Time-limited dispatching(TLD)analysis of the full authority digital engine control(FADEC)systems is an important part of the aircraft system safety analysis and a necessary task for the certification of commercial air...Time-limited dispatching(TLD)analysis of the full authority digital engine control(FADEC)systems is an important part of the aircraft system safety analysis and a necessary task for the certification of commercial aircraft and aeroengines.In the time limited dispatch guidance document ARP5107B,a single-fault Markov model(MM)approach is proposed for TLD analysis.However,ARP5107B also requires that the loss of thrust control(LOTC)rate error calculated by applying the single-fault MM must be less than 5%when performing airworthiness certification.Firstly,the sources of accuracy errors in three kinds of MM are analyzed and specified through a case study of the general FADEC system,and secondly a two-fault MM considering maintenance policy is established through analyzing and calculating the expected repair time when two related faults happen.Finally,a specific FADEC system is given to study on the influence factors of accuracy error in the single-fault MM,and the results show that the accuracy error of the single-fault MM decreases with the increase of short or long prescribed dispatch time,and the range values of short time(ST)and long time(LT)are determined to satisfy the requirement of accuracy error within 5%.展开更多
在定制化应用场景下亟需提升大语言模型(Large language models,LLMs)在特定垂直领域的语言理解和生成能力。本文提出一种适用于垂直领域的大语言模型系统开发范式——武信。其涵盖架构、数据、模型和训练等大语言模型系统的系列开发方...在定制化应用场景下亟需提升大语言模型(Large language models,LLMs)在特定垂直领域的语言理解和生成能力。本文提出一种适用于垂直领域的大语言模型系统开发范式——武信。其涵盖架构、数据、模型和训练等大语言模型系统的系列开发方法,利用人在回路的数据增强提升军事训练伤问答数据集的质量,采用梯度低秩投影(GaLore)策略对轻量级基座大语言模型进行高效全参微调。实验结果表明,所采用的全参微调方法在收敛性和准确性指标上优于主流的LoRA微调,所训练的武信大模型在军事训练伤防治专业知识理解、克服“幻觉”等方面优势明显,相关成果可为垂直领域问答大模型系统设计与应用提供参考。展开更多
基金supported by the National Natural Science Foundation of China(51705242)Shanghai Sailing Program(16YF1404900)the Fundamental Research Funds for the Central Universities(NS2015072)
文摘Time-limited dispatching(TLD)analysis of the full authority digital engine control(FADEC)systems is an important part of the aircraft system safety analysis and a necessary task for the certification of commercial aircraft and aeroengines.In the time limited dispatch guidance document ARP5107B,a single-fault Markov model(MM)approach is proposed for TLD analysis.However,ARP5107B also requires that the loss of thrust control(LOTC)rate error calculated by applying the single-fault MM must be less than 5%when performing airworthiness certification.Firstly,the sources of accuracy errors in three kinds of MM are analyzed and specified through a case study of the general FADEC system,and secondly a two-fault MM considering maintenance policy is established through analyzing and calculating the expected repair time when two related faults happen.Finally,a specific FADEC system is given to study on the influence factors of accuracy error in the single-fault MM,and the results show that the accuracy error of the single-fault MM decreases with the increase of short or long prescribed dispatch time,and the range values of short time(ST)and long time(LT)are determined to satisfy the requirement of accuracy error within 5%.
文摘在定制化应用场景下亟需提升大语言模型(Large language models,LLMs)在特定垂直领域的语言理解和生成能力。本文提出一种适用于垂直领域的大语言模型系统开发范式——武信。其涵盖架构、数据、模型和训练等大语言模型系统的系列开发方法,利用人在回路的数据增强提升军事训练伤问答数据集的质量,采用梯度低秩投影(GaLore)策略对轻量级基座大语言模型进行高效全参微调。实验结果表明,所采用的全参微调方法在收敛性和准确性指标上优于主流的LoRA微调,所训练的武信大模型在军事训练伤防治专业知识理解、克服“幻觉”等方面优势明显,相关成果可为垂直领域问答大模型系统设计与应用提供参考。
