RT-AD: Real-Time Anomaly Detection of Time-Series Data Based on Lightweight Periodic Transformers for Industrial Edge Devices

Real-time anomaly detection in industrial time-series data is essential for predictive maintenance in resourceconstrained edge environments. However, many existing Transformer-based approaches require substantial computation and memory, limiting their practicality on edge devices. This paper introduces RT-AD (Real-Time Anomaly Detection), a lightweight Transformer architecture designed to reduce processing overhead while effectively capturing periodic behaviors commonly observed in industrial systems. The model incorporates patch-based sequence reduction, efficient channel-correlation learning, and a periodic expert mechanism that adapts to diverse operational cycles. RT-AD employs a forecast-based anomaly detection paradigm where reconstruction errors between predicted and actual values serve as anomaly scores. Experiments across multiple public datasets and real-world HVAC data demonstrate that RT-AD maintains strong forecasting and anomaly detection capability while achieving fast inference suitable for deployment on low-power edge platforms. Comprehensive edge deployment experiments on Raspberry Pi and NVIDIA Jetson series validate real-time inference capability with latency under 50ms using INT8 quantization.