Fault detection in chemical plants using neural networks

The suitability of pattern recognition for process monitoring of chemical plants is discussed. Experiments in a miniplant, a pilot plant and simulation studies are carried out. While selecting the required test series of process variables when training neural networks, it is tried to use generalized forms of description to illustrate the system under question. It is therefore possible to combine data records originating from various sources. Thus, on the one hand, non-conforming operating conditions have to be simulated in a laboratory or technical system. On the other hand, simulation results might also be used to provide training on neural nets. This combination within the utilized data material allows for a dispense of preparing new physical-chemical models for each data-driven model. The prepared tool is subsequently used as a prototype for hydrogenation in a production system.