Publication: Análisis, optimización y evaluación de modelos de redes de neuronas artificiales para la clasificación y predicción de impactos de alta velocidad sobre distintos materiales
Loading...
Identifiers
Publication date
2010-02
Defense date
2010-02-12
Authors
Tutors
Journal Title
Journal ISSN
Volume Title
Publisher
Impact
Export
Abstract
Los problemas de impacto sobre estructuras tanto de baja y media velocidad, denominados Crashworthiness, como de alta velocidad, conocidos como impacto balístico, son de gran relevancia en el ámbito de la ingeniería estructural y de materiales avanzados. En ellos, se analiza y predice el comportamiento de materiales en condiciones extremas de trabajo, lo que permite el diseño y fabricación de nuevos componentes y estructuras sujetas a impactos. El diseño de estas estructuras es por lo general un proceso complejo y costoso, que requiere un alto consumo de recursos técnicos y humanos. Debido a que estas estructuras pueden ser partes esenciales de un sistema, y a que pueden sufrir durante el desempeño de su actividad impactos de otros cuerpos, es vital conocer como se comportarían tras sufrir el choque de un proyectil. Para intentar minimizar los daños producidos por el impacto, se han destinado muchos esfuerzos a estudiar el proceso de penetración de un proyectil sobre un blanco u objetivo. Entre las alternativas clásicas empleadas por los científicos destacan, en primer lugar las pruebas empíricas, y durante las últimas décadas las técnicas computacionales para la representación del proceso, conocidas como clásicas o tradicionales: simulación numérica y modelado analítico. No obstante, estas técnicas presentan ciertos problemas que impiden su aplicación en todos los escenarios posibles. Por ello, durante los últimos años ha comenzado a utilizarse, en numerosas ramas de la ingeniería industrial y estructural, metodologías multidisciplinares para mejorar y optimizar los procesos encargados de modelar este fenómeno. Entre ellas destaca el empleo de técnicas predictivas y de aprendizaje automático basadas en Redes Neuronales Artificiales (RNA). Esta investigación ha sido definida como una estructura integral de trabajo donde colaboran diferentes componentes para la resolución y optimización del problema ingenieril del impacto de sólidos: Las herramientas de simulación numérica y la experimentación de ensayos balísticos han sido los primeros componentes empleados. Se trata de dos elementos cruciales al servir de nexo de unión con los modelos de Perceptrón MultiCapa (PMC) diseñados. La simulación numérica ha permitido a partir del análisis del fenómeno físico del problema, y basándose en una serie de modelos matemáticos, recrear mediante simulaciones numéricas las características de los materiales y el comportamiento de los objetos que intervienen. Por tanto gracias a su empleo se han simulado una serie de ensayos balísticos encaminados a entrenar y validar los PMC. Las pruebas empíricas realizadas en laboratorio han servido para comprobar la verosimilitud de los modelos de simulación numérica con la realidad. El siguiente componente se corresponde con las técnicas de Inteligencia Artificial (IA) basadas en RNA, más concretamente de la arquitectura PMC. Su objetivo fundamental ha sido sustituir a los componentes anteriores para recrear el proceso de impacto de sólidos a alta velocidad. Estos modelos de RNA han sido empleados para resolver tres de los escenarios más representativos, i.e. conocer el resultado del impacto y estudiar el comportamiento de los objetos, demostrando su validez como solución alternativa a las técnicas tradicionales. La optimización de los modelos neuronales se ha llevado a cabo mediante la definición de una metodología de optimización, basada en una serie de actividades y tareas, orientadas a cumplir los objetivos e hipótesis expuestas al comienzo de esta tesis doctoral. Para homogeneizar el gran número de alternativas existentes y simplificar su aplicación sobre cada uno de los escenarios, éstas se han agrupado en tres grandes bloques o fases: patrones, arquitectura y algoritmo. En cada uno de ellos se han incluido y probado las variantes que según la literatura mejor se adaptan al problema tratado, tomándose únicamente aquellas que mejores resultados han mostrado. El ámbito de esta investigación es en consecuencia multidisciplinar, ya que la base teórica del problema a resolver proviene de la rama ingenieril de la mecánica de sólidos y para su resolución se aplicarían técnicas informáticas basadas en IA. De esta forma, la resolución del problema del impacto balístico, tratado en esta investigación, requiere de la colaboración de técnicas heterogéneas provenientes de diferentes ramas de la ciencia y la ingeniería en busca de un beneficio común. Como elemento final, la presente tesis doctoral aporta una metodología de optimización de modelos neuronales basados en la arquitectura PMC para diferentes escenarios del impacto balístico. La metodología de investigación seguida en esta tesis doctoral para alcanzar los objetivos y demostrar las hipótesis planteadas inicialmente ha consistido en: 1. Estado del arte. El principal objetivo de esta actividad se ha centrado en detallar los trabajos realizados con RNA dentro del ámbito del impacto de sólidos. Las principales cuestiones tratadas en este apartado han sido: Definición del problema a resolver, centrándose en los elementos fundamentales del mismo: proyectil y protección. Descripción física del problema, centrada en la física de los materiales que intervienen y la dinámica del fenómeno balístico. Análisis crítico de las principales alternativas clásicas para la resolución de problemas de impacto. Identificación de los principales parámetros que influyen en la capacidad de predicción de las RNA dentro del dominio tratado. Antecedentes de la aplicación de soluciones basadas en RNA dentro del dominio estudiado. 2. Simulación numérica de los ensayos. El comportamiento de los proyectiles y de las protecciones ante los impactos se ha recreado mediante herramientas de simulación numérica. Se ha generado un conjunto de casos de impacto con diferentes valores de los parámetros descriptivos de cada una de las estructuras, proyectil y protección, que intervienen en el problema. 3. Desarrollo de nuevos modelos de RNA. El trabajo realizado por el autor en este apartado ha consistido en crear nuevos diseños de RNA para solventar dos variantes del problema del impacto: clasificación del resultado y predicción de la velocidad final y masa residual. 4. Simulación de los nuevos modelos de RNA diseñados. Las RNA han sido probadas para distintas geometrías, tanto de proyectil como de panel protector, velocidades y materiales. 5. Análisis de posibles alternativas para la optimización del aprendizaje y de los resultados. Existen múltiples parámetros de una RNA, tanto de la arquitectura como del algoritmo de entrenamiento y de los datos empleados para extraer el conocimiento del dominio, que modifican su capacidad de generalización y que afectan por tanto a su precisión de predicción. La búsqueda de valores óptimos para estos parámetros, es un punto de relevancia dentro del trabajo ya que ha permitido optimizar tanto el aprendizaje como los resultados. 6. Evaluación y Validación. Las actividades de evaluación y validación incluidas han determinado la efectividad de las diferentes soluciones de RNA planteadas y demostrado la validez de la metodología de optimización propuesta. 7. Documentación y Conclusiones. Esta actividad ha consistido en la documentación de cada uno de los aspectos relativos a la tesis doctoral y presenta las conclusiones extraídas a raíz de la investigación. La evaluación de los modelos neuronales ha demostrado la viabilidad del PMC para resolver diferentes escenarios balísticos. Gracias a su utilización las herramientas tradicionales de simulación o modelado pueden utilizarse como complemento al PMC para mejorar el diseño de sistemas de protección. El empleo de un PMC optimizado permite predecir, tanto el resultado de un impacto, como el comportamiento de los objetos que intervienen, con un porcentaje de acierto superior al resto de propuestas evaluadas. Para ratificar estas conclusiones, el PMC ha sido entrenado y probado bajo diversas combinaciones geométricas y de materiales para el proyectil y la protección, y diferentes velocidades de impacto. La ventaja de emplear un PMC para resolver los diferentes escenarios es que los resultados se obtienen en tiempo real y por tanto el coste computacional es muy inferior al de las técnicas tradiciones de simulación y modelado. Además la comparación realizada con otras propuestas de maquinas de aprendizaje ha demostrado que un PMC, correctamente configurado, es la mejor alternativa para los diferentes escenarios balísticos planteados. Asimismo la metodología de optimización, orientada a la sistematización de los diferentes métodos y técnicas existentes a la hora de configurar y parametrizar una RNA, ha sido validada para diferentes escenarios balísticos. En consecuencia, gracias a las conclusiones obtenidas en esta investigación se podran recrear en el futuro de una manera fiable nuevos ensayos balísticos. Esto facilitará el posterior diseño de protecciones y estudio de las consecuencias asociadas a la introducción de diferentes configuraciones de proyectiles y paneles protectores con diferentes características. Por último, las deducciones obtenidas al emplear en primer lugar soluciones basadas en IA y en segundo lugar la aplicación de una metodología de optimización pueden extrapolarse a otras áreas como el diseño de estructuras sujetas a impactos de baja velocidad o la seguridad pasiva de vehículos.-----------------------------------------------------------------------------------------------------------------------------
The problems of impact of solids, both low and medium speed, the so-called Crashworthiness, as well as those of high speed, known as ballistic impact, are of great relevance in the field of structural engineering and of advanced materials. In these kind of problems the behavior of materials under extreme conditions of work is analyzed and predicted, allowing the design and manufacture of new components and structures subject to impact. The design of these structures is generally a complex and costly process that requires a high consumption of technical and human resources. Due to the fact that these structures can be essential parts of a system, and to the fact that they can suffer impacts from other bodies during the performance of their activity, it is vital to know how they would behave after suffering the shock of a projectile. To try to minimize the damage caused by the impact, many efforts have been directed to study the process of penetration of a projectile on a target or goal. Among the proposed alternatives highlight the empirical evidence and, in recent years, the computational techniques for the representation of the process, known as classic or traditional: numerical simulation and analytical modeling. However, these techniques have certain problems that prevent them from being implemented in all possible scenarios. Therefore, during recent years multidisciplinary methodologies have begun to be used in many branches of industrial and structural engineering, for improving and optimizing processes in charge of modeling this phenomenon. Among these multidisciplinary methodologies, the use of predictive and machine learning techniques based on Artificial Neural Networks (ANN) stands out. This research has been defined as an integral framework in which different components work together for the solution and optimization of the engineering problem of impact solids: The numerical simulation tools and the experimentation of ballistic tests have been the first components used, being two crucial elements as they serve as a link with the Multilayer Perceptron (MLP) models designed. The numerical simulation, has allowed from the analysis of the physical phenomenon of the problem, and based on a series of mathematical models, to recreate by means of numerical simulations, the material’s characteristics and the behavior of the objects involved. Thus, thanks to its use a series of missile tests aimed to train and validate the MLP have been simulated. Empirical tests conducted in the laboratory have been useful for checking the plausibility of the numerical simulation models with reality. The next component corresponds to the techniques of Artificial Intelligence (AI) based on ANN, more specifically the MLP architecture. Its main objective has been to replace the former components to recreate the impact process of solids at high speed. These ANN models have been used to solve three of the most representative scenarios, i.e. to know the impact’s result and to study the behavior of objects, proving its validity as an alternative solution to traditional techniques. Optimization of neural models has been carried out by defining an optimization methodology based on a series of activities and tasks aimed at fulfilling the targets and hypothesis outlined at the beginning of this doctoral thesis. To homogenize the large number of existing alternatives and to simplify its implementation on each of the scenarios, these are grouped into three large blocks or phases: patterns, architecture and algorithm. The variants that according to the literature best adapt to the problem treated, have been included and tested in each of them, only taking those that have shown better results. The scope of this research is accordingly multidisciplinary, due to the fact that the theoretical basis of the problem comes from the engineering branch of solid mechanics and IA-based computing techniques will be implemented for its resolution. Thus, the solution of the problem of ballistic impact discussed in this research requires the collaboration of heterogeneous techniques coming from different branches of science and engineering in search of a common benefit. Consequently, this doctoral thesis provides a methodology of optimization of neural models based on MLP architecture for different scenarios of ballistic impact. The research methodology followed in this doctoral thesis to achieve the objectives and demonstrate the hypotheses initially set out has consisted of: 1. State of the art. The main objective of this activity is to detail the works performed with ANN within the field of impact of solids. The main issues addressed in this section will be: Definition of the problem to be solved. This will involve a detailed description of the phenomenon of penetration treated, focusing on its key elements: projectile and protection. Physical description of the problem, focused on the physics of the materials involved and the dynamics of the ballistic phenomenon. Critical analysis of the main classic alternatives for solving problems of impact. Identification of the main parameters that influence on the predictive ability of the ANN within the domain treated. Precedents of the implementation of ANN-based solutions within the domain studied. 2. Numerical simulation of the tests. The behavior of projectiles and protections in the impacts has been recreated by means of numerical simulation tools. It has been generated a set of cases of impact with different values for the descriptive parameters of each of the structures, projectile and protection, which take part in the problem. 3. Development of new models of ANN. The work to be carried out by the author in this section will involve creating new ANN designs in order to solve two variants of the problem of impact: result’s classification and prediction of the final velocity and residual mass. 4. Simulation of the new ANN designed models. The ANN will be tested for different geometries, both projectile and protector panel, velocities and materials. 5. Analysis of possible alternatives for the optimization of learning and results. There are many parameters of an ANN, both of the architecture and of the training algorithm, which modify its generalization capacity and therefore affect its accuracy of prediction. The search for optimal values for these parameters is a point of importance in the work to be carried out as it will allow optimizing learning and performance. 6. Evaluation and Validation. The assessment and validation activities pretend to estimate the effectiveness of different ANN solutions raised and the validity of the optimization methodology proposed. 7. Documentation and Conclusions. In this activity each aspect of the thesis will be documented. The different alternatives of ANN will be described as well as their contributions to the results obtained. Finally, the conclusions drawn from the evaluation will be exposed. The evaluation of the neural models has demonstrated the viability of the MLP to solve different ballistic scenarios. Thanks to its use, traditional simulation or modeling tools can be used as a complement to the MLP to improve the design of the protection systems. The use of an optimized MLP allows predicting both the result of an impact and the behavior of the objects which take part with a higher success rate than the other proposals evaluated. To confirm these conclusions, the MLP has been trained and tested under various combinations of geometries and materials for the projectile and the protection, and different impact velocities. The advantage of using a MLP to solve the different scenarios is that the results are obtained in real time and the computational cost is thus much lower than the one corresponding to the traditional techniques of simulation and modeling. Furthermore the comparison with other proposals of learning machines has shown that the MLP is the best alternative for the different ballistic scenarios set out. Besides the optimization methodology aimed at the systematization of the different methods and techniques existing when it comes to configuring an ANN, has been validated for different ballistic scenarios. In summary, the conclusions obtained show that thanks to this research new ballistic tests can be recreated in the future in a reliable way. This will facilitate the subsequent design of protections and study of the consequences associated with the introduction of different configurations of projectiles and protective panels with different characteristics. Finally, the findings obtained by using in the first place AI-based solutions and in the second place the application of an optimization methodology can be extrapolated to other areas such as the design of structures subject to low-speed impacts or the passive safety of vehicles.
The problems of impact of solids, both low and medium speed, the so-called Crashworthiness, as well as those of high speed, known as ballistic impact, are of great relevance in the field of structural engineering and of advanced materials. In these kind of problems the behavior of materials under extreme conditions of work is analyzed and predicted, allowing the design and manufacture of new components and structures subject to impact. The design of these structures is generally a complex and costly process that requires a high consumption of technical and human resources. Due to the fact that these structures can be essential parts of a system, and to the fact that they can suffer impacts from other bodies during the performance of their activity, it is vital to know how they would behave after suffering the shock of a projectile. To try to minimize the damage caused by the impact, many efforts have been directed to study the process of penetration of a projectile on a target or goal. Among the proposed alternatives highlight the empirical evidence and, in recent years, the computational techniques for the representation of the process, known as classic or traditional: numerical simulation and analytical modeling. However, these techniques have certain problems that prevent them from being implemented in all possible scenarios. Therefore, during recent years multidisciplinary methodologies have begun to be used in many branches of industrial and structural engineering, for improving and optimizing processes in charge of modeling this phenomenon. Among these multidisciplinary methodologies, the use of predictive and machine learning techniques based on Artificial Neural Networks (ANN) stands out. This research has been defined as an integral framework in which different components work together for the solution and optimization of the engineering problem of impact solids: The numerical simulation tools and the experimentation of ballistic tests have been the first components used, being two crucial elements as they serve as a link with the Multilayer Perceptron (MLP) models designed. The numerical simulation, has allowed from the analysis of the physical phenomenon of the problem, and based on a series of mathematical models, to recreate by means of numerical simulations, the material’s characteristics and the behavior of the objects involved. Thus, thanks to its use a series of missile tests aimed to train and validate the MLP have been simulated. Empirical tests conducted in the laboratory have been useful for checking the plausibility of the numerical simulation models with reality. The next component corresponds to the techniques of Artificial Intelligence (AI) based on ANN, more specifically the MLP architecture. Its main objective has been to replace the former components to recreate the impact process of solids at high speed. These ANN models have been used to solve three of the most representative scenarios, i.e. to know the impact’s result and to study the behavior of objects, proving its validity as an alternative solution to traditional techniques. Optimization of neural models has been carried out by defining an optimization methodology based on a series of activities and tasks aimed at fulfilling the targets and hypothesis outlined at the beginning of this doctoral thesis. To homogenize the large number of existing alternatives and to simplify its implementation on each of the scenarios, these are grouped into three large blocks or phases: patterns, architecture and algorithm. The variants that according to the literature best adapt to the problem treated, have been included and tested in each of them, only taking those that have shown better results. The scope of this research is accordingly multidisciplinary, due to the fact that the theoretical basis of the problem comes from the engineering branch of solid mechanics and IA-based computing techniques will be implemented for its resolution. Thus, the solution of the problem of ballistic impact discussed in this research requires the collaboration of heterogeneous techniques coming from different branches of science and engineering in search of a common benefit. Consequently, this doctoral thesis provides a methodology of optimization of neural models based on MLP architecture for different scenarios of ballistic impact. The research methodology followed in this doctoral thesis to achieve the objectives and demonstrate the hypotheses initially set out has consisted of: 1. State of the art. The main objective of this activity is to detail the works performed with ANN within the field of impact of solids. The main issues addressed in this section will be: Definition of the problem to be solved. This will involve a detailed description of the phenomenon of penetration treated, focusing on its key elements: projectile and protection. Physical description of the problem, focused on the physics of the materials involved and the dynamics of the ballistic phenomenon. Critical analysis of the main classic alternatives for solving problems of impact. Identification of the main parameters that influence on the predictive ability of the ANN within the domain treated. Precedents of the implementation of ANN-based solutions within the domain studied. 2. Numerical simulation of the tests. The behavior of projectiles and protections in the impacts has been recreated by means of numerical simulation tools. It has been generated a set of cases of impact with different values for the descriptive parameters of each of the structures, projectile and protection, which take part in the problem. 3. Development of new models of ANN. The work to be carried out by the author in this section will involve creating new ANN designs in order to solve two variants of the problem of impact: result’s classification and prediction of the final velocity and residual mass. 4. Simulation of the new ANN designed models. The ANN will be tested for different geometries, both projectile and protector panel, velocities and materials. 5. Analysis of possible alternatives for the optimization of learning and results. There are many parameters of an ANN, both of the architecture and of the training algorithm, which modify its generalization capacity and therefore affect its accuracy of prediction. The search for optimal values for these parameters is a point of importance in the work to be carried out as it will allow optimizing learning and performance. 6. Evaluation and Validation. The assessment and validation activities pretend to estimate the effectiveness of different ANN solutions raised and the validity of the optimization methodology proposed. 7. Documentation and Conclusions. In this activity each aspect of the thesis will be documented. The different alternatives of ANN will be described as well as their contributions to the results obtained. Finally, the conclusions drawn from the evaluation will be exposed. The evaluation of the neural models has demonstrated the viability of the MLP to solve different ballistic scenarios. Thanks to its use, traditional simulation or modeling tools can be used as a complement to the MLP to improve the design of the protection systems. The use of an optimized MLP allows predicting both the result of an impact and the behavior of the objects which take part with a higher success rate than the other proposals evaluated. To confirm these conclusions, the MLP has been trained and tested under various combinations of geometries and materials for the projectile and the protection, and different impact velocities. The advantage of using a MLP to solve the different scenarios is that the results are obtained in real time and the computational cost is thus much lower than the one corresponding to the traditional techniques of simulation and modeling. Furthermore the comparison with other proposals of learning machines has shown that the MLP is the best alternative for the different ballistic scenarios set out. Besides the optimization methodology aimed at the systematization of the different methods and techniques existing when it comes to configuring an ANN, has been validated for different ballistic scenarios. In summary, the conclusions obtained show that thanks to this research new ballistic tests can be recreated in the future in a reliable way. This will facilitate the subsequent design of protections and study of the consequences associated with the introduction of different configurations of projectiles and protective panels with different characteristics. Finally, the findings obtained by using in the first place AI-based solutions and in the second place the application of an optimization methodology can be extrapolated to other areas such as the design of structures subject to low-speed impacts or the passive safety of vehicles.
Description
Keywords
Ensayo de materiales, Impacto balístico, Crashworthiness, Aprendizaje automático, Redes neuronales artificiales