Data-Driven Identification of Noise Covariances in Kalman Filtering for Virtual Sensing Applications

Szymon Greś, Michael Döhler, Vasilis Dertimanis, Eleni ChatziLecture Notes in Civil Engineering 2024 · DOI · OpenAlex

Abstract

Optimal Kalman filtering in virtual sensing requires precise knowledge of process ($Q$) and measurement ($R$) noise covariances, which are often unknown in practice. This paper introduces a streamlined, data-driven approach using subspace identification and linear regression for basis alignment. By extracting noise profiles directly from system residuals, the proposed method avoids the computational complexity of traditional iterative optimization techniques like Autocovariance Least-Squares (ALS). Validation on a 6-DOF mechanical system demonstrates that this "one-shot" method provides more accurate state estimates at unmeasured locations with significantly lower computational overhead.

Introduction: The "Black Magic" of Kalman Filtering

In structural health monitoring, "virtual sensing" allows engineers to estimate dynamic states—such as displacements in a bridge or machine component—at locations where physical sensors cannot be installed. While the Kalman filter is the standard tool for this task, its optimality depends heavily on the accurate specification of noise covariances ($Q$ and $R$).

In real-world applications, these parameters are rarely known. Engineers typically rely on heuristic tuning or complex, non-convex optimization algorithms like Autocovariance Least-Squares (ALS). These methods are often computationally expensive and numerically unstable. This research proposes a direct, data-driven alternative that identifies these noise parameters from system residuals, bridging the gap between physical models and raw data.

Methodology

The proposed technical strategy centers on subspace-based noise covariance estimation. The process involves several key steps:

  1. Subspace Identification: The N4SID algorithm identifies state sequences from input/output data.
  2. Basis Alignment: Subspace methods often identify systems in an arbitrary mathematical basis. The authors introduce a change-of-basis matrix ($\hat{T}$) using linear regression to align data-driven states with the existing physical model matrices ($A, B, C, D$).
  3. Residual Extraction: Once aligned, residuals are calculated as the difference between model predictions and observed data.
  4. Covariance Calculation: Noise covariances ($\hat{Q}, \hat{R}, \hat{S}$) are computed directly from the sample covariance of these residuals.

This "one-shot" approach eliminates the need for the iterative optimization loops required by methods like ALS.

Experimental Setup

The method was validated using a numerical 6-DOF mechanical chain-like system.

Schematic of the 6-DOF mechanical chain system used for validation, showing masses, springs, and sensor locations.

The system consists of a series of masses connected by springs and dampers. Inputs were simulated as white noise excitation at all degrees of freedom (DOFs), with 1% process noise. Acceleration measurements were taken at DOFs 1, 3, and 5, with 5% measurement noise. The primary goal was to predict displacements at the unmeasured locations (DOFs 2, 4, and 6).

Results and Discussion

The performance of the subspace-based method was compared against the "true" noise parameters, the ALS method, and a Latin Hypercube Sampling (LHC) random search.

Distribution of mean RMSE on virtual measurements for LHS samples, comparing the proposed method (red circle) against the true values (yellow square) and ALS (purple star).

Key Findings:

  • Superior Accuracy: In Monte Carlo simulations (1,000 samples), the proposed method consistently yielded a lower Root Mean Square Error (RMSE) for state predictions compared to the state-of-the-art ALS approach.
  • Statistical Robustness: The RMSE distribution for the subspace method was tighter and closer to the theoretical limit (the error achieved with true noise values) than the ALS method.
  • Consistency vs. Uniqueness: While the identified $Q, R, S$ triplet may not match the physical "true" values exactly, they are consistent for the system in innovation form. This ensures the resulting Kalman gain is optimal for filtering performance.
  • Computational Efficiency: By replacing iterative optimization with a regression-based residual analysis, the method is significantly more suitable for real-time or large-scale applications.

Conclusion and Future Work

The study demonstrates that a data-driven approach can effectively fill the gaps in physical models regarding noise statistics. While validated on a standard 6-DOF benchmark, future research will address the challenges of real-world structures, including non-linearities and non-stationary noise. Scaling the algorithm for large-scale mechanical systems, such as aircraft wings or full-scale bridges, remains a priority to confirm its computational advantages in high-dimensional state spaces.

Summary updated

BibTeX
@inproceedings{gres2024data,
  author = {Szymon Greś and Michael Döhler and Vasilis Dertimanis and Eleni Chatzi},
  title = {{Data-Driven Identification of Noise Covariances in Kalman Filtering for Virtual Sensing Applications}},
  booktitle = {Lecture Notes in Civil Engineering},
  year = {2024},
  doi = {10.1007/978-3-031-61425-5\_36},
}
Figure