Distributed Diagnosis of Continuous Systems: Global Diagnosis Through Local Analysis
Early detection and isolation of faults is crucial for ensuring system safety and efficiency. Online diagnosis schemes are usually integrated with fault adaptive control schemes to mitigate these fault effects, and avoid catastrophic consequences. These diagnosis approaches must be robust to uncertainties, such as sensor and process noise, to be effective in real world applications. Also, diagnosis schemes must address the drawbacks of centralized diagnosis schemes, such as large memory and computational requirements, single points of failure, and poor scalability. Finally, to be effective, fault diagnosis schemes must be capable of diagnosing different fault types, such as incipient (slow) and abrupt (fast) faults in system parameters. This dissertation addresses the above problems by developing: (i) a unified qualitative diagnosis framework for incipient and abrupt faults in system parameters; (ii) a distributed, qualitative diagnosis approach, where each diagnoser generates globally correct diagnosis results without a centralized coordinator and communicates minimal measurement information and no partial diagnosis results with other diagnosers; (iii) a centralized Bayesian diagnosis scheme that combines our qualitative diagnosis approach with a Dynamic Bayesian network (DBN)-based diagnosis scheme; and (iv) a distributed DBN-based diagnosis scheme, where the global DBN is systematically factored into structurally observable independent DBN factors that are decoupled across time, so that the random variables in each DBN factor are conditionally independent of those in all other factors, given a subset of communicated measurements that are converted into system inputs. This allows the implementation of the combined qualitative and DBN-based diagnosis scheme on each DBN factor, which operate independently with a minimal number of shared measurements to generate globally correct diagnosis results locally without a centralized coordinator, and without communicating any partial diagnosis results with other diagnosers. The correctness and effectiveness of these diagnosis approaches is demonstrated by applying the qualitative diagnosis approaches to the Advanced Water Recovery System developed at NASA Johnson Space Center; and the DBN-based diagnosis schemes to a complex, twelfth-order electrical system.