Early detection of clandestine tunnels beneath gas and oil pipelines is a critical challenge for energy security. Conventional methods only identify damage once the infrastructure has already been compromised, leaving room for theft, sabotage, and smuggling. Ground-penetrating radar (GPR) offers non-invasive imaging capability, but manual interpretation of radargrams is unfeasible for continuous surveillance along extensive corridors. Supervised approaches, on the other hand, require examples of tunnels that are scarce in practice. Faced with this limitation, an innovative solution based on unsupervised learning emerges: a pipeline that trains a denoising convolutional autoencoder exclusively with normal subsurface radargrams, without any labeled tunnels. During inference, anomalies are detected through reconstruction error.
The key advance of this methodology lies in a depth-constrained, top-k-based anomaly score. Instead of evaluating the entire radargram, it focuses on the band where tunnels can physically exist (1.5-3 meters), selecting only the highest reconstruction errors within that range. This physical principle, without the need for retraining or labels, raises the AUC from 0.986 to 0.994 and reduces missed detections from 74 to 17 over 634 tunnel windows, compared to the full-image score. Furthermore, it is observed that the optimal top-k fraction varies when applying the depth constraint: 1% is better on full images, while 5% works optimally once the range is restricted. The final system achieves an AUC of 0.994, F1 of 0.975, recall of 97.3%, and precision of 97.6% over 1600 real-field windows, with a false alarm rate of only 1.6%, without using tunnel labels for training, scoring, or threshold calibration.
This approach demonstrates that automated, scalable, and accurate surveillance of critical infrastructure is possible through unsupervised artificial intelligence. The combination of deep neural networks with domain-specific physical rules opens the door to tailored applications in sectors such as oil & gas, mining, and perimeter security. Deploying such solutions in production requires robust technological infrastructure: from AI for businesses that integrate anomaly detection models to cloud platforms that process large volumes of sensor data. A company like Q2BSTUDIO offers precisely that ecosystem, with custom software that can adapt these algorithms to each client's specific needs, combining artificial intelligence, cybersecurity to protect data, and AWS and Azure cloud services to ensure scalability and availability.
Additionally, the ability to generate early alerts and visualize trends is enhanced with business intelligence services such as Power BI, which allow operators to interpret model results in interactive dashboards. The integration of autonomous AI agents could even automate responses to confirmed detections, reducing reaction times. In a context where critical infrastructure security is increasingly a priority, solutions like the one described—based on unsupervised learning, physical constraints, and a comprehensive technology platform—represent the next step in preventive asset protection.
