Machine learning enhanced high dynamic range fringe projection profilometry for in-situ layer-wise surface topography measurement during LPBF additive manufacturing
Machine learning enhanced high dynamic range fringe projection profilometry for in-situ layer-wise surface topography measurement during LPBF additive manufacturing
复制标题
DOI：
10.1016/j.precisioneng.2023.06.015
复制
发表时间：
2023
期刊：
影响因子：
--
通讯作者：
Haolin Zhang;Chaitanya Krishna Prasad Vallabh;Xiayun Zhao
中科院分区：
文献类型：
--
作者：
Haolin Zhang;Chaitanya Krishna Prasad Vallabh;Xiayun Zhao
条纹投影轮廓术（FPP）是一种经济有效的非破坏性方法，通常用于测量物体的精细特征和重建3D形貌。然而，为了在基于激光粉末床融合（LPBF）的增材制造（AM）过程期间使用FPP方法来测量粉末床和打印层的动态形貌，由于构建室中的材料性质和环境条件的变化而存在独特的挑战。在这项工作中，我们的目标是通过整合我们最近开发的LPBF特定的FPP传感模型，该模型具有局部传感器校准和傅立叶滤波器辅助展开，以及基于设备的高动态范围（HDR）方法和机器学习（ML）辅助FPP数据分析，来提高FPP在LPBF AM期间测量逐层表面形貌的特定应用场景中的可扩展性，准确性和分辨率。首先，应用基于投影仪的HDR方法，通过投影不同强度的正弦条纹图案来减轻阴影和强度饱和问题。其次，提出了一个ML框架，用于提高表面形貌测量精度（直接测量点的RMSE从10.57 μm提高到7.49 μm，甚至4.35 μm），并提高目前受硬件限制的分辨率（横向从38 μm提高到5 μm，纵向从10 μm提高到1 μm）。几种不同类型的候选神经网络（NN）的训练和测试使用的原位FPP测量数据和异位标准光学显微镜表征数据。多个神经网络为基础的模型，结果和比较的能力，以提高FPP的最终结果（高度测量）的精度和分辨率。通过选择最佳性能的NN使能图像超分辨率模型，所提出的ML集成HDR FPP方法有望在LPBF-AM期间更有能力和更有效地测量印刷层的表面形貌，从而将现有的最先进的方法推向LPBF印刷缺陷的期望在线检测。
Fringe Projection Profilometry (FPP) is a cost-effective and non-destructive method, typically used for measuring finer features and reconstructing 3D topography of objects. However, to use the FPP method for measuring the dynamic topography of powder bed and printed layers during Laser Powder Bed Fusion (LPBF) based additive manufacturing (AM) process, unique challenges exist due to the varying material properties and ambient conditions in the build chamber. In this work, we aim to enhance the discernibility, accuracy, and resolution of FPP in the specific application scenario of measuring layer-wise surface topography during LPBF AM by integrating our recently developed LPBF-specific FPP sensing model that features localized sensor calibration and Fourier filter-aided unwrapping with an equipment-based High dynamic range (HDR) method and machine learning (ML) aided FPP data analysis. First, a projector based HDR method is applied to mitigate the shadowing and intensity saturation problems by projecting sinusoidal fringe patterns of varying intensities. Secondly, a ML framework is developed to improve the surface topography measurement accuracy (RMSE from 10.57 μm to 7.49 μm or even 4.35 μm for directly measurable points) and enhance resolution that is currently subjected to hardware limitations (from 38 μm to 5 μm laterally and from 10 μm to 1 μm vertically). Several different types of candidate neural networks (NNs) are trained and tested using the in-situ FPP measurement data and ex-situ standard optical microscopy characterization data. Multiple NN-based models are resulted and compared in terms of their ability to enhance the accuracy and resolution of FPP's end-result (height measurement). By selecting the best-performance NN enabled image super resolution model, the proposed ML integrated HDR FPP method is expected to measure the surface topography of printed layers during LPBF-AM more capably and efficiently, thus advancing the existing state-of-the-art methods towards the desired online inspection of LPBF print defects.