Robustness analysis method of cluster node level unmanned aerial vehicle based on regression prediction

Abstract

The invention discloses a cluster node level unmanned aerial vehicle robustness analysis method based on regression prediction, and the method comprises the steps: 1), a heterogeneous unmanned aerial vehicle cluster executes a corresponding subject in a set map environment of each subject, and records a data information log of each unmanned aerial vehicle under each frame of operation of each subject; setting a difficulty calculation method of the random environment for the unmanned aerial vehicle node under each subject and a performance calculation method of the unmanned aerial vehicle node; 2) based on the data information log of each unmanned aerial vehicle, identifying the relationship between nodes in the heterogeneous unmanned aerial vehicle cluster; 3) constructing a node capability network of the heterogeneous unmanned aerial vehicle cluster according to the relationship between nodes in each frame; 4) calculating the difficulty score of each unmanned aerial vehicle node under the corresponding subject and the performance score of each unmanned aerial vehicle node; performing function regression prediction on the unmanned aerial vehicle node based on the difficulty score and the performance score; and 5) evaluating the robustness of the heterogeneous unmanned aerial vehicle cluster according to the stationarity of the regression prediction function of each unmanned aerial vehicle node.

Description

Technical Field

[0001] The present invention belongs to the field of control engineering and simulation technology, and relates to simulation and robustness analysis technology of heterogeneous cluster systems, and in particular to a robustness analysis method of cluster node-level UAVs based on regression prediction. Background Art

[0002] Cluster systems are already widely used in fields such as disaster relief and environmental monitoring. Currently, cluster systems are trending towards large-scale, heterogeneous, and intelligent systems, enabling hundreds or even thousands of heterogeneous agents to collaborate and complete complex tasks. However, existing traditional evaluation frameworks and models fail to incorporate the characteristics of heterogeneous clusters and lack the ability to assess their intelligence. Furthermore, these methods primarily focus on the external performance of cluster algorithms, without deeply analyzing the internal intelligence mechanisms of the algorithms to identify the drivers of performance.

[0003] Currently, in the field of UAV robustness analysis, there are multiple solutions applied in industry and scientific research. Most common practices simulate error behavior by artificially imposing different types of attacks on the algorithm, such as randomly deleting UAV nodes, modifying cluster algorithm parameters, artificially increasing node load, etc., which actually changes the external and internal environment in which the cluster algorithm runs to observe the impact on the final operation results of the algorithm. Existing technologies mainly rely on manually setting attack types to measure the robustness of the cluster. However, the efficiency of manually setting attack types and attack-related parameters is low, and the difficulty of combining multiple attacks on the cluster is ignored during the setting process. Therefore, there is room for improvement in the existing technology in terms of robustness difficulty calculation. Traditional UAV robustness analysis is as follows:

[0004] 1) Mathematical Modeling and Simulation: Using Newton-Euler law and the Euler-Lagrange method, we derive the differential equations for drone dynamics and establish a mathematical model of the drone. Using simulation tools such as Matlab and Simulink, we build a simulation model of the drone system and conduct simulation experiments in various scenarios to evaluate the robustness of the drone.

[0005] 2) Control Algorithm Design: Control algorithms can be broadly categorized into linear control, nonlinear control, and intelligent control. Linear Control: Design linear controllers, such as linear quadratic regulators (LQRs), for attitude stabilization and trajectory tracking. Nonlinear Control: Employ nonlinear control techniques such as auto-disturbance rejection algorithms and feedback linearization controllers to improve the robustness of UAVs in complex environments. Intelligent Control: Leverage intelligent control methods such as deep learning and neural networks to optimize UAV control strategies and enhance their adaptability and robustness.

[0006] 3) Communication and Navigation Systems: Advanced communication technologies, such as the UAV swarm communication system, are used to ensure smooth communication between UAVs and ground control stations, improving the stability and robustness of the communication network. GPS and inertial navigation systems (INS) are used to provide UAVs with accurate navigation information, ensuring flight safety in complex environments.

[0007] 4) Fault Detection and Diagnosis: Sensors monitor the drone's operating status to detect and diagnose faults promptly. Design fault-tolerant control algorithms to enable the drone to continue its mission or land safely in the event of a fault.

[0008] 5) Environmental and Mission Adaptability Analysis: Analyze the UAV's adaptability to specific environments and missions. This includes meteorological conditions: analyzing the UAV's flight performance under different weather conditions, such as strong winds, heavy rain, and other extreme weather conditions; and terrain: evaluating the UAV's flight capabilities in different terrains, such as mountainous areas and urban areas.

[0009] The current technical solution closest to the present invention focuses on analyzing the robustness of the path planning algorithm for the entire cluster. This approach is feasible for analyzing the robustness of a single algorithm for the entire cluster, but it cannot be applied to robustness analysis at the node level targeting overall task performance. Specifically, this solution has the following drawbacks:

[0010] 1) Focusing on the robustness analysis of certain aspects of the cluster and ignoring the overall task: When conducting cluster robustness analysis, excessive focus on certain specific aspects of the cluster may lead to neglect of the overall task.

[0011] 2) Focusing on the robustness analysis of the entire cluster while ignoring individual nodes: In cluster robustness analysis, excessive focus on overall stability and anti-interference capabilities can lead to neglecting detailed examinations of individual nodes. This bias can lead to analysis results that fail to accurately reflect the vulnerability of individual nodes, especially in the face of specific attacks or failures, where the failure of a critical node can have a significant impact on the entire cluster system.

[0012] Existing cluster robustness analysis technologies are relatively mature in some aspects, but they still have certain limitations in terms of overall task-oriented robustness. The main defects are as follows:

[0013] 1. Lack of efficient representation of relationships between cluster nodes.

[0014] Existing solutions do not pay attention to the relationships between cluster nodes, which may lead to unbalanced task allocation among cluster nodes, poor robustness and other adverse consequences.

[0015] 2. Lack of effective modeling methods for heterogeneous cluster capability networks.

[0016] Existing solutions and technologies use random or scale-free networks to model cluster networks, lacking effective representation methods for heterogeneous cluster networks. This means that when processing and analyzing heterogeneous cluster networks, existing models and algorithms may not accurately capture and describe the network's structure and dynamic characteristics. This can lead to errors in simulating, predicting, and optimizing cluster behavior, compromising the accuracy and efficiency of decision-making. Specifically, this deficiency can manifest itself in difficulties in capturing the internal complexity of clusters, leading to efficiency issues and a lack of dynamics.

[0017] 3. Focusing on the robustness analysis of the cluster as a whole in specific aspects, it neither pays attention to the task performance of the cluster nor ignores the performance of individual nodes.

[0018] Existing robustness analysis techniques for drone swarms primarily focus on specific aspects of the swarm, such as its communication and path planning algorithms. These techniques design robustness analysis methods specifically targeting these specific aspects, lacking robustness analysis methods for swarm mission performance. Focusing solely on the durability or communication capabilities of individual drones while ignoring their roles and interactions in large-scale collaborative missions can lead to inaccurate predictions of swarm performance in practical applications. Furthermore, existing robustness analysis techniques for drone swarms focus on the swarm as a whole, ignoring the robustness of individual nodes. This bias can result in analysis results that fail to accurately reflect the vulnerability of individual nodes, particularly in the face of specific attacks or failures, where the failure of critical nodes can significantly impact the entire swarm system.

[0019] 4. Failure to fully analyze the massive data from large-scale heterogeneous drone swarms

[0020] Existing robustness analysis techniques for drone swarms are designed for low- to medium-sized swarm systems. When the number of nodes increases to a larger scale (e.g., a thousand drones), such analysis techniques struggle to capture the relationships, patterns, and regularities among the massive amounts of data generated by large-scale drone swarms during mission execution. Summary of the Invention

[0021] To address the challenges of the existing technology, the present invention aims to provide a robustness analysis method for swarm node-level drones based on regression prediction. This method leverages data mining techniques to fully utilize and process massive amounts of data. Using regression prediction methods, the method fits the "difficulty-performance" function of cluster nodes and performs a stationary analysis on this function, resulting in node-level robustness analysis results. This method can identify potential drone issues and improve their robustness. The present invention first establishes a heterogeneous swarm capability network. Based on this network, mathematical modeling, numerical calculations, and simulation experiments are performed to analyze the difficulty and performance of node-level drones, enabling node-level robustness analysis of the swarm system. By studying the performance of different nodes in the face of emergencies, it reveals which nodes or combinations exhibit strong stability and adaptability. Analyzing the robustness of a swarm helps understand its stability under different conditions and provides guidance for enhancing the system's anti-interference capabilities and optimizing its adaptability. The present invention focuses on establishing a robustness analysis model for drones and calculating robustness numerical results, thereby providing guidance for future optimization of the drones and verifying their effectiveness.

[0022] The technical solution of the present invention is:

[0023] A robustness analysis method for cluster node-level UAVs based on regression prediction, comprising the following steps:

[0024] 1) The heterogeneous drone cluster executes the corresponding subjects in the set map environment of each subject, and records the data information log of each drone in each frame of each subject. The difficulty calculation method for the random environment under each subject and the performance calculation method of the drone node are set;

[0025] 2) Based on the data information logs of each drone, identify the relationship between nodes in the heterogeneous drone cluster; wherein the nodes include drone nodes and scene nodes, each drone corresponds to a drone node, and each obstacle or target unit in the environment corresponds to a scene node;

[0026] 3) Based on the relationship between nodes in each frame, a node capability network of a heterogeneous UAV cluster is constructed;

[0027] 4) Based on the difficulty calculation method for drone nodes in each subject and the performance calculation method for drone nodes, calculate the difficulty score and performance score of each drone node in the corresponding subject; use the difficulty score of the drone node as the independent variable and the performance score as the dependent variable, and perform function regression prediction on the drone node based on the difficulty score and performance score of the drone node in each subject;

[0028] 5) Evaluate the robustness of the heterogeneous UAV cluster based on the stationarity of the regression prediction function of each UAV node.

[0029] Furthermore, the relationship includes patrol relationship, avoidance relationship, collision relationship, coverage relationship and coordination relationship.

[0030] Furthermore, the subject is collaborative security patrol, and the drone nodes include small patrol aircraft nodes and large patrol aircraft nodes;

[0031] 1) For the i-th small patrol aircraft node, its patrol area is area A, and the difficulty score under the collaborative security patrol is x i =x i,intensity +x i,dispersion ; where x i,intensity Indicates the density of warning areas and urban facilities in area A, x i,dispersion Indicates the degree of dispersion of the locations of urban facilities in Area A; performance score under coordinated security patrols Among them, n represents the total number of urban facilities in area A that the patrol aircraft successfully patrols, CityBuilding j represents the type of urban facilities patrolled during the jth patrol; weight represents the weighted reward points corresponding to the urban facilities; It represents the time reward score of the i-th small patrol aircraft node for the j-th patrol to the urban facilities;

[0032] 2) For large patrol aircraft nodes, their patrol area is Area B. The difficulty score under collaborative security patrol is the density of the guarded area and patrol targets in Area B. The performance under collaborative security patrol is the sum of the bonus points for the number of patrol targets in Area B that are successfully patrolled and the total time to complete the task.

[0033] Further, n represents the total number of urban facilities in area A, m represents the total number of warning zones in area A, distance_alert(j,k) represents the distance between the jth urban facility and the kth warning zone, min k＝1,2,…,m (distance(j,k)) represents the distance to the nearest warning zone of the jth urban facility; distance_building(j,k) represents the distance between the jth city facility and the kth city facility; t j It represents the time when the i-th small patrol aircraft node patrols the city facilities for the jth time.

[0034] Furthermore, the cubic function is selected as the objective function of regression prediction to perform function regression prediction on the drone nodes.

[0035] Furthermore, the method for function regression prediction of drone nodes is:

[0036] 1) The data points of each subject of each drone node (x p ,y p ) is normalized to the interval [0,100]; x p is the difficulty score of the p-th subject simulation of the drone node, y p is the performance score of the drone node in the pth subject simulation;

[0037] 2) The coefficients a, b, c, and d of the cubic function are randomly initialized within their respective value ranges;

[0038] 3) Randomly select a data point (x p ,y p ), change x p Substituting into the expression, we get

[0039] 4) Calculate the loss function that minimizes the residual sum of squares

[0040] 5. Calculate the partial derivatives of the loss function with respect to a, b, c, and d respectively

[0041] 6) Use gradients to update parameters Where α is the learning rate hyperparameter, which determines the step size of each update;

[0042] 7) If the stopping condition is reached, the iteration ends and the current a, b, c, d are output as the fitting result a * 、b * 、c * d * , and obtain the regression prediction function y=a corresponding to the UAV * x 3 +b * x 2 +c * x+d * , a * ≠0; otherwise return to step 3).

[0043] Furthermore, the method for calculating the stationarity of the regression prediction function of each drone node is as follows: first calculate the regression prediction function y=a of the drone node * x 3 +b * x 2 +c * x+d * The derivative y′=3a * x 2 +2b * x+c *, determine the maximum and average offset of y' relative to 0 in the domain; if the offset of y' relative to 0 is very small, it means that the change of y in the entire domain is very small; the maximum and average offsets are respectively the maximum value y' of the absolute value of y' in the domain max and the mean value y′ mean To express; then introduce the variance of n test regression prediction function y As another indicator to measure the degree of data dispersion, the robustness value of the drone node is obtained where y′ max It is the maximum value of the absolute value of the derivative of the regression prediction function y in the domain [lower_bound, upper_bound], y′ mean It is the average value of the absolute value of the derivative of the regression prediction function y integrated over the domain [lower_bound, upper_bound].

[0044] Furthermore, the subjects include coordinated patrol missions in land, sea and airspace, coordinated monitoring missions in land, sea and airspace, and coordinated rescue missions in land, sea and airspace.

[0045] A server, characterized in that it includes a memory and a processor, the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program includes instructions for executing the above method.

[0046] A computer-readable storage medium stores a computer program thereon, wherein the computer program implements the above method when executed by a processor.

[0047] The advantages of the present invention are as follows:

[0048] 1. This paper proposes a series of calculation methods for the relationships between cluster nodes, and clearly defines the types, quantification, and calculation methods of inter-node relationships. The inter-cluster node relationship refers to the interaction and collaboration mode between each node in the UAV cluster system, which reflects how each node shares information with other nodes in the cluster, collaborates to perform tasks, coordinates actions, etc. Cluster node relationships can be established through various channels such as communication, perception, and collaboration, and have a profound impact on the task execution efficiency, stability, and intelligence level of the entire cluster system.

[0049] 2. The present invention proposes an innovative method for constructing a heterogeneous cluster capability network, which can efficiently describe the collaboration, task allocation, and interdependence between nodes in the cluster, and can help optimize task planning, improve system efficiency, realize intelligent collaboration of drone clusters, and analyze cluster capabilities. A heterogeneous cluster capability network refers to a network model that reflects the capability relationship between nodes in a drone cluster system based on the different capabilities and attributes of each drone node; the network model uses drones as nodes, the capabilities of drones as node attributes, and the relationships between drones as edges. By connecting each node and its capabilities, the network model can effectively describe the collaboration, task allocation, and interdependence between nodes in the cluster; constructing this network can help optimize task planning, improve system efficiency, realize intelligent collaboration of drone clusters, and analyze cluster capabilities.

[0050] 3. This paper proposes a robustness analysis method for cluster node-level drones based on task performance. This method generates task-oriented robustness analysis results for each drone in the cluster. This analysis can help identify nodes with poor robustness in the cluster and guide cluster algorithm developers in targeted optimization of these nodes. The overall cluster mission objective is broken down to derive the subtasks required of each drone. Based on the cluster capability network, the attributes and relationships of node-level drones are extracted, and the completion of each drone's subtasks is evaluated and calculated. Based on this evaluation result, robustness analysis of the node-level drones is performed. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 This is a flowchart of the robustness analysis of large-scale heterogeneous cluster nodes based on data mining in the present invention.

[0052] Figure 2 Schematic diagram of a large-scale heterogeneous cluster node capability network.

[0053] Figure 3 A map for city patrol subjects.

[0054] Figure 4 Flowchart for function regression prediction.

[0055] Figure 5 This is the regression function fitting diagram of the patrol aircraft node with the highest robustness.

[0056] Figure 6 This is the regression function fitting diagram of the 10 patrol aircraft nodes with the highest robustness. DETAILED DESCRIPTION

[0057] The present invention will be described in further detail below with reference to the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0058] The present invention is a heterogeneous cluster node-level robustness analysis technology based on data mining, which mainly includes the following steps: "large-scale heterogeneous cluster simulation experiment", "definition, quantification and judgment of the relationship between cluster nodes", "construction of large-scale heterogeneous cluster node capability network", "cluster node-level robustness analysis based on regression prediction" and "validity verification". The specific process is as follows Figure 1 shown.

[0059] Step 1: Heterogeneous cluster simulation test

[0060] The present invention is a robustness analysis of the simulation test results of heterogeneous clusters, so it is necessary to first conduct a large-scale heterogeneous cluster system simulation test. Step 1 includes simulation tests set for each subject, including three tasks: collaborative patrol tasks, collaborative monitoring tasks, and collaborative rescue tasks in land, sea, and airspace, for a total of 9 subjects. Specifically, in the Unreal Engine 4 engine, based on the developed subject map environment, the heterogeneous cluster is allowed to perform specific subject tasks in this environment, record and organize the data generated by the cluster in each frame during the entire process of subject operation, and obtain the data information log of each drone in each frame.

[0061] Step 2: Define, quantify, and judge the relationships between cluster nodes

[0062] Based on the simulation test data log obtained in step 1, the data is processed and analyzed for determining the relationship between cluster nodes. Prior to this, the present invention first clearly defines and quantifies the relationship between cluster nodes.

[0063] Step 2.1 Definition of relationships between cluster nodes

[0064] The relationship between swarm nodes refers to the interaction and collaboration between individual nodes in a swarm drone system. It reflects how each node shares information, allocates tasks, and coordinates with other nodes within the swarm. Relationships between swarm nodes can be established through various channels, including communication, perception, and collaboration, and have a profound impact on the mission execution efficiency, stability, and intelligence level of the entire swarm system. Determining the relationships between drones in a swarm typically involves three steps: relationship definition, relationship quantification, and relationship judgment. These steps help accurately identify and measure interactions between nodes, enabling better coordination and optimization.

[0065] A relationship definition is a clear description of the interaction pattern between drone nodes and scene nodes in a swarm mission, indicating under what conditions a specific relationship exists between the two. This definition provides a rule-based basis for subsequent quantification and judgment. Common drone cluster node relationships include:

[0066] (1) Patrol Relationship: When a drone has relevant sensory sensors and uses its sensory capabilities to patrol and collect information about a target unit, it is considered to have a patrol relationship with the target unit. All drones have sensory sensors, and when the distance between them and the target unit is less than a set threshold, they are set to a patrol relationship.

[0067] (2) Avoidance relationship: During flight, when a drone detects that it is too close to another aircraft or obstacle and takes evasive measures to avoid collision, it is considered that the drone has an avoidance relationship with the aircraft or obstacle.

[0068] (3) Collision relationship: When a UAV comes into contact with another aircraft or obstacle, resulting in physical damage or failure, it is considered that the UAV is in a collision relationship with the target.

[0069] (4) Coverage relationship: When a drone’s detection range can continuously cover a specific area or target unit and provide real-time data support during the execution of a mission, it is considered that the drone has a coverage relationship with the area or target unit.

[0070] (5) Collaborative relationship: When multiple UAVs perform the same task and improve the overall task efficiency through information sharing or task allocation, it is considered that there is a collaborative relationship between these UAVs.

[0071] The above relationship definitions help to more accurately describe, manage, and understand the interaction between nodes in a complex drone swarm system, thereby optimizing task execution and system safety. The relationships between nodes are quantified based on step 2.2.

[0072] Step 2.2 Quantification of relationships between cluster nodes

[0073] Relationship quantification involves parameterizing the fuzzy concepts described in the relationship definition (such as "nearby," "collaboration," and "communication") with specific numerical parameters. For example, the "follow" relationship can be quantitatively described by the relative distance and duration between drones. This step provides clear quantitative criteria for the relationship definition, enabling the algorithm to make more effective judgments. The following is an example of relationship quantification for the relationship definition mentioned in step 2.1:

[0074] (1) Patrol relationship: When a UAV equipped with a sensing sensor patrols a target (building, ship, facility, etc.), and the detection range of the UAV covers a certain range of the target (e.g., more than 80%), the UAV is considered to have a patrol relationship with the target.

[0075] (2) Avoidance relationship: When a UAV detects that the distance to a scene node (other aircraft or obstacles) is less than the safe distance (e.g., 100 meters) during flight and takes active evasive measures (e.g., changing flight altitude or direction) to avoid collision, it is considered that there is an avoidance relationship between the UAV and the scene node.

[0076] (3) Collision relationship: When a drone comes into contact with a scene node (other aircraft or obstacles), resulting in physical collision or damage (such as blood loss or direct damage), the drone is considered to have a collision relationship with the scene node.

[0077] (4) Coverage relationship: When a drone is performing a mission and its detection range (such as sensor coverage reaching more than 90%) can continuously cover a specific area or target and provide real-time data support, it is considered that the drone has a coverage relationship with the area or target.

[0078] (5) Collaborative relationship: When multiple UAVs in the same mission effectively improve the overall mission efficiency (such as reducing the mission completion time by 30%) through information sharing or task allocation (such as mutual communication and coordinated actions), it is considered that there is a collaborative relationship between the UAVs.

[0079] These quantitative criteria help establish clear operation and decision-making criteria among cluster nodes, thereby more effectively evaluating the interaction between nodes and task execution.

[0080] Step 2.3 Determine the relationship between cluster nodes

[0081] Relationship determination involves analyzing the interactions between drone nodes in a cluster using algorithms based on defined relationship criteria and quantitative rules to determine whether a relationship exists between them. Common algorithms can be based on time series data, geometric calculations, or machine learning models.

[0082] In order to more clearly describe the node relationship judgment process for heterogeneous clusters, the node set involved in the cluster movement process is recorded as V = V I ∪V E , where V I represents the internal nodes of the drone cluster (i.e., drone nodes), represents the set of drones in the entire drone cluster, and V E Represents external nodes (i.e., scene nodes), which represent external nodes other than drones, such as buildings, pedestrians, enemy units, etc. E is used to represent the node relationship set. represents the i-th drone node at time t, R is the judgment function after the relationship quantification mentioned in step 2.2, r k (·) represents the judgment function after the kth relationship is quantified, r k->{True,False}, when the result is True, it means the relationship exists, and when the result is False, it means the relationship does not exist. In relationship judgment, time series data is often needed as a basis to judge whether there is a certain relationship between two nodes at the current moment. For example, the "leader-follower" relationship requires node data within 10 seconds as a basis for judgment, so use Represents the state of the i-th UAV node in the time interval [t-Δt, t].

[0083] Step 3: Heterogeneous cluster node capability network construction

[0084] The construction of a heterogeneous cluster node capability network primarily quantifies and formalizes the attributes, behaviors, and relationships of the drone nodes in the cluster based on historical data. Drone attributes are derived from historical data (logs) from the simulated environment, including their position, speed, health, and so on, at different time steps. Behavior is determined based on the values ​​of these attributes at different moments. For example, if a drone's speed is high at one moment and low at the next, it can be considered to be decelerating. The construction process begins by extracting entity node attributes from the drone's historical trajectory information frame by frame based on the data logs obtained in step 1, and constructing entity nodes. Then, using the relationship determination method described in step 2, relationships between entity nodes are determined based on the extracted drone entities and quantitative rules. This results in a large-scale heterogeneous cluster node capability network.

[0085] Building a cluster node capability network first requires extracting scene node parameters frame by frame, defining relationship types, and performing relationship quantification or selecting machine learning methods to determine node relationships. Based on this and the content described in step 2.3, the following algorithm is used to determine scene node relationships frame by frame and build a large-scale heterogeneous cluster node capability network:

[0086]

[0087]

[0088] Initialize the relationship set M and the time slice interval size Δt, where M is a two-dimensional list of sets, each of which is initialized to an empty set. Traverse the historical data frame by frame, select the i-th node and the j-th node, and if the two nodes are different, traverse the relationship judgment function list, read the historical data of the two nodes, and use the k-th relationship judgment function to judge. If a relationship is found, it is considered that the i-th node and the j-th node have a relationship, and the relationship is added to the relationship set M[i][j].

[0089] Through Algorithm 1, the cluster log data can be processed and sorted, and the relationship information of the entire cluster can be obtained, that is, the cluster node capability network, such as Figure 2 shown.

[0090] At this point, step 3 completes the construction of the large-scale heterogeneous cluster node capability network, laying the foundation for the subsequent cluster node-level robustness analysis based on data mining.

[0091] Step 4: Cluster node-level robustness analysis based on regression prediction

[0092] Robustness analysis of drone swarms focuses on their ability to maintain functionality and complete tasks in the face of external disturbances, internal failures, or environmental changes. This can be linked to observing how the task completion performance of swarm nodes changes as environmental conditions change. Specifically, for randomly generated external environments, we can calculate the difficulty of completing a specific task for a node in that environment and how well the swarm nodes perform under those conditions. Therefore, the robustness of drones can be modeled using difficulty as the independent variable and performance as the dependent variable. First, we define the difficulty of random environments for drone nodes under different subjects and calculate their performance under those environments. Then, we repeat simulations, generating data points for each drone (difficulty, performance) in each experiment. We then perform regression predictions on the multiple data points obtained from each individual drone across multiple trials. Finally, we calculate the stability of the regression prediction function as a robustness indicator for the swarm. In summary, this project analyzes the robustness of node-level drones based on task completion using regression prediction techniques.

[0093] Before conducting a robustness analysis, we need to make an assumption about the cluster system's task allocation: even if the external environment within the same subject changes slightly, the task allocation within the cluster system for each subject will remain unchanged. For example, in the first simulation of the collaborative security patrol subject, the system assigns patrol aircraft numbered 100 the task of "patrolling the objects to be patrolled within Area A." In the second simulation, the external environment may have slightly changed (such as a change in the type or number of objects to be patrolled within Area A), but the system will still assign this task to the drone. For this non-heuristic, non-random cluster task allocation method, we can perform a robustness analysis on a single drone node.

[0094] Taking the collaborative security patrol subject as an example, the following first analyzes the characteristics of this subject and defines the calculation method for the difficulty and performance of various types of drones in this subject;

[0095] Step 4.1 Analysis of the characteristics of typical subjects

[0096] (1) Course Process Introduction

[0097] In the collaborative security patrol subject, 100 small four-rotor patrol aircraft will first perform regional advance patrol missions, and then 900 large fixed-wing patrol aircraft will perform target approach patrol missions. The targets to be patrolled include various urban facilities, which are divided into static targets and dynamic targets. Static targets include urban buildings such as factories, helipads, and residential buildings; dynamic targets include helicopters, patrol cars, residents, etc. Among them, small patrol aircraft are responsible for patrolling static facilities, and large patrol aircraft are responsible for patrolling dynamic facilities. The initial positions of the targets to be patrolled and the warning areas are randomly generated. The map of this subject is as follows Figure 3 shown.

[0098] (2) Analysis of subject characteristics

[0099] The difficulty of small patrol aircraft in patrol subjects is mainly reflected in the distance between the patrol target and the warning area: the smaller the distance, the more likely the drone will enter the trigger radius of the warning area, resulting in patrol failure, and the more difficult it will be for the drone to patrol the target.

[0100] The performance of large patrol aircraft in patrol subjects is mainly reflected in the number of successful patrols to patrol targets: the more successful patrols, the better the performance of the large patrol aircraft.

[0101] (3) Calculation method of difficulty and performance

[0102] (a) Small patrol aircraft

[0103] Assume that the patrol aircraft node to be analyzed is assigned the task of "patrolling the objects to be patrolled in area A".

[0104] For small patrol aircraft nodes, the difficulty of this subject is defined as the sum of the density of the warning area and the location of urban facilities in area A, and the dispersion of the location of urban facilities: the higher the density of the warning area and urban facilities, that is, the smaller the distance between the warning area and the urban facilities, the higher the difficulty of patrolling a single urban facility; the greater the distance between urban facilities, the more dispersed the urban facilities are, and the higher the difficulty of fully patrolling all urban facilities. For the i-th drone, its difficulty x i The calculation formula is:

[0105] x i =x i,intensity +x i,dispersion

[0106] Among them, x i,intensity Indicates the density of warning areas and urban facilities in area A. The calculation formula is:

[0107]

[0108] n represents the total number of urban facilities in area A of the patrol aircraft, and m represents the total number of warning zones in area A. distance_alert(j,k) represents the distance between the jth urban facility and the kth warning zone, min k＝1,2,…,m (distance(j,k)) represents the distance to the nearest warning zone of the j-th urban facility.

[0109] x i,dispersion Indicates the degree of dispersion of the locations of urban facilities in area A. The calculation formula is:

[0110]

[0111] n represents the total number of urban facilities within the patrol aircraft's area A. distance_building(j,k) represents the distance between the jth urban facility and the kth urban facility.

[0112] For the patrol aircraft node, the performance score y under this subject is defined as i The score y is the sum of the number of urban facilities in area A that are successfully patrolled and the total time to complete the task. i The calculation formula is:

[0113]

[0114] Where n represents the total number of urban facilities in area A that the patrol aircraft successfully patrolled. j Indicates the type of urban facilities patrolled during the jth patrol; weight indicates the weighted reward points corresponding to a certain urban facility, both of which are 25. It represents the time bonus points for the jth patrol to the city facilities, and the calculation formula is:

[0115]

[0116] Among them, t j It represents the time of the jth patrol to the city facility, in seconds.

[0117] (2) Large patrol aircraft

[0118] Assume that the patrol aircraft node to be analyzed is assigned the task of "patrolling the objects to be patrolled in area B".

[0119] For large patrol aircraft nodes, the difficulty of this subject is defined as the density of the guarded area and the patrol target in area B. The higher the density of the guarded area and the patrol target, that is, the smaller the distance between the guarded area and the patrol target, the greater the difficulty of patrolling a single patrol target. i The calculation formula of ′ is:

[0120]

[0121] n' represents the total number of patrol targets in the patrol aircraft's B area, and m' represents the total number of warning zones. distance_alert(j',k') represents the distance between the j'th patrol target and the k'th warning zone, min k′＝1,2,…,m′ (distance(j′,k′)) represents the distance to the nearest warning zone of the j′th patrol target.

[0122] For patrol aircraft nodes, the performance under this subject is defined as the sum of the number of patrol targets in the B area that are successfully patrolled and the total time to complete the task. The score calculation formula is:

[0123]

[0124] Where n′ represents the total number of patrol targets in area B that the patrol aircraft successfully patrolled. i′ Indicates the type of patrol target encountered during the i′th patrol; weight indicates the weighted reward score corresponding to a patrol target, which are 5, 3, and 1 respectively. It represents the time bonus points for the i′th patrol to the city facilities. The calculation formula is:

[0125]

[0126] Where t′ j′ It represents the time of the j′th patrol to the city facilities, in seconds.

[0127] Step 4.2 Robustness analysis method based on regression prediction

[0128] The difficulty and performance scores of the node-level drone to be analyzed are calculated by the above method: For the p-th subject simulation experiment, calculate the difficulty x of this experimental scenario for the drone to be analyzed. p , and the performance score y of the drone to be analyzed in this experiment p , and get a point (x p ,y p ). If q simulation experiments are carried out on the UAV to be analyzed, q points (x p ,y p All the resulting points are plotted on the coordinate axes and a function regression prediction is performed. The fitted function represents how the performance of the drone under analysis changes with the difficulty level, and also implies the robustness level of the drone under analysis. Mathematical analysis of the fitted function can be used as the result of the robustness analysis of the drone under analysis.

[0129] Furthermore, this project only conducts robustness analysis on drone nodes whose scores are non-zero in at least 25% of the total number of tests. For example, suppose a drone has a score non-zero (greater than) zero in n out of 30 tests. If n ≥ 30 × 25% = 7.5, meaning the drone scored greater than zero in at least 8 of the 30 tests, robustness analysis is performed on it. Otherwise, robustness analysis is omitted and the robustness score is assigned to 0. For example, in the security patrol subject, a patrol drone scored 1000 in 4 of the 30 tests, with the remaining 26 scoring zero. This indicates that the robustness of this drone is extremely poor, and the regression prediction result of its "difficulty-performance" function will be a wildly oscillating curve, making it meaningless for analysis. Therefore, it is skipped and the robustness score is assigned to 0.

[0130] In order to balance model accuracy and computational efficiency, the present invention selects the cubic function as the objective function of regression prediction. The mathematical expression of the cubic function is:

[0131] y=ax 3 +bx 2 +cx+d(a≠0)

[0132] Where x and y are the independent and dependent variables, respectively. In this project, they represent the difficulty of the environment for the drone and the performance of the drone in this environment; a, b, c, and d are model parameters.

[0133] The specific function regression steps are shown in Figure 4 , the specific description is as follows:

[0134] ① Data preprocessing: All data points (x p ,y p ) is normalized to the interval [0,100]. Specifically, for a certain type of drone in a certain subject, the upper and lower bounds of the difficulty of this drone in all simulation experiments, as well as the upper and lower bounds of the performance, are counted, and the difficulty and performance are normalized according to the two sets of upper and lower bounds. For example, for a large patrol aircraft in the security patrol subject, 900 patrol aircraft are counted in n tests, with a total of 900*n data points (x p ,y p ), calculate the upper and lower bounds of patrol aircraft difficulty

[0135]

[0136]

[0137] and the upper and lower bounds of patrol aircraft performance

[0138]

[0139] Then, the difficulty and performance are normalized to the interval [0,100] according to the two sets of upper and lower bounds.

[0140]

[0141] ② Random initialization: a, b, c, and d are randomly initialized within their respective value ranges.

[0142] ③Data extraction: Randomly extract a data point (x p ,y p ), change x p Substituting into the expression, we get

[0143]

[0144] ④Calculate the loss: Calculate the minimum residual sum of squares (RSS) loss function, that is,

[0145] ⑤Calculate partial derivatives: Calculate the partial derivatives of the loss function with respect to a, b, c, and d respectively

[0146] ⑥ Gradient update: Use gradient to update parameters: Where α is the learning rate hyperparameter, which determines the step size of each update.

[0147] ⑦ Iteration judgment: If the stopping condition is reached (the maximum number of iterations is reached, or the loss is less than a certain threshold), the iteration ends and the current a, b, c, d are output as the fitting result a * 、b * 、c * d * ; otherwise return ③.

[0148] Step 4.3 Function stationarity calculation

[0149] The regression results of the "difficulty-performance" function of the drone to be analyzed are obtained through step 4.2:

[0150] y=a * x 3 +b * x 2 +c * x+d * (a * ≠0)

[0151] The stability of the "difficulty-performance" function corresponding to the analyzed drone is calculated as the robustness analysis result of the drone. In order to find the degree of change in the performance of the regression function within the definition domain, the derivative y′=3a of the regression function y of the drone is calculated. * x 2 +2b * x+c * , determine the maximum and average offset of y' relative to 0 in the domain. If the offset of y' relative to 0 is very small, it means that the change of y in the entire domain is very small. The maximum and average offsets are respectively the maximum value y' of the absolute value of y' in the domain. max and the mean value y′ mean To express.

[0152] Then introduce the variance of n test y As another indicator to measure the degree of data dispersion, the calculation method of the robustness value of the drone is obtained.

[0153]

[0154] where y′ max It is the maximum value of the absolute value of the derivative in the domain [lower_bound, upper_bound], where lower_bound represents the minimum value of the derivative domain and upper_bound represents the maximum value of the derivative domain. max The calculation formula is:

[0155] y′ max =max x∈[lower_bound,upper_bound] |y′|

[0156] y′ mean It is the average value of the absolute value of the derivative integrated over the domain [lower_bound, upper_bound], and its calculation formula is:

[0157]

[0158] Step 5: Validity Verification

[0159] Through steps 1-4, this paper has introduced a specific robustness analysis method. Step 5 verifies the effectiveness of this method. Specifically, using a developed clustering algorithm, we conduct a node-level robustness analysis of the cluster's performance in completing tasks in the security patrol subject. The following is a presentation of the analysis results.

[0160] (1) Among the 100 small patrol aircraft, the patrol aircraft node numbered 903 has the highest robustness. The variance of its performance score in 30 rounds of testing is 38.71. Its corresponding regression function is as follows: Figure 5 The derivative of this function is at x=4.51×10 -6 The maximum value y′ is reached at max =0.0677. Its average value y′ over the entire domain mean =0.0344.

[0161] (2) Among the 900 large patrol aircraft, the patrol aircraft node numbered 637 has the highest robustness. The variance of its performance score in 30 rounds of testing is variance = 68.68. Its corresponding regression function is as follows: Figure 6 The derivative of this function reaches its maximum value y′ at x=40.9169 max =0.0313. Its average value y′ over the entire domain mean =0.002.

[0162] (3) The regression functions corresponding to the 10 patrol aircraft nodes with the highest robustness are shown in Figure 6 .

[0163] It can be seen that the function corresponding to each node has extremely low fluctuations within the domain of definition, which means that the function has a very small change range within the domain of definition. This also means that the performance of the drone node changes very little as the difficulty changes, that is, these drone nodes have good robustness.

[0164] (4) Through the above series of pictures, the robustness of each drone node has been quantified and visualized during the execution of the subject. The function graph reflects the trend of drone nodes changing with the difficulty, and the bar chart further reveals the 30 specific drone nodes with the highest robustness. Robustness analysis not only helps us understand the main nodes and secondary nodes in task execution, but also provides a strong basis for the optimal configuration of the drone group. Subsequent work can be based on the results of this robustness analysis to optimize the drone's task allocation strategy and other aspects, thereby improving the robustness of the cluster system and drone nodes, and further improving the overall system performance. For example, the following aspects can be considered:

[0165] (a) Task allocation

[0166] Based on the robustness of each drone, more difficult tasks are prioritized for drones with higher robustness, reducing the risk of critical tasks being impacted by drone failures. For example, in the drone swarm system used by the algorithm, patrol aircraft numbered 903 and other more robust patrol aircraft can be assigned to patrol larger areas or more targets.

[0167] (b) Path planning

[0168] When planning paths, the robustness of each drone can be considered. Less robust drones can be assigned more conservative paths to avoid high-risk areas, while more robust drones can be assigned to more complex or risky tasks. For example, in the drone swarm system used by the algorithm, patrol drones numbered 903 and other more robust drones can be assigned to patrol areas with more dispersed urban infrastructure, as these areas are more challenging.

[0169] (c) Health management

[0170] The robustness of drones can be used to implement more refined health management strategies. For drones with lower robustness, the frequency of maintenance and inspections can be increased to prevent potential failures. For example, in the drone swarm system used by the algorithm, the maintenance frequency of less robust drones can be increased, while the frequency of patrol aircraft numbered 903 and other more robust drones can be reduced. This can reduce maintenance costs and improve maintenance efficiency.

[0171] While specific embodiments of the present invention have been disclosed for illustrative purposes, intended to facilitate understanding and implementation of the present invention, those skilled in the art will appreciate that various substitutions, variations, and modifications are possible without departing from the spirit and scope of the present invention and the appended claims. Therefore, the present invention should not be limited to the disclosure of the preferred embodiments, and the scope of protection claimed in the present invention shall be determined by the scope of the claims.

Claims

1. A robustness analysis method for cluster node-level UAVs based on regression prediction, comprising the following steps: 1) The heterogeneous drone cluster executes the corresponding subjects in the set map environment of each subject, and records the data information log of each drone in each frame of each subject; Set the difficulty calculation method for drone nodes in the random environment under each subject and the performance calculation method for drone nodes; 2) Based on the data information logs of each drone, identify the relationship between nodes in the heterogeneous drone cluster; wherein the nodes include drone nodes and scene nodes, each drone corresponds to a drone node, and each obstacle or target unit in the environment corresponds to a scene node; 3) Based on the relationship between nodes in each frame, a node capability network of a heterogeneous UAV cluster is constructed; 4) Based on the difficulty calculation method for drone nodes in each subject and the performance calculation method for drone nodes, calculate the difficulty score and performance score of each drone node in the corresponding subject; use the difficulty score of the drone node as the independent variable and the performance score as the dependent variable, and perform function regression prediction on the drone node based on the difficulty score and performance score of the drone node in each subject; 5) Evaluate the robustness of the heterogeneous UAV cluster based on the stationarity of the regression prediction function of each UAV node.

2. The method according to claim 1, characterized in that The relationships include patrol relationships, avoidance relationships, collision relationships, coverage relationships and coordination relationships.

3. The method according to claim 1, characterized in that The subject is collaborative security patrol, and the drone nodes include small patrol aircraft nodes and large patrol aircraft nodes; 1) For the i-th small patrol aircraft node, its patrol area is area A, and the difficulty score under the collaborative security patrol is x i =x i,intensity +x i,dispersion ; where x i,intensity Indicates the density of warning areas and urban facilities in area A, x i,dispersion Indicates the degree of dispersion of the locations of urban facilities in Area A; performance score under coordinated security patrols Among them, n represents the total number of urban facilities in area A that the patrol aircraft successfully patrols, CityBuilding j represents the type of urban facilities patrolled during the jth patrol; weight represents the weighted reward points corresponding to the urban facilities; It represents the time reward score of the i-th small patrol aircraft node for the j-th patrol to the urban facilities; 2) For large patrol aircraft nodes, their patrol area is Area B. The difficulty score under collaborative security patrol is the density of the guarded area and patrol targets in Area B. The performance under collaborative security patrol is the sum of the bonus points for the number of patrol targets in Area B that are successfully patrolled and the total time to complete the task.

4. The method according to claim 3, characterized in that n represents the total number of urban facilities in area A, m represents the total number of warning zones in area A, distance_alert(j,k) represents the distance between the jth urban facility and the kth warning zone, min k＝1,2,…,m (distance(j,k)) represents the distance to the nearest warning zone of the jth urban facility; distance_building(j,k) represents the distance between the jth city facility and the kth city facility; t j It represents the time when the i-th small patrol aircraft node patrols the city facilities for the jth time.

5. The method according to claim 1, wherein The cubic function is selected as the objective function of regression prediction to perform function regression prediction on the drone nodes.

6. The method according to claim 5, characterized in that The method for function regression prediction of drone nodes is: 1) The data points of each subject of each drone node (x p ,y p ) is normalized to the interval [0,100]; x p is the difficulty score of the p-th subject simulation of the drone node, y p is the performance score of the drone node in the pth subject simulation; 2) The coefficients a, b, c, and d of the cubic function are randomly initialized within their respective value ranges; 3) Randomly select a data point (x p ,y p ), change x p Substituting into the expression, we get 4) Calculate the loss function that minimizes the residual sum of squares 5) Calculate the partial derivatives of the loss function with respect to a, b, c, and d respectively 6) Use gradients to update parameters Where α is the learning rate hyperparameter, which determines the step size of each update; 7) If the stopping condition is reached, the iteration ends and the current a, b, c, d are output as the fitting result a * 、b * 、c * d * , and obtain the regression prediction function y=a corresponding to the UAV * x 3 +b * x 2 +c * x+d * , a * ≠0; otherwise return to step 3).

7. The method according to claim 6, characterized in that The method for calculating the stationarity of the regression prediction function of each drone node is as follows: first calculate the regression prediction function y = a of the drone node * x 3 +b * x 2 +c * x+d * The derivative y ′ =3a * x 2 +2b * x+c * , judge y ′ The maximum and average offsets relative to 0 in the domain; if y ′ The offset relative to 0 is very small, which means that the change of y in the entire domain is very small; the maximum offset and the average offset are expressed as y ′ The maximum value y′ of the absolute value in the domain of definition max and the mean value y′ mean To express; then introduce the variance of n test regression prediction function y As another indicator to measure the degree of data dispersion, the robustness value of the drone node is obtained where y′ max It is the maximum value of the absolute value of the derivative of the regression prediction function y in the domain [lower_bound, upper_bound], y′ mean It is the average value of the absolute value of the derivative of the regression prediction function y integrated over the domain [lower_bound, upper_bound].

8. The method according to claim 1, characterized in that The subjects include coordinated patrol missions in land, sea and airspace, coordinated monitoring missions in land, sea and airspace, and coordinated rescue missions in land, sea and airspace.

9. A server, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program, the computer program is configured to be executed by the processor, and the computer program comprises instructions for executing the method according to any one of claims 1 to 8.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the method according to any one of claims 1 to 8 is implemented.

Images