WO2019127945A1

WO2019127945A1 - Structured neural network-based imaging task schedulability prediction method

Info

Publication number: WO2019127945A1
Application number: PCT/CN2018/080419
Authority: WO
Inventors: 何敏藩; 邢立宁; 白国庆; 石建迈; 王锐; 谭旭; 陈剑; 黄勇; 熊彦; 甘文勇
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: 2018-04-24
Publication date: 2019-07-04
Also published as: CN108717572B; CN108717572A

Abstract

A structured neural network-based imaging task schedulability prediction method. Schedulability predication for an imaging task is completed by constructing and extracting a task planning result sample set eigenvalue, and constructing a structured neural network model so that the structured neural network model establishes a nonlinear mapping relation between a task eigenvalue and a satellite capability in a learning process. The method has advantages of a high parameter interpretation capability, effective settlement for defects present in a traditional feedforward neural network model, such as an unstructured model, a low convergence speed, difficulty in determining the number of neurons, and local minimum, and high prediction precision.

Description

Imaging task schedulability prediction method based on structured neural network

Technical field

The present invention relates to the technical field of neural network prediction, and in particular to a structured neural network based imaging task schedulability prediction method.

Background technique

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”.

The working modes and usage constraints of different remote sensing satellites are very complex, and generally have relatively independent mission planning systems. With the increasing number of remote sensing satellites and imaging tasks, how to use a large amount of historical data accumulated by different remote sensing satellite mission planning systems The advanced theoretical design imaging task schedulability prediction method in the fields of intelligence and operations research has very important theoretical significance and practical value.

The imaging task schedulability prediction can be expressed as a six-tuple <J _T , J _p , S, C, X, G>, that is, for the resource set S, the constraint set C, and the optimization target G, based on the scheduled task sample set data J _T is assigned to the decision variable X={x ₁ ,...,x _j } of the new task sample set J _p . The difficulty in solving this problem is mainly reflected in the following four aspects.

(1) The complexity of the task planning problem. Intelligent satellite mission planning has certain specialities in tasks, resources, constraints and optimization objectives. Common resource scheduling models and optimization methods are difficult to solve.

(2) The complexity and uncertainty of the scheduling algorithm. The randomness of the scheduling algorithm makes the scheduling result also uncertain, and also increases the difficulty of schedulability prediction.

(3) The complexity of task sample selection. Different satellites will accumulate a large amount of historical mission data during the orbital operation. How to select typical representative samples to improve the execution efficiency of the prediction algorithm is difficult.

(4) The complexity of sample feature extraction. Imaging tasks generally have both static and dynamic attribute characteristics: static attributes are mainly related to tasks that do not change with the change of the task set, such as the data type, resolution, priority, and demand observation duration of the imaging task. Meteorological conditions and imaging modes; dynamic attributes change with the set of tasks, such as describing resource competition between tasks and conflicts of observation opportunities. How to select the features that have a decisive influence on the prediction process among the various attributes is also very complicated.

In summary, a new imaging task schedulability prediction method is needed to overcome the above shortcomings and meet the demand.

Summary of the invention

The object of the present invention is to overcome the deficiencies of the prior art, and to provide a parameter interpretation capability capable of effectively solving the existence of a traditional feedforward neural network model such as model unstructured, slow convergence rate, difficulty in determining the number of neurons, and localization. A structured neural network-based imaging task schedulability prediction method with minimum defects and high prediction accuracy.

In order to achieve the above object, the technical solution provided by the present invention is to construct a structured neural network model by constructing and extracting feature values of a sample set of task planning results, so as to establish a relationship between task feature values and satellite capabilities during the learning process. Nonlinear mapping relationship to complete schedulability prediction of imaging tasks;

Specific steps are as follows:

S1, definition of scheduling scenarios and imaging tasks:

Scheduling scenario: {S _i =<J _i ,O _i ,C>|i=0,...,n}, where J _i is the task set assigned to satellite i, and SubJ _i represents the child of task set J _i Set, the set of imaging opportunities that each task in SubJ _i has for satellite i is W _i , and C is a set of constraints for satellite use;

Task _{_{job i = <p i, d}} i, w oi, w fi>, p i ∈ [1,8] for the priority, the greater the more important; d _i represents the duration of the imaging job _i, task job _i Observations must be arranged within a given time [w _oi , w _fi ];

Assuming that o _sj and o _ej represent the start time and end time of the imaging opportunity j, respectively, all the imaging opportunities of job _{i are} expressed as O _i ={<o _s1 , o _e1 , sl ₁ >,..., <o _sj ,o _{_{ej, sl j>, ...,}} <o sm, o em, sl m>}, where _j denotes SL placed job _i j corresponding angle side imaging opportunities, in the range of 0-180 degrees;

Suppose that the feature vector of job _i is defined as {f ₁ , f ₂ , f ₃ , f ₄ , f ₅ }, where

f ₁ :Duration _i =d _i ,

f ₂ :Priority _i =p _i ∈[1,8],

f ₅ :Conflict _i , indicating the conflict between job _i and other mission observation opportunities;

S2, constructing a structured neural network model;

S3. Determine an input layer neural node and an output layer neural node;

S4. Perform schedulability prediction of the imaging task after training through multiple sets of data.

Further, the structured neural network model in step S2 is integrated by multiple BP neural networks with different hidden layer nodes, and all the connection relationships between the nodes of each BP neural network are constructed based on the causal relationship of the actual actual system. .

Further, the input layer neural node corresponds to five characteristic values of {f ₁ , f ₂ , f ₃ , f ₄ , f ₅ }; the output layer neural node is a characteristic value Scheduled _i = {-1, 1}, and the task job _i passes If the scheduling enters the imaging scheme, Scheduled _i =1, indicating that the scheduling is successful; otherwise Scheduled _i = -1.

Further, the calculation process of Conflict _i is: first input O _i ={<o _s1 , o _e1 , sl ₁ >,..., <o _sj , o _ej , sl _j >,..., <o _sm , o _em ,sl _m >}, i=1,2,,n, satellite yaw average velocity v, Conflict _i =0; then each task k imaging opportunity belonging to all mission imaging opportunity sets in SubJ _i ow _k =< o _sk , o _ek , sl _k > and each task i imaging opportunity that does not belong to all task imaging opportunity sets in SubJ _i ow _i =<o _si ,o _ei ,sl _i >one-to-one comparison; if ow _k =< o _sk , o _ek , sl _k > and ow _i =<o _si , o _ei ,sl _i > partially overlap, then Conflict _i adds one; if ow _k =<o _sk ,o _ek ,sl _k >the corresponding side The product of the absolute value of the angular difference of the corresponding sway angle sl _i and the average velocity v of the satellite yaw in the swing angles sl _k and ow _i =<o _si ,o _ei ,sl _i > plus ow _k =<o _sk , o _ek , sl _k > end time o _{ek is} greater than ow _i = < o _si , o _ei , sl _i > start time, then Conflict _i plus one; if ow _i = < o _si , o _ei , sl _i > The absolute value of the angular difference of the corresponding side sway angle sl _k and the average yaw rate of the satellite yaw in the corresponding sway angles sl _i and ow _k =<o _sk , o _ek ,sl _k > The product of ow _i =<o _si ,o _ei ,sl _i >end time o _{ei is} greater than the start time of ow _k =<o _sk ,o _ek ,sl _k >, then Conflict _i adds one; The value of the Conflict _i .

The specific steps of the imaging task schedulability prediction in step S4 are as follows:

S4-1. Performing prediction by using multiple BP neural networks with different same to-be-predicted data through different hidden layer nodes;

S4-2, sorting the prediction success rate corresponding to different hidden layer nodes from high to low;

S4-3, creating multiple BP neural network integrations with a singular number, and the plurality of BP neural network integrations are respectively composed of BP neural networks that predict the success rate of the first row;

S4-4. The results of the integrated output of the plurality of BP neural networks obtained according to step S4-3 are voted by the majority voting method, thereby generating an imaging task schedulability prediction result.

The principle of this scheme is as follows:

Through the construction and extraction of the eigenvalues of the task planning result set, a structured neural network model for schedulability prediction of imaging tasks is adopted. The structured neural network model can establish the task eigenvalue and satellite capability during the learning process. The non-linear mapping relationship between them completes the schedulability prediction of the imaging task.

Compared with the prior art, the advantages are as follows:

1. Construct a structured neural network model based on multiple BP networks, which has a good ability to interpret model parameters.

2. When the actual scheduling results are fed back online, the schedulability prediction model can be updated, and the use of the task schedulability model makes the distributed bilevel programming problem easier to solve.

3. It can effectively solve various defects such as unstructured model, slow convergence rate, difficult to determine the number of neurons and local minimum in the traditional feedforward neural network model.

4. Input layer neural nodes correspond to five characteristic values of {f ₁ , f ₂ , f ₃ , f ₄ , f ₅ }, especially the introduction of two eigenvalues of f ₂ priority and f ₅ conflict condition, greatly improving BP neural Network prediction accuracy.

DRAWINGS

1 is a flowchart of a method for predicting schedulability of an imaging task based on a structured neural network according to the present invention;

2 is a prediction effect diagram of different attribute input changes with hidden layer nodes based on BP neural network in the present invention;

3 is a diagram showing prediction effects of various BP neural networks for different data sets in the present invention;

4 is a corresponding priority distribution of 2000 sets of test data in the present invention; and a scheduling success rate change pattern after a scheduling method;

5 is a graph showing a relationship between a task priority based on 2000 sets of test data and a predicted success rate of BP neural network in the present invention;

6 is a diagram showing a distribution of corresponding task conflicts of 2000 sets of task data and a change of scheduling success rate with task conflict degree after the scheduling method according to the present invention;

Figure 7 is a graph showing the relationship between the task conflict degree based on 2000 sets of test data and the predicted output success rate of the BP neural network in the present invention.

Detailed ways

The present invention will be further described below in conjunction with specific embodiments:

Referring to FIG. 1 , a structured neural network-based imaging task schedulability prediction method according to the embodiment includes the following steps:

S1, definition of scheduling scenarios and imaging tasks:

Assuming that o _sj and o _ej represent the start time and end time of the imaging opportunity j, respectively, all the imaging opportunities of job _{i are} expressed as O _i ={<o _s1 , o _e1 , sl ₁ >,..., <o _sj ,o _{_{ej, sl j>, ...,}} <o sm, o em, sl m>}, where _j denotes SL placed job _i j corresponding angle side imaging opportunities;

f ₁ :Duration _i =d _i ,

f ₂ :Priority _i =p _i ∈[1,8],

S2. Construct a structured neural network model:

The structured neural network model is integrated by multiple BP neural networks with different hidden layer nodes. All the connection relationships between the nodes of each BP neural network are constructed based on the causal relationship of the actual actual system.

S3. Determine the input layer neural node and the output layer neural node:

The input layer neural nodes correspond to five characteristic values of {f ₁ , f ₂ , f ₃ , f ₄ , f ₅ };

Among them, the calculation process of Conflict _i is: first input O _i ={<o _s1 ,o _e1 ,sl ₁ >,...,<o _sj ,o _ej ,sl _j >,...,<o _sm ,o _Em ,sl _m >}, i=1,2,,n, satellite yaw average velocity v, Conflict _i =0; then each task k imaging opportunity belonging to all mission imaging opportunity sets in SubJ _i ow _k =<o _Sk , o _ek , sl _k > and each task i imaging opportunity that does not belong to the set of imaging opportunities in SubJ _i ow _i =<o _si ,o _ei ,sl _i >one-to-one comparison; if ow _k =<o _Sk , o _ek , sl _k > and ow _i =<o _si , o _ei ,sl _i > partially overlap, then Conflict _i increases by one; if ow _k =<o _sk ,o _ek ,sl _k >the corresponding side sway sl _k and angles _{_{ow i = <o si, o}} ei, sl i> corresponding side swing angle difference angle sl _i and an absolute value of the satellite-side pendulum volume average velocity v plus ow _k = <o _sk, o The end time o _{ek of} _ek ,sl _k > is greater than the start time of ow _i =<o _si ,o _ei ,sl _i >, then Conflict _{i is} incremented by one; if ow _i =<o _si ,o _ei ,sl _i > The product of the absolute value of the angular difference between the corresponding sway angle sl _k and the average yaw velocity v of the satellite yaw angle sl _i and ow _k =<o _sk , o _ek ,sl _k > Adding the end time o _{ei of} ow _i =<o _si ,o _ei ,sl _i > is greater than the start time of ow _k =<o _sk ,o _ek ,sl _k >, then Conflict _{i is} incremented by one; finally the final Conflict is obtained. _i value.

The output layer neural node is the characteristic value Scheduled _i = {-1, 1}, and the task job _i is scheduled to enter the imaging scheme, then Scheduled _i =1, indicating that the scheduling is successful; otherwise, Scheduled _i = -1.

S4. Perform schedulability prediction of the imaging task after training through multiple sets of data; wherein the specific steps of performing schedulability prediction of the imaging task are as follows:

In the above, the two BP neural networks use the same 1900 sets of data for training, 100 sets of data for predictive testing, and the number of hidden layer nodes is uniformly determined by values between 1-49. 10 learnings were performed for each node number, and the predicted success rate was averaged. As shown in Fig. 2, the addition of the eigenvalue Conflict effectively improves the prediction success rate of the BP neural network. As the number of hidden layer nodes increases, the improvement effect becomes more stable, with an average increase of 2.5 percentage points.

Before and after the two BP neural networks respectively learn the number of hidden layer nodes with the best effect in Figure 2. The 2000 group data was divided into 20 parts on average, 100 sets of each group were used as test data, and the remaining 1900 sets were used as training data. Each group of data was studied 10 times, and the predicted success rate was averaged. As shown in Figure 3, the BP neural network with five attribute inputs plays a better prediction effect on different data sets.

In addition, as shown in FIG. 4, based on the 2000 group task data, all tasks with priority level 8 are successfully scheduled, and the scheduling success rate with priority level 1 is the lowest. However, the task scheduling success rate with the priority in the middle does not increase strictly with the increase of the priority.

As shown in FIG. 5, the task schedulability prediction success rate of task priority 8 is 1, and the priority is 1 corresponding success rate is also greater than 0.95. The lowest and highest priority become good discriminators in the task schedulability prediction process. While other priorities do not have a good classification effect, combined with Figure 4, the visible input feature value Priority is closely related to the output feature value Scheduled predictive output.

As shown in FIG. 6, the overall trend of scheduling success rate decreases as the degree of conflict increases. When the degree of conflict is greater than 14, no task scheduling succeeds;

As shown in FIG. 7, when the degree of conflict is greater than 12, the predicted success rate is greater than 95%.

Multi-BP neural network integration prediction based on the number of hidden layer nodes:

The basic data used in Figure 2 and the variation range of the hidden layer nodes are changed from 2 to 50. The prediction success rate of each BP neural network is as shown in Table 1:

Table 1

The integration of several different BP neural networks mentioned above is used to obtain the predicted effects as shown in Table 2:

Table 2

From Table 2, the performance of the task schedulability prediction component based on multi-BP neural network integration is more stable than that of a single BP neural network, and the highest prediction success rate is 91%.

Compared with the prior art, this embodiment has the following advantages:

The embodiments described above are only the preferred embodiments of the present invention, and are not intended to limit the scope of the present invention. Therefore, variations in the shapes and principles of the present invention are intended to be included within the scope of the present invention.

Claims

A structuring predictive method for imaging tasks based on structured neural network is characterized in that a structured neural network model is constructed by constructing and extracting feature values of the sample set of task planning results, so that task feature values are established during the learning process. A non-linear mapping relationship between satellite capabilities to complete schedulability predictions for imaging tasks;

Specific steps are as follows:

S1, definition of scheduling scenarios and imaging tasks:

Scheduling scenario: {S i =<J i ,O i ,C>|i=0,...,n}, where J i is the task set assigned to satellite i, SubJ i represents a subset of task set J i , Each of the tasks in SubJ i has a set of imaging opportunities for satellite i as W i , and C is a constellation set for satellite use;

Task job i = <p i, d i, w oi, w fi>, p i ∈ [1,8] for the priority, the greater the more important; d i represents the duration of the imaging job i, task job i Observations must be arranged within a given time [w oi , w fi ];

Assuming that o sj and o ej represent the start time and end time of the imaging opportunity j, respectively, all the imaging opportunities of job i are expressed as O i ={<o s1 , o e1 , sl 1 >,..., <o sj , o ej , Sl j >,..., <o sm , o em , sl m >}, where sl j represents the corresponding side angle of job i in imaging opportunity j;

Suppose that the feature vector of job i is defined as {f 1 , f 2 , f 3 , f 4 , f 5 }, where

f 1 :Duration i =d i ,

f 2 :Priority i =p i ∈[1,8],

f 3 :

f 4 :

f 5 :Conflict i , indicating the conflict between job i and other mission observation opportunities;

S2, constructing a structured neural network model;

S3. Determine an input layer neural node and an output layer neural node;

S4. Perform schedulability prediction of the imaging task after training through multiple sets of data.
The structured neural network-based imaging task schedulability prediction method according to claim 1, wherein the structured neural network model in step S2 is integrated by multiple BP neural networks with different hidden layer nodes, each All the connection relationships between the nodes of the BP neural network are constructed based on the causal relationship of the actual actual system.
The structured neural network-based imaging task schedulability prediction method according to claim 1, wherein the input layer neural nodes correspond to {f 1 , f 2 , f 3 , f 4 , f 5 } The eigenvalue; the output layer neural node is the eigenvalue Scheduled i = {-1, 1}, and the task job i is scheduled to enter the imaging scheme, then Scheduled i =1, indicating that the scheduling is successful; otherwise, Scheduled i = -1.
The method for predicting schedulability of an imaging task based on structured neural network according to claim 1, wherein the calculation process of the Conflict i is: first inputting O i ={<o s1 , o e1 , sl 1 >,...,<o sj ,o ej ,sl j >,...,<o sm ,o em ,sl m >},i=1,2,,n, satellite yaw average velocity v,Conflict i =0; each belongs task SUBJ i all tasks imaging opportunity set k of imaging opportunities ow k = <o sk, o ek, sl k> and task i imaging every opportunity not all tasks imaging opportunities SUBJ i in the set ow i = <o si ,o ei ,sl i >one-to-one comparison; if ow k =<o sk ,o ek ,sl k > and ow i =<o si ,o ei ,sl i > partially overlap, then Conflict i plus One; if ow k =<o sk ,o ek ,sl k >the corresponding sway angles sl k and ow i =<o si ,o ei ,sl i >the angle difference of the corresponding side angles sl i The product of the absolute value and the satellite yaw average velocity v plus ow k =<o sk , o ek ,sl k >the end time o ek is greater than the start time of ow i =<o si ,o ei ,sl i >, then Conflict i plus one; if ow i =<o si ,o ei ,sl i >the corresponding side angles sl i and o w k =<o sk ,o ek ,sl k >the absolute value of the angular difference of the corresponding sway angle sl k and the product of the satellite yaw average velocity v plus ow i =<o si ,o ei ,sl i The end time o ei of > is greater than the start time of ow k = <o sk , o ek , sl k >, then Conflict i is incremented by one; finally the final Conflict i value is obtained.
The structured neural network-based imaging task schedulability prediction method according to claim 1, wherein the specific steps of the imaging task schedulability prediction in the step S4 are as follows:

S4-1. Performing prediction by using multiple BP neural networks with different same to-be-predicted data through different hidden layer nodes;

S4-2, sorting the prediction success rate corresponding to different hidden layer nodes from high to low;

S4-3, creating multiple BP neural network integrations with a singular number, and the plurality of BP neural network integrations are respectively composed of BP neural networks that predict the success rate of the first row;

S4-4. The results of the integrated output of the plurality of BP neural networks obtained according to step S4-3 are voted by the majority voting method, thereby generating an imaging task schedulability prediction result.