Learning-Based Modeling Frameworks for Injection Molding Quality Prediction

Abstract

Injection Molding (IM) is one of the most important manufacturing processes for plastic products, and accurate quality prediction is the key to ensuring product consistency and reducing production costs. However, data-driven modeling methods for IM quality prediction still face two major challenges. On the one hand, under a single working condition with fixed machines, materials and molds, the inherent process complexity and various unpredictable fluctuations means that accurate modeling requires capturing dynamic features in production data for support, while traditional methods only focus on the mapping relationship between process parameter settings and final product quality. On the other hand, the switching of production conditions driven by changing product requirements necessitates the update of old models to enhance their adaptability, yet the collection of new quality data incurs high costs.

To address the above issues, this thesis proposes targeted modeling frameworks. First, a multi-source feature-fused backpropagation neural network (BPNN) model is designed, verifying the application value of multi-source data and serving as a baseline model to demonstrate the advantages of subsequent learning methods. Second, a novel teacher-student soft sensor framework is proposed, which takes the GRU-based autoencoder with attention mechanism (GRU-A-AE) as the teacher model to extract deep implicit temporal features and uses the BPNN as the student model to achieve fast quality prediction. This framework outperforms traditional methods in both accuracy and robustness, addressing the underutilization of high-frequency sensor data under single working conditions. This thesis also designs a semi-supervised learning-aided deep model migration framework (SSLA-DMM), which transfers the pre-trained parameters of old models and adopts a tri-training strategy to fully utilize unlabeled data, thus reducing the reliance on labeled data under new working conditions. All the proposed frameworks are tested on actual factory production data, which verifies the effectiveness of these frameworks.

Date of Award	2026
Original language	English
Awarding Institution	The Hong Kong University of Science and Technology
Supervisor	Furong GAO (Supervisor)