Studi Komparatif Metode Synthetic Data Generation untuk Predictive Maintenance Bearing Motor Induksi Tiga Fasa Menggunakan Vibration Analysis

Joey, Axel (2026) Studi Komparatif Metode Synthetic Data Generation untuk Predictive Maintenance Bearing Motor Induksi Tiga Fasa Menggunakan Vibration Analysis. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022221139-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (4MB) | Request a copy

Abstract

Motor induksi tiga fasa merupakan komponen kritis dalam berbagai sektor industri, di mana kegagalan bearing menjadi penyebab utama kerusakan yang dapat dideteksi melalui analisis sinyal getaran. Predictive maintenance (PdM) berbasis data menjadi solusi efektif untuk melakukan klasifikasi fault dan prediksi remaining useful life (RUL). Namun, implementasi PdM di industri sering terkendala menggunakan Data yang tidak lengkap. Oleh karena itu, penelitian ini bertujuan melakukan studi komparatif metode synthetic data generation untuk mengatasi data scarcity dalam penerapan PdM. Penelitian ini membandingkan hasil PdM berupa fault classification dan RUL prediction menggunakan model single-input-multi-output Convolutional Neural Network–Long Short-Term Memory (SIMO CNN-LSTM) yang dilatih dengan data getaran asli dan data sintetik dari dua metode, yaitu ARMA/ARIMA dan LSTM. Hasil penelitian menunjukkan bahwa F1-score tertinggi diperoleh model SIMO CNN-LSTM dengan SDG LSTM, yaitu 88,92%, atau 0,81 poin persentase lebih tinggi dibandingkan data getaran asli sebesar 88,11%. Untuk RUL prediction, RMSE terendah sebesar 0,1522 dan R² Score tertinggi sebesar 0,7118 dihasilkan oleh model SIMO CNN-LSTM yang dilatih menggunakan data sintetik ARIMA dengan differencing 1. Sementara itu, model dengan data getaran asli menghasilkan RMSE sebesar 0,1544 dan R² Score sebesar 0,7036, sehingga SDG ARIMA dengan differencing 1 menurunkan RMSE sebesar 0,0022 dan meningkatkan R² sebesar 0,0082. Hasil penelitian ini dapat membantu pelaku industri memilih metode SDG yang paling sesuai untuk menerapkan PdM.
======================================================================================================================================
Three-phase induction motors are critical components in industry, and bearing failure is a major source of damage that can be detected through vibration signal analysis. Data-driven predictive maintenance (PdM) supports fault classification and remaining useful life (RUL) prediction. However, PdM implementation is often constrained by incomplete data. Therefore, this study compares synthetic data generation (SDG) methods to address data scarcity in PdM. PdM performance is evaluated using a single-input multi-output Convolutional Neural Network–Long Short-Term Memory (SIMO CNN-LSTM) model trained with original vibration data and synthetic data generated using ARMA/ARIMA and LSTM. Vibration data were collected from a three-phase induction motor under different bearing conditions. To simulate incomplete data, the final 10% of the training dataset was removed, and the remaining partial data were used to generate synthetic vibration data. The original and synthetic datasets were then used to train and evaluate the SIMO CNN-LSTM model for fault classification and RUL prediction. Model performance was evaluated using classification metrics, including accuracy, precision, recall, F1-score, and confusion matrix, and regression metrics, including mean absolute error (MAE), root mean square error (RMSE), and R² Score. The highest F1-score was achieved by the SIMO CNN-LSTM model trained with LSTM-based SDG, reaching 88.92%, which is 0.81 percentage points higher than the original-data model at 88.11%. For RUL prediction, ARIMA-based SDG with differencing order 1 produced the lowest RMSE of 0.1522 and the highest R² Score of 0.7118.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	predictive maintenance, synthetic data generation, ARMA/ARIMA, LSTM, SIMO CNN-LSTM, remaining useful life prediction, fault classification, predictive maintenance, synthetic data generation, ARMA/ARIMA, LSTM, SIMO CNN-LSTM, remaining useful life prediction, fault classification
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK1007 Electric power systems control T Technology > TS Manufactures > TS174 Maintainability (Engineering) . Reliability (Engineering)
Divisions:	Faculty of Electrical Technology > Electrical Engineering
Depositing User:	Axel Joey
Date Deposited:	27 Jul 2026 08:50
Last Modified:	27 Jul 2026 08:50
URI:	http://repository.its.ac.id/id/eprint/137909

Actions (login required)

View Item