To overcome the limitations of the standard ellipsoidal unit neural networks, some new approaches used in ellipsoidal unit neural networks have been proposed. These new approaches address three main issues: firstly, to understand better and represent the nature of fault classification boundaries; secondly, to determine the network structure without the usual trial and error schemes; lastly, to avoid erroneous generalizations. The application in CSTR shows that the ellipsoidal unit networks can possess arbitrary nonlinear classifying ability, nonlinear interfacial describing ability, and obtain accurate and efficient diagnosis results.