Real-Time Multi-Class Disturbance Detection for Φ-OTDR Based on YOLO Algorithm.

Area of Science:

• Optoelectronics and Photonics
• Artificial Intelligence and Machine Learning
• Security and Surveillance Systems

Background:

• Distributed optical fiber sensing systems (DOFS) using phase-sensitive optical time-domain reflectometry (Φ-OTDR) are crucial for perimeter security.
• Real-time detection and classification of diverse intrusion events remain a challenge for existing DOFS technologies.

Purpose of the Study:

• To develop a real-time, multi-class disturbance detection algorithm for Φ-OTDR-based DOFS.
• To enhance the speed and accuracy of identifying and categorizing external intrusions.

Main Methods:

• A novel algorithm based on the YOLO (You Only Look Once) object detection framework was proposed.
• The Darknet53 network was employed to integrate event localization and classification into a single-stage process.
• A dataset of 5787 spatial-temporal sensing images from five event types was collected and utilized for model training.

Main Results:

• The YOLO-based algorithm achieved a detection speed of 22.83 frames per second (FPS) with 96.14% accuracy.
• The proposed method demonstrated a significant speed improvement: 44.90 times faster than Fast-RCNN and 3.79 times faster than Faster-RCNN.
• The algorithm successfully enabled real-time localization and classification of intrusion events from continuous sensing data.

Conclusions:

• The developed YOLO-based algorithm provides an effective solution for real-time, multi-class external intrusion detection and classification in Φ-OTDR-based DOFS.
• This approach significantly enhances the operational capabilities of fiber optic sensing systems for practical security applications.
• The study validates the potential of deep learning, specifically YOLO, in advancing the performance of distributed fiber vibration sensing.