基于时序数据库与深度学习的制丝实时数据应用研究  

Application study of real-time data in tobacco primary processing based on TSDB and deep learning

作　　者：顾茜 钱继春 GU Qian;QIAN Jichun(IT Department,China Tobacco Xiamen Industrial Co.,Ltd,Xiamen 361022,Fujian,China)

机构地区：[1]厦门烟草工业有限责任公司,信息技术部,福建厦门361022

出　　处：《中国烟草学报》2021年第3期104-113,共10页Acta Tabacaria Sinica

基　　金：福建中烟管理创新项目“基于大数据分析的CDT-5L型CTD自适应控制模型研究”(FJZY2019ZNCX019)。

摘　　要：【目的】研究气流烘丝机叶丝干燥工序出口叶丝含水率实时趋势预测的解决方案,提高产品工艺质量稳定性。【方法】基于时序数据库InfluxDB对实时生产数据进行存储、查询统计、分析告警,使用三次指数平滑算法、长短期记忆网络(LSTM)和序列到序列学习(Seq2Seq)对生产过程数据建模预测。【结果】①三次指数平滑算法集成在时序数据库中,可实现精度良好、准实时的叶丝出口含水率趋势预测;②在料头生产关键参数模仿学习方面,LSTM模型在整体性能上具有显著优势。【结论】采用时序数据库和深度学习算法结合的实时数据应用框架,横向集成最优算法模型,预测精度和应用效果良好。[Objective]This study aims to find solution of real-time trend prediction of leaf moisture content at the outlet of air flow drying machine,and improve the stability of product process quality.[Methods]The real-time data was stored,queried,counted,analyzed and alarmed based on the time-series database InfluxDB,and the data of production process was modeled and predicted by Holt-Winters Exponential Smoothing algorithms,LSTM and Seq2Seq.[Results]1)By integrating cubic exponential smoothing algorithm in time series database,accurate and quasi-real-time trend prediction of leaf moisture content can be realized;2)LSTM model has significant overall performance advantages for start stage imitation learning.[Conclusion]The real-time data application framework based on the combination of TSDB and deep learning with horizontal integration of optimal algorithm model has good prediction accuracy and high application value.

关 键 词：时间序列数据库 深度学习 工业物联网 LSTM模型 Seq2Seq模型 InfluxDB 边缘计算 

分 类 号：TS452[农业科学—烟草工业]