A Graph-Based Variational Autoencoder and Graph Entropy Framework for Early Fault Detection of Run-to-Failure Bearing

Al Hafizh, Dzulfahmi (2026) A Graph-Based Variational Autoencoder and Graph Entropy Framework for Early Fault Detection of Run-to-Failure Bearing. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007221082-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (9MB) | Request a copy

Abstract

Kegagalan tak terduga pada rolling element bearings menyebabkan penghentian produksi yang parah, sehingga deteksi degradasi dini pada bearing sangat penting untuk pemeliharaan prediktif. Penelitian ini bertujuan untuk mengembangkan kerangka kerja terintegrasi yang menggabungkan Graph-Based Variational Autoencoder (GVAE) dan Graph Entropy (GE) dan menggunakan fitur multidomain untuk deteksi dini kerusakan dan diagnosis mode kegagalan. Fitur multidomain, yang diekstraksi dari sinyal getaran, menyusun graf temporal untuk diproses oleh GVAE. GE kemudian dihitung dari ruang laten GVAE sebagai indikator kesehatan, dan algoritma Short-Term Month-over-Month (STMM) adaptif digunakan untuk menentukan titik awal degradasi. Secara bersamaan, cabang diagnosis paralel mengevaluasi envelope spectrum untuk melacak karakteristik harmonik kerusakan. Ambang batas STMM yang dioptimalkan, dengan history window=150 dan koefisien K=4.5, berhasil mengisolasi titik deteksi dini pada file ke-525 untuk pengujian kedua dan file ke-6160 untuk pengujian ketiga. Kerangka kerja ini mencapai akurasi 99.2%¬-−99.8% dengan tingkat alarm palsu 0.2%−1.5% dan tingkat alarm terlewat 0% pada seluruh skenario pemotongan data. Cabang diagnostik juga secara akurat mengidentifikasi potensi cacat lingkar luar (BPFO) tepat pada awal degradasi. Meskipun pengujian kedua menunjukkan monotonisitas (0.109) dan trendabilitas (0.421) yang lebih rendah dibandingkan pengujian ketiga (0.500 dan 0.830), kerangka kerja ini tetap terbukti tangguh dan andal untuk pemantauan kondisi bearing di industri.
=========================================================================================================================================
The unexpected failures in rolling element bearings cause severe downtime, making early degradation detection of bearing essential for predictive maintenance. This research aims to develop an integrated framework combining a Graph-Based Variational Autoencoder (GVAE) and Graph Entropy (GE) and using multidomain features for high-sensitivity early fault detection and concurrent fault diagnosis. Multidomain features, extracted from vibration signals, construct temporal graphs for the GVAE. GE is then computed from the GVAE latent space as a health indicator, and an adaptive Short-Term Month-over-Month (STMM) algorithm pinpoints degradation onset. Concurrently, a parallel diagnosis branch evaluates the envelope spectrum to track harmonic fault signatures. The optimized STMM, with history window=150 and the coefficient of K=4.5, successfully isolated early faults at the 525th file for IMS Second Test and the 6160th file for IMS Third Test. The framework achieved 99.2%−99.8% accuracy alongside 0.2%−1.5% false alarm rate and 0% missed alarm rate across all truncation scenarios. The diagnostic branch accurately identified potential outer race defects (BPFO) at the exact degradation onset. Although the second test exhibits lower monotonicity (0.109) and trendability (0.421) than the third test (0.500 and 0.830), the framework remains robust and reliable for industrial bearing condition monitoring.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi Dini Kerusakan, Diagnosis Mode Kegagalan, Graph-Based Variational Autoencoder, Graph Entropy, Fitur Multidomain ============================================================== Early Fault Detection, Fault Mode Diagnosis, Graph-Based Variational Autoencoder, Graph Entropy, Multidomain Features
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA169 Reliability (Engineering) T Technology > TA Engineering (General). Civil engineering (General) > TA169.5 Failure analysis T Technology > TA Engineering (General). Civil engineering (General) > TA355 Vibration.
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Dzulfahmi Al Hafizh
Date Deposited:	03 Aug 2026 07:37
Last Modified:	03 Aug 2026 07:37
URI:	http://repository.its.ac.id/id/eprint/142398

Actions (login required)

View Item