Tan Long Bin, Nhat Nguyen Dang Phuc
Institute of High Performance Computing (IHPC), A*STAR, 1 Fusionopolis Way, #16-16, Connexis North Tower, Singapore 138632, Singapore.
Polymers (Basel). 2022 Jul 12;14(14):2838. doi: 10.3390/polym14142838.
Thermoforming is a process where the laminated sheet is pre-heated to the desired forming temperature before being pressed and cooled between the molds to give the final formed part. Defects such as wrinkles, matrix-smear or ply-splitting could occur if the process is not optimized. Traditionally, for thermoforming of fiber-reinforced composites, engineers would either have to perform numerous physical trial and error experiments or to run a large number of high-fidelity simulations in order to determine satisfactory combinations of process parameters that would yield a defect-free part. Such methods are expensive in terms of equipment and raw material usage, mold fabrication cost and man-hours. In the last decade, there has been an ongoing trend of applying machine learning methods to engineering problems, but none for woven composite thermoforming. In this paper, two applications of artificial neural networks (ANN) are presented. The first is the use of ANN to analyze full-field contour results from simulation so as to predict the process parameters resulting in the quality of the formed product. Results show that the developed ANN can predict some input parameters reasonably well from just inspecting the images of the thermoformed laminate. The second application is to optimize the process parameters that would result in a quality part through the objectives of minimizing the maximum slip-path length and maximizing the regions of the laminate with a predesignated shear angle range. Our results show that the ANN can provide reasonable optimization of the process parameters to yield improved product quality. Overall, the results from the ANNs are encouraging when compared against experimental data. The image analysis method proposed here for machine learning is novel for composite manufacturing as it can potentially be combined with machine vision in the actual manufacturing operation to provide active feedback to ensure quality products.
层压板在被压制并在模具之间冷却以得到最终成型部件之前，先被预热到所需的成型温度。如果该过程未得到优化，可能会出现诸如皱纹、基体涂抹或层片分裂等缺陷。传统上，对于纤维增强复合材料的热成型，工程师要么必须进行大量的物理试错实验，要么运行大量的高保真模拟，以便确定能产生无缺陷部件的令人满意的工艺参数组合。这些方法在设备和原材料使用、模具制造成本以及工时方面都很昂贵。在过去十年中，一直存在将机器学习方法应用于工程问题的趋势，但在编织复合材料热成型方面却没有。本文介绍了人工神经网络（ANN）的两种应用。第一种是使用ANN分析模拟得到的全场轮廓结果，以便预测导致成型产品质量的工艺参数。结果表明，所开发的ANN仅通过检查热成型层压板的图像就能较好地预测一些输入参数。第二种应用是通过最小化最大滑移路径长度以及最大化具有预定剪切角范围的层压板区域的目标，来优化能产生高质量部件的工艺参数。我们的结果表明，ANN能够对工艺参数进行合理优化，从而提高产品质量。总体而言，与实验数据相比，ANN的结果令人鼓舞。这里提出的用于机器学习的图像分析方法在复合材料制造中是新颖的，因为它有可能在实际制造操作中与机器视觉相结合，以提供主动反馈来确保产品质量。
