Deep quality prediction model for manufacturing process: Feature reconstruction and global information fusion solution

Accurate quality prediction in intelligent manufacturing is challenged by high-dimensional redundant data, insufficient spatiotemporal fusion, and poor model interpretability. While recent deep learning advances have improved prediction, they rarely address systematic feature reconstruction from redundant multi-parameter data or integrate spatial and temporal learning within a single framework. To address this problem, this paper proposes a deep learning framework that fills these gaps by integrating ReliefF-based feature reconstruction with a hybrid CNN–BiLSTM–attention architecture. ReliefF selects critical parameters and reduces redundancy; CNN–BiLSTM captures unified spatial–temporal dependencies; and attention weights influential features to enhance sensitivity and interpretability. Experiments on a real-world yarn manufacturing time-series dataset show that the proposed model outperforms state-of-the-art methods (including CNN, BiLSTM, CNN–BiLSTM, SVM, and PSO–DBN), achieving R² of 0.9536 and MAE of 0.1415. This work advances the knowledge of deep learning for manufacturing quality prediction by introducing a feature reconstruction-guided spatiotemporal attention framework, offering a more accurate and interpretable solution for complex multi-parameter processes.