陈奇志1  赵沛舟2  赵海全2  蔡锦涛2  谢昌富3

（1.成都交大光芒科技股份有限公司，四川 成都 610041  

2.西南交通大学，四川 成都 611756

3.深圳市地铁集团有限公司，广东 深圳 518026）

摘要：针对传统轨道交通变压器种类多、内部结构复杂、运行工况多样，导致对其进行故障诊断较为困难的问题，提出深度学习融合线性预测倒谱系数（LPCC）和梅尔频率倒谱系数（MFCC）组合特征的轨道交通变压器故障诊断方法。首先，利用小波阈值去噪法对噪声信号预处理；然后，分别提取噪声信号的LPCC特征和MFCC特征，并组合成特征向量；最后，将组合特征向量输入到基于深度学习的CNN-LSTM模型，实现轨道交通变压器的故障诊断。实验结果表明，该文提出的LPCC-MFCC组合特征和CNN-LSTM模型对轨道交通变压器的故障诊断准确率可达99.48%，精度、召回率和F1分数均达到99.59%。

关键词：轨道交通变压器；故障诊断；噪声信号分析；特征提取；深度学习

中图分类号：TP3            文献标志码：A              文章编号：1674-2605(2023)03-0008-06

DOI：10.3969/j.issn.1674-2605.2023.03.008

Fault Diagnosis Method for Rail Transit Transformer                   Based on Deep Learning 

CHEN Qizhi1  ZHAO Peizhou2  ZHAO Haiquan2  CAI Jintao2  XIE Changfu3

(1.Chengdu Jiaoda Guangmang Technology Co., Ltd., Chengdu 610041, China 

2.Southwest Jiaotong University, Chengdu 611756, China 

3.Shenzhen Metro Group Co., Ltd., Shenzhen 518026, China)

Abstract: In response to the problem of multiple types, complex internal structures, and diverse operating conditions of traditional rail transit transformers, which makes fault diagnosis more difficult, a deep learning fusion of linear prediction cepstrum coefficient (LPCC) and Mel frequency cepstrum coefficient (MFCC) combined features is proposed for rail transit transformer fault diagnosis. Firstly, the wavelet threshold denoising method is used to preprocess the noisy signal; Then, LPCC and MFCC features of the noise signal are extracted separately, and combined to form feature vectors; Finally, the combined feature vectors are input into the CNN-LSTM model based on deep learning to achieve fault diagnosis of rail transit transformers. The experimental results show that the proposed LPCC-MFCC combination feature and CNN-LSTM model have an accuracy of 99.48% for rail transit transformer fault diagnosis, and the accuracy, recall rate, and F1 score all reach 99.59%.

Keywords: rail transit transformer; fault diagnosis; noise signa analysis; feature extraction; deep learning