A method and system for detecting full-wall-thickness defects in a non-ferromagnetic sheet

Abstract

The application relates to the technical field of nondestructive testing, and particularly discloses a nonferromagnetic sheet full-wall-thickness defect detection method and system, which comprises the following steps: a rotating electromagnetic field generated by applying two-phase quadrature multi-frequency signal excitation is applied, spatial magnetic field response data is synchronously collected, original response data is formed, time-domain signals of each spatial point are converted into two-dimensional time-frequency spectra, and a time-frequency-space three-dimensional joint data body is organized; the data body is subjected to high-order decomposition to separate a plurality of independent signal components; a spatial statistical variation coefficient and a maximum frequency centroid offset of the signal components are calculated; target signal components representing defect disturbance are screened out; defect depth is calculated based on time characteristics of the target components; defect two-dimensional morphology is reconstructed in combination with spatial characteristics of the target components; and a defect three-dimensional model is generated by improved inverse distance weighted interpolation and surface smoothing by fusing the defect depth and the defect two-dimensional morphology.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology, specifically to a method and system for detecting defects in the full wall thickness of nonferromagnetic thin plates. Background Technology

[0002] With increasingly stringent requirements for the safe operation and maintenance of non-ferromagnetic thin-walled structures in fields such as nuclear power and aerospace, non-destructive testing (NDT) technology is developing towards integration, intelligence, and quantification. Against this backdrop, detection methods integrating multi-frequency rotating electromagnetic field excitation, intelligent sensor arrays, and advanced signal processing algorithms have become a research frontier. In particular, the application of intelligent sensors makes it possible to synchronously and at high resolution acquire multi-dimensional response information of spatial magnetic fields, laying a solid data foundation for subsequent high-sensitivity identification of defects and accurate reconstruction of three-dimensional morphology across the entire wall thickness of thin plates.

[0003] For non-ferromagnetic thin plates, existing electromagnetic nondestructive testing technologies generally adopt single-frequency and unidirectional excitation methods, which makes it impossible for their detection sensitivity to simultaneously take into account defects of different depths and orientations within the plate. Furthermore, it is difficult to decouple and quantify the three-dimensional morphological information of defects from complex electromagnetic response signals, thus failing to achieve high-sensitivity detection and accurate three-dimensional visualization assessment of multiple types of defects across the entire wall thickness of such components. Summary of the Invention

[0004] The purpose of this invention is to provide a method and system for detecting defects in the full wall thickness of non-ferromagnetic thin plates, so as to solve the problems mentioned above.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate includes the following steps:

[0007] S1: Apply multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquire the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation to obtain the raw response data containing multiple frequency components and spatial position information.

[0008] S2: Convert the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions, and organize the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space.

[0009] S3: Perform high-order decomposition on the three-dimensional joint data volume of time-frequency-space, separating it into multiple independent signal components;

[0010] S4: From multiple independent signal components, based on the differences between the characteristics of the signal components in the frequency dimension and the known intrinsic electromagnetic response modes of the material, as well as the uniformity of their distribution in the spatial dimension, identify and screen the target signal components corresponding to the abnormal structural disturbances inside the thin plate.

[0011] S5: Based on the distribution characteristics of the target signal components in the time dimension, the location of the abnormal structure in the thickness direction of the thin plate is determined, and combined with the distribution characteristics in the spatial dimension, the two-dimensional morphology of the abnormal structure in the thin plate plane is reconstructed, and then the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate is synthesized.

[0012] As a further aspect of the present invention: S1 specifically includes:

[0013] Two sinusoidal signals containing the same fundamental frequency and multiple preset harmonic frequency components are generated in parallel using direct digital frequency synthesis technology. One of the signals is then introduced into a fixed 1 / 4-period delay to construct two multi-frequency excitation signals with equal amplitude, the same frequency components, and orthogonal phase.

[0014] Two multi-frequency excitation signals are respectively input into a pair of excitation coils orthogonally arranged in space, thereby generating a multi-frequency composite rotating electromagnetic field excitation in the test area on the thin plate surface. Simultaneously, a three-dimensional magnetic sensor array is used to collect the spatial magnetic field distribution signal generated by the induction to obtain the initial time domain response sequence.

[0015] Automatic gain control is applied to the initial time-domain response sequence to dynamically adjust the signal amplitude of each sampling channel to a preset range, forming raw response data containing multiple frequency components and spatial location information.

[0016] As a further aspect of the present invention: S2 specifically includes:

[0017] For the time-domain signal corresponding to each spatial sampling point in the original response data, continuous wavelet transform is used for time-frequency decomposition. The transform result is then resampled and gridded in a two-dimensional plane composed of the time axis and the logarithmic frequency axis to generate a two-dimensional time spectrum with uniform high resolution.

[0018] Amplitude normalization is performed point by point along the frequency dimension of the two-dimensional time spectrum to eliminate the amplitude ratio deviation between different frequency components caused by the difference in electromagnetic excitation intensity, and the normalized time spectrum is obtained.

[0019] The normalized time-frequency spectra of all spatial sampling points are stacked according to the order of the two-dimensional spatial coordinates corresponding to the normalized time-frequency spectra, and a fourth-order tensor is formed by the time dimension, the logarithmic frequency dimension, the first spatial dimension and the second spatial dimension, thereby constructing a three-dimensional joint data volume of time-frequency-space.

[0020] As a further aspect of the present invention: S3 specifically includes:

[0021] For the three-dimensional joint data volume of time-frequency-space, the non-negative tensor decomposition algorithm is used to perform the decomposition operation. The number of signal components to be decomposed is preset, and the feature matrices of each signal component in time, frequency and two spatial dimensions are randomly initialized.

[0022] Based on the preset iterative update rules, with the optimization objective of minimizing the overall error between the reconstructed data volume and the original data volume, each feature matrix is ​​updated alternately, and additional constraints on the amplitude smoothing of adjacent positions are applied to the feature matrix of the spatial dimension during the update process.

[0023] When the change in the overall error between two consecutive iterations is less than a preset threshold, the decomposition is determined to be converged, the iteration is stopped, and the multiple independent signal components obtained from the decomposition and their corresponding feature matrices of each dimension are output.

[0024] As a further aspect of the present invention: the decomposition operation of the time-frequency-space three-dimensional joint data volume is performed using a non-negative tensor decomposition algorithm, specifically including:

[0025] For the three-dimensional joint data volume of time-frequency-space, calculate the signal energy entropy value of all spatial points in the time dimension and frequency dimension respectively, and mark the dimension interval where the average energy entropy exceeds the preset threshold as the effective region;

[0026] The number of signal components to be decomposed is dynamically determined based on the ratio of the volume of the effective region to the preset empirical volume contributed by a unit physical process.

[0027] Based on the dynamically determined number of signal components, a deterministic method based on the energy centroid projection of the effective region is adopted to generate the initial feature matrix of each signal component in time, frequency and two spatial dimensions, and complete the random initialization.

[0028] As a further aspect of the present invention: S4 specifically includes:

[0029] Calculate the statistical coefficient of variation of each signal component on its spatial dimension feature matrix, and the maximum centroid offset of the signal component on its frequency dimension feature vector relative to the reference intrinsic response spectrum, to form a set of feature parameters;

[0030] The system analyzes the set of characteristic parameters of multiple signal components and automatically classifies them into several categories based on the distance between the parameters.

[0031] Select categories with statistical variation coefficients lower than a preset uniformity threshold and maximum centroid offset higher than a preset disturbance threshold. Merge all signal components in each category and remove signal components from other categories to obtain the target signal component.

[0032] As a further aspect of the present invention: the analysis of the feature parameter set of multiple signal components, and the automatic division into several categories based on the distance between the parameters, specifically includes:

[0033] For the set of feature parameters, calculate the similarity distance between the feature parameters of every two signal components and construct a distance matrix;

[0034] A cohesive hierarchical clustering method is used on the distance matrix. Starting from each signal component forming its own class, the two closest classes are iteratively merged based on the similarity distance from small to large until the preset local association threshold is met.

[0035] Record all categories formed during the iterative merging process and their contained signal component members, as several categories automatically divided based on the distance between parameters.

[0036] As a further aspect of the present invention: S5 specifically includes:

[0037] Based on the feature matrix of the target signal component in the time dimension, the time coordinate corresponding to the main energy peak of the feature matrix is ​​extracted, and according to the known group velocity propagation model and dispersion correction relationship of electromagnetic waves in the thin plate material, the time coordinate is converted into the precise depth position of the abnormal structure in the thickness direction.

[0038] Based on the feature matrix of the target signal components in the spatial dimension, an adaptive threshold segmentation method is used to generate an initial binary topography map of the abnormal structure at the corresponding depth position. Morphological closing operation is then performed on the initial binary topography map to fill internal holes and connect adjacent regions, resulting in a continuous two-dimensional topography.

[0039] The precise depth position is used as the third-dimensional coordinate, which is registered and superimposed with the planar coordinates of the two-dimensional shape. A continuous surface model is generated through three-dimensional meshing interpolation to complete the synthesis of three-dimensional position and shape information.

[0040] As a further aspect of the present invention: the synthesis of three-dimensional position and shape information specifically includes:

[0041] Based on the edge features of the two-dimensional morphology and its distribution density in the plane, a set of spatial grid points adapted to the morphological features is generated by non-uniform sampling.

[0042] The precise depth position is used as the elevation value and assigned to the corresponding planar coordinate point to form an initial point cloud containing three-dimensional coordinates;

[0043] Inverse distance weighted interpolation based on the weighting factor related to the two-dimensional topography gradient of each point in the initial point cloud is used to calculate the elevation value on the non-uniform spatial grid point.

[0044] Minimum curvature smoothing is applied to the interpolation results at all spatial grid points to generate a continuous surface model for characterizing the external contour of the defect, thus completing the synthesis.

[0045] A full-thickness defect detection system for non-ferromagnetic thin plates, specifically comprising:

[0046] The signal acquisition and acquisition module applies a multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquires the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation, thereby obtaining raw response data containing multiple frequency components and spatial position information.

[0047] The data processing and construction module converts the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions. It then organizes the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space.

[0048] The signal decomposition module performs high-order decomposition on the three-dimensional joint data volume of time-frequency-space, separating it into multiple independent signal components.

[0049] The target screening module identifies and screens target signal components corresponding to abnormal structural disturbances inside the thin plate from multiple independent signal components based on the differences between the characteristics of the signal components in the frequency dimension and the known intrinsic electromagnetic response modes of the material, as well as the uniformity of their distribution in the spatial dimension.

[0050] The defect reconstruction and visualization module determines the location of the abnormal structure in the thickness direction of the thin plate based on the distribution characteristics of the target signal components in the time dimension, and reconstructs the two-dimensional morphology of the abnormal structure in the thin plate plane by combining the distribution characteristics in the spatial dimension, and then synthesizes the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate.

[0051] The beneficial effects of this invention are:

[0052] (1) This invention applies a multi-frequency composite rotating electromagnetic field to simultaneously excite induced eddy currents that rotate in the plane and are distributed in layers according to frequency along the depth direction within a thin plate, enabling a single scan to effectively cover defects at different depths and in any direction. More importantly, by constructing a three-dimensional joint data volume of time, frequency, and space, and employing dynamic component decomposition based on energy entropy and adaptive signal filtering based on statistical coefficient of variation and maximum centroid offset, it is possible to effectively separate and accurately identify interference signals such as the uniform response of the material background and edge effects from disturbance signals caused by abnormal structures such as internal cracks and corrosion in multiple dimensions of time, frequency, and space. This series of signal processing techniques makes this method not only sensitive to surface defects, but also effectively detects tiny defects buried in the middle of the plate or even close to the back side, improving the detection depth range and defect detection capability, and overcoming the bottleneck of the sensitivity of conventional methods decreasing sharply with depth.

[0053] (2) This invention directly extracts time-dimensional features from the selected target signal components and performs precise time-depth conversion by combining the known group velocity and dispersion characteristics of the material, thus realizing the quantitative calculation of the defect depth location. Simultaneously, based on the spatial feature matrix of the target signal, a clear two-dimensional contour of the defect on the corresponding depth section is reconstructed through adaptive threshold segmentation and morphological processing. Finally, by introducing an improved inverse distance weighted interpolation algorithm with morphological gradient weights, the discrete depth information and two-dimensional morphology are fused with high-fidelity three-dimensional data and optimized for surface analysis, automatically generating a continuous and smooth three-dimensional surface model of the defect. This process minimizes the reliance on manual interpretation and can output key quantitative parameters such as the length, width, depth, and overall three-dimensional morphology of the defect, providing a direct and accurate data foundation for subsequent structural integrity assessment, remaining life prediction, and maintenance plan formulation, thus promoting the development of non-destructive testing from qualitative judgment to quantitative and visual analysis. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a flowchart of the method of the present invention;

[0056] Figure 2 This is a system block diagram of the present invention. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0058] Please see Figure 1 As shown, this invention provides a method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate, comprising the following steps:

[0059] S1: Apply multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquire the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation to obtain the raw response data containing multiple frequency components and spatial position information.

[0060] S2: Convert the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions, and organize the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space.

[0061] S3: Perform high-order decomposition on the three-dimensional joint data volume of time-frequency-space, separating it into multiple independent signal components;

[0062] S4: From multiple independent signal components, based on the differences between the characteristics of the signal components in the frequency dimension and the known intrinsic electromagnetic response modes of the material, as well as the uniformity of their distribution in the spatial dimension, identify and screen the target signal components corresponding to the abnormal structural disturbances inside the thin plate.

[0063] S5: Based on the distribution characteristics of the target signal components in the time dimension, the location of the abnormal structure in the thickness direction of the thin plate is determined, and combined with the distribution characteristics in the spatial dimension, the two-dimensional morphology of the abnormal structure in the thin plate plane is reconstructed, and then the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate is synthesized.

[0064] In S1, the excitation signal is first generated using direct digital frequency synthesis (DDS). This technique generates two digital sequences in parallel by looking up a pre-stored table of sine wave functions. The fundamental frequency and all harmonic frequency components contained in these two sequences are identical. Then, a fixed offset corresponding to 1 / 4 of the fundamental period is applied to the read address of one of the digital sequences, thus causing a fixed 1 / 4 period delay in the analog signal relative to the other after digital-to-analog conversion. The correctness of the phase orthogonality is verified by measuring whether the phase difference between the two output signals at the same frequency point is 90 degrees. Thus, two multi-frequency excitation signals with equal amplitude, identical frequency components, and strictly orthogonal phases are constructed.

[0065] Secondly, the two generated multi-frequency excitation signals are respectively connected to a pair of excitation coils arranged at a 90-degree angle in physical space. When current flows through the coils, the two magnetic field vectors are orthogonally superimposed in the test area above the thin plate surface, thereby coupling to generate a composite electromagnetic excitation field whose field strength vector direction rotates periodically with time and contains multiple frequency components. At the same time, a three-dimensional sensor array composed of multiple magnetically sensitive units arranged in a matrix is ​​placed in the excitation field to synchronously measure the three orthogonal spatial components of the induced magnetic field, obtaining a voltage sequence that varies with time, i.e., the initial time-domain response sequence.

[0066] Next, the acquired initial time-domain response sequence undergoes automatic gain control preprocessing. This processing is performed on each independent signal acquisition channel: first, the effective value of the signal amplitude within a time window of that channel is calculated, and this effective value is compared with a preset ideal target amplitude value; based on the comparison result, a control voltage is generated to adjust the gain factor of the preamplifier circuit of that channel. Through this feedback adjustment, the amplitude of the output signal of each channel is dynamically adjusted and stabilized near the preset target amplitude range, thereby overcoming the problem of inconsistent signal amplitude caused by small changes in lift-off height or differences in sensor sensitivity.

[0067] Finally, the data from all channels and all sampling times, after the aforementioned gain adjustment, are arranged and integrated according to their corresponding sensor spatial coordinates to form the raw response data. This data fully encompasses the magnetic field response induced by multi-frequency rotating electromagnetic field excitation and acquired via a spatial array. Its information dimensions cover multiple excitation frequency components and complete two-dimensional spatial position information, providing a foundation for subsequent in-depth analysis.

[0068] In S2, firstly, for each spatial sampling point in the original response data, its corresponding time-domain signal is decomposed into a two-dimensional time-frequency spectrum to generate a continuous wavelet transform. Specifically, a mother wavelet function with good localization properties is selected, translated along the time axis, and simultaneously scaled (inversely proportional to frequency) to generate a series of wavelet basis functions. The time-domain signal is then convolved with each scaled and translated wavelet basis function, and the integral result is the wavelet coefficient of the signal at the corresponding time point and scale (frequency). To obtain a logarithmically uniform distribution along the frequency axis that better conforms to electromagnetic signal analysis practices, the scale parameter is converted to a base-2 logarithmic frequency coordinate. Subsequently, on a two-dimensional regular grid spanned by uniformly discrete time point coordinates and logarithmic frequency point coordinates, bilinear interpolation is used to map and fill the calculated irregularly distributed wavelet coefficient values ​​to each grid point, thereby generating a two-dimensional time-frequency spectrum with uniform high resolution in both the time and logarithmic frequency dimensions.

[0069] Secondly, amplitude normalization is performed on each of the obtained two-dimensional time-frequency spectra to eliminate the proportional influence of energy differences in different frequency components of the excitation source on the signal amplitude. The method is as follows: for a single time-frequency spectra, along its logarithmic frequency dimension, find the maximum amplitude value among all time points corresponding to that dimension; then, divide the original amplitude value at each time-frequency point of the time-frequency spectra by the previously found maximum amplitude value corresponding to that frequency point. This operation ensures that for each time-frequency spectra, the maximum amplitude value is normalized to 1 in each of its own logarithmic frequency slices, while other values ​​are scaled proportionally, thus obtaining a normalized time-frequency spectra with uniform amplitude proportions.

[0070] Finally, the normalized time-frequency spectra of all spatial sampling points are organized according to their respective two-dimensional planar spatial coordinates. Each time-frequency spectrum is treated as an element and arranged and stacked according to the first dimension (e.g., the X-axis) and the second dimension (e.g., the Y-axis) of its spatial coordinates. Thus, the time dimension and logarithmic frequency dimension of a single time-frequency spectrum, together with the two spatial dimensions formed by the stacking of all time-frequency spectra, constitute a four-dimensional data structure, namely a fourth-order tensor. This tensor is the aforementioned time-frequency-space three-dimensional joint data volume, whose four index axes correspond to time, logarithmic frequency, the first spatial dimension, and the second spatial dimension, respectively, fully encapsulating the characteristic information of the signal in the entire spatiotemporal frequency domain.

[0071] In S3, the first step is to prepare for the high-order decomposition of the time-frequency-space three-dimensional joint data volume. The core of this preparation is determining the number of signal components to be decomposed and initializing relevant parameters. Specifically, the following steps are performed: For the three-dimensional joint data volume, the signal energy distribution is analyzed along its time and frequency dimensions. For the time dimension, at each discrete time point, the sum of squares of the signal amplitudes at all spatial locations is calculated to obtain the total energy at that time point. For the frequency dimension, at each discrete logarithmic frequency point, the sum of squares of the signal amplitudes at all spatial locations is similarly calculated to obtain the total energy at that frequency point. Based on the energy values ​​of each point in each dimension, its probability distribution is calculated, i.e., the proportion of the energy value of each point to the total energy value of that dimension. Subsequently, according to the definition of entropy in information theory, the energy probability of each point is multiplied by the logarithm of that probability (base 2), and the sum of the calculation results for all points in that dimension is taken as a negative value, thus obtaining the signal energy entropy value of all spatial points in that dimension. The average energy entropy is calculated separately along both the time and frequency dimensions. This average is compared with a threshold obtained beforehand through statistical analysis of typical defect-free sample data. Dimensional intervals where the average energy entropy exceeds this threshold (i.e., segments with more complex and uncertain energy distribution) are marked as "effective regions." Next, the number of discrete points covered by this effective region in the time-frequency plane is calculated, and this is considered the "volume of the effective region." This volume is divided by an "empirical volume," determined through numerous prior experiments, representing the average number of discrete points occupied by a single physical process (such as the uniform response of a material matrix) in the time-frequency domain. The quotient is rounded up, and the result is dynamically determined as the "number of signal components" to be found in this decomposition operation. Finally, based on this dynamically determined number, initialization is performed: for each signal component to be found, a feature matrix is ​​established along its time, frequency, and two spatial dimensions. The initial values ​​are not completely randomized but generated using a deterministic method. Specifically, within the "effective region" marked in step two, the energy centroid coordinates of the entire data volume on the time-frequency two-dimensional plane are calculated. Then, the energy centroid is projected onto the dimensions corresponding to each feature matrix to be initialized. Matrix elements near the projection point are assigned higher initial values ​​(e.g., random numbers between 0.8 and 1.0), while elements far from the projection point are assigned lower initial values ​​(e.g., random numbers between 0 and 0.2), thereby completing the initialization of each feature matrix.

[0072] Secondly, decomposition calculations are performed according to preset iterative update rules, with the goal of making the reconstructed data volume as close as possible to the original data volume. The reconstructed data volume is synthesized from the feature matrices of each signal component in the current iteration step, according to the multiplication and summation rules specified in the tensor decomposition model. The overall error is defined as the sum of squares of the differences between the values ​​of all elements in the original data volume and the values ​​of corresponding elements in the reconstructed data volume. In each iteration, the idea of ​​alternating least squares is used to update each feature matrix sequentially: when updating a certain feature matrix, the feature matrices of all other dimensions are temporarily fixed, and the updated value of the current feature matrix is ​​calculated by solving a least squares problem with the goal of minimizing the overall error. In this process, additional constraints are imposed on the update of the feature matrices representing the distribution of the two spatial dimensions: for each element in the spatial feature matrix, the smoothing contribution of the current value of the elements at its neighboring positions (such as the upper, lower, left, and right adjacent points) is incorporated into its updated value. Specifically, this is achieved by multiplying the average value of the neighboring elements by a smoothing coefficient less than 1, and then performing a weighted average with the originally calculated updated value, thereby constraining the smoothness of the spatial distribution and conforming to the continuity law of the physical field distribution. One iteration is completed after all feature matrices have been updated sequentially.

[0073] Finally, determine if the decomposition process has converged. After each complete iteration, calculate the overall error between the current reconstructed data volume and the original data volume. Record the overall error value of this iteration and the overall error value of the previous iteration, and calculate the absolute value of the difference between the two. Compare this absolute value with a preset minimum positive threshold (e.g., 0.001). If the absolute value is less than this threshold, the decomposition process is considered to have converged, and the iteration update stops. Otherwise, continue to the next round of iteration update. When the iteration stops, output the final feature matrices corresponding to each signal component in the time dimension, frequency dimension, and two spatial dimensions. These feature matrices collectively characterize the separated, independent signal components, each reflecting the characteristics of a potential, relatively independent physical process in the time-space-frequency domain.

[0074] In S4, firstly, to quantitatively evaluate the characteristics of each signal component, a set of characteristic parameters that characterize its spatial distribution uniformity and frequency characteristic deviation needs to be calculated. For each signal component, the calculation of its characteristic parameters is divided into two parts. The first part is to calculate its spatial distribution uniformity index, namely the statistical coefficient of variation. This coefficient is calculated as follows: extract the feature matrix of the signal component in two spatial dimensions, take the values ​​of all elements in the matrix as the dataset, and calculate the arithmetic mean of the dataset; then, calculate the square of the difference between each element value and the arithmetic mean, sum all the squared differences, divide by the number of data points minus one, and take the square root of the result to obtain the standard deviation; finally, divide the standard deviation by the arithmetic mean, and the resulting ratio is the statistical coefficient of variation of the signal component. The smaller this value, the more uniform the spatial distribution. The second part is to calculate its frequency characteristic deviation index relative to the intrinsic response of the material, namely the maximum centroid shift. The calculation process is as follows: first, obtain the standard response of a defect-free sample of the same material under the same testing conditions, and extract its frequency dimension characteristics as a reference intrinsic response spectrum. For each signal component to be evaluated, its frequency-dimensional eigenvector is taken. Each element in this vector is considered a "mass," and its corresponding logarithmic frequency coordinate is considered a "position." The centroid position of this vector is calculated by multiplying each position coordinate by its corresponding element value, summing all products, and then dividing by the sum of all element values. The same method is used to calculate the centroid position of the reference intrinsic response spectrum. The eigenvector value of the signal component to be evaluated at each logarithmic frequency point is compared point-by-point with the value of the reference intrinsic response spectrum at the corresponding point. The frequency point with the largest absolute difference between the two is found. The absolute value of the difference between the eigenvector value of the signal component to be evaluated and the reference spectrum value at that point is calculated, and this absolute value is divided by the value of the reference spectrum at that point to obtain a normalized offset. This process is repeated for all frequency points, and the maximum normalized offset is the maximum centroid offset. The larger this value, the more significant the difference between the frequency characteristics and the material's intrinsic mode. Each signal component is ultimately described by a pair of values: the statistical coefficient of variation and the maximum centroid offset.

[0075] Secondly, the set of feature parameter pairs of all signal components is analyzed, and they are automatically categorized based on the similarity between the parameters. The specific analysis process is as follows: First, a distance matrix reflecting the differences between all signal components is constructed. For each pair of signal components in the feature parameter set, the similarity distance between them is calculated. This distance calculation is not a simple straight-line distance, but rather uses a non-Euclidean metric combined with a bandwidth adaptive kernel function. Specifically, for each pair of feature parameters of two signal components, the absolute value of the difference between their statistical coefficients of variation and the absolute value of the difference between their maximum centroid offsets are calculated. These two absolute differences are divided by an adaptive bandwidth parameter, which is determined by the standard deviation of all signal components in the current dataset along this parameter dimension. The result of the division is substituted into a preset kernel function (e.g., a Gaussian function) to calculate the similarity score in both dimensions. These two scores are multiplied to obtain a joint similarity. Finally, the joint similarity is subtracted from 1 to obtain the similarity distance between the two signal components. This calculation is performed pairwise for all signal components, forming a symmetric distance matrix. The second step involves automatically grouping data using agglomerative hierarchical clustering based on the distance matrix. At the start of clustering, each signal component forms its own independent category. In each iteration, the pairwise distance between all current categories is calculated, employing the "minimum distance method," where the minimum similarity distance between any two signal components in two categories is used as the distance between those two categories. The pair of categories with the smallest distance is identified and merged into a new category. This merging process is repeated to form a tree-like clustering structure. The merging termination condition is controlled by a preset local association threshold, which is a decimal between 0 and 1, such as 0.2. The merging iteration stops when the minimum distance between the two categories to be merged exceeds this threshold. The third step involves recording and backtracking the entire iterative merging process. Starting from the initial state where each signal component is a separate category, until the stopping condition is met, the intermediate category composition formed after each merge is recorded. These categories and their members represent the categories automatically partitioned based on the distances between parameters.

[0076] Finally, target signal components are identified and extracted from all categories according to preset screening criteria. The screening criteria are based on two preset thresholds: a uniformity threshold and a perturbation threshold. The uniformity threshold is determined by statistically analyzing the statistical coefficient of variation (COP) of signal components corresponding to a large number of defect-free regions; for example, twice the average of these COPs is used as the threshold. The perturbation threshold is determined by analyzing the maximum centroid offset of signal components corresponding to weak artificial defect samples; for example, half the average of these offsets is used as the threshold. All categories obtained in step two are iterated over. For each category, the average COP of all its signal components and the average maximum centroid offset are calculated. Categories with an average COP lower than the uniformity threshold and an average maximum centroid offset higher than the perturbation threshold are selected. These selected categories are considered to exhibit relatively concentrated spatial distribution (a non-uniform background field) and frequency characteristics that significantly deviate from the intrinsic mode of the material, consistent with the characteristics of internal anomalous structural perturbations. All signal components contained in the selected categories are extracted and merged into a single entity, which is the target signal component. The signal components contained in the remaining unselected categories are considered to correspond to the uniform response of the material matrix, edge effects, or other non-target interference processes and are therefore discarded.

[0077] In S5, firstly, based on the feature matrix of the target signal components in the time dimension, the depth position of the defect in the thickness direction is determined. The element with the largest amplitude in this feature matrix is ​​extracted; the discrete-time index corresponding to this element is the time coordinate of the main energy peak, denoted as […]. To convert this time coordinate to physical depth, the propagation characteristics of electromagnetic waves in a specific thin-film material need to be considered. Definition The frequency at the excitation center is calculated based on the known electromagnetic parameters of the material. The group velocity at that location. Due to the dispersion effect of the broadband excitation signal, correction is required. A correction factor related to the material's dispersion characteristics is introduced. This factor is determined by comparing the theoretical transit time of a standard test block of known thickness with the difference between the actual measured peak time, specifically the average of the ratios of the two. Finally, the precise depth position of the defect's upper surface or main reflective interface from the test surface is determined. It is calculated using the following formula: This step uniquely maps the time information of the signal to the depth information of the defect.

[0078] Secondly, based on the feature matrices of the target signal components in two spatial dimensions, the two-dimensional projection shape of the defect in the depth plane is reconstructed. The two spatial feature matrices are then synthesized to obtain a comprehensive spatial energy distribution matrix. An adaptive threshold segmentation method is used to binarize the matrix: first, the matrix is ​​calculated... The average value of all elements in and standard deviation Set the segmentation threshold as follows: , where the coefficient Determined based on prior knowledge of the contrast between typical defects and the background, for example, a value of 2. The matrix... The value of each element and the threshold The image is compared, and values ​​greater than a threshold are set to 1 (suspected defect area), otherwise set to 0 (background area), thus generating an initial binary topography image. This initial image may contain noise points and discontinuous pores. Subsequently, morphological closing operations are performed on this binary image to optimize the topography: first, a dilation operation is performed using a square structuring element with a side length of 3 pixels, traversing all pixels with a value of 1 in the image and setting the values ​​of all pixels in its neighborhood to 1; then, an erosion operation is performed using the same structuring element, traversing the image, and only keeping a pixel with a value of 1 in the result if all pixels in its neighborhood have a value of 1, otherwise setting it to 0. This operation can effectively fill the small voids inside the defect area and connect adjacent defect areas broken by noise, finally obtaining a continuous and complete two-dimensional binary topography image of the defect. .

[0079] Finally, the precise depth information obtained above is fused with the two-dimensional topography information to synthesize a three-dimensional model of the defect. The specific process consists of four steps: First, based on the two-dimensional topography image... Edge features and region density are used to non-uniformly generate spatial mesh points for 3D reconstruction. First, identify... For all pixel regions with a value of 1, extract their contours. Near the contour lines and within the regions, sample planar coordinates at a higher density (e.g., one point per pixel); in background regions far from the contours, sample at a lower density (e.g., one point per four pixels), forming a set of planar grid point coordinates adapted to the shape features. The second step is to use the precise depth position calculated in the first step. As an elevation value, it is assigned to the topographic map. The actual planar coordinates corresponding to each pixel with a value of 1 are used to form a series of three-dimensional spatial points. The initial point cloud is constructed. The third step is to construct a non-uniform grid of points. An improved inverse distance-weighted interpolation method is used to estimate the elevation value. For any grid point to be determined... Its elevation value The calculation considers not only the distance between the grid point and other points in the initial point cloud, but also the influence of the topography gradient. First, the topography map at that grid point is calculated. Gradient magnitude after distance transformation .point elevation Calculated by the following formula: ;

[0080] in, This represents the number of points in the initial point cloud. Indicates the first Points Indicates the first Points. Distance weight: ; It is a point With point cloud points The planar Euclidean distance between them It is a small constant to prevent division by zero (e.g., 1e-6). Gradient weights: ; It is a point cloud. gradient magnitude at that point This is a parameter that controls the sensitivity of gradient weights. This formula assigns higher weights to point cloud points that are close to the point being calculated and have similar topographic gradient features, making the interpolation result more closely match the edge features of the defect. The fourth step involves interpolating the elevation values ​​for all grid points. Minimum curvature smoothing is performed. An iterative algorithm is used to adjust the elevation values ​​of each point, minimizing the sum of squared curvatures of the discrete surface formed by these points while maintaining a good fit to the original interpolation data, thus generating a smooth and continuous 3D surface model of the defect. At this point, the 3D spatial location and geometric morphology information of the full-thickness defect inside the non-ferromagnetic thin plate are synthesized.

[0081] Please see Figure 2 As shown, a full-thickness defect detection system for non-ferromagnetic thin plates specifically includes:

[0082] The signal acquisition and acquisition module applies a multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquires the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation, thereby obtaining raw response data containing multiple frequency components and spatial position information.

[0083] The data processing and construction module converts the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions. It then organizes the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space.

[0084] The signal decomposition module performs high-order decomposition on the three-dimensional joint data volume of time-frequency-space, separating it into multiple independent signal components.

[0085] The target screening module identifies and screens target signal components corresponding to abnormal structural disturbances inside the thin plate from multiple independent signal components based on the differences between the characteristics of the signal components in the frequency dimension and the known intrinsic electromagnetic response modes of the material, as well as the uniformity of their distribution in the spatial dimension.

[0086] The defect reconstruction and visualization module determines the location of the abnormal structure in the thickness direction of the thin plate based on the distribution characteristics of the target signal components in the time dimension, and reconstructs the two-dimensional morphology of the abnormal structure in the thin plate plane by combining the distribution characteristics in the spatial dimension, and then synthesizes the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate.

[0087] The working principle of this invention is as follows: First, two multi-frequency excitation signals with orthogonal phases and consistent spectra are generated using direct digital frequency synthesis technology. These signals drive orthogonally arranged excitation coils to generate a multi-frequency composite rotating electromagnetic field excitation on the surface of a thin plate. Simultaneously, a three-dimensional magnetic sensor array is used to collect spatial magnetic field distribution signals, which are then processed by automatic gain control to form raw response data. Second, continuous wavelet transform is performed on the time-domain signal at each spatial sampling point to generate a high-resolution two-dimensional time spectrum. After amplitude normalization, all time spectra are stacked according to spatial coordinates to construct a three-dimensional joint data volume of time, frequency, and space. Then, based on the energy entropy of this data volume in the time and frequency dimensions, the number of signal components is dynamically determined, and non-negative tensor decomposition with spatial smoothness constraints is employed. The algorithm separates the data volume into multiple independent signal components. Then, it calculates the spatial statistical variation coefficient and the maximum centroid offset relative to the intrinsic response spectrum of the material for each component. It classifies the components using agglomerative hierarchical clustering based on adaptive distance metric and selects categories with concentrated spatial distribution and frequency characteristics that significantly deviate from the intrinsic mode. The components are then merged to obtain the target signal. Finally, the time coordinates of the main peak are extracted from the time features of the target signal. The defect depth is calculated by combining the material group velocity and dispersion correction factor. At the same time, the two-dimensional morphology of the defect is reconstructed based on its spatial feature matrix through adaptive threshold segmentation and morphological closing operation. An improved inverse distance weighted interpolation with fused morphology gradient weights and minimum curvature smoothing are used to synthesize the three-dimensional position and continuous surface morphology of the defect.

[0088] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate, characterized in that, Includes the following steps: S1: Apply multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquire the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation to obtain the raw response data containing multiple frequency components and spatial position information. S2: Convert the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions, and organize the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space. S3: Perform nonnegative tensor decomposition on the time-frequency-space three-dimensional joint data volume, separating it into multiple independent signal components, specifically including: For the three-dimensional joint data volume of time-frequency-space, the non-negative tensor decomposition algorithm is used to perform the decomposition operation. The number of signal components to be decomposed is preset, and the feature matrices of each signal component in time, frequency and two spatial dimensions are randomly initialized. Based on the preset iterative update rules, with the optimization objective of minimizing the overall error between the reconstructed data volume and the original data volume, each feature matrix is ​​updated alternately, and additional constraints on the amplitude smoothing of adjacent positions are applied to the feature matrix of the spatial dimension during the update process. When the change in the overall error between two consecutive iterations is less than a preset threshold, the decomposition is determined to be converged, the iteration is stopped, and the multiple independent signal components obtained by the decomposition and their corresponding feature matrices of each dimension are output. S4: From multiple independent signal components, based on the differences between the characteristics of the signal components in the frequency dimension and the reference intrinsic response spectrum obtained in advance by detecting defect-free areas, as well as the uniformity of their distribution in the spatial dimension, the target signal components corresponding to the abnormal structural disturbance inside the thin plate are identified and screened. The specific calculation process of the difference is as follows: the standard response of the same material sample without defects is obtained in advance under the same detection conditions, and its frequency dimension characteristics are extracted as the reference intrinsic response spectrum. For the signal component to be evaluated, take its frequency dimension feature vector, regard the value of each element in the frequency dimension feature vector as mass, and regard its corresponding logarithmic frequency coordinate as position. Calculate the centroid position of the vector, that is, multiply each position coordinate by its corresponding element value, sum all the products and then divide by the sum of all element values. Calculate the centroid location of the reference intrinsic response spectrum; The eigenvector value of the signal component to be evaluated at each logarithmic frequency point is compared with the value of the reference intrinsic response spectrum at the corresponding point. The frequency point with the largest absolute difference between the two is found. The absolute value of the difference between the eigenvector value of the signal component to be evaluated and the reference spectrum value at the frequency point is calculated. This absolute value is then divided by the value of the reference spectrum at the frequency point to obtain the normalized offset. By iterating through all frequency points, the maximum normalized offset is the maximum centroid offset. The larger the maximum centroid offset, the more significant the difference between the frequency characteristics and the intrinsic mode of the material. S5: Based on the distribution characteristics of the target signal components in the time dimension, the location of the abnormal structure in the thickness direction of the thin plate is determined, and combined with the distribution characteristics in the spatial dimension, the two-dimensional morphology of the abnormal structure in the thin plate plane is reconstructed, and then the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate is synthesized.

2. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 1, characterized in that, S1 specifically includes: Two sinusoidal signals containing the same fundamental frequency and multiple preset harmonic frequency components are generated in parallel using direct digital frequency synthesis technology. One of the signals is then introduced into a fixed 1 / 4-period delay to construct two multi-frequency excitation signals with equal amplitude, the same frequency components, and orthogonal phase. Two multi-frequency excitation signals are respectively input into a pair of excitation coils orthogonally arranged in space, thereby generating a multi-frequency composite rotating electromagnetic field excitation in the test area on the thin plate surface. Simultaneously, a three-dimensional magnetic sensor array is used to collect the spatial magnetic field distribution signal generated by the induction to obtain the initial time domain response sequence. Automatic gain control is applied to the initial time-domain response sequence to dynamically adjust the signal amplitude of each sampling channel to a preset range, forming raw response data containing multiple frequency components and spatial location information.

3. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 1, characterized in that, S2 specifically includes: For the time-domain signal corresponding to each spatial sampling point in the original response data, continuous wavelet transform is used for time-frequency decomposition. The transform result is then resampled and gridded in a two-dimensional plane composed of the time axis and the logarithmic frequency axis to generate a two-dimensional time spectrum with uniform high resolution. Amplitude normalization is performed point by point along the frequency dimension of the two-dimensional time spectrum to eliminate the amplitude ratio deviation between different frequency components caused by the difference in electromagnetic excitation intensity, and the normalized time spectrum is obtained. The normalized time-frequency spectra of all spatial sampling points are stacked according to the order of the two-dimensional spatial coordinates corresponding to the normalized time-frequency spectra, and a fourth-order tensor is formed by the time dimension, the logarithmic frequency dimension, the first spatial dimension and the second spatial dimension, thereby constructing a three-dimensional joint data volume of time-frequency-space.

4. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 1, characterized in that, The decomposition operation of the three-dimensional joint data volume of time-frequency-space is performed using a non-negative tensor decomposition algorithm, specifically including: For the three-dimensional joint data volume of time-frequency-space, calculate the signal energy entropy value of all spatial points in the time dimension and frequency dimension respectively, and mark the dimension interval where the average energy entropy exceeds the preset threshold as the effective region; The number of signal components to be decomposed is dynamically determined based on the ratio of the volume of the effective region to the preset empirical volume contributed by a unit physical process. Based on the dynamically determined number of signal components, a deterministic method based on the energy centroid projection of the effective region is adopted to generate the initial feature matrix of each signal component in time, frequency and two spatial dimensions, and complete the random initialization.

5. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 1, characterized in that, S4 specifically includes: Calculate the statistical coefficient of variation of each signal component on its spatial dimension feature matrix, and the maximum centroid offset of the signal component on its frequency dimension feature vector relative to the reference intrinsic response spectrum, to form a set of feature parameters; The system analyzes the set of characteristic parameters of multiple signal components and automatically classifies them into several categories based on the distance between the parameters. Select categories with statistical variation coefficients lower than a preset uniformity threshold and maximum centroid offset higher than a preset disturbance threshold. Merge all signal components in each category and remove signal components from other categories to obtain the target signal component.

6. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 5, characterized in that, The analysis of the feature parameter set of multiple signal components, automatically classifying them into several categories based on the distance between the parameters, specifically includes: For the set of feature parameters, calculate the similarity distance between the feature parameters of every two signal components and construct a distance matrix; A cohesive hierarchical clustering method is used on the distance matrix. Starting from each signal component forming its own class, the two closest classes are iteratively merged based on the similarity distance from small to large until the preset local association threshold is met. Record all categories formed during the iterative merging process and their contained signal component members, as several categories automatically divided based on the distance between parameters.

7. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 1, characterized in that, S5 specifically includes: Based on the feature matrix of the target signal component in the time dimension, the time coordinate corresponding to the main energy peak of the feature matrix is ​​extracted, and according to the known group velocity propagation model and dispersion correction relationship of electromagnetic waves in the thin plate material, the time coordinate is converted into the precise depth position of the abnormal structure in the thickness direction. Based on the feature matrix of the target signal components in the spatial dimension, an adaptive threshold segmentation method is used to generate an initial binary topography map of the abnormal structure at the corresponding depth position. Morphological closing operation is then performed on the initial binary topography map to fill internal holes and connect adjacent regions, resulting in a continuous two-dimensional topography. The precise depth position is used as the third-dimensional coordinate, which is registered and superimposed with the planar coordinates of the two-dimensional shape. A continuous surface model is generated through three-dimensional meshing interpolation to complete the synthesis of three-dimensional position and shape information.

8. The method for detecting defects across the entire wall thickness of a non-ferromagnetic thin plate according to claim 7, characterized in that, The synthesis of the three-dimensional position and shape information specifically includes: Based on the edge features of the two-dimensional morphology and its distribution density in the plane, a set of spatial grid points adapted to the morphological features is generated by non-uniform sampling. The precise depth position is used as the elevation value and assigned to the corresponding planar coordinate point to form an initial point cloud containing three-dimensional coordinates; Inverse distance weighted interpolation based on the weighting factor related to the two-dimensional topography gradient of each point in the initial point cloud is used to calculate the elevation value on the non-uniform spatial grid point. Minimum curvature smoothing is applied to the interpolation results at all spatial grid points to generate a continuous surface model for characterizing the external contour of the defect, thus completing the synthesis.

9. A defect detection system for non-ferromagnetic thin plates with full wall thickness, characterized in that, The method for detecting defects in the full wall thickness of a non-ferromagnetic thin plate according to any one of claims 1-8 specifically includes: The signal acquisition and acquisition module applies a multi-frequency composite rotating electromagnetic field excitation to the test area of ​​the non-ferromagnetic thin plate, and simultaneously acquires the spatial magnetic field distribution signal induced in the thin plate by the rotating electromagnetic field excitation, thereby obtaining raw response data containing multiple frequency components and spatial position information. The data processing and construction module converts the time-domain signal corresponding to each spatial sampling point in the original response data into a two-dimensional time spectrum that is simultaneously expanded in the time and frequency dimensions. It then organizes the two-dimensional time spectrum of all spatial sampling points according to spatial coordinates to construct a three-dimensional joint data volume of time-frequency-space. The signal decomposition module performs nonnegative tensor decomposition on the time-frequency-space three-dimensional joint data volume, separating it into multiple independent signal components. Specifically, it includes: For the three-dimensional joint data volume of time-frequency-space, the non-negative tensor decomposition algorithm is used to perform the decomposition operation. The number of signal components to be decomposed is preset, and the feature matrices of each signal component in time, frequency and two spatial dimensions are randomly initialized. Based on the preset iterative update rules, with the optimization objective of minimizing the overall error between the reconstructed data volume and the original data volume, each feature matrix is ​​updated alternately, and additional constraints on the amplitude smoothing of adjacent positions are applied to the feature matrix of the spatial dimension during the update process. When the change in the overall error between two consecutive iterations is less than a preset threshold, the decomposition is determined to be converged, the iteration is stopped, and the multiple independent signal components obtained by the decomposition and their corresponding feature matrices of each dimension are output. The target screening module identifies and screens target signal components corresponding to abnormal structural disturbances inside the thin plate from multiple independent signal components based on the differences between the frequency characteristics of the signal components and the reference intrinsic response spectrum obtained in advance by detecting defect-free areas, as well as the uniformity of their distribution in the spatial dimension. The defect reconstruction and visualization module determines the location of the abnormal structure in the thickness direction of the thin plate based on the distribution characteristics of the target signal components in the time dimension, and reconstructs the two-dimensional morphology of the abnormal structure in the thin plate plane by combining the distribution characteristics in the spatial dimension, and then synthesizes the three-dimensional location and morphology information of the full-thickness defects inside the non-ferromagnetic thin plate.

Images