Abstract:

During the injection molding processes， the dimensions of molded parts were easily affected by the coupling of various complex factors. To improve prediction accuracy， a quality prediction method was proposed based on temporal convolutional networks （TCN）， Bidirectional gated recurrent units （BiGRU）， and squeeze-and-excitation （SE） attention mechanism （TCN-BiGRU-SE）. The TCN-BiGRU-SE network was utilized to extract deep features from time-series data， characterizing the dynamic changes during the injection molding processes. Quantitative feature values and dimensionless values from the injection and holding phases were extracted and stacked into a three-dimensional matrix， which was then dimensionally reduced using convolutional neural networks （CNN） to capture the changing trends at critical phases. By integrating high-frequency data， statistical features， and machine state information， an end-to-end deep prediction model was constructed for the prediction of molded part size. Comparative， ablation， and stability tests were conducted on the Foxconn injection molding dataset， along with generalization tests on three types of injection experimental datasets. The results show that the model outperforms other methods on multiple evaluation metrics， demonstrating strong robustness and generalization capability.

Key words: injection molding, quality prediction, time-series data, multi-feature fusion, deep learning

摘要：

注塑成形过程中，塑件尺寸易受多种复杂因素的耦合影响。为提高预测精度，提出一种基于时间卷积网络（TCN）-双向门控循环单元（BiGRU）-SE注意力机制（SE）的注塑质量预测方法（TCN-BiGRU-SE）。采用TCN-BiGRU-SE网络提取时序数据的深层特征，表征注塑过程中的动态变化；提取注射和保压阶段的定量特征值及量纲一值，堆叠形成三维矩阵，通过卷积神经网络（CNN）进行降维，捕捉关键阶段的变化趋势。通过融合高频数据、统计特征与机器状态信息，构建了一个端到端的深度预测模型，以实现对塑件尺寸的预测。在富士康注塑成形数据集上进行了模型对比、消融实验和稳定性检验，并在三类注塑实验数据集上进行了泛化性检验，结果表明，所建模型在多项评价指标上优于其他方法，具有良好的鲁棒性和泛化能力。

关键词: 注塑成形, 质量预测, 时序数据, 多特征融合, 深度学习