Event-based architecture for diagnosis in control advisory systems

Abstract Artificial intelligence research on diagnostic systems is focused mainly on static problems in which the object of interest is assumed to be in a stable condition that will remain stable until the diagnostic agent has completed its deliberation. Control systems, on the other hand, cannot afford the luxury of ‘suspended time’; faults must be detected, isolated and reconfiguration implemented within one computation cycle to prevent departure from stable control envelopes. A new class of system is emerging from recent programs directed toward vehicle operator aids for fighter aircraft, submarines and helicopters. These systems are neither static off-line aids nor real-time controllers; instead they are control advisory systems which span the time scales of both regimes. These systems interface with controllers to interpret the error codes and to conduct tests and implement reconfigurations. On the other hand, these systems also interact with the vehicle operator to prioritize their activity consistent with the operator's goals and to recommend diagnostic/emergency procedures. Pilot's Associate (PA), a complex system of software and cockpit displays currently being developed to aid Air Force fighter pilots, is used to exemplify a control advisor and to display the requirements for real-time equipment diagnosis in such a system. It is shown that the required functionality challenges the capabilities of current diagnostic algorithms. The PA diagnostic architecture is presented. This architecture is centred around the concept of a diagnostic event which begins when abnormal data are detected, and which is treated as an interruptable computational process which may exist in conjunction with other event processes in multiple fault situations. The architecture solves a number of problems such as interruptability, and integrating the different time-scales of the pilot and the local system controllers, and the architecture provides a means for fusing statistical fault detection methods such as Kalman filters with artificial intelligence techniques such as model-based reasoning and classification. The result is an approach which exploits the strengths of each technique and provides a mechanism for automated reasoning using both quantitative and qualitative information.