Monitoring techniques for an online neuro-adaptive controller

Abstract

The appeal of biologically inspired soft computing systems such as neural networks in complex systems comes from their ability to cope with a changing environment. Unfortunately, adaptability induces uncertainty that limits the applicability of static analysis to such systems. This is particularly true for systems with multiple adaptive components or systems with multiple types of learning operation. This work builds a paradigm of dynamic analysis for a neuro-adaptive controller where different types of learning are to be employed for its online neural networks. We use support vector data description as the novelty detector to detect unforeseen patterns that may cause abrupt system functionality changes. It differentiates transients from failures based on the duration and degree of novelties. Further, for incremental learning, we utilize Lyapunov functions to assess real-time performance of the online neural networks. For quasi-online learning, we define a confidence measure, the validity index, to be associated with each network output. Our study on the NASA F-15 Intelligent Flight Control System demonstrates that our novelty detection tool effectively filters out transients and detects failures; and our light-weight monitoring techniques supply sufficient evidence for an insightful validation.