The 20,000-ton combined train running has greatly promoted China’s heavy-haul railway transportation capability. The application of controllable train-tail devices could improve the braking wave of the train and brak...The 20,000-ton combined train running has greatly promoted China’s heavy-haul railway transportation capability. The application of controllable train-tail devices could improve the braking wave of the train and braking synchronism, and alleviate longitudinal impulse.However, the characteristics of the controllable train-tail device such as exhaust area, exhaust duration and exhaust action time are not uniform in practice, and their effects on the longitudinal impulse of the train are not apparent,which is worth studying. In this work, according to the formation of the Datong-Qinhuangdao Railway, the train air brake and longitudinal dynamics simulation system(TABLDSS) is applied to establish a 20,000-ton combined train model with the controllable train-tail device, and the braking characteristics and the longitudinal impulse of the train are calculated synchronously with changing the air exhaust time, exhaust area, and action lag time under initial braking. The results show that the maximum coupler force of the combined train will decrease with the extension of the continuous exhaust time, while the total exhaust time of the controllable train-tail device remains unchanged;the maximum coupler force of the combined train reduces by32.5% with the exhaust area increasing from 70% to 140%;when the lag time between the controllable train-tail device and the master locomotive is more than 1.5 s, the maximum coupler force of the train increases along with the time difference enlargement.展开更多
鉴于ROC曲线下面积(Area Under the ROC Curve,AUC)对数据分布的不敏感特性,面向AUC的对抗训练(AdAUC)近来已成为机器学习领域中抵御长尾分布下对抗攻击的有效范式之一。当前主流方法大多遵循基于平方替代损失的AUC对抗训练框架,并将成...鉴于ROC曲线下面积(Area Under the ROC Curve,AUC)对数据分布的不敏感特性,面向AUC的对抗训练(AdAUC)近来已成为机器学习领域中抵御长尾分布下对抗攻击的有效范式之一。当前主流方法大多遵循基于平方替代损失的AUC对抗训练框架,并将成对比较形式的AUC对抗损失重构为一个逐样本的随机鞍点优化问题,克服端到端的计算瓶颈。然而,面向复杂的实际应用场景,基于平方损失设计的AUC对抗训练框架恐难以适应多样的下游任务需求。此外,与传统对抗训练范式类似,面向AUC的对抗训练方法在提高模型对抗鲁棒性的同时,也会降低模型在正常样本上的AUC性能,而目前鲜有针对该问题的有效解决方案。鉴于此,本文对如何构建一般化的高效AUC对抗机器学习范式展开系统研究。首先,提出了一种基于标准化分数扰动的通用AUC对抗训练框架(NSAdAUC),在相对温和的条件下,该框架可通过直接扰动模型对样本的预测得分实现对AUC指标的攻击,且不依赖于特定的AUC替代损失。在此基础上,本文进一步指出鲁棒AUC误差可分解为标准AUC误差和边界AUC误差两项之和,并据此设计了一种基于排序感知对抗正则化的AUC对抗训练框架(RARAdAUC),同时兼顾模型的标准AUC和鲁棒AUC性能。为验证所提框架的有效性,在5个长尾基准数据集上进行了大量实验,结果表明所提NSAdAUC和RARAdAUC框架在多种对抗攻击下的鲁棒性均优于现有方法,可在平均意义上分别产生0.94%、5.52%的标准AUC和5.69%、5.41%的鲁棒AUC性能提升。展开更多
基金China National Railway Group Co.,Ltd(N2020J037).
文摘The 20,000-ton combined train running has greatly promoted China’s heavy-haul railway transportation capability. The application of controllable train-tail devices could improve the braking wave of the train and braking synchronism, and alleviate longitudinal impulse.However, the characteristics of the controllable train-tail device such as exhaust area, exhaust duration and exhaust action time are not uniform in practice, and their effects on the longitudinal impulse of the train are not apparent,which is worth studying. In this work, according to the formation of the Datong-Qinhuangdao Railway, the train air brake and longitudinal dynamics simulation system(TABLDSS) is applied to establish a 20,000-ton combined train model with the controllable train-tail device, and the braking characteristics and the longitudinal impulse of the train are calculated synchronously with changing the air exhaust time, exhaust area, and action lag time under initial braking. The results show that the maximum coupler force of the combined train will decrease with the extension of the continuous exhaust time, while the total exhaust time of the controllable train-tail device remains unchanged;the maximum coupler force of the combined train reduces by32.5% with the exhaust area increasing from 70% to 140%;when the lag time between the controllable train-tail device and the master locomotive is more than 1.5 s, the maximum coupler force of the train increases along with the time difference enlargement.
文摘鉴于ROC曲线下面积(Area Under the ROC Curve,AUC)对数据分布的不敏感特性,面向AUC的对抗训练(AdAUC)近来已成为机器学习领域中抵御长尾分布下对抗攻击的有效范式之一。当前主流方法大多遵循基于平方替代损失的AUC对抗训练框架,并将成对比较形式的AUC对抗损失重构为一个逐样本的随机鞍点优化问题,克服端到端的计算瓶颈。然而,面向复杂的实际应用场景,基于平方损失设计的AUC对抗训练框架恐难以适应多样的下游任务需求。此外,与传统对抗训练范式类似,面向AUC的对抗训练方法在提高模型对抗鲁棒性的同时,也会降低模型在正常样本上的AUC性能,而目前鲜有针对该问题的有效解决方案。鉴于此,本文对如何构建一般化的高效AUC对抗机器学习范式展开系统研究。首先,提出了一种基于标准化分数扰动的通用AUC对抗训练框架(NSAdAUC),在相对温和的条件下,该框架可通过直接扰动模型对样本的预测得分实现对AUC指标的攻击,且不依赖于特定的AUC替代损失。在此基础上,本文进一步指出鲁棒AUC误差可分解为标准AUC误差和边界AUC误差两项之和,并据此设计了一种基于排序感知对抗正则化的AUC对抗训练框架(RARAdAUC),同时兼顾模型的标准AUC和鲁棒AUC性能。为验证所提框架的有效性,在5个长尾基准数据集上进行了大量实验,结果表明所提NSAdAUC和RARAdAUC框架在多种对抗攻击下的鲁棒性均优于现有方法,可在平均意义上分别产生0.94%、5.52%的标准AUC和5.69%、5.41%的鲁棒AUC性能提升。
