General Information

---

Project Description

During the 3D printing process, a number of defects may arise, such as porosity (void space in the bulk of the material), spatial deformation, and dimensional inaccuracy. While there exist several techniques to detect anomalies such as these after a part has been printed, it would be optimal to discover these anomalies as soon as possible during the printing process, so that the printing process can be dynamically corrected to alleviate these problems and so that the model may be updated to improve future prints.

---

Weekly Log

Week 1:

The majority of this week was spent getting acclimated with the program, as well as understanding the problem. I studied papers in review and past research on the topic, and analyzed some of the existing mathematical models for interpolating the temperature information into useful data.

Week 2:

This week, I searched for ways to implement a neural network to characterize the porosity of the printed part. The main challenge was a lack of training data. We only had the final piecewise characterization of a single printed part to evaluate thousands of matrices worth of data. My current idea is to assign porosity values to each layer and use the layers as training data, or to assign values to each sequential volume as it is constructed. Once the training data is constructed, we will construct a neural network and begin to develop a characterization process.

Week 3:

This week was spent developing the actual neural net and deep learning architectures that will be used to generate our model. We discussed ways of dealing with problems, such as asynchronicity and data incompleteness, that may arise during training. I am currently constructing the actual neural net and trying to get reasonable accuracy on the training data.

Week 4:

The IR neural net is very close to being done and shows promising accuracy. The net applies several recurrent convolutions to sequences of IR images to obtain a prediction of part quality. Currently, the net produces ~86% accurate predictions, which is less accurate than statistical modeling with interpolants, but running the neural net on an input image sequence offers a significant time savings over computing these interpolants.

Week 5:

The pyrometer neural net is under construction. The net is similar in structure to the IR neural net, but it does not incorporate time dependency because the pyrometer is stationary with respect to the melt pool's frame of reference and so time variation may be attributed to noise. The net is not yet refined, but it is close to operating with similar accuracy to the IR net. Once both nets are constructed, I will write a program to cross-reference information from nearby timestamps to improve the overall accuracy.

Week 6:

The pyrometer neural net (named PyroNet) is done, and is significantly more accurate than the IR neural net (IRNet) at 94%. The breakthrough that improved the accuracy of the net - and which will be used on IRNet to improve its accuracy - is using a standard stochastic gradient descent optimizer rather than the typical Adam optimizer, in order to discover better local minima instead of becoming "stuck" at one. I also employed image augmentation to better represent failures in the dataset, by providing 700 augmented failures. These techniques significantly improved the accuracy of the network.

Week 7:

IRNet has been revised with the new techniques, and now exhibits almost 90% accuracy on the training data. The next step is to tighten up the performance of IRNet via hyperparameter tuning, and then to combine IRNet and PyroNet results in an adaptive-weighted decision-level fusion process.

Week 8:

Both neural nets have been tweaked slightly. PyroNet was trained on a wider variety of augmented data, which lowered testing accuracy to 92.7% but increased the range of data which can be classified with this accuracy. IRNet underwent various tunings and achieved a final accuracy of just over 90%. I also began writing a program to combine the outputs of both models in one weighted decision.

Week 9:

The final program (Titanet) has been compiled. It uses an adaptive decision-level fusion process as described by Osadciw and Veeramachaneni (2016) to synthesize the results of the neural networks when given an input, and come up with a final answer. With this, I had enough to write the final paper describing the overall structure and capabilities of the finished program.

---

Presentations

• First Presentation
• Second Presentation

---

Additional Information

• My Mentor
  • Dr. Weihong (Grace) Guo

---

Special Thanks

This project was funded by the National Science Foundation, grant number CCF-1852215.

Special thanks to Adriana Scanteianu, Max Wang, and Spencer Alara Yaculak for their support and assistance.