Data Normalization and Anomaly Detection in a Steel Plate-Girder Bridge Using LSTM

Abstract

Modal properties are recognized as indicators reflecting structural condition in structural health monitoring (SHM). However, changing environmental and operational variables (EOVs) cause variability in the identified modal parameters and subsequently obscure damage effects. To address the issue caused by EOV-related variability, this study investigated the variability of modal frequencies in long-term SHM of a steel plate-girder bridge. A Bayesian fast Fourier transform (FFT) method was used for operational modal analysis in a probabilistic viewpoint. Bayesian linear regression (BLR) and Gaussian process regression (GPR) models were utilized to capture the variability in the identified most probable values (MPVs) of modal frequencies as temperature-driven models, and the limitation of these models for data normalization with latent EOVs is discussed. To overcome the interference of latent EOVs indirectly, a long short-term memory (LSTM) network was established to trace the variability as an autocorrelated process, with a traditional seasonal autoregressive integrated moving average (SARIMA) model as a benchmark. Finally, an anomaly detection method based on residuals of one-step-ahead predictions by LSTM was proposed associating with the Mann-Whitney U-test.