Examining Condition Monitoring Data for Predictive Maintenance Based on Deep Reinforcement Learning (DRL)

Putri, Rizqi Wahyu Mustika (2025) Examining Condition Monitoring Data for Predictive Maintenance Based on Deep Reinforcement Learning (DRL). Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6010221029-Master_Thesis.pdf Restricted to Repository staff only until 1 April 2027. Download (3MB) | Request a copy

Abstract

Industry 4.0 encourages manufacturing companies to integrate the real and digital worlds, one of which is through the installation of sensors on production machine components to enable real-time monitoring. This is also implemented by one of the cement producing companies in Indonesia. However, in practice, real-world conditions often do not match the thresholds of the machine, resulting in sudden failure of the system. This study aims to develop a model that analyzes motor conditions, predicts the RUL of the motor, and determines the optimal maintenance schedule on a raw mill machine. Three main approaches are used: first, a K-Means-based diagnostic model for motor data clustering. Second, a prognostic model based on LSTM to predict the RUL of the motor. Third, optimal maintenance scheduling is carried out using DRL. The results show that clustering with K-Means successfully grouped data into maintenance mode, normal mode, and high risk mode on 3 separator, main, and fan motors with best value of Davis-Bouldin Index of 0.548; 0.7818; 0.8747 and high Calinski-Harabasz Index of 128411,6745; 78749,4388; 57859,6716; consecutively, indicating that the formed clusters are well separated. The LSTM prognostic model produces an MAE of 0,7678; MAPE of 103.1948%; and RMSE of 0.9887 indicating the model's difficulty in predicting accurately compared to CNN-LSTM model with the MAE of 0.089; MAPE of 42.9%; and RMSE of 0.162. The DRL implementation shows that the agent successfully learns with consistent reward increases and convergence at -15.387, yet the reward value is still negative indicationg the environment modeling still needs improvement, especially in RUL modeling and the provision of regret values in the reward calculation.
===================================================================================================================================
Industri 4.0 mendorong perusahaan manufaktur untuk mengintegrasikan dunia nyata dan digital, salah satunya melalui pemasangan sensor pada komponen mesin produksi agar dapat melakukan pemantauan secara real-time. Hal ini juga diterapkan oleh salah satu perusahaan produsen semen di Indonesia. Namun dalam praktiknya, kondisi dunia nyata seringkali tidak sesuai dengan ambang batas mesin sehingga mengakibatkan kegagalan sistem secara tiba-tiba. Penelitian ini bertujuan untuk mengembangkan model yang menganalisis kondisi motor, memprediksi RUL motor, dan menentukan jadwal perawatan yang optimal pada mesin raw mill. Tiga pendekatan utama digunakan: pertama, model diagnostik berbasis K-Means untuk pengelompokan data motor. Kedua, model prognostik berbasis LSTM untuk memprediksi RUL motor. Ketiga, penjadwalan perawatan yang optimal dilakukan dengan menggunakan DRL. Hasil penelitian menunjukkan bahwa pengelompokan dengan K-Means berhasil mengelompokkan data ke dalam mode perawatan, mode normal, dan mode risiko tinggi pada 3 motor separator, motor utama, dan motor fan dengan nilai Davis-Bouldin Index terbaik sebesar 0,548; 0,8747 dan Indeks Calinski-Harabasz yang tinggi yaitu 128411,6745; 78749,4388; 57859,6716; secara berurutan, menunjukkan bahwa klaster yang terbentuk terpisah dengan baik. Model prognostik LSTM menghasilkan MAE sebesar 0,7678; MAPE sebesar 103,1948%; dan RMSE sebesar 0,9887 yang menunjukkan kesulitan model dalam melakukan prediksi secara akurat dibandingkan dengan model CNN-LSTM dengan MAE sebesar 0,089; MAPE sebesar 42,9%; dan RMSE sebesar 0,162. Implementasi DRL menunjukkan bahwa agen berhasil belajar dengan peningkatan reward yang konsisten dan konvergensi sebesar -15,387, namun nilai reward masih negatif yang menunjukkan pemodelan lingkungan masih perlu perbaikan terutama pada pemodelan RUL dan pemberian nilai regret pada perhitungan reward.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Predictive Maintenance, Maintenance Scheduling, K-Means, LSTM, DRL
Subjects:	T Technology > TS Manufactures > TS173 Reliability of industrial products
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Industrial Engineering > 26101-(S2) Master Thesis
Depositing User:	RIZQI WAHYU MUSTIKA PUTRI
Date Deposited:	06 Feb 2025 06:21
Last Modified:	06 Feb 2025 06:21
URI:	http://repository.its.ac.id/id/eprint/118400

Actions (login required)

View Item