VIBRATION ANALYSIS METHOD TO DIAGNOSE DEFECTS OF ROTATING MACHINE WITH MACHINE LEARNING APPROACH

• Main Author: Dainuri, Akhmad
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/70664
• Institution: Institut Teknologi Bandung

The application of industry 4.0 is currently being intensively developed. One component that is widely discussed and studied is predictive maintenance. This method is proven to reduce maintance costs and decrease the losses due to sudden termination to production process. There are several kind of predictive maintenance techniques, one of them is vibration analysis. This technique is usually used to diagnose the condition of a rotating machine and determine the location of damage. Some kind of vibration analysis methods have been developed, there are time domain analysis, wavelet coefficient analysis, and frequency spectrum analysis. Time domain analysis and wavelet coefficient analysis are now being pretty much developed. However, they are still having some weaknesses and less effective to use in practical application. Another method that has been being widely used is frequency spectrum analysis. Fast Fourier Transform (FFT) is an algorithm that is used to decompose the frequencies in a vibration signal. Nevertheless, this algorithm has some weaknesses in case of machine learning classification, they are spectral leakage and frequency shifting due to measurement error. Spectral leakage makes the amplitude of a frequency lower than the real value and frequency shifting leads to error in reading of amplitude in feature extraction. Therefore, in this paper the vibration analysis is conducted with same perspective with classic frequency spectrum analysis, but it is done with little bit different approach. Generally, the process is divided into filtering in certain frequency of fundamental frequency and the harmonics. Next, Root Mean Square (RMS) is calculated for each filtered signal and it is normalized to get approximation amplitude of the fundamental frequency and harmoncis. This method is then named as FiltRA (Filtering and RMS Approximation). The values obtained from that method are used for process in machine learning. Vibration analysis conducted in this paper was divided into two parts, that are low frequency analysis and high frequency analysis. In low frequency analysis, the defects observed were unbalance, misalignment, and looseness. Then, in high frequency analysis, the observed defects were outer race, inner race, roller, and cage defect in a bearing. This separation was done because in many cases of high iv frequency defect, the amplitude of high frequency defect is disappeared because it is too small if compared with amplitude of low frequency defect. Therefore, by applying this division, it is expected that the defect in low and high frequency can be detected clearly. In classification process, low frequency analysis provided output with dominant defect while high frequency analysis could provide defects that appeared together. This thing could be conducted because in high frequency analysis, the defect frequency appeared in different value. But, in low frequency, the frequencies used were same, so it was hard to extract the defects appeared together. There were three different algorithms used, they were Multi Level Perceptron (MLP), KNearest Neighbor (KNN), and Random Forest (RF). After training and validation process, it was gained that RF algorithm had the best performance with 99,66% accuracy and 0,044 ms computational time to predict a vibration datum with defect in low frequency domain. Same result was obtained in analysis of bearing defect. It showed 100% accuracy for RF algorithm and 0,052 ms computational time. However, in bearing defect detection, all classification algorithms can generate 100% accuracy because the difference between the defects and normal condition was clearly separated.