Multitask Vehicle Signal Recognition With Dual-Speed Adaptive Weighting

In mixed traffic environments, the accurate identification of vehicular devices’ modulation schemes, communication protocols, and emitter device information directly affects perception capabilities toward surrounding vehicles and infrastructure. However, existing studies predominantly focus on single-dimensional information analysis, resulting in limited completeness and accuracy in signal feature interpretation. This paper proposes a multitask learning framework (DSR-CNN-LSTM) for collaborative identification of this information. Furthermore, to mitigate task conflicts and noise interference, a dual-rate adaptive weight adjustment strategy is developed to optimize model performance through dynamic balancing of task learning rates and gradient update speeds. Experimental results demonstrate the superior performance of the DSR-CNN-LSTM framework in complex communication environments: Modulation recognition accuracy shows improvements of 20.67%, 10.38%, and 9.96% on three open-source datasets, while the weighted average recognition accuracy for communication protocols and emitter device information achieves enhancements of 45.52%, 72.21%, and 11.11%, respectively. The proposed model outperforms existing methods in both recognition precision and anti-interference capabilities, providing novel technical insights and solutions for the advancement of intelligent connected vehicle technologies.