The application of neural networks driven by nonlinear model data in road roughness estimation

Abstract Road roughness significantly impacts vehicles’ transportation performance. The purpose of this study is to develop an innovative, cost-effective, and precise method for estimating road roughness based on acceleration sensors. Unlike other approaches, this method employs a nonlinear full-vehicle dynamic model and a high-performance Gaussian nonlinear autoregressive with external inputs (G-NARX) neural network to significantly enhance the accuracy, without additional costs. In this study, acceleration sensors would first capture unsprung mass acceleration signals, and the neural network is trained with acceleration and velocity inputs. Then, the trained model would generate the root mean square value of the power spectral density to estimate the road roughness. Test results reveal that the G-NARX Inputs neural network outperforms the NARX neural network significantly, with the classification accuracy improved by at least 26% and the overall accuracy exceeding 90%. In road tests, the nonlinear model used in this study showed a 20% accuracy improvement over linear models. And field experiments conducted on three common types of roads further validated the robustness and accuracy of this study’s estimation method.