Evaluating Anomaly Detection Techniques in Industrial Environments: A Comparative Analysis of Autoencoders, Deep SVDD, and Supervised 2D CNNs

As industrial systems become increasingly complex, the need to improve operational efficiency and ensure worker safety is more urgent than ever. Radar-based monitoring offers a promising solution, but the resulting high-dimensional data presents challenges for real-time analysis and anomaly detection. In this study, we propose a radar-based anomaly detection framework built on Orthogonal Time Frequency Space (OTFS) modulation, which transforms raw radar returns into informative spatio-temporal features. Our approach integrates three deep learning models—Deep SVDD, Autoencoders, and a supervised 2D Convolutional Neural Network (CNN)—to identify abnormal movements that deviate from typical worker behavior. To boost the performance of unsupervised methods, we introduce a dynamic thresholding mechanism that adjusts to shifts in environmental conditions, improving reliability in noisy and cluttered scenes. In evaluations using real-world radar data from industrial settings, the supervised 2D CNN achieved 99.9% accuracy, while all models recorded F1-scores between 0.98 and 0.99. Notably, Deep SVDD delivered the fastest inference time at 1.71 seconds, supporting the feasibility of real-time deployment. Additionally, lightweight Transformer-based models were compared, showing comparable accuracy but higher computational cost, reaffirming the practicality of the proposed designs for edge-oriented industrial sensing.