A heterogeneous unmanned aerial vehicle cooperative operation air safety assessment method

Abstract

The application discloses a kind of heterogeneous unmanned aerial vehicle cooperative operation air safety evaluation method, it is related to unmanned aerial vehicle cooperative operation safety technical field, comprising: obtaining the static parameters and running dynamic parameters of heterogeneous unmanned aerial vehicle;Dynamic and static parameter association matrix based on gradient boosting decision tree algorithm is constructed;Positioning accuracy probability matrix based on bayesian update is constructed;Real-time collection task and running dynamic data;Combined with environmental interference coefficient and task priority weight, the collision time prediction is optimized using LSTM, to calculate the probability of air collision;Trigger hierarchical safety risk alarm.The application realizes the fine quantification evaluation and hierarchical early warning of safety risk in the cooperative scenario of heterogeneous unmanned aerial vehicle, solves the problem of extensive evaluation and insufficient cooperation in the prior art.

Description

Technical Field

[0001] This invention relates to the field of safety technology for collaborative operation of unmanned aerial vehicles (UAVs), and in particular to an aerial safety assessment method for collaborative operation of heterogeneous UAVs. Background Technology

[0002] With the booming development of the low-altitude economy, the drone industry has experienced rapid growth, and its applications have broadly covered numerous fields such as emergency rescue, logistics and transportation, agricultural and forestry plant protection, urban governance, and aerial surveying, becoming an important force in promoting social productivity and optimizing public services. In this process, the operational capabilities of single-type drones are gradually becoming insufficient to meet the demands of complex tasks, and the collaborative operation mode of multi-type heterogeneous drones is becoming increasingly common. Drones of different weights, payloads, and performance parameters complement each other to jointly complete high-precision and high-efficiency tasks. For example, in emergency rescue scenarios, medium and large firefighting drones, with their large payloads and long endurance, undertake the main firefighting operations; similar transport drones are responsible for emergency material delivery and personnel transfer; and micro-light reconnaissance drones, relying on their flexibility and maneuverability, perform tasks such as disaster reconnaissance and target location. The three work together to form a complete rescue operation chain.

[0003] However, when heterogeneous drones operate in collaboration, the air safety risks become increasingly prominent due to the significant differences in flight performance (such as maximum speed, rate of climb, and maneuverability), mission characteristics (such as operating altitude and flight trajectory), and physical attributes (such as airframe size and weight) among the various models.

[0004] Large and medium-sized UAVs have low flexibility in flight trajectory adjustment due to their high inertia and slow turning response. While micro and lightweight UAVs are maneuverable, they have weak anti-interference capabilities and are easily affected by environmental factors such as airflow, causing them to deviate from their paths. When both types of UAVs operate in the same airspace, the timeliness of collision warnings and the coordination of avoidance actions are difficult to guarantee, significantly increasing the probability of mid-air collisions. At the same time, existing air safety management systems are mostly designed for single-type or homogeneous UAV groups, lacking a safety assessment mechanism specifically for the collaborative operation of heterogeneous UAVs. Traditional risk assessment methods cannot comprehensively consider the dynamic interaction relationships and potential conflict points between different UAV types, making it difficult to accurately quantify the safety risk level in collaborative operations.

[0005] This situation not only limits the full realization of the collaborative operation efficiency of heterogeneous drones, but also poses a serious threat to the overall operational safety of low-altitude airspace. Therefore, developing an air safety assessment method that can adapt to the characteristics of collaborative operation of heterogeneous drones has become an urgent need to promote the healthy development of the drone industry and ensure the safe and orderly operation of the low-altitude economy. Summary of the Invention

[0006] To address the issues of crude assessment and insufficient collaboration in existing UAV safety assessments, this invention proposes an aerial safety assessment method for heterogeneous UAV collaborative operations. Through multi-dimensional parameter fusion, dynamic correlation modeling, and differentiated risk management, it achieves refined quantitative assessment and graded early warning of safety risks in heterogeneous UAV collaborative scenarios.

[0007] This application discloses an aerial safety assessment method for heterogeneous unmanned aerial vehicle (UAV) cooperative operation, including the following steps:

[0008] S1. Obtain static and dynamic parameters of various heterogeneous UAVs;

[0009] S2. Based on the static and dynamic parameters, construct a correlation matrix between the dynamic and static parameters. The correlation matrix is ​​obtained by training a gradient boosting decision tree model and is used to quantify the correlation strength between the static and dynamic parameters.

[0010] S3. Construct a positioning accuracy probability matrix. Based on the fusion data of airborne GNSS and inertial measurement unit, establish a probability distribution model for different positioning error ranges, and introduce a Bayesian update mechanism for dynamic correction.

[0011] S4. Real-time collection of task and operational dynamic data, including task list, route planning, real-time positioning data, operational dynamic parameters and environmental data;

[0012] S5. Based on the correlation matrix, the positioning accuracy probability matrix and real-time data, calculate the probability of air collision, including constructing a safe distance threshold, calculating the probability of position overlap and the probability of time overlap, introducing an environmental interference coefficient and a task priority weight, and using a long short-term memory network to optimize the collision time prediction.

[0013] S6. Based on the stated mid-air collision probability, trigger an air safety risk alarm and execute a differentiated warning response strategy based on the risk level.

[0014] Preferably, the static parameters include weight, dimensions, maximum range, maximum load, maximum horizontal acceleration, maximum cruising speed, maximum rate of climb, maximum rate of descent, and motor horsepower;

[0015] The dynamic parameters include payload weight, range, cruising speed, horizontal acceleration, rate of climb, and rate of descent.

[0016] Preferably, the construction of the dynamic and static parameter correlation matrix includes:

[0017] Prepare a training dataset. The input features of the dataset are static parameters, and the output labels are dynamic parameters.

[0018] The model is trained using a federated learning mechanism. Each machine node uses local data to train the initial gradient boosting decision tree model, and the model parameters are uploaded to the central server for aggregation to generate a global model.

[0019] During the training of the gradient boosting decision tree model, physical mechanism constraints are integrated to ensure that the correlation between parameters conforms to physical laws.

[0020] Based on the trained gradient boosting decision tree model, the feature importance score of each static parameter to the dynamic parameter is calculated, normalized to the 0~1 interval as the correlation value, and the correlation direction is marked according to the positive and negative correlation between the parameters to generate a structured correlation matrix.

[0021] An online learning mechanism is introduced to continuously fine-tune the correlation matrix using dynamic parameters measured in real time by airborne sensors to adapt to changes in the UAV's status.

[0022] Preferably, the gradient boosting decision tree model training process includes:

[0023] The first decision tree is constructed using the mean of the dynamic parameters of all training samples as the initial prediction value.

[0024] Multiple rounds of iterative optimization are performed, and a new decision tree is trained in each round using the prediction residuals of the current model as new labels;

[0025] For each static parameter, calculate the information gain for all possible split points, and select the split point with the largest information gain for node partitioning;

[0026] The learning rate is set to 0.1, and training stops when the number of iterations reaches 100 or the sum of squared residuals of the model falls below a preset threshold.

[0027] Preferably, the construction of the positioning accuracy probability matrix specifically includes:

[0028] Based on the fusion data of airborne GNSS and inertial measurement unit, a prior probability distribution model of different positioning error ranges in the vertical and horizontal directions for different aircraft models is established.

[0029] A dynamic correction mechanism based on Bayesian update is introduced to update the probability matrix in real time by combining airborne multi-sensor fusion data.

[0030] Adaptive adjustment is achieved through multi-sensor fusion, GNSS quality indicators are monitored in real time, and inertial measurement unit data-assisted positioning is activated when signal quality deteriorates. Kalman filtering algorithm is used for position estimation optimization, and the probability distribution of positioning accuracy is dynamically adjusted based on the estimation error covariance.

[0031] Preferably, the real-time acquisition of task and dynamic data includes:

[0032] Build and manage dynamic task dependency chains, generate a task list containing unique task identifiers, execution aircraft type, task type, spatiotemporal constraints and priorities, then plan the task duration based on the maximum endurance of each aircraft type and reserve emergency return redundancy, and establish logical dependencies between tasks and task delay propagation models.

[0033] Perform dynamic route planning based on task collaboration, planning an initial route for each UAV including waypoint coordinates, altitude, and dwell time; dynamically adjust the route according to the real-time status of the task dependency chain, and the adjustment strategy includes changing the flight altitude, reducing waypoint dwell time, or triggering extended hovering commands;

[0034] Implement full-dimensional real-time positioning and dynamic parameter acquisition. Obtain real-time latitude and longitude coordinates through a GNSS and inertial measurement unit fusion positioning scheme. The sampling frequency is dynamically configured according to the mission's criticality. Systematically collect dynamic parameters and perform real-time diagnosis of parameter health.

[0035] Multi-source meteorological fusion is carried out by collecting meteorological data through a three-element fusion method of airborne sensors, ground meteorological stations and numerical forecasts, and a three-level fusion decision architecture of airborne sensors, edge computing and ground cloud platform is established.

[0036] Preferably, the calculation of the mid-air collision probability includes:

[0037] Construct horizontal and vertical collision detection distances;

[0038] Calculate the positional overlap probability as the average of the horizontal and vertical overlap probabilities;

[0039] Calculate the probability of time overlap based on relative motion speed and avoidance time window;

[0040] Construct an environmental interference coefficient measurement model;

[0041] Use LSTM networks to predict future velocity changes and optimize the calculation of meeting times;

[0042] Calculate the final collision probability including task priority weights.

[0043] Preferably, the environmental interference coefficient meter model is as follows:

[0044]

[0045] in, denoted as the environmental disturbance coefficient, v as the real-time wind speed, L as the visibility, and T as the turbulence intensity.

[0046] Preferably, the final collision probability is calculated using the following formula:

[0047]

[0048] The task priority weights are dynamically adjusted based on the task dependency chain and execution phase.

[0049] Preferably, the security risk alarm includes:

[0050] Safety risks are classified into four levels: Level 1: no risk; Level 2: low risk; Level 3: medium risk; and Level 4: high risk.

[0051] Warnings are triggered collaboratively based on parameter correlation matrix, positioning accuracy probability, task coordination relationship, and aircraft response characteristics.

[0052] Implement differentiated response strategies, including adjusting data transmission frequency, modifying flight routes, and triggering coordinated avoidance actions.

[0053] The beneficial effects of this invention are:

[0054] (1) The safety risk assessment is upgraded from extensive experience-based judgment to a refined and forward-looking quantitative assessment based on data and physical laws, which effectively improves the fit of the assessment.

[0055] (2) Through the dynamic correction of the positioning probability model and task collaborative decision-making, adaptive and low false alarm collaborative early warning is realized, and the false alarm rate is effectively reduced.

[0056] (3) A complete technical closed loop has been constructed from risk perception, assessment, early warning to model self-optimization, which has lifelong learning capabilities and continuously ensures the safety and efficiency of heterogeneous UAV collaborative operations. Attached Figure Description

[0057] Figure 1 This is a flowchart of the heterogeneous unmanned aerial vehicle (UAV) collaborative operation aerial safety assessment method according to an embodiment of the present invention. Detailed Implementation

[0058] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided with reference to the accompanying drawings and embodiments.

[0059] This application discloses an aerial safety assessment method for heterogeneous unmanned aerial vehicle (UAV) cooperative operation, the process of which is as follows: Figure 1 As shown, it includes the following steps:

[0060] S1. Obtain the static and dynamic parameters of various heterogeneous UAVs. The heterogeneous UAVs used in this embodiment include model A: DJIFlyCart30, model B: WK-1800 fire-fighting UAV, and model C: Zhuoyi D30T UAV.

[0061] The static parameters of a drone include its weight (self-weight), dimensions (length / width / height with arms extended / folded), maximum range, maximum payload, maximum horizontal acceleration, maximum cruise speed, maximum rate of climb, maximum rate of descent, and motor horsepower. The weight (self-weight) is obtained from the manufacturer's technical manual, the dimensions are measured based on the extended / folded state, and the other parameters are extracted through experimental testing or from the manual.

[0062] The weight of the drone can be obtained from the product technical manual, factory test report or official parameter document provided by the drone manufacturer. The following are the weight data of the drones used in this embodiment: Model A: 42.5 kg without battery, 65 kg with dual batteries; Model B: 35.68 kg; Model C: 21 kg.

[0063] The dimensions of the drone can be obtained from the product technical manual, factory inspection report, or official parameter document provided by the drone manufacturer. The following are the dimensions of the drone used in this embodiment:

[0064] Model A: Arm extended, blades folded: Length 1590 mm, Width 1900 mm, Height 947 mm; Arm extended, blades extended: Length 2800 mm, Width 3085 mm, Height 947 mm; Arm folded, blades folded: Length 1115 mm, Width 760 mm, Height 1027 mm.

[0065] Model B: Arm extended: Length 2070 mm, Width 1800 mm, Height 910 mm; Arm folded: Length 1100 mm, Width 950 mm, Height 910 mm.

[0066] Model C: Arm extended: Length 2100 mm, Width 1860 mm, Height 820 mm; Arm folded: Length 1200 mm, Width 1100 mm, Height 820 mm.

[0067] The maximum flight time of a drone can be obtained from the product technical manual, factory test report, or official parameter document provided by the drone manufacturer. The following are the maximum flight time data of the drones used in this embodiment: Model A: Maximum empty range is 28 kilometers, maximum flight time is 31 minutes when empty, maximum fully loaded range is 16 kilometers, maximum flight time is 18 minutes when fully loaded; Model B: Flight time is 57 minutes; Model C: Flight time is 56 minutes when empty, and 14 minutes when carrying a 30 kg payload.

[0068] The maximum payload of a drone can be obtained from the product technical manual, factory test report or official parameter document provided by the drone manufacturer. The following are the maximum payload data of the drones used in this embodiment: Model A: Maximum payload of 40 kg with a single battery and 30 kg with a dual battery; Model B: 65 kg; Model C: 30 kg.

[0069] Horizontal acceleration of drones The resultant external force in the horizontal direction Decision. According to Newton's second law: The net external force in the known horizontal direction Given the weight of the drone, the horizontal acceleration of the drone can be calculated. .in, This refers to the total mass of the drone, i.e., its own weight plus the payload weight. Because the requirement is for the drone's maximum horizontal acceleration, which is typically the performance limit under full load, the [specific value] at this point... It is the weight of the object plus the maximum load, i.e., the full load mass. It is the resultant force in the horizontal direction, that is, the difference between the horizontal thrust provided by the power system and the air resistance.

[0070] Determine the total thrust of the power system The drone's power comes from its motors, and the thrust needs to overcome both gravity during climb and air resistance during horizontal or vertical movement. Therefore, the total thrust can be calculated by reversing the maximum climb rate. The maximum climb rate is v. y (m / s) is the maximum velocity in the vertical direction. The resultant force in the vertical direction needs to overcome gravity and provide the acceleration for climbing, that is: ,in It is the climb angle, and g is taken as 9.8 m / s². 2 For multi-rotor UAVs, the total thrust is the sum of the thrust of each motor, and the horizontal thrust is provided by the horizontal component of the total thrust. That is... ,in, It is the angle between the rotor and the vertical direction; the maximum horizontal thrust corresponds to the maximum tilt angle.

[0071] Calculate horizontal air resistance When flying horizontally, the magnitude of air resistance is related to the speed of the drone, and can be approximated as: ,in, The density of air is approximately 1.225 kg / m³ under standard conditions. This is the horizontal projected area of ​​the drone, which can be calculated using the formula: Calculated; This refers to the drag coefficient, which is approximately 0.8 to 1.2 for multi-rotor drones. It is the horizontal speed, i.e., the maximum cruising speed. .

[0072] Calculate the maximum horizontal acceleration According to the formula for resultant force in the horizontal direction: Substituting the above parameters, we get: The final maximum horizontal acceleration of the UAV in this embodiment is: Model A: 7.2 m / s² 2 Model B: 4.8m / s 2 Model C: 7.5m / s 2 .

[0073] The maximum cruising speed of the drone can be obtained from the product technical manual, factory test report or official parameter document provided by the drone manufacturer. The following are the maximum cruising speed data of the drones used in this embodiment: Model A: 30 kg payload, maximum horizontal flight speed of 20 m / s in a zero-altitude, windless experimental environment; Model B: 10 m / s; Model C: 18 m / s.

[0074] The maximum climb rate of a drone can be obtained from the product technical manual, factory test report, or official parameter document provided by the drone manufacturer. The following are the maximum climb rate data of the drones used in this embodiment: Model A: 30 kg payload, maximum climb rate of 5 m / s in a zero-altitude, windless experimental environment; Model B: 5 m / s; Model C: 4 m / s.

[0075] The maximum descent rate of the drone can be obtained from the product technical manual, factory test report or official parameter document provided by the drone manufacturer. The following are the maximum descent rate data of the drones used in this embodiment: Model A: 30 kg payload, in a zero-altitude, windless experimental environment, the maximum vertical descent speed is 3 m / s, and the maximum tilt descent speed is 5 m / s; Model B: 3 m / s; Model C: 2 m / s.

[0076] The horsepower of a drone's motors can be calculated from the perspective of the drone's power requirements. Total lift in hovering state = Total gravity. , m It is the total mass of the drone. Take 9.8 m / s 2 Hovering power can be calculated using the formula... Received, among which It is the average airflow velocity of the propeller. This refers to the efficiency of the power system. During climbing, additional power is needed to overcome the increase in altitude; the total power is... ,in This is the maximum climb power. For multi-rotor drones, the power of a single motor is... = .

[0077] The final motor horsepower of the drones in this embodiment is as follows: Model A: 4.46 horsepower; Model B: 3.16 horsepower; Model C: 1.92 horsepower.

[0078] The dynamic data of the UAV in actual missions, including weight (including payload), range (measured average), cruise speed (measured average), horizontal acceleration (measured average), climb rate (measured average), and descent rate (measured average), are obtained through actual experimental testing.

[0079] The payload capacity of the drones of models A, B, and C is based on the maximum payload. The full payload data for model A drone with dual batteries is as follows: Model A: 95kg; Model B: 100.68kg; Model C: 51kg.

[0080] Flight range test: The drone was brought to a full charge (100% battery). The flight mode was changed to "maximum cruising speed" for constant flight, and the route was set to a straight round trip to reduce energy consumption during turns. The flight trajectory was recorded in real time using GPS, and the takeoff point and landing point when the battery was depleted were marked. The actual flight distance of the trajectory was measured using map software such as Google Earth or Gaode Maps, or the "cumulative flight distance" data exported from the flight control system was used. For each type of drone, the test was repeated 5 times under the same conditions, and the average value was taken as the flight range of that drone. The following are the flight range data for the three models A, B, and C in this example: Model A: 16km; Model C: 16km; Model C: 15.12km.

[0081] Cruise speed test: Mark two points 1000m apart in an open area, labeled A and B, ensuring there are no obstacles between them; control the drone to fly horizontally and at a constant speed from point A to point B, maintaining a constant altitude throughout the flight and avoiding any climbing or descending; record the time difference (Δt) between passing points A and B using a stopwatch, accurate to 0.1 seconds; speed v = 1000m / Δt. For each type of drone, repeat the test 5 times under the same conditions, and take the average value as the drone's cruise speed. The following are the cruise speed data for the three drone models A, B, and C in this embodiment: Model A: 20m / s; Model B: 10m / s; Model C: 18m / s.

[0082] Horizontal acceleration test: Hover the drone at the starting point of the horizontal test segment, for example, at a height of 10m, keeping the drone stationary; activate maximum power to accelerate forward, and record the speed value per second via the flight control system or GPS, recording the speed change over time; take the time period from 0 to maximum cruising speed, discarding the unstable data of the initial 0.5s acceleration phase, and calculate the acceleration a = (final velocity - initial velocity) / time; for each type of drone, repeat the test 5 times under the same conditions, and take the maximum value as the drone's horizontal acceleration. The following are the horizontal acceleration data for the three drone models A, B, and C in this embodiment: Model A: 7.2m / s 2 Model B: 4.8m / s 2 Model C: 7.5m / s 2 .

[0083] Climb rate test: Hover the drone 10m above the ground; activate the drone's maximum power to ascend vertically, while simultaneously recording the altitude per second using an altimeter; calculate the climb rate over 5 consecutive seconds as the altitude change rate = change in altitude / 5s; repeat the test 5 times for each drone while keeping the conditions constant, and take the average as the drone's climb rate. The following are the climb rate data for the three drone models A, B, and C in this embodiment: Model A: 5m / s; Model B: 5m / s; Model C: 4m / s.

[0084] Descent rate test: Hover the drone at a height of 100m to ensure sufficient descent distance; activate the drone's "maximum descent mode," disable climb power, retain only braking control, and descent vertically at a constant speed, recording the time from 100m to 10m; exclude the initial acceleration and terminal deceleration phases, and calculate the descent rate as 90m / descent time; repeat the test 5 times for each drone while keeping the conditions unchanged, and take the average as the drone's descent rate. The following are the descent rate data for the three drone models A, B, and C in this embodiment: Model A: 3m / s; Model B: 3m / s; Model C: 2m / s.

[0085] S2. Based on static and dynamic parameters, construct a correlation matrix between dynamic and static parameters. The correlation matrix is ​​obtained by training a gradient boosting decision tree model and is used to quantify the correlation strength between static and dynamic parameters.

[0086] The correlation matrix of dynamic and static parameters of UAVs is a core tool for quantifying the degree of mutual influence between the two types of parameters. It intuitively presents the correlation strength between parameters through the values ​​of the matrix elements, where values ​​range from 0 to 1, with "+" indicating positive correlation and "-" indicating negative correlation. This matrix construction breaks through the limitations of traditional empirical assignment, innovatively employing a hybrid modeling method that integrates Gradient Boosting Decision Tree (GBDT) and physical mechanism models to accurately capture the nonlinear correlations between parameters in a data-driven manner. This matrix plays a unique role in the intelligent decision-making core of collaborative safety assessment of heterogeneous UAVs. By quantifying the intrinsic connection between "static constitution" and "dynamic performance," it provides a core basis for forward-looking task allocation and precise risk tracing and early warning, realizing a paradigm shift in safety assessment from passive response to proactive anticipation.

[0087] In the matrix construction process, a comprehensive and reliable training sample set is first required. In this embodiment, the training set covers more than 1,000 sets of UAV test data covering multiple payloads, weather, and mission scenarios. The input features are selected from the above 9 static parameters, including inherent attribute indicators such as weight, size, maximum payload, and motor horsepower. The accuracy of these parameters has been verified through the manufacturer's technical manual and factory inspection report. The output labels correspond to the above 6 dynamic parameters, such as flight range and horizontal acceleration, which are collected in real time by airborne sensors. After recording one set of data every 5 seconds and removing outliers, the data is accurately matched with the static parameters.

[0088] To protect data privacy and achieve cross-machine knowledge sharing, this embodiment introduces a federated learning mechanism. Each machine node uses its own data for initial model training locally, and only uploads model parameters (such as split nodes and information gain) to the central server for aggregation to generate a globally strongly correlated model, ensuring that "data does not leave the domain, and knowledge can be shared." All parameters undergo rigorous normalization before training to eliminate the influence of units, ensuring the stability and accuracy of GBDT model training.

[0089] GBDT model training employs a systematic data-driven approach. First, the mean of the dynamic parameters of all training samples is used as the initial predicted value to construct the first decision tree; at this point, the model error is the initial residual. Subsequently, a multi-round iterative optimization process begins. In each iteration, a new decision tree is trained using the current model's predicted residual as the new label. The information gain is calculated for all possible split points of each static parameter, and the split point with the largest information gain is selected for node partitioning.

[0090] To enhance the physical credibility of the model, physical constraints are integrated during training. For example, when learning the relationship between "weight (including load)" and "horizontal acceleration," the model is forced to follow the general trend of Newton's second law (F=ma), ensuring that the data-driven discovery patterns do not contradict fundamental physical laws. This fusion approach enables the model to discover complex nonlinear patterns such as the "3 horsepower threshold effect" in the "motor horsepower-horizontal acceleration" relationship (below 3 horsepower, acceleration increases by 0.8 m / s² for every 0.5 horsepower increase; beyond that, the growth rate drops to 0.3 m / s²), while also ensuring its engineering applicability.

[0091] To avoid overfitting, a learning rate of 0.1 is set during training, allowing each new tree to progressively optimize the model with small weights. Training stops when the number of iterations reaches 100 or the sum of squared residuals falls below a preset threshold, forming a strong learner consisting of 100 decision trees. After training, based on the frequency of features used as split nodes in the decision trees and the information gain they bring, the feature importance score of each static parameter to the dynamic parameter is calculated, normalized to the 0-1 range as the correlation value, and the correlation direction is labeled according to the positive and negative correlation between parameters, ultimately generating a structured correlation matrix.

[0092] To enable continuous model evolution, this approach incorporates online learning and real-time correction mechanisms. The correlation matrix is ​​not fixed; dynamic parameters are measured in real-time using data fusion from airborne sensors, and a lightweight online learning module continuously fine-tunes the correlation matrix. For example, when a decrease in actual output thrust due to motor aging is detected, the correlation between "motor horsepower" and "climb rate" is automatically reduced, ensuring the model always maintains the optimal mapping to the current aircraft state.

[0093] The deeper value of this correlation matrix lies in its provision of quantitative risk insights and decision support for the collaborative operation of heterogeneous drones, specifically reflected in:

[0094] Optimize collaborative task allocation: During the task planning phase, the correlation matrix is ​​used as a "performance predictor". For example, in a subtask that requires high-speed maneuvering, the model with a high correlation (+0.8) between "motor horsepower" and "horizontal acceleration" will be selected first, rather than the model with a low correlation (+0.4), to ensure that the task capabilities are matched from the source and reduce operational risks.

[0095] Achieving precise, source-based risk warnings: When dynamic parameters are abnormal, the correlation matrix acts as a "fault diagnosis expert system." For example, after detecting a sharp drop in "driving range," the system queries the matrix in real time and finds that it is most strongly correlated with "maximum load" (-0.7). If load over-limit is also detected at the same time, it can accurately trigger a "load overload leading to driving range risk" warning and suggest "reducing load" instead of the inefficient "immediate return to base," improving the targeting of the warning response by 40%.

[0096] Compared to traditional empirical assignment methods, the correlation matrix constructed in this embodiment exhibits significant technical advantages. It not only accurately captures nonlinear correlations, avoiding systematic errors inherent in traditional linear methods, but also ensures data privacy through federated learning, guarantees model interpretability through physical constraints, and achieves lifelong evolution through online learning. Extensive sample validation in this embodiment demonstrates that the correlation degree value achieves a 92% fit with actual parameter changes, more than 30% higher than the 60% fit of traditional empirical assignment. This complete technical system provides reliable, adaptive, and continuously optimizing decision support for the safety assessment of heterogeneous UAV collaborative operations.

[0097] The dynamic and static parameter correlation matrices of the three models based on the GBDT algorithm in this embodiment are shown in Tables 1-3.

[0098] Table 1: Correlation Matrix of Dynamic and Static Parameters of UAV Type A

[0099]

[0100] Table 2: Correlation Matrix of Dynamic and Static Parameters of Model B UAV

[0101]

[0102] Table 3: Correlation Matrix of Dynamic and Static Parameters of Model C UAV

[0103]

[0104] S3. Construct a positioning accuracy probability matrix. Based on the fusion data of airborne GNSS and inertial measurement unit (IMU), establish a probability distribution model for different positioning error ranges, and introduce a Bayesian update mechanism for dynamic correction.

[0105] To address the differences in positioning accuracy during the collaborative operation of heterogeneous UAVs, this embodiment constructs a probability distribution table of vertical / horizontal positioning accuracy for different UAV models. The positioning accuracy probability matrix is ​​based on fused data from airborne GNSS and inertial measurement unit (IMU) data, establishing a probability distribution model for different positioning error ranges. This probability matrix allows for a quantitative assessment of the positioning reliability of each UAV model at different accuracy levels, providing a data foundation for subsequent collision risk analysis.

[0106] To address the issue of fixed probability matrices in traditional positioning accuracy methods, this embodiment introduces a dynamic correction mechanism based on Bayesian updates. This method combines airborne multi-sensor fusion data (GNSS + IMU + visual odometry) and updates the probability matrix in real time using Bayesian formulas. For example, in a specific implementation, when the C-type aircraft enters a densely built-up area and the visual odometry detects GNSS signal drift (error > 3m), the Bayesian update process is automatically triggered. The initial horizontal positioning accuracy (error ≤ 5m) probability P(A) = 0.91 (original matrix value), the conditional probability of the sensor detecting drift event B P(B|A) = 0.2 (low drift probability when error ≤ 5m), and P(B|¬A) = 0.8 (high drift probability when error > 5m) are calculated using Bayesian formulas. After the update, P(A|B) = (0.91 × 0.2) / (0.91 × 0.2 + 0.09 × 0.8) ≈ 0.72, thereby dynamically reducing the confidence level of the positioning accuracy. This mechanism effectively avoids collision misjudgments caused by fixed probabilities and improves the ability to adapt to environmental changes.

[0107] This embodiment further deepens the integration of the positioning accuracy probability matrix and collision probability calculation, achieving adaptive adjustment of accuracy through multi-sensor data fusion. It monitors key GNSS quality indicators such as the number of satellites, signal-to-noise ratio, and DOP value in real time. When a degradation in GNSS signal quality is detected, IMU data-assisted positioning is immediately activated, and a Kalman filter algorithm is used for position estimation optimization. Based on the estimation error covariance analysis results, the positioning accuracy probability distribution is dynamically adjusted to ensure that the probability value always reflects the current actual positioning status. In the collision probability calculation stage, the dynamically adjusted accuracy probability value is used for risk assessment, significantly improving the accuracy and timeliness of collision warnings.

[0108] By deeply integrating the positioning accuracy probability matrix into the collision determination distance calculation, a novel "probability-distance" coupled determination model is constructed. Taking the horizontal collision determination between models A and B as an example, the traditional calculation method uses a fixed error value: Collision determination distance = (1900mm / 2 + 1800mm / 2) + 5m = 6.85m. The improved calculation formula of this scheme is: Collision determination distance = (1900mm / 2 + 1800mm / 2) + ∑(positioning error value × corresponding probability). In the specific calculation process, ∑(positioning error value × corresponding probability) = 0.15×0.1 + 1×0.88 + 5×0.92 + 10×0.98 (weighted horizontal positioning probability of model A) + 0.21×0.1 + 1×0.75 + 5×0.90 + 10×0.95 (weighted horizontal positioning probability of model B). After calculation, the corrected collision determination distance is 7.23m, which is closer to the actual positioning risk characteristics than the original fixed error calculation method. Actual tests show that this probabilistic correction model reduces the false alarm rate by 25%, significantly improving the safety performance of heterogeneous UAV collaborative operation.

[0109] The positioning accuracy probability matrix and collision probability calculation are deeply integrated. When the horizontal distance between two drones approaches the safety threshold, the probability of position overlap is calculated by combining the product of the horizontal positioning accuracy probabilities of both drones, effectively reducing false alarms caused by positioning errors. Simultaneously, the positioning accuracy probability matrix provides a precise quantitative basis for positioning uncertainty for the entire risk assessment system, enabling collision warnings to make more scientific judgments based on actual positioning accuracy. This probability-based collision determination method not only improves reliability but also provides technical support for the safe collaborative operation of heterogeneous drones in complex environments. The vertical / horizontal positioning accuracy probability matrices for different drone models in this embodiment are shown in Tables 4-5.

[0110] Table 4: Probability Matrix of Vertical Positioning Accuracy for Different Models

[0111]

[0112] Table 5: Horizontal Positioning Accuracy Matrix for Different Models

[0113]

[0114] S4. Real-time collection of task and operational dynamic data, including task list, route planning, real-time positioning data, operational dynamic parameters, and environmental data.

[0115] Real-time task and operational dynamic data acquisition is a core data support component for the collaborative safety assessment of heterogeneous UAVs. This embodiment constructs a full-link, high-frequency, and multi-dimensional data acquisition system to systematically acquire task, flight path, status, and environmental data. In particular, through the dynamic parsing of task collaboration relationships (dependency chains and priorities), the task execution status is deeply coupled with real-time risk perception, realizing dynamic closed-loop control of risk assessment and early warning strategies. This provides an accurate, real-time, and context-aware data foundation for subsequent collision probability calculations and safety risk alarms, ensuring that risks are perceptible, decision-making, and controllable during collaborative operations.

[0116] First, we establish a drone task list and manage dynamic dependency chains. Managing the drone task list not only clarifies the collaborative division of labor among heterogeneous drone models, but also defines the logical dependencies between tasks and the dynamic risk transmission paths, serving as the core basis for assessing the rationality of task conflicts and resource allocation.

[0117] Basic mission information and spatiotemporal constraints. Mission identifiers employ a unique coding mechanism (e.g., Task-20230801-001), clearly specifying the aircraft type and number (A-01, B-02, C-03) and classifying the mission type (A-type for material transport, B-type for firefighting, C-type for disaster reconnaissance). Mission time parameters are dynamically redundantly planned based on the maximum endurance of each aircraft type: Aircraft A has a maximum endurance of 18 minutes, a mission duration of 14 minutes, and a 4-minute emergency return period; Aircraft B has a maximum endurance of 57 minutes, a mission duration of 50 minutes, and a 7-minute reserve; Aircraft C has a maximum endurance of 14 minutes, a mission duration of 10 minutes, and a 4-minute reserve. Through precise spatiotemporal constraint modeling, the risk of mission interruption or loss of control due to insufficient endurance estimation is mitigated from the outset.

[0118] Task coordination and dynamic risk assessment. A dependency chain logic of "reconnaissance-firefighting-deployment" is constructed: After aircraft C completes reconnaissance of the target area and marks the coordinates of the fire points, aircraft B performs firefighting operations based on the coordinates, and aircraft A subsequently deploys relief supplies. Task priorities are divided into levels according to urgency (e.g., levels 1-5). In emergency firefighting and rescue scenarios, aircraft A's supply transportation is level 1, aircraft B's firefighting is level 2, and aircraft C's reconnaissance is level 3.

[0119] An innovative dynamic risk assessment and adjustment mechanism is introduced, and a task delay transmission model is established: Risk increase factor = K × t. Where K is the task delay propagation coefficient (K=0.05 obtained by fitting historical data). t represents the delay time. For example, if the reconnaissance mission of aircraft C is delayed by 5 minutes, the collision risk of the firefighting mission of aircraft B increases by 0.25 times, automatically triggering a flight path adjustment for aircraft B (climbing 10m) and a reduction in the flight speed of aircraft A (from 20m / s to 15m / s). If the quality of the preceding mission is not up to standard (e.g., fire point marking error > 10 meters), the safety monitoring level of subsequent missions will be increased.

[0120] A high-real-time data acquisition and synchronization mechanism is implemented. The ground control center is responsible for mission planning, generating an initial mission list, and then distributing it to each UAV via a low-latency, highly reliable encrypted data link. UAVs transmit mission status ("In Progress," "Paused," "Completed," "Abnormal Interruption") in real time via status heartbeat packets, with an update frequency of no less than once per minute. Critical status changes (such as "Completed" or "Abnormal Interruption") are reported immediately using an event-driven mechanism, ensuring that the dynamic adjustment and risk reassessment of the mission chain are delayed by less than one second.

[0121] Secondly, dynamic route planning based on task collaboration is conducted. By combining the mission requirements and performance characteristics of each aircraft type, spatial constraints on flight trajectories are designed to provide a basis for assessing route conflict and deviation risks. However, the route planning is not statically preset but dynamically adjusted and conflict resolved according to the real-time status of task collaboration relationships. For example, in this embodiment:

[0122] The cargo transport route for Aircraft Type A is as follows: Cargo loading point (N30°10′, E120°05′, 30m) — Waypoint 1 (N30°12′, E120°07′, 50m, 0-second stop) — Waypoint 2 (N30°14′, E120°09′, 45m, 5-second stop) — Drop point (N30°15′, E120°10′, 40m). The segment lengths are 2500m, 1800m, and 500m respectively. This route is designed with dynamic altitude adjustment capability: when the mission chain indicates that the firefighting operation of Aircraft Type B is not completed, Aircraft Type A can be automatically instructed to temporarily raise its altitude from 45m to 60m at Waypoint 2 to avoid potential conflicts.

[0123] The firefighting route for Model B is as follows: Fire station (N30°02′, E119°58′, 60m) — Waypoint 1 (N30°03′, E119°59′, 80m, 15-second stop) — Waypoint 2 (N30°04′, E120°00′, 80m, 10-second stop) — Waypoint 3 (N30°03′, E119°59.5′, 70m, 5-second stop) — Return point. The operating altitude (80m) and stop time of this route are strongly correlated with the accuracy of the fire coordinates provided by Model C. When the coordinate accuracy is high (error <5m), the operation proceeds as planned; when the coordinate accuracy is average (error 5-10m), an extended circling command is automatically triggered, sacrificing mission efficiency for higher safety redundancy.

[0124] The disaster reconnaissance flight path for the C-type aircraft is as follows: Command Center (N30°08′, E120°02′, 100m) — Waypoint 1 (N30°09′, E120°03′, 120m, 40-second stop) — Waypoint 2 (N30°07′, E120°04′, 110m, 30-second stop) → Waypoint 3 (N30°08.5′, E120°01′, 100m, 20-second stop) — Return point. The waypoint dwell time is a key parameter that can be dynamically compressed. If an overall delay in the mission chain is detected, the C-type aircraft can be instructed to shorten its dwell time from 40 seconds to 20 seconds, prioritizing the acquisition of critical reconnaissance data to drive subsequent missions.

[0125] Then, a full-dimensional real-time positioning and dynamic parameter acquisition system is constructed, including real-time latitude and longitude coordinates, dynamic parameter system, and real-time health diagnosis.

[0126] Real-time latitude and longitude coordinates employ a deep fusion positioning scheme of GNSS and IMU, with adaptive Kalman filtering to correct instantaneous errors. The sampling frequency is dynamically configured according to mission criticality: A / B models maintain ≥1Hz during normal cruising, automatically increasing to 5Hz when entering conflict zones; C models performing reconnaissance missions maintain ≥5Hz at all times to ensure high-precision positioning requirements.

[0127] The system operates a dynamic parameter system and provides real-time health diagnostics. Flight range: Based on the formula Remaining range = (Current battery level / Full battery level) × Maximum range, updated every 10 seconds; Cruise speed: Sampled at 10Hz by the GNSS module and filtered by the flight control system; Horizontal acceleration: Acquired by the IMU, sampled at 10Hz, and compared in real-time with safety thresholds (7.2m / s² for Model A, 4.8m / s² for Model B); Climb / descent rate: Fusion of barometric altimeter and GNSS data, sampled at 5Hz, monitoring the risk of rapid ascent (over 5m / s) and rapid descent (over 3m / s).

[0128] Next, multi-source meteorological fusion and environmental interference quantification are carried out, including meteorological data collection and the establishment of environmental interference coefficient models.

[0129] Meteorological data acquisition employs a three-element fusion approach: airborne sensors, ground-based meteorological stations, and numerical weather prediction. Airborne temperature and humidity sensors and anemometers sample micro-environmental data at a frequency of 1 Hz; ground-based meteorological stations provide macro-level trends; and meteorological forecast data is used for forward-looking risk prediction. These three elements are fused and cross-validated through the Ground Control System (GCS) to deliver environmental early warnings.

[0130] Meteorological data, including real-time wind speed, visibility, and turbulence intensity, are collected using a ternary fusion method combining airborne sensors, ground meteorological stations, and numerical weather prediction. This provides a data foundation for environmental corrections in subsequent collision probability calculations.

[0131] Finally, a multi-level data fusion and task collaborative decision-making architecture is established. This architecture comprises three levels: airborne sensors (airborne terminal), edge computing (edge ​​terminal), and ground cloud platform (cloud terminal), ensuring real-time data processing and globally optimal decision-making.

[0132] Airborne: GNSS (10Hz), IMU (100Hz), and vision sensor (30fps) multi-source acquisition, with the airborne processor performing preliminary filtering, alignment, and timestamp synchronization.

[0133] At the edge: The onboard computing module (computing power ≥ 2 TOPS) generates a parameter health score (0-100 points) every 0.1 seconds. When the score is below 60 points, an early warning is immediately triggered, and a preliminary diagnosis result is reported (e.g., "IMU data oscillations have intensified, suspected to be due to a sudden gust of wind").

[0134] In the cloud: The ground control center integrates all UAV data, environmental data, and real-time mission status every second, updating the global risk map. When an anomaly occurs in the mission dependency chain (such as reconnaissance delay), the cloud platform performs collaborative replanning based on the global situation, such as assigning backup reconnaissance areas to UAV type B or adjusting the delivery sequence of UAV type A, mitigating cascading risks at the system level.

[0135] This data acquisition system deeply couples dynamic mission evaluation, precise route planning, real-time parameter monitoring, environmental interference quantification, and collaborative decision-making, building a complete, real-time, and intelligent data foundation for the collaborative operation of heterogeneous UAVs, and significantly improving the ability to perceive risks and proactively control complex mission scenarios.

[0136] S5. Based on the correlation matrix, positioning accuracy probability matrix and real-time data, calculate the probability of air collision, including constructing a safe distance threshold, calculating the probability of position overlap and the probability of time overlap, introducing environmental interference coefficient and task priority weight, and using a long short-term memory network (LSTM) to optimize the collision time prediction.

[0137] In-flight collision probability calculation is a core component of safety assessment for the collaborative operation of heterogeneous drones. It requires comprehensive consideration of the drone's physical size, precise positioning parameters, dynamic motion parameters, and task priority. By constructing a safe distance threshold, calculating the probability of positional overlap and temporal overlap, and combining environmental interference coefficients and task priority weights, the collision risk can be accurately quantified, providing a basis for subsequent safety alarms.

[0138] Establish safe distance thresholds. Based on the physical size and positioning parameters of the UAV, establish "safe distance thresholds" (i.e., collision detection distances) in the horizontal and vertical directions respectively, as the basic standard for judging whether there is a potential collision risk between the two aircraft.

[0139] The formula for calculating the horizontal collision judgment distance is: Horizontal collision judgment distance = (Horizontal half-width of model 1 + Horizontal half-width of model 2) + Maximum horizontal positioning error (5m). Based on the horizontal dimensions of each model (taken from static parameters), the specific calculation results are as follows: Model A (horizontal half-width 950mm, 1900mm / 2) and Model B (horizontal half-width 900mm, 1800mm / 2): (0.95 + 0.9) + 5 = 6.85m. Model A and Model C (horizontal half-width 930mm, 1860mm / 2): (0.95 + 0.93) + 5 = 6.88m. Model B and Model C: (0.9 + 0.93) + 5 = 6.83m.

[0140] The formula for calculating the vertical collision judgment distance is: Vertical collision judgment distance = (Vertical height of model 1 / 2 + Vertical height of model 2 / 2) + Maximum vertical positioning error (5m). Combining the vertical heights of each model (taken from static parameters), the specific calculation results are as follows: Model A (vertical height 947mm, half-height 473.5mm) and Model B (vertical height 910mm, half-height 455mm): (0.4735 + 0.455) + 5 = 5.9285m. Model A and Model C (vertical height 820mm, half-height 410mm): (0.4735 + 0.41) + 5 = 5.8835m. Model B and Model C: (0.455 + 0.41) + 5 = 5.865m.

[0141] When the real-time horizontal distance between the two aircraft is less than or equal to the horizontal collision detection distance, and the real-time vertical distance is less than or equal to the vertical collision detection distance, it is judged as a "potential collision risk" and the collision probability needs to be further calculated.

[0142] Calculate the probability of position overlap. The probability of position overlap is calculated based on the positioning accuracy probability matrix, reflecting the likelihood that the actual positions of the two aircraft will fall into the "collision determination space", and is divided into two dimensions: horizontal and vertical.

[0143] Horizontal overlap probability. The product of the horizontal positioning accuracy probabilities of the two aircraft within the "horizontal collision detection distance" is taken. Specific data are as follows: Aircraft A (probability 0.92 when horizontal accuracy ≤ 6.85m) vs. Aircraft B (probability 0.90 when horizontal accuracy ≤ 6.85m): 0.92 × 0.90 = 0.828. Aircraft A (probability 0.92 when horizontal accuracy ≤ 6.88m) vs. Aircraft C (probability 0.90 when horizontal accuracy ≤ 6.88m): 0.92 × 0.90 = 0.828. Aircraft B (probability 0.90 when horizontal accuracy ≤ 6.83m) vs. Aircraft C (probability 0.90 when horizontal accuracy ≤ 6.83m): 0.90 × 0.90 = 0.810.

[0144] Vertical overlap probability. The product of the vertical positioning accuracy probabilities of the two devices within the "vertical collision detection distance" is taken. Specific data are as follows: Device A (probability 0.85 when vertical accuracy ≤ 5.9285m) vs. Device B (probability 0.80 when vertical accuracy ≤ 5.9285m): 0.85 × 0.80 = 0.68. Device A (probability 0.85 when vertical accuracy ≤ 5.8835m) vs. Device C (probability 0.83 when vertical accuracy ≤ 5.8835m): 0.85 × 0.83 = 0.7056. Device B (probability 0.80 when vertical accuracy ≤ 5.865m) vs. Device C (probability 0.83 when vertical accuracy ≤ 5.865m): 0.80 × 0.83 = 0.664.

[0145] Overall positional overlap probability. The overall positional overlap probability is the average of the horizontal and vertical overlap probabilities (because horizontal and vertical collision risks are equally important). For example, the overall positional overlap probability between aircraft A and B is (0.828 + 0.68) / 2 = 0.754, the overall positional overlap probability between aircraft A and C is (0.828 + 0.7056) / 2 = 0.7668, and the overall positional overlap probability between aircraft C and B is (0.810 + 0.664) / 2 = 0.737.

[0146] Calculate the probability of time overlap. Combining the dynamic parameters of the UAV (speed, acceleration), calculate the time difference between the two UAVs entering the collision determination space and the avoidance time window to determine whether there is time overlap, and then determine the probability of time overlap (0 or 1).

[0147] Relative velocity and time of encounter. The relative velocity of the two aircraft is calculated using the difference in their horizontal velocity vectors, using the following formula:

[0148]

[0149] in, This represents the horizontal velocity component of the first aircraft. This represents the horizontal component of the velocity of the second aircraft. This represents the vertical component of the velocity of the first aircraft. This represents the vertical velocity component of the second aircraft.

[0150] Meeting time determination rule: If the initial distance between the two aircraft is less than or equal to the collision determination distance, the meeting time is 0 (determined as an immediate collision risk); otherwise, the meeting time = initial distance ÷ relative speed.

[0151] Avoidance time window. The avoidance time window is the shortest time required for two aircraft to complete the minimum avoidance maneuver. It is calculated based on horizontal acceleration, and the formula is:

[0152]

[0153] The horizontal acceleration is taken from dynamic parameters. Substituting the data, we can see that the avoidance time window between aircraft type A and aircraft type B is 1.51 seconds, the avoidance time window between aircraft type A and aircraft type C is 1.47 seconds, and the avoidance time window between aircraft type B and aircraft type C is 1.63 seconds.

[0154] Time overlap probability determination. If the time difference between the two machines entering the collision determination space is less than or equal to the avoidance time window, it is determined as "unable to avoid effectively", and the time overlap probability is recorded as 1; if the time difference is greater than the avoidance time window, it is determined as "able to avoid effectively", and the time overlap probability is recorded as 0.

[0155] An environmental disturbance coefficient calculation model is constructed. The environmental disturbance coefficient is based on a composite model of wind speed, visibility, and turbulence intensity, quantifying the impact of meteorological conditions on collision probability. The formula is as follows:

[0156]

[0157] in, , where v is the environmental interference coefficient, v is the real-time wind speed (m / s, value 0-20), L is the visibility (m, value 50-10000), and T is the turbulence intensity (dimensionless, 0-1, calculated by detecting air pressure fluctuations from an airborne barometer).

[0158] When the wind speed is 10 m / s, the visibility is 500 m, and the turbulence intensity is 0.3, substituting these values ​​into the formula yields the following result. The value is 0.721. This coefficient directly corrects for the probability of position and time overlap, making the collision probability more consistent with the actual environmental impact.

[0159] Collision time prediction optimization based on LSTM. Traditional encounter time calculation relies solely on the current speed, which is prone to errors due to dynamic speed changes. This embodiment introduces a Long Short-Term Memory (LSTM) network to predict the drone's speed change within the next 5 seconds, thus optimizing the encounter time calculation. Specifically, the model is trained using the speed sequence of the past 10 seconds as input and the predicted speed value for the next 5 seconds as output, with over 1000 sets of historical dynamic parameter samples used for training.

[0160] When calculating the relative motion between aircraft A and C, if LSTM predicts that aircraft A's speed will decrease from 20 m / s to 15 m / s in the next 2 seconds, while aircraft C's speed remains at 18 m / s, the traditional encounter time (based on the current speed of 20 m / s) is... The corrected meeting time (based on the predicted average velocity of 17.5 m / s) is: The extra seconds allow for more time to make an avoidance decision, increasing the success rate of avoidance by 35%.

[0161] When calculating the relative motion of aircraft A and B, LSTM predicts that aircraft A's speed will increase from 18 m / s to 22 m / s in the next 3 seconds, while aircraft B's speed will decrease from 16 m / s to 14 m / s. The traditional encounter time (based on the current speeds of aircraft A (18 m / s) and aircraft B (16 m / s) is... The corrected meeting time (based on the average predicted speed of aircraft type A of 20 m / s and the average predicted speed of aircraft type B of 15 m / s) is: The second can more promptly reflect the impact of speed changes on the encounter time, assisting in planning avoidance actions earlier or later.

[0162] In a scenario where aircraft B and C are in relative motion, LSTM predicts that aircraft B's speed will fluctuate from 15 m / s to 17 m / s and then back to 15 m / s over the next 4 seconds, while aircraft C's speed will steadily decrease from 19 m / s to 17 m / s. The traditional encounter time (based on the current speeds of aircraft B (15 m / s) and aircraft C (19 m / s)) is... The corrected meeting time (based on the average predicted speed of aircraft B (16 m / s) and aircraft C (18 m / s) is: The second measurement makes the encounter time calculation more closely match the dynamic changes in speed, improving the accuracy of avoidance decisions.

[0163] Calculation of the final collision probability including task priority weights. The final collision probability needs to incorporate task priority weights to quantify the impact of task importance on collision risk assessment. The formula is optimized as follows:

[0164]

[0165] Task priority weights are dynamically adjusted based on task dependency chains and execution phases, with the specific rules as follows:

[0166] High-priority tasks (such as material transportation for aircraft type A, level 1): The weight of the critical phase (such as 5 minutes before material delivery) is 1.2, and the weight of the normal phase (such as the cruise phase) is 1.0.

[0167] Low-priority tasks (such as C-type reconnaissance, level 3): When resource conflicts occur with high-priority tasks (such as airspace overlap), the weight is 0.8, reducing the impact on the collision probability of high-priority tasks.

[0168] Medium priority tasks (such as fire suppression for aircraft type B, level 2): ​​The default weight is 1.0, which is increased to 1.1 only during critical fire suppression phases (such as spraying the core fire point).

[0169] S6. Based on the collision probability, trigger an air safety risk alarm and execute a differentiated warning response strategy based on the risk level.

[0170] By leveraging the synergistic effects of parameter correlation matrix, positioning accuracy probability, task collaboration relationships, and aircraft response characteristics, a collaborative multi-dimensional risk assessment method is proposed. The parameter correlation matrix provides early warning of performance anomalies, the positioning accuracy probability matrix quantifies position uncertainty, the task collaboration relationship determines conflict priority, and the aircraft response characteristics provide differentiated obstacle avoidance strategies. Based on the UAV's operating parameters and in-flight collision probability, this embodiment classifies in-flight safety risks into four levels as shown in Table 6:

[0171] Table 6: Classification of Air Safety Risk Levels

[0172]

[0173] The warning triggering mechanism is based on the feedback data from "real-time task and dynamic runtime data collection," and dynamically triggers alarms through collaborative analysis:

[0174] Parameter correlation matrix monitoring: Real-time monitoring of abnormal correlations between static and dynamic parameters, such as the correlation between motor horsepower and horizontal acceleration deviating from the threshold.

[0175] Positioning accuracy assessment: By combining the positioning accuracy probability matrix, the impact of position uncertainty on collision risk is assessed.

[0176] Task dependency analysis: Based on task collaboration relationships, identify conflict risks caused by dependency chain anomalies.

[0177] Environmental coefficient correction: Quantify the amplification effect of meteorological conditions on risk level through environmental disturbance coefficient.

[0178] Based on the synergy of "parameter correlation matrix + positioning accuracy + task dependence + environmental coefficient", a differentiated response strategy as shown in Table 7 is formulated:

[0179] Table 7: Differentiated Response Strategies

[0180]

[0181] Warning information output and recording. The ground control center displays the risk level, the aircraft model involved, the triggering cause, and attaches a "risk source tracing label," including:

[0182] Parameter correlation anomalies (e.g., "abnormal correlation between motor horsepower and horizontal acceleration").

[0183] Positioning accuracy is affected (e.g., "deterioration of GNSS signal quality leads to increased positioning error").

[0184] Task-dependent issues (e.g., "delayed reconnaissance missions hinder firefighting operations").

[0185] Environmental disturbance factors (such as "instability caused by excessive turbulence intensity").

[0186] The drone's onboard indicator lights flash (red / yellow / green correspond to high / medium / low risk), automatically saving the alarm time, parameters of the involved drone model, meteorological data, and response measures, with the storage period extended to 90 days.

[0187] Alarm handling and optimization closed loop. Establish a complete closed-loop management system for "alarm-handling-review":

[0188] Real-time collaborative response: Based on the analysis results of parameter correlation matrix, positioning accuracy probability, task collaboration relationship and aircraft response characteristics, the corresponding collaborative response strategy is initiated.

[0189] Monthly in-depth analysis: Cluster analysis is performed on historical alarm data every month to identify high-frequency risk scenarios and module correlations.

[0190] Model parameter optimization: Optimize parameters based on analysis results, such as adjusting the correlation weights in the parameter correlation matrix.

[0191] Continuous strategy improvement: Through closed-loop feedback, continuously improve the coordination mechanism of parameter correlation matrix, positioning accuracy probability, task coordination relationship and aircraft response characteristics to enhance the overall risk assessment accuracy.

[0192] Multi-dimensional risk tracing. Based on collaborative analysis, multi-dimensional risk tracing reports are automatically generated:

[0193] Parameter correlation tracing: Identifying the root cause of abnormal static and dynamic parameter correlations.

[0194] Positioning accuracy assessment: Analyze the specific impact of positioning errors on collision probability.

[0195] Task chain diagnosis: Locating weak links in task dependencies.

[0196] Environmental factor quantification: accurately assess the contribution of each environmental element to the risk level.

[0197] By deeply integrating parameter correlation matrix, positioning accuracy probability, task collaboration relationship, and aircraft response characteristics, the proposed solution in this embodiment achieves a complete safety management and control system from risk perception, accurate assessment, collaborative early warning to optimization and improvement, providing comprehensive and multi-level reliable protection for the collaborative operation of heterogeneous UAVs.

[0198] This embodiment achieves holistic and adaptive optimization of safety assessment for heterogeneous UAV collaborative operations through deep collaboration of parameter correlation matrices, positioning accuracy probabilities, task coordination relationships, and UAV model response characteristics. These four parts do not operate independently but form an organic whole through closed-loop interaction, significantly improving risk perception, decision-making accuracy, and response efficiency. The collaborative workflow between the four parts is described in detail below:

[0199] The parameter correlation matrix provides a performance benchmark for positioning accuracy assessment. By quantifying the nonlinear relationship between static and dynamic parameters, the parameter correlation matrix provides a dynamic performance benchmark for positioning accuracy assessment. During collaborative operation, when an abnormal increase in the positioning error of a certain aircraft model is detected, the parameter correlation matrix is ​​queried in real time to analyze whether key static parameters related to the positioning accuracy of that aircraft model (such as motor horsepower, weight, and dimensions) have experienced performance degradation. For example, if the parameter correlation matrix shows that the correlation between "motor horsepower" and "horizontal acceleration" decreases from the benchmark value of +0.8 to +0.5, it indicates that the power system output may be attenuated due to aging or failure, thereby affecting positioning stability (such as increased IMU data oscillation). Based on this, a Bayesian update mechanism for the positioning accuracy probability matrix is ​​triggered to dynamically correct the positioning error probability distribution (such as lowering the probability of horizontal positioning accuracy ≤5m from 0.92 to 0.85), making the positioning assessment more consistent with the current aircraft status. At the same time, the "size-cruising speed" correlation (-0.5) in the parameter correlation matrix can also be used to predict the impact of airflow disturbances on positioning, providing prior knowledge for accuracy assessment.

[0200] Positioning accuracy probability influences collaborative task decision-making. The positioning accuracy probability matrix quantifies the positional uncertainty of each aircraft type, directly driving dynamic adjustments to collaborative task decisions. During task execution, the positioning accuracy probability of each aircraft type is calculated in real time and coupled with the task dependency chain for analysis. For example, when aircraft type C (reconnaissance) enters a densely built-up area, causing its horizontal positioning accuracy probability to drop from 0.90 to 0.75, the operation initiation of aircraft type B (firefighting) will be automatically delayed based on the task dependency relationship ("reconnaissance-firefighting-delivery" chain), and aircraft type C will be instructed to extend its waypoint dwell time to improve positioning reliability. Simultaneously, task priority weights are dynamically allocated based on positioning accuracy probability: if aircraft type A (material transport) is in a critical delivery phase and has high positioning accuracy (probability ≥ 0.90), its task weight remains at 1.2; if positioning accuracy decreases (probability ≤ 0.80), the weight drops to 1.0, potentially triggering a backup delivery plan. This positioning accuracy-based task decision-making mechanism ensures the robustness of collaborative tasks in uncertain environments.

[0201] Task dependencies guide aircraft response strategy selection. Task dependencies define the logical order, priority, and dynamic risk propagation path between tasks, directly guiding aircraft response strategy selection when safety risk warnings are issued. When a collision risk is detected, the root cause of the conflict is analyzed based on task dependencies, and differentiated response strategies are formulated. For example, in the "reconnaissance-firefighting-delivery" chain, if aircraft type B (firefighting) overlaps with aircraft type A (supply transport) due to a delay in advance reconnaissance, the higher-priority task (aircraft type A weight 1.2) will be prioritized, instructing aircraft type B to suspend operations and climb 20m to avoid collision, rather than forcing aircraft type A to change course. Simultaneously, the dynamic risk assessment model within the task dependencies (risk amplification factor = K × Δt) quantifies response urgency: if aircraft type C's reconnaissance is delayed by 5 minutes, the response time window for aircraft type B's avoidance action is compressed by 1.25 times, driving it to adopt more aggressive actions (such as maximum descent rate evacuation). This strategy ensures that response actions align with the overall mission objectives while minimizing cascading risks.

[0202] Aircraft response data is used to optimize the parameter correlation model. Aircraft response data (including the effectiveness of avoidance maneuvers and real-time changes in dynamic parameters) is continuously fed back to optimize the parameter correlation model through an online learning mechanism. For example, when aircraft A triggers a high-risk alarm and performs avoidance maneuvers with maximum horizontal acceleration (7.2 m / s²), the actual acceleration value, motor horsepower output, and environmental interference coefficient are recorded and compared with the predicted values ​​in the parameter correlation matrix. If the actual acceleration is consistently lower than the predicted value (e.g., only reaching 6.5 m / s²), it indicates that the correlation between "motor horsepower and horizontal acceleration" (+0.8) may be overestimated. This correlation is dynamically lowered to +0.7 using a lightweight online learning module, and other relevant parameters in the correlation matrix (such as the correlation between "weight and horizontal acceleration") are adjusted. Simultaneously, the response data is also used to optimize the physical constraint model: for example, when aircraft B avoids obstacles in a turbulent environment, if the actual climb rate is lower than the prediction, the efficiency factor η in the "motor horsepower-climb rate" correlation is corrected to make the model more closely reflect actual weather conditions. This feedback loop ensures the lifelong evolution of the parameter correlation matrix, improving the accuracy and adaptability of the overall security assessment.

[0203] Through the above collaborative work, this embodiment realizes a complete closed loop from performance benchmark provision, decision influence, strategy guidance to model optimization, and deeply integrates parameter correlation, positioning accuracy, task collaboration and aircraft response to build an adaptive and evolvable heterogeneous UAV collaborative safety assessment system.

[0204] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A heterogeneous UAV cooperative operation air safety assessment method, characterized in that, The method comprises the following steps: S1, obtaining static parameters and dynamic parameters of multiple heterogeneous unmanned aerial vehicles; S2, based on the static parameters and dynamic parameters, a dynamic and static parameter correlation matrix is constructed, the correlation matrix is obtained by gradient boosting decision tree model training, and is used to quantify the correlation strength between the static parameters and the dynamic parameters; S3, a positioning accuracy probability matrix is constructed, a probability distribution model of different positioning error ranges is established based on the fusion data of the airborne GNSS and the inertial measurement unit, and a Bayesian updating mechanism is introduced for dynamic correction; S4, real-time collection of task and running dynamic data, including task list, flight path planning, real-time positioning data, running dynamic parameters and environment data; S5, based on the correlation matrix, the positioning accuracy probability matrix and the real-time data, the air collision probability is calculated, including constructing a safety distance threshold, calculating the position overlap probability and the time overlap probability, introducing an environmental interference coefficient and a task priority weight, and using a long short-term memory network to optimize the collision time prediction; S6, according to the air collision probability, triggering the air safety risk alarm, and executing a differentiated warning response strategy based on the risk level.

2. The heterogeneous UAV cooperative operation air safety assessment method of claim 1, wherein, The static parameters include self-weight, size, maximum endurance, maximum load, maximum horizontal acceleration, maximum cruising speed, maximum climb rate, maximum descent rate and motor horsepower; The dynamic parameters include load weight, endurance mileage, cruising speed, horizontal acceleration, climb rate and descent rate.

3. The heterogeneous UAV cooperative operation air safety assessment method of claim 2, wherein, The construction of the dynamic and static parameter correlation matrix comprises: Prepare a training data set, the input features of the data set are static parameters, and the output labels are dynamic parameters; Use a federated learning mechanism to train the model, each model node uses local data to preliminarily train a gradient boosting decision tree model, and uploads the model parameters to a central server for aggregation to generate a global model; In the gradient boosting decision tree model training process, physical mechanism constraints are integrated to ensure that the correlation between parameters meets the physical laws; Based on the trained gradient boosting decision tree model, the feature importance score of each static parameter to the dynamic parameter is calculated, which is normalized to the interval 0~1 as the correlation degree value, and the correlation direction is labeled according to the positive and negative correlation between parameters to generate a structured correlation matrix; An online learning mechanism is introduced to continuously fine-tune the correlation matrix based on the dynamic parameters measured by the onboard sensors in real time to adapt to the changes in the state of the unmanned aerial vehicle.

4. The heterogeneous UAV cooperative operation air safety assessment method of claim 3, wherein, The gradient boosting decision tree model training process comprises: Taking the average of the dynamic parameters of all training samples as the initial prediction value, a first decision tree is constructed; Multiple rounds of iterative optimization are performed, and in each round of iteration, the prediction residual of the current model is used as a new label to train a new decision tree; Calculate the information gain of all possible split points for each static parameter, and select the split point with the maximum information gain for node division; The learning rate is set to 0.1, and the training is stopped when the number of iterations reaches 100 or the model residual sum of squares is lower than a preset threshold.

5. The heterogeneous UAV cooperative operation air safety assessment method of claim 4, wherein, The construction of the positioning accuracy probability matrix specifically comprises: Based on the fusion data of the airborne GNSS and the inertial measurement unit, a prior probability distribution model of different positioning error ranges in the vertical and horizontal directions of different models is established; A dynamic correction mechanism based on Bayesian update is introduced to update the probability matrix in real time combined with multi-sensor fusion data on board; Adaptive adjustment is achieved through multi-sensor fusion, real-time monitoring of GNSS quality indicators, starting of inertial measurement unit data aided positioning when signal quality decreases, and position estimation optimization using Kalman filtering algorithm, and dynamic adjustment of positioning accuracy probability distribution based on estimation error covariance.

6. The heterogeneous UAV cooperative operation air safety assessment method of claim 5, wherein, The real-time acquisition task and running dynamic data include: A dynamic task dependency chain is constructed and managed, a task list containing unique task identification, execution model, task type, space-time constraints and priority is generated, then the task duration is planned based on the maximum endurance of each model and an emergency return redundancy is reserved, and a logical dependency relationship between tasks and a task delay propagation model are established; Dynamic route planning based on task coordination is performed, an initial route containing waypoint coordinates, height and dwell time is planned for each UAV; the route is dynamically adjusted according to the real-time state of the task dependency chain, the adjustment strategy includes changing the flight height, compressing the waypoint dwell time or triggering the extended circling instruction; Full-dimensional real-time positioning and dynamic parameter acquisition are implemented, real-time latitude and longitude coordinates are obtained through GNSS and inertial measurement unit fusion positioning scheme, the sampling frequency is dynamically configured according to the task criticality; dynamic parameters are systematically collected, and the parameter health is diagnosed in real time; Multi-source weather fusion is performed, weather data is collected by using airborne sensors, ground weather stations and numerical prediction ternary fusion mode, and a three-level fusion decision architecture of airborne sensors, edge computing and ground cloud platform is established.

7. The heterogeneous UAV cooperative operation air safety assessment method of claim 6, wherein, The calculation of the air collision probability includes: The horizontal collision judgment distance and the vertical collision judgment distance are constructed; The position overlap probability is calculated as the average of the horizontal overlap probability and the vertical overlap probability; The time overlap probability is calculated based on the relative motion speed and the avoidance time window; An environmental interference coefficient model is constructed; The future speed change is predicted using the LSTM network to optimize the meeting time calculation; The final collision probability with task priority weight is calculated.

8. The heterogeneous UAV cooperative operation air safety assessment method of claim 7, wherein, The final collision probability is calculated by the following formula: The task priority weight is dynamically adjusted according to the task dependency chain and the execution stage.

9. The heterogeneous UAV cooperative operation air safety assessment method of claim 8, wherein, The safety risk warning includes: The safety risk is divided into four levels: level one is no risk, level two is low risk, level three is medium risk and level four is high risk; Based on the parameter correlation matrix, positioning accuracy probability, task coordination relationship and model response characteristics, the warning is triggered cooperatively; Differential response strategies are implemented, including adjusting the data return frequency, modifying the route, and triggering cooperative avoidance actions.

Images