E-YQP: A self-adaptive end-to-end framework for quality prediction in yarn spinning manufacturing

Abstract

In spinning manufacturing, quality prediction is crucial for stable production, quality assurance, and cost control. Establishing a learning-based yarn quality prediction model can leverage engineering knowledge from historical production data and reduce reliance on personnel experience. The model's input comprises variable-length cotton blending information (CBI) and discrete system control parameters (CP). For variable-length input, current studies commonly use statistical indicators based on the linear aggregation assumption to extract its fixed-length features, potentially leading to information loss due to specific design choices. Additionally, they tend to overlook the relationship between CBI and CP, thus limiting the effectiveness of feature fusion. In this study, based on the induction of the information flow in spinning manufacturing systems, we propose an Encoder-based Yarn Quality Prediction (E-YQP) framework. Within this framework, we introduce a blending-mapping attention mechanism to model the blending process of cotton, extracting features from the variable-length CBI. Additionally, we propose a gate-controlled structure to simulate the impact process of CP on CBI, achieving effective feature fusion and accurate quality prediction. We conduct experiments on the actual production datasets captured from a spinning mill. The results showed that the performance of our model is superior to those of other existing methods, with an R2 of 0.83, an MAE of 0.0369, an MSE of 0.0034 and a RMSE of 0.0542. We also carried out ablation studies and demonstrated the rationality and effectiveness of each component used in the proposed model. With the integration of the blending-mapping attention mechanism and gate-controlled structure, the E-YQP framework possesses flexible data adaptability, thus holding the potential for broader application in engineering prediction tasks involving the blending of raw materials and the conduction of control parameters.