In this paper we propose a Fault Detection and Isolation (FDI) filter design method for Spark Injection Engines. Starting from a detailed nonlinear mean-value representation of the engine, a LPV approximation is obtained based on a judicious convex interpolation of a family of linearized models. A LPV-FDI filter based on a bank of Luenberger observers is synthesized by ensuring guaranteed levels of disturbance rejection and fault detection and isolation. The resulting diagnostic setup is parameter-dependent and uses a set of engine parameters, assumed measurable on-line, as a scheduling vector. The effectiveness of the LPV-FDI framework is illustrated by numerical examples where the diagnostic capabilities of the proposed FDI architecture are proved.