On Innovation-Based Triggering for Event-Based Nonlinear State Estimation Using the Particle Filter

Abstract

Event-based sampling has been proposed as a general technique for lowering the average communication rate, energy consumption and computational burden in remote state estimation. However, the design of the event trigger is critical for good performance. In this paper, we study the combination of innovation-based triggering and state estimation of nonlinear dynamical systems using the particle filter. It is found that innovation-based triggering is easily incorporated into the particle filter framework, and that it vastly outperforms the classical send-on-delta scheme for certain types of nonlinear systems. We further show how the particle filter can be used to jointly precompute the future state estimates and trigger probabilities, thus eliminating the need for periodic observerto-sensor communication, at the cost of increased computational burden at the observer. For wireless, battery-powered sensors, this enables the radio to be turned off between sampling events, which is key to saving energy.

Improved event-based particle filtering in resource-constrained remote state estimation

“…For a closed-loop trigger, we would further need to calculate H k based on the posterior approximation at time k − 1. For the IBT, this would resort to approximating the mean of the predictive likelihood as follows [12]:…”

“…Using the precomputation or periodic downlink approach, H k is instead computed from some measure on the onestep prediction of the state distribution p (x k | Y 1:k−1 ), e.g. as shown in (12) for IBT. Using the BPF, an estimate of p (x k | Y 1:k−1 ) can be obtained directly from the particle propagation.…”

“…An illustration of the closedloop triggering strategy is given in Figure 1. Considering the error/communication trade-off, it is preferable to use a closed-loop trigger [8], [12], but to actually obtain the observer information at the sensor is a potential problem. Traditionally, it has either been solved by letting only the sensor-to-observer communication be event based [11], [13], or by running a local filter or predictor at the sensor [5], [6], [14]- [17].…”

“…How limitations such as these affect the usability of particle filters in event-based systems remains a largely unexplored area. The aim of this paper, which is an extension of our two conference papers [12], [22], is to shed some light on these issues and to present improvements to increase the applicability of particle filters under event-based sampling.…”