基于多源特征信息融合的油浸式变压器故障智能诊断模型

Intelligent transformer fault diagnosis model based on multi-source feature information fusion

变压器是电力系统关键设备之一,目前在运行的变压器大多属于油浸式变压器,其发生故障会严重影响系统的安全稳定运行,还可能造成巨大的经济损失,故变压器故障诊断是目前国内外亟待解决的问题。然而,现有研究普遍存在信息来源单一、诊断预测精度不高等问题,为此,本文提出一种基于多源信息融合技术和卷积神经网络的油浸式变压器故障智能诊断模型。首先选择其油色谱分析数据、高压套管的红外图像、放电超声波检测图谱以及特高频局放检测图谱共同作为变压器故障诊断的原始输入信息;其次,针对上述不同类型的原始输入数据,分别采用深度神经网络和卷积神经网络进行特征提取,将提取后的结果进行特征融合;最后综合融合后的特征分析结果进行故障诊断的分类。为避免不均衡样本集造成模型诊断的“倾向性”,本文还提出改进的交叉熵代替传统的损失函数,同时采用kappa系数作为预测结果的评价指标,有效提高了智能诊断模型的预测效果。并通过某电网检修公司提供的实际数据算例进行测试分析,结果验证了本文所提出方法的精准性和有效性。

Transformers are one type of the key power equipment of power system,and most of the transformers currently in operation are oil-immersed transformers.Their failure will not only have a great impact on the secure and stable operation of the system,but also cause serious economic losses.Therefore,transformer fault diagnosis technology is an urgent problem to be solved at home and abroad.However,the existing research generally has the problems of using single information source and having low diagnosis and prediction accuracy.Thus,an intelligent transformer fault diagnosis model is proposed in this paper,based on multi-source information fusion technology and convolutional neural network.Firstly,transformer oil color spectrum analysis data,infrared image of high-voltage transformer bushing,acoustic emission detection spectrum and ultra-high frequency partial discharge detection spectrum are selected as the original input information of transformer fault diagnosis.For different types of original inputs such as numerical and image data,deep neural network and convolutional neural network are used to extract the features of the original information,and the extracted features are fused.Finally,the fused feature analysis results are integrated for classification modeling of fault diagnosis.In order to avoid the tendency of model diagnosis caused by unbalanced sample set,an improved cross entropy is also proposed to replace the traditional loss function,and kappa coefficient is used as the evaluation index of prediction results,which effectively improves the prediction effect of intelligent diagnosis model.Through the test and analysis of the actual data of a certain power grid maintenance company,the results verify the accuracy and effectiveness of the method proposed in this paper.