Self-Supervised Anomaly Detection in Logistics Quality Data Streams

Abstract

The increasing digitalization of logistics operations has resulted in continuous streams of quality-related data generated from sensors, automation systems, and information platforms. Traditional anomaly detection approaches often rely on labeled datasets or static rules, which are impractical in dynamic logistics environments where quality deviations are rare, evolving, and context-dependent. This study proposes a self-supervised anomaly detection framework for monitoring logistics quality data streams in real time. The proposed approach learns normal operational patterns directly from unlabeled data by leveraging self-supervised learning techniques, enabling the detection of handling, transportation, storage, and sensor-related anomalies without prior fault annotations. The framework incorporates adaptive anomaly scoring and streaming data processing to address concept drift and operational variability. Experimental results demonstrate that the proposed model outperforms conventional unsupervised and supervised baselines in terms of detection accuracy, robustness, and early warning capability. The findings highlight the potential of self-supervised learning as an effective tool for proactive quality assurance in modern logistics and supply chain systems.