Bayesian Probabilistic Numerical Methods in Time-Dependent State Estimation for Industrial Hydrocyclone Equipment

Chris J. Oates, Jon Cockayne, Robert G. Aykroyd, Mark GirolamiJournal of the American Statistical Association 2017 · arXiv

TL;DR

This paper introduces a novel Bayesian approach to estimate the internal state of industrial equipment like hydrocyclones in real-time. By explicitly modeling numerical errors in the underlying physical simulations, it provides more reliable uncertainty quantification, preventing overconfident and potentially dangerous inferences. This method is implemented using sequential Monte Carlo techniques and demonstrated on experimental data.

What Problem Does This Paper Address?

Industrial processes often rely on complex equipment like hydrocyclones for tasks such as separating particles from liquids. Ensuring these systems operate correctly and safely requires continuous monitoring of their internal state. However, directly observing these internal states is often impossible due to physical constraints (e.g., reinforced walls, opacity). Electrical Impedance Tomography (EIT) offers a non-invasive solution by estimating the electrical conductivity field within the equipment.

The challenge lies in the fact that EIT relies on solving complex physical models (governing equations) repeatedly. These models are typically solved numerically, introducing discretization errors. Traditional methods often ignore these errors, leading to an overestimation of confidence in the results. This can be problematic in industrial settings where misinterpretations could lead to incorrect control decisions and potentially dangerous situations. The paper addresses this by developing a method that explicitly accounts for these numerical errors, providing a more accurate and trustworthy assessment of the system's state and its associated uncertainty.

Methods

The core of this work is the development of Bayesian Probabilistic Numerical Methods. Instead of treating numerical solvers as black boxes, these methods model the uncertainty introduced by the numerical approximation of the physical governing equations (partial differential equations, PDEs).

Key methodological components include:

  • Bayesian Inverse Problems: The problem of estimating the internal conductivity field from external measurements is framed as a Bayesian inverse problem. This allows for the incorporation of prior knowledge and the quantification of uncertainty in the estimated parameters.
  • Probabilistic Numerical Methods: This is the central novelty. The paper treats the numerical solution of the PDE as a "sub-inverse problem." It uses probabilistic models to represent the uncertainty in the solution of the PDE itself, arising from discretization. This is achieved through a probabilistic meshless method that models the solution as a Gaussian process.
  • Sequential Monte Carlo (SMC): To handle the time-evolving nature of the problem and the complexity of the probabilistic models, the authors employ a sequential Monte Carlo (SMC) framework, specifically a particle filter. This allows for efficient computation and data assimilation over time.
  • Modeling Discretization Error: The method explicitly models discretization error as an unknown random variable, which is then propagated through the computations. This is a key differentiator from standard approaches.
  • Markov Temporal Evolution: A Markovian prior model is developed for the temporal evolution of the conductivity field, which is computationally convenient for the filtering formulation.

Datasets

The paper uses data from a controlled laboratory experiment designed to mimic the behavior of a hydrocyclone. This experiment involved injecting a concentrated potassium chloride solution into tap water in a cylindrical tank with electrodes. The data consists of voltage measurements taken at regular time intervals. This dataset was previously analyzed using non-probabilistic methods.

Key Insights and Ideas

  • Numerical Error is Crucial: Ignoring numerical discretization error in the governing equations can lead to highly biased and overconfident inferences, potentially causing mis-control in industrial applications.
  • Principled Computation-Precision Trade-off: Probabilistic numerical methods allow for a principled trade-off between computational cost and the precision of the results, by explicitly accounting for the uncertainty introduced by numerical approximations.
  • Uncertainty Propagation: The proposed method effectively propagates discretization uncertainty through time-evolving computations, providing a more realistic assessment of the overall uncertainty.
  • Time-Dependent State Estimation: The framework is designed for dynamic systems, enabling real-time monitoring and estimation of how the internal state changes over time.

Key Results

  • Improved Uncertainty Quantification: The probabilistic numerical methods provide an inflated posterior variance compared to methods that ignore discretization error. This inflation is crucial for reflecting the true uncertainty and preventing erroneous conclusions.
  • Convergence with Resolution: As the number of collocation points (used for discretizing the PDE) increases, the probabilistic method's results converge to a reference solution obtained with a high-resolution solver.
  • Demonstrated Efficacy on Experimental Data: The method was successfully applied to real experimental data from a mixing process, showing its practical applicability.
  • Computation-Precision Trade-off: Figures 6 and 9 visually demonstrate the trade-off: increasing the number of collocation points (higher precision) leads to a decrease in the integrated posterior standard deviation, but the probabilistic method consistently provides a more realistic (and often larger) uncertainty estimate.

Limitations and Future Work

  • Markov Temporal Evolution Assumption: The current approach assumes a Markovian temporal evolution for the conductivity field. Physical considerations suggest that this might be too simplistic, as non-Markovian temporal structures could encode useful information. Investigating how to incorporate this information into the prior without sacrificing the filtering framework is a future direction.
  • No Temporal Component in Potential Field Prior: The prior model for the electrical potential field did not include a temporal component, which was a choice for algorithmic convenience. The effect of this simplification on the results warrants further investigation.
  • Combining with Emulation: The authors suggest that combining their probabilistic numerical methods with emulation techniques could further reduce computational costs, especially for complex models.
  • Encoding Partial Prior Information: The statistical justification for encoding only partial prior information into the Markov prior is discussed, and future work could explore the potential loss of estimation and predictive efficiency from this simplification.

Summary updated

Abstract

The use of high-power industrial equipment, such as large-scale mixing equipment or a hydrocyclone for separation of particles in liquid suspension, demands careful monitoring to ensure correct operation. The fundamental task of state-estimation for the liquid suspension can be posed as a time-evolving inverse problem and solved with Bayesian statistical methods. In this paper, we extend Bayesian methods to incorporate statistical models for the error that is incurred in the numerical solution of the physical governing equations. This enables full uncertainty quantification within a principled computation-precision trade-off, in contrast to the over-confident inferences that are obtained when all sources of numerical error are ignored. The method is cast within a sequential Monte Carlo framework and an optimised implementation is provided in Python.
BibTeX
@article{oates2017bayesian,
  author = {Chris J. Oates and Jon Cockayne and Robert G. Aykroyd and Mark Girolami},
  title = {{Bayesian Probabilistic Numerical Methods in Time-Dependent State Estimation for Industrial Hydrocyclone Equipment}},
  journal = {Journal of the American Statistical Association},
  year = {2017},
}
Figure