Enhancing Physics Informed Neural Network Remaining Useful Life Prediction using Prediction Interval

Napitu, Andreas (2025) Enhancing Physics Informed Neural Network Remaining Useful Life Prediction using Prediction Interval. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5007211083-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Prognostics and Health Management (PHM) sangat penting untuk memprediksi Remaining Useful Life (RUL) bantalan elemen gelinding pada mesin berputar, mengoptimalkan pemeliharaan, dan mencegah kegagalan dalam kondisi tekanan tinggi. Studi ini mengatasi tantangan kelebihan data pada kondisi nominal dan keterbatasan data degradasi dengan mengecualikan kondisi nominal dan memfokuskan pada fase awal degradasi (berkas 647 hingga 983 dalam dataset IMS) untuk meningkatkan akurasi RUL. Pendekatan Multi-Fidelity (MF) mengintegrasikan model eksponensial low-fidelity (LF) dengan data getaran high-fidelity (HF) menggunakan metode Lower Upper Bound Estimation (LUBE) untuk menghasilkan prediction interval yang mengakomodasi ketidakpastian. Model dilatih pada 50, 75, dan 100 titik data mulai dari First Predicted Time (FPT) untuk memprediksi RUL. Evaluasi menggunakan metrik PI menunjukkan model 50 titik memiliki presisi lebih tinggi dengan selisih End-of-Life (EOL) yang lebih kecil dan Coverage Width Criterion (CWC) lebih rendah (sekitar 0.2958), namun PICP hanya 0.3% yang menunjukkan ketidakmampuan menangkap ketidakpastian secara memadai dan gagal mencapai Minimum Credibility Cycle (MCC). Sebaliknya, model 75 dan 100 titik dengan PICP di atas 0.7 melampaui MCC dan mendekati High Credibility Cycle (HCC) meskipun memiliki interval lebih lebar dan prediksi EOL kurang presisi.
=================================================================================================================================
Prognostics and Health Management (PHM) is essential for predicting the Remaining Useful Life (RUL) of roller element bearings in rotating machinery, optimizing maintenance, and preventing failures under high-stress conditions. This study addresses challenges of nominal state data overload and limited degradation data by excluding nominal states and focusing on the initial degradation phase (files 647 to 983 in the IMS dataset) to enhance RUL accuracy. A Multi-Fidelity (MF) approach integrates a low-fidelity (LF) exponential model with high-fidelity (HF) vibration data, employing the Lower Upper Bound Estimation (LUBE) method to generate prediction intervals that accommodate uncertainty. The model was trained on 50, 75, and 100 data points starting from the First Predicted Time (FPT) to predict RUL. Evaluation using PI metrics revealed that the 50-point model demonstrated higher precision with smaller End-of-Life (EOL) discrepancies and a lower Coverage Width Criterion (CWC) of approximately 0.2958, but its Prediction Interval Coverage Probability (PICP) of 0.3% indicated insufficient uncertainty capture, failing to reach the Minimum Credibility Cycle (MCC). In contrast, the 75- and 100-point models, with PICP above 0.7, surpassed MCC and approached the High Credibility Cycle (HCC) despite wider intervals and less precise EOL predictions.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Prognostik, Multi-Fidelity, RUL, Kondisi Nominal, Prognostics, Multi-Fidelity, RUL, LUBE
Subjects:	Q Science Q Science > Q Science (General) Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Mechanical Engineering > 21201-(S1) Undergraduate Thesis
Depositing User:	Andreas Elmonangan Hananya Napitu
Date Deposited:	01 Aug 2025 02:03
Last Modified:	01 Aug 2025 02:03
URI:	http://repository.its.ac.id/id/eprint/125049

Actions (login required)

View Item