WO2019127946A1

WO2019127946A1 - Learning genetic algorithm-based multi-task and multi-resource rolling distribution method

Info

Publication number: WO2019127946A1
Application number: PCT/CN2018/080420
Authority: WO
Inventors: 邢立宁; 何敏藩; 白国庆; 吕欣; 王炯琦; 伍国华; 熊彦; 文翰; 甘文勇; 黄勇
Original assignee: 佛山科学技术学院; 佛山市有义家科技有限公司
Priority date: 2017-12-26
Filing date: 2018-03-26
Publication date: 2019-07-04
Also published as: CN108256671A

Abstract

A learning genetic algorithm-based multi-task and multi-resource rolling distribution method. According to the method, on the basis of a rolling time-domain control principle, a multi-task and multi-resource dynamic rolling distribution mechanism is constructed, comprising determining elements such as a prediction window, a rolling window, a distribution sub-problem, and a rolling mechanism. Task information is updated in real time by means of a current prediction window, and a current rolling window is determined on the basis of the prediction window. The distribution sub-problem is a local distribution problem constructed according to the current rolling window at each planning moment. The rolling mechanism is used for determining an execution position where a distribution scheme ends after the distribution sub-problem is solved and a next planning moment. By means of the rolling distribution mechanism, a complex dynamic distribution problem is converted into rolling updated static distribution problems. According to the method, by means of rolling dynamic planning, a complex dynamic distribution problem is decomposed into a plurality of simple and static scheduling problems, and then the optimal solutions of the sub-problems are combined to replace the optimal solution of the original problem, thereby reducing the solving difficulty of the original problem.

Description

一种基于学习型遗传算法的多任务多资源滚动分配方法Multi-task multi-resource rolling allocation method based on learning genetic algorithm

技术领域Technical field

本发明涉及遥感卫星技术领域，尤其涉及一种观测任务分配方法，特别是一种基于学习型遗传算法的多任务多资源滚动分配方法。The invention relates to the field of remote sensing satellite technology, in particular to an observation task allocation method, in particular to a multi-task multi-resource rolling allocation method based on a learning genetic algorithm.

背景技术Background technique

遥感卫星是(remote sensing satellite)用作外层空间遥感平台的人造卫星。用卫星作为平台的遥感技术称为卫星遥感。通常，遥感卫星可在轨道上运行数年。卫星轨道可根据需要来确定。遥感卫星能在规定的时间内覆盖整个地球或指定的任何区域，当沿地球同步轨道运行时，它能连续地对地球表面某指定地域进行遥感。所有的遥感卫星都需要有遥感卫星地面站，卫星获得的图像数据通过无线电波传输到地面站，地面站发出指令以控制卫星运行和工作。遥感卫星主要有气象卫星、“陆地卫星”和“海洋卫星”三种类型。Remote sensing satellites are artificial satellites used as remote sensing platforms for outer space. Remote sensing technology using satellites as a platform is called satellite remote sensing. Typically, remote sensing satellites can operate in orbit for several years. Satellite orbits can be determined as needed. Remote sensing satellites can cover the entire Earth or any designated area within a specified time. When operating along geosynchronous orbit, it can continuously remotely sense a designated area on the Earth's surface. All remote sensing satellites require a remote sensing satellite ground station. The image data obtained by the satellite is transmitted to the ground station via radio waves, and the ground station issues commands to control the satellite operation and operation. Remote sensing satellites mainly include three types: meteorological satellites, “land satellites” and “marine satellites”.

在未来十年到二十年里，我国可用的遥感卫星数目将急剧增加数百颗(如吉林一号卫星群，到2030年将实现138颗卫星在轨运行)，不同行业的众多用户每天提交的成像观测需求将突破数万条，如何将这些众多用户提交的大量成像任务有效地分配给不同遥感卫星，目前尚缺乏有效的理论和技术支撑。In the next ten to twenty years, the number of remote sensing satellites available in China will increase dramatically by hundreds (such as the Jilin No. 1 satellite group, which will achieve 138 satellites in orbit by 2030), and many users from different industries submit daily. The imaging observation needs will exceed tens of thousands. How to effectively allocate a large number of imaging tasks submitted by these many users to different remote sensing satellites, there is still no effective theoretical and technical support.

发明内容Summary of the invention

为了解决现有技术中的问题，本发明的目的是提供一种观测任务分配方法(多任务多资源动态滚动分配方法)，或者一种基于学习型遗传算法的多任务多资源滚动分配方法，所述方法采用多任务多资源动态滚动分配机制，通过把复杂的动态调度问题分解为多个简单的静态调度子问题，再对子问题的优化解进行组合，从而代替原问题的最优解，这样大大降低了原问题求解的难度。In order to solve the problems in the prior art, an object of the present invention is to provide an observation task allocation method (multi-task multi-resource dynamic scrolling allocation method), or a multi-task multi-resource rolling allocation method based on a learning genetic algorithm. The method adopts the multi-task multi-resource dynamic rolling allocation mechanism, and decomposes the complex dynamic scheduling problem into multiple simple static scheduling sub-problems, and then combines the optimal solutions of the sub-problems to replace the optimal solution of the original problem. Greatly reduce the difficulty of solving the original problem.

为此，本发明提供一种观测任务分配方法(多任务多资源动态滚动分配方法)，或者一种基于学习型遗传算法的多任务多资源滚动分配方法，所述方法采用层次化分布式自主协同任务架构，多星任务协调器将滚动窗口内的任务集合分配给下辖的多颗智能卫星，各智能卫星利用其星上调度器对被分配的新任务和已有任务进行调度，在当前滚动窗口的起始时刻，多星任务协调器对任务信息进行更新，删除上一滚动窗口内已经完成的任务以及在所述起始时刻正在执行的任务，并将上一滚动窗口内未分配的任务、以及在上一滚动窗口内到达的新任务组合成当前滚动窗口内的任务集合，且所述多星任务协调器将该任务集合向所述多颗智能卫星进行分配，其中，基于混合触发模式来确定滚动窗口的起始时刻，一方面，每隔一个时间段触发滚动分配，该时间段为恒定的或根据预先设定的规则动态变化；另一方面，在出现使***状态发生改变的事件或在受到人工干预时触发滚动分配。To this end, the present invention provides an observation task allocation method (multi-task multi-resource dynamic rolling allocation method), or a multi-task multi-resource rolling allocation method based on a learning genetic algorithm, which adopts hierarchical distributed autonomous coordination Task architecture, the multi-star task coordinator assigns the task set in the rolling window to a plurality of intelligent satellites under the jurisdiction, and each intelligent satellite uses its on-board scheduler to schedule the assigned new tasks and existing tasks, and currently scrolls. At the beginning of the window, the multi-star task coordinator updates the task information, deletes the tasks that have been completed in the previous scrolling window, and the tasks that are being executed at the starting time, and the unassigned tasks in the previous scrolling window. And the new tasks arriving within the last scrolling window are combined into a set of tasks within the current scrolling window, and the multi-star task coordinator allocates the set of tasks to the plurality of intelligent satellites, wherein the hybrid triggering mode is based To determine the starting moment of the scrolling window, on the one hand, triggering the scrolling assignment every other time period, the time period is constant Or the dynamic changes in the rules set in advance; on the other hand, when the system state changes or events triggered by rolling assigned when manual intervention.

优选地，一方面，所述时间段根据测控周期设置；另一方面，所述使***状态发生改变的事件包括：接收到应急观测任务，且积累的未分配应急观测任务达到五件或者是所述多星任务协调器下辖的智能卫星数的5％。Preferably, in one aspect, the time period is set according to a measurement and control cycle; on the other hand, the event that changes a state of the system includes: receiving an emergency observation task, and the accumulated unallocated emergency observation task reaches five pieces or is 5% of the number of intelligent satellites under the multi-star mission coordinator.

优选地，所述多星任务协调器包括地面站和地球静止轨道通信卫星，在测控周期之内，所述地面站进行任务分配；在测控周期之外，所述地球静止轨道通信卫星进行任务分配，且所述应急观测任务由所述智能卫星生成。Preferably, the multi-satellite task coordinator comprises a ground station and a geostationary orbit communication satellite, wherein the ground station performs task assignment within a measurement and control period; and the geostationary orbit communication satellite performs task assignment outside the measurement and control period And the emergency observation task is generated by the smart satellite.

优选地，各智能卫星的星上调度器的任务调度策略如下：Preferably, the task scheduling strategy of the onboard scheduler of each intelligent satellite is as follows:

(1)在T-驱动的调度时刻点，采用渐进式方法中的完全重调度策略，生成下一个周期时间区间内的新任务计划，T-驱动的调度时刻点是根据给定的时间间隔T来确定特定的调度时间点lT，0≤l≤L,LT≤H＜(L+1)T，每到达一个调度时间点lT，则计算生成后一调度区间[lT,(l+1)T]的任务计划，其中l为正整数，T为给定的时间间隔，L为最大T-驱动调度次数，H为总调度区间，(1) At the T-driven scheduling time point, the full rescheduling strategy in the progressive method is used to generate a new task plan in the next cycle time interval, and the T-driven scheduling time point is based on the given time interval T To determine a specific scheduling time point lT, 0 ≤ l ≤ L, LT ≤ H < (L + 1) T, each time a scheduling time point lT is reached, the calculation of the latter scheduling interval [lT, (l + 1) T Mission plan, where l is a positive integer, T is the given time interval, L is the maximum number of T-drive scheduling, and H is the total scheduling interval.

(2)在C ^*-驱动的重调度时刻点，采用修订式方法中的调度计划修复策略，当卫星运行在给定的调度区间内时，若在某一时刻t(0＜t＜H)，星上的应急观测任务累积量C _t超过给定的阈值C ^*时，则执行重调度计算，其中阈值C ^*为应急观测任务的临界累积数， (2) At the C ^* -driven rescheduling time point, the scheduling plan repair strategy in the revised method is adopted. When the satellite is operating in a given scheduling interval, if it is at a certain time t (0 < t < H) emergency satellite observation missions on the accumulated amount of C _t exceeds a given threshold value C ^*, calculation is performed rescheduling, wherein C ^* is the critical threshold cumulative emergency observation tasks,

除上述两种调度时刻点之外，不在任何其他时刻点进行调度。Except for the above two scheduling moments, scheduling is not performed at any other point in time.

优选地，在T-驱动的调度时刻点的调度算法如下：Preferably, the scheduling algorithm at the T-driven scheduling moment is as follows:

输入：Enter:

–已到达且在T-驱动调度时刻点之前未被调度的应急观测任务集合；

– a set of emergency observation tasks that have arrived and are not scheduled before the T-drive scheduling time point;

–已接收且在T-驱动调度时刻点之前未被调度的常规观测任务集合；

– a set of conventional observation tasks that have been received and are not scheduled before the T-Drive scheduling time point;

输出:Output:

--下一时间周期T内的调度计划；

- a scheduling plan within the next time period T;

具体步骤如下：Specific steps are as follows:

步骤11 分别从

和

中选取时间窗口是否落入下一个时间周期T内的常规观测任务和应急观测任务，生成待调度求解的常规观测任务集合

和应急观测任务集合

Step 11 separately from

with

Select whether the time window falls into the conventional observation task and the emergency observation task in the next time period T, and generate a conventional observation task set to be solved.

And emergency observation task set

步骤12 将

和

整合为一个观测任务集合； Step 12 will

with

Integrated into a collection of observation tasks;

步骤13 按照设定的启发式规则，对整合后的观测任务集合中的任务进行排序；Step 13 Sort the tasks in the integrated observation task set according to the set heuristic rules;

步骤14 按照排序，对所述整合后的观测任务集合中的任务一一进行调度，以确定是否将之加入到

中，直至所述整合后的观测任务集合中再无任务可加入

中， Step 14 According to the sorting, the tasks in the integrated observation task set are scheduled one by one to determine whether to join the tasks.

In the above, until the integrated observation task set has no more tasks to join

in,

步骤15 输出下一时间周期T内的调度计划

Step 15 Output the schedule in the next time period T

在C ^*-驱动的重调度时刻点的调度算法如下： The scheduling algorithm at the C ^* -driven rescheduling point in time is as follows:

输入：Enter:

—在本时间周期T内且晚于C ^*-驱动调度时刻点t的调度计划；

- a scheduling plan within this time period T and later than the C ^* -drive scheduling time point t;

—在调度时刻点t之前已到达且未调度的应急观测任务集合；

- a set of emergency observation missions that have arrived and are not scheduled before the scheduling time point t;

输出:Output:

—在时间t时已修订的调度计划，

- the revised schedule at time t,

具体步骤如下：Specific steps are as follows:

步骤21 根据观测时间窗口处于时间t到下一个T-驱动调度时刻点这一时间区间内的条件，从任务集合

中选取应急观测任务，生成新的任务集合

Step 21: According to the condition that the observation time window is in the time interval from the time t to the next T-drive scheduling time point, the task collection

Select emergency observation tasks to generate new task sets

步骤22 根据设定的启发式规则，对

中的应急观测任务进行排序； Step 22 According to the set heuristic rules,

Sort the emergency observation tasks in the middle;

步骤23 按照新的任务次序，一一选取

中的应急观测任务并对

进行修订，直至

中再无应急观测任务可加入

中， Step 23 Select one by one according to the new task order.

Emergency observation mission

Revise until

No more emergency observation tasks can be added

in,

步骤24 输出已修订的调度计划

Step 24 Output the revised schedule

优选地，多星任务协调器采用设定的滚动分配机制将滚动窗口内的任务集合分配给下辖的多颗智能卫星，所述滚动分配机制从用户偏好和任务场景特征信息中提取规则，并将所述规则应用于遗传算法的交叉操作和变异操作。Preferably, the multi-star task coordinator allocates a task set in the rolling window to a plurality of intelligent satellites under the control using a set scroll allocation mechanism, and the scroll allocation mechanism extracts rules from user preference and task scene feature information, and The rules are applied to the crossover and mutation operations of the genetic algorithm.

优选地，多星任务协调器采用设定的滚动分配机制将滚动窗口内的任务集合分配给下辖的多颗智能卫星，所述滚动分配机制包括以下步骤：Preferably, the multi-star task coordinator allocates a set of tasks in the rolling window to a plurality of intelligent satellites under the control using a set rolling allocation mechanism, and the rolling allocation mechanism comprises the following steps:

S1、适应度评价：采用基于TOPSIS方法的多目标函数适应度评价，同时在不同场景竞争度指标下存储并提取用户偏好，进行优化目标的权重分配；S1, fitness evaluation: multi-objective function fitness evaluation based on TOPSIS method, while storing and extracting user preferences under different scene competition indicators, and performing weight distribution of optimization targets;

S2、种群初始化：采用基于启发式规则的初始化种群生成，在提高初始种群质量基础上同时保证初始种群的随机分布；S2. Population initialization: Initial population generation based on heuristic rules is adopted to ensure the random distribution of the initial population while improving the initial population quality;

S3、选择操作：在对选择算子进行改进的同时，不断学习迭代过程中出现好解的个体构件组合知识、个体中任务的资源竞争度知识，并应用于接下来的交叉与变异操作；S3, selection operation: while improving the selection operator, continuously learn the individual component combination knowledge that has a good solution in the iterative process, the resource competition knowledge of the task in the individual, and apply to the following crossover and mutation operations;

S4、交叉操作：在个体构件组合知识的指导下，以不同概率进行交叉位置的选择，同时采用多次操作取最优的策略，保证交叉操作的有效性；S4, cross operation: under the guidance of individual component combination knowledge, the intersection position is selected with different probabilities, and multiple operations are used to take the optimal strategy to ensure the effectiveness of the cross operation;

S5、变异操作：在任务的资源竞争度知识的指导下，以不同概率选择个体变异操作位置，同时也采用多次操作取最优的策略，提高变异操作效率；S5. Mutation operation: Under the guidance of the resource competition knowledge of the task, the individual mutation operation position is selected with different probabilities, and the optimal strategy is adopted by multiple operations to improve the efficiency of the mutation operation;

S6、局部搜索：采用局部随机与确定搜索相结合的策略，其中确定性搜索基于个体平台剩余能力排序的任务交流规则进行；S6. Local search: a strategy combining local randomization and determined search, wherein the deterministic search is performed based on task communication rules of the remaining capabilities of the individual platform;

S7、种群替换：加入用户任务锁定规则生成替换种群，在保证用户对任务偏好提取的基础上，同时实现算法跳出局部最优的能力。S7. Replacing the population: adding the user task locking rule to generate the replacement population, and ensuring the user's ability to jump out of the local optimal while ensuring the user's extraction of the task preference.

本发明还提供一种基于学习型遗传算法的多任务多资源滚动分配方法，其采用滚动时域控制原理，构建多任务多资源动态滚动分配机制，包括确定预测窗口、滚动窗口、分配子问题和滚动机制要素；通过当前预测窗口对任务信息进行实时更新，在预测窗口的基础上确定当前滚动窗口；分配子问题是在每个规划时刻，根据当前滚动窗口构造的局部分配问题；滚动机制用于确定分配子问题求解后分配方案结束的执行位置和下一个规划时刻；通过滚动分配机制将复杂动态分配问题转化为滚动更新的静态分配问题。The invention also provides a multi-task multi-resource rolling allocation method based on a learning genetic algorithm, which adopts a rolling time domain control principle to construct a multi-task multi-resource dynamic scrolling allocation mechanism, including determining a prediction window, a rolling window, an allocation sub-problem and Scrolling mechanism elements; real-time updating of task information through the current prediction window, determining the current scrolling window based on the prediction window; the allocation sub-question is a local allocation problem constructed according to the current scrolling window at each planning moment; the scrolling mechanism is used for Determine the execution position and the next planning moment after the allocation sub-question is solved; and the complex dynamic allocation problem is transformed into the static allocation problem of rolling update through the rolling allocation mechanism.

优选的方案，通过当前预测窗口对任务信息进行实时更新的内容包括增加任务、删除任务或修改任务属性信息。In a preferred solution, the content of the task information updated in real time through the current prediction window includes adding a task, deleting a task, or modifying task attribute information.

进一步优选的方案，所述滚动分配机制是从用户偏好和任务场景特征信息中提取规则，并在优化过程中挖掘相关知识来指导后续优化过程，从而获得比较满意的优化效果。In a further preferred solution, the rolling allocation mechanism extracts rules from user preference and task scene feature information, and mines relevant knowledge in the optimization process to guide the subsequent optimization process, thereby obtaining a satisfactory optimization effect.

所述滚动式动态规划框架采用基于测控窗口为周期的滚动式动态调度方法，把复杂的动态调度问题分解为多个简单的静态调度子问题，再对子问题的优化解进行组合，从而代替原问题的最优解。The rolling dynamic programming framework adopts a rolling dynamic scheduling method based on the measurement and control window, and decomposes the complex dynamic scheduling problem into a plurality of simple static scheduling sub-problems, and then combines the optimized solutions of the sub-problems to replace the original The optimal solution to the problem.

所述构建动态任务状态更新机制是指在不同调度时刻及时更新任务状态，以获取更加完备的信息来指导规划调度；根据任务在不同时刻的状态可将任务分为已完成任务、正在执行任务、等待执行任务及新到达任务。The constructing the dynamic task state update mechanism refers to updating the task state in time at different scheduling moments to obtain more complete information to guide the planning and scheduling; according to the state of the task at different times, the task can be divided into completed tasks, executing tasks, Waiting for the task to be executed and the new arrival task.

所述滚动式动态任务规划触发模式包括周期触发模式、事件触发模式和混合触发模式。所述周期触发模式是针对常规任务每隔一段时间进行一次规划，该时间段为均匀恒定的或动态变化的，如测控周期触发或值班周期触发，基于值班周期触发模式能够保证规划频率的稳定性，操作实现较为简单，但是该模式当周期较长时，无法为高时效性任务提供及时的规划服务，且应对***状态变化的能力也是非常有限的；所述事件触发模式是指在出现使***状态发生改变的事件，或在受到人工干预时开始执行规划，如有新任务到达、卫星状态发生变化、决策部门提出规划需求等情况发生时，事件触发模式对规划环境具备较高的敏感性，能够及时处置应急任务，但由于规划频繁，当任务数量较多时，算法的时间复杂度较高，事件触发主要是任务触发模式；所述混合触发模式为事件触发与周期触发相结合的触发模式，它能够兼获两种触发模式的优点，具备及时处置应急任务的能力，并且算法的时间复杂度有限。多任务多资源滚动分配模式主要为以应急任务触发的对已有规划方案进行动态调整的动态重规划和以测控点周期性触发的滚动式任务规划的混合触发模式。The rolling dynamic task planning trigger mode includes a periodic trigger mode, an event trigger mode, and a hybrid trigger mode. The periodic trigger mode is planned for a regular task at regular intervals, and the time period is uniformly constant or dynamically changed, such as a measurement control cycle trigger or a duty cycle trigger, and the duty cycle trigger mode can ensure the stability of the planned frequency. The operation is relatively simple to implement, but the mode cannot provide timely planning services for high-aging tasks when the period is long, and the ability to cope with system state changes is also very limited; the event trigger mode refers to the system in the presence of the system. The event-triggered mode is highly sensitive to the planning environment when the state changes, or when the manual intervention is started, and the new task arrives, the satellite status changes, and the decision-making department proposes the planning requirements. The emergency task can be processed in time, but due to frequent planning, when the number of tasks is large, the time complexity of the algorithm is high, and the event triggering is mainly the task trigger mode; the mixed trigger mode is a trigger mode combining event triggering and periodic triggering. It combines the advantages of both trigger modes The ability to handle emergency tasks in a timely manner, and the time complexity of the algorithm is limited. The multi-task multi-resource rolling allocation mode is mainly a dynamic re-planning that dynamically adjusts an existing planning scheme triggered by an emergency task and a mixed triggering mode that is triggered by a rolling task task periodically triggered by a monitoring point.

所述滚动时间窗口的选取策略是将任务按照到达顺序划分为具有一定重叠、但随着规划时刻不断向前推进的任务集合，在每次规划时，仅对当前滚动窗口内的任务进行规划，随着规划时刻的推进，新任务被不断加入，而完成规划的任务则被逐渐删除，从而实现滚动窗口的更新。The selection strategy of the scrolling time window is to divide the tasks into a set of tasks that have certain overlaps but continue to advance with the planning time according to the order of arrival. In each planning, only the tasks in the current rolling window are planned. As the planning moment progresses, new tasks are continuously added, and the tasks for completing the planning are gradually deleted, thereby realizing the update of the rolling window.

通过采用以上技术方案，本发明的基于学习型遗传算法的多任务多资源滚动分配方法与现有技术相比，其有益效果为：By adopting the above technical solution, the multi-task multi-resource rolling allocation method based on the learning genetic algorithm of the present invention has the beneficial effects compared with the prior art:

本发明采用多任务多资源动态滚动分配机制，通过把复杂的动态调度问题分解为多个简单的静态调度子问题，再对子问题的优化解进行组合，从而代替原问题的最优解，这样大大降低了原问题求解的难度。The invention adopts a multi-task and multi-resource dynamic rolling allocation mechanism, and decomposes the complex dynamic scheduling problem into a plurality of simple static scheduling sub-problems, and then combines the optimized solutions of the sub-problems to replace the optimal solution of the original problem, so that Greatly reduce the difficulty of solving the original problem.

附图说明DRAWINGS

图1为本发明多任务多资源的滚动分配机制的示意图；1 is a schematic diagram of a rolling allocation mechanism of multi-task and multi-resource according to the present invention;

图2为本发明基于学习型遗传算法的多任务多资源分配框架图；2 is a multi-task multi-resource allocation framework diagram based on a learning genetic algorithm according to the present invention;

图3为本发明的动态任务状态分类图；3 is a classification diagram of a dynamic task state according to the present invention;

图4为本发明动态任务规划触发模式；4 is a dynamic task planning triggering mode of the present invention;

图5为本发明的滚动优化策略示意图。FIG. 5 is a schematic diagram of a rolling optimization strategy of the present invention.

图6示出层次化分布式自主协同任务规划体系结构。Figure 6 shows a hierarchical distributed autonomous collaborative task planning architecture.

具体实施方式Detailed ways

为使本发明的目的、技术方案和优点更加清楚明了，下面结合具体实例，对本发明进一步详细说明。应该理解，这些描述只是示例性的，而并非要限制本发明的范围。此外，在以下说明中，省略了对公知结构和技术的描述，以避免不必要地混淆本发明的概念。In order to make the objects, technical solutions and advantages of the present invention more comprehensible, the present invention will be further described in detail below with reference to specific examples. It is to be understood that the description is not intended to limit the scope of the invention. In addition, descriptions of well-known structures and techniques are omitted in the following description in order to avoid unnecessarily obscuring the inventive concept.

本发明研究对象为智能遥感卫星网络。智能遥感卫星(简称智能卫星)是指正在研发的或未来发展的新型遥感卫星。在本发明中，智能卫星一般具有自我感知(如资源发现和环境感知)、自主决策(如复杂任务处理、在线任务规划和遥感图像处理)和在线协同(多颗卫星协同完成复杂任务)等现有遥感卫星所不具备的智能化能力。智能遥感卫星网络是多要素构成的综合性信息***，由不同轨道、不同类型和不同性能的智能遥感卫星及地面支持***通过星地、星间链路构成的天地一体化综合***。该网络具有自主运行管理能力及智能化的信息获取、储存、处理和分发能力，能与陆、海、空基的信息***互联，实现信息的多元、立体共享。The research object of the invention is an intelligent remote sensing satellite network. Intelligent Remote Sensing Satellite (Intelligent Satellite) refers to a new type of remote sensing satellite that is being developed or developed in the future. In the present invention, intelligent satellites generally have self-perception (such as resource discovery and environment awareness), autonomous decision making (such as complex task processing, online task planning and remote sensing image processing), and online collaboration (multiple satellites cooperate to complete complex tasks), etc. There are intelligent capabilities not available in remote sensing satellites. The intelligent remote sensing satellite network is a comprehensive information system composed of multiple elements. The intelligent remote sensing satellites and ground support systems with different orbits, different types and different performances form a comprehensive system of heaven and earth formed by satellite and inter-satellite links. The network has the capability of autonomous operation management and intelligent information acquisition, storage, processing and distribution. It can interconnect with land, sea and space-based information systems to realize multi-dimensional and three-dimensional sharing of information.

智能卫星平台和载荷逐渐体现出高敏捷性、快速响应和组网协同等新特点。Intelligent satellite platforms and payloads are gradually showing new features such as high agility, rapid response and networking synergies.

第一、越来越复杂的成像模式迫切要求突破星上自主任务规划技术。高敏捷性使得智能卫星支持越来越复杂的成像模式，传统离线任务规划模式已不适用于智能卫星，迫切需要探索星上自主任务规划技术。First, the increasingly complex imaging model is urgently required to break through the autonomous mission planning technology on the star. The high agility makes intelligent satellites support more and more complex imaging modes. The traditional offline mission planning mode is not suitable for intelligent satellites. It is urgent to explore the on-board autonomous mission planning technology.

第二、越来越高的服务质量迫切要求突破星上自主任务规划技术。气象、地形和卫星姿态等成像条件都会对成像质量造成显著影响。由于成像条件无法在地面精确预测，智能卫星必须具有星上自主任务规划能力才能合理选择成像姿态和时机，从而选择最优观测姿态及时机实现对给定目标的高质量成像。Second, higher and higher service quality is urgently required to break through the autonomous mission planning technology on the star. Imaging conditions such as weather, terrain, and satellite attitude can have a significant impact on image quality. Since the imaging conditions cannot be accurately predicted on the ground, the intelligent satellite must have the on-board autonomous mission planning ability to reasonably select the imaging attitude and timing, so as to select the optimal observation attitude and timely machine to achieve high-quality imaging of a given target.

第三、越来越多的复杂任务迫切要求突破星上自主协同任务规划技术。不同用户的各种任务需求越来越复杂(如立体成像、移动目标跟踪等)，迫切需要多颗智能卫星之间根据实际情况自主地进行沟通和协同。Third, more and more complex tasks urgently require breakthroughs in autonomous satellite collaborative mission planning techniques. The various task requirements of different users are more and more complex (such as stereo imaging, moving target tracking, etc.), and it is urgent to communicate and coordinate independently among multiple intelligent satellites according to actual conditions.

综上，传统地面离线任务规划模式已无法解决智能卫星面临的各种挑战，必须在充分考虑高敏捷性、协同性、分布性和自主性等特点的基础上，研究面向智能卫星网络的自主协同任务规划技术。In summary, the traditional ground offline mission planning model can not solve the various challenges faced by intelligent satellites. It must study the independent cooperation for intelligent satellite networks based on the characteristics of high agility, synergy, distribution and autonomy. Mission planning techniques.

现有遥感卫星***一般采用集中任务规划模式，对所辖资源进行集中统一规划调度。当观测平台数量、任务数量和数传资源数量大幅增加时，集中任务规划问题的复杂度将呈指数级增长，现有架构已经难以满足要求。本发明提出对多星自主协同任务规划问题进行分层分解处理，建立有效的分层和协同运行机制。The existing remote sensing satellite systems generally adopt a centralized task planning mode to centrally and uniformly plan and schedule the resources under their jurisdiction. When the number of observing platforms, the number of tasks, and the number of digital resources are greatly increased, the complexity of centralized task planning will increase exponentially, and the existing architecture is difficult to meet the requirements. The invention proposes hierarchical decomposition processing of multi-star autonomous collaborative task planning problems, and establishes an effective hierarchical and cooperative operation mechanism.

本发明基于双层规划理论探索一种新型自主协同任务规划架构。通过任务可调度性预测构建顶层多平台多任务协同分配机制，根据任务特点和资源特性将多个任务分配至各颗卫星；通过智能优化和约束推理构建底层的自主任务规划机制，为分配到当前卫星的不同任务自主地安排观测活动。通过对多星自主协同任务规划问题的预先统筹分配和分布式处理，层次化分布式自主协同任务规划架构可极大降低问题求解的复杂度。The invention explores a novel independent collaborative task planning architecture based on the bi-level programming theory. The top-level multi-platform multi-task cooperative allocation mechanism is constructed by task schedulability prediction, and multiple tasks are assigned to each satellite according to task characteristics and resource characteristics. The underlying autonomous task planning mechanism is constructed through intelligent optimization and constraint reasoning, for the current assignment. The different tasks of the satellite autonomously arrange observational activities. Through the pre-aggregate allocation and distributed processing of multi-star autonomous collaborative task planning problems, the hierarchical distributed autonomous collaborative task planning architecture can greatly reduce the complexity of problem solving.

双层规划理论及模型在应对具有多层次特性的决策优化问题方面具备独特的适应性，也非常适合分布式协同机制下的多星自主协同任务规划问题。分布式协同强调的是各子问题之间通过顶层协调单元的信息交互。分布式协同机制下多星自主协同任务规划问题适合采用双层规划问题数学模型进行描述，相关建模求解技术可以借鉴：可将多星自主协同规划过程分为顶层多平台多任务协同分配与底层单平台的自主规划两个相互结合、紧密连接的决策过程(如图6)。图6中上层为多平台多任务动态分配，通过该分配过程将根据任务特点和资源特性将任务分配至各个观测资源。The bi-level programming theory and model have unique adaptability to deal with decision-making optimization problems with multi-level characteristics, and are also very suitable for multi-satellite collaborative task planning under distributed cooperative mechanism. Distributed collaboration emphasizes the information interaction between sub-problems through the top-level coordination unit. The multi-satellite collaborative task planning problem under the distributed coordination mechanism is suitable for describing the mathematical model of bi-level programming problem. The related modeling and solving technology can be used for reference: the multi-star independent collaborative planning process can be divided into the top-level multi-platform multi-task collaborative allocation and the bottom layer. The independent planning of a single platform is a two-in-one, tightly connected decision-making process (Figure 6). The upper layer in Figure 6 is a multi-platform multi-task dynamic allocation. Through this allocation process, tasks are assigned to individual observation resources according to task characteristics and resource characteristics.

层次化分布式自主协同任务规划模式，在集中式协同任务规划的基础上增加了一个总控级别的协调器，并且取消了多星联合调度器，对于每颗卫星的调度使用其专用的单星任务调度。协调器管控级别较高，协调器对任务进行任务约束解析，根据任务要求以及下辖观测资源的状态，通过特定分配算法将任务分配到每个观测资源上去，并将任务处理成单星调度器直接识别的元任务，再由调度器执行调度算法生成各自观测资源的观测方案。各单星调度器可以向协调器反馈单星调度结果，未完成的任务协调器可以依据其他卫星的状态进行再次分配，通过若干次的反馈再分配机制，可以促进分配方案的合理化，从而促进资源使用的更加高效。The hierarchical distributed autonomous collaborative task planning mode adds a master-level coordinator based on the centralized collaborative task planning, and cancels the multi-star joint scheduler. For each satellite, it uses its dedicated single star. Task scheduling. The coordinator has a high level of management and control. The coordinator performs task constraint analysis on the task. According to the task requirements and the state of the observing resources under the jurisdiction, the task is assigned to each observing resource through a specific allocation algorithm, and the task is processed into a single star scheduler. The directly identified meta-tasks are then executed by the scheduler to generate an observation scheme for the respective observed resources. Each single-star scheduler can feed back the single-star scheduling result to the coordinator. The unfinished task coordinator can be re-allocated according to the state of other satellites. Through several feedback redistribution mechanisms, the rationalization of the allocation scheme can be promoted, thereby promoting resources. Use more efficiently.

当新任务到达较少时，协调器可根据任务特征以及已有的单星任务执行方案将该任务分配给某个或某几颗卫星，从而触发这几颗卫星的调度流程，对于未有新分配任务的单星继续执行已有方案，从而实现多星的异步管控，增强了观测资源管理的灵活性。本模式在一定程度上避免了多星调度统一建模的复杂性，将调度问题的进行分层处理，增强了***的重用性，也增强了***可拓展性，如果临时增加或者临时减少协同卫星，只需要在多星任务协调器处进行修改。When the new task arrives less, the coordinator can assign the task to one or several satellites according to the task characteristics and the existing single-star task execution plan, thereby triggering the scheduling process of the satellites. The single star of the assigned task continues to execute the existing scheme, thereby realizing multi-star asynchronous control and enhancing the flexibility of observation resource management. This model avoids the complexity of unified modeling of multi-star scheduling to a certain extent, and hierarchically processes the scheduling problem, which enhances the system's reusability and enhances system scalability. If the satellite is temporarily added or temporarily reduced. , only need to be modified at the multi-star task coordinator.

层任务规划将复杂问题进行预先统筹分配、分布式处理，大大降低问题的求解复杂度，然而能否合理建立两个层次决策变量之间的衔接关系则是决定分层任务规划有效性的关键。本项目将通过研究合理的任务可调度性预测方法来解决这一衔接难题。任务可调度性预测的作用就是在顶层任务预规划阶段，预先估计下层各平台调度结果，以此作为任务分配的依据，避免后期调度结果反馈的滞后性导致前期任务分配的盲目性。Layer task planning pre-allocates and distributes complex problems, which greatly reduces the complexity of solving problems. However, whether the connection between two levels of decision variables can be reasonably established is the key to determining the effectiveness of hierarchical task planning. This project will solve this convergence problem by studying a reasonable task schedulability prediction method. The role of task schedulability prediction is to pre-estimate the scheduling results of the lower platforms in the pre-planning stage of the top-level tasks, so as the basis of task allocation, avoiding the lag of the feedback of the late scheduling results, leading to the blindness of the previous task assignment.

具体地，本发明集中研究如何将多个观测任务分配给多个观测资源(智能观测卫星)，以及在测控窗口之外，如何应对应急任务。至于每颗智能观测卫星自身的任务调度，由星上资源完成调度。需要指出的是，在本发明中，“分配”是指，协调器将常规观测任务和应急观测任务分配给相应的智能观测卫星，且相应智能观测卫星上的星上调度器将所述任务加入到智能观测卫星实际执行的任务序列中去。如果加入不成功，则视为为完成分配。未完成分配的任务将在下一分配时刻由多星任务协调器再次分配，或者将被舍弃。In particular, the present invention focuses on how to assign multiple observation tasks to multiple observation resources (smart observation satellites) and how to respond to emergency tasks outside of the measurement and control window. As for the task scheduling of each intelligent observation satellite itself, the scheduling is completed by the on-board resources. It should be noted that, in the present invention, "allocation" means that the coordinator assigns the conventional observation task and the emergency observation task to the corresponding intelligent observation satellite, and the on-board scheduler on the corresponding intelligent observation satellite adds the task. Go to the task sequence actually executed by the intelligent observation satellite. If the join is unsuccessful, it is considered to be the assignment. Tasks that have not completed the assignment will be reassigned by the multi-star task coordinator at the next assigned moment, or will be discarded.

本发明的多任务多资源滚动分配方法采用层次化分布式自主协同任务架构。多星任务协调器将滚动窗口内的任务集合分配给下辖的多颗智能卫星，各智能卫星利用其星上调度器对被分配的新任务和已有任务进行调度，在当前滚动窗口的起始时刻，多星任务协调器对任务信息进行更新，删除上一滚动窗口内已经完成的任务以及在所述起始时刻正在执行的任务，并将上一滚动窗口内未分配的任务、以及在上一滚动窗口内到达的新任务组合成当前滚动窗口内的任务集合，且所述多星任务协调器将该任务集合向所述多颗智能卫星进行分配，其中，基于混合触发模式来确定滚动窗口的起始时刻，一方面，每隔一个时间段触发滚动分配，该时间段为恒定的或根据预先设定的规则动态变化；另一方面，在出现使***状态发生改变的事件或在受到人工干预时触发滚动分配。The multi-task multi-resource rolling allocation method of the invention adopts a hierarchical distributed autonomous collaborative task architecture. The multi-star task coordinator assigns a set of tasks in the rolling window to a plurality of intelligent satellites under the jurisdiction, and each intelligent satellite uses its on-board scheduler to schedule the assigned new tasks and existing tasks, starting from the current scrolling window. At the beginning, the multi-star task coordinator updates the task information, deletes the tasks that have been completed in the previous scrolling window, and the tasks that are being executed at the starting time, and the tasks that were not assigned in the previous scrolling window, and The new tasks arriving in the last scrolling window are combined into a task set in the current scrolling window, and the multi-star task coordinator allocates the task set to the plurality of smart satellites, wherein the scrolling is determined based on the hybrid triggering mode The start time of the window, on the one hand, triggers the scroll allocation every other time period, the time period is constant or dynamically changes according to a preset rule; on the other hand, the event that causes the system state to change or is subject to Triggering is triggered when manually intervened.

需要指出的是，有可能存在上一周期内从智能卫星退回的任务，给等被退回的任务可能是原来安排在智能卫星的任务执行序列中，但是由于被新分配的任务所替换，从而被退回；或者可能是上一周期内分配给智能卫星，但是最终未能安排至智能卫星的任务执行序列中，从而被退回。此等被退回的任务也纳入到当期周期的分配中，重新进行分配。有利的是，在重新进行分配时，提高该等重新分配的任务的优先等级。但是，保持该等重新分配的任务的收益值或权重不变。It should be pointed out that there may be a task that is returned from the smart satellite in the previous week. The task that was returned may be originally arranged in the mission execution sequence of the intelligent satellite, but it is replaced by the newly assigned task. Returned; or may have been assigned to a smart satellite during the previous week, but was ultimately not scheduled into the mission execution sequence of the intelligent satellite and was returned. These returned tasks are also included in the allocation of the current period and re-allocated. Advantageously, the priority of the reassigned tasks is increased when the assignment is re-allocated. However, the revenue value or weight of the tasks that maintain these redistributions remains the same.

优选地，一方面，所述时间段根据测控周期设置；另一方面，所述使***状态发生改变的事件包括：接收到应急观测任务，且积累的未分配应急观测任务达到五件或者是所述多星任务协调器下辖的智能卫星数的5％。测控周期根据通常根据地面站及相应的通信卫星、通信车、通信飞机、海上测控船的位置及通信条件确定的可与智能卫星有效通讯的时间周期来确定，此外与智能卫星的具体轨道与圈次等相关。Preferably, in one aspect, the time period is set according to a measurement and control cycle; on the other hand, the event that changes a state of the system includes: receiving an emergency observation task, and the accumulated unallocated emergency observation task reaches five pieces or is 5% of the number of intelligent satellites under the multi-star mission coordinator. The measurement and control cycle is determined according to the time period that can be effectively communicated with the intelligent satellite, which is usually determined according to the location and communication conditions of the ground station and the corresponding communication satellite, communication vehicle, communication aircraft, maritime monitoring and control vessel, and in addition to the specific orbit and circle of the intelligent satellite. Secondary relevance.

各智能卫星的星上调度器的任务调度策略如下：The task scheduling strategy of the onboard scheduler of each intelligent satellite is as follows:

输入：Enter:

输出:Output:

--下一时间周期T内的调度计划；

- a scheduling plan within the next time period T;

具体步骤如下：Specific steps are as follows:

步骤11 分别从

和

和应急观测任务集合

Step 11 separately from

with

And emergency observation task set

步骤12 将

和

整合为一个观测任务集合； Step 12 will

with

Integrated into a collection of observation tasks;

中，直至所述整合后的观测任务集合中再无任务可加入

in,

步骤15 输出下一时间周期T内的调度计划

Step 15 Output the schedule in the next time period T

输入：Enter:

—在本时间周期T内且晚于C ^*-驱动调度时刻点t的调度计划；

—在调度时刻点t之前已到达且未调度的应急观测任务集合；

输出:Output:

—在时间t时已修订的调度计划，

- the revised schedule at time t,

具体步骤如下：Specific steps are as follows:

中选取应急观测任务，生成新的任务集合

Select emergency observation tasks to generate new task sets

步骤22 根据设定的启发式规则，对

Sort the emergency observation tasks in the middle;

步骤23 按照新的任务次序，一一选取

中的应急观测任务并对

进行修订，直至

中再无应急观测任务可加入

中， Step 23 Select one by one according to the new task order.

Emergency observation mission

Revise until

No more emergency observation tasks can be added

in,

步骤24 输出已修订的调度计划

Step 24 Output the revised schedule

如图1至图5所示，本发明基于演化与学习机制，构建了求解面向复杂优化问题的学习型智能优化方法：采用智能优化模型和知识模型相结合的集成建模思路，智能优化模型按照“邻域搜索”策略对待优化问题的可行空间进行搜索；知识模型从前期的优化过程中挖掘出一些有用知识，然后采用得到的知识来指导智能优化方法的后续优化过程。而本发明采用学习型遗传算法，求解多任务多资源滚动分配的复杂优化问题。As shown in FIG. 1 to FIG. 5, the present invention builds a learning-based intelligent optimization method for solving complex optimization problems based on evolution and learning mechanism: an integrated modeling idea combining intelligent optimization model and knowledge model, and an intelligent optimization model according to The “neighbor search” strategy searches for the feasible space of the optimization problem; the knowledge model mines some useful knowledge from the previous optimization process, and then uses the acquired knowledge to guide the subsequent optimization process of the intelligent optimization method. The invention adopts a learning genetic algorithm to solve the complex optimization problem of multi-task multi-resource rolling allocation.

本发明基于滚动时域控制原理，构建多任务多资源动态分配的滚动机制，将复杂动态分配问题转化为滚动更新的静态分配问题；定义能体现多任务多资源分配问题本质特征的若干类知识，构建能有效管理这些知识的知识模型；基于遗传算法设计多任务多资源分配问题可行方案构建的遗传优化模型；重点研究遗传优化模型和知识模型之间的集成和交互机制，最终形成将遗传优化模型和知识模型高效集成的学习型遗传算法。The invention is based on the rolling time domain control principle, constructs a rolling mechanism of multi-task multi-resource dynamic allocation, converts the complex dynamic allocation problem into a static allocation problem of rolling update, and defines several kinds of knowledge that can reflect the essential characteristics of the multi-task multi-resource allocation problem. Construct a knowledge model that can effectively manage this knowledge; design a genetic optimization model based on genetic algorithm to design a multi-task multi-resource allocation problem; focus on the integration and interaction between genetic optimization model and knowledge model, and finally form a genetic optimization model A learning genetic algorithm that integrates efficiently with knowledge models.

基于滚动时域控制原理，构建多任务多资源动态分配的滚动机制：确定预测窗口、滚动窗口、分配子问题和滚动机制等要素；在每个规划时刻，通过当前预测窗口对任务信息进行实时更新(如增加任务、删除任务或修改任务属性信息等)，在预测窗口的基础上确定当前滚动窗口；分配子问题是在每个规划时刻根据当前滚动窗口构造的局部分配问题，滚动机制用于确定分配子问题求解后分配方案结束的执行位置和下一个规划时刻。通过这种滚动分配机制可将复杂动态分配问题转化为滚动更新的静态分配问题，如图1所示。Based on the principle of rolling time domain control, the rolling mechanism of multi-task multi-resource dynamic allocation is constructed: determining the prediction window, rolling window, allocation sub-problem and scrolling mechanism; in each planning moment, the task information is updated in real time through the current prediction window. (such as adding tasks, deleting tasks or modifying task attribute information, etc.), determining the current scrolling window based on the prediction window; the allocation sub-question is a local allocation problem constructed according to the current scrolling window at each planning moment, and the scrolling mechanism is used to determine After the sub-problem is solved, the execution position of the end of the allocation scheme and the next planning moment are allocated. Through this rolling allocation mechanism, complex dynamic allocation problems can be transformed into static allocation problems of rolling updates, as shown in Figure 1.

在求解多任务多资源分配问题的学习型遗传算法中，从用户偏好和任务场景特征等信息中提取规则，并在优化过程中挖掘相关知识来指导后续优化过程，从而获得比较满意的优化效果。In the learning genetic algorithm for solving multi-task multi-resource allocation problem, rules are extracted from information such as user preferences and task scene features, and relevant knowledge is mined in the optimization process to guide the subsequent optimization process, so as to obtain satisfactory optimization results.

基于学***台剩余能力排序的任务交流规则进行。种群替换则加入用户任务锁定规则生成替换种群，在保证用户对任务偏好提取的基础上，同时实现算法跳出局部最优的能力。The framework of multi-task multi-resource allocation technology based on learning genetic algorithm is shown in Fig. 2. Among them, the fitness evaluation is based on the multi-objective function fitness evaluation based on TOPSIS method, and the users are stored and extracted under different scene competition indicators. Preference is given to the weight assignment of the optimization goal. The population initialization uses the heuristic rule-based initial population generation to improve the initial population quality while ensuring the random distribution of the initial population. The selection operation, while improving the selection operator, continuously learns the individual component combination knowledge that has a good solution in the iterative process, the resource competition knowledge of the individual task, and applies to the next intersection and mutation operation. Under the guidance of individual component combination knowledge, the crossover operation selects the intersection position with different probabilities, and adopts multiple operations to take the optimal strategy to ensure the effectiveness of the crossover operation. Under the guidance of the resource competition knowledge of the task, the mutation operation selects the individual mutation operation position with different probabilities, and also adopts multiple operations to take the optimal strategy to improve the efficiency of the mutation operation. The local search uses a combination of local randomization and determined search, wherein the deterministic search is performed based on the task communication rules of the individual platform residual capacity ranking. The population replacement is added to the user task locking rule to generate the replacement population. On the basis of ensuring the user's extraction of the task preferences, the algorithm can also jump out of the local optimal ability.

多任务多资源滚动分配需要建立滚动式动态规划框架、构建动态任务状态更新机制、设计滚动式动态任务规划触发模式和确定滚动时间窗口的选取策略等。Multi-task multi-resource rolling assignment needs to establish a rolling dynamic planning framework, construct a dynamic task state update mechanism, design a rolling dynamic task planning triggering mode, and determine the selection strategy of the rolling time window.

(1)滚动式动态规划框架。根据我国测控资源分布情况，拟采用基于测控窗口为周期的滚动式动态调度方法。其基本思想是把任务按照到达时间划分为具有一定重叠度，随着调度时刻不断向前推进的任务集合；在每次调度时，仅对当前滚动窗口内的任务进行规划。随着调度时刻的推进，新任务被不断加入，而完成调度的任务则被逐渐删除，从而实现滚动窗口的更新。滚动式动态规划通过把复杂的动态调度问题分解为多个简单的静态调度子问题，再对子问题的优化解进行组合，从而代替原问题的最优解，这样做的优点是降低了原问题求解的难度。(1) Rolling dynamic programming framework. According to the distribution of China's measurement and control resources, it is proposed to adopt a rolling dynamic scheduling method based on the measurement and control window. The basic idea is to divide the task according to the arrival time into a set of tasks with a certain degree of overlap and continuously advance with the scheduling time; in each scheduling, only the tasks in the current rolling window are planned. As the scheduling time advances, new tasks are continuously added, and the tasks that complete the scheduling are gradually deleted, thereby implementing the update of the rolling window. Rolling dynamic programming reduces the original problem by decomposing complex dynamic scheduling problems into multiple simple static scheduling sub-problems and then combining the optimal solutions of sub-problems to replace the optimal solution of the original problem. The difficulty of solving.

(2)动态任务状态更新机制。在不同调度时刻需要及时更新任务状态，以获取更加完备的信息来指导规划调度。根据任务在不同时刻的状态将其分为四类，如图3所示：已完成任务(Finished task)、正在执行任务(Executing task)、等待执行任务(Waiting task)及新到达任务(New task)。其中，t _R为调度开始时刻，当TF _i＜t _R时，任务i为已完成任务；当TS _i＜t _R＜TF _i时，任务i为正在执行的任务；当t _R＜TS _i时，任务i为等待执行的任务；当T _iA＝t _R时，任务i为新到达任务。 (2) Dynamic task status update mechanism. The task status needs to be updated in time at different scheduling times to obtain more complete information to guide the planning and scheduling. According to the state of the task at different times, it is divided into four categories, as shown in Figure 3: completed task (Executing task), waiting for executing task (Waiting task) and new arrival task (New task ). Where t _R is the scheduling start time, when TF _i <t _R , task i is the completed task; when TS _i <t _R <TF _i , task i is the task being executed; when t _R <TS _i Task i is a task waiting to be executed; when T _i A=t _R , task i is a new arrival task.

(3)滚动式动态任务规划触发模式。调度时刻安排是影响滚动优化策略应用效果的关键，它取决于多种因素(如用户需求、卫星指令上注时间计划、管控中心规划能力等)，可将这些因素归结为周期性因素和事件性因素，滚动式动态调度策略在制定规划方案时可采用周期触发模式、事件触发模式或者混合触发模式。(3) Rolling dynamic task planning trigger mode. Scheduling schedule is the key to affect the application effect of the rolling optimization strategy. It depends on various factors (such as user requirements, satellite command time planning, control center planning ability, etc.), which can be attributed to periodic factors and event Factors, the rolling dynamic scheduling strategy can adopt the periodic trigger mode, the event trigger mode or the mixed trigger mode when formulating the planning scheme.

①周期触发模式。如测控周期触发或值班周期触发，针对常规任务每隔一段时间进行一次规划，该时间可以是均匀恒定的，也可以是动态变化的。基于值班周期触发模式能够保证规划频率的稳定性，操作实现较为简单，但是该模式当周期较长时，无法为高时效性任务提供及时的规划服务，且应对***状态变化的能力也是非常有限的。1 cycle trigger mode. For example, when the measurement and control cycle is triggered or the duty cycle is triggered, the regular task is scheduled once every time, and the time may be uniform or dynamic, or may be dynamically changed. The duty cycle trigger mode can ensure the stability of the planning frequency, and the operation is simple. However, when the cycle is long, it can not provide timely planning services for high-aging tasks, and the ability to cope with system state changes is also very limited. .

②事件触发模式。即在出现使***状态发生改变的事件，或在受到人工干预时开始执行规划，如有新任务到达、卫星状态发生变化、决策部门提出规划需求等情况发生时。事件触发模式对规划环境具备较高的敏感性，能够及时处置应急任务，但由于规划频繁，当任务数量较多时，算法的时间复杂度较高。事件触发主要是任务触发模式。2 event trigger mode. That is, when an event occurs that changes the state of the system, or when the human intervention is initiated, the planning is started, if a new task arrives, the satellite state changes, and the decision-making department proposes planning requirements. The event trigger mode has high sensitivity to the planning environment and can handle emergency tasks in time. However, due to frequent planning, when the number of tasks is large, the time complexity of the algorithm is high. The event trigger is mainly the task trigger mode.

③混合触发模式。即事件触发与周期触发相结合的触发模式。它能够兼获两种触发模式的优点，具备及时处置应急任务的能力，并且算法的时间复杂度有限。3 mixed trigger mode. That is, the trigger mode combined with the event trigger and the periodic trigger. It has the advantages of both trigger modes, the ability to handle emergency tasks in a timely manner, and the time complexity of the algorithm is limited.

根据上述分析，多任务多资源滚动分配模式主要为以应急任务触发的对已有规划方案进行动态调整的动态重规划和以测控点周期性触发的滚动式任务规划的混合触发模式，如图4所示。According to the above analysis, the multi-task multi-resource rolling allocation mode is mainly a dynamic re-planning that dynamically adjusts the existing planning scheme triggered by the emergency task and a mixed triggering mode of the rolling task planning triggered by the monitoring point periodically, as shown in FIG. 4 . Shown.

(4)滚动时间窗口选取策略。滚动优化策略的基本思想是将任务按照到达顺序划分为具有一定重叠、但随着规划时刻不断向前推进的任务集合，称为滚动窗口。在每次规划时，仅对当前滚动窗口内的任务进行规划，随着规划时刻的推进，新任务被不断加入，而完成规划的任务则被逐渐删除，从而实现滚动窗口的更新，如图5所示。滚动优化策略的优点是能够将复杂的动态规划问题分解为多个简单的静态规划子问题，并以子问题优化解的组合代替原问题的最优解，从而降低了原问题求解的难度。(4) Rolling time window selection strategy. The basic idea of the rolling optimization strategy is to divide the tasks into a set of tasks that have a certain overlap, but continue to advance with the planning moment, called a rolling window. In each planning, only the tasks in the current rolling window are planned. As the planning time advances, new tasks are continuously added, and the tasks for completing the planning are gradually deleted, thereby implementing the rolling window update, as shown in FIG. 5. Shown. The advantage of the rolling optimization strategy is that it can decompose complex dynamic programming problems into multiple simple static programming sub-problems, and replace the optimal solution of the original problem with the combination of sub-problem optimization solutions, thus reducing the difficulty of solving the original problem.

上述的具体实施方式只是示例性的，是为了更好地使本领域技术人员能够理解本专利，不能理解为是对本专利包括范围的限制；只要是根据本专利所揭示精神的所作的任何等同变更或修饰，均落入本专利包括的范围。The specific embodiments described above are merely exemplary in order to enable those skilled in the art to understand the present invention and are not to be construed as limiting the scope of the invention; any equivalent changes made in accordance with the spirit of the disclosure. Or modifications, are included in the scope of this patent.

Claims

一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，采用层次化分布式自主协同任务架构，多星任务协调器将滚动窗口内的任务集合分配给下辖的多颗智能卫星，各智能卫星利用其星上调度器对被分配的新任务和已有任务进行调度，在当前滚动窗口的起始时刻，多星任务协调器对任务信息进行更新，删除上一滚动窗口内已经完成的任务以及在所述起始时刻正在执行的任务，并将上一滚动窗口内未分配的任务、以及在上一滚动窗口内到达的新任务组合成当前滚动窗口内的任务集合，且所述多星任务协调器将该任务集合向所述多颗智能卫星进行分配，其中，基于混合触发模式来确定滚动窗口的起始时刻，一方面，每隔一个时间段触发滚动分配，该时间段为恒定的或根据预先设定的规则动态变化；另一方面，在出现使***状态发生改变的事件或在受到人工干预时触发滚动分配。A multi-task and multi-resource rolling allocation method based on learning genetic algorithm, characterized in that a hierarchical distributed autonomous collaborative task architecture is adopted, and a multi-star task coordinator allocates a task set in a rolling window to multiple intelligences under the jurisdiction Satellites, each intelligent satellite uses its on-board scheduler to schedule assigned new tasks and existing tasks. At the beginning of the current scrolling window, the multi-star task coordinator updates the task information and deletes the previous scroll window. The completed task and the task being executed at the starting time, and the unassigned task in the previous scrolling window and the new task arriving in the previous scrolling window are combined into a task set in the current scrolling window, and The multi-star task coordinator allocates the task set to the plurality of smart satellites, wherein the start time of the scroll window is determined based on the hybrid trigger mode, and on the other hand, the scroll allocation is triggered every other time period, the time The segment is constant or dynamically changes according to pre-set rules; on the other hand, the system state changes when it appears Event or trigger a rolling distribution when subjected to human intervention.
根据权利要求1所述的一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，一方面，所述时间段根据测控周期设置；另一方面，所述使***状态发生改变的事件包括：接收到应急观测任务，且积累的未分配应急观测任务达到五件或者是所述多星任务协调器下辖的智能卫星数的5％。The multi-task multi-resource rolling allocation method based on the learning genetic algorithm according to claim 1, wherein, on the one hand, the time period is set according to a measurement and control cycle; on the other hand, the system state is changed. The events include: receiving emergency observation missions, and the accumulated unallocated emergency observation missions reach five or 5% of the number of intelligent satellites under the multi-star mission coordinator.
根据权利要求2所述的一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，所述多星任务协调器包括地面站和地球静止轨道通信卫星，在测控周期之内，所述地面站进行任务分配；在测控周期之外，所述地球静止轨道通信卫星进行任务分配，且所述应急观测任务由所述智能卫星生成。The multi-task multi-resource rolling allocation method based on learning genetic algorithm according to claim 2, wherein the multi-star task coordinator comprises a ground station and a geostationary orbit communication satellite, within a measurement and control period, The ground station performs task assignment; outside the measurement and control period, the geostationary orbit communication satellite performs task assignment, and the emergency observation task is generated by the smart satellite.
根据权利要求1所述的基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，各智能卫星的星上调度器的任务调度策略如下：The multi-task multi-resource rolling allocation method based on the learning genetic algorithm according to claim 1, wherein the task scheduling strategy of the on-board scheduler of each intelligent satellite is as follows:

(1)在T-驱动的调度时刻点，采用渐进式方法中的完全重调度策略，生成下一个周期时间区间内的新任务计划，T-驱动的调度时刻点是根据给定的时间间隔T来确定特定的调度时间点lT，0≤l≤L,LT≤H＜(L+1)T，每到达一个调度时间点lT，则计算生成后一调度区间[lT,(l+1)T]的任务计划，其中l为正整数，T为给定的时间间隔，L为最大T-驱动调度次数，H为总调度区间，(1) At the scheduling moment of the T-drive, the full rescheduling strategy in the progressive method is used to generate a new mission plan in the next cycle time interval, and the T-driven scheduling time point is based on the given time interval T To determine a specific scheduling time point lT, 0 ≤ l ≤ L, LT ≤ H < (L + 1) T, each time a scheduling time point lT is reached, the calculation of the latter scheduling interval [lT, (l + 1) T Mission plan, where l is a positive integer, T is the given time interval, L is the maximum number of T-drive scheduling, and H is the total scheduling interval.

(2)在C ^*-驱动的重调度时刻点，采用修订式方法中的调度计划修复策略，当卫星运行在给定的调度区间内时，若在某一时刻t(0＜t＜H)，星上的应急观测任务累积量C _t超过给定的阈值C ^*时，则执行重调度计算，其中阈值C ^*为应急观测任务的临界累积数， (2) At the C ^* -driven rescheduling time point, the scheduling plan repair strategy in the revised method is adopted. When the satellite is operating in a given scheduling interval, if it is at a certain time t (0 < t < H) emergency satellite observation missions on the accumulated amount of C _t exceeds a given threshold value C ^*, calculation is performed rescheduling, wherein C ^* is the critical threshold cumulative emergency observation tasks,

除上述两种调度时刻点之外，不在任何其他时刻点进行调度。Except for the above two scheduling moments, scheduling is not performed at any other point in time.
根据权利要求4所述的一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，A multi-task multi-resource rolling allocation method based on a learning genetic algorithm according to claim 4, wherein

在T-驱动的调度时刻点的调度算法如下：The scheduling algorithm at the T-driven scheduling moment is as follows:

输入：Enter:

已到达且在T-驱动调度时刻点之前未被调度的应急观测任务集合；
An emergency observation task set that has arrived and is not scheduled before the T-drive scheduling time point;

已接收且在T-驱动调度时刻点之前未被调度的常规观测任务集合；
a set of conventional observation tasks that have been received and are not scheduled before the T-drive scheduling time point;

输出:Output:

下一时间周期T内的调度计划；
a scheduling plan within the next time period T;

具体步骤如下：Specific steps are as follows:

步骤11分别从
和
中选取时间窗口是否落入下一个时间周期T内的常规观测任务和应急观测任务，生成待调度求解的常规观测任务集合
和应急观测任务集合
Step 11 respectively
with
Select whether the time window falls into the conventional observation task and the emergency observation task in the next time period T, and generate a conventional observation task set to be solved.
And emergency observation task set

步骤12将
和
整合为一个观测任务集合； Step 12 will
with
Integrated into a collection of observation tasks;

步骤13按照设定的启发式规则，对整合后的观测任务集合中的任务进行排序；Step 13 sorts the tasks in the integrated observation task set according to the set heuristic rules;

步骤14按照排序，对所述整合后的观测任务集合中的任务一一进行调度，以确定是否将之加入到
中，直至所述整合后的观测任务集合中再无任务可加入
中， Step 14 sorts the tasks in the integrated observation task set one by one according to the ordering to determine whether to join the task
In the above, until the integrated observation task set has no more tasks to join
in,

步骤15输出下一时间周期T内的调度计划
Step 15 outputs the scheduling plan in the next time period T

在C ^*-驱动的重调度时刻点的调度算法如下： The scheduling algorithm at the C ^* -driven rescheduling point in time is as follows:

输入：Enter:

在本时间周期T内且晚于C ^*-驱动调度时刻点t的调度计划；
a scheduling plan within this time period T and later than the C ^* -drive scheduling time point t;

在调度时刻点t之前已到达且未调度的应急观测任务集合；
a set of emergency observation tasks that have arrived and are not scheduled before the scheduling time point t;

输出:Output:

在时间t时已修订的调度计划，
The revised schedule at time t,

具体步骤如下：Specific steps are as follows:

步骤21根据观测时间窗口处于时间t到下一个T-驱动调度时刻点这一时间区间内的条件，从任务集合
中选取应急观测任务，生成新的任务集合
Step 21: according to the condition that the observation time window is in the time interval from the time t to the next T-drive scheduling time point, from the task set
Select emergency observation tasks to generate new task sets

步骤22根据设定的启发式规则，对
中的应急观测任务进行排序； Step 22 according to the set heuristic rules,
Sort the emergency observation tasks in the middle;

步骤23按照新的任务次序，一一选取
中的应急观测任务并对
进行修订，直至
中再无应急观测任务可加入
中， Step 23 selects one by one according to the new task order.
Emergency observation mission
Revise until
No more emergency observation tasks can be added
in,

步骤24输出已修订的调度计划
Step 24 Output the revised schedule
根据权利要求1所述的一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，多星任务协调器采用设定的滚动分配机制将滚动窗口内的任务集合分配给下辖的多颗智能卫星，所述滚动分配机制从用户偏好和任务场景特征信息中提取规则，并将所述规则应用于遗传算法的交叉操作和变异操作。The multi-task multi-resource scrolling allocation method based on learning genetic algorithm according to claim 1, wherein the multi-star task coordinator assigns a task set in the scroll window to the subordinate by using a set scroll allocation mechanism. The plurality of intelligent satellites, the rolling allocation mechanism extracts rules from user preference and task scene feature information, and applies the rules to the cross operation and the mutation operation of the genetic algorithm.
根据权利要求1所述的一种基于学习型遗传算法的多任务多资源滚动分配方法，其特征在于，多星任务协调器采用设定的滚动分配机制将滚动窗口内的任务集合分配给下辖的多颗智能卫星，所述滚动分配机制包括以下步骤：The multi-task multi-resource scrolling allocation method based on learning genetic algorithm according to claim 1, wherein the multi-star task coordinator assigns a task set in the scroll window to the subordinate by using a set scroll allocation mechanism. Multiple intelligent satellites, the rolling distribution mechanism includes the following steps:

S1、适应度评价：采用基于TOPSIS方法的多目标函数适应度评价，同时在不同场景竞争度指标下存储并提取用户偏好，进行优化目标的权重分配；S1, fitness evaluation: multi-objective function fitness evaluation based on TOPSIS method, while storing and extracting user preferences under different scene competition indicators, and performing weight distribution of optimization targets;

S2、种群初始化：采用基于启发式规则的初始化种群生成，在提高初始种群质量基础上同时保证初始种群的随机分布；S2. Population initialization: Initial population generation based on heuristic rules is adopted to ensure the random distribution of the initial population while improving the initial population quality;

S3、选择操作：在对选择算子进行改进的同时，不断学习迭代过程中出现好解的个体构件组合知识、个体中任务的资源竞争度知识，并应用于接下来的交叉与变异操作；S3, selection operation: while improving the selection operator, continuously learn the individual component combination knowledge that has a good solution in the iterative process, the resource competition knowledge of the task in the individual, and apply to the following crossover and mutation operations;

S4、交叉操作：在个体构件组合知识的指导下，以不同概率进行交叉位置的选择，同时采用多次操作取最优的策略，保证交叉操作的有效性；S4, cross operation: under the guidance of individual component combination knowledge, the intersection position is selected with different probabilities, and multiple operations are used to take the optimal strategy to ensure the effectiveness of the cross operation;

S5、变异操作：在任务的资源竞争度知识的指导下，以不同概率选择个体变异操作位置，同时也采用多次操作取最优的策略，提高变异操作效率；S5. Mutation operation: Under the guidance of the resource competition knowledge of the task, the individual mutation operation position is selected with different probabilities, and the optimal strategy is adopted by multiple operations to improve the efficiency of the mutation operation;

S6、局部搜索：采用局部随机与确定搜索相结合的策略，其中确定性搜索基于个体平台剩余能力排序的任务交流规则进行；S6. Local search: a strategy combining local randomization and determined search, wherein the deterministic search is performed based on task communication rules of the remaining capabilities of the individual platform;

S7、种群替换：加入用户任务锁定规则生成替换种群，在保证用户对任务偏好提取的基础上，同时实现算法跳出局部最优的能力。S7. Replacing the population: adding the user task locking rule to generate the replacement population. On the basis of ensuring the user's extraction of the task preference, the algorithm can simultaneously jump out of the local optimal capability.