Controller tuning optimization and standardization of a CNC-grinding machine

Tuning of controller parameters is a highly relevant part of the controller design of a system, because of its high influence on the controllers performance. Manual tuning is a time consuming and difficult process for complex control systems, which results in a high variance of the systems performance, due to the absence of a rigorous definition which parameters to select and how to measure the performance. Thus a standardized, safe and efficient autotuning process that results in optimal performance of the system is a desirable feature for industrial machines and processes. In the scope of this work an efficient data-driven algorithm, based on Bayesian optimization, is developed for the tuning of a cascaded control structure. The Bayesian optimization algorithm uses a Gaussian process regression to model the performance of the system as an unknown objective function. Each iteration of the algorithm, the posterior distribution of the Gaussian process is updated and a constrained expected improvement acquisition function is maximized by particle swarm optimization to find the next parameter candidates in a systematic way. The algorithm is tested first on the simulation of a state-of-the-art CNC grinding machine, using an identified model of the grinding axes. On the simulation the algorithm is compared with relay feedback tuning and safe Bayesian optimization (SafeOpt). Afterwards the algorithm is tested on different grinding axes of the real CNC grinding machine and the result is evaluated, using a grinding process of a PCBN workpiece. Furthermore the algorithm is tested to detect faulty conditions of the CNC grinding axes. While the constrained Bayesian optimization algorithm shows a significantly better performance than the relay feedback tuning, it converges faster than the tested SafeOpt algorithm for the simulation experiments. Using the constrained Bayesian optimization tuning algorithm on the real system illustrates limitations for this method, caused by noise, rough quantization and differences between the mechanics during tuning and during grinding. The constrained Bayesian optimization tuning copes with the latter after the introduction of an algorithm to detect the critical gains of an axis and a performance metric of the cost, depending on the determined critical gains. During grinding this altered constrained Bayesian optimization tuning shows satisfying results and thus can be used as standardized autotuning method for the CNC grinding machines. Furthermore faulty conditions of the machine can be detected, if its influence on the position or critical gains is significant enough