φ -OTDR event classification method based on an efficient lightweight 1D CNN

Authors:
Publisher website
Google Scholar
Abstract
Phase-sensitive optical time domain reflectometry ( φ -OTDR) systems frequently encounter a bottleneck where high-accuracy event recognition compromises real-time edge processing capabilities due to the computational demands of complex deep learning models. This paper presents an end-to-end classification framework that resolves this trade-off by integrating discrete first-order difference preprocessing with a lightweight 1D convolutional neural network (1D-CNN). By applying a differential operator to suppress non-stationary baseline drift and Z-score standardization for distributional stability, the proposed method extracts high-frequency transient features from raw time-series data. The subsequent 1D-CNN architecture, characterized by a pyramidal channel expansion and global adaptive average pooling, captures temporal dependencies with minimal structural redundancy. Experimental validation on a public benchmark dataset demonstrates that the model achieves an average accuracy of 99.84%, surpassing state-of-the-art vision transformer (ViT) and MobileNetV2 architectures. Furthermore, the model maintains a parameter footprint of only 42k and achieves an end-to-end inference latency of 0.517 ms on GPU. These results demonstrate that the proposed architecture offers an effective balance between precision and efficiency, enabling robust real-time monitoring on resource-constrained edge devices.
Funding Information
  • National Natural Science Foundation of China (62374015)
  • National Natural Science Foundation of China (62401039)
  • Basic and Applied Basic Research Foundation of Guangdong Province (2024A1515140020)

This publication has 14 references indexed in Scilit: