But when starting from measurement data, the main challenge is to find the underlying structure beneath the observations, or structure identification. As the philosophers say, this task of seeing the "idea world" is impossible, at least in principle - however, if appropriate assumptions are made, the structures can be determined within a reasonably chosen structural framework. So what is the general structure of measurement data? It is proposed that a multi-model framework consisting of sparsely coded linear submodels would be appropriate. Modeling of a specific system boils down to extraction of features from data having special structure. I have implemented a special "GGHA toolbox" for Matlab to make the developed algorithms better accessible. As a matter of fact, this same framework seems to give interesting results also when modeling cognitive phenomena.
The tools for managing the complexity are of utmost importance. The basic target in my studies has been to overcome the gap between the conceptual modeling and mathematical data processing, and to reach this, "smart" special-purpose representation formalisms are needed. For dynamic system modeling purposes, a language was developed that could be utilized as an interface between the high-level applications and the numerical data. This tool has been applied in various non-trivial tasks, and it has turned out that in many cases the process of system modeling can be completely automated.
My background is in computer science, and in that field I have been studying language compilers in general. Surprisingly, based on the theoretical studies, some interesting results concerning the analyzability of nonlinear dynamic systems were found.
What comes to my more "traditional" control engineering interests, pH process control needs to be mentioned; it turns out that the multi-model approach works fine when modeling this kind of a highly nonlinear and time-variant process. Fault detection is another topic that has been studied using different approaches: In the X-ray spectral analyzer case, statistical multivariate methods were applied, and in condition monitoring of electrical machines, support vector classification has successfully been applied. A longer line of research studies the possibilities of determining the state of different mineral manipulation processes using machine vision: Having huge amounts of pixel data available, clever compression and sensor fusion methods are needed.