基于时空相似性的即时学习在线建模  

Online modeling of just-in-time learning based on spatial-temporal similarity

作　　者：施锦涛 陈磊 秦凯 李振兴 郝矿荣[1,2] Shi Jintao;Chen Lei;Qin Kai;Li Zhenxing;Hao Kuangrong(Engineering Research Center of Digitized Textile&Apparel Technology,Ministry of Education,Donghua University,Shanghai 201620,China;College of Information Science and Technology,Donghua Unirersiy,Shanghai 201620,China)

机构地区：[1]东华大学数字化纺织服装技术教育部工程研究中心,上海201620 [2]东华大学信息科学与技术学院,上海201620

出　　处：《仪器仪表学报》2022年第6期185-193,共9页Chinese Journal of Scientific Instrument

基　　金：上海市自然科学基金面上项目(19ZR1402300);中央高校基本科研业务费专项资助。

摘　　要：流程工业数据具有较大的时变性以及非线性,传统的离线模型难以应对实际生产过程中的工况变化,而即时学习是在线建模的有效方法。已有研究对即时学习的相似度度量方法大多只侧重于样本的空间距离,忽略了工业数据时序性的特点。为此,提出基于时空相似性的即时学习建模方法。首先,将样本点延拓成样本序列,结合动态时间规整计算样本间的时序距离。其次,提出时空相似性度量准则,通过对时序距离和空间距离进行非线性加权,构建时空相似性度量指标。最后,提出基于时空相似性的即时学习在线建模方法。将所提算法应用于公共数据集及聚酯纤维聚合过程,拟合优度分别达到了91.6%和98.6%,实验结果验证了算法的有效性和优越性。Data in the process industry are highly time-varying and nonlinear. Traditional offline models can hardly cope with the changing working conditions in the actual production process, while the just-in-time learning(JITL) is an effective online modeling method. Most of the studied similarity measurements of JITL only focus on samples’ spatial distance, which ignore the time-series characteristics of industrial data. To address this issue, a JITL method based on spatial-temporal similarity is proposed. First, the sample point is extended into a sample sequence, and the temporal-sequence distance among samples is calculated by combining dynamic time warping. Then, the spatial-temporal similarity metric(SSM) is proposed, and the SSM is constructed by nonlinearly weighting the temporal and spatial distances. Finally, the online modeling method for just-in-time learning based on spatial-temporal similarity(SS-JITL) is proposed. The algorithm is applied to a public dataset and an actual polyester fiber polymerization process. Experiment results show that the goodness of fit reaches 91.6% and 98.6%, which demonstrates the effectiveness and superiority of the proposed algorithm.

关 键 词：时空相似性 即时学习 在线建模 流程工业 数据驱动 

分 类 号：TH86[机械工程—仪器科学与技术] TP274[机械工程—精密仪器及机械]