基于MHA-1DMini-Xception的光纤复合架空地线覆冰识别方法

Optical fiber composite overhead ground wire icing identification method based on MHA-1DMini-Xception

摘要:

基于深度学习的相位敏感光时域反射计(Φ-OTDR)智能信号处理技术已成为研究热点。本文针对光纤复合架空地线(OPGW)覆冰监测应用,提出了MHA-1DMini-Xception深度学习网络模型。通过改进Mini-Xception网络的深度可分离卷积模块,使其能在参数较少的情况下仍保持模型优良性能,且可有效提取一维时序信号特征；同时引入多头注意力机制,通过并行计算允许模型在不同的表示子空间内学习特征信息,从而增强了模型的表达能力,提升了网络的鲁棒性。将Φ-OTDR采集的OPGW覆冰振动信号划分为无覆冰、一级覆冰和二级覆冰三种类别数据,采用MHA-1DMini-Xception模型对覆冰数据集进行分类识别,对测试集样本的识别准确率达到了97.47%,性能优于CNN模型和Mini-Xception模型。本文方法为光纤振动信号的分类识别提供了有益参考。

Abstract:

Phase-sensitive optical time-domain reflectometer (Φ-OTDR) intelligent signal processing technology based on deep learning has become a research hotspot. In this paper, the MHA-1DMini-Xception deep learning network model is proposed for the application of optical fiber composite overhead ground wire (OPGW) icing monitoring and processing. By improving the deep separable convolution module of the Mini-Xception network, the model can maintain excellent performance with fewer parameters, and can effectively extract one-dimensional time series signal features. At the same time, the multi-head attention mechanism is introduced to allow the model to learn feature information in different representation subspaces through parallel computing, which enhances the expression ability of the model and improves the robustness of the network. The OPGW icing vibration signal collected by Φ-OTDR was divided into three categories of data: no icing, first-level icing and second-stage icing, and the MHA-1DMini-Xception model was used to classify and identify the icing dataset, and the recognition accuracy of the samples in the test set reached 97.47%, which was better than the CNN model and the Mini-Xception model. This method provided a valuable reference for the classification and identification of optical fiber vibration signals.