Deteksi Dini Kerusakan Bearing Menggunakan Pendekatan Unsupervised Berbasis Variational Autoencoder Dan Logarithmic Envelope Spectrum Pada Dataset Kaist

Arozak, Brilliant (2026) Deteksi Dini Kerusakan Bearing Menggunakan Pendekatan Unsupervised Berbasis Variational Autoencoder Dan Logarithmic Envelope Spectrum Pada Dataset Kaist. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221146-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Bearing merupakan komponen kritis pada mesin berputar yang rentan mengalami kerusakan akibat beban dan gesekan. Kerusakan bearing yang tidak terdeteksi sejak dini dapat menyebabkan kegagalan mesin secara tiba-tiba dan kerugian operasional yang besar. Penelitian ini bertujuan mendeteksi kerusakan awal bearing secara unsupervised menggunakan framework SimUFD yang menggabungkan Logarithmic Envelope Spectrum (LogES) sebagai teknik prapemrosesan dan Variational Autoencoder (VAE) sebagai model deteksi anomali. LogES diperoleh dengan menerapkan transformasi logaritma pada Envelope Spectrum untuk mengompresi rentang dinamis spektrum, sehingga komponen frekuensi karakteristik kerusakan yang lemah menjadi lebih terlihat. VAE dilatih secara unsupervised menggunakan data kondisi normal bearing, kemudian reconstruction error digunakan sebagai Health Indicator (HI) untuk mendeteksi penyimpangan dari kondisi normal. Pengujian dilakukan pada dataset KAIST Ball Bearing Run-to-Failure menggunakan bearing NSK 6205 yang beroperasi pada 1.780 RPM dengan frekuensi sampling 25.600 Hz selama 129 jam operasi. Hasil menunjukkan bahwa SimUFD mampu mendeteksi incipient fault pada t=85, yaitu 40 jam lebih awal dibandingkan failure visible point yang ditetapkan pada t=125 berdasarkan analisis statistik RMS. Performa SimUFD diukur dengan FAR=0%, MAR=15,56%, Accuracy=94,57%, dan F1-score=0,9157. Dibandingkan dengan metode pembanding WSEDisE (fault point t=125) dan Adaptive Kurtosis (fault point t=126), SimUFD unggul secara signifikan dalam hal deteksi dini. Ablation study menunjukkan bahwa kontribusi LogES lebih dominan dibandingkan arsitektur VAE dalam meningkatkan performa deteksi pada dataset dengan karakteristik operasi intermiten, dengan selisih MAR sebesar 54 poin persentase akibat penggunaan LogES dibandingkan ES biasa.
=========================================================================================================================================
Bearings are critical components in rotating machinery that are susceptible to damage due to load and friction. Undetected bearing faults can lead to sudden machine failure and significant operational losses. This study aims to detect incipient bearing faults in an unsupervised manner using the SimUFD framework, which combines the Logarithmic Envelope Spectrum (LogES) as a preprocessing technique and a Variational Autoencoder (VAE) as an anomaly detection model. LogES is obtained by applying a logarithmic transformation to the Envelope Spectrum to compress its dynamic range, making weak fault characteristic frequency components more prominent. The VAE is trained in an unsupervised manner using normal bearing condition data, and the reconstruction error is used as a Health Indicator (HI) to detect deviations from normal conditions. Experiments were conducted on the KAIST Ball Bearing Run-to-Failure dataset using an NSK 6205 bearing operating at 1,780 RPM with a sampling frequency of 25,600 Hz over 129 hours of operation. Results show that SimUFD successfully detects an incipient fault at t=85, which is 40 hours earlier than the failure visible point established at t=125 based on statistical RMS analysis. SimUFD achieves FAR=0%, MAR=15.56%, Accuracy=94.57%, and F1-score=0.9157. Compared to WSEDisE (fault point t=125) and Adaptive Kurtosis (fault point t=126), SimUFD demonstrates significantly superior early detection capability. An ablation study reveals that LogES contributes more to detection performance than the VAE architecture in datasets with intermittent operating characteristics, with a MAR difference of 54 percentage points when using LogES over standard ES.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Dini Kerusakan, Unsupervised Learning, Variational Autoencoder, Logarithmic Envelope Spectrum, Health Indicator, KAISTEarly Fault Detection, Variational Autoencoder, Logarithmic Envelope Spectrum, Health Indicator, KAIST.
Subjects:	T Technology > TJ Mechanical engineering and machinery > TJ174 Maintenance and repair of machinery
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Brilliant Arozak
Date Deposited:	03 Aug 2026 06:47
Last Modified:	03 Aug 2026 06:47
URI:	http://repository.its.ac.id/id/eprint/142399

Actions (login required)

View Item