This work contributes to develop a new algebraic and deterministic approach of fault diagnosis, which is not based on an a priori explicit model of the process and presents a new perspective based on the theory of distributions and the pseudospectra analysis of matrix pencils. Our approach is based on certain tools and developments in the theory of algebraic estimation usual in automatic signal processing communities. The considered systems form a particular class of uncertain linear dynamical systems. The treated faults are actuator and sensor faults modeled through structured signals. The main advantage of this new approach is to realize fault diagnosis only from knowledge of input and output measurements and without identifying explicitly model parameters. The method exploits the fact that a structured signal satisfies a differential equation. The decision is entirely based on the temporal evolution of the estimates of some constant features of the fault. The effectiveness of the proposed approach is illustrated through academic examples.