Characterisation of Clustered Cracks using an ACFM Sensor and Application of an Artificial Neural Network

The alternating current field measurement (ACFM) technique can be applied for surface-breaking fatigue crack detection and sizing; the link between the ACFM signal and crack size is well understood for individual cracks. However, the ACFM response to multiple clustered cracks is significantly different to that of isolated cracks. In railway rails the high wheel-rail forces can lead to rolling contact fatigue (RCF) cracks. Often cracks appear together in small clusters or in long stretches. The accurate characterisation of such fatigue cracks is essential for carrying out efficient and safe repair and maintenance. This paper presents a method for sizing the important sub-surface section of multiple cracks using ACFM via the application of an artificial neural network (ANN). The approach is demonstrated using a railway case study: a simulation-based dataset of signal response covering the range of RCF cracks typically seen in in-service railway tracks has been generated to give a thorough representation of the effect of clustered crack parameters on the ACFM response. A 5×5×2×1 multi-layer ANN has been optimised and trained using the validated simulation database to learn the inverse relationship between the crack pocket length (desired output) and the ACFM signal for a given cluster of RCF cracks. The network has been evaluated on a set of experimental data to size cracks of known dimensions from ACFM measurements and also on unseen simulation data. Results from both simulation and experiment show that the approach presented can be used to size clustered cracks to approximately the same degree of accuracy as is possible for isolated cracks.