Table of Contents Author Guidelines Submit a Manuscript
Journal of Electrical and Computer Engineering
Volume 2018, Article ID 1505137, 11 pages
https://doi.org/10.1155/2018/1505137
Research Article

State Fusion of Decentralized Optimal Unbiased FIR Filters

Key Laboratory of Advanced Process Control for Light Industry, Jiangnan University, Wuxi 214122, China

Correspondence should be addressed to Fei Liu; nc.ude.nangnaij@uilf

Received 2 November 2017; Accepted 1 February 2018; Published 3 June 2018

Academic Editor: Rajesh Khanna

Copyright © 2018 Xuefeng Fan and Fei Liu. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Abstract

The paper presents a decentralized fusion strategy based on the optimal unbiased finite impulse response (OUFIR) filter for discrete systems with correlated process and measurement noise. We extend OUFIR filter to apply in the model with control inputs. Taking it as local filters, cross covariance between any two is calculated; then it is expressed to the fast iterative form. Finally based on cross covariance, optimal weights are utilized to fuse local estimates and the overall outcome is obtained. The numerical examples show that the proposed filter exhibits better robustness against temporary modeling uncertainties than the fusion Kalman filter used commonly.

1. Introduction

State estimation plays an important role in many applications such as control, moving target tracking, and timekeeping and clock synchronization [1]. The existing filtering methods can be classified by two types: the infinite impulse response (IIR) filter and the finite impulse response (FIR) filter [2]. Specifically, the former uses all the historical measurements, and a special case is the Kalman filter (KF) [3], while the latter utilizes limited memory over the most recent time interval [4]. It is due to the difference of structure that FIR-type filters exhibit some useful engineering features such as bounded input/bounded output (BIBO) stability, round off errors [5], and better robustness against temporary uncertainties [6].

With higher requirements of flexibility and accuracy, information fusion filtering theory for multisensor systems has been studied and widely applied [7, 8]. For example, [9, 10] proposed the weighted fusion estimator according to the maximum likelihood and weight least square respectively with local filtering error cross covariance zero or not. In [11], an optimal decentralized fusion weighted by matrix was proposed to fuse local KF in the linear unbiased minimum variance (LUMV) sense. Furthermore, the situation of correlated multiplicative and additive noise was analyzed in [8], and the decentralized KF fusion problem for cross-correlated measurement noises was investigated in [12] which also discussed the cases with feedback from the fusion center to local sensors. Reference [13] focused on the decentralized fusion estimation problem for networked systems with random delays and packet losses, while [14] aimed at missing measurements and correlated noises. Reference [15] applied decentralized fusion to the maneuvering target tracking with multirate sampling and uncertainties in wireless sensor networks.

The fusion filtering methods nowadays take the KF as local filters mostly, so that they inherit the properties of KF, optimal but not robust under some uncertainties. Meanwhile there are few results considering decentralized FIR-type filter to deal with multisensor systems. Basically optimal FIR (OFIR) filter was obtained in [16] by minimum mean square error sense. Besides unbiased FIR (UFIR) filter was derived for real-time models and realized iteratively in [17], which can ignore the noise statistics but does not guarantee optimality. In between, [18] represents two forms of optimal unbiased FIR (OUFIR) filter, minimum variance unbiased FIR and embedded unbiasedness (EU) constraint on OFIR, and proved the identity of the two further. Contrary to the KF, the OUFIR filter is more robust and performs lower sensitivity to the initial condition [19]. In addition, there are a few new study and improved solutions on FIR filtering [2022] developed nowadays. Some practical applications were reported [23, 24] as well.

In this correspondence, we firstly develop OUFIR filter to an extension to handle the system with or without the control inputs universally. Then considering it as local filters, cross covariance between any two is determined by batch and iterative form under the LUMV sense. Finally an optimal decentralized state fusion filter is proposed, which shows better immunity against temporary model uncertainties than traditional KF. The rest of this paper is organized as follows. In Section 2, we describe the preliminaries of system model. Local OUFIR filters are developed in Section 3. Main results about fusion and cross covariance are given in Section 4. The simulation and conclusions are provided in Sections 5 and 6, respectively.

2. System Model and Preliminaries

Consider the following linear discrete time-invariant system with multiple sensors:where denotes the index of sensor, is the discrete time index, and denote the state and measurement vectors, respectively, is the known control signal, and the system matrices , , , and are known with appropriate dimensions. In this paper, the process noise and measurement noise are assumed to be correlated white noises with zero mean andwhere is the Kronecker delta function, i.e., if , , otherwise, . For notational convenience, we assign , when . That is, the process noise is correlated with each sensor noise which is also correlated with each other.

Firstly the original state-space model (1) and (2) needs to be transformed to be a form with uncorrelated noise. In this way, we add the right-hand side of (1) with a zero term aswhere , , , and is a coefficient matrix. To remove the relevance between and , the expectation term should be zero by definition. That is,which further gives us . It is not difficult to find that , which means that the new-defined noise term is zero mean also. Accordingly, the cross state noise variance can be computed asAs can be seen, with an auxiliary matrix , there is no correlation between and , and we consider the state transition equation (5) and measurement equation (2) subsequently.

By defining an estimation horizon as , where is the initial time step and is the horizon length, models (5) and (2) within the estimation horizon can be extended as [16] similarly,where the initial state is assumed to be known, and the extended vectors are specified as The involved extended matrices , , , , , and are all -dependent, which have the forms ofwhere we assignand .

Once the extended state-space models (8) and (9) are available, the problem considered in this paper can be formulated as follows. Given the state-space model (1) and (2), we would like to design local OUFIR filter for each sensor.

3. Local Optimal Unbiased FIR Filter

A demonstration of the FIR filter structure is provided in Figure 1. From the figure, FIR estimators explicitly employ most recent measurements unlike KF. It is in this way that some nice properties like better robustness are achieved. Based on it, for subsystem represented as (8) and (9), we design a local OUFIR filter in this section.

Figure 1: Operation time diagrams of the FIR structures.
3.1. Batch Form

The linear FIR filter for each sensor can be expressed as the following batch form [25] generally:for gain matrices and . It is noted that the filter defined as (17) handles all the measurements and input values collected within the estimation horizon at one time.

Introduce unbiasedness constraint (or deadbeat constraint) [26]:in the estimation (17) where can be specified asand , are the first vector rows of , respectively. By substituting (19) and (17) into (18), replacing the term with (9), and providing the averaging, one arrives at two following constraintsSubstituting (21) into (17) yieldsProvided , the instantaneous filtering error can be defined as , then the OUFIR filter is derived by solving the following optimization problemwhere should satisfy the constrain (20).

By replacing with (19) and with (22), and substituting the unbiasedness constrain (20), filtering error is obtainedEmploying the trace operation and providing the averaging, the optimization problem (23) can be rewritten asIt is not difficult to getwhere is given in (7), and is defined in (4). Referring to [19], we can obtain the following gain:where and

At this point, the local OUFIR filter with batch form is specified by (22) with filter gain provided as (26). As mentioned, this structure is not suitable for the subsequent fusion step, as measurements are operated at one time and there is no error covariance available to quantify the estimation accuracy. To address these issues, an equivalent iterative computational formula is given below.

3.2. Equivalent Iterative Computation

Introducing , (22) can be divided into two parts shown as follows:and consider each components separately.

By former transformations, for (29), recursive algorithm is represented easily asAssigning , this part becomes . For this estimate, we notice a iterative form proposed in [19], which is represented similarlywhere and are gain matrices. Besides, is the first vector rows of , substituting (31), which is given byThen putting (33) into (32), iterative form is

Combining (31) and (35), we arrive atwhich suggests easily that the prior estimate is . Referring to [27], we finally come up with a general iterative OUFIR filtering algorithm similarly whose pseudocode is summarized as Algorithm 1 in which is an auxiliary variable replaced with . Unlike the OFIR or KF, it is because of unbiasedness condition that the filter does not require initial information. Instead that can be defined at using a short batch form given in Table 1; is the number of system states. In the Algorithm 1, variable ranges from to . The true filtering result corresponds to .

Table 1: Computation of the initial state.
Algorithm 1: General iterative OUFIR filtering algorithm.

Once the local estimates are corrected, , an optimal fusion strategy based on cross covariance is employed to get the overall estimates, which is introduced below.

4. Decentralized Fusion Filter

In this section, we will investigate the decentralized fusion filters and first deduce the corresponding filtering error cross covariance between any two local subsystems with FIR structure. Due to inefficient computation, then we find its fast iterative form. Based on these, a multisensor decentralized fusion OUFIR filtering algorithm is finally presented.

4.1. Cross Covariance Matrices and Iterative Computation

For the sensor subsystem, filtering error is written as (24) where filtering gain is specified as (27). So the cross covariance matrix between the and the subsystem can be showed aswhere , are defined in (26).

The batch cross covariance given by (37) is exact but takes a large amount of computation. To find a recursive formula, one may combine given in Algorithm 1 and (5) with respect to an auxiliary variable , so can be rewritten asSubstitute given by (5) and (2) into (38), then transform it intoSo can now be computed iteratively over the recursive form ofwhere ranges from to , and the output is taken when similarly. The initial value is given in a short batch form by (37) asClearly the pseudocode is shown in Algorithm 2.

Algorithm 2: Iterative form for .
4.2. Decentralized OUFIR Filtering Algorithm

Based on the local filters in Algorithm 1 and the filtering error cross covariance matrices between any two local filters in Algorithm 2, we can get the decentralized fusion filter weighted by matrices in the LUMV sense [7]. Introduce the synthetically unbiased estimatorwhere are arbitrary matrices. From the unbiasedness assumption, we have and . Taking the expectation in (42) yieldswhere and denotes identity matrix. Then fusion estimation error is . Let ; the error variance matrix of the fusion estimator iswhere is an symmetric positive definite matrix where is from (40).

To find the optimal weights under the constrain (43), applying the Lagrange multiplier approach, the target function is defined aswhere . Set and note that ; we haveCombining (43) and (46) yields the matrix equation asNoted is a symmetric positive definite matrix and using the formula of the inverse matrix [28] we have

Pseudocode is summarized by Algorithm 3, and the corresponding structure is shown in Figure 2. As can be seen, each sensor subsystem estimates the state independently at time . Next fusion layer determines optimal matrix weights. As a basic requirement, the proposed decentralized fusion filter will have better accuracy than any local one. Different from fusion KF, the proposed method does not relay on the initial information about , and , and it performs better robustness.

Algorithm 3: Optimal decentralized fusion OUFIR filter.
Figure 2: The structure of decentralized OUFIR filter.

5. Simulation

In this section, we consider two basic models of moving target tracking system and 1 degree of freedom (1-DOF) torsion system. Fusion OUFIR filter is applied to above models under various conditions such as temporary modeling uncertainties and unknown noise statistics. Compared to fusion Kalman filter (conveniently FOUF and FKF denote as fusion OUFIR filter and fusion KF filter respectively next), it demonstrates better performance.

5.1. Moving Target Tracking System

With three sensors, a simplified model can be specified by (1) and (2) with , , and where , , control signal , and . And the measurement noise is represented as , which are correlated with the white process noise . The correlation coefficients are , , and . are white Gaussian noise with zero mean and variances , , and , respectively, and are independent of . Easily we can get and as

5.1.1. Estimation Accuracy with Correct Noise Variances

In this section, we assume that all the modeling parameters are known accurately and show that FOUF outperforms any local filter. The process was simulated over 500 points, and the estimation horizon length used is . The average root mean square error (RMSE) based on 100 Monte-Carlo (MC) runs is shown in Figure 3. The result proves that the proposed approach has better accuracy exactly.

Figure 3: RMSEs produced by local and fusion filters: (a) the first state and (b) the second state.
5.1.2. Sensitivity to Incorrect Noise Variance

The knowledge of noise is typically not known exactly in practice, which does not guarantee optimality. To figure out the effect of noise variances, we give an idea about the FKF and FOUF immunity to noise statistics’ errors in the worst case. Introducing error coefficient , designed covariances are substituted here with .

of FKF and FOUF over 100 MC runs for are shown in Figure 4 and also listed in Table 2. As can be seen, errors in the FKF grow rapidly with and even become unacceptable, while the proposed FOUF has much higher immunity against noise error when . Otherwise, performance of FOUF exhibits closely to FKFs.

Table 2: Performance of FOUF and FKF with respect to different .
Figure 4: Average against system noises uncertainties between FKF and FOUF.
5.1.3. Immunity to Errors in the Noise Statistics

To investigate the effect of temporary inaccurate statistics, the state noise variance is set to vary as follows: when and otherwise. And invariable value is used in all the algorithms. The average RMSEs based on 100 MC runs are provides by Figure 5. One can see that FOUF has much higher immunity against errors of noise statistic than FKF. Besides it is noted that the first state is more obvious about robustness performance.

Figure 5: RMSEs produced by FKF and FUF with imprecisely defined noise statistics: (a) the first state and (b) the second state.
5.1.4. Robustness against Model Uncertainties

To identify the robustness of each tracker against temporary model uncertainties, we setwhere when and otherwise. The process is simulated over 500 subsequent points with correct noise statistics. Comparing the FKF as a benchmark, Figures 6(a) and 6(b) illustrate the typical response under the model mismatch introduced. From figures, one infers that the FOUF’s peak value is smaller and recovers to normal faster than the FKF’s, particularly with the first state.

Figure 6: Estimation errors by FKF and FOUF filter against temporary model uncertainty in a gap of : (a) the first state and (b) the second state.
5.2. 1-DOF Torsion System

We consider 1-DOF torsion system [29] as the object whose schematic and parameters are provided by Figure 7 and Table 3. Referring from [29] of the equations of motion, the continuous-time model can be described bywhere and

Table 3: Physical parameters of 1-DOF torsion system, where EMI is equal to equivalent moment of inertia, and EVD is equivalent viscous damping.
Figure 7: Schematic of 1-DOF Torsion system.

Discretize this process with sampling time , and the discrete state-space model is achieved as where is introduced to describe the stochastic properties of states, is input signal, , and . The angular positions of two loads are measured using three groups of sensors, and the observation equation is given by , where , The correlation between state and measurement noises is described as , where the coefficients are The measurement noises variances are assigned as , , and .

5.2.1. Imprecisely Defined Noise Statistics

With inaccurateness, the state noise variance is set to vary as follows: when and otherwise. Meanwhile invariable value , is used in the filter. The average RMSEs based on 100 MC runs are provided by Figure 8. It suggests again that FOUF performs closely to the FKF under the correct noise, but it has better robustness against noise errors than FKF.

Figure 8: RMSEs produced by FKF and FOUF with imprecisely defined noise statistics: (a) the first state and (b) the second state.
5.2.2. Temporary Model Uncertainty

Similarly, we assume that changes unpredictably as when and otherwise, which makes vary simultaneously. Average RMSEs produced by FKF and FOUF over 100 Monte-Carlo simulations are shown Figure 9.

Figure 9: RMSEs produced by FKF and FOUF with temporary model uncertainty: (a) the first state and (b) the second state.

Clearly the result obtained in this case coincides with that in the previous model. At this point, one can conclude that proposed FOUF is less influenced by noise and more robust against temporary model mismatch.

6. Conclusion

Based on optimal fusion weighted by matrices in the LUMV sense, this paper presents a decentralized fusion OUFIR filter for discrete-invariant control system. Fundamentally the FOUF has higher accuracy than each local filter. The simulation results show that the proposed algorithm maintains the advantages of OUFIR filter and suggests lowlier sensitive to noise statistics and more robust against modeling uncertainties.

Conflicts of Interest

The authors declare that there are no conflicts of interest regarding the publication of this paper.

Acknowledgments

This paper is supported by the National Natural Science Foundation of China (NSFC 61773183).

References

  1. D. Simon, Optimal State Estimation: Kalman, H∞, and Nonlinear Approaches, Wiley-Interscience, 2006. View at Publisher · View at Google Scholar · View at Scopus
  2. V. R. Algazi, M. Suk, and C.-S. Rim, “Design of almost minimax FIR filters in one and two dimensions by WLS techniques,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 33, no. 6, pp. 590–596, 1986. View at Publisher · View at Google Scholar · View at Scopus
  3. R. E. Kalman, “A new approach to linear filtering and prediction problems,” Journal of Basic Engineering, vol. 82, no. 1, pp. 34–45, 1960. View at Publisher · View at Google Scholar
  4. S. Zhao, Y. S. Shmaliy, B. Huang, and F. Liu, “Minimum variance unbiased FIR filter for discrete time-variant systems,” Automatica, vol. 53, pp. 355–361, 2015. View at Publisher · View at Google Scholar · View at MathSciNet
  5. W. H. Kwon, P. S. Kim, and S. H. Han, “A receding horizon unbiased FIR filter for discrete-time state space models,” Automatica, vol. 38, no. 3, pp. 545–551, 2002. View at Publisher · View at Google Scholar · View at Scopus
  6. A. H. Jazwinski, Stochastic Process and Filtering Theory, Academic Press, New York, NY, USA, 1970.
  7. S.-L. Sun and Z.-L. Deng, “Multi-sensor optimal information fusion Kalman filter,” Automatica, vol. 40, no. 6, pp. 1017–1023 (2005), 2004. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  8. T. Tian, S. Sun, and N. Li, “Multi-sensor information fusion estimators for stochastic uncertain systems with correlated noises,” Information Fusion, vol. 27, pp. 126–137, 2016. View at Publisher · View at Google Scholar · View at Scopus
  9. N. A. Carlson, “Federated square root filter for decentralized parallel processes,” IEEE Transactions on Aerospace and Electronic Systems, vol. 26, no. 3, pp. 517–525, 1990. View at Publisher · View at Google Scholar · View at Scopus
  10. K. C. Chang, R. K. Saha, and Y. Bar-Shalom, “On optimal track-to-track fusion,” IEEE Transactions on Aerospace and Electronic Systems, vol. 33, no. 4, pp. 1271–1276, 1997. View at Publisher · View at Google Scholar · View at Scopus
  11. S.-L. Sun, “Multi-sensor optimal information fusion Kalman filters with applications,” Aerospace Science and Technology, vol. 8, no. 1, pp. 57–62, 2004. View at Publisher · View at Google Scholar · View at Scopus
  12. E. Song, Y. Zhu, J. Zhou, and Z. You, “Optimal Kalman filtering fusion with cross-correlated sensor noises,” Automatica, vol. 43, no. 8, pp. 1450–1456, 2007. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  13. N. Li, S. Sun, and J. Ma, “Multi-sensor distributed fusion filtering for networked systems with different delay and loss rates,” Digital Signal Processing, vol. 34, pp. 29–38, 2014. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  14. R. Caballero-Águila, I. García-Garrido, and J. Linares-Pérez, “Optimal fusion filtering in multisensor stochastic systems with missing measurements and correlated noises,” Mathematical Problems in Engineering, vol. 2013, Article ID 418678, 14 pages, 2013. View at Publisher · View at Google Scholar · View at Scopus
  15. X. Yang, W.-A. Zhang, L. Yu, and K. Xing, “Multi-rate distributed fusion estimation for sensor network-based target tracking,” IEEE Sensors Journal, vol. 16, no. 5, pp. 1233–1242, 2016. View at Publisher · View at Google Scholar · View at Scopus
  16. Y. S. Shmaliy, “Linear optimal FIR estimation of discrete time-invariant state-space models,” IEEE Transactions on Signal Processing, vol. 58, no. 6, pp. 3086–3096, 2010. View at Publisher · View at Google Scholar · View at MathSciNet
  17. Y. S. Shmaliy, “An iterative Kalman-like algorithm ignoring noise and initial conditions,” IEEE Transactions on Signal Processing, vol. 59, no. 6, pp. 2465–2473, 2011. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  18. S. Zhao, Y. S. Shmaliy, F. Liu, O. Ibarra-Manzano, and S. H. Khan, “Effect of embedded unbiasedness on discrete-time optimal FIR filtering estimates,” EURASIP Journal on Advances in Signal Processing, vol. 2015, no. 1, Article ID 83, 2015. View at Publisher · View at Google Scholar · View at Scopus
  19. S. Zhao, Y. S. Shmaliy, and F. Liu, “Fast Kalman-like optimal unbiased FIR filtering with applications,” IEEE Transactions on Signal Processing, vol. 64, no. 9, pp. 2284–2297, 2016. View at Publisher · View at Google Scholar · View at MathSciNet
  20. C. K. Ahn, S. Zhao, Y. S. Shmaliy, and H. Li, “On the l2-l∞ and H∞ performances of continuous-time deadbeat H2 FIR filter,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. PP, no. 99, pp. 1–1, 2017. View at Publisher · View at Google Scholar
  21. C. K. Ahn, Y. S. Shmaliy, P. Shi, and Y. Zhao, “Receding-Horizon l2-l∞ FIR filter with embedded deadbeat property,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 64, no. 2, pp. 211–215, 2017. View at Publisher · View at Google Scholar · View at Scopus
  22. C. K. Ahn and Y. S. Shmaliy, “New receding horizon fir estimator for blind smart sensing of velocity via position measurements,” IEEE Transactions on Circuits and Systems II: Express Briefs, vol. 65, no. 1, pp. 135–139, 2018. View at Publisher · View at Google Scholar
  23. Y. Xu, Y. S. Shmaliy, Y. Li, and X. Chen, “UWB-based indoor human localization with time-delayed data using EFIR filtering,” IEEE Access, vol. 5, pp. 16676–16683, 2017. View at Publisher · View at Google Scholar · View at Scopus
  24. R. Rasoulzadeh and A. M. Shahri, “Accuracy improvement of a multi-MEMS inertial measurement unit by using an iterative UFIR filter,” in Proceedings of the 25th European Navigation Conference, ENC 2017, pp. 279–286, che, May 2017. View at Publisher · View at Google Scholar · View at Scopus
  25. S. H. Han, W. H. Kwon, and P. S. Kim, “Quasi-deadbeat minimax filters for deterministic state-space models,” Institute of Electrical and Electronics Engineers Transactions on Automatic Control, vol. 47, no. 11, pp. 1904–1908, 2002. View at Publisher · View at Google Scholar · View at MathSciNet · View at Scopus
  26. W. H. Kwon and S. H. Han, Receding Horizon Control: Model Predictive Control for State Models, Springer Science & Business Media, 2006.
  27. Y. S. Shmaliy, S. H. Khan, S. Zhao, and O. Ibarra-Manzano, “General unbiased FIR filter with applications to GPS-based steering of oscillator frequency,” IEEE Transactions on Control Systems Technology, vol. 25, no. 3, pp. 1141–1148, 2017. View at Publisher · View at Google Scholar · View at Scopus
  28. N. Xu, Stochastic Signal Estimation and System Control, Beijing Industry University Press, Beijing, China, 2001.
  29. G. Fuchs, V. Šátek, V. Vopěnka, J. Kunovský, and M. Kozek, “Application of the Modern Taylor Series Method to a multi-torsion chain,” Simulation Modelling Practice and Theory, vol. 33, pp. 89–101, 2013. View at Publisher · View at Google Scholar · View at Scopus