Effective alerting for bridge monitoring via a machine learning-based anomaly detection method

To alert upcoming structural failure is a critical task for structural health monitoring of bridges. Traditional methods mainly rely on thresholds, which are often fixed values and may cause missing or too sensitive reports. Identifying abnormal data, locating the source of anomalies and delivering proportional alerts require new, dynamic, and robust algorithms running on massively streaming monitoring data. This article proposes a new machine learning-based anomaly detection method for historical data mining as well as real-time alerting. The method transforms one-dimensional time series into two-dimensional tensors, enabling the encoder-like model to simultaneously learn the changes in multiple sensors within and between temporal cycles in a two-dimensional space. Training and validation of the proposed method are presented with data from a bridge monitoring system in service, and comparisons against traditional threshold-based alerting method are made. The proposed method can accurately identify abnormalities beyond the traditional thresholds and effectively detect abnormal deviations of sensors, thus constituting as a promising module for real-time alerting systems of bridges.