神经网络在变压器故障诊断中的应用研究

Application of Neural Network in the Transformer Fault Diagnosis

为达到良好的故障诊断效果,神经网络训练样本通常选取具有代表性的、紧凑的样本,但是由于变压器内外部存在不确定因素的影响,现实中变压器溶解气体量跨度远大于训练样本跨度。小样本训练不能够满足变压器故障诊断的实际需求,而大样本量存在训练样本跨度范围大、样本分散问题,会降低神经网络的泛化能力。针对数据量大、样本分散,利用累积频率归一化对数据进行规范化处理,用441组故障数据进行网络训练和网络检测,构建了BP神经网络、RBF神经网络和GRNN神经网络故障诊断模型,对数据规范化前后仿真结果进行对比,结果表明,累积频率归一化的数据规范化方法可不同程度地提高神经网络的变压器故障诊断效果,综合诊断效果和混合故障类型诊断效果均优于传统的三比值法,具有实际应用价值。

To achieve good fault diagnosis effect, the training samples of the neural network are usually representative and compact. However, because of the influence of the transformer internal and external uncertainties, the real amount of dissolved gas in transformer span is greater than the span of the training sample. Small sample training cannot meet the practical requirements of transformer fault diagnosis and large sample size has the problem of wide range and dispersion. It will reduce the generalization ability of neural network. In view of the problems such as large amount of data and sample dispersion, the BP neural network, RBF neural network and GRNN neural network fault diagnosis model is built, the simulation results before and after data standardization are compared by using the cumulative frequency normalization processing on the data standardization and putting 441 groups of failure data into network training and network detection. The results are as follows： the cumulative frequency normalized data standardization method improves the neural network in different degrees of transformer fault diagnosis effect, and comprehensive diagnosis effects and mixed type fault diagnosis effects are superior to the traditional three ratio method. Therefore, it has the value of practical application.