【作者】 夏恒； 汤健； 张润雨； 陈佳昆； 潘晓彤； 余文； 乔俊飞；

【Author】 Heng Xia;Jian Tang;Runyu Zhang;Jiakun Chen;Xiaotong Pan;Wen Yu;Junfei Qiao;Faculty of Information Technology, Beijing University of Technology;Beijing Laboratory of Smart Environmental Protection;Departamento de Control Automatico, Centro de In vestigation de Estudios Avanzados, National Polytechnic Institute Mexico;

【机构】 北京工业大学信息学部； 智慧环保北京实验室； 墨西哥国立理工大学高级研究中心(CINVESTAV-IPN)；

【摘要】 针对城市固废焚烧污染物二噁英(DXN)浓度检测存在的高成本、大滞后、非全面以及动态工况漂移下模型更新真值缺失等问题,提出了全流程级联式智能建模与动态检测方法。在智能建模部分,首先,依据DXN生成、吸附和排放阶段的固有特性和数据属性,进行过程数值数据、记录运行数据和火焰视频数据预处理策略以获得有效输入;其次,依据生成阶段多模态数据特性构建基于深度特征融合与改进跨层全联接深度森林回归的DXN生成模型,依据吸附阶段过程数值数据超小样本特性构建基于潜空间几何旋转虚拟样本生成和T–S模糊森林的DXN吸附模型,依据排放阶段虚实数据特性构建基于仿真机理与宽度深林的DXN排放模型;最后,利用多阶段递进级联式策略获得全流程软测量模型。在动态检测部分,对DXN生成、吸附和排放阶段分别进行工况漂移识别,提出工况漂移下的动态检测与样本库维护模式,开发在线智能检测系统。基于检测数据、验证平台和现场应用验证了所提方法的有效性。更多还原

【Abstract】 Traditional measurement methods for Dioxin(DXN) suffer from high costs, significant lag times, only partial detection, and lack of true values for model updates under dynamic operating conditions. This article proposes a novel whole process cascade intelligent modeling and dynamic detection method. In the intelligent modeling part, preprocessing strategies for process numerical data, recorded operating data, and flame video data are carried out to obtain input features. Based on the multi-modal data characteristics of the generation stage, a DXN generation model based on deep feature fusion and improved deep forest regression based on cross-layer fully connected is constructed. Based on the ultra-small sample characteristics of process numerical data in the adsorption stage, a DXN adsorption model based on latent space geometric rotation virtual sample generation and T-S fuzzy forest is constructed. Based on the virtual-real data characteristics of the emission stage, a DXN emission model based on simulation mechanism and broad forest is constructed. Finally, a multi-stage progressive cascade strategy is used to obtain the DXN whole process soft measurement model. In the dynamic detection part, working condition drift identification is carried out for the DXN generation, adsorption, and emission stages respectively. A dynamic detection and sample library maintenance mode under concept drift is proposed and a online intelligent detection system is developed. The effectiveness of the proposed method are verified based on detection data, verification platform, and industrial field applications.更多还原