订阅最新相关文献

针对现有自动调制识别模型参数量大且抗干扰能力不足的问题,文中提出一种融合U-Net信号去噪与旋转位置编码的轻量化自动调制识别网络.首先,设计复数域U-Net编解码架构,通过残差学习实现对输入信号的端到端噪声抑制,其对称收缩-扩张路径可分离噪声与信号的时频能量分布,增强关键特征、抑制干扰;随后,提出面向I/Q信号的旋转位置编码策略,利用正余弦函数的正交性构建相位敏感的位置嵌入矩阵,增强模型对载波频偏和相位噪声的鲁棒性;进一步,构建多尺度特征融合机制,通过二维卷积网络提取相位信息和局部空间特征;再结合旋转位置编码的位置感知Transformer捕获长程时序依赖,并利用门控循环单元实现跨尺度特征交互,增强模型对调制模式的判别能力;最后,使用全连接层对提取的特征进行分类.实验结果表明:在RML2016.10A和RML2016.10B数据集上,所提模型平均识别准确率分别达到62.20%和64.84%,较基准模型IC-AMCNet提升2.50%～5.26%,模型参数量约为IC-AMCNet的6.9%.理论分析与消融实验验证了U-Net去噪与RoPE的协同效应,为轻量化调制识别模型设计提供了新的技术路径.

抗生素耐药性(antimicrobial resistance,AMR)已成为威胁公共健康与生态安全的重要环境问题,医疗废水(hospital wastewater,HWW)因抗生素负荷高、污染基质复杂,被认为是AMR产生与扩散的重要环境源.现有研究虽揭示抗生素残留、共存污染物与高密度微生物群落共同驱动耐药基因富集,但对可转移遗传风险的刻画仍明显不足:研究停留在耐药菌与耐药基因的丰度波动,难以定量表征水平基因转移通量;工程评价亦长期侧重微生物灭活率,易忽视细胞裂解、遗传载体暴露及风险形态转换.系统梳理了HWW中AMR的主要驱动因素与扩散机制,指出水平基因转移、正向选择与生物膜微环境异质性共同放大耐药性传播,并评述现有工艺控制遗传风险方面的局限.在此基础上,提出以可转移遗传信息为核心控制目标,建议以Ⅰ类整合子整合酶基因intI1作为表征可移动遗传元件活性与群落选择压力的代表性指示因子,构建"监测—预测—调控"的动态管理框架,为HWW中AMR风险评估与精准治理提供新的研究思路.

Antimicrobial resistance(AMR)poses a growing threat to modern medicine,and it is estimated to cause more than 39 million deaths and trillions of dollars in economic losses be-tween 2025 and 2050[1].In this context,phage therapy is transitioning from a"last-resort"intervention to a standardized clinical approach for the treatment of drug-resistant bacteria[2].

The global antimicrobial resistance(AMR)crisis,fueled by the cross-border dissemination of antibiotic resistance genes(ARGs),poses a critical and escalating threat to public health.Migratory birds,with their extensive mobility and broad ecological connectivity,effectively bridge human-influenced and natural ecosystems,thereby playing a key role in the worldwide circulation of ARGs.While significant advances have been made in detecting ARGs in birds and tracing their links to anthropogenic sources,research remains fragmented:birds are often regarded merely as passive samplers,without due consideration of the mechanistic processes through which ARGs are acquired via avian traits,evolve within gut microbiomes,and are disseminated across continents through migratory networks.This review synthesizes current evidence to position migratory birds not only as global vectors but also as evolutionary incubators of ARGs.Integrating ecological,microbiological,and One Health perspectives,we delineate the pathways of bird-mediated ARG transmission across multiple scales—from molecular events within the avian gut and individual-level traits,to population-level migratory networks,and up to global flyways.A particular focus is placed on the gut microbiome,where mobile genetic elements(MGEs)serve as crucial drivers of ARG reassortment,stabilization,and horizontal transfer,thereby bridging micro-scale genetic mechanisms with macro-scale ecological spread.Looking forward,we recommend future studies to quantify ARG fluxes at key stopover sites,clarify the role of under-represented migratory taxa,and develop predictive frameworks that integrate avian movement ecology with resistome surveillance.We advocate for integrated One Health strategies and propose the Flyway-Node-Resistome(FNR)Framework as a basis for future research and transnational governance,aimed at mitigating AMR spread while conserving migratory bird populations.

面向开放域、缺乏任务特定标注的数据场景,为提升无监督事件抽取中触发词与论元抽取的稳定性与完整性,提出一种基于规则和统计学习的开放域事件抽取(RSL-ODEE)模型.基于事件语义学构建开放域事件模式(ES-ODES),将事件形式化为由触发词及主体、客体、时间、空间、工具、频度、度量等要素构成的结构化语义单元,以增强抽取过程的可解释性与跨领域泛化能力;在此模式指导下,构建融合触发词识别、抽象语义表示(AMR)语义图论元搜索与开放域机器阅读理解(MRC)论元矫正的无监督事件抽取模型,实现论元抽取、补全与断裂合并.结果表明:在中文突发事件语料库(CEC)事件提取数据集上,该方法在事件触发词与事件论元抽取的 F1 值分别仅比最优的有监督模型低 10.9%与 17.5%,具有可比较性与竞争力;在无监督条件下,引入事件语义学并结合 AMR 的结构性论元抽取与 MRC 的补全纠错能力,可有效缓解论元缺失与断裂问题,从而提升开放域事件抽取效果.

在笛卡儿网格数值模拟中,物理量急剧变化的区域往往采用网格自适应加密(adaptive mesh refinement,AMR)以确保计算精度.然而,传统的基于梯度的自适应标准依赖于预定义的阈值,可能会导致过度细化,造成网格数量增加太多,影响计算效率.此外,现有深度强化学习(deep reinforcement learning,DRL)驱动的 AMR 方法主要面向有限元框架,难以直接迁移至有限差分体系,且自适应决策与网格光顺的耦合机制尚未纳入强化学习框架.针对这一问题,本文设计了强化学习智能体的观测空间,建立了基于积分形式的奖励函数,使智能体学习符合有限差分数值特性的自适应策略,并将相邻网格层级差纳入观测空间,构造光顺惩罚项嵌入奖励函数,使智能体在执行细化/粗化操作的同时主动满足光顺约束,实现自适应决策与网格光顺的同步执行.所提出的强化学习方法在基准模拟算例中进行了评估.结果表明,在有限的网格预算下,DRL 生成的网格能更有效地捕捉物理量梯度,从而产生更低的误差值.此外,智能体在执行细化操作的同时还能满足缓冲层条件,从而显著提高网格利用率和计算效率.将 DRL 方法与所开发的加权最小二乘悬挂新生单元插值算法结合,进行了静止和运动边界的非定常流动问题求解,验证了该方法的有效性和泛化能力.

随着仓储搬运智能化从"单点尝试"进入"规模化应用"阶段,移动机器人的部署规模也从数十台向千台级集群拓展,作业场景从单车型同任务向多车型、多任务类型、多优先级并发的"百态级"应用持续演进;与此同时,面向轻量化、碎片化场景的单体智能也正成为行业新的发展方向.以海康机器人为代表的头部企业,正同步推进技术布局,适配集群规模化与单体轻量化两类差异化需求.由此,系统调度面临的不再是简单的路径规划问题,而是算法复杂度、系统确定性验证与不同业务动态适配的多重挑战.

针对现有事件图构建由于AMR全图解析复杂或仅事件类型嵌入不足的问题,提出一种基于自适应对比融合事件图的事件抽取模型ACFGEE.通过结合BERT和GCN的多源事件编码层,对句子与简单事件图信息进行编码嵌入,获得句子语义和融合角色信息的事件语义;通过事件图注意力对比融合层,使得句子信息能充分融合简单事件图中信息;利用一阶段解码层同时实现触发词和论元识别,减少多阶段解码的误差传播;通过自适应损失加权层动态调整对比融合学习损失与一阶段解码损失,实现多任务学习的自适应优化.实验结果表明,ACFGEE在公开数据集FewFC和DuEE上,其触发词识别、触发词分类、论元识别和论元分类4个任务上的F1值分别提升1.7和1.1、1.9和1.3、1.6和1.3、3.0和2.1个百分点;同时,在风电装备故障领域数据集上验证了其实用性和有效性.

液氢温区的低温与制冷技术是能源动力、红外探测和科学研究领域的关键制冷技术,磁制冷技术由于在液氢温区的潜在效率高,有望成为液氢温区的重要制冷技术之一.基于多物理场仿真软件 COMSOL Multiphysics构建了液氢温区二维瞬态主动磁回热器仿真模型.针对液氢温区磁制冷系统中的主动磁回热器,深入揭示了磁制冷工质颗粒直径、磁场-流场时序等影响因子对液氢温区磁制冷性能的影响规律,评估了20～30 K温区填充实际磁制冷工质的主动磁回热器制冷性能.结果表明AMR制冷性能对于颗粒直径的敏感性程度高于吹风系数对其影响,这是由于颗粒直径从0.05 mm增大至0.40 mm时,AMR的比表面积下降至原来的1/8.本研究为理解液氢温区多级磁制冷系统的性能影响规律以及后续整机优化设计与性能研究提供理论指导.

In off-road environments,some problems like insufficient consideration of environmental trafficability,difficulty in model convergence,and limited generalization often exist in auton-omous mobile robot(AMR)path planning research based on reinforcement learning(RL).In order to solve these problems,a distributed proximal policy optimization dynamic path planning algorithm based on experience playback enhancement(Re-DPPO)is proposed,which realizes the safe,feasible,and optimal trafficability path planning of AMR between any starting and ending points in an off-road environment.First,in order to achieve AMR path planning that takes into account environmental trafficability in off-road environments,an AMR trafficability map that integrates multiple environmental factors was constructed.Second,to ensure that AMR can simultaneously consider path trafficability during real-time obstacle avoidance,a multi-dimensional comprehensive reward function was designed that integrates dynamic obstacle avoidance,goal point proximity,and trafficability evaluation.In addition,in response to the problems of difficult model convergence and low utilization of high-value samples in traditional deep reinforcement learning(DRL)algorithms,a distributed training architecture and a priority experience replay mechanism are introduced.Finally,comparative experiments were conducted between the Re-DPPO algorithm and five DRL algorithms,namely the proximal policy optimization(PPO),experience replay enhanced proximal policy optimization algorithm(Re-PPO),distributed proximal policy optimization(DPPO),deep Q-network(DQN),and double DQN,and the path planning test was carried out in four new scenarios.The results show that among the six algorithms,the Re-DPPO algorithm exhibits the best convergence performance and path planning success rate.This algorithm can prioritize planning the optimal trafficability path while ensuring accessibility,and demonstrates good generalization ability in all four new scenarios.