Work

Machine Learning (ML) Based Prediction of Defects in Extrusion-Type Additively Manufactured Parts

Year: 2023
Type: conference-paper
Abstract: Additive manufacturing (AM) is gradually occupying a unique place as a viable industrial manufacturing technology for parts with complex geometry and difficult-to-machine materials. The capability of AM to apply to such parts in various sectors of industry, including automotive, aerospace, and medical devices, stems from the layer-by-layer building process and manufacturing directly from computer-aided design (CAD) models. All AM processes share a common set of steps and characteristics but are distinct in the type and form of raw material, energy source, and supply to the target location to build the layers. Based on these differences, the AM processes have been classified into different types: material extrusion-based processes, powder bed fusion (PBF) processes, direct energy deposition (DED) processes, and photo-polymerization processes. All the AM processes induce certain defects in the parts, making the final mechanical properties inferior. A capability to edict the mechanical properties of the parts by identifying defects in layers can help in designing parts and planning for AM to reduce those defects. The present work investigates the feasibility of applying a machine learning (ML) based technique to identify typical AM defects. The idea of transfer learning of a pre-trained network, namely, AlexNet, with a limited-size dataset of defect images generated from the layer images has been studied. A novel yet simple and effective experimental setup has been devised to capture high-resolution images of the layers of the part in a fused deposition modeling (FDM) machine. From these layer images, a total two hundred defect images have been captured and converted into an input image dataset for transfer learning. After suitable modification of the pre-trained network, the training and validation gave an accuracy of about 62%. Two different techniques of hyper-parameter tuning were then conducted, through which the accuracy of training improved to more than 95%. Based on the success achieved, further possible tracks of future research have been suggested.
Language: English
Open Access status: hybrid
OpenAlex ID: W4387106980
FWCI: 2.267
Cites: 11
Cited by: 7
Related to: 10