A general modeling and solution method for spacecraft heuristic temporal planning

Abstract

The present invention proposes a general spacecraft heuristic temporal planning modeling and solving method, which is characterized by: fully combining the characteristics of aerospace engineering tasks and actual needs, proposing a general model of the spacecraft field and its mathematical expression, using a temporal network structure diagram to represent the time constraints and energy constraints of the spacecraft patrol detection state transfer, using a forward pruning strategy to propagate constraints and solve problems on the network structure diagram, and designing and implementing a heuristic control function based on the maximum time span to relax the problem to improve the solving efficiency. The domain model constructed by this method is more complete, and the temporal network structure diagram technology used reduces the complexity of the planning solution calculation, and greatly simplifies the complexity of the algorithm design.

Description

Technical field:

[0001] The present invention relates to the application of artificial intelligence mission planning technology in the field of aerospace engineering, and in particular to a method applicable to modeling and solving mission planning for all spacecraft. Technical background:

[0002] Mission intelligent planning technology is to obtain a complete set of problems by modeling a real problem, analyzing the constraints to transform the actual problem into a constraint-satisfiable problem for solution, and supporting the cross-scheduling of parallel tasks in action sequences. It has been widely used in planning and scheduling systems, temporal database systems, common sense reasoning, natural language processing and other fields, and has proposed a variety of formal descriptions and reasoning methods for time constraints. A complete mission intelligent planning process includes three processes: domain modeling, constraint representation, and problem solving. When applied to the field of aerospace engineering, the difficulties faced by mission planning technology include:

[0003] (1) Formal description languages, represented by Planning Domain Definition Language (PDDL), require that the external world is fully observable and that actions do not have durations, etc., based on strong constraints. From a mathematical perspective, this description can help people better understand the representation and characteristics of planning. However, classical planning places too many restrictions on real-world problems. Specifically, in aerospace engineering tasks, most of the above constraints are not valid. For example, the state space of problems involving numerical variables is usually infinite; information from the external world may be unobservable; when actions cannot give a definite effect, the results of state transitions may be random; external events may be affected by accidental events; many problems in engineering tasks do not require the system to stop running immediately after reaching a certain target state, but to execute a sequence of actions in a loop, or to avoid certain special states during the action process; concurrent execution of actions is allowed during the motion of the spacecraft, and the form of the planning solution is not limited to a linear sequence of actions, but a set of actions with durations.

[0004] (2) Temporal planning technology is needed to solve time and resource constraints. The latter introduces the idea of ​​constraint processing problems in graph theory into the solution of constraint satisfaction problems. It is an effective way and means to describe and reason about systems with time knowledge and time constraints. In recent years, it has been widely used in planning and scheduling systems, temporal database systems, common sense reasoning, natural language processing and other fields, and a variety of formal description and reasoning methods for time constraints have been proposed, including Allen's time interval algebra method, Vilain and Kautz's time point algebra method, Malik and Binford's linear inequality group method, Dean and McDermott's time graph method, and Dechter and Meiri's time constraint network method. For each different description method, there are corresponding constraint propagation methods to support it. The dynamic solution method has high computational complexity and large demand for computing resources. In the result, many irrelevant action plans will be output. In engineering tasks, the initial state and action effect of the spacecraft are generally clarified in the form of implementation plans. The conditions and consequences of state transitions can be logically inferred without the need to search the state space of all branches during the solution process.

[0005] (3) Several common temporal planners support solving temporal and resource constraints, including TLPlan, which proposed an algorithm that supports concurrent operations of forward state space search. The solution speed is very fast, but it requires manual coding of domain-specific control rules, which is difficult to program. Zeno is a type of partial order planner that supports quantitative evaluation of results, but the solution process lacks heuristic control and the system architecture has poor applicability. HSTS is applied to NASA's "Deep Space 1" and uses a step-by-step decomposition and repeated patching method to support rich resource description and time constraint processing, but it requires professional domain knowledge to guide the search, resulting in a very complex system and high planning cost. Summary of the invention:

[0006] In view of the above technical problems, the present invention provides a general spacecraft heuristic temporal planning modeling and solving method, which fully combines the characteristics of aerospace engineering tasks and actual needs, proposes a general model in the field of spacecraft and its mathematical expression, uses a temporal network structure diagram to represent the time constraints and energy constraints of the state transition of spacecraft patrol detection, adopts a forward pruning strategy to propagate constraints and solve problems on the network structure diagram, and designs and implements a heuristic control function based on the maximum time span to relax the problem, so as to improve the solving efficiency. The characteristics of this method are:

[0007] (1) In addition to the spacecraft information, the constructed domain model adds a state transfer function to transform the complex constraints in the engineering task into time-dependent numerical quantitative relationships, making the information expression of the domain model more complete.

[0008] (2) Define a temporal network structure graph to represent the mission objectives, where nodes represent atomic tasks, edges represent state transition functions, and the direction of edges represents the temporal relationship between atomic tasks. Construct a directed weighted graph to transform the engineering problem into a graph theory problem of solving a reachable path from the starting point to the end point. This method uses the known initial state and action effects of the space mission to avoid invalid state space search processing during constraint processing and reduce the complexity of planning solution calculations.

[0009] (3) Construct a queue to handle constraint propagation and consistency problems in the temporal network structure diagram, design a state space search method to implement forward pruning, and design a problem relaxation strategy to optimize the heuristic control process. This method closely fits the characteristics of the engineering task and does not require manual coding of domain-specific control rules. It greatly simplifies the complexity of algorithm design, completes the logical reasoning of simple temporal problems under multi-task activities, and realizes rapid constraint propagation and consistency judgment.

[0010] A general spacecraft heuristic temporal planning modeling and solution method includes the following steps:

[0011] 1. Build a domain model.

[0012] The temporal planning problem of a spacecraft is represented as a triple P = (∑, s0, G), where

[0013] ∑ represents the task activity, and ∑ = (S, A, δ) is defined, where S is the task state set, A is the action set, and δ is the state transition function;

[0014] s0 is the initial state of the task, s′ is any state in S;

[0015] G is a set of target states, where each target proposition has the form S g ={s′∈S|s satisfies G}.

[0016] Let the state transition function δ = <T a , exp>, where the temporal semantics of the action is T a = <st a , et a >, further explained as follows: the start and end time of action a are recorded as st a et a For the action execution condition c∈C a , if st c =et c =st a , then condition c is required to be established at the beginning of a, and such conditions are called “starting conditions”; if st c =et c =eta , then c is required to be established at the end of a, and such conditions are called “end conditions”; if st c =st a , et c =et a , then c is required to be in the open interval (st c , et c ), and such conditions are called “continuity conditions”. m ∪X, the effect of action a <[t]v←d>, if t=st a , then the effect occurs at the beginning of the action, and this type of effect is called a "starting effect"; if t = et a The effect occurs at the end of the action, and this type of effect is called an "end effect".

[0017] The cost function that represents the execution of an action is an expression composed of numerical variables and constants, which defines the action exp

[0018] The relationship between task execution and energy consumption.

[0019] Formal definition of action set A <dur a , C a , E a >, where dur a Indicates the duration of an action; C a It is the execution condition set of a (abbreviated as: condition set), which describes the conditions that must be met at the start of execution, the conditions that are met at the end of execution, and the conditions that are maintained during execution; E a It is the set of execution effects of action a (referred to as effect set), which includes the effects produced by action a at the start and end of execution;

[0020] For the condition c∈C a If it constrains a logical variable, it has the form <(st c , et c )v=d>, d∈Dom(v), if it constrains a numerical variable, then it has the form <(st c , et c )vop exp>'op∈{>,≥,<,≤=}a∈A'exp is an expression composed of numerical variables and constants, which is used to define the relationship between action execution and energy consumption. The numerical variables can be obtained from the external input parameters. f ∈E a 'If it affects logical variables, it has the form <[t]v←d>, if it affects numerical variables, it has the form <[t]v eop exp>, eop∈{=, +=, -=, *=, / =}°

[0021] The solution to the planning problem P is a sequence of actions (a1, ..., a k ) corresponds to the target state transition sequence (s0, ..., s k ), satisfying s1=δ(s0,a1),...,s k =δ(s k-1 , a k ), and s k is a subset of the target state set G.

[0022] 2. According to the task implementation plan and customized file information, assign initial values ​​to complete the instantiation of the engineering problem.

[0023] 3. Establish a heuristic solution method.

[0024] a. According to the conditional constraints, construct a temporal network structure diagram and determine the values ​​of each node and edge;

[0025] By decomposing an engineering task into a set of independent atomic tasks, the constraint satisfiability problem related to the temporal constraints can be represented as a temporal network for solution.

[0026] Taking the rover's patrol and exploration of the surface of extraterrestrial bodies as an example, the atomic mission refers to the various working modes of the rover to complete the patrol mission, including perception mode, charging mode, mobile mode, sleep mode, etc.

[0027] The temporal network structure diagram is a directed weighted graph. Each node in the network represents an independent atomic task ∑′, satisfying ∑′=(S′,A′,δ), and defines the task attributes and status; the set of nodes represents an engineering task P, satisfying the triple P=(∑′,s0,G).

[0028] Each edge is represented by a state transition function δ, satisfying δ=(T a , exp>, defines the time and energy constraints between different actions. The action start time st a ', action end time et a ′ and the constant of the action cost exp together constitute the weight of the weighted graph.

[0029] The direction of the edge defines the temporal relationship between atomic tasks and provides a basis for judging the consistency of constraint propagation. The number of nodes is defined as M and the number of directed edges is defined as K. The relationship between the two conforms to the following logical relationship:

[0030]

[0031] b. Use forward pruning strategy to propagate constraints and solve problems on temporal network structure graphs;

[0032] Construct a queue Q, and read the nodes of the directed weighted graph constructed in step a in sequence to satisfy Q = {∑′}. Constraint judgments are made on the energy constraints and state attributes of the nodes in a "first-in, first-out" order until the queue is empty. In the process, a forward pruning strategy is used to remove nodes and edges that do not meet the constraints to complete constraint propagation; a reachable path (not necessarily the shortest path) from the start point to the end point of the directed weighted graph is obtained, that is, the action sequence required to complete a series of atomic tasks is obtained.

[0033] The steps are as follows:

[0034] 1) Create a Closed table and an Open table to record the search space and the search boundary. The Closed table is used to record the nodes that the search algorithm has traversed, and the Open table is used to record several states on the search boundary.

[0035] 2) Read the engineering configuration file and determine the remaining energy E of node n n Is it satisfied to complete action A? n Minimum energy consumption requirement.

[0036] Define E′ n To estimate the energy consumed from executing from the current node n to the next node n+1, the action execution cost function exp defined in the state transfer function δ is called, E′ n =exp n+1 ; Define the total energy as E, and the remaining energy of the current node n as E n , i-node serial number;

[0037] Set judgment condition E n <E′ n .

[0038] If the result is true, continue to execute status judgment and go to the next step;

[0039] If the result is false, return to the previous node and exit the solution process, indicating that the energy does not meet the requirements for subsequent action execution and there is no feasible solution.

[0040] 3) Determine the state s of node n n Is it a subset of the target state G?

[0041] If satisfied, go to step c, perform constraint propagation and consistency judgment according to the problem relaxation strategy, and calculate the action set A of node n n , forming an available action set App = {A n}, n∈N+, and put the result into the Open table;

[0042] If not, trace back to the state s of node nn The action sequence to the initial state s0.

[0043] In programming implementation, it is necessary to call other configuration items such as light forecast and other specific dynamic parameters to convert variable parameters into numerical calculations.

[0044] See the process for details Figure 1 "Intelligent Planning Solver Algorithm Framework".

[0045] c. Design problem relaxation strategies to improve solution efficiency.

[0046] In the classic action planning model, the duration of the action is not considered and the energy constraint is taken as a constant. The number of actions completed per unit time is used to measure the quality of the planning results. In actual engineering problems, there are many factors that affect the planning results, including: the action has a duration and is subject to multiple constraints. The time constraints include the communication time window determined by the data transmission plan, the measurement and control time window determined by the orbit forecast and local time coordinates, and the charging time window determined by the spacecraft attitude control and illumination forecast results. The energy constraint is a dynamically changing value rather than a constant. It is necessary to input platform or payload related parameters for calculation to obtain a relatively accurate value. If all constraints are solved one by one, the calculation efficiency will be seriously affected when processing network planning diagrams with a large number of nodes.

[0047] In order to solve complex planning problems efficiently, the method discussed in the present invention adopts a problem relaxation strategy to improve the efficiency of solving. That is, by simplifying the problem constraints, relaxing the action constraints, using node n to solve the relaxed problem and obtain an estimate of the expected degree of node n. The closer the relaxed problem is to the original problem, the more accurate the estimate is. On the other hand, the simpler the relaxed problem is, the simpler the heuristic calculation will be.

[0048] Specific methods include:

[0049] The mission implementation plan has clear requirements for the premise, state and effect of various actions of the spacecraft. Therefore, the heuristic algorithm solves a deterministic selection problem. The node selected by the algorithm is a feasible solution that meets the execution conditions. Among them, the time-related constraints are represented by the start time. An atomic task includes four parameters: the earliest start time, the latest start time, the earliest end time, and the latest end time. The directed distance between nodes can measure the degree of realization from the current state to the target state. Therefore, the optimal distance of a task activity from the initial state to the target state depends on the critical path between the initial state atomic task and the target state atomic task.

[0050] 1) Expand the grammatical framework of classic action planning and add action duration dur to the domain model a Representing time intervals, it provides an important mathematical tool for reasoning about task temporal relations.

[0051] 2) In the temporal network structure diagram, a heuristic control function based on the maximum time span (Max-span heuristic) is defined using the earliest start time of the node and the duration of the action:

[0052] Satisfy S n ={s′∈S|s satisfies G}, where n∈N+

[0053] 3) Use the heuristic control function to select nodes, calculate the directed distance between nodes to measure the degree of realization from the current state to the target state; the selected node is a feasible solution that meets the execution conditions.

[0054] 4) Use the "target completion time" to measure the completion of task activities, and use the relaxation value of a given target state set to reflect the degree of solution of the current temporal planning problem.

[0055] During the execution process, the fewer wrong nodes the heuristic function h(n) selects, the less computational space the algorithm requires to solve the problem, and the higher the efficiency of problem solving.

[0056] This design simplifies the complexity of algorithm design, completes the logical reasoning of simple temporal problems under multi-task activities, and realizes rapid constraint propagation and consistency judgment.

[0057] 4. Submit the domain model, instance file and solution algorithm to the planner to obtain the spacecraft action sequence. Description of the drawings:

[0058] Figure 1 :Intelligent planning solution algorithm framework

[0059] Table 1: Analysis of task scenario examples Specific implementation method:

[0060] The present invention is further described in detail below in conjunction with the accompanying drawings and implementation examples. It should be understood that the specific implementation examples described herein are only used to explain the present invention and do not limit the present invention.

[0061] Set the task scenario as follows:

[0062] The rover is currently located at S1 (coordinates: -26.7, -1.2), and scientists have designated it to move to the target point (coordinates: -45.1, -1.1) for scientific exploration. The remaining navigation points between the current point and the target point are S2 (coordinates: -32.5, -0.9) and S3 (coordinates: -37.9, -0.9).

[0063] Mission activities include entering the measurement and control area, sensing, moving, charging, and leaving the measurement and control area. Among them, the energy consumption in the sensing mode is 0.1 ampere-hour / minute; the energy consumption in the moving mode is 1.38 ampere-hour / minute; entering and leaving the measurement and control area does not consume electricity.

[0064] Charging is an action with a variable duration, the maximum duration is 100 minutes, and the charging rate is 1 ampere-hour / minute. The premise for its execution is that it is in the light range, the current energy is less than 50 ampere-hours, and charging is allowed. The effect of its execution is to increase the energy of the patroller, and the increase value is the product of the actual execution time and the charging rate. Regardless of the solar azimuth, when the solar altitude angle is >15°, the patroller can be considered to be in the light range.

[0065] The implementation steps are as follows:

[0066] (1) Build a domain model.

[0067] (2) Based on the above task scenario description, assign initial values ​​to complete the instantiation of the engineering problem.

[0068] (3) Establish a heuristic solution method.

[0069] a. According to the conditional constraints, construct a temporal network structure diagram and determine the values ​​of each node and edge;

[0070] b. Use forward pruning strategy to propagate constraints and solve problems on the network structure graph;

[0071] c. Relax the problem based on the heuristic control function of the maximum time span to improve the solution efficiency.

[0072] (4) Start the task planning calculation and obtain the action sequence as shown in Table 1. In this example, the Sapa planner is used to process the domain model and instance model files.

[0073] Table 1: Analysis of task scenario examples

[0074]

Claims

1. A general spacecraft heuristic temporal planning modeling and solving method, characterized in that: The method comprises the following steps:

1. Build a domain model; The step 1 further comprises: The temporal planning problem of the spacecraft is represented as a triple P = (∑, s0, G), where ∑ represents the task activity. Define ∑ = {S, A, δ), S is the task state set, A is the action set, δ is the state transition function; s0 is the task initial state, s′ is any state in S; G is the target state set, where each target proposition has the form S g ={s′∈S|s satisfies G}; let the state transfer function δ = <T a , exp>, where the temporal semantics of the action is T a = <st a , et a >, the start and end time of action a are recorded as st a et a , exp represents the cost function of action execution, which is an expression composed of numerical variables and constants, defining the relationship between action execution and energy consumption; formally defining the action set A = 〈dur a , C a , E a 〉, among which, dur a Indicates the duration of an action; C a It is the set of execution conditions of a, referred to as the condition set, which describes the conditions that must be met at the start of execution, the conditions that are met at the end of execution, and the conditions that are maintained during execution; E a It is the set of execution effects of action a, referred to as effect set, which includes the effects of action a at the start and end of execution. The solution to the planning problem P is a sequence of actions (a1, ..., a k ) corresponds to the target state transition sequence (s0, ..., s k ), satisfying s1=δ(s0,a1),…,s k =δ(s k-1 , a k ), and s1...s k is a subset of the target state set G; 2. According to the task implementation plan and customized file information, assign initial values ​​to complete the instantiation of engineering problems; 3. Establish a heuristic solution method; The step three further comprises: a. According to the conditional constraints, construct a temporal network structure diagram and determine the values ​​of each node and edge; In the step 3, the method for constructing the temporal network structure diagram is as follows: let each node of the temporal network structure diagram represent an independent atomic task ∑′, satisfying ∑′=(S′, A′, δ), and define the task attributes and states; the set of nodes represents an engineering task P, satisfying the triple P=(∑′, s0, G); each edge is represented by a state transfer function δ, satisfying δ= <T a , exp>, defines the time and energy constraints between different actions; the action start time st a ', action end time et a ′ and the constant of the action cost exp together constitute the weight of the weighted graph; the direction of the edge defines the temporal relationship between atomic tasks and provides a basis for judging the consistency of constraint propagation. The number of nodes is defined as M and the number of directed edges is defined as K. The relationship between the two conforms to the following logical relationship: b. Use forward pruning strategy to propagate constraints and solve problems on temporal network structure graphs; c. Design problem relaxation strategy to improve solution efficiency; In step 3, the method for determining the optimal solution of the relaxation problem is as follows: 1) Add action duration dur to the domain model a Used to indicate time intervals; 2) In the temporal network structure diagram, a heuristic control function based on the maximum time span is defined using the earliest start time of the node and the duration of the action: Satisfy S n ={s′∈S|s satisfies G}, where n∈N+; 3) Select nodes using heuristic control functions, calculate the directed distance between nodes to measure the degree of realization from the current state to the target state; the selected node is a feasible solution that meets the execution conditions, where the time-related constraints are represented by the start time, and an atomic task includes four parameters: the earliest start time, the latest start time, the earliest end time, and the latest end time; 4) Use "target completion time" to measure the completion of task activities, and use the relaxation value of a given target state set to reflect the degree of solution of the current temporal planning problem; 4. Submit the domain model, instance file and solution algorithm to the planner to obtain the spacecraft action sequence.

2. The spacecraft heuristic temporal planning modeling and solving method according to claim 1, characterized in that: In step 3, the process of using the forward pruning strategy to perform constraint propagation and problem solving on the temporal network structure graph is as follows: Construct a queue Q, and read the nodes of the directed weighted graph constructed in step a in sequence, so that it satisfies Q = {∑′}; The energy constraints and state attributes of the nodes are constrained in a "first-in, first-out" order until the queue is empty. In the process, a forward pruning strategy is used to remove nodes and edges that do not meet the constraints to complete constraint propagation. A reachable path from the starting point to the end point of the directed weighted graph is obtained, that is, the action sequence required to complete a series of atomic tasks is obtained.

3. The spacecraft heuristic temporal planning modeling and solving method according to claim 1, characterized in that: In the step 3, the process of using the forward pruning strategy to perform constraint propagation and problem solving on the temporal network structure graph further includes: 1) Create a Closed table and an Open table, where the Closed table is used to record the nodes that the search algorithm has traversed, and the Open table is used to record several states on the search boundary; 2) Read the engineering configuration file and determine the remaining energy E of node n n Is it satisfied to complete action A? n Minimum energy consumption requirements; Definition of E′ n To estimate the energy consumed from executing from the current node n to the next node n+1, the action execution cost function exp defined in the state transfer function δ is called, E′ n =exp n+1 ; Define the total energy as E, and the remaining energy of the current node n as E n , i-node serial number; Set judgment condition E n <E′ n ; If the result is true, continue to execute status judgment and go to the next step; If the result is false, the solution process is exited, indicating that the energy does not meet the requirements for subsequent action execution and there is no feasible solution; 3) Determine the state s of node n n Is it a subset of the target state G? If satisfied, constraint propagation and consistency judgment are performed according to the problem relaxation strategy, and the action set A of node n is calculated. n , forming an available action set App = {A n }, n∈N+, and put the result into the Open table; If it is not satisfied, then trace back to the state s of node n n The action sequence to the initial state s0.

4. The spacecraft heuristic temporal planning modeling and solving method as claimed in claim 3, characterized in that: In step 3, the method for determining the optimal solution of the relaxation problem is as follows: 1) Add action duration dur to the domain model a Used to indicate time intervals; 2) In the temporal network structure diagram, a heuristic control function based on the maximum time span is defined using the earliest start time of the node and the duration of the action: Satisfy S n = [s′∈S|s satisfies G}, where n∈N+; 3) Select nodes using heuristic control functions, calculate the directed distance between nodes to measure the degree of realization from the current state to the target state; the selected node is a feasible solution that meets the execution conditions, where the time-related constraints are represented by the start time, and an atomic task includes four parameters: the earliest start time, the latest start time, the earliest end time, and the latest end time; 4) Use "target completion time" to measure the completion of task activities, and use the relaxation value of a given target state set to reflect the degree of solution of the current temporal planning problem.

Images