Perancangan Sistem Prognosa Kerusakan Pada Turbin Uap Dengan Menggunakan Model Machine Learning

Ashar, Moch. Alwi (2026) Perancangan Sistem Prognosa Kerusakan Pada Turbin Uap Dengan Menggunakan Model Machine Learning. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221053-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (7MB) | Request a copy

Abstract

Pembangkit Listrik Tenaga Uap (PLTU) sangat bergantung pada keandalan turbin uap sebagai penggerak utama sistem pembangkitan. Kondisi operasi yang melibatkan temperatur tinggi dan getaran kontinu menyebabkan komponen turbin mengalami degradasi secara bertahap sehingga diperlukan sistem prognostik yang mampu memantau kondisi kesehatan komponen dan memprediksi perkembangan degradasi untuk mendukung implementasi predictive maintenance. Penelitian ini bertujuan mengembangkan sistem prognostik berbasis data operasional Distributed Control System (DCS) menggunakan model Deep Ensemble Long Short-Term Memory (Deep Ensemble LSTM). Tahapan penelitian meliputi ekstraksi fitur dari sensor temperatur dan vibrasi, standardisasi data menggunakan metode Z-score, pembentukan Health Index melalui fusi data multi-sensor menggunakan Mahalanobis Distance, optimasi sinyal menggunakan Exponential Moving Average (EMA), serta pemodelan degradasi menggunakan Deep Ensemble LSTM yang dievaluasi melalui one-step prediction, recursive forecasting, dan analisis komparatif beberapa konfigurasi model. Hasil penelitian menunjukkan bahwa fusi data sensor temperatur dan vibrasi menggunakan Mahalanobis Distance mampu menghasilkan Health Index yang merepresentasikan perkembangan degradasi komponen secara konsisten. Optimasi sinyal menggunakan Exponential Moving Average (EMA) menghasilkan Health Index yang lebih halus melalui proses smoothing terhadap fluktuasi jangka pendek sehingga pola degradasi lebih konsisten dan lebih mudah dipelajari oleh model. Pada skenario one-step prediction, konfigurasi Deep Ensemble LSTM + EMA menghasilkan kinerja terbaik dengan nilai MAE sebesar 0,0182, RMSE sebesar 0,0244, dan MAPE sebesar 6,68%. Analisis komparatif menunjukkan bahwa optimasi EMA meningkatkan kualitas representasi Health Index, sedangkan penambahan fitur kinematik memberikan peningkatan performa yang lebih nyata pada recursive forecasting, khususnya pada mekanisme Rolling Recursive Forecast. Secara keseluruhan, konfigurasi Deep Ensemble LSTM dengan optimasi EMA dan penambahan fitur kinematik memberikan performa paling konsisten dalam memprediksi perkembangan degradasi komponen. Hasil tersebut menunjukkan bahwa peningkatan kualitas Health Index melalui EMA serta pengayaan informasi masukan menggunakan fitur kinematik memberikan kontribusi yang lebih besar dibandingkan penambahan mekanisme Self-Attention. Selain itu, model mampu mempertahankan tren degradasi pada horizon prediksi yang lebih panjang sehingga berpotensi mendukung implementasi condition-based maintenance dan predictive maintenance.
=======================================================================================================================================
Steam Power Plants (SPPs) rely heavily on the reliability of steam turbines as the primary driving components of power generation systems. Harsh operating conditions involving high temperatures and continuous vibration lead to gradual component degradation, highlighting the need for prognostic systems capable of monitoring component health and predicting degradation progression to support predictive maintenance. This study aims to develop a prognostic system based on Distributed Control System (DCS) operational data using a Deep Ensemble Long Short-Term Memory (Deep Ensemble LSTM) model. The proposed framework consists of feature extraction from temperature and vibration sensors, data standardization using the Z-score method, construction of a Health Index through multi-sensor data fusion based on Mahalanobis Distance, signal optimization using Exponential Moving Average (EMA), and degradation modeling using Deep Ensemble LSTM. The proposed model was evaluated using one-step prediction, recursive forecasting, and a comparative analysis of different model configurations. The results demonstrate that the fusion of temperature and vibration sensor data using Mahalanobis Distance successfully produces a Health Index that consistently represents the degradation progression of steam turbine components. Exponential Moving Average (EMA) optimization generates a smoother Health Index by reducing short-term fluctuations, resulting in a more consistent degradation trend that can be learned more effectively by the Deep Ensemble LSTM model. In the one-step prediction scenario, the Deep Ensemble LSTM with EMA achieved the best performance, with a Mean Absolute Error (MAE) of 0.0182, a Root Mean Squared Error (RMSE) of 0.0244, and a Mean Absolute Percentage Error (MAPE) of 6.68%.
A comparative analysis of all model configurations indicates that EMA optimization plays a crucial role in improving the quality of the Health Index representation, while the incorporation of kinematic features provides more noticeable performance improvements during recursive forecasting, particularly under the Rolling Recursive Forecast mechanism. Across all evaluation scenarios, the Deep Ensemble LSTM model incorporating EMA optimization and kinematic features demonstrated the most consistent performance in predicting component degradation progression. These findings indicate that enhancing the Health Index representation through EMA and enriching the input information with kinematic features contribute more significantly to prognostic performance than incorporating a Self-Attention mechanism. Furthermore, the proposed model is capable of maintaining the degradation trend over extended prediction horizons, demonstrating its potential to support the implementation of condition-based maintenance (CBM) and predictive maintenance.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Turbin Uap, Prognostik, Health Index, Mahalanobis Distance, Deep Ensemble Long Short-Term Memory, Predictive Maintenance, Steam Turbine, Prognostic
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ174 Maintenance and repair of machinery T Technology > TJ Mechanical engineering and machinery > TJ217.6 Predictive Control
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Moch. Alwi Ashar
Date Deposited:	30 Jul 2026 14:55
Last Modified:	30 Jul 2026 14:55
URI:	http://repository.its.ac.id/id/eprint/140936

Actions (login required)

View Item