Author: John W.S. Lee
In this study, efforts were made to classify defects in parts produced by injection molding processes. Three different modeling approaches, namely supervised learning models, Mahalanobis Distance model, and Variational AutoEncoder model, were implemented and their performances were compared.
The dataset used in this study was downloaded from the Korea AI Manufacturing Platform, also known as KAMP. Although the website is written in Korean, the contents of the dataset were mostly written in English. The dataset consisted of 7,996 rows with 44 columns. One of the columns represented the target label, PassOrFail.
The following is a summary of this study. For more detailed codes and notebooks used in this study, please refer to the notebook folder.
The dataset had 4 different injection-molded parts, namely CN7 and RG3, each with Left-Hand and Right-Hand components. The figure below shows the distribution of processing parameters for the parts with 4 different combinations. As shown in the figure, the processing parameters for CN7 and RG3 exhibited very different distribution, whereas the difference between the Left-Hand and Right-Hand components were not big. Therefore, it was reasonable to proceed with two separate models for CN7 and RG3.
For each type of injection-molded parts, the distributions of the processing parameters were compared for passed parts(i.e. good parts) and failed parts (i.e. defective parts). In the case of CN7, there seemed to be some difference in the distributions of the processing parameters for passed/failed parts . However, the difference for RG3 seems to be less obvious than that for CN7.
As mentioned above, 3 different machine learning approaches were implemented for the purpose of classifying the defective injection-molded parts. Detailed codes can be found in the notebook folder. Since there was a significant class imbalances, f1-score was used as the evaluation metric.
For CN7, the f1-scores for supervised learning models, Mahalanobis Distance model, and Variational AutoEncoder model were 0.67, 0.55, and 0.73, respectively.
For RG3, the f1-scores for supervised learning models, Mahalanobis Distance model, and Variational AutoEncoder model were 0, 0.3, and 0.24, respectively.
Clearly, the effectiveness of the models were different for CN7 and RG3 parts. Especially, it was surprising that the f1-score could be improved from 0 to 0.27 by switching from supervised learning models to Mahalanobis Distance model. It should be also noted that the choice of the thresholds for Mahalanobis Distance model and Variational AutoEncoder model played a significant roled in determining their performances.
Feature importances for CN7 parts were checked on 3 models (i.e, SVC, RandomForest, and LightGBM) using the models' built-in function and shap library.
The importance of each feature appeared to be slightly different based on the models and the methods used. It turned out that "Max Injection Speed", "Filling Time", "Mold Temperature 4", "Barrel Temperature 1", and "Plasticizing Position" were the processing parameters that models thought to be important.
This study demonstrated how various machine learning approaches performed in the classification of defective parts in injection molding. It turned out that the performance of each approach varied based on the type of datasets, CN7 and RG3 for this study. For CN7 parts, Variational AutoEncoder performed best, whereas Mahalanobis Distance model performed best for RG3. This suggests that it is important to try several machine learning approaches to find the best-performing approach for a given data.
To download the contents of this GitHub page on to your local machine, follow these steps:
-
Copy and paste the following link:
git clone https://github.com/johnwslee/injection_molding_analysis.gitto your Terminal. -
On your terminal, type:
cd injection_molding_analysis. -
Create a virtualenv by typing:
conda env create -f env.yml -
Activate the virtualenv by typing:
conda activate inj_env -
Run the notebooks in notebook folder in order.