Abstract In rocks that are polydeformed an approach which separates faults prior to stress inversion is more appropriate. The traditional stress inversion approach involving the concept of the best-fit stress tensor, e.g. a tensor which minimises the misfit between calculated and measured fault-striae data, often risks computing artificial stress tensors that are some form of average of mixed sets of real stress tensors. A new approach is proposed in which fault data are pre-processed to group the faults on the basis of their response to all possible orientations and magnitudes of applied stress. A computer method is described which utilises cluster analysis based on the right-dihedra method to divide dynamically-mixed fault populations to monophase subsets. This division is based on the ranked similarity coefficients of each fault pair from the raw data set. The data clusters form dynamically-homogeneous subsets, which are used for the composite right-dihedra solution. This solution is re-computed for the reduced stress tensor defined by the orientation of principal stress axes and the ratio of their magnitudes.