Prediction of Subgrade California Bearing Ratio using Falling Weight Deflectometer for Pavement Evaluation in Kenya

Abstract

Subgrade CBR is a very important parameter which gives an indication of the strength of the pavement’s foundation. Kenya’s Road Design Manual (Part 3) classifies the subgrade as S1 to S6 depending on their CBR range. The objective of this study was to predict subgrade CBR using FWD for pavement evaluation. A sample of 18 road sections were selected purposively for the study due to availability of data. The data required was FWD data, trench logs, laboratory results for the samples collected from the trenches and construction data. The FWD data was collected at intervals of about 100m using Primax FWD equipment. Trial pits were dug at intervals of about 5 km and the samples collected were tested in the laboratory to establish the soil properties. An average of three FWD readings were obtained at the exact chainage where the trial pits were dug. The materials were first characterized and classified using Unified Classification System (UCS). Lean clay (CL) and Elastic Silt (MH) were found to be the predominant soil classes. The average values of PI, OMC and MDD per road section were found to range from 7 to 23%, 12 to 27% and 1250 to 1821 kg/m3 respectively. The 4th geophone (D4), which was positioned at 600mm from the centre of the loading plate, was selected for developing the model based on the deflection bowls and surface modulus graphs. 11 models were developed by considering varying independent variables like deflections, contact pressure, air temperature, MDD, OMC, PI and Soil Class the dependent variables were the predicted CBR values. The data for all the road sections was split randomly into two; 80% was used to train the models and 20% to test the models. The models developed were Linear Regression models (LR) and machine learning models which were Random Forest (RF) and Artificial Neural Networks (ANN). Three models had the highest R2 score of 55 to 61% which means they explained the variability of lab CBR and were the best models. The same models had the lowest RMSE therefore making them the best models for predicting CBR. The CBR values predicted were correlated with the CBR values from the lab and found to have Pearson’s correlation coefficient of 0.81 to 0.84. The two machine learning models require specialized software like python to make predictions. The LR model developed was CBR = 5.2215 + (-0.0036 x D4) + (-0.0144 x Pc) + (-0.2522 x Tair) + (0.0233 x MDD) + (-0.3732 x OMC). The deflections, contact pressure and air temperature are values obtained from the FWD data. The study recommended that the developed predictive models can be subjected to additional validation by using them to predict CBR from other road sections in future pavement evaluations.