The demo code for "Advanced electromagnetic modeling in pulsed eddy current testing of oil well casings"
Authors: Qiang Feng, Yibing Shi, Aihua Tao, Zhipeng Li and Wei Zhang
Abstract: The timeliness and accuracy of electromagnetic (EM) response estimation are severely limited by EM modeling methods used for pulsed eddy current testing (PECT) of oil well casing. Traditional mathematical analysis methods require a high degree of prior mathematical knowledge and are computationally expensive. Meanwhile, contemporary deep learning-based purely data-driven methods lack a convincing physical explanation and are less robust. This work fills the gap by presenting a novel physics-informed neural network (PINN) surrogate model that combines efficiency and interpretability to address the challenging task of estimating the EM response of oil casing in PECT. Specifically, the EM physics laws, are embedded as prior knowledge into the objective loss function to supervise the deep neural networks (DNNs) training process. This significantly enhances the interpretability of DNNs and improves the accuracy of estimating the responses. Secondly, sub-neural networks are used to estimate the EM responses of different computational domains, which are separated according to their physical properties. An interface loss is designed to compensate for the significant discontinuity and incompatibility between the separate networks when estimating the results of double domains. Thus improving the accuracy and robustness of EM response estimation. The performance of the proposed PINN is validated by EM response datasets obtained from finite element analysis (FEA) methods. The comprehensive analysis demonstrated that the PINN can accurately estimate the EM responses in oil casing PECT with a coefficient of determination exceeding 0.95 (95%). Furthermore, the inference time of the PINN is more than 52 times faster than that of FEA.
- Dependent environment:
- Pytorch = 1.13.1
- numpy
- matplotlib
- pandas
- scikit-learn
- Run the code:
-
Double_AB_regular_2NN_adaptive.py: Simultaneous estimation of the pulsed eddy current electromagnetic response (magnetic vector potential A and magnetic flux density B) in the air and casing domains using an interface loss and an adaptive weight balancing algorithm.
-
Single_A1B1_regular_2NN_adaptive.py: Only estimation of the pulsed eddy current electromagnetic response (magnetic vector potential A and flux density B) in the air domain using an adaptive weighting algorithm.