A heterogeneous unmanned system cluster distributed task dynamic allocation method and system

Abstract

The application belongs to the technical field of heterogeneous unmanned system cluster cooperative control, and particularly relates to a heterogeneous unmanned system cluster distributed task dynamic allocation method and system, which comprises the following steps: collecting the multi-dimensional capability state of each unmanned platform in the cluster in real time, performing normalization preprocessing, performing linear regression on the capability state time sequence based on a sliding time window, and extracting a decay gradient vector; combining a task sensitivity weight vector to calculate a decay sensitivity coupling index; establishing an adaptive mutation threshold based on cluster statistical characteristics for the time difference change rate of the index, triggering preventive reallocation after continuous period confirmation; performing local cascading reallocation with the mutation platform as the center, using a shadow task mechanism to ensure task continuity and complete closed-loop verification. The application objectively determines the reallocation triggering time by cluster real-time statistical characteristics, eliminating the perception blind area of capability snapshot lagging behind the actual capability in the fixed period reallocation scheme.

Description

Technical Field

[0001] This invention relates to the field of collaborative control technology for heterogeneous unmanned system clusters. More specifically, this invention relates to a method and system for dynamic allocation of distributed tasks in heterogeneous unmanned system clusters. Background Technology

[0002] Heterogeneous unmanned system swarms consist of various types of unmanned platforms, including rotary-wing UAVs, unmanned vehicles, and unmanned surface vessels. These platforms differ significantly in terms of endurance, payload configuration, sensor performance, and communication capabilities. When performing complex collaborative operations such as area reconnaissance, target tracking, or material delivery, the system needs to rationally allocate different tasks or sub-tasks to the most suitable platform based on the current capabilities of each platform. The rationality of this task allocation directly determines the overall operational efficiency of the swarm and the final success rate of the missions.

[0003] However, in real-world, dynamic, and complex environments, the operating conditions faced by unmanned platforms are not static, and their key capabilities often exhibit complex nonlinear degradation characteristics. For example, in high-wind-speed environments, the battery discharge curve of a rotary-wing UAV may show a turning point of accelerated degradation within a certain charge range, causing the degradation rate of various capabilities to increase sharply in a short period of time. If the system fails to detect this sudden change in capabilities caused by environmental stress in a timely manner and still allows a platform with severely degraded capabilities to perform tasks that are highly dependent on those capabilities, a mismatch between the platform and the task will occur. This mismatch not only risks the failure of the current individual task but may also trigger a cascading effect due to the forced interruption of tasks by individual platforms, forcing adjacent platforms to temporarily take on additional loads, thereby accelerating the depletion of the capabilities of the entire local subgroup.

[0004] To address the task allocation problem in dynamic environments, Chinese patent application CN112270488A discloses a method, apparatus, and system for task allocation in UAV swarms. This solution primarily involves a central node periodically acquiring capability declaration data from multiple UAVs and dynamically mapping the decomposed subtasks to their corresponding UAVs according to a pre-defined allocation algorithm. In practical engineering applications, such solutions typically rely on fixed time intervals to poll and collect platform capability status data, using snapshots of capabilities at discrete time points as input for task allocation decisions.

[0005] While the aforementioned existing technologies can cope with a certain degree of capability changes, they still have significant limitations when facing nonlinear, abrupt changes in the environment. The capability snapshot assessment mechanism based on fixed-period polling used in this application has an unavoidable perception blind spot during the interval between two polls. Because this mechanism can only obtain static capability values ​​at periodic nodes, it cannot continuously characterize the decay gradient of platform capabilities over time. When the platform capability happens to experience the aforementioned inflection point of nonlinear decay trend within the polling interval, the system cannot detect the step increase in the capability degradation rate in time due to the perception lag of the fixed period. This will cause the system to mistakenly believe that the platform is still in the healthy state at the time of the previous polling cycle assessment, thus allowing it to continue executing the original high-load tasks. This situation will ultimately lead to the triggering of task reassignment lagging far behind the actual degradation state of the platform capability, causing an irreversible risk of platform and task cascading mismatch. Therefore, how to capture the dynamic degradation characteristics of the platform capability at the first moment when the inflection point of decay occurs, thereby breaking the perception blind spot of the fixed-period polling scheme and triggering preventive dynamic reassignment, has become an urgent technical problem to be solved in the field of heterogeneous unmanned swarm collaborative control. Summary of the Invention

[0006] To address the technical problem that the capability snapshot assessment mechanism used in existing fixed-period reallocation schemes has a perception blind spot, causing the capability assessment results to lag behind the actual state of unmanned platform capability evolution, and thus leading to the risk of mismatch between the task execution and the unmanned platform, this invention provides a method and system for dynamic allocation of distributed tasks in heterogeneous unmanned system clusters.

[0007] In a first aspect, the present invention provides a method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster, comprising: real-time acquisition of the multi-dimensional capability status of each unmanned platform in the cluster; normalization preprocessing of the original sampled values ​​of each capability dimension; linear regression of the normalized capability status time series based on a sliding time window to extract the decay gradient of each capability dimension and combine them into a decay gradient vector; calculating the decay sensitivity coupling index of each unmanned platform relative to each task based on the sensitivity weight vector of each task to the capability dimension and the decay gradient vector of each platform; establishing an adaptive mutation threshold based on the cluster statistical characteristics for the time difference change rate of the decay sensitivity coupling index, and triggering preventive reallocation after confirmation through multiple consecutive calculation cycles; performing local cascaded reallocation within the communication topology neighbor range with the mutated platform as the center, ensuring task continuity through a shadow task mechanism, and performing closed-loop verification after the reallocation is completed.

[0008] By adopting the above technical solution, this invention continuously extracts the decay gradient of the capability dimension based on a sliding time window and calculates the coupling index in combination with task sensitivity. It can capture the dynamic degradation characteristics of the platform capability at the first moment when the nonlinear decay inflection point appears, thereby breaking the perception blind spot and triggering preventive reallocation in time before irreversible platform and task mismatch occurs. Combined with shadow tasks and closed-loop verification mechanism, it avoids the risk of cascading exhaustion caused by the failure of a single task and improves the overall operation efficiency and task continuity of heterogeneous clusters in dynamic and complex environments.

[0009] Preferably, the normalization preprocessing of the original sampled values ​​of each capability dimension includes: obtaining the design full-scale rated value of each capability dimension of each unmanned platform; dividing the original sampled value of each capability dimension at the current moment by the design full-scale rated value of the corresponding dimension to obtain a dimensionless capability state value with a value range of [0,1].

[0010] By adopting the above technical solution, this invention not only eliminates the non-additivity problem between different physical dimensions, but also lays a legitimate and robust physical and mathematical foundation for subsequent cross-dimensional gradient weighted summation and unified threshold comparison.

[0011] Preferably, the step of performing linear regression on the normalized capability state time series based on a sliding time window to extract the decay gradient of each capability dimension includes: taking the dimensionless capability state values ​​of the most recent preset number of sampling points to form a time series; performing first-order least squares linear regression on the time series and extracting the regression slope as the instantaneous decay gradient of the capability dimension; the sliding window length is the number of sampling points covered by a preset multiple of the unmanned platform body control loop period.

[0012] By adopting the above technical solution, the present invention can ensure that the linear regression sample size is sufficient to suppress the interference of single sampling noise on gradient estimation, and can also avoid the trend response lag caused by excessively long windows. Thus, a better engineering balance is achieved between the smoothness of gradient estimation and the sensitivity of degradation response, and the system's perception delay of capability degradation rate is compressed to an extremely low level.

[0013] Preferably, the step of calculating the decay sensitivity coupling index of each unmanned platform relative to each task based on the sensitivity weight vector of each task to the capability dimension and the decay gradient vector of each platform includes: defining a sensitivity weight vector for each task to each capability dimension, wherein the sum of the weights of each dimension of the sensitivity weight vector is 1; and performing a weighted inner product of the absolute value of the decay gradient of each unmanned platform with the sensitivity weight of the corresponding task to obtain the decay sensitivity coupling index.

[0014] By adopting the above technical solution, the present invention realizes the directional correlation between platform-side capability decay information and task-side capability dependency information, so that mismatch risk assessment can focus on whether the capabilities that the platform is rapidly decaying are precisely the capabilities that the current task depends on most, thereby realizing the location of mismatch risk.

[0015] Preferably, defining a sensitivity weight vector for each capability dimension for each task includes: if there are records of execution failure cases of the same type of task in the system's historical task database, the trigger frequency of each capability dimension in the failure event is counted, and the trigger frequency ratio is normalized and used as the initial weight of the corresponding dimension; if there is insufficient historical failure case data in the system's historical task database, the initial weight of each capability dimension is taken as a uniform distribution.

[0016] By adopting the above technical solution, this invention mines execution failure cases in the system's historical task database and normalizes the initial weights by using the trigger frequency ratio of each capability dimension in the failure events, providing an objective and reliable data-driven basis for determining sensitivity weights; at the same time, when historical data is insufficient, uniform distribution is used as the initial value, ensuring that the system can still operate stably in cold start or low-sample scenarios, avoiding subjective bias caused by manually setting weights.

[0017] Preferably, the sensitivity weight vector is also updated through an environmental dynamic correction mechanism, including: when the real-time environmental parameters of the task execution area exceed the preset standard operating condition baseline for the task type, the ratio of the measured value of the current environmental parameter to the standard operating condition baseline value is taken as an amplification factor; after multiplicatively amplifying the weights of the affected dimensions, the entire weight vector is normalized so that the sum of the weights of all dimensions is always 1.

[0018] Preferably, the establishment of an adaptive mutation threshold based on cluster statistical characteristics for the time difference change rate of the decay sensitivity coupling index includes: calculating the time difference change rate of the decay sensitivity coupling index between two adjacent calculation cycles; taking the time difference change rates of all active unmanned platforms paired with tasks within the cluster to form a risk acceleration sample set; removing the sample with the largest value from the risk acceleration sample set to obtain a truncated sample set, calculating the mean and standard deviation of the truncated sample set, and setting the mean plus three times the standard deviation as the adaptive mutation threshold; if the number of samples in the risk acceleration sample set is less than a preset number threshold, it degenerates into using a fixed threshold mode within the historical sliding time window of the target unmanned platform as the mutation threshold.

[0019] Preferably, the step of triggering preventive reallocation after confirmation through multiple consecutive calculation cycles includes: setting a confirmation window, the length of which is the number of calculation cycles covered by a preset proportion of the decay gradient sliding window time span; and requiring that the rate of change of time difference exceeds the mutation threshold within the calculation cycles that continuously reach the confirmation window length before formally triggering preventive reallocation.

[0020] Preferably, the step of performing local cascaded redistribution within the communication topology neighbor range, centered on the mutation platform, includes: the mutation platform broadcasting a task transfer request message to the communication topology neighbors; the neighbor platforms calculating the decay sensitivity coupling index assumed after taking over the task, based on their current decay gradient vector and the sensitivity weight vector of the task, as the expected decay sensitivity coupling index and bidding price; selecting the platform with the smallest bidding price as the task recipient; if the difference in bidding prices among multiple candidate platforms is less than the standard deviation of the current expected decay sensitivity coupling index sample set of the cluster, then selecting the platform with the lightest remaining task load; if there is no suitable candidate platform within the one-hop neighbors, the search is extended to the two-hop neighbors, with a maximum search radius not exceeding three hops.

[0021] Secondly, the present invention provides a distributed dynamic task allocation system for a heterogeneous unmanned system cluster, comprising a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the aforementioned distributed dynamic task allocation method for a heterogeneous unmanned system cluster is implemented.

[0022] By adopting the above technical solution, a computer program is generated from the above-mentioned method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster and stored in a memory for loading and execution by a processor. Terminal devices are then created based on the memory and processor for convenient use.

[0023] The technical solution of the present invention has the following beneficial technical effects: This invention establishes a coupled mapping relationship between the decay gradient vector and the sensitivity weight vector, precisely focusing mismatch risk assessment on the core contradiction: whether the capability dimension currently decaying on the unmanned platform happens to be the capability dimension highly dependent on by the current task. This mapping mechanism enables unmanned platforms with different capability degradation modes to generate differentiated risk assessment results when facing the same task, thereby achieving mismatch risk localization at the algorithmic level.

[0024] Furthermore, this invention employs an adaptive mutation detection and continuous confirmation mechanism based on real-time statistical characteristics of heterogeneous unmanned system clusters, ensuring that the triggering timing of task reassignment is objectively driven by the actual operational status of the heterogeneous unmanned system cluster. Once the inflection point occurs where the capability decay trend of the unmanned platform reverses, the system can quickly confirm the mutation and initiate local cascaded reassignment. Compared to the unavoidable perception blind spots in traditional fixed-period reassignment schemes, this invention reduces perception latency through real-time dynamic monitoring, preventing cascaded mismatch between the unmanned platform and tasks caused by perception lag. Attached Figure Description

[0025] Figure 1 This is a flowchart of a method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to the present invention; Figure 2 This is a schematic diagram of the signal effect in decay gradient abrupt detection. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] The hardware environment of this embodiment includes a heterogeneous cluster of nine platforms consisting of six rotary-wing UAVs, two wheeled unmanned vehicles, and one unmanned surface vessel. This cluster simultaneously undertakes three types of tasks: area reconnaissance, target tracking, and material delivery, operating in a continuously dynamic environment. This embodiment of the invention discloses a method for dynamically allocating distributed tasks in a heterogeneous unmanned system cluster, referring to... Figure 1 This includes steps S1-S4: S1: Real-time acquisition of multidimensional capabilities and decay gradient modeling.

[0028] Preferably, as an example, the multi-dimensional capability status of each unmanned platform in the cluster is collected in real time. Normalization preprocessing is performed on the raw sampled values ​​of each capability dimension. Linear regression is then performed on the normalized capability status time series based on a sliding time window to extract the decay gradients of each capability dimension, which are combined into a decay gradient vector, including: In an optional embodiment, a real-time acquisition channel for multi-dimensional capability status is established for each unmanned platform in the cluster. For the aforementioned nine-platform heterogeneous cluster, the capability dimension set for rotary-wing UAVs is defined as five dimensions, including remaining energy ratio, payload availability, positioning accuracy, communication link quality, and sensor detection performance. The capability dimension set for wheeled unmanned vehicles is defined as four dimensions, including remaining energy ratio, actual maximum speed of the wheel set, positioning accuracy, and communication link quality. The capability dimension set for unmanned surface vessels is defined as five dimensions, including remaining energy ratio, inertial navigation system positioning accuracy, communication link quality, sea state adaptability margin, and sensor detection performance. Each platform's sensors continuously report raw data at a acquisition frequency no less than one-tenth of the platform's control loop frequency. Taking a rotary-wing UAV with a control loop frequency of 50 Hz as an example, its capability status acquisition frequency is set to 5 Hz, meaning raw data for each dimension is acquired every 0.2 seconds.

[0029] Furthermore, before performing decay gradient calculations, normalization preprocessing is performed on the raw sampled values ​​for each capability dimension. First, the design full-scale rated value for each capability dimension of each platform is obtained; this value is given by the platform's manufacturer specifications and is a fixed, known quantity. The raw sampled value of each capability dimension at the current moment is divided by the corresponding dimension's design full-scale rated value to obtain the range of values. The dimensionless capability state value.

[0030] Specifically, the dimensionless capability state value satisfies the following relationship:

[0031] in, For the first The first platform Each capability dimension at any time The original sampled values; This is the rated full-scale value for the corresponding dimension; This is the normalized dimensionless capability state value.

[0032] Understandably, this formula is a standard linear normalization, a well-known preprocessing method. After this processing, all five capability dimensions of rotary-wing UAVs, four capability dimensions of unmanned vehicles, and five capability dimensions of unmanned surface vessels are mapped to... The interval. The units are unified as dimensionless ratios, which eliminates the non-additivity problem between different physical units and lays a physically legitimate foundation for the subsequent cross-dimensional weighted summation of gradients.

[0033] To more clearly illustrate the role and calculation process of the dimensionless capability state value, the embodiments of the present invention will be explained through specific examples below: Taking the remaining energy dimension of a certain rotary-wing drone as an example, its designed full-scale rated capacity is 22000 mAh. The currently measured remaining capacity is 15840 mAh. Substituting these values ​​into the formula for calculation... .

[0034] Specifically, for each capability dimension of each platform, a decay gradient estimate is constructed within a sliding time window. The dimensionless capability state values ​​of the most recent preset number of sampling points are used to construct a capability state time series. A first-order least-squares linear regression is performed on this capability state time series, and the regression slope is extracted as the instantaneous decay gradient for that capability dimension.

[0035] Specifically, the instantaneous decay gradient satisfies the following relationship:

[0036] in, For the first The first platform Each capability dimension at any time The decay gradient, in units of ; For the first in the sliding window Each sampling time; This represents the dimensionless capability state value at the corresponding moment; This represents the number of sampling points within the sliding window.

[0037] The theory behind the construction of the above relation is as follows: The basic expression for the least squares method of univariate linear regression in existing statistics is as follows:

[0038] In the formula, Represents the slope of the regression line; Represents the total number of sample points; The first independent variable One observation value; The first variable representing the dependent variable One observation value.

[0039] Considering the real-world dynamic and complex environments in which heterogeneous unmanned system swarms operate, the static capability values ​​acquired in a single instance are often subject to sampling noise interference and limited by the fixed-period perception lag, making it difficult to continuously characterize the true degradation trend of platform capabilities over time. To overcome the perception blind spot and achieve a better engineering balance between the smoothness of gradient estimation and the sensitivity of degradation response, a sliding time window mechanism is introduced to constrain the discrete sampling sequence. Simultaneously, a cross-disciplinary parameter mapping relationship is established between the real physical time series and the underlying statistical equations: the total number of independent observation samples in the aforementioned basic formula... The mapping is the number of sampling points covered by the control loop cycle of the unmanned platform body in this invention. , the observed values ​​of the independent variable Mapped to discrete sampling times within the sliding window of this invention The observed values ​​of the dependent variable The mapping is the dimensionless capability state value after eliminating the non-additivity of physical quantities in this invention. Thus, the instantaneous decay gradient relationship for the degradation rate perception of heterogeneous unmanned platforms was constructed.

[0040] Since all dimensionless capability state values ​​are dimensionless ratios, the dimensions of the instantaneous decay gradient are unified as follows: The gradients of each dimension are physically additive. When the instantaneous decay gradient is negative, it indicates that the capability dimension is degrading. The larger the absolute value, the faster the degradation.

[0041] Understandably, the value of the sliding window length is determined by the number of sampling points covered by a preset multiple of the platform's control loop cycle, aiming to achieve a balance between the smoothness of gradient estimation and response sensitivity. Specifically, the engineering basis for this range is as follows: Firstly, the window duration must cover at least 10 control loop cycles to ensure that the linear regression sample size is sufficient to suppress the interference of single-sampling noise on gradient estimation. The standard error of the slope estimation for a 10-point linear regression is approximately the square root of the standard deviation of the single-point noise. The noise suppression effect is significant, reaching 0.316 times the normal value. However, the window duration should not exceed 20 control loop cycles; otherwise, the capability changes within the window will span multiple dynamic stages, and the linear regression slope will reflect the mixed average of multiple trends rather than the current instantaneous decay rate, leading to gradient estimation lag. A range of 10 to 20 times achieves an engineering balance between smoothness and response sensitivity; the specific value can be adjusted within this range based on the platform's measured noise level. For a rotary-wing UAV with a control loop frequency of 50 Hz and a capability sampling frequency of 5 Hz, the number of sampling points should be... The value is 15, corresponding to a three-second sliding time window.

[0042] To more clearly illustrate the role and calculation process of the instantaneous decay gradient, the embodiments of the present invention will now be explained through specific examples: If the dimensionless remaining energy ratio sequence of the most recent 15 sampling points of a certain rotary-wing UAV shows an approximately linear decreasing trend from 0.72 to 0.68, substituting it into the above formula for calculation, The results indicate that the platform's energy is decaying at a rate of 1.33 percentage points per second.

[0043] The instantaneous decay gradients of all capability dimensions of each platform are combined to form the decay gradient vector of that platform. .in This indicates the total number of capability dimensions for this type of platform.

[0044] Understandably, this decay gradient vector will serve as the core input for subsequent steps in determining platform capability abrupt changes and triggering preventative reallocation. For the aforementioned rotary-wing UAV, its decay gradient vector is five-dimensional. For wheeled unmanned vehicles, its decay gradient vector is four-dimensional. For unmanned surface vessels, its decay gradient vector is five-dimensional.

[0045] Thus, this step extracts the instantaneous decay gradient through linear regression with a sliding time window, compressing the system's perception delay of the capability degradation rate to the duration of a single sliding time window, which shortens the perception blind zone by an order of magnitude compared to the fixed-period redistribution scheme.

[0046] S2: Capability coupling matrix construction and decay sensitivity mapping.

[0047] Preferably, as an example, based on the sensitivity weight vector of each task to the capability dimension and the decay gradient vector of each platform, the decay sensitivity coupling index of each unmanned platform relative to each task is calculated, including: In an optional embodiment, for each task to be assigned in the current task pool of the cluster, a pair is defined. Sensitivity weight vector for each capability dimension ,satisfy , Different types of missions exhibit parallel and differentiated sensitivities across various capability dimensions. Area reconnaissance missions are highly sensitive to endurance and sensor detection performance, but have almost no dependence on payload delivery capabilities; target tracking missions are highly sensitive to positioning accuracy and communication link quality; and material delivery missions are highly sensitive to payload availability and positioning accuracy, but have a relatively high tolerance for communication link quality.

[0048] Specifically, the initial values ​​of the sensitivity weight vector are determined in the following two ways: First, if the system's historical task database contains records of failed execution cases for the same type of task, the trigger frequency of each capability dimension in failure events is calculated. The trigger frequency is the proportion of the normalized capability state value of that dimension falling below a preset lower threshold at the time of task failure, relative to the total number of failures for that type of task. This frequency proportion is then normalized and used as the initial weight for the corresponding dimension. Second, if the system's historical task database lacks sufficient historical failure case data, a uniform distribution is used as the initial weight for each capability dimension. This method relies entirely on a subsequent dynamic environmental correction mechanism to drive weight differentiation.

[0049] After establishing the initial weights, the sensitivity weight vector is updated through a dynamic environmental correction mechanism. When real-time environmental parameters of the task execution area, such as wind speed, sea state, and terrain complexity, exceed the preset standard operating condition baseline for that task type, the sensitivity weight of the capability dimension most affected is proportionally increased. The ratio of the current measured value of the real-time environmental parameter to the standard operating condition baseline value is used as an amplification factor. After multiplicatively amplifying the weights of the affected dimensions, the entire weight vector is normalized to ensure that the sum of the weights of all dimensions is always one. After completing the sensitivity weight vector update, in order to assess the risk level of the platform's decaying capability being the capability most relied upon by the current task, the absolute value of the platform's decay gradient needs to be weighted with the sensitivity weight of the corresponding task, thus leading to the mathematical requirement of calculating the decay sensitivity coupling index.

[0050] Specifically, the decay sensitivity coupling exponent satisfies the following relationship:

[0051] in, For the first Each platform at any time The decay sensitivity coupling index relative to the task, in units of ; For the task to the first Sensitivity weights for each capability dimension; For the first The first platform The absolute value of the normalized decay gradient for each capability dimension; This represents the total number of capability dimensions.

[0052] Understandably, since the preliminary steps have already performed normalization preprocessing on all capability dimensions, the gradient dimensions of each dimension are consistent, and the weighted summation is physically and logically valid. The larger the decay sensitivity coupling index value, the more it means that the capability that the platform is rapidly decaying is precisely the capability that the current task relies on most, that is, the higher the risk of mismatch between the platform and the task.

[0053] To more clearly illustrate the role and calculation process of the decay sensitivity coupling index, the following embodiments of the present invention will provide specific examples: Taking a regional reconnaissance mission as an example, if a total of 47 regional reconnaissance missions failed in the historical database, the following failures were identified: 21 times due to endurance (0.447 frequency); 15 times due to sensor detection performance (0.319 frequency); 7 times due to positioning accuracy (0.149 frequency); and 4 times due to communication link quality (0.085 frequency). The normalized initial weight vector, ordered by endurance, payload, sensor, communication, and positioning dimensions, is as follows: When the measured wind speed is 12 m / s and the baseline wind speed under standard operating conditions is 8 m / s, the amplification factor is 1.5, and the current weight of the endurance dimension is multiplied by 1.5. The weight of the endurance dimension is multiplied by 1.5 to obtain 0.671, while the weights of other dimensions remain unchanged. After normalization, the entire sensitivity weight vector is updated as follows: The normalized values ​​accurately reflect the significantly enhanced decisive impact of endurance on mission success or failure under high wind speed conditions. At this time, a rotary-wing UAV is performing an area reconnaissance mission, and its normalized decay gradient vector is... Among these, the battery life rapidly declines, while other dimensions remain relatively stable. The sensitivity weight vector for reconnaissance missions, initialized using historical data and corrected for the current environment, is... ,but:

[0054] This decay sensitivity coupling index will be compared horizontally with the same index of other platforms in the cluster to identify the platform and task pairing combination with the highest mismatch risk.

[0055] Thus, this step, by directionally associating platform-side capability decay information with task-side capability dependency information, enables risk assessment to focus on whether the decaying capability happens to be the capability that the current task depends on, thereby achieving the location of mismatch risk.

[0056] S3: Preventive redistribution triggering based on decay gradient mutation detection.

[0057] Preferably, as an example, an adaptive mutation threshold based on cluster statistical characteristics is established for the time difference rate of change of the decay sensitivity coupling index. After confirmation over multiple consecutive calculation cycles, a preventative reallocation is triggered, including: In an optional embodiment, a second-order variation characteristic of the decay sensitivity coupling index for each platform-task pairing is established for its time series. To assess the acceleration of mismatch risk, the rate of change of the decay sensitivity coupling index over time between two adjacent computation periods is calculated.

[0058] Specifically, the rate of change of time difference satisfies the following relationship:

[0059] in, This represents the rate of change of time difference, which physically characterizes the acceleration of platform mismatch risk. The decay sensitivity coupling index represents the current calculation cycle. This represents the decay sensitivity coupling index of the previous calculation cycle; The calculation period interval for the decay sensitivity coupling index is set to [value] in this embodiment. Consistent with the capacity sampling period.

[0060] Specifically, to determine whether this accelerated deterioration constitutes a mutation event requiring preventative reallocation, an adaptive mutation threshold based on cluster statistical characteristics is established. The temporal difference rate of change of all active pairings executing tasks within the cluster is taken to form the current moment's risk acceleration sample set, denoted as... To prevent extreme outliers from self-inflating the threshold calculation itself, the mean is calculated... and standard deviation In this case, a truncated statistical method is employed. After removing the sample with the largest value from the risk acceleration sample set, the mean and standard deviation are calculated based on the remaining truncated sample set. This truncation method is statistically equivalent to performing an outlier test on a single maximum value, ensuring that the threshold is determined by the normal background noise of the cluster, rather than being contaminated by the outlier itself. The threshold is then applied to the entire set, including the removed sample, for judgment. To establish the numerical boundary that triggers preventive reallocation, a mutation threshold is introduced.

[0061] Specifically, the mutation threshold satisfies the following relationship:

[0062] in, Indicates the mutation threshold; This is the truncated sample set after removing the maximum value; This represents the mean of the truncated risk acceleration sample set after removing the maximum value; It represents the standard deviation of the truncated risk acceleration sample set after removing the maximum value.

[0063] When the number of active pairs in a cluster is less than four, the statistical significance after truncation is insufficient due to the small sample size, degenerating into a fixed threshold mode. The mean of the rate of change of time difference within the historical sliding time window of the platform, plus three times the standard deviation, is taken as the individual adaptive threshold to ensure detection effectiveness in small cluster scenarios.

[0064] Understandably, this mutation threshold is set based on the three-sigma criterion in statistics. Under the assumption of normal distribution, the probability of an observation exceeding the mean plus three standard deviations is no more than 0.27%, which can be considered a significant anomalous event. In actual heterogeneous cluster operation, the risk acceleration of most pairs will fluctuate around a certain mean, and only platforms that have truly entered the inflection point of capability decay will produce anomalous accelerations that significantly deviate from the mean.

[0065] To more clearly illustrate the role and calculation process of the mutation threshold, the embodiments of the present invention will be explained through specific examples below: Taking the nine-platform cluster of this embodiment as an example, assume that there are eight active pairings in the cluster at a certain moment. The sample set of its time difference rate of change is as follows: All units are The last value was significantly high, indicating that the endurance of a certain rotary-wing drone had entered a point of accelerated decline. Using a truncated statistical method, the maximum value was removed. Then, based on the remaining seven samples Perform the calculation. The mean is obtained. The standard deviation is Substituting into the formula, the mutation threshold is calculated. At this time, the time difference rate of change of the rotary-wing UAV is... Significantly exceeding the mutation threshold The continuous confirmation mechanism is then initiated. Compared to the situation where the threshold is raised to 0.000717 by outliers without truncation, preventing triggering, the truncation statistical method effectively eliminates the self-contamination effect of outliers on the threshold.

[0066] Furthermore, to prevent false triggering caused by noise in a single sampling, a continuous confirmation mechanism is introduced: a confirmation window is set, and the length of the confirmation window is defined as follows. The requirement is to continuously reach the confirmation window length, i.e., continuously. Preventative reallocation is only formally triggered when the rate of change of time difference exceeds the mutation threshold within each calculation cycle. (Confirmation window length) The value is determined based on: the time span of the decay gradient sliding time window. The number of computation cycles covered. The confirmation window's time span is taken as the sliding window's time span. The engineering basis is: the confirmation window needs to be short enough to ensure the response speed of preventive reallocation; if the confirmation window exceeds the duration of the sliding time window... In the worst-case scenario, the total delay from the capability inflection point to triggering preventative reallocation will exceed 1.5 sliding time windows. This is comparable to the perception blind spot of a fixed-cycle scheme, thus losing its preventative advantage. The confirmation window also needs to be long enough to filter out transient noise. In actual measurements, the duration of abnormal peak values ​​in the time difference rate of change caused by single-sample noise is usually no more than 2 to 3 calculation cycles. The sliding time window duration can effectively distinguish between real abrupt changes and noise spikes. Taking a rotary-wing UAV as an example, if the sliding time window is 3 seconds, then the confirmation window time span is 1 second. With a calculation cycle interval of 0.2 seconds, the length of the confirmation window... The confirmation window length corresponds to five calculation cycles. That is, after five consecutive calculation cycles confirm that the rate of change of the time difference is greater than the mutation threshold, a preventive reallocation is formally triggered.

[0067] After triggering preventative reallocation, instead of a global reallocation across the entire cluster, a local cascaded reallocation is initiated: Centered on the mutated platform, within the one-hop neighbor range of its communication topology neighbors, a search is performed for candidate platforms capable of undertaking the task and possessing the lowest decay sensitivity coupling index. If no suitable candidate platform is found within the one-hop neighbors, the search is extended to the two-hop neighbors, with a maximum search radius not exceeding three hops, to control the communication overhead of preventative reallocation. Taking a nine-platform cluster as an example, assume the mutated rotorcraft UAV's one-hop communication neighbors are three other rotorcraft UAVs and one unmanned vehicle. First, the search is performed among these four communication topology neighbors to find the candidate platform with the lowest decay sensitivity coupling index. If all four communication topology neighbors are unable to undertake the task due to excessive workload or insufficient capability to meet the minimum requirements, the search is extended to the two-hop neighbors, up to a maximum of three hops.

[0068] It should be noted that the method for obtaining the calculation period interval includes: reading the sampling frequency configuration parameter of the platform capability status; taking the reciprocal of the sampling frequency and defining it as the calculation period interval. In this embodiment, this calculation period interval is completely consistent with the capability sampling period.

[0069] Thus, this step, through an adaptive mutation detection mechanism based on cluster statistical characteristics, ensures that the timing of preventive reallocation is entirely determined by the objective reality, eliminating the perception blind spot where the capability snapshot lags behind the capability reality. After the capability decay inflection point appears, mutation confirmation and reallocation can be quickly completed.

[0070] S4: Coordinated execution and closed-loop verification of local cascaded redistribution.

[0071] Preferably, as an example, local cascaded reassignment is performed within the communication topology neighbor range, centered on the mutation platform. Task continuity is ensured through a shadow task mechanism, and closed-loop verification is performed after the reassignment is completed, including: In an optional embodiment, after the mutation detection mechanism triggers a local cascaded reallocation, the mutation platform broadcasts a task transfer request message to its communication topology neighbors. This message includes a task description of the task to be transferred, an estimate of the remaining workload of the task, and a sensitivity weight vector for the task. Upon receiving the request, each neighboring platform calculates its expected decay sensitivity coupling index after taking on the task, based on its current capability state and decay gradient. The neighboring platform then sends this expected decay sensitivity coupling index back to the mutation platform as a bidding offer.

[0072] Specifically, after collecting the bids from all neighbors, the mutation platform selects the platform with the smallest expected decay sensitivity coupling index as the task recipient. To determine whether the differences in bids from multiple candidate platforms are within the range of statistical noise, a sample set of coupling indices reflecting the overall operating benchmark of the cluster needs to be maintained simultaneously, denoted as . Then, the standard deviation of the sample set of the coupling index is calculated and denoted as . The standard deviation has the following dimensions: The standard deviation is consistent with the dimensions of the bid price. This standard deviation provides a statistical noise benchmark reflecting the fluctuations in the current cluster state.

[0073] To more clearly illustrate the role and calculation process of the coupling index sample set, the following embodiments of the present invention will provide specific examples: If the bid prices of the three candidate platforms are as follows: , , The current cluster standard deviation, calculated using the coupling index sample set, is The biggest difference between the three is The difference is less than the standard deviation. Therefore, it was determined that the differences among the three were within the range of statistical noise. Instead, the current remaining task load of the three platforms was compared, and the platform with the lightest load was selected to take on the task.

[0074] If the maximum difference among the expected decay sensitivity coupling indices of multiple candidate platforms is less than the standard deviation of the sample set of these coupling indices, then these candidate platforms are considered to have no statistically significant difference in terms of mismatch risk. In this case, the current remaining task load of these candidate platforms is compared, and the platform with the lightest current remaining task load is selected to achieve load balancing. The dimensions of this decision boundary are completely self-consistent. and Direct comparison, and adaptively determined by real-time statistical characteristics of the cluster, without the need for any manual scaling factors.

[0075] Furthermore, during the mission transfer process, a shadow mission mechanism is employed to ensure mission continuity. Before the new receiving platform completes its mission takeover preparations, including route planning updates and sensor mode switching, the original mutation platform continues to execute the mission in degraded mode. The degree of degrade is objectively determined by the real-time capability status of the mutation platform itself, using the normalized value of the key capability dimension at the moment of triggering preventive reallocation as the benchmark, and the ratio of the corresponding normalized value at the current moment to that value as the degrade coefficient for mission execution.

[0076] Specifically, the downgrade coefficient satisfies the following relationship:

[0077] in, This indicates the power degradation ratio of tasks executed on the platform under the shadow task mechanism; This represents the normalized value of the key capability dimension at the current moment; This represents the normalized value of the key capability dimension at the moment when preventive reallocation is triggered.

[0078] Understandably, the platform executes tasks at its rated power multiplied by this degradation factor. This degradation factor directly binds the extent of the original platform's degradation to its real-time capability status, ensuring that tasks do not experience execution gaps while keeping energy consumption at a level that matches the current capability status.

[0079] Understandably, the degradation coefficient is calculated only if the normalized value of the critical capability dimension is greater than zero at the time of triggering preventative reallocation; this prerequisite is physically guaranteed by the mutation triggering logic. A minimum usable threshold is set for each capability dimension. This value is given by the platform's factory specifications. Taking the energy dimension of a rotary-wing drone as an example, 0.1 corresponds to a 10% battery protection threshold. When the normalized value of any key capability dimension falls below this minimum available threshold, the platform's forced exit mechanism is directly triggered. This forced exit mechanism has higher priority than the mutation detection process. Therefore, for a platform to enter the shadow mission process, its key capability dimension normalized value must meet the following requirement. This mechanism ensures from a physical standpoint that the denominator in the degradation coefficient calculation formula is never zero, eliminating the risk of algorithm crash and eliminating the need to introduce any numerical compensation terms.

[0080] To more clearly illustrate the function and calculation process of the degradation coefficient, the following embodiments of the present invention will be illustrated with specific examples: If the normalized value of the critical capability dimension is 0.65 at the time of triggering preventative reallocation, and this value is significantly higher than the minimum availability threshold of 0.1, it indicates that the platform is still available; however, the normalized value of the critical capability dimension has now dropped to 0.55. Substituting the above data into the degradation coefficient formula, the degradation coefficient is calculated to be 0.846. In engineering terms, this means the platform is executing the task at 84.6% of its rated power. Once the new platform confirms its readiness to take over, it completes the atomic transfer of task control through a handshake protocol, and the original mutated platform exits the execution of the task.

[0081] In an optional embodiment, the closed-loop verification mechanism is initiated after the task transfer is completed. It continuously monitors the expected decay sensitivity coupling index of the newly acquired platform. If, within a complete sliding time window period after the transfer (in this embodiment, the complete sliding time window period is 3 seconds), the new platform's expected decay sensitivity coupling index does not show a mutation trend exceeding the mutation threshold, the reassignment is considered successful. Simultaneously, the task's reassignment flag is turned off. If the new platform also shows a mutation trend, the task is upgraded to a high-risk task, and the search scope is expanded to include two-hop neighbors to re-execute the bidding process. Simultaneously, the task's persistent mismatch status is reported to the cluster management node, which then assesses whether the task's execution parameters need adjustment, such as reducing task accuracy requirements or extending task completion time, to adapt to the cluster's current overall capability level.

[0082] The entire dynamic allocation process is continuously executed in a loop during cluster operation. S1 continuously collects, normalizes, and updates the decay gradient vector; S2 continuously updates the decay sensitivity coupling index and the coupling index sample set; S3 continuously performs truncated mutation detection and triggers preventive reallocation after confirmation over multiple consecutive computation cycles; S4 performs local cascaded reallocation upon triggering and completes closed-loop verification within a full sliding time window. This cyclical mechanism ensures that cluster task allocation always tracks the real-time decay status of the platform's capabilities, completing preventive adjustments before cascaded mismatches occur.

[0083] It should be noted that the method for obtaining the minimum available threshold includes: consulting the factory specifications or engineering operation manuals of each unmanned platform, extracting the safety protection limit values ​​of each capability dimension, and defining the normalized value corresponding to the safety protection limit value as the minimum available threshold.

[0084] In this way, this step directly binds the degradation level of the original mutation platform to its real-time capability status through the shadow task mechanism, eliminating the execution vacuum during the task transfer transition period and ensuring the task continuity of the cluster during the redistribution transition.

[0085] Figure 2 The figure illustrates the signal effect of decay gradient abrupt change detection. It includes a measured data curve reflecting the real-time fluctuation of risk acceleration over time, and a horizontal baseline dashed line representing the set abrupt change threshold. The trend in the figure shows that for most of the time period, risk acceleration remains at a low level within normal background fluctuations, with its extreme values ​​consistently suppressed below the abrupt change threshold. However, at a specific moment, the measured data curve suddenly exhibits a sharp upward spike, representing a step-like increase in the acceleration of mismatch risk. This spike significantly exceeds the set abrupt change threshold baseline. As indicated in the figure, when the abnormal spike in risk acceleration significantly exceeds the abrupt change threshold, the system successfully captures the abrupt change signal and executes the corresponding trigger redistribution action command. This figure visually demonstrates that introducing a abrupt change threshold can effectively filter normal background noise and trigger preventative redistribution at the first sign of capability decay inflection points.

[0086] This invention also discloses a distributed task dynamic allocation system for heterogeneous unmanned system clusters, including a processor and a memory. The memory stores computer program instructions, and when the computer program instructions are executed by the processor, a distributed task dynamic allocation method for heterogeneous unmanned system clusters according to the present invention is implemented.

[0087] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0088] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster, characterized in that, include: The system collects multidimensional capability status data for each unmanned platform in the cluster in real time. Normalization preprocessing is performed on the raw sampled values ​​of each capability dimension. Linear regression is then performed on the normalized capability status time series based on a sliding time window to extract the decay gradient of each capability dimension, which is then combined into a decay gradient vector. Based on the sensitivity weight vector of each task to the capability dimension and the decay gradient vector of each platform, the decay sensitivity coupling index of each unmanned platform relative to each task is calculated. An adaptive mutation threshold based on cluster statistical characteristics is established for the time difference change rate of the decay sensitivity coupling index. After confirmation over multiple consecutive calculation cycles, preventative reallocation is triggered. Local cascaded reallocation is performed within the communication topology neighbor range, centered on the mutated platform. Task continuity is ensured through a shadow task mechanism, and closed-loop verification is performed after reallocation is completed.

2. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 1, characterized in that, The normalization preprocessing of the original sampled values ​​of each capability dimension includes: obtaining the design full-scale rated value of each capability dimension of each unmanned platform; dividing the original sampled value of each capability dimension at the current moment by the design full-scale rated value of the corresponding dimension to obtain a dimensionless capability state value with a value range of [0,1].

3. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 1, characterized in that, The method of performing linear regression on the normalized capability state time series based on a sliding time window to extract the decay gradient of each capability dimension includes: taking the dimensionless capability state values ​​of the most recent preset number of sampling points to form a time series; performing first-order least squares linear regression on the time series and extracting the regression slope as the instantaneous decay gradient of the capability dimension; the sliding window length is the number of sampling points covered by a preset multiple of the control loop period of the unmanned platform body.

4. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 1, characterized in that, The step of calculating the decay sensitivity coupling index of each unmanned platform relative to each task based on the sensitivity weight vector of each task to the capability dimension and the decay gradient vector of each platform includes: defining a sensitivity weight vector for each task to each capability dimension, wherein the sum of the weights of each dimension of the sensitivity weight vector is 1; and performing a weighted inner product of the absolute value of the decay gradient of each unmanned platform with the sensitivity weight of the corresponding task to obtain the decay sensitivity coupling index.

5. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 4, characterized in that, The definition of sensitivity weight vectors for each capability dimension for each task includes: if there are records of execution failure cases of the same type of task in the system's historical task database, the trigger frequency of each capability dimension in the failure event is counted, and the trigger frequency ratio is normalized and used as the initial weight of the corresponding dimension; if there is insufficient historical failure case data in the system's historical task database, the initial weights of each capability dimension are uniformly distributed.

6. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 5, characterized in that, The sensitivity weight vector is also updated through an environmental dynamic correction mechanism, including: when the real-time environmental parameters of the task execution area exceed the preset standard operating condition baseline for the task type, the ratio of the measured value of the current environmental parameter to the standard operating condition baseline value is taken as an amplification factor; after multiplicatively amplifying the weights of the affected dimensions, the entire weight vector is normalized so that the sum of the weights of all dimensions is always 1.

7. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 1, characterized in that, The establishment of an adaptive mutation threshold based on cluster statistical characteristics for the time difference change rate of the decay sensitivity coupling index includes: calculating the time difference change rate of the decay sensitivity coupling index between two adjacent calculation cycles; taking the time difference change rates of all active unmanned platforms paired with tasks within the cluster to form a risk acceleration sample set; removing the sample with the largest value from the risk acceleration sample set to obtain a truncated sample set, calculating the mean and standard deviation of the truncated sample set, and setting the mean plus three times the standard deviation as the adaptive mutation threshold; if the number of samples in the risk acceleration sample set is less than a preset number threshold, it degenerates into using a fixed threshold mode within the historical sliding time window of the target unmanned platform as the mutation threshold.

8. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 7, characterized in that, The step of triggering preventive reallocation after confirmation through multiple consecutive calculation cycles includes: setting a confirmation window, the length of which is the number of calculation cycles covered by a preset proportion of the decay gradient sliding window time span; and requiring that the rate of change of time difference exceeds the mutation threshold within the calculation cycles that continuously reach the confirmation window length before formally triggering preventive reallocation.

9. The method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to claim 1, characterized in that, The process of performing local cascaded redistribution within the communication topology neighbor range, centered on the mutation platform, includes: the mutation platform broadcasting a task transfer request message to its communication topology neighbors; neighboring platforms calculating the decay sensitivity coupling index assumed after taking over the task, based on their current decay gradient vector and the sensitivity weight vector of the task, as the expected decay sensitivity coupling index and bidding price; selecting the platform with the smallest bidding price as the task recipient; if the difference in bidding prices among multiple candidate platforms is less than the standard deviation of the current expected decay sensitivity coupling index sample set of the cluster, then selecting the platform with the lightest remaining task load; if there is no suitable candidate platform within the one-hop neighbor range, the search is extended to the two-hop neighbor range, with a maximum search radius not exceeding three hops.

10. A distributed task dynamic allocation system for heterogeneous unmanned system clusters, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions, which, when executed by the processor, implement a method for dynamic allocation of distributed tasks in a heterogeneous unmanned system cluster according to any one of claims 1-9.

Images