A machine learning framework for guided wave-based damage detection of rail head using surface-bonded piezo-electric wafer transducers

Due to repeated heavy loads, environmental conditions and non-frequent monitoring, the rail is subjected to heavy damage resulting in sudden failure. Hence, a frequent, faster, and efficient monitoring strategy is required. This paper attempts to investigate the application of guided wave (GW) generated through surface-bonded piezo-electric wafer transducer (PWT) to detect damages in rail at high frequencies. Firstly, a combined experimental and simulation study is presented in an effort to understand the dispersion characteristics of guided wave and its interaction with head damages in a relatively small rail specimen. The numerical simulation results are validated with those obtained from the experiments showing a good agreement between them. Secondly, a framework based on the machine learning algorithm is proposed to efficiently detect damage in rail head. Numerous inseparable guided wave modes are observed at higher frequencies implying the inability to detect damage through specific mode. Therefore, a machine learning framework is trained using time, frequency, and time–frequency domain features of the signal. Total 672 numerical simulations of different types of damage with different severity and location in the rail head are carried out to train and validate the model. It is found that GW generated through surface bonded PWTs is able to detect minimum defect size of 5% of head area with 1 mm thickness. Finally, the proposed framework is tested using simulation and experiment results of arbitrary damage in the rail head. The error in estimating severity was found to be in the range from 2.00% to 16.67%.