Model Selection ensuring Practical Identifiability for Models of Electric Drives with Coupled Mechanics

Abstract

Physically motivated models of electric drive trains with coupled mechanics are ubiquitous in industry for control design, simulation, feed-forward, model-based fault diagnosis etc. Often, however, the effort of model building prohibits these model-based methods. In this paper an automated model selection strategy is proposed for dynamic simulation models that not only optimizes the accuracy of the fit but also ensures practical identifiability of model parameters during structural optimization. Practical identifiability is crucial for physically motivated, interpretable models as opposed to pure prediction and inference applications. Our approach extends structural optimization considering practical identifiability to nonlinear models. In spite of the nonlinearity, local and linear criteria are evaluated, the integrity of which is investigated exemplarily. The methods are validated experimentally on a stacker crane.