Dual-stage deep learning model for distributed optical fiber sensing high-risk event identification

A dual-stage deep learning model is proposed to address the insufficient identification rate of high-risk events in distributed optical fiber sensing system. The model consists of two stage, named High-Recall Screening Convolutional Neural Networks (HRS-CNN) and Self-Attention CNN (SA-CNN). HRS-CNN extracts features from 1D vibration signals with a low classification threshold to minimize missed detections. SA-CNN employs an attention-enhanced 1DCNN architecture to suppress noise and concentrate on critical features. Evaluated on a dataset from a field test, the proposed model achieves 96.2% accuracy and an 89.9% recall rate for high-risk event, surpassing a single-stage CNN by 44.2%. The results underscore the architectural advantages of the proposed model for high-risk behavior pattern recognition, having the potential to enhance the reliability and applicability of real-time monitoring system for optical fiber lines.