PERANCANGAN DATA PIPELINE UNTUK MENDUKUNG SISTEM PERAMALAN DI PABRIK CIKARANG PT X

• Main Author: C F Napitupulu, Rezki
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/80034
• Institution: Institut Teknologi Bandung

PT X is a snack food manufacturing holding company with various branches in Indonesia. PT X carried out a massive transformation in the Cikarang factory that had just been acquired, starting from organizational transformation to system transformation. The production department of PT X needs a system that can use raw data in the field to forecast OEE and downtime values in the future. OEE is a metric that assesses how much potential a machine has. Therefore, this research will build a data pipeline that can pull raw data from the field into a data table that can be used for analysis purposes, with the main metric being OEE. The construction of the pipeline is done with Python scripts using the ETL (Extraction, Transformation, Loading) approach. The extraction stage pulls data from 2 sources, namely the influxdb database and the machine. The transformation stage is carried out by compiling the existing variables in order to reduce the OEE value, then the loading stage analyzes the forecasting of the OEE value and downtime. To provide an example of the use of pipeline data, the processed pipelines data is used to forecast OEE and downtime values. ARIMA, XGBoost and LSTM models were built for OEE data, and a multivariate model was built with AutoTS for downtime data. The evaluation results of the three forecasting methods, namely ARIMA, XGBoost, and LSTM, show different performance in predicting the data. In terms of Mean Squared Error (MSE), ARIMA has a value of 0.1017, XGBoost with 0.0642, and LSTM 0.0709. Similarly, with Root Mean Squared Error (RMSE), where ARIMA has a value of 0.3190, XGBoost 0.2527, and LSTM 0.2119. Overall, the evaluation shows that XGBoost has lower MSE and RMSE values than ARIMA and LSTM, indicating better performance in predicting the data. In terms of Mean Absolute Error (MAE), XGBoost also performs best with a value of 0.1763, while ARIMA has a value of 0.2736 and LSTM 0.2662. The evaluation results of the blackbox method selected the best model is the Autoregressive Matrix Model (MAR).