棉纺织技术2024，Vol.52Issue(12)：51-57,7.
深度网络模型驱动的纱线原料特征参数反演
Inversion of yarn raw material feature parameter driven by deep neural network model
摘要
Abstract
The inversion of yarn raw material feature parameter based on hybrid neural network and improved genetic algorithm was discussed.One-dimensional Convolutional Neural Network(1D-CNN)was used to cyclically extract deep level features of yarn raw materials,and combined with Long Short-Term Memory(LSTM)network to model and express multi process information in spinning.A forward model integrating CNN-LSTM architecture was constructed to predict yarn quality.A sliding window was set to divide the dataset into multiple subsequences.The prediction accuracy under different length windows was tested.The optimal window parameters were selected to train and encapsulate the forward model.On this basis,the prediction errors of yarn coefficient of variation(CV),yarn strength coefficient of variation(CV)and yarn breaking strength indexes were used as fitness functions.Improved Genetic Algorithm(IGA)was applied to establish an inversion model for optimizing the key characteristic parameters of yarn raw materials.The experimental results showed that compared with the LSTM model,the prediction error with the three yarn quality indexes of the CNN-LSTM forward model was significantly reduced.The mean relative error(MRE)of the IGA algorithm in inverting the parameters of average fiber length,moisture regain and micronaire value was all less than 5％.After optimizing the parameters through inversion,the prediction accuracy of the forward model for the three yarn quality indexes was all improved,indicating the effectiveness of the IGA algorithm in inversion.关键词
混合神经网络/一维卷积神经网络/长短期记忆网络/改良遗传算法/特征参数反演Key words
hybrid neural network/one-dimensional convolutional neural network/long short-term memory network/improved genetic algorithm/inversion of feature parameter分类
轻工业引用本文复制引用
陈明亮,章军辉,丁羽璇,刘禹希,钱宇晗..深度网络模型驱动的纱线原料特征参数反演[J].棉纺织技术,2024,52(12):51-57,7.基金项目
江苏省博士后科研资助计划(2020Z411) （2020Z411）
