Pengembangan Model Machine Learning Failure Diagnosis di Turbin PLTU

Faireza, Muhammad Hanifsyah (2026) Pengembangan Model Machine Learning Failure Diagnosis di Turbin PLTU. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221023-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (8MB) | Request a copy

Abstract

Pembangkit Listrik Tenaga Uap (PLTU) menyumbang sekitar 50,52% dari kapasitas listrik nasional Indonesia pada tahun 2023, di mana turbin uap menjadi komponen vital yang rentan mengalami kerusakan akibat kondisi operasi ekstrem. Kegagalan mendadak pada turbin uap dapat menyebabkan unplanned shutdown, kerugian ekonomi besar, dan gangguan stabilitas jaringan. Penelitian ini mengusulkan pipeline diagnosis tingkat kerusakan (failure diagnosis) terintegrasi menggunakan data historis Distributed Control System (DCS) dengan metode multivariat Jarak Mahalanobis berbasis healthy baseline (15 Januari – 10 April 2024). Indeks kontinu Jarak Mahalanobis yang dinormalisasi ke skala 0–1 dipetakan menjadi empat tingkatan kelas alarm diskret (normal, watch, alert, warning) sebagai target klasifikasi untuk mendeteksi enam jenis failure mode. Lima algoritma supervised learning (Random Forest, Decision Tree, XGBoost, ANN (MLP), dan SVM) dilatih menggunakan data sampel operasional berinterval 6 jam hasil time-sync resampling.
Hasil pengujian komparatif menunjukkan bahwa model Artificial Neural Network (ANN - MLP) memberikan performa paling stabil di seluruh failure mode dengan batas performa terendah tertinggi (floor performance 81,63%) dan deviasi terendah (σ_"F1" =4,65%), di mana model ANN dan XGBoost berhasil melampaui acceptance criteria (≥94%) pada moda erosion or pitting. Kendala utama model secara umum terletak pada ketidakmampuan mencapai target acceptance criteria secara merata akibat fenomena tumpang-tindih data (feature overlap) pada zona transisi degradasi fisis kontinu. Guna menjaga integritas korelasi temporal deret waktu (time-series), manipulasi oversampling standar seperti SMOTE sengaja tidak diterapkan. Sebagai gantinya, penelitian ini merekomendasikan penggunaan teknik Soft-Labeling/Fuzzy Classification, ekstraksi fitur domain frekuensi (FFT/PSD), serta arsitektur Deep Learning berbasis memori temporal (LSTM/GRU) untuk meningkatkan sensitivitas deteksi pada fase kritis di masa mendatang.
=======================================================================================================================================
Steam Power Plants (PLTU) contributed approximately 50.52% of Indonesia's national electrical capacity in 2023, where steam turbines serve as vital components susceptible to damage under extreme operating conditions. Sudden failures in steam turbines can cause unplanned shutdowns, severe economic losses, and grid instability. This study proposes an integrated failure diagnosis system using historical Distributed Control System (DCS) data with the multivariate Mahalanobis Distance method based on a healthy baseline (January 15 – April 10, 2024). The continuous Mahalanobis Distance index, normalized to a 0–1 scale, is mapped into four discrete alarm levels (normal, watch, alert, warning) as classification targets to detect six failure modes. Five supervised learning algorithms (Random Forest, Decision Tree, XGBoost, ANN (MLP), and SVM) were trained using operational sample data at 6-hour intervals derived from time-sync resampling.
Comparative testing results demonstrate that the Artificial Neural Network (ANN - MLP) model achieves the most stable performance across all failure modes, exhibiting the highest floor performance (81.63%) and the lowest standard deviation (σ_"F1" =4.65%), with both ANN and XGBoost successfully exceeding the acceptance criteria (≥94%) on the erosion or pitting mode. The primary limitation across all models lies in the inability to uniformly meet the acceptance criteria due to feature overlap at the transition zones of continuous physical degradation. To preserve the temporal correlation of the time-series data, standard oversampling techniques like SMOTE were intentionally omitted. Instead, this study recommends the implementation of soft-labeling/fuzzy classification, frequency-domain feature extraction (FFT/PSD), and temporal memory-based Deep Learning architectures (LSTM/GRU) to enhance detection sensitivity during critical phases in future developments.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Diagnosis Kegagalan, Jarak Mahalanobis, Klasifikasi, Machine learning, Turbin Uap PLTU, Classification, Failure Diagnosis, Mahalanobis Distance, Machine learning, Predictive Maintenance, Steam Turbine PLTU
Subjects:	T Technology > T Technology (General) > T57.5 Data Processing T Technology > T Technology (General) > T57.8 Nonlinear programming. Support vector machine. Wavelets. Hidden Markov models. T Technology > T Technology (General) > T58.62 Decision support systems T Technology > TJ Mechanical engineering and machinery > TJ174 Maintenance and repair of machinery T Technology > TJ Mechanical engineering and machinery > TJ266 Turbines. Turbomachines (General) T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7 Turbogenerators. Steam-turbines T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7870.23 Reliability. Failures T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Hanifsyah Faireza
Date Deposited:	04 Aug 2026 01:18
Last Modified:	04 Aug 2026 01:18
URI:	http://repository.its.ac.id/id/eprint/142520

Actions (login required)

View Item