Advancing Perimeter Security: Integrating DAS and CNN for Object Classification in Fiber Vicinity

Abstract

This paper presents an advanced perimeter protection system that integrates phase-sensitive Optical Time-Domain Reflectometry ( -OTDR) with Convolutional Neural Networks (CNNs) for real-time event classification near optical fibers. The proposed approach enhances traditional security methods by providing robust monitoring in challenging environments, such as low visibility and large-scale areas. We evaluated multiple signal preprocessing techniques, including Fast Fourier Transform (FFT), Redundant Discrete Fourier Transform (RDFT), Discrete Wavelet Transform (DWT), and Mel-Frequency Cepstral Coefficients (MFCC), to optimize classification accuracy and computational efficiency. While MFCC achieved the highest accuracy (85.61%), RDFT provided the best balance between performance (85.47%) and real-time feasibility, making it the preferred method for deployment. The system successfully differentiates events such as vehicle movement, fence manipulation, and construction work, while anomaly detection capabilities further enhance security by identifying irregular activities with minimal error. These findings demonstrate the potential of integrating fiber-optic sensing with deep learning to develop scalable, real-time perimeter protection solutions for critical infrastructure, border surveillance, and urban security.This paper presents an advanced perimeter protection system that integrates phase-sensitive Optical Time-Domain Reflectometry ( -OTDR) with Convolutional Neural Networks (CNNs) for real-time event classification near optical fibers. The proposed approach enhances traditional security methods by providing robust monitoring in challenging environments, such as low visibility and large-scale areas. We evaluated multiple signal preprocessing techniques, including Fast Fourier Transform (FFT), Redundant Discrete Fourier Transform (RDFT), Discrete Wavelet Transform (DWT), and Mel-Frequency Cepstral Coefficients (MFCC), to optimize classification accuracy and computational efficiency. While MFCC achieved the highest accuracy (85.61%), RDFT provided the best balance between performance (85.47%) and real-time feasibility, making it the preferred method for deployment. The system successfully differentiates events such as vehicle movement, fence manipulation, and construction work, while anomaly detection capabilities further enhance security by identifying irregular activities with minimal error. These findings demonstrate the potential of integrating fiber-optic sensing with deep learning to develop scalable, real-time perimeter protection solutions for critical infrastructure, border surveillance, and urban security.