Comparative additive manufacturing defect prediction accuracy with a few transfer learning implementations of deep learning models

This paper addresses the problem of a comprehensive quality assurance strategy for additively manufactured components with integrated in-situ inspection and artificial intelligence and machine learning (AIML) models. A custom test setup was created around a fused filament fabrication (FFF) 3D printing apparatus, incorporating sensors to capture real-time data concerning part quality. This image data was harnessed to formulate an AI/ML dataset for training a Convolutional Neural Network (CNN). A robust framework for predicting defects in real-time during the additive manufacturing process and validating the accuracy of AIML predictions has been presented. A test setup, generating a varied dataset, crafting AI/ML models, and optimizing the AI/ML model for precise defect prediction has been made. The proposed methodology’s practical applicability and potential to redefine quality assurance in additively manufactured parts have been presented. Results were compared between the Matlab AlexNet pre- trained model and the user-designed model on Google Colab, which has the capability of hyperparameter tuning. Performance parameters, including accuracy, loss, precision, and recall, were plotted over epochs to analyze the model’s merit after training. With the addition of hyperparameter tuning to the AlexNet, the best model was chosen as per the training accuracy. Much better accuracies were observed in the earlier epochs, with the initial loss being highly reduced compared to the non-hyperparameter-tuned model. The ResNet50 model, which did not have hyperparameter tuning capabilities, produced less accurate results. On the other hand, the loss was much lower with ResNet50 compared to both the AlexNet Matlab model and the AlexNet model on Google Colab without hyperparameter tuning. ResNet50, at the same time, gave accuracy comparable to that of the AlexNet Model with hyper-parameter tuning.