CN102056192A - WSN (wireless sensor network) intra-network data fusion method based on kernel density estimation and non-parameter belief propagation - Google Patents

Abstract

The invention discloses a WSN (wireless sensor network) intra-network data fusion method based on kernel density estimation and non-parameter belief propagation, which comprises data acquisition and data fusion. The data acquisition is that monitoring unions which are respectively composed of no less than three sensor nodes for gathering the monitoring data are constructed in a monitoring region, each monitoring union is corresponding provided with a union header node for collecting the monitoring data, the sensor nodes in each monitoring union are respectively used for gathering the monitoring data of an object entering the monitoring region; and the data fusion is that the gathered monitoring data are subjected to KDE (kool desktop environment) processing by the sensor nodes in the monitoring unions respectively, the processed data are transmitted and collected to the union header nodes through NBP (name bind protocol) processing, the collected data are subjected to gauss mixing by the union header nodes, the data after gauss mixing are subjected to Gibbs sampling fusion, and the fused result is acted as a characteristic of the monitoring data. The accuracy of the monitoring data can be improved under a noisy or an uncertain environment, and the accurate fusion characterization of the monitoring data of the multi-node unions can be realized.

Description

Based on data fusion method in the WSN net of Density Estimator and nonparametric belief propagation

Technical field

The present invention relates to the wireless sensor network field, is a kind of based on data fusion method in the WSN net of Density Estimator and nonparametric belief propagation specifically.

Background technology

Wireless sensor network (wireless sensor network, WSN) be a kind of distributed network that forms by the micro radio sensing device node of the low cost of a large amount of random distribution, the low-power consumption mode by self-organizing, the information of monitored target all has broad application prospects in the military and civilian field in its energy real-time perception, collection and the monitoring network overlay area.It is that (Monitoring Data is the basis that the realization target is accurately located tracking to wireless sensor network accurately and effectively for wireless sensor network, important application WSN), and data fusion is the key technology that the WSN target localization is followed the tracks of that target localization is followed the tracks of.Detecting the fusion of level data in the WSN net is meant in WSN multisensor distributed detection system, each sensor node adopts certain data processing method to carry out preliminary treatment earlier to obtain observation, send more representational compressed information to other transducer then, gather at a certain center at last and merge these information generation global detection and adjudicate.WSN constitutes monitoring alliance synergic monitoring by many sensing nodes and improves the target monitoring data accuracy, but the sensing node that is used to monitor is many more, and the network data transmission amount is big more, and consumed energy is big more; And the sensor node collection a large amount of raw data associated of WSN by dense distribution comprise bulk redundancy, invalid and information that confidence level is relatively poor, if these data preliminary treatment are not directly reached aggregation node, not only the network data transmission amount is big, consumed energy is many, and the accuracy of data is collected in influence.In order to improve data transmission efficiency, save network energy, to strengthen and collect data accuracy, must adopt an effective measure data in netting are merged.By improving the Monitoring Data quality, rationally characterize Monitoring Data, make full use of the multisensor monitoring information, each node Monitoring Data of monitoring alliance is carried out proper data merge, can make more accurate and effective of Monitoring Data.

From data circulating form, network node processing mode, data fusion mainly contains centralized, distributed fusion dual mode in the WSN net.Centralized data fusion mode is to send data query by converging (Sink) node, relevant multiple source (Sources) node sends the data to the Sink node, carry out data fusion by the Sink node again, under the WSN node distributes comparatively intensive situation, there is approximate redundant information in a plurality of Sources nodes to the data characterization of same incident, and transmitting redundancy information will consume more multipotency; The distributed data amalgamation mode is that the data that send of Sources node are when intermediate node is transmitted, intermediate node is checked packet content, carry out being sent to the Sink node again after the data fusion, improve the whole efficiency that network data is collected to a certain extent, reduce transmitted data amount, cut down the consumption of energy, the distributed data processing mode is the main mode of WSN data processing.

Different according to data fusion principle and effect, WSN-MTT detects level data blending algorithm except conventional method (as unruly-value rejecting method, thresholding setting method, weighted mean method, adaptive weighted method, least square method, least variance method, maximum-likelihood method), also has emerging intelligent data processing methods such as Kalman filtering, neural net, fuzzy clustering (FCM), the sample analysis of probability density function estimation point.Traditional data processing method principle is simple, and amount of calculation is less, but fusion accuracy is lower; And emerging intelligent data processing arithmetic principle relative complex, but the fusion accuracy height takes the intelligent data processing method may increase computational complexity; The data processing utilization is that node calculates and storage resources in the WSN net, along with processor computational speed and disposal ability improve constantly, net interior Data Fusion to a certain extent, the communication overhead that exchanges high energy consumption with the computational resource of low energy consumption for is a feasible program.Therefore, emerging intelligent data Processing Algorithm is used more and more in WSN data fusion field.The countries in the world scholar carries out big quantity research to emerging intelligent data processing method, wherein Abdel-Aziz A.M (2007) utilizes FCM, by on the corresponding observation space of different sensors institute, setting up the projection of multiple target motion state, single-sensor data association algorithm is generalized in the multi-sensor information fusion system, is implemented in multiobject data association and accurate tracking under the intensive clutter environment.Kong Fantian (2006) utilizes the average clustering algorithm of distributed K-to realize the grouping rationally fast of wireless sensor network node sensing data, in conjunction with based on data fusion method in the adaptive weighted net, the size of data based its respective weights value of node perceived after the grouping is carried out fusion treatment, reduce the network data redundancy, save the storage resources and the network bandwidth.And probability density function method of estimation KDE only has from sampled data itself, do not rely on feature extraction and environmental constraints, can approach the characteristic of arbitrary form density distribution, can carry out accurately robust representation to the node sample data, it is self-align that the Alexander T.Ihler of Massachusetts Institute Technology (2005) is applied to the WSN node to Density Estimator KDE, obtains better effects.In addition, nonparametric belief propagation (Nonparametric Belief Propagation, NBP) have the advantages that be fit to handle distributed computing environment information, the application of achieving success in aspect such as they are followed the tracks of in computer vision, procedure fault detects diagnosis, medical science detects diagnosis, the WSN node is self-align.For the WSN-MTT system, owing to have random noise disturbance, particular physical environment deviation, sensor node perception fragility, measure that inaccuracy, Network Transmission influence etc. are various to be difficult to avoid factor, systematic collection information has many uncertainties.Detection data, NBP handle the WSN multinode and monitor alliance information in the KDE sign WSN net if utilize, and may reduce to measure and disturb and noise effect, and raising Monitoring Data accuracy is for the interior data fusion of WSN net is brought good result.

Summary of the invention

The purpose of this invention is to provide a kind of under noise and uncertain environment, can improve the Monitoring Data accuracy based on data fusion method in the WSN of Density Estimator and the nonparametric belief propagation net.

Provided by the invention a kind of based on data fusion method in the WSN net of Density Estimator and nonparametric belief propagation, this method comprises data acquisition and data fusion; Wherein

Data acquisition is to make up in the monitored area by being no less than 3 monitoring alliances that form as the sensor node of gathering Monitoring Data, each monitoring alliance to should have one be used for compiling Monitoring Data, as leader's node of the leader of alliance, the sensor node in each monitoring alliance is respectively to entering the target collection Monitoring Data of monitored area;

Data fusion is that the sensor node in the monitoring alliance carries out the KDE processing to the Monitoring Data of gathering respectively, data after the processing transmit the leader's node that is pooled to the leader of alliance by NBP, by leader's node aggregated data is carried out Gaussian Mixture, the data after the Gaussian Mixture carried out Gibbs sampling fusion again, the result of fusion is as the sign of Monitoring Data.

In data fusion step:

The algorithm that a, KDE handle Monitoring Data is:

$\hat{p}\left(z_n\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{i}N(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)$

In the formula: z _nBe the target range Monitoring Data of n sensor node acquisition independent noise in the monitoring alliance,

Be Monitoring Data z _nEstimated value, N (z; μ _i, Λ _i) be gaussian kernel function, μ _iBe its average, Λ _iBe the variance of i Gaussian Profile, w _iBe the weight of i Gaussian Profile, M is the Gaussian Profile sum;

The algorithm of b, Gaussian Mixture is:

$B_k=\underset{n=1}{\overset{d}{\mathrm{\Π}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_i{N}_n(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)=\underset{j=1}{\overset{M^d}{\mathrm{\Σ}}}{\stackrel{\text{-}}{w}}_j{N}_j(z;\stackrel{\‾}{\mathrm{\μ}},\stackrel{\‾}{\mathrm{\Λ}})$

In the formula: d is monitoring alliance sensor node number, Be respectively gaussian kernel function

Weights, desired value and variance,

The algorithm that c, Gibbs sampling is merged is:

Suppose to have d Gaussian Mixture information to multiply each other, in each iteration, from one of them Gaussian kernel l of d Gaussian Mixture information _jIn sample, it is constant that other d-1 Gaussian Mixture information is all fixed one of them Gaussian kernel, being multiplied each other by a constant d-1 Gaussian kernel and the Gaussian kernel that obtains of sampling obtains new Gaussian kernel, and the weight of new Gaussian kernel is formed to determine by fixing d-1, l of each iteration renewal _j, carry out k time iteration continuously, by each last l _jCan determine a sample in the Gaussian Mixture product; Carry out dkM ²Inferior Gibbs sampling just obtains product

M sample independently

Above-mentioned sample is averaged, obtains the t fused data sign of target monitoring alliance Monitoring Data constantly.

In the WSN monitored area, sensor node is monitored the formula target that occurs at random, and the monitoring alliance that establishes certain target is c, c={s _n, n=1, L, d}, s _nBe alliance's interior nodes, the interior nodes s of alliance _nObtain the target range observation data z of independent noise _n, utilize KDE that observation data is estimated, the observation data estimated value of establishing node s is

Then its KDE is expressed as

$\hat{p}\left(z\right)=\underset{i}{\mathrm{\Σ}}{w}_i{K}_h(z-z_i)$

Suppose The obedience Gaussian Mixture distributes, and forms kernel function K by M Gaussian Profile addition _h(*) adopts following Gaussian kernel:

K _h(z)＝N(z；μ _i，hI)∝exp(-||z-μ _i|| ²/(2h))

Then:

$\hat{p}\left(z\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{i}N(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)$

μ wherein _iBe i Gaussian Profile average; w _iBe the weight of i Gaussian Profile after the normalization, generally determine according to the measured value size, and

The simplest situation is chosen as

Λ _iBeing the variance of i Gaussian Profile, is bandwidth h here _i, h _iWith rule of thumb (Rule Of Thume, ROT) select:

$\left\{\begin{array}{c}{h}_i=h_{\mathrm{ROT}}\≈1.05{\mathrm{\σ}}^2{N}^{-2/5}\\ {\mathrm{\σ}}^2=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{(z_i-\mathrm{\μ})}^2\\ \mathrm{\μ}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{z}_i\end{array}\right.$

Can see that by above-mentioned each node observation data can be used M gaussian density

Weighting merge and represent.

The monitoring alliance of each target is generally by a leader of alliance and a plurality of sensing node s among the WSN _n(n=1,2 ..., d) to form, d sensing node data in the monitoring alliance are sent to the leader of alliance and merge.

WSN sensing node The data nonparametric belief propagation NBP algorithm compiles fusion.NBP is that a kind of iterative approximation is found the solution probability graph model (Probabilistic Graphical Model, PGM) method of probabilistic inference problem.PGM expresses by graphic model and concerns based on probability correlation.Undirected PGM be expressed as G=(V, E), V={v wherein _sBe node (node) collection, E={ (v _S1, v _S2) represent that sideline (edge) collects between the node.Each node v _sRepresent a stochastic variable x _s, the sideline annexation is represented the influence relation between the stochastic variable between the node.NBP algorithm purpose is to find the solution each node x among the PGM _sThe posteriority conditional probability distribution, it comprises information transmission between the graph model node, to two processes of information partial products weighting.

Obtain WSN target monitoring alliance information above and merged product (Gaussian Mixture product), this product reflection monitoring alliance can obtain sample value and carry out the fusion of monitoring information the target acquisition data cases by the method for Direct Sampling from this Gaussian Mixture product.But in Practical Calculation, the determined Joint Distribution of Gaussian Mixture product is complicated, and the computational complexity of Direct Sampling method can be exponential increase, for reducing computation complexity, the condition distribution of considering each Gaussian kernel simultaneously is easy to obtain, adopt the Gibbs method of sampling below, the independent draws sample comes the estimated information product from the Gaussian Mixture product.

The Gibbs method of sampling is that (Markov chain Monte Carlo, MCMC) method have simple, the fast advantage of computational speed to a kind of most widely used Markov chain Meng Tekaer.Its basic thought is, samples iteratively from full condition distributes, and when iterations is enough big, just can obtain from the sample of uniting the posteriority distribution.Its basic sampling process is actually a vibrational transition process, at first given initial value arbitrarily, alternately upgrades each variable then, by upgrading, transfer to a new state, constantly iterative cycles finally converges on a stable state, thereby obtains the sample from the distribution of associating posteriority.Gibbs sampling hypothesis has d Gaussian Mixture information to multiply each other, in each iteration, from one of them Gaussian kernel of d Gaussian Mixture information, sample, it is constant that other d-1 Gaussian Mixture information is all fixed one of them Gaussian kernel, being multiplied each other by a constant d-1 Gaussian kernel and the Gaussian kernel that obtains of sampling obtains new Gaussian kernel, and the weight of new Gaussian kernel is formed to come definite by fixing d-1.

The Gibbs sampling algorithm is as follows:

1. for each j ∈ [1, d], according to

Select an initial value l _j∈ [1, M];

2. for each k ∈ [1, d] (k ≠ j):

(a) through type (3-16) calculates product

Average μ ^*With variance L ^*

(b) for each

Calculate

Average And variance

Calculate power

Value:

(c) basis Sampling obtains a new l _j

3. will 2. repeat k time iteration;

4. calculate product

Average

And variance

From Middle sampling obtains a sample; Repeat above-mentioned Gibbs sampling, obtain product M sample independently

(5) above-mentioned sample is averaged, obtains the t fused data sign of target monitoring alliance measured value constantly.

The present invention monitors aspects such as alliance's node data characterizes, the propagation of multinode Monitoring Data compiles, multinode Monitoring Data processing method and considers Monitoring Data blending algorithm in design WSN nets from WSN-MTT.

Characterize for WSN-MTT monitoring alliance node data, only have from sampled data itself, do not rely on feature extraction and environmental constraints, can approach the characteristic of arbitrary form density distribution according to KDE, can carry out accurately robust representation to the node sample data, be suitable for representing the sensor node sampled data of the complicated uncertain environment of WSN.

Propagation compiles for the multinode Monitoring Data, utilize the nonparametric conviction to propagate the characteristics that NBP handles distributed computing environment information, the monitoring alliance collaborative perception monitoring information transmission that multinode is constituted is pooled to the leader of alliance, obtain Gaussian Mixture information product the leader of alliance by NBP, reduce to measure and disturb and noise effect, improve the Monitoring Data accuracy.

For the Gaussian Mixture information that obtains the leader of alliance, utilize the characteristics that the Gibbs method of sampling is simple, computational speed is fast, Gaussian Mixture information product is carried out the Gibbs sampling merge, realize that the accurate fusion of multinode alliance Monitoring Data characterizes.

Description of drawings

Fig. 1 is a Monitoring Data blending algorithm flow chart in the WSN net of the present invention;

Fig. 2 is that the WSN monitoring node data NBP of alliance of the present invention transmits schematic diagram;

Fig. 3 is the iterative process of Gibbs sampling among the present invention;

Fig. 4 is an object of experiment movement locus schematic diagram among the present invention;

Fig. 5 is object of experiment crisscross motion track and measurement data when t=19-35s among the present invention;

Fig. 6 tests two kinds of methods to T among the present invention ₁Fusion results relatively;

Fig. 7 tests two kinds of methods to T among the present invention ₂Fusion results relatively;

Fig. 8 tests two kinds of methods to T among the present invention ₃Fusion results relatively;

Fig. 9 tests the average RMSE contrast of two kinds of methods to 3 targets among the present invention;

Figure 10 tests KDE-NBP to 3 target fusion results RMSE curves among the present invention;

Figure 11 tests the FCM method to 3 target fusion results RMSE curves among the present invention.

Embodiment

Provided by the invention a kind of based on data fusion method in the WSN net of Density Estimator and nonparametric belief propagation, this method comprises data acquisition and data fusion; Wherein

Data acquisition is to make up in the monitored area by being no less than 3 monitoring alliances that form as the sensor node of gathering Monitoring Data, each monitoring alliance to should have one be used for compiling Monitoring Data, as leader's node of the leader of alliance, the sensor node in each monitoring alliance is respectively to entering the target collection Monitoring Data of monitored area;

Data fusion is that the sensor node in the monitoring alliance carries out the KDE processing to the Monitoring Data of gathering respectively, data after the processing transmit the leader's node that is pooled to the leader of alliance by NBP, by leader's node aggregated data is carried out Gaussian Mixture, the data after the Gaussian Mixture carried out Gibbs sampling fusion again, the result of fusion is as the sign of Monitoring Data.

In data fusion step:

The algorithm that a, KDE handle Monitoring Data is:

$\hat{p}\left(z_n\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{i}N(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)$

In the formula: z _nBe the target range Monitoring Data of n sensor node acquisition independent noise in the monitoring alliance,

Be Monitoring Data z _nEstimated value, N (z; μ _i, Λ _i) be gaussian kernel function, μ _iBe its average, Λ _iBe the variance of i Gaussian Profile, w _iBe the weight of i Gaussian Profile, M is the Gaussian Profile sum;

The algorithm of b, Gaussian Mixture is:

$B_k=\underset{n=1}{\overset{d}{\mathrm{\Π}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_i{N}_n(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)=\underset{j=1}{\overset{M^d}{\mathrm{\Σ}}}{\stackrel{\text{-}}{w}}_j{N}_j(z;\stackrel{\‾}{\mathrm{\μ}},\stackrel{\‾}{\mathrm{\Λ}})$

In the formula: d is monitoring alliance sensor node number,

Be respectively gaussian kernel function

Weights, desired value and variance,

The algorithm that c, Gibbs sampling is merged is:

Suppose to have d Gaussian Mixture information to multiply each other, in each iteration, from one of them Gaussian kernel l of d Gaussian Mixture information _jIn sample, it is constant that other d-1 Gaussian Mixture information is all fixed one of them Gaussian kernel, being multiplied each other by a constant d-1 Gaussian kernel and the Gaussian kernel that obtains of sampling obtains new Gaussian kernel, and the weight of new Gaussian kernel is formed to determine by fixing d-1, l of each iteration renewal _j, carry out k time iteration continuously, by each last l _jCan determine a sample in the Gaussian Mixture product; Carry out dkM ²Inferior Gibbs sampling just obtains product M sample independently

Above-mentioned sample is averaged, obtains the t fused data sign of target monitoring alliance Monitoring Data constantly.

Fig. 1 is Monitoring Data blending algorithm flow chart in the WSN net.As seen from Figure 1, when WSN detects target and enters into the monitored area, establishing target tracking and monitoring alliance at first, to each monitoring alliance, alliance's interior nodes continuous several times detection of a target carries out KDE to measurement data and handles, and then the KDE data is sent to the leader of alliance based on the NBP nonparametric, by the leader of alliance the monitoring alliance information is carried out fusions of sample of Gaussian Mixture, Gibbs, the result of fusion nets the sign of interior multisensor node detection data as WSN.

The Monitoring Data fusion method is accurately to characterize WSN-MTT systematic sampling data, by NBP multinode monitoring alliance collaborative perception monitoring information is propagated and compiled, by the Gibbs method of sampling NBP aggregated data carried out Gaussian Mixture Gibbs sampling and merge by KDE in the WSN net, realizes the accurate fusion of multinode alliance Monitoring Data.

Follow the tracks of for the WSN target monitoring, Fig. 2 has described the k NBP transport process of certain target monitoring alliance node observation data constantly.Scheme at the middle and upper levels node and represent k target monitoring alliance node constantly, each node measured value is respectively z ₁, z ₂, L z _d, lower floor's node is the leader of alliance in continuous two moment.

According to the NBP algorithm principle, be engraved in the product of the information fusion of leader's node during objective definition k for all local evidences of this moment (KDE of monitoring alliance node measurement value estimates), then the information fusion of k moment leader's node can be expressed as

$B_k=\underset{n=1}{\overset{d}{\mathrm{\Π}}}m\left(z_n\right)=\underset{n=1}{\overset{d}{\mathrm{\Π}}}\hat{p}\left(z_n\right)$

The information fusion product that can get leader's node is:

$B_k=\underset{n=1}{\overset{d}{\mathrm{\Π}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_i{N}_n(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)$

More than two formulas represent that the information fusion product at the leader of alliance place is to be multiplied each other definite (being called the Gaussian Mixture product) by d Gaussian Mixture distribution, the result of Gaussian Mixture product is M ^dThe individual Gaussian Profile sum of products, wherein each Gaussian Profile product term is multiplied each other by d Gaussian Profile and gets.D Gaussian Profile result of product remains a Gaussian Profile, establishes this Gaussian Profile and is

Then

$\underset{n=1}{\overset{d}{\mathrm{\Π}}}{N}_n(z;{\mathrm{\μ}}_n,{\mathrm{\Λ}}_n)\∝N(z;\stackrel{\‾}{\mathrm{\μ}},\stackrel{\‾}{\mathrm{\Λ}})$

Its weights, bandwidth and desired value are determined by following formula:

$\stackrel{\‾}{w}=\frac{\underset{n=1}{\overset{d}{\mathrm{\Π}}}{w}_{n}N(z;{\mathrm{\μ}}_n,{\mathrm{\Λ}}_n)}{N(z;\stackrel{\‾}{\mathrm{\μ}},\stackrel{\‾}{\mathrm{\Λ}})}$ ${\stackrel{\‾}{\mathrm{\Λ}}}^{-1}=\underset{n=1}{\overset{d}{\mathrm{\Σ}}}{{\mathrm{\Λ}}_n}^{-1}$ ${\stackrel{\‾}{\mathrm{\Λ}}}^{-1}{\stackrel{\‾}{\mathrm{\μ}}}^{-1}=\underset{n=1}{\overset{d}{\mathrm{\Σ}}}{{\mathrm{\Λ}}_n}^{-1}{\mathrm{\μ}}_n$

In the formula:

Be node s _nMonitoring Data shared weight in the information product, value and node s _nAnd distance is inversely proportional between the target, and

Fig. 3 has described the iterative process of Gibbs sampling from 3 Gaussian Mixture information (each information is made up of 4 Gaussian Profile additions) product.

Among Fig. 3, d=3, M=4, Msg ₁, Msg ₂, Msg ₃The expression Gaussian Mixture information that need merge for convenience, is used l respectively _jValue represent the sequence number of each Gaussian kernel in j the Gaussian Mixture information, as l ₁The 1st Gaussian kernel in the 1st Gaussian Mixture information of=1 expression, l ₂The 3rd Gaussian kernel in the 2nd Gaussian Mixture information of=3 expressions.The Gibbs sampling is in each iteration, Gaussian kernel in fixing wherein 2 Gaussian Mixture information is constant, from a Gaussian kernel of other 1 Gaussian Mixture information, sample, the Gaussian kernel that is obtained by constant 2 Gaussian kernel and sampling multiplies each other and obtains new Gaussian kernel, and the weight of new Gaussian kernel (arrow is represented among figure (a) and (b), (c)) is formed (solid arrow is represented among figure (a) and (b), (c)) by fixing 2 and come definite.As Fig. 3 (a): l ₂=1, l ₃=4, i.e. fixing Msg ₂And Msg ₃In the 1st and 4 Gaussian kernel, Msg ₁Respectively from l ₁=1,2,3,4 condition extracts in distributing, and the new Gaussian kernel weight after multiplying each other is by fixing Msg ₂And Msg ₃In the decision of the 1st and 4 Gaussian kernel.Each iteration is upgraded a l _j, carry out k time iteration continuously, by each last l _jCan determine a sample in the Gaussian Mixture product.Like this, carry out dkM ²Inferior Gibbs sampling just obtains M the sample from the Gaussian Mixture product.Number of iterations is many more, and the sample that sampling obtains is accurate more, but needs higher relatively assessing the cost.Shown in Fig. 3 (d) is through to k iteration of each Gaussian Mixture information, and the Gaussian kernel information (shown in the solid line) in each definite Gaussian Mixture, these information are one of compositions of Gauss's product (among Fig. 3 (e) not the solid line of overstriking).

Experimental example:

Below the present invention's (KDE-NBP) detection data anastomosing algorithm performance is carried out emulation experiment.Emulation is carried out under the MATLAB7.8 programmed environment on a PC, and machines configurations is a Windows XP professional operating system, Intel (R) Core (TM) 2CPU, and T5200@1.60GHz, the 1G internal memory, the 80G hard disk, dominant frequency is 1.60GHz.The emulation experiment parameter is provided with as shown in table 1.

Table 1 simulated environment and parameter setting

Being located at has three moving targets in the monitored area, moving target selects two classical motion models of widely quoting ^{[59,69,79,85 ,]}(one is simple linear model, and one is the strong nonlinearity model) and a complex nonlinear model that designs voluntarily, three targets were repeatedly intersected in 50 sampling periods, and its state equation is respectively

T ₁：x ₁＝x _t-1+[0.5；0.5x _t-1(2，1)/(1+(x _t-1(2，1)) ²)]+[0；2cos(1.2(t-1)]+w _t

$T_2:x_t=x_{t-1}+[5\cos \left(0.8(t-1)\right);5\sin \frac{p}{4t}]+w_t$

T ₃：x _t＝x _t-1+[0.5；0.5]+w _t

In the formula: w _tBe state-noise.

Suppose that three targets are the formula target, its initial position is respectively (8,15), (25,6), (7,2), and the sensor node in the target monitoring radius can both record the noise distance measure of target:

$z_t=\left|\right|x_t-s_t\left|\right|+v_t,v_t~N(v;0,{\mathrm{\σ}}_v^2)$

Wherein || x _t-s _t‖ is the distance between t moment target and the monitoring sensor node, v _tFor measuring noise.

Use based on the detection data anastomosing algorithm of KDE-NBP the measured value of target is merged estimation.For contrast effect, adopt FCM data anastomosing algorithm and this paper to propose algorithm simultaneously three moving target sensing datas are experimentized.The fusion accuracy of two kinds of algorithms, the performance of data loss rate aspect are compared in emulation.

Fig. 4 is the target trajectory schematic diagram.As seen from Figure 4, three moving target repeatedly crisscross motions in 50 sampling periods, existing simple linear movement has complicated nonlinear motion again.Fig. 5 is the repeatedly track and the measurement data of (t=19-35) during the crisscross motion of target, compare for the effect that adopts KDE-NBP, FCM method that 3 targets are carried out data fusion Fig. 6～8, Fig. 9 is the average RMSE contrasts of two kinds of fusion methods to three targets, Figure 10～11 are respectively the fusion RMSE curve of two kinds of methods to three targets, and table 2 is that two kinds of methods merge performance relatively.

By Fig. 6～11, table 2 as seen, no matter be linear goal or complex nonlinear target, no matter be target range apart from each other or nearer apart, the data fusion precision of KDE-NBP method is generally than FCM method height; Two kinds of methods are respectively 0.4585m, 0.6954m at the average mean square error RMSE of the fusion in 50 moment, and the KDE-NBP method is compared with the FCM method, and fusion accuracy on average improves 34.1%.In whole experiment, the KDE-NBP method is not found to lose with phenomenon, and FCM method fusion results is extremely unstable, and each experiment all exists to lose follows, and the average mistake of three targets is 4.6% with rate; The KDE-NBP data fusion method that the present invention proposes, taking into full account the noise information that the WSN-MTT system exists at random disturbs, the deviation that particular physical environment is brought, the inaccuracy of measuring, the fragility of sensor node perception, the various factors that are difficult to avoid such as the influence of Network Transmission, the information of systematic collection has many probabilistic real world applications environment, utilize KDE only from sampled distance data itself, do not rely on the restriction of Feature Extraction and environment, can approach the characteristics of the density distribution of arbitrary form, accurately characterize the sampled data of WSN-MTT system; Simultaneously, compile monitoring alliance's data and Gibbs fusion based on NBP, made full use of multisensor synergic monitoring information, overcome bursty interference, eliminate and lose with phenomenon, the present invention has obtained good Monitoring Data syncretizing effect.

The comparison of two kinds of method data fusion of table 2 performance

Claims (2)

1. data fusion method in the WSN based on Density Estimator and nonparametric belief propagation nets, this method comprises data acquisition and data fusion, it is characterized in that:

Data acquisition is to make up in the monitored area by being no less than 3 monitoring alliances that form as the sensor node of gathering Monitoring Data, each monitoring alliance to should have one be used for compiling Monitoring Data, as leader's node of the leader of alliance, the sensor node in each monitoring alliance is respectively to entering the target collection Monitoring Data of monitored area;

Data fusion is that the sensor node in the monitoring alliance carries out the KDE processing to the Monitoring Data of gathering respectively, data after the processing are handled by NBP and are transmitted the leader's node that is pooled to the leader of alliance, by leader's node aggregated data is carried out Gaussian Mixture, the data after the Gaussian Mixture carried out Gibbs sampling fusion again, the result of fusion is as the sign of Monitoring Data.

2. according to claim 1 based on data fusion method in the WSN net of Density Estimator and nonparametric belief propagation, it is characterized in that in data fusion step:

The algorithm that a, KDE handle Monitoring Data is:

$\hat{p}\left(z_n\right)=\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_{i}N(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)$

In the formula: z _nBe the target range Monitoring Data of n sensor node acquisition independent noise in the monitoring alliance,

Be Monitoring Data z _nEstimated value, N (z; μ _i, Λ _i) be gaussian kernel function, μ _iBe its average, Λ _iBe the variance of i Gaussian Profile, w _iBe the weight of i Gaussian Profile, M is the Gaussian Profile sum;

The algorithm of b, Gaussian Mixture is:

$B_k=\underset{n=1}{\overset{d}{\mathrm{\Π}}}\underset{i=1}{\overset{M}{\mathrm{\Σ}}}{w}_i{N}_n(z;{\mathrm{\μ}}_i,{\mathrm{\Λ}}_i)=\frac{M^d}{j=1}{\stackrel{\‾}{w}}_j{N}_j(z;\stackrel{\‾}{\mathrm{\μ}},\stackrel{\‾}{\mathrm{\Λ}})$

In the formula: d is monitoring alliance sensor node number,

Be respectively gaussian kernel function Weights, desired value and variance,

The algorithm that c, Gibbs sampling is merged is:

Suppose to have d Gaussian Mixture information to multiply each other, in each iteration, from one of them Gaussian kernel l of d Gaussian Mixture information _jIn sample, it is constant that other d-1 Gaussian Mixture information is all fixed one of them Gaussian kernel, being multiplied each other by a constant d-1 Gaussian kernel and the Gaussian kernel that obtains of sampling obtains new Gaussian kernel, and the weight of new Gaussian kernel is formed to determine by fixing d-1, l of each iteration renewal _j, carry out k time iteration continuously, by each last l _jCan determine a sample in the Gaussian Mixture product; Carry out dkM ²Inferior Gibbs sampling just obtains product

M sample independently

Above-mentioned sample is averaged, obtains the t fused data sign of target monitoring alliance Monitoring Data constantly.

Images