School of Electrical and Information Engineering, Anhui University of Science and Technology
School of Computer Science and Engineering, Anhui University of Science and Technology
China Electronics Corporation

针对传统基于音频信号的电动机故障诊断方法获取电动机音频信号特征信息不足和故障诊断精度不高的问题，提出了一种基于优化的变分模态分解（VMD）和卷积神经网络CNN-双向长短期记忆（BiLSTM）的矿井提升电动机故障诊断方法。针对模态混叠和端点效应问题，采用鲸鱼算法（WOA）优化的VMD对电动机音频信号进行分解，将电动机音频信号分解为K个本征模态分量（IMF），经Pearson相关系数筛选后，提取主IMF分量的13维静态MFCC特征参数，为了获取信号的动态特征，提取13维静态MFCC的一阶差分和二阶差分系数，构成39维特征向量，从而把动静态特征结合起来，提高故障诊断性能。为了提高故障诊断精度，在CNN中引入BiLSTM层，CNN在空间维度上提取音频信号的局部特征，BiLSTM在时间维度上保留音频信号的双向时间序列信息，捕获音频信号长距离依赖关系，从而最大程度保留全局和局部特征。实验结果表明：(1) VMD分解的每个IMF分量都具有独立的中心频率且分布均匀，在频域上表现出稀疏性的特点，能够有效避免模态混叠问题；在IMF求解中，VMD分解通过镜像延拓的方式避免了经验模态分解（EMD）和集合经验模态分解（EEMD）中出现的端点效应问题。(2)基于13维静态MFCC特征的故障诊断准确率为97.5%，基于39维动静态MFCC特征的故障诊断准确率比基于13维静态MFCC特征的故障诊断准确率提高了1.11%。(3)基于CNN-BiLSTM诊断模型的准确率达到98.61%，与目前通用诊断模型CNN, BiLSTM和CNN-LSTM相比，准确率分别提高5.83%,4.17%和3.89%。

The traditional motor fault diagnosis method based on the audio signal is insufficient to obtain the feature information of the motor audio signal and the fault diagnosis precision is not high. In order to solve the above problems, a mine motor fault diagnosis method based on optimized variational mode decomposition (VMD) and convolutional neural network CNN bidirectional long short-term memory (BiLSTM) is proposed. The whale algorithm (WOA) optimized VMD is used to decompose the motor audio signal to address the issues of modal aliasing and endpoint effects. The motor audio signal is decomposed into K intrinsic mode functions (IMF). After Pearson correlation coefficient screening, the 13-dimensional static MFCC feature parameters of the main IMF component are extracted. In order to obtain the dynamic features of the signal, the first and second-order difference coefficients of the 13-dimensional static MFCC are extracted to form a 39-dimensional feature vector. By combining dynamic and static features, the performance of fault diagnosis can be improved. In order to improve the precision of fault diagnosis, a BiLSTM layer is introduced into the CNN. The CNN extracts local features of the audio signal in the spatial dimension. The BiLSTM preserves bidirectional time series information of the audio signal in the temporal dimension. It captures long-distance dependencies of the audio signal, thereby maximizing the preservation of global and local features. The experimental results show the following points. ① Each IMF component of VMD decomposition has an independent center frequency and uniform distribution, and exhibits sparsity in the frequency domain. It can effectively avoid modal aliasing problems. In IMF solving, VMD decomposition avoids endpoint effects in empirical mode decomposition (EMD) and ensemble empirical mode decomposition (EEMD) through mirror extension. ② The fault diagnosis accuracy based on 13-dimensional static MFCC features is 97.5%. The fault diagnosis accuracy based on 39-dimensional dynamic and static MFCC features is 1.11% higher than that based on 13-dimensional static MFCC features. ③ The accuracy of the diagnostic model based on CNN-BiLSTM reaches 98.61%, which is 5.83%, 4.17%, and 3.89% higher than the current universal diagnostic models CNN, BiLSTM, and CNN-LSTM, respectively.

国家自然科学基金项目（51874010）；淮南市科技计划项目（2021A243）；

李敬兆,何娜,张金伟等.基于VMD和CNN-BiLSTM的矿井提升电动机故障诊断方法[J].工矿自动化,2023,49(07):49-59.DOI:10.13272/j.issn.1671-251x.2022120065.

LI Jingzhao, HE Na, ZHANG Jinwei, et al. Fault diagnosis method for mine hoisting motor based on VMD and CNN-BiLSTM[J]. Journal of Mine Automation，2023，49(7)：49-59.