One primary goal in the application of process analytical technology tools is improved process monitoring and control. A second is to obtain a better understanding of how a normal process behaves (i.e. the normal dynamics). In order to perform feed-forward control, time series models of the process data are required. Such models could be developed on the basis of known physical/chemical knowledge of the system (i.e. first principal or mechanistic modeling). However, very often, this is not possible because of the lack of sufficient information. This leads to the need of system identification (SI). One class of models within SI is the state space models, linear models that relate the input of the system at time k to the output at time k via estimation of the so-called system states. State space models may be fitted using what is known as the subspace methods. Subspace methods are based on the projection of data on subspaces identified by, for example, the singular value decomposition of time-shifted data during a training phase. This paper introduces state space models, illustrates how subspace methods are closely related to known chemometric tools, and how they can be applied in, for example, model-based feed-forward process monitoring and control. The concepts are illustrated using a data set from an intrinsically nonlinear milk coagulation process that can be approximated well by a linear dynamic model using a small set of virtual (or principal) states. We present an alternative process-monitoring strategy where the dynamic components and boundary conditions of a developing milk coagulation batch are estimated in real-time and compared to normal operating conditions.