Soft sensors offer several practical advantages for the process control of comminution circuits by combining multiple conventional sensors and other data sources with real-time analysis using mechanistic, phenomenological or machine learning models. While there has been a growth in the use of machine learning techniques for developing soft sensors in minerals processing applications, this paper describes advantages and challenges of using soft sensor models developed on the basis of process knowledge. Recent experience with three new soft sensor technologies developed by the authors are discussed. The Mill Filling Inference Tool which has been implemented at several mine sites, compares SAG mill load cell and power draw measurements with outputs of a semi-mechanistic power model to estimate the mill ball and total filling in real time. The progressive wear of the mill liners is modelled, allowing the model to correct the load mass due to the mass loss automatically and to estimate the charge trajectory. This information enables the mill to be operated more confidently near its limits. A coarse-ore stockpile soft sensor as been recently developed using a 3D cellular automata-based model to track ore size throughout the pile, and estimating the live capacity and size at each stockpile feeder to the primary mills. Finally, a hydrocyclone soft sensor is being developed to monitor performance using available hydrocyclone models to estimate the size and flow balance around the cyclone in real time. The model-based soft-sensor approach aims to synthesise and analyse existing data and provide operators and engineers with process- and model-relevant information that can facilitate making better process decisions. Maintenance data has proven useful for model development as it allows soft sensors to automatically adjust and compensate for the effect of progressive wear, for example for mill liners, thus helping the models to maintain calibration throughout the equipment life.