A Yarn Quality Prediction Method Based on M-ESTIMATOR Robust Broad Learning System With Tightly Cascaded Feature Layers

Aiming at the problem that multilayer neural networks rely on large datasets and broad learning system (BLS) cannot cope well with outliers in data of yarn production, which leads to low accuracy and stability when used for predicting yarn quality, we propose a robust broad learning system with the ability to resist the interference of outliers and optimize its ability to extract features. First, the feature mapping groups of the BLS were connected in a tightly cascaded manner to improve the model’s ability to express the features of the parameters of raw cotton and machine operation. Second, five M-ESTIMATOR regression loss functions, such as the Tukey function with robust characteristics, were proposed to replace the L2 norm squared loss function of the BLS. Using the iterative reweighted least-squares idea, weights were assigned to samples with different errors, and the output weights were updated to alleviate the error feedback to the error of the outliers so that the model could better fit the data. The experimental data were obtained from a medium-sized textile factory. Experimental results showed that the error of the new model was less than 3.5% when predicting the yarn strength and 4.5% when predicting the hairiness H-value. The accuracy and stability were improved compared to those of the multilayer neural network and BLS. The prediction effect of the new model satisfy the application requirements of textile enterprises, thus providing a new method for textile enterprises to predict yarn quality, which can guide production practices to a certain extent.