与传统的振动信号故障检测方法相比，声音检测具有非接触测量，设备简单，信号易获取等显著优势。而以深度学习理论为基础的智能故障检测技术还能够降低对信号处理专业知识与专家经验的依赖，极大的降低了故障检测技术的门槛，扩大了实际应用场景。传统检测方法多采用小波分解方法，支持向量机，人工神经网络ANN等方法，这些方法将声音信号故障检测问题建模为分类问题，在信噪比较高和样本量充分的条件下检测精度较好，实际采集的声音信号一般含有复杂的噪声干扰，对检测精度影响较大，且声音信号的特征提取和融合方法有待改进。

针对现有方法存在的缺点，本论文提出基于改进的门控小波卷积网络去噪模型、基于谱-时信息融合的特征提取方法及基于改进MobileFaceNet的机械设备声音信号异常检测方法。实验采用MIMII和ToyADMOS数据集进行研究，并基于实验室轴承综合故障诊断模拟实验台完成相关方法验证。主要内容如下：

（1）在声音信号降噪阶段，应用了集小波卷积，动态硬阈值，门控注意力机制等模块为一体的门控小波卷积网络去噪模型，使得去噪后声音信号的信噪比指标提升明显。其中添加信噪比为10dB，6dB，3dB和0dB时，信噪比提升分别为6.9dB，5.5dB，4.1dB和4.8dB。

（2）在特征提取阶段，提出谱-时信息融合的机械设备声音信号特征提取方法，结合对数梅尔频谱的频谱特征和基于一维卷积神经网络提取的时域特征，弥补了单独使用梅尔频谱特征可能会损失高频分量中的有用特征的不足，并利用谱-时联合注意力模块自适应的聚合频谱和时域特征，获得了更好的特征表达。

（3）在故障检测阶段，通过改进MobileFaceNet的机械设备声音信号异常检测方法，显著提高了异常检测任务的性能，特别是平均AUC比基线最好的Glow模型提高了4.25%，pAUC提高了4.67%。在检测的鲁棒性方面，平均mAUC超过最佳对比方法15.65%。

（4）在HZXT-008型转子轴承综合故障诊断模拟实验台实测数据上进行验证，故障检测率为95.72%，AUC为0.96，证明本文提出的模型是有效的，在故障检测领域具有一定的实际意义。

关 键 词：声音信号，故障检测，去噪，深度学习，特征融合，小波卷积，门控机制

研究类型：应用研究

Compared with traditional vibration signal fault detection methods, sound detection has significant advantages such as non-contact measurement, simple equipment, and easy signal acquisition. The intelligent fault detection technology based on deep learning theory can also reduce the dependence on professional knowledge and expert experience in signal processing, greatly reducing the threshold of fault detection technology and expanding practical application scenarios. Traditional detection methods often use wavelet decomposition methods, support vector machines, artificial neural networks, and other methods. These methods model the problem of sound signal fault detection as a classification problem, and have good detection accuracy under conditions of high signal-to-noise ratio and sufficient sample size. The actual collected sound signals generally contain complex noise interference, which has a significant impact on detection accuracy. Moreover, the feature extraction and fusion methods of sound signals need to be improved.

In response to the shortcomings of existing models, this paper proposes a denoising model based on an improved gated wavelet convolutional network, a feature extraction method based on spectral time information fusion, and an anomaly detection method for mechanical equipment sound signals based on an improved MobileFaceNet. The experiment was conducted using MIMII and ToyADMOS datasets, and the relevant method validation was completed based on a laboratory bearing comprehensive fault diagnosis simulation test bench.The main content is as follows:

(1) In the denoising stage of sound signals, a gated wavelet convolutional network denoising model based on modules such as wavelet convolution, dynamic hard threshold, and gated attention mechanism was applied, which significantly improved the signal-to-noise ratio of the denoised sound signal. When adding signal-to-noise ratios of 10dB, 6dB, 3dB, and 0dB, the signal-to-noise ratio increases by 6.9dB, 5.5dB, 4.1dB, and 4.8dB, respectively.

(2) In the feature extraction stage, by combining the spectral features of the logarithmic Mel spectrum with the time-domain features extracted based on one-dimensional convolutional neural networks, the shortcomings of using Mel spectrum features alone may result in the loss of useful features in high-frequency components are compensated for. The spectral temporal joint attention module is used to adaptively aggregate the spectral and time-domain features, achieving better feature expression.

(3) In the fault detection phase, the performance of the anomaly detection task was significantly improved by improving MobileFaceNet's mechanical equipment sound signal anomaly detection method, especially with an average AUC increase of 4.25% and a pAUC increase of 4.67% compared to the baseline best model Glow. In terms of detection robustness, the average mAUC exceeds the best comparison method by 15.65%.

(4) The validation was conducted on the HZXT-008 rotor bearing comprehensive fault diagnosis simulation experimental platform, and the fault detection rate was 95.72%, with an AUC of 0.96. This proves that the model proposed in this paper is effective and has certain practical significance in the field of fault detection.

Key words: Sound signal，Fault diagnosis, Sound noise reduction, Deep Learning, Feature fusion, Wavelet convolution,Gated attention mechanism   

Thesis    : Application Researh

[1]谭逸. 基于声音的泵故障检测与定位方法研究[D].北京化工大学, 2022.

[2]姚勇. 基于声学信号的典型机械设备故障智能检测方法研究[D]. 贵州大学, 2019.

[3]郑近德,应万明,潘海洋,等.基于改进全息希尔伯特谱分析的旋转机械故障检测方法[J].机械工程学报, 2023, 59(01):162-174.

[4]曹增欢,王国锋,户满堂,等.一种变分模态分解和频域积分相结合的故障特征提取方法[J].机械科学与技术, 2022, 41(10):1496-1502.

[5]陈礼顺,张晗,陈雪峰,等.基于低秩稀疏分解算法的航空锥齿轮故障检测[J].振动与冲击, 2020, 39(12):103-112.

[6]李继猛,王慧,李铭,等.基于改进的自适应无参经验小波变换的滚动轴承故障检测[J].计量学报, 2020, 41(06):710-716.

[7]聂周,李迎春,王森,等.小波降噪与快速包络谱峭度相结合的轴承组件故障检测技术[J].飞控与探测, 2021,4(02):96-102.

[8]Hu Can, Xing Futang, Pan Shuhan, et al. Fault Diagnosis of Rolling Bearings Based on Variational Mode Decomposition and Genetic Algorithm-Optimized Wavelet Threshold Denoising[J]. Machines, 2022, 10(8): 649.

[9]Wang Junxia, Zhan Chang, Li Sanping, et al. Adaptive variational mode decomposition based on Archimedes optimization algorithm and its application to bearing fault diagnosis[J]. Measurement, 2022, 191: 110798.

[10]Chen Wuge, Li Jun, Wang Qian, et al. Fault feature extraction and diagnosis of rolling bearings based on wavelet thresholding denoising with CEEMDAN energy entropy and PSO-LSSVM[J]. Measurement, 2021, 172: 108901.

[11]李巍华,单外平,曾雪琼.基于深度信念网络的轴承故障分类识别[J].振动工程学报, 2016, 29(02):340-347.

[12]Ince Turker, Kiranyaz Serkan, Eren Levent, et al. Real-time motor fault detection by 1-D convolutional neural networks[J]. IEEE Transactions on Industrial Electronics, 2016, 63(11): 7067-7075.

[13]Zhao Minghang, Zhong Shisheng, Fu Xuyun, et al. Deep residual shrinkage networks for fault diagnosis[J]. IEEE Transactions on Industrial Informatics, 2019, 16(7): 4681-4690.

[14]Tong Yizhi, Wu Ping, He Jiajun, et al. Bearing fault diagnosis by combining a deep residual shrinkage network and bidirectional LSTM[J]. Measurement Science and Technology, 2021, 33(3): 034001.

[15]Liang Haopeng, Cao Jie, Zhao Xiaoqiang. Multi-scale dynamic adaptive residual network for fault diagnosis[J]. Measurement, 2022, 188: 110397.

[16]Wang Huan, Liu Zhiliang,Peng Dandan,et al.Interpretable convolutional neural network with multilayer wavelet for Noise-Robust Machinery fault diagnosis[J]. Mechanical Systems and Signal Processing,2023,195: 110314.

[17]Weng Chaoyang, Lu Baochun, Gu Qian. A multi-scale kernel-based network with improved attention mechanism for rotating machinery fault diagnosis under noisy environments[J]. Measurement Science and Technology, 2022, 33(5): 055108.

[18]Li Yibing, Zou Wei, Jiang Li. Fault diagnosis of rotating machinery based on combination of Wasserstein generative adversarial networks and long short term memory fully convolutional network[J]. Measurement, 2022, 191: 110826.

[19]Li Tianfu, Zhao Zhibin, Sun Chuang, et al. WaveletKernelNet: An interpretable deep neural network for industrial intelligent diagnosis[J]. IEEE Transactions on Systems, Man, and Cybernetics: Systems, 2021, 52(4): 2302-2312.

[20]Jing Yuan, Cao Shuwei, Ren Gangxing, et al. LW-Net: an interpretable network with smart lifting wavelet kernel for mechanical feature extraction and fault diagnosis. Neural Computing and Applications, 2022, 34(18): 15661-15672.

[21]Wang Dong, Chen Yikai, Shen Changqing, et al. Fully interpretable neural network for locating resonance frequency bands for machine condition monitoring[J]. Mechanical Systems and Signal Processing, 2022, 168: 108673.

[22]Li Guoqiang,Wu Jun,Deng Chao,et al. Self-supervised learning for intelligent fault diagnosis of rotating machinery with limited labeled data[J]. Applied Acoustics,2022,191.

[23]Zhao Xuezhi, Ye Bangyan. Feature frequency extraction algorithm based on the singular value decomposition with changed matrix size and its application in fault diagnosis[J]. Journal of Sound and Vibration, 2022, 526: 116848.

[24]Li Jintao, Chen Zhaobo, Li Shengbo. Selection of the number of effective singular values for noise reduction[J]. Mechanical Systems and Signal Processing, 2023, 191: 110175.

[25]樊高瞻,周俊,朱昆莉.基于改进形态-小波阈值降噪的轴承复合故障声学检测[J].振动与冲击, 2020, 39(12):221-226+288.

[26]李云飞,苏文胜,刘彬彬,等.基于多小波变换和奇异值分解的声发射信号降噪方法[J].中国特种设备安全, 2022, 38(03):17-20+45.

[27]刘志坚,赵浩益,刘航,等.基于改进阈值估计和改进阈值函数的局部放电信号降噪方法[J].振动与冲击, 2022, 41(22):1-13.

[28]叶昱清,丁军航,叶宁祁,等.基于EMD-NLMS的船舶发动机声信号降噪处理方法[J/OL].电子设计工程:1-7[2023-04-24].

[29]于金涛,赵树延,王祁.基于经验模态分解和小波变换声发射信号去噪[J].哈尔滨工业大学学报, 2011, 43(10):88-92.

[30]Sun Hongchun, Cao Xu, Wang Changdong, et al. An interpretable anti-noise network for rolling bearing fault diagnosis based on FSWT[J]. Measurement, 2022, 190: 110698.

[31]Shang Zuogang, Zhao Zhibin, Yan Ruqiang. Denoising Fault-Aware Wavelet Network: A Signal Processing Informed Neural Network for Fault Diagnosis[J]. Chinese Journal of Mechanical Engineering, 2023, 36(1): 9.

[32]Ye Zhuang, Yu Jianbo. Deep morphological convolutional network for feature learning of vibration signals and its applications to gearbox fault diagnosis[J]. Mechanical Systems and Signal Processing, 2021, 161: 107984..

[33]夏小飞,易林,饶夏锦,等.基于声学指纹分析的高压断路器机械故障诊断方法[J].高压电器,2021,57(10):66-76.

[34]韦娟,岳凤丽,仇鹏,等.基于EEMD的异常声音多类识别算法[J]. 华中科技大学学报: 自然科学版, 2018, 46(7): 117-121.

[35]孙玉伟,罗林根,陈敬德,等.基于声音特征与改进稀疏表示分类的断路器机械故障诊断方法[J]. 电网技术, 2022, 46(3): 1214-1222.

[36]王瑞,刘宾,周天润,等.基于协同表示的声振传感器网络车辆分类识别[J]. 上海交通大学学报, 2018 (1): 103-110.

[37]王林,扈海泽,方梦鸽. MFCC-小波神经网络电气主设备音频监控研究[J]. 电力与能源, 2019, 40(6): 660-663.

[38]赵凯,董明明,刘锋,等.基于声信号的履带机器人地面分类试验研究[J]. 北京理工大学学报, 2018, 38(9): 912-916.

[39]殷鹰,谢罗峰,黄泰博.基于深度学习的磁瓦内部缺陷声振检测方法[J]. 中国测试, 2020, 46(3): 32-38.

[40]钟鸣拓,蔡文郁.基于特征融合的海洋哺乳动物声音识别[J]. 电子科技, 2019, 32(5): 32-37.

[41]Yao Yong, Wang Honglei, Li Shaobo, et al. End-to-end convolutional neural network model for gear fault diagnosis based on sound signals[J]. Applied Sciences, 2018, 8(9): 1584.

[42]王峰,张海涛.基于正则化稀疏滤波的轴承声信号故障检测方法[J].噪声与振动控制, 2022, 42(02):114-118.

[43]刘艳杰,陈炳发,丁力平.基于声学特征的微电机故障检测方法[J].机械制造与自动化, 2022, 51(02):190-194.

[44]时培明,肖立峰,许学方,等.双通道小波核-卷积神经网络轧机设备轴承检测方法[J/OL].机械科学与技术:1-11[2023-05-01].

[45]孙启涛,罗智孙,梁好,等.基于多尺度特征提取的风机音频信号故障检测方法研究[J].机电工程, 2023, 40(01):39-46.

[46]贺志军,李军霞,张伟,等.基于MFCC特征和GWO-SVM的托辊故障检测[J].机床与液压, 2022, 50(15):188-193.

[47]万智勇. 基于自编码器的机械声音异常检测方法研究[D].广东工业大学, 2022.

[48]刘少康,武英杰,安伟伦,等.基于声音信号和改进MS-LMD的风电齿轮箱故障检测[J].振动与冲击,2021,40(11):230-239+251.

[49]申博文,王华庆,唐刚,等.基于MCKD与CEEMDAN的声信号故障特征提取方法[J].复旦学报(自然科学版), 2019, 58(03):385-392+400.

[50]Purohit Harsh, Tanabe Ryo, Ichige Kenji, et al. MIMII Dataset: Sound dataset for malfunctioning industrial machine investigation and inspection[J]. arXiv preprint arXiv:1909.09347, 2019.

[51]Koizumi Yuma, Saito Shoichiro, Uematsu Hisashi, et al. ToyADMOS: A dataset of miniature-machine operating sounds for anomalous sound detection[C]//2019 IEEE Workshop on Applications of Signal Processing to Audio and Acoustics (WASPAA). IEEE, 2019: 313-317.

[52]覃坚,费太勇,曲智国,等.基于优化小波阈值的轴承振动信号降噪算法[J].现代防御技术,2023,51(02):141-147.

[53]Shan Shuaijie, Liu Jianbao, Wu Shuguang, et al. A motor bearing fault voiceprint recognition method based on Mel-CNN model[J]. Measurement, 2023, 207: 112408.

[54]周玉蓉,张巧灵,于广增,等.基于声信号的工业设备故障诊断研究综述[J].计算机工程与应用,2023,59(07):51-63.

[55]康涛,王书,陈志刚,等.基于1D-CNN的滚动轴承端到端故障诊断方法[J].设备管理与维修,2021(13):139-141.

[56]Michau Gabriel, Frusque Gaetan, Fink Olga. Fully learnable deep wavelet transform for unsupervised monitoring of high-frequency time series[J]. Proceedings of the National Academy of Sciences, 2022, 119(8): e2106598119.

[57]冯贤洋,何荇兮,符礼丹,等.基于注意力机制的双向门控循环单元网络齿轮故障识别系统[J].汽车工程学报,2023,13(01):111-117.

[58]易清明,吕人毅,石敏,等.融合多尺度空洞卷积与反卷积的轻量化目标检测[J].华南理工大学学报(自然科学版),2022,50(12):41-48.

[59]Giri Ritwik, Tenneti Srikanth V, Cheng Fangzhou, et al. Self-supervised classification for detecting anomalous sounds[J]. 2020:46-50.

[60]Koizumi Yuta, Saito Shoichiro, Uematsu Hisashi, et al. Unsupervised detection of anomalous sound based on deep learning and the neyman–pearson lemma[J]. IEEE/ACM Transactions on Audio, Speech, and Language Processing, 2018, 27(1): 212-224.

[61]Dohi Kota, Endo Takashi, Purohit Harsh, et al. Flow-based self-supervised density estimation for anomalous sound detection[C]//ICASSP 2021-2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP). IEEE, 2021: 336-340.

[62]Kingma Diederik P, Dhariwal Prafulla. Glow: Generative flow with invertible 1x1 convolutions[J]. Advances in neural information processing systems, 2018, 31.

[63]Papamakarios George, Pavlakou Theo, Murray I. Masked autoregressive flow for density estimation[J]. Advances in neural information processing systems, 2017, 30.

[64]王伟,万晓刚.结合注意力机制和特征融合的小目标检测方法[J].西安工程大学学报,2022,36(06):115-123.

[65]黄雅静,廖爱华,丁亚琦,等.多尺度CNN结合自注意力特征融合机制的轴承故障诊断方法[J].智能计算机与应用,2022,12(09):37-44.

[66]李星燃,张立言,姚树婧.结合特征融合和注意力机制的微表情识别方法[J].计算机科学,2022,49(02):4-11.

[67]陈向民,韩梦茹,舒文伊,等.基于VMD与多尺度一维卷积神经网络的故障诊断方法[J].现代电子技术,2023,46(09):103-109.

[68]王哲豪,范丽丽,何前.基于MobileNet V2和迁移学习的番茄病害识别[J].江苏农业科学,2023,51(09):215-221.

[69]Chen Sheng, Liu Yang, Gao Xiang, et al. Mobilefacenets: Efficient cnns for accurate real-time face verification on mobile devices[C]//Biometric Recognition: 13th Chinese Conference, CCBR 2018, Urumqi, China, August 11-12, 2018, Proceedings 13. Springer International Publishing, 2018: 428-438.

[70]Deng Jiankang, Guo Jia, Xue Niannan, et al. Arcface: Additive angular margin loss for deep face recognition[C]//Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2019: 4690-4699.