Black box variational inference to adaptive kalman filter with unknown process noise covariance matrix☆
Introduction
State estimation is essential for a wide range of engineering applications from time series analysis, navigation guidance, and target tracking, to name but a few. The Kalman filter (KF), as an optimal state estimator under the linear and Gaussian assumptions, is one of the most widely used state estimation methods in practice. Despite its wide applicability, KF is highly sensitive to the accuracy of the noise statistics [1]. In practical applications, as is often the case, the statistical information about the system models may be unknown or inaccurate. For example, the covariance matrixes of the process noise and the measurement noise may be unknown [2], [3]. In the case of unknown covariance matrixes, the state estimation performance of KF will be decreased and it may even diverge [4]. Due to this fact, in the KF community, special attention has been paid to develop adaptive filters with model uncertainty. In this paper, we focus on the model uncertainty of the process noise; specifically, the process noise covariance matrix (PNCM) is unknown.
In recent decades, a variety of filters have been proposed to address the state estimation problem with unknown noise covariance matrixes, such as robust KF [5], [6], adaptive KF [3], [7], [8], [9], [10], [11], [12], [13], and others [14], [15]. In the robust KF, the estimation performance is optimized for the worst case to ensure a lower bound of the performance. In the adaptive KF, the system state along with the noise statistics need to be estimated simultaneously to adjust the system models online. Generally speaking, the adaptive KF methods can be divided into covariance matching, maximum likelihood (ML), Bayesian methods, and so on. Among the above methods, the Bayesian methods have been intensively studied due to their ability to calculate the joint posterior distribution of the system state and noise covariance matrix, which integrates all available information contained in the measurement data. Unfortunately, direct Bayesian inference for the joint posterior distribution is usually analytically intractable and it is necessary to resort to approximate inference.
More recently, the variational Bayesian (VB) [16], [17], [18] method has been introduced into the filtering framework to achieve approximate Bayesian inference. In [11], a VB adaptive Kalman filter (VB-AKF) is proposed to jointly estimate the state and the variances of the measurement noise where the prior distribution of the variance is governed by an inverse-Gamma distribution. In [19], the VB-AKF is further extended to estimate the measurement noise covariance matrix (MNCM) with an inverse-Wishart distribution as its prior distribution. In addition, the VB methods have been extended to the jump Markov linear systems [20], nonlinear adaptive filters [21], [22], and others [23], [24]. However, the VB-based adaptive filters mentioned above are only suitable for online state estimation in the presence of the unknown MNCM. To the best of our knowledge, far too little attention has been paid to the online state estimation problem in the presence of the unknown PNCM based on the Bayesian inference method. In [10], the online state estimation with inaccurate PNCM is treated, in which the predicted error covariance matrix (PECM) rather than the PNCM is approximated by the VB method. However, in [10], the initial value of PECM is determined by the inaccurate PNCM through the time prediction procedure of KF, resulting in a limited performance improvement compared with KF due to the fact that it does not make use of the historical estimation information. Besides, the proposed adaptive filter in [10] cannot give the estimated value of the unknown PNCM. In [25], the authors experimentally show that the state estimation performance of the proposed adaptive filter in [10] may be worse than that of the traditional KF in some cases, which indicates that there is still much room for improvement.
In this paper, we address the online state estimation problem in the presence of the unknown PNCM. The main contributions of this paper are summarized as follows:
- (1)
We prove that the probabilistic model for the estimation of the joint posterior distribution of the system state and the PNCM is non-conjugate, which means that the estimation cannot be performed by the traditional coordinate-ascent variational inference (CAVI).
- (2)
A novel adaptive KF based on the black box variational inference (BBVI) method is proposed. By simply decomposing the joint posterior distribution into a structured formulation, the proposed filter can calculate the state and PNCM posterior distributions efficiently. a) The proposed filter uses the intrinsically Bayesian robust KF (IBR-KF) to compute the state posterior distribution, such that only the posterior distribution of the PNCM needs to be approximated; b) The BBVI online inference for the posterior distribution of the PNCM is derived, and a special case, when the structure of the PNCM is known, is discussed.
This paper is organized as follows: Section 2 presents the linear state-space models in the presence of unknown PNCM and reviews the variational inference. Section 3 presents the proposed BBVI based adaptive KF, which is the main contribution of the paper, and some important issues of the proposed filter are discussed. Simulation results are then illustrated in Section 4, and finally, the conclusions are given in Section 5.
Section snippets
The state-space models
Consider the following linear and Gaussian discrete time state-space models:Here, is the discrete time index, is the state vector, and is the measurement vector. and denote the state transition matrix and the measurement matrix, respectively. The process noise vector and measurement noise vector nk are uncorrelated sequence. In this paper, we assume that Q is unknown and time-invariant. As is well
Black box variational inference adaptive kalman filter
In this section, we propose a novel adaptive KF based on the BBVI method. It should be noted that the adaptive property of the proposed filter is attached by the BBVI method. Thus, we refer to the proposed filter as a black box variational inference adaptive Kalman filter (BBVI-AKF) for convenience. Section 3.1 proves that the probabilistic model for estimating the system state and PNCM is non-conjugate. In Section 3.2, we introduce a structured posterior formulation for efficient
Simulations
In this section, a target tracking scenario is carried out to evaluate the performance of the proposed BBVI-AKF. The target dynamic is modeled by a CV model where the state vector is ; The measurement vector zk consists of xk and yk. The state-space parameters Fk and Hk are given as follows:where the time interval is the sample length is . The PNCM Q is defined by Eq. (50) where . The MNCM is set to
Conclusions
In this paper, we propose an adaptive Kalman filter in the presence of unknown PNCM based on BBVI method. In order to explain the difficulty in the estimation of the joint posterior distribution of the system state and the PNCM, we prove that the probabilistic model for the variational inference is non-conjugate, which means that the estimation cannot be performed by the CAVI method. To reduce the difficulty of the approximation, a structured posterior model has been introduced, by which the
Declaration of Competing Interest
We declare that we have no financial and personal relationships with other people or organizations that can inappropriately influence our work, there is no professional or other personal interest of any nature or kind in any product, service and/or company that could be construed as influencing the position presented in, or the review of, the manuscript entitled, “Black Box Variational Inference to Adaptive Kalman Filter with Unknown Process Noise Covariance Matrix”.
References (36)
- et al.
A new autocovariance least-squares method for estimating noise covariances
Automatica
(2006) - et al.
Fixed-point iteration gaussian sum filtering estimator with unknown time-varying non-Gaussian measurement noise
Signal Process.
(2018) - et al.
Kalman filtering through the feedback adaption of prior error covariance
Signal Process.
(2018) - et al.
Analysis of discrete-time kalman filtering under incorrect noise covariances
IEEE Trans. Autom. Control
(1990) Linear estimation in the presence of errors in assumed
IEEE Trans. Autom. Control
(1967)- et al.
Adaptive kalman filtering for INS/GPS
J. Geod.
(1999) - et al.
Performance analysis of the kalman filter with mismatched noise covariances
IEEE Trans. Autom. Control
(2016) - et al.
Robust estimation with unknown noise statistics
IEEE Trans. Autom. Control
(1999) - et al.
Finite-horizon robust kalman filtering for uncertain discrete time-varying systems with uncertain-covariance white noises
IEEE Signal Process. Lett.
(2006) On the identification of variances and adaptive kalman filtering
IEEE Trans. Autom. Control
(1970)
Adaptive sequential estimation with unknown noise statistics
IEEE Trans. Autom. Control
A novel adaptive kalman filter with inaccurate process and measurement noise covariance matrices
IEEE Trans. Autom. Control
Recursive noise adaptive kalman filtering by variational Bayesian approximations
IEEE Trans. Autom. Control
Adaptive filtering with unknown prior statistics
Proceedings of the Joint Automatic Control Conference
Adaptive kalman filtering by covariance sampling
IEEE Signal Process. Lett.
Minimax state estimation for linear stochastic systems with noise uncertainty
IEEE Trans. Autom. Control
Seam tracking monitoring based on adaptive kalman filter embedded elman neural network during high-power fiber laser welding
IEEE Trans. Ind. Electron.
Pattern Recognition and Machine Learning
Cited by (15)
Maximum correntropy criterion variational Bayesian adaptive Kalman filter based on strong tracking with unknown noise covariances
2023, Journal of the Franklin InstituteRobust forecasting-aided state estimation of power system based on extended Kalman filter with adaptive kernel risk-sensitive loss
2023, International Journal of Electrical Power and Energy SystemsCitation Excerpt :Although existing studies have superior performance in power system FASE with non-Gaussian noise, inherently, most studies are based on the assumption that process noise and the measurement statistics are known. However, even during quasi-steady state conditions, the magnitudes of the process noise and measurement noise will change, which means that the true values of the process noise covariance matrix and the measurement noise covariance matrix also change [34–36]. When sudden state change occurs, if the noise covariance matrices are not adapted to the latest change of the system state, the performance of state estimator will be affected [37].
A Gaussian mixture regression model based adaptive filter for non-Gaussian noise without a priori statistic
2022, Signal ProcessingCitation Excerpt :In other words, when their assumptions hold, actually the distributions and statistics of measurement noises are known and only the parameters in the assumed distribution are unknown, which need to be calculated. In many existing adaptive Kalman filters (AKFs), such as maximum likelihood [5], covariance matching [6], correlation [7], variational Bayesian (VB) methods [8–11], it is assumed that measurement noises obey a Gaussian distribution. This means that the Gaussian distribution of the noise is known and only its covariance is unknown.
Variational Bayesian based adaptive PDA filter in scenarios with unknown detection probability and heavy-tailed process noise
2021, Journal of the Franklin InstituteCitation Excerpt :However, this filter addresses the conjugation problem by estimating the state prediction covariance, and it therefore does not estimate the process noise. To realize the conjugation of the heavy-tailed process noise, [28] used the black box VB method, which has high computational complexity, to estimate the process noise covariance. To improve the accuracy of multi-target tracking in scenarios with non-Gaussian heavy-tailed noise, [29] proposed a student's t-mixture model based multi-Bernoulli filter.
A generalized data assimilation architecture of digital twin for complex process industrial systems
2024, Measurement Science and TechnologySelf-Tuning Process Noise in Variational Bayesian Adaptive Kalman Filter for Target Tracking
2023, Electronics (Switzerland)