Alarm method and system for air collision risk of low-altitude unmanned aerial vehicle

Abstract

The invention belongs to the technical field of unmanned aerial vehicle safety, and particularly relates to a low-altitude unmanned aerial vehicle air collision risk alarm method and system. The method comprises the following steps: predicting a future trajectory of an unmanned aerial vehicle by fusing an LSTM neural network of an Adam optimizer; aiming at two different scenes of cooperation and non-cooperation, a dynamic Reich protection area model and a Monte Carlo reachable domain model are respectively adopted to carry out collision risk identification; calculating an overlapping degree and detecting intrusive collision; and outputting a graded alarm signal according to the prediction time window and the risk severity. The problems of high false alarm rate and short early warning time in the prior art are effectively solved, accurate, advanced and differentiated early warning of the collision risk of the low-altitude unmanned aerial vehicle can be realized, and the safety management level of a low airspace is remarkably improved. The method is suitable for various unmanned aerial vehicle monitoring platforms and low-altitude traffic management systems.

Description

Technical Field

[0001] This invention belongs to the field of low-altitude safety control technology for unmanned aerial vehicles (UAVs), specifically relating to a method and system for mid-air collision risk alarm for low-altitude UAVs. Background Technology

[0002] With the rapid development of the low-altitude economy, the drone industry is also growing at an extremely fast pace. Low-altitude drones have shown great potential in areas such as logistics delivery and emergency rescue. However, due to the surge in the number of drones and limited airspace restrictions, the risk of mid-air collisions has become a key factor restricting the industry's development. The goal of future air traffic is to minimize the probability of drone traffic collisions by using observation and prediction to prevent drone collision accidents in advance.

[0003] Currently, the flight paths of low-altitude drones are usually pre-planned. Although there may be some deviations due to the pursuit of flexibility and dynamic adjustments, the overall flight path of low-altitude drones is a conduit route. A conduit route is a three-dimensional channel planned for drones to fly in urban low-altitude areas or specific regions. This concept aims to isolate drone flight activities within a specific airspace, similar to establishing an "aerial elevated road network" for drones, to ensure their safe, orderly, and efficient operation. Its core objective is to achieve physical or virtual isolation, minimizing the risk of collisions between drones and manned aircraft, buildings, ground personnel, and other drones.

[0004] Safety Target Level (TLS) is a core concept in low-altitude UAV collision risk identification and airspace management. It provides a quantitative benchmark for acceptable risk thresholds and is the ultimate constraint that all risk assessment methods must adhere to. The necessity of TLS lies primarily in providing clear acceptance criteria for collision risk models. The operation of low-altitude UAVs requires maintaining a Safety Target Level, but increasing airspace utilization may increase collision risk, causing the target safety level to fall below acceptable limits. Therefore, finding a method that balances target safety levels with improved airspace utilization is currently a key issue. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a method and system for low-altitude UAV aerial collision risk alarm that distinguishes between cooperative and cooperative operation scenarios.

[0006] The first aspect of the present invention provides a method for mid-air collision risk warning for low-altitude unmanned aerial vehicles, comprising: Construct cooperative and non-cooperative drone operation scenarios; the cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route; the non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. Predict trajectories and construct protected areas; in cooperative drone operation scenarios, historical trajectory data of cooperative drones is obtained and input into the prediction model to output the predicted trajectory, while a dynamic Reich model protected area is constructed; in non-cooperative drone operation scenarios, historical flight trajectory data of both types of drones are obtained, the trajectory of cooperative drones is predicted, and an reachability protected area is constructed for non-cooperative drones. Collisions are determined based on predicted trajectories; specifically, the spherical cap overlap method is used to determine collisions in cooperative drone operation scenarios, while the intrusion method is used to determine collisions in non-cooperative drone operation scenarios. An alarm will be issued based on the judgment result.

[0007] A second aspect of the present invention provides a low-altitude unmanned aerial vehicle (UAV) mid-air collision risk warning system, comprising: The scenario building module is used to construct cooperative and non-cooperative drone operation scenarios. The cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route. The non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. The trajectory prediction module is used to predict the trajectory of cooperative drones; The protected area construction module is used to construct dynamic Reich model protected areas in cooperative drone operation scenarios, and to construct reachable domain protected areas for non-cooperative drones. The collision detection module is used to determine collisions based on the predicted trajectory. Specifically, it uses the spherical cap overlap method to determine collisions in cooperative drone operation scenarios and the intrusion method to determine collisions in non-cooperative drone operation scenarios. The alarm issuing module is used to issue alarms based on the judgment results.

[0008] Beneficial effects: The low-altitude UAV collision risk alarm method and system of this invention can provide real-time early warning functions. By deeply integrating trajectory prediction algorithms with dynamically adaptive protected area models, it achieves accurate and early differentiated alarms for low-altitude UAV collision risks. The system constructs optimal protected area models for collision risk identification in two different UAV operation scenarios: cooperative and non-cooperative. For cooperative UAVs, a dynamic Reich model protected area is used, whose protection radius can be dynamically adjusted to accurately match the protected area range with the actual maneuverability of the UAV. For non-cooperative UAVs with unclear flight intentions, a Monte Carlo random sampling-based reachability convex hull model is used to generate a three-dimensional protected area covering its possible future location, effectively addressing the challenges posed by its uncertainty. The collision risk judgment process also has differentiated characteristics. In the cooperative scenario, the collision risk is quantified by calculating the overlap of the spherical caps between two dynamic protected areas; while in the non-cooperative scenario, an alarm is triggered by determining whether the predicted trajectory point of the cooperative UAV intrudes into the reachability convex hull of the non-cooperative UAV. Finally, a graded alarm signal is output based on the prediction time window and the severity of the risk.

[0009] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention are realized and obtained in accordance with the structures particularly pointed out in the description, claims and drawings.

[0010] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0011] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0012] Figure 1 A flowchart illustrating the low-altitude UAV aerial collision risk alarm method and system for distinguishing between cooperative and cooperative operation scenarios according to the present invention; Figure 2 This is a schematic diagram illustrating a collaborative drone operation scenario. Figure 3 This is a flowchart of the data processing process; Figure 4 A diagram of the model structure used for trajectory prediction; Figure 5This is a schematic diagram of a dynamic spherical protected area model (showing the dynamic changes in radius). Figure 6 This is a schematic diagram illustrating the principle of calculating the volume of a spherical cap. Figure 7 A schematic diagram of the 3D convex hull construction for the Monte Carlo reachable region; Figure 8 Flowchart for hierarchical alarm logic judgment; Figure 9 This is a flowchart for invasive detection and judgment. Figure 10 This is a schematic diagram of the reachable domain trajectory generation in a two-dimensional case. Detailed Implementation

[0013] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions 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, 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.

[0014] Specifically, the process is as follows: Figure 1 As shown, this embodiment provides a method for mid-air collision risk warning for low-altitude unmanned aerial vehicles, including: Construct cooperative and non-cooperative drone operation scenarios; the cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route; the non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. Predict trajectories and construct protected areas; in cooperative drone operation scenarios, historical trajectory data of cooperative drones is obtained and input into the prediction model to output the predicted trajectory, while a dynamic Reich model protected area is constructed; in non-cooperative drone operation scenarios, historical flight trajectory data of both types of drones are obtained, the trajectory of cooperative drones is predicted, and an reachability protected area is constructed for non-cooperative drones. Collisions are determined based on predicted trajectories; specifically, the spherical cap overlap method is used to determine collisions in cooperative drone operation scenarios, while the intrusion method is used to determine collisions in non-cooperative drone operation scenarios. An alarm will be issued based on the judgment result.

[0015] Specifically, the operational scenarios for cooperative drones are as follows: Figure 2As shown, two drones with clear flight intentions interact on a preset flight path. The non-cooperative drone scenario involves one cooperative drone interacting with a non-cooperative drone whose flight intentions are random. Cooperative drones are those that comply with the "Interim Regulations on the Management of Unmanned Aerial Vehicle Flights," have compliantly reported flight plans and dynamics, possess clear flight intentions, and can share situational awareness of their surroundings. Non-cooperative drones refer to drones that have not reported flight plans and dynamics as required and cannot share situational awareness of their surroundings; their characteristic is unclear flight intentions.

[0016] Furthermore, the prediction model is an Adam-LSTM model, and its training method includes: Build a dataset and use it as input to the Adam-LSTM model; The dataset is divided into training and testing sets to train and test the Adam-LSTM model, resulting in a trained Adam-LSTM model.

[0017] Specifically, constructing the dataset includes using the dataset as input to the Adam-LSTM model; The dataset is divided into training and testing sets, and the data processing flow is as follows: Figure 3 As shown, the first step is to clean the trajectory data of cooperative drones, filtering out illegal data and addressing issues such as duplicates, poorly lengthed trajectories, and missing trajectories. First, linear interpolation is performed to complete the data. Assume we have a data sequence Y(t), where... and Given two known data points, and a distance t between them, we can use linear interpolation to find the value of this point. The specific linear interpolation formula is shown below: : in and These are the values ​​of two known data points; and t represents the location of known data points; t represents the location of missing values. The formula works by calculating the missing value based on the slope of the line between two known data points. This method is used to find missing values. This can then be calculated using linear interpolation. Next, the original data length is filtered; the trajectory length is artificially defined to be greater than or equal to 120 seconds, and trajectories shorter than 120 seconds are discarded. The data is then standardized and normalized; these simple methods will not be elaborated upon here.

[0018] Furthermore, the LSTM model includes an optimizer Adam and a two-layer LSTM network, wherein the Adam optimizer controls the learning rate, and the two-layer LSTM network processes and outputs the results. A codec-decode structure is used to achieve single-step temporal prediction. The LSTM network layer structure is as follows: Figure 4 As shown, the input layer receives 60 1-second time steps and five-dimensional features at each time step. These five-dimensional features include three-dimensional position coordinates and vertical and horizontal velocities. The network contains the following core components: input feature matrix X∈R 60×5 Includes standardized motion parameters: in , , For normalized three-dimensional coordinates, , These represent the normalized values ​​of the horizontal and vertical velocities, respectively. max To achieve the system's preset maximum flight speed, the first LSTM layer has 256 hidden units and uses a full-sequence output mode. This layer extracts local motion features through a gating mechanism. The second LSTM layer contains 128 units and outputs the context vector at the final time step. ∈R 128 The vector aggregates the global motion trend as shown in the following formula: The fully connected decoder consists of two dense layers: a feature decoding layer containing 128 units, activated using the ReLU function, and then... Mapping to a high-dimensional feature space, the calculation principle is as follows: The multi-step prediction layer contains 180 units, and a linear transformation is used to generate the future trajectory.

[0019] Specifically, the model training loss function is as follows. This invention uses two loss functions: a position loss function and a position loss function. and ， Its mathematical form is as follows: As shown. Position loss function L pos It is obtained by calculating the mean squared three-dimensional Euclidean distance between the predicted trajectory point and the actual position.

[0020] In this embodiment, a dynamic Reich model is constructed based on position uncertainty and basic drone data as a collision detection template in a cooperative drone operation scenario. Specific parameters need to be constructed based on the characteristics of the cooperative drone. Therefore, the main factors to be considered are as follows: aircraft performance, including speed, acceleration, and response time, flight mode, and regulatory standards.

[0021] The method for constructing a dynamic Reich model protected area includes: Computer body physical size radius : in This indicates the diameter of the drone's propeller after it has been deployed. This indicates the diameter of the drone's fuselage.

[0022] In one embodiment, taking the DJI Mavic Air2 as an example, its dimensions are 183*253*77mm, with a diagonal length of approximately 685mm. After the propellers are deployed, the length is approximately 0.68m. The radius of a single propeller is approximately 16cm, and the diameter of the cover when rotating is 0.8m. Therefore, the basic physical dimension radius is calculated to be 0.74m using the above formula.

[0023] Next, we consider the required dynamic error compensation, such as GPS positioning error, operational delay, and wind disturbance, and calculate the radius of the inner protection zone. : .

[0024] This refers to GPS positioning error, which can be found on the drone's official website or obtained through experience. In one embodiment, according to data from DJI's official website, when the GPS is working normally, the vertical accuracy of this model is ±0.5m and the horizontal accuracy is ±1.5m. Therefore, 1.5m is selected as the GPS positioning error compensation.

[0025] Therefore, the position drift caused by control signal delay at maximum flight speed can be obtained by the following formula, V max For maximum flight speed, T s To control the signal delay duration, ; In one embodiment, according to data from DJI's official website, the control signal delay is generally 100-150ms, so 150ms is chosen as the control signal delay; the maximum speed is 19m / s; according to the calculation formula, the position drift caused by the control signal delay is 2.85m under this data.

[0026] This is the offset amount under conditions where the flight control PID strongly compensates for wind force. In this embodiment, The value can be obtained from DJI Labs, or it can be taken from empirical values. 。 For example, considering a level 5 wind, which is the maximum wind speed the aircraft can withstand (10 m / s), and its impact over 2.5 seconds, the 2.5 seconds refers to the maximum time required for the aircraft to reduce its speed to 0 at its maximum speed (2.375 seconds). It's also necessary to consider the flight controller's active wind compensation; according to DJI lab data, under strong PID compensation in the flight controller, an offset of approximately 2.1 meters is observed.

[0027] Next, calculate the radius of the inner protection zone. Considering the combined physical dimensions of the fuselage radius and the three displacement deviations, the total displacement deviation reached 7.19m. Therefore, 7.2m can be selected as the base radius of the Reich collision template for the first layer of protection of this UAV.

[0028] Next, we consider the construction of the second layer dynamic radius of the Reich collision template. In this part of the study, we need to consider factors such as speed and reaction time.

[0029] Specifically, calculate the theoretically dynamically changing total radius. : Where v is the real-time speed of the UAV, and the instantaneous speed can be calculated by the speed decomposition calculation method, which is the difference between the positions of adjacent trajectory points divided by the time step.

[0030] a max This is the maximum deceleration of the drone, a value derived from the drone's own properties.

[0031] Indicates the time required for physical braking: ; T represents the response time. ; ; This is the additional compensation needed for the actual response time. It's sensor delay. It is a control signal delay. It's a flight control system response delay.

[0032] In one embodiment, the real-time speed v of the drone is taken as 19 m / s, and the maximum deceleration selected based on the drone model data is 8 m / s. The time required for physical braking of a high-speed drone (e.g., v = 19 m / s) is... The actual delay is 2.375s. However, considering the sensor delay of 0.3s, the control signal delay of 0.2s, and the flight control system response delay of 0.5s, The total time is 1 second, therefore the response time T is 3.375 seconds.

[0033] Ultimately, the theoretically dynamically changing total radius can be obtained. Theoretically, the maximum dynamic radius that can be achieved is 18.56m + 7.2m = 25.76m. The maximum dynamic radius can be set to 26m to avoid collision templates with excessively large radii.

[0034] Therefore, we obtained the template parameters for this model, and the calculation process is the same when applied to other models. The final change in the dynamic Reich model is as follows: Figure 5 As shown.

[0035] Furthermore, in cooperative drone operation scenarios, the overlap α of the dynamic Reich model is calculated by combining the trajectory prediction results with the spherical cap volume method, so that collisions can be judged based on whether the overlap exceeds a threshold.

[0036] Specifically, the method for determining the overlap of the spherical cap includes: Assume the distance between the centers of the two spheres is Their radii are respectively and ; When the distance between the centers of two balls is d≥r1+r2, the two balls do not intersect and the overlap α=0. When d ≤ |r1-r2|, then one sphere is completely contained within another sphere, and the overlap α = 1; and When |r1-r2|<d<r1+r2, the two spheres partially overlap, and α is calculated using the following formula: in, The volume of a sphere with radius r1, The volume of a sphere with radius r². The volume of the overlapping portion of two spheres. The volume of the overlapping portion of two spheres can be the sum of the volumes of the two spheres.

[0037] In one embodiment, optionally, the collision threshold can be set to 0.2, that is, no collision occurs when the overlap is in (0, 0.2], and a collision occurs when it is greater than 0.2.

[0038] Furthermore, a sliding window mechanism is used when predicting the trajectory: The window length is 60 seconds, and the step size is 1 second; that is, there are 60 predicted trajectory points within 60 seconds. The input is after standardization. ,in It is a normalized three-dimensional coordinate. These represent the normalized values ​​of the horizontal and vertical velocities, respectively. The output is a sequence of three-dimensional coordinates for the next 60 seconds.

[0039] In one embodiment, in a cooperative drone operation scenario, the method for issuing an alarm based on the judgment result includes: dividing the total duration of the prediction window into several time windows; defining collision risk levels for different collision scenarios; generating alarm types of corresponding severity based on the different collision risk levels of each predicted time window; and selecting the alarm with the highest severity from among multiple alarm types.

[0040] In one embodiment, the total prediction window duration of 60s can be divided into three windows: the first window is for prediction time ≤ 15s, the second window is for prediction time > 15s and the third window is for prediction time > 30 ... second window is for prediction time > 30s and the third window is for prediction time > 60s and the second window is for prediction time > 30s.

[0041] Collision risk levels are ranked from highest to lowest as L1 > L2 > L3. The collision risk level can be determined based on different collision scenarios using the following steps: S1: Determine if there are 3 consecutive points in the window that have been detected as intrusions. If so, the collision risk level is L1; otherwise, execute S2. S2, determine whether there are at least 6 intrusions detected among the 10 consecutive prediction points in the window. If so, the collision risk level is L2; ​​otherwise, execute S3. S3: Determine whether at least 4 out of 10 consecutive prediction points within the window have detected intrusion. If so, the collision risk level is L3; otherwise, there is no collision risk.

[0042] Alarm types are ranked from highest to lowest severity as follows: Alarm γ > Alarm > Warning. Specifically, an alarm is triggered when the collision risk level is L3 or higher within the first 30 seconds (i.e., the first and second windows), and a warning is triggered when the collision risk level is L3 or higher within the last 30 seconds (i.e., the third window). Furthermore, alarms occurring within the first 15 seconds (i.e., the first window) of the first 30 seconds require the addition of γ, meaning alarm γL1 is the highest-level alarm type.

[0043] See Figure 8In a specific scenario, during a 60-second trajectory prediction, if there are three consecutive points in the first window, and six out of ten points exceed the threshold (alarm type γ, collision risk level L1 or L2), the final alarm result for this window is alarm γL1. In the second window, if there are three consecutive points, and four out of ten points exceed the threshold (alarm type WARNING, collision risk level L1 or L3), the final alarm result for this window is alarm L1. In the third window, if there is an overlap greater than 0.2, the alarm type is warning. Finally, the alarm with the highest severity is selected from multiple alarm types, and the overall state of the predicted drone is alarm γL1.

[0044] Furthermore, rolling prediction verification can be performed. Specifically, this includes: re-executing trajectory prediction every 10 seconds; verifying risk points beyond 30 seconds from the previous prediction, and deactivating the alarm if the overlap remains below 0.15 for 15 consecutive steps.

[0045] For example, after a 60-second prediction, the first predicted alarm result is obtained. Then, after 10 seconds, another 60-second prediction is performed, with a 50-second overlap between the two predictions. At this point, alarm points that were in a warning state 30 seconds or less in the previous prediction can be verified within this second prediction period.

[0046] For example, if the overlap between the first prediction and the second prediction exceeds the threshold in the 33s-38s interval, the alarm type is set to warning. In the second prediction, between 23s and 28s, if the overlap also exceeds the threshold, the accuracy of the first prediction is verified; otherwise, the reliability of the first prediction is considered low. If, within a 20s-50s window, the overlap of 15 consecutive points remains below 0.15, the alarm is deactivated. The advantage of this rolling prediction is that it allows for multiple verifications of the accuracy of the predicted alarms and warnings.

[0047] Furthermore, taking DJI Mavic Air 2 as an example again, the method for constructing an reachability protection zone for non-cooperative drones is to generate a dynamic reachability convex hull for non-cooperative drones based on the drone dynamics equations and the Monte Carlo random sampling method. This method is based on the current state of the non-cooperative drone, which in this paper refers to the position xyz, horizontal velocity, horizontal acceleration, vertical velocity, and vertical acceleration, and predicts the possible positions of the non-cooperative drone in each direction within the next 10 seconds.

[0048] Specifically, this includes: dividing the non-cooperative drone into three motion modes. The first motion mode is the turning mode, in which the acceleration of the non-cooperative drone is set to change along the final random direction of flight, simulating the trajectory of the drone flying straight after turning to a predetermined angle; the second motion mode is the directional mode, in which the acceleration of the non-cooperative drone will be along the current speed direction, simulating stable flight; the third motion mode is the braking mode, in which the non-cooperative drone will randomly generate a reverse acceleration, simulating the situation of the non-cooperative drone braking and then moving in the opposite direction. First, regarding turning, the trajectory of the drone in a planar turning situation is derived by referring to basic kinematic formulas and the minimum turning radius.

[0049] Calculate the minimum turning radius: Where φ = 9.8 m / s is the acceleration due to gravity. The load is selected as the square of the maximum overload factor, referring to Section 29.337 of Chapter C, Flight Load, in the "Airworthiness Regulations for Transport Category Rotorcraft" issued by the Civil Aviation Administration of my country. It is 3.5. Indicates the speed of the drone; Secondly, the straight-flight process after the turn, the orientation mode, and the braking mode are calculated using basic dynamics formulas. The following are the calculation formulas in the x-axis direction: in Indicates the duration of a flight; get Figure 10 Following the planar trajectory shown, the two-dimensional planar flight trajectory is extended to a three-dimensional state. The maximum vertical ascent and descent speeds of the UAV are defined. After the turning phase, a direction is randomly selected at 180° in the vertical direction for straight flight, resulting in multiple points of the farthest trajectory. These points are linked together as patches of the reachable domain, forming a convex hull of the reachable domain.

[0050] In one embodiment, based on the performance parameters of the DJI Mavic Air 2, the maximum vertical ascent speed and maximum descent speed were first specified as 4 m / s and 3 m / s, respectively. Then, it was stipulated that after the turning phase, the drone would fly straight in a random 180° direction vertically. After setting these parameters, a random trajectory was generated based on the Monte Carlo random sampling principle. Physical constraints were applied during trajectory generation, meaning the minimum altitude was no less than 5 m and the maximum altitude was no more than 240 m, limiting the simulated drone trajectory from exceeding the predetermined parameters. This resulted in a set of multiple data points composed of trajectory points. Next, the outermost points were collected as points for constructing a convex surface, and the points were linked as patches of the reachable region, thus forming a reachable region. The constructed reachable region is shown below. Figure 7 As shown.

[0051] In some embodiments, optionally, the difficulty of generating reachable domains can be optimized at the end. For a 3*3*3 network, two points are randomly reserved as points used to build the convex hull patch of the reachable domain, thereby reducing the device load.

[0052] Furthermore, the collision determination using the intrusion method is achieved through the following steps: Decompose the convex hull into a set of triangular facets, Face={ , , }, , , These are the coordinates of the vertices of the triangle; Calculate the outward-facing unit normal vector for each facet. ; By symbolic distance Determine the relative position of a point to the convex hull; if the predicted trajectory point... Satisfy all dough pieces <0 indicates a point Located inside the convex hull; if >0, then point Located outside the normal direction of the surface; if =0, then point It is located exactly on the dough; in Indicates from the vertex of the face Point of view The vector, This represents the length of the projection of the vector onto the direction of the normal vector n.

[0053] In non-cooperative drone operation scenarios: If the predicted trajectory points of cooperative drones intrude into the reachability protection zone of non-cooperative drones within a preset time (i.e., the points...),... If the location is inside the convex hull, an alarm will be triggered immediately. In the embodiment, the preset time can be 10 seconds, 15 seconds, etc. The specific judgment algorithm flow is as follows: Figure 9 As shown.

[0054] Based on the above embodiments, another embodiment of this application provides a low-altitude unmanned aerial vehicle (UAV) mid-air collision risk warning system, including: The scenario building module is used to construct cooperative and non-cooperative drone operation scenarios. The cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route. The non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. The trajectory prediction module is used to predict the trajectory of cooperative drones; The protected area construction module is used to construct dynamic Reich model protected areas in cooperative drone operation scenarios, and to construct reachable domain protected areas for non-cooperative drones. The collision detection module is used to determine collisions based on the predicted trajectory. Specifically, it uses the spherical cap overlap method to determine collisions in cooperative drone operation scenarios and the intrusion method to determine collisions in non-cooperative drone operation scenarios. The alarm issuing module is used to issue alarms based on the judgment results.

[0055] In this embodiment, the working method of each module is as described above, and will not be repeated here.

[0056] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus embodiments described above are merely illustrative; for example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0057] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0058] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0059] Based on the above-described preferred embodiments of the present invention, and through the foregoing description, those skilled in the art can make various changes and modifications without departing from the inventive concept. The technical scope of this invention is not limited to the contents of the specification, but must be determined according to the scope of the claims.

Claims

1. A method for mid-air collision risk warning for low-altitude unmanned aerial vehicles (UAVs), characterized in that, Includes the following steps: Construct cooperative and non-cooperative drone operation scenarios; the cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route; the non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. Predict trajectories and construct protected areas; in cooperative drone operation scenarios, historical trajectory data of cooperative drones is obtained and input into the prediction model to output the predicted trajectory, while a dynamic Reich model protected area is constructed; in non-cooperative drone operation scenarios, historical flight trajectory data of both types of drones are obtained, the trajectory of cooperative drones is predicted, and an reachability protected area is constructed for non-cooperative drones. Collisions are determined based on predicted trajectories; specifically, the spherical cap overlap method is used to determine collisions in cooperative drone operation scenarios, while the intrusion method is used to determine collisions in non-cooperative drone operation scenarios. An alarm will be issued based on the judgment result.

2. The low-altitude UAV collision risk alarm method as described in claim 1, characterized in that, The prediction model is an Adam-LSTM model, and its training method includes: Build a dataset and use it as input to the Adam-LSTM model; The dataset is divided into training and testing sets to train and test the Adam-LSTM model, resulting in a trained Adam-LSTM model. The Adam-LSTM model includes an optimizer Adam and a two-layer LSTM network; the Adam optimizer controls the learning rate, and the two-layer LSTM network processes and outputs the results, using an encoding-decoding structure to achieve single-step time series prediction.

3. The low-altitude UAV mid-air collision risk alarm method as described in claim 1, characterized in that, The method for constructing a dynamic Reich model protected area includes: Computer body physical size radius : in This indicates the diameter of the drone's propeller after it has been deployed. Indicates the diameter of the drone's fuselage; Calculate the radius of the inner protection zone : It is a GPS positioning error; Therefore, the position drift caused by control signal delay at maximum flight speed can be obtained by the following formula, V max For maximum flight speed, T s To control the signal delay duration, ； It is the offset amount when the flight controller's PID is strongly compensated to counteract wind force; Calculate the theoretically dynamically changing total radius. : Where v is the real-time speed of the drone. Where T is the maximum deceleration of the drone, and T is the response time. Indicates the time required for physical braking; and ； ； ； This is the additional compensation needed for the actual response time. It's sensor delay. It is a control signal delay. It's a flight control system response delay.

4. The low-altitude UAV mid-air collision risk alarm method as described in claim 1, characterized in that, The method for constructing reachability protection zones for non-cooperative drones involves generating dynamic reachability convex hulls for non-cooperative drones based on drone dynamics equations and Monte Carlo random sampling, including: The non-cooperative drone is divided into three motion modes. The first motion mode is the turning mode, in which the acceleration of the non-cooperative drone is set to change along the final random direction, simulating the trajectory of the drone flying straight after turning to the predetermined angle. The second motion mode is the directional mode, in which the acceleration of the non-cooperative drone will be along the current speed direction, simulating stable flight. The third motion mode is the braking mode, in which the non-cooperative drone will randomly generate a reverse acceleration, simulating the situation of the non-cooperative drone braking and then moving in the opposite direction. First, regarding turning, the trajectory of the drone in a planar turning situation is derived by referring to basic kinematic formulas and the minimum turning radius; Calculate the minimum turning radius: Where φ = 9.8 m / s is the acceleration due to gravity. It is the square of the maximum overload factor. Indicates the speed of the drone; Secondly, the straight-flight process after the turn, the orientation mode, and the braking mode are calculated using basic dynamics formulas. The following are the calculation formulas in the x-axis direction: in Indicates the duration of a flight; Extend the two-dimensional flight trajectory to a three-dimensional state, define the maximum vertical ascent and descent speeds of the UAV, and after the turning phase, randomly select a direction of 180° in the vertical direction to fly straight, obtaining multiple points of the farthest trajectory. Connect each point as a patch of the reachable domain to form a convex hull of the reachable domain.

5. The low-altitude UAV mid-air collision risk alarm method as described in claim 1, characterized in that, The method for determining the overlap of the spherical cap includes: Assume the distance between the centers of the two spheres is Their radii are respectively and ; When the distance between the centers of the two balls is d≥r1+r2, the overlap α=0; When d ≤ |r1-r2|, the overlap α = 1; and When |r1-r2|<d<r1+r2, the overlap α is calculated using the following formula: in, The volume of a sphere with radius r1, The volume of a sphere with radius r². The volume of the overlapping portion of the two spheres.

6. The low-altitude UAV mid-air collision risk alarm method as described in claim 1, characterized in that, The collision detection method using intrusion is achieved through the following steps: Decompose the convex hull into a set of triangular facets, Face={ , , }, , , These are the coordinates of the vertices of the triangle; Calculate the outward-facing unit normal vector for each facet. ; By symbolic distance Determine the relative position of a point with respect to the convex hull, where the point is... These are the predicted trajectory points. If the point... Satisfy all dough pieces <0 indicates a point Located inside the convex hull; if >0, then point Located outside the normal direction of the surface; if =0, then point It is located exactly on the dough; in Indicates from the vertex of the face Point of view The vector, This represents the length of the projection of the vector onto the direction of the normal vector n.

7. The low-altitude UAV collision risk alarm method as described in claim 1, characterized in that, The predicted trajectory uses a sliding window mechanism: Window length 60 seconds, step size 1 second; The input is after standardization. ,in It is a normalized three-dimensional coordinate. These represent the normalized values ​​of the horizontal and vertical velocities, respectively. The output is a sequence of three-dimensional coordinates for the next 60 seconds.

8. The low-altitude UAV mid-air collision risk alarm method as described in claim 1, characterized in that, Methods for issuing alarms based on the judgment results include: In cooperative drone operation scenarios: the total prediction window duration is divided into several time windows; collision risk levels for different collision scenarios are defined; alarm types of corresponding severity are generated based on the different collision risk levels of each predicted time window, and the alarm with the highest severity is selected from multiple alarm types to issue an alarm; In non-cooperative drone operation scenarios: If the predicted trajectory point of a cooperative drone intrudes into the reachable protection zone of a non-cooperative drone within a preset time, an alarm will be triggered immediately.

9. The low-altitude UAV mid-air collision risk alarm method as described in claim 7, characterized in that, Also includes: In collaborative drone operation scenarios, perform rolling prediction verification: The trajectory prediction is re-executed every 10 seconds; Verify the risk points beyond 30 seconds in the previous prediction. If the overlap is below 0.15 for 15 consecutive steps, the alarm is lifted.

10. A low-altitude unmanned aerial vehicle (UAV) mid-air collision risk warning system, characterized in that, include: The scenario building module is used to construct cooperative and non-cooperative drone operation scenarios. The cooperative drone operation scenario includes two cooperative drones with clear flight intentions interacting on a preset route. The non-cooperative drone operation scenario includes one cooperative drone interacting with one non-cooperative drone with random flight intentions. The trajectory prediction module is used to predict the trajectory of cooperative drones; The protected area construction module is used to construct dynamic Reich model protected areas in cooperative drone operation scenarios, and to construct reachable domain protected areas for non-cooperative drones. The collision detection module is used to determine collisions based on the predicted trajectory. Specifically, it uses the spherical cap overlap method to determine collisions in cooperative drone operation scenarios and the intrusion method to determine collisions in non-cooperative drone operation scenarios. The alarm issuing module is used to issue alarms based on the judgment results.

Images