Page 118 - 《应用声学》2021年第4期

第 40 卷 第 4 期 Vol. 40, No. 4
2021 年 7 月 Journal of Applied Acoustics July, 2021

⋄ 研究报告 ⋄

声纹的变压器放电与机械故障诊断研究 ∗

吴国鑫 詹花茂 † 李 敏

(华北电力大学 新能源电力系统国家重点实验室 北京 102206)

摘要：变压器中的一些放电和机械故障会产生异常声，可用于故障检测。据此，该文提出基于可听声的变压器
放电和机械故障诊断方法。针对机械故障声与变压器本体噪声特征相似易混淆的问题，提出改进小波包 -BP
神经网络算法，与传统小波包 -BP 神经网络算法相比声音识别率提高了 5.7%。为提高声音识别系统的泛化性，
提出基于梅尔对数频谱和卷积神经网络的声音识别算法。两种算法相互验证，提高了系统的可靠性。在真实
变压器油箱中模拟了不同类型放电和机械故障。实验结果表明，该文提出的两种方法能成功识别放电声和机
械故障声，声音识别率分别为 99.6% 和 97.57%。
关键词：变压器；可听声；小波包；BP 神经网络；卷积神经网络
中图法分类号: TM411 文献标识码: A 文章编号: 1000-310X(2021)04-0602-09
DOI: 10.11684/j.issn.1000-310X.2021.04.015

Research on transformer discharge and mechanical fault diagnosis
based on voiceprint

WU Guoxin ZHAN Huamao LI Min

(State Key Laboratory of Alternate Electrical Power System with Renewable Energy Sources,
North China Electric Power University, Beijing 102206, China)

Abstract: Some discharges and mechanical faults in the transformer will produce abnormal sounds, which
can be used for fault detection. Accordingly, a transformer fault diagnosis method based on audible sound is
proposed. Aiming at the problem that the sound of mechanical faults is similar to the background noise of
the transformer itself and is easy to be confused, an improved wavelet packet-BP neural network algorithm is
proposed. Compared with the traditional wavelet packet-BP neural network algorithm, the sound recognition
rate is increased by 5.7%. In order to improve the generalization of the sound recognition system, a sound
recognition algorithm based on log-Mel spectrum and convolutional neural network is proposed. The two
algorithms are mutually verified, which improves the reliability of the system. Different types of discharges and
mechanical failures are simulated in the real transformer tank. The test results show that the two methods
proposed in this paper can successfully identify the sounds of discharge and mechanical failure, and the sound
recognition rates are 99.6% and 97.57%, respectively.
Keywords: Transformer; Audible sound; Wavelet packet; BP neural network; Convolutional neural network

2021-02-26 收稿; 2021-04-09 定稿
北京市自然科学基金资助项目 (3202032)
∗
作者简介: 吴国鑫 (1997– ), 女, 黑龙江哈尔滨人, 硕士研究生, 研究方向: 变压器绝缘及状态检测技术。
† 通信作者 E-mail: zhanhm@ncepu.edu.cn