Soil Clustering and Anomaly Detection Based on EPBM Data Using Principal Component Analysis and Local Outlier Factor

This paper presents an unsupervised anomaly detection method to infer changing ground conditions in real time during tunneling. The proposed method combines dimensionality reduction and density-based outlier detection methods. The method uses features related to the cutter, thrust, and ground conditioning systems of an earth pressure balance tunnel boring machine (EPBM) as the input. The principal component analysis (PCA) was used to extract information on the ground conditions from selected EPBM features, reduce the dimensionality, and embed the data points in a geometrical space. The local outlier factor (LOF) was used to measure the degree of the anomaly of the projected data points. This study used a data set from the State Route 99 (SR99) tunnel construction in Seattle, WA. Based on this study, it can be concluded that with appropriately selected EPBM features, PCA can dynamically cluster EPBM data according to the ground conditions. Interestingly, the dynamic behavior of the projected data points does not substantially affect the soil clusters as they were always grouped. This study also demonstrates that LOF can be a sensible measure to detect changing ground conditions.