Integração de séries temporais fuzzy, representação por imagens e aprendizado federado na classificação de séries temporais.

Abstract

With the growth of data across various application domains, numerous studies and algorithms have been developed to provide increasingly comprehensive, fast, and accurate solutions. A significant portion of this data is collected over time, exhibiting a dependency on the order of measurements. Such data, known as time series, play a crucial role in several applications, making their analysis a topic of substantial importance. Among the tasks associated with time series, classification stands out, and several methods have been proposed over the years, particularly in recent decades. Given the large volume of data typically involved, many of these methods seek approaches that enhance both the efficiency and accuracy of the analysis. With advances in computational resources, artificial intelligence techniques such as artificial neural networks and fuzzy logic have proven especially effective for forecasting, classification, and control tasks. This research investigates and proposes new methodologies for time series classification, emphasizing the integration of fuzzy time series approaches, image-based representations, and federated learning. The first contribution involves applying similarity measures directly to fuzzified time series, aiming to reduce noise and highlight relevant patterns. The second approach introduces three distinct strategies for transforming both raw and fuzzified time series into images, enabling subsequent classification via Convolutional Neural Networks (CNNs). The third contribution, also based on image representations, explores the use of Federated Learning (FL) for distributed classification while preserving data privacy, demonstrating the adaptability of the proposed methods in decentralized environments. The fourth contribution, also based on Federated Learning and on the image representation of time series, applies multitask classification for the generation of a global model and the classification of time series. The results obtained highlight the competitiveness of the proposed approaches in comparison with reference methods, expanding the potential for applying and studying fuzzy time series, visual representations, and privacy-constrained scenarios. This work, therefore, contributes to the advancement of time series classification techniques from multiple methodological and practical perspectives.