Semi-supervised Transfer Learning Method for Anomaly Detection in Industrial Sensors

Abstract

The deployment of industrial sensors has become fundamental to intelligent manufacturing, enabling continuous monitoring and predictive maintenance across various factory environments. However, practical anomaly detection remains challenging due to two persistent obstacles: the scarcity of labeled fault data and significant differences in data distributions between operational domains. This study addresses these issues by proposing a novel semi-supervised transfer learning approach for industrial sensor anomaly detection. The method integrates adversarial domain adaptation with consistency regularization and dynamic pseudo-labeling to fully utilize both labeled and abundant unlabeled data from heterogeneous domains. Experiments are conducted on multiple real-world and benchmark sensor datasets, where the proposed framework is evaluated against supervised, unsupervised, and state-of-the-art domain adaptation baselines. Results show that the new approach achieves significantly improved detection accuracy, robustness to domain shifts, and exceptional data efficiency; for example, reliable anomaly detection can be achieved with as little as 2–5% labeled target data, minimizing annotation costs and deployment delays. The system demonstrates strong adaptability, effectively generalizing across diverse equipment, process lines, and environmental conditions. These findings highlight the practicality and scalability of the proposed framework, offering an efficient path toward reduced manual intervention and enhanced operational reliability in next-generation industrial environments. The study demonstrates that semi-supervised transfer learning has substantial potential in reducing barriers for data-driven maintenance and improving the intelligence and resilience of industrial systems.