This paper presents a new approach to the problem of off-line model-based fault diagnosis for multivariable uncertain systems. The proposed method uses the generalized structured singular value and is based oil frequency-domain model invalidation tools. A novel step-by-step methodology for off-line fault identification and symptom-aided diagnostic is developed. Fault detectability and isolability issues are discussed within the new framework. Experimental results obtained by using real data front a three-tank system are reported, showing the potential of the proposed techniques.