Fused Deposition Modeling Source Identification Using Smartphone Photographs and Deep Learning

Additive manufacturing (AM) offers unique opportunities for agile and flexible manufacturing; however, there are concerns that AM supply chains may present risks in the form of inferior materials, poor process control, or counterfeit parts. There is a need for new technologies capable of tracking and validating AM parts and materials. This study demonstrates a method to predict the source of AM parts using images captured by a smartphone and analyzed using deep learning models. For this study, 918 parts were manufactured using industrial fused deposition modeling (FDM) machines from six contract manufacturers. The parts include nine different designs representative of common industrial parts with diverse geometric features. A smartphone camera collected high-resolution photographs of each part with consistent lighting. The smartphone images show minimal visual differences between parts that are produced by different suppliers, and it is not possible for a human to identify the manufacturing source by viewing the images or inspecting the parts. A deep machine learning (ML) model was trained to predict the supplier that produced each part. The part images are 4000 × 3000 pixels in size, which is typical for a smartphone but much higher than typical computer vision deep learning models that use images that are 224 × 224 pixels. A novel sampling strategy allows the high-resolution data to be interpreted with deep learning. The source identification model achieves 99.9% accuracy identifying the origin of 459 testing parts across the six suppliers and the nine part designs. The model is data driven, recognizing patterns from raw smartphone images, removing the need for manually designed or specialized features extracted with computation geometry. The model also demonstrates design generalizability and can extrapolate to make predictions for part designs never previously seen by the model. When the training dataset consists of five part designs, the model achieves 98.8% accuracy when applied to four previously unseen part designs. The source identification model is also data efficient, enabling the rapid addition of new manufacturers into the system with minimal training data. When trained using only 10 parts per manufacturer, the model achieves over 90% accuracy. Increasing the training set to 30 parts per manufacturer allows the model to exceed 95% accuracy. The model can authenticate the source of parts without the cooperation of the supply chain enabling anti-counterfeiting methods by detecting if parts originate from different sources or if the supplier has altered the production process. This research demonstrates the potential of source identification using simple mobile imaging to track the quality and authenticity of AM parts.