基于频谱感知音频去噪的无监督机器异常声音检测  

Unsupervised Machine Abnormal Sound Detection Based on Spectrum Sensing Audio Denoising

作　　者：仇睿 张晨旭 姚瑶 李圣辰 邵曦[1] QIU Rui;ZHANG Chenxu;YAO Yao;LI Shengchen;SHAO Xi(School of Telecommunications and Information Engineering,Nanjing University of Posts and Telecommunications,Nanjing,Jiangsu 210003,China;School of Advanced Engineering,Xi'an Jiaotong-Liverpool University,Suzhou,Jiangsu 215123,China)

机构地区：[1]南京邮电大学通信与信息工程学院,江苏南京210003 [2]西交利物浦大学先进工程学院,江苏苏州215123

出　　处：《复旦学报（自然科学版）》2022年第5期513-519,共7页Journal of Fudan University：Natural Science

基　　金：国家自然科学基金(61936005,61872199,61872424);国家科技创新2030——“新一代人工智能”重大项目(2020AAA0106200)。

摘　　要：在工业自动化生产中,通过声音监测来判断机器运行状态是否正常是一种有效的方法。针对机器运行状态正常变化引起的误判和现实生产环境中存在大量的背景噪声干扰监测的问题,提出了一种基于频谱感知音频去噪的无监督机器异常声音检测方法。首先,利用开源音频集训练一个基于频谱感知的去噪系统,将含噪音频变换到频域后,在频域上感知噪声谱的统计特性并修改频谱形成增强谱,再转换回时域输出去噪音频。音频特征选择对数Mel谱作为特征,之后利用提供的开发集信息训练一个基于深度可分离卷积和倒残差结构的分类器并对音频的每帧计算分类预测值,然后对其求平均负对数计算音频异常分数和确定异常阈值,通过与异常阈值的比较进行异常检测。DCASE Challenge 2021 Task2数据集上的实验结果表明,进行去噪预处理的异常检测系统的检测性能与基线系统相比有所提升。In industrial automation production,it is a common and effective method to judge whether the machine is operating normally through sound monitoring.Aiming at the misjudgement caused by the normal changes of the machine's operating status and the problem of a large number of background noise interference to monitoring in the actual production environment,an unsupervised abnormal sound detection method based on spectrum sensing audio denoising is proposed.First,the open source audio set is used to train a denoising system based on spectrum sensing.After the noise-containing audio is transformed into the frequency domain,the statistical characteristics of the noise spectrum are sensed in the frequency domain and the spectrum is modified to form an enhanced spectrum,then converted back to the time domain output denoised audio.The Log-Mel spectrum is selected as the audio feature,then the provided development set information is used to train a classifier based on the depth separable convolution and inverse residual structure,and the classification prediction value is calculated for each frame of the audio,then the average negative logarithm of its audio anomaly score to determine anomaly threshold,and the anomaly detection is performed by comparing with the anomaly threshold.The experimental results on the DCASE2021 Challenge Task2 data set show that the detection performance of the anomaly detection system that performs denoising preprocessing is improved compared with the baseline system.

关 键 词：无监督 音频去噪 异常检测 深度学习 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程] TP277[自动化与计算机技术—控制科学与工程]