基于条件式Wasserstein生成对抗网络的电力变压器故障样本增强技术

Data Augmentation Method for Power Transformer Fault Diagnosis Based on Conditional Wasserstein Generative Adversarial Network

数据非均衡问题是制约机器学习技术在电力变压器故障诊断领域中应用效果的关键因素。为克服传统过采样方法未考虑数据整体分布信息的缺陷,提出了一种基于深度学习的故障数据增强方法,以实现样本库的类别均衡化目标。首先,建立梯度惩罚优化的条件式Wasserstein生成对抗网络模型以指导多类别故障样本的生成过程,并克服了原始生成对抗网络模型的训练不稳定问题;然后,构建以油中溶解气体无编码比值为特征参量的栈式自编码器诊断模型,并进一步提出了基于数据增强方法的设备故障诊断技术框架;最后,选用由准确率、F1度量以及G-mean组成的评价指标体系对类别均衡化前后的模型诊断效果进行评估对比。算例研究结果表明,相较于传统过采样方法,提出的故障样本增强方法能够更为有效地改善诊断模型对于多数类的分类偏好问题,提升其整体分类性能,可作为电力变压器故障诊断的重要数据预处理环节。

The imbalanced data problem is a key factor that restricts the application of machine learning technology in the field of power transformer fault diagnosis. In order to overcome the defect of not taking the holistic distribution information of the samples into account in the traditional oversampling method, this paper proposed a fault data augmentation method based on deep learning to achieve the balanced distribution of the samples in different classes. Firstly, the conditional Wasserstein generative adversarial network with gradient penalty was built to guide the generation of multi-category fault samples, and the training instability problem of the original generative adversarial network model was solved. Secondly, the diagnosis model based on stacked autoencoder was constructed by taking the non-code ratios of the dissolved gases in oil as characteristic parameters, and the technical framework of power transformer fault diagnosis based on data augmentation method was further designed. Finally, the evaluation index system consisting of the accuracy, F1 score, and G-mean was selected to compare and analyze the diagnosis effects of the classifier before and after the data augmentation. The case study result indicates that the proposed fault data augmentation method can more effectively improve the classification preference problem for the majority class of the fault diagnosis model and improve its overall classification performance compared with the traditional over-sampling methods, so it can be used as an important link in the data preprocessing of the power transformer fault diagnosis.