Performance comparison of machine learning algorithms for machine condition monitoring and predictive maintenance

This thesis presents a comprehensive comparative analysis of various machine learning (ML) and deep learning (DL) algorithms for predicting the Remaining Useful Life (RUL) of bearings, a critical component of predictive maintenance in industrial applications. The study develops a robust Python-based...

Descripción completa

Detalles Bibliográficos
Autor: Rougé Creus, Armand
Tipo de recurso: tesis de maestría
Fecha de publicación:2024
País:España
Institución:Universitat Politècnica de Catalunya (UPC)
Repositorio:UPCommons. Portal del coneixement obert de la UPC
Idioma:inglés
OAI Identifier:oai:upcommons.upc.edu:2117/420102
Acceso en línea:https://hdl.handle.net/2117/420102
Access Level:acceso abierto
Palabra clave:Vibration
Bearings (Machinery)
Health status indicators
Vibration Analysis, Health Indicator, RUL, Prognosis, Regression, Bearing Degradation, Fault Detection, Datasets
Vibració
Àrees temàtiques de la UPC::Enginyeria mecànica::Mecànica::Vibracions mecàniques
Descripción
Sumario:This thesis presents a comprehensive comparative analysis of various machine learning (ML) and deep learning (DL) algorithms for predicting the Remaining Useful Life (RUL) of bearings, a critical component of predictive maintenance in industrial applications. The study develops a robust Python-based framework designed to preprocess vibration data, construct Health Indicators (HI), and implement eleven distinct algorithms for RUL estimation. The research utilizes two well-known datasets: NASA and XJTU-SY, to validate the effectiveness of these algorithms across different degradation patterns and operational conditions. Through extensive sensitivity and evolution analyses, the study uncovers significant insights into the performance of each model, highlighting how algorithmic accuracy and reliability vary depending on the type and amount and type of data used for training. The Temporal Convolutional Network (TNC) algorithm, in particular, is identified as the most versatile and accurate across both datasets, demonstrating superior adaptability and consistency in predicting RUL. The findings of this research contribute to the ongoing advancement of predictive maintenance technologies, offering practical implications for improving industrial machinery uptime and efficiency, while also identifying key areas for further optimization and exploration in future studies.