Low-altitude safety emergency system based on large model application

Abstract

This invention relates to the field of low-altitude safety emergency technology and discloses a low-altitude safety emergency system based on a large-scale model application. The system includes a low-altitude situational awareness module, which constructs a dynamic situational prediction model based on historical flight data, collects low-altitude target trajectory coordinate sequences, radar echoes, and communication signal characteristics in real time, and outputs theoretical safety situation values. A multi-source data fusion analysis module performs a three-dimensional difference analysis on the theoretical and measured situation values, involving temporal trend deviation, airspace coverage offset, and event sequence matching degree, generating a target-level difference feature matrix. A risk propagation path deduction module inputs this matrix into a geographic topology analysis network, combines airspace node connectivity parameters with target locations, and generates a probability heat map of abnormal risk propagation paths to locate risk-affected areas. An intelligent response strategy generation module configures parameters according to the map, enables high-frequency monitoring in high-probability areas, and conducts communication disturbance tests on adjacent targets.

Description

Technical Field

[0001] This invention relates to the field of low-altitude safety emergency technology, specifically a low-altitude safety emergency system based on large-scale model applications. Background Technology

[0002] With the increasing prevalence of drones, low-altitude aircraft, and other aircraft, low-altitude activities are becoming more frequent, significantly increasing the complexity of airspace management. Current low-altitude safety and emergency response systems face multiple technical challenges. Traditional situational awareness methods often rely on single data sources, such as radar or communication signals, making it difficult to comprehensively capture target dynamics. The application of historical flight data is limited to simple statistical analysis, lacking dynamic predictive capabilities, leading to delayed responses to sudden anomalies.

[0003] At the data processing level, existing systems mostly analyze the differences between measured and theoretical data at a two-dimensional level, focusing only on single deviations in time or space and ignoring the impact of correlations in event sequences. This one-sided approach is prone to overlooking potential risks. For example, a small deviation in the time domain of a low-altitude target's trajectory may overlap with airspace coverage offsets, creating a chain of risks. However, current technologies cannot achieve three-dimensional collaborative analysis, resulting in incomplete extraction of difference features.

[0004] In the risk assessment phase, traditional methods rely on human experience to extrapolate risk propagation paths. However, due to the complexity of geographical topology, it is difficult to quickly locate the physical areas affected by risks. Especially in densely populated urban areas, the connectivity parameters of airspace nodes change dynamically, and manual analysis cannot update the risk probability distribution in real time, resulting in insufficient accuracy in locating risk areas and delaying response.

[0005] In terms of response strategies, existing systems mostly adopt fixed patterns to deal with risks, lacking flexibility in adjusting monitoring frequencies for high-probability risk areas. Intervention methods for adjacent node targets are limited, relying solely on drastic measures such as forced removal, which can easily lead to secondary security problems. Furthermore, the application of communication disturbance testing is not yet linked to risk levels; blindly applying disturbances may interfere with normal airspace activities and reduce system operational efficiency.

[0006] Low-altitude safety emergency response involves technical challenges such as multi-domain data fusion and real-time decision-making. Existing technologies have significant shortcomings in dynamic prediction, three-dimensional difference analysis, and intelligent response, making it difficult to meet the safety management and control needs in complex airspace environments. Therefore, it is necessary to build an integrated solution that integrates large-scale model applications. Summary of the Invention

[0007] The purpose of this invention is to provide a low-altitude safety emergency system based on large-scale model applications to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides a low-altitude safety emergency system based on large-scale model applications, the system comprising:

[0009] Low-altitude situational awareness module: Based on historical flight data, a dynamic situational prediction model is constructed, and the current low-altitude target's trajectory coordinate sequence, radar echo characteristics, and communication signal characteristics are collected in real time. The theoretical safety situation value is output through the dynamic situational prediction model.

[0010] Multi-source data fusion analysis module: performs three-dimensional difference analysis on the theoretical security situation value and the actual situation value measured by the monitoring terminal. The three-dimensional difference analysis includes time domain trend deviation, spatial domain coverage offset and event sequence matching degree, and generates a target-level difference feature matrix.

[0011] Risk propagation path simulation module: Input the difference feature matrix into the geographic topology analysis network, combine the spatial node connectivity parameters and target location information to generate a risk probability heat map of abnormal risk propagation paths, and locate the physical areas affected by the risk;

[0012] Intelligent response strategy generation module: Configures response parameters based on the risk probability heat map, including:

[0013] High-frequency situational monitoring mode will be activated for high-probability risk areas;

[0014] Test communication disturbances to adjacent node targets.

[0015] Preferably, the low-altitude situational awareness module specifically includes:

[0016] Historical data feature extraction: Multimodal decomposition processing is performed on historical flight data, including: using a multimodal large model to extract the feature ratio of continuous and abrupt components of track coordinates, establishing a correlation map between communication signal features and target types through semantic association analysis, and using time series alignment algorithms to match radar echo patterns under different scenarios;

[0017] Dynamic situation prediction model construction: The processed historical flight data is input into the hybrid prediction system, which includes:

[0018] A Transformer time-series prediction unit based on the target motion law is used to generate basic situation prediction values. A fully connected network with an embedded spatial attention mechanism corrects prediction biases caused by signal interference. An event feature compensator dynamically adjusts prediction weights based on the characteristics of real-time acquired communication signals.

[0019] The heterogeneous data acquisition unit deployed at the edge synchronously captures the direction change rate and velocity fluctuation parameters of the track coordinates, the short pulse characteristics and energy distribution of the radar echo, and the frequency jump gradient and time interval entropy value of the communication signal.

[0020] Theoretical value calculation: Input the real-time collected data into the dynamic situation prediction model to obtain the theoretical security situation value.

[0021] Preferably, the calculation of the theoretical security situation value includes performing:

[0022] Based on adaptive noise reduction processing during the target flight phase, monitoring noise caused by environmental and meteorological conditions is eliminated;

[0023] The correlation features of track, echo, and signal are integrated through a spatiotemporal feature fusion algorithm;

[0024] The output includes the theoretical security situation value within the normal scenario fluctuation range, and the theoretical security situation value is dynamically updated according to the target state.

[0025] Preferably, the multi-source data fusion analysis module specifically includes:

[0026] Time-domain trend deviation calculation: The theoretical security situation value and the measured situation value are compared by sliding within a preset time window. The time series alignment algorithm is used to match the asynchronously sampled situation sequence, calculate the trend deviation within each window, and generate a time-domain deviation vector.

[0027] Airspace coverage offset detection: The radar echo characteristics of theoretical and measured values ​​are decomposed using a spatial domain decomposition algorithm, the coverage area ratio is calculated, the coverage offset index of each type of target is extracted, and an airspace offset vector is constructed.

[0028] Event sequence matching degree evaluation: Based on the pattern matching algorithm, the feature distribution of theoretical values ​​and measured event sequences is matched, the timing synchronization error of communication signal mutation points is calculated, the distribution difference of event interval distribution is quantified, and an event matching degree vector is generated.

[0029] Difference feature matrix generation: The temporal deviation vector, spatial offset vector, and event matching degree vector are concatenated into tensors. Through normalization processing with feature importance weighting, the dimensional differences are eliminated, and a third-order difference feature matrix with dimensions of [target number × timestamp × difference type] is output.

[0030] Preferably, the spatial coverage offset detection specifically includes:

[0031] In the spatial domain analysis phase, radar echo features corresponding to the theoretical and measured situation sequences are first extracted. Multi-scale feature extraction is used to perform multi-dimensional regional analysis on each set of echo signals to extract coverage distribution features within a preset sensitive area interval. The sensitive area interval is selected to cover the activity range of typical low-altitude targets. After extraction, the coverage density of theoretical and measured data within the sensitive area interval is quantitatively calculated, and the offset index of each type of target is extracted based on the relative offset degree of the two. The coverage offset results of all types are summarized to construct a spatial offset vector.

[0032] Preferably, in the event sequence matching evaluation, the pattern matching algorithm adopts the edit distance algorithm to perform feature matching on the set of mutation points in the two sequences and identify timing synchronization errors.

[0033] Preferably, the risk propagation path deduction module specifically includes:

[0034] Airspace topology modeling: Construct an airspace node connection topology map based on target location information, label the communication connectivity parameters between each node, overlay the constraints of temporary no-fly zones on the topology map, and generate a geographic topology model including connectivity matrix and node influence matrix;

[0035] Risk propagation simulation: Mapping the difference feature matrix to the corresponding nodes of the spatial topology model; performing risk propagation inference based on graph neural networks, the calculation of risk propagation inference includes:

[0036] i. Calculate the attenuation factor of risk impact based on node connectivity parameters;

[0037] ii. Capture cross-regional risk correlation characteristics through multi-head attention mechanisms;

[0038] iii. The Monte Carlo method is used to simulate the diffusion path of risk impacts in the topological network;

[0039] Probability distribution generation: Statistically analyze the frequency of risk impacts in each region during simulated propagation, calculate the risk impact retention probability value by combining regional connectivity parameters, generate a risk probability heat map covering the entire airspace, and mark the set of suspicious regions whose probability values ​​exceed the preset risk retention probability threshold;

[0040] Physical region location: Perform spatial clustering analysis on the risk probability heat map to identify risk probability clusters; delineate the physical boundaries of risk impact based on the target location and the spatial topology connection relationship.

[0041] Preferably, the risk propagation path simulation module further includes outputs including suspicious target identifiers and the main risk propagation path, wherein:

[0042] The suspicious target identifier is based on the airspace nodes connected to the suspicious area set. The airspace nodes are bound to the actual targets to form a suspicious target identifier set, indicating potential risk sources or affected terminals.

[0043] The risk propagation master path is obtained by recording the node paths and their order during each round of risk diffusion in the Monte Carlo simulation. Among all simulated paths, the frequency of occurrence of each path is counted, and the path sequence with the highest cumulative frequency is selected as the risk propagation master path. The output risk propagation master path sequence is a structured and ordered list of nodes, reflecting the main propagation trajectory of risk information in the airspace.

[0044] Preferably, the spatial topology modeling specifically includes:

[0045] The target location information is normalized to construct a three-dimensional topology map containing the latitude and longitude coordinates and altitude levels of the nodes; the communication delay parameters and signal attenuation coefficients between each node are labeled; the boundary coordinates and time range of the temporary flight restriction zone are superimposed on the topology map to generate a geographic topology model containing the connectivity strength matrix and the influence weight of the nodes.

[0046] Preferably, after the main risk propagation path is generated, the process further includes executing:

[0047] Historical data of suspicious targets on the main path are backtracked to extract their flight path characteristics, echo characteristics and communication signal characteristics during their historical flight periods, and a historical feature profile of the targets is constructed.

[0048] By using a large model similarity calculation algorithm, the feature matching degree between the current main path of risk propagation and the path of historical risk events is compared, and path similarity assessment results are generated.

[0049] Based on the path similarity assessment results and the set of suspicious target identifiers, update the preset risk dwell probability threshold and adjust the display parameters of the anomaly probability heat map.

[0050] Compared with the prior art, the beneficial effects of the present invention are:

[0051] Through the collaborative operation of multiple modules, a comprehensive upgrade of low-altitude safety emergency management has been achieved. The low-altitude situational awareness module constructs a dynamic situation prediction model based on historical flight data. Combined with real-time collected multi-dimensional features, it can generate theoretical safety situation values ​​that are more realistic, changing the perception bias caused by the reliance on a single data source in traditional systems. The real-time collected track coordinate sequences, radar echo characteristics, and communication signal characteristics, after being processed by the model, can predict the target movement trend in advance, avoiding passive responses to emergencies.

[0052] The multi-source data fusion analysis module introduces three-dimensional difference analysis, which overcomes the limitations of existing two-dimensional analysis. It incorporates temporal trend deviation, spatial coverage offset, and event sequence matching degree into a unified framework, generating a target-level difference feature matrix that can more accurately capture subtle differences between data. This comprehensive difference analysis can identify potential risks that are difficult to detect with single-dimensional analysis. For example, if a small deviation of a target in the temporal domain occurs simultaneously with spatial coverage offset and event sequence mismatch, three-dimensional analysis can promptly detect and generate corresponding difference features, providing richer evidence for subsequent risk assessment.

[0053] The risk propagation path simulation module leverages a geographic topology analysis network to combine a difference feature matrix with airspace node connectivity parameters and target location information. The resulting risk probability heatmap visually presents the propagation trend of abnormal risks and accurately locates the physical areas affected by these risks. Compared to traditional manual simulation methods, this module can quickly handle complex geographic topological relationships, especially in densely populated urban airspace. It can update the impact of changes in node connectivity on risk propagation in real time, avoiding misjudgments of risk areas due to delays in manual calculations.

[0054] The intelligent response strategy generation module dynamically configures response parameters based on the risk probability heat map. It activates a high-frequency situational monitoring mode for high-probability risk areas, capturing more details before the risk escalates. Furthermore, it applies communication disturbance tests to adjacent target nodes, verifying target response characteristics and determining the presence of anomalies without interfering with normal airspace activities. This differentiated approach improves the targeting of emergency responses while reducing interference with normal low-altitude activities, addressing the inflexibility and secondary risks inherent in traditional fixed-mode responses. Attached Figure Description

[0055] Figure 1 This is a schematic diagram illustrating the working principle of the low-altitude safety emergency system based on large-scale model application described in this invention.

[0056] Figure 2 A flowchart for the low-altitude situational awareness module;

[0057] Figure 3 A flowchart for the multi-source data fusion and analysis module;

[0058] Figure 4 A flowchart for the risk propagation path simulation module;

[0059] Figure 5 This is a flowchart of the post-processing of the main risk propagation path. Detailed Implementation

[0060] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] Please see Figures 1-5 This invention provides a low-altitude safety emergency response system based on a large-scale model application. The system includes: a low-altitude situational awareness module, a multi-source data fusion and analysis module, a risk propagation path simulation module, and an intelligent response strategy generation module. These modules work collaboratively to achieve dynamic monitoring and emergency response for low-altitude safety. The specific implementation steps are as follows:

[0062] The low-altitude situational awareness module constructs a dynamic situational prediction model based on historical flight data, collecting real-time flight path coordinate sequences, radar echo characteristics, and communication signal characteristics of current low-altitude targets. The dynamic situational prediction model outputs a theoretical safety situation value. The multi-source data fusion analysis module performs a three-dimensional difference analysis between the theoretical safety situation value and the measured situation value from the monitoring terminal. This analysis includes temporal trend deviation, airspace coverage offset, and event sequence matching degree, generating a target-level difference feature matrix. The risk propagation path deduction module inputs the difference feature matrix into a geographic topology analysis network, combining airspace node connectivity parameters and target location information to generate a risk probability heat map of abnormal risk propagation paths, locating the physical areas affected by the risk. The intelligent response strategy generation module configures response parameters based on the risk probability heat map, including activating a high-frequency situational monitoring mode for high-probability risk areas and applying communication disturbance tests to adjacent node targets.

[0063] Example 1:

[0064] The operation of this embodiment begins with the extraction of historical data features. When performing multimodal decomposition on historical flight data, a large multimodal model is used to analyze the trajectory coordinates, separating continuous and abrupt components, and calculating the proportion of each component in the overall trajectory features. This distinguishes the trajectory features of the target under normal flight conditions and sudden changes in direction. Through semantic association analysis, various features contained in the communication signals, such as signal frequency, modulation method, and transmission rate, are linked to the type information of low-altitude targets to establish an association map, enabling different types of targets (such as fixed-wing aircraft, helicopters, and drones) to be identified through their communication signal features. Simultaneously, a time-series alignment algorithm is used to match radar echo patterns collected in different scenarios (such as clear weather, rainy weather, nighttime, and complex terrain areas) to identify the commonalities and differences in echo features across various scenarios, providing basic data for scenario adaptation in subsequent situation prediction.

[0065] The construction of a dynamic situation prediction model requires inputting historical flight data, processed by multimodal decomposition, into a hybrid prediction system. This hybrid prediction system comprises three core components: a Transformer time-series prediction unit based on target motion patterns, which learns from historical flight data about the target's velocity, acceleration, and turning patterns to generate basic situation prediction values, reflecting the theoretical motion of the target under conditions free from external interference; a fully connected network embedded with an airspace attention mechanism, which can monitor signal interference at different airspace locations and correct prediction deviations caused by signal interference. For example, when strong electromagnetic interference in a certain airspace causes radar echo distortion, the network can adjust the prediction results based on the normal signal characteristics of the surrounding airspace; and an event feature compensator, which captures changes in the communication signal characteristics of the current low-altitude target in real time, such as sudden weakening of signal strength or abnormal frequency jumps, and dynamically adjusts the prediction weights based on these characteristics to make the prediction results more consistent with the target's current actual communication status.

[0066] In the real-time data acquisition phase, multiple parameters are simultaneously acquired through heterogeneous data acquisition units deployed at the edge. For track coordinates, the focus is on capturing the direction change rate and velocity fluctuation parameters. The direction change rate describes the angle by which the target's heading changes per unit time, while the velocity fluctuation parameter reflects the magnitude of the target's flight speed change over a short period of time. For radar echoes, short pulse characteristics are acquired, including pulse width, pulse repetition frequency, and the energy distribution of the echo signal, i.e., the distribution of energy at different frequencies and time dimensions. For communication signals, the frequency jump gradient, i.e., the amount of frequency change per unit time, and the time interval entropy value are recorded. This value is used to measure the degree of disorder in the communication signal transmission interval; a higher entropy value indicates a more unstable signal transmission interval.

[0067] In the theoretical calculation phase, the real-time collected track coordinate sequence, radar echo characteristics, and communication signal characteristics are input into the constructed dynamic situation prediction model. During the calculation process, adaptive noise reduction processing based on the target's flight phase is first performed. This eliminates noise caused by environmental meteorological conditions (such as wind speed, precipitation, and cloud thickness) on the monitored data at different flight phases, including takeoff, climb, level flight, descent, and landing. For example, during takeoff, the strong signal from the target's engine may mask some environmental noise, while during level flight, noise generated by airflow disturbances is more pronounced. The adaptive noise reduction processing adjusts the noise reduction parameters according to the noise characteristics of different phases to ensure the accuracy of the data input into the model.

[0068] A spatiotemporal feature fusion algorithm integrates the correlation features of track, echo, and signal. From a temporal perspective, it matches the three types of data collected at the same time point to analyze the correlation between the target's motion state, physical attributes, and communication behavior at that moment. From a spatial perspective, it combines the target's location information to analyze the distribution patterns of the three types of data in different airspace locations; for example, whether the radar echo characteristics of a certain area match the track characteristics of common targets in that area. Through this spatiotemporal fusion, a comprehensive feature set reflecting the target's state is formed.

[0069] The final output theoretical safety situation value includes the normal scenario fluctuation range. This range is determined based on the situation value distribution of a large number of normal flight cases in historical data, and can reflect the range of changes in the target's situation value under normal flight conditions. Furthermore, the theoretical safety situation value is updated in real time as the target's state changes dynamically. When the target's flight speed, direction, altitude, communication mode, etc., change, the model will recalculate and output a new theoretical safety situation value to continuously reflect the target's current theoretical safety status.

[0070] Example 2:

[0071] The calculation of theoretical safety situation values ​​requires multiple processing steps. Adaptive noise reduction processing is performed based on the target's flight phase. The environmental and meteorological interference faced by the target differs at different flight phases. For example, during takeoff, the target is at low altitude and low speed, making it susceptible to clutter reflected from ground structures, and near-surface airflow disturbances may cause irregular fluctuations in radar echoes. During level flight, the target's altitude is stable, mainly affected by changes in high-altitude wind speed, which may cause slight drifts in the track coordinates. During descent, the target's speed gradually decreases, and its course is frequently adjusted. At this time, precipitation or fog will exacerbate radar signal attenuation, leading to blurred echo characteristics. To address these situations, adaptive noise reduction processing calls upon corresponding noise reduction parameters based on the target's current flight phase to filter noise from the monitoring data. For clutter interference during takeoff, an airspace filtering algorithm is used to retain signal components consistent with the target's motion trend and eliminate isolated strong reflection points. For wind speed interference during level flight, time smoothing is used to eliminate small fluctuations in the flight path within a short period of time. For the impact of weather attenuation during landing, the echo energy is compensated based on the signal attenuation patterns under the same weather conditions during the same period in history, so that the processed monitoring data is closer to the true state of the target.

[0072] The spatiotemporal feature fusion algorithm integrates the correlation features of track, echo, and signal. Track coordinate sequences record the target's motion trajectory in space, including information such as position, velocity, and direction, reflecting the target's dynamic position changes. Radar echo features reflect the target's physical attributes; for example, echo intensity is related to target volume, echo spectrum distribution is related to target material properties, and echo pulse width is related to target velocity. Communication signal features include the target's identification, communication frequency, and signal modulation method, which can be used to distinguish target type and communication status. In the time dimension, the spatiotemporal feature fusion algorithm binds track, echo, and signal data collected at the same time, establishing a multi-feature set corresponding to the timestamp. In the spatial dimension, based on the target's location information, it matches the track coordinates with the radar echo coverage area and communication signal propagation path at that location, analyzing the correlation of the three in spatial distribution. For example, when the target changes course, the direction parameters of its track coordinates change, and the Doppler frequency shift of the radar echo changes accordingly. The transmission direction of the communication signal may also adjust. The algorithm captures these synchronously changing features, forming a correlation feature vector. For data acquired asynchronously over time, the algorithm uses interpolation and alignment to map features from different acquisition times onto the same time axis, ensuring the continuity of the fused features in the time dimension.

[0073] The output theoretical safety situation value includes a normal scenario fluctuation range, which is determined based on the distribution of situation values ​​under normal flight conditions in historical flight data. Through analysis of numerous historical normal flight cases, the range of situation value changes under different flight stages, weather conditions, and target types is statistically analyzed. For example, during a clear-sky level flight phase, the theoretical safety situation value of a certain type of UAV typically fluctuates within a certain range, which is the normal fluctuation range for that scenario. The theoretical safety situation value is dynamically updated according to the target state. When the target state changes, such as a sudden increase in speed, deviation from the preset route, or abnormal communication signal switching frequency, the dynamic situation prediction model will recalculate the theoretical safety situation value. During the update process, the model prioritizes the most recently collected monitoring data, combining the characteristic correlations before and after the target state change to adjust the prediction parameters. For example, when the target transitions from level flight to climb, the model will call upon the motion law parameters of the climb phase, combining real-time collected speed increments and altitude change rates to regenerate the theoretical safety situation value and simultaneously update the normal scenario fluctuation range to reflect the situation value change characteristics during the climb phase. This dynamic update mechanism enables the theoretical security situation value to track changes in the target's state in real time, accurately reflecting the target's theoretical security level in the current state.

[0074] Example 3:

[0075] This embodiment begins with the calculation of time-domain trend deviation, using a preset time window to perform a sliding comparison between theoretical and measured security situation values. The length of the preset time window can be determined based on the movement speed of the low-altitude target and the sampling frequency of the monitoring data. The window length must ensure that the target's movement state reflects a certain trend within that time period, while avoiding blurring of trend characteristics due to an excessively long window. A time series alignment algorithm is used to process asynchronously sampled situation sequences. Since the acquisition time points of theoretical and measured values ​​may differ, the algorithm aligns the situation sequences by finding the optimal matching position between the two on the time axis. For example, when the sampling interval for theoretical values ​​is 2 seconds and the sampling interval for measured values ​​is 3 seconds, the algorithm will supplement the missing time point data through interpolation or by sampling to establish a correspondence between the two at the same time point. Subsequently, the trend deviation within each window is calculated, i.e., the degree of deviation between the theoretical value trend line and the measured value trend line, generating a time-domain deviation vector. Each element of this vector corresponds to the trend deviation of a time window.

[0076] In airspace coverage offset detection, the radar echo features of theoretical and measured values ​​are decomposed using a spatial domain decomposition algorithm. The echo features are decomposed into components of different spatial regions. By calculating the coverage range of each component, the ratio of the theoretical echo coverage area to the measured echo coverage area is obtained. For different types of low-altitude targets, such as small UAVs, medium-sized helicopters, and large transport aircraft, their coverage offset indices are extracted. This index comprehensively considers factors such as the offset of the center position of the echo coverage and the similarity of the coverage shape, thereby constructing an airspace offset vector. The elements in the vector correspond to the degree of coverage offset for different types of targets. In the spatial domain analysis stage, the radar echo features corresponding to the theoretical and measured situation sequences are first extracted. Multi-scale feature extraction is used to analyze each set of echo signals, examining the coverage distribution of the echoes at different spatial scales (such as 100m×100m and 500m×500m grids). Pre-defined sensitive area intervals are selected, covering the common activity ranges of typical low-altitude targets, such as the 5-kilometer radius around airport runways and areas below 1 kilometer altitude above urban centers. Echo coverage distribution characteristics within these areas are extracted, including coverage density and uniformity. The coverage density of theoretical and measured data within the sensitive area intervals is quantified, for example, by counting the number of echo points per unit area. Based on the relative offset between the two, such as the distance deviation of the coverage center and the overlap ratio of high-density areas, offset indices for each type of target are extracted. All coverage offset results are then summarized to form an airspace offset vector.

[0077] Event sequence matching evaluation is based on a pattern matching algorithm, which uses the edit distance algorithm to match the feature distributions of theoretical and measured event sequences. Event sequences include events such as communication signal mutations, abnormal track changes, and sudden speed changes. The edit distance algorithm measures the degree of matching by calculating the minimum number of editing operations (insertion, deletion, and replacement events) required to convert the theoretical event sequence into the measured event sequence. The algorithm identifies the matching degree of features of mutation point sets in the two sequences, such as the signal strength change and duration of communication signal mutations, thereby determining the timing synchronization error, i.e., the time difference between theoretical and measured mutation points. It also quantifies the distributional difference of event intervals, such as the difference between the standard deviation of theoretical and measured event intervals, and generates an event matching degree vector. The vector elements reflect the degree of matching of the event sequences in different dimensions.

[0078] When generating the difference feature matrix, the temporal deviation vector, spatial offset vector, and event matching degree vector are tensor-concatenated to form a preliminary three-dimensional matrix. Through feature importance-weighted normalization, different weight values ​​are assigned to each vector based on its actual impact on risk assessment. For example, temporal deviation has a greater impact on short-term risk and can be assigned a higher weight; spatial offset has a more significant impact on regional risk and is also assigned a corresponding weight. Normalization eliminates the differences in numerical range caused by different dimensions of the vectors, transforming all element values ​​to the same interval. The final output is a third-order difference feature matrix with dimensions [target number × timestamp × difference type]. The target number corresponds to a specific low-altitude target, the timestamp corresponds to different monitoring times, and the difference type is divided into three categories: temporal, spatial, and event sequence. Each element in the matrix represents the specific numerical value of a certain type of difference for a specific target at a specific timestamp.

[0079] Example 4:

[0080] This embodiment begins with airspace topology modeling, constructing a topology graph of airspace node connections based on target location information. In this process, the target location information is first normalized, converting latitude and longitude coordinates from different sources into a unified coordinate system. Simultaneously, height information is incorporated to construct a three-dimensional topology graph containing node latitude and longitude coordinates and height levels. For example, an airspace can be divided into multiple levels based on height, such as below 100 meters, 100-500 meters, and 500-1000 meters, with target locations within each level forming an independent node network. Communication delay parameters and signal attenuation coefficients between nodes are labeled. For instance, the communication delay between two nodes may vary depending on distance, and the signal attenuation coefficient is related to obstacles between nodes; if tall buildings obstruct the view, the attenuation coefficient will increase accordingly. The boundary coordinates and time range of temporary flight restriction zones are superimposed on the topology map. For example, the boundary coordinates of a temporary no-fly zone designated during a large event are determined by the vertices of a polygon, and the time range is accurate to the hour. Finally, a geographical topology model containing a connectivity strength matrix and node influence weights is generated. The values ​​in the connectivity strength matrix reflect the stability of the connections between nodes, and the node influence weights are assigned according to the importance of the node's location. For example, the influence weight of nodes near airports is higher than that of nodes in remote areas.

[0081] In the risk propagation simulation phase, the difference feature matrix is ​​mapped to the corresponding nodes in the spatial topology model. Each node obtains an initial risk value based on its own difference feature values. When performing risk propagation simulation based on graph neural networks, the attenuation factor of the risk impact is first calculated based on the node connectivity parameters. If the communication delay between two nodes is high and the signal attenuation is large, it indicates weak connectivity. The attenuation factor when the risk propagates from one node to another is large, resulting in a smaller risk value transmitted. A multi-head attention mechanism is used to capture cross-regional risk association characteristics. For example, when a risk occurs in a certain region, this mechanism can identify the potential impact on other regions that are indirectly connected to it, even if there are no direct node connections between these regions. The Monte Carlo method is used to simulate the diffusion path of risk impact in the topology network. By randomly generating a large number of possible propagation scenarios, in each scenario, the risk starts from the initial node and gradually spreads according to the connectivity relationship and attenuation factor between nodes. The order and time of the risk arriving at each node in each simulation are recorded.

[0082] In the probability distribution generation stage, the frequency of risk impacts occurring in each region during multiple simulated propagation is statistically analyzed. The more times a region is affected by risk, the greater the likelihood that the region faces risk. The risk impact retention probability value is calculated by combining regional connectivity parameters. If a region has strong connectivity with surrounding nodes, the probability of risk retention in that region is relatively high. A risk probability heat map covering the entire airspace is generated. Different colors are used to indicate the risk probability of each region: red represents high-probability areas, blue represents low-probability areas, and a set of suspicious regions with probability values ​​exceeding a preset risk retention probability threshold is marked. For example, if the threshold is set to a certain value, all consecutive regions with probabilities exceeding that value are marked as suspicious regions.

[0083] When locating physical regions, spatial clustering analysis is performed on the risk probability heat map. Clustering algorithms are used to merge areas that are close to each other and have similar risk probabilities into a single risk probability cluster. For example, multiple small high-probability areas in the city center may be aggregated into a large cluster. Based on the target location and the spatial topology, the physical boundary of the risk impact is delineated. The boundary line is usually determined along the node connection line at the edge of the cluster, ensuring that all high-probability risk areas are included.

[0084] The risk propagation path simulation module also outputs suspicious target identifiers and the main risk propagation path. Suspicious target identifiers are based on the airspace nodes connected to the set of suspicious areas, binding each airspace node to an actual low-altitude target. For example, if a node in a suspicious area corresponds to three drones, these three drones are included in the suspicious target identifier set. The main risk propagation path records the node paths and their order during each round of risk diffusion in the Monte Carlo simulation. The frequency of occurrence of each path is counted across all simulated paths, and the path sequence with the highest cumulative frequency is selected as the main risk propagation path. For example, if a path appears 300 times in 1000 simulations, significantly more than other paths, it is determined as the main path. The output path sequence is arranged sequentially by node number, clearly showing the main propagation trajectory of the risk from the initial node to each area.

[0085] After the main risk propagation path is generated, historical data of suspicious targets along the path is reviewed to extract their past flight trajectory characteristics, such as frequent entry into no-fly zones or sudden changes in speed and direction; radar echo characteristics, such as whether the echo intensity matches the target type; and communication signal characteristics, such as whether there are abnormal frequency jumps. A historical feature profile of the targets is constructed, and a large-scale model similarity calculation algorithm is used to compare the feature matching degree between the current risk propagation path and historical risk event paths. If the node connection patterns, risk spread speeds, and other characteristics are similar, it indicates that the current risk may be related to a historical event. Based on the path similarity assessment results and the set of suspicious target identifiers, the preset risk retention probability threshold is updated, and the display parameters of the anomaly probability heat map are adjusted, such as changing the color gradient to make high-risk areas more prominent.

[0086] Example 5:

[0087] The intelligent response strategy generation module uses a risk probability heat map as its core basis and configures corresponding response parameters to achieve dynamic response to low-altitude safety risks. The risk probability heat map visually presents the risk probability distribution of various areas in the airspace through different color gradients, with darker colors representing higher risk probabilities. These areas are often the focus of risk prevention and control.

[0088] To activate high-frequency situational awareness monitoring in high-probability risk areas, the specific boundaries of these areas must first be clearly defined. Based on the risk probability values ​​marked on the heat map and considering actual airspace management needs, the boundaries of high-probability risk areas are delineated. These boundaries typically coincide with the edge of the area corresponding to the preset probability threshold on the risk probability heat map. After entering high-frequency situational awareness monitoring mode, the parameters of the monitoring equipment within this area are adjusted, shortening the data acquisition interval. For example, the frequency of track coordinate acquisition is increased from once every 30 seconds in the normal mode to once every 5 seconds, ensuring that subtle changes in target movement can be captured. Simultaneously, the number of radar scans is increased, and the sampling density of radar echo characteristics is improved, resulting in more detailed information on short-pulse characteristics and energy distribution of the echoes. For communication signal monitoring, the monitoring range of the signal frequency band is expanded, and the calculation granularity of frequency jump gradients and time interval entropy values ​​is refined to more sensitively detect signal anomalies. During high-frequency monitoring, the heterogeneous data acquisition units deployed at the edge must operate at full capacity, simultaneously processing massive amounts of real-time acquired data and rapidly transmitting the data to the back-end processing system to ensure the timeliness of monitoring information. In addition, the high-frequency situation monitoring mode will also activate a multi-device collaborative monitoring mechanism, mobilize backup monitoring equipment in the surrounding area, and cross-verify the target from different angles and distances to reduce blind spots or errors that may exist in single-device monitoring. Through the mutual verification of multi-source data, the accuracy of situational awareness of high-probability risk areas will be improved.

[0089] To conduct communication disturbance tests on adjacent target nodes, the range of these adjacent targets must first be determined. Based on the risk probability heat map and airspace topology model output by the risk propagation path simulation module, low-altitude targets corresponding to nodes with direct communication connections around high-probability risk areas are identified; these targets are the adjacent target nodes. The specific method of communication disturbance testing includes sending interference signals at specific frequencies. The frequency of the interference signal is selected based on the frequency range of the target's usual communication signals, typically fluctuating slightly around its commonly used frequency to avoid causing widespread interference to normal communication in non-target areas. The strength of the interference signal needs to be precisely controlled, ideally causing just a slight response from the target's communication system, without causing communication interruption, while still allowing observation of the target's response mechanism to the disturbance. During the application of communication disturbance, changes in the target's communication signal characteristics are monitored simultaneously, recording whether the target adjusts its communication frequency, increases signal strength, or changes its communication protocol, etc. Simultaneously, changes in the target's flight path coordinates are tracked to observe whether its flight trajectory or speed changes due to communication disturbance. During the test, the impact range of the disturbance signal was monitored in real time to ensure that the disturbance was confined to the airspace of the adjacent target node and did not spread to other unrelated areas. Each communication disturbance test lasted for a certain period of time. After the test, the transmission of interference signals was stopped, and the target's communication signals and track characteristics were monitored to observe whether it returned to its pre-disturbance state. By comparing the target's state before and after the disturbance, a reference was provided for judging the target's risk level and formulating further handling strategies.

[0090] The intelligent response strategy generation module adjusts its strategy in real time based on the dynamic changes in the risk probability heat map when configuring response parameters. When the area of ​​a high-probability risk zone expands or the risk probability increases, the number of high-frequency situational monitoring devices and the monitoring frequency are increased, while the target range for communication disturbance testing is expanded. Conversely, when the risk probability decreases or the area shrinks, the monitoring frequency is reduced accordingly, decreasing the intensity and frequency of disturbance tests to minimize the impact on normal low-altitude flight activities while ensuring safety. Real-time data exchange is maintained between modules. The risk propagation path simulation module continuously updates the risk probability heat map, and the intelligent response strategy generation module dynamically optimizes response parameters based on the updated information, forming a closed-loop emergency response mechanism.

[0091] It should 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 comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A low-altitude safety emergency system based on a large model application, characterized in that, Comprise: Low-altitude situation awareness module: based on historical flight data to build a dynamic situation prediction model, real-time collection of current low-altitude target track coordinate sequence, radar echo characteristics and communication signal characteristics, through the dynamic situation prediction model output theoretical safety situation value; Multi-source data fusion analysis module: three-dimensional difference analysis of the theoretical safety situation value and the measured situation value of the monitoring terminal, the three-dimensional difference analysis includes time domain trend deviation, spatial coverage deviation and event sequence matching degree, and generates a difference feature matrix of the target level; Risk propagation path deduction module: input the difference feature matrix into the geographical topology analysis network, combine the airspace node connectivity parameters and target location information, generate a risk probability heat distribution map of the abnormal risk propagation path, locate the risk impact physical area, the risk propagation path deduction module specifically comprises: Airspace topology modeling: according to the target location information, build the connection relationship topology graph of the airspace nodes, label the communication connectivity parameters between nodes, superimpose the restriction conditions of the temporary flight restricted area in the topology graph, generate a geographical topology model including connectivity matrix and node influence matrix; Risk propagation simulation: map the difference feature matrix to the corresponding node of the airspace topology model; based on graph neural network to perform risk propagation deduction, the calculation of risk propagation deduction includes: i. Calculate the risk impact decay factor according to the node connectivity parameters; ii. Capture cross-regional risk correlation characteristics through multi-head attention mechanism; iii. Use Monte Carlo method to simulate the diffusion path of risk impact in the topology network; Probability distribution generation: statistics of the risk impact frequency of each region in the simulation propagation, combined with the region connectivity parameters to calculate the risk impact residence probability value, generate a risk probability heat distribution map covering the whole airspace, label the suspicious region set whose probability value exceeds the preset risk residence probability threshold; Physical area positioning: perform spatial clustering analysis on the risk probability heat distribution map to identify risk probability aggregation areas; according to the target location and the connection relationship of the airspace topology, the risk impact physical edge is drawn; Intelligent disposal strategy generation module: configure disposal parameters according to the risk probability heat distribution map, including: Enable high-frequency situation monitoring mode for high-probability risk areas; Apply communication disturbance test to adjacent node targets.

2. The large model application-based low-altitude safety emergency system according to claim 1, characterized in that, The low-altitude situation awareness module specifically comprises: Historical data feature extraction: multi-modal decomposition processing of historical flight data, including: using multi-modal large model to extract the feature proportion of continuous components and mutation components of track coordinates, establishing a correlation graph of communication signal characteristics and target type through semantic correlation analysis, and matching radar echo patterns in different scenes using time series alignment algorithm; Dynamic situation prediction model construction: input the processed historical flight data into a hybrid prediction system, the hybrid prediction system includes: Transformer time series prediction unit based on target motion law, used to generate basic situation prediction value, full connection network embedded with airspace attention mechanism, used to correct prediction deviation caused by signal interference, event feature compensator, used to dynamically adjust prediction weight according to real-time collected communication signal characteristics; The direction mutation rate and the speed fluctuation parameter of the track coordinates, the short pulse characteristic and the energy distribution of the radar echo, the frequency hopping gradient and the time interval entropy value of the communication signal are synchronously captured by the heterogeneous data acquisition units deployed on the edge side; The theoretical safety situation value is calculated by inputting the real-time collected data into the dynamic situation prediction model. 3.The low-altitude safety emergency system based on a large model application of claim 2, wherein, In the calculation of the theoretical safety situation value, the following are performed: Adaptive noise reduction processing based on the target flight phase is performed to eliminate monitoring noise caused by environmental meteorological conditions; The correlation features of the track, echo, and signal are integrated by a spatio-temporal feature fusion algorithm. The theoretical safety situation value including a normal scene fluctuation interval is output, and the theoretical safety situation value is dynamically updated according to the target state. 4.The low-altitude safety emergency system based on a large model application of claim 1, wherein, The multi-source data fusion analysis module specifically includes: Time-domain trend deviation calculation: The theoretical safety situation value and the measured situation value are compared in a preset time window, a time series alignment algorithm is used to match the non-synchronous sampled situation sequences, the trend deviation amount in each window is calculated, and a time-domain deviation vector is generated; Space-domain coverage offset detection: The radar echo features of the theoretical value and the measured value are decomposed by a space-domain decomposition algorithm, the coverage area ratio is calculated, the coverage offset index of each type of target is extracted, and a space-domain offset vector is constructed; Event sequence matching degree evaluation: The feature distribution of the theoretical value and the measured event sequence is matched based on a pattern matching algorithm, the time sequence synchronization error of the communication signal mutation point is calculated, the distribution difference degree of the event interval distribution is quantified, and an event matching degree vector is generated; Difference feature matrix generation: The time-domain deviation vector, the space-domain offset vector, and the event matching degree vector are tensor spliced, the dimension difference is eliminated through normalized processing of feature importance weighting, and a three-order difference feature matrix with a dimension of [target number x timestamp x difference type] is output. 5.The low-altitude safety emergency system based on a large model application according to claim 4, wherein, The space-domain coverage offset detection specifically includes: In the space-domain analysis stage, the corresponding radar echo features in the theoretical and measured situation sequences are first extracted, multi-dimensional region analysis is performed on each group of echo signals by using a multi-scale feature extraction, the coverage distribution features in a preset sensitive region interval are extracted, the sensitive region interval selects the activity range covering typical low-altitude targets, after the extraction is completed, the coverage density in the sensitive region interval of the theoretical and measured data is quantitatively calculated, and the offset index of each type of target is extracted based on the relative offset degree of the two, all types of coverage offset results are summarized, and a space-domain offset vector is constructed. 6.The low-altitude safety emergency system based on a large model application of claim 4, wherein, In the event sequence matching degree evaluation, an edit distance algorithm is used in the pattern matching algorithm to match the feature of the mutation point set in the two sequences and identify the time sequence synchronization error. 7.The low-altitude safety emergency system based on a large model application of claim 1, wherein, The risk propagation path deduction module further includes an output including a suspicious target identifier and a risk propagation main path, wherein: The suspicious target identifier binds the space-domain node to the actual target based on the space-domain node connected to the suspicious region set to form a suspicious target identifier set, indicating a potential risk source or an affected terminal. The risk propagation main path is obtained by recording the node path and its order experienced by each round of propagation in the risk diffusion process of Monte Carlo simulation, counting the occurrence frequency of each path in all simulation paths, selecting the path sequence with the highest cumulative occurrence frequency as the risk propagation main path, and outputting the risk propagation main path sequence as a structured and ordered node list reflecting the main propagation trajectory of the risk information in the airspace. 8.The low-altitude safety emergency system based on a large model application of claim 1, wherein, The airspace topology modeling specifically includes: The target position information is subjected to coordinate normalization processing, a three-dimensional topology graph containing node latitude and longitude coordinates and height levels is constructed, the communication delay parameters and signal attenuation coefficients between nodes are labeled, the boundary coordinates and time range of the temporary flight restriction area are superimposed in the topology graph, and a geographical topology model containing a connectivity strength matrix and node influence weight is generated. 9.The low-altitude safety emergency system based on a large model application according to claim 7, wherein, After the risk propagation main path is generated, the following is further included: The suspicious target on the main path is subjected to historical data backtracking, the track features, echo features and communication signal features of the historical flight period of the target are extracted, and a target historical feature archive is constructed; The feature matching degree of the current risk propagation main path and the historical risk event path is compared through a large model similarity calculation algorithm, and a path similarity evaluation result is generated; According to the path similarity evaluation result and the suspicious target identification set, the preset risk residence probability threshold is updated, and the display parameters of the abnormal probability heat distribution map are adjusted.

Images