Priors-Informed Machine Learning for Dynamic System Modelling: Mecanum Wheel as an Example

An accurate dynamic model holds critical significance for model based applications such as state estimation, system control, monitoring and optimization. While first principles based dynamic modelling allows for clear interpretability, it faces difficulties for complex systems under uncertainties. Purely data-driven machine learning based black box modelling is easy to implement but has low interpretability, poor generalization and often demands extensive data. This study advocates hybrid modelling seamlessly integrating prior knowledge into machine learning for better modelling performance. Specifically, we focus on assessing the impacts of different levels of model priors on machine learning based modelling performance utilizing the Neural Ordinary Differential Equation framework. Comparative simulations on a Mecanum wheel mobile robot show that an increased level of prior modelling information effectively enhances the modelling performance.