Methods of residual generation and evaluation for model-based fault diagnosis in uncertain linear dynamic systems are investigated with the aid of frequency domain approaches and H(infinity)-optimization techniques. Based on the factorization technique, the construction of residual generators in terms of a parametrization concept is proposed. With its help the fault detection and isolation problem with perfect model uncertainty decoupling is studied and, moreover, by its optimal approximation the problem of optimally robust residual generation is treated. The optimization problem is solved by H(infinity)-techniques. To increase the robustness of residual evaluation a frequency domain residual evaluation index is introduced, and optimal input adaptive fault thresholds are derived with respect to the frequency domain evaluation index. The results are illustrated in terms of an inverted pendulum example.