CN104936209A - Distributed filtering method based on adjustable weights - Google Patents

Abstract

A distributed filtering method based on an adjustable weights introduces an assessment algorithm of node confidence to obtain the target state estimation confidence of sensor network nodes, secondly, defines the loads of the nodes as the target estimation confidence degrees of the nodes in a sensor network, distributes the weights of the attacked nodes by a weight redistribution method, updates the target state estimation confidence of the sensor network nodes, and introduces the weights composed by the confidence into a consistency protocol to update the target state estimation values of the sensor nodes, thereby improving the estimation precision of the distributed filtering method and the estimation value consistency of the sensor nodes. The distributed filtering method based on the adjustable weights of the present invention enables the estimation precision of a wireless sensor network to the target tacking to be improved under a non-trust environment.

Description

A kind of Distributed filtering method adjustable based on weights

Technical field

The present invention relates to a kind of Distributed filtering method adjustable based on weights, belong to wireless sensor network and control technology field.

Background technology

The extensive use of distributed network, has extremely important directive significance to the development of multiple-sensor network, is widely used in fields such as military and national defense, social economy, network communication, Systematical control.Researchers explore between intelligent individuality how to complete complicated task by concertation control, information fusion from multiple angle.A basic problem in wireless sensor network research is exactly by designing suitable filtering algorithm, the information utilizing transducer to collect completes accurately to be estimated and the problem of following the tracks of, its key be to find a kind ofly take into account stability, convergence, ageing and energy-conservation algorithm for estimating carry out effective integration to the information between transducer.Research shows that based on conforming filtering algorithm be effective blending algorithm, and this kind of algorithm does not need fusion center, and node only needs to carry out information interaction with neighbor node, finally makes all-network node state estimated value reach unanimity, realizes high-precision estimation.This mechanism of carrying out communication with the node in contiguous range greatly reduces the energy consumption of network.

Multiple-sensor network has strong robustness, the preferably feature such as adaptivity and low in price cost, overcome the circumscribed defect of single-sensor self-ability, storage and disposal ability, make wireless sensor network have huge application market and application space.Such as, broadcast sowing extensive intensive sensor node at random by aircraft, zone of ignorance detected, target location is followed the tracks of, detect and locate.But, consider the restriction of the hardware resource of node in wireless sensor network own and be easily subject to the interference of external environment, node generation obstacle, communication failures or communication abnormality in wireless sensor network, there is larger error in the data that node perceived arrives, makes the tracking of wireless sensor network to dynamic object have poor real-time.The existence of this kind of unfavorable factor wireless sensor network is existed problems such as data are unreliable, loss of data so that many traditional algorithms also no longer adapt to.The solution proposed the earliest this kind of problem is that consistency algorithm is applied to wireless sensor network, and application weighted average-consensus algorithm estimates the mean value of all the sensors node input value, can solve the situation of sensor fault preferably.Also have the most widely used algorithm of a class to be Kalman's consistency filtering algorithm, introduce consistency filter and merge the estimated value of target, covariance sensor node, multiple information is distributed transmits the state estimation improving each node.But this kind of algorithm is all based upon on supposition sensors observe data reliable basis and carries out, and namely assert that the environment residing for transducer is trusty.In actual applications, sensor node is exposed to outside uncontrolled space, is very easily attacked, and information is easily tampered in transmittance process, forges, thus the precision affecting target following location and follow the tracks of.The meanwhile limitation of node construction itself and the power of the perception of environment to external world, make the perception of the node of zones of different diverse location to same target completely different, how to make the contribution margin that the larger estimation effect of perception strong node performance acts in whole estimation to improve it, weaken the node of perception difference to the impact of whole estimation, this problems demand solves.

Summary of the invention

The object of the invention is, in order to the effective integration problem of wireless sensor network system state estimation of being attacked under solving non-trusted environment, provide a kind of Distributed filtering method adjustable based on weights.

Technical scheme of the present invention is, a kind of Distributed filtering method adjustable based on weights, the described position of methods analyst sensor node in network topology structure and the influence power to network thereof, introduce the assessment algorithm of node believes degree, obtain the certainty factor of each sensor network nodes to Target state estimator; Secondly, use for reference the multiple load-carrying of Complex Networks Theory and distribute thought, by the load definition of node be in sensor network this node to the certainty factor value of target state estimator, the method of weight reassignment is adopted to distribute suffering the weights attacking node, upgrade the certainty factor of each sensor network nodes to Target state estimator, and the weights formed by this certainty factor to be introduced in consistency protocol more new sensor node and, to the state estimation of target, thus are improved the estimated accuracy of Distributed filtering algorithm and the consistency of sensor node estimated value.

Described method, according to node topology location in a network, obtains each node certainty factor value in a network, and the impact of node on neighbor node estimated value that certainty factor is larger is larger; Two nodes state estimation difference is at a time larger, and the mutual degree of support of two nodes is lower; The node that certainty factor is less, the weight that neighbor node merges this node state estimated value is less, and the weights making the node of estimated performance difference participate in merging diminish; The node that certainty factor is larger, participates in the aggregate contribution effect of consistency stage larger; The cooperation synergistic mechanism of coherency mechanism, makes each sensor node approach to actual value, thus improves network to the estimated accuracy of target.

A kind of Distributed filtering method adjustable based on weights of the present invention, embed micro-filter in the sensor, each transducer is used as a node, and calculation procedure is as follows:

(1) according to node topology location in a network, give the certainty factor that each node is different, node believes degree K (i) computing formula is as follows:

$S_i=\frac{B_1\left(i\right)}{B_2\left(i\right)}---\left(1\right)$

$K\left(i\right)=S\left(i\right)/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_{\mathrm{ij}}---\left(2\right)$

In formula, i represents each sensor node, i ∈ 1,2 ... n}; υ _ijfor the beeline between sensor node i and sensor node j; B ₁(i) for network all the sensors node is to the shortest path number through node i, B ₂(i) for sensor network all shortest path sum;

(2), after obtaining the certainty factor of each sensor node, calculate the weight that each sensor node participates in each node in fusing stage contiguous range, formula is as follows:

$p_i\left(t\right)=K\left(i\right)/\underset{i=1}{\overset{n}{\mathrm{\Σ}}}K\left(i\right)---\left(3\right)$

d _ij(t)＝1-2arctan(Δ(t))/π (4)

p _i(t+1)＝p _i(t)+Δp _i(t+1) (5)

${\mathrm{\Δp}}_i(t+1)=p_j\left(t\right)\·\frac{p_i\left(t\right)\·d_{\mathrm{ij}}\left(t\right)}{\underset{l\∈{\mathrm{\Γ}}_j}{\mathrm{\Σ}}{p}_l\left(t\right)\·d_{\mathrm{lj}}\left(t\right)}---\left(6\right)$

In formula, be the state estimation difference of two nodes, the k moment of estimated value for sensor node i in to(for) target; d _ijt () is that two internodal state estimation exist _tthe support in moment; Δ p _it () represents that node j distributes to the weights of its neighbor node i; Γ _jfor adjacent node collection (wherein, the Γ of node j _jdo not comprise node j).The node that weight is larger, the influence for neighbor node partial estimation value is larger.

A kind of Distributed filtering method adjustable based on weights of the present invention, specific implementation process is as follows:

(1) set algorithm greatest iteration step number step _max, all nodes of initialization to the priori estimates of goal systems and estimate covariance matrix, sytem matrix and other parameters;

(2) exist _tmoment sensor node detects new measured value z _i(t);

(3) each node filter gain of transducer is upgraded: L _i(t)=AP _i(t) H _i' (R+H _ip _ih _i') ^-1

(4) according to formula computing node V _iweight ratio in the entire network;

(5) weights are calculated: $c_{\mathrm{ij}}\left(t\right)=\frac{p_j\left(t\right)}{{\mathrm{\Σ}}_{j\∈N_i}{p}_j\left(t\right)};$

(6) each node state estimated value of calculating sensor network;

$x_i\left(t\right)=A{\hat{x}}_i\left(t\right)+L_i\left(t\right)(z_i\left(t\right)-H_i\left(t\right){\hat{x}}_i\left(t\right));{\hat{x}}_i(t+1)={\mathrm{\Σ}}_{j\∈N_i}{c}_{\mathrm{ij}}\left(t\right)x_j\left(t\right);$

(7) gain matrix of each sensor node is upgraded:

$P_i(t+1)=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}\left(t\right)[(A-L_j\left(t\right)H_j\left(t\right))P_j\left(t\right){(A-L_j\left(t\right)H_j\left(t\right))}^{\′}+L_j\left(t\right){\mathrm{RL}}_j{\left(t\right)}^{\′}]+Q;$

(8) if t is less than greatest iteration step number step _max, t=t+1, returns step (2); Otherwise algorithm terminates.

Under sensor network is in non-trusted environment, namely suffer to attack (if there is multiple sensor node to be attacked at a certain t sensor node i, then suppose some nodes of being attacked not neighbor node each other), above-mentioned algorithm steps is amended as follows:

I, the estimated value of t-1 moment node i and the estimated value between each neighbor node is utilized to calculate corresponding each internodal support angle value and assignment gives the degree of consistency value between t node i and each neighbor node, namely calculate and upgrade weights p _i(t).

II, in Distributed filtering algorithm incremental update step (5) and (6), parameter lambda is introduced _i, wherein, λ _i=1 represents that node i is not attacked, λ _i=0 represents node V _iattacked.Formula then in step (5) and (6) is rewritten as respectively:

$x_i\left(t\right)=A{\hat{x}}_i\left(t\right)+L_i\left(t\right)(z_i\left(t\right)-H_i\left(t\right){\hat{x}}_i\left(t\right))---\left(7\right)$

$P_i(t+1)=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}\left(t\right)[(A-L_j\left(t\right)H_j\left(t\right))P_j\left(t\right){(A-L_j\left(t\right)H_j\left(t\right))}^{\′}+L_j\left(t\right){\mathrm{RL}}_j{\left(t\right)}^{\′}]+Q---\left(8\right)$

In t, when having node to be attacked in sensor network, by the λ in formula (7) and formula (8) _ibe set to 0, mean and this node believes angle value is set to 0, namely make the estimation of any data to whole algorithm of this node cut little ice.When merging renewal, readjust the weights of each neighbor node, and meet ${\mathrm{\Σ}}_{j\∈N_i}{c}_{\mathrm{ij}}\left(t\right)=1.$

The invention has the beneficial effects as follows, under the present invention can improve non-trusted environment, wireless sensor network is to the estimated accuracy of target following.

Accompanying drawing explanation

Fig. 1 is the network topology structure schematic diagram that 30 sensor nodes are formed;

Fig. 2 is the averaged power spectrum application condition figure under random attack;

Fig. 3 is the nonuniformity averaged power spectrum application condition figure under random attack;

Fig. 4 is the averaged power spectrum application condition figure under selection is attacked;

Fig. 5 is the nonuniformity averaged power spectrum application condition figure under selection is attacked;

Fig. 6 is different attack rate, the random evaluated error comparison diagram attacking lower algorithm;

Fig. 7 is different attack rate, selects the evaluated error comparison diagram attacking lower algorithm;

Fig. 8 is the calculation process block diagram of the inventive method.

Embodiment

The wireless sensor network G that the embodiment of the present invention adopts 30 sensor nodes to form estimates a certain state by the goal systems of Gauusian noise jammer, and the state model of goal systems and observation model are:

x(t+1)＝Ax(t)+w(t) (8)

z _i(t)＝H _ix(t)+v _i(t) (9)

Wherein, x (t) ∈ R ^{m × 1}for goal systems is in the state vector of t, z _i(t) ∈ R ^{m × 1}for sensor node i is at the observation vector of t, A ∈ R ^{m × m}for goal systems state-transition matrix, H _i∈ R ^{m × m}for the observing matrix of sensor node i.W (t) ∈ R ^{m × 1}for goal systems is at the process noise of t, and meet w _t~ N (0, Q); v _i(t) ∈ R ^{m × 1}for sensor node i is in the observation noise of t, and meet v _i(t) ~ N (0, R).Covariance meets Cov (w _k, w _l)=Q δ (k-l), Cov (v _i(k), v _j(l))=R δ (i-j) δ (k-l), wherein, δ (τ)=1, τ=0, otherwise δ (τ)=0, τ=1.

Based on above model, in the process implemented, the method and the former algorithm that propose the present invention respectively carry out emulation experiment in targeted attack pattern and at random under attack mode, secondly, focus on when network exists different attack rate and carry out emulation experiment, and be analyzed in each node state estimated value, error, stability.

Fig. 1 is the network topology structure figure of 30 Node distribution in the present embodiment emulation, with small circle representative sensor node.Stochastic choice 3 network nodes are attacked, Fig. 2 nonuniformity averaged power spectrum application condition figure that to be the averaged power spectrum application condition figure under network attack state at random, Fig. 3 be under network attack state at random.Wherein, the state equation of target is as follows with the observational equation parameter choose of system: state value x=[10,15] ', state-transition matrix $A=\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right],$ Systematic procedure noise covariance matrix $Q=\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right],$ Initial condition $P_{\text{i}}\left(0\right)=\left[\begin{array}{cc}1& 0\\ 0& 1\end{array}\right].$ Sensor node observing matrix $H_i=\left[\begin{array}{cc}{c}_1& 0\\ 0& c_2\end{array}\right],$ Wherein c ₁∈ [0.2 ~ 0.8], c ₂∈ [0.2 ~ 0.8], sensor node observation noise covariance matrix is $R_i=\left[\begin{array}{cc}0.8+0.2a_1& 0\\ 0& 0.8+0.2a_2\end{array}\right],$ Wherein a ₁and a ₂it is the random number between [0 ~ 1].Algorithm iteration step number step is set _max=300.As can be seen from Fig. 2, Fig. 3, the method estimation effect after improvement is significantly improved.

Under identical parameters setting model, the certainty factor of node each in sensor network is sorted, select 3 high network nodes of certainty factor to attack, with the wireless sensor network that G is formed, state is estimated, carry out emulation experiment.Fig. 4 provides the evaluated error comparison diagram of algorithm in targeted attack situation, and Fig. 5 provides the nonuniformity averaged power spectrum application condition figure under targeted attack state.

Under Fig. 6, Fig. 7 provide the different attack rate situation of algorithm, to attack and the averaged power spectrum application condition figure of algorithm under targeted attack pattern at random.

As can be seen from the figure, no matter be the speed that whole estimated accuracy or each sensor node reach unanimity, the algorithm of proposition all will apparently higher than traditional algorithm.This explanation, at consistency protocol, more the new stage utilizes the estimation certainty factor of sensor node to target to merge the information of other sensor nodes, significantly improve the estimated performance of algorithm, and when network suffers external environmental interference (the present invention considers to apply white Gaussian noise to the part of nodes of Stochastic choice), the estimated performance that the method that the present invention proposes can keep relative stability.

Fig. 8 is the calculation process block diagram that the present invention is based on the adjustable Distributed filtering method of weights.

Claims (5)

1. based on the Distributed filtering method that weights are adjustable, it is characterized in that, described method introduces the assessment algorithm of node believes degree, obtains the certainty factor of each sensor network nodes to Target state estimator; Secondly, by the load definition of node be in sensor network this node to the certainty factor value of target state estimator, adopt the method for weight reassignment, distribute suffering the weights attacking node, upgrade the certainty factor of each sensor network nodes to Target state estimator, and the weights formed by this certainty factor to be introduced in consistency protocol more new sensor node and, to the state estimation of target, thus are improved the estimated accuracy of Distributed filtering method and the consistency of sensor node estimated value;

Described method, according to node topology location in a network, obtains each node certainty factor value in a network, and the impact of node on neighbor node estimated value that certainty factor is larger is larger; Two nodes state estimation difference is at a time larger, and the mutual degree of support of two nodes is lower; The node that certainty factor is less, the weight that neighbor node merges this node state estimated value is less, and the weights making the node of estimated performance difference participate in merging diminish; The node that certainty factor is larger, participates in the aggregate contribution effect of consistency stage larger; The cooperation synergistic mechanism of coherency mechanism, makes each sensor node approach to actual value, thus improves network to the estimated accuracy of target.

2. a kind of Distributed filtering method adjustable based on weights according to claim 1, it is characterized in that, described method embeds micro-filter in the sensor, and each transducer is used as a node, and calculation procedure is as follows:

(1) according to node topology location in a network, give the certainty factor that each node is different, node believes degree K (i) computing formula is as follows:

$S_i=\frac{B_1\left(i\right)}{B_2\left(i\right)}$

$K\left(i\right)=S\left(i\right)/\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{\mathrm{\υ}}_{\mathrm{ij}}$

In formula, i represents each sensor node, i ∈ 1,2 ... n}; υ _ijfor the beeline between sensor node i and sensor node j; B ₁(i) for network all the sensors node is to the shortest path number through node i, B ₂(i) for sensor network all shortest path sum;

(2), after obtaining the certainty factor of each sensor node, calculate the weight that each sensor node participates in each node in fusing stage contiguous range, formula is as follows:

$p_i\left(t\right)=K\left(i\right)/\underset{i=1}{\overset{n}{\mathrm{\Σ}}}K\left(i\right);$

d _ij(t)＝1-2arctan(Δ(t))/π；

p _i(t+1)＝p _i(t)+Δp _i(t+1)；

$\mathrm{\Δ}{p}_i(t+1)=p_j\left(t\right)\·\frac{p_i\left(t\right)\·d_{\mathrm{ij}}\left(t\right)}{\underset{l\∈{\mathrm{\Γ}}_j}{\mathrm{\Σ}}{p}_l\left(t\right)\·d_{\mathrm{lj}}\left(t\right)};$

In formula, be the state estimation difference of two nodes, the k moment of estimated value for sensor node i in to(for) target; d _ijt () is the supports of two internodal state estimation in t; Δ p _it () represents that node j distributes to the weights of its neighbor node i; Γ _jfor the adjacent node collection of node j.

3. a kind of Distributed filtering method adjustable based on weights according to claim 1, it is characterized in that, the specific implementation process of described method is:

(1) set algorithm greatest iteration step number step _max, all nodes of initialization to the priori estimates of goal systems and estimate covariance matrix, sytem matrix and other parameters;

(2) exist _tmoment sensor node detects new measured value z _i(t);

(3) each node filter gain of transducer is upgraded: L _i(t)=AP _i(t) H _i' (R+H _ip _ih _i') ^-1

(4) according to formula computing node V _iweight ratio in the entire network;

(5) weights are calculated: $c_{\mathrm{ij}}\left(t\right)=\frac{p_j\left(t\right)}{{\mathrm{\Σ}}_{j\∈N_i}{p}_j\left(t\right)};$

(6) each node state estimated value of calculating sensor network;

$x_i\left(t\right)=A{\hat{x}}_i\left(t\right)+L_i\left(t\right)(z_i\left(t\right)-H_i\left(t\right){\hat{x}}_i\left(t\right));{\hat{x}}_i(t+1)={\mathrm{\Σ}}_{j\∈N_i}{c}_{\mathrm{ij}}\left(t\right)x_j\left(t\right);$

(7) gain matrix of each sensor node is upgraded:

$P_i(t+1)=\underset{j=1}{\overset{n}{\mathrm{\Σ}}}{c}_{\mathrm{ij}}\left(t\right)[(A-L_j\left(t\right)H_j\left(t\right))P_j\left(t\right)(A-L_j\left(t\right)H_j\left(t\right){)}^{\′}+L_j\left(t\right){\mathrm{RL}}_j{\left(t\right)}^{\′}]+Q;$

(8) if t is less than greatest iteration step number step _max, t=t+1, returns step (2); Otherwise algorithm terminates.

4. a kind of Distributed filtering method adjustable based on weights according to claim 3, is characterized in that, described method is under sensor network is in non-trusted environment, and the realization flow of described method is:

(1) estimated value of t-1 moment node i and the estimated value between each neighbor node is utilized to calculate corresponding each internodal support angle value and assignment gives the degree of consistency value between t node i and each neighbor node, namely calculate and upgrade weights p _i(t);

(2) parameter lambda is introduced at Distributed filtering algorithm increment _i, wherein, λ _i=1 represents that node i is not attacked, λ _i=0 represents node V _iattacked; Then:

The each node state estimated value of calculating sensor network:

$x_i\left(t\right)=A{\hat{x}}_i\left(t\right)+{\mathrm{\λ}}_i{L}_i\left(t\right)(z_i\left(t\right)-H_i\left(t\right){\hat{x}}_i\left(t\right));$

The gain matrix upgrading each sensor node is:

$P_i(t+1)={\mathrm{\Σ}}_{j\∈N_i}{c}_{\mathrm{ij}}\left(t\right)[(A-{\mathrm{\λ}}_i{L}_j\left(t\right)H_j\left(t\right))P_j\left(t\right){(A-{\mathrm{\λ}}_i{L}_j\left(t\right)H_j\left(t\right))}^{\′}+{\mathrm{\λ}}_i{L}_j\left(t\right)R{L}_j{\left(t\right)}^{\′}]+Q$

In t, when having node to be attacked in sensor network, by above formula λ _ibe set to 0, mean that this node believes angle value is set to 0, namely make the estimation of any data to whole algorithm of this node cut little ice; When merging renewal, readjust the weights of each neighbor node, and meet

5. a kind of Distributed filtering method adjustable based on weights according to claim 3, is characterized in that, described non-trusted environment, is namely attacked at a certain t sensor node i; If there is multiple sensor node to be attacked, then suppose some nodes of being attacked not neighbor node each other.