Work

Modelos de árboles de decisión incrementales en flujos de datos aplicados al mantenimiento predictivo

Year: 2024
Type: dissertation
Abstract: The proliferation of sensors and connected devices in modern cyber-physical environments has led to an unprecedented surge in sequential data generation. This data, whether in the formof time series or continuous streams, carries crucial information about system behavior, performance, and health. This is of particular importance in the field of predictive maintenance, an application that has become very relevant in the industrial environment due to its objective of minimizing maintenance times and operations by using data analytics and machine learning to predict when equipment failures might occur. However, the inherent temporal dependencies, high dimensionality, and often unbounded nature of these data types pose significant challenges to traditional analytical approaches. There is a pressing need for efficient, scalable, and interpretable machine learning models capable of extracting actionable insights from this complex sequential data, particularly in the context of predictive maintenance for Industry 4.0 and emerging Industry 5.0 paradigms. This thesis addresses these challenges by developing novel machine learning methods for sequential data analysis, focusing on flexibility, scalability, and interpretability. The research tackles key issues in time series classification, online multi-label learning, and predictive maintenance applications, with incremental decision trees as the cornerstone where the different contributions of this thesis intersect. Incremental decision trees emerge as a promising foundation for addressing these challenges. These models offer a unique combination of adaptability, interpretability, and efficiency, making them well-suited for handling dynamic, high-dimensional data streams. Unlike traditional batch learning methods, incremental decision trees can continuously update their model structure as new data arrives, adapting to evolving patterns and concept drift. Their hierarchical structure provides clear decision paths, offering interpretable insights crucial for decision-support in several applications. This thesis leverages and extends the capabilities of incremental decision trees to develop novel approaches for complex sequential data analysis. In the realm of big data analytics, traditional supervised learning approaches often struggle with the sheer volume and complexity of data, particularly in industrial settings where labeling large datasets can be expensive or impractical. Non-standard data representations, such as multi-instance learning and multi-label learning, offer valuable solutions to these challenges. Multi-instance learning allows for the representation of complex objects as bags of instances, where only the bag is labeled, reducing the granularity of required annotations. Multi-label learning, on the other hand, enables the handling of multiple, potentially interdependent labels per instance, capturing the complexity of real-world scenarios where multiple conditions or states can coexist. By integrating these flexible representations with incremental decision trees, this thesis develops novel approaches that can effectively learn from weakly labeled or partially labeled data, significantly enhancing the applicability of machine learning in industrial big data contexts. In summary, the main contributions of this thesis are: ▶Development of novel models based on incremental decision trees that efficiently process data streams and utilize non-standard information representations: • A pioneering approach for time series classification that integrates multi-instance learning with incremental trees, enabling effective learning from flexible data in weakly-labeled environments (MIHT: Multi-Instance Hoeffding Tree). • An innovative proposal for online learning in multi-label data streams, leveraging incremental trees to dynamically capture label correlations (MLHAT: Multi-Label Hoeffding Adaptive Tree). These models are rigorously evaluated using diverse benchmarks from multiple real-world applications, validating their accuracy, performance, and adaptability to data streams and large-scale datasets in comparison to state-of-the-art methods. ▶ A comprehensive review of the predictive maintenance field from a machine learning perspective, providing valuable insights into key techniques, current trends, and gaps in predictive maintenance applications. ▶ Development of specialized models for predictive maintenance, building upon the theoretical foundations established for time series classification and online learning with non-standard representations. These models leverage incremental learning to continuously update predictions as new information is received, crucial for detecting emerging patterns and anticipating potential system failures: • A time-series classification model combining multi-instance representation and incremental trees to diagnose system degradation status in predictive maintenance scenarios (MIHATT: Multi-Instance Hoeffding Anytime Tree for Predictive Maintenance). • Multiple classification models based on online ensembles of incremental trees for multi-label learning, designed to identify multiple faults in dynamic industrial systems (OEMLHAT: Online Ensemble of Multi-Label Hoeffding Adaptive Trees for Predictive Maintenance). These models are evaluated on real-world industrial problems, demonstrating superior performance compared to previous approaches and providing valuable insights into failure causes, derived from the interpretability of the proposed models. This document introduces all global concepts necessary to understand the proposed contributions: provides a theoretical vision of the underlying fundamentals, presents complete specifications of the proposals, and details the experimental studies conducted. Furthermore, it offers access to published articles derived from this thesis and online repositories containing implementations and additional materials associated with the proposals. Finally, the main conclusions drawn from this thesis are presented, as well as possible research directions to continue the advancements of this thesis.
Author: Esteban Toscano, Aurora
Language: English
Open Access status: green
OpenAlex ID: W7020963586
Cites: 228
Cited by:
Related to: 10