Lightweight Anomaly Detection for Edge Intelligence through Compressed Representation Learning

Abstract

This paper addresses the challenges of anomaly detection in edge computing scenarios, including limited computational resources, strict response latency, and dynamic data changes. A lightweight anomaly detection framework is proposed to meet these demands. The method adopts a sliding window mechanism to model multivariate time series data in a structured manner. Feature compression is achieved through random projection. A compact discrimination module is constructed by integrating a nonlinear mapping and sparse connections. This design balances model expressiveness with computational efficiency. During inference, the method does not rely on large-scale parameter training. It enables efficient anomaly detection and adapts well to the low-resource, high-frequency operating conditions typical of edge devices. To systematically evaluate model performance, a series of sensitivity experiments is conducted. These experiments cover key factors such as learning rate, window length, noise intensity, anomaly ratio, and temporal perturbation. The model's adaptability and robustness are assessed using metrics such as Accuracy and F1-Score under different conditions. Experimental results show that the proposed method maintains stable detection capability while keeping inference latency below 10 ms. It effectively handles frequently changing data patterns in complex edge environments. The lightweight modeling strategy and perturbation-aware mechanism presented in this paper offer both methodological support and structural guidance for building deployable and reliable anomaly detection systems in edge intelligence.