Fault diagnosis in chemical processes with application of hierarchical neural networks

In this paper the chemical process fault diagnosis is considered. It is offered to apply a hierarchical neural network (NN) model built and trained with application of the expert analysis, which results in improvement of NN model architecture and training set selection. At a high level, the network serves for localization of the process faults, and at a low level, a set of networks identifies the causes of these faults. РСА is used for essential dimensionality reduction for the high level network. The decomposition of expert knowledge of the monitoring process determines the number of neurons in the output layer. The training set for that network includes projections of the fuzzy values of diagnostic parameters (the scores) onto the principal component space. These values characterize normal and abnormal states of the process and they are obtained from analysis of the expert information and HAZOP (hazard and operability) analysis of the process. Dimensions of the low level networks are rather low; therefore the scaled process variables can be fed directly into the network. Such an approach reduces the networks retraining time essentially. It also keeps the advantages of expert systems without their explicit introduction in the diagnostic system structure. In particular, if a new situation occurs, the faulty section of the process under control will be defined even in case the network of the low level does not identify this fault. The efficiency of the suggested method is shown on the example of the faults diagnosis of the hydrocarbons pyrolysis process.