Robust fault diagnosis of process systems using neural networks with ellipsoidal units

Suryanarasimham Venkata Kavuri, Purdue University

Abstract

The problem of fault diagnosis is important in process operations for maintaining a plant safely and efficiently. Neural networks are considered attractive as fault classifiers because of their ability to approximate arbitrary nonlinear decision boundaries. However, lack of control in the construction of the decision boundaries results in unintuitive extrapolations. To provide robustness to noise and to unknown disturbances, a network of ellipsoidal units is proposed. To overcome the problem of local minima traps and the difficulties in determining network structure, a two stage training procedure is adapted. In the first stage, unsupervised learning is performed that determines the network structure and initializes the network weights so that local minima traps may be avoided. Unsupervised learning using a fuzzy clustering scheme identifies groups of training patterns with similar characteristics. The fuzzy clustering procedure extracts a minimal number of features (hidden nodes) and thus avoid having to determine network structure by trial and error. Each of the clusters so formed is approximated by an ellipsoid each. The notion of a bounding ellipsoid is introduced to control the growth of ellipsoids. In the second stage, network weights are tuned using supervised learning to minimize the global error of misclassification. For large scale processes, network size, input size and the number of patterns are very large. Decomposition strategies are developed to reduce the network size, number of patterns and reduce the input size by decorrelation. An Amoco Model IV FCCU process has been considered for performing real-time diagnosis. A network of ellipsoidal units has been shown to successfully identify all the known faults as well as identify novel situations. Its superiority over other networks such as backpropagation network and RBF network has been demonstrated.

Degree

Ph.D.

Advisors

Venkatasubramanian, Purdue University.

Subject Area

Chemical engineering|Computer science

COinS