Methodologies for Engineering with Plug-and-Learn Components: Synthesis and Analysis Across Abstraction Layers

Abstract:

Effective engineering of complex devices often depends on the ability to encapsulate responsibility for tasks into modular components with specific responsibilities and clearly defined lines of communication. Under such conditions, one can determine what components or lines of communication are at fault for poor system performance because the system can be checked against modularized model specifications. CyberPhysical Systems (CPS) that modify themselves to improve performance or repair damage often rewrite the modular relationships that make system modeling or Verification and Validation (V&V) possible. In this project, we are exploring methods to create self--modifying CPS systems that, in addition to self--improving or self--repairing, are also capable of maintaining system models that support the use of V&V techniques. Our work specifically focuses on a Flapping--Wing Micro Air Vehicle (FW--MAV) that learns to recover from in--service wing damage by making system modifications at multiple levels of abstraction in the device. In our FW--MAV systems, controller repair is simultaneously applied at the levels of wing motion shape, bulk wing motion properties (frequency and shift), and control allocation. With simultaneous adaptation at three distinct levels of design abstraction, major disruptions in design model modularity is likely providing a perfect situation in which to study V&V in environments where modular interactions are being continuously destroyed and reformed. Year one results relate primarily to FW--MAV vehicle and construction. In addition to discussing and demonstrating our vehicle and associated models, we will discuss early successful results in combining learning methods with model inference to recover updated vehicle models as a side effect of controller repair learning. Using these methods, we have shown it possible to detect the magnitude and location of single--wing damage deficits by placing a symmetry restriction on allowable wing motion solutions that provide easily testable hypotheses about the nature of vehicle damage. Testing these hypotheses during vehicle operation directly reveals the faults and allows the system model to be updated to reflect the actual state of the FW--MAV.