A method and system for eddy current testing of surface cracks in a metal material

Abstract

The application provides a kind of metal material surface crack eddy current testing method and system, belong to testing field, method includes using double-frequency excitation probe to scan metal material surface, to detect crack depth characteristic with low-frequency excitation, high-frequency excitation detects lift-off effect, acquires composite eddy current response signal and separates out double-frequency corresponding component signal by time-frequency analysis reconstruction;From low-frequency component, extract instantaneous phase sequence and first-order to third-order time derivative, construct two-dimensional phase space trajectory and calculate curvature, combine to form real-time feature vector;From high-frequency component, extract lift-off effect parameters, generate lift-off influence weight function;Rely on the discrete point cloud manifold and Riemann metric of standard crack sample construction, calculate the initial geodesic distance of real-time feature vector to crack prototype set, adjust to get minimum geodesic distance after weight function, if the distance is less than the preset threshold, it is judged that there is crack on the surface of metal material. Solve the stable representation of feature sensitivity, establish accurate identification model.

Description

Technical Field

[0001] This application belongs to the field of testing, and in particular relates to a method and system for eddy current testing of surface cracks in metallic materials. Background Technology

[0002] Eddy current nondestructive testing (EDT), a detection method based on the principle of electromagnetic induction, is widely used for online monitoring and offline evaluation of early-stage surface and near-surface cracks in metal components due to its advantages such as non-contact operation, high sensitivity, and fast detection speed. However, in actual testing, the eddy current detection signal is extremely weak and often drowned out by various strong noises and interferences. Among these, the lift-off effect caused by minute fluctuations in the distance between the probe and the surface of the workpiece is the main source of interference. The resulting signal intensity can even far exceed that of micro-crack defects, leading to a low signal-to-noise ratio in traditional detection methods and a high likelihood of false alarms or missed detections. This severely restricts the application of eddy current testing technology in the field of weak defect identification.

[0003] Existing technologies typically employ strategies such as multi-frequency excitation, probe structure optimization, and complex signal processing algorithms. Dual-frequency or multi-frequency excitation techniques compensate for the differences in sensitivity to cracks and lift-off effects using eddy current signals of different frequencies, or distinguish defects from interference by analyzing the amplitude and phase characteristics of signal trajectories on impedance plane diagrams. These methods mostly rely on linear or quasi-linear models to suppress lift-off interference; however, the separation and compensation effects decrease when lift-off fluctuations are severe or when there is significant coupling with crack signals. Traditional identification methods based on single or combined features such as signal amplitude and phase fail to adequately utilize information about the signal evolution process, exhibiting poor feature stability and unable to stably represent the intrinsic properties of cracks under strong interference. Therefore, how to construct a stable feature representation that is sensitive to crack geometry but insensitive to interference such as lift-off under complex working conditions, and establish an accurate identification model, is a pressing technical challenge in the field of eddy current detection. Summary of the Invention

[0004] To address the problem that existing technologies cannot construct stable feature representations that are sensitive to crack geometry features and insensitive to disturbances such as lift-off under complex working conditions, and cannot establish accurate identification models.

[0005] In the first aspect, the present invention proposes a method for testing eddy current cracks on the surface of metallic materials, comprising: A dual-frequency excitation probe is used to scan the surface of a metallic material, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The probe also collects the composite eddy current response signal. Time-frequency analysis is performed on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal. The instantaneous phase sequence and first to third-order time derivatives are extracted from the first component signal. A two-dimensional phase space trajectory is constructed using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and the curvature of the trajectory is calculated. The instantaneous phase and the curvature are combined to form a real-time feature vector. Parameters representing the lift-off effect in real time are extracted from the second component signal, and a lift-off effect weighting function is generated based on the parameters. Based on a discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemann metric, the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold is calculated; the initial geodesic distance is adjusted using the lift-off influence weight function to obtain the minimum geodesic distance; when the minimum geodesic distance is less than a preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

[0006] Optionally, performing time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal includes: The composite eddy current response signal was processed using short-time Fourier transform, with a Hanning window set as the window function, a window length of 512 sampling points, and an overlap length of 256 sampling points. In the generated time spectrum, the center frequency bands of the first excitation frequency and the second excitation frequency are located respectively, and the spectral data within the two frequency bands are extracted. The inverse short-time Fourier transform is performed on the spectral data of the two extracted frequency bands to obtain the first component signal and the second component signal.

[0007] Optionally, the step of extracting the instantaneous phase sequence and first to third-order time derivatives from the first component signal, constructing a two-dimensional phase space trajectory using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculating the curvature of the trajectory includes: Apply the Hilbert transform to the first component signal to obtain the analytic signal, and calculate the instantaneous phase sequence of the analytic signal; The first, second, and third time derivatives of the instantaneous phase sequence are calculated iteratively using the central difference method. Based on the first-order time derivative, the second-order time derivative, the third-order time derivative, and the preset time constant, the curvature of the trajectory is calculated using the parametric curve curvature formula to obtain a curvature feature sequence.

[0008] Optionally, the step of extracting parameters representing the lift-off effect in real time from the second component signal and generating a lift-off effect weighting function based on the parameters includes: The lift-off influence weight function W(t) is generated by adding a regularization constant to a Gaussian function, and its expression is: ,in The envelope amplitude is extracted in real time from the second component signal. As the reference envelope amplitude, is the attenuation control coefficient, and C is the preset regularization constant.

[0009] Optionally, combining the instantaneous phase and the curvature to form a real-time feature vector includes: At each time point, the instantaneous phase at the current moment is taken as the first dimension feature, the phase trajectory curvature is taken as the second dimension feature, and the two scalar values ​​are combined into a two-dimensional column vector.

[0010] Optionally, the step of calculating the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold, based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemannian metric, includes: The feature vector set of the standard crack sample is used as discrete sampling points on the manifold; Construct a k-nearest neighbor graph with each sampling point as a node, connecting each node to its k nearest neighbors, and use the Euclidean distance between nodes as the initial edge weights; The k-nearest neighbor graph structure is used as a discrete approximation of the Riemannian manifold, and the Riemannian metric is implicitly defined by the local Euclidean distance structure. Dijkstra's algorithm is used to calculate the shortest path length between all pairs of nodes in the graph, which is then used as the initial geodesic distance between them.

[0011] Optionally, adjusting the initial geodesic distance using the lift-off influence weighting function to obtain the minimum geodesic distance includes: In the feature space, find the sampling point on the manifold that has the closest Euclidean distance to the current real-time feature vector; Query the initial geodesic distance from the nearest sampling point to the preset sampling points of multiple crack prototype sets; Multiply the initial geodesic distance by the lift-off influence weight function to obtain the adjusted geodesic distance; The minimum value of the adjusted geodesic distance is selected as the minimum geodesic distance.

[0012] Optionally, the step of determining that a crack exists on the surface of the metal material when the minimum geodesic distance is less than a preset criterion threshold includes: By testing known cracked samples and known uncracked samples, two sets of minimum geodesic distance distributions were obtained; The receiver operating characteristic curve is used to analyze the two sets of minimum geodesic distance distributions. The geodesic distance value corresponding to the maximum Youden index is selected as the preset criterion threshold so that the sum of the detection sensitivity and specificity is maximized.

[0013] On the other hand, the present invention also proposes a surface crack eddy current testing system for metallic materials, comprising the following modules: The reconstruction module is used to scan the surface of a metal material using a dual-frequency excitation probe, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The module also acquires the composite eddy current response signal generated by the probe. Time-frequency analysis is performed on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal. The generation module is used to extract the instantaneous phase sequence and first to third-order time derivatives from the first component signal, construct a two-dimensional phase space trajectory with the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculate the curvature of the trajectory; combine the instantaneous phase and the curvature to form a real-time feature vector; extract parameters representing the lift-off effect in real time from the second component signal, and generate a lift-off effect weighting function based on the parameters; The determination module is used to calculate the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and the preset Riemann metric; adjust the initial geodesic distance using the lift-off influence weight function to obtain the minimum geodesic distance; when the minimum geodesic distance is less than the preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

[0014] Preferably, performing time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal includes: The composite eddy current response signal was processed using short-time Fourier transform, with a Hanning window set as the window function, a window length of 512 sampling points, and an overlap length of 256 sampling points. In the generated time spectrum, the center frequency bands of the first excitation frequency and the second excitation frequency are located respectively, and the spectral data within the two frequency bands are extracted. The inverse short-time Fourier transform is performed on the spectral data of the two extracted frequency bands to obtain the first component signal and the second component signal.

[0015] Preferably, the step of extracting the instantaneous phase sequence and first to third-order time derivatives from the first component signal, constructing a two-dimensional phase space trajectory using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculating the curvature of the trajectory includes: Apply the Hilbert transform to the first component signal to obtain the analytic signal, and calculate the instantaneous phase sequence of the analytic signal; The first, second, and third time derivatives of the instantaneous phase sequence are calculated iteratively using the central difference method. Based on the first-order time derivative, the second-order time derivative, the third-order time derivative, and the preset time constant, the curvature of the trajectory is calculated using the parametric curve curvature formula to obtain a curvature feature sequence.

[0016] Preferably, the step of extracting parameters representing the lift-off effect in real time from the second component signal and generating a lift-off effect weighting function based on the parameters includes: The lift-off influence weight function W(t) is generated by adding a regularization constant to a Gaussian function, and its expression is: ,in The envelope amplitude is extracted in real time from the second component signal. As the reference envelope amplitude, is the attenuation control coefficient, and C is the preset regularization constant.

[0017] Preferably, the step of combining the instantaneous phase and the curvature to form a real-time feature vector includes: At each time point, the instantaneous phase at the current moment is taken as the first dimension feature, the phase trajectory curvature is taken as the second dimension feature, and the two scalar values ​​are combined into a two-dimensional column vector.

[0018] Preferably, the step of calculating the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold, based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemannian metric, includes: The feature vector set of the standard crack sample is used as discrete sampling points on the manifold; Construct a k-nearest neighbor graph with each sampling point as a node, connecting each node to its k nearest neighbors, and use the Euclidean distance between nodes as the initial edge weights; The k-nearest neighbor graph structure is used as a discrete approximation of the Riemannian manifold, and the Riemannian metric is implicitly defined by the local Euclidean distance structure. Dijkstra's algorithm is used to calculate the shortest path length between all pairs of nodes in the graph, which is then used as the initial geodesic distance between them.

[0019] Preferably, adjusting the initial geodesic distance using the lift-off influence weighting function to obtain the minimum geodesic distance includes: In the feature space, find the sampling point on the manifold that has the closest Euclidean distance to the current real-time feature vector; Query the initial geodesic distance from the nearest sampling point to the preset sampling points of multiple crack prototype sets; Multiply the initial geodesic distance by the lift-off influence weight function to obtain the adjusted geodesic distance; The minimum value of the adjusted geodesic distance is selected as the minimum geodesic distance.

[0020] Preferably, the step of determining that a crack exists on the surface of the metal material when the minimum geodesic distance is less than a preset criterion threshold includes: By testing known cracked samples and known uncracked samples, two sets of minimum geodesic distance distributions were obtained; The receiver operating characteristic curve is used to analyze the two sets of minimum geodesic distance distributions. The geodesic distance value corresponding to the maximum Youden index is selected as the preset criterion threshold so that the sum of the detection sensitivity and specificity is maximized.

[0021] This invention employs dual-frequency excitation technology to separate crack depth characteristics and lift-off effects at the signal level. A compensation mechanism for the lift-off effect is established using high-frequency components, thereby suppressing critical lift-off noise interference in eddy current detection. By deeply exploring the signal phase information in low-frequency components and constructing a feature vector containing instantaneous phase, phase change rate, and phase space trajectory curvature, a feature representation with higher discriminative power and stability for crack morphology is obtained. Combining manifold learning and geodesic distance criteria, and using the lift-off compensation mechanism to correct the discrimination process, accurate crack identification is ensured even under lift-off fluctuations, improving the accuracy, reliability, and anti-interference capability of eddy current detection of cracks on metallic material surfaces. Attached Figure Description

[0022] Figure 1 A flowchart of the first embodiment; Figure 2 A schematic diagram of the receiver's operational characteristic curve; Figure 3 This is a schematic diagram comparing the lift-off height with the detection accuracy. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0024] Firstly, this invention proposes a method for testing eddy current in cracks on the surface of metallic materials, such as... Figure 1 As shown, it includes the following steps: S1, a dual-frequency excitation probe is used to scan the surface of the metal material, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The composite eddy current response signal generated by the probe is also collected.

[0025] Use a signal generator to generate the first sinusoidal excitation signal. With the second sinusoidal excitation signal , Set to 15kHz. The frequency is set to 300kHz. Two signals are linearly superimposed using an analog adder circuit to form a composite excitation signal. This composite signal is amplified by a power amplifier to drive the differential eddy current probe coil. A two-dimensional or three-dimensional scanning frame controlled by a stepper motor moves the probe along a preset grid path at a constant speed to scan the surface of the metal material. The voltage signal induced by the probe's pickup coil is amplified by a preamplifier and digitized by a high-speed analog-to-digital converter (ADC) with a sampling frequency of at least 1.8MHz to obtain a discrete time series of the composite eddy current response signal.

[0026] S2, perform time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal.

[0027] The acquired composite eddy current response signal is processed using the Hilbert-Huang Transform (HHT) method. The Hilbert-Huang Transform is an adaptive time-frequency analysis method that first decomposes the signal into several intrinsic mode functions (IMFs) through empirical mode decomposition (EMD), and then performs a Hilbert transform on each component to obtain the instantaneous frequency and instantaneous amplitude. The ensemble empirical mode decomposition (EEMD) algorithm is then applied to decompose the composite signal into a set of IMFs. The instantaneous frequency of each IMF component is calculated and compared with a preset excitation frequency. and By comparison, the center frequency closest to the center frequency is identified. The IMF component is used as the first component signal, and the center frequency is closest to The IMF component is used as the second component signal and the signal is reconstructed.

[0028] In an alternative embodiment, performing time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal includes: The composite eddy current response signal was processed using short-time Fourier transform, with a Hanning window set as the window function, a window length of 512 sampling points, and an overlap length of 256 sampling points. In the generated time spectrum, the center frequency bands of the first excitation frequency and the second excitation frequency are located respectively, and the spectral data within the two frequency bands are extracted. The inverse short-time Fourier transform is performed on the spectral data of the two extracted frequency bands to obtain the first component signal and the second component signal.

[0029] Set the sampling frequency of the data acquisition system, for example, to 1.8MHz, to ensure distortion-free detection of high-frequency signals. Assume the first excitation frequency is used for depth detection. =15kHz, the second excitation frequency is used for lift-off detection. =300kHz. Short-time Fourier Transform (STFT) was performed on the acquired discrete-time sequence signal x[n]. The core parameters of the STFT were set as follows: the window function was selected as the Hanning window with good smoothness to reduce spectral leakage; the window length L was set to 512 sampling points, which achieved a balance between ensuring sufficient frequency resolution and time resolution; the overlap length was set to 256 sampling points, i.e., 50% overlap rate, to ensure smooth transition in time and avoid frame boundary effects.

[0030] STFT processing generates a complex matrix, i.e., the time spectrum. The next step is to locate and separate the energies of the two excitation frequencies within this time spectrum. The first excitation frequency... The frequency bin index corresponding to =15kHz is approximately 4.27, rounded down to 4; second excitation frequency The frequency bin index corresponding to 300kHz is approximately 85.33, rounded to 85. To fully detect signal energy, a frequency band is defined around the center frequency bin. For example, selecting 5 bins on each side of the center frequency bin, the first frequency band ranges from index -1 to 9, actually taken as 0 to 9, and the second frequency band ranges from index 80 to 90. Two new zero matrices with the same dimensions as the original time-spectrum are created, and the spectral data from the first and second frequency bands are copied to their corresponding positions. The inverse short-time Fourier transform (ISTFT) is performed on the two time-spectrum matrices containing only single-band data to reconstruct the clean first component time-domain signal. Second component time domain signal .

[0031] S3, extract the instantaneous phase sequence and first to third-order time derivatives from the first component signal, construct a two-dimensional phase space trajectory with the instantaneous phase as the first coordinate and the product of the first-order time derivative and the preset time constant as the second coordinate, and calculate the curvature of the trajectory; combine the instantaneous phase and the curvature to form a real-time feature vector.

[0032] Apply the Hilbert transform to the first component signal obtained from the reconstruction, and use the scipy.signal.hilbert function to calculate the analytic signal. ; through the Take the arctangent and perform a phase expansion operation to obtain a continuous instantaneous phase sequence. The numerical differentiation method is used to calculate... The time derivative is specifically calculated using a Savitzky-Gore filter based on local polynomial least squares fitting to determine the instantaneous frequency of the first derivative. Second derivative and third derivative This improves the noise resistance of derivative calculations. A time constant is set. Construct a set of trajectory points in a two-dimensional phase space ;Calculate the curvature based on the parametric curve formula Substituting the calculated derivatives, we obtain the trajectory curvature at each time point. ; Instantaneous phase With curvature Combined, they form the two-dimensional real-time feature vector at that moment. In an optional embodiment, the step of extracting the instantaneous phase sequence and first to third-order time derivatives from the first component signal, constructing a two-dimensional phase space trajectory using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculating the curvature of the trajectory includes: Apply the Hilbert transform to the first component signal to obtain the analytic signal, and calculate the instantaneous phase sequence of the analytic signal; The first, second, and third time derivatives of the instantaneous phase sequence are calculated iteratively using the central difference method. Based on the first-order time derivative, the second-order time derivative, the third-order time derivative, and the preset time constant, the curvature of the trajectory is calculated using the parametric curve curvature formula to obtain a curvature feature sequence.

[0033] For the first component signal Perform Hilbert transform to obtain orthogonal components Thus, an analytical signal is constructed. The instantaneous phase sequence is obtained by calculating the argument of the analytic signal. And perform a phase unwinding operation on the sequence to eliminate The jumps yield continuously changing phase functions. For the discretized phase sequence... Numerical differentiation is performed using the central difference method to improve accuracy. Let the sampling time interval be... For example, 0.55s, then the first derivative Second derivative Third derivative .

[0034] After obtaining the instantaneous phase and derivatives of each order, a two-dimensional phase space trajectory is constructed. This trajectory is a parametric curve. ,in , Preset time constant This is an adjustable parameter used to scale the second coordinate axis and optimize the trajectory shape. Its value can be set to an integer multiple of the sampling interval, for example... The curvature of the parametric curve Calculate using standard formulas. Provide the expressions and derivatives of x(t) and y(t). , , , Substituting the values, we obtain the formula for calculating curvature: Calculate for each time point n This generates a curvature feature sequence synchronized with the phase sequence.

[0035] In one embodiment, combining the instantaneous phase and the curvature to form a real-time feature vector includes: At each time point, the instantaneous phase at the current moment is taken as the first dimension feature, the phase trajectory curvature is taken as the second dimension feature, and the two scalar values ​​are combined into a two-dimensional column vector.

[0036] For each sampling point n acquired, a corresponding instantaneous phase value is calculated. and a curvature value To construct the feature vector, the two time-aligned scalars mentioned above are combined. Mathematically, the real-time feature vector V[n] corresponding to the nth sampling point is represented as... This is a two-dimensional column vector, where the first element is the instantaneous phase and the second element is the curvature.

[0037] For example, during the scanning process, The instantaneous phase calculated at sampling point i at time t is =2.345 radians, phase trajectory curvature is =1.572. Therefore, the eigenvector at that moment is... As the probe moves, at the next sampling time... This may result in new values, such as radian, This generates new feature vectors. The entire scanning process generates a sequence of two-dimensional vectors, which completely represents the trajectory of the measured signal in the phase-curvature feature space.

[0038] S4, extract parameters representing the lift-off effect in real time from the second component signal, and generate a lift-off effect weighting function based on the parameters.

[0039] The reconstructed second component signal Calculate the analytic signal using the Hilbert transform. ;extract instantaneous amplitude As a parameter representing the lift-off effect. Calculated when scanning defect-free regions. The mean is used as the reference value. The weighting function W(t) for lifting out influence is in Gaussian form.

[0040] In an optional embodiment, the step of extracting parameters representing the lift-off effect in real time from the second component signal and generating a lift-off effect weighting function based on the parameters includes: The lift-off influence weight function W(t) is generated by adding a regularization constant to a Gaussian function, and its expression is: ,in The envelope amplitude is extracted in real time from the second component signal. As the reference envelope amplitude, is the attenuation control coefficient, and C is the preset regularization constant.

[0041] Before the test begins, the eddy current probe is fully attached to the surface of a defect-free piece of the same metal material. The second component signal is then measured. The envelope amplitude is defined as the reference envelope amplitude. The envelope amplitude can be calculated. The modulus of the analytic signal is used to obtain the result, i.e. , where H is the Hilbert transform. For example, measured during calibration. =1.2V. As required, the attenuation control coefficient is set to 0.5 times the reference amplitude, i.e. =0.6V. The regularization constant C is set to a small positive value, such as C=0.01, to prevent division by zero errors in subsequent calculations caused by the weight function being zero.

[0042] During real-time scanning, the envelope of the second component signal is continuously extracted. As a parameter representing the lift-off effect, an increase in lift-off distance leads to a decrease in eddy current induction efficiency, thereby... Decrease. For example, at a certain time t, due to the lift-off effect, the measured real-time envelope amplitude is... =0.7V. Therefore, the weight function value is W(t) = 0.717. The weight value being less than 1 indicates a significant lift-off effect, requiring adjustments to subsequent distance calculations.

[0043] S5. Based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and the preset Riemannian metric, calculate the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold.

[0044] In the offline phase, standard crack samples with known sizes and shapes are scanned to extract feature vectors V corresponding to all crack signals, forming a high-density two-dimensional point cloud. Based on this point cloud, a k-nearest neighbor graph (k-NNGraph) is constructed, where each feature vector is a graph node, and each node is connected to its k nearest neighbors in Euclidean space. The edge weights are the Euclidean distances between nodes. This graph structure approximates the manifold containing the data, and the edge weights define the Riemannian metric. K-Means clustering is used to identify several clusters in the point cloud, and the centroid of each cluster is used as the crack prototype. Each centroid is replaced with the feature vector node in the point cloud closest to that centroid, forming a crack prototype set. During the online computation phase, for the currently acquired real-time feature vector V(t), a point is temporarily added to the pre-constructed k-NN graph and connected to the k nodes in the feature space with the closest Euclidean distance. The edge weights are Euclidean distances. The Dijkstra algorithm or A* search algorithm is used to compute V(t) to each prototype in the crack prototype set. shortest path distance That is, the geodesic distance; the minimum value among all geodesic distances is taken as the initial geodesic distance, i.e. .

[0045] In one embodiment, calculating the initial geodesic distance from the real-time feature vectors to a preset set of crack prototypes on the manifold, based on a discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemannian metric, includes: The feature vector set of the standard crack sample is used as discrete sampling points on the manifold; Construct a k-nearest neighbor graph with each sampling point as a node, connecting each node to its k nearest neighbors, and use the Euclidean distance between nodes as the initial edge weights; The k-nearest neighbor graph structure is used as a discrete approximation of the Riemannian manifold, and the Riemannian metric is implicitly defined by the local Euclidean distance structure. Dijkstra's algorithm is used to calculate the shortest path length between all pairs of nodes in the graph, which is then used as the initial geodesic distance between them.

[0046] Prepare a set of representative standard crack samples, for example, EDM grooves with depths of 0.2 mm, 0.5 mm, 1.0 mm, and 1.5 mm. Perform a comprehensive eddy current scan on the samples and extract feature vectors V[n] for all time moments according to the aforementioned steps. Collect all feature vectors extracted from all standard samples, for example, a total of 50,000 two-dimensional vectors, to form a point cloud dataset. Each point in the dataset is a two-dimensional feature vector representing a certain state of the standard crack signal.

[0047] Construct a k-nearest neighbor graph (k-NNGraph) based on this point cloud. Choose a suitable value for k, for example, k=12. For each point in the point cloud... Using algorithms such as kd-trees, the 12 nearest neighbors to the Euclidean distance are found in the feature space. In graph G, for With each neighboring point Create an undirected edge between them, with the weight w(i,j) of the edge set to the Euclidean distance between them. To calculate the geodesic distance between any two points, i.e., the shortest distance along the surface of the manifold, run Dijkstra's algorithm or Floyd-Warshall's algorithm on the graph for all node pairs. The output is a distance matrix of a neural network. Where N is the total number of points in the point cloud, and the matrix elements are... (a,b) stores the data from point... arrive The shortest path length is the pre-calculated initial geodesic distance.

[0048] For the crack prototype set built offline, each prototype node is selected from existing nodes in the point cloud, so its geodesic distance to other nodes can be directly obtained from... The results were obtained from a query. During online computation, for a real-time feature vector V(t), first find the k nodes in the point cloud with the closest Euclidean distance to it, add V(t) as a temporary node to the graph, and connect the k nodes, with the edge weights also set to Euclidean distance. Then, use Dijkstra's algorithm to calculate the shortest path length from V(t) to each prototype node, obtaining the geodesic distance. The minimum value among these is taken as the initial geodesic distance.

[0049] S6, The initial geodesic distance is adjusted using the lift-off influence weighting function to obtain the minimum geodesic distance. .

[0050] The initial geodesic distance is adjusted using the lift-off influence weight function W(t) calculated in the previous step. Adjust the formula to This achieves the principle of high-safety detection with minimal missed detections. In actual detection, when the lift-off effect occurs—that is, when the probe momentarily deviates from the workpiece surface—the actual crack eddy current response signal attenuates sharply. This causes the extracted feature vector to deviate from the actual crack prototype in the manifold space, resulting in an abnormally increased initial geodesic distance, which easily leads to missed detections of actual cracks. As the lift-off increases, the calculated W(t) decreases accordingly, with its value between 0 and 1. Multiplying the abnormally increased initial distance due to lift-off by a smaller W(t) is equivalent to applying a sensitivity compensation mechanism for signal attenuation, forcibly reducing the distance proportionally, making it more likely to fall back into the preset alarm threshold range, thereby minimizing the possibility of missed detections of actual cracks caused by lift-off jitter. The minimum value in the adjusted distance set is the minimum geodesic distance. .

[0051] In one embodiment, adjusting the initial geodesic distance using the lift-off influence weighting function to obtain the minimum geodesic distance includes: In the feature space, find the sampling point on the manifold that has the closest Euclidean distance to the current real-time feature vector; Query the initial geodesic distance from the nearest sampling point to the preset sampling points of multiple crack prototype sets; Multiply the initial geodesic distance by the lift-off influence weight function to obtain the adjusted geodesic distance; The minimum value of the adjusted geodesic distance is selected as the minimum geodesic distance.

[0052] A set of crack prototype points needs to be predefined in the point cloud of the manifold. For example, select five feature points that have the strongest signal response to cracks of 0.2mm, 0.5mm, 1.0mm, and 1.5mm respectively as the prototype set. In real-time detection, when a new feature vector V(t) is generated, the vector is projected onto the manifold, i.e., onto N pre-stored manifold sampling points. In the process, the point with the smallest Euclidean distance to V(t) is quickly found using a kd-tree. .

[0053] turn up Then, using the pre-calculated geodetic distance matrix Query The initial geodesic distances to the five crack prototypes are obtained. A distance vector is then generated. .

[0054] Assume the retrieved distance vector is [3.2, 5.1, 8.9, 12.0, 15.4]. Calculate the lift-influence weight function value W(t) at the current moment based on the second component signal. Assuming a slight lift occurs, W(t) is 0.85. Apply multiplicative compensation to each initial geodesic distance using the weight function, resulting in the adjusted distance vector: [2.72, 4.34, 7.57, 10.20, 13.09]. Select the minimum value from the adjusted distance vector as the output for the current moment; that is, the minimum geodesic distance is 2.72.

[0055] In one preferred embodiment, to avoid numerous false alarms caused by the multiplicative compensation mechanism when severe lift-off occurs without cracks, the adjustment of the initial geodesic distance using the lift-off influence weighting function employs a segmented processing method, including: Preset a lift-off critical weight threshold For example, set it to 0.3; When calculated in real time If the condition is determined to be mild or moderate, compensation logic is executed: using the formula... Adjust the distance to reduce its range to prevent weak cracks from being missed. When calculated in real time At this point, it is determined to be a severe lift-off, meaning the extracted first component feature is completely distorted. The geodesic distance calculation is stopped, and a preset penalty number is directly output as... This triggers an alert for abnormal device status due to probe misalignment, requiring a rescan of the area. The segmented strategy ensures high detection sensitivity while effectively eliminating false alarms caused by serious operational errors.

[0056] S7. When the minimum geodetic distance is less than the preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

[0057] By testing a validation sample set containing cracked samples and defect-free samples with lift-off interference, receiver operating characteristic (ROC) curves were plotted, and the adjusted geodesic distance value corresponding to the maximum Youden exponent was selected as the optimal criterion threshold. The Youdens Index is a comprehensive indicator for evaluating the overall accuracy of diagnostic tests or classification models. During real-time detection, it is the minimum geodesic distance calculated at each time point. and Compare. If If a crack is detected at the current probe position, an alarm signal will be output or a mark will be made at the corresponding position on the C-scan imaging image; otherwise, it will be determined that there is no crack.

[0058] In one embodiment, the step of determining that a crack exists on the surface of the metal material when the minimum geodesic distance is less than a preset criterion threshold includes: By testing known cracked samples and known uncracked samples, two sets of minimum geodesic distance distributions were obtained; The receiver operating characteristic curve is used to analyze the two sets of minimum geodesic distance distributions. The geodesic distance value corresponding to the maximum Youden index is selected as the preset criterion threshold so that the sum of the detection sensitivity and specificity is maximized.

[0059] The determination of the preset criterion threshold is an offline calibration process. Two types of test samples are prepared: one is a positive sample set containing multiple known cracks, and the other is a negative sample set of intact surface samples confirmed to be crack-free. These samples should be independent of the samples used to construct the manifold. A comprehensive scan of both types of samples is performed, and the acquired signals undergo a complete processing flow to calculate the minimum geodesic distance for each sample at all time points during the scan. For each positive sample, record the global minimum value in the corresponding sequence; for each negative sample, also record... The global minimum of the sequence. Two sets of distance data were obtained: the distance distribution with cracks and the distance distribution without cracks.

[0060] Receiver operating characteristic ROC curve analysis was performed using the two sets of data above. All distance values ​​were sorted from low to high, and each value was used as a candidate threshold. For each ,like If a sample has a crack, it is considered to have one; otherwise, it is considered to have no crack. Based on this, the true positive rate sensitivity is calculated as Sensitivity = TP / (TP + FN), and the true negative rate specificity is calculated as TN / (TN + FP). Here, TP represents true positives (samples where a crack actually exists and the test result also classifies them as having a crack); FN represents false negatives (samples where a crack actually exists but the test result classifies them as having no crack); TN represents true negatives (samples where there is no crack and the test result classifies them as having no crack); and FP represents false positives (samples where there is no crack but the test result classifies them as having a crack). The Youden index J is calculated as Sensitivity + Specificity - 1. All candidate thresholds are iterated through to find the threshold that maximizes the Youden index J. ,like Figure 2 As shown, the ROC curve of the criterion model of this invention is superior to that of a random benchmark. The optimal threshold selected based on the Youden index achieves an optimal balance between sensitivity and specificity, ensuring the accuracy and stability of crack detection under complex working conditions. For example, analysis reveals that when the threshold... When the threshold is set to 4.52, the sensitivity is 0.98 and the specificity is 0.99, at which point J=0.97 is the maximum value. Therefore, 4.52 is set as the preset criterion threshold. In subsequent actual testing, as long as the value calculated in real time is less than... 4.52, output an alarm signal indicating the presence of cracks.

[0061] An ablation experiment was conducted using the lift-off effect compensation module to verify the rationality of the lift-off influence weighting function in suppressing the lift-off effect in eddy current detection.

[0062] The experimental conditions involved testing 7075 aluminum alloy specimens containing standard electrical discharge machining (EDM) cracks ranging from 0.2 mm to 1.5 mm in depth. A dual-frequency eddy current probe was used, with the low frequency set to 15 kHz and the high frequency to 300 kHz. The data acquisition system had a sampling frequency of 1.8 MHz. The experimental group employed the complete scheme, adjusting the geodesic distance using a lift-off influence weighting function. The control group removed this compensation module, using the unadjusted initial minimum geodesic distance as the criterion; the remaining processing procedures were identical to the experimental group. Different lift-off heights were simulated by placing insulating pads of 0 mm, 0.5 mm, 1.0 mm, and 1.5 mm thickness between the probe and the specimen. The average crack detection accuracy of both methods at different lift-off heights was recorded.

[0063] The crack detection accuracy of the two methods at four different lift-off heights is shown in Table 1 below.

[0064] Table 1 Lifting height control group accuracy Accuracy of the experimental group 0mm 99.1% 99.3% 0.5mm 92.5% 98.8% 1.0mm 78.3% 98.1% 1.5mm 61.7% 97.5% To visually demonstrate the impact of lift-off height on the detection performance of the two methods, the crack detection accuracy at different lift-off heights is compared to... Figure 3 As shown.

[0065] At a height of 0 mm (no lift), both the control and experimental groups showed extremely high detection accuracy with no difference. However, as the lift height increased, the detection accuracy of the control group dropped sharply, falling to 61.7% at a lift height of 1.5 mm. This indicates that the uncompensated features were severely shifted under the influence of lift, leading to a large number of missed detections and false positives. In contrast, the detection accuracy of the experimental group remained above 97.5% throughout the entire lift height variation range, demonstrating extremely strong stability.

[0066] The lift-off effect weighting function uses the envelope of the high-frequency component signal to represent the signal attenuation caused by the lift-off effect in real time. By dividing the initial geodesic distance by this weight, the characteristic distance increased by lift-off is increased and calibrated, making the measurement results highly stable to lift-off changes. Under a lift-off condition of 1.0 mm, the accuracy improved from 78.3% to 98.1%, an improvement of 19.8 percentage points. This result proves that the proposed compensation mechanism can overcome the lift-off interference in eddy current detection and enhance the reliability and accuracy of crack detection under complex working conditions.

[0067] Secondly, the present invention also proposes a surface crack eddy current testing system for metallic materials, comprising the following modules: The reconstruction module is used to scan the surface of a metal material using a dual-frequency excitation probe, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The module also acquires the composite eddy current response signal generated by the probe. Time-frequency analysis is performed on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal. The generation module is used to extract the instantaneous phase sequence and first to third-order time derivatives from the first component signal, construct a two-dimensional phase space trajectory with the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculate the curvature of the trajectory; combine the instantaneous phase and the curvature to form a real-time feature vector; extract parameters representing the lift-off effect in real time from the second component signal, and generate a lift-off effect weighting function based on the parameters; The determination module is used to calculate the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and the preset Riemann metric; adjust the initial geodesic distance using the lift-off influence weight function to obtain the minimum geodesic distance; when the minimum geodesic distance is less than the preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

[0068] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this application.

[0069] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A method for testing eddy current cracks on the surface of metallic materials, characterized in that, Includes the following steps: A dual-frequency excitation probe is used to scan the surface of a metallic material, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The probe also collects the composite eddy current response signal. Time-frequency analysis is performed on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal. The instantaneous phase sequence and first to third-order time derivatives are extracted from the first component signal. A two-dimensional phase space trajectory is constructed using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate. The curvature of the trajectory is calculated. The instantaneous phase and the curvature are combined to form a real-time feature vector. Extract parameters representing the lift-off effect in real time from the second component signal, and generate a lift-off effect weighting function based on the parameters; Based on a discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemann metric, the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold is calculated; the initial geodesic distance is adjusted using the lift-off influence weight function to obtain the minimum geodesic distance; when the minimum geodesic distance is less than a preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

2. The method according to claim 1, characterized in that, The step of performing time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal includes: The composite eddy current response signal was processed using short-time Fourier transform, with a Hanning window set as the window function, a window length of 512 sampling points, and an overlap length of 256 sampling points. In the generated time spectrum, the center frequency bands of the first excitation frequency and the second excitation frequency are located respectively, and the spectral data within the two frequency bands are extracted. The inverse short-time Fourier transform is performed on the spectral data of the two extracted frequency bands to obtain the first component signal and the second component signal.

3. The method according to claim 2, characterized in that, The step of extracting the instantaneous phase sequence and first to third-order time derivatives from the first component signal, constructing a two-dimensional phase space trajectory using the instantaneous phase as the first coordinate and the product of the first-order time derivative and a preset time constant as the second coordinate, and calculating the curvature of the trajectory includes: Apply the Hilbert transform to the first component signal to obtain the analytic signal, and calculate the instantaneous phase sequence of the analytic signal; The first, second, and third time derivatives of the instantaneous phase sequence are calculated iteratively using the central difference method. Based on the first-order time derivative, the second-order time derivative, the third-order time derivative, and the preset time constant, the curvature of the trajectory is calculated using the parametric curve curvature formula to obtain a curvature feature sequence.

4. The method according to claim 1, characterized in that, The step of extracting parameters representing the lift-off effect in real time from the second component signal and generating a lift-off effect weighting function based on the parameters includes: The lift-off influence weight function is generated by adding a regularization constant to a Gaussian function, and its expression is as follows: in The envelope amplitude is extracted in real time from the second component signal. As the reference envelope amplitude, is the attenuation control coefficient, and C is the preset regularization constant.

5. The method according to claim 1, characterized in that, The step of combining the instantaneous phase and the curvature to form a real-time feature vector includes: At each time point, the instantaneous phase at the current moment is taken as the first dimension feature, the phase trajectory curvature is taken as the second dimension feature, and the two scalar values ​​are combined into a two-dimensional column vector.

6. The method according to claim 1, characterized in that, The calculation of the initial geodesic distance from the real-time feature vectors to the preset crack prototype set on the manifold, based on a discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and a preset Riemannian metric, includes: The feature vector set of the standard crack sample is used as discrete sampling points on the manifold; Construct a k-nearest neighbor graph with each sampling point as a node, connecting each node to its k nearest neighbors, and use the Euclidean distance between nodes as the initial edge weights; The k-nearest neighbor graph structure is used as a discrete approximation of the Riemannian manifold, and the Riemannian metric is implicitly defined by the local Euclidean distance structure. Dijkstra's algorithm is used to calculate the shortest path length between all pairs of nodes in the graph, which is then used as the initial geodesic distance between them.

7. The method according to claim 1, characterized in that, The step of adjusting the initial geodesic distance using the lift-off influence weight function to obtain the minimum geodesic distance includes: In the feature space, find the sampling point on the manifold that has the closest Euclidean distance to the current real-time feature vector; Query the initial geodesic distance from the nearest sampling point to the preset sampling points of multiple crack prototype sets; Multiply the initial geodesic distance by the lift-off influence weight function to obtain the adjusted geodesic distance; The minimum value of the adjusted geodesic distance is selected as the minimum geodesic distance.

8. The method according to claim 1, characterized in that, The condition that a crack exists on the surface of the metal material when the minimum geodesic distance is less than a preset criterion threshold includes: By testing known cracked samples and known uncracked samples, two sets of minimum geodesic distance distributions were obtained; The receiver operating characteristic curve is used to analyze the two sets of minimum geodesic distance distributions. The geodesic distance value corresponding to the maximum Youden index is selected as the preset criterion threshold so that the sum of the detection sensitivity and specificity is maximized.

9. A surface crack eddy current testing system for metallic materials, characterized in that, Includes the following modules: The reconstruction module is used to scan the surface of a metal material using a dual-frequency excitation probe, wherein the first excitation frequency is lower than the second excitation frequency. The first excitation frequency is used for sensitive detection of crack depth characteristics, and the second excitation frequency is used for sensitive detection of lift-off effect. The module also acquires the composite eddy current response signal generated by the probe. Time-frequency analysis is performed on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal. The generation module is used to extract the instantaneous phase sequence and the first to third order time derivatives from the first component signal, construct a two-dimensional phase space trajectory with the instantaneous phase as the first coordinate and the product of the first order time derivative and the preset time constant as the second coordinate, and calculate the curvature of the trajectory; and combine the instantaneous phase and the curvature to form a real-time feature vector. Extract parameters representing the lift-off effect in real time from the second component signal, and generate a lift-off effect weighting function based on the parameters; The determination module is used to calculate the initial geodesic distance from the real-time feature vector to the preset crack prototype set on the manifold based on the discrete point cloud manifold pre-constructed from the feature vector set of standard crack samples and the preset Riemann metric; adjust the initial geodesic distance using the lift-off influence weight function to obtain the minimum geodesic distance; when the minimum geodesic distance is less than the preset criterion threshold, it is determined that there is a crack on the surface of the metal material.

10. The system according to claim 9, characterized in that, The step of performing time-frequency analysis on the composite eddy current response signal to separate and reconstruct the first component signal and the second component signal includes: The composite eddy current response signal was processed using short-time Fourier transform, with a Hanning window set as the window function, a window length of 512 sampling points, and an overlap length of 256 sampling points. In the generated time spectrum, the center frequency bands of the first excitation frequency and the second excitation frequency are located respectively, and the spectral data within the two frequency bands are extracted. The inverse short-time Fourier transform is performed on the spectral data of the two extracted frequency bands to obtain the first component signal and the second component signal.

Images