LEARNING SYNTHESIS KNOWLEDGE FOR RAPID PRODUCT DESIGN

An agile manufacturing environment must capture and leverage available information for maximum benefit. This is especially applicable to a high product variety scenario where many of the new products, despite external dissimilarity, can be viewed as variants of old designs - sharing parts, manufacturing setups and cost models. This calls for intelligent organization of archived design data that provides effective support in a new product design situation. This paper presents a neural network based approach to support knowledge based synthesis of mechatronic components. Our domain of application is electric motor design and manufacture. We present strategies for rapidly developing approximate synthesis information for new designs from initial specifications by learning over archived design data. The role of the neural network architectures is both to identify relevant past designs as well as predict a specification set for the new proposed design. We argue that such learned representations of data are well suited to the flexible requirements of agile design and manufacture. Finally we present some results from implementing the learning approaches over data from a commercial vendor catalog.