A Multivariate Time Series Prediction Method for Automotive Controller Area Network Bus Data

• DOI: 10.3390/electronics13142707
• Publication Date: 2024-07-10T19:22:05Z
• Authors: Dan Yang; Shuya Yang; Junsuo Qu; Ke Wang
• Keywords: CAN bus

This study addresses the prediction of CAN bus data, a lesser-explored aspect within unsupervised anomaly detection research. We propose Fast-Gated Attention (FGA) Transformer, novel approach designed for accurate and efficient data. model utilizes cross-attention window to optimize computational scale feature extraction, gated single-head attention mechanism in place multi-head attention, shared parameters minimize size. Additionally, generalized unbiased linear approximation technique speeds up block computation. On three datasets—Car-Hacking, SynCAN, Automotive Sensors—the FGA Transformer achieves predicted root mean square errors 1.86 × 10−3, 3.03 30.66 with processing 2178, 2768, 3062 frames per second, respectively. The provides best or comparable accuracy speed improvement ranging from 6 170 times over existing methods, underscoring its potential data prediction.