为解决复杂多变环境下光纤入侵事件因噪声干扰识别困难、误报率高的问题，提出了基于改进的奇异谱分析和遗传算法优化的双向长短期记忆神经网络的入侵事件识别方法。首先，为了减少噪声对识别效果的影响采用改进的奇异谱分析法去噪，对入侵信号及其分量进行迭代奇异谱分析去噪，并利用信号贡献率的大小来确定信号重构的秩阶次，调节信号分量去噪的程度，实现光纤信号的去噪。然后，利用遗传算法优化神经网络结构参数，构建双向长短期记忆神经网络提取光纤信号空间特征，最后基于以上方法对攀爬、跑动、敲击、静态、大风、雨天6种实测信号进行入侵事件识别实验，实验结果表明，在双Mach-Zehnder光纤周界传感系统识别入侵事件过程中，改进的奇异谱分析相比普通的奇异谱分析，去噪信噪比有明显提高，平均信噪比提高了12.79 dB，平均均方根误差略有减少。遗传算法优化的双向长短期记忆神经网络较未优化神经网络平均识别率提高了5.7%，识别准确率最高可达98.1%。

This study aims to improve the recognition of perimeter intrusion events of the optical fiber sensing system under complex outdoor conditions. An intrusion event recognition method based on improved singular spectrum analysis and genetic algorithm optimized bidirectional long-short-term memory neural network (GA-BiLSTM) is proposed. First, the improved singular spectrum analysis was used to iteratively denoise the optical fiber sensing signal and its components. The signal contribution rate was used to determine the order of signal reconstruction, which controls the denoising process of the signal components, thereby completing the denoising of the optical fiber sensing signal. To recognize intrusion events, the genetic algorithm was used to optimize the parameters of the neural network. Subsequently, a bi-directional long-short-term memory neural network was constructed to extract the spatial characteristics of optical fiber signals. An intrusion event recognition experiment was carried out using the measured optical fiber sensing signals of six events, i.e., climbing, running, knocking, static, windy, and rainy days. The experimental results show that the improved singular spectrum analysis, when applied to the dual Mach-Zehnder fiber perimeter sensing system, exhibits superior denoising performance compared to ordinary singular spectrum analysis. The average signal-to-noise ratio of the consequent signal improved by 12.79 dB. However, the mean root mean square error was slightly reduced. Moreover, the GA-BiLSTM method increased the average recognition rate of intrusion events by 5.7%, with the recognition accuracy rate reaching up to 98.1%.