Identifying trending coefficients with an ensemble Kalman filter

Max Schwenzer, Giuseppe Visconti, Muzaffer Ay, Thomas Bergs, Michaël Herty, Dirk Abel

Abstract

This paper extends the ensemble Kalman filter (EnKF) for inverse problems to identify trending model coefficients. This is done by repeatedly inflating the ensemble while maintaining the mean of the particles. As a benchmark serves a classic EnKF and a recursive least squares (RLS) on the example of identifying a force model in milling, which changes due to the progression of tool wear. For a proper comparison, the true values are simulated and augmented with white Gaussian noise. The results demonstrate the feasibility of the approach for dynamic identification while still achieving good accuracy in the static case. Further, the inflated EnKF shows a remarkably insensitivity on the starting set but a less smooth convergence compared to the classic EnKF.

Introduction

In precision manufacturing, specifically milling, mathematical models are used to predict cutting forces, which are essential for adaptive control and preventing tool breakage. However, the coefficients in these models are not constant; as a tool wears down, the "true" coefficients drift over time. Standard identification methods like Recursive Least Squares (RLS) often struggle with non-linear models or measurement noise. While Bayesian filters like the Ensemble Kalman Filter (EnKF) excel at finding static parameters, they suffer from "ensemble collapse"—once the ensemble converges, the filter loses its ability to track moving targets.

Methodology: The Inflated EnKF (EnKF*)

To address the tracking problem, researchers developed a modified inflated EnKF (EnKF*). The approach utilizes the non-linear Kienzle force model, relating cutting force to chip thickness and width via exponential coefficients ($k_t$ and $m_t$).

The EnKF propagates an ensemble of 100 potential state vectors. To prevent collapse, the authors periodically "inflate" the ensemble by resetting its covariance (spread) to the initial state ($P_0$) while preserving the current mean. To ensure smooth transitions, only a subset (10%) of particles is inflated at each step. Box-constraints are also applied to keep parameters within physically realistic bounds.

Tracking of model coefficients $k_t$ and $m_t$ across static, ascending, and alternating scenarios. The inflated EnKF* (red) successfully follows the ground truth (dashed), whereas the classic EnKF (blue) collapses and fails to track trends.

Experimental Setup

The method was validated using simulated datasets of a straight milling operation on stainless steel (X5CrNi18-10). Three scenarios were tested:

  1. Static: Constant coefficients.
  2. Ascending: Gradual drift simulating tool wear.
  3. Alternating: Rapid changes in coefficients. Signals were augmented with white Gaussian noise (SNR = 15) to simulate factory conditions.

Results and Discussion

The performance was evaluated using Root-Mean-Square Error (RMSE) across 1,000 Monte Carlo simulations. The results, summarized in the table below, highlight the trade-offs between tracking capability and precision.

RMSE values for various EnKF* configurations compared to the classic EnKF across different scenarios.

In the static case, the classic EnKF achieved the highest accuracy (RMSE = 3.8), while the EnKF was slightly less precise (RMSE ~6.4) due to the jitter introduced by inflation. However, in trending scenarios, the classic EnKF failed significantly (RMSE = 23.2 for ascending, 50.7 for alternating). In contrast, the EnKF successfully tracked these trends with much lower errors (RMSE = 7.2 and 8.9, respectively).

A key insight is the "tunnel vision" of classic EnKF; it becomes so certain of its estimate that it stops exploring the search space. Inflation acts as a "reset button," forcing the filter to maintain uncertainty and continue tracking. While RLS often oscillated wildly under high noise, the EnKF* provided a much cleaner error profile.

Conclusion

The inflated EnKF* demonstrates a robust ability to track trending parameters in non-linear systems, even with poor initial guesses. While there is a trade-off in smoothness and a need for parameter tuning (inflation step and factor), the method significantly outperforms standard EnKF and RLS in dynamic environments. Future work will focus on validating these results with real-world measurement signals from physical milling machines.

Summary updated

Abstract

This paper extends the ensemble Kalman filter (EnKF) for inverse problems to identify trending model coefficients. This is done by repeatedly inflating the ensemble while maintaining the mean of the particles. As a benchmark serves a classic EnKF and a recursive least squares (RLS) on the example of identifying a force model in milling, which changes due to the progression of tool wear. For a proper comparison, the true values are simulated and augmented with white Gaussian noise. The results demonstrate the feasibility of the approach for dynamic identification while still achieving good accuracy in the static case. Further, the inflated EnKF shows a remarkably insensitivity on the starting set but a less smooth convergence compared to the classic EnKF.
BibTeX
@inproceedings{schwenzer2020identifying,
  author = {Max Schwenzer and Giuseppe Visconti and Muzaffer Ay and Thomas Bergs and Michaël Herty and Dirk Abel},
  title = {{Identifying trending coefficients with an ensemble Kalman filter}},
  booktitle = {IFAC-PapersOnLine},
  year = {2020},
  doi = {10.1016/j.ifacol.2020.12.1490},
}
Figure