Neural network applications for detecting process faults in packed towers

Artificial neural networks can be used as a fault diagnostic tool in chemical process industries. Connection strengths representing correlation between inputs (sensor measurements) and outputs (faults) are made to learn by the network using the back propagation algorithm. Results are presented for diagnostic faults in an ammonia–water packed distillation column. First, a 6-4-6 network architecture (six input nodes corresponding to the state variables and six output nodes corresponding to the six malfunctions) was chosen based on the minimum root-mean-square-error and mean absolute percentage error; and a maximum value of the Pearson correlation coefficient (CP). The values of the learning rate, momentum and the gain terms were taken as 0.8, 0.8 and 1.0, respectively. The detection of the designated faults by the network was good. Relative importance of the various input variables on the output variables was calculated based on the partitioning of connection weights which showed that bottoms temperature, overhead composition and overhead temperature are not much affected by the disturbances in feed rate, feed composition and vapor rate in the given range. This resulted in a simplified 3-4-6 net architecture with similar capabilities as the 6-4-6 net thereby reducing the number of computations.