Disclosure of Invention
The invention aims to provide a multi-source information fusion method based on weight pre-allocation, which eliminates the influence of adverse factors such as sensor errors, monitoring data loss, deviation and the like on a fusion result, establishes a distributed fusion framework based on weight pre-allocation and improves the fusion accuracy of sensors.
The technical scheme adopted by the invention is that a multi-source information fusion method based on weight pre-distribution is implemented according to the following steps:
step 1, selecting a Q test method to remove abnormal monitoring data of each sensor aiming at a system containing a plurality of sensors, and pre-distributing fusion weight to the data after removing abnormal values based on a distance criterion;
step 2, selecting a self-attenuation unscented Kalman filter UKF based on the Mahalanobis distance as a local state estimator, evaluating the squared Mahalanobis distance of the innovation vector, and taking corresponding measures to improve the adaptability and robustness of the UKF to the modeling error of the multi-sensor nonlinear random system to obtain a local state estimation result;
and 3, fusing the multi-sensor monitoring data based on a minimum variance linear weighting criterion to obtain a global state estimation result.
The present invention is also characterized in that,
the step 1 is as follows:
step 1.1, setting the state space model of each sensor to meet the following form:
in the formula, xtAnd xt+1Respectively representing the state values of the tested system at the time t and the time t + 1; f (-) is the system nonlinear state function; w is atIs zero mean Gaussian white noise with variance Q more than or equal to 0; z is a radical oft+1Is the measured value at the moment of the sensor t + 1; h (-) is a sensor nonlinear measurement function; e.g. of the typet+1Zero mean Gaussian white noise with variance R not less than 0 at the moment of t + 1;
step 1.2, taking the monitoring data of each sensor at the moment t as an example, setting
i=1,2,…,M
sObtaining a system monitoring result Z at the t moment for a system monitoring result at the t moment of the ith sensor
t:
In the formula, MsThe number of sensors;
step 1.3, monitoring result Z of the system at the time t
tArranging according to increasing order to obtain ascending sequence
And calculating a check value Q
1:
In the formula (I), the compound is shown in the specification,
and
the maximum value and the minimum value of the monitoring data at the time t are respectively,
and
respectively measuring results of the ith sensor at the moment t and the nearest monitoring result thereof;
step 1.4, according to ascending sequence
Determination of the test value Q from the number of measured values and the specified confidence level
2If Q is
1>Q
2Then the ith sensor measurement at time t is compared
If the abnormal value is regarded as an abnormal value and discarded, otherwise, the abnormal value is retained, and the steps are repeated on the processed monitoring data until the moment is monitoredAll abnormal values in the data are removed to obtain an abnormal-free monitoring data sequence
In the formula (I), the compound is shown in the specification,
the first monitoring value at the time t for finally obtaining the abnormal data, the same principle
And
second and Nth time points respectively representing the existence of finally obtained abnormal-free data
sThe monitored value.
Step 1.5, pre-distributing fusion weight to the data after the abnormal value is removed based on a distance criterion:
in the formula (I), the compound is shown in the specification,
pre-assigning a fusion weight to the ith sensor at time t,
for the ith monitoring value at the moment t with finally obtained abnormal-free data,
is the mean value of the finally obtained abnormal-free data, N
sIn order to obtain the final abnormal-free data,
and the k-th monitoring value at the moment t is the finally obtained abnormal-free data.
The step 2 is as follows:
step 2.1, taking the state estimation of the jth sensor as an example, calculating the initial mean square error matrix P of the state vector
0And a priori mean
In the formula, E [ Delta ]]Mean expectation, P, of Δ0For initial estimation of error variance, x0Is the initial value of the system state;
step 2.2, calculating an unscented transformation Sigma sampling point based on a sampling strategy:
in the formula (I), the compound is shown in the specification,
is the estimated value of the system state at the time t-1, xi
0,t-1And xi
i,t-1Respectively the 0 th and ith system unscented transformation Sigma sampling points at the time of t-1, n is the system state vector dimension, P is used for adjusting the distance between the sampling point and the original sample point, P is the state variable covariance matrix at the time of t-1,
the ith main diagonal element representing the square root matrix;
step 2.3, calculating the first-order statistical characteristic weight coefficient of the sampling point
And weight coefficients of second order statistical characteristics
Step 2.4, calculating a one-step prediction matrix based on Sigma sampling points at the moment t
Sum covariance matrix P
t|t-1:
ξi,t|t-1=f(ξi,t-1)(i=0,1,…,2n)
In the formula, wtIs the state noise at time t, QtThe state noise variance at time t.
Step 2.5, for the one-step prediction matrix in step 2.4
The UT transform was performed again to obtain a new Sigma point set as follows:
in the formula, xi'
0,t|t-1And ξ'
i,t|t-1With respect to one-step prediction matrices for time t-1, respectively
And the 0 th and i th systems of (a) transform the Sigma sample points without traces.
Step 2.6, substituting the new Sigma point set obtained in the step 2.5 into the measurement equation to obtain an observation predicted value z of the ith sensor at the moment t
i,t|t-1To the observed value z
i,t|t-1Weighted summation is carried out to obtain an observed prediction mean value
zi,t|t-1=h(ξi',t|t-1)
In the formula (I), the compound is shown in the specification,
for the observed variance at time t, e
tObserved noise at time t, R
tIs the observed noise variance at time t.
Step 2.7, State estimation covariance matrix
Is updated as:
step 2.8, calculating Kalman gain K
tUpdating the state variable estimate
Sum equation of state covariance P
t:
Step 2.9, introducing a time-varying adaptive fading factor lambdatTo prediction state covariance matrix Pt|t-1:
Step 2.10, with correction
Prediction state covariance matrix P instead of classical UKF
t|t-1Completing classical UKF to update local state estimation and obtaining estimation result of jth sensor at t moment
Step 2.11, performing state parallel estimation on all sensors in the system according to the steps 2.1-2.10 to obtain a local state estimation result
j=1,2,…,N。
The step 3 is as follows:
step 3.1, the global optimal state fusion value of the system at the moment t
Expressed as:
in the formula (I), the compound is shown in the specification,
and fusion weights representing the jth sensor estimation result at the time t.
Step 3.2, combining the fusion weight pre-distribution and linear weighting fusion criteria to obtain a global state estimation result:
in the formula (I), the compound is shown in the specification,
for the global state estimation result at the time t of the system
A weight is pre-assigned for the jth sensor estimate at time t,
and
respectively, the jth sensor local estimation result at the moment t and the fusion weight thereof.
The invention has the beneficial effects that the multisource information fusion method based on weight pre-distribution selects the Q test method to remove abnormal monitoring data of the sensor, and pre-distributes the fusion weight to the data after removing abnormal values based on the distance criterion; self-attenuation Unscented Kalman Filtering (UKF) based on the Mahalanobis distance is selected as a local state estimator, the adaptability and robustness of the UKF to modeling errors of a multi-sensor nonlinear random system are improved by evaluating the squared Mahalanobis distance of an innovation vector and taking corresponding measures to obtain a local state estimation result; and fusing the multi-sensor monitoring data based on a minimum variance linear weighting criterion to obtain a global state estimation result. The algorithm has good fusion effect and high precision, can still obtain accurate fusion results for the conditions of sensor faults, monitoring data loss and the like, and has strong referential property and practicability.
Detailed Description
The present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.
The invention discloses a multi-source information fusion method based on weight pre-distribution, which is implemented by combining a figure 1 according to the following steps:
step 1, selecting a Q test method to remove abnormal monitoring data of each sensor aiming at a system containing a plurality of sensors, and pre-distributing fusion weight to the data after removing abnormal values based on a distance criterion;
the step 1 is as follows:
step 1.1, setting the state space model of each sensor to meet the following form:
in the formula, xtAnd xt+1Respectively representing the state values of the tested system at the time t and the time t + 1; f (-) is the system nonlinear state function; w is atIs zero mean Gaussian white noise with variance Q more than or equal to 0; z is a radical oft+1Is the measured value at the moment of the sensor t + 1; h (-) is a sensor nonlinear measurement function; e.g. of the typet+1Zero mean Gaussian white noise with variance R not less than 0 at the moment of t + 1;
step 1.2, taking the monitoring data of each sensor at the moment t as an example, setting
i=1,2,…,M
sObtaining a system monitoring result Z at the t moment for a system monitoring result at the t moment of the ith sensor
t:
In the formula, MsThe number of sensors;
step 1.3, monitoring result Z of the system at the time t
tArranging according to increasing order to obtain ascending sequence
And calculating a check value Q
1:
In the formula (I), the compound is shown in the specification,
and
the maximum value and the minimum value of the monitoring data at the time t are respectively,
and
respectively measuring results of the ith sensor at the moment t and the nearest monitoring result thereof;
step 1.4, according to ascending sequence
Determination of the test value Q from the number of measured values and a specified confidence level (e.g. 90%)
2If Q is
1>Q
2Then the ith sensor measurement at time t is compared
Discarding the abnormal values, otherwise, retaining the abnormal values, repeating the steps on the processed monitoring data until all the abnormal values in the monitoring data at the moment are removed, and obtaining the abnormal-free monitoring data sequence
In the formula (I), the compound is shown in the specification,
the first monitoring value at the time t for finally obtaining the abnormal data, the same principle
And
second and Nth time points respectively representing the existence of finally obtained abnormal-free data
sThe monitored value.
Step 1.5, pre-distributing fusion weight to the data after the abnormal value is removed based on a distance criterion:
in the formula (I), the compound is shown in the specification,
pre-assigning a fusion weight to the ith sensor at time t,
for the ith monitoring value at the moment t with finally obtained abnormal-free data,
is the mean value of the finally obtained abnormal-free data, N
sIn order to obtain the final abnormal-free data,
and the k-th monitoring value at the moment t is the finally obtained abnormal-free data.
Step 2, selecting a self-attenuation unscented Kalman filter UKF based on the Mahalanobis distance as a local state estimator, evaluating the squared Mahalanobis distance of the innovation vector, and taking corresponding measures to improve the adaptability and robustness of the UKF to the modeling error of the multi-sensor nonlinear random system to obtain a local state estimation result;
the step 2 is as follows:
step 2.1, taking the state estimation of the jth sensor as an example, calculating the initial mean square error matrix P of the state vector
0And a priori mean
In the formula, E [ Delta ]]Mean expectation, P, of Δ0For initial estimation of error variance, x0Is the initial value of the system state;
step 2.2, calculating an unscented transformation Sigma sampling point based on a sampling strategy:
in the formula (I), the compound is shown in the specification,
is the estimated value of the system state at the time t-1, xi
0,t-1And xi
i,t-1Respectively the 0 th and ith system unscented transformation Sigma sampling points at the time of t-1, n is the system state vector dimension, P is used for adjusting the distance between the sampling point and the original sample point, P is the state variable covariance matrix at the time of t-1,
the ith main diagonal element representing the square root matrix;
step 2.3, calculating the first-order statistical characteristic weight coefficient of the sampling point
And weight coefficients of second order statistical characteristics
Step 2.4, calculating a one-step prediction matrix based on Sigma sampling points at the moment t
Sum covariance matrix P
t|t-1:
ξi,t|t-1=f(ξi,t-1)(i=0,1,…,2n)
In the formula, wtIs the state noise at time t, QtThe state noise variance at time t.
Step 2.5, for the one-step prediction matrix in step 2.4
The UT transform was performed again to obtain a new Sigma point set as follows:
in the formula, xi'
0,t|t-1And ξ'
i,t|t-1With respect to one-step prediction matrices for time t-1, respectively
And the 0 th and i th systems of (a) transform the Sigma sample points without traces.
Step 2.6, substituting the new Sigma point set obtained in the step 2.5 into the measurement equation to obtain an observation predicted value z of the ith sensor at the moment t
i,t|t-1To the observed value z
i,t|t-1Weighted summation is carried out to obtain an observed prediction mean value
zi,t|t-1=h(ξi',t|t-1)
In the formula (I), the compound is shown in the specification,
for the observed variance at time t, e
tObserved noise at time t, R
tIs the observed noise variance at time t.
Step 2.7, State estimation covariance matrix
Is updated as:
step 2.8, calculating Kalman gain K
tUpdating the state variable estimate
Sum equation of state covariance P
t:
Step 2.9, introducing a time-varying adaptive fading factor lambdatTo prediction state covariance matrix Pt|t-1:
Step 2.10, with correction
Predicted state covariance matrix instead of classical UKFArray P
t|t-1Completing classical UKF to update local state estimation and obtaining estimation result of jth sensor at t moment
Step 2.11, performing state parallel estimation on all sensors in the system according to the steps 2.1-2.10 to obtain a local state estimation result
j=1,2,…,N。
And 3, fusing the multi-sensor monitoring data based on a minimum variance linear weighting criterion to obtain a global state estimation result.
The step 3 is as follows:
step 3.1, under the condition of not considering the influence of the pre-distribution weight, the global optimal state fusion value of the system at the moment t
Expressed as:
in the formula (I), the compound is shown in the specification,
and fusion weights representing the jth sensor estimation result at the time t.
Step 3.2, combining the fusion weight pre-distribution and linear weighting fusion criteria to obtain a global state estimation result:
in the formula (I), the compound is shown in the specification,
for the global state estimation result at the time t of the system
A weight is pre-assigned for the jth sensor estimate at time t,
and
respectively, the jth sensor local estimation result at the moment t and the fusion weight thereof.
Fig. 2 is a diagram of the monitoring results of the system sensor of the invention, in which the black solid line is the real braking speed of a certain train, and the other four curves represent the four-way speed measurement results of the train. Fig. 3 shows a train speed fusion result graph obtained by the present invention, in which the black solid line is the real speed of the train, and other curves represent the monitoring data fusion results under different conditions. It can be clearly seen from the observation of fig. 3 that the method provided by the invention can effectively fuse the multi-path speed measurement results of the train, is influenced by behavior factors such as loss or abnormality of the measured data of the sensor, and has slightly higher precision than other two conditions because the measured data contains more effective information under the condition of meeting the actual precision requirement of the system.