In-situ surface porosity prediction in DED (directed energy deposition) printed SS316L parts using multimodal sensor fusion

Overview

• This study aims to relate the time-frequency patterns of acoustic emission (AE) and other multi-modal sensor data to the formation of pores in a hybrid directed energy deposition (DED) process.
• The researchers used an explainable AI method called LIME (Local Interpretable Model-Agnostic Explanations) to attribute certain high-frequency AE waveform signatures to two main causes of pore formation: spatter events and insufficient fusion between printing tracks.
• This approach allows real-time prediction of pore presence in every printed voxel (0.5 mm cube), a significant advance over previous efforts.

Plain English Explanation

The researchers wanted to understand how the sounds and other sensor data collected during a 3D printing process called directed energy deposition (DED) could be used to detect the presence of tiny pockets of empty space, or pores, in the printed metal parts. DED involves using a high-energy laser to melt and fuse metal powder into a desired shape.

By analyzing the high-frequency sound waves produced during the printing process, the researchers were able to identify two main causes of pore formation: 1) spatter, where small droplets of molten metal are ejected from the melt pool, and 2) insufficient fusion between adjacent printing tracks due to low heat input. The researchers used an AI technique called LIME to "explain" how these specific sound wave patterns were linked to the presence of pores.

This is an important advance because it allows the detection of pores in real-time as the part is being printed, rather than having to wait until the part is finished and inspected. Being able to predict pore formation in this way could help manufacturers make better parts with fewer defects.

Technical Explanation

The researchers collected synchronized data from multiple sensors, including force, acoustic emission (AE), vibration, and temperature, while printing and machining an SS316L metal sample. They then used a deep convolutional neural network to analyze the time-frequency patterns (spectrograms) of the sensor data and identify the presence of pores on the surface of the printed voxels (0.5 mm cubes).

The results showed that the sensor data collected during the DED printing process was more sensitive for detecting porosity compared to the data collected during the subsequent machining process. The classification test accuracy was 87% for the DED data.

Through the LIME analysis, the researchers were able to attribute certain high-frequency AE waveform signatures to the two main pathways for pore formation. For porous voxels with prevalent spatter events, the high-frequency AE band had about 27% higher energy content compared to other porous voxels. This suggests that the spatter events are linked to increased pore formation. Conversely, porous voxels with insufficient fusion between adjacent printing tracks had about 33% lower energy captured in the high-frequency AE waveforms, indicating a relatively lower laser-material interaction in the melt pool and hence poor overlap between tracks.

This approach enables real-time prediction of pore presence in every printed voxel, a significant advancement compared to previous efforts that could only assess porosity after the part was completed.

Critical Analysis

The researchers acknowledge that their approach is limited to detecting pores on the surface of the printed part, and does not capture internal pores. Additionally, the study was conducted on a single material (SS316L), and the findings may not generalize to other metal alloys used in DED processes.

Further research is needed to expand the application of this technique to different materials and part geometries, as well as to improve the accuracy and robustness of the pore detection. Integrating the real-time pore predictions with digital twin models of the DED process could also lead to significant advancements in process monitoring and control.

Conclusion

This study demonstrates the potential of using multi-modal sensor data, particularly acoustic emission signals, to detect and understand the formation of pores in a hybrid DED metal 3D printing process. The ability to predict pore presence in real-time as the part is being printed represents a significant advancement that could help manufacturers improve the quality and consistency of their 3D printed metal components.