Article


SHARE:  

Machine Learning: A Game Changer for AM Quality Assurance

By: Darren Beckett

Darren Beckett is chief technology officer, Sigma Labs Inc., Santa Fe, NM; www.sigmalabsinc.com.

Wednesday, November 11, 2020
 

Each of us has a preconception of what the term “machine learning” means. According to ExpertSystems.com, “Machine learning (ML) is an application of artificial intelligence (AI) that provides systems the ability to automatically learn and improve from experience without being explicitly programmed.”

Note where the definition mentions learning without being “explicitly” programmed. This makes it sound like the “rise of the machines” from the “Terminator” movies.

Machine learning implies that systems can learn from data, identify patterns and make decisions with minimal human intervention. ML focuses on the development of computer programs that can access data and then allow machines to use that data and learn for themselves. 

Machine learning in any environment, including metal AM, depends on gathering quality data and then using the right system to validate that data. Like most other processes, the old expression “garbage in garbage out” applies.

Does ML function in the real-world of AM? The answer is a resounding yes. As noted in a recent 3DPrinting.com report, “Porosity and other defects are a problem where it comes to parts printed with metal powder-bed fusion processes. One team of researchers from the U.S. Department of Energy’s Argonne National Laboratory and Texas A&M University has published its findings that may play a part in reducing these subsurface defects. The team has figured out a novel way of predicting the formation of subsurface porosity as well as measuring the temperature of a region at the very moment of printing of Ti-6Al-4V powder.

“Specifically, the researchers were interested in predicting the formation of keyhole pores,” the online article continues, “which are formed when a surplus of energy is concentrated in the metal melt pool for too long. These pores can act as stress concentrators in the final part’s structure, leading to part unpredictability under load, and potential failures. By use of a top-down high-speed thermal camera in the build area of the laser fusion printer they are able to log the relative temperature at a certain time during printing, and correlate that with additional scans from synchrotron x-ray imaging taken at the exact same moment. This helps to build a picture of how the thermal history of the fused part affects the microstructure of the printed metal.”

The Part/Process Quality Ecosystem

Think of the end-to-end part/process quality decision methodology as an ecosystem that functions best if it’s possible to distinguish between part and process quality. Effective AM-ML models are designed to recognize anomalies in specific parts, and also to monitor the quality of the process itself. This enables quality and consistent-improvement opportunities on AM parts. 

Maintaining the process of examining parts, training models and monitoring new builds forms a cyclic quality-decision ecosystem that contains three components:

  • The training process selects model types from the warehouse, trains them to recognize data patterns in the training builds and saves the trained models back to the warehouse. 
  • The QA process selects trained models appropriate to the material and the design of test builds, in order to evaluate ongoing process quality. 
  • When new designs and materials are used, the process cycles back to improve the ecosystem. Here we begin to outline the training process. 

The accompanying graphic illustrates an example of the part/process quality ecosystem. The AM system monitors the builds used during training and produces in-process quality metrics (IPQMs) for use as features of ML models. Parts of the builds are selected for nondestructive testing in the form of post-process computed tomography to digitally capture part porosity. Porosity geometric-location data from the analysis are registered and aligned with the process metrics, and the tuning and training process then uses the metrics as features in ML models. It also uses the anomaly data from the process analysis as labels for supervised learning. 

The training process is monitored by assessing the diagnostic accuracy with visualization and numerical scores. This information helps to further tune the models and their training parameters and complete the description of the training process, starting with the monitoring of builds and the detection of anomalies by physical analysis of the parts. The system then moves from the training process to the QA process, monitoring production builds and producing IPQMs that will be used as features of the trained models. The ML-QA process feeds the metrics to trained models in order to evaluate build quality, and visualization of the predictions is used to identify the location of potential anomalies and predict real defects.

The Part-Quality Decision Dashboard

Several tools in the part-quality decision dashboard help manufacturers evaluate the severity of the potential anomalies, while the 3D visualization provides overall patterns and an SPC chart allows for data-trend analysis. This completes the description of the QA process starting with the monitoring of builds, the prediction of potential anomaly locations, and combining this information to evaluate the severity of the potential anomalies. 

The QA process leverages trained models, selected from the model warehouse, and production metrics to predict potential anomalies. As such, the components of the part/process quality-decision ecosystem work together so that when new designs and materials are used, the process cycles back to improve the ecosystem. 3DMP

 

See also: Sigma Labs Inc.


Reader Comments

There are no comments posted at this time.

 

Post a Comment

* Indicates field is required.

YOUR COMMENTS * (You may use html to format)

 
YOUR NAME *
EMAIL *
WEBSITE