融合注意力机制卷积神经网络的扬声器异常声分类  

Diagnosis of abnormal sound in loudspeakers by integratedattention mechanism convolutional neural network

作　　者：周静雷[1] 王晓明 李丽敏[1] ZHOU Jinglei;WANG Xiaoming;LI Limin(School of Electronics and Information,Xi’an Polytechnic University,Xi’an 710048,China)

机构地区：[1]西安工程大学电子信息学院,陕西西安710048

出　　处：《西安工程大学学报》2024年第2期101-108,共8页Journal of Xi’an Polytechnic University

基　　金：国家自然科学基金(62203344);陕西省技术创新引导专项(2020CGXNX-009);陕西省自然科学基础研究计划(2022JM-322)。

摘　　要：针对扬声器异常声非线性、非平稳且易受外部噪声干扰,以及因特征冗余而导致扬声器异常声识别率偏低的问题,提出一种基于变分模态分解(variational mode decomposition, VMD)和一维卷积循环注意力网络(1DCNN-BiLSTM-Attention)相结合的扬声器异常声分类方法。首先,采集不同类型异常声信号,采用VMD对异常声信号进行分解并提取扬声器异常声特征,构建标签化的初始数据;其次,将特征数据输入至1DCNN-BiLSTM网络中进行初始化特征提取,利用注意力机制自适应优化网络对异常声特征的学习权重,提升网络对特征鉴别能力,并优化Dropout抑制网络在训练过程中存在的过拟合问题,构成1DCNN-BiLSTM-Attention分类网络;最后,将所提方法应用于扬声器异常声分类中。实验结果表明:该方法可以有效提取到扬声器异常声中的关键特征,平均分类准确率为99.17%,与VGG16、RF和DCNN相比,其准确率分别提高了13.14%、0.56%,12.34%。In response to the problem of the non-linear,non-stationary nature of speaker abnormal sound,as well as their susceptibility to external noise interference,and the low recognition rates,a speaker abnormal sound classification method with variational mode decomposition(VMD)and 1D convolutional recurrent attention network(1DCNN-BiLSTM-Attention)was proposed.Firstly,different types of abnormal sound signals were collected,and VMD was used to decompose the signals and extract the features of speaker abnormal sound,constructing labeled initial data.Secondly,the feature data was input into the 1DCNN-BiLSTM network for initial feature extraction.The attention mechanism was employed to adaptively optimize the network's learning weights for abnormal sound features,enhancing the networks discriminative capability.Additionally,dropout was optimized to suppress overfitting during the training process,resulting in the construction of the 1DCNN-BiLSTM-Attention classification network.Finally,the proposed method was applied to speaker abnormal sound classification.The experimental results demonstrate that this method effectively extracts key features from speaker abnormal sounds,with an average accuracy of 99.17%.Compared to VGG16,RF,and DCNN,the accuracy has been improved by 13.14%,0.56%,and 12.34%respectively.

关 键 词：异常声分类 变分模态分解 卷积神经网络 注意力机制 

分 类 号：TN643[电子电信—电路与系统]