基于边际效应带通阻隔与优选特征深度映射的高压电缆接头多源局放分离与识别方法

Partial Discharge (PD) monitoring of High Voltage (HV) cables, which has already been recommended by the international standards and organizations, has not been widely applied in the industrial application in China because of the great challenges, including: 1) PD could arise from a number of sources, the cable insulation itself, intermediate joints and terminations, and furthermore PD is also simultaneously interfered by the interference signals and other PD that generated by the apparatuses which are connected to the cable ends, so that it is challenging to separate PD signals from multiple sources; 2) There are different types of PD in the different parts of the cable system, and some of the PD types are with high similarity which are difficult to be distinguished. In order to overcome the challenges, the research will concentrate on the following two aspects: 1) A cable system will be set up in the lab, based on which the PD testing of the different types of artificial defects of the cable systems will be carried out. PD signals from multiple sources will be synchronously detected. The margin effect of PD signals under the conditions of short distance transmission and long distance transmission will be studied, based on which a novel band-pass filter which can block interference signals and PD from other areas of the cable system, will be established; 2) Aiming at to distinguish the different types of PD signals from the same area of the cable, feature extraction will be carried out to generate high dimensional new PD features. Then optimal feature selection, feature transformation and feature deep mapping will be studied based on the mapping and visualizing of chosen features group. A new combined optimal feature selection, feature transformation and deep learning based PD recognition method will be established. The research is intended to establish and improve the theories of PD extraction and pattern recognition from multiple sources of cable system, in terms of the recognition sequence and the recognition depth, improve the recognition accuracy, and lay a theoretical foundation for the pervasive industrial application of PD monitoring of HV cable systems.

高压电缆局部放电（简称局放）监测已被国际标准组织认可与推荐，但因监测难度较大，在我国工业界暂未广泛应用，其难点包括：1）电缆局放不但来自本体、中间接头、终端接头不同源头，还与其两端连接设备的局放和干扰混叠，多源信号分离难度大；2）电缆同区域局放也有多种类型，部分类型之间相似度高，识别困难。针对这两个难点：1）搭建模拟工业现场的电缆系统，设置典型人工缺陷，开展加压测试，多测点同步耦合多源头局放信号，研究局放近传输和远传输条件下的边际效应，建立一种新的边际效应带通阻隔法，阻隔区域外局放和干扰；2）对同一区域高相似度局放，构建高维度多元新特征，通过对优选特征的逐层映射与可视化，研究特征变换与深度映射对多类型局放的表征能力，建立一种新的特征寻优-特征变换-深度学习局放识别方法。本申请旨在从识别的层次与深度两个方面，建立电缆多源局放的分离方法和识别方法，提升识别精度，推动电缆局放应用基础科学的发展。

本项目的主要研究内容包括：研究基于边际效应数学模型的电缆接头局放多维参数带通阻隔方法；研究电缆多源局放新特征构建与核心表征特征优选；研究基于高维特征优化排列和深度学习的电缆接头多源局放识别方法，取得的主要成果有：.（1）搭建了含5种人工缺陷的高压电缆PD检测实验平台，并开展了PD加压测试。人工缺陷类型包括外半导电层尖端、绝缘层刀痕、绝缘层气隙、半导电层杂质和悬浮电位缺陷。在不同电压等级下，对每种缺陷类型获得了700组PD信号。.（2）通过仿真研究，搭建了高压电缆频变参数暂态模型，得到了局部放电信号沿电缆导体层和金属屏蔽层传播时的衰减特性，得到了局部放电信号特征参数在不同传输距离下的衰减数学模型。提出了基于边际效应的高压电缆局部放电带通阻隔方法。算例分析验证了该带通阻隔方法能有效剔除边界外其他类型的局部放电信号。.（3）构建了高压电缆PD多维候选特征集合，并研究了基于随机决策森林的高压电缆PD特征寻优。通过实验室数据进行了验证，分析得到了表征PD的新特征。不同类型PD信号的特征寻优结果表明，小波组合特征是不同类型PD识别的有效特征参数。.（4）研究了基于卷积神经网络的高压电缆PD深度学习模式识别。研究了不同网络层数、不同激活函数以及不同池化方式对CNN性能的影响，并将CNN识别效果与传统的浅层分类器SVM和BPNN进行了比较，结果表明三种方法中CNN总体识别准确率最高。CNN中的卷积层和池化层具有优异的特征学习能力，能捕捉到数据的细节特征，信息丢失少，对高相似度缺陷识别能力强、鲁棒性好。.本项目从识别的层次与深度两个方面，建立了电缆多源局放的分离方法和识别方法，提升了识别精度，对电缆局放应用基础科学的发展起到一定的推动作用。.基于该项目的成果，进一步获得了国家自然科学基金面上项目的连续资助，孵化了后续超过530万的电缆局放工业在线监测项目，并在工业现场发现5例局部放电案例，把论文写在了祖国的大地上。