A probabilistic system for autonomous tracking of excavated ore in mining would be highly valuable. Currently, mine sites can have a vast array of sensors; examples include excavator load cells, bucket volume estimators and haul truck strut pressures which can be used to gather information about the excavated ore at different locations. These sensors can be utilised efficiently to increase the accuracy of mine estimates by introducing a real-time ore tracking system. This requires the introduction of a probabilistic ore tracking framework to consistently fuse information from any sensor input. Current modelling practices are manually intensive, deterministic (no recognition of the quality of data) and generally not a complete end-to-end system. This paper will provide a novel method for tracking extensive ore properties (mass, volume) and intensive ore properties such as chemical composition (eg Fe per cent, SiO2 per cent, Al2O3 per cent). This method is proved to ensure that correct correlations between states are conserved as material is transferred, combined and/or separated at different locations with any configuration of sensor inputs. This ensures that consistency of estimates is maintained and other benefits such as real-time probabilistic reconciliation can be implemented. A method for combining material of different mass and chemical compositions is also developed to enable accurate and efficient operation of the system. Representative experiments on a small-scale and a larger field trial which mimic the mining processes of excavation, haul and unload to ROM Stockpile will be used to verify the approach.