Prediksi Remaining Useful Life (RUL) Gas Engine Pada Kompresor Gas Berbasis Deep Learning Dengan Data Temperatur

Alauddin, Daffa (2026) Prediksi Remaining Useful Life (RUL) Gas Engine Pada Kompresor Gas Berbasis Deep Learning Dengan Data Temperatur. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5009221071-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Gas engine merupakan salah satu peralatan kritis dalam industri minyak dan gas, sehingga prediksi remaining useful life (RUL) menggunakan deep learning sebelum terjadinya unplanned downtime sangat penting untuk menjaga ketersediaan operasional. Mayoritas studi yang ada saat ini masih bergantung pada data getaran dari rig simulasi ideal, sementara pemanfaatan data temperatur dari lapangan nyata masih terbatas. Penelitian ini bertujuan untuk mengevaluasi performa prediksi RUL menggunakan mean absolute error (MAE) dan root mean square error (RMSE), sekaligus menganalisis pengaruh kondisi outlier dan variasi kombinasi sensor terhadap hasil prediksi. Pendekatan data driven berbasis long short-term memory (LSTM) digunakan dalam penelitian ini, dengan principal component analysis (PCA) untuk membentuk health indicator (HI). Tiga variasi pengelompokan data sensor didefinisikan: Combustion (lima data sensor-sensor temperatur exhaust silinder), Systemic (data sensor temperatur jacket water dan sensor temperatur exhaust manifold bank kanan dan kiri), serta Global (gabungan sensor Combustion dan Systemic), yang masing-masing dievaluasi pada dua kondisi data, dengan dan tanpa pembersihan outlier. Hasil menunjukkan bahwa variasi Combustion tanpa pembersihan outlier mencapai performa terbaik dengan RMSE 42,89 jam dan MAE 35,27 jam dengan prediksi RUL menunjukkan end of life (EoL) pada jam ke-320 pada posisi 50% data dan jam ke-283 pada posisi 75% dengan EoL aktualnya adalah jam ke-280. Variasi ini juga menunjukkan kemampuan generalisasi pada kondisi operasi yang berbeda, dengan RMSE 24,00 jam dan MAE 20,25 jam setelah fine-tuning dengan prediksi RUL menunjukkan EoL pada jam ke 146 pada posisi 50% data dan jam ke 184 pada posisi 75% dengan EoL aktual berada pada jam ke 160. Penelitian ini menyimpulkan bahwa pengelompokan sensor dalam satu sistem yang terkorelasi kuat menghasilkan error prediksi paling rendah dan mampu mempertahankan performanya pada kondisi operasi berbeda. Pembersihan outlier tidak selalu menguntungkan karena metode ini dapat menurunkan error pada fusi sensor antar sistem dengan korelasi rendah dan jumlah sensor yang banyak namun dapat menyebabkan kegagalan model LSTM pada pengelompokan sensor dalam sistem tunggal.
=======================================================================================================================================
Gas engine is one of the critical equipment in the oil and gas industry, making remaining useful life (RUL) prediction using deep learning essential to prevent unplanned downtime and ensure operational availability. The majority of existing studies rely on vibration data from ideal simulation rigs, while the utilization of real field temperature data remains scarce. This study aims to evaluate RUL prediction performance using mean absolute error (MAE) and root mean square error (RMSE), while also investigating the effect of outlier conditions and sensor combination variations on prediction results. A data driven approach based on long short-term memory (LSTM) was employed, with principal component analysis (PCA) applied to construct the health indicator (HI). Three sensor groupings were defined: Combustion (five exhaust cylinder temperature sensors), Systemic (jacket water temperature and exhaust manifold temperature for both right and left banks), and Global (a combination of Combustion and Systemic sensors), each evaluated under two data conditions: with and without outlier removal. The results indicate that the Combustion variant without outlier removal achieved the best performance, with an RMSE of 42.89 hours and an MAE of 35.27 hours, RUL predictions indicated an End of Life (EoL) at 320 hours (using 50% of the data) and 283 hours (using 75% of the data), compared to the actual EoL of 280 hours. This variant also demonstrated generalization capability under different operating conditions, yielding an RMSE of 24.00 hours and an MAE of 20.25 hours after fine-tuning with RUL predictions indicating an EoL at 146 hours (50% data) and 184 hours (75% data), against an actual EoL of 160 hours. This study concludes that sensor groupings within a single, strongly correlated system yield the lowest prediction error and maintain performance across different operating conditions. Outlier removal is not universally beneficial. It reduces error in low-correlation, high-sensor-count fusion scenarios but can cause LSTM model failure when applied to single-system sensor groups.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deep Learning, Gas Engine, Machine Learning, Remaining Useful Life, Deep Learning, Gas Engine, Machine Learning, Remaining Useful Life
Subjects:	Q Science > QA Mathematics > QA336 Artificial Intelligence T Technology > T Technology (General) > T57.5 Data Processing T Technology > TJ Mechanical engineering and machinery > TJ762.E93 Exhaust systems T Technology > TJ Mechanical engineering and machinery > TJ785 Internal combustion engines. Spark ignition T Technology > TS Manufactures > TS174 Maintainability (Engineering) . Reliability (Engineering)
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Physics Engineering > 30201-(S1) Undergraduate Thesis
Depositing User:	Daffa' Alauddin
Date Deposited:	31 Jul 2026 03:16
Last Modified:	31 Jul 2026 03:16
URI:	http://repository.its.ac.id/id/eprint/140385

Actions (login required)

View Item