Accurate ultra-short-term prediction of the Earth rotation parameters(ERP)holds paramount impor-tance for real-time applications,particularly in reference frame conversion.Among them,diurnal rota-tion(UT1-UTC)which ca...Accurate ultra-short-term prediction of the Earth rotation parameters(ERP)holds paramount impor-tance for real-time applications,particularly in reference frame conversion.Among them,diurnal rota-tion(UT1-UTC)which cannot be directly estimated through Global Navigation Satellite System(GNSS)techniques,significantly affects the rapid and ultra-rapid orbit determination of GNsS satellites.Pres-ently,the traditional LS(least squares)+AR(autoregressive)and LS+MAR(multivariate autoregressive)hybrid methods stand as primary approaches for UT1-UTC ultra-short-term predictions(1-10 days).The LS+MAR hybrid method relies on the UT1-UTC and LOD(length of day)series.However,the correlation between LOD and first-order-difference UT1-UTC is stronger than that between LOD and UT1-UTC.In light of this,and with the aid of the first-order-difference UT1-UTC,we propose an enhanced LS+MAR hybrid method to UT1-UTC ultra-short-term prediction.By using the UT1-UTC and LOD data series of the IERS(International Earth Rotation and Reference Systems Service)EOP 14 C04 product,we conducted a thorough analysis and evaluation of the improved method's prediction performance compared to the traditional LS+AR and LS+MAR hybrid methods.According to the numerical results over more than 210 days,they demonstrate that,when considering the correlation information between the LoD and the first-order-difference UT1-UTC,the mean absolute errors(MAEs)of the improved LS+MAR hybrid method range from 21 to 934μs in 1-10 days predictions.In comparison to the traditional LS+AR hybrid method,the MAEs show a reduction of 7-53μs in 1-10 days predictions,with corresponding improvement percentages ranging from 1 to 28%.Similarly,when compared to the traditional LS+MAR hybrid method,the MAEs have a reduction of 5-42μs in 1-10 days predictions,with corresponding improvement percentages ranging from 4-20%.Additionally,when aided by GNSS-derived LOD data series,the MAEs of improved LS+MAR hybrid method experience further reduction.展开更多
针对不平衡文本分类问题中少数类样本在分类器上预测精度低的问题,提出了一种基于改进的Focal Loss损失函数和EDA(Easy Data Augmentation)文本增强技术的不平衡文本分类算法。在训练数据层面利用EDA文本增强技术对小样本数据进行增强;...针对不平衡文本分类问题中少数类样本在分类器上预测精度低的问题,提出了一种基于改进的Focal Loss损失函数和EDA(Easy Data Augmentation)文本增强技术的不平衡文本分类算法。在训练数据层面利用EDA文本增强技术对小样本数据进行增强;考虑到样本训练难易程度的动态变化,改进了Focal Loss损失函数平衡因子参数的设定方式;接着利用增强后的数据和改进后的损失函数结合较为简单且保留文本语序信息的DCNN模型进行分类模型的训练。在搜狗新闻数据集上,控制相同的参数进行对比实验,结果表明EDA技术和改进的Focal loss损失函数对于不平衡问题都有一定的改善作用,综合应用两种技术的算法获得了最好的表现。展开更多
基金supported by China Natural Science Fund,China(No.42004016)the science and technology innovation Program of Hunan Province,China(No.2023RC3217)+1 种基金Research Foundation of the Department of Natural Resources of Hunan Province(Grant No:20240105CH)HuBei Natural Science Fund,China(No.2020CFB329).
文摘Accurate ultra-short-term prediction of the Earth rotation parameters(ERP)holds paramount impor-tance for real-time applications,particularly in reference frame conversion.Among them,diurnal rota-tion(UT1-UTC)which cannot be directly estimated through Global Navigation Satellite System(GNSS)techniques,significantly affects the rapid and ultra-rapid orbit determination of GNsS satellites.Pres-ently,the traditional LS(least squares)+AR(autoregressive)and LS+MAR(multivariate autoregressive)hybrid methods stand as primary approaches for UT1-UTC ultra-short-term predictions(1-10 days).The LS+MAR hybrid method relies on the UT1-UTC and LOD(length of day)series.However,the correlation between LOD and first-order-difference UT1-UTC is stronger than that between LOD and UT1-UTC.In light of this,and with the aid of the first-order-difference UT1-UTC,we propose an enhanced LS+MAR hybrid method to UT1-UTC ultra-short-term prediction.By using the UT1-UTC and LOD data series of the IERS(International Earth Rotation and Reference Systems Service)EOP 14 C04 product,we conducted a thorough analysis and evaluation of the improved method's prediction performance compared to the traditional LS+AR and LS+MAR hybrid methods.According to the numerical results over more than 210 days,they demonstrate that,when considering the correlation information between the LoD and the first-order-difference UT1-UTC,the mean absolute errors(MAEs)of the improved LS+MAR hybrid method range from 21 to 934μs in 1-10 days predictions.In comparison to the traditional LS+AR hybrid method,the MAEs show a reduction of 7-53μs in 1-10 days predictions,with corresponding improvement percentages ranging from 1 to 28%.Similarly,when compared to the traditional LS+MAR hybrid method,the MAEs have a reduction of 5-42μs in 1-10 days predictions,with corresponding improvement percentages ranging from 4-20%.Additionally,when aided by GNSS-derived LOD data series,the MAEs of improved LS+MAR hybrid method experience further reduction.