Firstly, a Hybrid Approach: Models and Process Data Go Hand in Hand
"For our analysis, we have brought two things together: Firstly, the physical laws that we have represented in a model – that is to say, the expert knowledge about the thermo-dynamic and chemical processes. And secondly, the data that various sensors determine about the measuring process, for example temperature and pressure. We use these where no physical data is available," explains Dr. Karl-Heinz Küfer, division manager at the Fraunhofer ITWM. So far, such sensor data have already been used to monitor pro-cesses and to be able to react in good time if, for example, pressure or temperature deviate. The team around the two researchers is using machine learning methods to raise this "data treasure," including the training of artificial neural networks. Models and process data complement each other beneficially.
The application possibilities are not limited to the chemical industry: Rather, advantages can be expected wherever processes with a large number of influencing factors have to be controlled – and cannot be described by measurements or process data alone. In the long term, according to the researchers' plan, the system should be able to operate in real time.