On‐line fault/damage detection schemes for mechanical and structural systems

Abstract

AbstractOn‐line monitoring has become an attractive approach to detecting faults in a system during its service time. Once a fault is detected, necessary measures can be employed to ensure satisfactory system performance and to prevent drastic economic loss that may be caused by the fault. This paper presents a performance study of two fault detection/diagnosis approaches, namely the wavelet‐based and the neural‐network‐based on‐line schemes. A brief summary of the wavelet‐based approach and two theorems on the neural‐network‐based on‐line approximation scheme are given. The schemes are illustrated for a single‐degree‐of‐freedom (SDOF) oscillator and a three‐degree‐of‐freedom mechanical system with damageable springs. Only harmonic excitations and abrupt changes in the system stiffness due to breakage of springs caused by the first passage of a threshold value are considered. Extensive numerical studies show that both approaches may be successfully implemented on‐line to detect the times when faults occurred and to locate regions where faults occurred. In addition, the on‐line neural network scheme is able to provide diagnosis of the fault by correctly identifying the stiffness change. In contrast to the neural‐network‐based on‐line approximation, where information on the healthy system and both displacement and velocity response measurements are required, the wavelet approach seems less model‐dependent and only the acceleration measurement is required. This study demonstrates the great promise of these schemes for on‐line structural health monitoring. Copyright © 2003 John Wiley & Sons, Ltd.