Pavement condition detection using acceleration data collected by smartphones based on 1D convolutional neural network

Vibration-based pavement condition detection methods have advanced in recent years, and it has been proven to be feasible to identify pavement conditions by analysing acceleration data. In this study, a public participation solution is proposed, and a one-dimensional convolutional neural network (1D-CNN) is introduced to directly process acceleration signals, addressing the limitations of traditional machine-learning classification methods. In this study, a smartphone and bicycle were used as the experimental tools, and 422 samples of acceleration data across the X-, Y-, and Z-axes were collected, including four types of pavement conditions: bumpy pavement, speed bumps, smooth pavement, and potholes. Five types of 1D-CNN with different activation functions and network structures were designed to classify the data and were compared with machine learning algorithms, including support vector machine (SVM) and radial basis function (RBF) neural networks. The results show that a 1D-CNN, with three convolution layers and three pooling layers using the ReLU activation function, achieved the best classification performance, with a classification accuracy of 0.9976. Compared with SVM and RBF neural networks, CNN not only saves considerable time by eliminating manual feature extraction operations but also provides higher classification accuracy.