Automated weighing by sequential inference in dynamic environments

TL;DR

This paper presents a method for real-time mass estimation of a suspended object during filling using sequential inference techniques, demonstrating that physics-informed UKFs outperform particle filters in accuracy and speed.

Contribution

It introduces the use of physics-based UKFs for dynamic mass inference, showing improved performance over particle filters in a simulated environment.

Findings

UKFs with physics models provide rapid, accurate mass predictions.

Physics-informed inference outperforms non-physics methods.

UKFs outperform particle filters in this application.

Abstract

We demonstrate sequential mass inference of a suspended bag of milk powder from simulated measurements of the vertical force component at the pivot while the bag is being filled. We compare the predictions of various sequential inference methods both with and without a physics model to capture the system dynamics. We find that non-augmented and augmented-state unscented Kalman filters (UKFs) in conjunction with a physics model of a pendulum of varying mass and length provide rapid and accurate predictions of the milk powder mass as a function of time. The UKFs outperform the other method tested - a particle filter. Moreover, inference methods which incorporate a physics model outperform equivalent algorithms which do not.