This paper deals with a new causal graph-based fault detection scheme. The Causal Ordering Graph (COG) is herein involved to rise above difficulties of modeling dynamic systems due to both qualitative and quantitative reasonings. For the fault detection purpose, the generic model in natural (or integral) causality is apt to design the COG observers in Luenberger forms. Compared with classical observers-based fault detection, the developed scheme allows to get graphical representation taking in to account the physical aspects (the transit of information and the energy transfer) related to the system mechanism. Owing to energetic and structural aspects, the COG observer gain is determined via causal manipulations (causal loops, causal paths ... etc). The relevance of the suggested framework is proved through the simulation results on the hydraulic system with two tanks.