Aplicación de técnicas de Machine Learning basado en información sísmica para profundizar la probabilidad de terremotos mediante el uso de regresión logística y redes neuronales.
Fecha
2020-04
Título de la revista
ISSN de la revista
Título del volumen
Editor
Universidad de Guayaquil. Facultad de Ciencias Matemáticas y Físicas. Carrera de Ingeniería en Sistemas Computacionales.
Resumen
El objetivo principal del presente proyecto es predecir la posibilidad de que ocurra un
evento sísmico y las características de este mediante la aplicación de técnicas de Machine
Learning con la finalidad de minimizar daños dentro del territorio ecuatoriano, debido a
que, la aparición de un terremoto imprevisto es causante de daños estructurales y sociales
dependiendo de su magnitud, por lo que, se aplicaron las técnicas de Redes Neuronales y
Regresión Logística. Para aplicar aquellos algoritmos, se elaboró una base de datos que
contiene datos históricos provenientes de páginas web acerca de los sismos que sucedieron
en el Ecuador desde el año 1985 hasta el 2019 en las regiones Costa, Sierra y Oriente. En
total se recolectaron 658 sismos, los cuales fueron utilizados por las técnicas ya
mencionadas para la generación de nuevas predicciones y con ellas una probabilidad de
que estas en realidad puedan suceder. Los datos históricos, resultados y gráficos obtenidos
por ambas técnicas se puede visualizar por medio de una página web. Los dos algoritmos
demuestran que son viables al momento de predecir un evento sísmico, por ejemplo a través
de la aplicación de Redes Neuronales, la pérdida de valor de los datos de validación
(17.64%) es menor a los de la pérdida de valor de los datos de entrenamiento (23.8%) lo
que significa que no hay sobreajuste es decir que el proceso tendrá una menor probabilidad
en fallar al reconocer nuevos eventos sísmicos; mientras que el algoritmo de Regresión
Logística obtuvo una precisión del 72.25% siendo así un porcentaje mayor al 70% mínimo
requerido para poder aplicarse y el valor de la curva ROC o AUC es aproximadamente
70.4% dando a entender que los datos por encima de ese valor tienden a ser clasificadas
correctamente puesto que a medida que se acerquen más a 1, los datos tendrán una mejor
clasificación. Tomando en cuenta los resultados mencionados, se puede concluir que las
técnicas aplicadas en este proyecto son eficientes al momento de predecir eventos próximos
en base a datos históricos y que el uso de técnicas de aprendizaje automático resultan muy
útiles para resolver diversos problemas predictivos.
The main objective of this project is to predict the possibility of a seismic event and its characteristics through the application of Machine Learning techniques in order to minimize damages within the Ecuadorian territory, due to the fact that the occurrence of an unforeseen earthquake may cause structural and social damage depending on its magnitude, so the techniques of Neural Networks and Logistic Regression were applied. To apply those algorithms, a database was created, it contains historical data from web pages about the earthquakes that happened in Ecuador from 1985 to 2019 in the Costa, Sierra and Oriente regions. In total 658 earthquakes were collected, which were used by the aforementioned techniques for the generation of new predictions and with them a probability that these may actually happen. Historical data, results and graphs obtained from both techniques can be visualized through a web page. The two algorithms demonstrate that they are viable to predict a seismic event, for example through the application of Neural Networks, the loss of value of the validation data (17.64%) is less than those of the loss of value of the training data (23.8%) which means that there is no overfitting, meaning, the process will have a lower probability of failing to recognize new seismic events; while the Logistic Regression algorithm obtained an accuracy of 72.25%, thus being a percentage greater than the 70% minimum required to apply and the value of the ROC or AUC curve is approximately 70.4%, implying that the data above that value tend to be classified correctly since they get closer to 1, the data will have a better classification. Taking into account the mentioned results, it can be concluded that the techniques applied in this project are efficient when predicting upcoming events based on historical data and that the use of machine learning techniques are very useful to solve various predictive problems.
The main objective of this project is to predict the possibility of a seismic event and its characteristics through the application of Machine Learning techniques in order to minimize damages within the Ecuadorian territory, due to the fact that the occurrence of an unforeseen earthquake may cause structural and social damage depending on its magnitude, so the techniques of Neural Networks and Logistic Regression were applied. To apply those algorithms, a database was created, it contains historical data from web pages about the earthquakes that happened in Ecuador from 1985 to 2019 in the Costa, Sierra and Oriente regions. In total 658 earthquakes were collected, which were used by the aforementioned techniques for the generation of new predictions and with them a probability that these may actually happen. Historical data, results and graphs obtained from both techniques can be visualized through a web page. The two algorithms demonstrate that they are viable to predict a seismic event, for example through the application of Neural Networks, the loss of value of the validation data (17.64%) is less than those of the loss of value of the training data (23.8%) which means that there is no overfitting, meaning, the process will have a lower probability of failing to recognize new seismic events; while the Logistic Regression algorithm obtained an accuracy of 72.25%, thus being a percentage greater than the 70% minimum required to apply and the value of the ROC or AUC curve is approximately 70.4%, implying that the data above that value tend to be classified correctly since they get closer to 1, the data will have a better classification. Taking into account the mentioned results, it can be concluded that the techniques applied in this project are efficient when predicting upcoming events based on historical data and that the use of machine learning techniques are very useful to solve various predictive problems.
Descripción
PDF
Palabras clave
Sismos, Predicción,, Machine Learning, Redes Neuronales, Regresión Logística, Python