Neural network decomposition strategies for large-scale fault diagnosis.
Int J Control 1994 Jun; 59(3):767-792
Three decomposition strategies for ellipsoidal unit networks were presented as a means for reducing the impact of computational complexity. Neural networks using ellipsoidal activation functions were described. The computational demands of network training for the diagnosis of large scale processes led to the proposal of the decomposition strategies. In network decomposition, the network was divided into independently trained subnetworks for more efficient training. Training set decomposition allowed the identification and elimination of training patterns which had little effect on the training of a subnetwork. Principal component analysis was used to affect input space decomposition by reducing the number of inputs from irrelevant sensors. Together the network, training set, and input space decomposition strategies resulted in smaller subnetworks with fewer training patterns. Mathematical derivations of the orientation directions of a cluster, dimensionality reduction in cluster approximation, and residual variance were presented. The input layer, hidden and density approximation layers, classification layer, and output layer of the network classifier were described. The proposed and existing models were compared in realtime diagnosis of an Amoco model IV fluidized catalytic cracking (FCCU) process. Lists of known faults used for training and unknown faults used in testing were provided. The complexities of the network classifier were compared to those of networks with linear activation and radial basis functions (RBF). Except for changes in diesel feed composition, the ellipsoidal unit network did not mistake any of the unknown faults for known faults. The RBF model misdiagnosed five and the linear model ten of 16 unknown faults. The authors conclude that the decomposition techniques resulted in problem size reduction and improvements in identification of novel malfunctions.
NIOSH-Publication; NIOSH-Grant; Control-technology; Mathematical-models; Data-processing; Computer-models; Petroleum-industry; Information-processing
Chemical Engineering Purdue University West Lafayette, IN 47907
International Journal of Control
Purdue University West Lafayette, West Lafayette, Indiana