Physics-assisted machine learning is a powerful framework that enhances data efficiency by integrating the strengths of conventional machine learning with physical knowledge. This paper applies this concept and focuses on the design of a driver evaluator using physics-assisted unsupervised learning, which serves as a virtual reference generator that provides different driving modes for vehicles equipped with active actuators. A strategy that applies sensitivity analysis regarding the vehicle handling performance, aiming to reduce the computational workload of the clustering algorithms, is proposed. First, a bicycle model with nonlinear Pacejka’s tire models is established for the analysis of lateral dynamics. Next, mathematical interpretations of sensitivity analysis are derived to evaluate the contribution of physical parameters to the system response and build the reduced parameters set. Then, Gaussian mixture models are fitted to a database generated with the full parameters set and another with the reduced set, respectively. Finally, step-steer and constant radius tests are performed to assess the handling performance with respect to the two validated centroids. Comparisons of lateral dynamics and understeer characteristics indicate that the proposed method can accurately distinguish driving modes in a much faster manner compared to traditional machine learning. This methodology has significant potential for practical applications with large databases and more complex systems.
Loading....