第44卷第3期2023年6月Vol.44No.3June2023内燃机工程ChineseInternalCombustionEngineEngineering基于降噪–残差神经网络的发动机部分失火故障诊断庞皓乾1，张攀1，王文2，王彦军1，邹佳华1，高文志1（1.天津大学内燃机燃烧学国家重点实验室，天津300354；2.中国重汽集团青岛重工有限公司，青岛266111）PartialMisfireFaultDiagnosisofAnEngineBasedonNoiseReduction–ResidualNeuralNetworkPANGHaoqian1，ZHANGPan1，WANGWen2，WANGYanjun1，ZOUJiahua1，GAOWenzhi1（1.StateKeyLaboratoryofEngines，TianjinUniversity，Tianjin300354，China；2.SINOTRUCKQingdaoHeavyIndustryCo.，Ltd.，Qingdao266111，China）Abstract：Afaultdiagnosismethodof“noisereduction–residualneuralnetwork”basedonwaveletthresholddenoisingandresidualneuralnetworkwasproposedforpartialmisfirefaultdiagnosisofenginecylinders.Combiningthenoisereductionanddeeplearningalgorithm，thesignalwasdenoisedbywaveletthresholdandfedintotheresidualneuralnetworkforfaultdiagnosis.Unlikethepreviousresidualnetwork，theshortresidualblockwouldbeutilizedtofurtherpreventnetworkdegradation.Besides，thelargeconvolutionkernelwasalsousedtoexpandtheconvolutionfieldoflongdatainputandimprovetheabilitytoextractfaultcharacteristics.Experimentalresultsshowthatthismethodcannotonlyachievemorethan97%faultdiagnosisaccuracyfortheoperationconditionswithouttraining，butalsoachievehighdiagnosisaccuracyforthenoisysignalswithGaussiannoise.Theperformanceoftheproposedmethodisprovedtobemoresuperiorityandexcellentthanthoseofothermisfirefaultdiagnosisalgorithms.摘要：针对发动机单缸部分失火故障，提出基于小波阈值降噪和残差神经网络的“降噪–残差神经网络”故障诊断方法。通过降噪与深度学习算法相结合，将小波阈值降噪后的振动信号输入到残差神经网络进行故障诊断；使用短残差块进一步防止网络的退化，并利用大卷积核增大长数据输入的卷积视野，提高信号故障特征的提取能力。测试结果证明该方法不仅实现了未参与训练的运转工况97%以上的故障诊断准确率，而且对于加入高斯噪声后的含噪声信号也能实现较高的诊断准确率。通过与其他故障诊断网络进行对比证明了该方法的优越性。关键词：部分失火；故障诊断；振动信号；小波阈值降噪；...
