面向电力通信网的光纤感知事件智能识别方法研究

摘要

针对电力通信网光纤感知事件识别中存在的时空特征混淆、复杂噪声干扰及误报率高等问题，提出一种融合多域特征提取与智能分类的混合模型。首先，构建基于深度可分卷积与Log-Mel的时空特征提取框架，通过并行提取振动信号的时间依赖性与空间关联性，有效区分挖掘、敲击等时空特征相似事件；其次，引入EMA注意力机制动态优化特征权重，结合Deformable Attention ViT捕捉数据间的时空依赖，提升对复杂现场环境下产生的低信噪比事件的识别鲁棒性；最后，采用更低计算量的神经网络结构作为主体，提高检测实时性。我们在电网光纤监测情景下进行数据收集，共收集了6478个样本，捕获了29种类型的事件，该方法在数据集上的识别准确率达94.5%，显著优于传统方法。本研究为电力通信光缆的智能化运维提供了有益的技术参考。

Abstract

To address the confusion of spatio-temporal features, complex noise interference and high false-alarm rates in the recognition of optical fiber sensing events in power communication networks, a hybrid model that combines multi-domain feature extraction with intelligent classification is proposed. First, a spatio-temporal feature extraction framework based on depthwise separable convolution and Log-Mel spectrograms is constructed; by extracting the temporal dependency and spatial correlation of vibration signals in parallel, events with similar spatio-temporal characteristics, such as excavation and knocking, are effectively distinguished. Second, an EMA attention mechanism is introduced to dynamically optimize feature weights, and a Deformable Attention ViT is used to capture spatio-temporal dependencies among the data, improving the robustness of recognition for low signal-to-noise ratio events generated in complex field environments. Finally, a neural network structure with lower computational cost is adopted as the backbone to improve real-time detection performance. Data were collected in a power grid optical fiber monitoring scenario, yielding 6478 samples covering 29 event types. The proposed method achieves a recognition accuracy of 94.5% on this dataset, significantly outperforming conventional methods. This study provides a useful technical reference for the intelligent operation and maintenance of power communication optical cables.