Vibration-based classification of road damages: Gyroscope data and a simple neural network model

Abstract

A system for automatic detection of road damages is essential for logistics management. At the core of the system lies an algorithm for classification of damages. This work intends to establish such an algorithm. In the current proposal, the algorithm receives data of the probe vehicle rate of rotations, extracts statistics descriptive from the data to produce a feature vector, and finally, feeds the vector to a simple-three-layer neural network to classify the damages. The types of damages are limited to four cases: normal, pothole, speed bump, and expansion joint. For the development, a probe vehicle is used to produce 400 empirical data involving those four damage cases. By using Monte Carlo approach, the established neural network model is evaluated. The results show that the approach is able to classify the cases with 85% accuracy. The study also finds that the rates of pitch and roll to be the determining factors.