In order to predict the extreme load of the mechanical components during the entire life,an automatic method based on the bootstrapping technology(BT)is proposed to determine the most suitable threshold.Based on all t...In order to predict the extreme load of the mechanical components during the entire life,an automatic method based on the bootstrapping technology(BT)is proposed to determine the most suitable threshold.Based on all the turning points of the load history and a series of thresholds estimated in advance,the generalized Pareto distribution is established to fit the exceedances.The corresponding distribution parameters are estimated with the maximum likelihood method.Then,BT is employed to calculate the mean squared error(MSE)of each estimated threshold based on the exceedances and the specific distribution parameters.Finally,the threshold with the smallest MSE will be the optimal one.Compared to the kurtosis method and the mean excess function method,the average deviation of the probability density function of exceedances determined by BT reduces by 38.52%and 29.25%,respectively.Moreover,the quantile-quantile plot of the exceedances determined by BT is closer to a straight line.The results suggest the improvement of the modeling flexibility and the determined threshold precision.If the exceedances are insufficient,BT will enlarge their amount by resampling to solve the instability problem of the original distribution parameters.展开更多
文摘目标检测方法甚高速卷积神经网络(Faster Region-based Convolutional Neural Network,Faster RCNN)在训练过程中存在负样本远多于正样本的问题,即数据集不平衡问题。针对该问题,提出了一个综合定位误差和分类误差的判别函数用于判别难正样本,基于该函数和难负样本挖掘提出了改进的自助采样法,并提出了基于该自助采样的"五步训练法"用于训练Faster RCNN。与传统的Faster RCNN训练方法相比,五步法加强了对难样本的学习,提高了网络泛化能力,减少了误判;训练出的模型在Pascal VOC 2007数据集上测试的平均正确率均值(mean Average Precision,mAP)提高了2.4%,在FDDB(Face Detection Data Set and Benchmark)相同检出率下误检率降低了3.2%,且边框拟合度更高。
基金The National Science and Technology Pillar Program of China(No.2015BAF07B00)
文摘In order to predict the extreme load of the mechanical components during the entire life,an automatic method based on the bootstrapping technology(BT)is proposed to determine the most suitable threshold.Based on all the turning points of the load history and a series of thresholds estimated in advance,the generalized Pareto distribution is established to fit the exceedances.The corresponding distribution parameters are estimated with the maximum likelihood method.Then,BT is employed to calculate the mean squared error(MSE)of each estimated threshold based on the exceedances and the specific distribution parameters.Finally,the threshold with the smallest MSE will be the optimal one.Compared to the kurtosis method and the mean excess function method,the average deviation of the probability density function of exceedances determined by BT reduces by 38.52%and 29.25%,respectively.Moreover,the quantile-quantile plot of the exceedances determined by BT is closer to a straight line.The results suggest the improvement of the modeling flexibility and the determined threshold precision.If the exceedances are insufficient,BT will enlarge their amount by resampling to solve the instability problem of the original distribution parameters.
