Using GA-BP neural networks to analyze vertical bearing capacity of single rock-socketed pile

Abstract

With regard to the design of ultimate vertical bearing capacity of single rock-socketed pile, theoretical formula recommended by “Technical code for building foundations” (JGJ 94-2008)and static loading test are the two most popular methods. But results obtained from this two approaches have great differences. The average of relative difference (note: taking absolute value) is even up to 45.7%, and in most cases the values from the former are greater than that from the latter, as found from analyzing the static loading test data of 101 rock-socketed piles. By using founded GA-BP neural networks which taking 101 rock-socketed piles' static loading test data as a training sample, 10 predicted results have a good agreement with expected values from static loading tests. The average of relative difference is about 5%. GA-BP neural networks can also be used in analyzing the vertical bearing behaviors of rock-socketed pile. Using the GA-BP neural networks as mentioned above, varying the magnitude of ratio of rock-socketed depth to pile diameter, and keeping the rest of other parameters as unchanged constants, a curve of relationship between the variable ratio and output predicted vertical bearing capacity could be achieved, which revealed that for a given construction field and workmanship, a most reasonable value of the ratio could be found which corresponding to the maximum value of ultimate vertical bearing capacity, but no constant value has been found while changing the site condition.