Deep principal component analysis: An enhanced approach for structural damage identification. (November 2019)

• Record Type: Journal Article
• Authors: Silva, Moisés; Santos, Adam; Santos, Reginaldo; Figueiredo, Eloi; Sales, Claudomiro; Costa, João CWA
• Is Part Of: Structural health monitoring. Volume 18:Number 5/6(2019)
• Page Start: 1444
• Page End: 1463
• Publication Date: 2019-11
• Subjects: Structural health monitoring -- deep learning -- principal component analysis -- data normalization -- environmental variability -- operational variability; Structural health monitoring -- Periodicals; Structural stability -- Periodicals; Strength of materials -- Periodicals; Nondestructive testing -- Periodicals; Constructions -- Stabilité -- Périodiques; Résistance des matériaux -- Périodiques; Contrôle non destructif -- Périodiques; Electronic journals; 624.17
• DOI: 10.1177/1475921718799070
• Languages: English
• ISSNs: 1475-9217

Abstract

The structural health monitoring relies on the continuous observation of a dynamic system over time to identify its actual condition, detect abnormal behaviors, and predict future states. The regular changes in environmental factors have been reported as one of the main challenges for the application of structural health monitoring systems. These influences in the structural responses are in general nonlinear, affecting the damage-sensitive features in the most varied forms. The usual process to remove these normal changes is referred to as data normalization. In that regard, principal component analysis is probably the most studied algorithm in structural health monitoring, having numerous versions to learn strong nonlinear normal changes. However, in most cases, not all variability is properly accounted for via the existing nonlinear principal component analysis approaches, resulting in poor damage detection and quantification performances. In this article, a new paradigm based on deep principal component analysis, rooted in the deep learning field, is presented to overcome these limitations. This approach extracts the most salient underlying feature distributions by stacking multiple feedforward neural networks trained to learn an identity mapping of the input variables, where the network inputs are reproduced into the outputs. Similar to the traditional nonlinear principal component analysis–based approach, our approach identifies a nonlinear output-only model of an undamaged structure by comprising modal features into an internal bottleneck layer, which implicitly represents the independent environmental factors. The proposed technique is validated through the application on a progressively damaged prestressed concrete bridge and a three-span suspension bridge. The experimental results demonstrate that capturing the most slight nonlinear variations in the data can lead to improved data normalization and, consequently, better damage detection and quantification performances.