Fault detection using vibration data-driven models-a simple and well-controlled experimental example

Damage detection in structures has in recent years received great attention from both industry and academia. This has motivated the development of other techniques for damage identification. Some common techniques are based on monitoring of vibration signals. Depending on the type of damage to a structure, a change in the dynamic behavior is expected, which can be reflected in the vibration signals.

For instance, is possible to identify cracks in mechanical systems by analyzing their natural frequencies because the presence of cracks results in a reduction of these frequencies. However, this reduction is very small and can be hidden by environmental conditions. This drawback is leading studies that employ statistical methods to detect the damages.

In this work, an outlier test was used to identify a crack in a beam. A test rig composed of a clamped beam was built and the crack was represented by the introduction of a transversal cut in the beam. The outlier test used in this paper is based on the Mahalanobis distance, which measures the distance between a point and a distribution.

Therefore, if a point is far from this distribution, the measurement is statistically unlikely and assumed to be a result of damage. The results showed the capability of the method to identify the damage, but it also proved to be very sensitive to the boundary conditions. These very simple and extremely well-controlled laboratory experiments shed some light on (i) the importance of the changes of the boundary conditions when reassembling machine components and (ii) the influence of exchanging similar machine components when re-using data-driven models built from vibration signals for fault detection.