Multichannel Convolutional Neural Network With Multiscale and Hybrid Residual for Multisensor Prediction

Convolutional neural network (CNN) is greatly affected by many factors, such as the gradient loss, the adequacy of feature extraction and the number of channels and multi-sensor traffic volume data. To reduce the influence of these factors, a parallel multi-channel convolutional neural network with multi-scale and hybrid residual, namely MSPC-HRCN is designed to predict traffic volume by utilizing the complementarity of multi-sensor data, which improve feature extraction capabilities, and achieve accurate prediction of traffic volume. Firstly, for the loss of gradient propagation, the residual block is designed to propose a connection method between the shallow layer features and deep network, and the HRCN by connecting input with output is constructed to effectively extract features from multi-sensor data and suppress gradient loss. Secondly, the parallel multi-channel model is designed to construct a parallel multi-channel HRCN, namely PC-HRCN, which assigns a specific channel for each type of divided data feature, and automatically learns and integrates the features of different input multi-sensor data. Thirdly, a feature extraction model with multi-scale convolution is constructed, and the features of multi-sensor data are captured by dividing the input signal into different scales to effectively distinguish the valuable features in each channel. Finally, the output features of each parallel channel are cascaded through the cascade layer structure, so that the presented information can been better processed by combining the extracted features from different channels. The experiment results on Metro interstate traffic volume data show that the designed MSPC-HRCN model can effectively improve the ability of feature extraction and realize the accurate prediction of traffic volume by comparing to other methods.