This paper delves into the baseline design under the baseline parameterization model in experimental design, focusing on the relationship between the K-aberration criterion and the word length pattern (WLP) of regular...This paper delves into the baseline design under the baseline parameterization model in experimental design, focusing on the relationship between the K-aberration criterion and the word length pattern (WLP) of regular two-level designs. The paper provides a detailed analysis of the relationship between K5and the WLP for regular two-level designs with resolution t=3, and proposes corresponding theoretical results. These results not only theoretically reveal the connection between the orthogonal parameterization model and the baseline parameterization model but also provide theoretical support for finding the K-aberration optimal regular two-level baseline designs. It demonstrates how to apply these theories to evaluate and select the optimal experimental designs. In practical applications, experimental designers can utilize the theoretical results of this paper to quickly assess and select regular two-level baseline designs with minimal K-aberration by analyzing the WLP of the experimental design. This allows for the identification of key factors that significantly affect the experimental outcomes without frequently changing the factor levels, thereby maximizing the benefits of the experiment.展开更多
针对绝缘子缺陷检测算法具有较大的参数规模和计算量导致难以部署在边缘设备,模型剪枝后难以获得正确连接,且过度稀疏化训练导致模型精度大幅度下降等问题,提出一种基于DepGraph偏移正则化的绝缘子多缺陷检测轻量化算法。通过依赖图(Dep...针对绝缘子缺陷检测算法具有较大的参数规模和计算量导致难以部署在边缘设备,模型剪枝后难以获得正确连接,且过度稀疏化训练导致模型精度大幅度下降等问题,提出一种基于DepGraph偏移正则化的绝缘子多缺陷检测轻量化算法。通过依赖图(DepGraph)对改进后YOLOv7网络建立连接关系模型,再添加偏移正则化稀疏约束对其进行组级的稀疏训练,删除冗余的连接,得到参数规模和计算量更小的轻量型检测算法。将提出的模型压缩算法应用到绝缘子多缺陷检测任务中,实验结果表明,剪枝后模型相较于未剪枝模型的参数规模和计算量分别下降65.25%和65.98%,而平均准确率仅减少1.1个百分点,验证了DepGraph偏移正则化方案在绝缘子多缺陷检测任务中的有效性;在CIFAR-10数据集上进行实验,实验结果表明,在加速比为2.88时,所提算法仍可以保持93.69%的分类精度。使用TensorRT对该算法进行推理加速,并在Jetson Orin Nano平台上部署,经过TensorRT优化后模型的检测速度达到了35.24帧/s,符合在移动设备上部署的需求。展开更多
文摘This paper delves into the baseline design under the baseline parameterization model in experimental design, focusing on the relationship between the K-aberration criterion and the word length pattern (WLP) of regular two-level designs. The paper provides a detailed analysis of the relationship between K5and the WLP for regular two-level designs with resolution t=3, and proposes corresponding theoretical results. These results not only theoretically reveal the connection between the orthogonal parameterization model and the baseline parameterization model but also provide theoretical support for finding the K-aberration optimal regular two-level baseline designs. It demonstrates how to apply these theories to evaluate and select the optimal experimental designs. In practical applications, experimental designers can utilize the theoretical results of this paper to quickly assess and select regular two-level baseline designs with minimal K-aberration by analyzing the WLP of the experimental design. This allows for the identification of key factors that significantly affect the experimental outcomes without frequently changing the factor levels, thereby maximizing the benefits of the experiment.
文摘针对绝缘子缺陷检测算法具有较大的参数规模和计算量导致难以部署在边缘设备,模型剪枝后难以获得正确连接,且过度稀疏化训练导致模型精度大幅度下降等问题,提出一种基于DepGraph偏移正则化的绝缘子多缺陷检测轻量化算法。通过依赖图(DepGraph)对改进后YOLOv7网络建立连接关系模型,再添加偏移正则化稀疏约束对其进行组级的稀疏训练,删除冗余的连接,得到参数规模和计算量更小的轻量型检测算法。将提出的模型压缩算法应用到绝缘子多缺陷检测任务中,实验结果表明,剪枝后模型相较于未剪枝模型的参数规模和计算量分别下降65.25%和65.98%,而平均准确率仅减少1.1个百分点,验证了DepGraph偏移正则化方案在绝缘子多缺陷检测任务中的有效性;在CIFAR-10数据集上进行实验,实验结果表明,在加速比为2.88时,所提算法仍可以保持93.69%的分类精度。使用TensorRT对该算法进行推理加速,并在Jetson Orin Nano平台上部署,经过TensorRT优化后模型的检测速度达到了35.24帧/s,符合在移动设备上部署的需求。
