Research and application of pattern recognition LSTM based bridge data anomaly detection

Abstract

The effectiveness of bridge condition assessment relies on the reliability of monitoring data, which can have multiple types of anomaly patterns in complex environments, and the accuracy and timeliness of traditional pattern recognition anomaly detection schemes do not meet the needs of practice. To address the multiple types of anomalies in bridge monitoring data, nine features that fit the bridge data situation are constructed using high-frequency acceleration data, and anomaly detection is performed using a pattern recognition LSTM neural network that is sensitive to time series, and run in real time by Flux in the Ganjiang Special Bridge monitoring system. The experimental results show that this scheme achieves a high level of accuracy for each category of anomaly detection, with a 4.53% improvement in accuracy compared to the PRNN neural network scheme. The overall detection time of real-time samples in practice is about 1.10s, and the overall anomaly detection accuracy reaches 99.65%, which meets the need for timeliness and accuracy of bridge anomaly detection system in practice.