铁水硅含量是反映高炉冶炼过程中热状态变化的灵敏指示剂,但无法实时在线检测,造成铁水质量调控盲目.为此,提出一种基于动态注意力深度迁移网络(Attention deep transfer network, ADTNet)的高炉铁水硅含量在线预测方法.首先,针对传统...铁水硅含量是反映高炉冶炼过程中热状态变化的灵敏指示剂,但无法实时在线检测,造成铁水质量调控盲目.为此,提出一种基于动态注意力深度迁移网络(Attention deep transfer network, ADTNet)的高炉铁水硅含量在线预测方法.首先,针对传统深度网络静态建模思路无法准确描述过程变量与铁水硅含量之间的关系,提出一种基于注意力机制模块的输入过程变量与输出硅含量之间的动态关系描述方法;其次,为降低硅含量预测模型训练时对标签数据的依赖,考虑到铁水温度与硅含量数据之间的正相关性,利用小时级硅含量标签数据微调基于分钟级铁水温度数据预训练好的深度模型的结构,进而提高基于动态注意力深度迁移网络的硅含量预测精度;同时,为增强预测网络的可解释性,实时给出了基于动态注意力机制模块计算的每个样本各过程变量对铁水硅含量的贡献度;最后,基于某钢铁厂2号高炉的工业实验,验证了该方法的准确性、有效性和先进性.展开更多
为了解决传统火焰烟雾检测算法在森林树木遮挡与雨雾天气因素影响下存在漏检误检、准确性下降、小目标检测效果不佳的缺陷,提出了基于交叉可变特征融合和动态稀疏注意力YOLOv8的遥感森林野火检测模型。首先,针对火焰烟雾目标特征复杂的...为了解决传统火焰烟雾检测算法在森林树木遮挡与雨雾天气因素影响下存在漏检误检、准确性下降、小目标检测效果不佳的缺陷,提出了基于交叉可变特征融合和动态稀疏注意力YOLOv8的遥感森林野火检测模型。首先,针对火焰烟雾目标特征复杂的问题,在C2f模块中融合可变形卷积网络(DCNv3)实现特征融合,提升对网络图像中不同尺度火焰烟雾空间位置变化的细节感知能力,增强了网络在不同尺度下的特征表示能力。然后,在主干检测网络加入BiFormer注意力模块,达到抑制干扰信息,提升模型表征能力的效果。最后,引入小目标检测层,进一步提高了检测精度。改进后的算法相比于传统算法,mAP50值提高了1.3%,P值提高了1.5%,R值提高了0.4%。In order to solve the shortcomings of the traditional flame and smoke detection algorithm under the influence of forest tree occlusion and rain and fog weather factors, such as missing detection, false detection, reduced accuracy and poor detection effect of small targets, a remote sensing wildfire detection model based on cross-variable feature fusion and dynamic sparse attention YOLOv8 is proposed. Firstly, in order to solve the problem of complex features of flame smoke targets, the C2f module is fused with a Deformable Convolution Network v3 (DCNv3) to achieve feature fusion, which improves the detail perception ability of the spatial position changes of flame smoke at different scales in the network image, and enhances the feature representation ability of the network at different scales. Then, the attention module of BiFormer was added to the backbone detection network to suppress the interference information and improve the model representation ability. Finally, small object detection layer is introduced to further improve the detection accuracy. Compared with the traditional algorithm, the mAP50 value is increased by 1.3%, the P value is increased by 1.5%, and the R value is increased by 0.4%.展开更多
文摘铁水硅含量是反映高炉冶炼过程中热状态变化的灵敏指示剂,但无法实时在线检测,造成铁水质量调控盲目.为此,提出一种基于动态注意力深度迁移网络(Attention deep transfer network, ADTNet)的高炉铁水硅含量在线预测方法.首先,针对传统深度网络静态建模思路无法准确描述过程变量与铁水硅含量之间的关系,提出一种基于注意力机制模块的输入过程变量与输出硅含量之间的动态关系描述方法;其次,为降低硅含量预测模型训练时对标签数据的依赖,考虑到铁水温度与硅含量数据之间的正相关性,利用小时级硅含量标签数据微调基于分钟级铁水温度数据预训练好的深度模型的结构,进而提高基于动态注意力深度迁移网络的硅含量预测精度;同时,为增强预测网络的可解释性,实时给出了基于动态注意力机制模块计算的每个样本各过程变量对铁水硅含量的贡献度;最后,基于某钢铁厂2号高炉的工业实验,验证了该方法的准确性、有效性和先进性.
文摘为了解决传统火焰烟雾检测算法在森林树木遮挡与雨雾天气因素影响下存在漏检误检、准确性下降、小目标检测效果不佳的缺陷,提出了基于交叉可变特征融合和动态稀疏注意力YOLOv8的遥感森林野火检测模型。首先,针对火焰烟雾目标特征复杂的问题,在C2f模块中融合可变形卷积网络(DCNv3)实现特征融合,提升对网络图像中不同尺度火焰烟雾空间位置变化的细节感知能力,增强了网络在不同尺度下的特征表示能力。然后,在主干检测网络加入BiFormer注意力模块,达到抑制干扰信息,提升模型表征能力的效果。最后,引入小目标检测层,进一步提高了检测精度。改进后的算法相比于传统算法,mAP50值提高了1.3%,P值提高了1.5%,R值提高了0.4%。In order to solve the shortcomings of the traditional flame and smoke detection algorithm under the influence of forest tree occlusion and rain and fog weather factors, such as missing detection, false detection, reduced accuracy and poor detection effect of small targets, a remote sensing wildfire detection model based on cross-variable feature fusion and dynamic sparse attention YOLOv8 is proposed. Firstly, in order to solve the problem of complex features of flame smoke targets, the C2f module is fused with a Deformable Convolution Network v3 (DCNv3) to achieve feature fusion, which improves the detail perception ability of the spatial position changes of flame smoke at different scales in the network image, and enhances the feature representation ability of the network at different scales. Then, the attention module of BiFormer was added to the backbone detection network to suppress the interference information and improve the model representation ability. Finally, small object detection layer is introduced to further improve the detection accuracy. Compared with the traditional algorithm, the mAP50 value is increased by 1.3%, the P value is increased by 1.5%, and the R value is increased by 0.4%.
