Machine learning-based rail corrugation recognition: a metro vehicle response and noise perspective

Rail corrugation is a common problem in metro lines, and its efficient recognition is always an issue worth studying. To recognize the wavelength and amplitude of rail corrugation, a particle probabilistic neural network (PPNN) algorithm is developed. The PPNN is incorporated with the particle swarm optimization algorithm and the probabilistic neural network. On the basis of the above, the in-vehicle noise characteristics measured in the field are used to recognize normal rail wavelengths of 30 and 50 mm. A stepwise moving window search algorithm suitable for selecting features with a fixed order was developed to select in-vehicle noise features. Sound pressure levels at 400, 500, 630 and 800 Hz of in-vehicle noise are fed into the PPNN, and the average accuracy can reach 96.43%. The bogie acceleration characteristics calculated by the multi-body dynamics simulation model are used to recognize normal rail amplitudes of 0.1 and 0.2 mm. The bogie acceleration is decomposed by the complete ensemble empirical mode decomposition with adaptive noise, and a reconstructional signal is obtained. The energy entropy of the reconstructional signal is fed into the PPNN, and the average accuracy can reach 95.40%.This article is part of the theme issue 'Artificial intelligence in failure analysis of transportation infrastructure and materials'.

Automated summary

This study develops a rail corrugation recognition method based on a particle probabilistic neural network algorithm to identify the wavelength and amplitude of rail corrugation in metro lines. The method combines the particle swarm optimization algorithm with the probabilistic neural network, using in-vehicle noise characteristics and bogie acceleration characteristics to achieve high accuracy recognition. The average accuracies can reach 96.43% and 95.40% for rail wavelengths and amplitudes, respectively.