing logical contradictions or violating basic operational principles. Sixth, existing alert systems typically only output the alert level without providing the reasoning process and basis, making it difficult for commanders to understand the reasons for the alert and resulting in low trust levels. In summary, existing technologies cannot simultaneously satisfy the requirements of proactive prediction, rapid response, reliable results, and interpretability in alerts.

[0021] To address the aforementioned issues, this application's embodiments firstly address the problem of delayed early warning. It inputs key battlefield situational characteristics into a large-scale language model, leveraging the model's reasoning capabilities to proactively generate multiple hypotheses about future enemy-friendly contact points. This shifts from passively waiting for rule triggers to proactively predicting, advancing the early warning time by 2 to 5 minutes. Secondly, addressing the issues of poor rule adaptability and insufficient correlation mining, the large-scale model automatically analyzes enemy intentions, infers contact points, and assesses risks through a pre-defined thinking framework. This eliminates the need for manually defined rules and uncovers implicit correlations between multiple elements, improving early warning accuracy to 82% to 88%. Thirdly, addressing the problem of excessively long simulation time, this application employs a lightweight combat simulation model, retaining only key factors such as troop movement, fire coverage, and casualty calculations. It performs multi-branch parallel simulations for each hypothesis, assigning different random perturbations to each branch to simulate uncertainty. Execution is performed in parallel on a graphics processor, reducing the time for a single simulation from several minutes to 500 to 1000 milliseconds, meeting real-time early warning requirements. Then, addressing the issue of insufficient credibility in large-scale models, this application calculates the probability of contact by statistically analyzing the proportion of branches where contact occurs across all simulation branches under the same assumption. Utilizing multi-branch statistical averaging effectively suppresses the illusions and random errors of single-inference reasoning, outputting a probability value with statistical credibility. Finally, addressing the lack of interpretability in the results, the structured early warning information generated by this application not only includes the early warning level, contact point, time window, scale, expected casualties, and response recommendations, but also displays the thought process of the large-scale model, key simulation parameters, multi-branch result distribution, and credibility score, enabling commanders to understand the basis of the early warning and increasing credibility to 75% to 85%. Thus, this application comprehensively solves several problems existing in current technologies, achieving proactive, rapid, reliable, and interpretable situational early warning.

[0022] It should be noted that the early warning methods, systems, products, devices, and media provided in this application can be applied to the field of computer technology. The above are merely examples and do not limit the application areas of the early warning methods, systems, products, devices, and media provided in this application. Furthermore, the embodiments of this application may not limit the executing entity of the early warning; for example, the early warning method of the embodiments of this application can be applied to data processing devices such as terminal devices or servers. The terminal device can be an electronic device such as a computer or a personal digital assistant (PDA). The server can be a standalone server, a cloud server, or a cluster server composed of multiple servers.

[0023] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0024] The following embodiment illustrates an early warning method provided in this application. See also... Figure 1 ,Should Figure 1 A flowchart of an early warning method provided in this application embodiment, the method including: S101. Obtain key situational characteristics of the current battlefield.

[0025] At the outset of the proactive situational awareness (AIA) process, the system first needs to comprehensively, accurately, and in real-time acquire the key situational characteristics of the current battlefield. This step forms the data foundation for the entire AIA method and directly determines the reliability of subsequent large-scale model inference and simulation. Specifically, this step is achieved through a real-time situational awareness acquisition and fusion module. This module is responsible for accessing real-time situational data from various heterogeneous data sources and performing in-depth data fusion, cleaning, feature extraction, and historical data maintenance.

[0026] The acquisition of key situational characteristics of the current battlefield includes: accessing multi-source real-time situational data from sensor networks, intelligence systems, and command information systems, wherein the multi-source real-time situational data includes at least one of enemy and friendly troop positions, activity intensity, logistics supply lines, communication activities, and reconnaissance activities; performing time synchronization, coordinate alignment, and outlier detection on the multi-source real-time situational data to eliminate data redundancy and contradictions; extracting the key situational characteristics from the fused situational data, wherein the key situational characteristics include at least one of the following: enemy and friendly troop strength comparison, enemy activity intensity index, enemy logistics activity index, friendly defense situation index, terrain and meteorological factors; and maintaining a historical situational database with a sliding time window for trend analysis and anomaly detection.

[0027] First, in the multi-source data access phase, the system simultaneously connects to multiple data sources, including sensor networks, intelligence systems, and command information systems. Sensor networks include ground radar, airborne early warning aircraft, unmanned aerial vehicle (UAV) reconnaissance platforms, sonar detection systems, and ground vibration sensors. These sensors can capture raw data in real time, such as the enemy's troop location, movement speed, activity frequency, and firing points. Intelligence systems provide pre-analyzed intelligence products, such as the characteristics of enemy communication signals intercepted by electronic reconnaissance, enemy command and dispatch patterns analyzed by signals intelligence, enemy logistics movement obtained through human intelligence, and the scale of enemy troop concentrations determined by satellite imagery. Command information systems include our own troop deployments, operational plans, situation sharing maps, and summaries of enemy intelligence reported by higher authorities. These multi-source data specifically cover the following types: the coordinates of enemy and friendly forces, records of the intensity of enemy and friendly activities, dynamic changes in logistics supply lines, changes in the frequency and encryption level of communication activities, and the routes and intensity of reconnaissance activities. The system can simultaneously access structured data, such as troop statistics tables in a database, as well as semi-structured or unstructured data, such as intelligence report texts and reconnaissance image annotation information, ensuring the comprehensiveness of data sources.

[0028] After completing the access to multi-source data, the system immediately enters the data fusion and cleaning phase. Due to significant differences in time reference, coordinate system, data format, and accuracy between data from different sensors and intelligence sources, strict alignment and calibration are necessary. Specific operations include: time synchronization processing, where the system uses network time protocols or high-precision time synchronization equipment to unify the timestamps of all data sources to the same time baseline, with time synchronization accuracy controlled within 100 milliseconds, thereby ensuring the accuracy of subsequent time series analysis; coordinate alignment processing, where the system converts data from different coordinate systems, such as polar coordinates used by radar, geographic coordinates used by satellite imagery, and projected coordinates used by maps, to a common battlefield geographic coordinate system, such as the WGS84 coordinate system of latitude, longitude, and elevation, or military topographic map grid coordinates; outlier detection, where the system uses statistical methods or rule-based filters to automatically identify and remove abnormal data points that deviate significantly from the normal range due to sensor malfunctions, electromagnetic interference, data transmission errors, etc., such as unreasonable high speeds or position jumps at a certain moment; and redundancy elimination, where for duplicate information from multiple data sources reporting the same target or the same event, the system uses data association algorithms to fuse them, retaining the data with the highest confidence or the most recent timeliness, while removing duplicate records, thereby generating a unique and accurate situational fact.

[0029] Key situational features are extracted from the fused situational data, including: the ratio of friendly to enemy troop strength; the enemy activity intensity index, which is a weighted combination of reconnaissance frequency, maneuver speed, and fire activity frequency; the enemy logistics activity index, which is calculated based on the frequency of logistics supply line adjustments and the scale of material accumulation; the friendly defense situation index, which comprehensively assesses the density of defense lines, reserve positions, and firepower configuration; and terrain and meteorological factors, including terrain accessibility, visibility, and wind direction and speed.

[0030] In addition to extracting real-time situational features, the system also maintains a historical situational database with a sliding time window. This database typically stores all fused situational data and extracted feature sequences from the past 24 hours. The purpose of this historical database is to support trend analysis and anomaly detection. For example, by comparing the change curve of the enemy activity intensity index over the past 24 hours, the system can identify whether the activity intensity is gradually increasing or suddenly surging, thus determining whether the enemy is preparing for action; by analyzing the time series of the logistics activity index, it can discover whether the supply frequency has periodicity or abnormal peaks. Simultaneously, the historical database also provides contextual reference for subsequent large-scale model inference, helping the model understand the evolution of the situation. This database uses a first-in, first-out sliding window mechanism, with new data continuously written and older data exceeding 24 hours automatically archived or deleted, ensuring a controllable database size and sufficient data freshness. In summary, step S101, through four sub-steps—multi-source access, fusion cleaning, feature extraction, and historical maintenance—provides high-quality, timely, and multi-dimensional key situational feature inputs for the entire early warning method, laying a solid data foundation for subsequent large-scale model hypothesis generation and simulation.

[0031] S102. Input the key situational features into the large language model so that the large language model can reason based on a preset thinking framework and generate multiple hypotheses about the enemy-friend contact point.

[0032] After acquiring and fusing key battlefield situational characteristics, the system enters the core reasoning phase, which utilizes the powerful semantic understanding and generation capabilities of a large-scale language model to proactively generate multiple hypotheses about potential future points of contact between friendly and enemy forces. This step aims to address the lag issue in traditional passive rule-triggered early warning systems and provide clear and diverse verification targets for subsequent simulations.

[0033] Specifically, the system first extracts key situational features from step S101, including the comparison of enemy and friendly troop strength, enemy activity intensity index, enemy logistics activity index, friendly defense posture index, and terrain and meteorological factors, and organizes them into a structured text prompt according to a preset format, then inputs it into the large language model. The large language model uses a general-purpose model with strong reasoning capabilities, such as Qwen2-72B-Instruct, and is deployed on a high-performance inference server to ensure response speed.

[0034] To guide large language models in systematic and logically rigorous reasoning, and to prevent them from arbitrarily diverging or producing meaningless output, this method designs a structured thinking framework. This framework consists of three progressively advancing reasoning steps, forcing the model to analyze according to a specific cognitive path.

[0035] The first reasoning step is enemy intent analysis. The system instructs the large language model to answer the following three core questions based on the key situational features input: What are the current characteristics of the enemy's activities? For example, the system will prompt the model to pay attention to specific manifestations such as whether the enemy's reconnaissance frequency has suddenly increased, whether the speed of maneuver has significantly improved, whether the logistics supply lines have been frequently adjusted, and whether communication activities have been encrypted. These features can be directly obtained from the enemy activity intensity index and logistics activity index in step S101. Next, the model needs to infer what kind of enemy intent these activity features indicate. For example, an increase in enemy reconnaissance activities may indicate that they are conducting pre-battle reconnaissance, an increase in the speed of maneuver may indicate that they are gathering at the attack launch position, and frequent adjustments to the logistics supply lines may indicate that they are stockpiling supplies for a large-scale offensive. The model needs to integrate multiple features for cross-validation, rather than looking at a single indicator in isolation. Finally, the model needs to list possible enemy targets, namely, key terrain that the enemy may attempt to seize, important targets to destroy, or sections of defense that they may attempt to breach. These targets are usually related to high-value targets on the battlefield, such as friendly command posts, supply hubs, fire positions, or key passes. Through this step, the model completes a preliminary qualitative analysis of the enemy's current state and possible future moves.

[0036] The second reasoning step involves inferring possible enemy actions. After clarifying the enemy's intentions and potential objectives, the system instructs the large language model to further answer: What specific actions might the enemy take to achieve these objectives? For example, if the enemy's intention is to break through our defenses, possible actions include: concentrating armored forces for a frontal assault, implementing a flanking maneuver, preparing firepower before launching an infantry assault, or launching a surprise attack under cover of darkness and inclement weather. The model needs to consider terrain and meteorological factors to determine which actions are physically feasible. For example, a large-scale armored assault would be unreasonable on swampy terrain, and the model should automatically exclude such assumptions. Next, the model needs to predict the time window for these actions, i.e., how many minutes after the current moment the enemy is likely to launch an action. The prediction of the time window should be estimated based on factors such as the enemy's current distance, speed of movement, and preparation time. For example, if an enemy armored group is 20 kilometers from our defenses and advancing at a speed of 30 kilometers per hour, the time window is approximately 40 to 60 minutes. The model also needs to provide the specific location of the point of contact, outputting it in latitude and longitude coordinates. The point of contact should be inferred based on the enemy's movement route and our defensive deployment. For example, the enemy might choose the weakest section of the defensive line, the most advantageous terrain, or the least fortified area as the main breakthrough point. Through this step, the model transforms the abstract enemy intentions into concrete, verifiable spatial and temporal predictions.

[0037] The third reasoning step is our risk assessment. After predicting the enemy's possible actions, time windows, and points of contact, the system instructs the large language model to assess the risks we face at these points of contact. Specific questions include: What is our defensive posture at the predicted points of contact? The model needs to retrieve our defensive posture index extracted in step S101, including defensive line density, reserve position, and firepower configuration, and combine this with terrain and meteorological factors to determine whether we occupy favorable conditions. For example, if our defensive line density is high, our reserve mobility is strong, and our firepower coverage is comprehensive, the risk is low; conversely, the risk is high. Next, the model needs to assess the level of threat we may face, such as low risk, medium risk, or high risk. The assessment of the threat level can be based on a comprehensive judgment of multiple factors, such as the enemy's troop size, the comparison of our defensive strength, the advantages and disadvantages of terrain, and whether the warning time is sufficient. Finally, the model also needs to initially propose countermeasures that we may need to take, such as reinforcing the troop strength of a certain section of the defensive line, adjusting reserve deployment, conducting firepower counter-preparation in advance, or implementing tactical feints to confuse the enemy. Although specific countermeasures will be further optimized in subsequent simulations, suggestions provided in advance by the large model help commanders quickly understand the situation and form preliminary decision-making ideas.

[0038] After completing the three-step reasoning framework described above, the large language model enters the multi-hypothesis generation stage. This method does not simply generate a single most probable prediction, as battlefield situations are highly uncertain, and a single prediction is easily inaccurate due to incomplete information or model bias. Therefore, the system utilizes the sampling mechanism of the large language model to generate multiple different hypothetical scenarios. Specifically, during the generation process, the model sets the temperature parameter between 0.7 and 0.9. This parameter controls the randomness of the model's output: the lower the temperature, the more certain and conservative the output; the higher the temperature, the more diverse and exploratory the output. Within the temperature range of 0.7 to 0.9, the model can generate a certain degree of variation while maintaining basic rationality, thereby covering multiple possible enemy action paths. For each reasoning iteration, the model generates N different hypotheses, typically 3 to 5. Each hypothesis is a complete structured object containing at least the following five pieces of information: First, a description of the enemy's intentions, briefly explaining in natural language what the enemy wants to do, such as the enemy intending to outflank our defenses from the north; Second, the predicted coordinates of the enemy-friendly contact point, precisely representing the geographical location where the expected engagement will occur in latitude and longitude; Third, the predicted contact time window, expressed in minutes relative to the current time, such as 30 to 45 minutes; Fourth, the predicted contact scale, i.e., the expected number of enemy forces involved in the engagement, such as 20 tanks and 300 infantry; Fifth, a confidence score, ranging from 0 to 1, which is self-assessed by the large language model based on the completeness of the current situation information and the certainty of the reasoning. A higher score indicates that the model is more confident in the hypothesis.

[0039] After generating multiple hypotheses, the system performs a preliminary screening step to determine their reasonableness. Because large language models may rely on statistical regularities in the training data and ignore actual physical constraints or operational principles, they can produce clearly unreasonable hypotheses, such as predicting a large-scale armored assault by the enemy on impassable swampy terrain, or predicting an offensive by the enemy despite a significant numerical disadvantage. The system incorporates a set of basic operational principle constraints and geographical accessibility constraints to check each generated hypothesis. If a hypothesis violates any of these constraints, such as the contact point being located in the center of a lake or the enemy's action time window being shorter than the minimum maneuver time, it is judged as an unreasonable hypothesis and discarded. After screening, the system retains 3 to 5 of the most reasonable hypotheses for the next step of ultra-real-time multi-branch parallel simulation. This screening step significantly reduces the computational burden of subsequent simulations while ensuring that the hypotheses entering the simulation phase have basic physical and tactical credibility.

[0040] See Figure 2 , Figure 2 This is a schematic diagram illustrating the process by which the large language model in this application reasons and generates hypotheses according to a three-layer thinking framework. The process begins with the input of real-time situational data at the top and proceeds through three progressive reasoning steps.

[0041] The first step is enemy intent analysis. The system first observes the situational characteristics, extracting the enemy's current activity patterns from real-time situational data, including reconnaissance frequency, maneuver speed, and logistical adjustments; then it infers the enemy's intent, determining their possible operational objectives, such as preparing for an attack, defense, or feint. The second step is possible action inference. Based on the enemy's intent, the system further infers the specific actions the enemy might take: determining the action type and time window, such as a frontal assault or a flanking maneuver and the expected launch time; simultaneously determining the point of contact and scale, i.e., the predicted coordinates of the enemy-friendly engagement location and the number of enemy forces involved. The third step is friendly risk assessment. The system assesses friendly defensive capabilities at the predicted point of contact, including defensive line density, reserve positions, and firepower configuration; and predicts friendly expected casualties and the probability of the defensive line being breached.

[0042] After completing the three-layer reasoning, the system enters the multi-hypothesis generation stage. Multiple different hypotheses are generated using the sampling mechanism of the large model. The example in the figure generates three hypotheses with confidence levels of 0.78, 0.65, and 0.52, respectively. Each hypothesis contains complete predictive information. Finally, after a reasonableness screening process, unreasonable hypotheses that violate basic operational principles or terrain constraints are eliminated, and 3 to 5 reasonable hypotheses are output and sent to the subsequent simulation and deduction module for verification. This figure fully illustrates the systematic reasoning logic guided by the structured thinking framework of this application.

[0043] In summary, step S102 guides the large language model to perform systematic reasoning through a pre-defined three-layer thinking framework. It utilizes a sampling mechanism to generate multiple diverse hypotheses, which are then filtered for reasonableness, ultimately outputting a set of high-quality, verifiable hypotheses regarding future enemy-ally contact points. These hypotheses leverage the advantages of the large model's rapid association and correlation mining capabilities while also eliminating obvious logical errors through preliminary screening, laying a solid foundation for the next stage of simulation verification.

[0044] S103. For each of the above assumptions, a lightweight combat simulation model is used to perform multi-branch parallel simulations, wherein each simulation branch is given different random perturbations, and each simulation branch outputs simulation results representing the contact situation between the enemy and our side.

[0045] In step S102, the system has generated reasonable hypotheses about future enemy-friendly contact points using a large language model. Each hypothesis includes the enemy's intentions, contact point coordinates, time window, troop size, and confidence score. However, the reasoning of large models is essentially based on statistical laws and lacks accurate modeling of physical constraints, operational dynamics, and random factors, so its output may be biased or misleading. To quickly and reliably verify these hypotheses, this step introduces a lightweight combat simulation engine to independently perform ultra-real-time multi-branch parallel simulations for each hypothesis.

[0046] First, this step employs a lightweight combat simulation model, rather than the complex, parameter-rich models found in traditional combat simulations. Traditional simulations often contain thousands of parameters, involving ballistic flight trajectories, armor penetration depth calculations, electromagnetic spectrum simulations, and logistical supply details, resulting in single simulations taking minutes or even hours, failing to meet real-time early warning requirements. This system, tailored to the characteristics of early warning scenarios, highly simplifies the simulation model, retaining only the key operational factors that determine whether enemy or friendly forces engage and the resulting casualties. Specifically, the retained model elements include the following five aspects.

[0047] First, the system models the positions and movements of enemy and friendly forces. It employs a simplified kinematic model to describe the movement of both sides' forces, ignoring complex factors such as vehicle steering dynamics and the subtle effects of terrain undulations on speed. Based solely on current troop positions, preset routes, average speeds, and terrain mobility correction factors, it linearly extrapolates future positions at various points in time. For example, if enemy armored forces are moving along a road towards a coordinate point at a speed of 30 kilometers per hour, they advance 500 meters per minute. This simplification is sufficient to determine whether either side is likely to enter their respective firing range within a specific time window.

[0048] Second, the fire coverage model. Traditional simulations require detailed calculations of ballistic trajectories, flight times, ammunition dispersion, target acquisition probabilities, etc., which this module simplifies into a geometric model. The system pre-determines the effective ranges of various weapon units; for example, the direct-fire range of a tank gun is 2 kilometers, and the maximum range of a howitzer is 15 kilometers. During the simulation, when enemy forces enter the geometric range of a friendly fire unit, it is considered that the fire unit can engage the enemy; otherwise, it cannot. This geometric model ignores secondary factors such as the influence of complex weather on ballistics, ammunition stockpiles, and fire response time, but it can quickly determine the feasibility of fire support.

[0049] Third, the casualty calculation model. This module uses a simplified version of the Lanchester equations. The Lanchester equations are a classic mathematical model for combat simulation, describing the rate of attrition of forces on both sides during engagement. Traditional simulations require calculating hits and damage per round, resulting in a huge computational burden. This module directly uses the discretized form of the Lanchester square law or linear law to calculate the expected number of casualties per minute based on the number of forces on both sides, firepower effectiveness coefficients, and engagement time step. This simplified model can significantly improve calculation speed without sacrificing too much accuracy.

[0050] Fourth, the logistics constraint model. This module uses a linear model to approximate the constraints of logistics on combat operations. For example, each unit is assumed to have initial ammunition and fuel reserves. Each fire strike or movement a certain distance consumes corresponding resources; when resources fall below a threshold, the unit's attack power or mobility decreases linearly. The system does not simulate complex details such as supply convoy scheduling and transport route planning; it only uses a simple resource consumption formula to determine whether the unit can maintain high-intensity combat operations.

[0051] Fifth, the communication and command delay model. This module uses a fixed delay model to simulate the time required for the command and control chain, from target detection, intelligence reporting, and order issuance to troop execution. Traditional simulations may require simulating random factors such as communication network topology, message collisions, and commander decision-making time. This module directly sets a fixed delay value, such as 30 seconds or 1 minute, assuming a fixed time difference between situational awareness and troop response. This simplification is sufficient to capture the impact of command delays on the operational rhythm while avoiding complex network simulations.

[0052] Building upon the model design, this step further employs a hyper-real-time simulation mechanism. Hyper-real-time refers to the simulation speed being far faster than the actual passage of time. Specific parameter settings are as follows: the simulation time step is set to 1 minute, while traditional simulations often use step sizes in the seconds or even milliseconds. Using a 1-minute step means the system calculates the changes in the troop status of both sides every minute, significantly reducing the number of calculations. The simulation time range is set to the next 30 minutes, corresponding to the contact time window predicted by the large model in step S102, as most enemy actions become clear within 30 minutes. For slower-changing scenarios such as land battlefields, the simulation time range can be extended to 120 minutes. Due to the lightweight model and large step size, the computational complexity of the entire simulation is only O(n), where n is the number of troop units participating in the simulation, typically within a few thousand, allowing a single simulation to be completed in milliseconds, thousands of times faster than traditional simulations.

[0053] To handle inherent battlefield uncertainties, such as the timing of enemy actions, the speed of our response, and random tactical choices, this step simultaneously launches multiple simulation branches for each hypothesis, typically 5 to 10 branches. Each branch uses the exact same lightweight simulation model and initial conditions, but is given different random perturbations. Types of random perturbations include: enemy action delays (e.g., the enemy's actual attack might be delayed by 30 seconds or advanced by 20 seconds); variations in our response time (e.g., faster response when command links are clear, slower response when interference is severe); and random tactical choices (e.g., whether the enemy continues with a direct assault or turns to a flanking maneuver when encountering resistance). These random perturbations are applied to the simulation parameters using a pseudo-random number generator, ensuring that each branch represents a possible future scenario. By running multiple branches, the system can cover the probability distribution of uncertainties, rather than outputting only a single deterministic result.

[0054] Multiple simulation branches are executed in parallel on a graphics processing unit (GPU). Because the simulation calculations for each branch are independent and the lightweight model itself has low computational cost, the GPU's massively parallel architecture can process all branches simultaneously. For example, if 10 branches are initiated for a given hypothesis, the GPU can complete all simulations in almost equal time, with the total time typically controlled between 500 and 1000 milliseconds. This parallel capability enables the system to verify all hypotheses within seconds, truly achieving real-time alerts.

[0055] After each simulation branch is completed, it outputs a set of structured simulation results, specifically including the following six pieces of information: First, whether enemy-ally contact occurred, represented by a Boolean value of yes or no, indicating whether substantial firepower contact occurred between the enemy and friendly forces within the simulation time window under the random perturbation of this branch. Second, the time of contact, in minutes, representing the time from the current moment to the moment of the first contact. If no contact occurred in this branch, this item is empty or marked as infinity. Third, the location of the contact, given in latitude and longitude coordinates, i.e., the geographical point where the first engagement occurred. Fourth, the scale of the contact, i.e., the number of enemy forces involved in the engagement, such as the number of tanks or infantry. Fifth, the expected number of friendly casualties, expressed as a specific number or percentage, i.e., the number of personnel expected to be lost by our side during the contact. Sixth, the probability of our defensive line being breached, ranging from 0 to 1, indicating whether the enemy successfully breached our preset defensive line in the simulation results of this branch, or to what extent. These six results together constitute a branch's quantitative verification conclusion of the hypothesis.

[0056] Taking a specific hypothesis as an example, assuming contact is predicted to occur in 45 minutes with a confidence level of 0.7, the system initiates multiple extrapolation branches for this hypothesis, including a baseline branch, a 5-minute delay branch, a 5-minute advance branch, and other action branches. After each branch runs independently, it enters the statistical aggregation stage. The aggregation results show: the probability of contact is 100%, the average contact time is 45 minutes, the expected number of our casualties is 75, and the probability of the defense line being breached is 12%. The model decision-making stage compares the aggregation results with preset thresholds: the 100% contact probability is greater than 60%, satisfying the condition; the expected number of casualties (75) is greater than 50, also satisfying the condition. Therefore, the system determines that a red alert has been triggered.

[0057] See Figure 3 , Figure 3 This is a flowchart illustrating the structure and operation of the lightweight combat simulation module of this application. The process is divided into four parts from top to bottom: input, simulation engine core, simulation execution, and output. The input part includes hypothesis input and situational data, where the hypothesis comes from the large model generation module, and the situational data comes from the real-time fusion module. The simulation engine core is uniformly scheduled by the simulation engine controller and contains four key models: a troop movement model using simplified kinematics to describe changes in troop positions; a fire coverage model using a geometric model to determine the firing range; a casualty calculation model based on simplified Lanchester equations to calculate troop losses; and a logistics constraint model using a linear model to simulate supply limitations. The simulation execution part adopts a time-stepping mechanism with a step size of 1 minute. The system simultaneously launches 8 to 12 parallel simulation branches for each hypothesis, each branch introducing different random perturbations and utilizing GPUs for accelerated computation to achieve ultra-real-time simulation. The output part generates the simulation results for each branch, including whether enemy-friendly contact occurred, the time and location of the contact, and the expected number of friendly casualties. The entire flowchart clearly demonstrates the three major features of lightweight design, multi-branch parallelism, and GPU acceleration, ensuring that simulation can be completed within milliseconds and meeting the requirements for real-time early warning.

[0058] In summary, through step S103, the system transforms the subjective assumptions generated by the large language model into objective inference results from multiple independent simulation branches. These results include both deterministic information, such as whether, when, and where contact occurred, and uncertain information, such as the range of variation under random perturbations, providing a rich data foundation for subsequent statistical aggregation and early warning decision-making. The lightweight model and parallel execution ensure that the entire verification process is completed within seconds, fully meeting the requirements for online real-time early warning.

[0059] S104. Based on the deduction results of all deduction branches under the same assumption, determine the corresponding contact probability.

[0060] After completing the multi-branch parallel simulation in step S103, each hypothesis generates results for multiple simulation branches. Each branch independently outputs whether enemy-ally contact has occurred, the contact time, the contact location, the enemy troop size, the expected friendly casualties, and the probability of the defensive line being breached. These branch results differ due to the introduction of different random perturbations, and directly using the result of a single branch for early warning decisions would be highly unpredictable. Therefore, this step statistically aggregates all simulation branch results under the same hypothesis, extracts statistically significant quantitative indicators, and uses these indicators to determine whether an early warning is triggered and to determine the warning level.

[0061] First, the system performs aggregate calculations on the results of multiple inference branches under the same assumption. The core metric is the probability of contact, which is calculated as follows: count the number of branches that report enemy-friendly contact among all inference branches under that assumption, divide that number by the total number of inference branches under that assumption, and the resulting ratio is the probability of contact. For example, if 10 inference branches are run under a certain assumption, and 7 of them report enemy-friendly contact, then the probability of contact is 70%. This probability directly reflects the likelihood of the predicted enemy-friendly contact event actually occurring, considering various random perturbations.

[0062] In addition to the probability of contact, the system also calculates the confidence interval for the contact location. Specifically, it collects the contact location coordinates output by all branches where contact has occurred, treats these coordinates as a two-dimensional geographical distribution, calculates the arithmetic mean of all coordinates at its center point, and multiplies the scatter radius (the standard deviation of the distance from each coordinate point to the center point) by a confidence coefficient. For example, the system can output a 95% confidence ellipse, indicating that there is a 95% probability that the actual contact point falls within this elliptical region. This confidence interval provides commanders with the range of uncertainty regarding the contact location, facilitating adjustments to troop deployment.

[0063] The system also calculates the expected value and variance of our casualties. The expected value of casualties is calculated by taking the arithmetic mean of the expected number of our casualties output from all the extrapolation branches under this assumption, regardless of whether contact occurred in that branch. For branches where no contact occurred, the casualty value is usually zero or extremely low. The variance is used to measure the degree of dispersion between the casualty results of each branch. The larger the variance, the more uncertain the casualty results are, and the more cautious we need to be.

[0064] After completing the above statistical aggregation, the system enters the early warning threshold judgment stage. The system presets two core early warning thresholds: the contact probability threshold and the expected casualty threshold. The specific judgment logic is as follows: when the contact probability is greater than 60%, and the expected casualty value on our side is greater than the preset casualty threshold, the system determines that an early warning needs to be triggered. Both conditions must be met simultaneously, because even if the contact probability is high, if the expected casualties are small, no special early warning may be necessary; conversely, even if the expected casualties are large, if the contact probability is low, the actual likelihood of it occurring is low, and an immediate early warning is not required.

[0065] Based on the numerical range of the probability of contact, the system further classifies warning levels to allow commanders to quickly assess the urgency of the threat. When the probability of contact is less than 40%, it is classified as a green warning level, i.e., no warning, indicating that the risk of contact between friendly and enemy forces is low under current assumptions, and no special action is required. When the probability of contact is greater than or equal to 40% but less than 60%, it is classified as a yellow warning level, i.e., a low-risk warning, indicating a certain possibility of contact, and commanders are advised to remain vigilant. When the probability of contact is greater than or equal to 60% but less than 80%, it is classified as an orange warning level, i.e., a medium-risk warning, indicating a high possibility of contact, and commanders are advised to prepare for response. When the probability of contact is greater than or equal to 80%, it is classified as a red warning level, i.e., a high-risk warning, indicating that contact is almost certain to occur, and immediate defensive measures are required.

[0066] Finally, the system generates structured early warning information. This information includes not only the calculated early warning level, the center coordinates of the confidence interval of the predicted enemy-friendly contact point (i.e., the contact location), the predicted contact time window (i.e., the average plus or minus the standard deviation of the contact time for each branch), the predicted contact scale (i.e., the average number of enemy troops), and the expected casualties of our side, but also automatically generates recommended countermeasures. These countermeasures are automatically generated based on the weakest points in our defenses with the highest probability of being breached, combined with effective defensive strategies verified in the simulation. Examples include suggestions to reinforce a section of the defense line, adjust reserve deployment, or implement preemptive fire counter-preparations. Furthermore, the early warning information includes a complete reasoning process, including the thought chain reasoning steps of the large model in step S102 and the key parameter settings of the simulation in step S103, making the early warning results highly interpretable and easy for commanders to understand and trust.

[0067] S105. Generate early warning information based on the probability of contact occurrence corresponding to each hypothesis.

[0068] After statistically aggregating all extrapolation branches under the same assumption, and calculating the probability of contact, expected casualties, confidence interval for contact location, contact time window, and probability of the defense line being breached, the system enters the final early warning information generation stage. That is, based on the aforementioned quantitative indicators and combined with preset early warning thresholds, it outputs structured and interpretable early warning results.

[0069] In one possible implementation, after completing statistical aggregation and early warning level determination, this system not only outputs the early warning conclusion but also provides detailed explanations and trust assessments to address the lack of interpretability in existing early warning systems. Specifically, the system uses interpretability and trust assessment modules to visualize the reasoning process, calculate trust scores, and provide natural language explanations of the early warning basis.

[0070] First, the system displays the complete reasoning process. This includes three levels of visualization: First, the reasoning process of the large model's thought chain, presenting the input and output of each step of the three-step thinking framework executed by the large language model in step S102 in flowchart or text form, allowing commanders to see the complete logical chain from enemy intention analysis to contact point inference to risk assessment. Second, key parameters and results of the simulation, such as the parameter settings of the lightweight simulation model used in step S103, including troop movement speed, fire coverage radius, casualty calculation coefficient, etc., as well as charts such as enemy and friendly troop change curves and casualty rate evolution curves generated during the simulation, helping commanders understand how the simulation derives specific values. Third, the distribution of results from multi-branch simulations, such as histograms of contact probability, scatter plots and confidence ellipses of contact locations on electronic maps, and box plots of contact time, visually demonstrating the magnitude and distribution of uncertainty.

[0071] Secondly, the system calculates a confidence score for the early warning results, ranging from 0 to 1. This score integrates three dimensions: the reasonableness score of the large model assumptions, directly taken from the confidence score of each assumption in step S102, reflecting the model's confidence in its own reasoning; the consistency of simulation results, measured by calculating the variance of the output results of each branch under the same assumption, such as contact location and contact time; the smaller the variance, the more concentrated the results of each branch, and the higher the consistency; and the historical early warning accuracy rate, which is the statistical accuracy rate of this type of early warning over a past period, i.e., the number of actual enemy-friendly contacts after the early warning divided by the total number of early warnings triggered. The result consistency score is obtained by weighted summation of the location consistency score, time consistency score, and contact consistency score. Among them, the location consistency score is calculated based on the variance of the contact location coordinates of each branch, with a higher score for a smaller variance; the time consistency score is calculated based on the variance of the contact time of each branch; and the contact consistency score is equal to the number of branches where contact occurred divided by the total number of branches. The weights of the three sub-scores can be equal or preset according to the scenario.

[0072] The system calculates a trust score using a weighted summation method. The specific formula is: Trust score = 0.4 multiplied by the assumption reasonableness score + 0.3 multiplied by the result consistency score + 0.3 multiplied by the historical early warning accuracy score. This score is output to the commander as a supplementary reference for deciding whether to accept the early warning conclusion.

[0073] Finally, the system automatically generates a natural language explanation of the warning's rationale. This explanation does not use a fixed, templated sentence structure, but is dynamically generated based on the actual reasoning process of this warning. It includes key situational characteristics such as whether the enemy's activity intensity index has increased abnormally, inferred enemy intentions such as a possible flanking maneuver from the north, and risk assessments such as the weakness of our forces in a certain section of the defensive line. By reading this explanation, commanders can clearly understand why the warning was triggered, thereby increasing their trust in the system and their willingness to adopt it.

[0074] In one possible implementation, to overcome the shortcomings of fixed enemy actions and mechanical friendly reactions in traditional linear simulations, a red-blue agent module can be optionally integrated into the multi-branch parallel simulation. This module introduces two agents with autonomous decision-making capabilities, simulating the enemy and friendly forces respectively, and dynamically adjusting their strategies according to changes in the situation during the simulation, thereby making the simulation results closer to the game process in real combat.

[0075] The red agent represents the enemy. Before the simulation begins, the red agent automatically generates multiple possible action plans based on the current battlefield situation and preset combat objectives. These plans must comprehensively consider the enemy's constraints, including the number of remaining troops, the status of logistical supply lines, the integrity of communication networks, and the impact of terrain on maneuverability. For example, if the enemy has sufficient troops and smooth logistics, the red agent may choose a frontal assault; if troops are damaged or resupply is difficult, it may choose a flanking maneuver or harassment. The red agent uses a simplified game theory model to evaluate the expected payoff of each plan and selects the optimal action plan as the enemy's initial strategy in this simulation branch. During the simulation, the red agent continuously monitors changes in the simulation situation, such as adjustments to friendly defensive forces or the movement of reserves, and dynamically adjusts its action plan accordingly, such as shifting from a frontal assault to a flanking maneuver to maximize its breakthrough success rate.

[0076] The blue agent represents our side. Corresponding to the red agent, the blue agent automatically generates multiple possible defense plans based on predictions of enemy actions and our own defensive posture and available resources. Our defensive posture includes defensive line density, reserve force positions, and firepower configuration, while available resources cover troop reserves, ammunition and fuel, and air support capabilities. The blue agent also uses a simplified game theory model to evaluate the effectiveness of each defense plan in response to the enemy's current and anticipated actions, selecting the optimal defense plan as our initial response strategy in that branch. During the simulation, the blue agent adjusts our defensive deployment in real time according to the actual changes in the red agent's actions, such as reinforcing threatened areas, moving reserves to key directions, or implementing tactical feints. This adversarial simulation between the red and blue sides makes the simulation no longer a one-way, pre-scripted execution, but a two-way interactive, strategic game process, the results of which better reflect the uncertainty and adaptability of the real battlefield.

[0077] For example, the specific process of a red-blue agent adversarial simulation can be seen below. The initial situation is that there are five enemy ships 150 kilometers away from our island. The red agent represents the enemy, and its decision-making process includes analyzing the current situation conditions, generating three possible action plans, evaluating the benefits of each plan through a simplified game theory model, and finally selecting the optimal strategy, namely a full-scale attack. The blue agent represents our side, and its decision-making process is based on the predicted enemy actions, assessing the enemy's available resources, generating three defense plans, and selecting the optimal strategy, namely a combined defense, after effectiveness evaluation. The adversarial timeline records key events: at T=0, the enemy is 150 kilometers away; at T=20 minutes, the enemy radar covers our target; at T=45 minutes, the enemy opens fire; at T=50 minutes, one enemy ship is hit, at which point the red agent dynamically adjusts its strategy according to the battlefield situation; at T=60 minutes, the enemy radar covers our target again. The simulation results output includes the simulation timeline, our casualties, and the probability of the defense line being breached. The entire process reflects the adaptive adversarial and dynamic strategy adjustment of both the red and blue sides during the simulation.

[0078] Through adaptive extrapolation by red-blue agents, each branch of the simulation becomes a dynamic, adversarial exercise. The outputs—whether contact occurs, when, where, and the extent of casualties—are based on rational decisions made by both sides, rather than simple linear extrapolation. This significantly enhances the reliability and tactical value of the simulation results, providing a more solid basis for subsequent early warning decisions.

[0079] See Figure 4 , Figure 4 The overall architecture diagram provided for the embodiments of this application shows the overall architecture of the active situation warning system based on large models and parallel inference, which is divided into five layers from top to bottom.

[0080] The data input layer receives multi-source situational data, including real-time data from sensors and intelligence systems such as radar, satellites, drones, and reconnaissance troops.

[0081] The real-time situational awareness layer is handled by Module 1, which integrates and merges multi-source data and extracts key situational features.

[0082] The hypothesis generation layer is responsible for generating the hypothesis of the two major models in Module 2. It reasones according to the preset thinking framework and generates 3 to 5 hypotheses about the future contact points between enemies and friends.

[0083] The simulation and verification layer comprises three collaborative modules. Module 3, the lightweight simulation engine, provides a simplified combat simulation model; Module 4, the multi-branch simulation, runs 8 to 12 simulation branches in parallel, with each branch introducing different random perturbations; Module 5, the red-blue agent adversarial exercise, introduces adaptive adversarial simulations between red and blue agents to enhance the realism of the simulation.

[0084] The assessment and decision-making layer includes Module 6, Result Evaluation and Early Warning, and Module 7, Interpretability Assessment. Module 6 statistically aggregates the results of multi-branch inferences, calculates the probability of contact occurrence, and determines the early warning level; Module 7 generates a trust score and a natural language early warning description.

[0085] The output layer is used to push the final structured early warning information to commanders or decision-making systems. The entire architecture embodies a complete closed loop of data flow from input to output, realizing proactive prediction, rapid verification, and reliable early warning.

[0086] The following example illustrates the method provided in this application: proactive early warning of our armored formation breaking through enemy defenses on the land battlefield.

[0087] This embodiment uses a land battlefield as an application scenario to detail the specific implementation process of the proactive situational awareness system based on large models and parallel simulations provided in this application. For example... Figure 1 As shown, the system includes seven core modules: real-time situational awareness and fusion module, large model hypothesis generation module with a thinking framework, lightweight combat simulation engine, ultra-real-time multi-branch inference module, red-blue intelligent agent and adaptive inference module, result evaluation and early warning decision module, and interpretability and trustworthiness evaluation module.

[0088] The mission scenario is as follows: On a land battlefield, our armored group, consisting of 50 tanks, 100 armored vehicles, and 200 infantry, is assembled approximately 20 kilometers in front of the enemy's defensive line. The enemy's defensive line consists of three defensive positions with a total strength of approximately 500 personnel. The core requirement of our command is to anticipate the enemy's defensive posture and potential breach points in advance, in order to adjust the offensive deployment in a timely manner.

[0089] In terms of system configuration, this embodiment uses the Qwen2-72B-Instruct large language model, deployed on a high-performance inference server. The lightweight simulation model supports parallel simulations of up to 500 troop units. Considering the higher complexity of land warfare compared to naval warfare, the number of simulation branches is set to 12. The warning threshold is defined as a probability of contact greater than 50% and an expected casualty count greater than 100. Due to the relatively slow change in the land battlefield situation, the simulation timeframe is set to the next 120 minutes.

[0090] The specific implementation process includes the following six stages: The first stage is situational awareness acquisition and feature extraction. The real-time situational awareness acquisition module accesses multi-source data from ground radar, UAVs, reconnaissance troops, etc., performs data fusion and key feature extraction, identifies situational factors such as the activity intensity and logistical status of our armored formation, and provides accurate input for subsequent reasoning.

[0091] The second phase involves generating the large-scale model's hypotheses. The large-scale model reasones according to a three-step framework: First, it analyzes our intentions, judging from information such as the accelerated approach of armored formations, increased reconnaissance activities, and adjustments to logistical supply lines, that we are preparing to launch an attack; second, it infers possible actions, predicting that we will launch an armored assault within the next 60 to 90 minutes, with the point of contact located on the enemy's frontal line; third, it assesses the enemy's risk, concluding that the enemy's defensive capabilities are limited and they face medium to high risk. Based on this reasoning, the large-scale model generates 3 to 5 different hypothetical scenarios, each including the predicted contact time, the size of the enemy force, and the model's own confidence score.

[0092] The third phase is ultra-real-time multi-branch simulation. For each generated hypothesis, the system launches a lightweight combat simulation model for simulation. The simulation parameters are set to a time step of 1 minute and a simulation range of 120 minutes. The system executes 12 simulation branches in parallel, each branch introducing different random perturbations, such as delays in our actions and changes in enemy reaction time. The simulation model includes key modules such as troop movement, fire coverage, casualty calculation, and logistical constraints. Each branch outputs results such as the probability of contact, the distribution of contact time and location, and the expected casualty values.

[0093] The fourth stage is result aggregation and evaluation. The system statistically aggregates the results of the 12 branch simulations under the same assumption, calculates the probability of contact (i.e., the proportion of branches where contact occurred to the total number of branches), statistically analyzes the distribution characteristics of contact time and location, and calculates the expected casualties and the probability of the defense line being breached. The aggregated contact probability and expected casualties are compared with preset thresholds to determine whether an early warning is triggered. In this embodiment, the contact probability exceeds 50% and the expected casualties are greater than 100, thus meeting the early warning conditions.

[0094] The fifth stage is a red-blue agent adversarial simulation. To further enhance the realism of the simulation, this embodiment introduces a red-blue agent adversarial simulation. The red agent represents our side and analyzes the optimal offensive strategy based on the current situation; the blue agent represents the enemy and generates the optimal defensive strategy based on predicted our actions. Both agents dynamically adjust their strategies according to changes in the situation during the simulation, making the simulation results closer to the real game adversarial process.

[0095] The sixth stage is interpretability and trustworthiness assessment. The system generates structured, interpretable warning information, demonstrating the three-step reasoning process of the large model, displaying key parameters and statistical results of the simulation, and calculating a trustworthiness score. The trustworthiness score is obtained by weighted summation of the hypothesis reasonableness score, the result consistency score, and the historical accuracy score, with weights of 0.4, 0.3, and 0.3, respectively. This embodiment ultimately generates a natural language warning description, explaining the reasons for the warning in detail and providing corresponding suggestions.

[0096] The system takes only 3 seconds in total, achieving a second-level response from situation input to early warning output; the consistency of results across all inference branches reaches 88%, indicating that the multi-branch results are highly stable and reliable; the trust score is 0.86, belonging to the highly reliable level. Compared with traditional passive rule-triggered schemes, this application achieves proactive prediction, issuing early warnings 60 minutes in advance, with an advance time range of 60% to 80%, fully verifying the technical advantages of this application.

[0097] The method provided in this application is illustrated below through an example from another application scenario: proactive early warning when an enemy fleet approaches our island defense line in a certain sea area. This embodiment, based on Embodiment 1, switches the application scenario from a land battlefield to a naval warfare scenario, specifically proactive early warning when an enemy fleet is preparing to attack our island defense line.

[0098] The mission scenario involves our island defenses facing a direct threat from an enemy fleet in a certain sea area. The current situation is that the enemy fleet is approximately 150 kilometers from our islands and consists of one destroyer, three frigates, and two supply ships. Our defenses are comprised of island air defense positions, coastal artillery positions, and a maritime patrol fleet. Our command needs to anticipate the enemy fleet's offensive intentions and potential firing points in advance.

[0099] Regarding system configuration, the following differences exist compared to Example 1. Due to the smaller number of troop units and relatively open combat space in naval warfare scenarios, the number of simulation branches has been reduced to eight. The early warning threshold has been adjusted to a probability of contact greater than 60% and an expected casualty count greater than 50. Considering the fast pace of naval warfare, the simulation timeframe is set to the next 60 minutes. Other configurations, such as the large model type and simulation engine capabilities, remain consistent with Example 1.

[0100] The key operational process is divided into six stages. The first stage is situational awareness acquisition and feature extraction. The real-time situational awareness acquisition module accesses multi-source data from radar, satellites, UAVs, etc., performs data fusion and key feature extraction, focusing on identifying features such as the activity intensity, logistical supply status, and changes in the course of the enemy fleet, providing input for subsequent inference.

[0101] The second stage involves generating the large-scale model's hypotheses. The large-scale model reasons according to a three-step framework: First, it analyzes the enemy's intentions, judging from information such as the enemy fleet's accelerated approach, increased reconnaissance activities, and adjustments to logistical supply lines, that the enemy is preparing to launch an attack; second, it infers possible actions, predicting that the enemy will launch naval gunfire within the next 30 to 60 minutes, with the main point of contact being our island's air defense positions; third, it assesses our own risks, concluding that our air defense positions have limited defensive capabilities and face moderate risk. Based on the above reasoning, the large-scale model generates 3 to 5 different hypothetical scenarios, each including the predicted contact time, the size of the enemy force, and a confidence score.

[0102] The third phase is ultra-real-time multi-branch simulation. For each generated hypothesis, the system launches a lightweight combat simulation model for simulation. The simulation parameters are set to a time step of 1 minute and a simulation range of 60 minutes. The system executes 8 simulation branches in parallel, each branch introducing different random perturbations, such as enemy action delays and changes in friendly reaction time. The simulation model includes key modules such as troop movement, fire coverage, casualty calculation, and logistical constraints. Each branch outputs results such as the probability of contact, the distribution of contact time and location, and the expected casualty values.

[0103] The fourth stage is result aggregation and evaluation. The system statistically aggregates the results of the eight branches under the same assumption, calculates the probability of contact (i.e., the proportion of branches where contact occurs to the total number of branches), statistically analyzes the distribution characteristics of contact time and location, and calculates the expected casualties and the probability of the defense line being breached. The aggregated contact probability and expected casualties are compared with preset thresholds to determine whether an early warning is triggered. In this embodiment, the contact probability exceeds 60% and the expected casualties are greater than 50, thus meeting the early warning conditions.

[0104] The fifth stage is a red-blue agent adversarial simulation. To enhance the realism of the simulation, this embodiment introduces a red-blue agent adversarial simulation. The red agent represents the enemy and analyzes the optimal offensive strategy based on the current situation; the blue agent represents our side and generates the optimal defensive strategy based on predicted enemy actions. Both agents dynamically adjust their strategies according to changes in the situation during the simulation, making the simulation results closer to real naval combat.

[0105] The sixth stage is interpretability and trustworthiness assessment. The system generates structured, interpretable early warning information, demonstrating the three-step reasoning process of the large model, showcasing key simulation parameters and statistical results, and calculating a trustworthiness score. The trustworthiness score is obtained by weighted summation of the hypothesis reasonableness score, the result consistency score, and the historical accuracy score, with weights of 0.4, 0.3, and 0.3, respectively. Finally, a natural language early warning description is generated, detailing the potential attack window of the enemy fleet, the expected area of ​​fire, and the countermeasures that our side should take, such as strengthening the combat readiness of air defense positions, deploying forward patrol fleets to intercept, or evacuating non-combat personnel on islands.

[0106] Compared with the prior art, the technical solution of this application has the following advantages: First, it enables proactive prediction, providing early warnings 2 to 5 minutes in advance. Large-scale models rapidly generate hypotheses and quickly verify them through simulations, enabling early warnings to be issued before enemy actions, representing a 60% to 80% increase in advance time compared to the 5 to 15-minute lag of passive rule-triggered warnings.

[0107] Second, the accuracy of early warnings has been significantly improved, while the false alarm rate has been greatly reduced. By integrating large-scale model correlation analysis with simulation physical constraint verification, the accuracy of early warnings has increased from 45% to 60% to 82% to 88%, and the false alarm rate has decreased from 25% to 40% to 8% to 12%.

[0108] Third, the adaptive capability is significantly enhanced. The system does not require manual rule definition; the large model automatically learns the correlation between situation and intent, and adapts in real time to changes in the combat environment or enemy tactics. The rule update cycle has been shortened from several days to several weeks to real time.

[0109] Fourth, the comprehensiveness of early warning has been improved. Large-scale model multi-step reasoning and simulation multi-branch verification can uncover implicit correlations between situational elements, increasing the number of threat types that can be detected from 3 to 5 to 8 to 12.

[0110] Fifth, it supports online real-time early warning. The lightweight simulation model and GPU parallel inference reduce the time of a single inference from 5 to 30 minutes to 500 to 1000 milliseconds, a reduction of over 95%.

[0111] Sixth, the early warning results are explainable, increasing commanders' trust. By demonstrating the thought process, key simulation parameters, and trust scores, commanders understand the basis for the early warning, increasing their trust from 40%-50% to 75%-85%.

[0112] The above are some specific implementations of the early warning method provided in the embodiments of this application. Based on this, this application also provides a corresponding early warning system. The system provided in the embodiments of this application will be described below from the perspective of functional modularity. Figure 5 This is a structural diagram of an early warning system provided in an embodiment of this application.

[0113] The system includes: Acquisition unit 110 is used to acquire key situational characteristics of the current battlefield. The reasoning unit 111 is used to input the key situation features into the large language model so that the large language model can reason based on a preset thinking framework and generate multiple hypotheses about the enemy-friend contact point. The simulation unit 112 is used to perform multi-branch parallel simulations for each of the above assumptions using a lightweight combat simulation model, wherein each simulation branch is given different random perturbations, and each simulation branch outputs simulation results characterizing the contact situation between the enemy and our side. Unit 113 is used to determine the corresponding contact probability based on the deduction results of all deduction branches under the same assumption. The generation unit 114 is used to generate early warning information based on the contact occurrence probability corresponding to each hypothesis.

[0114] This application also provides corresponding devices and computer storage media for implementing the early warning scheme provided in this application.

[0115] The device includes a memory and a processor. The memory stores instructions or code, and the processor executes the instructions or code to enable the device to perform the early warning method described in any embodiment of this application.

[0116] The computer storage medium stores code, and when the code is executed, the device running the code implements the early warning method described in any embodiment of this application.

[0117] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems or apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple, and relevant parts can be referred to the method section.

[0118] It should be understood that in this application, "at least one" refers to one or more items, and "more" refers to two or more items. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one" or similar expressions refer to any combination of these items, including any combination of singular or plural items. For example, "at least one" of a, b, or c can represent: a, b, c, a and b, a and c, b and c, or a and b and c, where a, b, and c can be single or multiple.

[0119] It should be understood that the terms center, longitudinal, transverse, up, down, front, back, left, right, vertical, horizontal, top, bottom, inside, outside, etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this application.

[0120] It should be noted that, unless otherwise explicitly specified and limited, the terms installation, connection, and linking should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0121] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the statement "including a…" does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0122] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.

[0123] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method of early warning, characterized in that, include: Obtain key situational characteristics of the current battlefield; The key situational features are input into a large language model, which then performs reasoning based on a preset thinking framework to generate multiple hypotheses about the point of contact between the enemy and ourselves. For each of the aforementioned assumptions, a lightweight combat simulation model is used to perform multi-branch parallel simulations, where each simulation branch is given different random perturbations, and each simulation branch outputs simulation results representing the contact situation between the enemy and our side. Based on the deduction results of all deduction branches under the same assumption, the corresponding contact probability is determined; Early warning information is generated based on the probability of contact corresponding to each hypothesis.

2. The method according to claim 1, characterized in that, The preset thinking framework includes a first reasoning step, a second reasoning step, and a third reasoning step; the first reasoning step is used to instruct the large language model to analyze the enemy activity features extracted from the key situation features in order to determine the enemy's intentions. The second reasoning step is used to instruct the large language model to infer the expected enemy-friendly contact point based on the enemy's intention; The third reasoning step is used to instruct the large language model to assess the risk level faced by our side at the expected enemy-friend contact point based on the expected enemy-friend contact point.

3. The method according to claim 1, characterized in that, The simulation results include our expected number of casualties; Before generating the warning information based on the contact probability corresponding to each hypothesis, the following steps are also included: For each hypothesis, calculate the arithmetic mean of the expected number of casualties on our side output by all branches of the simulation under the hypothesis, and use it as the corresponding expected value of our side's casualties. The generation of early warning information based on the contact probability corresponding to each hypothesis includes: For each hypothesis, the probability of contact occurring corresponding to the hypothesis is compared with a first preset threshold, and the expected value of our side's casualties corresponding to the hypothesis is compared with a second preset threshold. When the probability of the contact occurring is greater than or equal to the first preset threshold, and the expected value of our side's casualties is greater than or equal to the second preset threshold, an early warning message is generated.

4. The method according to claim 3, characterized in that, The simulation results also include: whether enemy-friendly contact occurs, the time when enemy-friendly contact occurs, the location where enemy-friendly contact occurs, the size of enemy forces, and the probability of our defenses being breached; The generated early warning information includes: For each hypothesis, the warning level is determined based on the probability of contact occurring corresponding to the hypothesis; the predicted map coordinates of the contact point are determined based on the contact location output by the branch in all simulation branches under the hypothesis that the contact occurs; the predicted contact time window is determined based on the contact time output by the branch that the contact occurs; and the predicted contact scale is determined based on the enemy force size output by the branch that the contact occurs. The warning information is obtained by integrating the warning level, the predicted point of contact between the enemy and ourselves, the predicted contact time window, the predicted contact scale, and the expected value of our side's casualties.

5. The method according to claim 4, characterized in that, The warning information also includes a trust score, which is determined by the following methods: Calculate the variance of the contact location output by all inference branches under the same assumption, and determine the location consistency score based on the variance; calculate the variance of the contact time output by all inference branches under the same assumption, and determine the time consistency score based on the variance; calculate the ratio of the number of branches that output enemy-friendly contact to the total number of branches in all inference branches under the same assumption, and use this as the contact consistency score; perform a weighted summation of the location consistency score, the time consistency score, and the contact consistency score to obtain the result consistency score; count the number of actual enemy-friendly contact after triggering an early warning in the historical early warning records, and calculate the ratio of this number to the total number of historical early warnings, and use this as the historical early warning accuracy score; obtain the confidence score output by the large language model for each assumption, and use this as the assumption reasonableness score; The confidence score is obtained by weighting and summing the hypothesis rationality score, the result consistency score, and the historical early warning accuracy score.

6. The method according to claim 1, characterized in that, The use of a lightweight combat simulation model for multi-branch parallel simulation includes: An adaptive adversarial simulation is introduced, with a red team agent and a blue team agent participating. The red team agent simulates the enemy, generating and selecting the optimal offensive strategy based on the current situation, combat objectives, troop strength, logistics, and communication constraints. The blue team agent simulates the friendly force, generating and selecting the optimal defensive strategy based on predicted enemy actions, the friendly force's defensive posture, and available resources. During the simulation, the red and blue team agents dynamically adjust their strategies according to real-time situational changes to achieve adaptive adversarial simulation, making the simulation results closely resemble the actual combat confrontation process.

7. The method according to claim 1, characterized in that, After generating multiple hypotheses about the point of contact between the enemy and friendly forces, the process also includes: Unreasonable assumptions that do not conform to basic operational principles and are subject to terrain inaccessibility constraints are eliminated from the multiple assumptions, and a predetermined number of assumptions are retained for subsequent simulations.

8. An early warning system, characterized in that, include: The acquisition unit is used to acquire key situational characteristics of the current battlefield. The reasoning unit is used to input the key situational features into the large language model, so that the large language model can reason based on a preset thinking framework and generate multiple hypotheses about the enemy-friend contact point; The simulation unit is used to perform multi-branch parallel simulations for each of the above assumptions using a lightweight combat simulation model, wherein each simulation branch is given different random perturbations, and each simulation branch outputs simulation results characterizing the contact situation between the enemy and our side. The determination unit is used to determine the corresponding contact probability based on the deduction results of all deduction branches under the same assumption; The generation unit is used to generate early warning information based on the contact probability corresponding to each hypothesis.

9. A computer program product, characterized in that, The computer program product stores instructions that, when executed on a terminal device, cause the terminal device to perform the early warning method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the early warning method as described in any one of claims 1-7.

Images