Predicting Positioning Error and Finding Features for Large Industrial Robots Based on Deep Learning

In this study, we evaluated the motion accuracy of a large industrial robot and its compensation method and constructed an off-line teaching operation based on three-dimensional computer aided design data. In this experiment, we used a laser tracker to measure the coordinates of the end effector of the robot. Simultaneously, the end-effector coordinates, each joint angle, the maximum current of the motors attached to each joint, and rotation speed of each joint were measured. This servo information was converted into image data as visible information. For each robot movement path, an image was created; the horizontal axis represented the movement time of the robot and the vertical axis represented the servo information. A convolutional neural network (CNN), a type of deep learning, was used to predict the positioning error with high accuracy. Subsequently, to identify the features of the positioning error, the image was divided into several analysis areas, one of which was filled with various colors and analyzed by the CNN. If the prediction accuracy of the CNN decreased, then the analysis area would be identified as a feature. Thus, the features of the Y -axis positioning error were observed for teaching each joint angle in the opposite direction just after the start of the motion, overshoot of the rotational joint current, and the change in the swivel joint current.