摘要
针对谐波减速器故障程度难以精确量化以及不同故障位置无法在同一尺度下定量分析的问题,提出一种改进深度残差网络(ResNet)结合多核支持向量数据描述(MKSVDD)的谐波减速器多状态同尺度下的定量评估方法。该方法首先提出一种新的谐波减速器多状态同尺度定量评估框架,并对微弱故障敏感的声发射信号进行连续小波变换构建二维时频图数据集;其次提出卷积注意力模块改进ResNet以充分挖掘二维时频图的深层特征;再引入多核核函数改进支持向量数据描述,基于谐波减速器正常状态的深层特征构建MKSVDD健康状态评估模型;然后,计算不同故障程度的特征相对于正常状态球心的距离,构建评估指标,通过拟合得到定量评估曲线;此外,根据谐波减速器的结构和声发射信号传播机理,提出相对距离补偿方案以构建多状态评估指标,实现谐波减速器不同健康状态在同一尺度下的定量评估。通过搭建谐波减速器实验台,对未知故障程度的数据进行多组对比实验的结果表明,改进后的深度残差网络提取到的特征更聚集,所提方法能实现谐波减速器不同故障位置在同一尺度下的定量分析,且评估误差不超过3.2%,有效完成谐波减速器多状态同尺度的定量评估。
For the problems,it is difficult to accurately quantify the fault degree of harmonic reducer,and same-scale quantitative analysis of different fault locations cannot be carried out,a same-scale quantitative assessment method is proposed for multiple states of harmonic reducer based on improved deep residual network(ResNet)and multi-kernel support vector data description(MKSVDD).First,a new same-scale quantitative assessment framework for harmonic reducer is proposed for multiple states,perform continuous wavelet transformation on acoustic emission signals that are sensitive to weak faults,and the two-dimensional time-frequency diagram dataset can be constructed.Then,convolutional block attention module is used to improve ResNet to fully exploit the deep features of the two-dimensional time-frequency diagram.Furthermore,a multi-kernel function is introduced to improve support vector data description,and an MKSVDD assessment model for healthy states is constructed based on the deep features of the harmonic reducer under the normal state.Next,the assessment indicators are constructed by calculating the distance of the features at different fault levels relative to the center of the sphere in the normal state,and then the quantitative assessment curve is obtained by fitting the distance indicators.In addition,according to the structure of the harmonic reducer and the transmission mechanism of acoustic emission signals,the relative distance compensation scheme is proposed to construct the multiple states assessment index,thereby realizing the quantitative assessment for harmonic reducer of different healthy states under the unified scale.By establishing the harmonic reducer test bench and conducting multiple sets of comparative experiments using data of unknown fault levels,the results show that the features extracted by the improved deep residual network are more concentrated.It is also shown that the proposed method can achieve same-scale quantitative analysis of different fault levels,with an error of assessment no more than 3.2%,effectively completing the same-scale quantitative assessment of harmonic reducer in multiple states.To address the difficulty in accurately quantifying the fault degree of harmonic reducers and the inability to perform same-scale quantitative analysis for different fault locations,a same-scale quantitative assessment method is proposed for multiple states of harmonic reducer based on improved deep residual network(ResNet)and multi-kernel support vector data description(MKSVDD).First,a new same-scale quantitative assessment framework for multiple states of harmonic reducer is proposed,and continuous wavelet transform is applied to acoustic emission signals sensitive to weak faults to construct a two-dimensional time-frequency image dataset.Then,a convolutionalattention module is used to improve ResNet in order to fully extract the deep features of the two-dimensional time-frequency images.Furthermore,a multi-kernel function is introduced to enhance the support vector data description,and an MKSVDD health state assessment model is constructed based on the deep features of the harmonic reducer in the normal state.Next,the distance between the features of different fault degrees and the center of the hypersphere under the normal condition is calculated to construct the assessment indicators,and the quantitative assessment curve is obtained by fitting these indicators.In addition,based on the structure of the harmonic reducer and the propagation mechanism of acoustic emission signals,a relative distance compensation scheme is proposed to construct the multi-state assessment indicator,thereby achieving quantitative assessment of different health states for harmonic reducer under a unified scale.Through the establishment of a harmonic reducer test bench and multiple comparative experiments on data with unknown fault degrees,the results show that the features extracted by the improved deep residual network are more compact.The proposed method enables same-scale quantitative assessment of different fault locations,with an assessment error not exceeding 3.2%,effectively completing the same-scale quantitative assessment of harmonic reducer in multiple states.
作者
孙宇林
罗双
康守强
王玉静
刘连胜
Sun Yulin;Luo Shuang;Kang Shouqiang;Wang Yujing;Liu Liansheng(Heilongjiang Province Key Laboratory of Pattern Recognition and Information Perception,Harbin University of Science and Technology,Harbin 150080,China;Automatic Test and Control Institute,Harbin Institute of Technology,Harbin 150001,China)
出处
《仪器仪表学报》
北大核心
2025年第6期304-316,共13页
Chinese Journal of Scientific Instrument
基金
国家自然科学基金项目(52375533)
黑龙江省自然科学基金项目(PL2024F018)
山东省自然科学基金项目(ZR2023ME057)
哈尔滨市制造业科技创新人才项目(2023CXRCCG017)资助。
关键词
谐波减速器
卷积注意力机制
多核支持向量数据描述
多故障状态
定量评估
harmonic reducer
convolutional attention mechanism
multi-kernel support vector data description
multiple fault states
quantitative assessment
