Research Update

Defect Detection in AM Process (LPBF)

Toward in-situ flaw detection in laser powder bed fusion additive manufacturing through layerwise imagery and machine learning [1]

Figure 1. Sampling

Snow et al. investigated the machine learning approach to process layerwise surface monitoring data to detect flaws in the LPBF. 6 different images states with different light conditions were used as input data to identify flaws in the additively manufactured parts. Convolutional neural networks (CNN) is adapted to classify the existence of anomalies. The datasets are labeled with the X-ray computed tomography (XCT) data as ground truth. The input data was concatenated from various combinations of optical images. This work investigated the classification accuracy based on the size of defects. When the size of flaws is smaller than 100 μm, the classification accuracy is only reaching 68%. However, the proposed approach shows the accuracy of 87.3% for defects that have the size above 200 μm in diameter. Here is the research question: Can we construct multimodal learning approach with point cloud data and statistical features together to identify defects?

Deep learning of variant geometry in layerwise imaging profiles for additive manufacturing quality control [2]

Figure 2. Schematic flowchart of the proposed method

Imani et al. proposed a machine learning-based detection of layerwise defects focused on variant geometry in each layer of the AM process. In the AM process (especially LPBF), there is a case where there is a dramatic change in build geometries from one layer to the next. This proposed method introduces layerwise region of interest (ROI) estimation to overcome these difficulties. Based on the slice data extracted from the design file (CAD), ROIs are derived from the surface image. Divide the ROIs into smaller pieces and extract useful spatial information through geometry analysis. The extracted spatial information expressed in three-color information (RGB) was obtained to find the characteristics for detecting defects in the deep learning model. This paper has compared the proposed algorithm with traditional algorithms that take statistical features (mean, median, variance, skewness, kurtosis, minimum, maximum, and range) as input. Here is the research question: Can the classified label be replaced with statistical features? (in other words, is there any way to make a generalized correlation between defect and statistical features?)

References:

[1] Snow, Z., Diehl, B., Reutzel, E. W., & Nassar, A. (2021). Toward in-situ flaw detection in laser powder bed fusion additive manufacturing through layerwise imagery and machine learning. Journal of Manunfacturing Systems, 59, 12-26.

[2] Imani, F., Chen, R., Diewald, E., Reutzel, E., & Yang, H. (2019). Deep learning of variant geometry in layerwise imaging profiles for additive manufacturing quality control. Journal of Manufacturing Science and Engineering, 141(11), 111001.