Abstract
This study investigates anomaly detection in multimodal Internet of Things time series with emphasis on inertial channels X, Y, and Z. A complete and reproducible data processing and modeling pipeline is proposed, covering preprocessing, feature construction, weak labeling, and comparative evaluation of classical and hybrid classical-quantum (CQ) approaches. Experiments are conducted on a real-world dataset comprising 384,614 records with a sampling interval of 10 seconds. The results demonstrate that even simple window-based statistical features provide strong separability between normal and anomalous states. A Random Forest classifier achieves near-perfect performance (ROC-AUC = 0.9982), significantly outperforming both an unsupervised Isolation Forest baseline and the proposed hybrid CQ model (ROC-AUC = 0.8135). Although the hybrid approach does not surpass classical methods in the current configuration, it exhibits stable discriminative capability and confirms the applicability of quantum- enhanced feature transformations. The study establishes a strong practical baseline and identifies key limitations and directions for improving hybrid quantum pipelines in IoT anomaly detection.