Use of quantum algorithms for classification of rolling bearing damage

• Ano de defesa: 2023
• Autor(a) principal: AICHELE FIGUEIROA, Diego Andrés
• Orientador(a): MOURA, Márcio Jose das Chagas
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• dARK ID: ark:/64986/001300000drg3
• Idioma: eng
• Instituição de defesa: Universidade Federal de Pernambuco
• Programa de Pós-Graduação: Programa de Pos Graduacao em Engenharia de Producao
• País: Brasil
• Palavras-chave em Português: Engenharia de produção; Aprendizado de máquina quântico; Conjuntos de dados de rolamentos; Gerenciamento de prognóstico e saúde; Diagnóstico de falhas
• Link de acesso: https://repositorio.ufpe.br/handle/123456789/54969

In Production Engineering, there is an area of research on reliability, maintenance, and risks in Production Systems. In this area, the objectives include maintenance policies aimed at predicting component failures and consequently minimizing unexpected downtimes of complex systems. Reliability engineering has successfully utilized machine learning to predict and categorize equipment and machine states. In this thesis, we aim to compare different machine learning algorithms and quantum machine learning in the context of reliability engineering. Specifically, we compare 8 models created with quantum machine learning, using three different bearing datasets: Case Western Reserve University, Machinery Failure Prevention Technology, and Paderborn University. These datasets mainly consist of accelerometer vibration data from the bearings, which must be processed to fit the different quantum machine learning algorithms. We focus on fault detection in rotating equipment, using quantum computing techniques to analyze sensor data installed on the equipment. Our approach considers the healthy state, inner ring faults, and outer ring faults of the bearings, with the aim of achieving fault detection. We compare established circuit designs, such as Real Amplitudes and Quantum Convolutional Circuits, as well as hybrid models that combine quantum circuits and neural networks using the library of TensorFlow with Cirq in python. We obtain results on classical computers using analytical calculations that simulate a quantum computer without experiencing the "quantum noise" delivered by the hardware. The most effective way to input data's key features is through a ZFeatureMap circuit, which assigns a qubit to each extracted feature from the data. Quantum convolutional circuits yield better results than other parameterized circuits, while hybrid models provide higher accuracy than their counterparts that do not utilize neural networks. The achieved accuracy rates on the training dataset are around 96%, suggesting that the parameterized quantum circuits used in this master's thesis yield stable results.