Experimental study of unpaved road identification by optimized LSTM

Abstract Identification of Road Unevenness on Unpaved Surfaces is crucial in facilitating the health status evaluation of engineering vehicle suspension systems and enhancing the optimization of active control strategies. However, the deformation of tires and road surfaces and the interaction between them pose a significant challenge in accurately identifying road roughness. This paper proposes a method for identifying soft road roughness grades based on a dual-input LSTM (long short-term memory) network. Firstly, a whole vehicle dynamics model of a mining dump truck with deformed tire-soft road coupling was established, and the vertical acceleration response of the wheel center under the consideration of coupled road and tire deformation was collected. The data features were then transformed using Hilbert envelopes. Finally, a dual-input LSTM network model was designed to achieve the road unevenness of multiple mixed scenarios using typical vehicle speed and wheel center acceleration as inputs. Grade-accurate recognition was achieved. The results show that the overall identification accuracy reaches 97%, with high robustness and recognition accuracy under various conditions of vehicle speed, load, damping, and random changes in sampling duration. In comparison to alternative methodologies, the optimized LSTM model attains a recall rate and accuracy of 97.92% and 98.2%, respectively, across various road grade classifications, thereby demonstrating exceptional performance. Finally, field experiments with mining trucks were conducted, and the acceleration of heavy-duty mining dump trucks was collected and verified using the analytical model. The results indicate that this method has good identification accuracy under conditions of vehicle speeds ranging from 10–30 km h −1 , both empty and fully loaded. The proposed method significantly enhances the environmental adaptability of engineering vehicles in complex dynamic scenarios through high-fidelity soft terrain roughness identification.