Cascaded Machine Learning Models For Anomaly Detection At The Edge : Optimizing Computational Efficiency in Multi-stage Anomaly Detection for Edge Devices

Sammanfattning

Time series anomaly detection is a crucial aspect of predictive maintenance and fault diagnosis in industrial applications. In edge computing environments, such as connected vehicles, real-time anomaly detection is challenging due to computational constraints. While deep learning models have demonstrated high accuracy in anomaly detection, their resourceintensive nature makes them unsuitable for direct deployment on edge devices. To address this, cascaded machine learning models offer a promising solution, balancing computational efficiency with detection performance. This study investigates a three-stage cascaded machine learning approach for anomaly detection at the edge. The key challenge is to develop a framework that progressively refines predictions while minimizing computational overhead. Despite the growing interest in anomaly detection for edge computing, existing solutions often rely on either lightweight models with limited accuracy or complex deep learning architectures that are impractical for real-time edge deployment. This research bridges the gap by systematically evaluating a range of machine learning models to identify an optimal trade-off between efficiency and accuracy. A variety of machine learning models were explored, including statistical, probabilistic, and deep learning methods. Through extensive experimentation, it was found that the Gaussian mixture model provides a cost-effective solution for anomaly detection on edge devices, offering a reasonable balance between computational efficiency and accuracy. Meanwhile, for cloud-based processing, a fully connected variational autoencoder demonstrated the highest anomaly detection accuracy, making it suitable for offloaded inference in resource-rich environments. Additionally, an adaptive learning mechanism is being integrated to dynamically adjust the offloading threshold during inference, allowing the system to react to varying conditions in real time. To quantify its performance, a regret bound will be computed to measure how far the adaptive decisions deviate from the optimal threshold. By structuring the detection pipeline into a cascaded model and incorporating adaptive thresholding, this approach enables scalable, efficient, and intelligent anomaly detection tailored to resource-constrained edge environments.