Answers are grounded in this paper and cite their sources.
This paper proposes a feedforward neural network (FFNN) based machine learning (ML) technique for simultaneous sizing of cuboidal flaws from three-dimensional (3D) magnetic flux leakage (MFL) images. The ML algorithm was implemented in Python using Keras and scikit libraries and used rectified linear unit (ReLU) activation function in the hidden layer. Dipole modeling was used for predicting the 3D components (tangential, Hx; transverse, Hy and normal, Hz) of MFL images of cuboidal slots in 12.0 mm thick carbon steel plates. A total of 384 datasets and 7 significant features were considered for slots of different size combinations (10.0–17.0 mm length, 0.5–5.0 mm width, and 1.0–8.0 mm depth). The number of nodes in the hidden layer and the epochs used in the ML technique were optimized. The optimized technique was found to be efficient for simultaneously predicting length, width, and depth of surface breaking cuboidal slots with maximum root mean square errors (RMSEs) of ± 0.025 mm, ±0.053 mm, and ± 0.102 mm for the lengths, widths, and depths, respectively. Effect of Gaussian white noise on the performance of the ML technique was evaluated and the predictions of slot dimensions were found to be within the acceptable range. The technique was further evaluated on experimental MFL images and could simultaneously predict a slot of 15.0 mm length, 2.1 mm width, and 6.1 mm depth in a 12.0 mm thick carbon steel plate with a maximum variation of 4.8%, while a slot of 15.0 mm length, 1.0 mm width, and 3.3 mm depth in a 12.0 mm thick carbon steel plate with a maximum variation of 10.0%.
No takes yet. Share an insight, caveat, or question.
Singh et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: