Inductor core classification screening method based on pattern recognition

Abstract

The invention provides an inductance magnetic core classification screening method based on pattern recognition, and the method comprises the steps: obtaining original image data containing pore distribution and defect forms, and obtaining an image set covering different scale ranges through a pre-established multi-layer resolution scanning mode; performing preliminary segmentation on the multi-scale characteristics of the pores according to the image set, distinguishing a pore region and a non-pore region in the image, and determining basic contour information of the pores; performing feature extraction on the topological structure data of the pore distribution, analyzing the connection strength and the distribution density of key nodes in the pore network, and determining potential influence factors of the pore network on the magnetic core performance; and according to the determined potential influence factor, in combination with the multi-scale distribution feature of the pore network, adjusting the initial magnetic conductivity prediction feature vector to obtain an adjusted magnetic conductivity prediction feature vector.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to an inductance magnetic core classification and screening method based on pattern recognition. BACKGROUND

[0002] In the field of modern industrial manufacturing, as a core component, the performance of inductance magnetic cores is directly related to the stability and efficiency of electronic devices, and has strategic significance. The microstructure inside the magnetic core, especially the defects such as pores, cannot be ignored in terms of its influence on magnetic performance. Therefore, how to accurately evaluate these defects and predict performance has become a key issue to improve product quality and reliability. However, existing methods often seem inadequate when faced with the complexity of internal defects of the magnetic core. Many traditional techniques rely more on a single detection method or a fixed analysis model, making it difficult to adapt to the diversity of defect morphology and distribution, especially when faced with the variability and complex network structure of pore distribution, lacking sufficient flexibility and depth of analysis. This limitation leads to an incomplete capture of defect features, which in turn affects the reliability of performance prediction. In-depth, the core challenges in this field mainly focus on two interrelated technical factors. On the one hand, the multi-scale characteristics of pore distribution make the size and morphology of defects extremely uncertain, and traditional image processing methods are difficult to accurately distinguish and quantify these differences, resulting in limited precision in feature extraction. On the other hand, this multi-scale distribution further affects the connectivity between pores, forming a complex network structure, and existing algorithms often fail to adjust strategies in a timely manner when dealing with this dynamic network relationship, causing evaluation bias of the overall impact of defects. These two factors interact with each other, making it difficult to accurately predict the impact of defects on the performance of the magnetic core. Therefore, how to dynamically adjust the image segmentation and feature extraction strategy under the influence of multi-scale changes in pore distribution and complex network structure of connectivity, and improve the accuracy of magnetic core performance level prediction through this adjustment, has become a key problem that needs to be overcome in this research. SUMMARY

[0003] The present application provides an inductance magnetic core classification and screening method based on pattern recognition, mainly comprising:

[0004] Obtain the original image data containing the distribution of pores and the morphology of defects, and use the pre-established multi-layer resolution scanning method to obtain a set of images covering different scale ranges;

[0005] According to the image set, the multi-scale characteristics of the pores are preliminarily segmented, and the pore area and non-pore area in the image are distinguished to determine the basic outline information of the pores;

[0006] For the determined basic contour information of the pores, the spatial position relationship between the pores is obtained, and a pore connectivity network model is constructed by using a graph theory analysis method in combination with the characteristics of connectivity complexity and network structure variation, to obtain topological structure data of the pore distribution;

[0007] The topological structure data of the pore distribution is subjected to feature extraction, the connection strength and distribution density of key nodes in the pore network are analyzed, and potential influencing factors of the pore network on the performance of the magnetic core are determined;

[0008] According to the determined potential influencing factors, in combination with the multi-scale distribution characteristics of the pore network, an initial permeability prediction feature vector is adjusted to obtain an adjusted permeability prediction feature vector;

[0009] Based on the adjusted permeability prediction feature vector, magnetic core performance pattern recognition is performed, and the magnetic core performance level is preliminarily evaluated according to the recognized magnetic core performance pattern. If the value of a certain dimension of the feature vector deviates from a preset range, the value is calibrated, and an intermediate result of performance prediction is determined. If the value of a certain dimension of the feature vector does not deviate from the preset range, the current prediction result is directly output as the intermediate result;

[0010] By multi-dimensional comparison of the determined intermediate result, the correlation mapping relationship between the pore distribution and the performance level is obtained, and in combination with the internal microstructure of the magnetic core, a performance level prediction value is generated;

[0011] The deviation information of the performance level prediction value and the actual performance data is obtained. If the deviation exceeds a preset range, the threshold parameter optimization is performed from the image segmentation link to determine an accurate segmentation result, to obtain an updated performance prediction value, and to obtain a magnetic core performance level classification result.

[0012] Further, the original image data containing the pore distribution and the defect morphology is obtained, and a multi-layer resolution scanning mode is used to obtain an image set covering different scale ranges, including:

[0013] The magnetic core is scanned layer by layer to obtain a three-dimensional voxel data set containing pore position coordinates, pore diameters and defect contours. The three-dimensional voxel data set is labeled, the pore region, the defect region and the normal material region are distinguished based on the gray value difference, the connected component analysis is performed, and the volume parameters of the independent pores and the geometric morphology parameters of the defects are determined. Based on the volume parameters and the geometric morphology parameters, the space distribution matrix is constructed by assigning values in the three-dimensional space coordinate system, the ratio of the total pore volume in the partition to the partition volume is calculated as the pore density distribution, the ratio of the total defect area to the partition volume is calculated as the defect aggregation degree, and the image set is generated.

[0014] Further, the preliminary segmentation of the stomata multi-scale characteristics according to the image set distinguishes the stomata region and the non-stomata region in the image, and determines the basic contour information of the stomata, including:

[0015] The gray scale histogram is extracted from the image set, the initial segmentation threshold is calculated based on the bimodal distribution characteristics, and a multi-scale threshold set is generated; binaryzation is performed on each image, pixels below the multi-scale threshold set are marked as stomata candidate regions, and pixels above the threshold are marked as non-stomata regions; morphological opening operation is applied to remove noise points, and a binary image is generated; the eight-connected region labeling algorithm is used on the binary image to label the stomata candidate regions, the area and perimeter parameters of the connected regions are calculated, the connected regions with an area less than a preset threshold are removed, and a labeled image of the effective stomata region is generated; the boundary pixel coordinate sequence of the effective stomata region is extracted, a closed contour point set is formed through boundary tracking method, and the basic contour information of each stomata is determined.

[0016] Further, for the determined basic contour information of the stomata, the spatial position relationship between the stomata is obtained, the connectivity complexity and the network structure change characteristics are combined, the graph theory analysis method is used to construct a stomata connectivity network model, and the topological structure data of the stomata distribution is obtained, including:

[0017] The centroid coordinates of each stomata in the stomata contour information are calculated, and a stomata pair distance matrix is generated based on the Euclidean distance formula; based on the distance matrix and the connectivity determination threshold, a binary adjacency matrix is constructed, and a value of 1 indicates stomata connectivity; a graph structure is constructed based on the binary adjacency matrix, the stomata are nodes, the positions with a value of 1 are edges, the nodes are traversed through a depth-first search algorithm, the connected subgraphs and node sets are labeled, and the topological structure data of the stomata distribution is generated.

[0018] Further, the topological structure data of the stomata distribution is extracted, the connection strength and distribution density of the key nodes in the stomata network are analyzed, the potential impact factor of the stomata network on the magnetic core performance is determined, including:

[0019] Based on the topological structure data, the connectivity of each stomata node is calculated, and the nodes with a connectivity greater than a preset multiple of the average value are marked as high-connectivity nodes; for the high-connectivity nodes, the perpendicular distance from the centroid to the center line of the magnetic flux path is calculated, and the nodes located in the magnetic flux dense area are determined as key nodes; the product of the connection strength value and the distribution density value of the key nodes is calculated as the potential impact factor.

[0020] Further, according to the determined potential impact factor, the multi-scale distribution characteristics of the stomata network are combined, the initial permeability prediction feature vector is adjusted, and an adjusted permeability prediction feature vector is obtained, including:

[0021] Based on the potential influencing factors and material parameters, an initial magnetic permeability prediction feature vector is constructed, the potential influencing factors are mapped to vector components, an extended feature vector is generated, the local pore density, the number of connected domains and the distribution uniformity of the multi-scale distribution features of the pore network are calculated and added as new components to generate a multi-scale feature vector, and based on the weight coefficients of the potential influencing factors, the multi-scale feature vector is normalized to generate an adjusted magnetic permeability prediction feature vector.

[0022] Further, based on the adjusted magnetic permeability prediction feature vector, core performance pattern recognition is performed, the core performance level is preliminarily evaluated according to the recognized core performance pattern, and an intermediate result of performance prediction is determined, including:

[0023] Taking the adjusted magnetic permeability prediction feature vector as input, the performance categories are divided by support vector machine algorithm to generate preliminary evaluation results of superior products, qualified products and unqualified products, and the dimension values of the adjusted magnetic permeability prediction feature vector are checked and compared with the normal range threshold, the dimensions deviating from the threshold are calibrated by linear interpolation to generate a calibrated feature vector as an intermediate result.

[0024] Further, by comparing the determined intermediate result in multiple dimensions, the correlation mapping relationship between the pore distribution and the performance level is obtained, the performance level prediction value is generated in combination with the microstructure inside the core, including:

[0025] The correlation between the pore distribution feature parameters and the historical performance level is calculated by Pearson correlation coefficient to generate a correlation mapping matrix, the grain size, the grain boundary density and the phase composition ratio of the image set are combined to construct a performance prediction weight matrix, a comprehensive score is calculated, and a performance level prediction value is generated based on the score interval.

[0026] Further, the deviation information of the performance level prediction value and the actual performance data is obtained to generate a core performance level classification result, including:

[0027] The deviation is calculated by comparing the performance level prediction value and the actual performance level, if the deviation exceeds the error tolerance threshold, the pore connectivity network model and the adjusted magnetic permeability prediction feature vector are updated to generate a corrected performance level prediction value, until the deviation meets the tolerance, and the core performance level classification result is output.

[0028] The technical scheme provided by the embodiment of the application can include the following beneficial effects:

[0029] The application discloses an inductance magnetic core classification and screening method based on pattern recognition, and aims at a service scene problem that the performance of a magnetic core is influenced by the complexity of air hole distribution and connectivity. A multi-scale image set is acquired through high-resolution imaging, a network model of air hole connectivity is constructed by combining threshold segmentation and graph analysis, the connection strength and distribution density of key nodes are extracted, and the potential influence of the air hole network on the magnetic flux path and permeability is revealed. Based on this, a performance pattern recognition model is established by adjusting the permeability prediction feature vector, and a performance grade prediction value is generated through multidimensional comparison. If the prediction deviation exceeds the standard, the image segmentation threshold is backtracked and optimized, the network model and feature vector are updated, and an accurate prediction result is obtained through iteration. Through multi-scale air hole characteristic analysis and a dynamic calibration mechanism, the application realizes high-precision prediction of the performance of a magnetic core, and provides efficient technical support for the design and quality control of a magnetic core material. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 A flowchart of the inductance magnetic core classification and screening method based on pattern recognition. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical scheme and advantages of the application clearer, the application is described in detail below with reference to the drawings and specific embodiments.

[0032] As Figure 1 the inductance magnetic core classification and screening method based on pattern recognition specifically can include:

[0033] In step S101, original image data containing air hole distribution and defect morphology is acquired, and a multi-layer resolution scanning mode is used to obtain an image set covering different scale ranges.

[0034] The X-ray tomography equipment is used to scan the magnetic core layer by layer. During the scanning process, data is collected according to preset resolution parameters. Each layer of scanning data contains the pore position coordinates, pore diameter value and defect profile information of the layer. After the scanning is completed, a three-dimensional voxel data set containing spatial position information is obtained. For the obtained three-dimensional voxel data set, an image segmentation algorithm is used to label each voxel. According to the gray value difference, the pore area, the defect area and the normal material area are distinguished. The labeled voxel data is subjected to connected component analysis to obtain the volume parameters of each independent pore and the geometric morphological parameters of the defect. According to the volume parameters of each independent pore and the geometric morphological parameters of the defect, the corresponding positions in the three-dimensional coordinate system are assigned. The pore position is assigned to the pore volume value, and the defect position is assigned to the defect area value. A spatial distribution matrix is constructed. By partitioning the spatial distribution matrix according to different spatial scales, the ratio of the total pore volume in each partition to the partition volume is calculated as the pore density distribution, and the ratio of the total defect area to the partition volume is calculated as the defect aggregation degree, to obtain an image set of the microstructure inside the magnetic core at different scales.

[0035] Specifically, the application of X-ray tomography technology in magnetic core detection is based on the difference in penetration ability of different density materials. When X-rays pass through the magnetic core material, the normal magnetic material area will produce strong attenuation, while the pore and defect areas have lower density, and the X-ray attenuation degree is significantly weakened. This attenuation difference forms different gray values on the detector, thereby realizing the imaging of the internal structure.

[0036] In one possible implementation, the scanning equipment uses cone beam CT technology. The magnetic core is rotated 360 degrees between the X-ray source and the detector by rotating the scanning table. One projection image is collected every 0.5 degrees of rotation. A total of 720 projection data are obtained after completing a complete rotation. These projection data are processed by a filtered back projection reconstruction algorithm to form a cross-sectional image. By continuous scanning in the vertical direction, a complete three-dimensional voxel data set is constructed.

[0037] It should be noted that the image segmentation algorithm uses a method combining threshold segmentation and region growing when processing three-dimensional voxel data. First, the initial threshold is determined according to the gray histogram, and the voxels are divided into three categories: background, normal material and defect. For the region with fuzzy boundary, the region growing algorithm is used to start from the seed point and expand according to the gray similarity of adjacent voxels until the stop condition is met. The connected component analysis aggregates the voxels that are adjacent in space and have the same attributes into independent three-dimensional objects according to the 26-neighbor connection criterion.

[0038] Specifically, in the construction of the spatial distribution matrix, each matrix element corresponds to a spatial unit inside the magnetic core. When there is a pore in the unit, the element value is set to the pore volume value; when there is a defect, the element value is set to the defect surface area value; and the normal material region is assigned a value of zero. This quantitative representation method makes subsequent statistical analysis possible.

[0039] In an embodiment, the multi-scale analysis is implemented using a sliding window method. Three different sizes of cubic windows are set, 2mm x 2mm x 2mm, 5mm x 5mm x 5mm and 10mm x 10mm x 10mm, corresponding to the micro, meso and macro scales. The window slides in the spatial distribution matrix, and each time the moving step is half the size of the window, ensuring that adjacent windows have 50% overlap. In each window, the local pore density is obtained by counting the ratio of the total pore volume to the window volume, and the local defect aggregation degree is obtained by counting the ratio of the total defect area to the window volume. The advantage of this multi-scale analysis method is that it can capture structural features at different levels simultaneously. The micro scale reveals the distribution of individual pores and defects, the meso scale reflects the structural uniformity of the local region, and the macro scale shows the overall quality distribution trend.

[0040] Step S102, according to the image set, the preliminary segmentation of the pore multi-scale characteristics is carried out, and the pore region and the non-pore region in the image are distinguished to determine the basic outline information of the pore.

[0041] The gray level histogram of each image is extracted from the image set. According to the bimodal distribution characteristics in the histogram, the initial segmentation threshold is calculated by using the maximum inter-class variance method. The corresponding threshold values are calculated for images of different scales to obtain a multi-scale threshold set. According to the multi-scale threshold set, each image is binarized. The pixels with a gray value lower than the threshold value are marked as pore candidate regions, and the pixels with a gray value higher than the threshold value are marked as non-pore regions. Small noise points are removed by morphological opening operation to obtain a binary image after preliminary segmentation. For the binary image after preliminary segmentation, the eight-connected domain labeling algorithm is used to independently label the pore candidate regions. The area of each connected domain is obtained by counting the number of pixels in each connected domain, and the perimeter parameter is obtained by calculating the number of boundary pixels of the connected domain. If the area of the connected domain is less than the area threshold value determined according to the image resolution and the minimum detectable pore size, the connected domain is removed from the pore candidate region to obtain a labeled image of the effective pore region. According to the labeled image of the effective pore region, the boundary pixel coordinate sequence of each pore is extracted. The pixel position is sequentially recorded along the pore boundary by using the Moore neighborhood boundary tracking method. From any point on the boundary, the adjacent pixels are traversed in a clockwise direction to form a closed contour point set, and the basic outline information of each pore is determined.

[0042] Specifically, the bimodal distribution of the gray scale histogram is an important basis for the segmentation of the pore image. In the magnetic core image, the pixel gray scale values of the normal material region are concentrated in a higher value range, while the pore region shows a lower gray scale value due to the strong X-ray penetration. This difference is manifested as two obvious peaks in the histogram, with the low gray scale peak corresponding to the pores and the high gray scale peak corresponding to the normal material.

[0043] Specifically, the maximum between-class variance method calculates the between-class variance of two classes of pixels under each possible gray scale threshold by traversing all possible gray scale thresholds. When the between-class variance reaches the maximum, it indicates that this threshold can best distinguish the two classes of pixels. For an image containing 256 gray scales, the algorithm tests from gray scale 1 to 255 one by one, each time divides the image into two classes, counts the number of pixels and the average gray scale of each class, calculates the between-class variance, and finally selects the gray scale value that makes the variance the maximum as the segmentation threshold.

[0044] In an embodiment, the implementation process of the morphological opening operation includes two basic operations of erosion and dilation. The erosion operation uses a 3x3 structure element to shrink the pore candidate region in the binary image and eliminate isolated noise and small protrusions. The subsequent dilation operation uses the same structure element to restore the size of the main pore region. This combination of erosion followed by dilation not only removes noise but also maintains the basic morphological features of the pores. The core of the eight-connected domain labeling algorithm is to identify spatially adjacent pixels of the same type. The algorithm starts from the top left corner of the image and scans each row. When an unmarked pore pixel is encountered, a new label value is assigned, and the eight neighborhood positions of the pixel are checked. If there is an unmarked pixel of the same type in the neighborhood, the same label value is assigned, and the expansion is continued through recursion or queue method until all pixels in the connected region are labeled. The area calculation is achieved by counting the number of pixels corresponding to each label value, and the perimeter is obtained by counting the boundary pixels.

[0045] It should be noted that the determination of the area threshold depends on the spatial resolution of the imaging system. If each pixel represents an actual size of 10 microns and the minimum detectable pore diameter is set to 50 microns, the corresponding minimum area is 20 pixels. Connected domains below this threshold are considered as noise or artifacts and are removed. The implementation of the Moore neighborhood boundary tracking method starts with finding the starting boundary point of the pore region, usually the top left corner boundary pixel. From this point, the algorithm checks the eight neighborhood directions in a specific search order, giving priority to the next boundary pixel in the clockwise direction. After moving to a new boundary pixel, the search starting direction is updated to ensure progress along the outside of the pore contour. When the tracking path returns to the starting point, the extraction of a closed contour is completed.

[0046] Step S103, for the determined basic contour information of the pores, the spatial position relationship between the pores is obtained, and a pore connectivity network model is constructed by using a graph theory analysis method in combination with the characteristics of connectivity complexity and network structure change, to obtain topological structure data of the pore distribution.

[0047] According to the basic contour information of the pores, the centroid coordinates of each pore are calculated, the centroid position is obtained by averaging all pixel coordinates in the contour, the spatial distance between the centroids of any two pores is calculated by using the Euclidean distance formula, and a distance matrix containing distance values of all pore pairs is constructed. Based on the distance matrix, a connectivity determination threshold is determined according to the material physical properties and the average size of the pores, if the distance between two pores is less than the threshold, the corresponding position in the adjacency matrix is assigned a value of 1, indicating that the two pores have a connectivity relationship, otherwise the value is 0, and a binary adjacency matrix describing the connectivity relationship of the pores is obtained. For the graph structure constructed based on the binary adjacency matrix, each pore is a node in the graph, the position with a value of 1 in the adjacency matrix corresponds to the edge between the nodes, the connected nodes are traversed starting from any node by using a depth-first search algorithm, the visited nodes are marked, the number of independent connected subgraphs and the node set contained in each subgraph are identified, and the topological structure data of the pore distribution is obtained.

[0048] Specifically, the calculation of the centroid coordinates is based on the position information of all pixel points inside the contour.

[0049] Specifically, for a closed pore contour, each pixel inside the contour has corresponding horizontal and vertical coordinate values. The horizontal coordinate of the centroid is obtained by adding the horizontal coordinates of all internal pixels and dividing by the total number of pixels; the vertical coordinate is calculated in the same way. This averaging process makes the centroid position represent the spatial position of the entire pore, even if the pore shape is irregular, the centroid can still reflect its geometric center. As the most intuitive spatial distance measurement method, the Euclidean distance is obtained by taking the square root of the sum of the coordinate differences of two points in a two-dimensional plane. When dealing with hundreds of pores, the distance matrix has a symmetric property, and the diagonal elements are zero, indicating the distance of the pore to itself. Each element in the matrix records the exact distance value between the corresponding pore pairs, which directly reflects the distribution density of the pores in the material.

[0050] In one possible implementation, the determination of the connectivity determination threshold needs to consider multiple factors. The physical properties of the material determine the critical distance between pores. When the distance between two pores is less than the material's fracture toughness scale, the stress concentration effect will interact with each other. At the same time, the average size of the pores is also an important reference. Usually, the threshold is set to be in the range of 1.5 to 2 times the average pore diameter. This setting can identify the true connectivity relationship and avoid misjudging independent pores far apart as connected. The construction process of the binary adjacency matrix converts continuous distance values into discrete connectivity relationships. The 1 in the matrix indicates that there is an edge connection between two pore nodes, and 0 indicates no connection. This binary representation simplifies subsequent graph structure analysis, and complex spatial relationships are converted into clear topological connections.

[0051] It should be noted that the depth-first search algorithm uses a recursive strategy when traversing the graph structure. Starting from an unvisited node, the algorithm first marks the node as visited, and then checks all its adjacent nodes. For each unvisited adjacent node, the same visiting process is recursively performed. When all adjacent nodes of a node have been visited, the algorithm backtracks to the previous layer to continue searching. This depth-first traversal method can completely explore each connected subgraph. The connected subgraphs identified by the depth-first search reveal the clustering characteristics of the pore distribution. Independent connected subgraphs represent spatially isolated pore groups, and the node set within the subgraph reflects the degree of pore clustering in the local area. Topological structure data not only contains the number of connected subgraphs, but also records the node composition, internal connection density of each subgraph, and the shortest distance between subgraphs. These information together constitute a complete topological description of the pore distribution.

[0052] In step S104, the topological structure data of the pore distribution is extracted, the connection strength and distribution density of the key nodes in the pore network are analyzed, and the potential influencing factors of the pore network on the performance of the magnetic core are determined.

[0053] According to the binary adjacency matrix in the topological structure data, the connectivity of each pore node, i.e. the number of other nodes directly connected to the node, is counted, the average value of the connectivity of all nodes is calculated, and if the connectivity of a node is greater than the average value multiplied by a preset multiple, the node is marked as a high-connectivity node to obtain a high-connectivity node set. For each node in the high-connectivity node set, the main conduction path of the magnetic flux in the material is determined according to the geometric structure of the magnetic core, the vertical distance from the node centroid to the center line of the conduction path is calculated, and if the distance is less than the influence range threshold determined according to the size of the magnetic core and the magnetic permeability of the material, it is determined that the node is located in a magnetic flux dense area and affects the conduction of the magnetic flux, and is determined as a key node to obtain a key node list. Based on the key node list, the connection strength value of each key node is calculated, the local connection density is obtained by counting the number of edges connected to the node divided by the total number of edges in the subgraph where the node is located, and the distribution density value of the key node in a unit volume is calculated, and the product of the connection strength value and the distribution density value is taken as the potential influence factor of the node on the performance of the magnetic core.

[0054] Specifically, the connectivity, as a basic concept in graph theory, reflects the local importance of network nodes. In the pore network, the connectivity of a node is equal to the number of edges directly connected to it.

[0055] Specifically, by traversing each row of the adjacency matrix, the number of elements with a value of 1 can be obtained to obtain the connectivity of the corresponding node. When a certain pore has a communication relationship with the surrounding 5 pores, and the average connectivity of the network is 2, the connectivity of the pore is 2.5 times the average value, indicating that it forms a dense communication structure in the local area. The identification of high-connectivity nodes uses statistical methods. After calculating the arithmetic mean of the connectivity of all nodes, a multiple threshold is set, usually 1.5 to 2 times. This relative threshold setting method can adapt to pore networks of different densities and avoid misjudgment that may be caused by absolute numerical judgment. High-connectivity nodes are often located at the center of the pore aggregation area and play the role of a connection hub.

[0056] In one possible implementation, the determination of the main conduction path of the magnetic flux is based on the geometric structure of the magnetic core and the magnetic circuit principle. For a ring-shaped magnetic core, the magnetic flux circulates along the center line of the ring; for a bar-shaped magnetic core, the magnetic flux conducts along the length direction. The conduction path can be determined in advance by finite element simulation, or approximated according to the geometric center line of the cross-sectional shape of the magnetic core. This path represents the area with the highest magnetic flux density. The calculation of the vertical distance involves spatial geometric relationships. For nodes and paths in three-dimensional space, the shortest distance from a point to a straight line is calculated by vector projection method. First, determine the point on the path closest to the node, and then calculate the Euclidean distance between the two points. This distance value directly reflects the degree of interference of the pore on the conduction of the magnetic flux.

[0057] It should be noted that the determination of the influence range threshold comprehensively considers the size and material properties of the magnetic core. The higher the permeability of the material, the easier the magnetic flux bypasses the obstacle, and the relatively smaller the influence range. In a typical ferrite magnetic core, the influence range is about one-tenth of the cross-sectional size of the magnetic core. The pores within this range will significantly change the local magnetic flux distribution. The calculation of the connection strength value reflects the relative importance of the node in the local network. By counting the number of connection edges of the key node and dividing by the total number of edges in the connected subgraph where the node is located, a ratio between 0 and 1 is obtained. The larger this ratio, the more densely connected the node is in the local network. The distribution density is obtained by dividing the space grid and counting the number of key nodes in each grid. The potential influence factor, as a comprehensive index, combines the structural characteristics and spatial distribution. The connection strength reflects the topological properties of the pore network, and the distribution density reflects the degree of spatial aggregation. The product of the two can comprehensively evaluate the influence of the pores on the performance of the magnetic core. The larger the value of the influence factor, the more serious the obstruction of the magnetic flux conduction, and the more obvious the decrease in the effective permeability of the magnetic core.

[0058] Step S105, according to the determined potential influence factor, combining the multi-scale distribution characteristics of the pore network, adjusting the initial permeability prediction feature vector to obtain the adjusted permeability prediction feature vector.

[0059] According to the potential influence factor value and the physical parameters of the magnetic core material, an initial permeability prediction feature vector is constructed, which includes the basic permeability value obtained from the material database, the ratio of the total volume of the pores to the total volume of the material obtained by image analysis, and the average pore size obtained by dividing the total volume of all pores by the number of pores. The influence factor of each key node is mapped to the corresponding vector component according to its spatial position to obtain an extended feature vector containing the distribution information of the influence factor. For the extended feature vector, different scale feature parameters are calculated from the obtained pore distribution data. The local pore density at the microscale is obtained by counting the number of pores per unit volume. The number of connected domains at the mesoscale is obtained by counting the number of independent connected subgraphs. The overall distribution uniformity at the macroscale is obtained by calculating the reciprocal of the standard deviation of the pore spatial distribution. These multi-scale feature parameters are added as new vector components to the extended feature vector to form a multi-scale feature vector. Based on the multi-scale feature vector, the weight coefficients of the corresponding components are determined according to the numerical value of the influence factor. The normalized weight value is obtained by dividing the influence factor by the sum of all influence factors. The adjusted permeability prediction feature vector is obtained by multiplying the vector components by the corresponding weight coefficients and then normalizing the component values by dividing by the square root of the square sum of all components.

[0060] Specifically, the construction of the permeability prediction feature vector is based on the principle of multi-dimensional information fusion. The material-based permeability is usually obtained from a material database. For common ferrite materials, the initial relative permeability value is between 1000 and 5000. This value represents the material's magnetic performance benchmark in an ideal defect-free state.

[0061] Specifically, the calculation of the total volume fraction of pores is achieved through image analysis. In the segmented binary image, the number of all pore pixels is counted and then divided by the total number of image pixels to obtain the two-dimensional projection area ratio. For three-dimensional cases, the volume fraction is obtained by cumulative statistics of multi-layer slice images. When the pore volume accounts for 5% of the total material volume, the permeability will decrease significantly. The determination of the average pore size involves statistical methods. The volumes of all identified pores are accumulated and then divided by the total number of pores to obtain the average value. This parameter reflects the typical size of defects, and large-size pores have a more obvious hindering effect on magnetic flux.

[0062] In one possible implementation, the mapping of the influence factor to the vector component adopts a spatial coding method. The magnetic core space is divided into several regions, each corresponding to the position of a component of the vector. When a certain key node is located in a specific region, its influence factor value is assigned to the corresponding vector component. This mapping preserves the spatial distribution information. The extraction of multi-scale feature parameters embodies the idea of hierarchical analysis. Microscale focuses on local details, and the local density value is obtained by counting the number of pores in a small window and then dividing by the window volume. Mesoscale reflects the structural organization, and the number of connected domains is directly obtained from the graph theory analysis result, representing the aggregation degree of pores. Macro-scale evaluates the overall uniformity by calculating the standard deviation of all local density values and then taking the reciprocal. The smaller the standard deviation, the more uniform the distribution, and the larger the reciprocal value.

[0063] It should be noted that the determination of the weight coefficient adopts normalization processing to ensure numerical stability. Each influence factor is divided by the sum of all influence factors to obtain a weight value between 0 and 1. The component with a large weight dominates in the final feature vector, reflecting the importance of the region's influence on permeability. The purpose of vector normalization is to eliminate the dimensional differences between different dimensions. By calculating the square root of the sum of squares of all components, i.e., the Euclidean norm of the vector, and then dividing each component by this norm value. The components of the normalized feature vector are within the same numerical range, facilitating subsequent machine learning model processing. This multi-level, multi-scale feature construction method organically combines physical parameters, spatial distribution, and topological structure information to form a comprehensive feature representation reflecting the influence of pores on permeability. The adjusted feature vector not only contains the number information of defects but also encodes their spatial distribution patterns and network connectivity characteristics.

[0064] In step S106, the magnetic core performance mode is identified based on the adjusted permeability prediction feature vector, the magnetic core performance level is preliminarily evaluated according to the identified magnetic core performance mode, if a dimension value of the feature vector deviates from the preset range, the dimension value is calibrated, the intermediate result of performance prediction is determined, and if a dimension value of the feature vector does not deviate from the preset range, the current prediction result is directly output as the intermediate result.

[0065] The adjusted permeability prediction feature vector is used as input to establish a magnetic core performance classification boundary through a support vector machine algorithm. According to the position of the feature vector in the multi-dimensional space, the performance category to which it belongs is determined, and the magnetic core performance is divided into three levels: superior product, qualified product and unqualified product, to obtain a preliminary evaluation result of the magnetic core performance level. According to the preliminary evaluation result, each dimension value of the feature vector is checked one by one, and each dimension value is compared with the upper and lower threshold values of the normal range of each dimension obtained according to the historical qualified product statistics. If a dimension value is less than the lower threshold value or greater than the upper threshold value, it is determined that the dimension deviates from the preset range, and the index and deviation degree value of the deviated dimension are recorded. For the dimensions deviating from the preset range, linear interpolation method is used for calibration. The difference between the deviation value and the corresponding threshold value is calculated, and then divided by the threshold value to obtain the deviation rate. The calibration coefficient is determined according to the deviation rate, which is 1 minus the absolute value of the deviation rate. The original dimension value is multiplied by the calibration coefficient to obtain the calibrated value, which replaces the corresponding dimension in the original feature vector to form the calibrated feature vector, which is used as the intermediate result of performance prediction. If all dimension values are within the preset range, the current feature vector is directly used as the intermediate result.

[0066] Specifically, the application of the support vector machine algorithm in the magnetic core performance classification is based on the principle of finding the optimal classification hyperplane. The algorithm maps the feature vector to a high-dimensional space and constructs a decision boundary with the maximum interval between samples of different performance levels. For magnetic core performance evaluation, each dimension of the feature vector represents a different physical characteristic, such as permeability, pore distribution density, connectivity index, etc.

[0067] Specifically, the algorithm training process uses historical magnetic core sample data, each sample has a clear quality label. Superior products usually have a combination of high permeability, low pore density and uniform distribution; the feature values of qualified products are in the middle range; unqualified products show abnormal deviation in some dimensions. By learning these patterns, the algorithm establishes a classification boundary in multi-dimensional space.

[0068] In one possible implementation, when a new feature vector is input, the algorithm calculates the distance of the vector to each classification boundary. The nearest boundary determines the performance category of the magnetic core. This judgment method based on geometric distance has good interpretability and stability. The determination of the threshold range uses statistical methods. By collecting a large amount of feature data of qualified products, the mean and standard deviation of each dimension are calculated. The normal range is usually set as the interval of the mean plus or minus twice the standard deviation, which can cover more than 95% of normal samples.

[0069] For example, if the historical mean of the pore density dimension is 0.03 and the standard deviation is 0.01, the normal range of this dimension is 0.01 to 0.05.

[0070] It should be noted that the quantification of the deviation degree is crucial for subsequent calibration. When the value of a certain dimension is 0.07, which exceeds the upper limit of 0.05, the deviation value is 0.02. The deviation rate is calculated as 0.02 divided by 0.05, which is 0.4, indicating that it deviates from the normal upper limit by 40%. This relative measurement method makes the deviation of different dimensions comparable. The core of the linear interpolation calibration method is gradual adjustment. The calculation of the calibration coefficient follows the principle of conservation to avoid over-correction. When the deviation rate is 0.4, the calibration coefficient is 1 minus 0.4, i.e. 0.6. The original value 0.07 multiplied by 0.6 gives the calibrated value 0.042, which is returned to the normal range. This method not only corrects the abnormal value, but also retains part of the original measurement information. The selective execution of the calibration process reflects the flexibility of the algorithm. Only the deviated dimensions need to be calibrated, and the normal dimensions remain unchanged. This differential treatment ensures the overall integrity of the feature vector and avoids unnecessary interference with normal data. Through this multi-level evaluation and calibration mechanism, the performance level of the magnetic core can be accurately identified, and abnormal features can be reasonably corrected.

[0071] Step S107, by comparing the determined intermediate results in multiple dimensions, the correlation mapping relationship between the pore distribution and the performance level is obtained, and the performance level prediction value is generated in combination with the internal microstructure of the magnetic core.

[0072] According to the determined intermediate results, the pore distribution feature parameters are extracted, including pore distribution density, connectivity and uniformity of spatial distribution. At the same time, the corresponding feature parameters of magnetic cores with different performance levels are obtained from the historical database. The correlation strength between each feature parameter and the performance level is determined by calculating the Pearson correlation coefficient, and the correlation mapping matrix is obtained. Based on the correlation mapping matrix, in combination with the internal microstructure parameters of the magnetic core, including the average grain size, grain boundary density distribution and phase composition ratio represented by the proportion of different gray area obtained by pre-analysis of image texture, a performance prediction weight matrix is constructed, wherein

[0073] W ij = α · Gi + β · B i + γ · C i

[0074] , W ij represents the weight value of the i-th row and j-th column in the performance prediction weight matrix, G i represents the grain size parameter of the i-th region, B i represents the grain boundary density distribution parameter of the i-th region, C i represents the phase composition ratio parameter represented by the proportion of different gray regions in the i-th region, α, β, γ respectively represent the weight coefficients corresponding to the grain size, grain boundary density and phase composition ratio; after multiplying each dimension value in the intermediate result by the corresponding weight and accumulating, the comprehensive score value is obtained, which can be calculated by the following formula:

[0075]

[0076] , S represents the comprehensive score value, D i represents the value of the i-th dimension in the intermediate result, W i represents the weight value corresponding to the i-th dimension, n represents the total number of dimensions, and the formula calculates the comprehensive score by multiplying each dimension value by the corresponding weight and accumulating, realizing quantitative prediction and evaluation of the performance of the magnetic core; according to the score intervals of superior products, qualified products and unqualified products determined by historical qualified product statistics, the comprehensive score is converted into a specific performance grade prediction value.

[0077] Specifically, Pearson correlation coefficient as a statistical indicator to measure the degree of linear correlation between two variables plays a key role in the evaluation of magnetic core performance. The value of this coefficient ranges from -1 to 1, close to 1 indicates strong positive correlation, close to -1 indicates strong negative correlation, and close to 0 indicates no linear correlation. In the analysis of magnetic core quality, the pore density and performance grade usually show a negative correlation, with a correlation coefficient of about -0.7 to -0.9, meaning that the higher the pore density, the lower the performance grade.

[0078] Specifically, the calculation process involves the covariance of two groups of data divided by the product of their standard deviations. For the correlation analysis of pore density and performance grade, a large number of historical sample data need to be collected. Each sample contains the measured pore density value and the corresponding performance grade score. Through statistical calculation, the influence degree of different characteristic parameters on performance can be quantified.

[0079] In one possible implementation, the construction of the correlation mapping matrix is based on the comprehensive analysis of multiple characteristic dimensions. In addition to the pore density, the connectivity reflects the degree of networking of defects, and the spatial distribution uniformity embodies the aggregation characteristics of defects. The correlation coefficients of these parameters with the performance level constitute a multi-dimensional matrix, and each element represents the influence weight of a specific feature on performance. Image texture analysis plays a fundamental role in extracting microstructure parameters. The measurement of grain size is achieved by identifying the grain boundaries in the image. Grain boundaries appear as lines of gray level discontinuity in imaging, and the grain outlines can be outlined by edge detection algorithms. The area of each grain is counted, and the average grain size is obtained through the equivalent circle diameter calculation. Smaller grain size usually means better magnetic performance.

[0080] It should be noted that the calculation of the grain boundary density distribution involves the statistics of the grain boundary length per unit area. By thinning the grain boundary image, the grain boundaries are converted into single-pixel-wide lines, and then the total length of the lines is counted. In areas with high grain boundary density, the movement of magnetic domain walls is more seriously hindered, affecting the magnetic permeability. The phase composition ratio is obtained by gray histogram analysis. Different phases exhibit different gray levels in X-ray imaging. The main phase usually occupies the main gray interval, while the second phase or impurity phase appears as other gray peaks. By counting the pixel proportion of each gray interval, the volume fraction of different phases can be estimated. The construction of the performance prediction weight matrix considers both the correlation strength and the physical mechanism. The features with larger absolute correlation coefficients obtain higher weights. At the same time, according to the magnetic theory, the pores have the most direct hindering effect on the magnetic flux, so the weight of the pore-related features is usually the highest. The grain size affects the domain structure, and the weight is second. The calculation of the comprehensive score adopts the weighted summation method. When the pore density feature value is 0.3, the weight is 0.4; the grain size feature value is 0.5, the weight is 0.3; and other features are calculated in turn, and finally a comprehensive score between 0 and 1 is obtained. According to historical statistics, 0.8 and above is excellent, 0.6 to 0.8 is qualified, and 0.6 and below is unqualified.

[0081] Step S108, the deviation information of the performance level prediction value and the actual performance data is obtained, if the deviation exceeds the preset range, the threshold parameter optimization is performed back to the image segmentation link, the accurate segmentation result is determined, the updated performance prediction value is obtained, and the magnetic core performance level classification result is obtained.

[0082] According to the obtained performance grade prediction value, the difference between the prediction value and the actual performance grade converted from the permeability value and the hysteresis loss value measured on the same batch of magnetic core samples by using the standard test method is calculated as the deviation information. If the absolute value of the deviation exceeds a preset error tolerance threshold, a parameter optimization process is started. The threshold value of image segmentation is adjusted by using the gradient descent method. The current threshold value is fine-tuned in the direction of deviation reduction. The image segmentation is performed again by using the adjusted threshold value to obtain more accurate pore profiles. The pore connectivity network is reconstructed based on the new segmentation results, and the updated topological structure data and influence factors are calculated. Based on the updated topological structure data and influence factors, the magnetic permeability prediction feature vector is regenerated. The updated feature vector is input into the performance prediction process to obtain a corrected performance grade prediction value. If the corrected deviation is within the tolerance range, the current prediction result is output as the magnetic core performance grade classification result. If the deviation still exceeds the range, the iteration optimization is continued until the accuracy requirement is met.

[0083] Specifically, the standard test method plays a role of a benchmark reference in the magnetic core performance evaluation. The permeability test usually adopts a ring sample method. The magnetic core to be tested is wound with a specific number of turns of an excitation coil and an induction coil. A sinusoidal alternating excitation signal is applied, and the amplitude and phase of the induced voltage are measured. By calculating the ratio of the excitation magnetic field strength to the induced magnetic flux density, the relative permeability value of the material is obtained.

[0084] Specifically, the measurement of hysteresis loss is based on the area calculation of the hysteresis loop. Under different frequencies and magnetic flux densities, the magnetization curve of the magnetic core is recorded to form a complete hysteresis loop. The area enclosed by the loop represents the energy loss per unit volume per unit period. The measured permeability and loss values are classified according to the industry standard. Generally, the magnetic core with a permeability higher than 4000 and a loss lower than a specific threshold is a superior product.

[0085] In a possible implementation, the calculation of the deviation information needs to consider the weights of different performance indicators. When the prediction is a superior product but the actual result is a qualified product, the grade deviation is 1. When the prediction is a superior product but the actual result is an unqualified product, the grade deviation is 2. This quantification method makes the deviation have a clear physical meaning and comparability. The application of the gradient descent method in threshold optimization is based on the error back propagation principle. The performance prediction deviation is taken as the objective function, and the image segmentation threshold is taken as the optimization variable. The direction and step size of threshold adjustment are determined by calculating the partial derivative of the deviation with respect to the threshold value. If the deviation is positive, it means that the prediction is too high, and the threshold value needs to be reduced to identify more pores. Otherwise, the threshold value needs to be increased.

[0086] It should be noted that the magnitude of threshold adjustment needs to be carefully controlled. Too large adjustment can cause algorithmic oscillation and failure to converge; too small adjustment results in slow convergence. A typical adjustment strategy is to use adaptive step size, with initial step size being large and gradually decreasing as the iteration number increases. When the deviation changes of consecutive iterations are smaller than a set value, it is considered that the optimal solution has been approached. The process of reconstructing the pore connectivity network fully utilizes the optimized segmentation results. The new threshold value can more accurately identify the pore boundaries, especially for the fuzzy areas with gray values near the boundaries. These areas often contain small connected channels, which have an important influence on magnetic flux conduction. More accurate segmentation makes the network topology closer to the real situation. The convergence judgment of iterative optimization involves multiple aspects. In addition to whether the deviation is within the tolerance range, the iteration number limit also needs to be considered to avoid infinite loop. When the maximum iteration number is reached and still not converged, the current optimal result is output and labeled with confidence. This mechanism ensures the practicality and robustness of the algorithm. Through this closed-loop optimization mechanism, the system can adaptively adjust the parameters to compensate for the deficiencies of the initial settings. Each iteration is based on the feedback of actual test data, making the prediction results gradually approach the true value. The magnetic core performance grade classification result combines the efficiency of image analysis and the accuracy of actual test, realizing fast and reliable quality evaluation.

[0087] The above describes only the specific embodiments of the present specification, and those skilled in the art can clearly understand that, for the convenience and brevity of description, the specific working process of the above-described system, module and unit can refer to the corresponding process in the foregoing method embodiments, which will not be described here. It should be understood that the protection scope of the present specification is not limited to this. Any person skilled in the art can easily think of various equivalent modifications or replacements within the technical range disclosed in the present specification, and these modifications or replacements should be covered within the protection scope of the present specification.

Claims

1. A method for classifying and screening inductor cores based on pattern recognition, characterized in that, The method includes: The raw image data containing pore distribution and defect morphology is acquired, and an image set covering different scale ranges is obtained by using a pre-established multi-layer resolution scanning method. Preliminary segmentation of stomatal multi-scale characteristics is performed based on the image set to distinguish stomatal regions from non-stomatal regions in the image and determine the basic outline information of stomata. Based on the determined basic outline information of stomata, the spatial relationship between stomata is obtained. Combining the characteristics of complex connectivity and variable network structure, graph theory analysis is used to construct a stomatal connectivity network model to obtain the topological data of stomatal distribution. Feature extraction was performed on the topological data of pore distribution, and the connection strength and distribution density of key nodes in the pore network were analyzed to determine the potential influencing factors of the pore network on the core performance. Based on the identified potential influencing factors and combined with the multi-scale distribution characteristics of the pore network, the initial permeability prediction feature vector is adjusted to obtain the adjusted permeability prediction feature vector. Based on the adjusted permeability prediction feature vector, the core performance pattern is identified, and the core performance level is initially evaluated according to the identified core performance pattern to determine the intermediate results of performance prediction. By comparing the determined intermediate results in multiple dimensions, the correlation mapping relationship between porosity distribution and performance level is obtained. Combined with the internal microstructure of the magnetic core, a performance level prediction value is generated. Obtain the deviation information between the predicted performance level and the actual performance data, and generate the core performance level classification result.

2. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The process of acquiring raw image data containing pore distribution and defect morphology, and using a pre-established multi-resolution scanning method to obtain an image set covering different scale ranges, includes: The magnetic core is scanned layer by layer to obtain a three-dimensional voxel dataset containing the coordinates of pore locations, pore diameters, and defect contours. The three-dimensional voxel dataset is labeled, and pore regions, defect regions, and normal material regions are distinguished based on grayscale differences. Connectivity analysis is performed to determine the volume parameters of independent pores and the geometric morphology parameters of defects. Based on the volume parameters and the geometric morphology parameters, values ​​are assigned in a three-dimensional spatial coordinate system to construct a spatial distribution matrix. The ratio of the total pore volume to the volume of the partition is calculated as the pore density distribution, and the ratio of the total defect area to the volume of the partition is calculated as the defect aggregation degree, generating an image set.

3. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The preliminary segmentation of stomatal multi-scale characteristics based on the image set, distinguishing stomatal regions from non-stomatal regions in the image, and determining the basic contour information of the stomata includes: Gray-level histograms are extracted from the image set, and an initial segmentation threshold is calculated based on the bimodal distribution feature to generate a multi-scale threshold set. Each image is binarized, and pixels below the multi-scale threshold set are marked as stomatal candidate regions, while pixels above the threshold are marked as non-stomatal regions. Morphological opening operations are applied to remove noise, generating a binary image. An eight-connected-domain labeling algorithm is used on the binary image to label stomatal candidate regions, and the area and perimeter parameters of the connected domains are calculated. Connected domains with areas smaller than a preset threshold are removed, generating labeled images of effective stomatal regions. The boundary pixel coordinate sequences of the effective stomatal regions are extracted, and a closed contour point set is formed using a boundary tracking method to determine the basic contour information of each stoma.

4. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, Based on the determined basic outline information of stomata, the spatial relationships between stomata are obtained. Considering the complexity of connectivity and the variable network structure, a stomatal connectivity network model is constructed using graph theory analysis methods to obtain the topological data of the stomatal distribution, including: Calculate the centroid coordinates of each stomata in the stomata contour information, and generate a stomata pair distance matrix based on the Euclidean distance formula; construct a binary adjacency matrix based on the distance matrix and the connectivity threshold, where a value of 1 indicates that the stomata are connected; construct a graph structure using the binary adjacency matrix, with stomata as nodes and positions with a value of 1 as edges, traverse the nodes using a depth-first search algorithm, mark the connected subgraphs and node sets, and generate the topological structure data of the stomata distribution.

5. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The topological data of the pore distribution are used to extract features, analyze the connection strength and distribution density of key nodes in the pore network, and determine the potential influencing factors of the pore network on the core performance, including: Based on the topological data, the connectivity of each pore node is calculated, and nodes with connectivity greater than a preset multiple of the average value are marked as high connectivity nodes. For the high connectivity nodes, the vertical distance from the centroid to the center line of the magnetic flux path is calculated, and nodes located in the dense magnetic flux region are identified as key nodes. The product of the connectivity strength value and the distribution density value of the key nodes is calculated as a potential influencing factor.

6. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The process of adjusting the initial permeability prediction feature vector based on the determined potential influencing factors and the multi-scale distribution characteristics of the pore network to obtain the adjusted permeability prediction feature vector includes: Based on the potential influencing factors and material parameters, an initial permeability prediction feature vector is constructed. The potential influencing factors are mapped to vector components to generate an extended feature vector. The local pore density, number of connected domains, and distribution uniformity of the multi-scale distribution characteristics of the pore network are calculated and added as new components to generate a multi-scale feature vector. Based on the weight coefficients of the potential influencing factors, the multi-scale feature vector is normalized to generate an adjusted permeability prediction feature vector.

7. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The process involves identifying core performance patterns based on the adjusted permeability prediction feature vector, conducting a preliminary assessment of core performance levels based on the identified patterns, and determining intermediate performance prediction results, including: Using the adjusted permeability prediction feature vector as input, the performance categories are divided by the support vector machine algorithm to generate preliminary evaluation results of superior, qualified and unqualified products; the dimension value of the adjusted permeability prediction feature vector is checked and compared with the normal range threshold. The dimension that deviates from the threshold is calibrated by linear interpolation to generate a calibrated feature vector as an intermediate result.

8. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The process involves performing multi-dimensional comparisons of determined intermediate results to obtain the correlation mapping relationship between porosity distribution and performance level. Combined with the internal microstructure of the magnetic core, a performance level prediction value is generated, including: The correlation between the porosity distribution characteristic parameters and historical performance levels is calculated using the Pearson correlation coefficient to generate a correlation mapping matrix. Combined with the grain size, grain boundary density, and phase composition ratio of the image set, a performance prediction weight matrix is ​​constructed, a comprehensive score is calculated, and a performance level prediction value is generated based on the score interval.

9. The method for classifying and screening inductor cores based on pattern recognition according to claim 1, characterized in that, The process of obtaining the deviation information between the predicted performance level and the actual performance data, and generating the core performance level classification result, includes: Compare the predicted performance level with the actual performance level to calculate the deviation; if the deviation exceeds the error tolerance threshold, update the pore connectivity network model and the adjusted permeability prediction feature vector to generate a corrected performance level prediction value, until the deviation meets the tolerance, and output the core performance level classification result.

Images