Page 81 - IJAMD-2-2

International Journal of AI for
Materials and Design Prediction of AM defect based on DL

Table 6. Data analytics based on the regular DNN after data normalization (min‑max normalization)
Structures (hidden layers) Algorithms FPR (%) FNR (%) ACC (%)
c(8, 4) rprop− 25.00 46.15 61.90
c(8, 5) rprop− 12.50 53.85 61.90
c(8, 6, 3) rprop+ or sag 12.50 53.85 61.90
Abbreviations: ACC: Accuracy; DNN: Deep neural network; FNR: False negative rate; FPR: False positive rate

Table 7. Data analytics based on the regular DNN after z‑score standardization
Structures (hidden layers) Algorithms FPR (%) FNR (%) ACC (%)
c(7, 4) slr 25.00 15.38 80.95
c(8, 4) slr 37.50 7.69 80.95
c(8, 5) rprop+ 12.50 23.08 80.95
c(8, 6) rprop- 25.00 23.08 76.19
c(8, 6, 3) rprop- or sag 0.00 61.54 61.90
Abbreviations: ACC: Accuracy; DNN: Deep neural network; FNR: False negative rate; FPR: False positive rate; slr: Smallest learning rate.

Table 8. Data analytics based on traditional ANN
Data pre‑processing Structures (number of nodes in the hidden layer) FPR (%) FNR (%) ACC (%)
Min-max normalization c(6) 0.00 53.85 66.67
c(10) 12.50 53.85 61.90
c(14) 0.00 53.85 66.67
z-score standardization c(6) 75.00 69.23 28.57
c(10) 12.50 38.46 71.43
c(14) 25.00 53.85 57.14
Abbreviations: ACC: Accuracy; ANN: Artificial neural network; FNR: False negative rate; FPR: False positive rate.

6. Conclusion and future research Acknowledgments

Data analytics based on four DL techniques (the Elman The author thanks the support from Mississippi State
neural network, the Jordan neural network, the DNN- University, Mississippi, USA.
DBN, and regular DNN based on the four algorithms
(“rprop+”, “rprop-“, “sag”, and “slr”) were conducted Funding
on a small dataset with unbalanced data of the LPBF. None.
After z-score standardization was employed, the regular
DNN could obtain a better ACC than the three other DL Conflict of interest
techniques (Elman, Jordan, and the DNN-DBN), while The author declares no conflicts of interest.
most of the relevant FPR and FNR values were much
lower than those of the three DL techniques. Future Author contributions
research should consider DL-based data analytics and
defect prediction of other defects (e.g., keyholes and This is a single-authored article.
cracks) while considering the material properties of Ethics approval and consent to participate
LPBF. This will improve the DL model’s applicability and
enable a more comprehensive defect classification in LPBF Not applicable.
processes. DL-based data analytics on more AM datasets
(small or big) and the prediction of related defects are also Consent for publication
directions for future work. Not applicable.

Volume 2 Issue 2 (2025) 75 doi: 10.36922/IJAMD025060005