05 de septiembre de 2022
Resumen:
Industry 4.0 is already a reality. Industrial systems involved in the value chain are more and more complex and implementing data-driven diagnostic and prognostic strategies can become a complex, costly, and even irrational task. Simplifying this task in a clear, simple, intuitive, and inclusive way is the main objective of this thesis, which aims to democratize access to anomaly detection and prognostic methodologies to those who do not have advanced knowledge in data analytics. The proposed methodologies are not intended to replace custom techniques designed specifically for a particular type of system, but to provide robust, simple, and fast inclusive solutions to those industry sectors that are in incipient stages of intelligent data-driven asset management solutions. Patterns, Smart Indicators, or Digital Twins, are some of the concepts that will appear throughout this thesis and give names to the different types of knowledge extracted from data coming from sensors and SCADA systems. The scarcity of data, exacerbated by the rapid changes in the operating modes of modern systems, encourages the use of techniques that generate scalable models capable of being updated modularly. The use of unsupervised learning techniques makes possible the rapid integration of the proposed diagnostic and prognosis models without the need for a detailed fault log or an exhaustive list of the different system operation modes. Features used to define the behavior of a system were chosen based on common characteristics of any time series that describes the state of a system. This thesis presents two anomaly detection methodologies and a prognosis methodology. The first anomaly detection methodology is based on the characterization of behaviors using patterns and intelligent indicators, which will later be the pillars of the third methodology focused on prognosis. The second methodology seeks to complement the first anomaly detection methodology using a completely new approach inspired by the Digital Twins paradigm. This methodology focuses on the study of interactions between Digital Twin behaviors to obtain a crucial knowledge for the detection of contextual anomalies in collaborative systems. Each of the proposed methodologies is supported by one or more case studies that illustrate the integration process into a real system. Each of these examples aims to explain the steps to be followed when implementing each methodology, understand the type of knowledge generated, and how the results obtained can be integrated into a decision-making process.
Descriptores: Matemáticas, Estadística, Métodos de Distribución Libre y no Paramétrica, Teoría Estocástica y Análisis de Series Temporales
Cita:
P. Calvo-Báscones (2022), Inclusive methodologies for anomaly detection and prognosis of industrial systems based on behavior patterns, smart indicators and digital twin ecosystems. Madrid (España).
