발간년도 : [2023]
| 논문정보 |
|
| 논문명(한글) |
[Vol.18, No.1] Predicting CNC Processing Defects Using Nonsupervised Learning Algorithms |
|
| 논문투고자 |
Yong-Hee Han |
|
| 논문내용 |
Using various unsupervised learning algorithms (k-means, Gaussian mixture model, and autoencoder) that can be applied to unlabeled data, this study predicted whether CNC (computerized numerical control) workpiece is defective or not and compared accuracy, precision, recall, and F1 score. For k-means, only precision is shown to be the best with a slight difference, while accuracy and F1 scores are inferior to Gaussian mixture model and autoencoder algorithm. In addition, it was confirmed that the difference in F1 score and accuracy between Gaussian mixture model and autoencoder is relatively insignificant, while the execution time of neural network-based autoencoder is more than 17 times that of k-means and Gaussian mixture model. Gaussian mixture model's relatively good performance in metrics and short running time have practical advantages such as reducing the probability of CNC breakage and reducing the cost of building a real system due to rapid failure prediction, so the Gaussian mixture model model is most effective in predicting defects. In addition, visualization using t-SNE (t-distributed static neighbor embedding) algorithm confirmed that in this problem, k-means and autoencoder showed strength in distinguishing samples with large differences in attributes, while Gaussian mixture model showed strength in separating samples with relatively small differences in attributes. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
