A non-destructive testing method and system for metal powder injection molding green bodies

Abstract

This invention belongs to the field of nondestructive testing technology, specifically relating to a method and system for nondestructive testing of metal powder injection molding blanks. The method includes: constructing a high-order response tensor containing spatial and excitation parameter dimensions; extracting local geometric features of a three-dimensional geometric model to generate static geometric prior information; adjusting the sparse regularization term by combining the peak-entropy ratio dynamic significance index and the static prior information during tensor orthogonal decomposition iteration; setting a threshold based on the peak-entropy ratio of the converged component basis to filter defect components; reconstructing the initial defect representation data volume; and locating internal defects through multi-scale wavelet packet transform and cross-scale correlation criteria. This invention achieves high-precision three-dimensional spatial localization of minute internal defects in the green blank by effectively removing strong background interference.

Description

Technical Field

[0001] This invention relates to the field of nondestructive testing technology. More specifically, this invention relates to a method and system for nondestructive testing of metal powder injection molding blanks. Background Technology

[0002] Metal powder injection molding is an advanced near-net-shape manufacturing technology suitable for the large-scale production of high-precision metal parts with complex three-dimensional geometries. In this process, the primary product formed by injection molding of metal powder mixed with a polymer binder is usually called a green compact. If hidden defects such as pores, microcracks, powder agglomeration, or uneven binder distribution form inside the green compact during injection molding, these defects are usually difficult to eliminate on their own during subsequent debinding and high-temperature sintering. Instead, they may further expand with the release of thermal stress, leading to a decrease in the mechanical properties of the finished product, and in severe cases, causing batch scrap. Because the green compact is a multiphase composite material, it is relatively fragile and often has a complex and variable three-dimensional geometry. Its response to external multi-field excitation signals often exhibits obvious spatial non-uniformity and variability.

[0003] Currently, acquiring full-field spatial response data by performing multi-field excitation scanning on the test object, constructing a high-dimensional tensor based on the extracted feature parameters, and performing tensor decomposition is a common method for achieving feature separation in non-destructive testing. Chinese patent application CN121577750A discloses a non-destructive testing method and system for weld defects in pressure vessels using multimodal ultrasonic fusion. This method utilizes ultrasonic equipment to acquire massive amounts of radio frequency data containing spatial coordinates and detection parameters. Multidimensional feature vectors are arranged according to spatial position to construct a three-dimensional feature tensor, which is then decomposed using Tucker decomposition to obtain the core tensor. Subsequently, weighted summaries and feature enhancements are performed using physical modeling. Based on these techniques, massive response data containing multidimensional spatial coordinates and multiple excitation parameters can be processed to a certain extent, achieving multidimensional feature fusion and completing defect detection and identification.

[0004] However, while the aforementioned technical solutions can process massive amounts of multidimensional high-order response data and extract defect features to some extent, in actual nondestructive testing practice, when performing three-dimensional full-field inspection on complex geometric green blanks, conventional data processing and feature extraction methods often rely on global unified analysis or simple dimensionality reduction when faced with such high-order data. This makes it difficult to fully utilize the constraint effect of the test object's own three-dimensional geometric prior structure on the response distribution. In the boundary regions of curved surfaces or deep interiors far from the surface, weak early defect signals are easily submerged by geometric morphology interference, material background fluctuations, or overall evolution trends. Due to the lack of an intelligent separation mechanism that can combine local spatial geometric features with signal change patterns, existing technologies still have shortcomings in stripping away strong background interference, discovering hidden defect features, and reconstructing weak defects with high signal-to-noise ratio representations, which can easily lead to missed detections, misjudgments, and three-dimensional spatial positioning errors. Summary of the Invention

[0005] To address the technical problem of missed detection and misjudgment in the detection of weak defects in complex geometric green blanks, this invention provides solutions in the following aspects.

[0006] In a first aspect, the present invention provides a non-destructive testing method for metal powder injection molding blanks, comprising: S1, acquiring full-field three-dimensional response data of the metal powder injection molding blank under a preset excitation field, and constructing a high-order response tensor including spatial dimension and excitation parameter dimension; establishing a three-dimensional geometric model of the metal powder injection molding blank under test, and extracting local geometric features to generate static geometric prior information; S2, constructing an iterative optimization process by orthogonally decomposing the high-order response tensor, constructing an initial sparse regularization term based on the static geometric prior information in the first iteration; and in subsequent iterations, calculating the peak entropy ratio of each orthogonal component basis of the previous decomposition in the excitation parameter dimension as a dynamic... Significance index: A joint adjustment factor is constructed by combining dynamic significance index and static geometric prior information. The sparse regularization term for tensor reconstruction is adjusted by back mapping according to the joint adjustment factor until iterative convergence, resulting in orthogonal component basis and core tensor. S3: Based on the peak entropy ratio of each orthogonal component basis after iterative convergence, a discrimination threshold is set, and defect-related orthogonal component basis with a peak entropy ratio higher than the discrimination threshold is selected. An initial defect representation data volume is generated by reconstructing the defect-related orthogonal component basis. Multi-scale wavelet packet transform is performed on the initial defect representation data volume, and defects inside the metal powder injection green body are identified and located according to the local modulus maxima and cross-scale correlation criteria.

[0007] This invention constructs an intelligent separation mechanism that combines local spatial geometric features with signal variation patterns by extracting static geometric prior information and introducing a sparse regularization term for tensor reconstruction that dynamically adjusts the peak-entropy ratio. This fully utilizes the constraint effect of the three-dimensional geometric prior structure on the response distribution, improving the ability to extract hidden and weak defect features under strong background interference. By combining multi-scale wavelet packet transform and cross-scale correlation criteria, the probability of missed detection and misjudgment is reduced, achieving high-precision three-dimensional spatial positioning of minute defects inside the green blank to be tested.

[0008] Preferably, acquiring the full-field three-dimensional response data of the metal powder injection preform under a preset excitation field to construct a high-order response tensor containing spatial and excitation parameter dimensions includes: using a mechanical scanning arm to drive an air-coupled array ultrasonic probe to scan the metal powder injection preform under test along a set surface two-dimensional grid trajectory, emitting multi-band ultrasonic excitation pulses at each grid node and acquiring ultrasonic reflected echo signals; extracting the echo amplitude time series of the ultrasonic reflected echo signals, converting the echo flight time into depth parameters according to the pre-calibrated equivalent material sound velocity, and converting the depth parameters into three-dimensional voxel coordinates by combining the spatial position of the scanning node, the probe attitude, and the sound beam propagation direction; and performing feature matching and combination of the three-dimensional voxel coordinates, excitation frequency parameters, and echo amplitude to construct a four-dimensional numerical matrix containing three-dimensional spatial dimensions and one-dimensional excitation parameter dimensions as a high-order response tensor.

[0009] This invention extracts the echo flight time and combines it with the position and attitude of the scanning node to calculate the three-dimensional voxel coordinates, thereby constructing a high-order response tensor of a four-dimensional numerical matrix, which preserves the complete physical mapping information of the evolution of the full-field three-dimensional response data with space and excitation parameters.

[0010] Preferably, extracting local geometric features to generate static geometric prior information includes: extracting the normal vectors of each spatial node on the surface of the three-dimensional geometric model and calculating the surface curvature distribution features; calculating the Euclidean distance from each spatial node inside the three-dimensional geometric model to the nearest model surface; mapping the surface curvature distribution features to the voxel mesh inside the three-dimensional geometric model and combining them with the Euclidean distance to generate static geometric prior information representing the boundary variation features and spatial depth of the green body structure of the metal powder injection test piece.

[0011] This invention calculates curvature distribution features by extracting normal vectors and combining them with Euclidean distance parameters, thereby realizing a multidimensional representation of local geometric features and providing a reliable geometric constraint basis for accurately characterizing the boundary changes and spatial depth of the object under test.

[0012] Preferably, the calculation of the peak-entropy ratio of each orthogonal component basis in the excitation parameter dimension of the previous decomposition as a dynamic significance index includes: extracting the amplitude sequence of each orthogonal component basis in the excitation parameter dimension, calculating the ratio of the fourth central moment to the square of the second central moment of the amplitude sequence to obtain the kurtosis value; using a normalization function to transform the amplitude sequence into a probability distribution sequence, calculating the Shannon entropy of the probability distribution sequence; and dividing the kurtosis value by the sum of the Shannon entropy and a preset zero-prevention constant to obtain the peak-entropy ratio representing the signal mutation characteristics and energy concentration degree.

[0013] This invention accurately characterizes the abrupt change features and energy concentration of a signal by comprehensively calculating the kurtosis and Shannon entropy of the amplitude sequence, thereby improving the sensitivity of feature separation in evaluating the response of local anomalous signals using dynamic significance indicators.

[0014] Preferably, a joint adjustment factor is constructed by combining the dynamic significance index and the static geometric prior information. The sparse regularization term for tensor reconstruction is adjusted by inverse mapping of the joint adjustment factor, including: normalizing the dynamic significance index, combining it with the Euclidean distance parameter in the static geometric prior information, and constructing a space-excitation joint adjustment factor through tensor product operation; inverse mapping the joint adjustment factor to a sparse regularization penalty coefficient and inputting it into the sparse regularization term, so that the sparse regularization penalty coefficient corresponding to the region with a larger peak-to-entropy ratio and a larger Euclidean distance parameter is lower.

[0015] Preferably, a screening threshold is set based on the peak entropy ratio of each orthogonal component basis after iterative convergence, and defect-related orthogonal component basis with a peak entropy ratio higher than the screening threshold is selected. This includes: arranging the peak entropy ratios of each orthogonal component basis in descending order of numerical value to generate a peak entropy ratio sequence; calculating the gradient difference between adjacent values ​​in the peak entropy ratio sequence, finding the position where the gradient changes most abruptly, and setting the peak entropy ratio value corresponding to the position after the position of the largest abrupt change as the screening threshold; retaining orthogonal component basis with a peak entropy ratio greater than the screening threshold as defect-related orthogonal component basis, recording the index corresponding to the defect-related orthogonal component basis, and removing the remaining orthogonal component basis with a peak entropy ratio less than or equal to the screening threshold.

[0016] This invention achieves adaptive dynamic calibration of the discrimination threshold by calculating the gradient difference between the descending peak entropy ratio and the sequence and locking the position of the maximum mutation, thus ensuring the objectivity and accuracy of screening defect-related orthogonal component bases.

[0017] Preferably, the initial defect representation data volume is subjected to multi-scale wavelet packet transform, and the defects inside the metal powder injection green compact to be tested are identified and located according to the local modulus maxima and cross-scale correlation criteria. This includes: performing three-level wavelet packet decomposition on the initial defect representation data volume, finding local modulus maxima points in the high-frequency wavelet coefficients of each decomposition scale; uniformly mapping the coordinates of the local modulus maxima points at each decomposition scale to the original three-dimensional voxel coordinate system, calculating the spatial overlap of local modulus maxima points between adjacent decomposition scales; identifying local modulus maxima points with a spatial overlap greater than a set proportion as real defect response points, and locating the defects according to the original three-dimensional voxel coordinates.

[0018] Preferably, the orthogonal decomposition of the higher-order response tensor is constructed as an iterative optimization process, including: constructing the objective optimization problem based on the tensor alternating least squares method framework, initializing the factor matrix of each dimension using the higher-order singular value decomposition method; after each round of factor matrix update, performing orthogonalization processing on the factor matrix of each dimension so that the column vectors of the updated factor matrix of each dimension satisfy the orthogonal constraint.

[0019] Preferably, the initial defect representation data volume is generated by reconstructing the defect-related orthogonal component basis, including: extracting the component indices of the defect-related orthogonal component basis in the excitation parameter dimension; retaining the core sub-tensor elements in the core tensor corresponding to the component indices, setting the core sub-tensor elements corresponding to the unretained component indices to zero, and generating the processed core tensor; performing tensor shrinkage and pattern multiplication operations on the processed core tensor and the orthogonal component basis in each spatial dimension to generate the initial defect representation data volume.

[0020] Secondly, the present invention provides a non-destructive testing system for metal powder injection green bodies, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned non-destructive testing method for metal powder injection green bodies is implemented.

[0021] By adopting the above technical solution, a computer program for the above-mentioned non-destructive testing method for metal powder injection molding is generated and stored in a memory so that it can be loaded and executed by a processor. A terminal device can then be manufactured based on the memory and the processor for convenient use.

[0022] The beneficial effects of this invention are as follows:

[0023] This invention constructs a high-order response tensor containing spatial and excitation parameter dimensions, and extracts static geometric prior information using a three-dimensional geometric model, achieving a fusion representation of the response characteristics and spatial structural features of the green billet under test. During the iterative optimization of orthogonal decomposition, the peak-entropy ratio of the orthogonal component basis is used as a dynamic significance index. This index, combined with geometric prior information, adaptively adjusts the sparse regularization term in the tensor reconstruction. This effectively reduces the sparse constraint strength of components far from the surface and with high significance, thus more completely preserving the weak defect responses hidden in strong background interference, improving the stability, signal-to-noise ratio, and feature separation capability of the tensor decomposition results. By screening defect-related components using a threshold and performing inverse transform reconstruction, combined with multi-scale wavelet packet transform and cross-scale feature criteria, the response representation capability of small defects is enhanced, achieving high-sensitivity identification and high-precision three-dimensional spatial positioning of internal defects in the green billet, improving the accuracy and reliability of non-destructive testing results. Attached Figure Description

[0024] Figure 1 This is a flowchart of a non-destructive testing method for metal powder injection green bodies according to the present invention; Figure 2 This is a schematic diagram comparing the probability distribution of the orthogonal component basis in the excitation parameter dimension in this invention; Figure 3 This is a schematic diagram of the orthogonal component base peak entropy ratio sequence and difference gradient distribution in this invention; Figure 4This is a schematic diagram comparing the number of defects detected and the number of false alarms for each detection scheme in this invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0026] This invention discloses a non-destructive testing method for metal powder injection molding green bodies, referring to... Figure 1 This includes steps S1-S3: S1. Obtain response data and extract geometric prior information.

[0027] In an optional embodiment, the full-field three-dimensional response data of the metal powder injection preform under test is acquired under a preset excitation field and constructed as a high-order response tensor containing spatial and excitation parameter dimensions. A three-dimensional geometric model of the metal powder injection preform under test is established, and local geometric features are extracted and represented as static geometric prior information.

[0028] A non-contact air-coupled array ultrasonic probe mounted on a mechanical scanning arm is used to perform multi-frequency ultrasonic excitation scanning on the metal powder injection preform under test, and the corresponding ultrasonic echo response data is acquired. For each discrete excitation frequency, a set of three-dimensional echo amplitude volume data is acquired according to the same surface scanning trajectory. Then, the three-dimensional echo amplitude volume data at different excitation frequencies are stacked along the excitation parameter dimension to construct a fourth-order response tensor containing three spatial dimensions and one excitation frequency dimension. This three-dimensional geometric model is provided by a computer-aided geometric model of the metal powder injection preform under test or by a geometric mesh model registered with the detection coordinate system. The geometric model is used to extract static geometric prior information such as curvature, normal vector, and Euclidean distance to the model surface. The ultrasonic response data and geometric prior information are spatially registered through a unified coordinate system.

[0029] When extracting static geometric prior information, the shortest Euclidean distance from each voxel in the higher-order response tensor space to the model surface is calculated to generate a three-dimensional distance field matrix. Then, a complementary Gaussian mapping or linear normalization mapping, which is monotonically positively correlated with distance, is used to convert the distance values ​​into depth weight values ​​between 0 and 1, ensuring that the depth weight value increases with distance from the surface. This complementary Gaussian mapping is represented as depth weight equal to 1 minus the distance decay term, or it can be implemented using an equivalent monotonically increasing function after normalization, avoiding the inconsistency of directly interpreting the Gaussian decay function as having a larger weight with greater distance. This three-dimensional depth weight matrix, together with the curvature normalization result, constitutes quantitative static geometric prior information. Before combining the curvature distribution characteristics with the Euclidean distance parameters, the curvature characteristics of the model surface are first extended to the three-dimensional voxel mesh of the higher-order response tensor using nearest surface point mapping, normal extension, or spatial interpolation, so that each internal voxel corresponds to a curvature prior value and a depth distance value, before weighted fusion.

[0030] In one implementation, the full-field three-dimensional response data of the metal powder injection preform under a preset excitation field is acquired and constructed into a high-order response tensor containing spatial and excitation parameter dimensions. Specifically, a mechanical scanning arm drives a non-contact air-coupled array ultrasonic probe to scan the metal powder injection preform along a set surface two-dimensional grid trajectory, emitting multi-band ultrasonic excitation pulses at each grid node; ultrasonic reflected echo signals at each node are acquired, the echo amplitude time series is extracted, and the echo flight time is converted into a depth parameter according to the pre-calibrated equivalent material sound velocity of the metal powder injection preform; combined with the spatial position of the scanning node in a unified detection coordinate system, the probe attitude, and the sound beam propagation direction, the depth parameter is converted into three-dimensional voxel coordinates; the three-dimensional voxel coordinates, excitation frequency parameters, and echo amplitudes are feature-matched and combined to construct a four-dimensional numerical matrix containing three-dimensional spatial dimensions and one-dimensional excitation parameter dimensions as a high-order response tensor.

[0031] During the acquisition of full-field three-dimensional response data, a 6-axis mechanical scanning arm equipped with a broadband non-contact air-coupled array ultrasonic probe is driven. The excitation frequency parameters of the ultrasonic probe cover multiple discrete frequency points within the effective bandwidth of the air-coupled probe. The scanning process proceeds at a constant speed along a pre-set surface two-dimensional grid trajectory, with the grid node step size resolution set between 0.1mm and 1mm. For example, a step size of 0.5mm constitutes... Surface scanning dot matrix.

[0032] At each two-dimensional grid coordinate node, the probe sequentially emits ultrasonic excitation pulses according to a preset discrete frequency sequence, for example, setting 10 step frequency channels; each frequency channel corresponds to the acquisition of one echo amplitude time series. Therefore, the frequency dimension is not obtained by simply stacking or resampling three-dimensional image sequences, but rather by echo response data repeatedly acquired and registered at different excitation frequencies. For the first excitation frequency, a set of three-dimensional amplitude volume data is obtained by converting the echo amplitude time series from all surface scanning nodes; for... to The acquisition and registration process is repeatedly executed at each excitation frequency, and... The three-dimensional amplitude data are stacked sequentially along the excitation frequency dimension to form a four-dimensional numerical matrix.

[0033] The sampling rate of the ultrasonic acquisition card is preferably set to 50MHz to 100MHz, for example, 100MHz; the time window is preferably set to 0 to 20μs. Taking a pre-calibrated equivalent material sound velocity of 3500m / s as an example, based on the pulse-echo two-way propagation relationship, the sampling time is converted into depth coordinates, with a maximum representable depth of 35mm, and the original number of time sampling points is... The original depth sampling interval was 0.0175 mm. To reduce computational load, the original 2000 depth sampling points were compressed into 500 depth voxels using equal-interval grouping averaging, peak preservation, or interpolation resampling methods, resulting in an effective depth interval of 0.07 mm. When the surface scan point array is... When the depth voxel count is 500 and the excitation frequency channel count is 10, the constructed dimension is The four-dimensional numerical matrix data structure is used as the higher-order response tensor for subsequent iterative orthogonal decomposition.

[0034] In one implementation, for establishing a three-dimensional geometric model of the metal powder injection molding blank to be tested, local geometric features are extracted and represented as static geometric prior information. Specifically, the computer-aided geometric model of the metal powder injection molding blank to be tested is loaded, the normal vectors of each spatial node on the model surface are extracted, and the surface curvature distribution characteristics are calculated; the Euclidean distance from each spatial node inside the model to the nearest model surface is calculated; the curvature distribution characteristics are mapped to the voxel mesh inside the model and combined with the Euclidean distance to generate static geometric prior information representing the boundary variation characteristics and spatial depth of the metal powder injection molding blank to be tested.

[0035] When establishing a three-dimensional geometric model of the metal powder injection preform to be tested and representing prior information, the computer-aided geometric model of the metal powder injection preform to be tested is loaded through a standard interface, such as a file with the extension .sTL or .STEP. The spatial region of the computer-aided geometric model is discretized into a mesh according to the spatial sampling interval of the test data, so that the resolution in the X and Y directions is consistent with the surface scanning step size, and the resolution in the Z direction is consistent with the depth voxel interval after resampling; for example, the sampling interval in the X and Y directions is 0.5 mm, and the effective depth interval in the Z direction is 0.07 mm. All triangular facets or boundary voxel nodes on the model surface are traversed, the unit normal vector of each surface node is extracted, and the local Gaussian curvature and mean curvature of each node are calculated using a differential geometry algorithm. Surface curvature features include mean curvature, Gaussian curvature, or a combination of mean curvature and Gaussian curvature, resulting in curvature intensity values. This curvature distribution characteristic represents the degree of boundary unevenness and variation of the tested metal powder injection green body structure. Before participating in weighted fusion, this surface curvature feature undergoes absolute value conversion, truncation, and normalization processing, transforming the processed surface curvature feature into a value within the range of Gaussian curvature. The dimensionless curvature parameter of a closed interval, for example, for a certain acute-angled edge region, obtains the normalized curvature intensity value.

[0036] Using a pre-built spatial index structure, the Euclidean distance parameter from the coordinate point of a three-dimensional voxel inside the model to the nearest surface of the green model is calculated. This Euclidean distance parameter represents the spatial depth of any point within the material, and its preferred detection range is set between 0 mm and 50 mm. For example, the nearest Euclidean distance from a voxel in the central region of the model to the surface is 12.5 mm.

[0037] When combining curvature distribution characteristics and Euclidean distance, the absolute value of curvature is taken as the curvature intensity value. This curvature intensity value is then truncated and normalized according to a preset upper limit of curvature or the maximum curvature within the sample to obtain the value at which the curvature intensity is within the range of the sample. Normalized curvature parameters for closed intervals; simultaneously, dividing the Euclidean distance by the maximum effective detection depth within the model or using complementary Gaussian mapping to obtain the normalized curvature parameters for closed intervals; A normalized depth parameter is defined as a closed interval that increases with distance. Then, the normalized curvature parameter and the normalized depth parameter are combined using a weighted fusion formula. For example, setting the curvature weight coefficient to 0.4 and the distance weight coefficient to 0.6, the normalized curvature parameter and the normalized depth parameter are superimposed to calculate the depth parameter at each spatial node. A single scalar value representing a closed interval is used as static geometric prior information.

[0038] S2. The iterative optimization process of constructing tensor orthogonal decomposition.

[0039] In an optional embodiment, the orthogonal decomposition of the higher-order response tensor is constructed as an iterative optimization process. In the first iteration, an initial sparse regularization term is constructed based on static geometric prior information. In subsequent iterations, the peak entropy ratio of each orthogonal component basis from the previous decomposition in the excitation parameter dimension is calculated as a dynamic significance index. The dynamic significance index and the static geometric prior information are combined to construct a joint adjustment factor through tensor product. The sparse regularization term for tensor reconstruction is adjusted by inverse mapping based on the joint adjustment factor, reducing the sparsity constraint of components with high dynamic significance index and located far from the surface of the 3D geometric model in the next decomposition, until the iteration converges, obtaining the orthogonal component basis and the core tensor.

[0040] The objective optimization problem is constructed based on the tensor alternating least squares method framework, and the factor matrices of each dimension are initialized using higher-order singular value decomposition or equivalent initialization. After each round of factor matrix update, orthogonalization is performed on the factor matrices of each dimension to ensure that the column vectors of the updated factor matrices satisfy orthogonal constraints, guaranteeing that the output results can be used as orthogonal component bases. In the first iteration, an initial sparse regularized penalty coefficient tensor is constructed based on static geometric prior information, and the size of the initial sparse regularized penalty coefficient tensor matches that of the original spatial excitation frequency domain reconstruction tensor. This penalty coefficient tensor is used to constrain the tensor reconstruction results or to constrain the component contribution tensor, rather than directly requiring element-wise multiplication with the low-rank core tensor. When the specific optimization implementation applies constraints to the core tensor domain, the penalty coefficient tensor of the original spatial excitation frequency domain is first converted into a low-rank penalty coefficient tensor by factor matrix projection, block averaging, or same-scale aggregation, and then the low-rank penalty coefficient tensor participates in the regularization update of the core variables.

[0041] In each subsequent iteration, the amplitude sequence is extracted from each column of the excitation parameter dimension factor matrix obtained from the previous decomposition. The ratio of the fourth central moment to the square of the second central moment of the amplitude sequence is calculated to obtain the kurtosis value. Then, the amplitude sequence is normalized to a probability distribution sequence and the Shannon entropy is calculated. The kurtosis value is divided by the sum of the Shannon entropy and the zero constant to obtain the peak-entropy ratio.

[0042] in, Peak entropy ratio; Kurtosis value; Shannon entropy; To prevent zero constant.

[0043] It should be noted that in the peak-entropy ratio, "peak" refers to the kurtosis value, which is not the maximum absolute amplitude at a single point. The orthogonal component basis used to calculate the peak-entropy ratio is the column vectors in the excitation parameter dimension factor matrix, with each column vector corresponding to an excitation dimension component index. The spatial dimension factor matrix is ​​mainly used for the subsequent inverse transformation reconstruction of the defect representation data volume.

[0044] The peak-entropy ratio of each component is normalized to obtain a dynamic significance index. This dynamic significance index is then combined with the normalized depth parameter from the static geometric prior information to construct a joint spatial and excitation regulation factor. A sparse regularization penalty coefficient tensor is obtained based on the inverse mapping of the joint regulation factor. The update logic for the penalty coefficient is as follows: the larger the peak-entropy ratio and the larger the normalized depth parameter, the more likely the corresponding component is a deep abrupt defect response far from the surface, and the lower the sparse regularization penalty coefficient corresponding to the deep abrupt defect response. Specifically, the calculation method for the next-generation penalty coefficient corresponding to the current region and component is as follows: first, the regulation constant is multiplied by the sum of the normalized peak-entropy ratio value, the normalized depth value, and 1; then, the basic penalty coefficient is divided by this sum to obtain the next-generation penalty coefficient, thus ensuring that components with high dynamic significance and far from the surface receive weaker sparsity suppression in the next decomposition.

[0045] At the end of each iteration, calculate the relative error of the Frobenius norm of the core tensor between the two updates. If the relative error is less than... If the iteration converges, the loop stops, and the three spatial dimension factor matrices and one excitation parameter dimension factor matrix after stabilization are output as orthogonal component bases. Simultaneously, the core tensor obtained from the decomposition is also output. It should be noted that the Frobenius norm is a norm that measures the overall size of a matrix or tensor. Typically, the Frobenius norm is equal to the square root of the sum of the squares of all elements of the matrix or tensor, and it is used to represent the overall difference between two consecutive matrix or tensor update results.

[0046] In one implementation, the peak-entropy ratio of each orthogonal component basis from the previous decomposition along the excitation parameter dimension is used as a dynamic significance index. Specifically, the amplitude sequence of each orthogonal component basis along the excitation parameter dimension is extracted; the ratio of the fourth central moment to the square of the second central moment of the amplitude sequence is calculated to obtain the kurtosis value; the amplitude sequence is transformed into a probability distribution sequence using a normalization function; the Shannon entropy of the probability distribution sequence is calculated; and the kurtosis value is divided by the sum of the Shannon entropy and a preset zero-prevention constant to obtain the peak-entropy ratio, which represents the signal abrupt change characteristics and the degree of energy concentration.

[0047] In the calculation of peak-entropy ratio as a dynamic significance index, for each orthogonal component basis output from the previous decomposition iteration, the discrete amplitude sequence of the orthogonal component basis in the excitation parameter dimension is extracted. Assuming the number of ultrasonic multi-band excitation channels is 10, the length of the discrete amplitude sequence is 10. Statistical central moments are calculated element-wise for each discrete amplitude sequence. These statistical central moments include the second and fourth central moments deviating from the mean; the second central moment is the sequence variance. The kurtosis variable of the component basis sequence is calculated by dividing the fourth central moment by the square of the second central moment. Kurtosis can detect the peak-heavy tail characteristic of the amplitude sequence. For components with abrupt defect responses, the kurtosis value is greater than the reference value for the normal background; for example, the kurtosis value of a uniform background noise component is equal to 3, while the kurtosis value of a component containing defect resonance characteristics is... The value is 8.5. The amplitude sequence is mapped to a probability distribution sequence with a sum of 1 using either the absolute value proportion method or the square energy proportion method, ensuring that the probability value of each frequency response component is within the range of 1. Within the interval. Using information theory formulas, the Shannon entropy variable of the probability distribution sequence is calculated. The Shannon entropy variable represents the degree of disorder or concentration of signal energy in each frequency band. A more concentrated energy distribution corresponds to a lower Shannon entropy. For example, the calculated Shannon entropy... The value is 1.1. Enter the zero-prevention constant, for example, the zero-prevention constant. for The peak-entropy ratio is calculated by dividing the kurtosis value by the sum of the Shannon entropy and the zero-prevention constant, generating an index for the degree of abrupt change and energy accumulation state of the fused signal. Substituting these parameters into the calculation yields the peak-entropy ratio. The larger the peak-entropy ratio, the more prominent the local anomalies or sudden reflection characteristics of the corresponding component basis under multi-band ultrasonic excitation.

[0048] Reference Figure 2 The data changes of the anomalous component distribution and the background component distribution are shown. As the number of discrete excitation frequency channels increases, the anomalous component distribution shows a significant spike near channel 4, exhibiting a probability distribution characteristic of highly concentrated energy. The background component distribution is relatively smooth and uniform across all discrete excitation frequency channels, without significant probability spikes. By comparing the data change trends of the background component distribution and the anomalous component distribution, it is demonstrated that the anomalous component has prominent burst reflection characteristics in a specific frequency band, verifying the rationality of using the peak-entropy ratio to measure the degree of local component anomalies.

[0049] In one implementation, a joint adjustment factor is constructed by combining the joint dynamic significance index and static geometric prior information through tensor product. The sparse regularization term of the tensor reconstruction is then adjusted based on the inverse mapping of this joint adjustment factor, reducing the sparsity constraint in the next decomposition for components with high dynamic significance indexes located far from the surface of the 3D geometric model. Specifically, the dynamic significance index is normalized and combined with the Euclidean distance parameter from the static geometric prior information to construct a joint spatial and excitation adjustment factor through tensor product operation. This joint adjustment factor is then converted into a sparse regularization penalty coefficient through inverse mapping. This sparse regularization penalty coefficient is input into the sparse regularization term of the tensor reconstruction, resulting in lower sparse regularization penalty coefficients for regions with larger peak-to-entropy ratios and farther distances from the model surface, thereby updating and replacing the constraints of the previous iteration.

[0050] When adjusting the sparse regularization term of the tensor reconstruction, the dynamic significance index obtained from the previous generation of iterative calculations, namely the set of peak-entropy ratios of various components, is normalized to a maximum-minimum linear scaling method. Within the interval, normalized parameters are obtained. The Euclidean distance parameter is extracted from the static geometric prior information and divided by the global maximum detection depth within the model to convert it into a dimensionless spatial depth normalization coefficient. For example, if the global maximum detection depth is 25mm, the normalized parameter corresponding to a Euclidean distance of 15mm for a certain coordinate point within the model is 0.6.

[0051] Tensor product operations are performed between the spatial depth normalization coefficient and the significance parameter of the excitation component to generate a joint spatial and excitation adjustment factor. Then, a sparse regularization penalty coefficient is calculated based on this joint adjustment factor, instead of directly multiplying the reciprocal of the significance index by the static geometric prior information. The update logic for the penalty coefficient is as follows: the adjustment constant is multiplied by the sum of the peak entropy ratio normalization value, the depth normalization value, and 1; then, the base penalty coefficient is divided by this sum to obtain the next-generation penalty coefficient corresponding to the current region and component. The adjustment constant is preferably set to 5 to 10, for example, 5. Under this operation mechanism, if the peak entropy ratio normalization value corresponding to a certain internal region component is 0.9, indicating strong defect mutation characteristics, and the depth normalization value is 0.6, indicating a deep internal region far from the model surface, then the sparse regularization penalty coefficient will decrease from the base value of 0.1 to 0.027. The attenuated sparse regularization penalty coefficient can reduce the sparse filtering of deep strong mutation defects during iteration, so that internal anomalies far from the surface of the metal powder injection green body are preserved in the next tensor reconstruction iteration.

[0052] The size of the penalty coefficient tensor obtained by the inverse mapping of the joint regulation factor is consistent with the size of the object on which the penalty coefficient tensor acts. When the penalty coefficient tensor acts on the reconstructed data volume or component contribution tensor in the original spatial excitation frequency domain, the size of the penalty coefficient tensor is consistent with the higher-order response tensor, for example, a size of... When the penalty coefficient tensor acts on the core tensor of orthogonal tensor decomposition, the penalty coefficient tensor is first transformed into a function using factor matrix projection, component index aggregation, or block averaging. The core tensor is then used to measure the values, and the transformed penalty coefficient tensor is matched element-wise with the core tensor variables. This ensures that the penalty coefficient tensor and the actual optimization variables maintain dimensionality consistency.

[0053] S3. Screen defect-related component bases and locate internal defects.

[0054] In an optional embodiment, a discrimination threshold is set based on the peak-entropy ratio of the orthogonal component basis after iterative convergence. Defect-related orthogonal component basis with a peak-entropy ratio higher than the discrimination threshold is selected. An initial defect representation data volume is generated by reconstructing the defect-related orthogonal component basis. Multi-scale wavelet packet transform is performed on the initial defect representation data volume. Defects inside the metal powder injection green body are identified and located based on the local modulus maxima and cross-scale correlation criteria.

[0055] After iterative convergence, the peak entropy ratio values ​​of each orthogonal component basis in the excitation parameter dimension are extracted and sorted in descending order to generate a peak entropy ratio sequence. The gradient difference between adjacent values ​​in the peak entropy ratio sequence is calculated, and the position of the maximum gradient abrupt change is found. The peak entropy ratio corresponding to the position after the maximum abrupt change is set as the discrimination threshold. Subsequently, orthogonal component basis with peak entropy ratios greater than the discrimination threshold is retained as defect-related orthogonal component basis, and the position number of the retained component in the core tensor is extracted as an index. The position number as an index at least includes the component index of the defect-related orthogonal component basis in the excitation parameter dimension. During the inverse transformation reconstruction, the core sub-tensor elements in the core tensor corresponding to the component indices in the excitation parameter dimension are retained, and the core sub-tensor elements corresponding to the excitation parameter dimension component indices that were not retained are set to zero. Then, tensor shrinkage and pattern multiplication operations are performed with each spatial dimension orthogonal component basis to generate the initial defect representation data volume. This completes the reconstruction of the initial defect representation data volume based on the defect-related excitation dimension components.

[0056] The initial defect representation data volume is subjected to three-level discrete wavelet packet decomposition to obtain a three-dimensional matrix of wavelet coefficients at multiple scales from low frequency to high frequency. The voxel point with the local maximum absolute value in the three-dimensional matrix of wavelet coefficients at each scale is identified as the local modulus maxima. First, the coordinates of the local modulus maxima at each decomposition scale are uniformly mapped to the original three-dimensional voxel coordinate system, or the coordinates of the local modulus maxima at each decomposition scale are resampled to the same spatial resolution as the initial defect representation data volume. Then, the spatial coordinates of each local modulus maxima are compared along the increasing scale direction. If a spatial coordinate neighborhood is marked as a local modulus maxima in the wavelet coefficients of at least two adjacent scales, the cross-scale correlation criterion is satisfied. The original three-dimensional voxel coordinates are determined to be the location of micropore defects or crack defects inside the metal powder injection green body. All original three-dimensional voxel coordinates that satisfy the cross-scale correlation criterion are recorded and output, completing the defect identification and location.

[0057] In one implementation, a screening threshold is set based on the peak-entropy ratio of the orthogonal component bases after iterative convergence, and defect-related orthogonal component bases with peak-entropy ratios higher than the screening threshold are selected. Specifically, the peak-entropy ratios of each orthogonal component base after iterative convergence are sorted in descending order of numerical value to generate a peak-entropy ratio sequence; the gradient difference between adjacent values ​​in the peak-entropy ratio sequence is calculated, the position where the gradient changes abruptly the most is found, and the peak-entropy ratio corresponding to the position after the position of the largest change is set as the screening threshold; orthogonal component bases with peak-entropy ratios greater than the screening threshold are retained as defect-related orthogonal component bases, the corresponding indices of the orthogonal component bases are recorded, and the remaining component bases with peak-entropy ratios less than or equal to the screening threshold are removed.

[0058] After the iterative decomposition reaches the convergence tolerance condition, a defect feature screening and removal mechanism is established based on the obtained orthogonal component basis. All orthogonal component bases are extracted. Assuming the total number of bases set by tensor orthogonal decomposition is 50, corresponding to scalar values ​​of peak entropy ratio, these values ​​are arranged in descending order to generate a monotonically decreasing peak entropy ratio sequence of length 50. For example, the values ​​at the beginning of the sequence are 12.4, 9.8, 9.1, 3.2, 2.8, 1.5… The gradient variable of the difference between adjacent values ​​in the descending sequence is calculated using a first-order discrete difference function, forming a gradient distribution vector of length 49. The gradient distribution vector is traversed and scanned to find the global extremum location where the gradient value undergoes the largest abrupt change. Based on the aforementioned example sequence, the gradient difference between the third value 9.1 and the fourth value 3.2 reaches 5.9, which is greater than the gradient reduction of other values ​​in the sequence, such as 0.7 or 0.4. This pinpoints the sequence node indices where the mutation occurs, establishing the lower bound data point causing the mutation, i.e., the value 3.2 corresponding to the fourth position, as the discrimination threshold for distinguishing defect responses from normal structural noise. After threshold calibration, the first three orthogonal component bases with peak-to-entropy ratios greater than the discrimination threshold 3.2 are identified as orthogonal component bases carrying defect characteristics. The component indices of the first three orthogonal component bases in the excitation parameter dimension are extracted and recorded. The core sub-tensor elements corresponding to the excitation parameter dimension component indices in the core tensor are retained, while the core sub-tensor elements corresponding to the excitation parameter dimension component indices that are not retained are set to 0. Subsequently, tensor shrinkage and pattern multiplication operations are performed on the processed core tensor and the orthogonal component bases in each spatial dimension to reconstruct the initial defect representation data volume.

[0059] Reference Figure 3 The data demonstrates the changes in peak entropy ratio and difference gradient as the orthogonal component basis index increases. The peak entropy ratio distribution exhibits high values ​​in the initial region where the orthogonal component basis index is small, then drops sharply at a specific index position, subsequently remaining at a low and stable level across most of the remaining orthogonal component basis index range. The difference gradient distribution forms a significant maximum abrupt peak at the corresponding steep drop position. Through these two trends—the tortuous descent of the peak entropy ratio distribution and the abrupt global extremum of the difference gradient—it is demonstrated that the boundary between the small number of component basis bases containing high-intensity anomalous signals and the large number of remaining component basis bases dominated by background noise can be clearly defined. This verifies the reliability of using the position following the maximum abrupt point of the difference gradient to set the discrimination threshold and achieve effective separation of defect features.

[0060] In one implementation, a multi-scale wavelet packet transform is performed on the initial defect representation data volume. Defects inside the metal powder injection green compact are identified and located based on the local modulus maxima and cross-scale correlation criteria. Specifically, the initial defect representation data volume undergoes three-level wavelet packet decomposition. Local modulus maxima points are found in the high-frequency wavelet coefficients at each decomposition scale. After uniformly mapping the coordinates of the local modulus maxima points at each decomposition scale to the original three-dimensional voxel coordinate system, the spatial overlap of local modulus maxima points between adjacent scales is calculated. Local modulus maxima points with a spatial overlap greater than a set proportion are identified as true defect response points, and the internal defects of the metal powder injection green compact are located based on the original three-dimensional voxel coordinates.

[0061] In the defect localization and cross-scale feature fusion identification stage, the initial defect representation data volume after denoising and reconstruction is used as the original input. A wavelet basis with compactly supported orthogonal properties is selected, such as a wavelet basis from the Dobessi wavelet family, to perform a three-dimensional third-level wavelet packet transform. Multi-resolution analysis generates scale parameters. , and The corresponding multi-level high-frequency sub-band wavelet coefficient three-dimensional matrix. A matrix of preferably sized values ​​is set within the high-frequency spatial voxel data of each decomposition level. or A three-dimensional sliding spatial detection window is used for the voxel, and the absolute values ​​of wavelet coefficients within the window are compared pixel by pixel. Only when the coefficient of the center voxel is greater than the 26 or 124 three-dimensional neighboring pixels surrounding the center voxel is the spatial coordinate determined and recorded as a local modulus maxima at the current decomposition scale. To filter out spurious extrema caused by residual system perturbations, a cross-scale correlation metric based on spatial evolution consistency is input.

[0062] First, set the scale parameters , and The candidate maxima's 3D coordinates are uniformly mapped to the original 3D voxel coordinate system according to the corresponding scale factor, or resampled to a spatial resolution consistent with the initial defect representation data volume. Then, scale parameters are extracted one by one. and scale parameters The candidate maxima's three-dimensional coordinates are used to calculate the extracted coordinates and scale parameters. The Euclidean distance between corresponding maxima is used. If the spatial displacement deviation falls within the allowable three-dimensional alignment tolerance radius, preferably set to 1 to 2 voxel dimensions, and the equivalent error is controlled between 0.5 mm and 1 mm, then a cross-scale spatial coincidence is determined. A threshold for the cross-scale positional coincidence degree is set, preferably between 60% and 100%. For example, a lower limit for the coincidence ratio is required to be 66.7%, meaning that a certain modulus maxima must maintain neighborhood coincidence across at least two levels in all three-level scales. Only spatial nodes that simultaneously pass the high-frequency local modulus maxima detection and meet the lower limit requirement for cross-scale spatial coincidence degree will be confirmed by the non-destructive testing system as internal defects in the metal powder injection green body. The three-dimensional spatial coordinates of the defect confirmation point in a unified detection coordinate system are derived. For example, the centroid coordinates of a defect body are located as X=24.5 mm, Y=18.0 mm, and Z=8.5 mm, enabling visualization and location of the defect.

[0063] The experiment used a standard test piece of iron-based metal powder injection molding containing known, pre-defined internal minute defects as the test object. Hardware specifications included a six-axis mechanical scanning arm equipped with a multi-band non-contact air-coupled array ultrasonic probe, with a step size of 0.5 mm for continuous scanning. The acquisition frequency covered 0.5 MHz to 5 MHz. Each frequency channel acquired raw depth sampling points, which were then resampled and unified to 500 depth voxels, ultimately constructing a... The experiment used a high-order response tensor. Four control groups were set up. The basic control group used only conventional high-order tensor orthogonal decomposition and global fixed threshold extraction. Ablation group one added static geometric prior information based on surface curvature and spatial depth to the basic control group. Ablation group two added a tensor reconstruction sparse regularization adjustment module based on the joint peak-entropy ratio to ablation group one. The complete scheme group superimposed a localization algorithm based on a three-level wavelet packet transform and cross-scale correlation criteria.

[0064] Data was obtained through independent, repeated blind tests on 50 micropores and cracks at different depths embedded within a standard test piece. The basic control group detected 32 defects with 19 false alarms; the reconstructed defect signal-to-noise ratio (SNR) was 14.2 dB, and the average spatial positioning error was 1.7 mm. Ablation group one detected 39 defects, reducing false alarms to 11, improving the SNR to 18.5 dB, and achieving an average positioning error of 1.3 mm. Ablation group two successfully detected 46 defects, reducing false alarms to 3, improving the SNR to 23.6 dB, and decreasing the average positioning error to 0.8 mm. The complete solution group detected 49 real defects, achieving a detection rate of 98%, with zero false alarms under blind testing conditions on the standard test piece; the reconstructed signal SNR reached 28.4 dB, and the average absolute spatial positioning error was 0.4 mm.

[0065] Reference Figure 4This study demonstrates the evolution of defect detection and false alarm counts among the basic control group, ablation group 1, ablation group 2, and the complete scheme group. The basic control group exhibits a relatively low number of detections accompanied by a high number of false alarms. With the sequential introduction of static geometric prior information, sparse regularization adjustment module, and cross-scale correlation criterion, ablation groups 1 and 2 show an increasing number of defect detections and a decreasing number of false alarms. The complete scheme group achieves the highest number of defect detections and suppresses the number of false alarms to zero. The stepwise increase in the number of detections and the stepwise decrease in the number of false alarms demonstrate that the processing mechanisms such as static geometric prior information and sparse regularization adjustment work synergistically to filter out interference and retain weak defect features. This verifies that the combined multi-scale wavelet modulus maxima and spatial location coincidence test can effectively remove spurious extrema caused by residual system disturbances, thereby achieving a highly sensitive and accurate three-dimensional localization effect for minute defects inside the green blank.

[0066] The inclusion of static geometric prior information provides boundary constraints, reducing the false alarm rate in shallow regions. The joint peak-entropy ratio and spatial depth sparse regularization term reduces the probability of deep, weak defects being smoothly filtered out during tensor iteration, ensuring the complete preservation of deep, hidden anomalies far from the surface of the metal powder injection green body. This is reflected in an increase in the number of deep defect detections and an improvement in the signal-to-noise ratio. The joint verification criterion of input multi-scale wavelet mode maxima and spatial location coincidence utilizes the consistency of state evolution to remove spurious extrema caused by residual system disturbances. Under standard test conditions, this reduces the number of false alarms while achieving high-precision sub-millimeter-level three-dimensional defect localization.

[0067] This invention also discloses a non-destructive testing system for metal powder injection green bodies, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a non-destructive testing method for metal powder injection green bodies according to the present invention.

[0068] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0069] It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept, and these all fall within the scope of protection of this invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A non-destructive testing method for metal powder injection molding green bodies, characterized in that, include: S1. Obtain the full-field three-dimensional response data of the metal powder injection preform under the preset excitation field, and construct a high-order response tensor containing spatial dimension and excitation parameter dimension; establish a three-dimensional geometric model of the metal powder injection preform under test, and extract local geometric features to generate static geometric prior information. S2. The orthogonal decomposition of the high-order response tensor is constructed as an iterative optimization process. In the first iteration, an initial sparse regularization term is constructed based on static geometric prior information. In subsequent iterations, the peak entropy ratio of each orthogonal component basis of the previous decomposition in the excitation parameter dimension is calculated as a dynamic significance index. A joint adjustment factor is constructed by combining the dynamic significance index and the static geometric prior information. The sparse regularization term for tensor reconstruction is adjusted by back-mapping according to the joint adjustment factor until the iteration converges, and the orthogonal component basis and the core tensor are obtained. S3. Based on the peak-entropy ratio of each orthogonal component basis after iterative convergence, a discrimination threshold is set, and defect-related orthogonal component basis with a peak-entropy ratio higher than the discrimination threshold is selected; the initial defect representation data volume is reconstructed based on the defect-related orthogonal component basis; multi-scale wavelet packet transform is performed on the initial defect representation data volume, and defects inside the metal powder injection green body are identified and located based on the local modulus maxima and cross-scale correlation criteria.

2. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, To acquire the full-field three-dimensional response data of the metal powder injection preform under a preset excitation field and construct a high-order response tensor containing spatial and excitation parameter dimensions, the following steps are taken: using a mechanical scanning arm to drive an air-coupled array ultrasonic probe to scan the metal powder injection preform along a set surface two-dimensional grid trajectory, emitting multi-band ultrasonic excitation pulses at each grid node and acquiring ultrasonic reflected echo signals; extracting the echo amplitude time series of the ultrasonic reflected echo signals, converting the echo flight time into depth parameters according to the pre-calibrated equivalent material sound velocity, and converting the depth parameters into three-dimensional voxel coordinates by combining the spatial position of the scanning node, the probe attitude, and the sound beam propagation direction; and combining the three-dimensional voxel coordinates, excitation frequency parameters, and echo amplitudes through feature matching to construct a four-dimensional numerical matrix containing three-dimensional spatial dimensions and one-dimensional excitation parameter dimensions as the high-order response tensor.

3. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, Extracting local geometric features to generate static geometric prior information includes: extracting the normal vectors of each spatial node on the surface of the 3D geometric model and calculating the surface curvature distribution features; calculating the Euclidean distance from each spatial node inside the 3D geometric model to the nearest model surface; mapping the surface curvature distribution features to the voxel mesh inside the 3D geometric model and combining them with the Euclidean distance to generate static geometric prior information representing the boundary variation features and spatial depth of the green body structure of the metal powder injection test piece.

4. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, The peak-entropy ratio of each orthogonal component basis from the previous decomposition in the excitation parameter dimension is calculated as a dynamic significance index. This includes: extracting the amplitude sequence of each orthogonal component basis in the excitation parameter dimension; calculating the ratio of the fourth central moment to the square of the second central moment of the amplitude sequence to obtain the kurtosis value; using a normalization function to transform the amplitude sequence into a probability distribution sequence; calculating the Shannon entropy of the probability distribution sequence; and dividing the kurtosis value by the sum of the Shannon entropy and a preset zero-prevention constant to obtain the peak-entropy ratio, which represents the signal abrupt change characteristics and the degree of energy concentration.

5. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, A joint adjustment factor is constructed by combining dynamic significance index and static geometric prior information. The sparse regularization term for tensor reconstruction is adjusted by inverse mapping of the joint adjustment factor. This includes: normalizing the dynamic significance index, combining it with the Euclidean distance parameter in the static geometric prior information, and constructing a joint adjustment factor for space and excitation through tensor product operation; and inverse mapping the joint adjustment factor to a sparse regularization penalty coefficient as input to the sparse regularization term, so that the sparse regularization penalty coefficient is lower for regions with larger peak-to-entropy ratios and larger Euclidean distance parameters.

6. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, Based on the peak-entropy ratio of each orthogonal component basis after iterative convergence, a screening threshold is set to select defect-related orthogonal component basis with a peak-entropy ratio higher than the screening threshold. This includes: sorting the peak-entropy ratios of each orthogonal component basis in descending order of numerical value to generate a peak-entropy ratio sequence; calculating the gradient difference between adjacent values ​​in the peak-entropy ratio sequence, finding the position where the gradient changes most abruptly, and setting the peak-entropy ratio corresponding to the position after the position of the largest abrupt change as the screening threshold; retaining orthogonal component basis with a peak-entropy ratio greater than the screening threshold as defect-related orthogonal component basis, recording the index corresponding to the defect-related orthogonal component basis, and removing the remaining orthogonal component basis with a peak-entropy ratio less than or equal to the screening threshold.

7. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, Multi-scale wavelet packet transform is performed on the initial defect representation data volume. Defects inside the metal powder injection green body are identified and located based on the local modulus maxima and cross-scale correlation criteria. This includes: performing three-level wavelet packet decomposition on the initial defect representation data volume; finding local modulus maxima points in the high-frequency wavelet coefficients of each decomposition scale; mapping the coordinates of local modulus maxima points at each decomposition scale to the original three-dimensional voxel coordinate system; calculating the spatial overlap of local modulus maxima points between adjacent decomposition scales; identifying local modulus maxima points with a spatial overlap greater than a set proportion as true defect response points; and locating the defects based on the original three-dimensional voxel coordinates.

8. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, The orthogonal decomposition of the higher-order response tensor is constructed as an iterative optimization process, including: constructing the objective optimization problem based on the tensor alternating least squares method framework, initializing the factor matrices of each dimension using the higher-order singular value decomposition method; after each round of factor matrix update, orthogonalization processing is performed on the factor matrices of each dimension so that the column vectors of the updated factor matrices of each dimension satisfy the orthogonal constraint.

9. The non-destructive testing method for metal powder injection molding green bodies according to claim 1, characterized in that, The initial defect representation data volume is generated by reconstructing the defect-related orthogonal component basis, including: extracting the component indices of the defect-related orthogonal component basis in the excitation parameter dimension; retaining the core sub-tensor elements in the core tensor corresponding to the component indices, setting the core sub-tensor elements corresponding to the unretained component indices to zero, and generating the processed core tensor; performing tensor shrinkage and pattern multiplication operations on the processed core tensor and the orthogonal component basis in each spatial dimension to generate the initial defect representation data volume.

10. A non-destructive testing system for metal powder injection molding green bodies, characterized in that, include: A processor and a memory, wherein the memory stores computer program instructions that, when executed by the processor, implement a non-destructive testing method for metal powder injection green bodies according to any one of claims 1-9.

Images