Optical Fiber Vibration Signal Identification Method Based on Improved YOLOv4

Area of Science:

• Sensor Technology
• Signal Processing
• Machine Learning

Background:

• Traditional early-warning systems process optical fiber vibration signals sequentially.
• Integrating endpoint detection and pattern recognition is crucial for improving efficiency.
• Current methods face challenges in real-time processing and accuracy.

Purpose of the Study:

• To develop an improved You Only Look Once v4 (YOLOv4) model for simultaneous endpoint detection and pattern recognition of optical fiber vibration signals.
• To enhance the speed and accuracy of vibration signal identification in security systems.
• To address the limitations of existing target detection algorithms in terms of complexity and dataset imbalance.

Main Methods:

• Implemented a deep separable convolution (DSC) network to reduce YOLOv4 model complexity.
• Utilized a focal loss function to address unbalanced sample classification.
• Visualized real-time collected signals as oscillograph and time-frequency datasets for detection.
• Combined endpoint detection and pattern recognition into a unified target detection algorithm.

Main Results:

• Achieved high mean Average Precision (mAP) of 98.50% and 93.48% on different datasets.
• Reached impressive frames per second (FPS) of 84.8 and 69.9, indicating significant speed improvements.
• The improved algorithm demonstrated a detection speed approximately 20 times faster than other optical fiber vibration signal recognition algorithms.
• Outperformed the standard YOLOv4 model in both mAP and FPS metrics.

Conclusions:

• The enhanced YOLOv4 algorithm effectively integrates endpoint detection and pattern recognition for optical fiber vibration signals.
• The proposed method significantly improves detection accuracy and real-time processing speed for early-warning systems.
• This approach reduces false alarms and enhances the overall performance of security systems by enabling rapid and accurate identification of external intrusions.