SILVA, B. S.; SILVA, Bruno Sousa da.
Résumé:
The objective of this work is to develop, apply and analyze the results of an artificial
neural network (ANN) developed using a matrix simulation software with a package
for artificial neural networks, which is capable of performing a monthly cost estimate
in an earthworks work. in cut and fill services. The principle is to understand the
importance of each input and output parameter; organize all collected data and
normalize these data so that they vary within the range of 0 to 1. Several
architectures were studied, varying the number of layers, neurons and activation
function, with a lower percentage of error, to verify the volume prediction capacity
of cut and fill one to be used in the following month and, with this, reach the monthly
disbursement value from the amount paid per cubic meter excavated or compacted.
Thus, a medium to high standard horizontal condominium project was chosen,
consisting of projects and budgets, in addition to other data provided for the input
parameters. The chosen input parameters were stake location, distance from point to
stake zero location, design cut volume, design backfill volume, month of execution,
and large. Meanwhile, the output parameters were cut volume and landfill conducted.
The results of the ANNs were compared to the actual values in two ways: in the first,
the cut volumes and predicted landfill were compared from pile to pile, in the
second, the compared volume was separated into months and by street, no longer by
pile. The best performing architecture for RNA1, discovery of 02 (two) layers, 10
(ten) neurons and Tansig activation function. Another neural network showed results
similar to those of RNA1, however, with 03 (three) layers and 15 (fifteen) neurons
and, therefore, requiring a greater effort from the machine to reach the same average
result. The average error values greater than the admitted range found in the landfill
volume can be explained by the small amount of works to analyze and compare, as
well as the small amount of input parameters available.
