基于改进卷积神经网络的变压器有载分接开关故障自适应识别方法  

Modified Convolutional Neural Network-based Adaptive Fault Identification for Transformer On-load Tap Changer

作　　者：高志刚 GAO Zhigang(Guoneng Inner Mongolia Hulunbuir Power Generation Co.,Ltd.,Hulunbuir 021025,China)

机构地区：[1]国能内蒙古呼伦贝尔发电有限公司,内蒙古呼伦贝尔021025

出　　处：《电工技术》2024年第20期66-70,共5页Electric Engineering

摘　　要：常规变压器有载分接开关故障自适应识别多采用改进半监督阶梯网络算法,但由于无法解决网络梯度爆炸问题,最终的故障识别精度较低,因此提出基于改进卷积神经网络的变压器有载分接开关故障自适应识别方法。依据变压器有载分接开关的基本组成结构,利用小波包分解算法与信号的频域识别向量挖掘其中的故障特征参量,采用一维卷积神经网络对特征参量进行融合处理,并引入注意力机制改进与优化网络结构参数,进而构建故障自适应识别模型,通过残差结构解决网络结构的梯度爆炸问题,求取输入样本的故障综合评分,确定样本的所属故障类型,由此实现故障自适应识别。实例应用结果显示,所提方法能够有效识别有载分接开关故障,识别结果与实际一致,并且F1_score最高值达到0.97,因此所提方法具备较高的识别精度。The conventional adaptive identification of on-load tap changer faults in transformers often uses improved semi supervised ladder network algorithms,but due to the inability to solve the problem of network gradient explosion,the final fault identification accuracy is low.To this end,this work studied anomaly detection schemes for switchgear.Wavelet packet decomposition algorithm and frequency domain recognition vector of signal are used to mine the fault characteristic parameters.Utilizing conventional networks,features were integrated and network structures were updated.Then a fault adaptive recognition model was constructed,and the gradient explosion problem of the network structure was solved through residual structure.The comprehensive fault score of the input sample was obtained to determine the type of fault the sample,thereby achieving adaptive fault recognition.The example application results showed that the proposed method can effectively identify the fault of the on-load tap-changer,and the identification results are consistent with the actual results,and the highest F1_score value is 0.97,which has a high identification accuracy.

关 键 词：改进卷积神经网络 变压器有载分接开关 故障自适应识别 识别精度 

分 类 号：TH113.1[机械工程—机械设计及理论]