Method for realizing enemy behavior intention reasoning based on HMM

Abstract

The invention belongs to the technical field of target intention recognition, and particularly relates to a method for realizing enemy behavior intention reasoning based on an HMM (Hidden Markov Model). The method can be divided into three steps of prediction, updating and maximum posteriori estimation, specifically, a series of clutter observations are generated according to a Poisson point process model, and meanwhile, HMM observation is generated according to a certain detection probability pD in the real state of the current moment; predicting the next state, and obtaining the existence probability q of the HMM at the current moment and the probability quality function P of each state according to the prediction equation of the Bernoulli filter; next, q and P are updated according to an updating equation and a likelihood function; and finally, the state corresponding to the maximum P is the estimation state at the current moment. According to the method, intention reasoning is realized through the HMM and the single Bernoulli filter, the method is simpler and more convenient, and reasoning is more accurate.

Description

Technical Field

[0001] The invention belongs to the technical field of target intention recognition, and in particular relates to a method for realizing enemy behavior intention reasoning based on HMM. Background Art

[0002] Situation understanding is the process of interpreting the current situation based on the situation feature vector generated by situation awareness and combining the military knowledge of experts in the field to identify the enemy's intentions and combat plans. Battlefield target combat intention reasoning has always been the focus of commanders at all levels. It is a hot issue in the field of situation assessment and an important basis for commanders to decide on the next combat action. Therefore, how to accurately and efficiently identify radar working modes or behavioral intentions has important theoretical significance and military value. Previous intention reasoning methods are mostly based on traditional Bayesian networks, deep learning algorithms, etc., which are relatively complex to implement. Summary of the invention

[0003] In view of the above problems, the present invention proposes a method for realizing enemy behavior intention reasoning based on HMM.

[0004] The technical solution of the present invention is:

[0005] A method for realizing enemy behavior intention reasoning based on HMM includes the following steps:

[0006] S1. Collect the attribute data of the enemy radar, determine the real intention attribute of the target, that is, determine all possible intentions of the enemy, and discretize the data to obtain a discretized data set; analyze the attribute data of the target to obtain the transfer matrix between the states of the target; use the data set to obtain the fuzzy matrix of the state generation observation;

[0007] S2. Perform Bernoulli filtering on the data, and perform data prediction and update based on the clutter observation and the HMM observation generated by the real state at the previous moment. Thus, the most likely state at the current moment is obtained, that is, the enemy's intention (such as tracking, attack, etc.);

[0008] Among them, the specific process of Bernoulli filtering is:

[0009] The posterior probability density function of the initial moment obtained from the discretized data set obtained from S1 is:

[0010]

[0011] Then the posterior probability density function at time k-1 can be expressed as:

[0012]

[0013] Among them, S k-1represents the hidden Markov state at time k-1. There are two cases: the target does not exist. Or for a single target, S k-1 ={s k-1}, Z k-1 represents the set of all observations from time 1 to time k-1, q k-1|k-1 and P k-1|k-1 (s k-1 ) represents the posterior probability of the HMM existence and the posterior probability mass function of the HMM state.

[0014] q k-1|k-1 =P(S k-1 |=1|Z k-1 ) (3)

[0015] P k-1|k-1 (s k-1 )=P(S k-1 ={s k-1}|Z k-1 ) (4)

[0016] The Bernoulli filter propagates these two quantities over time using a prediction equation and an update equation.

[0017] Prediction equation:

[0018]

[0019] in, Represents the transition density of state S. There are two cases:

[0020]

[0021]

[0022] π(s k |s k-1 )=P(S k ={s k}S k-1 ={s k-1}) indicates the state {s k-1} to state {s k The transition probability between} is given by the transition matrix, p b represents the probability of HMM being born, that is, the probability of starting the plan at time k, p s It represents the probability of HMM survival, that is, the probability that the plan at the previous moment is not cancelled. The posterior probability density function f of each state at time k can be predicted from time k-1 k|k-1 (S k |Z k-1 ), that is, the probability q of HMM existence is obtained k|k-1And the probability distribution P of each state k|k-1 (s k ).

[0023] Update equation:

[0024]

[0025] Among them, represents the observation likelihood function, which can be obtained from the clutter observation and the fuzzy matrix, and f(Z k |Z k-1 ) is defined as follows:

[0026]

[0027] By updating the equation, the prediction results can be calibrated to obtain f k|k (S k |Z k ), the updated result is q k|k and P k|k (s k )

[0028] S3, based on the updated q k|k Make a judgment, if maxP k|k (s)>1-q, then HMM exists, and then P k|k (s k ) to perform maximum a posteriori estimation and obtain the state s with the highest probability at time k k ; If maxP k|k (s)<1-q, then HMM does not exist and it is judged as a false alarm;

[0029] S4, loop steps S2-S3. If time k ≥ threshold (i.e. current time t), end the loop and output the current state. At this point, intention reasoning is completed and all possible intentions of the enemy are obtained.

[0030] The present invention realizes intention reasoning through HMM and single Bernoulli filter, the method is simpler and the reasoning is more accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 Flowchart for reasoning about intent.

[0032] Figure 2 It is the state transition diagram of the simulation scenario.

[0033] Figure 3 This is the simulation result diagram of intention reasoning. DETAILED DESCRIPTION

[0034] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings and simulation examples:

[0035] The present invention can be divided into three steps: obtaining a discretized data set, Bernoulli filtering, and judgment estimation. Specifically, the state transfer matrix, fuzzy matrix, and the probability of HMM existence at the initial moment and the probability distribution of each state are obtained by using the obtained discretized data set, and then a series of clutter observations are generated according to the Poisson point process model. At the same time, the real state at the previous moment will also be detected with a certain probability p. D To generate an HMM observation. Then predict the next state, and use the prediction equation of the Bernoulli filter to obtain the probability q of the existence of the HMM at the current moment and the probability mass function P of each state. The next step is to update q and P according to the update equation and likelihood function. Determine whether the HMM exists or not. If it does, the state corresponding to the maximum P is the estimated state at time k. Repeat the Bernoulli filtering until time k ≥ threshold.

[0036] like Figure 1 As shown, the specific method of the present invention is:

[0037] S1. Analyze the collected enemy data to obtain the transfer matrix between the target states; use the data set to obtain the fuzzy matrix of the state generation observation;

[0038] S2. Generate observations

[0039] Usually the actual observation model Z k By clutter model C k and theoretical measurement model O k Two parts, with Z k Represents all observations from time 1 to time k, represented by Z k ={Z1,Z2,…,Z k}. Use C k represents the clutter observation value, O k represents the observations generated by the HMM. If the HMM does not exist, then Z k =C k ; If HMM exists, then Z k =C k ∪O k .

[0040] Since the number of clutter observations follows a Poisson distribution, the number of measurements at each moment is |Z k |Unsure.

[0041] In the present invention, the clutter itself obeys a uniform distribution f in region V c (c) = 1 / V and the clutter sequence is randomly arranged, and the clutter measurement at time k is Clutter quantity m k It follows a Poisson distribution with parameter λ, m k~Po(λ). Then the probability distribution of clutter observation is:

[0042]

[0043] The probability density function of the theoretical measurement model is:

[0044]

[0045] The number of theoretical observations |O k |Subject to parameter p D (s k ) is a Bernoulli distribution, that is, the target s k The probability of being detected is p D (s k ); and g s (o k |s k ) is the likelihood function, indicating that the target state is s k The observation value is o under the condition k The probability of is given by the fuzzy matrix.

[0046] S3, Bernoulli filter

[0047] predict:

[0048] According to formula (4)(5)(6), we can get:

[0049]

[0050] q k|k-1 =p b (1-q k-1|k-1 )+p s q k-1|k-1 (12)

[0051]

[0052] The posterior probability density function f at the known k-1 time k-1|k-1 (S k-1 |Z k-1 ), predict the probability q of HMM existence at time k k|k-1 And the probability distribution P of each state k|k-1 (s k ).

[0053] renew:

[0054] According to the predicted q k|k-1 and P k|k-1 (s k ) and all observations at previous moments are updated.

[0055] Z={z1,z2,…,z m}, There are two forms, one One S = {s}. In these two cases, we have:

[0056]

[0057]

[0058] “\” represents the difference set. According to formula (7), we have:

[0059]

[0060]

[0061] Update the probability q of HMM existence at time k k|k And the probability distribution P of each state k|k (s k ).

[0062] S4, maximum a posteriori estimate of current state

[0063] If maxP k|k (s)>1-q, then HMM exists, and the maximum posterior is taken for the updated probability density function:

[0064]

[0065] Get the estimated state at time k Then, based on the correspondence between state s and the enemy's behavioral intention, the most likely task for the enemy to perform at the current moment, such as tracking, attacking, avoiding, etc., is obtained to complete the reasoning.

[0066] If maxP k|k (s)<1-q, it is judged as a false alarm, that is, the HMM is considered to be non-existent.

[0067] Loop step S3 until k is greater than or equal to the current time t, and output the most likely state Intention inference is complete.

[0068] Simulation Example

[0069] Based on the intercepted radar signal information and the comprehensive intelligence information from other sources, the present invention hopes to infer the enemy's combat intentions through the Bernoulli filtering method based on HMM. A simulation scenario is established, with a total of 11 states represented by 0 to 11, where 0 means that HMM does not exist, that is, The rest represent 10 different states, and there is a Markov state transition diagram between each state, such as Figure 2As shown, the transfer matrix can be obtained.

[0070] Set a true path that is more in line with the actual situation and the posterior probability density function f at time 1 1|1 (S1|Z1), starting from time 1, the prediction and update steps are repeated until time 100, which is one estimation. This simulation performs 100 estimations and obtains the most likely estimation path. The specific results are as follows Figure 3 shown.

Claims

1. A method for realizing enemy behavior intention reasoning based on HMM, characterized in that: The following steps are involved: S1. Collect the attribute data of the enemy radar, determine the real intention attribute of the target, that is, determine all possible intentions of the enemy, and discretize the data to obtain a discretized data set; analyze the attribute data of the target to obtain the transfer matrix between each state of the target; Using the data set, the fuzzy matrix of the state generation observation is obtained; S2, perform Bernoulli filtering on the data, predict and update the data based on the clutter observation and the HMM observation generated by the real state at the previous moment, so as to obtain the most likely state at the next moment, that is, the enemy's intention; The specific process is: The posterior probability density function of the initial moment obtained from the discretized data set obtained from S1 is: Then the posterior probability density function at time k-1 is expressed as: Among them, S k-1 represents the hidden Markov state at time k-1. There are two cases: the target does not exist. Or for a single target, S k-1 ={s k-1 }, Z k-1 represents the set of all observations from time 1 to time k-1, q k-1|k-1 and P k-1|k-1 (s k-1 ) represents the posterior probability of HMM existence and the posterior probability mass function of HMM state: q k-1|k-1 =P(S k-1 |=1|Z k-1 ) (3) P k-1|k-1 (s k-1 )=P(S k-1 ={s k-1 }|Z k-1 ) (4) The Bernoulli filter propagates these two quantities over time using a prediction equation and an update equation, specifically: Prediction equation: in, Representing the transition density of state S, we have: π(s k |s k-1 )=P(S k ={s k }S k-1 ={s k-1 }) indicates the state {s k-1 } to state {s k The transition probability between} is given by the transition matrix, p b represents the probability of HMM being born, that is, the probability of starting the plan at time k, p s It represents the probability of HMM survival, that is, the probability that the plan at the previous moment is not cancelled; the posterior probability density function f of each state at time k is predicted from time k-1 k|k-1 (S k |Z k-1 ), that is, the probability q of HMM existence is obtained k|k-1 And the probability distribution P of each state k|k-1 (s k ); Update equation: Among them, represents the observation likelihood function, which is obtained by the clutter observation and the fuzzy matrix. k |Z k-1 ) is defined as follows: By updating the equation, the prediction results are calibrated to obtain f k|k (S k |Z k ), the updated result is q k|k and P k|k (s k ); S3, based on the updated q k|k Make a judgment, if maxP k|k (s)>1-q, then HMM exists, and then P k|k (s k ) to perform maximum a posteriori estimation and obtain the state s with the highest probability at time k k ; If maxP k|k (s)<1-q, then HMM does not exist and it is judged as a false alarm; S4. Loop steps S2-S3. If time k ≥ threshold, end the loop and output the current state. The intention reasoning is completed.

Images