17 December 2025 in IEEE Access
Modern vehicles generate high-frequency, multivariate data streams from the Controller Area Network (CAN) bus, the fundamental digital nervous system for in-vehicle communication. Although its primary role is real-time control of vehicle subsystems, the CAN bus also serves as a rich source of derived information on vehicle state and driver behavior. These data are particularly valuable for Vehicle-to-Everything (V2X) applications such as driver identification, predictive maintenance, and cooperative perception. However, transmitting raw data is often impractical due to bandwidth limitations in wireless networks. To address this challenge, we propose BinaryCAN, a novel framework that learns compact latent representations directly from automotive time series using autoencoders with binary bottlenecks. Our approach leverages the Straight-Through Estimator (STE) to enforce discrete embeddings while preserving task-relevant semantics.We evaluated three architectures, including a new binary encoder based on the graph neural network (GNN), on downstream tasks such as fuel supplier classification, oil brand identification, speed regression, and driver fingerprinting. Experimental results demonstrate that our learned embeddings not only maintain strong performance in these tasks while achieving substantial compression compared to the raw data, but in some cases even surpass raw data performance. Our work suggests that learned binary embeddings can potentially complement or enhance traditional V2X messaging by enabling more efficient transmission of semantically rich information.