DIMACS
DIMACS REU 2018

General Information

me
Student: Xinru Liu
Office: CoRE 450
School: Wheaton College(MA)
E-mail: liu(underscore)xinru(at)wheatoncollege(dot)edu
Project: Characterizing the Quality of 3D-Printed Parts using Spatiotemporal Data Analytics

Project Description

Additive manufacturing, or three-dimensional (3D) printing, is a promising technology that enables the direct fabrication of complex shapes and eliminates the waste associated with traditional manufacturing. A major issue that adversely affects its performance, and hence wider adoption, is that material solidification in the printing process yields geometric shape deviation. This study focuses on the analysis of 3D surface measurements of 3D-printed parts. Dimension, contour, and surface roughness are measured and represented in image data. We are developing methods for extracting spatiotemporal patterns from the measurement images and then correlating the patterns with part quality.

Weekly Log

Week 1:
This is my first week for DIMACS. On the first day, I met with my peer Alex and my mentor Dr. Guo. Dr. Guo introduced the backgroud of our project and showed us the Wide-Area 3D Measurement System to measure the 3D-printing object. I spent some time reading the manual and learning how to use the software. I spent most of the time reading research paper related to our topic. I also read some tutorials about the data-driven techniques that we might use later, like principal component analysis and tensor decomposition. We ended up with making slides for the next week presentation.
Week 2:
This is my second week in Rutgers. We did the presentation on Monday to introduce what our project is. We started to print more object using 3D-printer by changing input parameter, like layer thickness, speed and fill percentage. We found that the less the fill percentage, the better quality the object will be. We exported the data matrices from the point height measurements. One challenge that we met is there would be some missing data when we did the measurement due to the inconsistency of the system and we need to fill those data. We haven't found a good way to fix this problem, but we are thinking of some smoothing techniques, such as LOWESS smoothing. I found that collecting good data and cleaning data is crucial when you start a project. Only when the data is intepretable can we go on to the next step. I also read some papers and did some PCA(principal component analysis) on the matrics using R. I realized that vectorizing those matrices will lose the spacial correlation between each point. In the future we will try to do the multi-linear PCA on those matrices and intepret the result. This weekend I will also do some physical feature extraction, like fitting the data from the curve of the dome with the orignal reference, to see how quality changes with different input parameter. On Tuesday, we listened to a seminar from a speaker, Nina Fefferman. Her talk was about trans-disciplinary adventures in the mathematical biology of networks, which interested me a lot. Besides her intelligence and passion on her field, she also had very strong presenation skills to explain her research in a comprehensible way. I think I can learn a lot from her as a scientist.
Week 3:
This week we finished doing the non-linear least square between the curve of the object and the function curve of the original input parameter. The result verified our assumption that the less the fill of the object, the better it fits the function curve. We also output the PCA of the seven samples we gathered. The first two PCs explain most proportion of the variance, which are 70% and 15% respectively. After reconstructing the image from the two PCs, we are trying to find what features that each PC explains. We had a meeting with Dr. Guo on Thursday when we decided the model that we would build. The plan for next week will be 1) gathering the feacture of surface roughness and build a cluster with the curve fit residual that we gathered last week. 2) trying to build a model between the quality and input parameters with PCs. We had a great time on culture day(Friday). It is fun to learn different culture. I enjoy listening to stories of students from different backgrounds.
Week 4:
This week we first gathered some surface roughness of the object and then did the clustering based on the roughness and the profile curve residual standard error. Then we tried to build some regression model on the predict variables, like PCs, fill and speed. Alex found that when doing linear regression on profile curve residual and predict variables, includng PC4 will increase adjusted R^2 dramatically. I was trying to build logistic regression but have not finished. We met with Dr. Guo and we found that we need to redefine the quality of the object. Next week we will 1) collect more profile curves to make the data more comprehensive 2) build training, testing and validation dataset to make better prediction. This week we also had IBM trip. The company looked nice and some of the talks are interesting.
Week 5:
This week we stuck on accurately fit the profiles curve to a reference curve. We first want to manually fit but found it was too time consumming. Alex found a method called the dynamtic time warping which provided both a distance measure that is insensitive to local compression and stretched and the warping which optimally deforms one of the two input series onto the other. I used this method to output the distances between the reference curve the test curves, which was not bad. Alex was also trying the loess function to find the residuals. This week I also rewrote most of my r code to make it run multiple files more efficiently. Next week we will try to fit different models.
Week 6:
This week we started to fit different models. We ran the multinomial logistic regression, classification decision tree and support vector machine and used cross validation to evaluate the models. Given such a small sample size, the result of decision tree is surprisingly accurate. We will be continuing to print more samples to make our model more general and interpretable. I also started to make the outline for the presentation next week (can't believe the time past so fast!) We will meet Dr. Guo next Monday to discuss our result.
Week 7:
We fitted two multilinear regression models between the two response variables which evaluate the quality, profile residual and roughness, with other predict variables. The adjusted R-square of profile residual is pretty high, while R^2 of roughness is not ideal. We think that might because the places that we extracted roughness data were inconsistent with different samples. We finally ran the Uncorrelated Multilinear Principal Component Analysis(UMPCA). The regression model with UMPCA boosted the r^2. This week we also did the presentation. We will finish the paper in the next two weeks.
Week 8 & 9:
I didn't have much progress in week 8 since I was preparing GRE test which was on Thursday. We met Dr. Guo on Monday in the last week and decided to add another data source to our project to form a more unified picture of the 3D printing quality. We went to Rutgers makerspace and used 3D scanner, another 3D measurement system, to measure all our of our 3D parts. And then we exported the geometric data and implemented PCA on them. Since we only had few days left, we decided to continue our project after the program. We finished our paper and packed everything on the last day. Many thanks to DIMACS for everything they have done for me in the last two months.

References

[1] Zhao, Li-fang WuXiao-hua GuoLi-dong, and Meng Jian. “A Quality Evaluation Scheme to 3D Printing Objects Using Stereovision Measurement.” SpringerLink, Springer, Dordrecht, 13 Sept. 2017, link.springer.com/chapter/10.1007/978-3-319-71598-8_41.

[2] Khaleghi, Bahador, et al. “Multisensor Data Fusion: A Review of the State-of-the-Art.” Information Fusion, vol. 14, no. 1, 2013, pp. 28–44., doi:10.1016/j.inffus.2011.08.001.

[3] Khaleghi, Bahador, et al. “Multisensor Data Fusion: A Review of the State-of-the-Art.” Information Fusion, vol. 14, no. 1, 2013, pp. 28–44., doi:10.1016/j.inffus.2011.08.001.

[4] Lu, Haiping, et al. “Uncorrelated Multilinear Principal Component Analysis for Unsupervised Multilinear Subspace Learning.” IEEE Transactions on Neural Networks, vol. 20, no. 11, 2009, pp. 1820–1836., doi:10.1109/tnn.2009.2031144.

[5] Giorgino, Toni. “Computing and Visualizing Dynamic Time Warping Alignments InR: ThedtwPackage.” Journal of Statistical Software, vol. 31, no. 7, 2009, doi:10.18637/jss.v031.i07.

Presentations

Additional Information