A Big Data-Based Power Chip Test Data Processing System and Method

Abstract

This invention relates to the field of power chip testing technology and discloses a power chip testing data processing system and method based on big data. The method performs spatial and temporal correction on raw time-series observation data synchronously acquired by multiple testing devices to form a three-dimensional observation data volume. A virtual sensing grid is constructed within this volume, and a continuously spatially distributed gridded parameter field is generated through interpolation and iterative resampling. The local geometric features of this field are identified and adaptively divided into feature sub-regions, mapped to a graph structure, and a hierarchical feature map is generated through topological agglomeration. This map drives a parameter generation network to synthesize a high-resolution parameter field. The synthesized field is then fused with the original gridded parameter field across scales to obtain an enhanced observation data volume. Physical decoupling and signal reconstruction are performed based on a physical constraint model, and the test results are output. This invention achieves intelligent identification and enhanced reconstruction of the internal features of the test data, improving the resolution and physical reliability of the results.

Description

Technical Field

[0001] This invention relates to the field of power chip testing technology, specifically to a power chip testing data processing system and method based on big data. Background Technology

[0002] Large-scale production testing of power chips relies on multiple parallel testing devices to simultaneously acquire massive amounts of time-series observation data under various operating conditions. Existing technologies mainly analyze data from individual test channels independently, or perform simple statistical fusion and signal processing after time-series alignment of multi-channel data. These methods treat the testing devices as discrete observation points, processing separate data sequences. Faced with high-dimensional, asynchronous raw data, existing solutions struggle to construct a unified data model that can comprehensively describe the continuous spatiotemporal evolution of the chip's state, and cannot effectively analyze the distribution and correlation of parameters in the chip's physical space.

[0003] Existing data processing techniques typically operate on a pre-defined uniform grid, lacking mechanisms to identify the inherent physical characteristics and structural boundaries within the data field. This leads to the smoothing of features in regions of dramatic parameter changes, potentially causing the algorithm's results to deviate from actual physical laws. Furthermore, conventional methods cannot adaptively divide different analysis regions based on the data's inherent characteristics. The core technological bottleneck in achieving high-fidelity, high-resolution test data reconstruction lies in how to automatically identify key feature structures from discrete observations, perform structured representation and enhancement based on these structures, and then combine this with strict physical constraints. Therefore, a new method is needed that can construct continuous parameter fields, intelligently identify their feature structures, and achieve deep integration of data-driven approaches and physical models. Summary of the Invention

[0004] The purpose of this invention is to provide a power chip test data processing system and method based on big data, so as to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides a power chip test data processing method based on big data, the method comprising: The system acquires raw time-series observation data synchronously collected by multiple parallel testing devices in various preset power chip operating condition cycles, performs spatial registration and time synchronization correction on the raw time-series observation data, and generates a registered three-dimensional observation data volume. A virtual sensing grid is constructed inside the three-dimensional observation data volume, and the grid nodes are filled by interpolation and iterative resampling to generate a gridded parameter field with continuous spatial properties. Based on the local geometric properties of the gridded parameter field, key geometric features are identified and marked. Based on the distribution of key geometric features in the gridded parameter field, multiple feature sub-regions are adaptively divided. The boundary conditions and internal node data of each feature sub-region are mapped to a preset graph structure, which is then transformed into an initial feature graph composed of nodes and edges. Perform topological agglomeration and feature propagation operations on the initial feature map to generate an agglomerated feature map with hierarchical features; A parameter generation network is driven by agglomerative feature maps to synthesize a synthetic parameter field with enhanced resolution. The synthetic parameter field and the gridded parameter field are fused across scales to obtain the fused enhanced observation data volume. The physical constraint model based on the power chip performs physical decoupling and signal restoration on the enhanced observation data volume, and outputs the final test data results.

[0006] Preferably, the step of performing spatial registration and temporal synchronization correction on the original time-series observation data to generate a registered three-dimensional observation data volume includes: Obtain the spatial location identifier and internal clock identifier corresponding to each test device; The relative spatial transformation matrix between each test device is calculated based on the spatial location identification code, and the original time series observation data are unified to the same spatial coordinate system; The timing drift between each test device is calculated based on the internal clock identifier code, and the timestamps of all original timing observation data are aligned using a timing interpolation algorithm. The data, which is aligned in both space and time, is rearranged according to time slices and spatial location dimensions to construct a three-dimensional observation data volume with time slice index, spatial dimension index, and observation parameter index as axes. Each data point is then associated with its source test device identifier and corrected global spatiotemporal coordinates.

[0007] Preferably, the step of constructing a virtual sensing mesh within the three-dimensional observation data volume, and filling the mesh nodes through interpolation and iterative resampling to generate a meshed parameter field with continuous spatial properties includes: Based on the spatial boundary and data density distribution of the three-dimensional observation data volume, a virtual grid topology covering the entire data space is defined; Using the actual observation points in the three-dimensional observation data volume as constraint points, the initial parameter values ​​of each node of the virtual mesh are calculated using the radial basis function interpolation algorithm; Based on the initial parameter values, perform a preset number of iterative resampling operations. In each iteration, adjust the node parameter values ​​according to the statistical distribution and gradient information of the observation point data in the neighborhood of the grid node, until the adjacent differences of the grid node parameter values ​​converge to within the preset threshold. The virtual grid nodes that are eventually stabilized through iteration, along with their parameter values ​​and spatial topological connections, are stored together to form a gridded parameter field.

[0008] Preferably, the step of identifying and marking key geometric features within a meshed parameter field based on its local geometric properties includes: Calculate the parameter gradient vector and Hessian matrix at each grid node in the gridded parameter field; Based on the combination of the gradient vector magnitude and the eigenvalues ​​of the Hessian matrix, the grid nodes are classified into flat region points, ridge feature points, valley feature points, or saddle points. Spatial clustering is performed on nodes classified as ridge feature points or valley feature points to form discrete feature line segments; Feature line segments are traced and merged according to their spatial orientation and connection relationship to form a complete key geometric feature structure, and a unique structure identifier code and type label are assigned to each key geometric feature structure.

[0009] Preferably, the adaptive division of multiple feature sub-regions based on the distribution of key geometric features in the meshed parameter field includes: Calculate the spatial influence range of each key geometric feature structure in the gridded parameter field; Based on the overlapping relationship of the spatial influence range, the gridded parameter field is divided into regions by growth to ensure that each grid node belongs to the background region of the nearest key geometric feature structure or its influence. A set of mesh nodes belonging to the same key geometric feature structure and its influence range is defined as a feature sub-region. For each feature sub-region, the identifier code of its core key geometric feature structure, the regional spatial range mask, and the parameter statistical summary of the mesh nodes inside the region are recorded.

[0010] Preferably, the step of mapping the boundary conditions and internal node data of each feature sub-region to a preset graph structure, transforming it into an initial feature graph composed of nodes and edges, includes: Each grid node within the feature sub-region is mapped to a data node in the graph structure. The attributes of the data node include its parameter values, spatial coordinates, and node type. Based on the spatial topological connection relationship of the grid nodes in the gridded parameter field, establish edge connections between the corresponding data nodes; The boundary conditions of the feature sub-region, including the set of boundary nodes and boundary parameter constraints, are encoded as boundary constraint nodes in the graph structure, and special constraint edges are established between them and the corresponding boundary data nodes. By integrating data nodes, boundary constraint nodes, and the edge connections between them, an initial feature map is constructed, with feature sub-regions as the unit.

[0011] Preferably, the step of performing topological agglomeration and feature propagation operations on the initial feature map to generate agglomerated feature map with hierarchical features includes: In the initial feature map, node agglomeration is performed based on the parameter similarity and spatial proximity between data nodes. Multiple data nodes that meet the agglomeration conditions are merged into a super node. The attributes of the super node are generated by aggregating the attributes of the nodes before merging. During the agglomeration process, the edges connecting the merged nodes are reconnected or their weights are reset according to the weights and types of the edges. The condensation process stops once the preset condensation level is reached or the condensation termination condition is met. The supernodes generated during the aggregation process, together with the remaining original data nodes, boundary constraint nodes, and updated edge connections, constitute the aggregated feature map, and the node mapping relationship from the initial feature map to the aggregated feature map is recorded.

[0012] Preferably, the step of using agglomerated feature maps to drive a parameter generation network to synthesize a synthetic parameter field with enhanced resolution includes: The node attributes and edge connectivity in the agglomerative feature map are used as the input to the parameter generation network; The parameter generation network contains multiple sequentially connected graph convolutional and pooling layers, which are used to perform multiple rounds of feature transformation and information aggregation on the input graph data; The last layer of the parameter generation network is the graph upsampling layer, which maps the learned high-dimensional features back to the grid nodes that match the spatial resolution of the original gridded parameter field based on the node mapping relationship. Through forward propagation of the parameter generation network, a new set of synthetic parameter values ​​with higher discriminative power is generated for each grid node, and the synthetic parameter values ​​constitute a synthetic parameter field in space.

[0013] Preferably, the step of fusing the synthetic parameter field and the gridded parameter field across scales to obtain the fused enhanced observation data volume includes: At multiple preset spatial scales, Gaussian pyramid decomposition is performed on the synthetic parameter field and the gridded parameter field respectively to obtain parameter sub-bands at each scale; For each spatial scale, the synthetic parameter field and the corresponding parameter sub-band of the gridded parameter field are weighted and fused. The weights are determined by the local contrast and information entropy of each sub-band at the scale. The fused parameter subbands at each scale are synthesized using the Laplacian pyramid reconstruction algorithm to generate a single-scale fused parameter field. The fused parameter field is reorganized according to the spatial and temporal dimensions of the three-dimensional observation data volume, replacing the original parameter data to form the fused enhanced observation data volume.

[0014] The physical constraint model based on the power chip performs physical decoupling and signal reconstruction on the enhanced observation data volume, and outputs the final test data results, including: A physical constraint model is established to describe the mapping relationship between the internal electrical parameters of the power chip and the external observation parameters. The physical constraint model includes circuit topology constraints and energy conservation constraints. Using the enhanced observation data volume as the observation input and the physical constraint model as the equality constraint, a parameter decoupling optimization problem is constructed. An iterative optimization algorithm is used to solve the parameter decoupling optimization problem, resulting in a set of estimated internal electrical parameters of the decoupled power chip that satisfy physical constraints. Based on the estimated internal electrical parameters and the known transfer function of the test device, the standard test signal sequence that should be observed under ideal test conditions is reconstructed and output as the final test data result.

[0015] Preferably, the present invention also includes a power chip test data processing system based on big data, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the power chip test data processing method based on big data as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are: By analyzing the local geometric properties of the gridded parameter field, key feature structures within it are identified and labeled, and feature sub-regions are adaptively divided based on the distribution of these structures. The boundary conditions and node data of each sub-region are mapped to a graph structure, forming an initial feature map. Topological agglomeration and feature propagation are performed on this map to generate a condensed feature map with hierarchical features. This condensed map is used to drive a parameter generation network to synthesize a synthetic parameter field with enhanced resolution. This approach allows the data augmentation process to be directly controlled by the abstract topological and semantic features extracted from the data field itself. The synthesized result improves spatial resolution while fully preserving the geometric morphology and relative relationships of key feature structures in the original data, avoiding feature blurring or distortion problems caused by conventional uniform interpolation or fixed kernel function methods.

[0017] A virtual sensing mesh is constructed within the 3D observation data volume. A gridded parameter field with continuous spatial properties is generated through interpolation and iterative resampling, establishing a unified mathematical representation from discrete asynchronous observation points to a continuous spatiotemporal physical field. The synthesized parameter field output by the parameter generation network is fused across scales with this basic gridded parameter field. Based on the physical constraint model of the power chip, physical decoupling and signal reconstruction are performed on the fused enhanced observation data volume. This combines the high-resolution generation capability of the data-driven method with the constraint correction capability of first-principles physics. The introduction of the physical model imposes regularization constraints on the generation process, ensuring that the final output test data results, while presenting rich details, strictly conform to the basic electrical and thermodynamic laws of chip operation in terms of numerical values ​​and dynamic evolution, overcoming the defect that the results of a purely data-driven method may violate physical common sense. Attached Figure Description Figure 1 This is a schematic diagram illustrating the working principle of the power chip test data processing method based on big data as described in this invention. Figure 2 A flowchart for generating the registered 3D observation data volume; Figure 3 A flowchart for identifying and labeling key geometric features; Figure 4 A comparison chart of "node aggregation attribute aggregation" in power chip test data processing; Figure 5 This is a comparison chart of the electrical parameters before and after physical decoupling for the power chip test data. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1This invention provides a method for processing power chip test data based on big data. The method includes: acquiring raw time-series observation data synchronously collected by multiple parallel testing devices in preset cycles of various power chip operating conditions; performing spatial registration and time-series synchronization correction on the raw time-series observation data to generate a registered three-dimensional observation data volume; constructing a virtual sensing mesh inside the three-dimensional observation data volume; filling the mesh nodes through interpolation and iterative resampling to generate a meshed parameter field with continuous spatial attributes; identifying and marking key geometric feature structures within the meshed parameter field based on its local geometric attributes; adaptively dividing multiple feature sub-regions according to the distribution of key geometric feature structures in the meshed parameter field; mapping the boundary conditions and internal node data of each feature sub-region to a preset graph structure, transforming it into an initial feature map composed of nodes and edges; performing topological agglomeration and feature propagation operations on the initial feature map to generate a condensed feature map with hierarchical features; and using the condensed feature map to drive a parameter generation network to synthesize a synthetic parameter field with enhanced resolution. The synthetic parameter field and the gridded parameter field are fused across scales to obtain the fused enhanced observation data volume. The physical constraint model based on the power chip performs physical decoupling and signal restoration on the enhanced observation data volume, and outputs the final test data results.

[0020] In one embodiment of the present invention, see [reference] Figure 2 The spatial registration and timing synchronization correction of the power chip test data processing method based on big data are performed as follows: multiple parallel test devices collect voltage and current waveforms as raw timing observation data in a preset power chip operating cycle. Each parallel test device is assigned a unique spatial location identifier and internal clock identifier upon deployment. The spatial location identifier records the installation position and attitude angle of the parallel test device in the three-dimensional coordinate system of the test platform, while the internal clock identifier is associated with the factory calibration parameters and initial start timestamp of the timing crystal inside each parallel test device.

[0021] In some embodiments, the relative spatial transformation matrix between each parallel test device is calculated based on the spatial location identifier code. The installation position coordinates and attitude angles contained in the spatial location identifier code are used to construct the direction cosine matrix and translation vector from the local coordinate system of each parallel test device to the global test platform coordinate system. Matrix multiplication is used to uniformly transform the local spatial coordinates attached to the raw time-series observation data collected by each parallel test device to the global test platform coordinate system, ensuring that all spatial sampling points are located within the same spatial reference frame. The timing drift between each parallel test device is calculated based on the internal clock identifier code. The crystal oscillator frequency deviation and initial time difference provided by the internal clock identifier code are used to model the clock drift. For each time segment to be aligned, the relative time offset between the parallel test devices is estimated by solving a least-squares problem.

[0022] It is understandable that by aligning the timestamps of all original time-series observation data using a timing interpolation algorithm, and resampling the data sequence acquired by the parallel test device with a non-reference clock using a cubic spline interpolation function, the data points are interpolated onto a globally time-stamped sequence defined by the reference clock, thereby eliminating sampling time differences caused by clock asynchrony. The data with unified spatial coordinates and aligned timestamps is then rearranged and organized according to the global time slice index, the spatial dimension index of the parallel test device, and the indexes of observation parameters such as voltage and current, constructing a three-dimensional observation data volume in the form of a three-dimensional array. The first dimension of the three-dimensional observation data volume corresponds to time evolution, the second dimension corresponds to different spatial observation positions, and the third dimension corresponds to different observed electrical parameters. Each data point in the three-dimensional observation data volume is associated with its source parallel test device identifier and its corrected global time and global spatial coordinates. The global spatial coordinates are calculated from the spatial transformation matrix, and the global time coordinates are given by the timestamps after timing interpolation.

[0023] Optionally, the mathematical model for calculating the timing drift can be expressed as finding a time drift such that the response waveforms of two parallel testing devices to the same excitation signal have maximum cross-correlation on the time axis. The objective function used to estimate the timing drift is... It can be constructed as follows: ; Where: symbol Represents the estimated time drift, symbol This represents the signal sequence acquired by the parallel test device used as a time reference, with the symbol... This represents the signal sequence acquired by the parallel test device to be calibrated, with the symbol... This represents the time offset variable to be solved, and the summation operation is performed within the overlapping time window.

[0024] In some embodiments, the spatial position identification code is obtained by measuring the optical target sphere mounted on each parallel test device using a laser tracker, thereby obtaining high-precision spatial position and attitude data. The internal clock identification code is generated when the parallel test device is powered on and initialized. The main controller sends a global start pulse and time packet via a precision clock synchronization protocol. The parallel test device records the initial difference between the internal clock and the global clock and stores it as part of the internal clock identification code. During the process of unifying the spatial coordinate system, for the local coordinate system of the parallel test device with rotational relationships, quaternions or rotation matrices are used to represent the attitude and perform synthetic calculations. When constructing the 3D observation data volume, for grid points that are not covered due to different spatial positions, the data channels corresponding to their observation parameter indices are filled with invalid value markers in the 3D observation data volume, to be processed in subsequent virtual mesh interpolation steps.

[0025] In one embodiment of the present invention, see [reference] Figure 3 The construction of the gridded parameter field and the identification of key geometric features are performed as follows: Based on the generated registered 3D observation data volume, the data points may be unevenly distributed in space and have gaps. A virtual grid topology covering the entire data space is defined according to the spatial boundary and data density distribution of the 3D observation data volume. The spatial boundary is determined by the maximum and minimum values ​​of the global spatial coordinates after correction by all parallel testing devices. The data density distribution is obtained by statistically analyzing the observation points using a kernel density estimation algorithm. The virtual grid topology uses a regular hexahedral grid, with a higher grid resolution in areas of high data density and a lower resolution in areas of low data density, thus forming an adaptive variable resolution grid structure.

[0026] Using actual observation points in the 3D observation data volume as constraint points, the radial basis function interpolation algorithm is employed to calculate the initial parameter values ​​of each node in the virtual mesh. The shape parameters are dynamically adjusted based on the average distance between adjacent observation points. For each virtual mesh node, the radial basis function interpolation algorithm finds all actual observation points within a certain radius around it. The parameter values ​​and spatial coordinates of these observation points are used as input to solve a system of linear equations to obtain a set of weighting coefficients. Finally, a weighted sum is calculated to obtain the initial parameter value of the mesh node. Based on the initial parameter values, a preset number of iterative resampling iterations is performed (set to ten). In each iteration, the node parameter values ​​are adjusted based on the statistical distribution and gradient information of the observation point data within the mesh node's neighborhood. The neighborhood is defined as a cubic region centered on the node with a side length of three grid units. The statistical distribution includes the mean and variance of the observation point parameter values ​​within the neighborhood, and the gradient information is calculated from the three-dimensional finite difference of the observation point parameter values ​​within the neighborhood.

[0027] The adjustment rule for iterative resampling is to move the current parameter value of a grid node towards the mean of the parameter values ​​of its neighboring observation points. The step size is proportional to the difference between the current parameter value and the neighborhood mean. A stability term based on neighborhood variance is introduced to avoid over-adjustment. For regions with significant gradient information, the step size is appropriately adjusted along the gradient direction. The iterative process continues until the adjacent differences in the parameter values ​​of the grid nodes converge to within a preset threshold. The adjacent difference is defined as the absolute value of the change in parameter value of the same grid node in two adjacent iterations. The preset threshold is set to one ten-thousandth of the initial parameter value range. The finally iteratively stable virtual grid nodes, their parameter values, and spatial topological connections are stored together to form a gridded parameter field. The spatial topological connections record the node indices of the eight vertices of each hexahedral grid cell.

[0028] In some embodiments, the parameter gradient vector and Hessian matrix at each grid node in the gridded parameter field are calculated. The parameter gradient vector is calculated using the central difference method on the regular hexahedral structure of the gridded parameter field, and the Hessian matrix is ​​obtained by calculating the gradient of each component of the gradient vector, i.e., by performing a second-order difference operation. Grid nodes are classified according to the combination of the gradient vector magnitude and the eigenvalues ​​of the Hessian matrix. Nodes with a gradient vector magnitude lower than 10% of the global average magnitude are initially classified as candidate points for flat regions, while nodes with a magnitude higher than this threshold proceed to subsequent feature analysis. The eigenvalues ​​of the Hessian matrix are solved using the Jacobi iteration method, and the signs and relative magnitudes of the eigenvalues ​​are analyzed: if two eigenvalues ​​are negative and their absolute values ​​are significantly greater than the third eigenvalue, the node is classified as a ridge feature point; if two eigenvalues ​​are positive and their absolute values ​​are significantly greater than the third eigenvalue, the node is classified as a valley feature point; if the eigenvalues ​​have both positive and negative signs and their absolute values ​​are similar, the node is classified as a saddle point; nodes that do not meet the above conditions are classified as points in flat regions.

[0029] Nodes classified as ridge or valley feature points are spatially clustered to form discrete feature segments. A density-based clustering algorithm is used, with the clustering distance threshold set to twice the average grid spacing. Feature segments are then tracked and merged based on their spatial orientation and connectivity to form complete key geometric feature structures. The tracking algorithm starts from any unvisited feature segment origin and searches for adjacent segments along the segment's extension direction where spatial distance and direction angle variations are within tolerance, connecting them until no connectable segments are found, thus forming a continuous feature structure. Each complete key geometric feature structure is assigned a unique structure identifier and type label, recorded as either "ridge" or "valley".

[0030] It is understandable that the specific calculation process of radial basis function interpolation involves solving a linear system. For a local region with m constrained observation points, the parameter values ​​of the virtual mesh node p need to be solved. The weighting coefficients that need to be solved The following system of equations must be satisfied: ; Where: symbol The symbol represents the value of the parameter to be determined at virtual mesh node p. The symbol represents the weight coefficient corresponding to the i-th constrained observation point. Denotes the selected radial basis function, symbol This represents the spatial position vector of the virtual mesh node p and the spatial position vector of the i-th constrained observation point. The Euclidean distance between them, symbol The shape parameter represents the Gaussian kernel function. Weighting coefficients. It is determined by solving a system of linear equations consisting of parameter values ​​at m constrained observation points.

[0031] Optionally, the statistical distribution calculation of observation point data within the neighborhood during iterative resampling includes not only the neighborhood of the grid node itself in the current iteration, but also the actual observation point data falling within that neighborhood in the original 3D observation data volume, to ensure that the gridding process is always constrained by the original measured data. In some embodiments, eigenvalue analysis of the Hessian matrix is ​​performed after removing candidate points in flat regions to reduce computational load. The merging operation of feature segments allows for small gaps at connection points, with the gap tolerance set at 1.5 times the average grid spacing. It can be understood that the spatial distribution of ridge and valley feature points identified through the local geometric properties of the gridding parameter field often corresponds to the boundaries or conduction paths of parameter abrupt changes in the power chip test data. The complete extraction of key geometric feature structures provides a direct basis for the subsequent adaptive partitioning of feature sub-regions.

[0032] In one embodiment of the present invention, the adaptive partitioning of feature sub-regions and the construction of the initial feature map are performed as follows: based on a gridded parameter field containing complete key geometric feature structures that has been labeled. The spatial influence range of each key geometric feature structure in the gridded parameter field is calculated. For linear key geometric feature structures such as ridges or valleys, the spatial influence range is defined by calculating the Euclidean distance from each grid node in the gridded parameter field to the nearest point on the key geometric feature structure, and a distance threshold is set. The distance threshold is adaptively determined based on the attenuation characteristics of the parameter gradient in the local region traversed by the key geometric feature structure. A larger distance threshold is used for regions with gentle parameter gradient changes, and a smaller distance threshold is used for regions with drastic parameter gradient changes. For complex key geometric feature structures, the spatial influence range is the minimum value of the distance field to all the line segments that make up the structure.

[0033] The gridded parameter field is divided into regions based on the overlapping relationships of their spatial influence ranges. Starting from each grid node on a key geometric feature structure, the growth is constrained by the boundary of its spatial influence range, incorporating neighboring grid nodes that meet the conditions. Each grid node is assigned to the nearest key geometric feature structure or its influenced background region. This assignment is determined by comparing the distances from each grid node to each key geometric feature structure. If the distance from a grid node to all key geometric features structures is greater than the distance threshold defined for each of those key geometric features at that node, then that grid node belongs to the background region. A set of grid nodes belonging to the same key geometric feature structure and its influence range is defined as a feature sub-region. The background region is also defined as one or more separate feature sub-regions. For each feature sub-region, the core key geometric feature structure identifier, the region's spatial range mask, and a parameter statistical summary of the grid nodes within the region are recorded. The parameter statistical summary includes the mean, variance, maximum, and minimum values ​​of the node parameters within the region.

[0034] Each grid node within a feature sub-region is mapped to a data node in the graph structure. The attributes of each data node include parameter values ​​inherited from the gridding parameter field, spatial coordinates, and node type. The node type is determined based on whether the grid node was classified as a flat region point, ridge feature point, valley feature point, or saddle point in the previous step. Edge connections are established between corresponding data nodes based on the spatial topological connectivity of the grid nodes in the gridding parameter field. The adjacency relationship of the regular hexahedral grid in the gridding parameter field defines the edge connections between data nodes. If two grid nodes share an edge in the grid topology, an undirected edge is established between the corresponding two data nodes.

[0035] In some embodiments, the boundary conditions of the feature sub-regions are encoded as boundary constraint nodes in a graph structure. The boundary conditions include a set of boundary nodes and boundary parameter constraints. The set of boundary nodes is a set of mesh nodes located at the edge of the spatial extent mask of the feature sub-region. The boundary parameter constraints are continuity conditions or known fixed values ​​that the parameter values ​​at these boundary nodes must satisfy. A boundary constraint node is created for each boundary condition, and the attributes of the boundary constraint node store the constraint type and constraint value it represents. Special constraint edges are established between the boundary constraint nodes and their corresponding boundary data nodes. These constraint edges have a different type identifier than ordinary connection edges and carry constraint information in the graph structure, such as equality constraints or inequality constraints. The data nodes, boundary constraint nodes, and the edge connections between them are integrated to form an initial feature graph based on the feature sub-regions. The initial feature graph is stored in memory in the form of an adjacency list or adjacency matrix.

[0036] It is understandable that the spatial influence range can be calculated using a diffusion process-based model. This model determines the influence intensity distribution of key geometric features by solving an anisotropic diffusion equation, with the influence range defined by regions where the intensity exceeds a certain threshold. The diffusion equation takes the form: ; Where: symbol Indicates the intensity of influence at spatial location x and time t, with the symbol... It is a diffusion tensor related to the local geometric properties of the meshing parameter field at location x, with the symbol... It is the gradient operator, symbol It is a divergence operator. The locations of key geometric features are set as initial influence sources. The equation is solved numerically until a steady state is reached, and the steady-state solution is obtained. The area exceeding the set threshold is defined as the spatial influence range of the key geometric feature structure.

[0037] Optionally, the regional spatial extent mask is stored as a three-dimensional Boolean array, with the array dimensions consistent with the meshing parameter field. Grid positions within the feature sub-region are marked as true values, otherwise as false values. In some embodiments, for the background region as a feature sub-region, its core key geometric feature structure identifier is recorded as a null value or a specific identifier. The parameter statistical summary of the grid nodes within the feature sub-region can be dynamically calculated and appended to the supernode representing the feature sub-region after the graph structure is constructed, but it is stored separately as regional metadata during the initial feature map construction phase. It can be understood that the introduction of boundary constraint nodes allows the graph structure to not only carry the topological relationships of the observation data of the feature sub-region but also embed physical or geometric boundary conditions, providing a physically meaningful constraint framework for subsequent graph-based analysis and processing.

[0038] In one embodiment of the present invention, the implementation method of performing topological agglomeration and feature propagation operations on the initial feature map and using the agglomerated feature map to drive the parameter generation network is as follows. In the initial feature map, node agglomeration is performed based on the parameter similarity and spatial proximity between data nodes. Parameter similarity is measured by calculating the Euclidean distance between the parameter values ​​in the attributes of two data nodes, and spatial proximity is measured by calculating the Euclidean distance between the spatial coordinates in the attributes of two data nodes. A comprehensive similarity threshold is set. When the weighted sum of parameter similarity and spatial proximity between two data nodes exceeds the threshold, it is determined that the agglomeration condition is met. Multiple data nodes that meet the agglomeration condition are merged into a supernode. The attributes of the supernode are generated by aggregating the attributes of the nodes before merging. The aggregation operation includes taking the mean of numerical parameters, taking the mode of categorical node types, and taking the centroid of spatial coordinates.

[0039] During the aggregation process, edges connecting merged nodes are reconnected or have their weights reset based on their weights and types. Edges connecting two data nodes merged into the same supernode are removed from the aggregation feature map. Edges connecting a merged data node to an unmerged data node are reconnected to the corresponding supernode. Constraint edges connecting a merged data node to a boundary constraint node are also reconnected to the corresponding supernode, with their constraint type and value remaining unchanged. The weights of the reconnected edges are adjusted based on their original weights and the topological relationships of the connected nodes before and after aggregation.

[0040] The agglomeration process stops after reaching a preset agglomeration level or meeting agglomeration termination condition. The preset agglomeration level is set to three levels, and the agglomeration termination condition is set to the number of remaining nodes in the graph being less than 10% of the initial number of nodes. The supernodes generated during the agglomeration process, together with the remaining original data nodes, boundary constraint nodes, and updated edge connections, constitute the agglomeration feature map. The node mapping relationship from the initial feature map to the agglomeration feature map is recorded. This mapping relationship records which original data nodes in the initial feature map each supernode was merged from. See Table 1.

[0041] Table 1: Shows the attribute aggregation table of a supernode during the node aggregation process. Node type Original node ID Parameter value Spatial coordinates (x, y, z) Node type label Data Node D_101 3.25 (1.0,2.0,0.5) flat area points Data Node D_102 3.31 (1.1,2.1,0.5) flat area points Data Node D_103 3.18 (1.0,2.0,0.6) flat area points supernode S_15 3.25 (1.03,2.03,0.53) flat area points In some embodiments, the comprehensive similarity calculation for determining whether two data nodes satisfy the aggregation condition is performed using the following formula: ; Where: symbol The symbol represents the overall similarity between data node i and data node j. and These represent the feature vectors related to the parameter values ​​in the attribute vectors of data node i and data node j, respectively, with symbols... and Let i and j represent the spatial coordinate vectors of data node i and data node j, respectively. Represents the eigenvector The Euclidean norm, Represents the eigenvector The Euclidean norm, symbol It is a weighting coefficient between 0 and 1 used to balance parameter similarity and spatial proximity, with the sign... It is a positive scaling parameter used to adjust the degree of influence of spatial distance. When When the similarity exceeds the preset threshold, data node i and data node j satisfy the aggregation condition.

[0042] The node attributes and edge connections in the agglomerative feature map are used as input to the parameter generation network. Node attributes are converted into feature vectors, and edge connections are converted into adjacency matrices. The parameter generation network contains multiple sequentially connected graph convolutional and pooling layers to perform multiple rounds of feature transformation and information aggregation on the input graph data. The graph convolutional layers use spectral domain-based graph convolution operations to aggregate the feature information of each node and its neighboring nodes. The pooling layers downsample based on the existing supernode hierarchy in the agglomerative feature map, gradually abstracting higher-level graph features. The last layer of the parameter generation network is a graph upsampling layer. The graph upsampling layer backmaps the learned high-dimensional features to grid nodes that match the spatial resolution of the original gridded parameter field according to the node mapping relationship. The node mapping relationship records which original data nodes in the initial feature map correspond to each node in the agglomerative feature map.

[0043] The parameter generation network generates a new set of synthetic parameter values ​​with higher discriminative power for each grid node through forward propagation. During forward propagation, node features are transferred between graph convolutional layers without linear transformation. Pooling layers compress the graph structure and extract global features. The final upsampling layer combines global and local features to regress synthetic parameter values ​​for each target grid node. The synthetic parameter values ​​constitute a synthetic parameter field in space, and the gradient changes of the synthetic parameter field at geometric feature boundaries are more significant than those of the original gridded parameter field.

[0044] It can be understood that the operation of graph convolutional layers can be described as iterative propagation and transformation of the node feature matrix and the normalized adjacency matrix. The training process of the parameter generation network uses a sample map constructed from a large amount of historical power chip test data, and the loss function is designed to minimize the difference between the synthetic parameter field output by the network and the high-precision reference parameter field. Optionally, the edge weight reset strategy can be updated based on the attribute similarity of the nodes connected by the edge, with edges connecting nodes with similar attributes receiving higher weights after reconnection. In some embodiments, the preset agglomeration level can be dynamically calculated according to the scale of the initial feature map, for example, setting the agglomeration level to an integer such that the final number of supernodes is approximately the square root of the initial number of nodes. It can be understood that the agglomeration feature map provides a multi-level abstract representation of the data, and the parameter generation network uses this hierarchical feature to learn and generate a new parameter field that enhances the contrast of the feature regions, which provides a high-resolution synthetic data source for subsequent cross-scale fusion.

[0045] See Figure 4This is a comparative chart of "node agglomeration and attribute aggregation" in power chip test data processing, a professional analysis tool used in the topology agglomeration stage. The parameter value of supernode S_15 (approximately 3.25) is the average of the parameter values ​​of the three original nodes, and the X / Y / Z coordinates are the centroids of the original node coordinates. While there are slight differences in the parameter values ​​and coordinates of the original nodes, the supernodes achieve "smooth aggregation" of attributes after agglomeration, preserving the spatial and parameter characteristics of the original nodes. This chart visually verifies the effectiveness of "node agglomeration": by agglomerating the attributes of similar original nodes, supernodes both compress the data size and retain the core features of the original data, providing a concise and effective data foundation for subsequent generation of hierarchical feature maps and driving parameter generation networks.

[0046] In one embodiment of the present invention, cross-scale fusion and physical decoupling signal restoration are performed as follows, based on the generated synthetic parameter field and the original gridded parameter field. Gaussian pyramid decomposition is performed on the synthetic parameter field and the gridded parameter field at multiple preset spatial scales to obtain parameter sub-bands at each scale. The preset spatial scales include the original scale and scales after two, four, and eight downsampling operations. Gaussian pyramid decomposition is achieved by iteratively applying Gaussian kernel convolution and downsampling operations, with each downsampling halving the image size. For each spatial scale, the parameter sub-bands corresponding to the synthetic parameter field and the gridded parameter field are weighted and fused. The weights are determined by the local contrast and information entropy of each sub-band at the scale. The local contrast is obtained by calculating the relative difference between the maximum and minimum values ​​in the neighborhood of each pixel in the sub-band, and the information entropy is obtained by calculating the Shannon entropy of the grayscale value distribution of the sub-band data.

[0047] The fused parameter subbands at each scale are combined into a single-scale fused parameter field using the Laplacian pyramid reconstruction algorithm. The Laplacian pyramid reconstruction starts at the coarsest scale, adding the upsampled low-resolution fused subbands to the high-frequency detail subbands of the next scale, progressively upsampling and adding until the original resolution is restored. The fused parameter field is then reorganized according to the spatial and temporal dimensions of the 3D observation data volume, replacing the original parameter data to form the enhanced fused observation data volume. During the reorganization process, each 2D slice of the fused parameter field is filled into the corresponding position in the 3D observation data volume based on its spatial and temporal slice indices, overwriting or replacing the original observation parameter values.

[0048] In some embodiments, a physical constraint model is established to describe the mapping relationship between the internal electrical parameters of the power chip and the external observation parameters. The physical constraint model includes circuit topology constraints and energy conservation constraints. The circuit topology constraints are established based on the equivalent circuit diagram of the power chip, representing the relationship between internal node voltages and branch currents as a system of linear equations. The energy conservation constraints require a balance between input power, output power, and power loss, expressed as nonlinear equations. Using the augmented observation data volume as the observation input and the physical constraint model as the equality constraint condition, a parameter decoupling optimization problem is constructed. The objective function of the optimization problem is to minimize the mean square error between the parameter sequence observed in the augmented observation data volume and the parameter sequence predicted by the internal electrical parameters. The equality constraint condition is the physical constraint model equation. An iterative optimization algorithm is used to solve the parameter decoupling optimization problem to obtain a set of decoupled internal electrical parameter estimates of the power chip that satisfy the physical constraints. The iterative optimization algorithm uses the Lagrange multiplier method combined with gradient descent, updating the internal electrical parameter estimates in each iteration to simultaneously reduce the objective function value and satisfy the constraint conditions.

[0049] Based on the estimated internal electrical parameters and the known transfer function of the test device, the standard test signal sequence that should be observed under ideal test conditions is reconstructed as the final test data output. The transfer function of the test device describes the amplitude attenuation and phase delay characteristics of the test device hardware for the ideal signal. The compensated signal is obtained by calculating the inverse model of the transfer function using the estimated internal electrical parameters. The specific calculation of the weighting coefficients in the weighted fusion is determined using the following formula: ; Where: symbol This represents the fusion weight of the synthesized parametric field subband at the current scale and pixel location, with the sign... The symbol represents the local contrast of the synthesized parameter field subband within the local window of the current pixel. The symbol represents the information entropy of the synthesized parameter field subband within the local window of the current pixel. The symbol represents the local contrast of the meshed parameter field subbands within the same local window. This represents the information entropy of the gridded parametric field sub-band within the same local window. The fusion weight of the gridded parametric field sub-band is then... .

[0050] It is understandable that the Gaussian kernel convolution template size used in the Gaussian pyramid decomposition process increases with scale to maintain smoothness and consistency. A 3x3 template is used at the finest scale, and the template size doubles with each additional scale layer. The high-frequency detail subbands in the Laplacian pyramid reconstruction are obtained by subtracting the Gaussian-smoothed, downsampled, and then upsampled version from the original scale subbands. The circuit topology constraints in the physical constraint model are specifically represented by Kirchhoff's current law and voltage law equations, while the energy conservation constraint is specifically represented by the integral of the product of the input port voltage and current being equal to the sum of the integrals of the products of the output port voltage and current plus the integral of the internal power loss. Optionally, the calculation window size for local contrast is inversely proportional to the resolution of the current pyramid scale; a larger window is used at coarser scales to capture regional features. The iterative optimization algorithm can introduce a regularization term to prevent the internal electrical parameter estimates from overfitting the observation noise; the regularization term uses the L2 norm of the internal electrical parameters. In some embodiments, the test device transfer function is obtained through prior calibration experiments and stored in the form of a frequency response matrix; the inverse model calculation involves matrix inversion or pseudo-inversion operations. It is understandable that cross-scale fusion fully utilizes the high-resolution advantage of the synthetic parameter field at the feature boundary and the stable characteristics of the gridded parameter field in the uniform region. The physical decoupling process maps the fused enhanced observation data back to the intrinsic electrical parameter space of the power chip, thereby outputting standard test data results that are not affected by the characteristics of the test device.

[0051] See Figure 5 This is a comparison chart of electrical parameters before and after physical decoupling of power supply chip test data, a professional analysis tool used in the final stage of data processing. Before decoupling, the parameter values ​​fluctuated significantly, reflecting the influence of noise and interference on the original observation data. After decoupling, all parameter values ​​became more reasonable, demonstrating the effectiveness of physical decoupling in removing interference. This chart visually verifies the core function of "physical decoupling": by combining the physical constraint model of the power supply chip, it restores the coupled observation data to true electrical parameters that conform to circuit topology and energy conservation, providing accurate test data support for power supply chip performance evaluation and fault diagnosis.

[0052] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0053] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A power chip test data processing method based on big data, characterized in that, include: The system acquires raw time-series observation data synchronously collected by multiple parallel testing devices in various preset power chip operating condition cycles, performs spatial registration and time synchronization correction on the raw time-series observation data, and generates a registered three-dimensional observation data volume. A virtual sensing grid is constructed inside the three-dimensional observation data volume, and the grid nodes are filled by interpolation and iterative resampling to generate a gridded parameter field with continuous spatial properties. Based on the local geometric properties of the gridded parameter field, key geometric features are identified and marked. Based on the distribution of key geometric features in the gridded parameter field, multiple feature sub-regions are adaptively divided. The boundary conditions and internal node data of each feature sub-region are mapped to a preset graph structure, which is then transformed into an initial feature graph composed of nodes and edges. Perform topological agglomeration and feature propagation operations on the initial feature map to generate an agglomerated feature map with hierarchical features; A parameter generation network is driven by agglomerative feature maps to synthesize a synthetic parameter field with enhanced resolution. The synthetic parameter field and the gridded parameter field are fused across scales to obtain the fused enhanced observation data volume. The physical constraint model based on the power chip performs physical decoupling and signal restoration on the enhanced observation data volume, and outputs the final test data results.

2. The power chip test data processing method based on big data according to claim 1, characterized in that, The process of spatial registration and temporal synchronization correction of the original time-series observation data to generate a registered three-dimensional observation data volume includes: Obtain the spatial location identifier and internal clock identifier corresponding to each test device; The relative spatial transformation matrix between each test device is calculated based on the spatial location identification code, and the original time series observation data are unified to the same spatial coordinate system; The timing drift between each test device is calculated based on the internal clock identifier code, and the timestamps of all original timing observation data are aligned using a timing interpolation algorithm. The data, which is aligned in both space and time, is rearranged according to time slices and spatial location dimensions to construct a three-dimensional observation data volume with time slice index, spatial dimension index, and observation parameter index as axes. Each data point is then associated with its source test device identifier and corrected global spatiotemporal coordinates.

3. The power chip test data processing method based on big data according to claim 2, characterized in that, The process of constructing a virtual sensing mesh within the three-dimensional observation data volume, and filling the mesh nodes through interpolation and iterative resampling to generate a meshed parameter field with continuous spatial properties includes: Based on the spatial boundary and data density distribution of the three-dimensional observation data volume, a virtual grid topology covering the entire data space is defined; Using the actual observation points in the three-dimensional observation data volume as constraint points, the initial parameter values ​​of each node of the virtual mesh are calculated using the radial basis function interpolation algorithm; Based on the initial parameter values, perform a preset number of iterative resampling operations. In each iteration, adjust the node parameter values ​​according to the statistical distribution and gradient information of the observation point data in the neighborhood of the grid node, until the adjacent differences of the grid node parameter values ​​converge to within the preset threshold. The virtual grid nodes that are eventually stabilized through iteration, along with their parameter values ​​and spatial topological connections, are stored together to form a gridded parameter field.

4. The power chip test data processing method based on big data according to claim 3, characterized in that, The process of identifying and labeling key geometric features within a gridded parameter field based on its local geometric properties includes: Calculate the parameter gradient vector and Hessian matrix at each grid node in the gridded parameter field; Based on the combination of the gradient vector magnitude and the eigenvalues ​​of the Hessian matrix, the grid nodes are classified into flat region points, ridge feature points, valley feature points, or saddle points. Spatial clustering is performed on nodes classified as ridge feature points or valley feature points to form discrete feature line segments; Feature line segments are traced and merged according to their spatial orientation and connection relationship to form a complete key geometric feature structure, and a unique structure identifier code and type label are assigned to each key geometric feature structure.

5. The power chip test data processing method based on big data according to claim 4, characterized in that, The process of adaptively dividing multiple feature sub-regions based on the distribution of key geometric features in the gridded parameter field includes: Calculate the spatial influence range of each key geometric feature structure in the gridded parameter field; Based on the overlapping relationship of the spatial influence range, the gridded parameter field is divided into regions by growth to ensure that each grid node belongs to the background region of the nearest key geometric feature structure or its influence. A set of mesh nodes belonging to the same key geometric feature structure and its influence range is defined as a feature sub-region. For each feature sub-region, the identifier code of its core key geometric feature structure, the regional spatial range mask, and the parameter statistical summary of the mesh nodes inside the region are recorded.

6. The power chip test data processing method based on big data according to claim 5, characterized in that, The step of mapping the boundary conditions and internal node data of each feature sub-region to a preset graph structure, transforming it into an initial feature graph composed of nodes and edges, includes: Each grid node within the feature sub-region is mapped to a data node in the graph structure. The attributes of the data node include its parameter values, spatial coordinates, and node type. Based on the spatial topological connection relationship of the grid nodes in the gridded parameter field, establish edge connections between the corresponding data nodes; The boundary conditions of the feature sub-region, including the set of boundary nodes and boundary parameter constraints, are encoded as boundary constraint nodes in the graph structure, and special constraint edges are established between them and the corresponding boundary data nodes. By integrating data nodes, boundary constraint nodes, and the edge connections between them, an initial feature map is constructed, with feature sub-regions as the unit.

7. The power chip test data processing method based on big data according to claim 6, characterized in that, The step of performing topological agglomeration and feature propagation operations on the initial feature map to generate an agglomerated feature map with hierarchical features includes: In the initial feature map, node agglomeration is performed based on the parameter similarity and spatial proximity between data nodes. Multiple data nodes that meet the agglomeration conditions are merged into a super node. The attributes of the super node are generated by aggregating the attributes of the nodes before merging. During the agglomeration process, the edges connecting the merged nodes are reconnected or their weights are reset according to the weights and types of the edges. The condensation process stops once the preset condensation level is reached or the condensation termination condition is met. The supernodes generated during the aggregation process, together with the remaining original data nodes, boundary constraint nodes, and updated edge connections, constitute the aggregated feature map, and the node mapping relationship from the initial feature map to the aggregated feature map is recorded.

8. The power chip test data processing method based on big data according to claim 7, characterized in that, The method of using condensed feature maps to drive a parameter generation network to synthesize a synthetic parameter field with enhanced resolution includes: The node attributes and edge connectivity in the agglomerative feature map are used as the input to the parameter generation network; The parameter generation network contains multiple sequentially connected graph convolutional and pooling layers, which are used to perform multiple rounds of feature transformation and information aggregation on the input graph data; The last layer of the parameter generation network is the graph upsampling layer, which maps the learned high-dimensional features back to the grid nodes that match the spatial resolution of the original gridded parameter field based on the node mapping relationship. Through forward propagation of the parameter generation network, a new set of synthetic parameter values ​​with higher discriminative power is generated for each grid node, and the synthetic parameter values ​​constitute a synthetic parameter field in space.

9. The power chip test data processing method based on big data according to claim 8, characterized in that, The process of fusing the synthetic parameter field and the gridded parameter field across scales to obtain the fused enhanced observation data volume includes: At multiple preset spatial scales, Gaussian pyramid decomposition is performed on the synthetic parameter field and the gridded parameter field respectively to obtain parameter sub-bands at each scale; For each spatial scale, the synthetic parameter field and the corresponding parameter sub-band of the gridded parameter field are weighted and fused. The weights are determined by the local contrast and information entropy of each sub-band at the scale. The fused parameter subbands at each scale are synthesized using the Laplacian pyramid reconstruction algorithm to generate a single-scale fused parameter field. The fused parameter field is reorganized according to the spatial and temporal dimensions of the three-dimensional observation data volume, replacing the original parameter data to form the fused enhanced observation data volume. The physical constraint model based on the power chip performs physical decoupling and signal reconstruction on the enhanced observation data volume, and outputs the final test data results, including: A physical constraint model is established to describe the mapping relationship between the internal electrical parameters of the power chip and the external observation parameters. The physical constraint model includes circuit topology constraints and energy conservation constraints. Using the enhanced observation data volume as the observation input and the physical constraint model as the equality constraint, a parameter decoupling optimization problem is constructed. An iterative optimization algorithm is used to solve the parameter decoupling optimization problem, resulting in a set of estimated internal electrical parameters of the decoupled power chip that satisfy physical constraints. Based on the estimated internal electrical parameters and the known transfer function of the test device, the standard test signal sequence that should be observed under ideal test conditions is reconstructed and output as the final test data result.

10. A power chip test data processing system based on big data, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the power chip test data processing method based on big data as described in any one of claims 1 to 9.

Images