Probabilistic Support of Operators
by Miroslav Kárny
The need for a completeness of advances technical solutions shifts interests of control theory to higher hierarchical levels up to the highest one humanly operated. Often, human operators supervising complex processes are left without a proper guidance and outcomes of their actions are strongly dependent on their experience and/or their mental state. This motivated a consortium (formed by the University of Reading, United Kingdom, Trinity College Dublin, Ireland, Institute of Information Theory and Automation AS, Compureg Plzen and Kovohute Rokycany, all Czech Republic) to put together a project that tries to create and verify industrially an adequate support of a generic nature.
A novel framework for creating of robust, fast, user-friendly advisory system has been proposed. The system supports operators of complex processes in making decisions. It guides them through high-dimensional noisy heterogeneous data spaces to successful configurations. It also warns them against dangerous configurations. Feasibility of the project was demonstrated in the 1st phase LTR Esprit project supported by European Commission within 4th Framework. Now the project will focus on reaching a full scale applicability of the tool called ProDACTool.
This implies addressing the following subtasks: (i) completion of the consistent novel probabilistic theory of dynamic adaptive clustering and theory of advising (ii) creating and thorough tests of the full set of robust and fast algorithms covering both experience accumulation and on-line advising to operators (iii) preparation of an adequate interaction sessions suitable for process staff (iv) implementation of the advisory systems with an intelligent graphical interface tailored to industrial environment (v) demonstration of the power of the tool on a full scale, cold-rolling mill application (vi) demonstration of the generic nature of the tool on data taken from other sites and fields (transportation, glass industry and medicine).
The project combines all measures to built and run a probabilistic, dynamical advisory system permanently on a cold-rolling mill. The care devoted to speed, internal consistency and robustness of information processing will make the system usable for other applications. Its work-packages will address:
- Theory: extension of Bayesian decision-making to dynamic adaptive operator support; robustness analysis; design of ergonomics; adaptation of data preparation techniques to the addressed problem; approximation techniques for high-dimensional spaces (acceleration of MT algorithm, solution of shadows cancelling problem, stability and robustness of algorithmic kernel).
- Algorithms: robust, factorised mixture estimation dealing with mixed continuous-discrete data; parallelisation of critical steps and adaptation of robust technique of stabilised forgetting; extensive ‘quality assurance’ tests on simulated and real data; benchmarking with respect to available competitors; implementation to industrially applicable software environment.
- Implementation: data collection with the necessary changes of sensors and information system; data analysis; preparation of intelligent industrial graphical interface; industrial tests.
- Application: full-scale use on a cold-rolling mill and evaluation of the impact; preparation of tools for other applications (other rolling sites, transportation, glass industry and medicine).
The generic nature of the advisory system will be tightly observed by creating robust and fast implementation that can function beyond boundaries of the application considered. This will be especially reflected in the quality assurance of the implemented system (consisting of a suite of software modules). It must meet the highest standards of speed, performance efficiency and robustness. For strategic evaluations, a robust parallel design will be used.
Effectiveness of the tool is to be guaranteed via a sophisticated graphical front-end, whose logic will respect both the underlying theory and the adapted know-how on man-machine interface.
The underlying theory can be characterised as an extension of dynamic Bayesian decision making under uncertainty applied to a (mixed) mixture models. The power of the decisive novel efficient estimation algorithm is illustrated on a simple search for structure of a transition probability matrix (TPM) using very limited amount of data. Figure shows the mixture fitted to matrix with 2500 entries using 1000 data. It nicely determines areas of TMP with significantly changing entries.
Please contact:
Miroslav Kárny - CRCIM
Tel: +420 2 6605 2274
E-mail: school@utia.cas.cz
