An Adaptive Connectionist Front-End for Automated Fettling

  • Authors:

  • John Keat;Velupillai Balendran;Kandiah Sivayoganathan;Tony Sackfield

  • Affiliations:

  • Manufacturing Automation Research Group, Department of Mechanical and Manufacturing Engineering, The Nottingham Trent University, Burton Street, Nottingham NG1 4BU, UK;Manufacturing Automation Research Group, Department of Mechanical and Manufacturing Engineering, The Nottingham Trent University, Burton Street, Nottingham NG1 4BU, UK;Manufacturing Automation Research Group, Department of Mechanical and Manufacturing Engineering, The Nottingham Trent University, Burton Street, Nottingham NG1 4BU, UK;Department of Mathematics and Statistics, The Nottingham Trent University, Burton Street, Nottingham NG1 4BU, UK

  • Venue:
  • Integrated Computer-Aided Engineering
  • Year:
  • 1999

Quantified Score

Hi-index 0.00

Visualization

Abstract

The manufacturing of metal castings produces objects with extraneous features due to the nature of the process. These features, called runners and risers, are removed by a hazardous manual process known as fettling. They differ in form, size and presence from casting to casting. Unfettled castings tend to be geometrically imprecise and often composed of surfaces of a free-form nature. This makes the automation of the process of fettling difficult as a fixed datum cannot be easily defined. Previous methods to automate the process have centred on directing a robotic tool through a predefined path, irrespective of the nature, or even absence, of any rogue features. We present here a new connectionist model, IvOR (Invariant Object Recogniser), for the adaptive recognition and assessment of free-form 3-D objects. It has been derived as the front-end to an intelligent automated fettling system. IvOR has two distinct phases for learning and recognition. For both phases the initial stages are the same, a surface type map (termed an HK map) is created from the local mean (H) and Gaussian (K) curvatures within the range data representation of an object. The HK map is then passed through a focusing pre-processor to remove positional and scale variance within the input image. In the learning stage the "focused" image is learnt by an ART2a Network. In the recognition stage rotational variance (relative to the master representation) is removed by a rotational match network based on the principles of Adaptive Resonance Theory (ART). Assessment for rogue features is done by anovel local matching method based on the local comparison of network weights. IvOR allows adaptation of object representations and learning of new objects, online, with recognition and assessment at speeds acceptable for a manufacturing process.