Incipient diagnosis of multiple faults in chemical processes via hierarchical artificial neural network

The authors describe a method for discriminating causes of multiple faults in chemical processes via artificial neural networks. A suitable hierarchical artificial neural network (HANN) is introduced. The HANN is composed of two stages of artificial neural networks with the structure of a multilayer network. Input nodes at the first stage receive the process measurements. Each output from the network in the first stage corresponds to the individual causes of the presumed fault and the normal situation. The second stage is composed of multilayer neural networks with each network connected to the outputs of the first stage. The networks in the second stage confirm the causes of the multiple faults. To improve the ability of the associate memory of the network in the first stage, the gains in the sigmoid functions of the artificial neurons in the first stage were optimized. The causes of multiple faults were clearly discriminated by the proposed HANN.<<ETX>>