Identifying effective AE parameters for damage evaluation of concrete in headwork: a combined cluster and random forest analysis of acoustic emission data

Agricultural concrete structures are damaged by environmental factors. To maintain such structures, it is necessary to properly determine the mechanical properties and degree of damage suffered by concrete using core tests. In previous studies, the degree of damage has been evaluated by acoustic emissions (AE) detected in compressive stress fields. The process of fracture in damaged concrete has been evaluated by analyzing various AE parameters such as AE hits, energy, and frequency. The usefulness of many AE parameters for evaluating the fracture process is evident, but the most effective AE parameters have not yet been identified. In this study, the relationship between the stress level of concrete under compression and AE parameters was investigated using regression analysis with random forests to identify the most important AE parameters. Whether the accuracy of the regression analysis could be improved by clustering AE waves was also investigated. For this purpose, core samples, severely damaged by frost, were drilled out from a concrete headwork and subjected to compressive strength tests using the AE method. After monitoring the uniaxial compression tests, statistics for seven AE parameters were calculated for every 20 × 10 −6 increment of strain. Then, AE waves were classified into three clusters based on three parameters: peak amplitude, peak frequency, and centroid frequency. The accuracy of the regression analysis was compared using non-clustered and clustered data. The peak frequencies of cluster 1 and cluster 3 were significantly higher than that of cluster 2. This result suggests that cluster 1 and cluster 3 can be attributed to macro- or mezzo-scale damage. The regression analysis’ results showed that R 2 was higher (0.720 as compared to 0.620), and RMSE and MAE were lower in cluster 1 and cluster 3 (high-peak-frequency clusters) than in non-clustered cases. Therefore, cluster analysis can be expected to improve the accuracy of AE testing. Finally, the importance of AE parameters using random forests was calculated. The most important parameter was determined to be rise time, and the second was the centroid frequency. These results suggest that these two parameters can be used to clarify compressive fracture behavior of damaged concrete.