Neural Networks for Sensor Fault Correction in Structural Control

Abstract

Two neural network architectures are proposed for use in structural control applications: a Failure Detection Neural Network and a Failure Accommodation Neural Network. The Failure Detection Network monitors structural responses and automatically detects sensor failures that can reduce control performance and effectiveness, while the Failure Accommodation Network accounts for the failed sensors. Together, the networks are a step toward development of an expert diagnostic system for structural applications. Examples of two simple structures are used to illustrate the features of the networks. Sensor failures are simulated during control operation, and the ability of the networks to detect and accommodate the failures is examined. The numerical results reveal that these networks show promise for automated intelligent fault detection, identification, classification, and accommodation, and as such may have potential use in real civil structures. Although the networks have been used to detect and account for sensor faults alone, they may also be trained for other kinds of failures. Thus, they may have potential for incorporation into an intelligent structural monitoring system.