Modelagem com deep learning para previsão do RMSVA da suspensão de veículos com base nos parâmetros da suspensão, velocidade e condições do pavimento

Abstract

Numerical simulations in Simulink deliver accurate suspension responses but become costly when thousands of parameter combinations must be swept. This work introduces a deep learning (DL) surrogate that predicts the Root Mean Square Vertical Acceleration (RMSVA) from suspension parameters, vehicle speed, and road roughness (International Roughness Index – IRI), cutting computation time without sacrificing fidelity. The dataset covers nine vehicle classes and 36 track profiles per speed (30–120 km/h), with IRI sampled every 100 m and five-sample temporal windows. The architecture combines multi-scale 1D convolutions, an attention mechanism to weight relevant segments, and vehicle-class embeddings, trained with leave-one-vehicle-out validation. The model achieved an average coefficient of determination (R2) of about 0.95 and an average Root Mean Squared Error (RMSE) near 0.10, closely matching the reference responses. The resulting tool accelerates suspension scenario exploration and design decisions by delivering comfort metrics rapidly.