基于神经网络的变压器故障声识别研究

Research on Transformer Fault Sound Recognition Based on Neural Network

作为新兴的变压器故障识别技术,声音识别对于变压器故障的识别尤为重要,然而在声音识别过程中,常常受到环境中其他声音的影响而降低识别的准确率,基于此,文章提出一种基于MFCC声音特征提取以及人工神经网络(ANN)模型相结合的变压器故障声音识别的方法,为了提高模型的训练精度,文章对比分析了Levenberg-Marquardt算法、Bayesian Regularization算法以及Scaled Conjugate Gradient算法的收敛性与准确性,选取收敛速度快、误差较小的Levenberg-Marquardt算法来实现ANN模型的误差反向传播并完成故障诊断的验证,验证结果表明,文章所采用的模型对于100个验证样本数据的预测准确率为92%,最终证实模型能够很好的应用于变压器故障的声识别。

As a new transformer fault identification technology,sound recognition is significant for transformer fault identification.However,in the process of sound recognition,the accuracy of transformer fault recognition is often reduced by the influence of other sounds in the environment.This paper proposes a sound recognition method for transformer fault based on MFCC sound feature extraction and the ANN model.To improve the training accuracy of the model,the convergence and accuracy of the Levenberg-Marquardt algorithm,the Bayesian Regularization algorithm,and the Scaled Conjugate Gradient algorithm are compared and analyzed.The Levenberg-Marquardt algorithm with fast convergence speed and small error is selected to realize the error backpropagation of the ANN model and verification of fault diagnosis.The verification results show that the model's prediction accuracy is 92%for the data of 100 verification samples,which finally proves that the model can be well applied to the acoustic identification of transformer faults.