A Novel Approach for Anomaly Detection in Vibration-Based Structural Health Monitoring Using Autoencoders in Deep Learning

Abstract

Structural health monitoring (SHM) has been widely employed in civil infrastructures for a number of years. Real-time monitoring of civil projects involves the utilization of diverse sensors. Nevertheless, accurately assessing the actual condition of a structure can pose challenges due to the existence of anomalies in the collected data. Abnormalities in this context often arise from a variety of factors, including extreme weather conditions, malfunctioning sensors, and structural impairments. The existing condition of anomaly detection is significantly impeded by this disparity. Online detection of anomalies in SHM data plays a crucial role in promptly assessing the status of structures and making informed decisions. In vibration-based SHM, enhanced frequency domain decomposition (EFDD) is one of the most used methods in the frequency domain. The signal output obtained from EFDD also includes the frequencies of the structures, which is a holistic evaluation. The findings of frequency measurements are influenced by the presence of structural damages. Extracting damage-sensitive characteristics from structural response has emerged as a complex task. Deep learning approaches have garnered growing interest due to their capacity to efficiently extract high-level abstract features from raw data. Within the scope of the study, a novel approach based on anomaly detection of changes in the signal output obtained using the EFDD was developed with autoencoders in deep learning. The performance of the novel approach was examined depending on different noise ratios (0%, 0.5%, 1%, 1.5%, and 2.0%) using the Z24 Bridge dataset. In the autoencoder training model, an autoencoder model containing a 4 Conv1D layer encoder–decoder as 128 × 64 × 64 × 128 was designed. By using the signal data of the first singular values obtained with the EFDD method, grouping was made with the labels “training data (1260 pieces),” “undamaged new data (250 pieces),” and “damaged new data (320 pieces).” In addition, the upper limit of the reconstruction error was calculated as 810 using the training data in the autoencoder model. The filtered reconstruction error values obtained were compared under different noise levels. At the end of the study, it was concluded that the novel approach works effectively under different noises and can be used in anomaly detection.