Data-driven online prediction and control method for injection molding product quality

Abstract

Injection molding is a complex, non-linear production process where product quality depends on variable and interacting process parameters. In continuous mass production, fluctuations in process parameters make it impossible to ensure the stability of product quality. Existing quality control mainly relies on historical experience for manual adjustments, often leading to high scrap rates, reduced productivity, and resource consumption. Therefore, this paper proposes a data-driven quality control method for injection molded products, which uses a feature prediction model to forecast the process parameters for the next production cycle. An optimization algorithm based on the quality prediction model is used to fine-tune the process parameters, providing operators with a reasonable parameter scheme in advance. First, a time series feature prediction model is proposed based on the multiscale retention module in the Retentive Network (RetNet) model. This model integrates Empirical Mode Decomposition (EMD) and a Mixed Domain Attention Module (MDAM). The model extends features across different time dimensions using EMD to explore the time-series relationships. Additionally, a new MDAM is designed to identify key process features adaptively. Second, a quality prediction model based on the Extreme Gradient Boosting (XGBoost) algorithm is built on the feature prediction model. The output of the prediction model is used to calculate fitness. At the same time, the Dung Beetle Optimization algorithm is employed for efficient reverse search to adjust the process parameters precisely. Finally, the effectiveness of the proposed method in injection product quality control is validated through the computational analysis of two injection molding datasets, providing strong support and solutions for real-time quality management in this field.