随机配置网络(SCN)通过引入不等式约束来限制输入权重和偏置的赋值,随着节点数量增加,网络能够逼近任意的数学函数和数据模型。在构建SCN的过程中,由于网络本身性质以及样本数据的不适定性和病态条件等问题会引起网络的过拟合,故提出一种基于Dropout技术的改进型SCN模型(Dropout-SCN)来自适应地约束输出权重分布和大小,以此来提高网络模型的识别精度。光纤数据验证的结果表明:与传统的SCN和L2范数正则化的SCN模型相比, Dropout-SCN模型具有更低的测试误差,有效地减缓了网络过拟合问题,提高了对光纤预警系统(OFPS)中光纤入侵信号的识别准确率。

为了对分布式光纤上的入侵信号类型进行准确识别,提出了一种基于集合经验模态分解(EEMD)结合随机向量函数链接(RVFL)神经网络的光纤入侵信号的特征提取与识别算法。算法步骤为:对采集到的光纤入侵信号作预处理操作,包括最小-最大规范化处理和利用db3小波去除信号的低频噪声;采用EEMD方法对入侵信号进行分解,得到5组本征模态函数(IMF);计算各IMF分量的能量占比,并依据方差分析法筛选出3组特征向量;将特征向量送入RVFL神经网络进行训练并对入侵信号进行识别。实验结果显示:该方法能正确识别不同入侵信号的类型,具有较高的准确率。

Optical fiber pre-warning system (OFPS) is widely utilized in pipeline transport fields. The intrusions of OFPS need to be located. In this system, the original signals consist of noises, interferences, and intrusion signals. Here, noises are background and harmless interferences possessing with high power, and the intrusion signals are the main target of detection in this system. Hence, the study stresses on extracting the intrusion signals from the total ones. The proposed method can be divided into two parts, constant false alarm rate (CFAR) and dilation and erosion (DE). The former is applied to eliminate noises, and the latter is to remove interferences. According to some researches, the feature of noise background accords with the CFAR spatial detection. Furthermore, the detection results after CFAR can be presented as a binary image of time and space. Besides, interferences are relatively disconnected. Consequently, they can be eliminated by DE which is introduced from the image processing. To sum up, this novel method is based on CFAR and DE which can eliminate noises and interferences effectively. Moreover, it performs a brilliant detection performance. A series of tests were developed in Men Tou Gou of Beijing, China, and the reliability of proposed method can be verified by these tests.

针对相位敏感光时域反射计(Φ-OTDR)光纤预警系统对一维信号进行模式识别产生的误报和较低的识别效率, 提出基于形态学方法提取时空二维信号特征, 并利用相关向量机(RVM)分类器对事件进行分类识别的方法。首先, 将Φ-OTDR采集到的时空二维信号当作图像, 根据信号在图像上的特征采用图像处理的方法对不同入侵事件信号进行阈值分割。然后, 基于本文提出的特征提取方法, 利用不同事件区域在幅值、面积、形状以及区域间隔上的差别提取不同信号特征。最后, 利用相关向量机分类器对不同事件信号进行识别并采用"一对一"的多分类策略。对3种管道安全事件进行了实验。实验结果表明, 本文提出方法的识别精度能够达到97.8%, 而算法时间不到1 s。 与传统模式识别方法相比, 提出的算法大幅度地改善了系统性能, 且简便易行, 能够满足Φ-OTDR光纤预警系统在线实时监测的要求。