Long-Short Term Memory Network-Based Monitoring Data Anomaly Detection of a Long-Span Suspension Bridge

Structural health monitoring (SHM) systems have been widely applied in long-span bridges and a large amount of SHM data is continually collected. The harsh environment of sensors installed at structures causes multiple types of anomalies such as outlier, minor, missing, trend, drift, and break in the SHM data, which seriously hinders the further analysis of SHM data. In order to achieve anomaly detection from a large amount of SHM data, this paper proposes a long-short term memory (LSTM) network-based anomaly detection method. Firstly, the proposed method reduces the workload for preparing training sets. Secondly, the purpose of real-time anomaly detection can be met. Thirdly, the problem of high alarm rate can be avoided by utilizing double thresholds. To validate the effectiveness of the proposed method, a case study of finite element model simulation is firstly introduced, which illustrates the detailed implementation process. Finally, acceleration data from the SHM system of a long-span suspension bridge located in Jiangyin, China is employed. The results show that the proposed method can detect anomaly with high accuracy and identify abnormal accidents such as a ship collision quickly.

Automated summary

This paper proposes an anomaly detection method for structural health monitoring (SHM) data based on LSTM network. The method reduces workload for preparing training sets, achieves real-time anomaly detection, and avoids high alarm rate by utilizing double thresholds. The case study and validation results with actual data show that the proposed method can accurately detect abnormal events.