关键词: 数据增强, 道路塌陷, 风险评估, 深度学习, 神经网络

Abstract:

To prevent road collapse accidents and minimize losses, it is essential to conduct scientific and effective risk assessments of urban roads. To address the imbalance in ground collapse datasets, a synthetic minority oversampling technique (SMOTE) and a generative adversarial network (GAN) were integrated to enhance the original positive accident samples. Deep learning models including convolutional neural networks (CNN) and multilayer perceptron (MLP) were employed for training, and the trained models could be used for risk identification of urban roads. Experimental validation was conducted using ground collapse incident data from Foshan City (2014-2018). The results demonstrate that data augmentation can significantly improve the model's ability to recognize minority samples (with an average recall rate increase of up to 20%) and solve the overfitting problem when training models with imbalanced data.

Key words: data augmentation, road collapse, risk assessment, deep learning, neural network