一种用于机器声音异常检测的ARViTrans方法
2025-08-03陈龙黄小伟郭法滨陆亚师

doi:10.19734/j.issn.1001-3695.2024.10.0365
ARViTrans method for machine sound anomaly detection
ChenLonga,b,Guo Fabina,bt,Huang Xiaoweia,b,Lu Yashia,b (a.Schoolofifcaellgeeamp;igtdslqpntStatesEaoFultredicoamp;tee University,Hefei ,China)
Abstract:Inordertosolvetheproblems thattheexisting machine soundanomalydetection methodsonly focus onthe single featuresof the time,frequencyorchanneldimensions,ignoringthemutualconnectionbetweenthespectralfeaturesandthe timeseriesinformation,andtheinitial featurelossleads toinaccurate fitingof thesampledatadistribution,thuscausing a highanomaly mised detectionrateand1 alarmrate,this paper proposedARViTrans,amachine sound anomalydetection methodthatintegratedatentionmechanismsandskipconnections.Firstlythispaperproposedathree-dimensionalicient coordinateatentionmechanismstocollaborativelycapturethetimedomain,frequencydomainandchanneldimensionfeatures through thedecouplingoperationofthefeaturespace.Secondly,itusedMobileViTasthebackbonenetworkanddesignedthe RES-MoViT module toreplacethe original MobileViT module.Skipconnections captured the information between the input andoutputand beterfitthesampledata distribution.Thegradientrefluxreduced therepeatedlearningof similarfeatureparametersandimprovedtheparameterutilizationeficiency.Finally,itcomparedtheexperimentalresultsontheMMdataset with the AE and MobileNetV2 of the DCASE Task2 baseline system. The AUC improves by 10.14% and 10.26% ,respectively.The pAUC improves by 13.40% and 6.50% ,respectively. The experimental results indicate that the proposed method caneffectivelycapturethemutualconnectionbetween featuresofdiferent dimensions while maintainingalowmodelcomplexity,improve the accuracy of anomaly detection and reduce the 1 alarm rate.
Key words:anomaly detection;MobileViT;attention mechanism;residual connection;unsupervise
0 引言
近年来,工业机器设备的状态监测在工厂自动化领域中发挥着至关重要的作用[1]。异常事件不仅会导致生产效率降低、产品质量下降,还可能危及工人的生命安全。因此,实时监控机器的运行状态,确保工厂生产的正常运行成为了必不可少的措施。声音监测由于成本低、安装简便且能覆盖视频监控的盲区,在监控领域得到了广泛应用。目前,异常声音检测技术已被广泛用于工业设备监控[3]、生物医学安全[4]和海上设备监控[5]等多个领域。
在机器异常声音检测的研究中,通常对机器声音的原始信号进行预处理,再进行时域、频域或时频分析,最后采用基于分类或重构的无监督学习方法来检测异常。然而,机器声音异常检测面临多个挑战:一方面,声音信号包含丰富的时序和频谱信息,而现有检测方法往往仅关注时域、频域或通道维度的单一特征,未能有效捕捉声谱信息与时间序列间的复杂关系,从而导致对异常事件的识别能力不足。例如,Wang等人[围绕特征融合和信息损失问题,提出了一种结合高效通道注意机制的轻量化MobileNetV3模型,以增强对机器异常声音的检测性能。Bai等人[7]提出了一种双路Transformer网络架构,通过交替在时间和频率维度上建模时频谱中的声学特征来有效地进行异常检测。Almudevar等人[8]将时频谱的每个时间帧作为ViT的输入,开发了一种基于ViT的嵌入提取器,以增强特征提取的全面性。另一方面,基于重构的方法通过重构样本并使用重构误差作为异常检测分数,由于训练过程中不涉及异常声音,若训练得到的特征也适用于异常声音,这种方法的有效性可能会受限[2]。……
