A crack depth prediction method using transient geometric features of three-dimensional circular array MFL signals aided LSTM

It is of great significance to improve the detection accuracy of initial fatigue cracks on the rail top surface for early warning of rail damage. The crack depth is especially difficult to measure directly in common methods. This study proposes an approach using transient characteristics of three-dimensional array MFL signals aided by Long Short-Term Memory (LSTM) algorithm to predict the crack depth. First, the characteristics of MFL signals are analyzed in the finite element simulation model when crack depth changes. Then, transient geometric features of the circular array MFL signals are extracted as a set of input feature vectors. Finally, the LSTM neural network algorithm is used to predict the crack depth. The experiment results show that the average prediction accuracy of crack depth is about 93.6%. It indicates that the method can explore more signal features and improve the crack depth prediction accuracy.