Perawatan Prediktif Berbasis Optimum Model Hybrid Temporal Convolutional Network Untuk Estimasi Remaining Useful Life Bearing

Oktamutia, Tsabitha Kanaya (2026) Perawatan Prediktif Berbasis Optimum Model Hybrid Temporal Convolutional Network Untuk Estimasi Remaining Useful Life Bearing. Masters thesis, Institut Teknologi Sepuluh Nopember.

Text 6007241057-Master_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Rolling element bearing merupakan komponen mekanis kritis yang mengalami degradasi progresif selama operasi, sehingga estimasi Remaining Useful Life (RUL) yang akurat menjadi kunci pemeliharaan prediktif berbasis kondisi. Penelitian ini mengusulkan kerangka hybrid Temporal Convolutional Network (TCN) untuk estimasi RUL bearing pada dataset NASA IMS melalui skenario cross-set generalization (Set 1 dan 2 sebagai data latih, Set 3 sebagai blind test). Fitur multi-domain diseleksi menggunakan korelasi peringkat Spearman (didominasi indikator spektral, ρ>0,80), dan label RUL dibangun melalui skema piecewise berbasis First Predicting Time (FPT) yang mempertahankan RUL konstan selama fase sehat sebelum meluruh eksponensial pasca-FPT. Kerangka menggunakan transfer learning dua tahap: TCN dilatih sebagai ekstraktor fitur temporal, dibekukan, lalu representasi latennya menjadi masukan XGBoost, RNN, dan LSTM, dibandingkan terhadap padanan standalone-nya. TCN-RNN teridentifikasi optimum di antara ketiga model hybrid berdasarkan ambang alarm industrial 75% degradasi (RMSE=0,279; NASA Score=107,29; status EARLY SAFE), namun Standalone XGBoost mencapai akurasi tertinggi keseluruhan (RMSE=0,143; R²=0,470) di antara ketujuh model. Validasi Leave-One-Set-Out mengungkap model standalone justru mendominasi seluruh tiga fold IMS, dengan RMSE TCN Baseline melonjak ke 2,344 saat diuji lintas moda kegagalan berbeda (Fold-C). Validasi lanjutan pada delapan skenario gabungan (tiga fold LOSO dan lima unit PRONOSTIA) mengonfirmasi pola yang bergantung dataset dimana standalone menang mutlak di IMS (0 dari 3), sementara arsitektur TCN baseline menang di seluruh unit PRONOSTIA (5 dari 5), diduga akibat interaksi ukuran window/dilasi TCN yang identik dengan perbedaan granularitas akuisisi data sekitar 60 kali lipat antar dataset. Audit kalibrasi fase-sehat pada kedua dataset mengonfirmasi kualitas kalibrasi dan ketepatan alarm merupakan properti independen. Temuan ini menggeser kontribusi penelitian dari klaim superioritas model hybrid menjadi wawasan metodologis: keunggulan arsitektur temporal kompleks bersifat kondisional terhadap karakteristik akuisisi data, bukan keunggulan default berbasis kompleksitas arsitektur semata.
=======================================================================================================================================
Rolling element bearings are critical mechanical components subject to progressive degradation during operation, making accurate Remaining Useful Life (RUL) estimation key to condition-based predictive maintenance. This study proposes a hybrid Temporal Convolutional Network (TCN) framework for bearing RUL estimation on the NASA IMS dataset through a cross-set generalization scenario (Sets 1 and 2 for training, Set 3 as blind test). Multi-domain features were selected via Spearman rank correlation (dominated by spectral indicators, ρ>0.80), and RUL labels were constructed through a piecewise scheme based on First Predicting Time (FPT), keeping RUL constant during the healthy phase before exponential decay post-FPT. The framework adopts two-stage transfer learning: TCN is trained as a temporal feature extractor, frozen, and its latent representation feeds XGBoost, RNN, and LSTM heads, benchmarked against standalone counterparts. TCN-RNN was identified as optimal among the three hybrid models under the 75%-degradation industrial alarm framework (RMSE=0.279; NASA Score=107.29; EARLY SAFE status), yet Standalone XGBoost achieved the highest overall accuracy (RMSE=0.143; R²=0.470) among all seven models. Leave-One-Set-Out validation revealed standalone models dominated all three IMS folds outright, with TCN Baseline's RMSE surging to 2.344 when tested across differing failure modes (Fold-C). Extended validation across eight combined scenarios (three LOSO folds and five PRONOSTIA units) confirmed a dataset-dependent pattern: standalone models won outright on IMS (0 of 3), whereas the TCN baseline architecture won across all PRONOSTIA units (5 of 5), plausibly due to interaction between TCN's fixed window/dilation size and a roughly 60-fold difference in data acquisition granularity between datasets. Healthy-stage calibration audits on both datasets confirmed calibration quality and alarm timeliness as independent properties. These findings shift the study's contribution from claiming hybrid-model superiority toward a methodological insight: the advantage of complex temporal architectures is conditional on data acquisition characteristics, not a default benefit of architectural complexity alone.

Item Type:	Thesis (Masters)
Uncontrolled Keywords:	Audit Kalibrasi Model, Prognostik Bearing, Remaining Useful Life, Temporal Convolutional Network, Transfer Learning, Bearing Prognostics, Model Calibration Audit
Subjects:	T Technology > TJ Mechanical engineering and machinery
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21101-(S2) Master Thesis
Depositing User:	Tsabitha Kanaya Oktamutia
Date Deposited:	03 Aug 2026 09:33
Last Modified:	03 Aug 2026 09:33
URI:	http://repository.its.ac.id/id/eprint/141540

Actions (login required)

View Item