CN106156790A - A kind of distributed collaborative algorithm being applied to sensor network and data syncretizing mechanism - Google Patents

Abstract

nullThe present invention relates to sensing data processing technology field，A kind of distributed collaborative algorithm being applied to sensor network and data syncretizing mechanism，Including sensor-level data fusion and data fusion at central level，Method is: sensor-level data fusion realizes the data prediction to sensor acquisition and extracts the different characteristic amount of system under test (SUT)，The system under test (SUT) characteristic quantity of data fusion at central level integration sensor-level data fusion obtains the state estimation of system under test (SUT) more accurately，The method of the present invention uses the result of the modified result sensor-level data fusion of data fusion at central level，Method for Single Sensor System precision can be solved low、The problem that information is single and data fusion is computationally intensive，Can be widely applied to multi-source image be combined、Robot and communication in intelligent instrument system、Battlefield and UAV、Graphical analysis and understanding、Object detecting and tracking、In automatic target detection.

Description

A kind of distributed collaborative algorithm being applied to sensor network and data syncretizing mechanism

Technical field

The present invention relates to sensing data processing technology field, a kind of distributed association being applied to sensor network Make algorithm and data syncretizing mechanism.

Background technology

The information that multiband information or the different classes of sensor of single-sensor are provided by data fusion is in addition comprehensive, Eliminate redundancy that may be present and contradiction between multi-sensor information, in addition complementary, improve promptness that remote sensing information extracts and Reliability, improves the service efficiency of data.Data fusion is applied in military target identification field the earliest, the most extensively should Be used in that multi-source image is compound, robot and communication in intelligent instrument system, battlefield and UAV, graphical analysis and understanding, target Detect and in tracking, automatic target detection.

Existing Data fusion technique is usually sensor-level data fusion result for data fusion process at central level, and The result of data fusion at central level is directly as final fusion results, and this technology belongs to the Data fusion technique of open loop, tool Have that fusion results precision is low, the shortcoming of poor reliability.

Summary of the invention

Present invention aim to address that in data fusion process, data fusion at central level cannot be used for revising sensor progression The fusion accuracy caused according to the result merged is low, the problem of poor reliability, it is proposed that a kind of distribution being applied to sensor network Formula cooperation algorithm and data syncretizing mechanism, mainly include (1) sensor-level data fusion；(2) data fusion at central level.

Described sensor-level data fusion specifically includes following steps:

(21). obtained by sensor and sufficiently measure initial data, and reject the abnormal data in measurement, for non-thread Property sensor measurement model, model can be made to carry out linearisation near certain nominal value, the most again by linear measurement model Solve；

(22). assuming that linear sensor measurement model is s=H θ, H is the input quantity of sensor, H be a full rank N × The observing matrix (N ＞ P) of P, H is made up of observation sample, if s=[s [0], s [1] ..., s [N-1]]^TIt it is the model of sensor Output, s is the matrix of N × 1, x=[x [0], x [1] ..., x [N-1]]^TBeing the measurement amount of sensor output, θ is biography to be asked The parameter model of sensor, θ is the matrix of P × 1；

(23). obtain measurement amount x of sensor output by linear weighted function mode and passed by linear sensor measurement model Difference J (θ) of the s that sensor linear measurement model solution goes out, if W is the diagonal matrix of N × N, W=C^-1, C represents zero mean noise Covariance matrix；

$J\left(\mathrm{\θ}\right)=\underset{n=0}{\overset{N-1}{\Σ}}w\left(n\right){\left(x\right(n)-s(n\left)\right)}^2={(x-H\mathrm{\θ})}^{T}W(x-H\mathrm{\θ})={(x-H\mathrm{\θ})}^T{C}^{-1}(x-H\mathrm{\θ})$

(24). solve J (θ) function gradient to θ variable；

$\frac{\∂J\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}=-2H^{T}Wx+2H^{T}WH\mathrm{\θ}=-2H^T{C}^{-1}x+2H^T{C}^{-1}H\mathrm{\θ}$

(25). make the gradient of J (θ) function be equal to 0, can be in the hope of the estimated value of linear sensor measurement model

$\widehat{\mathrm{\θ}}={\left(H^{T}WH\right)}^{-1}{H}^{T}Wx={\left(H^T{C}^{-1}H\right)}^{-1}{H}^T{C}^{-1}x$

(26) if. during Ce Lianging, noise is incoherent, then C estimator is diagonal matrix, makes C [n]=diag (σ₀, σ₁..., σ_n),Wherein H [n] is the matrix of n × p, h^T[n] is the matrix of 1 × p, X [n]=[x [0], x [1] ..., x [n]]^T, then have estimatorC [n] is to make an uproar The covariance matrix of sound, least squares estimatorCovariance matrix be

(27). calculate estimatorRenewal with the covariance ∑ [n] of least squares estimator；

The renewal of estimator is calculated as follows:

Wherein

The renewal of the covariance of least squares estimator is as follows:

∑ [n]=(I-K [n] h^T[n])∑[n-1]。

(28). the state estimation that heterogeneity or same nature sensor are obtained is carried out Dempster-Shafer number Initial condition is provided according to being fused to data fusion at central level.

Described data fusion at central level comprises the following steps:

(31). set up state equation and the measurement equation of the system that is observed.

If system state equation is x_k=Ax_k-1+Bu_k-1+w_k-1, measurement equation is y_k=Cx_k-1+v_k, wherein x_kBe by (x_{1, k}, x_{2, k}..., x_{N, k}) composition N × 1 state variable, u_kIt is (u_{1, k}, u_{2, k}... u_{L, k}) input vector of L × 1 that forms, y_kBy (y_{1, k}, y_{2, k}... y_{M, k}) observation vector of M × 1 that forms, w_kIt is by (w_{1, k}, w_{2, k}... w_{N, k}) composition N × 1 process Noise vector, v_kIt is by (v_{1, k}, v_{2, k}... v_{M, k}) the measurement noise vector of M × 1 that forms, A is the state transfer square of N × N Battle array, B is the input relational matrix of N × L, and C is that M × N measures relational matrix；

(32). on the basis of-1 state variable optimal estimation value of kth, utilize the state equation of system to predict kth The prior state value in moment.

P_k ^-=AP_k-1A^T+ Q, whereinRepresent the priori estimates of kth moment state variable,Represent the optimal estimation value of kth-1 moment state variable,Represent kth moment priori to estimate The covariance matrix of meter error,Represent the covariance matrix of kth moment Posterior estimator error, Q=E [w_kw_k ^T] represent the covariance that process noise is vectorial.

(33). utilize actual measured value correction previous step to predict the state priori estimates obtainedObtain system mode Optimal estimation value

K_k=P_k ^-C^T(CP_k ^-C^T+R)^-1,P_k=(I-K_kC)P_k ^-, wherein R=E [v_kv_k ^T] represent Measure the covariance of noise vector.

(34). in order to avoid matrix inversion operation, use scalar facture successively to process, it is assumed that each measures noise Between relevant, carry out Orthogonal Decomposition, R=MDM to measuring noise covariance R^T, wherein M is orthogonal matrix, D=diag (σ₁, σ₂..., σ_M) it is diagonal matrix, measurement equation is carried out matrixing M^Ty_k=M^TCx_k-1+M^Tv_k, define new observation vector y_k'= M^Ty_k, calculation matrix C '=M^TC, measures noise vector v_k'=M^Tv_k, the new covariance measuring noise vector is R=E [v_k′v_k ′^T]=E [M^Tv_kv_k ^TM]=M^TRM=D, remembers calculation matrixy_k'=(y_{1, k}', y_{2, k}' ..., y_{M, k}'),P_{0, k}=P_k ^-, then can carry out state with following scalar algorithm and be corrected with mean squared error matrix:P_{M, k}=(I-k_{M, k}c_m′)P_{M-1, k}, wherein m =1,2 ..., M,P_k=P_{M, k}。

(35). use data fusion result at central level to go to revise the result of sensor-level data fusion.

The beneficial outcomes that the present invention brings: the present invention uses the modified result sensor-level data of data fusion at central level to melt That close as a result, it is possible to solve the problem that Method for Single Sensor System precision is low, information is single and data fusion is computationally intensive, can be extensive Be applied to that multi-source image is compound, robot and communication in intelligent instrument system, battlefield and UAV, graphical analysis and understanding, mesh Mark detects and in tracking, automatic target detection.

Accompanying drawing explanation

The present invention is further detailed explanation with detailed description of the invention below in conjunction with the accompanying drawings.

Fig. 1 is data fusion the general frame of the present invention.

Fig. 2 is the flow chart of sensor-level data fusion.

Fig. 3 is the flow chart that sensor measurement model is set up.

Fig. 4 is the flow chart rejecting sensor abnormality measurement data.

Fig. 5 is the flow chart of data fusion at central level.

Detailed description of the invention

Fig. 1 is data fusion overall pattern of the present invention, and wherein sensor-level data fusion is distributed in different tested by fusion The different classes of sensor of object obtains being observed the measurement noise variance of the Partial State Information of system and sensor, at central level Data fusion provides initial condition, and data fusion at central level passes through the Fuzzy Kalman Filter result to sensor-level data fusion It is filtered processing the status information of the system under test (SUT) obtaining optimum, uses the modified result sensor-level of data fusion at central level The result of data fusion.

Fig. 2 is the flow chart of sensor-level data fusion, and wherein same tested state is measured by homogeneity sensor, carries The credibility of high tested state, different tested states is measured by Heterogeneous Sensor, it is provided that more fully system mode letter The system state data that breath, homogeneity sensor and Heterogeneous Sensor are measured is central authorities by Dempster-Shafer data fusion DBMS merges provides initial condition.Sensor-level data fusion specifically includes following steps:

(21). initially setting up the measurement model of sensor, particular flow sheet is shown in Fig. 3, for nonlinear sensor measurement Model, can make model carry out linearisation near certain nominal value, solve by linear measurement model, by passing Sensor obtains and sufficiently measures initial data, and rejects the abnormal data in measurement.For time-invariant system, sensor abnormality is surveyed The idiographic flow that amount data are rejected is shown in Fig. 4, can first obtain average α and the variances sigma of tested state, then by sensor measurement number Average again and variance after rejecting according to the data outside interval (α-3 σ, α+3 σ)；

(22). assuming that linear sensor measurement model is s=H θ, H is the input quantity of sensor, H be a full rank N × The observing matrix (N ＞ P) of P, H is made up of observation sample, if s=[s [0], s [1] ..., s [N-1]]^TIt it is the model of sensor Output, s is the matrix of N × 1, x=[x [0], x [1] ..., x [N-1]]^TBeing the measurement amount of sensor output, θ is biography to be asked The parameter model of sensor, θ is the matrix of P × 1；

(23). obtain measurement amount x of sensor output by linear weighted function mode and passed by linear sensor measurement model Difference J (θ) of the s that sensor linear measurement model solution goes out, if W is the diagonal matrix of N × N, W=C^-1, C represents zero mean noise Covariance matrix；

$J\left(\mathrm{\θ}\right)=\underset{n=0}{\overset{N-1}{\Σ}}w\left(n\right){\left(x\right(n)-s(n\left)\right)}^2={(x-H\mathrm{\θ})}^{T}W(x-H\mathrm{\θ})={(x-H\mathrm{\θ})}^T{C}^{-1}(x-H\mathrm{\θ})$

(24). solve J (θ) function gradient to θ variable；

$\frac{\∂J\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}=-2H^{T}Wx+2H^{T}WH\mathrm{\θ}=-2H^T{C}^{-1}x+2H^T{C}^{-1}H\mathrm{\θ}$

(25). make the gradient of J (θ) function be equal to 0, can be in the hope of the estimated value of linear sensor measurement model

$\widehat{\mathrm{\θ}}={\left(H^{T}WH\right)}^{-1}{H}^{T}Wx={\left(H^T{C}^{-1}H\right)}^{-1}{H}^T{C}^{-1}x$

(26) if. during Ce Lianging, noise is incoherent, then C estimator is diagonal matrix, makes C [n]=diag (σ₀, σ₁..., σ_n),Wherein H [n] is the matrix of n × p, h^T[n] is the matrix of 1 × p, X [n]=[x [0], x [1] ..., x [n]]^T, then have estimatorC [n] is to make an uproar The covariance matrix of sound, least squares estimatorCovariance matrix be

(27). calculate estimatorRenewal with the covariance ∑ [n] of least squares estimator；

The renewal of estimator is calculated as follows:

Wherein

The renewal of the covariance of least squares estimator is as follows:

∑ [n]=(I-K [n] h^T[n])∑[n-1]。

Fig. 5 is the flow chart of data fusion at central level, specifically includes following steps:

(31). set up state equation and the measurement equation of the system that is observed, for the nonlinear state being observed system Equation and measurement equation, the method migration that can use local linearization is that linear system solves.

If system state equation is x_k=Ax_k-1+Bu_k-1+w_k-1, measurement equation is y_k=Cx_k-1+v_k, wherein x_kBe by (x_{1, k}, x_{2, k}..., x_{N, k}) composition N × 1 state variable, u_kIt is (u_{1, k}, u_{2, k}... u_{L, k}) input vector of L × 1 that forms, y_kBy (y_{1, k}, y_{2, k}... y_{M, k}) observation vector of M × 1 that forms, w_kIt is by (w_{1, k}, w_{2, k}... w_{N, k}) composition N × 1 process Noise vector, v_kIt is by (v_{1, k}, v_{2, k}... v_{M, k}) the measurement noise vector of M × 1 that forms, A is the state transfer square of N × N Battle array, B is the input relational matrix of N × L, and C is that M × N measures relational matrix；

(32). on the basis of-1 state variable optimal estimation value of kth, utilize the state equation of system to predict kth The prior state value in moment.

P_k ^-=AP_k-1A^T+ Q, whereinRepresent the priori estimates of kth moment state variable,Represent the optimal estimation value of kth-1 moment state variable,Represent kth moment priori to estimate The covariance matrix of meter error,Represent the covariance matrix of kth moment Posterior estimator error, Q=E [w_kw_k ^T] represent the covariance that process noise is vectorial.

(33). utilize actual measured value correction previous step to predict the state priori estimates obtainedObtain system mode Optimal estimation value

K_k=P_k ^-C^T(CP_k ^-C^T+R)^-1,P_k=(I-K_kC)P_k ^-, wherein R=E [v_kv_k ^T] represent Measure the covariance of noise vector.

(34). in order to avoid matrix inversion operation, use scalar facture successively to process, it is assumed that each measures noise Between relevant, carry out Orthogonal Decomposition, R=MDM to measuring noise covariance R^T, wherein M is orthogonal matrix, D=diag (σ₁, σ₂..., σ_M) it is diagonal matrix, measurement equation is carried out matrixing M^Ty_k=M^TCx_k-1+M^Tv_k, define new observation vector y_k'= M^Ty_k, calculation matrix C '=M^TC, measures noise vector v_k'=M^Tv_k, the new covariance measuring noise vector is R=E [v_k′v_k ′^T]=E [M^Tv_kv_k ^TM]=M^TRM=D, remembers calculation matrixy_k'=(y_{1, k}', y_{2, k}' ..., y_{M, k}'),P_{0, k}=P_k ^-, then can carry out state with following scalar algorithm and be corrected with mean squared error matrix:P_{M, k}=(I-k_{M, k}c_m′)P_{M-1, k}, wherein m =1,2 ..., M,P_k=P_{M, k}。

(35). use data fusion result at central level to go to revise the result of sensor-level data fusion.

Those of ordinary skill in the art it will be appreciated that embodiment described here be to aid in reader understanding this Bright implementation, it should be understood that protection scope of the present invention is not limited to such special statement and embodiment.Ability The those of ordinary skill in territory can make various its without departing from essence of the present invention according to these technology disclosed by the invention enlightenment Its various concrete deformation and combination, these deformation and combination are the most within the scope of the present invention.

Claims (3)

1. the distributed collaborative algorithm being applied to sensor network and data syncretizing mechanism, it is characterised in that include sensing Device DBMS merges and data fusion at central level, and wherein sensor-level data fusion is distributed in different measurand not by fusion Generic sensor obtains being observed the measurement noise variance of the Partial State Information of system and sensor, melts for data at central level Closing and provide initial condition, the result of sensor-level data fusion is filtered by data fusion at central level by Fuzzy Kalman Filter Ripple processes the status information of the system under test (SUT) obtaining optimum, uses the modified result sensor-level data of data fusion at central level to melt The result closed.

2. according to a kind of distributed collaborative algorithm being applied to sensor network described in claim 1 and data syncretizing mechanism, It is characterized in that, described sensor-level data fusion comprises the following steps:

(21). obtained by sensor and sufficiently measure initial data, and reject the abnormal data in measurement, for nonlinear Sensor measurement model, can make model carry out linearisation near certain nominal value, carry out by linear measurement model Solve；

(22). assuming that linear sensor measurement model is s=H θ, H is the input quantity of sensor, and H is the N × P of a full rank Observing matrix (N ＞ P), H is made up of observation sample, if s=[s [0], s [1] ..., s [N-1]]^TThe model being sensor is defeated Going out, s is the matrix of N × 1, x=[x [0], x [1] ..., x [N-1]]^TBeing the measurement amount of sensor output, θ is sensing to be asked The parameter model of device, θ is the matrix of P × 1；

(23). obtain measurement amount x of sensor output and by linear sensor measurement model sensor by linear weighted function mode Difference J (θ) of the s that linear measurement model solution goes out, if W is the diagonal matrix of N × N, W=C^-1, C represents the association side of zero mean noise Difference matrix；

$J\left(\mathrm{\θ}\right)=\underset{n=0}{\overset{N-1}{\mathrm{\Σ}}}w\left(n\right){\left(x\right(n)-s(n\left)\right)}^2={(x-H\mathrm{\θ})}^{T}W(x-H\mathrm{\θ})={(x-H\mathrm{\θ})}^T{C}^{-1}(x-H\mathrm{\θ})$

(24). solve J (θ) function gradient to θ variable；

$\frac{\∂J\left(\mathrm{\θ}\right)}{\∂\mathrm{\θ}}=-2H^{T}Wx+2H^{T}WH\mathrm{\θ}=-2H^T{C}^{-1}x+2H^T{C}^{-1}H\mathrm{\θ}$

(25). make the gradient of J (θ) function be equal to 0, can be in the hope of the estimated value of linear sensor measurement model

$\widehat{\mathrm{\θ}}={\left(H^{T}WH\right)}^{-1}{H}^{T}Wx={\left(H^T{C}^{-1}H\right)}^{-1}{H}^T{C}^{-1}x$

(26) if. during Ce Lianging, noise is incoherent, then C estimator is diagonal matrix, makes C [n]=diag (σ₀, σ₁..., σ_n),Wherein H [n] is the matrix of n × p, h^T[n] is the matrix of 1 × p, X [n]=[x [0], X [1] ..., x [n]]^T, then have estimatorC [n] is the association of noise Variance matrix, least squares estimatorCovariance matrix be

(27). calculate estimatorRenewal with the covariance ∑ [n] of least squares estimator；

The renewal of estimator is calculated as follows:

Wherein

The renewal of the covariance of least squares estimator is as follows:

∑ [n]=(I-K [n] h^T[n])∑[n-1]。

(28). the state estimation that heterogeneity or same nature sensor are obtained is carried out Dempster-Shafer data and melts It is combined into data fusion at central level and initial condition is provided.

3. according to a kind of distributed collaborative algorithm being applied to sensor network described in claim 1 and data syncretizing mechanism, It is characterized in that, described data fusion at central level comprises the following steps:

(31). set up state equation and the measurement equation of the system that is observed.

If system state equation is x_k=Ax_k-1+Bu_k-1+w_k-1, measurement equation is y_k=Cx_k-1+v_k, wherein x_kIt is by (x_{1, k}, x_{2, k}... x_{N, k}) composition N × 1 state variable, u_kIt is (u_{1, k}, u_{2, k}... u_{L, k}) input vector of L × 1 that forms, y_kBy (y_{1, k}, y_{2, k}... y_{M, k}) observation vector of M × 1 that forms, w_kIt is by (w_{1, k}, w_{2, k}... w_{N, k}) composition N × 1 process noise Vector, v_kIt is by (v_{1, k}, v_{2, k}... v_{M, k}) the measurement noise vector of M × 1 that forms, A is the state-transition matrix of N × N, and B is The input relational matrix of N × L, C is that M × N measures relational matrix；

(32). on the basis of-1 state variable optimal estimation value of kth, utilize the state equation of system to predict the kth moment Prior state value.

P_k ^-=AP_k-1A^T+ Q, whereinRepresent the priori estimates of kth moment state variable,Represent The optimal estimation value of kth-1 moment state variable,Represent kth moment prior estimate error Covariance matrix,Represent the covariance matrix of kth moment Posterior estimator error, Q=E [w_kw_k ^T] represent the covariance that process noise is vectorial.

(33). utilize actual measured value correction previous step to predict the state priori estimates obtainedObtain the optimum of system mode Estimated value

K_k=P_k ^-C^T(CP_k ^-C^T+R)^-1,P_k=(I-K_kC)P_k ^-, wherein R=E [v_kv_k ^T] represent and measure The covariance of noise vector.

(34). in order to avoid matrix inversion operation, use scalar facture successively to process, it is assumed that each is measured between noise Relevant, carry out Orthogonal Decomposition, R=MDM to measuring noise covariance R^T, wherein M is orthogonal matrix, D=diag (σ₁, σ₂..., σ_M) it is diagonal matrix, measurement equation is carried out matrixing M^Ty_k=M^TCx_k-1+M^Tv_k, define new observation vector y_k'=M^Ty_k, survey Moment matrix C '=M^TC, measures noise vector v_k'=M^Tv_k, the new covariance measuring noise vector is R=E [v_k′v_k′^T]=E [M^Tv_kv_k ^TM]=M^TRM=D, remembers calculation matrixy_k'=(y_{1, k}', y_{2, k}' ..., y_{M, k}'), P_{0, k}=P_k ^-, then can carry out state with following scalar algorithm and be corrected with mean squared error matrix:P_{M, k}=(I-k_{M, k}c_m′)P_{M-1, k}, wherein m =1,2 ..., M,P_k=P_{M, k}。

(35). use data fusion result at central level to go to revise the result of sensor-level data fusion.