A cooperative control system and method for unmanned swarms based on multi-source heterogeneous data
By real-time monitoring of computing power levels through built-in benchmark testing units in unmanned cluster nodes, segmenting atomic control primitives and constructing bipartite graphs for node matching, and using high-computing-power nodes to compensate for low-computing-power nodes, the computational heterogeneity problem in multi-source heterogeneous unmanned cluster systems is solved, achieving efficient multi-source heterogeneous data collaborative control and improving cluster performance and stability.
Patent Information
- Application Number
- CN202611462462.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-20
- Publication Date
- 2026-11-17
AI Technical Summary
In existing technologies, multi-source heterogeneous unmanned swarm systems have significant differences in sensor configuration, onboard computing power, communication capabilities and energy supply, resulting in computing heterogeneity problems. Existing cooperative control algorithms have failed to effectively achieve the coordination of multi-source heterogeneous data and have failed to unify decision-making on task requirements, environmental information and underlying resource scheduling.
By real-time monitoring of instantaneous computing power levels through built-in benchmark testing units in unmanned nodes, atomic control primitives are divided using computing power dispersion, a bipartite graph is constructed for node matching, different sampling strategies are executed, and virtual control quantities generated by high-computing-power nodes are used to compensate low-computing-power nodes, establishing a collaborative control architecture with a closed-loop chain, including modules for computing power analysis, task division, allocation, compensation, and balance judgment.
It enables effective collaboration of multi-source heterogeneous data under conditions of extremely asymmetrical computing power, improves the overall cluster performance, avoids resource mismatch and latency, and ensures system stability and security.
Smart Images

Figure CN123113585A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of unmanned aerial vehicle (UAV) collaborative control technology, specifically to an unmanned aerial vehicle (UAV) swarm collaborative control system and method based on multi-source heterogeneous data. Background Technology
[0002] With the rapid development of unmanned systems technology, unmanned swarm systems, composed of diverse platforms such as drones, unmanned vehicles, unmanned surface vessels, and unmanned underwater vehicles, have shown broad application prospects in military and civilian fields such as military reconnaissance, disaster relief, environmental monitoring, and regional patrol. Compared with a single unmanned platform, unmanned swarms can achieve wider spatial coverage and stronger mission execution capabilities through distributed collaborative operations. However, in practical applications, unmanned swarms are usually composed of heterogeneous unmanned platforms of different types and generations. Each node has significant differences in sensor configuration, onboard computing power, communication bandwidth, and energy supply. Heterogeneity is both the source of the swarm's functional diversity and a serious challenge to the design of the collaborative control system.
[0003] In traditional technologies, different unmanned platforms exhibit significant differences in sensor configuration, onboard computing power, communication capabilities, and energy supply. Traditional cooperative control algorithms often assume that the computing power of each node is equal. However, in reality, a high-end drone may be equipped with a GPU for real-time inference, while another inexpensive drone may only have microcontroller-level computing power. This "computational heterogeneity" is particularly pronounced in multi-source data fusion scenarios, where high-computing-power nodes can handle high-dimensional data, while low-computing-power nodes may only be able to process data from a single sensor. Current research severely lacks a solution for achieving effective multi-source heterogeneous data collaboration under conditions of extreme computing power asymmetry. More subtly, traditional methods optimize perception, communication, and computation in a fragmented manner, failing to unify decision-making based on task requirements, environmental information, and underlying resource scheduling. Summary of the Invention
[0004] The purpose of this invention is to provide an unmanned swarm collaborative control system and method based on multi-source heterogeneous data, so as to solve the problems raised in the prior art.
[0005] To achieve the above objectives, the present invention provides the following technical solution:
[0006] A method for unmanned cluster cooperative control based on multi-source heterogeneous data, the method comprising the following steps:
[0007] S100: Each unmanned node in the unmanned cluster has a built-in benchmark testing unit. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the CPU utilization and core temperature in real time, and calculate the instantaneous computing power level.
[0008] The calculation of the instantaneous computing power level specifically involves:
[0009] S101. A lightweight benchmark unit is built into each unmanned node i to measure the peak number of floating-point operations F when the system starts. i and memory bandwidth B i Real-time monitoring of CPU utilization. i (t) and core temperature T i (t), the instantaneous computing power water level is calculated using the following formula:
[0010] ;
[0011] Among them, W i (t) represents the instantaneous computing power level of the i-th unmanned node at the current moment, k1 represents the operating system scheduling loss coefficient, k2 represents the temperature frequency attenuation coefficient, T0 represents the rated operating temperature, and e represents the natural constant. k1 and k2 are actively set by the staff based on system work data. The instantaneous computing power level can be updated in real time during any control cycle, making the computing power of each node in the unmanned cluster comparable.
[0012] S200: When controlling an unmanned cluster, calculate the total information volume of the global task, divide the total information volume into atomic control primitives according to the instantaneous computing power level, set the required computing power threshold for each atomic control primitive, and output the primitive-demand mapping table to be allocated.
[0013] The output primitive-demand mapping table to be assigned is specifically as follows:
[0014] S201. The ground station calculates the total information content of the unmanned swarm's global mission using the following formula:
[0015] ;
[0016] Among them, E tot Let p(x) represent the total information content, and p(x) represent the prior probability distribution of various sensor data in the fusion process.
[0017] The total information volume of the global task is proportionally divided into M atomic control units based on the instantaneous computing power level of the unmanned nodes; the computing power dispersion is calculated using the instantaneous computing power level, using the following formula:
[0018] ;
[0019] Where D represents the computing power dispersion, W max W represents the maximum instantaneous computing power level. min W represents the minimum instantaneous computing power level. avg This represents the average value of all instantaneous computing power levels, and ε represents the minimum value to prevent division by zero.
[0020] The number of atomic control primitives can be calculated using the computational power dispersion formula:
[0021] ;
[0022] Where N represents the number of unmanned nodes in the unmanned cluster, and round represents rounding to the nearest integer;
[0023] S202, Each atomic control primitive in the segmentation is accompanied by a required computational power threshold H. j The required computing power threshold is set by staff based on the amount of global task information contained in the atomic control primitives, and constraint H is set. j ≤max{W i (t)}, ensuring that there is at least one unmanned node that can process the corresponding atomic control primitive, and the required computing power threshold represents the computing power required to process the corresponding atomic control primitive; generate a primitive-demand mapping table.
[0024] The more uneven the distribution of computing power in the cluster, the finer the primitives are divided. The "dispersion-driven granularity" design ensures that strong nodes will not be blocked by a single complex primitive due to the large size of the task block, and weak nodes will not be unable to handle any primitives due to the coarseness of the task block, thus avoiding resource misallocation caused by "one-size-fits-all" approach from the source.
[0025] S300: Construct a bipartite graph using unmanned nodes and atomic control primitives. Allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. Calculate the mismatch potential energy of each edge as the edge weight. Iterate to calculate the minimum global total potential energy and output the optimal allocation scheme.
[0026] The specific steps for iteratively calculating the optimal allocation scheme for the minimum global total potential energy output are as follows:
[0027] S301. Construct a bipartite graph, with the left vertex being the set of unmanned nodes and the right vertex being the set of atomic control primitives. The edges represent the relationships between unmanned nodes and atomic control primitives. When the instantaneous computing power of an unmanned node is greater than or equal to the required computing power threshold, it means that it can undertake the corresponding atomic control primitive.
[0028] The mismatch potential energy of each edge is calculated as the edge weight, using the following formula:
[0029] ;
[0030] Among them, M ij (t) represents the edge weight between unmanned node i and atomic control primitive j, L ij This represents the communication link quality coefficient, which is set by staff based on the signal-to-noise ratio.
[0031] Existing load balancing algorithms often use absolute difference, which leads to high-water mark nodes having much larger weights than low-water mark nodes. This causes them to dominate the matching direction during iterative optimization, ignoring the matching needs of weaker nodes. The ratio form of this invention ensures that any node-primitive pair competes on the same numerical scale, guaranteeing matching fairness.
[0032] S302. Atomic control primitives are allocated to unmanned nodes according to edges. The edges actually used for allocation are used as allocation edges. The sum of the weights of all allocated edges is calculated using distributed iteration as the global situation energy. The allocation scheme that minimizes the global situation energy is output as the optimal allocation scheme. Each node only exchanges matching intentions rather than the original data, and converges to a local optimal stable state after iterating the asynchronous optimal response. This makes the method robust even when the communication topology changes dynamically, avoiding the risk of single point of failure of the central node.
[0033] S400: For each unmanned node, the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio are used to judge the data semantic compression ratio, filter high computing power nodes and low computing power nodes, and execute different sampling strategies.
[0034] The specific steps for implementing different sampling strategies are as follows:
[0035] S401. For each unmanned node, calculate the sum of the required computing power thresholds H of all atomic control primitives allocated to the unmanned node in the optimal allocation scheme. sum,i The semantic compression ratio of the data is calculated using the following formula:
[0036] R i (t) = W i (t) / H sum,i ;
[0037] Among them, R i (t) represents the data semantic compression ratio;
[0038] If R i If (t) > 1, it is determined that the corresponding unmanned node has excessive computing power and is a high computing power node. Then, oversampling interpolation is performed. The oversampling interpolation means improving the point cloud voxel detail of the local lidar while retaining the high-frequency time series.
[0039] If R i If (t) < 1, it is determined that the computing power is tight and it is a low computing power node. Sparse gating is then performed. The sparse gating means that the data frames are gradually thinned from the edge to the center according to the gradient decreasing outward from the center of the image within the sensor's field of view, but the lowest Nyquist sampling rate within the field of view is guaranteed. The pruned data stream will no longer contain all redundant information, but will only retain the feature channels that are strongly correlated with the primitives of the task it is responsible for, which greatly reduces the computational burden.
[0040] If Ri When (t) = 1, normal sampling is performed. By determining the execution of different strategies, nodes with different capabilities collect data with different precision, but all serve the global collaborative goal. This not only relieves the computing power pressure on weak nodes, but also maximizes the perception potential of strong nodes.
[0041] S500 uses unmanned node sensors and state observation theory to generate its own state and the estimated actual state of neighboring low computing power nodes for high computing power nodes, and receives the observed state reported by neighboring low computing power nodes, calculates the situational deviation gradient between the estimated actual state and the observed state, and generates virtual control quantities to compensate for the basic control commands of low computing power nodes.
[0042] The specific steps for generating virtual control quantities to compensate for control commands of low-computing-power nodes are as follows:
[0043] S501, high-computing-power nodes collect their own status X via sensors. b The predicted actual state Y of neighboring low-computing-power nodes k Receive the observation status Z reported by neighboring low-computing-power nodes k The low-computing-power node is determined by calculating the straight-line spatial distance between the low-computing-power node and the high-computing-power node, and the straight-line spatial distance is less than the upper limit of the physical communication radius; neighboring low-computing-power nodes report their observation status measured by sensors to the high-computing-power node within the control cycle of the global task; the status includes a three-dimensional spatial position vector and heading;
[0044] S502. Calculate the situation deviation gradient between the predicted actual state and the observed state of neighboring low-computing-power nodes. The formula is: G bk =Y k -Z k Among them, G bk Indicates the gradient of the situational deviation;
[0045] High-computing-power nodes encode the situational deviation gradient to form additional virtual control quantities. When low-computing-power nodes execute the assigned atomic control primitives in the global task, the virtual control quantities are superimposed on the basic control instructions. After receiving the virtual control quantities, the low-computing-power nodes directly inject them into their own underlying attitude control loop as external inertial damping compensation.
[0046] By enabling strong nodes to utilize their surplus computing power to reverse-engineer the kinematic drift error caused by data pruning in weak nodes, and generating and forcibly injecting compensation instructions, the heterogeneous computing power is transformed from a "burden" into a "compensation potential." After receiving the virtual control quantity, low-computing-power nodes do not require any state calculation or coordinate transformation, completely avoiding the latency and error accumulation caused by insufficient computing power of weak nodes in traditional heterogeneous fusion.
[0047] S600. In each control cycle, select a temporary verification node to calculate the derivative of the current system energy functional. Use the derivative of the system energy functional to determine whether the system control is converging toward the equilibrium point. If not, trigger reverse elastic braking.
[0048] The specific steps for determining whether the system control converges to the equilibrium point by using the derivative of the energy functional of the entire system are as follows:
[0049] S601. Within each control cycle, assess the instantaneous computing power level of all unmanned nodes and select the median value as a temporary verification node; calculate the derivative of the current system energy functional at the temporary verification node, using the following formula:
[0050] ;
[0051] Among them, S sys (t) represents the energy functional of the entire system. Then, the derivative of the current energy functional of the entire system is obtained by differentiating the energy functional of the entire system. W m (t) represents the average instantaneous computing power level of the unmanned cluster at the current moment, C i (t) represents the actual compensation residual of the unmanned node, which represents the vector magnitude of the difference between the situational deviation gradient and the virtual control quantity actually injected by the low computing power node; a and b represent the positive definite damping coefficients; the positive definite damping coefficients are actively set by the staff in this application based on the experiment;
[0052] The energy functional of this invention incorporates "uniformity of computing power distribution" and "tracking error" into the same scalar framework. The former measures the balance of computing power distribution in the cluster, while the latter measures the deviation of weak nodes. The derivative of the sum of the two is negative, which means that the positive contribution of computing power outweighs the negative contribution of error, and the system is converging towards stability, providing an unprecedented stability criterion for heterogeneous clusters.
[0053] S602, if If the value is less than 0, it proves that the system is converging towards a stable equilibrium point, and the S400 and S500 strategies are effective. Continue execution.
[0054] like If the value is ≥0, a reverse elastic braking is immediately triggered. This reverse elastic braking means forcibly reverting the semantic compression ratio of all unmanned node data to 1 and clearing the virtual control quantities injected by low-computing-power nodes to ensure physical safety as a priority. Once a system instability trend is detected, all nodes are immediately forced to revert to full data mode and the virtual control quantities are cleared. This "hard braking" mechanism ensures that even if the intelligent adaptation strategy fails, the system can still return to a conservative but safe baseline state, eliminating the risk of loss of control.
[0055] S700: When reverse elastic braking is triggered, the actual compensation residual is judged and the correction coefficient is calculated; the correction coefficient is substituted into the next control cycle to optimize the instantaneous computing power level.
[0056] The specific steps for substituting the residual vector as a correction term into the next control cycle to optimize the instantaneous computing power level are as follows:
[0057] S701. When reverse elastic braking is triggered, the actual compensation residual of the low computing power node is extracted, and a preset tolerance threshold δ is set. When the actual compensation residual ≤ δ, a correction coefficient P is assigned. i (t) = 1; if the actual compensation residual > δ, the penalty procedure is initiated; the tolerance threshold is preset by experts, and the preferred value in this application is 0.5;
[0058] S702. The penalty procedure involves calculating a correction coefficient using the actual compensated residual and the tolerance threshold, as shown in the formula:
[0059] P i (t) = 1 - α × (C) i (t)-δ);
[0060] Wherein, α represents the penalty slope, which is fixed by the staff based on experience, and is preferably 0.1 in this application;
[0061] The correction factor is substituted into the instantaneous computing power level calculation for the next control cycle to correct the number of peak floating-point operations: F i new (t + Δt) = F i +P i (t); F i new (t+△t) represents the corrected peak floating-point operation count; the corrected peak floating-point operation count is used to participate in the calculation of the instantaneous computing power level to correct the instantaneous computing power level of the next control cycle.
[0062] An unmanned swarm cooperative control system based on multi-source heterogeneous data, comprising a computing power analysis module, a task partitioning module, an allocation module, an allocation computing power judgment module, a computing power compensation module, a balance judgment module, and a correction module;
[0063] The computing power analysis module is used to build a benchmark testing unit in each unmanned node of the unmanned cluster. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the central processing unit utilization and core temperature in real time, and calculate the instantaneous computing power level.
[0064] The task segmentation module is used to calculate the total information volume of the global task when controlling an unmanned cluster, and to segment the total information volume into atomic control primitives according to the instantaneous computing power level. A necessary computing power threshold is set for each atomic control primitive, and the primitive-demand mapping table to be allocated is output.
[0065] The allocation module is used to construct a bipartite graph using unmanned nodes and atomic control primitives, allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold, calculate the mismatch potential energy of each edge as the edge weight, and output the optimal allocation scheme by iteratively calculating the minimum global total potential energy.
[0066] The allocated computing power judgment module is used to judge the data semantic compression ratio for each unmanned node by using the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio, to filter high computing power nodes and low computing power nodes, and to execute different sampling strategies.
[0067] The computing power compensation module is used to generate the state of the high computing power node and the estimated actual state of the neighboring low computing power nodes by using unmanned node sensors and state observation theory, and to receive the observation state reported by the neighboring low computing power nodes, calculate the situational deviation gradient between the estimated actual state and the observation state, and generate virtual control quantities to compensate the basic control commands of the low computing power nodes.
[0068] The balance judgment module is used to select a temporary verification node to calculate the derivative of the current whole system energy functional in each control cycle, and judge whether the system control is converging towards the equilibrium point based on the derivative of the whole system energy functional. If not, it triggers reverse elastic braking.
[0069] The correction module is used to judge the actual compensation residual when the reverse elastic braking is triggered, calculate the correction coefficient, and substitute the correction coefficient into the next control cycle to optimize the instantaneous computing power level.
[0070] The allocation module includes a bipartite graph construction unit and an optimal solution generation unit;
[0071] The bipartite graph construction unit is used to construct a bipartite graph using unmanned nodes and atomic control primitives. Atomic control primitives are allocated to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. The mismatch potential energy of each edge is calculated as the edge weight.
[0072] The optimal solution generation unit is used to output the optimal allocation scheme by iteratively calculating the minimum global total potential energy.
[0073] Compared with the prior art, the beneficial effects of the present invention are:
[0074] 1. All existing technological efforts revolve around "how to ensure that heterogeneous nodes do not fall behind when working collaboratively," i.e., peak shaving and valley filling, and eliminating differences. This invention takes the opposite approach, proposing for the first time that heterogeneity is not a problem but a resource. High-performance nodes, after completing their own tasks, still have surplus cognitive resources, which are systematically used to compensate for the perceptual deficiencies and motion drift of low-performance nodes. The strength of strong nodes is not only reflected in their high execution efficiency but also in their ability to support weaker nodes, thereby improving the overall cluster performance.
[0075] 2. A collaborative control architecture designed specifically for multi-source heterogeneous unmanned clusters was constructed, which is a closed-loop system of "perception-allocation-adaptation-compensation-verification-correction". Its core idea, "heterogeneous computing power is the compensating potential", fundamentally reverses the passive logic of "eliminating differences" in traditional heterogeneous collaboration. Attached Figure Description
[0076] Figure 1 This is a schematic diagram of the structure of an unmanned cluster cooperative control method based on multi-source heterogeneous data according to the present invention. Detailed Implementation
[0077] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0078] Example: Figure 1 As shown, the present invention provides a technical solution.
[0079] A method for unmanned cluster cooperative control based on multi-source heterogeneous data, the method comprising the following steps:
[0080] S100: Each unmanned node in the unmanned cluster has a built-in benchmark testing unit. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the CPU utilization and core temperature in real time, and calculate the instantaneous computing power level.
[0081] The calculation of the instantaneous computing power level specifically involves:
[0082] S101. A lightweight benchmark unit is built into each unmanned node i to measure the peak number of floating-point operations F when the system starts. i and memory bandwidth B i Real-time monitoring of CPU utilization. i (t) and core temperature T i (t), then normalized, and the instantaneous computing power water level is calculated using the following formula:
[0083] ;
[0084] Among them, W i (t) represents the instantaneous computing power level of the i-th unmanned node at the current moment, k1 represents the operating system scheduling loss coefficient, k2 represents the temperature frequency attenuation coefficient, T0 represents the rated operating temperature, and e represents the natural constant. k1 and k2 are actively set by the staff based on system work data. The instantaneous computing power level can be updated in real time during any control cycle, making the computing power of each node in the unmanned cluster comparable.
[0085] S200: When controlling an unmanned cluster, calculate the total information volume of the global task, divide the total information volume into atomic control primitives according to the instantaneous computing power level, set the required computing power threshold for each atomic control primitive, and output the primitive-demand mapping table to be allocated.
[0086] The output primitive-demand mapping table to be assigned is specifically as follows:
[0087] S201. The ground station calculates the total information content of the unmanned swarm's global mission using the following formula:
[0088] ;
[0089] Among them, E tot Let p(x) represent the total information content, and p(x) represent the prior probability distribution of various sensor data in the fusion process.
[0090] The total information volume of the global task is proportionally divided into M atomic control units based on the instantaneous computing power level of the unmanned nodes; the computing power dispersion is calculated using the instantaneous computing power level, using the following formula:
[0091] ;
[0092] Where D represents the computing power dispersion, W max W represents the maximum instantaneous computing power level. min W represents the minimum instantaneous computing power level. avg This represents the average value of all instantaneous computing power levels, and ε represents the minimum value to prevent division by zero.
[0093] The number of atomic control primitives can be calculated using the computational power dispersion formula:
[0094] ;
[0095] Where N represents the number of unmanned nodes in the unmanned cluster, and round represents rounding to the nearest integer; the information content of each of the M atomic control primitives is the same;
[0096] S202, Each atomic control primitive in the segmentation is accompanied by a required computational power threshold H. j The required computing power threshold is set by the staff based on the amount of global task information contained in the atomic control primitives. In this invention, the staff selects formula H. j =q×ρ j +ρ min ×γ requires a computational power threshold, where q represents the amount of information contained in each atomic control primitive, and ρ j ρ represents the operational density coefficient. min γ represents the minimum computational density coefficient, and γ represents the communication unpacking constant.
[0097] Set constraint H j ≤max{W i (t)}, ensuring that there is at least one unmanned node that can process the corresponding atomic control primitive, and the required computing power threshold represents the computing power required to process the corresponding atomic control primitive; generate a primitive-demand mapping table.
[0098] The more uneven the distribution of computing power in the cluster, the finer the primitives are divided. The "dispersion-driven granularity" design ensures that strong nodes will not be blocked by a single complex primitive due to the large size of the task block, and weak nodes will not be unable to handle any primitives due to the coarseness of the task block, thus avoiding resource misallocation caused by "one-size-fits-all" approach from the source.
[0099] S300: Construct a bipartite graph using unmanned nodes and atomic control primitives. Allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. Calculate the mismatch potential energy of each edge as the edge weight. Iterate to calculate the minimum global total potential energy and output the optimal allocation scheme.
[0100] The specific steps for iteratively calculating the optimal allocation scheme for the minimum global total potential energy output are as follows:
[0101] S301. Construct a bipartite graph, with the left vertex being the set of unmanned nodes and the right vertex being the set of atomic control primitives. The edges represent the relationships between unmanned nodes and atomic control primitives. When the instantaneous computing power of an unmanned node is greater than or equal to the required computing power threshold, it means that it can undertake the corresponding atomic control primitive.
[0102] The mismatch potential energy of each edge is calculated as the edge weight, using the following formula:
[0103] ;
[0104] Among them, M ij (t) represents the edge weight between unmanned node i and atomic control primitive j, L ij This represents the communication link quality coefficient, which is set by staff based on the signal-to-noise ratio.
[0105] Existing load balancing algorithms often use absolute difference, which leads to high-water mark nodes having much larger weights than low-water mark nodes. This causes them to dominate the matching direction during iterative optimization, ignoring the matching needs of weaker nodes. The ratio form of this invention ensures that any node-primitive pair competes on the same numerical scale, guaranteeing matching fairness.
[0106] S302. Atomic control primitives are allocated to unmanned nodes according to edges. The edges actually used for allocation are used as allocation edges. The sum of the weights of all allocated edges is calculated using distributed iteration as the global situation energy. The allocation scheme that minimizes the global situation energy is output as the optimal allocation scheme. Each node only exchanges matching intentions rather than the original data, and converges to a local optimal stable state after iterating the asynchronous optimal response. This makes the method robust even when the communication topology changes dynamically, avoiding the risk of single point of failure of the central node.
[0107] S400: For each unmanned node, the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio are used to judge the data semantic compression ratio, filter high computing power nodes and low computing power nodes, and execute different sampling strategies.
[0108] The specific steps for implementing different sampling strategies are as follows:
[0109] S401. For each unmanned node, calculate the sum of the required computing power thresholds H of all atomic control primitives allocated to the unmanned node in the optimal allocation scheme. sum,i The semantic compression ratio of the data is calculated using the following formula:
[0110] R i (t) = W i (t) / H sum,i ;
[0111] Among them, R i (t) represents the data semantic compression ratio;
[0112] If R i If (t) > 1, it is determined that the corresponding unmanned node has excessive computing power and is a high computing power node. Then, oversampling interpolation is performed. The oversampling interpolation means improving the point cloud voxel detail of the local lidar while retaining the high-frequency time series.
[0113] If R i If (t) < 1, it is determined that the computing power is tight and it is a low computing power node. Sparse gating is then performed. The sparse gating means that the data frames are gradually thinned from the edge to the center according to the gradient decreasing outward from the center of the image within the sensor's field of view, but the lowest Nyquist sampling rate within the field of view is guaranteed. The pruned data stream will no longer contain all redundant information, but will only retain the feature channels that are strongly correlated with the primitives of the task it is responsible for, which greatly reduces the computational burden.
[0114] If Ri When (t) = 1, normal sampling is performed. By determining the execution of different strategies, nodes with different capabilities collect data with different precision, but all serve the global collaborative goal. This not only relieves the computing power pressure on weak nodes, but also maximizes the perception potential of strong nodes.
[0115] S500 uses unmanned node sensors and state observation theory to generate its own state and the estimated actual state of neighboring low computing power nodes for high computing power nodes, and receives the observed state reported by neighboring low computing power nodes, calculates the situational deviation gradient between the estimated actual state and the observed state, and generates virtual control quantities to compensate for the basic control commands of low computing power nodes.
[0116] The specific steps for generating virtual control quantities to compensate for control commands of low-computing-power nodes are as follows:
[0117] S501, high-computing-power nodes collect their own status X via sensors. b The predicted actual state Y of neighboring low-computing-power nodes k Receive the observation status Z reported by neighboring low-computing-power nodes k The low-computing-power node is determined by calculating the straight-line spatial distance between the low-computing-power node and the high-computing-power node, and the straight-line spatial distance is less than the upper limit of the physical communication radius; neighboring low-computing-power nodes report their observation status measured by sensors to the high-computing-power node within the control cycle of the global task; the status includes a three-dimensional spatial position vector and heading;
[0118] S502. Calculate the situation deviation gradient between the predicted actual state and the observed state of neighboring low-computing-power nodes. The formula is: G bk =Y k -Z k Among them, G bk Indicates the gradient of the situational deviation;
[0119] High-computing-power nodes encode the situational deviation gradient to form additional virtual control quantities. When low-computing-power nodes execute the assigned atomic control primitives in the global task, the virtual control quantities are superimposed on the basic control instructions. After receiving the virtual control quantities, the low-computing-power nodes directly inject them into their own underlying attitude control loop as external inertial damping compensation.
[0120] By enabling strong nodes to utilize their surplus computing power to reverse-engineer the kinematic drift error caused by data pruning in weak nodes, and generating and forcibly injecting compensation instructions, the heterogeneous computing power is transformed from a "burden" into a "compensation potential." After receiving the virtual control quantity, low-computing-power nodes do not require any state calculation or coordinate transformation, completely avoiding the latency and error accumulation caused by insufficient computing power of weak nodes in traditional heterogeneous fusion.
[0121] S600. In each control cycle, select a temporary verification node to calculate the derivative of the current system energy functional. Use the derivative of the system energy functional to determine whether the system control is converging toward the equilibrium point. If not, trigger reverse elastic braking.
[0122] The specific steps for determining whether the system control converges to the equilibrium point by using the derivative of the energy functional of the entire system are as follows:
[0123] S601. Within each control cycle, the instantaneous computing power level of all unmanned nodes is assessed, and the intermediate value is selected as a temporary verification node. The derivative of the current energy functional of the entire system is calculated at the temporary verification node. Only the numerical value is used in the calculation; dimensions are not considered. The formula is:
[0124] ;
[0125] Among them, S sys (t) represents the energy functional of the entire system. Then, the derivative of the current energy functional of the entire system is obtained by differentiating the energy functional of the entire system. W m (t) represents the average instantaneous computing power level of the unmanned cluster at the current moment, C i (t) represents the actual compensation residual of the unmanned node, which represents the vector magnitude of the difference between the situational deviation gradient and the virtual control quantity actually injected by the low computing power node; a and b represent the positive definite damping coefficients; the positive definite damping coefficients are actively set by the staff in this application based on the experiment;
[0126] The energy functional of this invention incorporates "uniformity of computing power distribution" and "tracking error" into the same scalar framework. The former measures the balance of computing power distribution in the cluster, while the latter measures the deviation of weak nodes. The derivative of the sum of the two is negative, which means that the positive contribution of computing power outweighs the negative contribution of error, and the system is converging towards stability, providing an unprecedented stability criterion for heterogeneous clusters.
[0127] S602, if If the value is less than 0, it proves that the system is converging towards a stable equilibrium point, and the S400 and S500 strategies are effective. Continue execution.
[0128] like If the value is ≥0, a reverse elastic braking is immediately triggered. This reverse elastic braking means forcibly reverting the semantic compression ratio of all unmanned node data to 1 and clearing the virtual control quantities injected by low-computing-power nodes to ensure physical safety as a priority. Once a system instability trend is detected, all nodes are immediately forced to revert to full data mode and the virtual control quantities are cleared. This "hard braking" mechanism ensures that even if the intelligent adaptation strategy fails, the system can still return to a conservative but safe baseline state, eliminating the risk of loss of control.
[0129] S700: When reverse elastic braking is triggered, the actual compensation residual is judged and the correction coefficient is calculated; the correction coefficient is substituted into the next control cycle to optimize the instantaneous computing power level.
[0130] The specific steps for substituting the residual vector as a correction term into the next control cycle to optimize the instantaneous computing power level are as follows:
[0131] S701. When reverse elastic braking is triggered, the actual compensation residual of the low computing power node is extracted, and a preset tolerance threshold δ is set. When the actual compensation residual ≤ δ, a correction coefficient P is assigned. i (t) = 1; if the actual compensation residual > δ, the penalty procedure is initiated; the tolerance threshold is preset by experts, and the preferred value in this application is 0.5;
[0132] S702. The penalty procedure involves calculating a correction coefficient using the actual compensated residual and the tolerance threshold, as shown in the formula:
[0133] P i (t) = 1 - α × (C) i (t)-δ);
[0134] Wherein, α represents the penalty slope, which is fixed by the staff based on experience, and is preferably 0.1 in this application;
[0135] The correction factor is substituted into the instantaneous computing power level calculation for the next control cycle to correct the number of peak floating-point operations: F i new (t + Δt) = F i +P i (t); F i new (t+△t) represents the corrected peak floating-point operation count; the corrected peak floating-point operation count is used to participate in the calculation of the instantaneous computing power level to correct the instantaneous computing power level of the next control cycle.
[0136] An unmanned swarm cooperative control system based on multi-source heterogeneous data, comprising a computing power analysis module, a task partitioning module, an allocation module, an allocation computing power judgment module, a computing power compensation module, a balance judgment module, and a correction module;
[0137] The computing power analysis module is used to build a benchmark testing unit in each unmanned node of the unmanned cluster. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the central processing unit utilization and core temperature in real time, and calculate the instantaneous computing power level.
[0138] The task segmentation module is used to calculate the total information volume of the global task when controlling an unmanned cluster, and to segment the total information volume into atomic control primitives according to the instantaneous computing power level. A necessary computing power threshold is set for each atomic control primitive, and the primitive-demand mapping table to be allocated is output.
[0139] The allocation module is used to construct a bipartite graph using unmanned nodes and atomic control primitives, allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold, calculate the mismatch potential energy of each edge as the edge weight, and output the optimal allocation scheme by iteratively calculating the minimum global total potential energy.
[0140] The allocated computing power judgment module is used to judge the data semantic compression ratio for each unmanned node by using the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio, to filter high computing power nodes and low computing power nodes, and to execute different sampling strategies.
[0141] The computing power compensation module is used to generate the state of the high computing power node and the estimated actual state of the neighboring low computing power nodes by using unmanned node sensors and state observation theory, and to receive the observation state reported by the neighboring low computing power nodes, calculate the situational deviation gradient between the estimated actual state and the observation state, and generate virtual control quantities to compensate the basic control commands of the low computing power nodes.
[0142] The balance judgment module is used to select a temporary verification node to calculate the derivative of the current whole system energy functional in each control cycle, and judge whether the system control is converging towards the equilibrium point based on the derivative of the whole system energy functional. If not, it triggers reverse elastic braking.
[0143] The correction module is used to judge the actual compensation residual when the reverse elastic braking is triggered, calculate the correction coefficient, and substitute the correction coefficient into the next control cycle to optimize the instantaneous computing power level.
[0144] The allocation module includes a bipartite graph construction unit and an optimal solution generation unit;
[0145] The bipartite graph construction unit is used to construct a bipartite graph using unmanned nodes and atomic control primitives. Atomic control primitives are allocated to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. The mismatch potential energy of each edge is calculated as the edge weight.
[0146] The optimal solution generation unit is used to output the optimal allocation scheme by iteratively calculating the minimum global total potential energy.
[0147] Example:
[0148] Suppose a certain unmanned cluster consists of 6 unmanned nodes (numbered 1-6) performing a "cooperative area search and target capture" task. The heterogeneous platform is configured as follows:
[0149] 1. High-end drone with a peak floating-point operation count of 120 and a memory bandwidth of 80.
[0150] 2. Medium-sized UAVs, peak floating-point operations 60, memory bandwidth 50;
[0151] 3. Medium-sized drones, peak floating-point operations 55, memory bandwidth 45;
[0152] 4. Small drones, peak floating-point operations 20, memory bandwidth 25;
[0153] 5 small drones, peak floating-point operations 18, memory bandwidth 22;
[0154] 6 micro-drones, peak floating-point operations 8, memory bandwidth 12;
[0155] Preset parameters: k1=0.05, k2=0.008, T0=40℃, control cycle=0.1s, upper limit of physical communication radius=500m, positive definite damping coefficients a=0.5, b=0.8; α=0.1, δ=0.5;
[0156] The instantaneous computing power levels were calculated to be 1.332, 1.095, 1.136, 0.76, 0.781, and 0.643, respectively.
[0157] Assuming that in the current collaborative search task, the prior probability distribution of data fusion from various sensors is as follows: LiDAR data accounts for 40%, visual image data accounts for 35%, and inertial navigation data accounts for 25%, the total information content is calculated to be 1.559; the computing power dispersion is 0.718; and the number of atomic control primitives is calculated to be 12.
[0158] The staff divided the 12 atomic control primitives into three categories: 1-4 require a computing power threshold of 0.601, 5-8 require a computing power threshold of 0.450, and 9-12 require a computing power threshold of 0.387; the total original requirement is 5.752.
[0159] When constructing the bipartite graph, the mismatch potential energy is calculated as the edge weight. For example, the edge weights between node 1 and all primitives are 0.359 for 1-4, 0.470 for 5-8, and 0.523 for 9-12. Through iteration, the final allocation scheme is output as follows: nodes 1-1, 2, 3, 5; nodes 2-4, 6; nodes 3-7, 8; nodes 4-9, 10; nodes 5-11; nodes 6-12.
[0160] The semantic compression ratios of all nodes are 0.591, 1.042, 1.262, 0.982, 2.018, and 1.662, respectively; the nodes with high computing power are identified as 2, 3, 5, and 6, and the nodes with low computing power are identified as 1 and 4.
[0161] For example, the neighboring low-computing-power nodes of node 2 are 1 and 4;
[0162] Node 4 reported the observation status as position = (195, 155, 48) and heading = 0.28 radians;
[0163] The estimated actual state is: position = (195.5, 173.7, 50), heading 0.35 radians;
[0164] Calculate the situation deviation gradient = (0.5, 18.7, 2.0, 0.07); use the situation deviation gradient encoding to generate a virtual control quantity to compensate node 4.
[0165] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.
Claims
1. A method for unmanned cluster cooperative control based on multi-source heterogeneous data, characterized in that: The method includes the following steps: S100: Each unmanned node in the unmanned cluster has a built-in benchmark testing unit. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the CPU utilization and core temperature in real time, and calculate the instantaneous computing power level. S200: When controlling an unmanned cluster, calculate the total information volume of the global task, divide the total information volume into atomic control primitives according to the instantaneous computing power level, set the required computing power threshold for each atomic control primitive, and output the primitive-demand mapping table to be allocated. S300: Construct a bipartite graph using unmanned nodes and atomic control primitives. Allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. Calculate the mismatch potential energy of each edge as the edge weight. Iterate to calculate the minimum global total potential energy and output the optimal allocation scheme. S400: For each unmanned node, the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio are used to judge the data semantic compression ratio, filter high computing power nodes and low computing power nodes, and execute different sampling strategies. S500 uses unmanned node sensors and state observation theory to generate its own state and the estimated actual state of neighboring low computing power nodes for high computing power nodes, and receives the observed state reported by neighboring low computing power nodes, calculates the situational deviation gradient between the estimated actual state and the observed state, and generates virtual control quantities to compensate for the basic control commands of low computing power nodes. S600. In each control cycle, select a temporary verification node to calculate the derivative of the current system energy functional. Use the derivative of the system energy functional to determine whether the system control is converging toward the equilibrium point. If not, trigger reverse elastic braking. S700: When reverse elastic braking is triggered, the actual compensation residual is judged and the correction coefficient is calculated; the correction coefficient is substituted into the next control cycle to optimize the instantaneous computing power level.
2. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: The primitive-demand mapping table to be allocated output in S200 is specifically as follows: The total information volume of the global task is divided into M atomic control units proportionally based on the instantaneous computing power level of the unmanned nodes; the computing power dispersion is calculated using the instantaneous computing power level, and the number of atomic control units is calculated using the computing power dispersion. Each atomic control primitive in the segmentation comes with a required computational power threshold H. j Set constraint H j ≤max{W i (t)}, W i (t) Instantaneous computing power level; Required computing power threshold represents the computing power necessary to process the corresponding atomic control primitive; Generate primitive-demand mapping table.
3. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: The construction of the bipartite graph in S300 using unmanned nodes and atomic control primitives is specifically as follows: Construct a bipartite graph, with the left vertex being the set of unmanned nodes and the right vertex being the set of atomic control primitives. Edges represent the relationships in which unmanned nodes can undertake atomic control primitives. When the instantaneous computing power level of an unmanned node is greater than or equal to the required computing power threshold, it means that it can undertake the corresponding atomic control primitive. Calculate the mismatch potential energy as the edge weight using the instantaneous computing power level of the unmanned node and the required computing power threshold of the atomic control primitive for each edge.
4. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: The optimal allocation scheme for outputting the minimum global total potential energy through iterative calculation in S300 is specifically as follows: The atomic control primitives are allocated to unmanned nodes according to the edges. The actual edges used for allocation are used as the allocation edges. The sum of the weights of all allocated edges is calculated using distributed iteration as the global state energy. The allocation scheme that minimizes the global state energy is output as the optimal allocation scheme.
5. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: The specific implementation of different sampling strategies in S400 is as follows: For each unmanned node, calculate the sum of the required computing power threshold H of all atomic control primitives allocated to the unmanned node in the optimal allocation scheme. sum,i The semantic compression ratio of the data is calculated using the following formula: R i (t)=W i (t) / H sum,i ; Among them, R i (t) represents the data semantic compression ratio; If R i If (t) > 1, it is determined that the corresponding unmanned node has excessive computing power and is a high computing power node. Then, oversampling interpolation is performed. The oversampling interpolation means improving the point cloud voxel detail of the local lidar while retaining the high-frequency time series. If R i If (t) < 1, it is determined that the computing power is tight, indicating a low-computing-power node. Sparsity gating is then implemented. This sparsity gating means gradually thinning the data frames from the edge to the center based on the gradient decreasing outwards from the image center within the sensor's field of view, while ensuring the lowest Nyquist sampling rate within the field of view. If R i When (t) = 1, normal sampling is performed.
6. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: The specific steps for compensating low-computing-power nodes by generating virtual control quantities in S500 are as follows: High-performance computing nodes collect their own status X b The predicted actual state Y of neighboring low-computing-power nodes k Receive the observation status Z reported by neighboring low-computing-power nodes k The low-computing-power node is determined by calculating the straight-line spatial distance between the low-computing-power node and the high-computing-power node, and the straight-line spatial distance is less than the upper limit of the physical communication radius; neighboring low-computing-power nodes report their observation status measured by sensors to the high-computing-power node within the control cycle of the global task; the status includes a three-dimensional spatial position vector and heading; The gradient of the situational deviation between the predicted actual state and the observed state of neighboring low-computing-power nodes is calculated using the formula: G bk =Y k -Z k Among them, G bk Indicates the gradient of the situational deviation; High-computing-power nodes encode the situational deviation gradient to form additional virtual control quantities. When low-computing-power nodes execute the assigned atomic control primitives in the global task, the virtual control quantities are superimposed on the basic control instructions. After receiving the virtual control quantities, the low-computing-power nodes directly inject them into their own underlying attitude control loop as external inertial damping compensation.
7. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: In S600, determining whether the system control converges towards the equilibrium point by using the derivative of the energy functional of the entire system is specifically as follows: Within each control cycle, the instantaneous computing power level of all unmanned nodes is assessed, and the median value is selected as a temporary verification node. The derivative of the current energy functional of the entire system is then calculated at the temporary verification node. The derivative of the energy functional of the whole system is composed of the instantaneous computing power level and the actual compensation residual; the actual compensation residual represents the vector magnitude of the difference between the situation deviation gradient and the virtual control quantity actually injected by the low computing power node; a and b represent positive definite damping coefficients; the positive definite damping coefficients are actively set by the staff according to the experiment in this application; like If the value is less than 0, it proves that the system is converging towards a stable equilibrium point, and the S400 and S500 strategies are effective. Continue execution. like If the value is ≥0, reverse elastic braking is immediately triggered. Reverse elastic braking means forcibly reverting the semantic compression ratio of all unmanned node data to 1 and clearing the virtual control quantity injected by low computing power nodes.
8. The unmanned cluster cooperative control method based on multi-source heterogeneous data according to claim 1, characterized in that: In S700, the correction coefficient is substituted into the next control cycle to optimize the instantaneous computing power level as follows: When reverse elastic braking is triggered, the actual compensation residual of the low-computing-power node is extracted, and a preset tolerance threshold δ is used. When the actual compensation residual is ≤ δ, a correction coefficient P is assigned. i (t) = 1; If the actual compensated residual is greater than δ, a penalty procedure will be initiated. The penalty procedure involves calculating a correction coefficient using the actual compensation residual and tolerance threshold, substituting the correction coefficient into the instantaneous computing power level calculation for the next control cycle to correct the peak floating-point operation count, and using the corrected peak floating-point operation count to participate in the instantaneous computing power level calculation to correct the instantaneous computing power level for the next control cycle.
9. An unmanned swarm cooperative control system based on multi-source heterogeneous data, characterized in that: The unmanned swarm collaborative control system includes a computing power analysis module, a task partitioning module, an allocation module, an allocation computing power judgment module, a computing power compensation module, a balance judgment module, and a correction module; The computing power analysis module is used to build a benchmark testing unit in each unmanned node of the unmanned cluster. The benchmark testing unit is used to collect the peak number of floating-point operations and memory bandwidth at startup, collect the central processing unit utilization and core temperature in real time, and calculate the instantaneous computing power level. The task segmentation module is used to calculate the total information volume of the global task when controlling an unmanned cluster, and to segment the total information volume into atomic control primitives according to the instantaneous computing power level. A necessary computing power threshold is set for each atomic control primitive, and the primitive-demand mapping table to be allocated is output. The allocation module is used to construct a bipartite graph using unmanned nodes and atomic control primitives, allocate atomic control primitives to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold, calculate the mismatch potential energy of each edge as the edge weight, and output the optimal allocation scheme by iteratively calculating the minimum global total potential energy. The allocated computing power judgment module is used to judge the data semantic compression ratio for each unmanned node by using the required computing power threshold of all allocated atomic control primitives and the data semantic compression ratio, to filter high computing power nodes and low computing power nodes, and to execute different sampling strategies. The computing power compensation module is used to generate the state of the high computing power node and the estimated actual state of the neighboring low computing power nodes by using unmanned node sensors and state observation theory, and to receive the observation state reported by the neighboring low computing power nodes, calculate the situational deviation gradient between the estimated actual state and the observation state, and generate virtual control quantities to compensate the basic control commands of the low computing power nodes. The balance judgment module is used to select a temporary verification node to calculate the derivative of the current whole system energy functional in each control cycle, and judge whether the system control is converging towards the equilibrium point based on the derivative of the whole system energy functional. If not, it triggers reverse elastic braking. The correction module is used to determine the actual compensation residual and calculate the correction coefficient when the reverse elastic braking is triggered. The correction coefficient is substituted into the next control cycle to optimize the instantaneous computing power level.
10. The unmanned cluster collaborative control system based on multi-source heterogeneous data according to claim 9, characterized in that: The allocation module includes a bipartite graph construction unit and an optimal solution generation unit; The bipartite graph construction unit is used to construct a bipartite graph using unmanned nodes and atomic control primitives. Atomic control primitives are allocated to unmanned nodes to form the edges of the bipartite graph using the required computing power threshold. The mismatch potential energy of each edge is calculated as the edge weight. The optimal solution generation unit is used to output the optimal allocation scheme by iteratively calculating the minimum global total potential energy.
