Implementing system identification for a vtol unmanned aerial vehicle

This work was motivated by a need for practical baseline system identification tools for UAV platform developers. Basis functions derived from physics can be tedious to develop, and errors in parameter estimation can lead to model divergence from measurements. Alternatively, grey and black-box modeling can be effective tools for data-driven system identification, enabling rapid prototyping of controllers. In this paper, the authors implement multiple algorithms to be used for system identification for an unconventional UAV. We demonstrate linear regression as a baseline model and Gaussian Processes as a black box model to represent the reduced order dynamics of the UAV. Additionally, we compare these models to a physics-based function model derived from first principles. The goal is to demonstrate a system identification workflow that can be utilized to learn a black-box model of a given data set, or fit a basis function model. In system identification we tackle the problem of learning propulsion system dynamics separately from flight dynamics. This decouples the learning problem and allows for the application of well known linear regression tools to both tasks. We show that the linear basis function model, the physics-based model, and the Gaussian Process model can be fit to experimental data with high accuracy, however long term prediction requires additional optimization. On simulation data the physics based model is able to replicate the dynamics after multi-step optimization, implying we have successfully learned the system dynamics given the correct set of basis functions. On experimental data, the long term prediction diverges for all models, and optimization is required. The Gaussian Process model provides a more robust tool for long term propagation, but future improvements are needed to improve the long term accuracy. As a matter of practical implementation, we have also shown that a discrete-time formulation using re-sampled raw logged data eliminates the need for taking derivatives or following prescribed regimes for gathering input data. Furthermore, data acquired from low-cost off-the-shelf electronics is shown to be comparable in quality to high-precision Vicon data, when implemented via appropriate filtering techniques.