Rock Mass Anomaly Detection with a Variational Autoencoder in Tunnel Boring Machine Data

This study presents an unsupervised machine learning approach to rock mass anomaly detection in tunnel boring machine operational data using a Variational Autoencoder combined with bidirectional Long Short‐Term Memory cells. The model was trained on TBM data from the Brenner Base Tunnel project, with minor faults and geotechnically relevant fault zones removed to ensure a rock mass anomaly‐free training set. By reconstructing input data from a compressed latent space, the Variational Autoencoder distinguishes between normal and abnormal patterns based on reconstruction errors. Anomalies, such as significant changes in rock mass conditions, are identified when reconstruction errors exceed a threshold based on a skewness‐adjusted boxplot. Testing on three distinct sections of TBM data demonstrated the model's effectiveness in reliably detecting minor faults and major fault zones, though occasional delays in detection were noted. This approach underscores the potential of Variational Autoencoder‐based rock mass anomaly detection in mechanized tunneling and paves the way for real‐time, data‐driven decision‐making in future TBM tunneling projects.