Abstract

Processing plants in the industry are heavily instrumented with a large variety of sensors. The original purpose of the instrumentation is monitoring and control. In the last two decades the data being measured and stored has found a new application in the form of soft sensors. These are pieces of software that process together several measurements and that can be used for calculating new quantities based on the interaction among those measurements. They are, typically, developed by highly qualified experts. Classical control has progressed from simple PID-controllers to e.g. linear-model-based dynamic matrix optimisation or multivariable controllers for ensuring the fastest path to a new operating point. But more and more new ways of plant operations are imposed onto the operators. Not just securing operating points, but pursuing goals like energy efficient operation or contribution margin optimised production control have become equally important. Plant operators are thus faced with a dynamic process that has paradigm shift from two sides: they are assessed against flexible objectives like low energy consumption or sustainable use of feedstock while on the other hand needing to cope with permanently changing process hardware. This leads to a greater demand on the flexibility of the control, advice and optimisation methods. Processing plants are by no means static environments. On the contrary, they undergo steady process changes, development and updates, their components get older, contaminated and thus represent a highly dynamic environment. Due to their lack of self-learning or self-adapting capabilities and in order to cope with the changes, the soft sensor models have to be adapted by the experts which is an expensive procedure. They have to be recalibrated or even worse redesigned with often the same effort that was required for their original development. Therefore the methods applied in the soft sensors should possess high degree of adaptivity and the ability to react to changing environments. This talk will discuss the current limitations of existing approaches and challenges on the way to development of robust adaptive predictive modelling environments based on open, flexible, extendable architectures encompassing many self-* features required in modern applications. The discussed concepts will cover local learning, ensemble methods, meta-learning, concept drift, robust adaptation mechanisms etc. in the context of complex adaptive systems and eventual departure from the engineering maxim of “simple is beautiful” towards the biological statement of “complexity is not a problem”.