Convolutional neural network-based data anomaly detection method using multiple information for structural health monitoring

Abstract

Structural health monitoring (SHM) is used worldwide for managing and maintaining civil infrastructures. SHM systems have produced huge amounts of data, but the effective monitoring, mining, and utilization of this data still need in-depth study. SHM data generally includes multiple types of anomalies caused by sensor faults or system malfunctions that can disturb structural analysis and assessment. In the routine data pre-processing, multiple signal processing techniques are required to detect the anomalies, respectively, which is inefficient. The large variations of extracted features from massive SHM data make the data anomaly detection techniques prone to be over-processed or under-processed. Even with expert intervention, the parameter tuning, associated with multiple data preprocessing methods, is still a challenge, making the procedure expensive and inefficient. In addition, one data anomaly detection technique frequently mis-detects other types of anomaly. In this work, we focus on the anomaly detection in the stage of data pre-processing that little work has been done based on the real-world continuous SHM data with multiclass anomalies. We proposed a novel data anomaly detection method based on a convolutional neural network (CNN) that imitates human vision and decision making. First, we split raw time series data into sections, and visualized the data in time and frequency domain, respectively. Then each section's images were stacked as a single dual-channel image and labeled according to graphical features (multi-2D image space expression). Second, a CNN was designed and trained for data anomaly classification, during which the descriptions and representations of the anomalies' features were learned by convolution. To validate our work, we considered the effects of balanced and imbalanced training sets and training ratios on actual acceleration data from an SHM system for a long-span cable-stayed bridge. The results show that our method could detect the multipattern anomalies of SHM data efficiently with high accuracy. The proposed dual-information CNN-based design makes this detection process readily scalable, faster, and more accurate, thereby providing a novel perspective with strong potential for SHM data preprocessing.