Metodologia para monitoramento de maquinário em uso crítico por meio de AIoT com foco em manutenção preditiva

Abstract

The growing demand for reliability, availability, and efficiency of industrial assets has driven the development of advanced predictive maintenance strategies within the context of Industry 4.0. In this scenario, the integration of sensors, Internet of Things (IoT) platforms, Artificial Intelligence (AI) techniques, and Machine Learning (ML) algorithms becomes essential for estimating the equipment health condition (Health State – HS) and predicting the Remaining Useful Life (RUL). However, a large portion of the approaches found in the literature focuses on isolated components or standardized experimental datasets, with limited validation in real operational environments and limited integration with distributed data processing architectures. In this context, this thesis proposes a structured methodology for monitoring and managing the health of machinery operating in critical applications, based on the integration of concepts from Prognostics and Health Management (PHM), AioT architectures, and system development methods. The methodology was developed based on a systematic literature review and structured according to the principles of Design Science Research, resulting in a methodological workflow composed of stages that include the identification of the critical system, data collection and processing, health state inference, remaining useful life prediction, and generation of operational alerts. The applicability of the proposed methodology was evaluated through two complementary applications. The first application uses real data obtained from a railway track geometry measurement system, in which sparse time series were modeled using statistical techniques and Long ShortTerm Memory (LSTM) recurrent neural networks. The results demonstrated the capability to estimate infrastructure degradation progression, identify operational regime changes, and predict the crossing of health state thresholds, enabling consistent RUL estimations even under low temporal granularity conditions. The second application consists of the implementation of an IoT-based monitoring system for stationary generators operating on the Otto cycle, integrating embedded sensors, MQTT communication protocol, processing in a flow-based development platform, and data storage in a SQL database. The developed solution enables structured acquisition of operational data, remote updating of inference models, and continuous evaluation of the fuel pump operating condition, demonstrating the feasibility of applying PHM to equipment operating in environments with connectivity constraints. The results obtained from both applications highlight the flexibility, scalability, and applicability of the proposed methodology, validating its capability to integrate data acquisition, degradation modeling, and maintenance decision support. As a scientific contribution, this work proposes a replicable methodological framework for the development of machine health monitoring solutions in critical applications, combining concepts of PHM, IoT, and machine learning in an integrated predictive maintenance approach.