Keywords: neural network model, connection weight approach, sensitivity analysis, feedforward backpropagation, light structures, expansive soils, potential damage influences.

In this paper a Neural Network Model (NNM) was used to develop a ranking of the potential damage influences (PDI) for light structures on expansive soils in Victoria. These influences include geology (G), the Thornthwaite moisture index (TMI), vegetation cover (VC), construction wall type (CW), construction foundation type (CF), geographical region (GR) and the age of the building when first inspected. A limited number of studies have been performed on damage prediction to light structures, but no predictive damage model has been developed. This study focused on Victoria, Australia. The Geological map of Victoria [1] shows that three quarters of the soils in Victoria are highly expansive.

The NNM was chosen to predict the rank of importance of the PDI because it is robust and fault tolerant and can deal with noisy data found in the database. In this paper, feedforward backpropagation (FB) was adopted since it minimizes the amount of work and time devoted to model development by letting the computer do more of the work [2]. The connection weight approach (CWA) is used to calculate the important weights and sensitivity analysis (SA) is used to prove that the NNM is viable.

The CWA approach was adopted in this paper as it is shown in [3] that it is the most accurate method for quantifying ranking of importance in the NNM. The highest-ranking PDI is age followed by G, CF, VC, CW, TMIO, GR and TMIN. From the analysis done by using CWA and SA, it was found that both approaches give similar results. Input variables with larger amount of weights represent greater intensities of signal transfer which is more important in the prediction process compared to variables with smaller weights [4]. Negative connection weights represent inhibitory effects on neurons and decrease the value of the predicted response as in the case of TMIO, GR and TMIN above. Positive connection weights represent excitatory effects on neurons and increase the value of the predicted response as in the case of Age, G, CF, VC and CW.

From the results, it can be seen that age gives the highest positive values and is rank first in the rank of importance. It is a possibility that the connection weight for age is higher than the connection weights for the rest of the PDI because nearly all the PDI relies on age such as TMIO, TMIN, GR and G. It is possible that age is calculated more than once due to that. G is an important factor in the damage to structure. Since three quarters of the soils in Victoria are highly expansive, it is not a surprise that G ranks second after age.

TMIN and TMIO rank sixth and eighth respectively. Although they give the same function where both the TMIs measure the climate classification, their connection weight give totally different results where the TMIN is much less than the TMIO. From the results obtained it is proven that the TMI has changed over the years [5,6]. It is not surprising to see the construction types to be the third and fifth ranking of the PDI. This is because the designs for walls and footings were designed with regard to the TMIO. Codes such as AS2870 [7] should incorporate allowances in footing and wall design to accommodate change in the TMI over the life of the structure.

The NNM proved to be useful in many other domains. From the analysis, the rank of importance of the PDI was obtained using FB. Thus a model of the ranking of the importance for damage to light structures founded on expansive soil can be constructed.

1

Geology of Australia ore deposits, "The Geological map of Victoria", Australia, 1965.

2

Smith, M., "Neural Networks for statistical modelling", International Thompson Computer Press, USA. 1996.

3

Olden, J.D., Joy, M.K., Death, R.G., "An accurate comparison of methods for quantifying variable importance in artificial neural networks using simulated data", Ecological Model, 178, 389-397, 2004. doi:10.1016/j.ecolmodel.2004.03.013

4

Olden, J.D., Jackson, D.A., "Illuminating the 'black box': a randomization approach for understanidng variable contributions in artificial neural networks", Ecological Model, 154,135-150, 2002. doi:10.1016/S0304-3800(02)00064-9

5

McManus, K.J., Lopes, D., Osman, N.Y., "The influence of drought cycles on the thornthwaite moisture index contours in Victoria Australia", An International Conference on Problematic Soils, 2, 357, 2003.

6

McManus, K.J., Lopes, D., Osman, N.Y., "The effect of Thornthwaite Moisture Index Changes in Ground movement predictions in Australian Soils", Proceedings 9th Australia New Zealand Conference on Geomechanics, 2, 555-561, 2004.

7

Standards Australia, "AS2870-1996 - Residential slabs and footings-construction", Standards Australia International, Australia. 1996.

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 £80 +P&P)