Clustering Based Multi Sensor Data Fusion for Honeycomb Detection in Concrete

We use three clustering algorithms to aggregate a three-modal non-destructive testing data set into defect and not-defect groups. Our data set consist of impact-echo, ultrasound (US) and ground penetrating radar data collected on a large concrete slab with embedded simulated honeycombing defects. US performs best in defect discriminating and sizing, however the false positive rate is still high. We fuse the data set using K-Means, Fuzzy C-Means and DBSCAN clustering at feature-level. We discern that DBSCAN improves the detectability up to 10 %. A discussion of its advantages over commonly used K-Means and Fuzzy C-Means clustering are provided.