PSI - Issue 77

A. Polanský et al. / Procedia Structural Integrity 77 (2026) 529–536 Adam Polanský / Structural Integrity Procedia 00 (2026) 000 – 000

530

2

structural integrity. It is affected by many factors discussed in Section 3. In this paper, we focus on the possibilities of using convolutional neural networks (CNNs) to monitor layer quality. CNNs are a subtype of neural networks designed for effective image processing. The history of CNNs dates to the 80s, when the first models were created [1]. Lately, as architecture became more sophisticated, hardware was still a limitation. The most significant breakthrough came in 2012 when a model called AlexNet [2] was introduced and succeeded in the ImageNet competition [3]. This came with using effective GPU CUDA cores instead of CPUs. These days, CNNs are widely used for computer vision tasks such as image classification [4,5], detection [6,7], and segmentation [8,9]. The basic principle is using a convolution layer, which performs edge, texture, shape, and more complex detections. Nowadays, a wide range of CNNs exists for specific applications in the computer vision field. The main contributions of the paper are as follows: • We proposed an end-to-end protocol for automatic assessment of the SLM quality layer based on convolutional neural networks. • We compared three different CNN models in two different testing protocols. • We published all our code at: https://github.com/AdamPolansky3/Using-machine-learning-in-for-quality control-additive-manufacturing

Fig. 1. Visualization of SLM process

2. Related Works There are different approaches to monitoring powder bed quality. In this part, related works are introduced. Chen et al. used EfficientNet-B7 [10] for image classification of powder bed images with and without defects, reaching 99.16% accuracy. In the second phase, several CNN models were trained for the segmentation of common defects with the best accuracy of 94.38 % [11]. Scime and Beuth divided the input image into so-called patches, which were used for training within a multi-scale CNN. This approach allows for performing defect segmentation; however, it’s limited by a predefined patch size [12]. Zhao and team used two CCD cameras and a laser projector to achieve a 3D map of the so-called point cloud. This was converted to a 2D image and processed for segmentation using a CNN. This proved to be reliable. However, the potential drawback of implementing this system is the requirement for structural adjustments to enable installation of the cameras and projector [13].