基于深度学习模型的电力变压器故障声音诊断方法研究

Study on the fault diagnosis method of power transformer by sound signals based on deep learning model

利用声音信号对电力变压器进行故障诊断是一种不停机的设备维护方法,可以诊断变压器故障类型或预测故障产生的时间。声音诊断相对于其他诊断方式有许多优点,但是传统的声音自动诊断方法识别率不够理想。为了探索深度学习技术在声音故障诊断方面的可行性,本文采集了真实变压器在正常状态、老化和放电两种故障运行状态下发出的声音信号,对信号分别进行了声谱图转换和梅尔对数谱图的转换,输入一种高效轻量级卷积神经网络--Mobile Net深度学习模型中开展了训练。训练结果表明,将卷积神经网络应用在变压器故障声音诊断上能够得到较高的准确率,尤其是采用梅尔对数谱图对三种状态下识别准确率均能达到99%以上,而采用声谱图进行训练对放电类型的故障识别率较高,老化故障识别率不够理想。

Fault diagnosis of power transformers by sound is a nonstop equipment maintenance method,which can diagnose the type of transformer faults or predict the time of fault will happen. Fault diagnosis by sound has many advantages over other methods,but the recognition rate of traditional automatic sound diagnosis is not high enough. In order to explore the feasibility of deep learning technology in fault diagnosis by sound,the sound signals of real transformer running in normal state and two kinds of failure state including aging state and discharge state were collected. The signals were transformed respectively to the spectrogram and Mel Logarithmic spectrum. All the spectrums images were imported in a highly efficient lightweight Convolutional Neural Networks named Mobile Net for training. The results show that the Convolutional Neural Network training can get very high recognition accuracy for all the three kinds of mode. The recognition accuracy rate can be over 99% by Mel Logarithmic spectrum image,the accuracy is also pretty good with spectrogram for discharge mode,but not satisfied for the aging mode.