Abstract: The moving loads borne by bridges during long-term service are core indicators for assessing their operational status and remaining service life. However, existing load identification methods generally rely on high-fidelity finite element models, are sensitive to noise, and struggle to simultaneously reconstruct multi-parameters—such as vehicle weight and speed—under conditions of sparse sensor deployment. To address these issues, this paper proposes a lightweight multi-task learning model combining Convolutional Neural Networks (CNN) and Gated Recurrent Units (GRU). Utilizing only mid-span and quarter-span accelerations as inputs, the model first extracts local time-frequency features via the CNN and then combines them with the GRU to capture global temporal dependencies. The architecture creates parallel outputs for vehicle gross weight and driving speed to achieve the joint prediction of total vehicle load and velocity. Research results indicate that this method achieves a rapid convergence speed and maintains the Mean Squared Error (MSE) of prediction at a low level. The proposed CNN-GRU model demonstrates high prediction accuracy and generalization capability, retaining strong robustness even under certain noise levels. Consequently, it provides a feasible, lightweight, and intelligent solution for bridge moving load identification.