【作者】 周涛；

【导师】 汪永超； 曹明松；

【作者基本信息】 四川大学 ， 机械（专业学位）， 2023， 硕士

【摘要】 随着烟草产业的发展,生产过程数字化、智能化需求不断增加。烘丝是卷烟工艺中最重要的工序之一,它主要通过烘丝机去除烟丝中多余的水分,使烟丝松散卷曲,增加弹性提高填充能力,并且干燥可以让烟丝中的青杂味物质挥发排出,使香气显露,味道醇厚。气流式烘丝机是常用的烘丝设备,它基于烘丝工艺设置烟丝流量、工艺气体温度、天然气流量等参数来控制烟丝干燥过程,烟丝出口含水率会直接影响成品烟丝的品质。目前大多数卷烟厂所使用的气流式烘丝机为十多年前的进口设备,由于设备老化、控制系统落后,很大程度上依赖于人工经验辅助烘丝,可靠性和准确性已无法满足质量要求。但是气流式烘丝机价格昂贵,更换设备将面临投入大、停产时间长等问题,而且烘丝过程中众多工艺参数相互耦合、协同作用,难以建立有效的数学模型,传统的控制方法无法保证出口含水率的准确和稳定。因此如何基于现有设备进行智能化改造,增强烘丝稳定性提升烟丝质量,成为了烟草工业研究的重点方向。在此背景下本文与四川中烟合作,以成都卷烟厂的HDT气流式烘丝机为研究对象,针对这一类陈旧设备进行智能控制方法研究,在不影响生产、不更换设备的前提下,提出了一种基于深度学习的气流式烘丝机智能控制方法,取得了较好的应用效果,极大降低了智能化改造成本。主要的研究工作如下:(1)剖析HDT气流式烘丝机的组成和原理,分析烟丝干燥工艺过程、工艺参数和控制系统之间的关系,对烘丝过程监测数据进行采集、预处理,并进一步开展特征工程,以增强监测数据与烟丝出口含水率之间的隐性关系。(2)从精准把握烟丝出口含水率变化趋势的角度,本文提出了一种基于局部峰值编码循环网络(Local Peak-Coded Recurrent Network,LPCRN)的出口含水率实时预测方法。LPCRN以“编码门”和“记忆门”为核心,有效表征出口含水率的波动情况,并通过记忆传递实现对长时间序列的处理,利用固定时间差的多变量预测方法实现未来120秒的出口含水率序列预测。在实例分析中LPCRN的预测效果明显优于LSTM,能够表征HDT气流式烘丝机的工艺参数与出口含水率之间的映射关系,消除了控制参数与质量指标间的时间差。(3)从HDT气流式烘丝机智能控制的角度,本文提出了一种注意力判断时间卷积网络(Attention Discriminant Time Convolutional Network,ADTCN),利用注意力机制的判别能力对烘丝过程监测数据中每个位置进行重要程度识别,同时实现了基于权重的元素筛选模式的时间卷积。并且将ADTCN应用于Encoder-Decoder框架,形成ED-ADTCN,将LPCRN预测的出口含水率替换掉烘丝过程监测数据中出口含水率监测值,输入ED-ADTCN编码为隐层输出,再利用转置卷积和反池化构成的解码器恢复为标准参数数据,根据烘丝过程监测数据与标准参数数据的差值计算出各个控制参数的调整时机和调整量,从而构成基于ED-ADTCN的控制参数智能调整算法,在实例分析中ED-ADTCN表现出强大的调整能力,对调整时机和调整量的判断能力明显优于操作人员。在实际生产中,ED-ADTCN控制下的出口含水率准确性和稳定性也更优秀,更好的保证了烘丝质量。(4)基于上述研究成果,采用Python和C#混合编程,设计并实现了HDT气流式烘丝机智能控制系统,包括数据处理与算法模型、交互与展示界面,着重分析了烘丝实时控制功能以及历史批次数据查询与分析功能。该系统具有交互性强、页面简洁明了、运行稳定等优点,实现了烘丝过程全自动化控制,提升了烘丝质量,减少了操作人员的干预。更多还原

【Abstract】 With the development of tobacco industry,the demand for digital and intelligent production process is increasing.Drying is one of the most important processes in the cigarette production process.It mainly removes excess water from the tobacco yarn through the drying machine to loosen and curl the tobacco yarn,increase elasticity and improve filling capacity,and drying allows the unpleasant substances in the tobacco yarn to evaporate and discharge,so that the aroma is revealed and the flavor is mellow.The air-flow drying machine is the common wire drying equipment,it is based on the drying process to set the tobacco flow rate,process gas temperature,natural gas flow rate and other parameters to control the drying process,the tobacco export moisture content will directly affect the quality of the finished tobacco.Currently,most of the cigarette factories are using air flow drying machines imported more than a decade ago,due to the aging of the equipment,control system backward,relying largely on manual experience to assist in drying,reliability and accuracy can no longer meet the quality requirements.However,the air flow drying machine is expensive,and replacing the equipment will face problems such as large investment and long downtime,and many process parameters in the drying process are coupled and synergistic,and it is difficult to establish an effective mathematical model,and the traditional control method cannot guarantee the accuracy and stability of the moisture content of the export.Therefore,how to carry out intelligent transformation based on the existing equipment to enhance the stability of drying and improve the quality of tobacco,has become a key direction of research in the tobacco industry.In this context,this paper cooperates with Sichuan Cigarette and takes the HDT air flow drying machine of Chengdu Cigarette Factory as the research object,and conducts research on the intelligent control method for this kind of obsolete equipment,and proposes a deep learning-based intelligent control method for air flow drying machine without affecting the production and replacing the equipment,which achieves a better application effect and greatly reduces the cost of intelligent transformation.The main research work is as follows:(1)Dissect the composition and principle of the HDT air flow drying machine,analyze the relationship between the tobacco drying process,process parameters and control system,collect and pre-process the monitoring data of the yarn drying process,and further carry out feature engineering to enhance the implicit relationship between the monitoring data and the moisture content of the tobacco export.(2)From the perspective of accurately grasping the trend of tobacco export moisture content,a real-time export moisture content prediction method based on Local Peak-Coded Recurrent Network(LPCRN)is proposed in this paper.The LPCRN uses the "coding gate" and "memory gate" as the core to effectively characterize the fluctuation of the export water content,and uses the memory transfer to process the long time series,and uses the multivariate prediction method with fixed time difference to predict the export water content series for the next 120 seconds.The prediction effect of LPCRN is significantly better than that of LSTM in the example analysis,and it can characterize the mapping relationship between process parameters and exit moisture content of HDT air flow drying machine.(3)From the perspective of intelligent control of HDT air flow drying machine,an Attention Discriminant Time Convolutional Network(ADTCN)is proposed in this paper,which uses the discriminative ability of attention mechanism to identify the importance of each position in the monitoring data of wire drying process,and at the same time The ADTCN is used to identify the importance of each position in the baking process monitoring data by using the discriminative ability of the attention mechanism,and at the same time,it achieves the temporal convolution of the weightbased element screening pattern.In addition,ADTCN is applied to the EncoderDecoder framework to form ED-ADTCN,which replaces the exit moisture content monitoring value in the drying process monitoring data with the exit moisture content monitoring value predicted by LPCRN,encodes it into the ED-ADTCN as the hidden layer output,and then recovers it into the standard parameter data using the decoder composed of transposed convolution and inverse pooling,and calculates the difference between the drying process monitoring data and the standard parameter data according to the The adjustment timing and adjustment amount of each control parameter are calculated based on the difference between the monitoring data and the standard parameter data,thus constituting an intelligent adjustment algorithm for control parameters based on ED-ADTCN,which shows strong adjustment capability in the example analysis,and the judgment ability of the adjustment timing and adjustment amount is significantly better than that of the operator.In actual production,the accuracy and stability of the exit moisture content under ED-ADTCN control are also better,which better ensures the quality of the dried silk.(4)Based on the above research results,an intelligent control system for HDT air flow drying machine is designed and implemented using a mixture of Python and C# programming,including data processing and algorithm model,interaction and display interface,focusing on analysis of the real-time control function of yarn drying and the historical batch data query and analysis function.The system has the advantages of strong interactivity,concise and clear pages,and stable operation,etc.It realizes the fully automatic control of the drying process,improves the quality of drying,and reduces the intervention of operators.更多还原