Resumen de Integration of fdi and dx techniques within consistency-based diagnosis with possible conflicts
Consistency-based Diagnosis, CBD, is one of the most widely used approaches to Model-based Diagnosis within the Artificial Intelligence, AI, community, The set of these AI approaches are usually known as the DX approach. Focusing on CBD, the results for fault localization are promising in several domains, besides being supported by a solid theoretical ground. Nevertheless, the approach faces some problems regarding: (i) fault diagnosis in dynamic systems, and (ii) computational complexity for on-line fault identification.
On the other hand, Model-based Diagnosis techniques coming from the Control Theory field, known as FDI approach, are well-established, and mainly focus on robustness in the detection and identification of components and isolated processes using numerical models. But these techniques lack the advantages of CBD for the localization stage.
Possible Conflicts approach, PCs, is an off-line dependency compilation technique from the DX community similar to the Analytical Redundancy Relations from the FDI community. PCs compute subsystems within a system model as minimal subsets of equations with the analytical redundancy property to detect and isolate faults. Consistency-based Diagnosis using PCs is based on iterative on-line simulation of these subsystems, that can become conflicts if a discrepancy is found between estimated and observed behavior.
Recently, the BRIDGE community has provided to researchers from both fields with a common framework for sharing results and techniques, at least for static systems. The main goal of this Thesis deals with, based on the BRIDGE framework, exploring and suggesting new ways to integrate FDI techniques within the CBD framework using Possible Conflicts to improve the overall diagnosis problem for nonlinear dynamic systems. This Thesis' contributions apply to three main topics:
- The first topic is related with the PCs computation algorithms. We improved ways to derive PCs taking into account cyclical structures that can appear in systems models halting the local propagation process. We also developed new ways for systematic computation of PCs using bond graph models and Temporal Causal Graphs as intermediate structure. This new way of PCs computation allows: (i) to derive PCs in a completely automatic way; (ii) to take into account the temporal information included in Temporal Causal Graphs for isolation purposes; and (iii) to compare the PCs approach with similar approaches from the FDI and the DX communities to make easier the integration of techniques.
- The second topic is related with the improvements in the detection stage of the CBD with PCs. In this Thesis, several FDI techniques have been studied, concluding that state observers are the best technique to estimate the values of state variables required as initial conditions for simulation, allowing to increase the robustness about uncertainties in Possible Conflicts simulations. Focusing on BRIDGE and the works about equivalence of techniques in the FDI approach, Possible Conflicts have been proposed to design state observers, providing a way to ease the integration of techniques. Then, a new way of integration has been proposed and tested obtaining satisfactory results.
- The third topic addresses the problems related with the fault identification task. Straightforward on-line model-based fault identification using numerical models is infeasible for large, complex systems, due to the inherent computational complexity and convergence problems faced by the optimization task. In this part of the Thesis we tackled the fault identification task by using the output of the fault isolation and the minimal subsystems associated with Possible Conflicts, to generate minimal parameter estimators that allow computationally efficient on-line fault identification without compromising the estimation results. One of the main advantages of this proposal is that it can be used to improve on-line fault identification in any diagnosis systems, and to prove so, we integrated the approach into different well-defined diagnosis systems obtaining satisfactory results.
Finally, the correctness and effectiveness of the contributions presented on this Thesis are demonstrated on two practical systems: the Reconfigurable Thermohydraulic Laboratory Plant, and the Reverse Osmosis System of the Advanced Water Recovery System developed at NASA Johnson Space Center.
