In sheet metal forming, the quality of a formed part is strongly influenced by the local lubrication conditions on the blank. Fluctuations in lubrication distribution can cause failures such as excessive thinning and cracks. Predicting these failures in real-time for the entire part is still a very challenging task. Machine learning (ML) based digital twins and advanced computing power offer new ways to analyze manufacturing processes inline in the shortest possible time. This study presents a digital twin for simulating a deep drawing process that incorporates an advanced ML model and optimization algorithm. Convolutional neural networks with RES-SE-U-Net architecture, were used to capture the full friction conditions on the blank. The ML model was trained with data from a calibrated finite element model. The ML model establishes a correlation between the local friction conditions across the blank and the quality of the drawn part. It accurately predicts the geometry and thinning of the formed part in real-time by assessing the friction conditions on the blank. A particle swarm optimization algorithm incorporates the ML model and provides tailored recommendations for adjusting local friction conditions to promptly correct detected quality deviations with minimal amount of additional lubricant. Experiments show that the ML model deployed on an industrial control system can predict part quality in real-time and recommend adjustments in case of quality deviation in 1.6 s. The error between prediction and ground truth is on average 0.16 mm for geometric accuracy and 0.02 % for thinning.

中文翻译：

---

板材成形中数字孪生的实时质量预测和摩擦局部调整


在金属板材成形中，成形零件的质量很大程度上受到毛坯上局部润滑条件的影响。润滑分布的波动会导致过度稀疏和裂纹等故障。实时预测整个零件的这些故障仍然是一项非常具有挑战性的任务。基于机器学习 (ML) 的数字孪生和先进的计算能力提供了在尽可能短的时间内在线分析制造流程的新方法。本研究提出了一种用于模拟深拉工艺的数字孪生，其中结合了先进的机器学习模型和优化算法。采用 RES-SE-U-Net 架构的卷积神经网络用于捕获毛坯上的完整摩擦条件。机器学习模型使用来自校准有限元模型的数据进行训练。 ML 模型建立了毛坯局部摩擦条件与拉深零件质量之间的相关性。它通过评估毛坯上的摩擦条件，准确地实时预测成型零件的几何形状和减薄情况。粒子群优化算法结合了 ML 模型，并提供了调整局部摩擦条件的定制建议，以便以最少量的额外润滑剂迅速纠正检测到的质量偏差。实验表明，部署在工业控制系统上的机器学习模型可以实时预测零件质量，并在出现质量偏差时在 1.6 秒内提出调整建议。预测与地面真实情况之间的几何精度误差平均为 0.16 毫米，细化误差为 0.02%。