~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~噪声与振动柴油机设计与制造DesignandManufactureofDieselEngine2023年第2期第29卷(总第183期)基于卷积神经网络的发动机齿轮啸叫识别方法邹佳烨(上海新动力汽车科技股份有限公司,上海200438)摘要:传统发动机生产线对发动机齿轮啸叫的识别基于人工判定,即对每台装配该齿轮系的发动机进行试车,人工识别齿轮啸叫问题,但人工识别工作量大、效率较低。因此,提出录制发动机在试车时产生的音频数据,利用短时傅里叶变换对其进行预处理,转换成色谱图,再用二值化和最大类间方差法(OTSU算法)对色谱图进行信息缩减,最后通过建立卷积神经网络的算法模型,智能筛选出有齿轮啸叫的发动机。结果表明:模型筛选啸叫齿轮的准确率为96%左右,该模型可以降低人工识别的误判率和工作量,提高识别效率。关键词:齿轮啸叫;噪声−振动−声振粗糙度(NVH);神经网络;噪声识别TherecognitionmethodofenginegearwhinebasedonconvolutionalneuralnetworkZOUJiaye(ShanghaiNewPowerAutomotiveTechnologyCo.,Ltd.,Shanghai200438,China)Abstract:Therecognitionofenginegearwhineintraditionalengineproductionlineisbasedonla⁃bor,whichinvolvestestingeachengineequippedwiththisgeartraintomanuallyidentifythegearsqueal⁃ingproblem,resultinginheavymanualidentificationworkloadandlowefficiency.Therefore,itispro⁃posedtorecordtheaudiodatageneratedbytheengineduringtestrun,preprocessitusingshort⁃timeFou⁃riertransform,andconvertitintoacolormapimage.ThenthebinarizationandOTSUalgorithmsareusedtoreducetheinformationofthecolormapimage.Finally,byestablishinganalgorithmmodelofcon⁃volutionalneuralnetwork,theenginegearisintelligentlyscreenedforwhine.Theresultsshowthattheaccuracyofthemodelinscreeningoutthewhinegearsis96%.Itcanreducethemisjudgmentrateandworkloadofmanualrecognition,improvetherecognitionefficiency.Keywords:gearwhine;neuralnetwork;noise−vibration−harshness(NVH);noiseidenti⁃ficationDOI:10.3969/j.issn.1671-0614.2023.02.0080前言随着汽车制造技术的发展,人们不仅仅满足于汽车动力的提升,对汽车舒适性的要求也日益增加,因此各大汽车厂商对汽车的噪声−振动−声振粗糙度(NVH)要求也越来越高,投入到NVH的研发费用也日益...