A new unsupervised method for damage detection in structures: Cor-PCAE

Despite its widespread adoption in Structural Health Monitoring (SHM), Moving Principal Component Analysis (MPCA) still has some significant drawbacks in terms of human bias, thus limiting its ability to detect minor damage in structures. This article proposes a new fully unsupervised method called Correlation Principal Component Auto-Encoder (Cor-PCAE), which enhances the anomaly (i.e. damage) detection performance of the MPCA method. The new Cor-PCAE method optimises computational resources by splitting data into a sequence of sensor correlations and generates a suitable sensor correlation model with a deep autoencoder. The Cor-PCAE method is applied to detect minor to moderate damage in a Fibre Reinforced Polymer (FRP) bridge subjected to walking tests. Three damaged conditions were considered, denoted by minor damage (DM1 & DM2) and moderate damage (DM3). The acceleration data (DTS1 & DTS2) were used to compare the performance of Cor-PCAE against existing methods such as Moving Principal Component Analysis (MPCA) and Combined MPCA-Multiple Linear Regression (MPCA-MLR). The results show that Cor-PCAE’s detection performance is faster than MPCA for both datasets. The new Cor-PCAE method also reduces the probability of misclassifying damage as a healthy condition. Lastly, by observing the probability of misclassification from three sensor configurations (“Far”, “Distributed”, and “Near”), the new Cor-PCAE method consistently achieves a relatively low probability of misclassifying damage, even in less ideal sensor configurations. This article contributes to the development of faster and more efficient damage detection methods for SHM of existing structures. • A new Cor-PCAE method for damage detection in structures is proposed. • Cor-PCAE removes human bias and enables an efficient correlation feature extraction. • Efficient Cor-PCAE allows for a faster anomaly detection than MPCA and MPCA-MLR methods. • Cor-PCAE has a lower probability of misclassifying damage as a “healthy" condition. • Cor-PCAE exhibits better performance, despite a less-than-ideal sensor configuration.