An Φ-OTDR event recognition method based on Transformer

Abstract

In recent years, CNN models have become a mainstream deep learning framework and have been widely used in research on improving the accuracy of $\Phi$-OTDR event recognition systems. However, a new deep learning framework, the Transformer model, has recently been proposed to replace CNN. In this article, an $\Phi$-OTDR event recognition method based on the transformer model is proposed. It has been compared with the CNN model, and experiments have shown that the transformer model performs better than CNN on pretrained models, with an accuracy rate of 99.64%. When the model is pruned to only 2 layers, its accuracy can still be saved, and the model size is reduced to 1.49M.

Topics & Concepts