Backpropagation Neural Network with Feature Sensitivity Analysis: Pothole Prediction Model for Flexible Pavements using Traffic and Climate Associated Factors
Dante Laroza Silva, Kevin Lawrence M. de Jesus
Abstract
Different industries were transitioning to the utilization of Industry 4.0 tools such as Artificial Intelligence (AI) techniques particularly Neural Network Modelling. The neural network application in pavement management is a serviceable tool to forecast the occurrence of pavement distress which leads to efficient allocation of budget for pavement maintenance. The aim of this research is to create a model for predicting the percentage occurrence of pothole in a flexible pavement and investigating the relative importance of the traffic and climate – associated factors. A database of 55 datasets was obtained including input parameters such as temperature, precipitation, pavement age, total Average Annual Daily Traffic (AADT) and AADT of heavy vehicles and target output which is the pothole % occurrence. Using the database, a backpropagation – neural network model (BP–NN) was established to forecast the percentage occurrence of potholes in a flexible pavement. The weights and biases generated from the simulation of the best model was employed to achieve a comprehensive sensitivity analysis to explore the effect of the input parameters to the pothole % occurrence. Results show that the BP-NN can effectively predict the pothole % occurrence in a flexible pavement and the best model generated has a topology of 5-12-1 (input-hidden-output). Additionally, the AADT of heavy vehicles is the most influential parameter affecting pothole % occurrence in flexible pavements based on the sensitivity analysis performed using Garson's algorithm.