Multi sensor data fusion approach for automatic honeycomb detection in concrete

• We laid out a novel conceptual framework for multi-sensor defect classification.
• No well-established technique for honeycomb detection.
• New methods to extract information (features) from each sensor data are presented.
• We introduce two conceptually simple yet effective fusion algorithms.
• A quantitative evaluation is performed by comparing ROC curves.

We present a systematic approach for fusion of multi-sensory nondestructive testing data. Our data set consists of impact-echo, ultrasonic pulse echo and ground penetrating radar data collected on a large-scale concrete specimen with built-in honeycombing defects. From each data set, the most significant signatures of honeycombs were extracted in the form of features. We applied two simple data fusion algorithms to the data: Dempster’s rule of combination and the Hadamard product. The performance of the fusion rules versus the single-sensor testing was evaluated. The fusion rules exhibit a slight improvement of false alarm rate over the best single sensor.