Model-free road friction estimation using machine learning

Four different machine learning methods (a convolutional neural network, a shallow neural network, a long short-term memory network and an ensemble of bagged decision trees) were trained on simulation data to provide model-free estimates of tyre-road friction properties using readily available sensor signals. The convolutional neural network and shallow neural network had the best performance on a previously unseen ensemble of test data. When typical noise was added to the predictors’ input values, the accuracy of the predictions decreased. To avoid this, the predictors were re-trained on noisy data, making them much more robust to noisy input data and showed marked improvement in root mean square error (RMSE) performance. Again, the convolutional neural network and shallow neural network had the best performance. This shows that building a model-free tyre-road friction predictor is possible and can yield promising results.