Phase-sensitive optical time domain reflectometer (Φ-OTDR) leverages Rayleigh scattering for long-distance, real-time monitoring, continuously sensing along the fibre to detect events like environmental disturbances. However, existing detection methods are often compromised by dynamic noise and signal drift, leading to reduced accuracy. This study introduces Lite-PhiOTDR, a streamlined detection network for Φ-OTDR. Initially, a trend-separation and denoising module using one-dimensional convolution is developed. Large-kernel convolution removes low-frequency trends, while a small convolutional network suppresses high-frequency noise, resulting in an input signal that is zero-mean, smooth, and possesses a high signal-to-noise ratio. Additionally, a lightweight feature-extraction structure is implemented to characterise various time-series patterns, including short-term mutations, periodic disturbances, and slow drifts. By employing depthwise separable convolution and a lightweight design, the proposed method significantly reduces parameter count and computational complexity, thereby enhancing detection accuracy and enabling feasible embedded deployment.
Jing Wang (Thu,) studied this question.