Centralized multi-sensor fusion filtering method based on characteristic function

文档序号：799583 发布日期：2021-04-13 浏览：27次中文

阅读说明：本技术 基于特征函数的集中式多传感器融合滤波方法 (Centralized multi-sensor fusion filtering method based on characteristic function ) 是由袁洢苒文成林裘奕婷李明媚徐晓滨于 2020-12-21 设计创作，主要内容包括：本发明公开了一种基于特征函数的集中式多传感器融合滤波方法,本发明通过对多个传感器采用集中的设计方式,将所有传感器收集到的信息都传输到融合中心之后,再对所有数据进行特征函数滤波处理。在时间充足的情况下,不遗漏任何信息。通过集中式融合方式,可以在很大程度上提高滤波估计精度,集中式设计考虑到了所有可能的情况,不考虑信息的丢包和延迟,收集到了所有的信息,能够得到非常高的估计精度,能在样本数量不太大、时间充足又要求高精度的非线性系统甚至强非线性系统中得到很好的应用。(The invention discloses a centralized multi-sensor fusion filtering method based on a characteristic function. In case of sufficient time, no information is missing. The filtering estimation precision can be improved to a great extent by a centralized fusion mode, the centralized design considers all possible situations, does not consider packet loss and delay of information, collects all information, can obtain very high estimation precision, and can be well applied to a nonlinear system which has small sample quantity, sufficient time and requires high precision, and even a strong nonlinear system.)

1. A centralized multi-sensor fusion filtering method based on a characteristic function is applied to a space target tracking system and comprises the following steps:

(1) designing a state space target tracking model, wherein the model of the target doing uniform linear motion in space is as follows:

where x (k) is the system state vector, y_i(k) Is the output vector of the ith group of sensors; w (k) and v_i(k +1) are the process noise and measurement noise vectors, respectively, for which the eigenfunctions are known, and whose distribution is F_w(x)、F_v(x) (ii) a A (k +1, k) is a known state transition matrix, G (k +1, k) is a known process drive matrix, h_i(. cndot.) is a continuous smooth nonlinear function, i ═ 1,2, …, N;

(2) and (3) calculating a measurement equation after the information of all the sensors is centralized at the moment of k + 1:

(2a) calculating the measurement value of each sensor according to a space target tracking model, and centralizing all the measurement values to obtain y (k + 1);

(2b) calculating a bol measurable nonlinear function of each sensor according to a space target tracking model, and concentrating all nonlinear functions to obtain h (x (k + 1));

(2c) calculating the measurement noise of each sensor according to a space target tracking model, and concentrating all the noises to obtain v (k + 1);

(2d) combining (2a), (2b) and (2c) to obtain a centralized measurement equation;

(3) under the CFF framework, the state model of the target and the centralized measurement model are combined to calculate the state estimation value of the centralized sensor at the moment of k +1

(3a) According to the target tracking model, calculating the predicted value of the target state from k to k +1

(3b) According to (3a), calculating the measurement predicted value of the sensors after concentration from the moment k to the moment k +1

(3c) Computing residual information of the post-concentration sensor from the measurement equations and (3b)

(3d) Calculating a state transition matrix A (k +1, k) according to a kinematic formula;

(3e) calculating an initial state estimate

(3f) Calculating a state error equation e (k) according to the state equation and the equation (3 e);

(3g) calculating a state error recurrence equation e (k +1) according to (3 f);

(3h) simultaneously solving characteristic functions on two sides of the equation (3 g);

(3i) establishing a known target feature function

(3j) Establishing a weight function matrix U (t) of the filter;

(3k) establishing a filter parameter index J according to (3h), (3i) and (3J)₀(k+1)；

(3l) establishing a filter performance index function J (k +1) according to the (3 k);

(3m) simplifying the parameters in (3 l);

(3n) obtaining a simplified filter performance index J' (k +1) according to (3l) and (3 m);

(3o) establishing a filter gain matrix K (K +1) to be estimated of the sensors after concentration;

(3p) solving a first order partial derivative of K (K +1) according to (3 n);

(3q) solving a second-order partial derivative of K (K +1) according to (3 n);

(3r) obtaining a filter gain matrix K (K +1) of the sensors after concentration according to (3o), (3p) and (3 q);

(3s) calculating the state estimation value of the centralized sensor according to (3a), (3c) and (3r)

Technical Field

The invention belongs to the field of space target tracking of a nonlinear dynamic system, and particularly relates to the field of space target tracking of which a state model is linear and a measurement model is a strong nonlinear system, which can be used for optimizing the real-time position and speed of a target in the space target tracking process.

Background

The filtering method is an important method in state estimation, and the state estimation has very wide application in the fields of fault diagnosis, target tracking, signal processing, computer vision, communication, navigation and the like.

The traditional Kalman filtering is only suitable for a system in which a state model and a measurement model are linear and noise is white Gaussian noise. When the noise of the system is no longer white gaussian noise or the system is no longer a linear system, the conventional kalman filtering method is no longer applicable. In practical application systems, most system models are nonlinear or non-gaussian, and therefore, for a nonlinear system or a system with non-gaussian noise, various filters are extended on the basis of a kalman filter in order to realize state estimation of the nonlinear system or the system. For example, an Extended Kalman Filter (EKF) is used, but the EKF can only achieve a second-order approximation at most, and discarded information of high-order terms brings certain errors to a filtering result; unscented Kalman Filter (UKF) and volume kalman filter (CKF) are all approximate through getting the point, to nonlinear gauss system, although EKF, UKF and CKF's application is all comparatively extensive, its nonlinear approximation all can cause certain error to can't be with these three kinds of wave filters application to nonlinear gauss system, the limitation is great. Particle Filters (PF) developed later, for non-linear non-gaussian systems, although they are better solved in theory, their implementation depends on a large number of particle samples, so that the computational complexity is very high, and the degradation phenomenon of particles during resampling can reduce the speed and accuracy of filtering, affecting practical applications.

For a large number of strong nonlinear observation systems, the existing filtering method is still difficult to solve well. The feature function filter (CFF) developed recently is only for a system in which a state model is linear, and has no requirement on a measurement model thereof, and also has no requirement on whether noise is gaussian or non-gaussian, so that the CFF is expected to solve the problem that the measurement model is a strongly non-linear system.

Although eigenfunction filtering is theoretically superior to any other nonlinear filtering method, it is implemented for only one filter. In practical use, the error can be reduced as far as possible, and cannot be completely eliminated. The main reasons for the low filtering accuracy of the feature function are from four parts, first, inaccurate data collection. For a target moving in space, the position and the speed of the target in the x, y and z directions are changed in real time, and if only one sensor is used for measuring the target, the change condition of the target cannot be comprehensively and accurately captured. Second, random noise settings are inaccurate. When modeling a system, random setting of noise is too ideal, and in an actual space dynamic system, as the speed increases, the target is greatly influenced by external environment and some random factors, so that a large deviation exists between an actual error and a random error set by the system. Third, the complexity of the system model is high. For a target with spatial motion, the state comprises six state variables of position and speed in the directions of x, y and z axes, and the complexity is too high compared with a common model. In the actual filtering process, each state generates a certain error when being updated, and the error may be larger by comprehensively considering the six state models. Fourth, the performance of the sensor itself changes. Due to aging of parts inside the sensor or changes of parameters caused by the service life or damage caused by moisture, the inaccuracy of the measurement result can be caused. Fifth, in a practical system, when the number of sensors is too large and the time is limited, packet loss and delay must be considered, and all information that should be collected is not collected, which also reduces the filtering accuracy.

Disclosure of Invention

In order to overcome the disadvantages of the prior art, the present invention provides a centralized filtering method for fusing a plurality of sensors.

The invention measures the space target from different directions by distributing a plurality of sensors at different positions in the space, adopts a centralized design, and centralizes all the information to carry out filtering processing after all the sensors transmit the information to the fusion center so as to realize real-time tracking of the position and the speed of the target. The filtering method of the fusion of a plurality of sensors can greatly improve the estimation accuracy under the condition of sufficient time.

In order to achieve the purpose, the invention adopts the technical scheme that:

the invention comprises the following steps:

(1) designing a state space target tracking model, wherein the model of the target doing uniform linear motion in space is as follows:

A (k +1, k) is a known state transition matrix, G (k +1, k) is a known process drive matrix, h_i(. cndot.) is a continuous smooth nonlinear function, i ═ 1,2, …, N.

(2) And (3) calculating a measurement equation after the information of all the sensors is centralized at the moment of k + 1: