A neuro-inspired computational model for adaptive fault diagnosis

Abstract

Fault diagnosis is a key process to ensure reliable and cost-effective performance of time-critical engineered systems. This article develops a data-driven computational model for adaptive fault diagnosis by drawing an analogy with the neurobiological process of conscious attention—a dynamic process that brings only the most novel 0.01% of the signals we receive with our five senses to our conscious experience. A model of conscious attention based on the theory of dynamic core hypothesis is first outlined, followed by a computational model that mimics key stages of the conscious attention process. Convolutional neural networks serve as a basis for modeling perceptual categorization and concept formation through automatic feature extraction, due to their analogy with the processes of neural group selection and reentry in the brain. Further, the process of incremental learning and its impact on signal novelty are modeled via transfer learning. The model is tested on the NASA C-MAPSS turbofan engine model, which indicated 95–99% fault diagnosis accuracy. This study aims at familiarizing the engineering community with the neurobiological process of conscious attention and its applications for adaptive process monitoring and improvement in engineered systems.