This paper aims to provide a parametric design for robust flight controller of the model-scale helicopter. The main contributions lie in two aspects. Firstly,under near-hovering condition,a procedure is presented for ...This paper aims to provide a parametric design for robust flight controller of the model-scale helicopter. The main contributions lie in two aspects. Firstly,under near-hovering condition,a procedure is presented for simplification of the highly nonlinear and under-actuated model of the model-scale helicopter. This nonlinear system is linearized around the trim values of the chosen flight mode,followed by decomposing this high-order linear model into three lower-order subsystems according to the coupling properties among channels.After decomposition,the three subsystems are obtained which include the coupling subsystem between the roll( pitch) motion and the lateral( longitudinal) motion,the subsystem of the yaw motion and the subsystem of the vertical motion. Secondly,by using eigenstructure assignment,the problem of flight controller design can be converted into solving two optimization problems and the linear robust controllers of these subsystems are designed through solving these optimization problems. Besides, this paper contrasts and analyzed the performances of the LQR controller and the parametric controller. The results demonstrate the effectiveness and the robustness against the parametric perturbations of the parametric controller.展开更多
针对岩石薄片图像超分辨率重建过程中因纹理复杂导致现有重建方法效果不理想的问题,提出面向岩石薄片图像的超分辨率网络模型(super-resolution denoising diffusion probability model of rock slice,rsDDPMSR).针对传统上采样方法往...针对岩石薄片图像超分辨率重建过程中因纹理复杂导致现有重建方法效果不理想的问题,提出面向岩石薄片图像的超分辨率网络模型(super-resolution denoising diffusion probability model of rock slice,rsDDPMSR).针对传统上采样方法往往会导致伪影和低分辨率图像先验信息利用不充分的问题提出分层特征增强网络(layered feature enhancement network,LFE-Net),利用双通路网络对平稳小波变换分解后的高频与低频分量进行分层特征增强.为引导扩散模型的生成方向并提供丰富先验信息,将经过LFE-Net增强后的低分辨率特征与目标高分辨率加噪图像特征通道拼接作为扩散模型的条件输入.在U-Net的基础上设计了双编码器多尺度噪声预测网络(ACA-U-Net)有效处理岩石薄片多尺度信息并在跳跃连接中引入时间感知的自适应交叉注意力机制适配扩散模型不同去噪阶段的特征分布变化增强模型对关键区域的关注程度,有效提升图像重建细节.实验结果表明,rsDDPMSR在2×、4×、8×放大倍数下,峰值信噪比(PSNR)和结构相似度(SSIM)相比于CAMixerSR、SDFlow、IDM和SR3等主流重建方法具有更优的重建效果.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
文摘This paper aims to provide a parametric design for robust flight controller of the model-scale helicopter. The main contributions lie in two aspects. Firstly,under near-hovering condition,a procedure is presented for simplification of the highly nonlinear and under-actuated model of the model-scale helicopter. This nonlinear system is linearized around the trim values of the chosen flight mode,followed by decomposing this high-order linear model into three lower-order subsystems according to the coupling properties among channels.After decomposition,the three subsystems are obtained which include the coupling subsystem between the roll( pitch) motion and the lateral( longitudinal) motion,the subsystem of the yaw motion and the subsystem of the vertical motion. Secondly,by using eigenstructure assignment,the problem of flight controller design can be converted into solving two optimization problems and the linear robust controllers of these subsystems are designed through solving these optimization problems. Besides, this paper contrasts and analyzed the performances of the LQR controller and the parametric controller. The results demonstrate the effectiveness and the robustness against the parametric perturbations of the parametric controller.
文摘针对岩石薄片图像超分辨率重建过程中因纹理复杂导致现有重建方法效果不理想的问题,提出面向岩石薄片图像的超分辨率网络模型(super-resolution denoising diffusion probability model of rock slice,rsDDPMSR).针对传统上采样方法往往会导致伪影和低分辨率图像先验信息利用不充分的问题提出分层特征增强网络(layered feature enhancement network,LFE-Net),利用双通路网络对平稳小波变换分解后的高频与低频分量进行分层特征增强.为引导扩散模型的生成方向并提供丰富先验信息,将经过LFE-Net增强后的低分辨率特征与目标高分辨率加噪图像特征通道拼接作为扩散模型的条件输入.在U-Net的基础上设计了双编码器多尺度噪声预测网络(ACA-U-Net)有效处理岩石薄片多尺度信息并在跳跃连接中引入时间感知的自适应交叉注意力机制适配扩散模型不同去噪阶段的特征分布变化增强模型对关键区域的关注程度,有效提升图像重建细节.实验结果表明,rsDDPMSR在2×、4×、8×放大倍数下,峰值信噪比(PSNR)和结构相似度(SSIM)相比于CAMixerSR、SDFlow、IDM和SR3等主流重建方法具有更优的重建效果.
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。
