AutoMixer: A Lightweight and Scalable Industrial 5.0 Safety Assurance Model with Multi-Scale Adaptive Dual-Attention

Lei Wang, Huaming Wu, Wei Yu, Jun Wang

Keywords: Spatio-temporal prediction, Efficiency optimization, Sequence frequency decomposition, Dual cross-attention

Abstract: With the rapid growth of intelligent transportation and industrial automation, traffic safety management and industrial system safety generate vast amounts of spatio-temporal data. These data offer rich temporal and spatial patterns for analysis but pose significant challenges, including dynamic traffic patterns, high-dimensional sensor data, and complex anomalies in industrial systems. Traditional methods struggle to capture nonlinear accident patterns, handle noisy sensor data, or model intricate multi-variable interactions, especially in real-time scenarios. Although deep learning and large-scale models have improved the accuracy of accident prediction and anomaly detection, their reliance on complex spatial operations and large parameter sizes creates computational bottlenecks, limiting scalability in large-scale and real-time safety applications. Therefore, we propose AutoMixer, a lightweight and scalable model that avoids explicit spatial modeling. It uses a dual cross-attention module to identify coupled trend and periodic features in multi-resolution spatio-temporal data. Extensive experiments demonstrate that AutoMixer consistently outperforms state-of-the-art baselines, achieving 7% higher detection accuracy while effectively handling large-scale node distributions and high-frequency data. AutoMixer provides a practical and deployable solution for real-time accident detection and industrial system safety analysis, enhancing computational efficiency and applicability in resource-constrained environments, thus optimizing performance for large-scale traffic and industrial safety tasks.

Primary Area: learning on time series and dynamical systems

Submission Number: 5517