A target intention recognition method based on dynamic Bayesian network

Abstract

The present invention discloses a target intention recognition method based on a dynamic Bayesian network, and belongs to the technical field of target intention recognition. The present invention uses raw data and a scoring search algorithm to construct a dynamic Bayesian network, uses the Bayesian scoring criterion BIC and an adaptive genetic algorithm in the scoring process, and uses the idea of ​​ensemble learning in the feedback strategy to complete the correction of the edge direction. The present invention can process the time series information and uncertain information obtained from complex situations, and at the same time introduces a feedback strategy to solve the reverse edge problem that affects the recognition accuracy in the process of using raw data to build a dynamic Bayesian network. The present invention can be used for the intention recognition processing of aerial targets, and the corresponding raw data is the flight status data of the target.

Description

Technical Field

[0001] The present invention belongs to the technical field of target intention recognition, and in particular relates to a target intention recognition method based on a dynamic Bayesian network. Background Art

[0002] Bayesian network is a probabilistic graph model. Its model structure can be used to express the dependency relationship between random variables. The joint probability distribution between variables is represented by a directed acyclic graph (DAG) structure, where the nodes in the graph represent each random variable and define the conditional probability distribution for the variable of a given parent node in the graph. The main problem solved by the structural learning of Bayesian network is to find the probability dependence of each node on its parent node set and obtain the corresponding conditional probability distribution function. Existing research has been able to prove that it is NP-hard to obtain the structure of such a network from data, and there are still many problems to be solved. By introducing the temporal relationship between nodes on the basis of Bayesian network, a dynamic Bayesian network can be constructed. Combining the structure of the network with time information enables Bayesian network to have the ability to process time series data, which is more suitable for various scenarios in life. Summary of the invention

[0003] The present invention provides a target intention recognition method based on a dynamic Bayesian network, which can be used to improve the processing speed and accuracy of target intention recognition.

[0004] The technical solution adopted by the present invention is:

[0005] A target intention recognition method based on a dynamic Bayesian network, the method comprising the following steps:

[0006] Step 1: Collect the target's attribute data to determine the target's true intention attribute, wherein the target's attribute data includes but is not limited to: the target's movement status information, location information and device information;

[0007] Step 2: Discretize the numerical data in the attribute data of the acquisition target;

[0008] Step 3: sampling the collected attribute data of the target, obtaining input data matching the input of the dynamic Bayesian network, and obtaining a training data set based on the multiple input data;

[0009] Step 4: Based on the set learning strategy, the dynamic Bayesian network is trained to learn the structure diagram between the attribute data of the training target and the intent attribute to be identified, so as to obtain a trained dynamic Bayesian network;

[0010] Among them, the training and learning of the dynamic Bayesian network based on the set learning strategy is specifically as follows:

[0011] 1) Divide the training data set into multiple data groups, each data group is defined as D n , n represents the number of the data set group;

[0012] 2) Construct the initial prior network B0 and the initial transfer network B based on the mutual information between nodes and the temporal mutual information → , and the initial prior network B0 and the initial transfer network B → Encoding;

[0013] 3) Based on each data group D n , through the adaptive genetic algorithm to the current prior network B k,n and transfer network B k→n Perform structural learning, where the subscript k represents the number of learning times;

[0014] In the adaptive genetic algorithm, the crossover probability P of each generation is c With mutation probability P m The update formula is set to:

[0015]

[0016]

[0017] Among them, P c1 , P c2 represents the crossover probability of each individual of the two individuals undergoing crossover, P m1 , P m2 represents the mutation probability of each of the two individuals undergoing mutation, f max Indicates the maximum fitness value in the population; f avg represents the average fitness value of all populations in each generation; f′ represents the larger fitness value of the two individuals that need to cross or mutate;

[0018] 4) Calculate the current prior network B k,n and transfer network B k→n The BIC score (the score of the Bayesian scoring criterion) is denoted as score(B k,n +B k→n );

[0019] Traverse the score of all data groups (B k,n +B k→n ), the score of each data group (B k,n +B k→n ) all satisfy score(B k,n +B k→n ,D)<score0, then execute step 5); otherwise, continue to execute step 3);

[0020] 5) Traverse the scores of all groups obtained in the last two times (B k,n +B k→n )、score(B k-1,n +B k-1→n ), for any group of scores with the same data group number, if they all meet score(B k,n +B k→n )<score(B k-1,n +B k-1→n ), then the most recently obtained prior network B k,n and transfer network B k→n Get a trained dynamic Bayesian network; otherwise, continue to execute step 2) based on the scores of the edges in the current network structure;

[0021] Step 5: Sample the attribute data to be identified, and then input the sampled data into the dynamic Bayesian network trained in step 4, and obtain the target intention recognition result based on its output.

[0022] Furthermore, in step 4, the initial prior network B0 and the initial transfer network B → The encoding is as follows:

[0023] The encoding rule for the prior network B0 is: define the node flag as 1, arrange the nodes in ascending order according to the node number, and for the same node number, the node at time t is in front and the node at time t-1 is behind. If the flag of the node is the parent node of the current node, the code of the current node is 1, otherwise it is 0;

[0024] Transfer network B → The encoding rule is: define the node flag as 1, arrange the nodes in ascending order according to the node number, and for the same node number, the node at time t is in front and the node at time t-1 is in the back. If the flag of the node is the current parent node, the code of the current node is 1, otherwise it is 0.

[0025] Further, in step 4, based on each data group D n , through the adaptive genetic algorithm to the current prior network B k,n and transfer network B k→n When performing structural learning, the crossover processing includes: if the current position to be crossed appears in the prior network, the set crossover strategy is executed in the non-temporal edge coding; if the current position to be crossed appears in the transfer network, the set crossover strategy is executed in the temporal edge coding.

[0026] Furthermore, in step 4, the current prior network B is trained by an adaptive genetic algorithm. k,n and transfer network B k→nStructural learning includes:

[0027] The BIC score of the network is used as the population fitness to calculate the population fitness of the current generation;

[0028] (1) Calculate the fitness of the current generation of population;

[0029] (2) retain the individuals with the highest fitness and the individuals that have been updated most recently;

[0030] (3) Select parent individuals from the retained individuals to evolve the single-point crossover strategy and calculate the crossover probability P c ;

[0031] Among them, the parent individuals are the individuals corresponding to the two largest fitness values ​​among the retained individuals, and the individuals selected each time are not repeated with the historically selected individuals;

[0032] (4) Calculate the mutation probability P m , and obtain variants;

[0033] (5) If the number of new individuals is greater than the specified value, execute step (6); otherwise, return to step (3); wherein the new individuals include the new individuals obtained by crossover and mutation;

[0034] (6) If the maximum fitness exceeds the specified number of iterations θ e If there is no numerical change, the two-point crossover strategy is executed and step (7) is continued; otherwise, step (7) is directly executed;

[0035] (7) Based on the number of new individuals K, select the first K individuals with the smallest fitness and replace them with the new individuals, and continue to execute step (8);

[0036] (8) Determine whether the end condition is met. If so, end the learning. Otherwise, repeat steps (1) to (7).

[0037] The termination condition is: the maximum fitness value converges, or the number of learning times reaches the preset maximum number of iterations.

[0038] Furthermore, in step 4, the score of the edge in the network structure is specifically:

[0039]

[0040] Among them, W represents the score of the edge in the network structure, N′ represents the number of groups in the data set, and N kn M represents the number of non-sequential edges in the network structure obtained after the kth training of the nth group of training data. knIt represents the matrix representation of the network structure obtained after the kth training of the nth group of training data, D represents the total sample size of the training data set, and ⊙ is the Hadamard product.

[0041] Furthermore, the target is an aerial target.

[0042] The technical solution provided by the present invention brings at least the following beneficial effects:

[0043] The present invention uses raw data and a scoring search algorithm to construct a dynamic Bayesian network. In the scoring process, the Bayesian Information Criterion (BIC) and an adaptive genetic algorithm are used. In the feedback strategy, the idea of ​​ensemble learning is used to complete the correction of the edge direction. The present invention can process the time series information and uncertain information obtained from complex situations. At the same time, the feedback strategy is introduced to solve the reverse edge problem that affects the recognition accuracy in the process of using raw data to construct a dynamic Bayesian network. The present invention can be used for the intention recognition processing of aerial targets, and the corresponding raw data is the flight status data of the target. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0045] Figure 1 is a schematic diagram of executing decomposition of a candidate parent node in an embodiment of the present invention, Figure 1 In the figure, (a) represents the case of selecting edges from all candidate edges, (b) represents the case of selecting edges only from sequential edges, and (c) represents the case of selecting edges only from non-sequential edges;

[0046] Figure 2 is a schematic diagram of a node transfer network in an embodiment of the present invention;

[0047] Figure 3 1 is a schematic diagram of a crossover operator execution flow of the IAGAFB-DBN algorithm used in an embodiment of the present invention;

[0048] Figure 4 Schematic diagram of two mutation operators used by the IAGAFB-DBN algorithm in an embodiment of the present invention. Figure 4 In the figure, (a) indicates a single-point mutation, and (b) indicates a multi-point mutation;

[0049] Figure 5 It is a schematic diagram of the processing process of the adaptive genetic algorithm in an embodiment of the present invention.

[0050] Figure 6 It is a schematic diagram of the processing process of the IAGAFB-DBN algorithm in an embodiment of the present invention. DETAILED DESCRIPTION

[0051] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0052] In view of the uncertainty of the behavioral characteristics and physical characteristics of the target, the uncertainty of the flight rules and the uncertainty of the action ability that appear in the process of aerial target intention recognition, an embodiment of the present invention provides a target intention recognition method based on a dynamic Bayesian network, which is used to process the time series information and uncertain information obtained from complex situations. At the same time, the introduction of a feedback strategy solves the reverse edge problem that affects the recognition accuracy in the process of using raw data to build a dynamic Bayesian network. The embodiment of the present invention uses raw data and a scoring search algorithm to build a dynamic Bayesian network, uses the Bayesian Information Criterion (BIC) and an adaptive genetic algorithm in the scoring process, and uses the idea of ​​ensemble learning in the feedback strategy to complete the correction of the edge direction.

[0053] The embodiment of the present invention is mainly based on a two-step strategy and an adaptive genetic algorithm, and proposes an Improved Adaptive Genetic Algorithm with Feedback-Dynamic Bayesian Network (IAGAFB-DBN) structure learning algorithm based on an adaptive genetic algorithm with a feedback strategy.

[0054] The structural learning of dynamic Bayesian networks can be transformed into the learning of prior networks and transfer networks under the conditions of stationarity and Markov properties. Under the condition of a complete data set, the scoring function of the Bayesian network can be decomposed into independent factors about the variable and all its parent nodes. If the parent-child structural relationship of a node changes, it will not affect the scoring results of other nodes. Therefore, when calculating the scoring function, it is only necessary to calculate the sufficient statistical factors of the local structure related to each node. Based on this idea, a two-step strategy can be implemented for the two characteristics of the temporal edges and non-temporal edges of the transfer network. Figure 1 The decomposition and execution strategies are presented. Figure 1 The situation shown in (a) is the existing strategy. Figure 1(b) and (c) show the decomposition and execution strategies adopted by the embodiments of the present invention. After decomposing the edges into sequential edges and non-sequential edges, the search space can be reduced and the time complexity of the algorithm operation can be reduced. On this basis, an adaptive genetic algorithm and a BIC scoring function are used to learn the optimal dynamic network structure model.

[0055] In the embodiment of the present invention, the main steps of the algorithm for learning the network structure are:

[0056] 1. Use mutual information and temporal mutual information to construct the initial prior network and the initial transfer network.

[0057] Under the condition of satisfying the Markov assumption, the state of the node at time t is only related to the state of the node at time t-1.

[0058] Let Y ij (t-1) = {X k (t-1),k≠i,j},X i (t-1) and X j (t) in Y ij The conditional mutual information formula under (t-1) condition is:

[0059]

[0060] Among them, Y ij (t-1) represents the condition set at time t-1. The subscripts i and j represent the node numbers of two different nodes, corresponding to the two nodes currently calculating the mutual information. k (t-1) represents the node state of node k at time t-1. Similarly, X i (t-1) represents the node state of node i at time t-1, X j (t) represents the node state of node j at time t, H(X) is the entropy of the discrete random variable X, and the threshold ε>0 is set. If there is It means that node i and node j are independent of each other. Based on this, the initial parent node set of each node is obtained to construct the initial transfer network population of the genetic algorithm for iterative evolution.

[0061] The initial prior network can be constructed using mutual information. The mutual information between variables X and Y is defined as:

[0062]

[0063] Among them, p XY (x,y) represents the joint probability density of variables X and Y, p X (x),p Y(y) represent the marginal probability density functions of variables X and Y respectively. The greater the mutual information, the greater the correlation between variables X and Y. When using this mutual information representation method to simulate the data obtained in the real environment, the joint density estimation will be difficult and the accuracy will be low. Copula entropy can be used to calculate the correlation between variables X and Y, which can avoid the estimation of the joint probability density function and reduce the time complexity when constructing the initialization network. The mutual information calculation method based on Copula entropy can be written as

[0064]

[0065] Where N represents the number of nodes and ω(x,y) represents the conditional probability function

[0066]

[0067] Set the threshold ε>0. If the mutual information value between variables X and Y exceeds the threshold, it is considered that there is an edge between node X and node Y. However, the direction of the edge cannot be determined at this time. The direction of the edge can be added using the standard that makes the original network contain no directed cycles.

[0068] 2. BIC score analysis.

[0069] BIC is a method for scoring and selecting models, which is applicable to models fitted under the maximum likelihood estimation framework. The BIC scoring formula is:

[0070]

[0071] Among them, the BIC score F BIC The first term of (G,D) is the likelihood function term, m ijk Indicates that the i-th X variable (node) has the j-th value X in the value combination of its parent node set. i = the number of instances of k, q i Indicates the number of values ​​of the parent node combination of the X variable, r i represents the number of values ​​of the X variable, m represents the total number of samples, N represents the number of attributes (i.e. the total number of nodes), G represents the network to be scored, and D represents the total sample size of the data set. During model training, the increase in the number of parameters will lead to an increase in the likelihood function, making the model more suitable for the training data. However, too many parameters will also increase the complexity of the network model, making the learned network structure almost converge to a complete graph. Therefore, the second term is needed as a penalty term for the model parameters to prevent overfitting caused by too many parameters in the model.

[0072] 3. Prior net and transfer net coding.

[0073] In order to make the network better adapt to the iteration, crossover and mutation of the genetic algorithm, the network structure needs to be encoded. The encoding stage is divided into the encoding of the prior network and the encoding of the transfer network. Therefore, the encoding of the corresponding network can be written as C = C0 + C → , the specific encoding rules are as follows Figure 2 shown.

[0074] Coding of C0 (a priori network coding): In the coding of the a priori network, only the relationship between nodes needs to be considered. Take the coding of the nodes in time slice t as an example. The coding rules are: the node flag is 1; the node with smaller serial number is in front, the node with larger serial number is in the back, the node at time t is in front, and the node at time t-1 is in the back; the node flag is 1 if it is the parent node of the current node, otherwise it is 0. Figure 2 The C0 encoding is 0000|1000|0100|1000.

[0075] C → Coding (transfer network coding): All nodes at time t-1 have no parent node, so only the coding of nodes at time slice t is considered. Set the coding rules: the node flag is 1; the node with smaller serial number is in front, and the node with larger serial number is in the back; the node at time t is in front, and the node at time t-1 is in the back; the node flag is 1 if it is the current parent node, otherwise it is 0. Figure 2 C → The encoding is: 1000|1100|0010|0101.

[0076] 4. Adaptive genetic algorithm.

[0077] Two important parameters in genetic algorithms are crossover probability (P c ) and mutation probability (P m ), where P c Determines the abundance of the population, P m Determines whether the genetic algorithm can jump out of the local optimal value. How to determine these two parameters has become the main research direction of many researchers using genetic algorithms. c It will destroy the good genes, and too large P m This will cause the genetic algorithm to degenerate into a random search algorithm.

[0078] M.Srinivas proposed that when the fitness of the population is relatively concentrated, the crossover probability should be increased; when the fitness of the population is relatively dispersed, the crossover probability should be reduced. Whether the fitness of the population is dispersed or not is measured by the maximum, minimum and average fitness. However, when the individual fitness is close to or equal to the maximum fitness, the crossover rate and mutation rate are close to or equal to zero, which is not conducive to the evolution stage in the genetic algorithm. It will make the elite individuals in the population in the early stage of evolution basically unchanged, thus causing the entire algorithm to fall into a local optimal solution. Based on the above theory, the existing crossover probability (Pc ) and mutation probability (P m ) is calculated as.

[0079]

[0080]

[0081] Among them, f max is the maximum fitness value in the population; f avg is the average fitness value of all populations in each generation; f′ is the larger fitness value of the two individuals that need to cross; P c1 , P m1 represents the crossover probability and mutation probability of one of the two individuals, P c2 , P m2 represents the crossover probability and mutation probability of the other of the two individuals, and f′ is the fitness value of the individual to be mutated. By dynamically changing the crossover probability and mutation probability, it can be ensured that when the individual fitness of the population reaches the maximum, the crossover rate and mutation rate will not become 0, and its best individual still has the ability to evolve probabilistically, ensuring that the algorithm can jump out of the local optimal solution.

[0082] It can be seen from the formula that when the individual fitness is lower than the average fitness, the mutation probability and the inheritance probability are constant values. c1 ,P m1 A larger value will lead to the elimination of high-quality genes carried by individuals with low fitness, increasing the possibility of the algorithm falling into a local optimal solution. Based on this, the embodiment of the present invention adjusts P c and P m The update strategy is improved, and the specific update formula is:

[0083]

[0084]

[0085] use As a judgment, it can serve the purpose of nonlinear adaptation. If f avg Increase, The increase is faster than linear growth, so it can more clearly describe the concentration or dispersion of population fitness.

[0086] The elite retention strategy is used in the iterative process of the genetic algorithm. If the highest fitness value in the new generation population after the genetic operation is less than the highest fitness value of the population before the genetic operation, the individual with the highest fitness value in the previous generation population will be added to the new generation. The elite retention strategy can avoid the sudden increase in the calculated crossover probability during actual use, which will destroy the high-quality genes, and also ensure the convergence ability of the algorithm.

[0087] 5. Calculate the crossover operator and mutation operator.

[0088] The IAGAFB-DBN algorithm uses a league selection mechanism, which selects better individuals from the current iteration population as parents to perform crossover and mutation operations. In addition, a two-point crossover operator is used to ensure the diversity of the population. The specific implementation steps are as follows: Figure 3 shown.

[0089] There are two types of mutation operators. In the early stage of evolution, when the fitness of individuals in the population is relatively dispersed, a single-point mutation operator is used to randomly select a site in the mutated individual for inversion operation. As the iterative evolution proceeds, the fitness of the group is relatively concentrated, and the population is almost full of elite individuals. A multi-point mutation operator is used, and the mutated individual will randomly assign 0 or 1 to all sites, so that the best individual can still evolve, ensuring that the algorithm can jump out of the local optimal solution. The specific two mutation operations are as follows: Figure 4 As shown, Figure 4 In the figure, (a) represents a single-point mutation and (b) represents multiple-point mutation.

[0090] The IAGAFB-DBN algorithm adds nonlinear mutation probability and crossover probability adaptive functions to the genetic algorithm, which can more accurately describe the changes in crossover probability and mutation probability when the fitness of the population is dispersed and concentrated. At the same time, it avoids the problem that the crossover rate or mutation rate approaches zero when the individual fitness is greater than or equal to or close to the maximum fitness of the group during the evolution using the linear adaptive function. This ensures the evolution process of the algorithm. IAGAFB-DBN uses mutual information based on Copula entropy when constructing the initial network, which avoids the problem of difficult calculation of edge joint probability in some practical problems and speeds up the construction of the initial network.

[0091] The genetic algorithm process used in the IAGAFB-BDN algorithm is as follows Figure 5 shown.

[0092] (1) Input data set group D i , and the input population size θ pn , specify the number of iterations θ e (also called retention parameter), the maximum number of iterations θ max , mutation probability P m , and the crossover probability P c ;

[0093] Preferably, the parameter θ e The value range can be set to

[0094] (2) Construct the initial prior network B0 based on the mutual information of Copula, and construct the initial transfer network B based on the temporal mutual information → , and encode;

[0095] (3) Calculate the population fitness F BIC (G,D i );

[0096] (4) Elite retention strategy retains the best individual: α best_fitness andα elite , where α best_fitness Represents the fitness F BIC (G,D i )The largest individual, α elite Indicates the most recently updated individual;

[0097] (5) Select the best parent individuals to evolve the single-point crossover strategy and calculate the crossover probability P c ;

[0098] Among them, the parent individuals are the individuals corresponding to the two current largest fitnesses, and the objects selected each time are not repeated with the objects that have been selected;

[0099] (6) Update P m Get the variant;

[0100] That is, based on the mutation probability P m Determine whether a mutation has occurred, and then mutate one or some positions of the code to generate a new individual.

[0101] (7) If the number of new individuals (new individuals obtained by crossover and mutation) is greater than M (specified value), execute step (8); otherwise, return to step (5);

[0102] (8) Determine whether the process has entered the final stage of the iteration. If so, execute step (9); otherwise, execute step (10);

[0103] In this step, when the best fitness exceeds the specified number of iterations θ e If there is no change, it is considered to have entered the final stage of the iteration.

[0104] (9) Execute the two-point crossover strategy. After completion, continue to step (10);

[0105] (10) Replace the poor individuals (i.e., the first M individuals with the smallest fitness) with the new individuals. After the replacement is completed, execute step (11);

[0106] (11) Determine whether to end the iteration: α best_fitness Convergence or the number of iterations reaches θ max If yes, then end; otherwise, return to step (3).

[0107] 6. Feedback strategy.

[0108] When constructing the initial network, the mutual information can only determine the existence of an edge between two nodes but cannot indicate the direction of the edge (the direction of the edge has a clear meaning in some practical problems). At the same time, when using the BIC scoring function for genetic algorithm search, the direction of the edge is not judged. Therefore, a feedback strategy is needed to correct the direction of the edge in order to improve the fit between the learned network structure and the data.

[0109] In the embodiment of the present invention, the specific feedback strategy is: the data sets generated in the simulation scenario are grouped according to the time series, N′ is defined to represent the number of data groups, and each set of training data sets is subjected to structural learning using an adaptive genetic algorithm and a BIC score search to obtain a new network structure, and based on the score W of the edge in the obtained network structure, the reliability of the edge obtained in the structural learning is indicated, wherein the calculation formula of W is:

[0110]

[0111] Among them, N′ represents the number of groups in the data set, N ki represents the number of non-temporal edges in the network learned by the proposed IAGAFB-BDN algorithm for the i-th group of training data, M ki It represents the matrix representation of the network structure learned by the proposed IAGAFB-DBN algorithm for the i-th group of training data, D represents the total sample size of the data set, and ⊙ is the Hadamard product.

[0112] Set the threshold θ to calculate the obtained W, for the matrix W obtained by the k-th calculation k (k represents the number of iterations), any element W k If [q,p] is less than θ, then let the element W k [q,p]=0; otherwise, W k [q,p]=1. Finally, the normalized W k The network diagram is constructed by matrix, and finally a new Bayesian network is obtained, denoted as BN k .

[0113] Calculate the score of the Bayesian network obtained by grouping the data set (BN k ,D) and the score score0 of the initial network obtained by mutual information. If there is score(BN k ,D)>score0, then the initially constructed network is updated. The specific update strategy is: first, transpose the obtained W matrix to obtain W T , determine the initial network B0 and WT For every element in, if B0[q,p]=W T [qp], then let B0[p,q]=1,B0[q,p]=0, complete the update of the initial network B0, and iterate until score(BN k ,D)<score0. The final network is obtained.

[0114] like Figure 6 As shown in Figure 1, the specific process of the IAGAFB-DBN strategy includes: inputting each group of the data set, and performing the following processing on each group:

[0115] (1) Based on the input data set grouping, the initial network is obtained, including the prior network B0 and the transfer network B → ;

[0116] (2) Based on the IAGAFB-DBN strategy (also known as BIC+adaptive genetic algorithm), the structure learning is performed to obtain the network of the nth learning (iteration): the prior network B k,n and transfer network B k→n , where the subscript k represents the number of learning times (iterations), and n represents the number of data set groups, n = 1, 2, ..., N';

[0117] (3) Determine whether score(B k,n +B k→n ,D)<score0, if not, continue to step (2); if yes, then, execute step (4);

[0118] (4) Determine whether score(B k,n +B k→n ,D)<score(B k-1,n +B k-1→n , D), if yes, then execute step (5), otherwise, according to the currently calculated W, the initial network (B0, B → ) for network updates.

[0119] The improved Bayesian network structure learning algorithm solves the problem of the frequent occurrence of reverse edges due to the inability to judge the direction of edges during the structure learning process. On the basis of the original algorithm, the strategy of initial network construction is improved to make the initial network construction faster. At the same time, after the structure learning is completed, a feedback strategy is added to complete the correction and update of the edges by calculating the edge score function W, thereby obtaining the optimal network structure.

[0120] The process of using the target intention recognition method based on dynamic Bayesian network provided by the embodiment of the present invention for processing aerial targets specifically includes:

[0121] Step 1: Simulate various attribute data of aerial targets obtained from sensors, including longitude, latitude, distance from the observation point, flight height, flight speed, radar model (RM), radar status (RS), communication system status (CSS), azimuth during flight, heading angle during flight, and whether there is acceleration during flight. Among them, longitude, latitude, distance from the observation point, flight height, flight speed, azimuth during flight, and heading angle during flight are numerical types; radar model, radar status, communication system status, and whether there is acceleration are enumeration types. The specific data descriptions are shown in Tables 1 and 2.

[0122] Table 1 A set of numerical data

[0123] property unit Numeric Longitude ° 102°54′ Latitude ° 30°05′ Distance km 310.0 Height km 15.8 Velocity m / s 220.0 Azimuth mil 2230.0 Heading Angle ° 12.0

[0124] Table 2 A set of non-numeric data

[0125]

[0126] Step 2: Discretize the numerical data.

[0127] A discretization method is: Height (0: low altitude; 1: medium altitude; 2: high altitude), Velocity (0: low speed; 1: medium speed; 2: high speed), Radar Model (0: early warning control radar; 1: terrain following radar; 2: fire control radar; 3: other), RadarStatus (0: shutdown; 1: guidance; 2: tracking; 3: search; 4: unknown), Communication System Status (0: off; 1: on), Distance (0: near; 1: medium; 2: far), Acceleration (0: uniform speed; 1: acceleration). In addition, the maneuvering mode (Maneuver) of the aerial target can be simulated through the azimuth, heading angle and longitude and latitude coordinates of the aerial target during flight, which can be discretized into (0: level flight; 1: turn; 2: dive; 3: ascent; 4: S-type; 5: O-type; 6: left detour; 7: right detour; 8: somersault; 9: ascending turn). After discretization, the data attributes have a total of 9 dimensions, including 8-dimensional aerial target attributes (the last two items in Table 1 are not used) and 1-dimensional intention attribute (Intention).

[0128] In the example of the present invention, Intention is set to: 0: Patrol; 1: Early warning detection / command; 2: Electronic reconnaissance; 3: Electronic interference; 4: Attack; 5: Strike. Of course, when applied to other scenarios, specific settings can be made according to actual needs, and the present invention does not limit this.

[0129] Step 3: Sample the data to ensure that it fits the input format of the model.

[0130] Step 4: Learn according to the IAGAFB-DBN algorithm and train to obtain the structural diagram between the attributes of the air target and the intention of the air target.

[0131] Step 5: Use the obtained structure diagram to complete the reasoning and obtain the intention of the aerial target.

[0132] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0133] The above are only some embodiments of the present invention. For those skilled in the art, several modifications and improvements can be made without departing from the creative concept of the present invention, which all belong to the protection scope of the present invention.

Claims

1. A target intention recognition method based on dynamic Bayesian network, It is characterized in that The following steps are involved: Step 1: Collect the target's attribute data to determine the target's true intention attribute, wherein the target's attribute data includes but is not limited to: the target's movement status information, location information and device information; Step 2: Discretize the numerical data in the attribute data of the acquisition target; Step 3: sampling the collected attribute data of the target, obtaining input data matching the input of the dynamic Bayesian network, and obtaining a training data set based on the multiple input data; Step 4: Based on the set learning strategy, the dynamic Bayesian network is trained to learn the structure diagram between the attribute data of the training target and the intent attribute to be identified, so as to obtain a trained dynamic Bayesian network; Among them, the training and learning of the dynamic Bayesian network based on the set learning strategy is specifically as follows: 1) Divide the training data set into multiple data groups, each data group is defined as D n , n represents the number of the data set group; 2) Construct the initial prior network B based on the mutual information between nodes and the temporal mutual information 0 and the initial transfer network B → , and the initial prior network B 0 and the initial transfer network B → Encoding; 3) Based on each data group D n , through the adaptive genetic algorithm to the current prior network B k,n and transfer network B k→n Perform structural learning, where the subscript k represents the number of learning times; In the adaptive genetic algorithm, the crossover probability P of each generation is c With mutation probability P m The update formula is set to: Among them, P c1 , P c2 represents the crossover probability of each individual of the two individuals undergoing crossover, P m1 , P m2 represents the mutation probability of each of the two individuals undergoing mutation, f max Indicates the maximum fitness value in the population; f avg represents the average fitness value of all populations in each generation; f′ represents the larger fitness value of the two individuals that need to cross or mutate; 4) Calculate the current prior network B k,n and transfer network B k→n The BIC score is denoted as score(B k,n +B k→n ); Traverse the score of all data groups (B k,n +B k→n ), the score of each data group (B k,n +B k→n ) all satisfy score(B k,n +B k→n ,D)<score 0 , then execute step 5); otherwise, continue to execute step 3); 5) Traverse the scores of all groups obtained in the last two times (B k,n +B k→n )、score(B k-1,n +B k-1→n ), for any group of scores with the same data group number, if they all meet score(B k,n +B k→n )<score(B k-1,n +B k-1→n ), then the most recently obtained prior network B k,n and transfer network B k→n Get a trained dynamic Bayesian network; otherwise, continue to execute step 2) based on the scores of the edges in the current network structure; Step 5: Sample the attribute data to be identified, and then input the sampled data into the dynamic Bayesian network trained in step 4, and obtain the target intention recognition result based on its output.

2. The method according to claim 1, It is characterized in that In step 4, the initial prior network B 0 and the initial transfer network B → The encoding is as follows: For the prior network B 0 The encoding rule is: define the node flag as 1, arrange the nodes in ascending order according to the node number, and for the same node number, the node at time t is in front and the node at time t-1 is behind. If the node flag is the parent node of the current node, the code of the current node is 1, otherwise it is 0; Transfer network B → The encoding rule is: define the node flag as 1, arrange the nodes in ascending order according to the node number, and for the same node number, the node at time t is in front and the node at time t-1 is in the back. If the flag of the node is the current parent node, the code of the current node is 1, otherwise it is 0.

3. The method according to claim 1 or 2, It is characterized in that In step 4, based on each data group D n , through the adaptive genetic algorithm to the current prior network B k,n and transfer network B k→n When performing structural learning, the crossover processing includes: if the current position to be crossed appears in the prior network, the set crossover strategy is executed in the non-temporal edge coding; if the current position to be crossed appears in the transfer network, the set crossover strategy is executed in the temporal edge coding.

4. The method according to claim 3, It is characterized in that In step 4, the current prior network B is trained using an adaptive genetic algorithm. k,n and transfer network B k→n Structural learning includes: The BIC score of the network is used as the population fitness to calculate the population fitness of the current generation; (1) Calculate the fitness of the current generation of population; (2) retain the individuals with the highest fitness and the individuals that have been updated most recently; (3) Select parent individuals from the retained individuals to evolve the single-point crossover strategy and calculate the crossover probability P c ; Among them, the parent individuals are the individuals corresponding to the two largest fitness values ​​among the retained individuals, and the individuals selected each time are not repeated with the historically selected individuals; (4) Calculate the mutation probability P m , and obtain variants; (5) If the number of new individuals is greater than the specified value, execute step (6); otherwise, return to step (3); wherein the new individuals include the new individuals obtained by crossover and mutation; (6) If the maximum fitness exceeds the specified number of iterations θ e If there is no numerical change, the two-point crossover strategy is executed and step (7) is continued; otherwise, step (7) is directly executed; (7) Based on the number of new individuals K, select the first K individuals with the smallest fitness and replace them with the new individuals, and continue to execute step (8); (8) Determine whether the end condition is met. If so, end the learning. Otherwise, repeat steps (1) to (7). The termination condition is: the maximum fitness value converges, or the number of learning times reaches the preset maximum number of iterations.

5. The method according to claim 4, characterized in that the number of iterations θ e The value range is set to: in, θ max Indicates the maximum number of iterations.

6. The method according to claim 1, characterized in that, in step 4, the score of the edge in the network structure is specifically: in, W represents the score of the edge in the network structure, N′ represents the number of groups in the data set, and N kn M represents the number of non-sequential edges in the network structure obtained after the kth training of the nth group of training data. kn It represents the matrix representation of the network structure obtained after the kth training of the nth group of training data, D represents the total sample size of the training data set, and ⊙ represents the Hadamard product.

Images