Abstract:

Objective The aim of this study is to address the challenges of diverse and high-dimensional data, as well as low prediction accuracy, in the multi-step and multi-variable cotton yarn production process. A feature selection method is proposed based on copula entropy (CE), combined with a one-dimensional convolutional neural network (1D-CNN) and K-nearest neighbors (KNN) algorithm, to predict four yarn quality indicators: coefficient of variation of strip stem uniformity (CV_m), hairiness index (H value), thin yarns (-50%), and thick yarns (+50%). This approach emphasizes the importance and necessity of enhancing prediction accuracy and model efficiency in yarn production.

Method This study proposed a CE-based feature selection method and a yarn quality prediction model based on 1D-CNN and KNN. Initially, CE was utilized as the basis for feature selection, quantifying the relationship between variables and targets, and selecting the top seven variables with the highest correlation for input into subsequent prediction models to achieve model lightweighting. Subsequently, a 1D-CNN-KNN model was constructed, where 1D-CNN extracted features from the variables, and KNN is utilized to fit the yarn quality indicators.

Results In order to verify the effectiveness of the proposed method, real ring spinning production data from a textile mill was used as an example. The copula entropy of raw cotton performance quantities and yarn quality indexes were calculated, and in order to determine the optimal number of features, cross-validation was adoptd to evaluate the average performance of the model with different numbers of features, and the top 7 features were finally selected as the key variables. The key variables were used as inputs to compare the performance differences between the proposed 1D-CNN-KNN and 1D-CNN, support vector regression(SVR), KNN, LightGBM, and Transformer models. The experimental results showed that the proposed model had higher prediction accuracy for yarn quality. Specifically, for the four yarn quality indicators of the mean absolute error, root mean square error, and cotticient of determination of the proposed models were improved by 18.4%, 16.5%, and 23.8%, respectively. Due to the small sample of experimental data, in order to verify the generalization performance of the proposed model, different training set samples were set to discuss the generalization ability of the model under small samples, and the experiments showed that the proposed model's comprehensive fitting ability was better than other models, and the performance was more stable. In addition, by comparing the results of different feature selection methods on the model performance, the copula entropy-based method had the highest prediction accuracy and the prediction time was shortened by 36.5% on average, improving the production prediction efficiency. Detailed analysis showed that CE-based feature selection effectively reduced the data dimensionality while retaining key information related to yarn quality. The 1D-CNN component of the model was able to capture complex patterns and features from the selected variables, which were then fitted by a KNN algorithm. The combination of these techniques produced an efficient and accurate predictive model for yarn quality metrics.

Conclusion The proposed CE-based feature selection method combined with the 1D-CNN-KNN model has proven effective in improving the prediction accuracy and efficiency of yarn quality indicators in the cotton yarn production process. This approach has the potential to contribute to the optimization of yarn production processes and the improvement of yarn quality in the textile industry. However, the current experimental data is only for cotton yarns, which can be extended to quality prediction tasks for more yarn types, such as chemical fiber and hemp, in the future. Thus, a model with better generalization can be constructed to further improve the feature selection and model training process to achieve better performance.

Key words: cotton yarn quality prediction, multidimensional feature, copula entropy, K-nearest neighborhood algorithm, one-dimensional convolutional neural network