CN1382997A - Precise tracking method based on nerve network for moving target - Google Patents

Abstract

The invented method for accurate tracking the maneuvering target based on neural net adopts double parallel structure being composed of two filters. The difference of states between the parallel double filters is picked-up as the character vector of the state of target to be estimate, which is input to the neutral net system. Based on the character vector input, the neural net system, which has been trained off-line before tracking estimation, outputs the predicted value of the acceleration variance of the target to be estimated at present time. Adaptive filtering the said predicted value based on the current statistical model obtains the fast and accurate tracking the target to be estimate. The invented method possesses favorable adaptive tracking ability.

Description

Maneuvering target precise tracking method based on neural network

Technical field:

The present invention relates to a kind of maneuvering target precise tracking method based on neural network (NN), be used for the maneuvering target hi-Fix and the prediction of systems such as intelligent transportation, robot, aviation electronics, Defence Against Ballistic Missile and precise guidance, belong to the intelligent information processing technology field.

Background technology:

Accurate tracking problem at motor-driven single goal, many maneuvering target model and adaptive tracing algorithm have been proposed respectively since the eighties both at home and abroad, comprise the differential polynomial model, at the uniform velocity with even acceleration model, the time correlation model, semi-Markov model and maneuvering target " current " statistical model (Zhou, H.and K.S.P.Kumar.Acurrent statistical model and adaptive algorithm for estimating maneuvering targets.ALAAJournal of Guidance, Control, and Dynamics, 1984,7 (5), 596-602.).Track algorithm mainly is to adopt adaptive tracking method." current " statistical model and adaptive tracking method are applicable to strong maneuvering target, have the advantage that response is fast, tracking accuracy is high, but then precision are lower to non-maneuver tracking.Because each model can only reflect certain side of realistic objective, such as at the uniform velocity with even acceleration model can only reflect target at the uniform velocity and uniformly accelerated motion, therefore the contradiction between response speed and tracking accuracy just becomes the difficult point of maneuvering target tracking.In order to overcome this difficult point, interactive multi-model adaptive tracking method (Blom, H.A.P., and Bar-Shalom, Y., The interacting multiple model algorithms for systems with Markovian swithchingcoefficients.IEEE Tran.Automatic Control, 1988,33 (8), 780-783) be suggested, though this method has improved the performance of following the tracks of to a certain extent, the speed of its tracking response and precision are from real system, and the demand of particularly accurate tracking system also has very big distance.Because the unpredictability of target travel characteristic, seek a model and an adaptive algorithm that the various kinetic characteristics of target are all had a good tracking accuracy and attracting numerous researchers always.

Summary of the invention:

The objective of the invention is at the deficiencies in the prior art, a kind of new maneuvering target precise tracking method based on neural network is proposed, make the tracking of maneuvering target when keeping adaptive tracing algorithm CSMAT method advantage, can also increase substantially tracking accuracy, can accomplish motor-driven and nonmaneuvering target non-motor-driven or weak maneuvering target " comprehensively " the self-adaptation high precision tracking.

For realizing such purpose, the present invention is based on " current " statistical model and BP (Error Back Propagation, direction of error is propagated) neural network, in conjunction with the neural network information fusion technology, the estimated information of full use dbjective state, a kind of maneuvering target parallel adaptive track algorithm (NIFPAT) based on neural network of novelty has been proposed, adopt the double filter parallel organization, utilize the total state feedback, to adapt to the motion change of target, realize that various motion states are had good adaptive tracing ability by BP network Adjustment System variance.

Maneuvering target precise tracking method based on neural network proposed by the invention, comprise feature extraction, network training and three basic steps of fusion tracking: 1, status flag extracts

Utilize the difference of the output of parallel double filter to carry out the extraction of status flag.Adopt two wave filters to form a two parallel organization, one of them wave filter adopts current statistical model, chooses maximum acceleration variance adapting to the various variations of target, and keeps motor-driven quick response; Another wave filter also adopts " current " statistical model, but its acceleration variance then is to determine size according to the output result of neural network, and initial value is got minimum acceleration variance, desirable 1.

Because the difference of variance, the status tracking precision of double filter comprises that position, speed and acceleration are also inequality, promptly has certain error.According to the tracking results of double filter, ask the norm of its error, and utilize sign function to handle, can obtain to treat the status flag vector of estimating target.Give nerve network system with the proper vector of extracting as input.2, network training

Before Tracking Estimation, nerve network system off-line training is good.Find that by a large amount of emulation the proper vector Changing Pattern of position, speed and acceleration is respectively μ to various motor-driven levels of target and forms of motion ₁(X) ∈ [L ₁, H ₁], μ ₂(X) ∈ [L ₂, H ₂] and μ ₃(X) ∈ [L ₃, H ₃].Wherein L represents minimum, and H represents mxm..The present invention utilizes the proper vector Changing Pattern μ of position, speed and acceleration ₁∈ [0.05,0.2], μ ₂∈ [0.2,0.6] and μ ₃∈ [0.1,0.4] selects for use the BP neural network to regulate, and its learning sample is chosen as follows: when each component of proper vector is in low value, think that target is in non-maneuvering condition, network is output as one and approaches 0 value; When each component of proper vector is in the high value, think that target is in strong maneuvering condition, network is output as one and approaches 1.0 value; When each component of proper vector changes between low value and high value, think that target is in weak maneuvering condition, network is exported the value between (0,1) automatically.3, state merges and adaptive tracing

At each proper vector for the treatment of estimating target that is constantly obtained, utilize the neural network that has trained, with the input of proper vector, and utilize the output O of neural network as neural network _NN∈ [1,1] is with different variances Adapt to the strong and weak motor-driven and non-motor-driven variation of target.Its output is the acceleration variance predicted value that current time is treated estimating target.Utilize this predicted value, carry out auto adapted filtering based on current statistical model and can obtain treating following the tracks of fast and accurately of estimating target.

The present invention has utilized the space time information of dbjective state, can not only more effective tracing machine moving-target (comprising that acceleration is motor-driven), and can follow the tracks of nonmaneuvering target exactly, has comprehensive adaptive tracing ability to the various motion states of target, and can farthest improve tracking accuracy, especially speed, acceleration precision have improved the fault freedom of system simultaneously.

Neural network feature extracting method and parallelism wave filter structure that the present invention adopts, can avoid based on the measurement noise that the double filter cascaded structure produced of statistical filtering method coloured relevant, estimate problems such as relevant, be a kind of effective and simple and direct motor-driven and nonmaneuvering target adaptive tracking method.

The present invention can be used for maneuvering target hi-Fix and prediction, on national defence, can be used for systems such as aviation electronics, Defence Against Ballistic Missile, precise guidance, can be used for fields such as air traffic control, safety inspection, remote sensing, crashproof, navigation and robot vision aspect civilian.

Description of drawings:

Fig. 1 is the maneuvering target tracking method theory diagram that the present invention is based on neural network.

Aimed acceleration aircraft pursuit course in Fig. 2 embodiment of the invention.

Embodiment:

For understanding technical scheme of the present invention better, be further described below in conjunction with accompanying drawing and specific embodiment.

The maneuvering target tracking method principle that the present invention is based on neural network as shown in Figure 1.Among the figure, wave filter F ₁And F ₂Formed a two parallel organization.At first echoed signal is through F ₁And F ₂Carry out filtering, F ₁The variance of wave filter is chosen maximum acceleration variance and is changed to adapt to various target maneuvers.F ₂The acceleration variance value of wave filter then is definite according to the output result of neural network.F ₁And F ₂The difference of filter state carry out feature extraction by characteristic extracting module, the proper vector of extraction is given nerve network system as input.Before Tracking Estimation, nerve network system off-line training is good.According to the proper vector of input, the relatively accurate acceleration variance predicted value of estimating object is treated in one of neural network output, feeds back to wave filter F ₂Wave filter F ₂According to the acceleration variance predicted value of neural network output, adopt adaptive tracing algorithm based on " current " statistical model, provide and treat the estimation fast and accurately of estimating object.The present invention adopts the double filter parallel organization, utilizes the total state feedback, to adapt to the motion change of target, has the good adaptive tracing ability to various motion states by BP network Adjustment System variance.

Embodiment 1, status flag extract

In order to overcome the influence of noise,, adopt F for the status information blending algorithm to characteristic quantity ₁And F ₂The error norm of wave filter current time valuation is as the input of neural network NN, and its component is

In the formula

With

Be wave filter F ₁And F ₂At k constantly based on echo (to position, speed and the acceleration filter value of target t.2, network training

Selected feature samples such as table 1.

Table 1 feature samples

????Input	????(0，0，0)	????(0，0，1)	????(0，1，0)	????(0，1，1)	????(1，0，0)	????(1，0，1)	????(1，1，0)	????(1，1，1)
									????net 0	????0.10	????0.75	?????0.60	????0.85	????0.40	????0.75	????0.60	????0.95

Because of three layers of BP network have the learning ability of approaching any nonlinear function, choose the BP network that contains a hidden layer, wherein input number of nodes is 3, the output node number is 1, the number of hidden nodes is 7, learning rate is 0.7, and the inertia scale factor is 0.9, and the network iteration can obtain satisfied convergence effect after 1500 steps.3, state merges and adaptive tracing

Wave filter F ₁And F ₂All adopt the adaptive tracing algorithm based on " current " statistical model, its state equation, measurement equation, filtering formula and state covariance matrix are respectively:

X(k+1)＝ΦX(k)+U a+Gw(k)

Z(k)＝HX(k)+v(k)

X(k|k-1)＝ΦX(k-1|k-1)+U a(k)

P(k|k-1)＝ΦP(k-1|k-1)Φ ^T+GQ(k-1)G ^T

K(k)＝P(k|k-1)H ^T[HP(k|k-1)H ^T+R] ^-1

X(k|k)＝X(k|k-1)+K(k)[Z(k)-H(k)X(k|k-1)]

P(k|k)＝[I-K(k)H]P(k|k-1) $Q_1\left(k\right)=2\mathrm{\&alpha;}{\mathrm{\&sigma;}}_{a1}^2{Q}_0$ $Q_2\left(k\right)=2\mathrm{\&alpha;}{\mathrm{\&sigma;}}_{a2}^2{\mathrm{<figure-callout\; id="0"\; label="Q"\; filenames="A0211206100092.png"\; state="\{\{state\}\}">Q</figure-callout>}}_0$

O wherein _FNNOutput for fuzzy neural network; α is motor-driven frequency; Q ₀Be the normal value matrix relevant with sampling period T with α.When object module employing " current " statistical model, initial parameter sees Table 2.

Table 2 initial parameter

	Type of sports	Initial position (km)	Initial velocity (m/s)	Initial acceleration (m/s 2)
					Target 1 target 2	The uniform motion galloping motion	(70，70，3) (100，30，3)	(-400，-400，0) (-400，-50，0)	(0，0，0) ((0，10，5)，0，0)

Wherein target 2 in the acceleration change process of x direction is: 0m/s (0-100s), and 10m/s (100-150s), 5m/s (150-300s), two targets intersect at the R=40km place, motor-driven frequency alpha=0.05, the sampling period is 1s.Radar accuracy is respectively apart from 0.4 ° in 85m, range rate 15m/s, 0.4 ° of pitching and orientation; Radar detedtion probability is 0.99, and clutter density λ is

Simulation result such as Fig. 2, Fig. 2 are the aimed acceleration aircraft pursuit course.Among the figure, "-" is true value, and ". " is the classic method filter value, and "-" is this method filter value.

The present invention is based on the maneuvering target precise tracking method of neural network and the comparative result of classic method and see Table 3, the error ratio of compression is to describe filtering improves degree with respect to the average behavior of measuring performance evaluation index.Method of the present invention as seen from Table 3 has higher tracking performance.

Table 3 error ratio of compression EN

Target	Method	Distance	Speed	Acceleration	The orientation	Pitching
							1 1 2 2	Classic method the inventive method classic method the inventive method	?0.4219 ?0.4001 ?0.6211 ?0.5423	?0.5454 ?0.3551 ?0.6932 ?0.4112	?0.6923 ?0.5443 ?0.7412 ?0.5811	?0.6721 ?0.6322 ?0.7522 ?0.7133	?0.7016 ?0.6511 ?0.7911 ?0.7678

Claims (1)

1, a kind of maneuvering target precise tracking method based on neural network is characterized in that comprising feature extraction, network training and three basic steps of fusion tracking:

1) status flag extracts: adopt two wave filters to form two parallel organizations, wave filter adopts current statistics mould

Type, one of them chooses maximum acceleration variance to adapt to object variations and to keep motor-driven fast

The speed response, another acceleration variance is determined size according to the output result of neural network, initial value is got

Minimum acceleration variance, the difference of extracting the state of parallel double filter is to treat the state spy of estimating target

Levy vector;

2) network training: the proper vector Changing Pattern that utilizes position, speed and acceleration

μ ₁∈ [0.05,0.2], μ ₂∈ [0.2,0.6] and μ ₃∈ [0.1,0.4] selects for use the BP neural network to regulate, its

Learning sample is chosen for: when each component of proper vector is in low value, think that target is in non-motor-driven shape

Attitude, network are output as one and approach 0 value; When each component of proper vector is in high value, think order

Mark is in strong maneuvering condition, and network is output as one and approaches 1.0 value; When each component of proper vector exists

When changing between low value and the high value, think that target is in weak maneuvering condition, network is exported (0,1) automatically

Between value;

3) state merges and adaptive tracing: at each proper vector for the treatment of estimating target that is constantly obtained,

Utilize the neural network that has trained,, and utilize refreshing the input of proper vector as neural network

Output O through network _NN∈ [1,1] is with different variances

Adapt to target

Strong and weak motor-driven and non-motor-driven variation, utilize the current time of output to treat the acceleration variance of estimating target

Predicted value is carried out auto adapted filtering based on current statistical model, obtain treating estimating target fast and

The accurate tracking.

Images