11:00 AM - 1:00 PM
[HTT20-P07] Identifying the condition of deck slab from ground-penetrating radar data using machine learning
Keywords:bridge deck slab, ground-penetrating radar, machine learning, deterioration
RC bridge has a big problem of sudden deterioration of concrete deck slab. Early detection using ground-penetrating radar (GPR) is required for safe use. However, the record of GPR is affected by many target conditions such as repair history of surface and base layer, condition of cracks, moisture of the surface layer and deck slab, and deterioration of deck slab. Signal responses of radio wave has different affections from these conditions.
Therefore, identifying only a target condition needs to analyze survey records from many ways. This requires many works and high techniques. Thus, we need something assistive technologies such as machine learning to improve the accuracy of detection.
Travel time, reflection amplitude and frequency of upper boundary of concrete slab and upper reinforcing steel are used for attribute in this study. AlexNet is mainly used for CNN. The number of records for training and predictions are from 258 to 444, which is the same number for correct and wrong. We divided the ratio of 7 and 3 for training and predictions for analyzing. Finally, we estimated high or low correlativity for the result of observing deck slabs without surface and base layers.
The frequency of reflection signal of reinforcing steel has the highest correlativity of detecting deteriorated area. Also, the reflection amplitude of upper face of deck slab is the second and the travel time of upper face of deck slab is the third. The correlativity of amplitude from the lower face of deck slab is also high. However, the correlativity of the depth of reinforcing steel is low. A series of results in this study will be used to improve the accuracy of multiple analysis using GPR.
Therefore, identifying only a target condition needs to analyze survey records from many ways. This requires many works and high techniques. Thus, we need something assistive technologies such as machine learning to improve the accuracy of detection.
Travel time, reflection amplitude and frequency of upper boundary of concrete slab and upper reinforcing steel are used for attribute in this study. AlexNet is mainly used for CNN. The number of records for training and predictions are from 258 to 444, which is the same number for correct and wrong. We divided the ratio of 7 and 3 for training and predictions for analyzing. Finally, we estimated high or low correlativity for the result of observing deck slabs without surface and base layers.
The frequency of reflection signal of reinforcing steel has the highest correlativity of detecting deteriorated area. Also, the reflection amplitude of upper face of deck slab is the second and the travel time of upper face of deck slab is the third. The correlativity of amplitude from the lower face of deck slab is also high. However, the correlativity of the depth of reinforcing steel is low. A series of results in this study will be used to improve the accuracy of multiple analysis using GPR.