Method for predicting the quality of cotton yarn using a data-driven hybrid generalized regression neural network with small samples

Aiming at the problems of low prediction accuracy, poor stability, and dependence on sample size of backpropagation (BP) artificial neural networks commonly used in current yarn quality prediction research, a prediction method combining grid search (GS) and generalized regression neural network (GRNN) is proposed. Through the hyperparameter optimum finding function of GS to select the best smoothing factor of GRNN, realizing the automatic optimization of model performance, while achieving the effect of reducing complexity and enhancing generalization ability, the prediction behavior of the model for various quality characteristics of yarns is improved. On this basis, in order to mitigate the influence of features that may be redundant or less contributive to the prediction target in the yarn multi-correlation parameters, sequential backward selection (SBS) is introduced as a dimensionality reduction method to determine the optimal feature subset to obtain the main influencing factors in the design input variables. This hybrid method is compared with multiple linear regression (MLR), BP neural network (BPNN), and support vector regression (SVR). The results show that under the condition of a small amount of spinning data, the average prediction error of GS-GRNN is about 29.31% lower than BPNN, and the prediction accuracy of the base model can be effectively improved by applying SBS to optimize the feature dimension (about 37.2% higher on average), moreover, compared with SVR and MLR, this hybrid model also possesses a better comprehensive prediction effect, which verifies the stable and excellent mapping ability of SBS-GS-GRNN under the small sample set.