Ar Rahman, Fajril (2026) Pengembangan Model Prognostik Data-Driven Untuk Estimasi Health indicator Dan Remaining Useful Life Pada Boiler Feed Pump PLTU Menggunakan LSTM Autoencoder Dan Bayesian Gaussian Process Regression. Masters thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 6007241042-Master_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
6007241042-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (12MB) | Request a copy |
Abstract
Keandalan operasi PLTU sangat bergantung pada performa peralatan kritikalnya, salah satunya Boiler Feed Pump (BFP) yang berfungsi menyuplai air bertekanan tinggi ke dalam boiler. Gangguan pada BFP dapat menyebabkan penurunan keandalan unit, derating, hingga forced outage. Oleh karena itu, penerapan predictive maintenance (PdM) berbasis kondisi menjadi penting untuk menghindari kegagalan tak terencana. Penelitian ini bertujuan mengembangkan model prognostik berbasis data aktual industri untuk membentuk Health indicator (HI) dan memperkirakan Time to maintenance (TTM) pada BFP. Tantangan utama penelitian ini adalah karakter data operasi aktual yang fluktuatif, mengandung noise, dipengaruhi perubahan beban, serta tidak memiliki data run-to-failure yang lengkap. Penelitian sebelumnya umumnya masih banyak menggunakan data laboratorium atau dataset benchmark yang lebih terkontrol, sehingga belum merepresentasikan kondisi operasi aktual industri. Penelitian ini menggunakan data historis aktual BFP dengan parameter multi-sensor, meliputi vibrasi, temperature, arus, flow, dan speed. Pembentukan HI dilakukan menggunakan Long Short-Term Memory Autoencoder berdasarkan reconstruction error terhadap kondisi sehat. Prediksi perkembangan HI dilakukan menggunakan Bayesian Gaussian Process Regression dengan strategi recursive one-step ahead forecasting, historical anchor, cutoff, dan Expanding Tail. Threshold kondisi ditentukan berdasarkan pendekatan standar displacement ISO 7919, dengan threshold alarm sebesar 68.21, threshold maintenance sebesar 85, dan threshold fault sebesar 100. Hasil evaluasi HI menunjukkan nilai monotonicity sebesar 0.039, trendability Spearman sebesar 0.828, trendability Pearson sebesar 0.813, dan prognosability normalized sebesar 0.960. Perbandingan model menunjukkan bahwa Bayesian GPR memiliki performa terbaik dengan RMSE mean sebesar 2.811 dan MAE mean sebesar 2.411, serta mampu melanjutkan tren kenaikan HI menuju fault threshold. Hasil prognostik menunjukkan bahwa BFP telah melewati threshold alarm dan memasuki zona persiapan maintenance. Titik fault diproyeksikan terjadi pada 2 November 2026. Dengan maintenance planning lead time lima bulan, waktu perencanaan maintenance direkomendasikan pada 2 Juni 2026. Dengan demikian, model yang dikembangkan dapat digunakan sebagai sistem pendukung keputusan dalam perencanaan pemeliharaan berbasis kondisi pada BFP PLTU.
=============================================================================================================================
The reliability of a coal-fired power plant is highly dependent on the performance of its critical equipment, including the Boiler Feed Pump (BFP), which is responsible for delivering high-pressure feedwater to the boiler. Disturbances in the BFP may reduce unit reliability, cause derating, and potentially lead to forced outage. Therefore, the implementation of condition-based predictive maintenance (PdM) is essential to prevent unplanned failures. This study develops a data-driven prognostic model using actual industrial data to generate a Health indicator (HI) and estimate the Time to maintenance (TTM) of the BFP. The main challenge addressed in this study is the fluctuating and noisy nature of actual operational data, which is influenced by load variations. Previous studies have generally relied on laboratory data or controlled benchmark datasets, which do not fully represent actual industrial operating conditions. This study uses actual historical BFP data consisting of multi-sensor parameters, including vibration, temperature, current, flow, and speed. The HI is constructed using a Long Short-Term Memory Autoencoder based on reconstruction error from healthy operating conditions. The future progression of the HI is then predicted using Bayesian Gaussian Process Regression with recursive one-step-ahead forecasting, historical anchor, cut-off, and Expanding Tail strategies. The condition thresholds are determined based on the ISO 7919 displacement standard, with an alarm threshold of 68.21, maintenance threshold of 85, and a fault threshold of 100. The HI evaluation results show a monotonicity value of 0.039, Spearman trendability of 0.828, Pearson trendability of 0.813, and normalized prognosability of 0.960. The model comparison shows that Bayesian GPR achieves the best performance, with a mean RMSE of 2.811 and a mean MAE of 2.411, while also being able to continue the increasing HI trend toward the fault threshold. The prognostic results indicate that the BFP has passed the alarm threshold and entered the maintenance preparation zone. The fault point is projected to occur on 2 November 2026. With a five-month maintenance planning lead time, the recommended maintenance planning date is 2 June 2026. Therefore, the developed model can be used as a decision-support system for condition-based maintenance planning of BFPs in coal-fired power plants.
Actions (login required)
 |
View Item |