Abstract: On-line estimation plays an important role in process control and monitoring. Obtaining a theoretical solution to the simultaneous state-parameter estimation problem for non-linear stochastic systems involves solving complex multi-dimensional integrals that are not amenable to analytical solution. While basic sequential Monte-Carlo (SMC) or particle filtering (PF) algorithms for simultaneous estimation exist, it is well recognized that there is a need for making these on-line algorithms non-degenerate, fast and applicable to processes with missing measurements. To overcome the deficiencies in traditional algorithms, this work proposes a Bayesian approach to on-line state and parameter estimation. Its extension to handle missing data in real-time is also provided. The simultaneous estimation is performed by filtering an extended vector of states and parameters using an adaptive sequential-importance-resampling (SIR) filter with a kernel density estimation method. The approach uses an on-line optimization algorithm based on Kullback-Leibler (KL) divergence to allow adaptation of the SIR filter for combined state-parameter estimation. An optimal tuning rule to control the width of the kernel and the variance of the artificial noise added to the parameters is also proposed. The approach is illustrated through numerical examples.
Introduction
In industrial process control, such as bioreactors or chemical plants, accurately estimating hidden states and unknown parameters in real-time is a critical challenge. While mathematical models provide a framework for these estimations, they often rely on parameters that are unknown or vary over time. Traditional methods like the Extended Kalman Filter (EKF) frequently fail when applied to highly non-linear systems. Particle Filters (PF) offer a more robust alternative by using a "cloud" of random samples (particles) to represent uncertainty, yet they face significant hurdles regarding sample degeneracy and variance inflation.
Comparative Methodologies
The choice of an estimation framework involves balancing computational efficiency with accuracy. Existing methods generally fall into two categories: Bayesian approaches and Maximum Likelihood (ML) approaches.
As summarized in the table above, ML methods like On-line EM are asymptotically efficient but computationally expensive ($O(N^2)$ per update) and prone to local optimality. Standard Bayesian methods using Artificial Dynamics are computationally lighter ($O(NT)$) but suffer from distribution alteration and are notoriously difficult to tune. This research addresses these specific Bayesian drawbacks by introducing an adaptive mechanism.
The Adaptive SIR Filter Framework
The proposed Adaptive Sequential Importance Resampling (Ad-SIR) filter integrates several key techniques to overcome the limitations of standard particle filters:
- Artificial Dynamics (ADA): Parameters are treated as slowly evolving states by introducing small amounts of Gaussian noise.
- Kernel Smoothing: To counteract the "blurriness" or variance inflation typically caused by ADA, a kernel density estimation method is employed. This shrinks particles toward their mean, maintaining a tight distribution.
- KL Divergence Optimization: The core innovation lies in the automated tuning of the kernel width ($h_t$). Rather than relying on heuristics, the authors derived an on-line optimization rule that minimizes the Kullback-Leibler (KL) divergence—a measure of information loss—between the predicted particle cloud and the actual measurement.
- Handling Missing Data: The algorithm is extended to handle sensor failures. When measurements are missing, the filter skips the update step and utilizes previous optimal tuning parameters to maintain estimation integrity.
Experimental Results
The Ad-SIR filter was evaluated using two complex synthetic benchmarks: a non-linear, non-Gaussian univariate system and an autonomous non-linear system involving squared states and cosine terms.
Key findings include:
- Robustness: In scenarios with up to 50% missing measurements, the algorithm successfully converged to true parameter values, albeit with a slight delay compared to full-data scenarios.
- Efficiency: The Bayesian Ad-SIR filter demonstrated significant speed advantages over ML methods. Processing 1,000 time steps with 20,000 particles took approximately 210 seconds on a standard desktop, making it suitable for high-speed industrial controllers.
- Comprehensive Estimation: Beyond model parameters, the filter uniquely estimated process and measurement noise variances ($Q_t$ and $R_t$) simultaneously.
Conclusion and Future Work
The Ad-SIR filter provides a computationally efficient and robust solution for real-time state-parameter estimation in non-linear systems. While it effectively manages sample degeneracy and missing data, future research is required to extend this framework to hybrid systems with discrete mode jumps and to address the "curse of dimensionality" in extremely high-dimensional parameter spaces. Additionally, while numerical results are promising, a formal mathematical proof of convergence for this adaptive scheme remains an open challenge.
Summary updated