Defect Detection for Additive Manufacturing with Machine Learning and Markov Decision Process

Additive Manufacturing (AM) is a quickly evolving manufacturing technique in recent years. One of the most essential steps is the quality control of it. This involves the defect detection of the products, which is one of the bottlenecks that affects the high quality of AM products. One promising solution to this problem is to detect the defects in-situ and make decisions on the fly. We adopted Machine Learning (ML) algorithms for defect detection and develop a Markov Decision Process (MDP) model to make decisions for AM process. Our main purpose is to save costs and time through early termination or parameter adjustment of the printing process.

In chapter 1, we developed a scheme based on ML models trained. Then these models are applied to detect defects in actual production. It will save training time and costs associated with many prints for each design by using synthetic 3D point clouds rather than experimental data. Besides, a new concept called “patch” to capture macro-level information about nearby points for ML training and implementation is introduced here. Numerical comparisons of prediction results on experimental data with different shapes showed that the proposed ML-based scheme outperformed the existing Z-difference method in the literature.

In chapter 2, we introduced a new curvature feature based on the models in chapter 1, i.e., Discrete Mean Curvature Measure, to capture macro-level information beyond the distances and incorporated it into the training data for ML algorithms. This new curvature feature was demonstrated to well improve the defect detection performance (the F-measure: as high as 94%).

In chapter 3, we proposed an MDP model serving as the feedback mechanism to form a whole closed-loop in-situ monitoring system for AM process. We defined the general framework of the MDP model for this purpose. Then we use Value Iteration method to solve a series of cases of the model. Some numerical analysis of the solution is also implemented to conclude some interesting phenomena in the process of solving.

In the future, it can be planned to collect experimental data to validate if the method is an applicable monitoring system for real production.