| Penulis/Author |
FRANSISKUS TATAS DWI ATMAJI (1); Prof. Ir. Jamasri, Ph.D., IPU., ASEAN Eng. (2); Ir. Hari Agung Yuniarto, S.T., M.Sc., Ph.D., IPU., ASEAN Eng. (3) ; Dr. Ir. I Made Miasa, S.T., M.Sc (4) |
| Abstrak/Abstract |
Bearing failures account for approximately 40–70% of malfunctions in rotating machinery, with shaft
misalignment recognized as a critical yet underexplored root cause of these failures. Despite its practical
importance, the direct impact of parallel shaft misalignment on bearing fault prediction remains unaddressed,
mainly in the existing literature. To fill this gap, the present study proposes a novel experimental framework
utilizing a custom-designed Bearing–Shaft Misalignment Simulator, specifically engineered to induce and control
varying degrees of parallel misalignment under realistic operating conditions. This setup enables systematic
analysis of bearing vibration behaviour, representing a significant methodological advancement over prior
studies that primarily rely on synthetic or public datasets. Six classification models—five machine learning al
gorithms (Multilayer Perceptron, Random Forest, Decision Tree, K-Nearest Neighbors, and Adaptive Boosting)
and one deep learning model (Long Short-Term Memory, LSTM)—were evaluated for classifying four levels of
misalignment severity. The results reveal a strong positive correlation between the magnitude of misalignment
and vibration intensity, highlighting the escalation of dynamic instability in the bearing system. Statistical time-
domain feature extraction notably improved the performance of classical models, with KNN achieving a
maximum accuracy of 92.9%. In contrast, the LSTM model, trained directly on raw time-series data, out
performed all other models, achieving a classification accuracy of 99.7%.This study contributes a novel dataset,
an original misalignment simulation platform, and a comprehensive comparative analysis of modelling ap
proaches. The findings highlight the crucial role of parallel shaft misalignment in bearing degradation and
demonstrate the superior capability of deep learning for early fault detection, representing a significant
advancement in condition-based maintenance for industrial applications. |
