Comparison of Various Methodologies to Detect Anomalies in a Time Series Data Taken from a Tunnelling Project

A major concern in urban mechanised tunnelling projects is avoiding damage to the existing buildings and the tunnel boring machine (TBM), which may be adjusted by an advanced precise excavation simulation. Because a realistic simulation must account for multiple interactions between the boring machine and the subsurface, an exact representation of the ground’s geological profile must be created beforehand. Due to the limited monitoring and sampling, several geologic anomalies may have been overlooked when sketching the geologic profile. As a result, the geological profile should be updated alongside the construction phases when new information becomes available. To accomplish this, one can use the boring machine’s recorded data to detect any irregularities in the drilling process caused by changes in geological conditions. This research compares various cutting-edge anomaly detection approaches on time series. Due to a large amount of sensor data, the visualization of multiple sensors/features over time was first performed, and the critical features with the highest impact on the detection process for identifying anomalies were selected. Anomaly detection techniques include isolation forest, k-Means, k-Means Sequential Time Series Cluster, Auto-Regression Integrated Moving Average (ARIMA), and Convolutional Neural Network (CNN) Auto-encoders are among the main aspects. The methods presented here were applied to a given data set from an actual tunnelling operation in Germany to locate the location of some concrete slabs in a relatively homogeneous ground. The obtained results agree well with the exact location of anomalies. The performance of various methods is evaluated through error quantification measures.