CN107194155A - A kind of threat assessment modeling method based on small data set and Bayesian network - Google Patents
A kind of threat assessment modeling method based on small data set and Bayesian network Download PDFInfo
- Publication number
- CN107194155A CN107194155A CN201710299124.8A CN201710299124A CN107194155A CN 107194155 A CN107194155 A CN 107194155A CN 201710299124 A CN201710299124 A CN 201710299124A CN 107194155 A CN107194155 A CN 107194155A
- Authority
- CN
- China
- Prior art keywords
- parameter
- node
- value
- threat
- data set
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
Classifications
-
- G—PHYSICS
- G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
- G16Z—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS, NOT OTHERWISE PROVIDED FOR
- G16Z99/00—Subject matter not provided for in other main groups of this subclass
Abstract
The invention provides a kind of threat assessment modeling method based on small data set and Bayesian network, for the threat assessment modeling problem under the conditions of small data set, constrained by the presence that Bayesian network side is introduced in the structural modeling stage, parameter monotonicity constraint is introduced in the parameter model stage, it compensate for small data set and include the problem of information is insufficient to cause modeling accuracy difference, the threat assessment modeling problem only under the conditions of small data set provides a feasible solution route, and provide the method referred to for the small data set modeling problem in other fields, with relatively broad application prospect.
Description
Technical field
Impend the modeling method of assessment the present invention relates to a kind of utilization machine learning techniques.
Background technology
The key link that fast and accurately threat assessment is actual operation is completed to threat source.On the one hand, it is intricate,
In fast changing operational environment, the factor such as enemy's interference and sensor performance limitation causes to obtain information content or observes data volume
It is insufficient；On the other hand, sometimes for the tactical mission for completing " first opposing decision-making, first enemy's strike ", it has to do not filled still in data
Timesharing, which impends, to be assessed and warfare decision.Therefore, a kind of assessment that can be impended under insufficient data qualification is needed badly
Insufficient data are referred to as small data set by the method for modeling here.The main thought of current threat assessment modeling method is basis
Object select threatening factors are assessed, and then these threatening factors are integrated.Main method has neutral net, supporting vector
Machine, Bayesian network, vague collection, expert system, fuzzy reasoning, linear weighted function and analytic hierarchy process (AHP) etc..In above-mentioned method
In, only Bayesian network and SVMs have the ability being modeled under the conditions of small data set, with SVMs
Compare, Bayesian network has the basic probability theory of similar human brain, explainable knowledge expression mode and flexible inference machine
System.Yu Zhouyi etc. is in document《Algorithm of Threat Level Assessment research based on Bayesian network》In have studied based on Bayesian network
Inference pattern and reasoning algorithm, but the Bayesian network model being directed to rule of thumb subjectivity provide, lack knot
Close data and carry out the process of model construction, while not having the ability of processing small data set.Qin Zheng et al. is in document《Based on mould
Paste the state threat assessment of dynamic bayesian network》In have studied data obfuscation process, data extract and threaten comment
The reasoning process estimated, but equally it is not directed to the Bayesian network model building process of small data set.
The content of the invention
In order to overcome the deficiencies in the prior art, the present invention provides a kind of threat assessment modeling method based on small data set.
The technical solution adopted for the present invention to solve the technical problems comprises the following steps：
Step 1, the threat data in battle field information is pre-processed, including target velocity V, entering angle sum A, enemy and we
Apart from Dis and targets'threat T, wherein, target velocityVuFor unmanned plane speed；
Entering angle sumFor the angle and unmanned plane velocity attitude and target wire clamp of target velocity direction and score
Angle sum；Enemy and we's distanceFor the distance between target and unmanned plane；Targets'threat
Step 2, using threat data as node, using Acquirement of field knowledge threat data bayesian network structure about
Beam, improves BIC scoringsWherein,
mijkIt is to meet X in sample datai=k, its father node π (Xi)=j number of samples,M is that sample data is total
Amount, e is the presence probability of corresponding sides in structural constraint, and δ is regulation parameter, and δ takes 10~100, i to represent i-th of node in network
Variable, n is node variable number, qiFor the father node value status number of i-th of node, riFor the child node value shape of i-th of node
State number；
Step 3, using K2 Algorithm Learning threat assessment network topology structures, step is as follows：
Step a, the but search the figure without side of n node variable is included since one；
Step b, according to node sequence ρ (X1, X2... Xi..., Xn) scan for, in nodes Xi, nodes X will be comeiBefore
Variable be combined, and therefrom select nodes XiFather node collection π (Xi), find out the father's section for making improvement BIC scorings reach maximum
Point set π (Xi)；
Step c, judges whether i is equal to n, if equal, algorithm terminates, output gained network structure；Otherwise, i values add
Return to step b in the lump；
Step 4, the network topology structure obtained using step 3 obtains the monotonicity constraint of network parameter, for by n change
Measure X={ X1, X2... XnComposition Bayesian network G, if the nodes X in GiHave r kinds value and value state set for 1,
2 ..., k ... r }, his father set of node π (Xi) there are q value and value state set for { 1,2 ..., j ... q }, make θijk
=p (Xi=k | π (Xi)=j), then there is θj1≤θij2≤......≤θijrOr θij1≥θij2≥......≥θijr；
Step 5, according to the normalization of parameter, i.e., father node value state is identical and one group of different parameter of child node value
It is being added and for 1, interval getparms
Step 6, it is believed that parameter is obeyed in the interval that step 5 is provided and is uniformly distributed, setting parameter θ obeys U (θ1, θ2),
[θ1, θ2] it is parameter value scope, with B (α1, α2) equivalent U (θ1, θ2), i.e.,Obtain
The hyper parameter of parameter beta prior distributionWithWherein, it is equally distributed to expectEqually distributed second moment
Step 7, the result of parameter learning is obtained using Bayesian EstimationN in formulaijkTo meet father
Node value is sample number when j, child node value are k, NijSample number when for father node value being j, [αi, αj] it is beta
The hyper parameter of distribution.
The beneficial effects of the invention are as follows：For the threat assessment modeling problem under the conditions of small data set, by being built in structure
The presence constraint on Bayesian network side is introduced in mode step section, parameter monotonicity constraint is introduced in the parameter model stage, compensate for small
Data set includes the problem of information is insufficient to cause modeling accuracy difference, not the threat assessment modeling only under the conditions of small data set
Problem provides feasible solution route, and provides and refer to for the small data set modeling problem in other fields
Method, with relatively broad application prospect.
Brief description of the drawings
Fig. 1 is unmanned plane battlefield scenario figure；
Fig. 2 is the presence constraint schematic diagram on Bayesian network side；
Fig. 3 is the intermediate result schematic diagram of this paper structure learning algorithms；
Fig. 4 is threat assessment BN structural representations obtained by 20 groups of data volumes, wherein, (a) is to threaten to comment obtained by this paper algorithms
Estimate BN structures, (b) is threat assessment BN structures obtained by original K2 algorithms；
Fig. 5 is the threat assessment BN structural representations obtained by 5000 groups of data of K2 algorithms.
Embodiment
The present invention is further described with reference to the accompanying drawings and examples, and the present invention includes but are not limited to following implementations
Example.
The invention provides a kind of threat assessment modeling method based on small data set, specifically comprising following steps：
Step 1：The pretreatment of threat data
Selection of the data prediction comprising deterrent and the discretization of threat data.Thought from battlefield as shown in Figure 1
It is fixed, threaten target to include radar and air defense position in scenario, we has man-machine commander's unmanned plane to carry out anti-pneumatics to enemy position
System.The complexity of battlefield surroundings determines that the factor related to threat is very more, and selected part deterrent of the present invention is used as prestige
The side of body assesses the factor that modeling considers, the method that selected deterrent and discretization is given below.
(1) target velocity (Velocity, V)：Velocity magnitude is refered in particular to, unmanned plane speed is designated as Vu
(2) entering angle sum (Angel, A)：Target velocity direction and the angle and unmanned plane velocity attitude and mesh of score
Graticule angle sum
(3) enemy and we's distance (Distance, Dis)：The distance between target and unmanned plane
(4) targets'threat (Threat, T), target is to our combat unit threat degree
Step 2：BIC scorings are improved using structural constraint.
(1) constrain as shown in Figure 2 according to the presence on the Bayesian network side of Acquirement of field knowledge.Wherein, dotted line represents two
The probability on side is not present between individual node, the solid line with arrow represents the probability that there is side.
(2) BIC score functions are improved by the presence constraint on the side provided in Fig. 2.With reference to the decomposability of score function,
The presence constraint on the side in Fig. 2 is dissolved into the way of a kind of similar partial structurtes priori in Structure learning scoring, and then
BIC scorings after to improvement are as shown in formula (1).
Wherein, mijkIt is to meet X in sample datai=k, its father node π (Xi)=j number of samples,
M is sample data total amount, and e is the presence probability of corresponding sides in structural constraint, and δ is regulation parameter, it is proposed that δ is between 10 to 100
Value, i is the subscript of nodes variable, and n is node variable number, qiFor father node value status number, riFor child node value
Status number.
Step 3：With reference to the BIC scorings after improvement, using K2 Algorithm Learning threat assessment network topology structures.Specific steps
It is as follows：
Step a：The but search the figure without side of n node variable is included since one.
Step b：According to node sequence ρ (X1, X2... Xi..., Xn) scan for, with XiExemplified by, nodes X will be comeiBefore
Variable be combined, and therefrom select XiFather node collection π (Xi), by calculating Vnew=BIC ((Xi, π (Xi)) | D), find out
Make Vnew=BIC ((Xi, π (Xi)) | D) reach maximum father node collection π (Xi), it should be noted that：X1Father node collection be empty
Collection.Here scoring is the improvement BIC scorings in step 2.
Step c：Judge whether i is equal to n, if equal, algorithm terminates, output gained network structure.Otherwise, i is performed
=i+1 return to step b.
Step 4：The network topology structure obtained using step 3 obtains the monotonicity constraint of network parameter.One by n change
Measure X={ X1, X2... XnComposition Bayesian network G, without loss of generality, if the nodes X in GiThere are r kinds value and value state
Collection is combined into { 1,2 ..., k ... r }, 1≤k≤r, his father set of node π (Xi) have q value and value state set for 1,
2 ..., j ... q }, 1≤j≤q.Make θijk=p (Xi=k | π (Xi)=j), then there is θij1≤θij2≤......≤θijrOr θij1
≥θij2≥......≥θijr。
Step 5：With reference to monotonicity constraint and the normative interval getparms of parameter.The normative of parameter contains
Justice is：Father node value state is identical and that one group of parameter that child node value is different is added and be 1.Specific conversion process can
Reference formula (2).
Step 6：Got parms the hyper parameter of beta prior distribution with reference to the interval of parameter in step 5.Formula (2) is given
The span of parameter is gone out, in the case of without other priori, then it is believed that parameter is obeyed in interval and is uniformly distributed.If ginseng
Number θ obeys U (θ1, θ2), [θ1, θ2] it is parameter value scope.How problem uses B (α if being converted into1, α2) equivalent U (θ1, θ2), i.e.,：
The problem of using described by moments estimation solution formula (3), shown in specific steps such as formula (4)~(6).
Wherein, [α1, α2] it is the hyper parameter that beta to be asked is distributed, M1The expectation being distributed for beta, M2Two be distributed for beta
Rank square, m1Expect to be equally distributed, m2For equally distributed second moment.
Step 7：The hyper parameter of the beta prior distribution obtained with reference to step 6, parameter learning is obtained using Bayesian Estimation
Result.Specific method refers to N in formula (7), formulaijkIt is j, sample when child node value is k to meet father node value
Number, NijSample number when for father node value being j, [αi, αj] it is the hyper parameter that beta is distributed.
Virtual implementation is carried out to this algorithm by Computer Simulation, specific implementation process is as follows：
Step 1 obtains battle field information and carries out discretization, enter by being emulated to the battlefield task scenario shown in Fig. 1
And threat information data are obtained, as shown in table 1.
Part threat information data of the table 1 after data prediction
Step 2 is as shown in Figure 2 using the bayesian network structure constraint of Acquirement of field knowledge.Target threat node is No. 1
Node, the entering angle sum and target velocity node of target range, target and unmanned plane are respectively 2,3, No. 4 nodes.
BIC scorings after the improvement that structural constraint and formula (1) in step 3 combination step 2 are provided, using K2 algorithms
Learning network topological structure.
Step a：But searched for since one comprising 4 node variables the figure without side.
Step b：According to node sequence ρ (X1, X2, X3, X4) scan for, X1Father node collection be empty set, since i=2, lead to
Cross calculating Vnew=BIC ((Xi, π (Xi)) | D), finding out makes Vnew=BIC ((Xi, π (Xi)) | D) reach maximum father node collection π
(Xi), scoring here is the improvement BIC scorings in step 2, and D is sample data.
Step c：Judge whether i is equal to 4, if equal, algorithm terminates, return to gained network structure.Otherwise, i is performed
=i+1 return to step b.
Fig. 3 gives the intermediate result of structure learning algorithm, and the network structure finally given is as shown in Figure 4.Fig. 4 (a) gives
The threat assessment BN structures obtained using this paper algorithms are gone out.Fig. 4 (b) gives the threat assessment BN knots that classical K2 algorithms are obtained
Structure.Fig. 5 gives the threat assessment BN structures obtained by classics K2 algorithms when sample data is 5000.Found by comparative analysis,
Compared with classical K2 algorithms, threat assessment BN structure of this paper algorithms obtained by when data volume is 20 and data volume when that is 5000 institute
Structure more closely, explanation this paper algorithms can preferably adapt to small data set, resulted under the conditions of small data set
Higher Structure learning precision.
The threat assessment BN structures monotonicity constraint getparms that step 4 is obtained with reference to step 3, such as formula (8)~
(10) shown in.
Step 5 obtains the span of parameter, tool using formula (2) and with reference to the parameter monotonicity calculating that step 4 is provided
Shown in body such as formula (11).
Step 6 solves the super ginseng of parameter beta priori using the parameter interval obtained in formula (3)~(6) and step 5
Number, it is specific as shown in formula (12)~(13).
Step 7 calculates threat assessment BN ginsengs using the hyper parameter of the parameter beta priori obtained in formula (7) and step 6
Number, concrete outcome is as shown in table 2~4.Table 2 is the parameter obtained by maximum- likelihood estimation when data volume is 5000, and table 3 is several
According to amount be 20 when maximum- likelihood estimation parameters obtained, table 4 be data volume be 20 when this paper algorithm parameters obtaineds.
Maximal possibility estimation parameters obtained when the data volume of table 2 is 5000
Maximal possibility estimation parameters obtained when the data volume of table 3 is 20
This paper algorithm parameters obtaineds when the data volume of table 4 is 20
Found by comparative analysis, when data volume is 20, compared with maximum- likelihood estimation, obtained by this paper algorithms
Parameter closer to data volume be 5000 when obtained by parameter, illustrate that the BN that this paper algorithms are more suitable under the conditions of small data set joins
Number problem concerning study, with higher study precision.
Claims (1)
1. a kind of threat assessment modeling method based on small data set and Bayesian network, it is characterised in that comprise the steps：
Step 1, the threat data in battle field information is pre-processed, including target velocity V, entering angle sum A, enemy and we's distance
Dis and targets'threat T, wherein, target velocityVuFor unmanned plane speed；Into
Angle sumFor target velocity direction and score angle and unmanned plane velocity attitude and score angle it
With；Enemy and we's distanceFor the distance between target and unmanned plane；Targets'threat
Step 2, threat data is constrained using the bayesian network structure of Acquirement of field knowledge threat data, changed as node
Enter BIC scoringsWherein, mijkIt is
X is met in sample datai=k, its father node π (Xi)=j number of samples,M is sample data total amount, e
For the presence probability of corresponding sides in structural constraint, δ is regulation parameter, and δ takes 10~100, i to represent i-th of node variable in network,
N is node variable number, qiFor the father node value status number of i-th of node, riFor the child node value status number of i-th of node；
Step 3, using K2 Algorithm Learning threat assessment network topology structures, step is as follows：
Step a, the but search the figure without side of n node variable is included since one；
Step b, according to node sequence ρ (X1,X2,...Xi...,Xn) scan for, in nodes Xi, nodes X will be comeiChange before
Amount is combined, and therefrom selects nodes XiFather node collection π (Xi), find out the father node collection for making improvement BIC scorings reach maximum
π(Xi)；
Step c, judges whether i is equal to n, if equal, algorithm terminates, output gained network structure；Otherwise, i values add in the lump
Return to step b；
Step 4, the network topology structure obtained using step 3 obtains the monotonicity constraint of network parameter, for by n variable X
={ X1,X2,...XnComposition Bayesian network G, if the nodes X in GiHave r kinds value and value state set for 1,2 ...,
K ... r }, his father set of node π (Xi) to have q value and value state set be { 1,2 ..., j .q. }, makes θijk=p (Xi=k
|π(Xi)=j), then there is θij1≤θij2≤......≤θijrOr θij1≥θij2≥......≥θijr；
Step 5, according to the normalization of parameter, i.e., father node value state is identical and sum that one group of parameter that child node value is different is added
For 1, interval getparms
Step 6, it is believed that parameter is obeyed in the interval that step 5 is provided and is uniformly distributed, setting parameter θ obeys U (θ1,θ2), [θ1,
θ2] it is parameter value scope, with B (α1, α2) equivalent U (θ1,θ2), i.e.,Get parms
The hyper parameter of beta prior distributionWithWherein, it is equally distributed to expectEqually distributed second moment
Step 7, the result of parameter learning is obtained using Bayesian EstimationN in formulaijkTo meet father node
Value is sample number when j, child node value are k, NijSample number when for father node value being j, [αi,αj] be distributed for beta
Hyper parameter.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710299124.8A CN107194155A (en) | 2017-05-02 | 2017-05-02 | A kind of threat assessment modeling method based on small data set and Bayesian network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710299124.8A CN107194155A (en) | 2017-05-02 | 2017-05-02 | A kind of threat assessment modeling method based on small data set and Bayesian network |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN107194155A true CN107194155A (en) | 2017-09-22 |
Family
ID=59873680
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201710299124.8A Pending CN107194155A (en) | 2017-05-02 | 2017-05-02 | A kind of threat assessment modeling method based on small data set and Bayesian network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN107194155A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446858A (en) * | 2018-03-27 | 2018-08-24 | 中国航空无线电电子研究所 | Vacant lot intimidation estimating method based on particular network structure |
| CN110348708A (en) * | 2019-06-26 | 2019-10-18 | 北京理工大学 | A kind of ground target dynamic threats appraisal procedure based on extreme learning machine |
| CN112163680A (en) * | 2020-09-28 | 2021-01-01 | 湘潭大学 | Cognitive computation-based wind power fault operation and maintenance management method |
Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103488886A (en) * | 2013-09-13 | 2014-01-01 | 清华大学 | State threat assessment method based on fuzzy dynamic Bayesian network |
| CN104217251A (en) * | 2014-08-12 | 2014-12-17 | 西北工业大学 | Equipment failure Bayesian network prediction method based on K2 algorithm |
-
2017
- 2017-05-02 CN CN201710299124.8A patent/CN107194155A/en active Pending
Patent Citations (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN103488886A (en) * | 2013-09-13 | 2014-01-01 | 清华大学 | State threat assessment method based on fuzzy dynamic Bayesian network |
| CN104217251A (en) * | 2014-08-12 | 2014-12-17 | 西北工业大学 | Equipment failure Bayesian network prediction method based on K2 algorithm |
Non-Patent Citations (3)
| Title |
|---|
| 邸若海等: "基于单调性约束的离散贝叶斯网络参数学习", 《系统工程与电子技术》 * |
| 邸若海等: "基于改进BIC评分的贝叶斯网络结构学习", 《系统工程与电子技术》 * |
| 邸若海等: "小数据集BN建模方法及其在威胁评估中的应用", 《电子学报》 * |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN108446858A (en) * | 2018-03-27 | 2018-08-24 | 中国航空无线电电子研究所 | Vacant lot intimidation estimating method based on particular network structure |
| CN108446858B (en) * | 2018-03-27 | 2022-03-08 | 中国航空无线电电子研究所 | Air-to-ground threat assessment method based on specific network structure |
| CN110348708A (en) * | 2019-06-26 | 2019-10-18 | 北京理工大学 | A kind of ground target dynamic threats appraisal procedure based on extreme learning machine |
| CN112163680A (en) * | 2020-09-28 | 2021-01-01 | 湘潭大学 | Cognitive computation-based wind power fault operation and maintenance management method |
| CN112163680B (en) * | 2020-09-28 | 2024-03-08 | 湘潭大学 | Wind power failure operation and maintenance management method based on cognitive calculation |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN107798199B (en) | Hydroelectric generating set parameter closed-loop identification method | |
| CN106873628A (en) | A kind of multiple no-manned plane tracks the collaboration paths planning method of many maneuvering targets | |
| CN112131786A (en) | Target detection and distribution method and device based on multi-agent reinforcement learning | |
| CN111008449A (en) | Acceleration method for deep reinforcement learning deduction decision training in battlefield simulation environment | |
| CN107194155A (en) | A kind of threat assessment modeling method based on small data set and Bayesian network | |
| CN111626500A (en) | Path planning method based on improved artificial bee colony algorithm | |
| CN106780058A (en) | The group dividing method and device of dynamic network | |
| Özsoy et al. | Estimating the parameters of nonlinear regression models through particle swarm optimization | |
| CN109976189A (en) | A kind of intelligence naval vessels automatic cruising analog simulation method | |
| Su et al. | Robot path planning based on random coding particle swarm optimization | |
| Chen et al. | Particle swarm optimization based on genetic operators for sensor-weapon-target assignment | |
| CN110110475B (en) | Extended Kalman filtering method based on online learning fading factors | |
| CN105487546A (en) | Deep space probe autonomous mission planning time constraint geometric processing method | |
| Xu et al. | Artificial moment method for swarm robot formation control | |
| Panda et al. | Model reduction of linear systems by conventional and evolutionary techniques | |
| Yu et al. | Unmanned aerial vehicle path planning based on TLBO algorithm | |
| Xu et al. | Algorithms and applications of intelligent swarm cooperative control: A comprehensive survey | |
| CN116680509A (en) | Dynamic matching method for multi-spacecraft escape-tracking game task | |
| CN116738334A (en) | Air multi-target threat assessment method based on DBN and TOPSIS method | |
| CN106951957A (en) | Particle swarm optimization algorithm, multicomputer method for parallel processing and system | |
| Wang et al. | An improved NSGA-II algorithm for UAV path planning problems | |
| CN116400726A (en) | Rotor unmanned aerial vehicle escape method and system based on reinforcement learning | |
| Guoqiang et al. | Study of RBF neural network based on PSO algorithm in nonlinear system identification | |
| CN115951695A (en) | Dynamic tactical control domain resolving method based on three-party game in air combat simulation environment | |
| CN108763604A (en) | A kind of radial base neural net point collocation solving the response of the statics of composite structure containing interval parameter |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| WD01 | Invention patent application deemed withdrawn after publication | ||
| WD01 | Invention patent application deemed withdrawn after publication |
Application publication date: 20170922 |