Detección de anomalías en procesos industriales mediante análisis de conectividad y modelos predictivos de Machine Learning.
Loading...
Date
2025
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
Este trabajo aborda el problema de la detección temprana de anomalías en procesos industriales, lo cual es fundamental para prevenir fallas críticas que afectan la continuidad operativa, la eficiencia productiva y los costos operacionales de la empresa. En particular, identificar condiciones anormales antes de que ocurran eventos críticos permite aplicar una estrategia preventiva. Desde el punto de vista técnico, el desafío radica en reconocer patrones que preceden a una falla, los cuales no siempre son evidentes mediante el monitoreo individual de sensores. En este contexto, se busca anticipar los cortes de hoja en una planta de celulosa mediante el análisis de las relaciones dinámicas entre variables dentro del proceso industrial y la detección de comportamientos anómalos previos al evento, con el objetivo de encontrar la causa raíz de un corte de hoja. La hipótesis central plantea que el uso de técnicas de análisis de conectividad y de modelos autoregresivos permite identificar sensores precursores y comprender la propagación de anomalías en la maquinaria. El estudio se basa en variaciones industriales del proceso de secado de celulosa de la planta CMPC Santa Fe, monitoreadas por 96 sensores con muestreo cada cinco minutos. En una primera etapa, se aplica el enfoque de conectividad bidireccional de Diebold–Yilmaz mediante modelos VAR-LASSO, con el objetivo de detectar patrones en las relaciones entre sensores y cuantificar su influencia sobre una serie binaria que representa la ocurrencia del corte de hoja. Posteriormente, se implementa un esquema de detección de anomalías utilizando modelos regresivos de Machine Learning, incluyendo K-Nearest Neighbors, Regresión Lineal, Random Forest, Support Vector Regression y XGBoost. Los modelos son entrenados bajo condiciones normales del proceso y evaluados en períodos cercanos al evento, comparando los errores de predicción mediante la métrica MSE. Los resultados permiten identificar sensores y zonas del proceso con mayor capacidad de detección ante condiciones anómalas previas a los cortes de hoja, aportando información relevante para el monitoreo preventivo. El trabajo demuestra el potencial en el área industrial del uso combinado de análisis de conectividad y aprendizaje automático como herramienta para la detección temprana de fallas y la disminución de sus consecuencias.
This work addresses the problem of early anomaly detection in industrial processes, which is essential to prevent critical failures that affect operational continuity, production efficiency, and the company’s operational costs. In particular, identifying abnormal conditions before critical events occur enables the implementation of a preventive strategy. From a technical perspective, the challenge lies in recognizing patterns that precede a failure, which are not always evident through the individual monitoring of sensors. In this context, this study aims to anticipate sheet breaks in a pulp mill by analyzing the dynamic relationships among variables within the industrial process and by detecting anomalous behaviors prior to the event, with the objective of identifying the root cause of a sheet break. The central hypothesis states that the use of connectivity analysis techniques and autoregressive models allows the identification of precursor sensors and the understanding of anomaly propagation throughout the machinery. The study is based on industrial variations from the pulp drying process at the CMPC Santa Fe plant, monitored by 96 sensors with a five-minute sampling interval. In the first stage, the bidirectional connectivity approach of Diebold–Yilmaz is applied using VAR-LASSO models, with the objective of detecting patterns in sensor relationships and quantifying their influence on a binary series representing the occurrence of a sheet break. Subsequently, an anomaly detection scheme is implemented using regression-based Machine Learning models, including K-Nearest Neighbors, Linear Regression, Random Forest, Support Vector Regression, and XGBoost. The models are trained under normal process conditions and evaluated during periods close to the event, with prediction errors compared using the MSE metric. The results allow the identification of sensors and process areas with greater detection capability under anomalous conditions preceding sheet breaks, providing relevant information for preventive monitoring. This work demonstrates the potential of the combined use of connectivity analysis and machine learning in the industrial domain as a tool for early failure detection and the reduction of its consequences.
This work addresses the problem of early anomaly detection in industrial processes, which is essential to prevent critical failures that affect operational continuity, production efficiency, and the company’s operational costs. In particular, identifying abnormal conditions before critical events occur enables the implementation of a preventive strategy. From a technical perspective, the challenge lies in recognizing patterns that precede a failure, which are not always evident through the individual monitoring of sensors. In this context, this study aims to anticipate sheet breaks in a pulp mill by analyzing the dynamic relationships among variables within the industrial process and by detecting anomalous behaviors prior to the event, with the objective of identifying the root cause of a sheet break. The central hypothesis states that the use of connectivity analysis techniques and autoregressive models allows the identification of precursor sensors and the understanding of anomaly propagation throughout the machinery. The study is based on industrial variations from the pulp drying process at the CMPC Santa Fe plant, monitored by 96 sensors with a five-minute sampling interval. In the first stage, the bidirectional connectivity approach of Diebold–Yilmaz is applied using VAR-LASSO models, with the objective of detecting patterns in sensor relationships and quantifying their influence on a binary series representing the occurrence of a sheet break. Subsequently, an anomaly detection scheme is implemented using regression-based Machine Learning models, including K-Nearest Neighbors, Linear Regression, Random Forest, Support Vector Regression, and XGBoost. The models are trained under normal process conditions and evaluated during periods close to the event, with prediction errors compared using the MSE metric. The results allow the identification of sensors and process areas with greater detection capability under anomalous conditions preceding sheet breaks, providing relevant information for preventive monitoring. This work demonstrates the potential of the combined use of connectivity analysis and machine learning in the industrial domain as a tool for early failure detection and the reduction of its consequences.
Description
Tesis presentada para optar al título de Ingeniero/a Civil Matemático/a.
Keywords
Anomalías, Procesos de manufactura, Disminución del riesgo, Sensores, Machine learning