NEURAL-NETWORK DECOMPOSITION STRATEGIES FOR LARGE-SCALE FAULT-DIAGNOSIS

To overcome the limitations of the black-box character of the standard neural network approaches, a network with ellipsoidal units has recently been proposed. This novel approach addresses three main issues: (a) to understand better and represent the nature of fault classification boundaries; (b) to determine the network structure without the usual trial and error schemes; and (c) to avoid erroneous generalizations. In this paper, we develop the ellipsoidal units approach further by addressing the problem of real-time large-scale fault diagnosis. For such applications neural networks become very large and complex, making the training and interpretation tasks time consuming and difficult. Networks with ellipsoidal units naturally lend themselves to the development of decomposition techniques that result in the training of smaller networks with fewer training patterns. Three decomposition strategies, namely, network decomposition, training set decomposition, and input space decomposition, have been developed for large-scale industrial processes. The results for the real-time diagnosis of an Amoco model IV FCCU simulation case study are discussed. Network size and diagnostic performance are compared with alternative approaches, such as backpropagation networks and radial basis function networks.