Keywords: artificial neural network (ANN), ravelling, porous asphalt, material, prediction, model.

The most unacceptable structural damage of porous asphalt top layers is ravelling. Therefore it is important to predict when porous asphalt top layer will achieve a critical level of ravelling so as to allocate funds for necessary maintenance. Because porous asphalt as current asphalt for motorways has been applied in many places and since a good maintenance strategy is not simple to obtain, it is also economically beneficial. That is why Artificial Neural Network (ANN) is the solution when it can uncover the most complex systems behaviour and let road authorities to make a better decision. ANNs are typically used for modelling complex relations in situations where insufficient knowledge of the system under investigation is available for the use of conventional models.

Porous asphalt is the topmost of a number of layers within road structure which differs from conventional materials in that it contains no fine material, only large single size stones, allowing a high percentage of air voids. The presence of air voids in the asphalt allows surface water to quickly drain below the road surface, offering markedly reduced spray and improved visibility. The major drawback of porous asphalt layers is that they are very sensitive to ravelling. The contact between tires and asphalt layers after a course of time make the top stones to come loose, which is called ravelling. Since this detriment can develop itself explosively fast, it happens often that the surface layers should be replaced relatively quickly after the first observation of that. Which factors are exactly responsible for this is still unknown.

The database provided by SHRP-NL research program was used in this study. The Strategic Highway Research Program Netherlands (SHRP-NL) has been performed between 1990 and 2000 and had initiatives from SHRP program. The database covers a period of ten years on 250 test sections, located on in-service roads. A total of 150 sections contained relevant information of porous asphalt. The severity of ravelling has been categorized in ravelling low, moderate high which are explained as follows: Low: 6-10% stone losing per m, Moderate: 10-20% stone losing per m, High: stone losing per m

For this paper, the author used a multi-layered neural network called QNET which uses a back-propagation algorithm. Previously described SHRP-NL data was initially analyzed using QNET with the objective of finding an ANN model as function of the available input parameters.

Model I: Input layer consist of 19 data neurons as follows:

• Input 1,4,7,10,13 (RL): Ravelling Low from 1993 to 1997 (%)
• Input 2,5,8,11,14 (RM): Ravelling Moderate from 1993 to 1997 (%)
• Input 3,6,9,12,15 (RH): Ravelling High from 1993 to 1997 (%)
• Input 16 (WD): Average of Warm Days between the years 1993 and 1997
• Input 17 (CD): Average of Cold Days between the years 1993 and 1997
• Input 18 (RP): Average Rainfall Precipitation between years 1993 and 1997 (mm)
• Input 19 (TP): average Traffic Percentage of heavy vehicles daily on test sections
• Input 20 (TH): maximum Thickness of asphalt
• Input 21 (R): maximum Roughness of asphalt
• Input 22 (Age): Age of the pavement at 1998

And output layer consists of 3 data neurons as follows: Output (1, 2, 3): Low, Moderate and High ravelling at year 1998 (%)

Results of Model I: Cold days are days with minimum temperature below zero and Warm days are days with a maximum temperature of more than 25^oCA training data set compromising 115 random sections out of 150 available (77%) was initially chosen for the learning stage. After training stage, the remaining 35 sections were used to validate the model. Model I is able to predict ravelling low, moderate and high with correlation factors 0.986, 0.926, and 0.976. ANN model quantifying the relative contribution of each input neuron to the computed output value. Hence it is possible to investigate the most relevant factors affecting ravelling in porous asphalt top layers.

Model II: Input layer consist of WD, CD, RP, TP, TH, R, Age from Model I with the same output of Model I

Results of Model II: It provided sensitivity analysis indicating the relative contribution of factors related to climate (58%), traffic factor (14%), thickness (6%), roughness (12%) and age (10%) for ravelling low and high but 46%, 15%, 15% and 11% for ravelling moderate. it allowed also analyzing the parameters interaction by colour contours which illustrated that heavy traffic, low thickness and high roughness cause ravelling on old asphalt especially in cold rainy days.

Model III: it was developed to analyze the relationship between material composition properties and ravelling. Properties are included asphalt density (DS), bitumen (BIT) percentage and void space (VS) percentage in asphalt mixture, stone type (ST) and cumulative percentage(C#) of stones which can pass the sieve with 16 mm to 63 m size (in 10 different size).

Results of Model III: C500m and smaller belongs to stones which are almost dust and according to model for avoiding ravelling this kind of stones should use less.

Optimized version of Model III analysed even more material relations. Since the ANN models explain complex relationships between material properties and ravelling which is with conventional model likely impossible. ANN proved to be a powerful technique opening great opportunities for development of ANN models for other asphalt detriments which influence technical maintenance strategies.

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 £95 +P&P)