To locate the moisture damage area in asphalt pavement, an automatic detection method by leveraging the state-of-the-art deep learning approach and continuous wavelet transform (CWT) method from high frequency GPR signal was presented. A numeric simulation and a laboratory experiment were employed to analyze the response of GPR signal in moisture damage area. GPR signals were transferred to time-frequency domain by using a CWT method, before comparing CWT results among normal pavement, moisture and bridge joints. Moisture damage datasets, including 8215 traces of moisture damages, 8215 traces of normal pavement and 7643 traces of bridge joints, were extracted from GPR field survey of bridge pavement with 2.3 GHz antenna. The maximum and mean values of CWT results were obtained from the dataset and used to normalize the CWT dataset. Typical classification models, including Resnet50, ResNet18, ShuffleNet, EfficientNet, NasNetmobile and DenseNet, were trained and compared on the CWT dataset. The result indicates that the CWT features among normal pavement, moisture damage and bridge joints have significant differences. Moreover, the energy distribution in the CWT image can represent the moisture damage depth and width information, proving that the CWT method is effective for extracting moisture damage features from GPR signal. The overall performance of ResNet50 model is best with high accuracy and high efficiency in moisture detection compared to other models. Experimental results proved the promising performance of our approach in detecting moisture damage in asphalt pavement, and this method could also be used for subsurface target detection in other GPR application.
Read full abstract