Testing Updations

Real-time anomaly detection in streaming sensor data is critical for applications ranging from industrial monitoring to autonomous vehicles, yet most high-accuracy models are too computationally expensive for deployment on edge devices. In this work, we propose a lightweight transformer-based architecture optimized for low-latency anomaly detection in multivariate time series. Our model reduces parameter count by 68% compared to standard transformer baselines through a combination of sparse attention and depth-wise separable convolutions, while maintaining competitive detection accuracy. We evaluate our approach on three public benchmark datasets (SWaT, WADI, and SMD) and demonstrate that it achieves an F1 score within 2% of state-of-the-art methods while running 4.3x faster on embedded hardware. We further introduce a novel early-exit mechanism that allows the model to make confident predictions on clearly normal samples without executing the full network, reducing average inference time by an additional 31%. Our results suggest that carefully designed lightweight architectures can bring near-state-of-the-art anomaly detection capabilities to resource-constrained edge environments, enabling broader deployment of real-time monitoring systems in industrial and IoT settings.