Using smart edge devices and Big Data analytics for predictive maintenance

• Autor(a) principal: Nunes, Pedro Alexandre Almeida
• Data de Publicação: 2024
• Idioma: eng
• Título da fonte: Repositório Científico de Acesso Aberto de Portugal (Repositórios Cientìficos)
• Texto Completo: http://hdl.handle.net/10773/42065
• topic: Predictive maintenance; Generalized predictive models; Smart edge devices; Generalized fault trees

In recent decades, significant emphasis has been placed on improving practices related to industrial equipment maintenance. With the evolution and widespread adoption of the internet of things, various predictive models that utilize data from industrial sensors to forecast the useful life of industrial assets, have been proposed. Predictive maintenance policies aim to leverage these data and predictive models to optimize maintenance scheduling and reduce associated costs. Despite concerted efforts, the adoption of such policies by companies is not as widespread as one might expect. This is due, in part, to certain challenges. Namely, predictive maintenance requires a flexible platform to allow the implementation and customization of predictive models and functionalities in a wide range of industrial environments. Furthermore, existing predictive models in the literature are typically tailored to specific machines or components, limiting their scope of application. Furthermore, their validation is often conducted in controlled environments using synthetic data generated by simulators, which fall short of representing the unpredictable nature of real-world data. Additionally, the complexity of some models renders them as black boxes to maintenance teams, making it difficult to discern the parameters influencing a given prediction. The primary objective of this thesis is to address these challenges. To achieve this, a novel approach called generalized fault trees (GFTs) was proposed and validated in real-world use cases at Bosch and OLI, companies of the Aveiro region. This approach stands out for its versatility in addressing various industrial maintenance issues and for providing a user-friendly graphical representation. The results obtained demonstrate significant reductions in maintenance costs. In addition to GFTs, a new model, combining this approach with a deep learning network, designated k-LSTM-GFT, was proposed for estimating the remaining useful life of industrial assets. It showed superior performance compared to state-of-the-art approaches on benchmark datasets containing multi- ple failure modes. Beyond predictive models, a decentralized architecture based on the use of smart edge devices is proposed for implementing maintenance-related functionalities and collecting industrial data on the shop floor. This architecture is designed to make the maintenance solution adaptable to various distinct scenarios, allowing for generalization not only at the model level but also at the deployment level, thereby attesting its practical utility.