深度学习在故障诊断中的应用要点.doc

分类号： 单位代码： 密 级： 学 号： 故障诊断论文作业 题目：深度学习在电力变压器故障诊断应用 姓 名： 黄贤存 学 号： 20152283405 研究方向：基于深度学习的图像处理 二○一六年四月 深度学习在电力变压器故障诊断应用 摘要：由于电力变压器发生故障时油色谱在线监测数据无标签，工程现场往往会得到大量无标签故障样本，而传统的故障诊断方法在对变压器故障类型进行判别时往往无法充分利用这些无标签故障样本。该文基于深度学习神经网络[1](deep learning neural network，DLNN)，构建了相应的分类模型，分析并用典型数据集对其分类性能进行测试。在此基础上提出一种电力变压器故障诊断新方法，它能够有效利用大量电力变压器油色谱在线监测无标签数据和少量故障电力变压器油中溶解气体分析[2] (dissolved gas-in-oil analysis，DGA)实验数据进行训练，并以概率形式给出故障诊断结果，具有更优的故障判别性能，能够为变压器的检修提供更为准确的参考信息。 关键词：故障诊断；电力变压器；深度学习神经网络；溶解气体分析 Abstract：As oil chromatography online-monitoring data is unlabeled during power transformer failure，project sites tend to get a large number of unlabeled fault samples． However，traditional diagnosis methods often fail to make full use of those unlabeled fault samples in judging transformer fault types． Based on deep learning neural network ( DLNN )，a corresponding classification model was established，whose classification performance was analyzed and tested by typical datasets． On this basis，a new fault diagnosis method of power transformer was further proposed，in which a large number of unlabeled data from oil chromatogram on-line monitoring devices and a small number of labeled data from dissolved gas-in-oil analysis (DGA) were fully used in training process． It could generate fault diagnosis result in the form of probabilities，and provide more accurate information for the maintenance of power transformer because of its better performance in fault diagnosis． Keywords：fault diagnosis; power transformer; deep learning neural network; dissolved gas-in-oil analysis(1) (2) 式中:和 通常取为 sigmoid 函数;W 表示输入层与隐含层之间的权值矩阵表示隐含层与输出层之间的权值矩阵;p 表示隐含层的偏置向量;q 表示输出层的偏置向量。为下面表示方便，将 AE 的参数记为 θ。 图 1AE 结构 Fig. 1 AE structure 假设训练样本集 S = {x 1 ，…，x n }，预训练 AE 的过程实质上就是利用 S 对参数 θ 进行训练的过程。为此，我们首先需定义一