A novel fault detection/diagnosis technique for linear dynamic systems is proposed. In comparison with existing schemes, the proposed method achieves fault detection/diagnosis using neither observer residuals nor parameter estimation errors; instead, it relies on the estimation of the underlying modal parameters of the system. The estimated modal parameters are compared with pre-calculated characteristic patterns of the system, which are represented as a set of root loci in terms of the physical system parameters. The modal parameter estimation is carried out using a numerically robust least-squares algorithm based on singular value decomposition. A pattern recognition technique employing linear multiprototype distance functions is used to classify the faults according to the variations of physical parameters. The proposed method has been applied to a simulated DC servo system where faults are introduced as abrupt changes in physical system parameters. It is shown that the proposed scheme is capable of diagnosing most of changes in physical system parameters.