Predict the Disposition of Failed Units after Test in Manufacturing Line TEAM 7 - MOHD IKHWAN * MUSLIHUDDIN

INTRODUCTION

Manufacturing Production line process start point is from Assembly of product, then to Product testing and lastly it will go to Packing Process. Our project will be focus on testing where Product Engineering will involve more. There are 3 test stage which is Module Setup Test, Temperature Compensation Test and Functional Test.

This project will be focus more on Functional Test. Production Modules will go through testing for 4 hours and module label as PASS or FAIL after complete test. Those raw data parameter resul will be stored in databases. PASS unit will proceed to next stage PACKING but those FAIL module will be put ON-HOLD under Engineering for troubleshoot/ or Debugging.

PROBLEM STATEMENT

Assembled products will be tested as part of the manufacturing process, and each one will be labelled PASS or FAIL. There will be a database entry for this information. Technician will examine the data and provide disposition for the module, such as rework or retest. Due to human error, sometimes technician might provide wrong disposition, thus will result in extra cost and extra time required before the module can have correct disposition to PASS the test.

PROJECT AIM

To build reliable predictive model to be able to predict accurately the disposition of the failed units such as Retest, Replace TOSA, Replace PCBA, Run PCBA visual inspection, Run Tech access (Rel3) etc of Product Code TL5400ZFN at Functional Test without any technician intervention and will reduce human error in providing FAIL module disposition.

DATASET DESCRIPTION

Dataset was obtain from the Production Line database cloud and it contain 2078 unique Module Serial Number with 35 columns of testing parameters. The Target Column named 'Disposition' contain 4 main class which is 'Retest from BH', Retest from MTC', Retest from Setup' and "Rework Change Component'. Refer example Figure 1.

Table 1: Example of dataset

Dataset have 35 numerical features and 1 categorical output. The output label need to be encoded to numerical value in order to be feed into Machine Learning algorithm

DATA CLEANING

Before proceed to do modelling, we did data cleaning on raw dataset. The action as per below:

Removing missing data using 'dropna()' function
The Output label is in string, so encoded it to numerical value using LabelEncoder function from Scikit-Learn library.

MODELLING

Use 3 classifier model, Decision Tree, Support Vector Machine and Artificial Neural Network algorithms. We are using K-Cross validation K=10 in this model. The ratio of the split is 70% training set and 30% test set. All 35 features included in this model. Below are the classification report of each Classifier model.

Decision Tree (DT)

Fig 1: Decision Tree Classification report

Artificial Neural Network (ANN)

Fig 2: Artificial Neural Network Classification Report

Support Vector Machine (SVM)

Fig 3: Support Vector Machine Classification Report

Based from 3 model above, SVM showing highest model accuracy 71.40%, followed by Decision Tree 71.05% and lastly Neural Network 69.39%. SVM will be selected since it have higher accuracy score.

ROC curve comparison between 2 classifier

Fig 4: ROC curve - comparison between DT and SVM

MODELLING (features selection)

In order to make more comparison, we did a feature selection on the dataset. Method such as Univariate Feature Selection (SelectKBest) and Recursive Feature Elimination (RFE) was used and the accuracy score summarize in Table 2 as below:

After applying SelectKBest and RFE method, only 15 features were selected out of 35 features in total.

Table 2: SelectKBest features ranking score

Fig 5: RFE feature selection method

Table 3: Comparison on accuracy score for All Features (35), Univariate Features Selection and Recursive Feature Elimination

From Table 2, Decision Tree Accuracy score highest 71.05% when we not applying any features selection method, and accuracy is significantly reduce when apply Select Kbest method. On the other hand, Neural Network model accuracy score has no significant difference between All features and features selection method. Support Vector Machine showing no change in Accuracy Score when applying both (Kbest and RFE) features selection method.

CONCLUSION

Overall, Model which have Highest accuracy score is Support Vector Machine (SVM). We will use SVM as our project model since it has highest accuracy score 72.10%.

Appreciate

Credits:

Created with images by 安琦王 - "Automobile manufacturing" • A Stockphoto - "Close up of the hand men hold tool repairs electronics manufacturing Services,Repair of electronic devices, tin soldering parts."