The low-frequency oscillation(LFO)has occurred in the train-network system due to the introduction of the power electronics of the trains.The modeling and analyzing method in current researches based on electrified ra...The low-frequency oscillation(LFO)has occurred in the train-network system due to the introduction of the power electronics of the trains.The modeling and analyzing method in current researches based on electrified railway unilateral power supply system are not suitable for the LFO analysis in a bilateral power supply system,where the trains are supplied by two traction substations.In this work,based on the single-input and single-output impedance model of China CRH5 trains,the node admittance matrices of the train-network system both in unilateral and bilateral power supply modes are established,including three-phase power grid,traction transformers and traction network.Then the modal analysis is used to study the oscillation modes and propagation characteristics of the unilateral and bilateral power supply systems.Moreover,the influence of the equivalent inductance of the power grid,the length of the transmission line,and the length of the traction network are analyzed on the critical oscillation mode of the bilateral power supply system.Finally,the theoretical analysis results are verified by the time-domain simulation model in MATLAB/Simulink.展开更多
现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络...现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络(GNN)对作文的多尺度特征进行联合学习的中文AES方法。首先,利用GNN分别获取作文在句子级别和段落级别的篇章特征;然后,将这些篇章特征与作文的全局语义特征进行联合特征学习,实现对作文更精准的评分;最后,构建一个中文AES数据集,为中文AES研究提供数据基础。在所构建的数据集上的实验结果表明,所提方法在6个作文主题上的平均二次加权Kappa(QWK)系数相较于R2-BERT(Bidirectional Encoder Representations from Transformers model with Regression and Ranking)提升了1.1个百分点,验证了在AES任务中进行多尺度特征联合学习的有效性。同时,消融实验结果进一步表明了不同尺度的作文特征对评分效果的贡献。为了证明小模型在特定任务场景下的优越性,与当前流行的通用大语言模型GPT-3.5-turbo和DeepSeek-V3进行了对比。结果表明,使用所提方法的BERT(Bidirectional Encoder Representations from Transformers)模型在6个作文主题上的平均QWK比GPT-3.5-turbo和DeepSeek-V3分别高出了65.8和45.3个百分点,验证了大语言模型(LLMs)在面向领域的篇章级作文评分任务中,因缺乏大规模有监督微调数据而表现不佳的观点。展开更多
混合专家模型(mixture of experts,MoE)是一种神经网络模型架构,其特点是在模型中引入路由网络与专家子网络,进而代替原始的稠密网络。在推理过程中,MoE架构通过路由网络选择每次需要激活的专家子网络,仅激活其中部分专家完成给定任务...混合专家模型(mixture of experts,MoE)是一种神经网络模型架构,其特点是在模型中引入路由网络与专家子网络,进而代替原始的稠密网络。在推理过程中,MoE架构通过路由网络选择每次需要激活的专家子网络,仅激活其中部分专家完成给定任务。由于采用稀疏激活机制,混合专家模型同与其性能相当的稠密模型相比,大幅减少了训练和推理过程的计算开销,使得在给定计算成本下扩展模型规模成为可能。展开更多
Power electronic traction transformers(PETTs)will be increasingly applied to locomotives in the future for their small volume and light weight.However,similar to conventional trains,PETTs behave as constant power load...Power electronic traction transformers(PETTs)will be increasingly applied to locomotives in the future for their small volume and light weight.However,similar to conventional trains,PETTs behave as constant power loads and may cause low-frequency oscillation(LFO)to the train-network system.To solve this issue,a mathematical model of the PETT is firstly proposed and verified based on the extended describing function(EDF)method in this paper.In the proposed model,the LLC converter is simplified to an equivalent circuit consisting of a capacitor and a resistor in parallel.It is further demonstrated that the model can apply to various LLC converters with different topologies and controls.Particularly,when the parameter differences between cells are not obvious,the PETT can be simplified to a single-phase rectifier(i.e.,conventional train)by equivalent transformation.Based on the model of PETT,the system low-frequency stability and influential factors are analyzed by using the generalized Nyquist criterion.Lastly,the correctness and accuracy of theoretical analyses are validated by off-line and hardware-in-the-loop simulation results.展开更多
为满足列车实时数据协议(train real-time data protocol,TRDP)承载的时间关键型业务对通信确定性的严苛要求,提出一种时间敏感网络(time-sensitive networking,TSN)与TRDP融合的时间敏感列车通信网络方案。该方案通过构建分层的时间敏...为满足列车实时数据协议(train real-time data protocol,TRDP)承载的时间关键型业务对通信确定性的严苛要求,提出一种时间敏感网络(time-sensitive networking,TSN)与TRDP融合的时间敏感列车通信网络方案。该方案通过构建分层的时间敏感列车通信网络架构,将多样化列车车载应用系统服务需求经过TRDP映射到TSN,并利用TSN的时间同步和流量调度能力保障时间敏感列车业务流量的端到端确定性。基于此架构,设计了基于最早截止时间优先策略的门控列表生成算法,以保障关键流量的可调度性。仿真实验结果表明,该方案显著提升了TRDP周期性数据的发送时间准确性与周期稳定性,并在混合背景流量下大幅降低了端到端时延抖动,为构建下一代高可靠、高效率的列车通信网络提供了理论支撑。展开更多
情感分类是自然语言处理领域的热点研究问题之一,方面级的文本情感分类旨在识别文本不同方面间的情感极性。针对方面级情感分类模型存在特征提取能力弱、方面词与上下文间交互不充分的问题,提出基于交互对抗网络的方面级情感分类模型(As...情感分类是自然语言处理领域的热点研究问题之一,方面级的文本情感分类旨在识别文本不同方面间的情感极性。针对方面级情感分类模型存在特征提取能力弱、方面词与上下文间交互不充分的问题,提出基于交互对抗网络的方面级情感分类模型(Aspect-level Sentiment classification model based on Interactive Adversarial Networks,ASIAN)。首先,通过Transformer的双向表征编码器模型作为编码器,将方面词和上下文进行单独建模提取隐含层特征。其次,构建交互注意力网络,将隐含层特征进行交互学习。最后,对交互信息进行联合学习,做交叉熵损失、回传参数。此外,ASIAN添加了对抗训练旨在进一步优化分类效果。在SemEval-2014任务4中的Laptop、Restaurant数据集和ACL-2014的Twitter数据集上,ASIAN与大多数基线模型相比有较高的分类准确率。展开更多
基金This work was supported by the Applied Basic Research Program of Science and Technology Plan Project of Sichuan Province of China(No.2020YJ0252).
文摘The low-frequency oscillation(LFO)has occurred in the train-network system due to the introduction of the power electronics of the trains.The modeling and analyzing method in current researches based on electrified railway unilateral power supply system are not suitable for the LFO analysis in a bilateral power supply system,where the trains are supplied by two traction substations.In this work,based on the single-input and single-output impedance model of China CRH5 trains,the node admittance matrices of the train-network system both in unilateral and bilateral power supply modes are established,including three-phase power grid,traction transformers and traction network.Then the modal analysis is used to study the oscillation modes and propagation characteristics of the unilateral and bilateral power supply systems.Moreover,the influence of the equivalent inductance of the power grid,the length of the transmission line,and the length of the traction network are analyzed on the critical oscillation mode of the bilateral power supply system.Finally,the theoretical analysis results are verified by the time-domain simulation model in MATLAB/Simulink.
文摘现有基于预训练语言模型(PLM)的作文自动评分(AES)方法偏向于直接使用从PLM提取的全局语义特征表示作文的质量,却忽略了作文质量与更细粒度特征关联关系的问题。聚焦于中文AES研究,从多种文本角度分析和评估作文质量,提出利用图神经网络(GNN)对作文的多尺度特征进行联合学习的中文AES方法。首先,利用GNN分别获取作文在句子级别和段落级别的篇章特征;然后,将这些篇章特征与作文的全局语义特征进行联合特征学习,实现对作文更精准的评分;最后,构建一个中文AES数据集,为中文AES研究提供数据基础。在所构建的数据集上的实验结果表明,所提方法在6个作文主题上的平均二次加权Kappa(QWK)系数相较于R2-BERT(Bidirectional Encoder Representations from Transformers model with Regression and Ranking)提升了1.1个百分点,验证了在AES任务中进行多尺度特征联合学习的有效性。同时,消融实验结果进一步表明了不同尺度的作文特征对评分效果的贡献。为了证明小模型在特定任务场景下的优越性,与当前流行的通用大语言模型GPT-3.5-turbo和DeepSeek-V3进行了对比。结果表明,使用所提方法的BERT(Bidirectional Encoder Representations from Transformers)模型在6个作文主题上的平均QWK比GPT-3.5-turbo和DeepSeek-V3分别高出了65.8和45.3个百分点,验证了大语言模型(LLMs)在面向领域的篇章级作文评分任务中,因缺乏大规模有监督微调数据而表现不佳的观点。
文摘混合专家模型(mixture of experts,MoE)是一种神经网络模型架构,其特点是在模型中引入路由网络与专家子网络,进而代替原始的稠密网络。在推理过程中,MoE架构通过路由网络选择每次需要激活的专家子网络,仅激活其中部分专家完成给定任务。由于采用稀疏激活机制,混合专家模型同与其性能相当的稠密模型相比,大幅减少了训练和推理过程的计算开销,使得在给定计算成本下扩展模型规模成为可能。
基金supported in part by the National Natural Science Foundation of China(52125705)in part by the Natural Science Foundation of Hunan Province(2022JJ40066)。
文摘Power electronic traction transformers(PETTs)will be increasingly applied to locomotives in the future for their small volume and light weight.However,similar to conventional trains,PETTs behave as constant power loads and may cause low-frequency oscillation(LFO)to the train-network system.To solve this issue,a mathematical model of the PETT is firstly proposed and verified based on the extended describing function(EDF)method in this paper.In the proposed model,the LLC converter is simplified to an equivalent circuit consisting of a capacitor and a resistor in parallel.It is further demonstrated that the model can apply to various LLC converters with different topologies and controls.Particularly,when the parameter differences between cells are not obvious,the PETT can be simplified to a single-phase rectifier(i.e.,conventional train)by equivalent transformation.Based on the model of PETT,the system low-frequency stability and influential factors are analyzed by using the generalized Nyquist criterion.Lastly,the correctness and accuracy of theoretical analyses are validated by off-line and hardware-in-the-loop simulation results.
文摘为满足列车实时数据协议(train real-time data protocol,TRDP)承载的时间关键型业务对通信确定性的严苛要求,提出一种时间敏感网络(time-sensitive networking,TSN)与TRDP融合的时间敏感列车通信网络方案。该方案通过构建分层的时间敏感列车通信网络架构,将多样化列车车载应用系统服务需求经过TRDP映射到TSN,并利用TSN的时间同步和流量调度能力保障时间敏感列车业务流量的端到端确定性。基于此架构,设计了基于最早截止时间优先策略的门控列表生成算法,以保障关键流量的可调度性。仿真实验结果表明,该方案显著提升了TRDP周期性数据的发送时间准确性与周期稳定性,并在混合背景流量下大幅降低了端到端时延抖动,为构建下一代高可靠、高效率的列车通信网络提供了理论支撑。
文摘情感分类是自然语言处理领域的热点研究问题之一,方面级的文本情感分类旨在识别文本不同方面间的情感极性。针对方面级情感分类模型存在特征提取能力弱、方面词与上下文间交互不充分的问题,提出基于交互对抗网络的方面级情感分类模型(Aspect-level Sentiment classification model based on Interactive Adversarial Networks,ASIAN)。首先,通过Transformer的双向表征编码器模型作为编码器,将方面词和上下文进行单独建模提取隐含层特征。其次,构建交互注意力网络,将隐含层特征进行交互学习。最后,对交互信息进行联合学习,做交叉熵损失、回传参数。此外,ASIAN添加了对抗训练旨在进一步优化分类效果。在SemEval-2014任务4中的Laptop、Restaurant数据集和ACL-2014的Twitter数据集上,ASIAN与大多数基线模型相比有较高的分类准确率。
