Learning to Reconstruct Crack Profiles for Eddy Current Nondestructive Testing

Shaohua Li, Ayesha Anees, Yu Zhong, Zaifeng Yang, Yong Liu, Rick Siow Mong Goh, En‐Xiao LiuarXiv 2019 · DOI · arXiv

Abstract

Eddy current testing (ECT) is one of the most popular Nondestructive Testing (NDT) techniques, especially for conductive materials. Reconstructing the crack profile from measured EC signals is one of the main goals of ECT. This task is highly challenging, as the EC signals are nonlinear responses resulted from the presence of cracks, and reconstructing the crack profile requires establishing the forward model of the nonlinear electromagnetic dynamics and solving its inverse problem, which is an ill-posed numerical optimization problem. Instead of solving the inverse problem numerically, we propose to directly learn the inverse mapping from EC signals to crack profiles with a deep encoder-decoder convolutional neural network named EddyNet. EddyNet is trained on a set of randomly generated crack profiles and the corresponding simulated EC responses generated from a realistic forward model. On the held-out test data, EddyNet achieved a mean absolute error of 0.198 between predicted profiles and ground truth ones. Qualitatively, the geometries of predicted profiles are visually similar to the ground truth profiles. Our method greatly reduces the usual reliance on domain experts, and the reconstruction is extremely fast both on GPUs and on CPUs. The source code of EddyNet is released on https://github.com/askerlee/EddyNet.

Introduction

In critical industries such as aerospace and petrochemicals, Nondestructive Testing (NDT) is essential for maintaining structural integrity. Eddy Current Testing (ECT) is a widely used NDT method for inspecting conductive components. While detecting the presence of a crack is relatively straightforward, reconstructing its precise 2D or 3D profile from electromagnetic signals is a complex "inverse problem."

Schematic of the Eddy Current Testing setup showing an induction coil over a planar conductor with a 3D crack.

Traditionally, this reconstruction involves iterative numerical optimization to find a crack profile whose simulated signal matches the observed one. However, this process is computationally expensive and often ill-posed, meaning multiple physical configurations could yield similar signals. To overcome these limitations, we propose EddyNet, a deep learning framework that learns a direct mapping from EC signals to crack profiles.

The EddyNet Architecture

EddyNet utilizes an encoder-decoder convolutional neural network (CNN) architecture. The model is designed to treat the inverse problem as an image-to-image translation task, shifting the computational burden from the inference phase to the training phase.

The EddyNet architecture, featuring an encoder that compresses 6-channel EC signals into a latent code and a decoder that reconstructs the crack profile.

Encoder and Decoder

The encoder consists of five convolutional layers that process 2D EC signals. These signals are captured at three different frequencies, with the real and imaginary components of each frequency treated as separate input channels (totaling 6 channels). The encoder compresses these features into a 128-dimensional latent code.

The decoder then uses five deconvolutional (transpose convolution) layers to expand the latent code into a 40x12 pixel image representing the crack profile. A key innovation is the inclusion of an attention mechanism in the final layer, which aggregates 20 feature channels to help the model reconstruct finer geometric details.

Optimization and Activation

To improve training stability and performance, EddyNet employs the Mish activation function and the Ranger optimizer (a combination of RAdam and LookAhead). These choices help the model navigate the complex loss landscape of electromagnetic inverse problems more effectively than standard ReLU or Adam optimizers.

Data Generation

Training a deep model for ECT requires a large amount of labeled data, which is difficult to obtain experimentally. We generated a synthetic dataset of 20,000 crack-signal pairs:

  1. Crack Profiles: Random binary matrices (40x12) were generated and smoothed using a 3x3 median filter. This smoothing process creates more physically realistic, continuous shapes that are easier for a CNN to learn.
  2. Forward Modeling: EC signals were simulated using a realistic forward model based on the Boundary Element Method and the reciprocal theorem, ensuring the synthetic data reflects real-world physics.

Results

EddyNet was evaluated on a held-out test set (20% of the data).

  • Accuracy: The model achieved a Mean Absolute Error (MAE) of 0.198 on binarized profiles, significantly outperforming baseline guesses.
  • Visual Fidelity: Predicted profiles closely match the ground truth in terms of general geometry and location.
  • Inference Speed: Reconstruction takes only 5 to 15 milliseconds on standard hardware (CPU or GPU). This is orders of magnitude faster than traditional numerical methods, which typically take minutes.

Conclusion

EddyNet demonstrates that deep learning can effectively solve the challenging inverse problem in Eddy Current Testing. By learning a direct mapping from signals to profiles, it enables near-instantaneous crack reconstruction with high accuracy. Future work will focus on incorporating more physically realistic crack generation and validating the model against experimental laboratory data to account for real-world sensor noise.

Summary updated

Abstract

Eddy current testing (ECT) is one of the most popular Nondestructive Testing (NDT) techniques, especially for conductive materials. Reconstructing the crack profile from measured EC signals is one of the main goals of ECT. This task is highly challenging, as the EC signals are nonlinear responses resulted from the presence of cracks, and reconstructing the crack profile requires establishing the forward model of the nonlinear electromagnetic dynamics and solving its inverse problem, which is an ill-posed numerical optimization problem. Instead of solving the inverse problem numerically, we propose to directly learn the inverse mapping from EC signals to crack profiles with a deep encoder-decoder convolutional neural network named EddyNet. EddyNet is trained on a set of randomly generated crack profiles and the corresponding simulated EC responses generated from a realistic forward model. On the held-out test data, EddyNet achieved a mean absolute error of 0.198 between predicted profiles and ground truth ones. Qualitatively, the geometries of predicted profiles are visually similar to the ground truth profiles. Our method greatly reduces the usual reliance on domain experts, and the reconstruction is extremely fast both on GPUs and on CPUs. The source code of EddyNet is released on https://github.com/askerlee/EddyNet.
BibTeX
@misc{li2019learning,
  author = {Shaohua Li and Ayesha Anees and Yu Zhong and Zaifeng Yang and Yong Liu and Rick Siow Mong Goh and En‐Xiao Liu},
  title = {{Learning to Reconstruct Crack Profiles for Eddy Current Nondestructive Testing}},
  year = {2019},
  eprint = {1910.08721},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1910.08721},
}
Figure