Learning Time-optimal Anti-swing Trajectories for Overhead Crane Systems

Abstract

Considering both state and control constraints, minimum-time trajectory planning (MTTP) can be implemented in an ‘offline’ way for overhead crane systems [1]. In this paper, we aim to establish a real-time trajectory planning model by using machine learning approaches to approximate those results obtained by MTTP. The fusion of machine learning regression approaches into the trajectory planning module is new and the application is promising for intelligent mechatronic systems. In particular, we first reformulate the considered trajectory planning problem in a three-segment form, where the acceleration and deceleration segments are symmetric. Then, the offline MTTP is applied to generate a database of minimum-time trajectories for the acceleration stage, based on which several regression approaches including Extreme Learning Machine (ELM) and Backpropagation Neural Network (BP) are adopt to approximate MTTP results with high accuracy. More important, the resulting model only contains a set of parameters, rather than a large volume of offline data, and thus machine learning based approaches could be implemented in low-cost digital signal processing chips required by industrial applications. Comparative evaluation results are provided to show the superior performance of the selected regression approach.