Real-Time Moving Horizon Estimation for Advanced Motion Control, Application to Friction State and Parameter Estimation. (Glijdende-horizon-schatters voor gevorderde bewegingscontrole, met toepassing op de schatting in reële tijd van wrijvingstoestanden en parameters)
Real-Time Moving Horizon Estimation for Advanced Motion Control, Application to Friction State and Parameter Estimation.
Friction is a nonlinear phenomenon that is present in almost all motion systems. Friction often limits systems performance by causing tracking and positioning errors and limit cycles. When these errors are unacceptable, the effect of friction needs to be compensated by the system controller, which estimates the friction force and feed this estimate back to the drives. Both model-free and model-based friction force estimation approaches exist. This thesis focuses on the modeling of friction and model-based friction estimation. Accurate model-based friction estimation is obtained provided that the friction model structure includes all major physical characteristics, and accurate real-time estimation of the model states and parameters is performed. This thesis presents a novel advanced friction model and a moving horizon estimator that allows to estimate model parameters and states in real-time. Moving horizon estimation (MHE) is an optimal control approach aiming to find the states and parameters of the system that are most consistent with current and past input-output data and the available system model. Moreover, real-time MHE is a gradient-based estimation technique that greatly benefits from a model that is differentiable with regards to state and parameter. Accurate friction models, which include all essential friction characteristics, as the generalized Maxwell-slip (GMS) model, are hybrid models with switching state conditions between presliding and sliding motion. To overcome these switching conditions, a smoothed version of the GMS model, called S-GMS, which consists of a set of differential equations well suited for gradient-based estimation is developed. Similar to the GMS model, the S-GMS model is a multi-state model that also describes all essential friction characteristics. A MHE friction observer is implemented for both the S-GMS model and the standard single-state LuGre model. Experimental state and parameter estimation shows the benefit of the multi-state S-GMS in presliding regime, where complex hysteresis behavior occurs.Moreover, a real-time embedded MHE friction observer is implemented for the S-GMS model via the automatic code generation tool ACADO and validated on a high-precision direct-drive linear motor. A sampling time in the millisecond range is achieved.