In protein refolding processes, the scarce availability of online measurements hampers effective process monitoring. In this work, we developed a mechanistic soft-sensor for protein refolding based on online intrinsic fluorescence measurements of tryptophan and tyrosin. In validation experiments using two model proteins, lactate dehydrogenase (LDH) and galactose oxidase, the soft-sensor showed accurate estimates for the prediction of the total sum of folding products (NRMSE < 6.1%) by calculating the changing rate of the average emission wavelength. For refolding of the enzyme LDH it was possible to obtain separate predictions of native protein and insoluble aggregates. The soft-sensor design was further extended by a model-based observer approach using particle filtering to incorporate kinetic formulations as well as physical constraints. The novel approach enabled the analysis of kinetic mechanisms during rapid reaction dynamics and can therefore be seen as an enabler to achieve a better understanding of kinetic refolding mechanisms.