Alghiffari, Muhammad Abdul Aziz (2026) Sistem Prediksi Clogging Injektor Menggunakan LSTM Dengan Ekstraksi Fitur Statistik Pada Data Exhaust Gas Temperature 18 Silinder Di PLTMG Sumbawa. Other thesis, Institut Teknologi Sepuluh Nopember.
![[thumbnail of 2040221021-Undergraduate_Thesis.pdf]](http://repository.its.ac.id/style/images/fileicons/text.png) |
Text
2040221021-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (11MB) | Request a copy |
Abstract
Exhaust gas Temperature (EGT) merupakan temperatur gas buang hasil proses pembakaran pada setiap silinder mesin gas yang digunakan sebagai indikator kualitas pembakaran dan distribusi bahan bakar. Pada mesin MAN 18V51DF di PLTMG Sumbawa, sistem monitoring menggunakan set point alarm sebesar 450°C pada masing-masing silinder. Namun, pendekatan berbasis ambang batas (threshold) tersebut hanya memberikan peringatan ketika temperatur telah mencapai nilai alarm, sehingga perubahan temperatur yang terjadi secara bertahap akibat indikasi awal clogging injektor berpotensi tidak terdeteksi lebih dini. Oleh karena itu, penelitian ini bertujuan mengembangkan sistem prediksi berbasis deep learning untuk mendukung mekanisme early warning terhadap indikasi clogging injektor menggunakan data EGT dari 18 silinder. Metode yang digunakan meliputi data cleaning, penanganan missing value, penyaringan noise menggunakan Moving Average, normalisasi Min-Max Scaling, serta ekstraksi Statistical-Based Features (SBF) berbasis sliding window yang terdiri atas nilai mean, maximum, minimum, dan standard deviation. Selanjutnya dilakukan hyperparameter tuning dan perbandingan empat arsitektur deep learning, yaitu Single LSTM, Stacked LSTM, Bidirectional LSTM, dan Gated Recurrent Unit (GRU), untuk memperoleh model dengan performa terbaik. Hasil penelitian menunjukkan bahwa Single LSTM menghasilkan nilai RMSE dan MAE terbaik, yaitu masing-masing sebesar 0,001110 dan 0,000867. Namun, berdasarkan pengujian robustness pada seluruh silinder menggunakan metrik RMSE, MAE, dan koefisien determinasi (R²), Stacked LSTM menunjukkan performa yang lebih konsisten sehingga dipilih sebagai model implementasi. Pengujian forecasting menghasilkan rata-rata MAE sebesar 0,93°C, MAPE sebesar 0,24%, dan standar deviasi error sebesar 0,50°C pada seluruh silinder. Model yang dikembangkan berhasil diintegrasikan ke dalam sistem monitoring berbasis website untuk menampilkan hasil prediksi dan memberikan early warning secara real-time. Hasil penelitian menunjukkan bahwa kombinasi Statistical-Based Features dan Stacked LSTM berpotensi mendukung penerapan predictive maintenance pada mesin gas industri.
==================================================================================================================================
Exhaust gas Temperature (EGT) is the temperature of exhaust gases produced during the combustion process in each cylinder of a gas engine and is commonly used as an indicator of combustion quality and fuel distribution. In the MAN 18V51DF gas engine at PLTMG Sumbawa, the monitoring system applies an alarm set point of 450°C for each cylinder. However, this threshold-based monitoring approach only provides a warning after the EGT reaches the predefined alarm value, making it difficult to detect the gradual temperature changes caused by the early stages of injector clogging. Therefore, this study aims to develop a deep learning-based prediction system to support an early warning mechanism for injector clogging using EGT data collected from 18 cylinders. The proposed method consists of data cleaning, missing value handling, noise reduction using a Moving Average filter, Min-Max Scaling normalization, and Statistical-Based Features (SBF) extraction through a sliding window approach, including the mean, maximum, minimum, and standard deviation. Hyperparameter tuning was subsequently performed, followed by a comparison of four deep learning architectures, namely Single Long Short-Term Memory (LSTM), Stacked LSTM, Bidirectional LSTM, and Gated Recurrent Unit (GRU), to determine the best-performing model. The evaluation results show that the Single LSTM achieved the lowest prediction errors, with an RMSE of 0.001110 and an MAE of 0.000867. However, robustness evaluation across all cylinders using RMSE, MAE, and the coefficient of determination (R²) demonstrated that the Stacked LSTM provided more consistent performance and was therefore selected for system implementation. Forecasting evaluation yielded an average MAE of 0.93°C, a MAPE of 0.24%, and a standard deviation of prediction error of 0.50°C across all cylinders. The developed model was successfully integrated into a web-based monitoring system capable of displaying prediction results and providing real-time early warnings. The findings indicate that the combination of Statistical-Based Features and Stacked LSTM has strong potential to support predictive maintenance in industrial gas engines.
Actions (login required)
 |
View Item |