Open Access Open Access  Restricted Access Subscription or Fee Access

Robust Hammering Echo Analysis for Concrete Assessment with Transfer Learning



This paper endeavors to develop robust echo pattern analysis system to discern defect-induced anomalous echoes, which can adapt to changing parameters in impactecho test, such as the echo spectrum variation induced by the change in the target concrete types. Conventional computerized impact-echo analysis systems are built on the assumption that the (pre-collect) training echo signal and (unseen) testing echo data reside in the same feature space with identical distribution characteristic. However, in the real-world scenarios, such assumption does not hold. That is, the echo signal is often affected by practical factors, such as selection of instruments, i.e. hammer and transducers, and the change in testing specimen. From the viewpoint of pattern recognition, these factors make the posterior distribution of the test hammering echo signal drift from that of the pre-collected training echoes (with normal/defective conditions), thus will degrade the accuracy of the echo pattern analysis models. In this study, we introduce transfer learning strategy to deal with distribution drift issue between training and input echo signal. For instance, proposed system can distil discriminant echo patterns collected from (training) normal concrete and then generalize the information to detect flaws in lightweight concrete. To validate those schemes, we prepared two kinds of air-coupled transducers and three types of concretes (normal/enhanced/lightweight) with various types of inner void defects. Condition assessment was conducted under mismatched training/testing settings. Extensive experimental results reveal that transfer learning approach can greatly improve robustness of impact-echo.

Full Text:



  • There are currently no refbacks.