Short-Training Damage Detection Method for Axially Loaded Beams Subject to Seasonal Thermal Variations

Abstract

Vibration-based damage features are widely adopted in the field of structural health monitoring (SHM), and particularly in the monitoring of axially loaded beams, due to their high sensitivity to damage-related changes in structural properties. However, changes in environmental and operating conditions often cause damage feature variations which can mask any possible change due to damage, thus strongly affecting the effectiveness of the monitoring strategy. Most of the approaches proposed to tackle this problem rely on the availability of a wide training dataset, accounting for the most part of the damage feature variability due to environmental and operating conditions. These approaches are reliable when a complete training set is available, and this represents a significant limitation in applications where only a short training set can be used. This often occurs when SHM systems aim at monitoring the health state of an already existing and possibly already damaged structure (e.g., tie-rods in historical buildings), or for systems which can undergo rapid deterioration. To overcome this limit, this work proposes a new damage index not affected by environmental conditions and able to properly detect system damages, even in case of short training set. The proposed index is based on the principal component analysis (PCA) of vibration-based damage features. PCA is shown to allow for a simple filtering procedure of the operating and environmental effects on the damage feature, thus avoiding any dependence on the extent of the training set. The proposed index effectiveness is shown through both simulated and experimental case studies related to an axially loaded beam-like structure, and it is compared with a Mahalanobis square distance-based index, as a reference. The obtained results highlight the capability of the proposed index in filtering out the temperature effects on a multivariate damage feature composed of eigenfrequencies, in case of both short and long training set. Moreover, the proposed PCA-based strategy is shown to outperform the benchmark one, both in terms of temperature dependency and damage sensitivity.

Keywords

• structural health monitoring
• unsupervised learning
• environmental variations
• principal component analysis
• short baseline
• tie-rods