针对FasterVit网络存在的注意力机制失衡、池化策略缺陷导致部分重要特征无法保留和损失函数不能全面考虑所有类别的信息导致学习到的特征比较分散等问题,提出了一种基于CFasterVit-三并联分支融合注意力机制(triple-parallel fusion at...针对FasterVit网络存在的注意力机制失衡、池化策略缺陷导致部分重要特征无法保留和损失函数不能全面考虑所有类别的信息导致学习到的特征比较分散等问题,提出了一种基于CFasterVit-三并联分支融合注意力机制(triple-parallel fusion attention model,TFAM)与余弦均匀流形逼近与投影(cosineuniform manifold approximation and projection,COS-UMAP)模型的滚动轴承故障诊断方法。该模型由FasterVit-TFAM网络、COS-UMAP降维算法和激活函数类距均值标准差损失函数(class-distance mean standard deviation loss,CMSD)-Softmax组成。首先,提出了一种新的注意力机制TFAM,并与FasterVit网络结合,提升了FasterVit网络信息关注的均衡性和表征能力;其次,将基于COS-UMAP降维算法取代FasterVit网络全连接层前最后一次池化操作,有效筛选并保留多维数据中的重要特征;最后,将类距均值标准差损失函数替换Softmax激活函数中的交叉熵损失函数,更全面地学习特征并提高模型的泛化性。西安交通大学滚动轴承数据集滚动轴承故障试验结果表明,TFAM注意力机制和其他注意力机制相比诊断准确率最大提升8.0%,COS-UMAP对比其他降维算法诊断准确率最大提升15.8%,CMSD对比交叉熵损失函数诊断准确率提升0.5%,所提模型对故障样本的识别准确率达到了99.6%,相比FasterVit提升了1.4%,相较于其他网络模型最大提升7.8%;东南大学滚动轴承数据集仿真验证试验结果表明,所提模型对故障样本识别率达98.6%,相比FasterVit提升了2.2%,平均每轮训练时间缩短了16.92 s,对比其他网络模型最大提升12.2%,有效提高了滚动轴承故障诊断模型的准确率和泛化性能。展开更多
In this work it is shown that the kinetic energy and the exchange-correlation energy are mutual dependent on each other.This aspect is first derived in an orbital-free context.It is shown that the total Fermi potentia...In this work it is shown that the kinetic energy and the exchange-correlation energy are mutual dependent on each other.This aspect is first derived in an orbital-free context.It is shown that the total Fermi potential depends on the density only,the individual parts,the Pauli kinetic energy and the exchange-correlation energy,however,are orbital dependent and as such mutually influence each other.The numerical investigation is performed for the orbital-based non-interacting Kohn-Sham system in order to avoid additional effects due to further approximations of the kinetic energy.The numerical influence of the exchange-correlation functional on the non-interacting kinetic energy is shown to be of the orderof a few Hartrees.For chemical purposes,however,the energetic performance as a function of the nuclear coordinates is much more important than total energies.Therefore,the effect on the bond dissociation curve was studied exemplarily for the carbon monoxide.The data reveals that,the mutual influence between the exchange-correlation functional and the kinetic energy has a significant influence on bond dissociation energies and bond distances.Therefore,the effect of the exchange-correlation treatment must be considered in the design of orbital-free density functional approximations for the kinetic energy.展开更多
文摘针对FasterVit网络存在的注意力机制失衡、池化策略缺陷导致部分重要特征无法保留和损失函数不能全面考虑所有类别的信息导致学习到的特征比较分散等问题,提出了一种基于CFasterVit-三并联分支融合注意力机制(triple-parallel fusion attention model,TFAM)与余弦均匀流形逼近与投影(cosineuniform manifold approximation and projection,COS-UMAP)模型的滚动轴承故障诊断方法。该模型由FasterVit-TFAM网络、COS-UMAP降维算法和激活函数类距均值标准差损失函数(class-distance mean standard deviation loss,CMSD)-Softmax组成。首先,提出了一种新的注意力机制TFAM,并与FasterVit网络结合,提升了FasterVit网络信息关注的均衡性和表征能力;其次,将基于COS-UMAP降维算法取代FasterVit网络全连接层前最后一次池化操作,有效筛选并保留多维数据中的重要特征;最后,将类距均值标准差损失函数替换Softmax激活函数中的交叉熵损失函数,更全面地学习特征并提高模型的泛化性。西安交通大学滚动轴承数据集滚动轴承故障试验结果表明,TFAM注意力机制和其他注意力机制相比诊断准确率最大提升8.0%,COS-UMAP对比其他降维算法诊断准确率最大提升15.8%,CMSD对比交叉熵损失函数诊断准确率提升0.5%,所提模型对故障样本的识别准确率达到了99.6%,相比FasterVit提升了1.4%,相较于其他网络模型最大提升7.8%;东南大学滚动轴承数据集仿真验证试验结果表明,所提模型对故障样本识别率达98.6%,相比FasterVit提升了2.2%,平均每轮训练时间缩短了16.92 s,对比其他网络模型最大提升12.2%,有效提高了滚动轴承故障诊断模型的准确率和泛化性能。
基金The project was supported by the Fund for Scientific Research in Flanders (FWO-Vlaanderen) for Research Grant G021115N.
文摘In this work it is shown that the kinetic energy and the exchange-correlation energy are mutual dependent on each other.This aspect is first derived in an orbital-free context.It is shown that the total Fermi potential depends on the density only,the individual parts,the Pauli kinetic energy and the exchange-correlation energy,however,are orbital dependent and as such mutually influence each other.The numerical investigation is performed for the orbital-based non-interacting Kohn-Sham system in order to avoid additional effects due to further approximations of the kinetic energy.The numerical influence of the exchange-correlation functional on the non-interacting kinetic energy is shown to be of the orderof a few Hartrees.For chemical purposes,however,the energetic performance as a function of the nuclear coordinates is much more important than total energies.Therefore,the effect on the bond dissociation curve was studied exemplarily for the carbon monoxide.The data reveals that,the mutual influence between the exchange-correlation functional and the kinetic energy has a significant influence on bond dissociation energies and bond distances.Therefore,the effect of the exchange-correlation treatment must be considered in the design of orbital-free density functional approximations for the kinetic energy.
