Eddy Current Testing (ECT) is a well-established method of Nondestructive Testing (NDT) for ensuring the integrity of conductive materials. While defect detection is wellestablished, the transition to accurate quantitative sizing-predicting the precise dimensions of a flaw-remains a significant challenge, yet it is crucial for structural health assessment. This transition is hindered by two main obstacles: the difficulty in obtaining high-fidelity ground-truth data for model training, and the inherent class imbalance of inspection data, where defect-free regions vastly outnumber defective ones, biasing traditional machine learning models. This study proposes a comprehensive framework to overcome both challenges. First, we introduce a geometry-guided labeling methodology that leverages a spherical parameterization of the defect to create a dense, point-by-point depth map, ensuring highly accurate training data. Second, we propose a hybrid classification-regression architecture. This model first employs a Gradient Boosting classifier to accurately distinguish between defective and non-defective signal regions. Subsequently, a specialized Random Forest regressor, trained exclusively on data from defective regions, performs the final depth prediction. This two-stage approach prevents the regression model from being biased by the overwhelming number of zero-depth samples. The proposed hybrid model’s performance was rigorously evaluated and compared against both a standard direct-regression model. The results demonstrate the best performance of the hybrid architecture, achieving a significant improvement in predictive accuracy for defect depth. The primary contribution of this work is an integrated methodology that combines physics-based geometric labeling with a specialized machine learning architecture, offering a robust and accurate solution for quantitative defect characterization in real-world NDT applications.

EMAT technology for Non Destructive testing is an important method for materials testing in several industries. In EMAT tools, a key issue is the EMAT coils design and implementation. Depending on the type of inspection, the coil type should be selected, and then, its dimensions should be calculated. This paper describes a methodology to select, design and implement EMAT coils based on Lorentz Force for applications such as thickness measurement and crack detection. Unlike previous works that focus on a single coil topology, this study integrates coil selection, dimensional design, COMSOL-based radiation-pattern simulation and experimental validation within a single workflow. Four Lorentz-force coil designs are covered: PCB spiral (CSPCB), 3D-printed spiral (CS3D), PCB meander-line (CMPCB) and 3D-printed meander-line (CM3D). Key design parameters are explicitly addressed: number of turns N, outer and inner radii R and $r_0$, track width w and spacing s for spiral coils, and meander length and inter-trace distance for meander-line coils. Simulation verification is performed in COMSOL Multiphysics by evaluating the von Mises stress along a semicircular path around the coil to obtain the angular radiation pattern. Experimentally, polar radiation patterns are measured at 500 kHz, 1.9 MHz and 4 MHz on a steel specimen, matching the simulation frequencies, with maximum amplitudes of 32.2, 46.4, 47.9 and 10.6 mV for CSPCB, CS3D, CMPCB and CM3D, respectively, showing consistent agreement between simulated and measured lobe shape and directivity. This work also uses an analogy with radio frequency antennas to better understand the operation of coils through the concept of radiation patterns, in this case in solid materials such as steel.

Eddy Current Testing (ECT) is a non-destructive testing method for identifying irregularities in electrically conductive materials using electromagnetic induction. While ECT is precise at detecting flaws, accurately determining their size remains challenging. The two main obstacles are obtaining high-quality ground-truth data for training machine learning models and handling class imbalance, since defect-free regions greatly outnumber those with defects, leading to biased predictive models. To address these challenges, this work presents a hybrid classification-regression framework based on a Hurdle model formulation to quantify defect size. The framework is tested on an unbalanced laboratory database. In this approach, defect morphology is reconstructed as the product of defect probability and estimated depth. The probability is obtained from a classifier model, while depth estimation is obtained from a regression model. The algorithms considered are a gradient-boosted classifier with probability calibration and a random-forest regressor. To evaluate the proposed framework, its performance is assessed in two ways: (1) spatial reconstruction–predicting the shape of a physically connected (contiguous) region held out from the training data–with a Mean Absolute Error (MAE) between 0.32 and 0.54 mm over all scan points of the held-out region at lift-off distances between 0.25 and 1.0 mm, and (2) Leave-One-Experiment-Out (LOEO) cross-validation across 21 different flaws (defect locations), yielding a median MAE of 0.29 mm over all scan points. Restricted to defective points and grouped by lift-off distance, the median of the MAE ranged from 0.82 to 1.16 mm. The LOEO strategy also examines the effects of flaw size (defect diameter), vertical size (defect depth), and sensor distance (lift-off) on the model’s predictive ability. Results show that deeper defects are associated with lower regression accuracy, whereas defect diameter and lift-off are not statistically significantly related to framework performance.

Centrifugal pumps handle fluids with properties that evolve due to changes in composition. Essential pump fluid properties, such as viscosity and density, significantly influence the performance of multi-stage pumps. This study establishes a correlation between vibration levels and fluid properties. It introduces an approach based on feature extraction and Artificial Neural Networks (ANN) to predict these properties.The methodology begins by gathering mechanical vibration data from a nine-stage pump operating with water or oil. Mechanical vibration signal processing involves applying the Discrete Fourier Transform, selecting frequency bands, and extracting the Root Mean Square as a feature for each frequency band. Extracted features, the pump pressure gain, and the pump rotation speed are the input selected features for an ANN model.The ANN architecture is a Multi-Layer Perceptron Artificial Neural Network (MLP-ANN). The model hyperparameters selection consists of establishing an optimization space and the optimization implementation based on a Tree-structured Parzen Estimator algorithm. Hyperparameter includes the number of layers, the activation functions kind for each layer, and neurons per layer. The obtained fitting results demonstrate errors of less than 3 [cP] within a viscosity range of 1 to 200 [cP] and errors less than 9 [kg/m ³ ] within a density range of 860 to 1000 [kg/m ³ ].