Defect Detection in Metal Additive Manufacturing

In-process inspection offers a path forward to insuring components produced via metal additive manufacturing (AM) such as directed energy deposition (DED). However, a simple anomaly detection ML technique driven from observational data may be insufficient for tracking the performance of a system intended to produce arbitrary geometry. The methodology chosen for a defect monitoring system must be given information about the intent of the machine’s current activity (e.g. by being fed data from the CAM system in addition to process monitoring data) in order to sufficiently judge the state of the system. The key innovation proposed for this project is the ability to cleanly automate the entire specimen preparation and analysis process.  For the success of machine learning tasks, much attention must be paid to training a joint embedding of the different data types in latent space that adequately represent the combined feature space. As preliminary work, the team developed an AM state prediction deep learning (DL) model that was applied to AM anomaly detection.   With PINE’s Hybrid Additive Manufacturing Repair (HAMR) we can unify the specimen creation, data collection, and model training in one unified system. By using ML techniques, we can develop a real-time defect detection method that could inform AM engineers to intervene and prevent further defects from happening, reducing waste of time, energy and resources.