Online Prediction And System Development Of Drying Quality Based On Deep Learning Network

Drying is the most critical and important part of the tobacco processing process,and the process has a significant impact on the aroma,smoke and taste characteristics of the product,and the quality of the drying process directly determines the quality of the product.This also determines that the key to efficient and stable operation of the production process is to predict the product quality quickly and accurately when the actual production conditions such as raw material characteristics and production conditions change,and to coordinate the regulation of the process parameters to ensure the stability of the quality of the drying process.At present,the research on the prediction of cigarette-drying production quality is mainly based on traditional mathematical models or expert systems,but not from the perspective of online data of cigarette production and the adaptiveness of the process quality prediction model,so it is difficult to fully explore the time-series correlations in the actual cigarette-drying production process data and improve the adaptiveness and prediction accuracy of the process quality prediction model.To address the above problems,this thesis constituted an encoding component by deep TCN with the help of sequence-to-sequence learning structure for extracting multi-source timing features,mining long-range timing "memory",and using residual link association to fuse high-order convolutional features and low-order original process features to compensate for distortion information,and at the same time,for focusing on key moment information.The decoding component was composed of a residual LSTM network,which was used to extract the timing information in the process quality synchronously,and the residual link correlation was used to fuse the process parameter feature information to complete the process feature information extraction,so as to build a universal model for the prediction of the cigarette-drying process quality.On this basis,to solve the problem that in actual cigarette-drying production,due to the adjustment of equipment parameters,environmental changes or changes in material characteristics and other production conditioned,resulting in large differences between the training samples and the original training sample set under different production conditions,the network structure and some parameters of the universal model built by migration learning were shared with the online prediction model to achieve online updating of the model.Furthermore,the actual data prediction cases of the cigarette-drying process were used.Finally,according to the actual processing needs of the cigarette-drying process,we developed an accurate prediction system for yarn-drying process quality,which realized the functions of process index monitoring,process quality online prediction and process abnormality warning,and carried out application verification.The results of this thesis demonstrate the feasibility and effectiveness of the application of migration learning in adaptive prediction of the cigarette-drying process quality,and also provide a new idea for online process-oriented process shop quality prediction based on deep learning,which provides a basis for subsequent work on optimization of production process parameters.