Tese: Vibration monitoring of mechanical systems using deep and shallow learning on edge-computers
Aluno(a) : Carolina de Oliveira ContenteOrientador(a): Helon Ayala
Área de Concentração: Mecânica Aplicada
Data: 12/05/2022
Link para tese/dissertação: https://doi.org/10.17771/PUCRio.acad.59831
Resumo: Structural health monitoring has been the focus of recent developments in the field of vibration-based assessment and, more recently, in the scope of internet of things as measurement and computation becomes distributed. Data has become abundant even though the transmission is not always feasible at higher frequencies needed for proper assessment, especially in remote applications such as pipelines, subsea, and smart fleets. It is thus important to devise data-driven model workflows that ensure the best compromise between model accuracy for condition assessment and also the computational resources needed for embedded solutions, a topic that has not been widely used in the context of vibration-based measurements. Vibration measurements are used to monitor static and rotating machines since they are sensitive to faulty conditions. In this context, the present research proposes two approaches for two applications being a static and a rotating one: in the first one a modeling workflow able to reduce the dimension of autoregressive models built on the basis of many acceleration sensors was proposed. The three-story building example was used to demonstrate the effectiveness of the method, together with ways assess the best compromise between accuracy and model size. Hoping to point future research directions of embedded computing, predictive analytics, and vibration based structural health monitoring, in order to ensure that the models created can be conveniently deployed while optimizing costs for computing infrastructure; and the second approach where a test rig composed of rotating inertias and slender connecting rods is used, and the monitoring solution was tested in an embedded GPU-based platform. Principal component analysis and deep autoencoder models were implemented to distinguish nominal and abnormal friction states effectively. Shallow models perform better concerning running time, memory usage, and accuracy in detecting faulty conditions, which is essential for monitoring rotating machines in real-time.