An adaptive neural network for on-line learning and diagnosis of process faults

Techniques to enable a radial basis function (RBF) network to exhibit online learning properties for process fault diagnosis are described. These methods were demonstrated in an application of the RBF network to the diagnosis of a range of both sudden and gradual faults in a simulated continuous stirred tank reactor (CSTR) process. The network was able to recognise and learn new fault conditions recursively, and also to successfully diagnose faults that had been previously encountered. The results demonstrate the potential of the approach for online fault diagnosis applications in real processes. The network can be considered as a type of process fault model which maps the process measurement space to a fault classification space. The network centres represent points in the process measurement space which correspond to transient and steady-state features of faults on the process. Further work includes investigating techniques for examining these features, either qualitatively or quantitatively, to provide a greater understanding of the network operation and the effects of process faults