سیستم کنترل یادگیری تکراری پیشرفته در ربات صنعتی به کار برده می شود

This paper proposes a model-based iterative learning control algorithm for time-varying systems with a high convergence speed. The convergence of components of the tracking error can be controlled individually with the algorithm. The convergence speed of each error component can be maximised unless robustness for noise or unmodelled dynamics is needed. The learning control algorithm is applied to the industrial Stäubli RX90 robot. A linear time-varying model of the robot dynamics is obtained by linearisation of the non-linear dynamic equations. Experiments show that the tracking error of the robot joints can be reduced to the desired level in a few iterations.

Laser welding has several advantages over conventional welding, e.g., the high depth-to-width ratio of the weld, the relatively small heat input and the high processing speed. To obtain defect free welds, the laser beam should typically track the weld seam with an accuracy in the order of 0.1 mm (Duley, 1998) at speeds beyond 100 mm/s. The demands on the orientation of the laser beam with respect to the weld seam are in the order of several degrees and not as restrictive as the demands on the linear tracking accuracy. Nevertheless, the required linear tracking accuracy puts high demands on the manipulator that moves the laser beam with respect to the weld seam. The industrial applicability of laser welding will be increased considerably by the use of commercially available six-axes industrial robots, as these robots can access complicated three-dimensional seam geometries. However, using standard industrial controllers the tracking accuracy of these robots appears to be insufficient for laser welding at high speeds. Since the (dynamic) repeatability of the robots is much better than their tracking accuracy it is expected that the tracking performance of these robots can be improved considerably with iterative learning control (ILC). ILC is a learning control technique for systems making repetitive movements. Each trial a feedforward is computed based on measurements of the error in the previous trials, such that the error converges to a small value. Since the pioneering work of Arimoto, Kawamura, and Miyazaki (1984) a vast amount of applications and implementations of ILC have been proposed. An overview of the work until 1998 is given by Moore (1998).

This paper presents a model-based iterative learning control algorithm for time-varying systems with a high convergence speed. The algorithm is based on the lifted system description. Singular value decomposition of the lifted system matrix is used to decouple the system equations. Analyses show that robustness of the learning controller for non-repetitive disturbances and model errors can be created at cost of convergence speed of the error components. A learning controller is proposed to control the components of the error individually. The convergence of part of the error components is limited by insufficient model knowledge or noise, while the convergence speed of other errors can be maximised. Using the standard industrial controller the tracking accuracy of the Stäubli RX90 robot is insufficient for laser welding at high velocities. Therefore, it is investigated if the performance can be increased with learning control. An LTV model of the robot is obtained by linearising its non-linear dynamics for small perturbations around a nominal trajectory. Disturbances and unmodelled robot dynamics are expected to affect the high-frequency error components. The proposed learning scheme is implemented for the robot such that the high-frequency components of the error are not compensated, while the remaining components are eliminated in one run. Experimental results show that most of the low-frequency errors can indeed be eliminated in one run, while the high-frequency components remain constant. After five iterations the accuracy of the tip-motion, estimated from the joint motion, is almost sufficient for high demanding laser welding tasks. Trying to compensate for the high-frequency error components as well results in diverging error components, probably due to modelling errors.