Pulsed eddy current (PEC) testing is a non-destructive testing (NDT) method, which is highly suitable for thickness measurement, and sub-surface defect detection. In this paper, PEC testing is employed to simultaneously detect defect and sample thickness. Firstly, Finite-element method (FEM) is conducted to reveal the signal characteristics of different defect sizes and depths on the aluminum plate, as well as its thickness. Experimental verification is conducted to verify the simulation results. It reveals that the signal variations associated with defect size and depth are relatively smaller compared to the impact of thickness changes. Both logarithmic function and discrete Fourier transform (DFT) are utilized as signal processing method to improve signal differentiation between defect characteristics. Notably, the processed signal through DFT produces both amplitude and phase spectra, with a focus on the enriched features within the high-frequency portion of the signals post-DFT. Lastly, all the raw and processed data are fed into deep learning (DL) models. The optimized Res2net18 network with post-DFT data demonstrates superior performance. It achieves classification accuracies of 98.0%, 100%, and 96.0% for defect size, thickness of the sample, and defect depth, respectively. The results show the proposed DL network has faster convergence and stability in comparison to the Res2Net network and DenseNet121, and there are 1.5% and 2.0%, 5.4% and 9.7% classification accuracy improvement in defect size and depth. The predict values based on Optimized Res2Net Network also present consistent results with validation and test results.

Lift-off effect, resulting in detection error, which is a significant challenge that needs to be addressed in defect detection using eddy current testing (ECT). A novel array eddy current sensor, with each unit composing of 4 excitation coils and a receiving coil, is proposed to implement fast and large-area defect scanning for metal sheet without lift-off interference. In addition, theoretical analysis based on Biot-Savart law is applied to illustrate the feasibility of the proposed sensor in eliminating lift-off effect. However, the proposed array sensor has also an angular sensitivity which means there are different signal responses for different defect orientations, but the feature is not conducive to evaluate the defect depths without knowing the angle difference as a prior, which can be addressed by deep learning (DL) algorithm. Both simulation and experiment are implemented to verify the reliability and feasibility of proposed array sensor. The proposed preprocessed method is utilized to attain significant signal feature differences for various defect depths. In addition, the improved ResNet18 network is proposed to train acquired experimental data, and its superior performance is demonstrated. The results show that the improved ResNet18 network with preprocessed data achieves fastest training speed compared to other methods in this paper. At same time, the proposed DL network achieves high accuracy of 98.1% and 95.4% for defects depth for aluminum sheet and stainless-steel sheet, respectively. Additionally, when considering 16 labels combinations of lift-off distances and angle differences during the process of inspection, the accuracy of lift-off classification for two materials are 97.2% and 96.9%, respectively.

Pulsed eddy current (PEC) testing is a non-destructive testing (NDT) technique that is widely used in the industry. The ability to detect corrosion in metal materials is crucial for ensuring safety and mitigating potential hazards. This paper provides a solution from hardware to software to show the ability of deep learning (DL) method to process PEC data under multiple complex distortions, and we also deliver a real-time DL edge computing system for metal thickness recognition. There are three key contributions. Firstly, we constructed a new dataset generated from our customdesigned PEC device with a diverse range of features. To better simulate real-world scenarios, measurement covers various thickness, lift-off, position, insulation and weather jacket conditions. Secondly, we adapted 1D convolutional neural network (CNN) to process the time series PEC data. This approach achieves high accurate thickness recognition, and the prediction is not affected by distortions, such as lift-off and edge effects. Lastly, we integrated a compact and tiny CNN into the STM32 microcontroller within our device. This edge computing system achieves real-time, accurate and low-latency thickness recognition. Our study represents a significant advancement towards the development of automated thickness recognition technologies in the NDT based on embedded DL. Our dataset and DL models are publicly available at Kaggle ^a and GitHub ^b .

The detection of lift-off variation presents a significant challenge in many eddy current testing (ECT) applications, where it is often considered an undesirable signal. To address this issue, the paper proposes a novel 4-coil excitation sensor, which can effectively suppress the lift-off effect when detecting defects of varying depths in planar structures. The proposed sensor design leverages opposite excitation directions between any two arbitrary adjacent excitation coils, which can effectively cancel out the lift-off signal due to a defect-free plate. However, this 4-coil excitation design has an angular sensitivity – i.e., different signal response for different angles between the same defect and the sensor. This angular sensitivity will cause trouble to identify the depth of defects without knowing the angle in prior. To overcome this challenge, the current study introduces deep learning (DL) models as a potential solution. Signals representing different defect depths and angles between the sensor and the defect are acquired using an experimental platform in our lab. The real and imaginary parts of the induced voltage on the receiver coil labelled with depths and angles are utilized to train a 1-D convolutional neural network (1D CNN) model, in which different convolution kernel sizes are used to extract more information and speed up convergence. The high performance of the proposed structure of 1D CNN were compared with convolutional long-short term memory (LSTM) model, recurrent neural network (RNN) model, LSTM+RNN model and ResNet18 model. The proposed 1D CNN model not only achieves good performances with a higher accuracy of 99.8% but also requires a smaller model size and less computational time.

Steel microstructure determines the mechanical properties of steel such as hardness and tensile strength. To monitor steel microstructure evolution during thermal processing such as hot rolling, online measurement of microstructure is vital to steel quality control. Previous studies using multi-frequency eddy-current testing to monitor the change of electromagnetic properties and infer steel microstructures have achieved remarkable success, yet due to the high-frequency excitation signal, the inspected thickness of steel microstructures is constrained on the surface due to the skin effect. This study introduces the pulsed eddy-current (PEC) method to characterise the electromagnetic property profile along the depth of thick steel plates. The low-frequency square wave signal enables PEC to reflect multi-layer characteristics in a single measurement with good penetration ability. An explicit analytical equation for calculating the time domain PEC response is derived for the first time which shows advantages in terms of speed and gaining in-depth physical insights. A custom PEC system has been designed and built with a TMR sensor as the receiver. Theoretical and numerical analysis are performed. A time domain intersection point on the Tau curve is found to be a useful indication of the properties of a planar structure. Experimental results on layered planar structures of various electromagnetic properties are analysed and validated.