A Human Intrusion Monitoring Strategy Based on Distributed Sensor Networks to Tackle Label Noise
Ning Xu, Yiwei Li, Fuyang Chen, Ruihang Xu · IEEE Transactions on Instrumentation and Measurement · 2025
This article proposes a human intrusion monitoring strategy based on distributed fiber-optic sensor (DFOS) systems, specifically designed to address the label noise challenges caused by the nonlinear spatial attenuation from varying external environments and uncertainties from human intrusions. Utilized for dynamic measurements of the physical environment, DFOS systems comprise various sensor nodes that capture anomalous vibration signals. Due to the distributed nature of DFOS systems, adjacent nodes frequently detect overlapping vibration signals. This, coupled with unknown interference events and variable delays in data acquisition, poses significant challenges to the accurate labeling of data. A dual-enhanced convolutional neural network (CNN)–Transformer co-teaching (DCTC) strategy is proposed that integrates the robust features of CNNs and Transformers to effectively handle the multiscale dependencies and uncertainties. By using CNNs to capture spatial features and Transformers to model long-range temporal dependencies, the DCTC method enhances the capability to discern and isolate true anomalies from noisy data. The DCTC employs a co-teaching strategy between two neural networks which facilitates mutual learning and improves resistance to label noise. This is further augmented by Bootstrapping Loss, which adjusts the training loss to prioritize reliable labels, thereby enhancing the overall robustness of the model. The practicality of the DCTC strategy is demonstrated through real-field experiments using DFOS in a high-speed train human intrusion monitoring system. These experiments validate the method’s ability to significantly enhance monitoring accuracy in environments with label noise.