Sistemas inteligentes evolutivos: novas abordagens para detecção e classificação de eventos

Abstract

Evolving Machine Learning, particularly when based on data streams, has proven to be a highly relevant tool in various fields. This approach enables computational models of physical systems to incorporate behavioral patterns observed in data streams, adopting an evolving perspective. Thus, models subject to evolution can start with little or no prior knowledge of the application domain and, through continuous monitoring, detect, internalize, and refine new patterns and information over time. Given the importance of this type of learning, the intelligent systems literature presents a variety of approaches aimed at the continuous processing of data throughout the system’s lifespan. However, there are still significant challenges to overcome, especially regarding performance degradation due to the increasing complexity of modern systems. In this context, this work aims to propose new evolving approaches and formulations for the task of event classification in data streams, focusing on clustering techniques and approximation residual evaluation. Three new approaches were developed, based on vector quantization and residual evaluation, which stand out for their competitive performance, simplicity, and speed, thanks to efficient and low-complexity learning procedures. To validate the proposals, experiments were conducted on three public benchmarks in the field of industrial processes, aiming to assess performance and compare the results with the most advanced methods available in the literature. The results demonstrate that, in general, the proposed approaches are competitive with reference techniques, presenting a balance between accuracy, efficiency, and practicality. In summary, this work contributes to the advancement of evolving machine learning, offering innovative and promising solutions for data stream processing in complex and dynamic scenarios.