Unmanned aerial vehicle cluster collaborative search algorithm based on probability distribution diagram updating rule under communication interference

Abstract

The invention discloses an unmanned aerial vehicle cluster collaborative search algorithm based on a probability distribution diagram updating rule under communication interference, and aims to search a moving target carrying a communication delay interference device in a large area by using a cluster formed by a small number of unmanned aerial vehicles. Searching a target in a specific area is an important task in the field of unmanned aerial vehicle application. According to the method, a search algorithm is locally deployed for individuals in an unmanned aerial vehicle cluster, a target probability distribution diagram of the whole region is generated and updated in real time by means of end-side calculation and communication of the unmanned aerial vehicle, and a search strategy is designed to guide the unmanned aerial vehicle to complete a search task as soon as possible. In the scene, a set of unmanned aerial vehicle cluster collaborative search algorithm system can be formed by matching the algorithm front end to estimate target parameters and the rear end sensor target identification. Through analogue simulation, the algorithm can quickly realize expected functions.

Description

I. Technical Field

[0001] This invention patent belongs to the category of algorithms and is applied in the field of drones, involving electronic and communication technologies. II. Background Technology

[0002] Recently, the development of unmanned aerial vehicle (UAV) technology has led to UAVs undertaking an increasing number of tasks, including search missions. Typical UAV search missions aim to have UAVs patrol a designated area to locate targets. UAVs performing search missions are generally equipped with sensors to identify specific targets; for example, they may carry optical imaging devices and rely on image processing algorithms to identify people or objects. On the one hand, when the search area is large, especially when it is much larger than the sensor's sensing range, the sensing area can be expanded by increasing the number of UAVs to reduce search time, and targets can be searched for over a wide area through UAV swarm collaboration. On the other hand, considering search costs, it is desirable to complete the search mission within a tolerable time using as few UAVs as possible. Therefore, UAV swarm collaborative search algorithms aim to find the optimal solution between time cost and UAV cost, achieving the best search results with minimal resources or relying on existing resources.

[0003] Different search scenarios present different problems for search algorithms. In adversarial scenarios, targets may carry communication jamming devices to hinder communication between drones and disrupt the collaborative search process. Drones affected by communication jamming will experience delays in sending and receiving data, making it impossible for them to acquire information from other drones in real time. Generally, the communication jamming devices carried by targets are limited in power and have a limited interference range, and outside the interference range, they have little impact on communication between drone swarms. In adversarial search processes involving communication jamming, the target parameters are unknown; therefore, search algorithms can incorporate parameter estimation methods to estimate the target parameters in order to discover the target more quickly.

[0004] During mission execution, drone swarms can be deployed in two ways, differing in the presence or absence of a central server to schedule individual drones. If a central server exists, the drones in the swarm communicate with it, the central server aggregates information, makes decisions, and issues commands to the drones. If no central server exists, the drones in the swarm communicate with each other, and each drone makes decisions locally. Considering communication interference, drones in the final search phase need to identify targets within the interference zone. Individuals in the interference zone will experience delays while waiting for commands from the central server, and the central server will also experience delays in receiving information from interfered individuals. Therefore, a central server deployment mode will reduce the search accuracy and efficiency in the final search phase. Deploying the search algorithm locally on the drones can avoid problems caused by communication delays. III. Summary of the Invention

[0005] This invention patent provides a cooperative search algorithm for UAV swarms to solve the problem of multi-target cooperative search in communication interference scenarios. The algorithm is deployed on all individual UAVs in the swarm, relying on a target parameter extraction algorithm to estimate the target's interference range and movement speed, and combining this with the sharing of information such as position and probability distribution maps among the UAVs to achieve local guidance for the UAVs.

[0006] The core of this invention is the combination of the design of probability distribution map update rules under disturbed scenarios and the design of a search algorithm based on the probability distribution map. 1. Probability Distribution Map Update Rules:

[0007] The probability distribution map is constructed using a grid-based method, establishing a Cartesian coordinate system with north-south and east-west axes. The grid spacing significantly impacts search accuracy. Overly coarse grids increase errors in the sensor's perception area during deployment, leading to missed detections; conversely, overly fine grids increase the processor's computational burden, straining computing resources. This is detrimental to environments with limited computing resources and limited payloads for UAVs, thus reducing response speed. Therefore, it is recommended to design the grid size by comprehensively considering the target's size and the UAV's flight parameters, minimizing computational resource consumption while ensuring accuracy.

[0008] In the actual construction and updating of the probability distribution map, the grid values ​​are not calculated based on probability values, but rather on threat values. That is, the actual calculation involves the threat coefficient distribution map. Threat coefficients and probability distributions are interchangeable; the probability distribution map can be obtained by dividing each grid in the threat coefficient distribution map by the sum of all grids in the entire threat coefficient distribution map. The reason for not directly using the probability distribution map is that the sum of all grids in the probability distribution map is always 1, making local operations and calculations difficult. Since there may be more than one target within the area, the algorithm's search decision is mainly influenced by the local area adjacent to the drone; using the concept of threat coefficients simplifies the calculation.

[0009] The probability distribution map (or threat coefficient distribution map, hereinafter referred to as the distribution map) is first constructed using the search area size as a parameter. The grid cells in the distribution map typically take values ​​between 0 and 1, but may take values ​​greater than 1 when a high-threat event occurs. A larger cell value indicates a greater threat level. Initially, since the target location and state are unknown, a uniform value of 1 is assigned. A value of 1 indicates that the grid location is in an undetermined state, neither safe nor posing a high threat.

[0010] The drone's update of the distribution map is divided into local updates (primary updates) and communication updates (secondary updates). The drone first updates the distribution map based on its own status, and then updates it a second time by combining information from the drone swarm after communication. Based on the usage scenario, the algorithm designs five update methods in the primary update stage: natural update method, undisturbed update method, disturbed status change update method, sensor update method, and high-threat area reset method.

[0011] ① Natural Update Rule: Since the target may be in motion, a natural update method needs to be introduced to update the distribution map globally.

[0012] In updated low-threat areas, considering the possibility of the target moving to that location, a threat reassessment is necessary. In high-threat areas, considering the possibility of the target moving to a different location, the range needs to be expanded. Therefore, it is necessary to combine the estimated target velocity and the distribution map update time interval to determine the range of local updates, and then traverse each grid cell to take the maximum value within the local update range as the new value.

[0013] ② Uninterrupted Update Rule: When the drone is not interfered with, it means that the drone is outside the interference range. Conversely, at this time, the update rule is updated based on the drone's status.

[0014] Centered on the drone, the area within a circle with the estimated target interference range radius can be considered as having no target. Therefore, reducing the threat value of this area in the distribution map is considered, and this area is called the low-threat zone. The reduction in threat value can be represented by a two-dimensional normal distribution with the drone's coordinates as the mean point. The closer to the drone's location, the greater the threat value reduction. The drone's coordinates are set to 0, and cells beyond the estimated target interference range are not updated. The variance of the normal distribution should be related to the confidence level of the target interference range estimate. When the interference range estimate is reliable, a larger variance is expected to reduce the threat value within the area more efficiently; when the interference range estimate is unreliable, a smaller variance is expected.

[0015] The possibility of a target still existing within the update range should be kept, and the variance should be relatively small. Note that the sum of the distributions within the update range is not necessarily 1.

[0016] ③ Interference Status Change Update Rules: When the UAV's interference status changes, it indicates that the UAV has just entered the target's interference range or has just...

[0017] Upon leaving the target interference range, the UAV, depending on the change method, identifies a vague area at the estimated target interference range as a high-threat region. This high-threat region can be a wide annular strip, the width of which is related to the confidence level of the estimated target interference range. Higher confidence results in a narrower strip. When the UAV's interference state changes, the annular region at the estimated target interference range needs to be defined based on the estimated target speed and its confidence level. Generally, assuming the UAV's speed is greater than the target speed, when the UAV first enters the interference state, the semi-circular region at the estimated target interference range along the UAV's speed direction is more likely to be designated as the updated region; when the UAV first leaves the interference state, the semi-circular region at the estimated target interference range along the opposite direction of the UAV's speed is more likely to be designated as the updated region. The high-threat zone needs a value greater than 1 to reflect a high threat level. Since the threat value of the grid within the statistical range will be weighted by distance when the drone first enters the interference zone (the weight decreases with distance), to ensure the high-threat zone dominates the search process, its value can be set to be greater than the reciprocal of the distance weight to balance the impact of closely spaced grids on the drone's search strategy. High-threat zones may be defined multiple times during the search. When high-threat zones are repeatedly defined, the threat values ​​need to be accumulated for more accurate target location.

[0018] ④ Sensor update rules: Sensors serve as a reliable means of target identification and a last resort for target discrimination, regardless of whether it is a drone.

[0019] The sensor's operation is unaffected by its state. If the sensor does not detect a target, it updates the threat value within the sensing area to 0, creating a safe zone. If the sensor detects a target, it broadcasts the target's location and awaits further instructions.

[0020] ⑤ High-threat area reset rule: When a drone detects a target, the high-threat area where the target is located needs to be reset to reduce the threat to the target.

[0021] To mitigate the threat level of the target area and prevent it from further hindering drones searching the vicinity, the relevant area's value needs to be reset to a pending state (value 1).

[0022] The above describes a method for updating the distribution map in one step. Following the update interval, the drone updates its local distribution map in real time, broadcasts its current status using a pre-defined communication interval, and receives information from other drones in the cluster. After receiving the current distribution map information from other drones, the drone performs a second update on its local distribution map. First, high-threat areas in all distribution maps are integrated, retaining all high-threat units. Then, low-threat areas in the distribution map are integrated, retaining the values ​​of the low-threat areas. If a high-threat value (greater than 1) and a low-threat value (less than 1) conflict, the low-threat value is assigned. Since the states of individuals within the drone cluster differ, different individuals may have inconsistent values ​​for a particular local location. In this case, the update rule for low-threat values ​​in this invention is more reliable than the update rule for high-threat values, therefore, low-threat values ​​should be prioritized. Only when all drones consider a unit to be in a pending state (value 1) or a high-threat state (greater than 1) is it assigned a high-threat value.

[0023] 2. Guidance algorithm based on probability distribution map:

[0024] The drone determines its current flight direction based on the local distribution map. Intuitively, the drone will move towards the area with the highest threat value. Therefore, the search algorithm aims to translate the local distribution map (two-dimensional) into a threat coefficient distribution vector (one-dimensional, hereinafter referred to as the distribution vector) indexed by the flight direction, and then find the direction with the highest threat value in the table for guidance. When searching for the direction with the highest threat value, since the distribution map update caused by the search process has an angular range, searching for a continuous range of directions with a higher overall threat value can reduce the overall threat more quickly than searching for a single direction with a high threat value. Therefore, the algorithm should perform local sliding window summation after obtaining the distribution vector and before searching for the direction with the highest threat value, hoping to obtain the optimal benefit value for guidance in a certain direction. This summation should also highlight the threat value at the center angle of the sliding window, so windowing operations can also be performed within the sliding window.

[0025] In harassed application scenarios, the update rules for the target distribution map differ between harassed and undisturbed scenarios, as does the target's expected location (whether within or outside the estimated harassment range). Therefore, different guidance strategies should be selected. Furthermore, for the drone swarm as a whole, it is desirable to improve collaboration efficiency and avoid redundant searches to minimize the overall threat value of the distribution map. Therefore, the algorithm will consist of two parts: a distribution vector calculation method and a proximity penalty method.

[0026] ① Distribution vector calculation module:

[0027] Considering the overall drone swarm, it needs to focus on both high-threat areas and a large unsearched area, avoiding a situation where all drones ignore high-threat areas or focus solely on them. The algorithm employs a proximity principle, using distance weighting to encourage drones to focus on nearby areas. If a drone is close enough to a high-threat area, it prioritizes searching that area; conversely, if a drone is far from a high-threat area, there's no need to travel there and ignore the current area.

[0028] Furthermore, considering that the search area covered by different flight directions within a region varies depending on the individual drone's location, a natural directional weighting effect occurs, making drones more inclined to fly towards the center of the area (for example, if all 10 grid vectors are in an undetermined state, and the drone is in the third grid from the left, with 2 grids to its left and 7 grids to its right, the total threat value on the right is naturally higher than on the left). Therefore, it is necessary to define a boundary for converting the distribution map to the distribution vectors (e.g., specifying that only threats within 2 grids are considered). This would both eliminate the drone's tendency to be centered on the area and naturally guide the drone back into the search area when it is far away from it.

[0029] Taking all the above into consideration, the algorithm will design different local threat value calculation methods based on the interference status of individual drones.

[0030] When a drone is in an undisturbed state, there is a high probability that there are no targets within the radius of its estimated target interference range. Therefore, to avoid the impact of uneven distribution within this range on the drone, the threat value within this range is not considered. A wide annular region is considered outside this range for calculating the threat value distribution vector. The range of this region needs to fully consider adjacent potential high-threat areas. When a drone first discovers a high-threat area, it is relatively close to that area, so it is not expected to leave that area. If the drone is in an undisturbed state and a new threat area is defined just as it transitions from an undisturbed state, it should change direction and re-search for high-threat areas in the opposite direction of its flight. Considering the diffusion caused by the natural updating of high-threat areas and the drone's turning radius, it is recommended that the width of this annular region be no less than the drone's turning radius. In this case, the distance from the grid cell within the annular region to the inner ring boundary can be used as a weight to weight the grid threat value.

[0031] When a UAV is under interference, and the target is within the area estimated as the radius of the interference range, searching nearby areas is more likely to find the target. Furthermore, delays caused by communication interference make it difficult for the UAV to receive external information, leading to inaccurate distribution maps outside the estimated radius. In this situation, the UAV should focus on searching for nearby targets rather than other areas, and consider converting the distribution map to a distribution vector within the estimated radius. Targeted guidance can be achieved by repeatedly entering this area and by addressing distribution differences within the range caused by sensor update methods. In this case, the threat value of the grid can be weighted using the distance from the target grid cell to the UAV.

[0032] ② Proximity penalty module:

[0033] When two drones are too close, their distribution vector calculations will result in a significant overlap in their distribution map ranges, potentially leading to similar guidance directions. This proximity and similar guidance directions greatly increase the likelihood of overlapping distribution map update ranges, especially when their flight directions are also similar. This is detrimental to reducing the overall threat level within the search area. Consider introducing proximity penalties to guide the two drones and differentiate their guidance strategies.

[0034] For a given drone, if another drone is too close, consider adding a weight in the opposite direction to the other drone when generating the distribution vector. Consider using this opposite angle as the mean and applying a normal distribution to weight the threat values ​​in the distribution map. The variance of the normal distribution will be related to the distance between the two drones; the closer they are, the smaller the variance, and the more the algorithm will tend to adjust the weights to move them further apart. Since the weighting directions of the two drones are opposite, this could potentially change their guidance directions.

[0035] It is important to emphasize that when at least one of the two drones is in a jammed state, due to latency, the drones do not penalize each other's distribution vectors. In this situation, the jammed drone is relatively close to the target, and imposing a penalty could significantly impact target search. Furthermore, the jammed drone only performs sensor updates and natural updates to the distribution map, resulting in a smaller search range. This also reduces the probability and cost of two nearby drones repeatedly searching in this state. Therefore, when the target is in the late stages of the search (in a jammed state), proximity penalties will not be considered.

[0036] The above-mentioned distribution vector calculation method and proximity penalty method can ultimately calculate the directional gain from the distribution map, and guide the UAV to complete the turning flight based on the maximum directional gain. IV. Description of the attached drawings

[0037] Figure 1This is a schematic diagram illustrating an algorithm application scenario. The diagram demonstrates the search for ground targets as an example. The drone enters from one side of the predetermined search area and searches for moving ground targets within that area. Collaboration is achieved through real-time communication during the search. Sensors utilize optical imaging equipment to identify ground targets.

[0038] Figures 2 to 79 All examples provide demonstrations of the algorithm's performance, which can intuitively show the algorithm's effects and the design details described in the invention content section. Figures 2 to 79 Corresponding to 4 scenarios ( Figures 2 to 19 Corresponding to scenario 1; Figures 20 to 34 Corresponding scenario 2; Figures 35 to 55 Corresponding scenario 3; Figures 56 to 79 Corresponding to scenario 4), the actual effects of the algorithm proposed in this invention patent in the predetermined scenario are shown in sequence. Figures 2 to 79 The example scenario corresponds to Figure 1 After entering the search area from one side, the drone searches for ground targets. Ground targets exhibit both circular and linear motion. Circular motion ensures the target's trajectory does not extend beyond the search area; targets moving in linear motion will return along their original path after reaching the boundary. In the diagram, squares represent drones, thick arrows originating from squares indicate the drone's current direction of movement, and thin arrows indicate the drone's current guidance direction. Crosses represent ground targets, and thick arrows originating from crosses indicate the ground target's direction of movement. Within the search area, red indicates a value of 1, blue indicates a value of 0, and high-threat areas are not shown because their values ​​exceed 1. The drone's initial position is randomly assigned in a uniform distribution along one boundary. Furthermore, upon detecting a target, the drone will stop updating its local distribution map and, using the communication delay interval, transmit the target detection information to other drones in the cluster before finally docking at the ground target. When a drone docks at a ground target, both the drone and the target are considered inactive; the drone no longer participates in the search mission, the ground target no longer causes communication interference, and it will not be recognized by other drones.

[0039] The common parameters across the four scenarios include: a search area of ​​240m x 240m; a grid granularity of 1m x 1m; four drones in the cluster; a fixed flight speed of 10m / s; a communication interval of 1 second; a guidance interval of 0.5 seconds; an optical imaging sensor with an imaging area of ​​20m x 20m at a distance of 20m from the flight direction; a drone turning radius of 20m; two ground targets; a ground target speed of 3m / s; a ground target communication interference radius of 50m; a ground target communication interference delay of 5 seconds; one ground target moving in a straight line; and another ground target moving in a circle with a radius of 100m. All distance-related parameters can be scaled up or down proportionally. The different parameters for the four scenarios are as follows: Figures 2 to 19 The estimated target interference range and target velocity are taken as true values; Figures 20 to 34 The estimated target interference range is taken as the true value, and the estimated target velocity is taken as (1 + 26.7%) times the true value; Figures 35 to 55 The estimated target interference range is taken as (1-8%) times the true value, and the estimated target speed is taken as the true value; Figures 56 to 79 The target interference range estimate is taken as (1-8%) times the true value, and the target velocity estimate is taken as (1+26.7%) times the true value. The confidence level for the target interference range estimate is set to 75%, and the confidence level for the target velocity estimate is set to 1 by default. In the interference state change update method, the initial width of the high-threat annular region is one grid cell. The high-threat region reset method resets all high-threat cells connected to the target location. In the search algorithm, the local angle summation range and the penalty application range are both 45 degrees, and the distribution vector is windowed using a logarithmic method during the summation phase.

[0040] Figures 2 to 79 The algorithm demonstrates that it can guide UAVs to perform search functions and has a certain tolerance for errors in target estimation parameters. V. Detailed Implementation Methods

[0041] The algorithm proposed in this invention is deployed in each individual drone in a swarm. Each drone is required to have the ability to estimate the interference range and speed of the target. The estimation algorithms for these two parameters are not within the scope of this patent, but they can also be estimated using a distribution map. Furthermore, the drones in the swarm need to have the ability to communicate with each other and perform self-localization, transmitting their respective positions and distribution map information so that the drones can integrate the overall swarm situation locally.

[0042] Drones make decisions locally based on the information they receive. It can be observed that undisturbed drones in a swarm, due to uninterrupted communication, will obtain the same distribution map after one and two updates. Therefore, the local decisions made by drones in this state can be predicted by other undisturbed drones in the swarm. Based on this, it can be said that the undisturbed drone swarm is completely unified, with no unexpected situations. However, disturbed drones, unaware of the real-time information of other drones, and unable to obtain the real-time information of disturbed drones, are isolated from the swarm. Once they recover from the disturbance, they will reintegrate their local historical information into the swarm. Because the information of isolated drones is delayed, relying on delayed information will lead to decision-making errors. Therefore, it is recommended that neither the drone swarm nor disturbed drones use delayed information in decision-making. Delayed information should only be used for information that does not affect search decisions, such as whether the target has been found.

[0043] During algorithm implementation, since multiple algorithm modules are involved, the temporal relationships need to be clearly defined. Distribution Figure 1 The time intervals between the first and second updates can be inconsistent, but if they need to be executed simultaneously, the first update must be performed first, followed by the second update, to ensure the timeliness of the information provided to other individuals within the cluster. Furthermore, the distribution graph update must be performed before calculating the distribution vector to guarantee timeliness. In addition, the time interval between the first and second updates should be kept as short as possible during implementation to improve computational accuracy.

[0044] The following Algorithm 1 presents the flowchart of the UAV cooperative search algorithm:

[0045]

[0046]

[0047] Algorithm 1 does not constrain the performance of the UAV, and its parameters are unrelated to the UAV's flight and control modes. It only generates the distribution map and calculates the distribution vector based on the UAV's current state (coordinates, sensor status, and disturbance status). Other UAV parameters, including speed, flight direction, cluster size, and heterogeneity, do not affect the algorithm's execution or the final result. However, since this scenario involves moving targets, it is recommended that the UAV's flight speed be greater than the target's speed to detect the target more quickly. In addition, the search area should be larger than the target's interference range to ensure that the estimated information of this parameter is used in decision-making. The estimation algorithm involved in step 3 is not within the scope of this patent discussion; the estimation method can be selected according to the actual situation. The method for updating the UAV coordinates in step 8 does not require actual calculation. It should be combined with the flight control method of the specifically deployed UAVs. It only needs to obtain the current UAV's coordinates, flight speed, and flight direction through external sensors. The information of other UAVs in the cluster obtained through communication in step 11 only includes information without delay at the current moment: if the local UAV is not disturbed, it will receive information from other UAVs in the cluster that are not disturbed; if the local UAV is disturbed, it will not receive information from any other UAVs. In the implementation of update ⑤, due to the possibility of multiple interconnected threat zones, fragmented single threat zones, and multiple targets within a single threat zone, the method for eliminating high-threat zones can be flexibly chosen. Step 14 in the algorithm uses the method of resetting the high-threat connected region where the target is located. The windowing method in step 18 is not limited in the algorithm and can be selected as appropriate. The update methods of updates ②-④ can calculate the local update template with the UAV as the center based on the estimated parameter values. When calling the update, the distribution map is updated by rotating and converting the coordinates based on the current flight direction of the UAV. An example of the distribution map update algorithm for steps 11 to 13 is shown in Algorithm 2 below.

[0048]

[0049] In Algorithm 1, steps 5 and 11 correspond to the undisturbed update rule; steps 6 and 12 correspond to the disturbed state change update method; steps 7 and 13 correspond to the sensor update method; step 10 corresponds to the natural update rule; step 14 corresponds to the high-threat area reset rule; steps 16, 17, and 19 correspond to the distribution vector calculation method; and step 18 corresponds to the proximity penalty method.

Claims

1. A UAV swarm cooperative search algorithm based on probability distribution map updating rules under communication interference, characterized in that, The application relates to a probability distribution map updating rule and a probability distribution map-based guidance algorithm. The probability distribution map updating rule is used for locally integrating current and historical state information of a UAV cluster. The probability distribution map-based guidance algorithm is used for calculating a UAV guidance direction according to local probability distribution map information of the UAV.

2. The UAV swarm cooperative search algorithm based on probability distribution map updating rules under communication interference according to claim 1, characterized in that, The probability distribution map updating rule comprises: a natural updating rule used for expanding a moving target candidate position region; an undisturbed updating rule used for determining a low-threat region by using a target interference range estimation value under communication interference; a disturbed state change updating rule used for determining a high-threat region by using the target interference range estimation value under the communication interference; a sensor updating rule used for determining a safe region by using an airborne sensor sensing range; a high-threat region resetting rule used for resetting a high-threat region where a target is located as a pending state after the target is found.

3. The UAV swarm cooperative search algorithm based on probability distribution map updating rules under communication interference according to claim 1, characterized in that, The probability distribution map-based guidance algorithm comprises: a distribution vector calculation module used for converting the local probability distribution map of the UAV into a benefit vector indexed by a direction, so as to provide a basis for guidance; a near punishment module used for planning trajectories between adjacent UAVs, and guiding the UAVs to avoid repeated search.

Images