Abstract
Additive manufacturing, a layer-by-layer method to build up parts, has gained serious attention lately; however, there are part qualification and certification bottlenecks due to various defects observed in printed parts. To overcome this issue, deep learning, a fast-growing technology, is leveraged to detect anomalies automatically. Due to the nature of deep learning algorithms, which require thousands of samples to be trained, transfer learning is offered. In this study, laser melting of 316L stainless steel to provide proof-of-concept data and training for transfer learning, with the goal to scale up the results of laser melting to proper metal laser-based additive manufacturing technologies like laser powder bed fusion and directed energy deposition. In this study, four pre-trained state-of-the-art algorithms on ImageNet are employed to be trained for additively manufactured part defect detection while having a limited dataset. The results demonstrated that pre-trained EfficientNetB7 and InceptionV3 algorithms showed promising results, successfully classifying the test data with an accuracy of more than 94%.
