Anomaly Detection in Distributed Strain Sensing Data: A case study of machine learning applications of the Götaälv bridge

Abstract

The Götaälv Bridge in Gothenburg, Sweden, was a vital part of the region’s infrastructure network, and ensuring its operational safety was essential as it approached the end of its service life. To support this, the bridge was instrumented with an advanced distributed fiber optic sensing system to monitor strain and structural behavior during its final years of operation, which contributed to keeping the bridge safely in service until it was replaced and decommissioned in 2021. This thesis explores unsupervised machine learning methods for anomaly detection in structural health monitoring, using distributed strain sensing data from the Götaälv Bridge. The primary focus is on evaluating the effectiveness of two distinct approaches: time-series anomaly detection using Long Short-Term Memory (LSTM) neural networks, and nonparametric clustering of strain profiles to detect new behaviors in the bridge using a Dirichlet Process Gaussian Mixture Model (DPGMM). Following an exploratory data analysis (EDA) and a literature review, the LSTM model is trained to reconstruct normal strain patterns over time, enabling the identification of anomalies through deviations in reconstruction error. In parallel, the DPGMM approach applies dimensionality reduction through principal component analysis (PCA) and probabilistic clustering to group similar strain profiles, revealing shifts in the bridges operational behavior and potential damages. Both methods are validated using known events, such as construction activities near the bridge, which provoked displacements of the bridges supports, and through comparison with alarms for abnormally high values raised by the old system. The results demonstrate one of the main limitations of the LSTM model: it requires high quality, healthy training data to reliably detect anomalies in new data. In this case, such data was not available in sufficient quantities because the system was installed on the bridge late in its service life, resulting in insufficient data for effective training. In contrast, by fully leveraging the distributed nature of the monitoring system, the DPGMM successfully identified changes in the bridge’s behavior by detecting shifts in the strain profiles along the monitored beams without requiring any healthy baseline data, making it well suited for applications on existing structures.