Publicación: Detección de fallas mecánicas mediante "Machine Learning", utilizando el clasificador "Random Forest"
Cargando...
Archivos
Fecha
2022
Editor/a
Director/a
Tutor/a
Coordinador/a
Prologuista
Revisor/a
Ilustrador/a
Derechos de acceso
info:eu-repo/semantics/openAccess
Título de la revista
ISSN de la revista
Título del volumen
Editor
['Universidad Nacional de Educación a Distancia (España)', 'Universidad Politécnica de Madrid. Departamento de Ingeniería Mecánica']
Resumen
En este trabajo se presentará una Inteligencia Artificial (IA) para el seguimiento de defectos de origen mecánico (desequilibrio, desalineación y holgura mecánica), además de la condición sin defectos, a partir de señales de vibración. El uso de "Machine Learning" puede considerarse un instrumento dentro de la Inteligencia Artificial para el diagnóstico de fallas mecánicas en máquinas rotativas. En este trabajo se abordó el modelo de "Machine Learning", a través de "Supervised Machine Learning" y utilizando "Classification Method" con algoritmos de clasificación por "Random Forest". El modelo se entrenó con el 70% de los datos disponibles en la "Base de Datos" y el 30% se utilizó para la validación de la prueba. Para la condición "Sin Defecto", el porcentaje de precisión fue del 99% y para las excitaciones "Desequilibrio" y "Holgura Mecánica", fueron del 98,1% y 99,3%, respectivamente. Esto se debe a la gran cantidad de señales disponibles para la prueba (532, 283 y 557) respectivamente. Para "Desalineamiento", el porcentaje de respuestas correctas fue del 69,6%. También fue influenciado por los porcentajes de error de 11,2% y 19,2%, por "Desequilibrio" y "Holgura Mecánica", respectivamente. Esto se debe al bajo número de señales disponibles para la prueba, solo 28, y porque las excitaciones están relacionadas con frecuencia de rotación (fr) y sus armónicos, lo que provoca "Confusión". De manera similar al razonamiento anterior, "Holgura Mecánica + Desalineación" presentó el porcentaje de acierto del 46,9%, con una contribución de error del 40,4% por "Desequilibrio". Esto también se debe a las mismas razones que antes, siendo aún menor el número de señales disponibles para las pruebas, es decir, solo 18. Para aumentar el asertividad, es necesario tener un mayor número de señales, o aplicar el "Data Augmentation" para aumentar el número de señales o incluso extraer más parámetros discriminativos en el modelo. Los resultados muestran que la metodología propuesta permitió la detección de fallas supervisadas en máquinas rotativas, siendo una herramienta prometedora para ser aplicada en la Industria 4.0.
An Artificial Intelligence (AI) will be presented for monitoring defects of mechanical origin (imbalance, misalignment, and mechanical backlash), in addition to the condition without defect, from vibration signals. The use of "Machine Learning" can be considered an instrument within Artificial Intelligence for the diagnosis of mechanical failures in rotating machines. In this work, the "Machine Learning" model was approached, through "Supervised Machine Learning" and using "Classification Method" with classification algorithms by "Random Forest". The model was trained with 70% of the data available in the "Database" and 30% were used for test validation. For the "No Defect" condition, the percentage of success was 99% and for the "Unbalance" and "Mechanical Backlash" excitations, they were 98.1% and 99.3%, respectively. This is due to the significant number of signals available for testing (532, 283 and 557) respectively. For "Misalignment", the percentage of correct answers was 69.6%. It was also influenced by the error percentages of 11.2% and 19.2%, due to "Unbalance" and "Mechanical Backlash", respectively. This is due to the low number of signals available for testing, only 28, and because the excitations are related to fr and its harmonics, which causes "Confusion". In a similar way to the previous reasoning, "Mechanical clearance + Misalignment" presented the percentage of correctness of 46.9%, with an error contribution of 40.4% due to "Unbalance". This is also due to the same reasons as before, with the number of signals available for tests being even smaller, that is, only 18. To increase assertiveness, it is necessary to have a greater number of signals, or to apply the "Data Augmentation" method to increase the number of signals or even extract more discriminative parameters in the model. The results show that the proposed methodology allowed the detection of supervised failures in rotating machines, being a promising tool to be applied in Industry 4.0.
An Artificial Intelligence (AI) will be presented for monitoring defects of mechanical origin (imbalance, misalignment, and mechanical backlash), in addition to the condition without defect, from vibration signals. The use of "Machine Learning" can be considered an instrument within Artificial Intelligence for the diagnosis of mechanical failures in rotating machines. In this work, the "Machine Learning" model was approached, through "Supervised Machine Learning" and using "Classification Method" with classification algorithms by "Random Forest". The model was trained with 70% of the data available in the "Database" and 30% were used for test validation. For the "No Defect" condition, the percentage of success was 99% and for the "Unbalance" and "Mechanical Backlash" excitations, they were 98.1% and 99.3%, respectively. This is due to the significant number of signals available for testing (532, 283 and 557) respectively. For "Misalignment", the percentage of correct answers was 69.6%. It was also influenced by the error percentages of 11.2% and 19.2%, due to "Unbalance" and "Mechanical Backlash", respectively. This is due to the low number of signals available for testing, only 28, and because the excitations are related to fr and its harmonics, which causes "Confusion". In a similar way to the previous reasoning, "Mechanical clearance + Misalignment" presented the percentage of correctness of 46.9%, with an error contribution of 40.4% due to "Unbalance". This is also due to the same reasons as before, with the number of signals available for tests being even smaller, that is, only 18. To increase assertiveness, it is necessary to have a greater number of signals, or to apply the "Data Augmentation" method to increase the number of signals or even extract more discriminative parameters in the model. The results show that the proposed methodology allowed the detection of supervised failures in rotating machines, being a promising tool to be applied in Industry 4.0.
Descripción
Categorías UNESCO
Palabras clave
detección de Fallas, "Machine Learning", "Random Forest", inteligencia artificial
Citación
Centro
E.T.S. de Ingenieros Industriales
Departamento
Mecánica