Pemodelan Predictive Maintenance Berbasis Machine Learning dengan Long Short-Term Memory untuk Prediksi Remaining Useful Life Gas Compressor Menggunakan Variabel Proses Pressure

Nugraha, Muhammad Rifqy Aji (2026) Pemodelan Predictive Maintenance Berbasis Machine Learning dengan Long Short-Term Memory untuk Prediksi Remaining Useful Life Gas Compressor Menggunakan Variabel Proses Pressure. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009221097-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Gas compressor adalah peralatan penting di industri minyak dan gas. Terjadinya kegagalan, mengakibatkan shutdown mendadak dan kerugian operasional besar. Penelitian ini bertujuan memprediksi Remaining Useful Life (RUL) pada reciprocating compressor K-101-A dengan metode direct forecasting berbasis Long Short-Term Memory (LSTM). Variabel proses yang digunakan adalah tekanan dari empat sensor: engine jacket water pressure, engine oil pressure, compressor jacket water pressure, dan compressor lube oil pressure. Penelitian ini dilakukan melalui beberapa tahap, mulai dari identifikasi data proses dan maintenance, imputasi missing value dengan PCHIP, segmentasi run to failure (RTF), analisis statistik deskriptif, dan analisis korelasi Pearson. Pembentukan Health Indicator (HI) berbasis PCA, penyusunan dataset untuk direct RUL forecasting, hyperparameter tuning, pelatihan model HI LSTM, analisis PDF dan Confidence Interval (CI), cut testing, serta pembuatan dashboard monitoring RUL berbasis MATLAB. Hasil PCA menunjukkan bahwa komponen utama pertama (PC1) mewakili 97,35% variasi data sensor dan digunakan sebagai dasar pembentukan HI. Model terbaik ditemukan melalui grid search dengan konfigurasi M15 (window 4, dropout 0,30, 32 hidden units), menghasilkan RMSE 4,5216 jam, MAE 3,8112 jam, dan MSE 20,4452 jam². Analisis CI 95% menunjukkan coverage 92,157%, dan cut testing pada RTF 4 menghasilkan absolute error 3,622–4,091 jam. Penelitian ini menunjukkan bahwa kombinasi HI berbasis PCA dan model direct HI-LSTM dapat menjadi dasar predictive maintenance berbasis machine learning pada gas compressor, serta mendukung SDG 9 tentang industri, inovasi, dan infrastruktur.
==================================================================================================================================
Gas compressors play a vital role in the oil and gas industry, and their failures can lead to unexpected shutdowns and major operational losses. In this study, we focus on predicting the Remaining Useful Life (RUL) of reciprocating compressor K-101-A. We use a direct forecasting method based on Long Short-Term Memory (LSTM) and rely on pressure data from four sensors: Engine Jacket Water Pressure, Engine Oil Pressure, Compressor Jacket Water Pressure, and Compressor Lube Oil Pressure. The research process covers several hours: identifying process and maintenance data, filling in missing values with PCHIP, segmenting data by Run to failure (RTF), performing descriptive statistics and Pearson correlation analysis, building a Health Indicator (HI) using PCA, preparing the dataset for direct RUL forecasting, tuning model parameters, training the HI-LSTM model, analyzing probability distributions and confidence intervals, conducting cut tests, and creating a MATLAB-based dashboard for RUL monitoring. Results show that the first principal component (PC1) captures 97.35% of the total sensor data variance, making it the basis for HI construction. The best model was obtained through grid search with configuration M15 (window 4, dropout 0.30, 32 hidden units), achieving an RMSE of 4.5216 hours, MAE of 3.8112 hours, and MSE of 20.4452 hours². The 95% CI analysis yielded a coverage of 92.157%, while cut testing on RTF 4 produced an absolute error of 3.622–4.091 hours. This study demonstrates that the combination of PCA based HI and a direct HI-LSTM model can serve as a foundation for machine learning-based predictive maintenance on gas compressors, while also supporting SDG 9 on industry, innovation, and infrastructure.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Gas Compressor, Machine Learning, Predictive Maintenance, Pressure, Remaining Useful Life
Subjects:	Q Science Q Science > Q Science (General) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Industrial Technology > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Rifqy Aji Nugraha
Date Deposited:	30 Jul 2026 07:19
Last Modified:	30 Jul 2026 07:19
URI:	http://repository.its.ac.id/id/eprint/139860

Actions (login required)

View Item