Modular Diagnosis

Abstract

In monolithic diagnosis, the reconstruction of the system behavior is carried out in a single step (monolithically) based on the simulation of the system reaction that is consistent with the given diagnostic problem. However, this is not in general the best way to perform reconstruction, both because the search space, besides the system observation, is constrained by the component models only and because no support for parallel computation is provided. A better way to proceed is to recursively break down the diagnostic problem into several smaller subproblems, each one relevant to a subpart of the system, namely a cluster. The subproblem associated with a cluster is obtained by projecting the observer, the observation, and the initial state on the cluster, which in general results in an open subsystem. Such a decomposition is specified by a reconstruction graph that, in its simplest form, is a tree that mirrors the decomposition of the original diagnostic problem. Starting from the leaves of the graph, reconstruction is carried out for lower-level clusters, while the resulting active spaces are joint to obtain the active space relevant to upper-level clusters. This is iterated until the root of the reconstruction graph is reached: the resulting active space is equivalent to the active space we would obtain by means of a monolithic reconstruction, as detailed in Chapter 4. In its more general form, the reconstruction graph is a DAG, where each node may represent several clusters that share the same diagnostic subproblem. This way, the method allows the parallel reconstructions of clusters, whilst supporting reuse of intermediate active spaces.