Highway subgrade stability prediction model based on depth separation convolutional fusion network

Abstract The automatic data collection of monitoring collects abnormal information on the pavement in real time with the help of detection sensors. At present, it is impossible to grasp the quality monitoring of pavement in large-scale pavement construction and maintenance or to gradually form regular research contents and correct conclusions to guide the construction of pavement construction projects. To promote the ability of high-precision highway maintenance and detection and solve the situation of false detection or missing detection of road defects, it is necessary to establish a monitoring mechanism of multi-scale feature fusion. The improved depth separation convolution model proposed in this paper innovatively adopts the vector space form of adjacent hypervoxels in dealing with the edge position of subgrade defects and an accurate segmentation feature scheme. It aims to facilitate the calculation of irregular area scope and the analysis of normal object motion law and reduce the time consumed in the direct search for the nearest neighbor of point data. According to the error between the current prediction result and the real result, the contribution of each neuron to the error is calculated in reverse. Then, according to the contribution, we update each neuron’s weight to reduce the error gradually and improve the accuracy of the output result. The method can quickly and effectively identify the adjacent supervoxels in the convolution kernel, effectively reduce the number of parameters of the network, and alleviate the problems of overfitting and long training time to improve computational efficiency. The experimental results indicate that the proposed data structure and algorithm can better distinguish the defect features of highway roadbeds. Compared with the traditional methods, the retrieval efficiency of accelerated feature objects is improved by 17.2% and 10.5%, respectively. Further, the image analysis quality is improved by 20.7%, and the complete defect prediction is gradually formed. The research results of this paper further improve the efficiency of the neural network model and maintain the accuracy of data, which not only meets the needs of highway subgrade detection but also promotes the application of large-scale image processing technology. Therefore, it has great market application value.