Comparison of Machine Learning Methods for Evaluating Pavement Roughness Based on Vehicle Response

The roughness of a road pavement a ects safety, ride comfort and road durability. A useful 6 indicator for evaluating roughness is the weighted longitudinal pro le (wLP). In this paper, we 7 compare three machine learning models for estimating the wLP when vehicle response information, 8 i.e. accelerometer and wheel speed data is collected from common in-vehicle sensors. We applied a 9 multilayer perceptron, support vector machine (SVM) and random forest for testing their e ective1 0 ness in estimating the key indices of wLP, namely range and standard deviation. We trained these 11 models from a set of features extracted from vehicle response simulations on accurate replications 12 of roads with various roughness problems. In contrast to other research, we validated the models 13 with measurements collected with a probe vehicle. Our results show that we can accurately detect 14 roughness phenomena. The SVM produced the best results, although the models achieved rather 15 similar performance. However, we found di erences regarding the model robustness when reducing 16 the size of the training feature set. The proposed method enables road network monitoring to 17 be achieved by conventional passenger cars, which can be seen as a practical supplement to the 18 prevalent road measurements with cost-intensive mobile devices. 19 Keywords: Roughness, road pavement, support vector machine, random forest, neural network, 20 weighted longitudinal pro le, machine learning, in-vehicle sensors, vehicle dynamics, simulation