Intelligent recognition of subgrade damages combining GPR signals and CNN models

Road collapse incidents resulting from subgrade damages are recurrent occurrences. The implementation of periodic nondestructive testing and assessment of road subgrades, facilitated by ground penetrating radar (GPR) technology, proves instrumental in timely identification of subgrade damages during their nascent stages, thereby furnishing essential groundwork for prompt intervention. Conventional interpretation of GPR-generated roadbed damage images predominantly relies on manual scrutiny, entailing inherent drawbacks such as misjudgments and suboptimal efficiency. For the extraction of GPR subgrade damage image features, this study proposes a methodology using convolutional neural networks (CNNs) to overcome these limitations. The suggested network design includes convolutional, pooling, and fully connected layers. The process initiates with input images undergoing feature extraction through five convolutional layers and five pooling layers. Subsequently, categorization is performed on the gathered features using the fully connected layer. The conclusive recognition outcomes are then generated through the softmax layer. Upon subjecting the model to a simulated dataset, comprising prevalent roadbed afflictions such as subgrade cavity inflation, subgrade cavity water filling, subgrade voiding, and subgrade aquifer issues, the proposed method achieves a commendable recognition accuracy of 93.48%. This empirical validation attests to the efficacy of the proposed approach in automated subgrade damage identification, marking a notable advancement in the field.