Speaker
Description
This work reports the first operational deployment of a neural network-based anomaly detection system for maritime container inspection using muon scattering tomography. While maritime container inspection is essential for infrastructure security, it remains highly complex due to the wide range of structural cargo configurations in which anomalies are likely to occur. In addition, the physical principles governing muon scattering impose fundamental constraints on the data which, combined with the scarcity of experimental measurements, significantly limit the availability of representative anomalous samples. Under these conditions, anomaly detection models must operate within a physically consistent framework to compensate for the lack of real detector measurement data, enabling the system to distinguish between genuine anomalies and extreme but physically plausible cargo configurations.
To overcome the combined challenges of structural cargo variability, data constraints, and limited experimental measurements, a synthetic dataset of physically realistic scenarios was constructed in cooperation with customs authorities. This dataset accurately replicates operational conditions and complies with transportation regulations. In addition, dedicated measurement campaigns were conducted under operational conditions to complement and experimentally validate the synthetic dataset using real detector measurements. The proposed anomaly detection system integrates unsupervised and semi-supervised learning approaches to identify structural inconsistencies in cargo by modeling the distribution of benign transport data and establishing a statistically robust baseline of normality. Furthermore, we demonstrate that combining experimental measurement data with models pre-trained on simulated data substantially enhances generalization through sim-to-real transfer. This approach enables unsupervised anomaly detection without requiring prior specification of contraband materials.