Date of Graduation


Document Type


Degree Type



Statler College of Engineering and Mineral Resources


Lane Department of Computer Science and Electrical Engineering

Committee Chair

Bojan Cukic


Neural networks are popular models used for online adaptation to accommodate system faults and recuperate against environmental changes in real-time automation and control applications. However, the adaptivity limits the applicability of conventional verification and validation (V&V) techniques to such systems. We investigated the V&V of neural network-based online adaptive systems and developed a novel validation approach consisting of two important methods. (1) An independent novelty detector at the system input layer detects failure conditions and tracks abnormal events/data that may cause unstable learning behavior. (2) At the system output layer, we perform a validity check on the network predictions to validate its accommodation performance.;Our research focuses on the Intelligent Flight Control System (IFCS) for NASA F-15 aircraft as an example of online adaptive control application. We utilized Support Vector Data Description (SVDD), a one-class classifier to examine the data entering the adaptive component and detect potential failures. We developed a "decompose and combine" strategy to drastically reduce its computational cost, from O(n 3) down to O( n32 log n) such that the novelty detector becomes feasible in real-time.;We define a confidence measure, the validity index, to validate the predictions of the Dynamic Cell Structure (DCS) network in IFCS. The statistical information is collected during adaptation. The validity index is computed to reflect the trustworthiness associated with each neural network output. The computation of validity index in DCS is straightforward and efficient.;Through experimentation with IFCS, we demonstrate that: (1) the SVDD tool detects system failures accurately and provides validation inferences in a real-time manner; (2) the validity index effectively indicates poor fitting within regions characterized by sparse data and/or inadequate learning. The developed methods can be integrated with available online monitoring tools and further generalized to complete a promising validation framework for neural network based online adaptive systems.