噪声背景下梅尔频率倒谱系数与多注意力网络在电机故障诊断中的应用  

Motor fault diagnosis using mel frequency cepstral coefficients and multi-attention network under strong noise background

作　　者：宋恩哲[1] 朱仁杰 靖海国 姚崇[1] 柯赟 SONG Enzhe;ZHU Renjie;JING Haiguo;YAO Chong;KE Yun(College of Power and Energy Engineering,Harbin Engineering University,Harbin 150001,China;CSSC Marine Power Co.,Ltd.,Zhenjiang 212001,China)

机构地区：[1]哈尔滨工程大学动力与能源工程学院,黑龙江哈尔滨150001 [2]中船动力镇江有限公司,江苏镇江212002

出　　处：《哈尔滨工程大学学报》2025年第3期475-485,共11页Journal of Harbin Engineering University

基　　金：中央高校基本科研业务费专项资金项目(3072022QBZ0301,3072022JC2704).

摘　　要：针对电机实际工作过程中存在噪声干扰导致故障诊断精度下降的问题,本文提出了一种基于梅尔频率倒谱系数动态特征与多注意力融合卷积神经网络的故障诊断方法。通过梅尔频率倒谱系数动态特征提取噪声信号中的低频信息,并结合卷积注意力模块的自适应调节能力及多特征融合策略进一步减少噪声对故障诊断的干扰。通过电机台架数据验证了该方法在噪声条件下诊断的可行性,然而该方法受梅尔频率倒谱系数参数与网络结构的直接影响,因此具体分析了不同参数条件对抗噪性能的影响。实验结果表明:在信噪比-10 dB噪声背景下,梅尔频率倒谱系数动态特征与多注意力融合卷积神经网络相结合的故障诊断方法仍保持90%以上的诊断精度。A fault diagnosis method leveraging the dynamic features of Mel frequency cepstral coefficients and a multi-attentional fusion convolutional neural network is proposed to address the issue of noise interference in the operation of electric machines.Noise interferences can degrade fault diagnosis accuracy.Thus,the dynamic features of Mel frequency cepstral coefficients are used to extract the low-frequency information from the noisy signals.Moreover,the adaptive adjustment capability of the convolutional attention module and the multi-feature fusion strategy are combined to further reduce the interference of noise on the fault diagnosis.The feasibility of the method under noisy conditions is verified using motor bench data.However,the method is directly affected by the parameters of the Mel frequency cepstral coefficients and the network structure.Therefore,the impact of different parameter conditions on the noise immunity performance is specifically analyzed.The experimental results demonstrate that under the noise background of signal to noise ratio-10 dB,the combination of the dynamic features of the Mel frequency cepstral coefficients and the multi-attentional fusion convolutional neural network offers superior noise immunity to other fault diagnosis methods over 90%.

关 键 词：电机 故障诊断 噪声环境 梅尔频率倒谱系数 卷积神经网络 多尺度 卷积注意力模块 特征融合 

分 类 号：U672[交通运输工程—船舶及航道工程]