Neural network based framework for fault diagnosis in batch chemical plants

In this work, an artificial neural network (ANN) based framework for fault diagnosis in batch chemical plants is presented. The proposed FDS consists of an ANN structure supplemented with a knowledge based expert system (KBES ) in a block-oriented configuration. The system combines the adaptive learn ing diagnostic procedure of the ANN and the transparent deep knowledge repr esentation of the KBES. The information needed to implement the FDS include s a historical database of past batches, a hazard and operability (HAZOP) a nalysis and a model of the batch plant. The historical database that includ es information related to normal and abnormal operating conditions is used to train the ANN structure. The deviations of the on-line measurements from a reference profile are processed by a multi-scale wavelet in order to det ermine the singularities of the transients and to reduce the dimensionality of the data. The processed signals are the inputs of an ANN. The ANNs outp uts are the signals of the different suspected faults. The HAZOP analysis i s useful to build the process deep knowledge base (KB) of the plant. This b ase relies on the knowledge of the operators and engineers about the proces s and allows the formulation of artificial intelligence algorithms. The cas e study corresponds to a batch reactor. The FDS performance is demonstrated through the simulation of different process faults. The FDS proposed is al so compared with other approaches based on multi-way principal component an alysis. (C) 2000 Elsevier Science Ltd. All rights reserved.