Subproject M1: Systematic over-instrumentation
The project Systematic over-instrumentation concentrates, firstly, on methods for extending the instrumentation (sensors, actuators) in the early phase of process maturation. For later on, when the maturation arrives at an advanced stage, also methods for the reduction of the instrumentation are a topic of M1.
To obtain a better process understanding, the process and its sub-processes are generously equipped with sensors that promise to deliver informative data, and with actuators that allow for a strong influence on the process. We callthis massive deployment of sensors and actuators ‘over-instrumentation’. Bayesian Optimal Experimental Design (BOED) methods will be investigated and extended in terms of the choice and parameterisation of additional instrumentation. Also, methods of Optimal Sensor Placement (OSP) will be extended to optimise parameterisation other than location, and also to consider actuators.
To accelerate the relevant calculations, neural networks will be used to learn the reasoning of BOED as well as OSP in order to perform fast, approximate computation of these methods. Particular attention will be paid to approaches that allow to integrate formalised expert knowledge into BOED and OSP. Hidden variables of the process, which are causes of other process variables, play an important role to understand a process; they are called Latent Confounders (LCs). Obviously, the possibility of observing (or influencing) LCs with sensors (or actuators) seems to be a strong means to better control the process. For the discovery of LCs, a kind of ‘Do-operator’ will be investigated, based on feeding in low-amplitude modulations of process quantities via process actuators. By analysing the propagation of these signals throughout the process, causal relations and LCs should be discovered, and adequate sensors (or actuators) may be used to observe (or influence) the confounders.
Additionally, Variational Autoencoders (VAE) will be investigated to discover Latent Variables (LVs), with the aim of modifying the VAE so that the LVs get the meaning of LCs that can be instrumented. During the advanced stages of iterative improvement, reductions in the initial over-instrumentation become increasingly important. For example, if some of the improved sub-processes could operate without a dedicated control loop, or with fewer or cheaper sensors, the instrumentation can be reduced.
Over-instrumentation is essentially used as an intermediate measure to enable rapid understanding, exploration, and advances in the initially immature process. Methods of Sensitivity Analysis (SA), calculated by data-driven methods of arbitrary Polynomial Chaos Expansions (aPCE), will be used as a methodological complement to causality analysis to evaluate the importance of variables and their instrumentation. For the example process Stamp Forming, the methods of over-instrumentation are to be tested, validated, and improved.