Uso de técnicas de agrupamento e rede neural em sinistros com máquinas agrícolas nas rodovias federais

• Ano de defesa: 2018
• Autor(a) principal: Macedo, Deivielison Ximenes Siqueira
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Palavras-chave em Português: Redes SOM; Segurança; Colisões
• Link de acesso: http://www.repositorio.ufc.br/handle/riufc/32006

With the agriculture growth and the need to improve production increasingly tractor use is necessary, along with the growth of the quantity of machines, comes the increase in the number of cases that resulted in accidents, but there is a lack of information on the current situation in the country and how to prevent these occurrences. The neural network is an important tool to tackle adversity, because they take into account all the variables inherent in the different problem of some conventional methods, which only explore the situation according to their limitations, and propose solutions to resolve the problem. The objective was apply the neural networks through the SOM (Self Organized Map) networks, on the indicators of the accidents occurred in Brazilian federal highways involving tractors, identifying the presented patterns and using techniques of hierarchical and non hierarchical groupings to determine the groups of accidents more similar to each other for each region and in the country. The work consisted of a partnership between the Laboratory of Agricultural Accident Investigations-LIMA and the 16th Superintendence of the Federal Highway Police in Fortaleza, which provided the TAB Traffic Accident Bulletin for the period from 2007 to 2016. In TAB the indicators evaluated were: type of accident, cause of the accident, time of day in which the accident occurred, the clinical condition of the victims, track layout and weather conditions at the time of the accident, to the networks for regions was also included the Federative unit. The use of neural networks has been performed by the self organized maps, hierarchical clustering method used was the dendrogram, the non-hierarchical was the K-means coefficient and was used DB index to validate and assist in determining the number of groups, and all methods were obtained through Matlab software. The classes of indicators that more neurons were activated rear-end collision, inattention, clear sky/sun without victims, straight track and in the morning. The most active neurons to the indicator status were Pará (North), Bahia (Northeast), Goiás and Mato Grosso do Sul (Midwest), Minas Gerais (Southeast) and Rio Grande do Sul (South). Situations that provided greater quantity of injured or fatalities were trampling and the rear-end collision in the evening. The combined use of cluster analysis with SOM networks proved to be an excellent combination for representation and analysis of accidents by working with multiple variables, allowing a more complete exploratory analysis of accidents.