Special Genetic Identification Algorithm with smoothing in the frequency domain

• We extended our identification with a smoothing technique in frequency domain.
• We identify nominal parameters of a PT1-system and a real robot for test purpose.
• We compare the computation cost, accuracy and precision with a standard algorithm.
• Smoothing in frequency domain decreases the computational cost.
• Smoothing in frequency domain increases the accuracy and precision.

Due to the increase in speed and lightweight construction, modern robots vibrate significantly during motion. Thus, accurate mechanical modeling and detailed controller behavior is essential for accurate path planning and control design of robots. For the suppression of undesired vibrations detailed models are used to develop robust controllers. Least square identification methods require deep insight in the analytical equations and thus are not very suitable for identification of different highly nonlinear robot models. Recently, we presented our genetic parameter identification in Brussels, Ludwig and Gerstmayr (2011). It minimizes the error of measured and simulated quantities. Highly efficient models in the multibody system tool HOTINT lead to short computational times for various simulations with different parameters. The simulation models can easily be assembled by engineers without a detailed knowledge of the underlying multibody system. As drawback of genetic optimization, many sub-minima were detected. Many simulations were required for the determination of the global minimum. Our current approach was to extend our previous algorithm. Measured and simulated quantities are transformed into the frequency domain. In contrast to previous work, Ludwig and Gerstmayr (2013), amplitude spectra of measured and simulated quantities are smoothed prior to the L2-norm computation. The presented method is tested using small scale test problems as well as real robots. Smoothing in the frequency domain leads to a smaller number of simulations needed for obtaining higher accuracy. It turns out that the presented algorithm is more accurate and precise than a standard algorithm and reduces the computational cost.