ROBUSTNESS EVALUATION OF COMBINED TIME-SERIES MODELING, PARAMETRIC MINIMUM-DISTANCE CLASSIFICATION, AND HYPOTHESIS TESTING IN FAULT IDENTIFICATION OF ROTATING MACHINERI

• Main Author: PRAMUDITA SAPUTRA (NIM : 13107108); Pembimbing : Ir. Ign. Pulung Nurprasetio, MSME., GABRIEL
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/15744
• Institution: Institut Teknologi Bandung

Maintenance is one important thing to keep engine performance. Perfect design and appropriate maintenance will produce longer lifetime and optimum production capacity. One of maintenance type that is used in the industry is condition-based maintenance (CBM). In CBM, maintenance recommendation will be given based on machinery condition. The steps needed in CBM are data acquisition, data processing, and maintenance decision. One popular approach in CBM employs vibration data that is analyzed in the frequency. However, this method may produce inaccurate conclusion, particularly if the data is contaminated heavily by noise. <br /> <br /> <br /> In this undergraduate thesis, we try a new method in processing vibration data. The method combines time series modeling (TSM) and nearest neighbor classification method (NNCM) in identifying common faults in rotating machineries (Dianviviyanthi and Nurprasetio, 2000). The main objective of this research is to explore the robustness of this method in detecting two faults, i.e., unbalance versus unbalance with mechanical looseness. <br /> <br /> <br /> The method is started with a learning stage to get the time series parameters. AutoRegressive (AR) or AutoRegressive Moving Average (ARMA) models are fitted to a set of data coming from a certain fault. For the unbalance fault, an AR(4) was found to be the best model, whereas for the unbalance with mechanical looseness, the best model is AR(12). The best model is the result of order determination procedure using Bayesian Information Criterion (BIC) analysis. <br /> <br /> <br /> Learning stage was conducted using 30 independent data set whereas 26 other data set is used to show robustness. The distribution of the distances is depicted in the form of boxplots, the form of which is complemented by Hotelling’s T2 statistic test to specify the fault type. <br /> <br /> <br /> The result of the experiment showed that the method succesfully identify the simulated faults, namely the pure unbalance and the combined unbalance and mechanical loosenes. Therefore, the objective of showing the reliability has been achieved and the method is deemed satisfactory to be used as a complement to the well-known frequency domain methods.