Modelling of a 3D crane system using parallel dynamic neural networks

J. Deng, V. M. Becerra

    Research output: Chapter in Book/Report/Conference proceedingConference contribution


    Cranes are widely used in factories, shipboards, ports and construction sites. It is usually expected that the payload can be brought to its destination and be still at the end of its journey. Because the payload exhibits a pendulumlike swing motion, it is difficult to position the payload precisely. In order to achieve fast and precise payload positioning, control methodologies must be developed. In this paper, the authors have used a laboratory scale crane that is able to position a load in three dimensions. The 3D crane is fully interfaced to MATLAB and SIMULINK. The crane has three manipulated inputs, which are the voltages applied to three DC motors, and five measurements obtained via optical encoders. The payload position in cartesian coordinates (xc; yc; zc) can be inferred from the measurements by means of kinematic equations. As the 3D crane is a highly nonlinear multi-input multi-output system, its accurate modelling is an important step for the successful application of advanced control methodologies. It is intended to apply to this crane nonlinear control methods which employ empirical models based on Hopfield-type dynamic neural networks. It is imperative to find a effective method to identify the 3D crane system. Since it is highly nonlinear, it is very difficult to identify this system using a single Hopfield-type dynamic neural network structure. Given the input-output dimensions of the system, the training problem exhibits a large amount of local minima, which makes gradient-based methods difficult to be applied. Even using a genetic algorithm to choose the weights of a single dynamic neural network has proven difficult. Therefore, a new dynamic neural network structure is proposed in order to identify this 3D Crane. The method uses up to three parallel dynamic neural networks to identify the 3D Crane system sequentially. This dynamic neural network structure is trained by following the following stages: (1) The first neural network is trained using a set of input trajectories (ux(t); uy(t); uz(t)) over a period of time t ε [to, tf ] and the measured payload positions (xc(t); yc(t); zc(t)) and it gives the outputs (x1(t), y1(t), z1(t)). (2) The second neural network is trained to reproduce the residual trajectories (xc - x1, yc - y1, zc - z1) and it gives the outputs (x2, y2, z2). (3) The third neural network is trained to reproduce the residual trajectories (xc- x1 - x2, yc - y1 - y2, zc - z1 - z2) and it gives the outputs (x3, y3, z3). The results (x1+x2+x3, y1+y2+y3, z1+z2+z3) are compared with (xc, yc, zc). Each neural network is trained based on a genetic algorithm with real encoding. The experimental results show that this method is effective to identify this highly nonlinear system.
    Original languageEnglish
    Title of host publicationCybernetic intelligence, challenges and advances
    Subtitle of host publication2nd IEEE Systems, Man & Cybernetics United Kingdom & Republic of Ireland Chapter Conference
    EditorsB. P. Amavasai, A. G. Hessami, R. J. Mitchell
    Place of PublicationReading
    PublisherThe University of Reading
    Publication statusPublished - 2003


    Dive into the research topics of 'Modelling of a 3D crane system using parallel dynamic neural networks'. Together they form a unique fingerprint.

    Cite this