Banca de DEFESA: MARIA JULIA DE OLIVEIRA HOLANDA

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
STUDENT : MARIA JULIA DE OLIVEIRA HOLANDA
DATE: 18/11/2022
TIME: 09:00
LOCAL: Pos Graduação Engenharia Civil
TITLE:

Collapsible and Expansive Soils in Brazil: Classification of Susceptibility to Occurrence Using Artificial Neural Networks

KEY WORDS:

Artificial Neural Networks. Expansion. Collapse

PAGES: 117
BIG AREA: Engenharias
AREA: Engenharia Civil
SUMMARY:

Collapsible and expansive soils are problematic soils in Civil Engineering. Identifying, classifying and understanding the hydro-geomechanical behavior requires laboratory and field testing procedures and the use of computer models that take into account the tensional state in which the soil is and to which it will be subjected, relating the volume variation due to the change in moisture. Artificial Neural Networks (ANN) constitute an important tool for this purpose, through the correlation between the predictors and the properties to be estimated. The research aims to identify the degree of probability and classify the susceptibility of occourence of collapsible and expansive soils in Brazil based on geotechnical variables and pedological, geological and climatological variables. Based on information from the Special Soils Database (BANDASE) of the research group on unsaturated soils (GSNsat) at UFPE, using ANN through Neural Design, the networks are developed and their performance is analyzed according to binary classification tests. Three networks are elaborated. The first network (PE04) created from 87 PE samples (57 training, 17 selection and 17 testing), has 4 input variables (% sand, % clay, plasticity and activity indices) and has classification accuracy of 76.5%. The second network (PE07) considers the same 87 samples with 7 input variables (the 4 variables of the PE04 network, in addition to climate, pedology and geology) and obtains an accuracy of 88.2%. The third network (BR03) developed with 393 samples (237 training, 78 selection and 78 testing), uses 3 input variables (climate, pedology and geology) and has an accuracy of 89.7%. The generalization of the prediction patterns of the networks presents accuracy rates of 91.11% and 81.95% for the PE04 and BR03 networks and are applied by interpolating samples within the same domain that they were developed; while the PE07 network, was validated with samples from the Northeast and later extrapolating to samples from all over Brazil, had a decrease in the accuracy rate: 65.5% and 56.7% respectively. With the best network (BR03) a soil classification program is developed and a probabilistic map of the occurrence of collapsible and expansive soils, “available for free” in a web environment. The ANN PE04 and PE07 allowed the production of estimates for the identification and classification of collapsible and expansive soils in the State of PE with satisfactory accuracy, establishing a good correlation between the variables and the BR03 network and ratified the importance of the variables of origin and soil formation for classify places prone to these phenomena achieving better classification results.

COMMITTEE MEMBERS:
Externa à Instituição - ANA PATRICIA NUNES BANDEIRA - UFCA
Externo à Instituição - FERNANDO ARTUR NOGUEIRA SILVA - UNICAP
Externo à Instituição - JOAQUIM TEODORO ROMAO DE OLIVEIRA - UNICAP
Externa à Instituição - KATIA VANESSA BICALHO - UFES
Presidente - 1654153 - MARIA ODETE HOLANDA MARIANO