Spatiotemporal Parameter Estimation of Thermal Treatment Process via Initial Condition Reconstruction Using Neural Networks

In this paper the design of control systems of periodical thermal treatment processes (TTP) with distributed parameters modeled by partial differential equations (PDEs) is considered. The main problem to decide is the estimation of the initial charge parameters—size, humidity, temperature and the relative load, which are all immeasurable. The investigation is based on first-principle models of the internal and external heat-exchange. Initially, after deriving the PDEs is created a representative TTP set by simulation using relevant combinations of charging parameters. To obtain their real estimates, the only measurable heating medium temperature, informative only during the first TTP stage, is applied. A cluster of N-nearest neighborhoods is found around the charge experimental temperature curve. A local situation-based dynamic neural network is learned to assess the charging parameters. They are implemented to define the optimal heating time of the current charge using another static neural network. Finally some aspects of industrial application of the proposed approaches are discussed.