A Layer-Wise Surface Deformation Defect Detection by Convolutional Neural Networks in Laser Powder-Bed Fusion Images

Abstract

Surface deformation is a multi-factor, laser powder-bed fusion (LPBF) defect that cannot be avoided entirely using current monitoring systems. Distortion and warping, if left unchecked, can compromise the mechanical and physical properties resulting in a build with an undesired geometry. Increasing dwell time, pre-heating the substrate, and selecting appropriate values for the printing parameters are common ways to combat surface deformation. However, the absence of real-time detection and correction of surface deformation is a crucial LPBF problem. In this work, we propose a novel approach to identifying surface deformation problems from powder-bed images in real time by employing a convolutional neural network-based solution. Identifying surface deformation from powder-bed images is a significant step toward real-time monitoring of LPBF. Thirteen bars, with overhangs, were printed to simulate surface deformation defects naturally. The carefully chosen geometric design overcomes problems relating to unlabelled data by providing both normal and defective examples for the model to train. To improve the quality and robustness of the model, we employed several deep learning techniques such as data augmentation and various model evaluation criteria. Our model is 99% accurate in identifying the surface distortion from powder-bed images.

Keywords: LPBF; convolutional neural network; deep learning; machine learning; metal additive manufacturing; surface deformation.

Figure 1

CAD design of the test specimen.

Figure 2

Area of a layer (No of pixels) exposed to laser power vs layer numbers of test bar1.

Figure 3

An overview of EOSTATE monitoring suite.

Figure 4

Raw OT image, PB image after laser exposure and PB image after re-coating.

Figure 5

Picture of bar1 at various layers during printing. The layer number is displayed on top of each layer image.

Figure 6

The powder-bed image of layer 20 before laser exposure.

Figure 7

The powder-bed image of layer 20 after laser exposure.

Figure 8

Accuracy of models on the data sets.

Figure 9

Confusion matrices of models on all data sets.

Figure 10

The difference in the shape of the test specimen at different layers.

Figure 11

Feature maps of normal layer image 300 after each convolutional layer.

Figure 12

Feature maps of defective layer image 500 after each convolutional layer.

Figure 13

Feature maps of defective layer image 440 after each convolutional layer.

Figure 14

Feature maps of defective layer image 450 after each convolutional layer.

Figure 15

The accuracy of the model on various learning rates.

Figure 16

The loss of the model on various learning rates.

Figure 17

Classification report of model3 on different learning rates.