Recognition of voids in pavement based on GPR time-frequency features and sparse autoencoder

Ground Penetrating Radar is widely used in non-destructive pavement engineering. However, its data interpretation requires human experience and considerable time. To address this issue, a void recognition method was proposed to accurate identify voids area based on GPR time-frequency features and sparse autoencoder (SAE). Datasets was constructed by the field tests. After standardization and resampling of the pre-processed GPR signals, 18 time-domain and 12 frequency-domain features representing voids were extracted. These feature parameters were used as inputs to train the sparse autoencoder, and the performance was compared with the multi-layer perceptron (MLP) and support vector machine (SVM) models on the same dataset. The results showed that SAE achieved the highest recognition performance, with an accuracy of SAE (90.75%) > SVM (90.08%) > MLP (53.65%). The SAE model can achieve automatic recognition of onsite data, which provides a basis for accurate maintenance of pavement voids.