Abstract:

Even though the detection of road surface roughness has reached standardization, the rapid, high-frequency and low-cost intelligent detection method of IRI has also been widely studied for constructing the smart cities and the intelligent transportation facilities. However, the detection accuracy and detection effectiveness of IRI based on different intelligent devices have not been deeply studied. Firstly, this paper conducted road driving experiments using two intelligent IRI detection devices that was developed by the authors’ research group. One device is called driving data collection Smartphone APP, and another one is called road driving data collection Intelligent Terminal Device. The data of driving test vehicle, such as vibration acceleration, speed, GPS location and so on, was collected during the driving experiment, and four vibration acceleration indicators that can best reflect the impact of IRI were determined by using a random forest model. Next, three prediction models of IRI were established using three neural networks: recurrent neural network RNN, gated recurrent unit GRU, and long short-term memory network LSTM. The detection accuracy of IRI using different devices and different prediction models was compared. The results showed that, LSTM model achieved the best robustness and highest prediction accuracy among three neural network models. The R² values of predicted IRI is 0.864 and 0.789 for the Intelligent Terminal Device and Smartphone APP respectively, which means that the detection accuracy of Intelligent Terminal Device was higher than that of Smartphone APP. The results of this paper have theoretical significance and application value for improving the informatization level of IRI detection and monitoring, as well as enhancing the scientific level of road maintenance decision-making.

Key words: road pavement, intelligent terminal device, smartphone APP, IRI, vehicle vibration, characterization index, predictive accuracy