In model-based diagnosis (MBD), structural models can provide useful information for fault diagnosis and fault-tolerant control design. In particular, they are known for supporting the design of analytical redundancy relations (ARRs) which are widely used to generate residuals for diagnosis. On the other hand, systems are increasingly complex whereby it is necessary to develop decentralized architectures to perform the diagnosis task. Decentralized diagnosis is of interest for on-board systems as a way to reduce computational costs or for large geographically distributed systems that require to minimizing data transfer. Decentralized solutions allow proper separation of industrial knowledge, provided that inputs and outputs are clearly defined. This paper builds on the results of [1] and proposes an optimized approach for decentralized fault-focused residual generation. It also introduce the concept of Fault-Driven Minimal Structurally-Overdetermined set (FMSO) ensuring minimal redundancy. The method decreases communication cost involved in decentralization with respect to the algorithm proposed in [1] while still maintaining the same isolation properties as the centralized approach as well as the isolation on request capability. 1. Introduction With increasing complexity of industrial processes, the requirement for reliability, availability and security is growing significantly. Fault detection and isolation (FDI) are becoming a major issue in industry. The structural approach constitutes a general framework to provide information when the system becomes complex. The main aim of the structural approach application is to identify the subsets of equations which include redundancy. The system structure analysis, originally developed for the decomposition of large systems of equations for their hierarchical resolution, was adopted by the Fault Detection and Isolation (FDI) community [2, 3]. Structural concepts are used for analysis of system monitor ability using the concept of complete matching on a graph. Decentralized diagnosis has received considerable attention to deal with distributed systems or with systems that may be too large to be diagnosed by one centralized site. In the same way, the decentralized solution allows proper separation of industrial knowledge, provided that inputs and outputs are clearly defined.