Publication: Aplicación de redes neuronales convolucionales a la monitorización del nivel de desgaste de herramientas en sistemas industriales de taladrado
Loading...
Date
2022
Editor
Director
Advisor
Coordinator
Commentator
Reviewer
Illustrator
Access rights
info:eu-repo/semantics/openAccess
Journal Title
Journal ISSN
Volume Title
Publisher
Universidad Nacional de Educación a Distancia (España), Universidad Politécnica de Madrid. Departamento de Ingeniería Mecánica
Abstract
La industria aeronáutica es un sector estratégico a nivel global y en constante crecimiento. Con esta premisa resulta interesante buscar formas de conseguir ahorros recurrentes en la fabricación de las aeronaves. En este contexto, uno de los procesos más comunes es el taladrado, en el que los altos estándares de calidad del sector hacen que las herramientas se encuentren, generalmente, infrautilizadas, ya que su reemplazo atiende a valores nominales de tiempo de vida útil con un alto margen de seguridad. La estimación del nivel de desgaste de las herramientas en tiempo real puede permitir un mayor aprovechamiento de las mismas. Este estudio se encuentra enmarcado en un proceso industrial de taladrado automático en el cual se fabrican componentes para aviones comerciales en la fábrica de Getafe de Airbus. La principal información que se extrae mediante monitorización indirecta, para no afectar al proceso, es la potencia eléctrica consumida por el husillo. Con el objetivo de extraer características relacionadas con el nivel de desgaste de la herramienta en tiempo real, la señal de potencia eléctrica ha sido procesada mediante redes neuronales convolucionales entrenadas bajo la técnica comparativa de las redes de tipo siamesas. El método propuesto palía el problema de la escasez de datos de entrenamiento, y los resultados de los experimentos realizados demuestran su capacidad para generar un espacio de características ordenado en función del nivel de desgaste de la herramienta.
The aeronautical industry is a strategic, and constantly growing, sector at a global level. With this in mind, it is interesting to look for ways to achieve recurrent savings in the aircrafts manufacturing. Within this context, one of the most common processes is drilling, in which the high-quality standards of the sector mean that cutting tools are generally underutilized since they are replaced according to nominal values of their service life with wide security margins. The tool real-time wear level estimation can lead to a better and longer use of the tools. This study is framed in an industrial process of automatic drilling in which different component of commercial aircrafts are manufactured in the Getafe factory of Airbus. The main process information extracted is the electrical power consumed by the spindle, through indirect monitoring so to not to affect the process. In order to extract wear level-related features in real time, the power signal has been processed by convolutional neural networks under the comparative training technique of Siamese networks. This proposed method alleviates the problem of shortage of training data, and the results of the conducted experiments demonstrate its capacity to generate a features space so that is sorted according to the tool wear level.
The aeronautical industry is a strategic, and constantly growing, sector at a global level. With this in mind, it is interesting to look for ways to achieve recurrent savings in the aircrafts manufacturing. Within this context, one of the most common processes is drilling, in which the high-quality standards of the sector mean that cutting tools are generally underutilized since they are replaced according to nominal values of their service life with wide security margins. The tool real-time wear level estimation can lead to a better and longer use of the tools. This study is framed in an industrial process of automatic drilling in which different component of commercial aircrafts are manufactured in the Getafe factory of Airbus. The main process information extracted is the electrical power consumed by the spindle, through indirect monitoring so to not to affect the process. In order to extract wear level-related features in real time, the power signal has been processed by convolutional neural networks under the comparative training technique of Siamese networks. This proposed method alleviates the problem of shortage of training data, and the results of the conducted experiments demonstrate its capacity to generate a features space so that is sorted according to the tool wear level.
Description
UNESCO Categories
Keywords
predicción desgaste, taladrado, procesamiento señal, redes neuronales, convolucionales, redes siamesas
Citation
Center
E.T.S. de Ingenieros Industriales
Department
Mecánica