CN103648108B - Sensor network distributed consistency object state estimation method - Google Patents

Abstract

The invention provides a sensor network distributed consistency object state estimation method. The method, based on information transmission between observation nodes of a sensor network, enables dynamic function division to be carried out on sensor nodes in the network, and observation node sets participating in consistency state estimation are adaptively optimized and selected in real time; on the basis of a distributed maximum a posteriori theory, weighting processing is performed on object prior information and measurement information; and with the influence of covariance of state estimation errors of different observation nodes on the calculation of average consistency state taking into consideration, and effective information consistency processing is performed, the distributed state estimation accuracy of each observation node can rapidly approach to the centralized estimation accuracy, state maintenance of blind nodes on an object is guaranteed, and the cases of endless emergence of new tracks and uncertainty of the tracks and the like can be effectively prevented.

Description

Sensor network distribution type concordance Target state estimator method

Technical field

The present invention relates to the information fusion system of sensor network, more particularly, to a kind of sensor network distribution type concordance Target state estimator method, belongs to sensor information process field.

Background technology

Distributed method is mutually transmitted based on information effective between node realizes resource-sharing, has height in catenet Fault-tolerance, be easily installed and extend etc. advantage, therefore the research of distributing sensor network with apply in receive much concern.

For the application of sensor network multi-target distributed tracking, in numerous distributed state estimation methods, it is based on The algorithm for estimating of consistency theory, by the way of iteration, constantly updates local estimation using the effective information of neighbor node, from And make the overall situation estimation that each node can alone after all information taken in calculating network, such as the state of all nodes Estimate global mean value.It is critical that coherence method does not need TOCOM total communication to can be achieved with the consistent state estimation of the whole network, And approximate convergence is in centralized estimated result.Therefore, spy is not had to the communication topology of network based on conforming estimation framework Different requirement, principle is applied to the sensor network of any random connection.Additionally, in the application of fully distributed multiple target tracking In, although some nodes cannot cannot detect certain target to all targets or in certain time by direct detection, examine Consider and between node, need cooperative cooperating, each node is necessary to keep the state estimation to all targets, is otherwise constantly gone out Show the practical problems such as flight path is chaotic, new flight path emerges in an endless stream.It is also desirable to be gone through based on all targets after the measurement of node acquisition target History flight path is realized measuring the correct interconnection with flight path.And in Uniform estimates framework, each node remains all targets State estimation, this characteristic is inherently especially suitable for the application of fully distributed multiple target tracking.

The coherence methods such as traditional Kalman's concordance filtering (Kalman consensus filter, KCF) assume institute There is node all can observe each target, nodes all in network are considered as phase to the contribution weights that overall mean state is estimated With, and have ignored the impact when calculating average homogeneity character state for the different node state estimation difference covariances, not complete in network Estimated accuracy and the convergence rate of algorithm can be badly influenced under full-mesh or the more scene of blind node.Additionally, in the modern times In war and actual life, the sensing capability of microsensor is typically apart from limited and anisotropic (such as video sensing Device, directional microphone, radar etc.), that is, the performance of sensor relies on the distance between observation station and target and direction simultaneously.Cause This, the node in network can not keep the observation to all targets at any time, exist at any time in network and cannot detect target Blind node.And, the change of sensing model will have influence on the target observation quality of node, cover to network detection, node is worked in coordination with Control, Target state estimator distributed is followed the tracks of key technology and brought new challenge.

In the target following practical application of larger sensor network, only minority node of single moment can observe and pass through The target of monitored area, and the node in network generally all has the routing channel reporting monitoring information upwards, therefore, in distance In limited catenet, the single moment only needs minority observer nodes to realize the distributions estimation of target can be known The accurate flight path of target, just need not can meet the practical application of target following by all node TOCOM total communications.

It is true that constantly moving in monitored area with target, observer nodes are dynamic changes, by node between enter The effective data transfer of row is it is possible to obtain the information about firms of single moment observer nodes.Additionally, for apart from limited sensing Device network complete distributed consensus Target state estimator problem, can be based on distributed MAP estimation (maximum a Posteriori, MAP) theory, the information observation quality in conjunction with sensor and state estimation quality are consistent through enough times Property information transmission and iterative processing, the distributions estimated accuracy of each observer nodes can approach centralized optimum Kalman's filter Wave method (centralized Kalman filter, CKF).This namely the present invention thinking source.

Content of the invention

It is an object of the invention to provide a kind of high-precision dynamic self-adapting distributed consensus method for estimating state.For Reach above-mentioned purpose, the present invention based on the information transmission between observer nodes in sensor network it is proposed that a kind of based on point The adaptive information weighting coherency state method of estimation of cloth maximum a posteriori probability, schematic diagram is as shown in figure 1, include：Sensing Device node obtains target measurement information；Network node Partition of role；Set up concordance set of node；Calculate local parameter of consistency； The observer nodes consistency information processes and merges；Target state estimator；Dbjective state is predicted.

Technical scheme and be embodied as measure：

For the ease of illustrating, do following agreement：

Communication connection between sensor network interior joint can be by non-directed graphTo represent, wherein S={ S₁, S₂..., S_NContain all of summit of in figure, represent the communication node in network, and gatherThen contain in figure all of Side, represents the feasible communications conduit that in network, different nodes are set up.Additionally, withRepresent all and S_iThere is direction communication The set of the node connecting, that is,In each node and S_iCertain a line in pie graph, is all S_iNeighbor node.No Individual node S is assumed in harm_iOnly there is a sensor, observe target in t, then S_iReferred to as observer nodes, its measurement can table It is shown asWherein m_iFor sensor S_iMetric data dimension, its state and measurement model are represented by

x_t+1=Φ x_t+w_k, k=0,1,2 ..., (1)

z_i=H_ix_t+v_i, k=0,1,2 ..., (2)

Wherein,For state-transition matrix, process noise For sensor S_iCan Time-varying observing matrix,For the white Gaussian noise of zero-mean, variance isOrderFor amount Measurement information matrix, also can time-varying.It is pointed out that observing matrix H_iIt is not row full rank, that is, have m_i＜ p.Pass Sensor S_iError variance with regard to target prior estimate is expressed asIts information matrix is defined as

It is pointed out that the targeted problem of the present invention does not assume that state x of target_iFor each z_iBeing all can Observation, but consider in whole networkFor dbjective state x_tThere is observability, that is, network covers completely Lid, the single moment, at least one sensor can observe the target of monitored area.Additionally, all biographies in network might as well be assumed The communication radius of sensor node be not less than sensing 2 times of radius it means that, observe that the node of target is certainly adjacent each other simultaneously Occupy, only need a step communication to get final product mutual transmission information.The purpose of method provided by the present invention is, for sensor network target Tracking problem, constantly moves with target, and the role of network node members is continually changing, by effective information transmission and place Reason, realizes observer nodes and the blind node of the neighbours distributed consensus state estimation to target in any single filtered time instant.

The sensor node S of target T will be observed with t below_iAs a example, to the concrete steps in technical scheme and reality The mode of applying is described in detail.

1. sensor node obtains target measurement information

Sensor node is obtained measurement information and refers to be obtained the local measurement z with regard to target by target echo_iAnd measurement Information matrix B_i, whereinR_iMeasure the variance of the zero mean Gaussian white noise obeyed for sensor, subscript i is sensing The identity of device.

2. network node Partition of role

Whether the role that target and node are served as in Target state estimator is observed according to sensor node in network, will The node that t detects target is referred to as observer nodes, and the neighbor node (no target detected) of all observer nodes is referred to as blind Node, the node of other no target detecteds is referred to as sleeping nodes.Wherein, observer nodes execution principal states estimation work, blind Node is kept by the information receiving observer nodes and updates local dbjective state, to prevent from occurring in flight path handshaking The phenomenon such as new flight path emerges in an endless stream, flight path is failed to understand.

3. set up concordance set of node

T node S_iAfter detecting target, exchanged by the information between observer nodes, set up the holding of this moment to mesh The sensor node collection that mark state consistency is estimated, referred to as concordance set of node, it is designated as C.The measure of being embodied as is：

(1) sensor S_iGet the measurement z of t target T_iAnd its measurement information matrix B_i；

(2) sensor S_iThe state estimation of broadcast t-1 moment target TAnd self identification (ID), if T is new mesh Mark, then broadcast the measuring value with regard to T；

(3) sensor S_iReceive the broadcast message from neighbours' observer nodes；

(4) sensor S_iSet up concordance set of node C, including all observer nodes of t and its blind node of neighbours, its In, the number of observer nodes is N '.

Due to all nodes neighbours each other in concordance set of node, by a step information transmission, each observer nodes energy Receive the status information with regard to target for other observer nodes.Meanwhile, in concordance set of node all observer nodes the blind section of neighbours Point also can know target status information, and these dbjective states being derived from different observer nodes reach unanimity, and this will be rear Continue bright middle elaboration.

4. calculate parameter of consistency

Without loss of generality, before t being filtered, the prior state of target T is designated asPrior information matrix is designated asSensor S_iParameter of consistencyAndCalculation as follows：

$U_i^0=\frac{1}{N^{\′}}{Y}_i^{-}\left(t\right)+H_i^T{B}_i{H}_i---\left(3\right)$

$u_i^0=\frac{1}{N^{\′}}{Y}_i^{-}\left(t\right)x_i^{-}\left(t\right)+H_i^T{B}_i{z}_i---\left(4\right)$

5. the consistency information processes and merges

Sensor S_iIt is calculated parameter of consistencyAndAfterwards, combining information transmission, willAndCarry out unification Process.Set concordance iterationses as K, concordance rate factor is ζ, the measure of being embodied as is：

Three below step from the beginning of k=1, to k=K, loop iteration K time

(1) to neighbor nodeSend parameter of consistencyWith

(2) receive neighbor nodeParameter of consistencyWith

(3) it is based on consistency theory, update consistency parameter：

Finally, through K iteration, sensor S_iObtain parameter of consistencyAndLimit according to consistency theory, if iteration Number of times is enough, and the parameter of consistency acquired in each observer nodes will gradually reach unanimity.

6. Target state estimator

Sensor S_iAfter completing consistency information process, using following formula, dbjective state is estimated：

${\hat{x}}_i^{+}\left(t\right)={\left(U_i^K\right)}^{-1}{u}_i^K---\left(7\right)$

${\hat{Y}}_i^{+}\left(t\right)=N^{\′}\·U_i^K---\left(8\right)$

Other sensors node in concordance set of node C all realizes the estimation to dbjective state for the t according to above formula.

7. dbjective state prediction

Bonding state equation, the sensor node in concordance set of node carries out pre- respectively to dbjective state and information matrix Survey, concrete accounting equation is as follows：

${\hat{x}}_i^{-}(t+1)=\mathrm{\Φ}{\hat{x}}_i^{+}\left(t\right)---\left(9\right)$

$Y_i^{-}(t+1)={(\mathrm{\Φ}{\left(Y_i^{+}\left(t\right)\right)}^{-1}{\mathrm{\Φ}}^T+Q)}^{-1}---\left(10\right)$

The foregoing describe the information transmission of certain observer nodes single filtered time instant Nei and the concrete mode of process and step, warp Cross effective consistency on messaging and process it is achieved that the distributed consensus state to target for all the sensors in observer nodes set Estimate.And, the establishment stage in concordance set of node, each observer nodes broadcast the state value of previous moment target to network In it is achieved that observer nodes and its blind node of the neighbours Uniform estimates to dbjective state.It is so designed that, be easy to subsequent time and pass Sensor measures the interconnection with existing targetpath, and effectively prevents from leading to follow the tracks of the situation of failure because of target fast reserve, it is to avoid The phenomenon such as in network, new flight path emerges in an endless stream, flight path is failed to understand.

Compared with prior art, the present invention has the advantages that：

(1) estimated accuracy is high

Compared with traditional Uniform estimates method, it is theoretical, using sensing that the present invention is based on distributed maximum a posteriori probability Information transmission between device node, by introducing information weighting strategy, is established sensor observation information quality and is believed with concordance Contact between breath process.Compared with traditional Uniform estimates method, distributed consensus method provided by the present invention Estimated accuracy is closer to centralized optimal estimation method.

(2) state consistency efficiency high

For the Target state estimator in the single moment, method proposed by the invention only observes the biography of target in minority Carry out consistency information process, in conjunction with effective concordance rate factor it is only necessary to minority iteration several times between sensor node Realize the quick uniform convergence to Target state estimator for a small range all the sensors, also improve whole state estimation simultaneously The real-time of process.

(3) follow the tracks of reliability high

Method proposed by the invention achieves each moment observer nodes and its blind node of neighbours to dbjective state Cause property is estimated, is conducive to subsequent time sensor to measure data interconnection with existing targetpath, and effectively prevent because target fast Speed is motor-driven to be led to follow the tracks of the situation of failure, it is to avoid the phenomenon such as in network, new flight path emerges in an endless stream, flight path is not clear, improves target Follow the tracks of reliability.

(4) calculating is low with communication energy consumption

In single filtered time instant, Uniform estimates method proposed by the present invention only minority observe target node it Between produce communication and calculate energy consumption, without TOCOM total communication and concordance iteration, therefore energy consumption is low, and network cost is little, more applicable Sensor network in energy constraint.

(5) flexible design, realization are simple

Coherency state method of estimation proposed by the present invention can by information transmission simple between sensor node Lai Self-adaptative adjustment concordance set of node, need not predict the prior information of network memberses, have the characteristics that flexible design；Meanwhile, one Cause property status method is not related to the algorithm of complexity, under existing hardware technology level, can directly be adopted by digit chip Digital method is realized, and does not have special requirement to hardware performance, is easy to integrated chip.

(6) applied widely

On the one hand, as traditional coherency state method of estimation, the present invention has the relatively broad scope of application, can For the small-scaled wire fixed sensor network such as video, radar；On the other hand, because this Uniform estimates method can be every Individual sampling instant real-time adaptive adjusts working node, and extensibility is good, and this makes the invention especially suitable for having and dynamically opens up Flutter the large-scale wireless sensor network of structure.

Brief description

Fig. 1 is that coherency state estimates flow chart.

Fig. 2 is that sensor network target follows the tracks of simulating scenes figure.

Fig. 3 is the network topological diagram that 50 sensor nodes are constituted.

Fig. 4 is the variation diagram with concordance iterationses for the root-mean-square error meansigma methodss of each method of estimation.

Specific embodiment

With reference to the accompanying drawings and examples the present invention is described in further detail, embodiment 1 is a network all standing When monotrack example of simulation system, for describing sensor network coherency state estimation side proposed by the present invention in detail The detailed process of method (being designated as DMAP-KCF) and systematic function.

Embodiment 1

Target following simulating scenes as shown in Fig. 2 monitored area be 100m × 100m planar rectangular, sensor node with Machine is distributed, and target is arbitrarily mobile in region.The present embodiment intends distributed consensus state estimation side more proposed by the present invention Method DMAP-KCF and the performance of Kalman concordance filtering method KCF, centralized optimum kalman filter method CKF, adopt 50 Secondary Monte Carlo simulation takes and is all worth to test comparative result.Specific dbjective state parameter given below, sensor parameters and Parameter of consistency.

Dbjective state parameter

Target arbitrarily moves in monitored area, and its initial position and velocity attitude determine, velocity magnitude is in [1m/s at random 3m/s] interval in determine at random.Dbjective state can be described as x=(x, y, v_x, v_y) four dimensional vectors, wherein (x, y) is target position Put, (v_x, v_y) it is speed.Shown in the motion model of target such as formula (1), wherein, process-noise variance be set to Q=diag (5,5, 1,1).

For Target state estimator model, the model shown in formula (1) and state-transition matrix and process is equally adopted to make an uproar Sound.The initial time of target motion, each observer nodes is set to consistent to the original state of target and initial variance.Wherein, Initial prior estimate error variance is set toTarget true initial state value is added one The random noise that zero-mean variance is is as initial prior state

Sensor parameters

50 sensor nodes are randomly dispersed in monitored area, all the sensors isomorphism, and communication radius are 41m, sensing Radius is 20m, and its network topology structure is as shown in Figure 3；Sensor can observe any be in sensing radius within target, And can find range and angle measurement, measure as bivector. simultaneously.Measuring value z_iLinear measurement model shown in formula (2) to determine, Wherein noise variance R_i=10I₂, wherein I₂For two-dimentional unit vector.Observing matrix H_iAnd state-transition matrix Φ is defined as

$H_i=\left[\begin{array}{cccc}1& 0& 0& 0\\ 0& 1& 0& 0\end{array}\right],\mathrm{\Φ}=\left[\begin{array}{cccc}1& 0& 1& 0\\ 0& 1& 0& 1\\ 0& 0& 1& 0\\ 0& 0& 0& 1\end{array}\right]$

Parameter of consistency

Processing stage the consistency information of state estimation, if not particularly pointing out, the concordance of setting acquiescence changes Generation number K=5, concordance rate factor is set toWherein m is the number of observer nodes in the current filter moment.

Analysis of simulation result

Fig. 4 gives the situation of change with concordance iterationses for the averaged power spectrum precision of DMAP-KCF, KCF and CKF.? This, iterationses K gradually rises to 10 by 1, and other specification is default value.Wherein, averaged power spectrum definition of accuracy is all moment Estimation root-mean-square error meansigma methodss, can be expressed as

$\mathrm{ARMSE}=\frac{1}{N_f}\underset{k=1}{\overset{N_f}{\mathrm{\Σ}}}\sqrt{{\left|\right|\stackrel{\‾}{x}\left(k\right)-\mathrm{\ξ}\left(k\right)\left|\right|}^2}$

Wherein, N_fFilter length, ξ (k) is the actual position of k moment target,For k moment observer nodes to target State estimation average, is represented by

$\stackrel{\‾}{x}\left(k\right)=\frac{1}{N^{\′}}\underset{i=1}{\overset{N^{\′}}{\mathrm{\Σ}}}{\hat{x}}_i\left(k\right)$

Wherein, N ' is the number of k moment observer nodes,For the state estimation to target for the k moment sensor.

As seen from Figure 4, during K=1, DMAP-KCF has higher estimated accuracy than KCF, and close to centralized side Method CKF.With the continuous increase of K, the root-mean-square error of DMAP-KCF is gradually lowered, and Fast Convergent is to CKF estimated accuracy.K When >=8, the precision of two kinds of Distributed fusion algorithms tends to be steady.Generally, DMAP-KCF greatly improves distribution with respect to KCF Formula estimated accuracy.Trace it to its cause, method DMAP-KCF being primarily due to present invention offer considers between different observer nodes Priori redundancy, be assigned with suitable weights to prior information and measurement information it is achieved that consistent raw estimated accuracy is quick Approach centralized optimum kalman filter method.

Claims (1)

1. a kind of sensor network distribution type concordance Target state estimator method, is a kind of distributed filter for target following Wave method, is obtained measurement information, network node Partition of role, is set up concordance set of node, calculates locally by sensor node Parameter of consistency, the observer nodes consistency information are processed realizes sensor network with fusion, Target state estimator, dbjective state prediction In network, partly preferably node keeps dynamically consistent accurate estimation to dbjective state in real time；

It is characterized in that, described sensor node obtain measurement information refer to obtain by target echo local with regard to target Measure z_iWith measurement information matrix B_i, whereinR_iMeasure the variance of the zero mean Gaussian white noise obeyed for sensor, Subscript i is the identity of sensor；

Whether described network node Partition of role refers to, observe target and node in target according to sensor node in network The role serving as during state estimation, the node that t is detected target is referred to as observer nodes, neighbours' section of all observer nodes Point (no target detected) is referred to as blind node, and the node of other no target detecteds is referred to as sleeping nodes, and wherein, observer nodes are held Row principal states estimation work, blind node is kept by the information receiving observer nodes and updates local dbjective state；

Described concordance set of node of setting up refers to, observer nodes send the status information bag (state containing target previous moment Estimated valueAnd the identity of observer nodes), observer nodes and blind node receive the status information bag from neighbor node, The local dbjective state of blind node updates, all observer nodes of current time and all blind node consistence of composition sets of node, section Sensor in point set remains the state estimation of target, and the wherein sum of t observer nodes is N '；

The local parameter of consistency of described calculating refers to target prior information and measurement information are weighted processing, and calculates local Consistency information matrixWith local consistency information vectorCalculation is respectively Wherein,For prior estimate information matrix,P_iT () is state Estimation error variance, H_iFor measurement matrix；

The described observer nodes consistency information is processed and is referred to fusion, starts to k=K from k=1, following three step circulations are changed For K time：Observer nodes send and receive and be derived from the consistency information bag of identity, observer nodes containing local parameter of consistency The consistency information bag of neighbours' observer nodes, the local parameter of consistency of renewal Wherein, execution k=k+1 after each iteration is complete, K are concordance iterationses, and ζ is one Cause property rate factor, finally, each observer nodes obtain approximately uniform parameter of consistencyWith

Described Target state estimator refers to, in conjunction with parameter of consistency and observer nodes number, to dbjective state and information matrix It is updated,WhereinThe target state estimator state obtaining for t filtering,For corresponding state estimation information matrix；

Described dbjective state prediction refers to reference to current target state estimation and motion model, to subsequent time target State and its information matrix are predicted,Wherein,For the dbjective state prediction in t+1 moment, Φ is state-transition matrix,For t+1 moment Target state estimator The prediction of information matrix, Q is the process-noise variance in target movement model.