Research on Vehicle Rollover Risk Prediction Based on CNN-LSTM and Unscented Kalman Filter Algorithm
Xiaoqiang Tan, Zefan Li, Xingyu Wang, Kai Liu, Chengbao Zhang, Guangqiang Wu
Abstract
SUVs are highly prone to rollover accidents during high-speed cornering due to their elevated center of gravity, often leading to severe accidents and increased fatality rates. Accurately predicting rollover risk in real time is challenging due to the complexity of vehicle dynamics. This article proposes a novel method for predicting rollover risk within the next second by integrating advanced deep learning techniques with phase plane analysis. This approach begins with a rollover risk assessment model based on roll angle and roll rate phase plane analysis, which considers the dynamic height between the vehicle’s center of mass and the roll axis, estimated using the unscented Kalman filter (UKF). To ensure accurate and timely prediction of rollover risk, a hybrid convolutional neural network (CNN) and long short-term memory (LSTM) model are introduced. The CNN module extracts spatial features from vehicle dynamics data, while the LSTM module captures temporal dependencies. By leveraging this combined architecture, the model effectively predicts changes in roll angle and roll rate, enabling prompt assessment of rollover risk. Simulation results indicate that the proposed CNN-LSTM model can predict roll angle and roll rate up to one second in advance, achieving root mean square errors (RMSEs) of 0.0065 rad and 0.0164 rad/s, respectively, after denormalization. These findings demonstrate the effectiveness of the proposed model in forecasting vehicle rollover risk, highlighting its potential to enhance the safety and reliability of SUV rollover prevention systems.