Time-delay neural network for monitoring of abnormal situations in continuous industrial processes. Case Study: Evaporation station of a pulp mill*

Monitoring abnormal situations in continuous chemical process industries is a worldwide challenge. The occurrence of this kind of event is common, however its detection is generally after its development into a faulty condition. The earlier it is detected, the greater the chance to guarantee safe, economical and clean operations. This study develops a reliable and automatic system to detect and diagnose abnormal situations. It works as a temporal pattern classifier, which is based on a dynamic neural network, namely a Time Delay Neural Network (TDNN). The proposed methodology was tested on a real benchmark from an evaporation station. An initial comparison showed its better performance over the static Multi-Layer Perceptron (MLP) neural network. Its generalization capacity in distinguishing normal and abnormal operating regions was attested, and a final inspection showed its ability to absorb transitions between them. The global average rates of correct classification amount to 94.9% and 94.1%, respectively.