Deployment of an Unsupervised Anomaly Detection Model Using Anomalib and PyTorch, Is it feasible on a low-powered edge-device?

Abstract

The deployment of pre-trained unsupervised anomaly detection models on low-cost and low-powered edge devices, specifically the Raspberry Pi 5, is a promising approach for cost effective and scalable solution for real-time monitoring in production environments. This thesis investigates the plausibility and performance of running such models on the RP5, focusing on their ability to accurately detect anomalies in real-time. This thesis addresses the challenges with hard hardware limitations, software configuration, dataset creation and model performance in an edge environment. To enable the training and validation of the model a custom dataset consisting of mugs stained with food coloring to act as anomalies. While the model successfully ran on the RP5 the inference results demonstrated a lack in accuracy with false positives and negatives as-well as a cycle time of 2000-3000 ms per image, was deemed to slow for real-time applications. Although the findings suggest that with further optimizations, such reducing the resolution of input data and further developing the inference script, the cycle time could be significantly reduced. As well as improving the accuracy by reducing the prevalence of false positives and negatives. Thus the model could be an effective solution for real-time anomaly detection.