The derivation of bright-grey box models for electric drives with coupled mechanics, such as stacker cranes, robots and linear gantries is an important step in control design but often too time-consuming for the ordinary commissioning process. It requires structure and parameter identification in repeated trial and error loops. In this paper an automated genetic programming solution is proposed that can cope with various features, including highly non-linear mechanics (friction, backlash). The generated state space representation can readily be used for stability analysis, state control, Kalman filtering, etc. This, however, requires several special rules in the genetic programming procedure and an automated integration of features into the defining state space form. Simulations are carried out with industrial data to investigate the performance and robustness.

No abstract

For many problems in the field of control design phenomenological models are required so that the need for parameter identification of given model structures arises. These models can be combined with observers to derive the system states in operation in addition to the parameters. However, identification and observation are limited in accuracy due to the restriction to existing series sensors. In the case of electric drives it is possible that due to elasticities in the structure part of the system is vibrating while the position sensor measures a nearly constant position. In this paper, the use of additional acceleration sensors is investigated in terms of identifiability and observability, which are installed at different points of the structure. The analysis is traced back to measures on the sensitivity matrix, where the integrating behaviour of the plant and the combination of different sensor types (position, velocity, acceleration) require special consideration. An industry-like stacker crane is used as a testbed for validation. It is shown that both identifiability and observability can be improved by the additional sensors in many cases. There is a good agreement between the expected and measured frequency response when the acceleration sensor is mounted on the first or second mass. Deviations only occur when mounted on the load suspension device, which is assumed to be the third mass.

Physically motivated models of electromechanical motion systems are required in several applications related to control design, model-based fault detection and simply interpretation. Often, however, the high effort of modelling prohibits these model-based methods in industrial applications. Therefore, all approaches of automatic modelling / model selection are naturally appealing. An intuitive approach is to identify the parameters of several model candidates and to select the one with the best fit on unseen data. A shortcoming of this approach is that the chosen model may be one with high complexity in which some of the physically interpretable parameters are not practically identifiable and uncertain. Also, ambiguities in selecting the model structure would not be disguised resulting in false confidence in a chosen model. Designing a reasonable set of candidate models requires that distinguishability of models can be checked prior to the identification procedure.This paper proposes a strategy for frequency domain model selection. The resulting model is tailored to ensure practical identifiability of all parameters for the given excitation. The analysis is based on local sensitivity calculated for the frequency domain cost function. Also, the paper describes distinguishability analysis of candidate models utilizing transfer function coefficients and Markov parameters. Model selection and distinguishability analysis are applied to a class of models as they are commonly used to describe servo control systems. It is shown in experiments on an industrial stacker crane that model selection works with little user interaction, except from defining normalized hyperparameters and ensuring that the resulting model is sound. Distinguishability is analysed systematically for all models that result from rearranging actuator, sensor and spring-damper elements along a chain of discrete masses. It can be proved or disproved for almost all combinations of potential models.

No abstract