Intelligent method and system for integrated circuit package current monitoring and analysis

Abstract

The present invention provides an intelligent method and system for monitoring and analyzing integrated circuit package current, which relates to the field of current monitoring technology, including obtaining current timing data of an integrated circuit to be tested under different working conditions, decomposing the calibrated sampled data into a characteristic coefficient matrix through adaptive wavelet transform; constructing a deep feature analysis model based on the characteristic coefficient matrix, using a graph attention network to combine the physical structure information of the integrated circuit with the current characteristics to construct a dynamic feature map, identifying abnormal current characteristics and locating abnormal areas through multi-layer graph convolution operations; inputting the abnormal current characteristics and abnormal area information into a digital twin analysis system, determining the corresponding relationship between the abnormal current characteristics and the physical parameters of the circuit through iterative calculation, combining the comparative analysis of the measured data with the simulation results, outputting the fault location results and the fault development trend prediction, and generating a diagnostic analysis report.

Description

Technical Field

[0001] The present invention relates to current monitoring technology, and in particular to an intelligent method and system for monitoring and analyzing current of integrated circuit packaging. Background Art

[0002] Reliability monitoring of integrated circuits during the packaging process and in actual applications is a key link to ensure their normal operation. At present, current monitoring is mainly used to evaluate the working status of integrated circuits. Traditional current monitoring methods mainly include the following: one is to measure the current through a series sampling resistor, but this method will interfere with the normal operation of the circuit; the second is to use Hall sensors for non-contact measurement, but the measurement accuracy and spatial resolution are low; the third is to use a current probe array for scanning, but the measurement efficiency is low and it is easy to damage the device under test.

[0003] However, a single current acquisition method cannot guarantee the accuracy and reliability of the measurement, and the traditional data analysis method lacks the ability to deeply explore the abnormal characteristics of the current. In addition, as integrated circuits develop towards high density and multi-function, their internal structures are becoming increasingly complex, and traditional current monitoring methods can no longer meet the needs of accurately locating and predicting faults. Especially in the packaging process, due to the high integration of the circuit structure, traditional monitoring methods face huge challenges in spatial resolution and time response. Therefore, it is urgent to develop new intelligent current monitoring and analysis methods. Summary of the invention

[0004] The embodiments of the present invention provide an intelligent method and system for monitoring and analyzing integrated circuit package current, which can solve the problems in the prior art.

[0005] According to a first aspect of the embodiments of the present invention,

[0006] An intelligent method for monitoring and analyzing integrated circuit package current is provided, comprising:

[0007] The current timing data of the integrated circuit to be tested under different working states is obtained through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel obtains the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, and the two-dimensional magnetic field distribution data is reconstructed into a current distribution map through an electromagnetic field inversion algorithm, and the sampling data of the main acquisition channel is calibrated and compensated according to the current distribution map, and the calibrated sampling data is decomposed into a characteristic coefficient matrix through an adaptive wavelet transform;

[0008] A deep feature analysis model is constructed based on the feature coefficient matrix. The deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer, and a space dimension convolution layer. The time domain features, frequency domain features, and space features in the feature coefficient matrix are extracted through a three-dimensional convolution operation, and are input into a graph attention network. The physical structure information of the integrated circuit is combined with the current features by using the graph attention network to construct a dynamic feature map. The abnormal current features are identified and the abnormal area is located through multi-layer graph convolution operations.

[0009] The abnormal current characteristics and abnormal area information are input into a digital twin analysis system, which includes a physical model of an integrated circuit and a multi-physical field coupling simulation model. An interactive influence relationship among current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model. The correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the prediction of the fault development trend are output, and a diagnostic analysis report is generated.

[0010] In an optional embodiment,

[0011] The two-dimensional magnetic field distribution data is reconstructed into a current distribution diagram through an electromagnetic field inversion algorithm, and the sampling data of the main acquisition channel is calibrated and compensated according to the current distribution diagram. The calibrated sampling data is decomposed into a characteristic coefficient matrix through adaptive wavelet transform, including:

[0012] Dividing the integrated circuit to be tested into a metal interconnect layer, a device layer and a substrate layer, applying current continuity constraints, interlayer interface constraints and current density upper limit constraints to the metal interconnect layer, the device layer and the substrate layer, wherein the current continuity constraints make the current density divergence in each layer zero, the interlayer interface constraints make the current density component perpendicular to the interlayer interface continuous, and the current density upper limit constraints are defined according to the process parameters of each layer;

[0013] Establishing a relationship equation between the two-dimensional magnetic field distribution data and the current density distribution, discretizing the relationship equation into a linear equation group, constructing an objective function including the measured magnetic field and the calculated magnetic field error terms, performing conjugate gradient iteration on the objective function, determining the iteration step size through one-dimensional search and updating the search direction, and obtaining a current distribution diagram of the integrated circuit to be tested;

[0014] Establishing a mapping relationship between the measured current of the main acquisition channel and the current distribution diagram, wherein the mapping relationship includes a proportionality coefficient and a measurement error; using a recursive least squares algorithm to update the proportionality coefficient in real time, substituting the updated proportionality coefficient into the mapping relationship, and calibrating and compensating the measured current of the main acquisition channel in combination with the current distribution under the nominal working state to obtain calibrated and compensated current data;

[0015] Selecting a wavelet basis function with a compact support property, constructing a wavelet packet dictionary containing basis functions of different scales and different frequency bands, and calculating the wavelet coefficients of the current data after calibration and compensation;

[0016] An optimal wavelet basis is determined based on information entropy, and the calibrated and compensated current data is decomposed in multiple scales using the optimal wavelet basis to obtain detail components and approximate components in different frequency bands. The detail components and approximate components are arranged in scale order to construct a characteristic coefficient matrix.

[0017] In an optional embodiment,

[0018] Determine the optimal wavelet basis based on information entropy, use the optimal wavelet basis to perform multi-scale decomposition on the calibrated and compensated current data, obtain detail components and approximate components in different frequency bands, and arrange the detail components and approximate components in scale order to construct a characteristic coefficient matrix, including:

[0019] Constructing a candidate wavelet basis set, and calculating the wavelet entropy corresponding to each basis function in the candidate wavelet basis set, wherein the wavelet entropy is calculated by the negative sum of the square and logarithm product of the wavelet coefficient;

[0020] Arrange the basis functions in the candidate wavelet basis set in ascending order according to wavelet entropy, select the basis functions with the highest order to construct a new candidate basis set, and perform orthogonal processing on the basis functions in the new candidate basis set;

[0021] Repeat the wavelet entropy calculation, ascending order arrangement and orthogonalization processing until the wavelet entropy change value corresponding to the basis function in the new candidate basis set is less than a preset threshold, thereby obtaining an optimal wavelet basis;

[0022] Performing wavelet decomposition on the current data using the optimal wavelet basis to obtain multi-layer detail coefficients and approximate coefficients, calculating the correlation between detail components of adjacent layers, determining the number of decomposition layers according to the correlation, stopping the decomposition when the correlation is less than a set correlation threshold, and obtaining the final detail components and approximate components;

[0023] The detail components and approximate components are constructed into an initial characteristic coefficient matrix in hierarchical order, each row of the characteristic coefficient matrix is ​​normalized, the mean and standard deviation of each row are calculated, and the characteristic coefficient matrix is ​​standardized using the mean and standard deviation to obtain the final characteristic coefficient matrix.

[0024] In an optional embodiment,

[0025] The time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted through a three-dimensional convolution operation and input into a graph attention network. The physical structure information of the integrated circuit is combined with the current features by using the graph attention network to construct a dynamic feature map. The abnormal current features are identified and the abnormal area is located through multi-layer graph convolution operations, including:

[0026] The time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted by a three-dimensional convolution operation, the importance of the features is weighted by an adaptive weight adjustment module, a feature fusion coefficient is calculated according to the mutual information between the features, and feature splicing is performed based on the feature fusion coefficient to obtain a fusion feature;

[0027] Constructing a heterogeneous graph based on the physical topological structure of the integrated circuit layout, constructing the functional modules, signal paths and power networks in the integrated circuit as vertices, establishing directed edge connections according to the physical connection relationship, and concatenating the fusion feature with the vertex type code as the initial feature of the vertex;

[0028] Construct a multi-head attention mechanism for bidirectional propagation, construct a feature transformation matrix based on the initial features of the vertex type, adjust the feature propagation strength through the feature transformation matrix and the edge weight matrix, and calculate the attention coefficient between vertices by combining the self-attention features and mutual attention features of the vertices;

[0029] Generate a local subgraph representation for the vertices in the heterogeneous graph based on a random walk algorithm, and fuse the local subgraph representation with the global features corresponding to the attention coefficients between the vertices to obtain enhanced vertex features;

[0030] Constructing a multi-layer graph convolutional network, inputting the enhanced vertex features into the multi-layer graph convolutional network, fusing local current features and global current features in each layer of graph convolution operation, transferring the current features through residual connections, and obtaining vertex current features;

[0031] The vertex current feature is input into the dynamic threshold detection module, and the abnormal threshold is calculated based on the distribution density of the vertex current feature to identify the abnormal vertices. At the same time, the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area, and the detection result of the current anomaly is output.

[0032] In an optional embodiment,

[0033] The vertex current feature is input into the dynamic threshold detection module, and the abnormal threshold is calculated based on the distribution density of the vertex current feature to identify the abnormal vertex. At the same time, the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area. The detection results of the output current abnormality include:

[0034] Acquire vertex current feature vectors, determine a cutoff distance based on a K-nearest neighbor search, substitute the cutoff distance into a Gaussian kernel function to calculate a similarity matrix between vertices, and perform row normalization on the similarity matrix to obtain a local density of vertex current feature vectors;

[0035] Calculate the Euclidean distance between vertices to obtain the vertex distance matrix. For each vertex, determine the target vertex set whose local density is greater than the current vertex. Extract the distance values ​​between the current vertex and each vertex in the target vertex set from the vertex distance matrix, and obtain the minimum value among the distance values ​​as the relative distance of the current vertex. Take the product of the local density of the vertex and the relative distance as the abnormality score of the current vertex.

[0036] A kernel density estimation method with an adaptive kernel bandwidth is used to perform probability density modeling on the anomaly score, a dynamic threshold is calculated in combination with the mean and standard deviation of the anomaly score, and vertices with anomaly scores lower than the dynamic threshold are marked as anomaly candidate points;

[0037] Based on the spatial distribution of the abnormal candidate points, a local neighborhood density matrix is ​​constructed, and the abnormal candidate points are clustered using a density clustering method to remove noise points to obtain an abnormal vertex set;

[0038] Calculate the attention weight between the abnormal vertex and other vertices, input the current feature vector of the abnormal vertex into the radial basis kernel function to obtain feature similarity, multiply the feature similarity and the attention weight to obtain the propagation strength between nodes, and construct an abnormal propagation transfer matrix according to the propagation strength;

[0039] Perform random walk iterative operation on the abnormal propagation transfer matrix, introduce a restart probability in each iteration to return to the initial abnormal vertex, record the propagation value of each vertex after each iteration, stop the iteration when the difference between the propagation values ​​of two adjacent iterations is less than a preset threshold, and use the final propagation value as the abnormality degree of each vertex;

[0040] Calculate the gradient value of the abnormality degree to the vertex current feature vector, calculate the sensitivity of the vertex current feature according to the gradient value, weight the sensitivity and the abnormality degree to obtain the vertex abnormality contribution value, select the vertex with the largest abnormality contribution value as the initial point, calculate the physical distance between other vertices and the initial point and the degree of difference in abnormality contribution values, merge the vertices whose physical distance is less than a first preset threshold and whose abnormality contribution value difference is less than a second preset threshold into an abnormal area, and output the positioning result of the abnormal area.

[0041] In an optional embodiment,

[0042] In the multi-physical field coupling simulation model, the interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established, the corresponding relationship between abnormal current characteristics and circuit physical parameters is determined through iterative calculation, and the fault location results and fault development trend prediction are output by combining the comparative analysis of measured data and simulation results, including:

[0043] Collecting current data and temperature data of the integrated circuit as initial measurement data, wherein the initial measurement data includes initial current distribution and initial temperature distribution;

[0044] Establishing a multi-physics field coupling model for an integrated circuit, wherein the multi-physics field coupling model includes a current distribution model, a temperature distribution model, and an electromagnetic field distribution model; setting the conductivity as a function of temperature, setting the Joule heat power density and the external heat source power density as heat source terms of the temperature field, setting the magnetic field force as an influence term of the current distribution, and obtaining a coupling equation group describing the interaction of physical fields;

[0045] Substituting the initial temperature distribution into the coupled equations for solving, obtaining the current distribution result, the temperature distribution result and the magnetic field distribution result through iterative calculation; comparing the obtained current distribution result, the temperature distribution result and the magnetic field distribution result with the initial measurement data to obtain the current deviation value and the temperature deviation value;

[0046] Establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining actual physical parameters of the circuit through the physical parameter inversion model;

[0047] Matching the actual physical parameters with a pre-established fault feature library to obtain an initial fault state, wherein the initial fault state includes fault location information and fault physical features;

[0048] Establish a fault state transfer model based on the actual physical parameters and the initial fault state, use the fault state transfer model to predict the fault state quantity and actual physical parameters at the next moment, substitute the physical parameters corresponding to the predicted state into the multi-physical field coupling model, and calculate the current distribution prediction value, temperature distribution prediction value and magnetic field distribution prediction value under the predicted state;

[0049] The predicted current distribution value, temperature distribution predicted value and magnetic field distribution predicted value under the predicted state are compared with the measured data to obtain the predicted deviation. The confidence interval of the predicted result is determined based on the predicted deviation, and the predicted result containing the fault location information, physical characteristics and development trend is output.

[0050] In an optional embodiment,

[0051] Establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining the actual physical parameters of the circuit through the physical parameter inversion model includes:

[0052] Collecting current deviation information and temperature deviation information of the integrated circuit, extracting current density gradient, current direction deflection angle, current fluctuation amplitude and phase delay according to the current deviation information to construct a current deviation feature vector, and extracting temperature gradient, thermal diffusion coefficient and thermal impedance according to the temperature deviation information to construct a temperature deviation feature vector;

[0053] Performing Fourier transform on the temperature deviation feature vector to obtain a temperature frequency domain feature, and combining the temperature frequency domain feature with the current deviation feature vector to form a multi-dimensional feature vector;

[0054] Using the multidimensional feature vector to train a multi-scale feature extraction network, the multi-scale feature extraction network extracts local features and global features through a residual connection structure, and calculates weight coefficients of the local features and the global features through an attention mechanism;

[0055] Constructing an Ohm's law constraint function, a Joule heat constraint function, and an electromagnetic induction constraint function into a physical constraint module, and combining the physical constraint module with the loss function of the multi-scale feature extraction network to form a constraint loss function;

[0056] Calculating the influence factor of the physical parameter on the current deviation information and the temperature deviation information, and combining the influence factor with the constraint loss function to construct a parameter sensitivity model;

[0057] Based on the parameter sensitivity model, multiple inversion models are trained, and the output results of the multiple inversion models are dynamically weighted to determine the actual physical parameters of the circuit.

[0058] According to a second aspect of the embodiments of the present invention,

[0059] Provided is an intelligent system for monitoring and analyzing integrated circuit package current, comprising:

[0060] The first unit is used to obtain the current timing data of the integrated circuit to be tested under different working states through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel obtains the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, and reconstructs the current distribution map of the two-dimensional magnetic field distribution data through an electromagnetic field inversion algorithm, calibrates and compensates the sampling data of the main acquisition channel according to the current distribution map, and decomposes the calibrated sampling data into a characteristic coefficient matrix through an adaptive wavelet transform;

[0061] The second unit is used to construct a deep feature analysis model based on the feature coefficient matrix, wherein the deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer and a space dimension convolution layer, extracts the time domain features, frequency domain features and space features in the feature coefficient matrix through a three-dimensional convolution operation, and inputs them into a graph attention network, uses the graph attention network to combine the physical structure information of the integrated circuit with the current features to construct a dynamic feature map, and identifies abnormal current features and locates abnormal areas through multi-layer graph convolution operations;

[0062] The third unit is used to input the abnormal current characteristics and abnormal area information into the digital twin analysis system, which includes a physical model of the integrated circuit and a multi-physical field coupling simulation model. The interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model, and the correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the fault development trend prediction are output, and a diagnostic analysis report is generated.

[0063] According to a third aspect of the embodiments of the present invention,

[0064] An electronic device is provided, comprising:

[0065] processor;

[0066] a memory for storing processor-executable instructions;

[0067] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0068] According to a fourth aspect of the embodiments of the present invention,

[0069] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0070] In this embodiment, through a dual-channel parallel acquisition system, combined with the lossless replication of the current mirror circuit and the two-dimensional magnetic field distribution data of the array magnetic sensor, and using the electromagnetic field inversion algorithm to reconstruct the current distribution map, the accurate measurement and calibration of the integrated circuit working current is achieved. This method can not only capture weak current changes, but also provide spatial distribution information of the current, overcome the limitations of the traditional single-point measurement method, and improve the accuracy and comprehensiveness of current monitoring. Based on the deep feature analysis model of the improved three-dimensional convolutional neural network and the graph attention network, the time domain, frequency domain and spatial characteristics of the current signal can be extracted, and the dynamic feature map can be constructed in combination with the physical structure information of the integrated circuit. This multi-dimensional feature extraction and analysis method can more effectively identify abnormal current characteristics and accurately locate abnormal areas, thereby improving the accuracy of fault diagnosis. The digital twin analysis system establishes the interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution through the physical model of the integrated circuit and the multi-physical field coupling simulation model. Combining the measured data and simulation results, we can deeply analyze the correspondence between abnormal current characteristics and circuit physical parameters, which can not only accurately locate the fault, but also predict the development trend of the fault. This makes fault analysis no longer limited to simple fault location, but can provide a more comprehensive diagnostic analysis report, which helps to prevent potential fault risks in advance. BRIEF DESCRIPTION OF THE DRAWINGS

[0071] Figure 1 A schematic diagram of a flow chart of an intelligent method for monitoring and analyzing integrated circuit package current according to an embodiment of the present invention;

[0072] Figure 2 Schematic diagram of the structure of an intelligent system for monitoring and analyzing integrated circuit package current according to an embodiment of the present invention. DETAILED DESCRIPTION

[0073] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0074] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0075] Figure 1 FIG. 1 is a flow chart of an intelligent method for monitoring and analyzing integrated circuit package current according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0076] S101. Acquire the current timing data of the integrated circuit to be tested under different working states through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel acquires the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, reconstructs the current distribution map of the two-dimensional magnetic field distribution data through an electromagnetic field inversion algorithm, calibrates and compensates the sampling data of the main acquisition channel according to the current distribution map, and decomposes the calibrated sampling data into a characteristic coefficient matrix through an adaptive wavelet transform;

[0077] S102. Constructing a deep feature analysis model based on the feature coefficient matrix, the deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer and a space dimension convolution layer, extracting the time domain features, frequency domain features and space features in the feature coefficient matrix through a three-dimensional convolution operation, and inputting them into a graph attention network, using the graph attention network to combine the physical structure information of the integrated circuit with the current features to construct a dynamic feature map, and identifying abnormal current features and locating abnormal areas through multi-layer graph convolution operations;

[0078] S103. Input the abnormal current characteristics and abnormal area information into the digital twin analysis system, which includes a physical model of the integrated circuit and a multi-physical field coupling simulation model. In the multi-physical field coupling simulation model, an interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established. The correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the fault development trend prediction are output, and a diagnostic analysis report is generated.

[0079] First, prepare the IC chip samples to be tested and build a test platform. The platform includes a dual-channel parallel acquisition system, in which the main acquisition channel consists of a high-precision current sampling module and a current mirror circuit, and the auxiliary acquisition channel consists of an array magnetic sensor and a data acquisition module. The current mirror circuit is constructed with high-bandwidth, low-noise operational amplifiers and precision resistors to ensure lossless replication of current signals. The array magnetic sensor uses a giant magnetoresistance (GMR) sensor or a tunnel magnetoresistance (TMR) sensor to achieve high-sensitivity measurement of the magnetic field on the chip surface.

[0080] Next, a preset test stimulus is applied to the integrated circuit to put it in different working states, such as standby, working, peak load, etc. A dual-channel parallel acquisition system is used to synchronously acquire current timing data and magnetic field distribution data. The main acquisition channel acquires the working current of the circuit to be tested in real time and converts it into a voltage signal for digital storage. The auxiliary acquisition channel acquires the two-dimensional magnetic field distribution data on the chip surface. Then, the acquired two-dimensional magnetic field distribution data is processed. The magnetic field data is reconstructed into a current distribution map using an electromagnetic field inversion algorithm. This can be achieved using finite element analysis software or other professional electromagnetic field simulation tools. For example, the chip surface can be divided into 100x100 grids, and the corresponding current density is calculated based on the magnetic field strength value of each grid to obtain a current distribution map. According to the reconstructed current distribution map, the current sampling data acquired by the main acquisition channel is calibrated and compensated to eliminate the errors caused by the current mirror circuit and the test system. For example, if the current distribution map shows that there is current leakage in the edge area of ​​the chip, the data acquired by the main channel can be corrected according to the magnitude of the leakage current.

[0081] Next, the calibrated current sampling data is feature extracted. Adaptive wavelet transform is used to decompose the current time series data into multi-level wavelet coefficients and construct a characteristic coefficient matrix. Adaptive wavelet transform can automatically select the best wavelet basis function and decomposition level according to the characteristics of the current signal. For example, the current signal can be decomposed into 5 levels, each level corresponds to a different frequency range, so as to extract the time-frequency characteristics of the current signal.

[0082] Based on the extracted feature coefficient matrix, a deep feature analysis model is constructed. The model adopts an improved three-dimensional convolutional neural network (3D-CNN) structure, which includes a time dimension convolution layer, a frequency dimension convolution layer, and a space dimension convolution layer. The time dimension convolution layer is used to extract the time domain features of the current signal, the frequency dimension convolution layer is used to extract the frequency domain features of the current signal, and the space dimension convolution layer is used to extract the spatial features of the current signal. For example, a 3x3x3 convolution kernel can be used to perform a convolution operation on the feature coefficient matrix to extract features in different dimensions.

[0083] The features extracted by the 3D convolution operation are input into the Graph Attention Network (GAT). Using the Graph Attention Network, the physical structure information of the integrated circuit (such as circuit topology, component layout, etc.) is combined with the current features to construct a dynamic feature map. Through multi-layer graph convolution operations, abnormal current features are identified and abnormal areas are located. For example, if the current feature of a node is significantly different from the normal situation, it can be marked as an abnormal node, and the abnormal area can be determined based on the location information of the node.

[0084] The identified abnormal current characteristics and abnormal area information are input into the digital twin analysis system. The system includes a physical model of the integrated circuit and a multi-physics coupling simulation model. In the multi-physics coupling simulation model, the interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established.

[0085] Through iterative calculation, the corresponding relationship between abnormal current characteristics and circuit physical parameters is determined. For example, the parameters in the circuit model can be adjusted to make the simulation results as consistent as possible with the measured data, thereby determining the physical parameters corresponding to the abnormal current characteristics. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and fault development trend prediction are output, and a diagnostic analysis report is generated.

[0086] In this embodiment, through dual-channel parallel acquisition and electromagnetic field inversion technology, the current distribution of the chip can be measured more accurately and the errors caused by traditional current monitoring methods can be eliminated. Through the deep feature analysis model and digital twin analysis system, abnormal current characteristics can be identified, the fault area can be accurately located, and the fault development trend can be predicted. Through early warning and diagnosis of faults, potential fault risks can be effectively prevented and the reliability and safety of integrated circuits can be improved.

[0087] In an optional implementation, the two-dimensional magnetic field distribution data is reconstructed into a current distribution map through an electromagnetic field inversion algorithm, the sampling data of the main acquisition channel is calibrated and compensated according to the current distribution map, and the calibrated sampling data is decomposed into a characteristic coefficient matrix through adaptive wavelet transform, including:

[0088] Dividing the integrated circuit to be tested into a metal interconnect layer, a device layer and a substrate layer, applying current continuity constraints, interlayer interface constraints and current density upper limit constraints to the metal interconnect layer, the device layer and the substrate layer, wherein the current continuity constraints make the current density divergence in each layer zero, the interlayer interface constraints make the current density component perpendicular to the interlayer interface continuous, and the current density upper limit constraints are defined according to the process parameters of each layer;

[0089] Establishing a relationship equation between the two-dimensional magnetic field distribution data and the current density distribution, discretizing the relationship equation into a linear equation group, constructing an objective function including the measured magnetic field and the calculated magnetic field error terms, performing conjugate gradient iteration on the objective function, determining the iteration step size through one-dimensional search and updating the search direction, and obtaining a current distribution diagram of the integrated circuit to be tested;

[0090] Establishing a mapping relationship between the measured current of the main acquisition channel and the current distribution diagram, wherein the mapping relationship includes a proportionality coefficient and a measurement error; using a recursive least squares algorithm to update the proportionality coefficient in real time, substituting the updated proportionality coefficient into the mapping relationship, and calibrating and compensating the measured current of the main acquisition channel in combination with the current distribution under the nominal working state to obtain calibrated and compensated current data;

[0091] Selecting a wavelet basis function with a compact support property, constructing a wavelet packet dictionary containing basis functions of different scales and different frequency bands, and calculating the wavelet coefficients of the current data after calibration and compensation;

[0092] An optimal wavelet basis is determined based on information entropy, and the calibrated and compensated current data is decomposed in multiple scales using the optimal wavelet basis to obtain detail components and approximate components in different frequency bands. The detail components and approximate components are arranged in scale order to construct a characteristic coefficient matrix.

[0093] In integrated circuit testing, due to the influence of process deviation, environmental noise and other factors, the measured current of the main acquisition channel often has errors, affecting the accuracy of the test results. This embodiment provides an integrated circuit current calibration and feature extraction method based on magnetic field inversion and adaptive wavelet transform, which can effectively improve the accuracy and efficiency of current measurement.

[0094] First, the integrated circuit to be tested is layered and divided into metal interconnect layer, device layer and substrate layer. A series of physical constraints are imposed on each layer. The current continuity constraint ensures that the flow of current in each layer complies with the laws of physics, the interlayer interface constraint ensures the continuity of the current transition between layers, and the current density upper limit constraint sets the maximum value of the current density according to the process parameters of each layer to prevent unreasonable current distribution. For example, the upper limit of the current density of the metal interconnect layer can be set to 10^7A / cm², the device layer to 10^6A / cm², and the substrate layer to 10^5A / cm².

[0095] Next, the two-dimensional magnetic field distribution data around the integrated circuit to be tested is collected through a specific device. Then, the relationship equation between the two-dimensional magnetic field distribution data and the current density distribution is established. The equation is discretized to form a linear system of equations. In order to solve this system of equations, an objective function is constructed, which contains the error terms of the measured magnetic field and the calculated magnetic field. The conjugate gradient iteration method is used to solve the objective function, and the step size and search direction of each iteration are determined by one-dimensional search, and finally the current distribution diagram of the integrated circuit to be tested is obtained. For example, assuming that the measured magnetic field data is B=[0.1, 0.2, 0.3], and the initial current density is J=[1, 1, 1], after multiple iterations, J=[1.2, 1.8, 2.7] is finally obtained.

[0096] Then, a mapping relationship between the measured current of the main acquisition channel and the current distribution diagram is established. This mapping relationship includes a proportional coefficient and a measurement error. The proportional coefficient is updated in real time using a recursive least squares algorithm, and the updated proportional coefficient is substituted into the mapping relationship. The measured current of the main acquisition channel is calibrated and compensated in combination with the current distribution under the nominal working state.

[0097] In order to further extract the characteristics of the current data, a wavelet basis function with compact support properties is selected. A wavelet packet dictionary containing different scales and wavelet basis functions is constructed. For example, Daubechies wavelet is selected as the basis function, and a wavelet packet dictionary containing scales 1 to 4 and frequency bands 1 to 4 is constructed. The wavelet coefficients of the current data after calibration compensation are calculated.

[0098] According to the information entropy criterion, the optimal wavelet basis is selected from the constructed wavelet packet dictionary. The current data after calibration and compensation is decomposed at multiple scales using the selected optimal wavelet basis to obtain detail components and approximate components in different frequency bands. For example, the current data is decomposed into 4 scales to obtain 4 detail components and 1 approximate component. Finally, these detail components and approximate components are arranged in scale order to construct the characteristic coefficient matrix. Assuming that the decomposed approximate component is A4=[1, 2, 3], the detail component is D4=[0.1, 0.2, 0.3], D3=[0.2, 0.4, 0.6], D2=[0.3, 0.6, 0.9], D1=[0.4, 0.8, 1.2], the characteristic coefficient matrix is ​​[A4, D4, D3, D2, D1].

[0099] In this embodiment, the current distribution is reconstructed by the magnetic field inversion algorithm, and the current distribution under the nominal working state is combined for calibration compensation, which effectively eliminates the measurement error and improves the accuracy of current measurement. The adaptive wavelet transform can select the optimal wavelet basis according to the characteristics of the current data, effectively extract the characteristic information of the current data, and construct the characteristic coefficient matrix, providing a reliable data basis for subsequent fault diagnosis and analysis. The recursive least squares algorithm is used to update the proportional coefficient in real time, which reduces the amount of calculation and improves the test efficiency.

[0100] In an optional implementation, determining an optimal wavelet basis based on information entropy, using the optimal wavelet basis to perform multi-scale decomposition on the calibrated and compensated current data to obtain detail components and approximate components in different frequency bands, and arranging the detail components and approximate components in scale order to construct a characteristic coefficient matrix includes:

[0101] Constructing a candidate wavelet basis set, and calculating the wavelet entropy corresponding to each basis function in the candidate wavelet basis set, wherein the wavelet entropy is calculated by the negative sum of the square and logarithm product of the wavelet coefficient;

[0102] Arrange the basis functions in the candidate wavelet basis set in ascending order according to wavelet entropy, select the basis functions with the highest order to construct a new candidate basis set, and perform orthogonal processing on the basis functions in the new candidate basis set;

[0103] Repeat the wavelet entropy calculation, ascending order arrangement and orthogonalization processing until the wavelet entropy change value corresponding to the basis function in the new candidate basis set is less than a preset threshold, thereby obtaining an optimal wavelet basis;

[0104] Performing wavelet decomposition on the current data using the optimal wavelet basis to obtain multi-layer detail coefficients and approximate coefficients, calculating the correlation between detail components of adjacent layers, determining the number of decomposition layers according to the correlation, stopping the decomposition when the correlation is less than a set correlation threshold, and obtaining the final detail components and approximate components;

[0105] The detail components and approximate components are constructed into an initial characteristic coefficient matrix in hierarchical order, each row of the characteristic coefficient matrix is ​​normalized, the mean and standard deviation of each row are calculated, and the characteristic coefficient matrix is ​​standardized using the mean and standard deviation to obtain the final characteristic coefficient matrix.

[0106] Exemplarily, a section of motor running current data is collected with a sampling frequency of 10 kHz, a total of 10,000 data points are collected, and the value range is between 0 and 10A.

[0107] Construct a candidate wavelet basis set. Select commonly used wavelet basis functions, such as Daubechies wavelet (dbN), Symlets wavelet (symN), Coiflets wavelet (coifN), etc., where N represents the order of the wavelet. For example, select db4, db8, sym8, and coif5 as candidate wavelet basis functions.

[0108] Calculate the wavelet entropy corresponding to each candidate wavelet basis. Taking the db4 wavelet as an example, first use the db4 wavelet to perform wavelet decomposition on the current data to obtain the detail coefficients and approximate coefficients of each layer. Then, calculate the square value of each coefficient and multiply the square value by the logarithm of the coefficient. Finally, take the negative value of the sum of the products of all coefficients to obtain the wavelet entropy corresponding to the db4 wavelet. Perform the same calculation for other candidate wavelet bases. Assume that the wavelet entropies of db4, db8, sym8, and coif5 are calculated to be 0.8, 0.7, 0.9, and 0.6, respectively.

[0109] Arrange the candidate wavelet bases in ascending order according to the wavelet entropy. According to the wavelet entropy calculated in the previous step, the candidate wavelet bases are sorted as coif5, db8, db4, and sym8.

[0110] Select the top-ranked basis functions to construct a new candidate basis set. Select several top-ranked basis functions, such as coif5 and db8.

[0111] Orthogonalize the basis functions in the new candidate basis set. Use the Gram-Schmidt orthogonalization method to orthogonalize coif5 and db8 to obtain two orthogonal wavelet basis functions.

[0112] Repeat the wavelet entropy calculation, ascending order arrangement and orthogonalization processing until the wavelet entropy change value corresponding to the basis function in the new candidate basis set is less than the preset threshold. For example, the preset threshold is 0.01. Repeat the above steps, assuming that the final wavelet basis is the orthogonalized coif5, whose wavelet entropy is 0.58, and the difference with the wavelet entropy of the previous round of 0.6 is less than 0.01, which meets the conditions. Therefore, the orthogonalized coif5 is selected as the optimal wavelet basis.

[0113] The optimal wavelet basis is used to perform wavelet decomposition on the current data. The orthogonalized coif5 wavelet is used to perform multi-scale decomposition on the current data to obtain multi-layer detail coefficients and approximate coefficients.

[0114] Calculate the correlation between detail components of adjacent layers. For example, calculate the correlation between the detail coefficients of the first layer and the second layer using the Pearson correlation coefficient.

[0115] Determine the number of decomposition layers based on the correlation. Set a correlation threshold, such as 0.1. If the correlation between detail coefficients of adjacent layers is less than the threshold, stop decomposition. Assume that when decomposing to the 4th layer, the correlation between the detail coefficients of the 4th layer and the 3rd layer is 0.08, which is less than the threshold, then stop decomposition, and finally get 4 layers of detail coefficients and 1 approximate coefficient.

[0116] The detail components and approximate components are constructed in hierarchical order to construct the initial feature coefficient matrix. The 4-layer detail coefficients and 1 approximate coefficient are arranged in hierarchical order into a matrix, with each row representing the coefficient of one layer.

[0117] Normalize each row of the feature coefficient matrix. Divide the coefficient of each row by the maximum absolute value of the row so that the coefficient of each row falls between -1 and 1.

[0118] Compute the mean and standard deviation of each row. Compute the mean and standard deviation of each row of the eigenvalue matrix.

[0119] The characteristic coefficient matrix is ​​standardized using the mean and standard deviation. The coefficient of each row is subtracted from the mean of the row and then divided by the standard deviation of the row to obtain the final characteristic coefficient matrix.

[0120] In this embodiment, by selecting the optimal wavelet basis based on information entropy, the key information in the current signal can be better captured, thereby improving the accuracy of feature extraction. The use of wavelet multi-scale decomposition and standardization processing can effectively reduce the impact of noise and outliers and enhance the robustness of features. Determining the number of decomposition layers by correlation can avoid over-decomposition, reduce computational complexity, and improve analysis efficiency.

[0121] In an optional implementation, the time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted by a three-dimensional convolution operation and input into a graph attention network, the graph attention network combines the physical structure information of the integrated circuit with the current features to construct a dynamic feature map, and the abnormal current features are identified and the abnormal area is located by multi-layer graph convolution operations, including:

[0122] The time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted by a three-dimensional convolution operation, the importance of the features is weighted by an adaptive weight adjustment module, a feature fusion coefficient is calculated according to the mutual information between the features, and feature splicing is performed based on the feature fusion coefficient to obtain a fusion feature;

[0123] Constructing a heterogeneous graph based on the physical topological structure of the integrated circuit layout, constructing the functional modules, signal paths and power networks in the integrated circuit as vertices, establishing directed edge connections according to the physical connection relationship, and concatenating the fusion feature with the vertex type code as the initial feature of the vertex;

[0124] Construct a multi-head attention mechanism for bidirectional propagation, construct a feature transformation matrix based on the initial features of the vertex type, adjust the feature propagation strength through the feature transformation matrix and the edge weight matrix, and calculate the attention coefficient between vertices by combining the self-attention features and mutual attention features of the vertices;

[0125] Generate a local subgraph representation for the vertices in the heterogeneous graph based on a random walk algorithm, and fuse the local subgraph representation with the global features corresponding to the attention coefficients between the vertices to obtain enhanced vertex features;

[0126] Constructing a multi-layer graph convolutional network, inputting the enhanced vertex features into the multi-layer graph convolutional network, fusing local current features and global current features in each layer of graph convolution operation, transferring the current features through residual connections, and obtaining vertex current features;

[0127] The vertex current feature is input into the dynamic threshold detection module, and the abnormal threshold is calculated based on the distribution density of the vertex current feature to identify the abnormal vertices. At the same time, the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area, and the detection result of the current anomaly is output.

[0128] Get the current data of the IC to be tested and convert it into a characteristic coefficient matrix. For example, collect the current data of the chip at 100 time points under different working conditions, each time point contains the current information of 50 frequency bands, and the current distribution of 10x10 regions on the chip, forming a 100x50x10x10 characteristic coefficient matrix.

[0129] The three-dimensional convolution operation is used to extract the time domain, frequency domain and spatial features in the feature coefficient matrix. For example, three different three-dimensional convolution kernels are used to extract the features of the time dimension, frequency dimension and spatial dimension respectively. The first convolution kernel size is 3x1x1, the step size is 1, and the time series features are extracted; the second convolution kernel size is 1x3x1, the step size is 1, and the frequency domain features are extracted; the third convolution kernel size is 1x1x3, the step size is 1, and the spatial features are extracted.

[0130] The extracted time domain, frequency domain and space features are weighted in importance. For example, an adaptive weight adjustment module is used to dynamically adjust the weight of each feature according to its contribution in anomaly detection. Assume that the initial weights of time domain features, frequency domain features and space features are 0.4, 0.3 and 0.3 respectively. By analyzing the training data, it is found that the time domain features are more important for identifying specific types of abnormal currents, so the weights are dynamically adjusted to 0.5, 0.25 and 0.25.

[0131] Calculate the mutual information between features and calculate the feature fusion coefficient based on the mutual information. For example, calculate the mutual information between time domain features, frequency domain features, and spatial features to obtain a 3x3 mutual information matrix. Then, normalize the mutual information matrix to obtain the feature fusion coefficient. Assuming that the mutual information of time domain features and frequency domain features is the highest, their corresponding fusion coefficients are also the highest.

[0132] Based on the feature fusion coefficient, feature splicing is performed to obtain a fusion feature. For example, the weighted time domain, frequency domain and space features are spliced ​​according to the calculated fusion coefficient to form a new feature vector as the fusion feature.

[0133] Construct a heterogeneous graph based on the physical topology of the integrated circuit layout. Construct the functional modules, signal paths, and power networks in the integrated circuit as vertices, and establish directed edge connections based on the physical connection relationships. For example, functional modules such as CPU, memory, and bus are used as vertices, and the connection relationships between them are used as directed edges.

[0134] The fused features are concatenated with the vertex type code as the initial features of the vertex. For example, the types of CPU, memory, and bus are encoded as 0, 1, and 2 respectively, and the fused features are concatenated with the corresponding type codes to form the initial feature vector of each vertex.

[0135] Build a multi-head attention mechanism for bidirectional propagation. Build a feature transformation matrix based on the initial features of the vertices. For example, use a linear transformation to map the initial feature vector to multiple different feature spaces, each of which corresponds to an attention head.

[0136] The feature propagation strength is adjusted by the feature transformation matrix and the edge weight matrix. For example, the feature transformation matrix is ​​multiplied by the edge weight matrix to obtain the adjusted feature propagation matrix.

[0137] The attention coefficients between vertices are calculated by combining the self-attention features and mutual attention features of the vertices. For example, the attention weights between each vertex and other vertices are calculated to indicate the degree of association between them.

[0138] A local subgraph representation is generated for the vertices in the heterogeneous graph based on a random walk algorithm. For example, starting from each vertex, a random walk is performed to generate a local subgraph containing its neighboring nodes.

[0139] The local subgraph representation is fused with the global feature corresponding to the attention coefficient between vertices to obtain the enhanced vertex feature. For example, the feature vector of the local subgraph is weighted and summed with the global attention coefficient to obtain the enhanced vertex feature.

[0140] Construct a multi-layer graph convolutional network. Input the enhanced vertex features into the multi-layer graph convolutional network. For example, use a two-layer graph convolutional network, where each layer aggregates and updates neighbor node information.

[0141] In each layer of graph convolution operation, local current features and global current features are fused. For example, the current features of the vertex itself and the current features of its neighboring nodes are weighted averaged as the fused current features.

[0142] The current features are transmitted through residual connections. For example, the current features of the previous layer are added to the current features of the current layer as the final current features.

[0143] The vertex current features are input into the dynamic threshold detection module. The abnormal threshold is calculated based on the distribution density of the vertex current features to identify abnormal vertices. For example, according to the statistical distribution of the current features, a dynamic threshold is set, and vertices exceeding the threshold are identified as abnormal vertices.

[0144] At the same time, the attention coefficient between vertices and the gradient information of vertex current features are used to locate abnormal areas. For example, the propagation path and impact range of abnormal current are determined based on the attention coefficient and current gradient.

[0145] Output the detection result of current anomaly. For example, output the ID, location, degree of anomaly and other information of the abnormal vertex.

[0146] In this embodiment, the accuracy of abnormal current detection is improved: by combining the physical structure information of the integrated circuit and the current characteristics, the abnormal current can be identified more accurately, reducing false positives and false negatives. By using the graph attention network and the dynamic threshold detection module, the area where the abnormal current is located can be accurately located, providing more accurate information for subsequent fault diagnosis. Through multi-layer graph convolutional networks and residual connections, current characteristics can be effectively extracted and transmitted, improving detection efficiency.

[0147] In an optional implementation, the vertex current feature is input into a dynamic threshold detection module, an abnormal threshold is calculated based on the distribution density of the vertex current feature to identify abnormal vertices, and the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area, and the detection result of the output current abnormality includes:

[0148] Acquire vertex current feature vectors, determine a cutoff distance based on a K-nearest neighbor search, substitute the cutoff distance into a Gaussian kernel function to calculate a similarity matrix between vertices, and perform row normalization on the similarity matrix to obtain a local density of vertex current feature vectors;

[0149] Calculate the Euclidean distance between vertices to obtain the vertex distance matrix. For each vertex, determine the target vertex set whose local density is greater than the current vertex. Extract the distance values ​​between the current vertex and each vertex in the target vertex set from the vertex distance matrix, and obtain the minimum value among the distance values ​​as the relative distance of the current vertex. Take the product of the local density of the vertex and the relative distance as the abnormality score of the current vertex.

[0150] A kernel density estimation method with an adaptive kernel bandwidth is used to perform probability density modeling on the anomaly score, a dynamic threshold is calculated in combination with the mean and standard deviation of the anomaly score, and vertices with anomaly scores lower than the dynamic threshold are marked as anomaly candidate points;

[0151] Based on the spatial distribution of the abnormal candidate points, a local neighborhood density matrix is ​​constructed, and the abnormal candidate points are clustered using a density clustering method to remove noise points to obtain an abnormal vertex set;

[0152] Calculate the attention weight between the abnormal vertex and other vertices, input the current feature vector of the abnormal vertex into the radial basis kernel function to obtain feature similarity, multiply the feature similarity and the attention weight to obtain the propagation strength between nodes, and construct an abnormal propagation transfer matrix according to the propagation strength;

[0153] Perform random walk iterative operation on the abnormal propagation transfer matrix, introduce a restart probability in each iteration to return to the initial abnormal vertex, record the propagation value of each vertex after each iteration, stop the iteration when the difference between the propagation values ​​of two adjacent iterations is less than a preset threshold, and use the final propagation value as the abnormality degree of each vertex;

[0154] Calculate the gradient value of the abnormality degree to the vertex current feature vector, calculate the sensitivity of the vertex current feature according to the gradient value, weight the sensitivity and the abnormality degree to obtain the vertex abnormality contribution value, select the vertex with the largest abnormality contribution value as the initial point, calculate the physical distance between other vertices and the initial point and the degree of difference in abnormality contribution values, merge the vertices whose physical distance is less than a first preset threshold and whose abnormality contribution value difference is less than a second preset threshold into an abnormal area, and output the positioning result of the abnormal area.

[0155] Exemplarily, the current characteristics of each node in the power system are obtained to form a feature vector. For example, in a power system with 100 nodes, the current characteristics of each node can be the real-time current value, historical current mean, current variance, etc. of the node, thus forming a 100-dimensional current feature vector. Then, a suitable truncation distance is determined using the K nearest neighbor search algorithm. The K nearest neighbor search algorithm means that for each node, the K nearest neighbor nodes are found, and then the average distance from these neighbor nodes to the node is calculated, and this average distance is used as the truncation distance of the node. For example, set K=5, for node A, find the 5 nearest nodes to it, calculate the average distance from these 5 nodes to A, and obtain the truncation distance of A. After calculating the truncation distances of all nodes, substitute these truncation distances into the Gaussian kernel function to calculate the similarity matrix between nodes. The Gaussian kernel function can measure the similarity between the current characteristics of two nodes, and the truncation distance, as a parameter of the Gaussian kernel function, can control the range of similarity calculation. After obtaining the similarity matrix, each row is normalized so that the sum of each row is 1, and the local density of the node current feature vector is obtained. The local density reflects the degree of aggregation of current features around each node.

[0156] According to the current feature vector of each node, the Euclidean distance between nodes is calculated to obtain the node distance matrix. The Euclidean distance can measure the degree of difference between the current characteristics of two nodes. For each node, find all other nodes with a local density greater than the current node to form the target node set. Extract the distance between the current node and each node in the target node set from the node distance matrix, find the minimum value of these distances, and use it as the relative distance of the current node. The relative distance represents the distance from the current node to the area with higher local density. Multiply the local density of each node by the relative distance to obtain the abnormality score of the node. The higher the abnormality score, the more abnormal the current characteristics of the node.

[0157] The kernel density estimation method with adaptive kernel bandwidth is used to model the probability density of the anomaly scores of all nodes. The kernel density estimation method can estimate the probability density function of the data according to the distribution of the data. Adaptive kernel bandwidth refers to automatically adjusting the bandwidth of the kernel function according to the data density, so that the bandwidth is small in places with high density and large in places with low density. According to the probability density distribution of the anomaly scores, the mean and standard deviation of the anomaly scores are calculated. Combined with the mean and standard deviation, the dynamic threshold is calculated. For example, the dynamic threshold can be set to the mean minus two times the standard deviation. Nodes with anomaly scores below the dynamic threshold are marked as anomaly candidates.

[0158] According to the spatial distribution of abnormal candidate points, a local neighborhood density matrix is ​​constructed. The local neighborhood density matrix represents the degree of spatial aggregation of abnormal candidate points. The density clustering method is used to cluster abnormal candidate points. The density clustering method can aggregate points connected by density to form different clusters. Noise points are eliminated, that is, those abnormal candidate points that do not belong to any cluster are eliminated to obtain the final abnormal node set.

[0159] Calculate the attention weight between the abnormal node and other nodes. The attention weight indicates the degree of influence of the abnormal node on other nodes. For example, the attention weight can be calculated using a graph neural network-based method. Input the current feature vector of the abnormal node into the radial basis kernel function to obtain the feature similarity. The radial basis kernel function can measure the similarity of the current features of two nodes. Multiply the feature similarity by the attention weight to obtain the propagation strength between nodes. The propagation strength indicates the strength of the abnormal information of the abnormal node to propagate to other nodes. According to the propagation strength, construct the abnormal propagation transfer matrix.

[0160] Perform random walk iterations on the abnormal propagation transfer matrix. In each iteration, return to the initial abnormal node with a certain probability (restart probability). Record the propagation value of each node after each iteration. The propagation value represents the cumulative degree of propagation of abnormal information in the network. When the difference between the propagation values ​​of two adjacent iterations is less than the preset threshold, stop the iteration. The final propagation value is used as the abnormal degree of each node.

[0161] Calculate the gradient value of the abnormality degree to the node current feature vector. The gradient value represents the sensitivity of the current feature vector to the abnormality degree. Calculate the sensitivity of the node current feature according to the gradient value. Weight the sensitivity and the abnormality degree to obtain the node abnormality contribution value. Select the node with the largest abnormality contribution value as the initial point. Calculate the difference between the physical distance and the abnormality contribution value between other nodes and the initial point. Merge the nodes whose physical distance is less than the first preset threshold and whose abnormality contribution value difference is less than the second preset threshold into an abnormal area. Output the positioning result of the abnormal area. For example, assuming that the abnormality contribution value of node A is the largest, A is used as the initial point. If the physical distance between node B and A is less than the first preset threshold, and the difference between the abnormality contribution values ​​of B and A is less than the second preset threshold, B and A are merged into one abnormal area.

[0162] In this embodiment, multiple factors such as local density, relative distance, abnormal score, abnormal degree and sensitivity of node current characteristics are comprehensively considered, so that abnormal nodes can be identified and abnormal areas can be located more accurately. The kernel density estimation method and density clustering method with adaptive kernel bandwidth are adopted, which can effectively handle the interference of noise data and abnormal data and improve the robustness of the method. This method can clearly explain the causes of abnormal nodes and abnormal areas, for example, which nodes contribute the most to the abnormality, which areas have the most sensitive current characteristics, etc., and provide valuable information for fault diagnosis and maintenance of power systems. For example, by analyzing the current characteristics of abnormal nodes, the type and cause of the fault can be determined.

[0163] In an optional implementation, the interactive influence relationship among current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model, the corresponding relationship between abnormal current characteristics and circuit physical parameters is determined by iterative calculation, and the fault location result and fault development trend prediction are outputted in combination with comparative analysis of measured data and simulation results, including:

[0164] Collecting current data and temperature data of the integrated circuit as initial measurement data, wherein the initial measurement data includes initial current distribution and initial temperature distribution;

[0165] Establishing a multi-physics field coupling model for an integrated circuit, wherein the multi-physics field coupling model includes a current distribution model, a temperature distribution model, and an electromagnetic field distribution model; setting the conductivity as a function of temperature, setting the Joule heat power density and the external heat source power density as heat source terms of the temperature field, setting the magnetic field force as an influence term of the current distribution, and obtaining a coupling equation group describing the interaction of physical fields;

[0166] Substituting the initial temperature distribution into the coupled equations for solving, obtaining the current distribution result, the temperature distribution result and the magnetic field distribution result through iterative calculation; comparing the obtained current distribution result, the temperature distribution result and the magnetic field distribution result with the initial measurement data to obtain the current deviation value and the temperature deviation value;

[0167] Establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining actual physical parameters of the circuit through the physical parameter inversion model;

[0168] Matching the actual physical parameters with a pre-established fault feature library to obtain an initial fault state, wherein the initial fault state includes fault location information and fault physical features;

[0169] Establish a fault state transfer model based on the actual physical parameters and the initial fault state, use the fault state transfer model to predict the fault state quantity and actual physical parameters at the next moment, substitute the physical parameters corresponding to the predicted state into the multi-physical field coupling model, and calculate the current distribution prediction value, temperature distribution prediction value and magnetic field distribution prediction value under the predicted state;

[0170] The predicted current distribution value, temperature distribution predicted value and magnetic field distribution predicted value under the predicted state are compared with the measured data to obtain the predicted deviation. The confidence interval of the predicted result is determined based on the predicted deviation, and the predicted result containing the fault location information, physical characteristics and development trend is output.

[0171] Exemplarily, obtain an integrated circuit to be tested and prepare corresponding measuring equipment. Taking a 16-bit microcontroller as an example, use a high-precision current probe and an infrared thermal imager to collect its current data and temperature data during operation. Assume that the collected initial current distribution data is the current value of each pin, for example, the current of pin 1 is 10mA, the current of pin 2 is 5mA, and so on. The initial temperature distribution data is the temperature value of each point on the surface of the chip, for example, the center temperature of the chip is 40°C, and the temperatures of the four corners are 35°C, 36°C, 37°C and 38°C respectively.

[0172] A multi-physics coupling model of integrated circuits is established. The model includes a current distribution model, a temperature distribution model, and an electromagnetic field distribution model. The current distribution model describes the flow of current in various parts of the circuit, the temperature distribution model describes the temperature distribution of various parts of the chip, and the electromagnetic field distribution model describes the electromagnetic field distribution around the chip. In order to reflect the interaction between the three physical fields, the conductivity of the material is set as a function of temperature to represent the effect of temperature on the current distribution. The Joule heat power density (the product of the square of the current and the resistance) and the external heat source power density are used as heat source terms of the temperature field to represent the effect of the current distribution on the temperature distribution. The magnetic field force (the cross product of the current and the magnetic field) is set as the influence term of the current distribution to represent the effect of the electromagnetic field on the current distribution. These settings ultimately constitute a set of coupled equations that describe the interaction of the three physical fields.

[0173] Substitute the collected initial temperature distribution data into the coupled equations established in the previous step. Solve the equations through numerical iterative calculation methods (such as finite element method or finite difference method) to obtain the current distribution, temperature distribution and magnetic field distribution results of each part of the chip. Compare the calculated current distribution results with the initial measured current distribution data to obtain the current deviation value. Similarly, compare the calculated temperature distribution results with the initial measured temperature distribution data to obtain the temperature deviation value. For example, assuming that the calculated current result of pin 1 is 12mA, the current deviation value of pin 1 is 2mA. Assuming that the calculated chip center temperature result is 42°C, the chip center temperature deviation value is 2°C.

[0174] The physical parameter inversion model is established using the current deviation value and the temperature deviation value. The purpose of this model is to correct the physical parameters in the circuit model, such as the resistivity and thermal conductivity of the material, according to the deviation between the calculated results and the measured results. The physical parameters are adjusted through an optimization algorithm (such as a genetic algorithm or a particle swarm algorithm) to minimize the deviation between the calculated results and the measured results. Finally, the actual physical parameters of the circuit are determined.

[0175] The actual physical parameters obtained by inversion are matched with the pre-established fault feature library. The fault feature library stores the physical parameter change characteristics corresponding to various known fault types. For example, an open circuit in a certain section of wire will cause the current on the wire to be zero and the resistance to be infinite. Through matching, the initial fault state of the integrated circuit can be determined, including the fault location information and the physical characteristics of the fault. For example, the matching result shows that the resistance of a certain section of wire increases, and it is preliminarily determined that the wire has a poor contact fault.

[0176] Based on the actual physical parameters and the initial fault state, a fault state transition model is established. This model describes how the fault develops and changes over time. For example, a wire with poor contact may further increase its resistance due to oxidation, eventually leading to an open circuit. The model can be used to predict the fault state and corresponding physical parameters at the next moment. For example, it is predicted that the resistance of the wire will increase by 15% and the temperature will rise by 3°C at the next moment.

[0177] Substitute the predicted physical parameters into the previously established multi-physics field coupling model to calculate the current distribution, temperature distribution, and magnetic field distribution under the predicted state. Compare the predicted current distribution, temperature distribution, and magnetic field distribution with the actual measurement data to obtain the prediction deviation. Determine the confidence interval of the prediction result based on the size of the prediction deviation. For example, if the prediction deviation is within a certain range, the prediction result is considered to be credible.

[0178] Finally, the output includes prediction results including fault location information, physical characteristics and development trends.

[0179] In this embodiment, through multi-physics field coupling simulation and parameter inversion, the fault location and type can be identified more accurately to avoid misdiagnosis. Through the fault state transition model, the development trend of the fault can be predicted, and preventive measures can be taken in advance to avoid further deterioration of the fault. Through accurate fault diagnosis and prediction, unnecessary maintenance work can be reduced and maintenance costs can be reduced.

[0180] In an optional embodiment, establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining the actual physical parameters of the circuit through the physical parameter inversion model includes:

[0181] Collecting current deviation information and temperature deviation information of the integrated circuit, extracting current density gradient, current direction deflection angle, current fluctuation amplitude and phase delay according to the current deviation information to construct a current deviation feature vector, and extracting temperature gradient, thermal diffusion coefficient and thermal impedance according to the temperature deviation information to construct a temperature deviation feature vector;

[0182] Performing Fourier transform on the temperature deviation feature vector to obtain a temperature frequency domain feature, and combining the temperature frequency domain feature with the current deviation feature vector to form a multi-dimensional feature vector;

[0183] Using the multidimensional feature vector to train a multi-scale feature extraction network, the multi-scale feature extraction network extracts local features and global features through a residual connection structure, and calculates weight coefficients of the local features and the global features through an attention mechanism;

[0184] Constructing an Ohm's law constraint function, a Joule heat constraint function, and an electromagnetic induction constraint function into a physical constraint module, and combining the physical constraint module with the loss function of the multi-scale feature extraction network to form a constraint loss function;

[0185] Calculating the influence factor of the physical parameter on the current deviation information and the temperature deviation information, and combining the influence factor with the constraint loss function to construct a parameter sensitivity model;

[0186] Based on the parameter sensitivity model, multiple inversion models are trained, and the output results of the multiple inversion models are dynamically weighted to determine the actual physical parameters of the circuit.

[0187] For example, the implementation process of the physical parameter inversion technology based on current and temperature deviation starts with the current and temperature data collection on the circuit surface. A dual-channel parallel sampling architecture is adopted, in which the current mirror circuit realizes non-destructive measurement by copying the working current of the circuit to be measured, and the mirror ratio can be adjusted to meet the requirements of different ranges. The magnetic induction array is composed of highly sensitive lateral Hall elements, which are used to obtain the spatial distribution of the magnetic field on the circuit surface, and then reconstruct a two-dimensional current distribution map. This dual-channel structure ensures the measurement accuracy without interfering with the normal operation of the circuit.

[0188] After acquiring the raw data, the current deviation features are extracted. The current density gradient reflects the rate of change of the current size per unit area, which is obtained by calculating the current difference and distance ratio of adjacent sampling points. The current direction deflection angle describes the angle between the actual current flow direction and the designed direction, reflecting the degree of distortion of the current path. The current fluctuation amplitude characterizes the current stability, which is calculated by the difference between the real-time current value and the mean. The phase delay reflects the timing characteristics of signal transmission, which is represented by the time displacement of the signal waveform. These features together constitute the current deviation feature vector.

[0189] Temperature data acquisition uses micro-focus infrared thermal imaging technology, with a spatial resolution of micrometers. The temperature distribution map of the circuit surface is obtained through real-time scanning, and the spatial gradient of the temperature field is calculated to characterize the direction and rate of heat propagation. The thermal diffusion coefficient describes the ability of heat to diffuse in the material and is determined based on the temperature change characteristics over time. Thermal impedance characterizes the heat dissipation performance of the material and is calculated based on the relationship between temperature rise and power consumption. Fast Fourier transform is performed on these temperature features to convert time domain information into frequency domain features, revealing the periodic law of temperature change.

[0190] In the feature fusion stage, an adaptive normalization method is used to process the features of different physical quantities. The normalization parameters are dynamically adjusted according to the feature distribution to ensure that the current features and temperature features are within the same magnitude range. Through nonlinear mapping, various features are converted to a unified feature space to construct a multidimensional feature vector, which contains time domain, frequency domain and spatial distribution information.

[0191] The feature extraction network adopts an improved residual structure and realizes multi-scale feature acquisition through multiple parallel branches. Residual connection can alleviate the gradient vanishing problem of deep networks and improve training stability. The attention mechanism includes two parts: channel attention and spatial attention, which dynamically adjust the importance weight of features. Channel attention focuses on the importance of different channels in the feature map, while spatial attention highlights the feature response of key areas.

[0192] The physical constraint module converts the basic laws of electromagnetism into constraints for network training. The Ohm's law constraint ensures that the relationship between current density and electric field strength conforms to the laws of physics, taking into account the spatial distribution characteristics of material conductivity. The Joule heat constraint establishes the correspondence between current distribution and heat source distribution, including the steady-state and transient processes of heat conduction. The electromagnetic induction constraint ensures that the electromagnetic field distribution meets the requirements of Maxwell's equations.

[0193] Parameter sensitivity analysis determines the importance of each physical parameter through perturbation theory. A small disturbance is applied to each parameter, the system response change is recorded, and the sensitivity coefficient is calculated. The larger the sensitivity coefficient, the more significant the impact of the parameter on the system behavior. Based on the results of sensitivity analysis, the optimization weights of each parameter in the inversion process are dynamically adjusted.

[0194] The final parameter inversion adopts a model integration strategy to train multiple neural networks with different structures in parallel. This includes convolutional neural networks that are good at feature extraction, graph neural networks that are suitable for processing topological relationships, and fully connected networks with powerful nonlinear mapping capabilities. Each model independently learns the mapping relationship from features to physical parameters, and the output results are fused through adaptive weights to obtain the final estimated value. The weight coefficients are dynamically updated based on the performance of each model on the verification data to ensure optimal performance.

[0195] In this embodiment, high-precision lossless inversion of integrated circuit physical parameters is achieved. By integrating current deviation and temperature deviation characteristics, a multi-physical field coupling analysis model is established. Combined with physical constraints and multi-model integration strategies, the reliability and stability of the inversion results are significantly improved. The system has real-time monitoring and early warning capabilities, which can detect potential fault hazards in advance and effectively prevent circuit failures. It is suitable for integrated circuits of different processes and structures, and has good versatility and engineering practical value. Through parameter sensitivity analysis and adaptive weight optimization, the efficiency of the inversion process is improved, making the system easier to implement in engineering and industrial applications. Integrated circuit manufacturers can use this technology to achieve full-process monitoring of product quality and improve yield and reliability levels.

[0196] Figure 2 FIG. 1 is a schematic diagram of the structure of an intelligent system for monitoring and analyzing integrated circuit package current according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0197] The first unit is used to obtain the current timing data of the integrated circuit to be tested under different working states through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel obtains the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, and reconstructs the current distribution map of the two-dimensional magnetic field distribution data through an electromagnetic field inversion algorithm, calibrates and compensates the sampling data of the main acquisition channel according to the current distribution map, and decomposes the calibrated sampling data into a characteristic coefficient matrix through an adaptive wavelet transform;

[0198] The second unit is used to construct a deep feature analysis model based on the feature coefficient matrix, wherein the deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer and a space dimension convolution layer, extracts the time domain features, frequency domain features and space features in the feature coefficient matrix through a three-dimensional convolution operation, and inputs them into a graph attention network, uses the graph attention network to combine the physical structure information of the integrated circuit with the current features to construct a dynamic feature map, and identifies abnormal current features and locates abnormal areas through multi-layer graph convolution operations;

[0199] The third unit is used to input the abnormal current characteristics and abnormal area information into the digital twin analysis system, which includes a physical model of the integrated circuit and a multi-physical field coupling simulation model. The interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model, and the correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the fault development trend prediction are output, and a diagnostic analysis report is generated.

[0200] According to a third aspect of the embodiments of the present invention,

[0201] An electronic device is provided, comprising:

[0202] processor;

[0203] a memory for storing processor-executable instructions;

[0204] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0205] According to a fourth aspect of the embodiments of the present invention,

[0206] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0207] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0208] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein with equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent method for monitoring and analyzing integrated circuit package current, characterized in that: include: The current timing data of the integrated circuit to be tested under different working states is obtained through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel obtains the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, and the two-dimensional magnetic field distribution data is reconstructed into a current distribution map through an electromagnetic field inversion algorithm, and the sampling data of the main acquisition channel is calibrated and compensated according to the current distribution map, and the calibrated sampling data is decomposed into a characteristic coefficient matrix through an adaptive wavelet transform; A deep feature analysis model is constructed based on the feature coefficient matrix. The deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer, and a space dimension convolution layer. The time domain features, frequency domain features, and space features in the feature coefficient matrix are extracted through a three-dimensional convolution operation, and are input into a graph attention network. The physical structure information of the integrated circuit is combined with the current features by using the graph attention network to construct a dynamic feature map. The abnormal current features are identified and the abnormal area is located through multi-layer graph convolution operations. The abnormal current characteristics and abnormal area information are input into a digital twin analysis system, which includes a physical model of an integrated circuit and a multi-physical field coupling simulation model. An interactive influence relationship among current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model. The correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the prediction of the fault development trend are output, and a diagnostic analysis report is generated.

2. The method according to claim 1, characterized in that The two-dimensional magnetic field distribution data is reconstructed into a current distribution diagram through an electromagnetic field inversion algorithm, and the sampling data of the main acquisition channel is calibrated and compensated according to the current distribution diagram. The calibrated sampling data is decomposed into a characteristic coefficient matrix through adaptive wavelet transform, including: Dividing the integrated circuit to be tested into a metal interconnect layer, a device layer and a substrate layer, applying current continuity constraints, interlayer interface constraints and current density upper limit constraints to the metal interconnect layer, the device layer and the substrate layer, wherein the current continuity constraints make the current density divergence in each layer zero, the interlayer interface constraints make the current density component perpendicular to the interlayer interface continuous, and the current density upper limit constraints are defined according to the process parameters of each layer; Establishing a relationship equation between the two-dimensional magnetic field distribution data and the current density distribution, discretizing the relationship equation into a linear equation group, constructing an objective function including the measured magnetic field and the calculated magnetic field error terms, performing conjugate gradient iteration on the objective function, determining the iteration step size through one-dimensional search and updating the search direction, and obtaining a current distribution diagram of the integrated circuit to be tested; Establishing a mapping relationship between the measured current of the main acquisition channel and the current distribution diagram, wherein the mapping relationship includes a proportionality coefficient and a measurement error; using a recursive least squares algorithm to update the proportionality coefficient in real time, substituting the updated proportionality coefficient into the mapping relationship, and calibrating and compensating the measured current of the main acquisition channel in combination with the current distribution under the nominal working state to obtain calibrated and compensated current data; Selecting a wavelet basis function with a compact support property, constructing a wavelet packet dictionary containing basis functions of different scales and different frequency bands, and calculating the wavelet coefficients of the current data after calibration and compensation; An optimal wavelet basis is determined based on information entropy, and the calibrated and compensated current data is decomposed in multiple scales using the optimal wavelet basis to obtain detail components and approximate components in different frequency bands. The detail components and approximate components are arranged in scale order to construct a characteristic coefficient matrix.

3. The method according to claim 2, characterized in that Determine the optimal wavelet basis based on information entropy, use the optimal wavelet basis to perform multi-scale decomposition on the calibrated and compensated current data, obtain detail components and approximate components in different frequency bands, and arrange the detail components and approximate components in scale order to construct a characteristic coefficient matrix, including: Constructing a candidate wavelet basis set, and calculating the wavelet entropy corresponding to each basis function in the candidate wavelet basis set, wherein the wavelet entropy is calculated by the negative sum of the square and logarithm product of the wavelet coefficient; Arrange the basis functions in the candidate wavelet basis set in ascending order according to wavelet entropy, select the basis functions with the highest order to construct a new candidate basis set, and perform orthogonal processing on the basis functions in the new candidate basis set; Repeat the wavelet entropy calculation, ascending order arrangement and orthogonalization processing until the wavelet entropy change value corresponding to the basis function in the new candidate basis set is less than a preset threshold, thereby obtaining an optimal wavelet basis; Performing wavelet decomposition on the current data using the optimal wavelet basis to obtain multi-layer detail coefficients and approximate coefficients, calculating the correlation between detail components of adjacent layers, determining the number of decomposition layers according to the correlation, stopping the decomposition when the correlation is less than a set correlation threshold, and obtaining the final detail components and approximate components; The detail components and approximate components are constructed into an initial characteristic coefficient matrix in hierarchical order, each row of the characteristic coefficient matrix is ​​normalized, the mean and standard deviation of each row are calculated, and the characteristic coefficient matrix is ​​standardized using the mean and standard deviation to obtain the final characteristic coefficient matrix.

4. The method according to claim 1, characterized in that: The time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted through a three-dimensional convolution operation and input into a graph attention network. The physical structure information of the integrated circuit is combined with the current features by using the graph attention network to construct a dynamic feature map. The abnormal current features are identified and the abnormal area is located through multi-layer graph convolution operations, including: The time domain features, frequency domain features and spatial features in the feature coefficient matrix are extracted by a three-dimensional convolution operation, the importance of the features is weighted by an adaptive weight adjustment module, a feature fusion coefficient is calculated according to the mutual information between the features, and feature splicing is performed based on the feature fusion coefficient to obtain a fusion feature; Constructing a heterogeneous graph based on the physical topological structure of the integrated circuit layout, constructing the functional modules, signal paths and power networks in the integrated circuit as vertices, establishing directed edge connections according to the physical connection relationship, and concatenating the fusion feature with the vertex type code as the initial feature of the vertex; Construct a multi-head attention mechanism for bidirectional propagation, construct a feature transformation matrix based on the initial features of the vertex type, adjust the feature propagation strength through the feature transformation matrix and the edge weight matrix, and calculate the attention coefficient between vertices by combining the self-attention features and mutual attention features of the vertices; Generate a local subgraph representation for the vertices in the heterogeneous graph based on a random walk algorithm, and fuse the local subgraph representation with the global features corresponding to the attention coefficients between the vertices to obtain enhanced vertex features; Constructing a multi-layer graph convolutional network, inputting the enhanced vertex features into the multi-layer graph convolutional network, fusing local current features and global current features in each layer of graph convolution operation, transferring the current features through residual connections, and obtaining vertex current features; The vertex current feature is input into the dynamic threshold detection module, and the abnormal threshold is calculated based on the distribution density of the vertex current feature to identify the abnormal vertices. At the same time, the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area, and the detection result of the current anomaly is output.

5. The method according to claim 4, characterized in that The vertex current feature is input into the dynamic threshold detection module, and the abnormal threshold is calculated based on the distribution density of the vertex current feature to identify the abnormal vertex. At the same time, the attention coefficient between vertices and the gradient information of the vertex current feature are used to locate the abnormal area. The detection results of the output current abnormality include: Acquire vertex current feature vectors, determine a cutoff distance based on a K-nearest neighbor search, substitute the cutoff distance into a Gaussian kernel function to calculate a similarity matrix between vertices, and perform row normalization on the similarity matrix to obtain a local density of vertex current feature vectors; Calculate the Euclidean distance between vertices to obtain the vertex distance matrix. For each vertex, determine the target vertex set whose local density is greater than the current vertex. Extract the distance values ​​between the current vertex and each vertex in the target vertex set from the vertex distance matrix, and obtain the minimum value among the distance values ​​as the relative distance of the current vertex. Take the product of the local density of the vertex and the relative distance as the abnormality score of the current vertex. A kernel density estimation method with an adaptive kernel bandwidth is used to perform probability density modeling on the anomaly score, a dynamic threshold is calculated in combination with the mean and standard deviation of the anomaly score, and vertices with anomaly scores lower than the dynamic threshold are marked as anomaly candidate points; Based on the spatial distribution of the abnormal candidate points, a local neighborhood density matrix is ​​constructed, and the abnormal candidate points are clustered using a density clustering method to remove noise points to obtain an abnormal vertex set; Calculate the attention weight between the abnormal vertex and other vertices, input the current feature vector of the abnormal vertex into the radial basis kernel function to obtain feature similarity, multiply the feature similarity and the attention weight to obtain the propagation strength between nodes, and construct an abnormal propagation transfer matrix according to the propagation strength; Perform random walk iterative operation on the abnormal propagation transfer matrix, introduce a restart probability in each iteration to return to the initial abnormal vertex, record the propagation value of each vertex after each iteration, stop the iteration when the difference between the propagation values ​​of two adjacent iterations is less than a preset threshold, and use the final propagation value as the abnormality degree of each vertex; Calculate the gradient value of the abnormality degree to the vertex current feature vector, calculate the sensitivity of the vertex current feature according to the gradient value, weight the sensitivity and the abnormality degree to obtain the vertex abnormality contribution value, select the vertex with the largest abnormality contribution value as the initial point, calculate the physical distance between other vertices and the initial point and the degree of difference in abnormality contribution values, merge the vertices whose physical distance is less than a first preset threshold and whose abnormality contribution value difference is less than a second preset threshold into an abnormal area, and output the positioning result of the abnormal area.

6. The method according to claim 1, characterized in that In the multi-physical field coupling simulation model, the interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established, the corresponding relationship between abnormal current characteristics and circuit physical parameters is determined through iterative calculation, and the fault location results and fault development trend prediction are output by combining the comparative analysis of measured data and simulation results, including: Collecting current data and temperature data of the integrated circuit as initial measurement data, wherein the initial measurement data includes initial current distribution and initial temperature distribution; Establishing a multi-physics field coupling model for an integrated circuit, wherein the multi-physics field coupling model includes a current distribution model, a temperature distribution model, and an electromagnetic field distribution model; setting the conductivity as a function of temperature, setting the Joule heat power density and the external heat source power density as heat source terms of the temperature field, setting the magnetic field force as an influence term of the current distribution, and obtaining a coupling equation group describing the interaction of physical fields; Substituting the initial temperature distribution into the coupled equations for solving, obtaining the current distribution result, the temperature distribution result and the magnetic field distribution result through iterative calculation; comparing the obtained current distribution result, the temperature distribution result and the magnetic field distribution result with the initial measurement data to obtain the current deviation value and the temperature deviation value; Establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining actual physical parameters of the circuit through the physical parameter inversion model; Matching the actual physical parameters with a pre-established fault feature library to obtain an initial fault state, wherein the initial fault state includes fault location information and fault physical features; Establish a fault state transfer model based on the actual physical parameters and the initial fault state, use the fault state transfer model to predict the fault state quantity and actual physical parameters at the next moment, substitute the physical parameters corresponding to the predicted state into the multi-physical field coupling model, and calculate the current distribution prediction value, temperature distribution prediction value and magnetic field distribution prediction value under the predicted state; The predicted current distribution value, temperature distribution predicted value and magnetic field distribution predicted value under the predicted state are compared with the measured data to obtain the predicted deviation. The confidence interval of the predicted result is determined based on the predicted deviation, and the predicted result containing the fault location information, physical characteristics and development trend is output.

7. The method according to claim 6, characterized in that Establishing a physical parameter inversion model based on the current deviation value and the temperature deviation value, and determining the actual physical parameters of the circuit through the physical parameter inversion model includes: Collecting current deviation information and temperature deviation information of the integrated circuit, extracting current density gradient, current direction deflection angle, current fluctuation amplitude and phase delay according to the current deviation information to construct a current deviation feature vector, and extracting temperature gradient, thermal diffusion coefficient and thermal impedance according to the temperature deviation information to construct a temperature deviation feature vector; Performing Fourier transform on the temperature deviation feature vector to obtain a temperature frequency domain feature, and combining the temperature frequency domain feature with the current deviation feature vector to form a multi-dimensional feature vector; Using the multidimensional feature vector to train a multi-scale feature extraction network, the multi-scale feature extraction network extracts local features and global features through a residual connection structure, and calculates weight coefficients of the local features and the global features through an attention mechanism; Constructing an Ohm's law constraint function, a Joule heat constraint function, and an electromagnetic induction constraint function into a physical constraint module, and combining the physical constraint module with the loss function of the multi-scale feature extraction network to form a constraint loss function; Calculating the influence factor of the physical parameter on the current deviation information and the temperature deviation information, and combining the influence factor with the constraint loss function to construct a parameter sensitivity model; Based on the parameter sensitivity model, multiple inversion models are trained, and the output results of the multiple inversion models are dynamically weighted to determine the actual physical parameters of the circuit.

8. An intelligent system for monitoring and analyzing integrated circuit package current, used to implement the method according to any one of claims 1 to 7, characterized in that: include: The first unit is used to obtain the current timing data of the integrated circuit to be tested under different working states through a dual-channel parallel acquisition system, wherein the main acquisition channel uses a current mirror circuit to perform lossless replication and real-time sampling of the working current of the integrated circuit to be tested, and the auxiliary acquisition channel obtains the two-dimensional magnetic field distribution data on the surface of the integrated circuit to be tested through an array magnetic sensor, and reconstructs the current distribution map of the two-dimensional magnetic field distribution data through an electromagnetic field inversion algorithm, calibrates and compensates the sampling data of the main acquisition channel according to the current distribution map, and decomposes the calibrated sampling data into a characteristic coefficient matrix through an adaptive wavelet transform; The second unit is used to construct a deep feature analysis model based on the feature coefficient matrix, wherein the deep feature analysis model adopts an improved three-dimensional convolutional neural network structure, which includes a time dimension convolution layer, a frequency dimension convolution layer and a space dimension convolution layer, extracts the time domain features, frequency domain features and space features in the feature coefficient matrix through a three-dimensional convolution operation, and inputs them into a graph attention network, uses the graph attention network to combine the physical structure information of the integrated circuit with the current features to construct a dynamic feature map, and identifies abnormal current features and locates abnormal areas through multi-layer graph convolution operations; The third unit is used to input the abnormal current characteristics and abnormal area information into the digital twin analysis system, which includes a physical model of the integrated circuit and a multi-physical field coupling simulation model. The interactive influence relationship between current distribution, temperature distribution and electromagnetic field distribution is established in the multi-physical field coupling simulation model, and the correspondence between the abnormal current characteristics and the physical parameters of the circuit is determined through iterative calculation. Combined with the comparative analysis of the measured data and the simulation results, the fault location results and the fault development trend prediction are output, and a diagnostic analysis report is generated.

9. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 7.

10. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 7 is implemented.

Images