PSI - Issue 42

Satyajit Dey et al. / Procedia Structural Integrity 42 (2022) 943–951 / Structural Integrity Procedia 00 (2019) 000–000 Satyajit Dey et al

944

2

Fig. 1. Various types of manufacturing defects in additively manufactured 316L sample

of manufacturing defects, deterring the widespread use of AM in the safety-critical industries such as aerospace and nuclear ([2], [3]). Common manufacturing defects in AM include lack of fusion, porosity and microcracks, as shown in Figure 1. These defects originate from di ff erent manufacturing process related sources and have varied e ff ects on the properties of the material. To understand the root cause of these defects and those subsequent e ff ects on the mechanical properties of materials, it is important to accurately detect and classify these defects. In this work, we have developed a novel deep learning-based model capable of automatically detect and classify various types of defects in additively manufactured 316L stainless steel sample. The developed model will contribute to the development of in-situ defect monitoring system for AM as well as to the development of robust predictive models to correlate manufacturing defects with material deformation behavior.

Nomenclature

ANN Artificial Neural Network CNN Convolutional Neural Network loF lack of Fusion pores porosity

2. Image Segmentation using UNet

Image segmentation is a popular technique used for feature identification in an image [4]. In this technique, regions of an image are classified into di ff erent types of objects. This is achieved by classifying the pixels of each region into di ff erent types, thereby partitioning the image into several segments. Image segmentation techniques consist of two broad types:

• traditional computer vision approaches (threshold based [5], edge based [6], clustering based [7] etc.) • artificial neural network based approach [8]

In this work, we propose an ANN approach to detect defects in components fabricated using additive manufactur ing. ANN based methods are increasingly being used for image segmentation in the medical industry for diagnosis of medical images [9]. [10], [11] and [12] explored application of ANN based image segmentation methods for detecting weld defects while [13], [14] and [15] applied deep learning models for prediction of defects in additive manufactur ing. [16] provides a comprehensive review of application of deep learning in additive manufacturing. An image segmentation model is developed using a CNN architecture. The model is then trained using a set of images. The trained model can then be used to detect image features, which in this case are manufacturing defects. Figure 2 shows a schematic of the important steps involved in training the CNN model. Two datasets are created from the acquired images containing the defects: training dataset and validation dataset. The training dataset is then fed into the CNN model for training. The updated model parameters are then validated by using the validation dataset. Training of the model is continued until the expected accuracy or the maximum number of iterations is reached.