By using an estimator, it is possible to substitute the measuring of difficult-to-measure process variable in which process the estimation relies on information about other process variables that are measured by sensors (so called easy-to-measure variables). In order to perform a continuous estimation, we need a mathematical model of the process which connects the chosen easy-to-measure variables with a difficult-to-measure variable that is being estimated. As there is usually not enough preliminary knowledge about the process, the process model building is based on the available measuring data. The basis is a general model structure which can approximate any function of input-output mapping, including presumptions of model dimension and estimation of its parameters based on the available measuring data. In economic terms, the most suitable model process building is the one based on plant data contained in the process data base. When modeling on these data the usual approaches to model building very often face serious mathematical problems and result in a model with unsatisfactory prediction properties. To build a model using plant data, the thesis proposes methods based on input space projection into a latent space and on regression on these new (latent) variables. From a mathematical point of view, this model presents a composition of two multivariable functions, the parameters of which are optimised through separate criterion functions, which allows better model structuring for a specific modeling task and more reliable model parameters estimation. As model properties are mostly defined by way of input space mapping, the thesis focuses on this aspect of model building. Several selected methods of input mapping are investigated in two cases of model process building.