Fusão de dados em análise de séries temporais binárias para detecção e localização de falhas em máquinas ferramentas

Abstract

Computer Numerical Control (CNC) machine tools are essential in modern manufacturing processes due to their high precision and speed. However, these characteristics are susceptible to hidden faults and degradation, reducing operational reliability. To increase the Mean Time Between Failures (MTBF), detection devices are used to monitor the operational state of machine components, enabling the analysis and processing of information to maximize reliability. Many industries adopt Predictive Maintenance (PdM), which uses intelligent monitoring to prevent future failures. In recent years, Machine Learning (ML) has proven to be fundamental in failure prediction, contributing significantly to PdM. ML techniques based on sensor data, including deep learning approaches such as Autoencoders (AE), have shown promising results. The analysis of sensor time series data, especially binary data, has received significant attention, enabling efficient detection of faults or abnormal activities and facilitating immediate maintenance actions. This work aims to develop accurate and efficient fault detection methods to improve the reliability and efficiency of industrial production systems. The research includes a comprehensive analysis of ML models for fault detection, focusing on binary time series data. In addition to evaluating model performance, a device was developed to capture sensor data and an experimental apparatus was built to safely simulate fault conditions. This thesis proposed an innovative approach that combines the fusion of multivariate binary time series with autoencoder models, demonstrating—through practical experiments with real data—that it is possible to detect faults in machine tools with high precision, while also contributing theoretically to the advancement of binary time series analysis in the context of predictive maintenance.