For elastic drive systems the motor shaft position does not correspond to the load position. Therefore, load side position feedback is usually used in the control to achieve correct positioning of the system. This paper proposes an alternative approach, that allows for a correct positioning using the motor position feedback for control. This is achieved through the online correction of the reference position using a robust Extended Kalman Filter. Using the load feedback it estimates an idealised, offset-free motor position. The correction factor calculated from the difference between the measured and the estimated motor position is then used for reference position correction (RPC). The proposed RPC method is implemented on a belt-driven test bed and evaluated for different movements, load oscillations and significant parameter errors to show its functionality and its suitability for industrial application.

Several techniques related to control design rely on parametric system models. In the industry of servo control commissioning these techniques are not well established, mostly because success hinges on the selection of a suitable model. Automatic model selection in view of control design requires a control-relevant criterion for identification and nomination of the best model. If in addition a bright-grey box model is required, the dominant physical effects of the system under study should be included in the model automatically. In this paper the v-gap metric is compared with a robust-control-relevant identification criterion with known controller in view of control relevance and feasibility of the identification. A focus is laid on servo control design and experiments are performed on a storage and retrieval system with off-the-shelf industrial components. It is found that the theoretical properties of both criteria are not as different as one might expect. Practically, both criteria are not easy to use but the identification with known controller emphasises certain frequencies more dominantly than the v-gap metric making it even more difficult to obtain universal plant models under realistic conditions.

Real-time, high-rate Extended Kalman Filter (EKF) execution must include considerations of its computational load, which can pose a challenge for the implementation depending on the specific observer rate. While methods for the reduction of the computational load exist, this paper seeks to circumvent the problem, by reducing the EKF sampling rate. This is explored for the case of multi-rate sensor fusion for drive control applications where at least one sensor sampling rate exceeds the control cycle rate. A lower rate EKF is implemented where, in contrast to a single-rate EKF approach, none of the higher rate measurements are neglected, but instead collected and sequentially processed during each EKF execution. Two formulations based on this concept are introduced. The first optimises the estimation error, accepting a significant increase in computational load, while the second seeks the best compromise between estimation error and computational load.

Physically motivated parametric models are the basis of several techniques related to control design. Industrial model-based controller tuning methods include pole placement, symmetric optimum and damping optimum. The challenge is that the resulting model-based controller is satisfactory only if the underlying model is appropriate. Typically, a set of potential models is known a priori, but it is not known, which model should be used. So, the critical question in model-based controller tuning is that of model selection. Existing approaches for model selection are mostly based on maximizing accuracy, but there is no reason why the most accurate model should also be the optimal model for control design. Given the overall aim to design a high-performance controller, in this paper the best model is considered as the one that has the potential to give a model-based controller the highest performance. The proposed method identifies parametric candidate models for control design. Then, a nonparametric model is used to predict the actual performance of the various controllers on the real system. A validation with two industry-like testbeds shows success of the method.