A SEMANTICALLY CONSTRAINED NEURAL-NETWORK FOR MANUFACTURING DIAGNOSIS

In an earlier work (Ransing et al. 1995), we represented the causal re lationship in a defect-metacause-rootcause form. This representation w as perceived to be of considerable importance to the research communit y as well as industry, as it is applicable to any form of manufacturin g process. Based on this representation we proposed 'A Semantically Co nstrained Bayesian Network' for the diagnostic problems (Lewis and Ran sing 1997). In this paper, we develop another popular Artificial Intel ligence tool, 'Feedforward Neural Network', for such diagnostic proble ms. The network is constrained to defect-metacause-rootcause topology and it has been shown that metacause concepts can be successfully asso ciated with the hidden nodes. The errors are calculated at both the ou tput layer and the hidden layer. Although the learning process is base d on the back-propagation algorithm with a momentum term, the weight c hanges would occur at a link connecting a node only if at least one of the nodes connected to it in the preceding layer has non-zero activat ion. The theoretical analysis of such constrained learning is given an d it is shown that the network behaviour is acceptable for the diagnos tic problems considered.