Thesis DESARROLLO DE HERRAMIENTAS DE VISUALIZACIÓN DE APLICACIONES DE MACHINE LEARNING EN MODELOS DE COMPORTAMIENTO DE PAVIMENTOS
Date
2021-03
Authors
DÍAZ BRITO, CAMILA PAZ
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Los Sistemas de Gestión de Pavimentos (SGP) son conjuntos de herramientas analíticas que ayuda a la toma de decisiones de redes de pavimentos; su fin principal es hacer un seguimiento y continua evaluación de los activos viales disponibles para que proporcionen seguridad y confort a los usuarios (Kargah-Ostadi, 2014; Hernan Solminihac et al., 2018). Un SGP debe desarrollar programas de priorización de mantenimiento y rehabilitación de pavimentos dentro del activo vial disponible, el cual dependerá del nivel de decisión que requiera el sector: nivel de red para umbrales de condición general y nivel de proyecto con datos más detallados y precisos. En ambos casos se requieren modelos de comportamiento precisos que predigan el deterioro de los pavimentos en función del tiempo y definir estrategias de mantenimiento y rehabilitación óptimas. (Osorio-Lird et al., 2018).
Generalmente, los modelos de comportamiento se desarrollan usando fórmulas funcionales continuas mediante técnicas clásicas como las regresiones y enfoques probabilísticos como cadenas de Markov; sin embargo, estos modelos son subjetivos, con dificultad de calibrar a condiciones distintas a las originales, y, además, requieren información histórica del pavimento (Hernan Solminihac et al., 2018). Actualmente, hay mayor interés por desarrollar modelos aplicando técnicas de Aprendizaje Automático (ML, Machine Learning) puesto que es una herramienta de programación, basado en métodos estadísticos, que utiliza datos de ejemplos o experiencias pasadas para optimizar un criterio de rendimiento automático y tiene la capacidad de adaptarse cuando las condiciones cambian en el tiempo (Alpaydin, 2010; R. Y. Choi et al., 2020). Las herramientas de ML más comunes son: Red Neuronal Artificial (ANN, Artificial Neural Network), regresión lineal y logística, Árbol de Decisión (DT, Decision Tree), Bosque Aleatorio (RF, Random Forest), Máquina de Vectores de Soporte (SVM, Support Vector Machine), Programación Genética (GP, Genetic Programming), entre otros.
Las herramientas de ML han proporcionado una solución conveniente y precisa para problemas de todos los campos, proyectándola como una buena alternativa para modelos de deterioro de pavimentos. El aumento en la investigación genera varias alternativas; por lo que no es simple decidir cuál será la herramienta de ML más conveniente para modelar el deterioro de pavimentos, cuál ayudará a tener un modelo más preciso y generalizable y cuál será la herramienta que más se ajusta a las condiciones y objetivos particulares.
El presente estudio hará una búsqueda detallada de las herramientas de ML aplicadas a modelos de deterioro estudiando las metodologías de modelamiento, las consideraciones tomadas, las ventajas y desventajas, los requisitos y los resultados de los modelos. Para luego, generar un software útil y versátil que permitirá definir la herramienta más adecuada para predecir teniendo en consideración el nivel de estudio, tipo de pavimento, proyección de predicción, variables de entrada y salida más usadas, datos disponibles, etc. Además, permitirá visualizar las oportunidades de investigación para aplicar herramientas de ML en modelos de condición de pavimentos
The Pavement Management Systems (PMS) are sets of analytical tools that help to make decisions about pavement networks; its main purpose is to monitor and continually evaluate the road assets available so that they provide safety and comfort to users (Kargah-Ostadi, 2014; Hernan Solminihac et al., 2018). An SGP must develop prioritization programs for maintenance and rehabilitation of pavements within the available road asset, which will depend on the level of decision required by the sector: network level for general condition thresholds and project level with more detailed and accurate data. In both cases, precise behavioral models are required to predict the deterioration of pavements as a function of time and define optimal maintenance and rehabilitation strategies. (Osorio-Lird et al., 2018). Generally, behavior models are developed using continuous functional formulas using classical techniques such as regressions and probabilistic approaches such as Markov chains; However, these models are subjective, with difficulty to calibrate to conditions other than the original ones, and, in addition, they require historical information on the pavement (Solminihac et al., 2018). Currently, there is greater interest in developing models by applying Machine Learning (ML) techniques since it is a programming tool, based on statistical methods, that uses data from examples or past experiences to optimize an automatic performance criterion and has the ability to adapt when conditions change over time (Alpaydin, 2010; RY Choi et al., 2020). The most common ML tools are Artificial Neural Network (ANN), linear and logistic regression, Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Genetic Programming (GP), among others. ML tools have provided a convenient and necessary solution to problems in all fields, projecting it as a good alternative for pavement deterioration models. The increase in research generates several alternatives, so it is not easy to decide which will be the most convenient ML tool to model the pavement deterioration, what will be more precise and generalizable model and which will be the tool that best suits the particular condition and objectives. This study wull make a detailed search to ML tools applied to impairment models, studying the modeling methodologies, the considerations taken, the advantages and disadvantages, the requirements and the results to the models. In order to later, generate a useful and versatile software that defines the most appropriate tool to predict taking into account the study level, type of pavement, prediction projection, most used input and output variables, available data, etc. visualize research opportunities to apply ML tools in pavement performance models
The Pavement Management Systems (PMS) are sets of analytical tools that help to make decisions about pavement networks; its main purpose is to monitor and continually evaluate the road assets available so that they provide safety and comfort to users (Kargah-Ostadi, 2014; Hernan Solminihac et al., 2018). An SGP must develop prioritization programs for maintenance and rehabilitation of pavements within the available road asset, which will depend on the level of decision required by the sector: network level for general condition thresholds and project level with more detailed and accurate data. In both cases, precise behavioral models are required to predict the deterioration of pavements as a function of time and define optimal maintenance and rehabilitation strategies. (Osorio-Lird et al., 2018). Generally, behavior models are developed using continuous functional formulas using classical techniques such as regressions and probabilistic approaches such as Markov chains; However, these models are subjective, with difficulty to calibrate to conditions other than the original ones, and, in addition, they require historical information on the pavement (Solminihac et al., 2018). Currently, there is greater interest in developing models by applying Machine Learning (ML) techniques since it is a programming tool, based on statistical methods, that uses data from examples or past experiences to optimize an automatic performance criterion and has the ability to adapt when conditions change over time (Alpaydin, 2010; RY Choi et al., 2020). The most common ML tools are Artificial Neural Network (ANN), linear and logistic regression, Decision Tree (DT), Random Forest (RF), Support Vector Machine (SVM), Genetic Programming (GP), among others. ML tools have provided a convenient and necessary solution to problems in all fields, projecting it as a good alternative for pavement deterioration models. The increase in research generates several alternatives, so it is not easy to decide which will be the most convenient ML tool to model the pavement deterioration, what will be more precise and generalizable model and which will be the tool that best suits the particular condition and objectives. This study wull make a detailed search to ML tools applied to impairment models, studying the modeling methodologies, the considerations taken, the advantages and disadvantages, the requirements and the results to the models. In order to later, generate a useful and versatile software that defines the most appropriate tool to predict taking into account the study level, type of pavement, prediction projection, most used input and output variables, available data, etc. visualize research opportunities to apply ML tools in pavement performance models
Description
Keywords
APRENDIZAJE AUTOMATICO (Inteligencia Artificial) , COMPUTADORES -- PROCESAMIENTO DE DATOS , PAVIMENTOS -- SOFTWARE PARA COMPUTADOR
