Análisis de vibraciones y correlación con variables operacionales, aplicado a motorreductor de accionamiento de polea en la gran minería.
Loading...
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad de Concepción
Abstract
La siguiente memoria de título se desarrolla con información proporcionada por una Minera dedicada al rubro de extracción del cobre, ubicada a 3000 msnm en la región de Antofagasta, Chile. En la búsqueda de nuevas estrategias en el área mantenimiento, nace la oportunidad de evaluar modelos predictivos implementando algoritmos de inteligencia artificial y análisis de datos para la detección de fallas de sus activos físicos. La progresiva integración de modelos de machine learning en la industria moderna ha desencadenado una serie de impactos positivos. Estos incluyen una significativa automatización de procesos, lo que no solo aumenta la eficiencia, sino que también reduce los costos operativos. Además, estos modelos contribuyen a reforzar la seguridad en los entornos industriales al identificar patrones y prevenir potenciales riesgos.
Se da inicio al trabajo con una contextualización del área de monitoreo de condiciones, una búsqueda exhaustiva de información técnica del caso de estudio, recolección y extracción de datos monitoreados por medio de sensores de vibraciones y variables operacionales. Este caso corresponde a un sistema motorreductor de accionamiento de poleas, el cual es clasificado como un sistema crítico, debido a que es el único medio que conecta el mineral desde el chancado primario al secundario. Se describen los componentes que lo integran como motores, reductores y poleas. Para la extracción de datos se seleccionan aquellos que presenten mediciones cercanas en el tiempo y representen mecánicamente el comportamiento de la máquina.
Se revisa el historial de fallas del sistema identificando los modos de falla, por medio del análisis de vibraciones, se identifican las variables más susceptibles al cambio del comportamiento de la máquina en presencia de falla. De la misma forma se realiza un análisis de correlaciones entre las variables seleccionadas para comprender el comportamiento físico del sistema.
Utilizando la base de datos del monitoreo de condiciones junto a variables operacionales se desarrollan modelos de detección de falla con enfoque de aprendizaje supervisado. Estos modelos se diseñan con el propósito de clasificar el comportamiento normal y detectar la presencia de fallas. Se emplean los clasificadores PCA-SVC y RNN-LSTM, debido a su buen rendimiento en casos de estudios similares publicadas en revistas científicas.
Se evalúan los desempeños de los modelos mediante sus métricas de rendimiento y porcentajes de falsas alarmas. Para todos los modos de falla, el modelo RNN-LSTM es el que obtiene los mejores resultados. La implementación adecuada de modelos de machine learning implica un profundo conocimiento del problema, un preprocesamiento de datos y una cantidad de registros adecuada, lo que ayuda a determinar la elección del modelo más eficiente.
Los resultados de índices de falsa alarma marcan la diferencia en la preponderancia entre un modelo y otro. El modelo PCA-SVC presenta valores de hasta un 35% de falsa alarma, mientras que el modelo RNN-LSTM muestra un máximo de un 16% en los modos de falla estudiados.
The following title report is developed, with information provided by a mining company dedicated to copper extraction, located at 3000 meters above sea level in the Antofagasta region, Calama, Chile. In the search for new strategies in the maintenance area, the opportunity arises to evaluate predictive models implementing artificial intelligence algorithms and data analysis to detect failures in their physical assets. The progressive integration of machine learning models in modern industry has unleashed a series of positive impacts. These include significant process automation, which reduces operating costs and increases efficiency. In addition, these models contribute to strengthening safety in industrial environments by identifying patterns and preventing potential risks. The work begins with a context of the condition monitoring area, an exhaustive search for technical information of the case study, and the collection and extraction of data monitored through vibration sensors and operational variables. This case corresponds to a pulley drive gear motor system, which is classified as a critical system because it is the only means that connects the ore from the primary to the secondary crushing. The components that make it up, such as motors, reducers, and pulleys, are described. For data extraction, those that present measurements close in time and mechanically represent the behavior of the machine are selected. The system's failure history is reviewed, identifying the failure modes, through vibration analysis, and the most susceptible variables to the change in the machine's behavior in the presence of failure are identified. In the same way, a correlation analysis is carried out between the selected variables to understand the physical behavior of the system. Using the condition monitoring database, fault detection models are developed with a supervised learning approach. These models are designed to classify normal behavior and detect the presence of faults. The PCA-SVC and RNN-LSTM classifiers are used, based on their implementation and good performances in similar case studies. The performances of the models are evaluated through their performance metrics and false alarm percentages. For all failure modes, the RNN-LSTM model obtains the best results. The proper implementation of machine learning models involves a deep understanding of the problem, data preprocessing, and an adequate number of records, which helps determine the choice of the most efficient model. The results of false positive and false negative rates make the difference in the predominance between one model and another. The PCA-SVC model presents values of up to 35% false alarm, while the RNN-LSTM model shows 16% in the studied failure modes.
The following title report is developed, with information provided by a mining company dedicated to copper extraction, located at 3000 meters above sea level in the Antofagasta region, Calama, Chile. In the search for new strategies in the maintenance area, the opportunity arises to evaluate predictive models implementing artificial intelligence algorithms and data analysis to detect failures in their physical assets. The progressive integration of machine learning models in modern industry has unleashed a series of positive impacts. These include significant process automation, which reduces operating costs and increases efficiency. In addition, these models contribute to strengthening safety in industrial environments by identifying patterns and preventing potential risks. The work begins with a context of the condition monitoring area, an exhaustive search for technical information of the case study, and the collection and extraction of data monitored through vibration sensors and operational variables. This case corresponds to a pulley drive gear motor system, which is classified as a critical system because it is the only means that connects the ore from the primary to the secondary crushing. The components that make it up, such as motors, reducers, and pulleys, are described. For data extraction, those that present measurements close in time and mechanically represent the behavior of the machine are selected. The system's failure history is reviewed, identifying the failure modes, through vibration analysis, and the most susceptible variables to the change in the machine's behavior in the presence of failure are identified. In the same way, a correlation analysis is carried out between the selected variables to understand the physical behavior of the system. Using the condition monitoring database, fault detection models are developed with a supervised learning approach. These models are designed to classify normal behavior and detect the presence of faults. The PCA-SVC and RNN-LSTM classifiers are used, based on their implementation and good performances in similar case studies. The performances of the models are evaluated through their performance metrics and false alarm percentages. For all failure modes, the RNN-LSTM model obtains the best results. The proper implementation of machine learning models involves a deep understanding of the problem, data preprocessing, and an adequate number of records, which helps determine the choice of the most efficient model. The results of false positive and false negative rates make the difference in the predominance between one model and another. The PCA-SVC model presents values of up to 35% false alarm, while the RNN-LSTM model shows 16% in the studied failure modes.
Description
Tesis presentada para optar al título de Ingeniero Civil Mecánico
Keywords
Vibración Mediciones, Maquinaria minera, Motores (Mecánica)