I am trying to distinguish between fasteners with cracks and those without (blanks). Measurements on both produce several parameters that can be used to characterize the fastener. The approach we are trying is to treat the blanks as a cluster in the parameter space with a centroid and covariance matrix. We want to establish a limit on the distance (Mahalanobis) such that x% of the blanks will lie within the limit y% of the time. In this way, it is very much like a tolerance limit on a population except that we have the square of a normally (hopefully) distributed variable instead of the variable itself.