Feature Learning using Deep Variational Autoencoder for Prediction of Defects in Car Engine

Machine learning has been employed to automatically detect the defects on car engines in several studies. One of crucial challenges on applying machine learning is on the amount of defect data collected is often large with high dimensional data, making manual detection inefficient and inaccurate. The other problem is on the missing data as oftentimes the collected data are incomplete. In this paper, we employ machine learning frameworks for engine defect detection. It comprises the data pre-processing stage which includes imputing missing value data, then performing feature correlation using the Pearson method, and selecting the features to use. After that, the label encoder and standard scaler are carried out. The experimental process begins with creating a baseline, then continues with improving imbalance data using SMOTE, and feature reconstruction using variational autoencoder (VAE). Afterwards, for classification, we employ convolutional neural networks (CNN). The proposed method achieved precision 99.63%. We collect engine quality dataset of 224,239 data with 90 features from major automobile manufacturing in Indonesia. This showed that SMOTE and Variational Autoencoder dimensional reconstruction method can overcome defect predictions in car engine defect data with data imbalance conditions. This novel methodology distinguished our study from prior methods and shows considerable increases in precision and recall matrix.