Aplicação de técnicas de agrupamento e rede neural artificial em acidentes com máquinas agrícolas

Abstract

The use of machines has become increasingly common; however, this has increased the number of accidents, which occur not only on the properties, but also on roads, highways and cities. Accidents involving tractors generate alarming data because, in some situations, they result in serious injuries or death, justifying the need to study them, identifying patterns, understanding the causes and minimizing the risks. Therefore, the objective of this study was to evaluate a dataset of accidents with agricultural machines on public roads and rural properties to form similar groups, characterization and identification of patterns using artificiais neurais networks, self-organizing maps, combined with techniques hierarchical and non-hierarchical data clustering, dendrogram, db index and k-means coefficient. Data were obtained from a LIMA survey database, with information from news of accidents involving tractors throughout the Brazilian territory in the period from 2013 to 2021. The news of accidents were classified according to their region: South region, Southeast region, Midwest Region, Northeast Region and North Region. Period of occurrences: dawn, morning; afternoon and night. Number of victims: no victims; injured victims, fatal victims and not informed. Location of accident: Via published, rural property and not informed. Causes of the accident: Irregular terrain, power take-off, improper use of the implement, speed incompatible with the road, unsafe overtaking, lack of attention, sleeping at the wheel, drunkenness, disobedience to signaling and other causes of accidents. Types of accidents: Imprisonment of limbs, loss of limbs, collision, being run over, overturning, falling from the tractor, other and not informed. Types of implements: Plow, harrow, rotary hoe, sprayer, seeder, harvester, cart, other implements and tractor. Operator's Age Range: Less than 10 years old, 11 to 20, 21 to 30, 31 to 40, 41 to 50, 51 to 60, 61 to 70 and over 71 years old. The class of each indicator was coded forming a matrix of quantitative data for analysis. For analysis, clustering techniques and SOM neural network were used. After determining the most suitable network configuration, training was automated, as per programming of the software input functions used. For training and simulation of the artificial neural network (SOM network), after coded indicators, the MATLAB 2016 software was used, where the somtoolbox tools were used, responsible for training the neural network and simulation. The clustering technique combined with the self-organizing neural network of maps proved to be a powerful tool in the analysis of accident fingers, allowing a safer and more complete exploratory analysis.