Probabilistic advisory system for operators can help with diagnostics of rolling mills

Abstract

Advisory system for operators of complex industrial processes has been developed and improved by an international team of scientists and people from industry since 2000. Main purpose of the advisory system is to help operator set up manually adjustable parameters of an industrial process, with the aim to reach required production quality. Industrial process is taken for a stochastic process and input signals of its control system are taken for random variables. Based on Bayesian probability theory, a software toolbox was created for handling mixtures of probability density functions describing behavior of the process. Advisory system was tested and pilot application was installed on rolling mills producing metal strips. During the tests, an idea emerged to exploit verified probabilistic approach for complicated diagnostic tasks too. This diagnostics is intended for recognition of process malfunction which cannot be easily revealed by analysis of particular single signals only but analysis in multidimensional data space must be involved instead. Main principle of the advanced diagnostic method consists in finding a representation of process behavior in a short history by a mixture of probability density functions called historical mixture. Process behavior in the latest time period is represented by actual mixture. Difference between historical and actual mixtures is evaluated by calculation of Kullback-Leibler divergence. Mixtures and divergences are calculated repeatedly in time and a big change in the divergence value can be used as a source of alarm for non-standard process behavior.