Redes neurais com topologias otimizadas aplicadas na modelagem de dados geotécnicos e pluviométricos para predição de deslizamentos de solo






The problems related to landslides caused by rain on slopes have been a reality in several cities of Brazil and parts of the world. The government and research institutions have sought, in partnership, solutions to minimize the damage caused by such accidents, from studies of correlation between rainfall and landslide occurrences enabling the development of a warning system that can help save lives and avoid property damage, by providing information to the responsible agencies and susceptible populations. Currently, there are locations in the country where the rainfall-landslide correlation was established for a warning system, based on precipitation envelope curve tting from a database of rainfall and landslide occurrences. The Municipality of Vitória (ES), where the same methodology was applied, has a signicant number of geotechnical problems. Neural networks are an interesting tool for geo-engineering due to its success in modeling various nonlinear problems to generate maps and prediction in the eld of geotechnical engineering. The use of metaheuristics for optimization of neural network topologies is well documented, but is often not applied. The present work applies neural network Multi-Layer Perceptron (MLP) topology, which is optimized by two metaheuristics called Particle Swarm Optimization (PSO) and Genetic Algorithms (GA) to predict landslides in Vitória-ES, using a database of precipitation series, geotechnical parameters and landslide occurrences. The algorithms also perform feature selection and they are compared with data mining methods. The results of both techniques and performances are compared with a reference neural network, that uses all available attributes. We compared the results obtained by neural networks with the method of power function envelope curve tting, performing hypothesis test on event records set aside for testing


ciencia da computacao redes neurais otimização por enxame de partículas algoritmos genéticos chuva deslizamentos de solo geotecnia vitória-es neural networks particle swarm optimization genetic algorithms rainfall landslide geotechnics vitoria-es

Documentos Relacionados