Road Pavement Identification based on Acceleration Signals of Off-road Vehicles Using the Batch Normalized Recurrent Neural Network

In this paper, a type of recurrent neural network, long-short term memory (LSTM) network is applied to identify different classes of sequential vehicle acceleration signals, which underlies active suspension control strategies. The architecture of this deep learning model is a two-layer stacked multi-inputs and single-output LSTM network. And for training acceleration, the batch normalization technique is used on layer inputs. The training and testing datasets for the model consist of three class of labeled vertical acceleration signals of off-road vehicles driving on three types of road. The training result shows that the proposed model has reliable accuracy among the testing dataset and can be seen as reliable basis for active suspension control strategies.