Signal Anomaly Detection of Bridge SHM System Based on Two-Stage Deep Convolutional Neural Networks

Abstract

Identifying and removing anomalies of sensor signals existing in the bridge structural health monitoring (SHM) system is conductive to correctly assessing the operation status of the monitored bridge. A data augmentation strategy of first-order derivation operation and equal-length sequence segmentation was proposed to extract more abundant features of signal anomalies. To reduce the impact of redundant information in the augmented data on the training efficiency of supervised learning, based on statistical analysis and ranking importance measurement, feature dimension reduction was carried out on the augmented sample dataset. Aiming at the sample dataset after dimensionality reduction, a two-stage deep convolutional neural network model that can effectively identify different signal anomaly patterns was established. The experimental results demonstrated that the proposed method can enhance the recognition accuracy on signal anomaly patterns when comparing to the effect from direct training on the original dataset.