Imaging steel plate defects by planar electromagnetic tomography with deep convolutional neural network

Area of Science:

• Materials Science
• Non-destructive Testing
• Computational Imaging

Background:

• Accurate detection of metal material defects is crucial for production and safety.
• Damage to metal materials alters local magnetic permeability, a key factor in electromagnetic tomography (EMT).
• The inverse problem in EMT for ferromagnetic materials is ill-posed and ill-conditioned, challenging defect detection.

Purpose of the Study:

• To propose an improved deep learning model, P-LeNet, for enhanced EMT defect detection and image reconstruction.
• To address the challenges posed by the ill-posed nature of EMT and the properties of ferromagnetic materials.
• To improve the accuracy and robustness of reconstructing metal material defects.

Main Methods:

• Developed P-LeNet, a convolutional neural network-based model for EMT.
• Established a nonlinear mapping between induced voltage measurements and material property distribution.
• Utilized multi-scale feature extraction to enhance reconstruction.
• Evaluated image reconstruction quality using correlation coefficient and image error.

Main Results:

• P-LeNet demonstrated superior imaging accuracy, artifact suppression, and overall performance compared to traditional algorithms in numerical simulations.
• The model exhibited strong anti-noise ability when subjected to Gaussian white noise.
• P-LeNet showed good generalization ability when tested with random samples.
• Experimental validation using a nine-coil planar EMT sensor confirmed the model's effectiveness.

Conclusions:

• The proposed P-LeNet model offers a significant advancement in EMT-based metal defect detection and image reconstruction.
• The deep learning approach effectively overcomes the inherent challenges of the EMT inverse problem for ferromagnetic materials.
• P-LeNet shows considerable potential for practical applications in non-destructive testing and material evaluation.