Abstract Summary
Structural health monitoring (SHM) systems have become a common practice to investigate the integrity and safety of civil infrastructures. By embedding sensor networks on vital structures, raw monitoring data from the physical world could be captured. Due to harsh environments and/or electromagnetic interference, sensors on structures are prone to produce unusual and erroneous readings, often known as outliers. These faulty, erroneous, corrupted outliers are not uncommon, and their existence may greatly influence the results of structural health evaluation and lead to false alarms. Therefore, soon after measurement, there is a high demand for exe-cuting data cleaning for SHM data. In this study, we propose a novel robust gross outlier re-moval method, termed Hankel-structured robust principal component analysis (HRPCA), to effectively remove the unwanted gross outliers embedded in the monitoring data of structural dynamic responses. Different from the deep-learning-based approaches that possess only outli-er identification or anomaly classification ability, HRPCA is a rapid and integrated methodolo-gy for data cleaning, which enables outlier detection, outlier identification, and recovery of fault data. It capitalizes on the fundamental duality between the sparsity of the signal and the rank of the structured matrix. Using of the annihilating filter-based fundamental duality, the originally redundant yet relevant structural dynamic response data could be modeled as lying in a low-dimensional (low-rank) subspace with additional Hankel structure, which allows for better sep-aration of gross outliers embedded in the monitoring data. Then the outlier removal issue turns into a matrix factorization problem, which could be successfully solved by robust principal component analysis (RPCA). To validate the denoising capability of HRPCA, real-world moni-toring data with varying degrees of outliers (e.g., single outlier, multiple outliers, and periodic outliers) is used to illustrate the efficiency of the proposed method.
