Keywords: neural network, piles, bearing capacity, Brazilian soils, load tests, geotechnical.

In order to guarantee quality and safety for civil engineering works, building foundations are increasingly tested and the corresponding experimental results carefully interpreted, especially with respect to the loading conditions that they will be submitted in the field. There are many technical procedures developed in geotechnical engineering to estimate the bearing capacity of foundations. In the case of deep foundations (piles) the most known methods are based on results from static and dynamic pile load tests.

Dynamic load tests have become widely used, largely because their execution is much more economic and relatively simple, when compared to the traditional static load tests. However, interpretation of a dynamic test is complex, since the signals collected during execution must be analysed using methods from the theory of wave propagation to indirectly estimate the bearing capacity of the tested pile.

A different approach is presented in this work, based on neural network techniques. It relies directly on the available experimental data to simulate the problem at hand and to make bearing capacity estimates even simultaneously with the data collection process in the field.

A set of 124 reports of dynamic pile load tests carried out in Brazil between 1996 and 1999 was used to estimate the bearing capacity of both concrete and steel piles. Several variables were selected as input parameters to the different neural networks, in an attempt to identify and to quantify their relative importance and influence on the total mobilised pile resistance.

A feed-forward network with one hidden layer was used in this work. It is composed of an input layer, with a variable number of source nodes, and an output layer with 1 neuron only, corresponding to the total pile bearing capacity. The hidden layer has 1 to 15 neurons, depending on the several neural network types tested in this investigation.

The algorithm employed to train the neural networks was the L.M. Levenberg- Marquardt method, that has shown to be faster than other training algorithms such as the error back-propagation algorithm. Each neural network was initialised at least 5 times, i.e. was trained and tested with different synaptic weights, randomly selected by the computational program at 5 different times. This procedure was taken in order to mitigate the effects of non optimum local minima in the network error surface. The stop criterion used in the network training was either 50 iterations in the LM algorithm or a minimum gradient error less than 0.0001. From analyses with several neural network types, it could be concluded that the maximum energy transferred to the pile, the maximum pile displacement and the plastic displacement component play the most significant roles in the overall behaviour of the pile. Other parameters, such as the geometric characteristics of the pile (length and cross section area), also tend to improve the bearing capacity estimates.

It was also observed that the number of STP blows, herein used as a measure of the soil resistance along the shaft and at the pile tip depth, did not have a very important effect on the final results, probably because the soil mechanical properties have already been implicitly considered through other input parameters related to the energy and pile displacements.

The present study suggests that prediction of the bearing capacity of piles through neural network techniques may bring important advantages to the geotechnical engineer. Estimates with errors within the range of 10 few parameters from the dynamic load tests, without the need to analyse the whole force x velocity curve, as in the traditional CAPWAP method, to get meaningful results. Moreover, it is possible to use neural networks as engineering tools to control the quality of the field tests, with a great economy in time and execution costs, since their results may be interpreted and analysed online, during the proper execution of the dynamic pile load test.

purchase the full-text of this paper (price £20)

go to the previous paper
go to the next paper
return to the table of contents
return to the book description
purchase this book (price £78 +P&P)