Study on separation identification of cement stabilized crushed stone mixture based on convolutional neural network

With the vigorous development of China's transportation industry, the mileage of high-grade highways based on semi rigid base layers has been increasing year by year. However, the commonly used material for semi rigid base layers, cement stabilized crushed stone mixture (hereinafter referred to as water stabilized mixture), often experiences segregation during mixing, transportation, and paving. Separation of water stabilized mixture can greatly reduce the service life of roads and cause damage to people's property, the traditional separation detection method that relies on manual experience has problems of low detection efficiency and low recognition accuracy. In order to solve these problems and assist in the modernization of road construction, this article proposes a separation recognition method for water stabilized mixtures based on deep learning. Firstly, a database of segregation diseases of water stabilized mixture was built. Secondly, the control tests were set up by standard fine-tuning and feature extraction, and four different optimizers were set up respectively. By comparing accuracy, loss, precision, recall and F 1 -score at the end of the pre-trained network, the overall recognition effect of ResNet-101 as the network model was better. Thirdly, the ResNet-101 model was optimized by SpotTune, replacing cross entropy loss with focus loss, adding PReLU to the pre-trained network and a BN layer to the top layer of the pre-trained network, and using 1×1. Convolutional replacement of the fully connected layer. Finally, build a web side water stabilized mixture segregation recognition platform, and its stability was verified in practical engineering. • A new CNN-based method for separation identification of cement stabilized crushed stone mixtures was proposed. • An optimized model design based on the ResNet-101 network was proposed. • A water stable mixture segregation identification system was designed and applied.