OTDR Event Detection Method Based on Improved 1D UNet

Abstract

Optical time domain reflectometer (OTDR) is the most basic and widely used equipment in optical fiber detection. Its performance and trace analysis ability play a decisive role in the maintenance of optical fiber. Traditional OTDR event detection methods rely on manual definition of pulse characteristics, require professional prior knowledge, and require high signal-to-noise ratio. The traditional preprocessing methods such as smoothing and denoising have some actual signal characteristics weakened or even disappeared. UNet is the most classical U-structured network model applied to medical image segmentation. It can learn a very robust model for edge extraction by using a small amount of data. Inspired by this, we propose the first OTDR event detection method based on the improved 1D UNet, which makes full use of the convolution neural network to automatically extract signal features. It can be applied to small sample data sets and it can accurately identify multiple types of events such as power injection, reflection, drop, end and echo events, with an average detection rate of 90%. Compared with the EXFO FastReporter software widely used in the industry, our method shows a stronger ability to resist noise interference, and the detection of echo events in high noise areas reaches 89%.