基于LB-IDBO-HKELM样本扩充与参量优选的变压器故障诊断

Fault Diagnosis of Transformers Based on LB-IDBO-HKELM with Sample Augmentation and Parameter Optimization

中文关键词:变压器  故障诊断  样本扩充  参量优选

keywords:transformer  fault diagnosis  sample augmenta-tion  parameter optimization

中文摘要:针对油中溶解气体法应用于变压器故障诊断时存在样本分布不均衡、诊断特征选择标准模糊等的问题，提出一种融合数据扩充与特征优选的混合智能诊断模型（localized randomized affine shadowsampling-borutashap-improved dung beetle optimizer-hybrid kernel extreme learning machine，LB-IDBO-HKELM）。首先，通过构建局部随机仿射阴影采样（localized randomized affine shadowsampling，LoRAS）算法对少数类故障样本实施过采样处理，在有效平衡各类别样本分布的同时，规避了传统过采样方法易引入噪声的缺陷；其次，采用 BorutaShap 算法对变压器油中溶解气体比值类候选特征开展动态优选，筛选出最优诊断特征集，提升模型的可解释性；再次，提出改进蜣螂优化算法（improved dung beetle optimizer，IDBO），通过增强算法全局搜索能力优化混合核极限学习机（Hybrid kernel extreme learning machine，HKELM）的核参数，构建自适应故障分类模型；最后，从样本扩充、特征优选和参数优化3个维度，将所提融合模型与 8 种典型诊断算法进行系统性对比验证。实验对比结果表明，LB-IDBO-HKELM模型在诊断精度、结果一致性（Kappa 系数、Macro-F1、Macro-AUC）及诊断效率方面均表现最优，不仅实现了高精度故障诊断，同时还满足变压器在线监测与诊断的时效性要求。

Abstract:To solve the problems of unbalanced sample distribution and ambiguous diagnostic feature selection criteria in transformer fault diagnosis based on the dissolved gas-in-oil method，a hybrid intelligent diagnosis model integrating sample augmentation and feature optimization is proposed，namely Localized Randomized Affine Shadowsampling-BorutaShap-Improved Dung Beetle Optimizer-Hybrid Kernel Extreme Learning Machine （LB-IDBO-HKELM）. First，a Localized Randomized Affine Shadowsampling （LoRAS） algorithm is developed to perform oversampling on minority fault samples. This method effectively balances the sample distribution across different fault categories while overcoming the drawback of traditional oversampling methods that are prone to introducing noise. Second，the BorutaShap algorithm is employed for dynamic selection of candidate features derived from transformer oil-dissolved gas ratios，and an optimal diagnostic feature set is obtained to enhance the model’s interpretability. Third，an Improved Dung Beetle Optimizer （IDBO） is proposed，which optimizes the kernel parameters of the Hybrid Kernel Extreme Learning Machine （HKELM） by strengthening the global search capability，thereby establishing an adaptive fault classification model. Finally，systematic comparative experiments are conducted between the proposed fusion model and eight typical diagnostic algorithms from three aspects: sample augmentation，feature optimization，and parameter optimization. The results demonstrate that the LB-IDBO-HKELM model exhibits the optimal performance in diagnostic accuracy，result consistency （Kappa coefficient，Macro-F1，Macro-AUC），and diagnostic efficiency. It not only achieves a high fault diagnosis accuracy of 96.15%，but also satisfies the real-time requirements for online transformer monitoring and diagnosis.