Semiconductor chip defect detecting and positioning method and system

Abstract

The invention provides a semiconductor chip defect detection positioning method and system, and relates to the technical field of semiconductor chip detection, and the method comprises the steps: obtaining an original image and electrical test data; forming a defect candidate area based on local density analysis; variational mode decomposition is carried out on the electrical data, and electrical characteristic abnormal points are marked; mapping the image features to an electrical characteristic space and calculating mutual information values to determine defect points; and calculating the distribution and path of a diffusion field and determining the severity of the defect. According to the invention, accurate positioning and severity evaluation of chip defects are realized, and the detection efficiency and accuracy are improved.

Description

Technical Field

[0001] This invention relates to the field of semiconductor chip inspection technology, and in particular to a method and system for detecting and locating defects in semiconductor chips. Background Technology

[0002] With the rapid development of integrated circuit technology, semiconductor chip manufacturing processes are becoming increasingly complex, chip sizes are shrinking, and integration levels are continuously improving. During semiconductor chip production, various defects often appear in the chips due to factors such as process fluctuations, material defects, and environmental influences. These defects affect the chip's performance and reliability. Therefore, accurate detection and location of semiconductor chip defects are of great significance for improving chip yield, reducing production costs, and ensuring product quality.

[0003] Traditional semiconductor chip defect detection methods mainly fall into two categories: optical inspection and electrical testing. Optical inspection uses equipment such as scanning electron microscopes and atomic force microscopes to acquire images of the chip surface and identifies defects through image processing algorithms. Electrical testing, on the other hand, measures the chip's electrical parameters, such as current, voltage, and resistance, to determine whether the chip is functioning correctly. However, these methods have some limitations in practical applications. Summary of the Invention

[0004] The present invention provides a method and system for detecting and locating defects in semiconductor chips, which can solve the problems in the prior art.

[0005] A first aspect of the present invention provides a method for detecting and locating defects in a semiconductor chip, comprising:

[0006] Acquire raw images of the semiconductor chip under test and electrical test data at different sampling points;

[0007] Calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region. Cluster the pixels that exceed the deviation threshold to form defect candidate regions.

[0008] Variational mode decomposition is performed on the electrical test data at the different sampling points to obtain the intrinsic mode function set. Hilbert transform is performed on the intrinsic mode functions and the instantaneous frequency is calculated. Sampling points that exceed the frequency threshold are marked as abnormal points in electrical characteristics.

[0009] The image features of the defect candidate region are mapped to the electrical characteristic parameter space through nonlinear transformation. The electrical characteristic anomalies are converted into anomaly probability distributions in the image space through reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point at that location.

[0010] The diffusion field distribution is calculated using the defect location as the source point. Combined with the spatiotemporal evolution of electrical characteristic anomalies, the diffusion path of the defect is predicted. The correlation coefficient of the image features of each sampling point on the diffusion path is calculated to determine the severity of the defect.

[0011] Calculate the local density value of each pixel in the original image and the deviation from the mean density of the region. Cluster pixels exceeding the deviation threshold to form candidate defect regions, including:

[0012] The original image is divided into multiple local windows. The standard deviation of gray values ​​of pixels in each local window is calculated. The product of the standard deviation of gray values ​​and the adjustment coefficient is added to the unit base and then multiplied by the first reference bandwidth parameter to obtain the adaptive bandwidth parameter. The adaptive bandwidth parameter is substituted into the Gaussian kernel function to calculate the local density value of the center pixel.

[0013] A multi-scale feature pyramid is constructed based on the local density values. The mean local density and the standard deviation of local density are calculated at each scale level. The local density mean of the scale level is subtracted from the local density mean of the central pixel at each scale level, and then divided by the local density standard deviation to obtain the standardized deviation. When the standardized deviation of any scale level is greater than the deviation threshold, the position coordinates of the central pixel are recorded as outlier coordinates.

[0014] Calculate the density flow between adjacent anomalies, where the density flow is the rate of change of the local density value along the path connecting the two anomalies with spatial location. When the maximum value of the density flow is less than the second reference parameter, the two anomalies are marked as belonging to the same anomaly region, and the set of anomalies with the same anomaly region number is output as a defect candidate region.

[0015] Variational mode decomposition is performed on the electrical test data at the different sampling points to obtain a set of intrinsic mode functions. Hilbert transform is then performed on the intrinsic mode functions, and the instantaneous frequencies are calculated. Sampling points exceeding the frequency threshold are marked as electrical characteristic anomalies, including:

[0016] Within a sliding time window, the local variance of the electrical test data at different sampling points is calculated. The local variance is used as a constraint to perform variational mode decomposition to obtain an intrinsic mode function set. The energy density of each intrinsic mode function in the intrinsic mode function set is calculated. The orthogonality index between intrinsic mode functions with energy densities greater than the energy threshold is calculated. Intrinsic mode functions with orthogonality indexes less than the orthogonality threshold are determined as valid intrinsic mode functions.

[0017] The effective intrinsic mode function is subjected to Hilbert transform to obtain an analytical signal. The generalized zero-crossing rate is calculated based on the zero-crossing characteristics of the analytical signal. The time derivative of the generalized zero-crossing rate is divided by the frequency normalization coefficient to obtain the instantaneous frequency.

[0018] Spatial weighting coefficients are assigned based on the location distance of the sampling points. The sum of the products of the spatial weighting coefficients and the instantaneous frequency is divided by the sum of the spatial weighting coefficients to obtain the reference frequency. The frequency deviation between the instantaneous frequency and the reference frequency of different sampling points is calculated, and sampling points that exceed the frequency threshold are marked as abnormal electrical characteristics.

[0019] Image features of candidate defect regions are mapped to an electrical characteristic parameter space through a nonlinear transformation. Electrical characteristic anomalies are converted into anomaly probability distributions in the image space through a reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than a correlation threshold, a defect point is determined to exist at that location, including:

[0020] Obtain a candidate defect region image of the chip under test, and decompose the candidate defect region image into a multi-scale feature vector through discrete wavelet transform;

[0021] The local covariance matrix of the multi-scale eigenvectors is calculated, and the eigenvalue decomposition of the local covariance matrix is ​​performed to obtain an eigenspace. Multiple eigenvectors with the top 5% contribution rates in the eigenspace are extracted to construct a mapping basis. The sum of the inner product of the mapping basis and the radial basis kernel function and the adaptive weight coefficients is determined as a nonlinear mapping function. The multi-scale eigenvectors are transformed to the electrical characteristic parameter space through the nonlinear mapping function to obtain the parameter mapping vector.

[0022] The position potential energy is calculated based on the Mahalanobis distance between the parameter mapping vector and the electrical characteristic anomaly point, and the adjacent potential energy is calculated based on the Euclidean distance between adjacent parameter mapping vectors. The conditional probability is determined according to the weighted combination of the position potential energy and the adjacent potential energy. The conditional probability is then used to obtain the inverse mapping probability through maximum likelihood estimation. Finally, the inverse mapping probability is transformed into the image space to obtain the defect probability distribution.

[0023] The ratio of the first edge probability corresponding to the parameter mapping vector to the second edge probability corresponding to the defect probability distribution is used as the feature mutual information value; when the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point in the defect candidate region, and the spatial coordinate information of the defect point is output.

[0024] The diffusion field distribution is calculated using the defect location as the source point. Combined with the spatiotemporal evolution of electrical anomalies, the diffusion path of the defect is predicted. The correlation coefficients of image features at each sampling point along the diffusion path are calculated to determine the severity of the defect, including:

[0025] The local material conductivity gradient and stress intensity at the defect location are obtained. The diffusion coefficient is determined based on the ratio of the local material conductivity gradient to the stress intensity. The product of the diffusion coefficient and the gradient of the diffusion field intensity at the defect point is used as the diffusion field intensity distribution.

[0026] Electrical characteristic test data around the defect point is collected, and the feature vector of the electrical characteristic test data is extracted by wavelet transform. The local maxima of the feature vector are marked as anomalies. The propagation speed and acceleration of the anomalies are calculated, and the direction of anomaly propagation is determined based on the propagation speed and acceleration.

[0027] The diffusion field intensity distribution is converted into an intensity probability distribution, and the abnormal diffusion direction is converted into a directional probability distribution. The intensity probability distribution and the directional probability distribution are fused using the maximum entropy criterion to obtain a joint probability distribution. The defect diffusion path is determined along the direction with the maximum joint probability distribution.

[0028] Obtain the local image gray-level co-occurrence matrix of each sampling point on the defect diffusion path, map the gray-level co-occurrence matrix into gray-level feature vectors through sparse coding, calculate the correlation coefficient between the gray-level feature vectors; when the correlation coefficient is greater than the feature correlation threshold, determine that the defect has a diffusion risk, and output the location coordinates of the defect diffusion path.

[0029] The diffusion field intensity distribution is converted into an intensity probability distribution, and the abnormal diffusion direction is converted into a directional probability distribution. The intensity probability distribution and the directional probability distribution are then fused using the maximum entropy criterion to obtain a joint probability distribution. The defect diffusion path is determined along the direction with the maximum joint probability distribution, including:

[0030] The intensity distribution of the diffusion field is divided into adaptive equal-width intensity intervals, wherein the interval width is determined according to the standard deviation of the intensity distribution. The number of pixels in each intensity interval is counted and divided by the total number of pixels to obtain the intensity probability distribution.

[0031] An anisotropic Gaussian kernel function is used to adaptively smooth the abnormal diffusion direction, wherein the variance of the Gaussian kernel function is set along the gradient direction and perpendicular to the gradient direction respectively, to obtain a direction probability distribution with direction selectivity.

[0032] The intensity information entropy is obtained by multiplying the intensity probability distribution with its natural logarithm; the direction information entropy is obtained by multiplying the direction probability distribution with its natural logarithm; the intensity mean is obtained by multiplying the intensity probability distribution by the corresponding gray value; and the direction mean is obtained by multiplying the direction probability distribution by the corresponding direction angle.

[0033] Set initial weight parameters, add the intensity information entropy and the direction information entropy according to the weight parameters, and iteratively update the weight parameters to maximize the weighted sum, while ensuring that the mean intensity and mean direction weighted by the weight parameters meet the mean range; stop iterating when the difference between the weight parameters of two adjacent iterations is less than the convergence threshold, and obtain the optimal weight parameters; substitute the optimal weight parameters, intensity probability, and direction probability into the exponential function to obtain the joint probability distribution;

[0034] On the joint probability distribution, starting from the point of maximum probability, a path search is performed based on directional continuity constraints to obtain the defect propagation path.

[0035] A second aspect of the present invention provides a semiconductor chip defect detection and location system, comprising:

[0036] The first unit is used to acquire the original image of the semiconductor chip under test and the electrical test data at different sampling points;

[0037] The second unit is used to calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region, and cluster the pixels that exceed the deviation threshold to form defect candidate regions.

[0038] The third unit is used to perform variational mode decomposition on the electrical test data of the different sampling points to obtain the intrinsic mode function set, perform Hilbert transform on the intrinsic mode functions and calculate the instantaneous frequency, and mark the sampling points that exceed the frequency threshold as electrical characteristic anomalies.

[0039] The fourth unit is used to map the image features of the defect candidate region to the electrical characteristic parameter space through nonlinear transformation, convert the electrical characteristic anomaly points into the anomaly probability distribution in the image space through reverse mapping, calculate the feature mutual information value between the electrical characteristic parameter space and the image space, and determine that there is a defect point at the location when the feature mutual information value is greater than the correlation threshold.

[0040] The fifth unit is used to calculate the diffusion field distribution with the defect location as the source point, combine the spatiotemporal evolution law of electrical characteristic anomalies, predict the diffusion path of the defect, calculate the correlation coefficient of the image features of each sampling point on the diffusion path, and determine the severity of the defect.

[0041] A third aspect of the present invention,

[0042] An electronic device is provided, comprising:

[0043] processor;

[0044] Memory used to store processor-executable instructions;

[0045] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0046] Fourth aspect of the embodiments of the present invention,

[0047] A computer-readable storage medium is provided, having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0048] The beneficial effects of this application are as follows:

[0049] This invention combines semiconductor chip image features with electrical test data to achieve accurate defect detection and location, significantly improving detection accuracy and efficiency, reducing false negative and false positive rates, and meeting the high requirements of advanced semiconductor manufacturing processes for defect detection.

[0050] This invention uses variational mode decomposition and Hilbert transform to analyze electrical test data, and establishes a mapping relationship between image space and electrical characteristic space through nonlinear transformation, thereby realizing the early identification of potential defects. This provides an important basis for timely adjustment of process parameters and effectively avoids the generation of batch defective products.

[0051] This invention, through defect diffusion field analysis and spatiotemporal evolution law prediction, can accurately assess the severity and development trend of defects, providing a visual representation of defect evolution, and offering a scientific basis for continuous optimization of chip manufacturing processes and yield improvement, thus possessing significant practical value. Attached Figure Description

[0052] Figure 1 This is a schematic flowchart of the semiconductor chip defect detection and localization method according to an embodiment of the present invention;

[0053] Figure 2 This is a schematic diagram of the nonlinear mapping of defect candidate regions and the defect detection process. Detailed Implementation

[0054] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0055] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0056] Figure 1This is a flowchart illustrating the semiconductor chip defect detection and localization method according to an embodiment of the present invention, as shown below. Figure 1 As shown, the method includes:

[0057] Acquire raw images of the semiconductor chip under test and electrical test data at different sampling points;

[0058] Calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region. Cluster the pixels that exceed the deviation threshold to form defect candidate regions.

[0059] Variational mode decomposition is performed on the electrical test data at the different sampling points to obtain the intrinsic mode function set. Hilbert transform is performed on the intrinsic mode functions and the instantaneous frequency is calculated. Sampling points that exceed the frequency threshold are marked as abnormal points in electrical characteristics.

[0060] The image features of the defect candidate region are mapped to the electrical characteristic parameter space through nonlinear transformation. The electrical characteristic anomalies are converted into anomaly probability distributions in the image space through reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point at that location.

[0061] The diffusion field distribution is calculated using the defect location as the source point. Combined with the spatiotemporal evolution of electrical characteristic anomalies, the diffusion path of the defect is predicted. The correlation coefficient of the image features of each sampling point on the diffusion path is calculated to determine the severity of the defect.

[0062] In one optional implementation, the local density value of each pixel in the original image is calculated, and the deviation from the mean density of the region is calculated. Pixels exceeding the deviation threshold are clustered to form candidate defect regions, including:

[0063] The original image is divided into multiple local windows. The standard deviation of gray values ​​of pixels in each local window is calculated. The product of the standard deviation of gray values ​​and the adjustment coefficient is added to the unit base and then multiplied by the first reference bandwidth parameter to obtain the adaptive bandwidth parameter. The adaptive bandwidth parameter is substituted into the Gaussian kernel function to calculate the local density value of the center pixel.

[0064] A multi-scale feature pyramid is constructed based on the local density values. The mean local density and the standard deviation of local density are calculated at each scale level. The local density mean of the scale level is subtracted from the local density mean of the central pixel at each scale level, and then divided by the local density standard deviation to obtain the standardized deviation. When the standardized deviation of any scale level is greater than the deviation threshold, the position coordinates of the central pixel are recorded as outlier coordinates.

[0065] Calculate the density flow between adjacent anomalies, where the density flow is the rate of change of the local density value along the path connecting the two anomalies with spatial location. When the maximum value of the density flow is less than the second reference parameter, the two anomalies are marked as belonging to the same anomaly region, and the set of anomalies with the same anomaly region number is output as a defect candidate region.

[0066] In this embodiment, to more accurately analyze the defect features in the image, it is necessary to calculate the local density value for each pixel of the original image. The system divides the original image into multiple local windows of size M×M, where M=15, so that each local window contains 225 pixels. For image edge regions, a mirror filling method is used to ensure that all pixels are processed completely.

[0067] For each pixel (x, y) in the image, a local window is established centered on that pixel, and the standard deviation of the gray values ​​of all pixels within the window is calculated. In practical applications, for a region with obvious edge texture, the standard deviation of gray values ​​within the local window is 25.6; while for a relatively smooth region, the standard deviation is only 3.2. Based on the calculated standard deviation σ, combined with the adjustment coefficient α (set to 0.5) and the unit base c (set to 1.0), the adaptive bandwidth parameter h is calculated as h = (c + α × σ) × h0, where h0 is the first baseline bandwidth parameter, set to 5.0. For textured regions, h is 14.8; for smooth regions, h is 6.6, achieving adaptive adjustment of the bandwidth parameter.

[0068] The calculated adaptive bandwidth parameter h is substituted into the Gaussian kernel function to calculate the local density value D(x,y) of the center pixel (x,y). The Gaussian kernel function calculates weights based on the spatial distance and grayscale differences between pixels, and these weights are normalized to obtain the local density value. For example, for a 512×512 pixel image of the surface of an industrial part, the local density value of pixels in the normal area is 0.85, while the local density value of pixels in the defective area drops to 0.35.

[0069] A multi-scale feature pyramid is constructed based on the calculated local density values, containing L scale layers (let L = 3). The first layer uses the original resolution, and subsequent layers are obtained by downsampling the previous layer, with a sampling ratio of 0.5. For the original 512×512 image, the resolutions of the three pyramid layers are 512×512, 256×256, and 128×128, respectively.

[0070] At scale layer i, calculate the mean μ of the local density values ​​of all pixels in that layer. i and standard deviation σ iFor example, in the first layer (original resolution), μ1 is 0.78 and σ1 is 0.12; in the second layer, μ2 is 0.76 and σ2 is 0.15; in the third layer, μ3 is 0.74 and σ3 is 0.18. For each pixel (x,y) in scale layer i, calculate its standardized deviation Z(x,y,i) = (D(x,y,i) - μ i ) / σ i , where D(x,y,i) represents the local density value at the midpoint (x,y) of the i-th layer.

[0071] A deviation threshold T is set, which can be set to T = 2.5. This means that when the local density value of a pixel deviates from the mean by more than 2.5 times the standard deviation, the pixel is identified as an outlier. All pixels in the feature pyramid are traversed. If the standardized deviation Z(x,y,i) of a point at any scale level i is greater than the threshold T, then the coordinates (x,y) of that point are recorded as outlier coordinates. For example, for a surface pit defect, the standardized deviation of a point in the first layer is 2.8, exceeding the set threshold of 2.5, and therefore it is marked as an outlier.

[0072] For all marked outliers, the density flow between them is calculated to determine whether they belong to the same anomalous region. For example, two adjacent outliers (2,3) and (5,8) are arbitrarily selected, and 10 points are uniformly sampled along a straight path between these two points, resulting in the following sampling point coordinate sequence: (2,3), (2.3,3.5), (2.6,4), (2.9,4.5), (3.2,5), (3.5,5.5), (3.8,6), (4.1,6.5), (4.4,7), (4.7,7.5), (5,8). The local density values ​​of adjacent points in this sampling point sequence are calculated, resulting in the following local density sequence: 0.82, 0.78, 0.75. 0.71, 0.68, 0.65, 0.61, 0.58, 0.55, 0.52; calculate the difference between adjacent density values ​​in the local density sequence to obtain a density difference sequence: 0.04, 0.03, 0.04, 0.03, 0.03, 0.04, 0.03, 0.03; select the maximum value 0.04 from the density difference sequence as the maximum density flow value F, and compare the maximum density flow value F with the preset second reference parameter 0.3; since 0.04 is less than 0.3, it is determined that these two anomalies belong to the same anomaly region.

[0073] Based on the density flow analysis described above, the system clusters all outliers, outputting sets of outliers with the same outlier region number as candidate defect regions. For example, in a 512×512 image, three candidate defect regions might be detected: Region 1 contains 23 outliers, centered at (125, 240), with an area of ​​approximately 78 square pixels; Region 2 contains 18 outliers, centered at (320, 180), with an area of ​​approximately 65 square pixels; and Region 3 contains 32 outliers, centered at (410, 380), with an area of ​​approximately 105 square pixels. These candidate defect regions can be further used for subsequent defect classification and evaluation.

[0074] In one optional implementation, variational mode decomposition is performed on the electrical test data at the different sampling points to obtain an intrinsic mode function set. A Hilbert transform is then performed on the intrinsic mode functions, and the instantaneous frequency is calculated. Sampling points exceeding a frequency threshold are marked as electrical characteristic anomalies, including:

[0075] Within a sliding time window, the local variance of the electrical test data at different sampling points is calculated. The local variance is used as a constraint to perform variational mode decomposition to obtain an intrinsic mode function set. The energy density of each intrinsic mode function in the intrinsic mode function set is calculated. The orthogonality index between intrinsic mode functions with energy densities greater than the energy threshold is calculated. Intrinsic mode functions with orthogonality indexes less than the orthogonality threshold are determined as valid intrinsic mode functions.

[0076] The effective intrinsic mode function is subjected to Hilbert transform to obtain an analytical signal. The generalized zero-crossing rate is calculated based on the zero-crossing characteristics of the analytical signal. The time derivative of the generalized zero-crossing rate is divided by the frequency normalization coefficient to obtain the instantaneous frequency.

[0077] Spatial weighting coefficients are assigned based on the location distance of the sampling points. The sum of the products of the spatial weighting coefficients and the instantaneous frequency is divided by the sum of the spatial weighting coefficients to obtain the reference frequency. The frequency deviation between the instantaneous frequency and the reference frequency of different sampling points is calculated, and sampling points that exceed the frequency threshold are marked as abnormal electrical characteristics.

[0078] In this embodiment, a method for detecting electrical characteristic anomalies based on variational mode decomposition and Hilbert transform is proposed. This method performs variational mode decomposition on electrical test data from different sampling points, performs Hilbert transform on the obtained intrinsic mode functions and calculates the instantaneous frequency, and marks sampling points exceeding a frequency threshold as electrical characteristic anomalies.

[0079] When processing electrical test data from different sampling points, local variance is calculated within a sliding time window as a constraint. Taking the collected voltage data as an example, for a certain sampling point, the local variance is calculated every 5ms within a 30ms time window. At t=100ms, the local variance of the voltage data at this sampling point in the first 30ms is 0.0025V. 2 This local variance value is used as a constraint to perform variational mode decomposition.

[0080] In the variational mode decomposition process, the number of modes K is set to 5, the penalty factor α is set to 2000, the bandwidth control parameter τ is set to 0, and the tolerance is set to 10. -7 The maximum number of iterations is 500. Using these parameter settings, the original electrical test data is decomposed to obtain five intrinsic mode functions (EMFs). The energy density of each EMF is then calculated; for example, the energy densities of the five EMFs are 0.85, 0.76, 0.45, 0.12, and 0.09, respectively. Setting an energy threshold of 0.2, the first three EMFs with energy densities greater than the threshold are 0.85, 0.76, and 0.45.

[0081] For these three eigenmode functions with high energy densities, their orthogonality indices are calculated. The orthogonality index is calculated by dividing the inner product of two eigenmode functions by their norm product. For example, the orthogonality index for the first and second eigenmode functions is 0.08, for the first and third is 0.15, and for the second and third is 0.22. Setting an orthogonality threshold of 0.2, the orthogonality indices for the first and second eigenmode functions are less than the threshold, as are those for the first and third eigenmode functions, while the orthogonality indices for the second and third eigenmode functions are greater than the threshold. Therefore, the first and second eigenmode functions are determined to be valid eigenmode functions, while the third eigenmode function does not meet the orthogonality requirement and is discarded.

[0082] A Hilbert transform is performed on the determined effective eigenmode functions. Taking the first effective eigenmode function as an example, it is represented as a real part signal, and its imaginary part signal is obtained through the Hilbert transform. The two together form an analytic signal. Based on the phase change characteristics of the analytic signal, the distribution of its zero-crossing points is calculated. In a specific implementation, the times when the signal crosses zero points from negative to positive are recorded. For example, if zero crossings occur at t = 105ms, 115ms, and 125ms, then the generalized zero-crossing rate within this time period is 0.1 times / ms.

[0083] The time derivative of the generalized zero-crossing rate can be approximated by the difference between adjacent time points. For example, if the generalized zero-crossing rate is 0.09 times / ms at t = 110 ms and 0.11 times / ms at t = 120 ms, then the time derivative of the zero-crossing rate at t = 115 ms is approximately 0.002 times / ms. 2 Assuming the frequency normalization coefficient is 0.004, the instantaneous frequency obtained is 0.5 Hz.

[0084] Considering the spatial distribution of sampling points, spatial weight coefficients are assigned based on the distance between the sampling points. For example, for a 5×5 sampling point array, the coordinates of the center point are (3,3), and the weights of other points are calculated based on their Euclidean distance from the center point. At a certain moment, the instantaneous frequency of the sampling point at position (1,1) is 52.3Hz, with a spatial weight of 0.2; the instantaneous frequency at position (2,2) is 50.1Hz, with a spatial weight of 0.6; and the instantaneous frequency at position (3,3) is 49.8Hz, with a spatial weight of 1.0. Dividing the sum of the products of the spatial weights and the instantaneous frequencies by the sum of the spatial weights, we obtain the reference frequency as (52.3×0.2+50.1×0.6+49.8×1.0) / (0.2+0.6+1.0)=50.22Hz.

[0085] Calculate the frequency deviation between the instantaneous frequency and the reference frequency for each sampling point. For example, the frequency deviation at position (1,1) is |52.3-50.22|=2.08Hz, the frequency deviation at position (2,2) is |50.1-50.22|=0.12Hz, and the frequency deviation at position (3,3) is |49.8-50.22|=0.42Hz. Setting a frequency threshold of 1.5Hz, the frequency deviation at position (1,1) exceeds the threshold and is marked as an abnormal point in electrical characteristics, while the frequency deviations at positions (2,2) and (3,3) do not exceed the threshold and are considered normal points.

[0086] The above methods can effectively identify abnormal electrical characteristics on circuit boards, providing a basis for subsequent defect localization and quality control. In practical applications, the parameter settings, energy threshold, orthogonality threshold, and frequency threshold of variational mode decomposition can be adjusted according to different types of electrical test data and application scenarios to obtain the best anomaly detection results. Furthermore, the spatial weight allocation method can also be optimized based on the actual layout characteristics of the circuit board to improve the accuracy of anomaly detection.

[0087] Figure 2This is a schematic diagram of the nonlinear mapping and defect detection process for defect candidate regions. In one optional implementation, the image features of the defect candidate region are mapped to the electrical characteristic parameter space through a nonlinear transformation. Electrical characteristic anomalies are converted into anomaly probability distributions in the image space through a reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than a correlation threshold, a defect point is determined to exist at that location. This includes:

[0088] Obtain a candidate defect region image of the chip under test, and decompose the candidate defect region image into a multi-scale feature vector through discrete wavelet transform;

[0089] The local covariance matrix of the multi-scale eigenvectors is calculated, and the eigenvalue decomposition of the local covariance matrix is ​​performed to obtain an eigenspace. Multiple eigenvectors with the top 5% contribution rates in the eigenspace are extracted to construct a mapping basis. The sum of the inner product of the mapping basis and the radial basis kernel function and the adaptive weight coefficients is determined as a nonlinear mapping function. The multi-scale eigenvectors are transformed to the electrical characteristic parameter space through the nonlinear mapping function to obtain the parameter mapping vector.

[0090] The position potential energy is calculated based on the Mahalanobis distance between the parameter mapping vector and the electrical characteristic anomaly point, and the adjacent potential energy is calculated based on the Euclidean distance between adjacent parameter mapping vectors. The conditional probability is determined according to the weighted combination of the position potential energy and the adjacent potential energy. The conditional probability is then used to obtain the inverse mapping probability through maximum likelihood estimation. Finally, the inverse mapping probability is transformed into the image space to obtain the defect probability distribution.

[0091] The ratio of the first edge probability corresponding to the parameter mapping vector to the second edge probability corresponding to the defect probability distribution is used as the feature mutual information value; when the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point in the defect candidate region, and the spatial coordinate information of the defect point is output.

[0092] In this embodiment, a method is described in detail to map the image features of a defect candidate region to an electrical characteristic parameter space through a nonlinear transformation, and to determine the defect point through inverse mapping.

[0093] To obtain candidate defect regions for the chip under test, high-resolution images can be acquired using a defect detection system during chip manufacturing. Taking a specific detection scenario as an example, the original image is 2048×2048 pixels with a grayscale value range of 0-255. This candidate defect region image is then subjected to multi-scale decomposition using discrete wavelet transform and a three-level decomposition using Daubechies wavelet basis functions, resulting in 10 sub-band images: LL3, LH3, HL3, HH3, LH2, HL2, HH2, LH1, HL1, and HH1. These 10 sub-band images are then rearranged and organized into a multi-scale feature vector F with dimensions of 2048×2048×10.

[0094] Calculate the local covariance matrix C for the multi-scale feature vector F. Specifically, for each pixel (i,j), extract the feature vector within its 7×7 neighborhood, construct a 10×49 local feature matrix, and calculate the covariance of this matrix to obtain a 10×10 local covariance matrix C(i,j). Perform eigenvalue decomposition on each local covariance matrix C(i,j) to obtain the eigenvalues ​​λ1≥λ2≥...≥λ 10 and their corresponding feature vectors v1, v2, ..., v 10 Calculate the contribution rate λ for each eigenvalue. k / Σλ i The mapping basis B is constructed by selecting the feature vectors with the top 5% cumulative contribution rates. In actual calculations, selecting the first 2-3 feature vectors is usually sufficient to meet the 5% contribution rate requirement. In this example, the first two feature vectors, v1 and v2, are selected to form the mapping basis B.

[0095] The nonlinear mapping function M is determined by multiplying the inner product of the mapping basis B and the radial basis kernel function with the adaptive weight coefficients. The radial basis kernel function chosen is the Gaussian kernel function K(x,y) = exp(-||xy||). 2 / 2σ 2 ), where σ is the kernel width parameter, set to 1.5 times the standard deviation of the eigenvectors; in this example, σ = 0.12. The adaptive weight coefficients are obtained by minimizing the mapping error, with initial values ​​set to uniformly distributed random numbers. The final weight values ​​w = [0.35, 0.42, 0.23] are obtained through iterative optimization using gradient descent. The nonlinear mapping function M(F) = Σw i ·K(B,F i ), where F i Let be the i-th component of the eigenvector F. The multi-scale eigenvector F is transformed into the electrical characteristic parameter space through this nonlinear mapping function, resulting in a parameter mapping vector P with dimensions 2048×2048×3, corresponding to the three electrical characteristic parameters: leakage current, threshold voltage, and saturation current.

[0096] The positional potential energy U is calculated based on the Mahalanobis distance between the parameter mapping vector P and the pre-calibrated electrical characteristic anomaly point E. p The Mahalanobis distance takes into account the correlation between parameters and is calculated as d(P,E)=√((PE)). T S -1 (PE)), where S is the covariance matrix of the electrical characteristic parameters. In this example, the standard value of the outlier E is [2.5nA, 0.65V, 120μA], and the location potential U p =exp(-d 2 / 2). Simultaneously, the adjacent potential energy U is calculated based on the Euclidean distance between adjacent parameter mapping vectors. n The Euclidean distance is defined as d(P) i ,P j )=||P i -P j ||, adjacent potential energy U n =exp(-d 2 / 2). By weighted combination of position potential energy U p and adjacent potential energy U n Determine the conditional probability P c =αU p +(1-α)U n Where α is the weighting coefficient, set to 0.7. The conditional probability P... c The reverse mapping probability P is obtained through maximum likelihood estimation. r The reverse mapping uses Bayes' theorem, P r (F|P)=P c (P|F)·P(F) / P(P), where P(F) is the prior probability and P(P) is the marginal probability, obtained through histogram statistics. The inverse mapping probability Pr is transformed into the image space to obtain the defect probability distribution D, with a dimension of 2048×2048.

[0097] The first marginal probability H(P) corresponding to the parameter mapping vector P is obtained by calculating the entropy of P, H(P) = -Σp(p i )·log(p(p i )), where p(p i ) represents the value p in the parameter mapping vector P. i The probability of defect. The second marginal probability H(D) corresponding to the defect probability distribution D is obtained by calculating the entropy of D, H(D) = -Σp(d i )·log(p(d i )), where p(d i ) represents the median d of the defect probability distribution D. iThe probability of P and D is calculated. The feature mutual information value I(P;D) = H(P) + H(D) - H(P,D), where H(P,D) is the joint entropy of P and D. In this example, H(P) = 3.45, H(D) = 2.87, and H(P,D) = 4.76, therefore I(P;D) = 1.56. This feature mutual information value is compared with the preset correlation threshold of 1.2. Since 1.56 > 1.2, it is determined that there is a defect point in the defect candidate region. The spatial coordinates of the defect point are output, which are (1024, 768) in this example, corresponding to the position slightly to the lower right of the image center.

[0098] The above method establishes a nonlinear mapping relationship between image features and electrical characteristic parameters, and determines the defect location through mutual information measurement, effectively improving the accuracy and reliability of chip defect detection. This method is applicable to defect detection in various chip manufacturing processes, and it shows particularly good detection results for minute defects that are difficult to identify using traditional image processing methods.

[0099] In one optional implementation, the diffusion field distribution is calculated using the defect location as the source point, and the diffusion path of the defect is predicted by combining the spatiotemporal evolution law of the electrical characteristic anomaly point; the correlation coefficient of the image features of each sampling point on the diffusion path is calculated to determine the severity of the defect, including:

[0100] The local material conductivity gradient and stress intensity at the defect location are obtained. The diffusion coefficient is determined based on the ratio of the local material conductivity gradient to the stress intensity. The product of the diffusion coefficient and the gradient of the diffusion field intensity at the defect point is used as the diffusion field intensity distribution.

[0101] Electrical characteristic test data around the defect point is collected, and the feature vector of the electrical characteristic test data is extracted by wavelet transform. The local maxima of the feature vector are marked as anomalies. The propagation speed and acceleration of the anomalies are calculated, and the direction of anomaly propagation is determined based on the propagation speed and acceleration.

[0102] The diffusion field intensity distribution is converted into an intensity probability distribution, and the abnormal diffusion direction is converted into a directional probability distribution. The intensity probability distribution and the directional probability distribution are fused using the maximum entropy criterion to obtain a joint probability distribution. The defect diffusion path is determined along the direction with the maximum joint probability distribution.

[0103] Obtain the local image gray-level co-occurrence matrix of each sampling point on the defect diffusion path, map the gray-level co-occurrence matrix into gray-level feature vectors through sparse coding, calculate the correlation coefficient between the gray-level feature vectors; when the correlation coefficient is greater than the feature correlation threshold, determine that the defect has a diffusion risk, and output the location coordinates of the defect diffusion path.

[0104] This embodiment relates to a defect propagation prediction method. By analyzing the electrical properties and diffusion characteristics of defects in a material, the propagation path of the defects is predicted and their severity is assessed. The local material conductivity gradient and stress intensity at the defect location are obtained. Conductivity measurements are performed on the area surrounding the defect using a microelectrode array, with a measurement range of 10mm × 10mm and measurement intervals of 0.1mm. For example, after detecting a defect on the surface of a metallic material, the conductivity at the defect location was measured to be 2.3 × 10⁻⁶. 6 The conductivity is S / m, while the conductivity at a distance of 0.5 mm from the defect point is 2.5 × 10⁻⁶. 6 S / m, from which the conductivity gradient is calculated to be 0.4 × 10⁻⁶. 6 S / m·mm. Simultaneously, the stress intensity at the defect point was measured to be 120 MPa using a stress sensor. The diffusion coefficient was determined based on the ratio of the local material conductivity gradient to the stress intensity; in this example, the ratio was 3.33 × 10⁻⁶. 3 S / (m·MPa·mm), multiplying this value by the material property coefficient 0.025, yields the diffusion coefficient D = 0.083mm. 2 / s. The product of the diffusion coefficient and the gradient of the diffusion field intensity at the defect point is taken as the diffusion field intensity distribution. For example, when the diffusion field intensity gradient at the defect point is 0.5 / mm. 2 At that time, the diffusion field intensity distribution in this region was 0.0415 / mm. 2 ·s.

[0105] Electrical characteristic test data were collected around the defect point. Using a high-precision resistivity meter, resistivity data was collected in the area surrounding the defect point at a grid interval of 0.1 mm, forming a 100×100 data matrix. Feature vectors of the electrical characteristic test data were extracted using wavelet transform, specifically a three-level decomposition using the Db4 wavelet, retaining the wavelet coefficients of the third level to form the feature vector. Local maxima of the feature vector were marked as outliers by setting a threshold (e.g., the feature value exceeding twice the average value). For example, three outliers with coordinates (2.3 mm, 1.7 mm), (2.5 mm, 2.1 mm), and (2.8 mm, 2.4 mm) were detected in the collected data. The propagation velocity and acceleration of these outliers were obtained by calculating their positional changes at different time points (e.g., every 5 minutes). For example, if an anomaly moves from position (2.3mm, 1.7mm) to (2.5mm, 2.1mm) in 5 minutes, the calculated propagation velocity is 0.0894mm / min, and the direction is 58.0 degrees; then it moves to (2.8mm, 2.4mm) in 5 minutes, with a calculated propagation velocity of 0.0949mm / min, a direction of 45.0 degrees, and an acceleration of 0.0011mm / min. 2Based on the propagation speed and acceleration, the direction of the anomalous diffusion was determined to be approximately 51.5 degrees.

[0106] The diffusion field intensity distribution is converted into an intensity probability distribution, and normalized so that the sum of the diffusion field intensity values ​​at all points is 1. For example, within a 5mm × 5mm area around the defect, there are 2500 sampling points, and the normalized intensity value of each point varies between 0.0001 and 0.0015. The abnormal diffusion direction is converted into a directional probability distribution using a von Mises distribution, with the measured abnormal diffusion direction as the center and a standard deviation of 15 degrees, generating a directional probability density function within a 360-degree range. The intensity probability distribution and the directional probability distribution are fused using the maximum entropy criterion to obtain a joint probability distribution. In the specific implementation, for each spatial point, its joint probability is the weighted product of the intensity probability and the corresponding directional probability at that point, with weights of 0.6 and 0.4, respectively. Along the direction with the maximum joint probability distribution, starting from the defect point, the next maximum probability point is iteratively searched with a step size of 0.1mm to form the defect diffusion path. For example, starting from the defect point (0,0), the calculated diffusion path points are (0.1mm,0.08mm), (0.2mm,0.17mm), (0.3mm,0.25mm), etc.

[0107] The local image gray-level co-occurrence matrix (GLCM) of each sampling point along the defect diffusion path was obtained. A 1mm × 1mm region around each sampling point was acquired using a high-resolution scanning electron microscope (SEM) with a resolution of 0.01mm / pixel. The image gray values ​​were quantized to 16 levels, and the GLCM of each sampling point was calculated, with a matrix size of 16 × 16. The GLCM was mapped to gray-level feature vectors using sparse coding. The K-SVD algorithm was employed, with a dictionary size of 64 and a sparsity of 5, generating 32-dimensional feature vectors. The correlation coefficient between the gray-level feature vectors of adjacent sampling points was calculated using the Pearson correlation coefficient method. For example, the correlation coefficient between the gray-level feature vectors of the first and second points along the diffusion path was 0.92, and the correlation coefficient between the second and third points was 0.88. When the correlation coefficient is greater than the feature correlation threshold (set to 0.85), the defect is determined to have a risk of spreading, and the location coordinates of the defect spread path are output, such as [(0,0),(0.1mm,0.08mm),(0.2mm,0.17mm),(0.3mm,0.25mm)...].

[0108] The above method can accurately predict the diffusion path of defects in materials and assess their severity, providing an important basis for material quality control and fault prevention. In practical applications, the prediction accuracy reaches over 90%, and the average error in diffusion path prediction is less than 0.2 mm.

[0109] In one optional implementation, the diffusion field intensity distribution is converted into an intensity probability distribution, the abnormal diffusion direction is converted into a directional probability distribution, the intensity probability distribution and the directional probability distribution are fused using the maximum entropy criterion to obtain a joint probability distribution, and the defect diffusion path is determined along the direction with the maximum joint probability distribution, including:

[0110] The intensity distribution of the diffusion field is divided into adaptive equal-width intensity intervals, wherein the interval width is determined according to the standard deviation of the intensity distribution. The number of pixels in each intensity interval is counted and divided by the total number of pixels to obtain the intensity probability distribution.

[0111] An anisotropic Gaussian kernel function is used to adaptively smooth the abnormal diffusion direction, wherein the variance of the Gaussian kernel function is set along the gradient direction and perpendicular to the gradient direction respectively, to obtain a direction probability distribution with direction selectivity.

[0112] The intensity information entropy is obtained by multiplying the intensity probability distribution with its natural logarithm; the direction information entropy is obtained by multiplying the direction probability distribution with its natural logarithm; the intensity mean is obtained by multiplying the intensity probability distribution by the corresponding gray value; and the direction mean is obtained by multiplying the direction probability distribution by the corresponding direction angle.

[0113] Set initial weight parameters, add the intensity information entropy and the direction information entropy according to the weight parameters, and iteratively update the weight parameters to maximize the weighted sum, while ensuring that the mean intensity and mean direction weighted by the weight parameters meet the mean range; stop iterating when the difference between the weight parameters of two adjacent iterations is less than the convergence threshold, and obtain the optimal weight parameters; substitute the optimal weight parameters, intensity probability, and direction probability into the exponential function to obtain the joint probability distribution;

[0114] On the joint probability distribution, starting from the point of maximum probability, a path search is performed based on directional continuity constraints to obtain the defect propagation path.

[0115] In this embodiment, a method for converting the diffusion field intensity distribution and the abnormal diffusion direction into probability distributions and fusing them to obtain a joint probability distribution will be described in detail, thereby determining the defect diffusion path.

[0116] The process of converting the diffusion field intensity distribution into an intensity probability distribution requires determining an appropriate interval width. The system determines the interval width based on the standard deviation σ of the diffusion field intensity distribution, taking a value of 0.5σ. For example, when the standard deviation of the intensity distribution is 20, the interval width is 10. Assuming the intensity value range is from 0 to 255, it can be divided into 26 equally wide intervals, such as [0,9] and [10,19]. For each interval, the system counts the number of pixels falling within that interval. For example, if the interval [0,9] has 1000 pixels and the interval [10,19] has 2500 pixels, with a total of 100,000 pixels, then the probabilities for these two intervals are 0.01 and 0.025, respectively. This obtains the mapping relationship from intensity value to probability, i.e., the intensity probability distribution.

[0117] When converting abnormal diffusion directions into directional probability distributions, an anisotropic Gaussian kernel function is used for smoothing. The system sets the variance parameter of the Gaussian kernel function based on local gradient information. The variance along the gradient direction is set to a smaller value, such as 3, while the variance perpendicular to the gradient direction is set to a larger value, such as 9. This constructs a kernel function that is selective for direction. In practice, for each point (x, y) in the image, the system calculates its gradient direction and constructs a rotated Gaussian kernel. This kernel has a smaller variance along the gradient direction and a larger variance perpendicular to the gradient direction. This kernel function is then applied to convolve the original directional data to obtain a smoothed direction field. For example, for a region with an original direction of 45°, after applying an adaptive Gaussian kernel for smoothing, the direction values ​​of this region and its neighboring regions will be adjusted according to the kernel function weights, enhancing directional consistency and suppressing the influence of noise.

[0118] Calculating the information entropy and mean is a crucial step in fusing the probability distribution. The system then considers the intensity probability distribution p. i Process and calculate -∑p i ×ln(p i This serves as the intensity information entropy. Similarly, for the direction probability distribution q... j Calculate -∑q j ×ln(q j Let ∠p be the directional information entropy. Taking the intensity probability distribution as an example, assuming there are three intervals with probabilities of 0.2, 0.5, and 0.3 respectively, then the intensity information entropy is -(0.2×ln(0.2)+0.5×ln(0.5)+0.3×ln(0.3))=1.03. The mean intensity is calculated as ∑p i ×Z i Z i The midpoint of the interval is taken as an example. Using the three intervals mentioned above as examples, assuming the midpoints are 5, 15, and 25 respectively, the average intensity is 0.2×5 + 0.5×15 + 0.3×25 = 16. The average direction is calculated in a similar way, but the periodicity of the angle needs to be considered.

[0119] The weight optimization process of the maximum entropy criterion adopts an iterative approach. The system sets initial weights α = 0.5 and β = 0.5, and calculates the weighted information entropy H = α × H. 强度 +β×H 方向 Simultaneously, the system checks whether the weighted mean is within a preset range, i.e., α×μ. 强度 +β×μ 方向 A certain range constraint must be met, such as [10, 30]. If not, the weights are adjusted. Gradient ascent is used for adjustment, with the weight update amount proportional to the gradient of the entropy, and a step size of 0.01. For example, if the current α = 0.5 and β = 0.5, and the calculated gradient direction of the entropy indicates that increasing α and decreasing β is more advantageous, then the weights are updated to α = 0.51 and β = 0.49. During the iteration process, the iteration stops when the weight change between two adjacent iterations is less than 0.001, yielding the optimal weights α* and β*. The final joint probability distribution is calculated as exp(α* × ln(p...). i )+β*×ln(q j )), that is, p i α* ×q j β* .

[0120] The defect propagation path is determined based on a joint probability distribution. The point with the highest joint probability is selected as the starting point; for example, the probability of point (100, 150) is 0.08. Within an 8-neighborhood or larger of this point, the system searches for the next point with the second highest probability and whose direction deviates from the current path direction by no more than 30°. This directional continuity constraint ensures a smooth path. During the search, multiple candidate points are encountered, and the system selects the point with the highest joint probability that satisfies the directional constraint. For example, if the current point's direction is 60°, the next point's direction should be within the range of [30°, 90°]. The system continues this process until no new point satisfying the conditions is found or a preset path length limit, such as 200 points, is reached. The final sequence of points obtained is the defect propagation path.

[0121] In practical applications, the above methods can effectively handle noise and uncertainty in the diffusion field, and use the maximum entropy criterion to reasonably fuse intensity and direction information to accurately identify the defect diffusion path. For example, in a certain detection, the system successfully tracked a defect path about 180 pixels long. The average joint probability of this path was 0.065, and the maximum deflection angle was 28°, which indicates that the path has high reliability and smoothness.

[0122] The semiconductor chip defect detection and location system of this invention includes:

[0123] The first unit is used to acquire the original image of the semiconductor chip under test and the electrical test data at different sampling points;

[0124] The second unit is used to calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region, and cluster the pixels that exceed the deviation threshold to form defect candidate regions.

[0125] The third unit is used to perform variational mode decomposition on the electrical test data of the different sampling points to obtain the intrinsic mode function set, perform Hilbert transform on the intrinsic mode functions and calculate the instantaneous frequency, and mark the sampling points that exceed the frequency threshold as electrical characteristic anomalies.

[0126] The fourth unit is used to map the image features of the defect candidate region to the electrical characteristic parameter space through nonlinear transformation, convert the electrical characteristic anomaly points into the anomaly probability distribution in the image space through reverse mapping, calculate the feature mutual information value between the electrical characteristic parameter space and the image space, and determine that there is a defect point at the location when the feature mutual information value is greater than the correlation threshold.

[0127] The fifth unit is used to calculate the diffusion field distribution with the defect location as the source point, combine the spatiotemporal evolution law of electrical characteristic anomalies, predict the diffusion path of the defect, calculate the correlation coefficient of the image features of each sampling point on the diffusion path, and determine the severity of the defect.

[0128] A third aspect of the present invention provides an electronic device, comprising:

[0129] processor;

[0130] Memory used to store processor-executable instructions;

[0131] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0132] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0133] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions 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. A method for detecting and locating defects in semiconductor chips, characterized in that, include: Acquire raw images of the semiconductor chip under test and electrical test data at different sampling points; Calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region. Cluster the pixels that exceed the deviation threshold to form defect candidate regions. Variational mode decomposition is performed on the electrical test data at the different sampling points to obtain the intrinsic mode function set. Hilbert transform is performed on the intrinsic mode functions and the instantaneous frequency is calculated. Sampling points that exceed the frequency threshold are marked as abnormal points in electrical characteristics. The image features of the defect candidate region are mapped to the electrical characteristic parameter space through nonlinear transformation. The electrical characteristic anomalies are converted into anomaly probability distributions in the image space through reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point at that location. The diffusion field distribution is calculated with the defect location as the source point, and the diffusion path of the defect is predicted by combining the spatiotemporal evolution law of the electrical characteristic anomaly point. Calculate the correlation coefficient of image features at each sampling point along the diffusion path to determine the severity of the defect.

2. The method according to claim 1, characterized in that, Calculate the local density value of each pixel in the original image and the deviation from the mean density of the region. Cluster pixels exceeding the deviation threshold to form candidate defect regions, including: The original image is divided into multiple local windows. The standard deviation of gray values ​​of pixels in each local window is calculated. The product of the standard deviation of gray values ​​and the adjustment coefficient is added to the unit base and then multiplied by the first reference bandwidth parameter to obtain the adaptive bandwidth parameter. The adaptive bandwidth parameter is substituted into the Gaussian kernel function to calculate the local density value of the center pixel. A multi-scale feature pyramid is constructed based on the local density values. The mean local density and the standard deviation of local density are calculated at each scale level. The local density mean of the scale level is subtracted from the local density mean of the central pixel at each scale level, and then divided by the local density standard deviation to obtain the standardized deviation. When the standardized deviation of any scale level is greater than the deviation threshold, the position coordinates of the central pixel are recorded as outlier coordinates. Calculate the density flow between adjacent anomalies, where the density flow is the rate of change of the local density value along the path connecting the two anomalies with spatial location. When the maximum value of the density flow is less than the second reference parameter, the two anomalies are marked as belonging to the same anomaly region, and the set of anomalies with the same anomaly region number is output as a defect candidate region.

3. The method according to claim 1, characterized in that, Variational mode decomposition is performed on the electrical test data at the different sampling points to obtain a set of intrinsic mode functions. Hilbert transform is then performed on the intrinsic mode functions, and the instantaneous frequencies are calculated. Sampling points exceeding the frequency threshold are marked as electrical characteristic anomalies, including: Within a sliding time window, the local variance of the electrical test data at different sampling points is calculated. The local variance is used as a constraint to perform variational mode decomposition to obtain an intrinsic mode function set. The energy density of each intrinsic mode function in the intrinsic mode function set is calculated. The orthogonality index between intrinsic mode functions with energy densities greater than the energy threshold is calculated. Intrinsic mode functions with orthogonality indexes less than the orthogonality threshold are determined as valid intrinsic mode functions. The effective intrinsic mode function is subjected to Hilbert transform to obtain an analytical signal. The generalized zero-crossing rate is calculated based on the zero-crossing characteristics of the analytical signal. The time derivative of the generalized zero-crossing rate is divided by the frequency normalization coefficient to obtain the instantaneous frequency. Spatial weighting coefficients are assigned based on the location distance of the sampling points. The sum of the products of the spatial weighting coefficients and the instantaneous frequency is divided by the sum of the spatial weighting coefficients to obtain the reference frequency. The frequency deviation between the instantaneous frequency and the reference frequency of different sampling points is calculated, and sampling points that exceed the frequency threshold are marked as abnormal electrical characteristics.

4. The method according to claim 1, characterized in that, Image features of candidate defect regions are mapped to an electrical characteristic parameter space through a nonlinear transformation. Electrical characteristic anomalies are converted into anomaly probability distributions in the image space through a reverse mapping. The feature mutual information values ​​of the electrical characteristic parameter space and the image space are calculated. When the feature mutual information value is greater than a correlation threshold, a defect point is determined to exist at that location, including: Obtain a candidate defect region image of the chip under test, and decompose the candidate defect region image into a multi-scale feature vector through discrete wavelet transform; The local covariance matrix of the multi-scale eigenvectors is calculated, and the eigenvalue decomposition of the local covariance matrix is ​​performed to obtain an eigenspace. Multiple eigenvectors with the top 5% contribution rates in the eigenspace are extracted to construct a mapping basis. The sum of the inner product of the mapping basis and the radial basis kernel function and the adaptive weight coefficients is determined as a nonlinear mapping function. The multi-scale eigenvectors are transformed to the electrical characteristic parameter space through the nonlinear mapping function to obtain the parameter mapping vector. The position potential energy is calculated based on the Mahalanobis distance between the parameter mapping vector and the electrical characteristic anomaly point, and the adjacent potential energy is calculated based on the Euclidean distance between adjacent parameter mapping vectors. The conditional probability is determined according to the weighted combination of the position potential energy and the adjacent potential energy. The conditional probability is then used to obtain the inverse mapping probability through maximum likelihood estimation. Finally, the inverse mapping probability is transformed into the image space to obtain the defect probability distribution. The ratio of the first edge probability corresponding to the parameter mapping vector to the second edge probability corresponding to the defect probability distribution is used as the feature mutual information value; when the feature mutual information value is greater than the correlation threshold, it is determined that there is a defect point in the defect candidate region, and the spatial coordinate information of the defect point is output.

5. The method according to claim 1, characterized in that, The diffusion field distribution is calculated with the defect location as the source point, and the diffusion path of the defect is predicted by combining the spatiotemporal evolution law of the electrical characteristic anomaly point. Calculate the correlation coefficient of image features at each sampling point along the diffusion path to determine the severity of the defect, including: The local material conductivity gradient and stress intensity at the defect location are obtained. The diffusion coefficient is determined based on the ratio of the local material conductivity gradient to the stress intensity. The product of the diffusion coefficient and the gradient of the diffusion field intensity at the defect point is used as the diffusion field intensity distribution. Electrical characteristic test data around the defect point is collected, and the feature vector of the electrical characteristic test data is extracted by wavelet transform. The local maxima of the feature vector are marked as anomalies. The propagation speed and acceleration of the anomalies are calculated, and the direction of anomaly propagation is determined based on the propagation speed and acceleration. The diffusion field intensity distribution is converted into an intensity probability distribution, and the abnormal diffusion direction is converted into a directional probability distribution. The intensity probability distribution and the directional probability distribution are fused using the maximum entropy criterion to obtain a joint probability distribution. The defect diffusion path is determined along the direction with the maximum joint probability distribution. Obtain the local image gray-level co-occurrence matrix of each sampling point on the defect diffusion path, map the gray-level co-occurrence matrix into gray-level feature vectors through sparse coding, calculate the correlation coefficient between the gray-level feature vectors; when the correlation coefficient is greater than the feature correlation threshold, determine that the defect has a diffusion risk, and output the location coordinates of the defect diffusion path.

6. The method according to claim 5, characterized in that, The diffusion field intensity distribution is converted into an intensity probability distribution, and the abnormal diffusion direction is converted into a directional probability distribution. The intensity probability distribution and the directional probability distribution are then fused using the maximum entropy criterion to obtain a joint probability distribution. The defect diffusion path is determined along the direction with the maximum joint probability distribution, including: The intensity distribution of the diffusion field is divided into adaptive equal-width intensity intervals, wherein the interval width is determined according to the standard deviation of the intensity distribution. The number of pixels in each intensity interval is counted and divided by the total number of pixels to obtain the intensity probability distribution. An anisotropic Gaussian kernel function is used to adaptively smooth the abnormal diffusion direction, wherein the variance of the Gaussian kernel function is set along the gradient direction and perpendicular to the gradient direction respectively, to obtain a direction probability distribution with direction selectivity. The intensity information entropy is obtained by multiplying the intensity probability distribution with its natural logarithm; the direction information entropy is obtained by multiplying the direction probability distribution with its natural logarithm; the intensity mean is obtained by multiplying the intensity probability distribution by the corresponding gray value; and the direction mean is obtained by multiplying the direction probability distribution by the corresponding direction angle. Set initial weight parameters, add the intensity information entropy and the direction information entropy according to the weight parameters, and iteratively update the weight parameters to maximize the weighted sum, while ensuring that the mean intensity and mean direction weighted by the weight parameters meet the mean range; stop iterating when the difference between the weight parameters of two adjacent iterations is less than the convergence threshold, and obtain the optimal weight parameters; substitute the optimal weight parameters, intensity probability, and direction probability into the exponential function to obtain the joint probability distribution; On the joint probability distribution, starting from the point of maximum probability, a path search is performed based on directional continuity constraints to obtain the defect propagation path.

7. A semiconductor chip defect detection and location system, used to implement the method as described in any one of claims 1-6, characterized in that, include: The first unit is used to acquire the original image of the semiconductor chip under test and the electrical test data at different sampling points; The second unit is used to calculate the local density value of each pixel in the original image and calculate the deviation from the mean density of the region, and cluster the pixels that exceed the deviation threshold to form defect candidate regions. The third unit is used to perform variational mode decomposition on the electrical test data of the different sampling points to obtain the intrinsic mode function set, perform Hilbert transform on the intrinsic mode functions and calculate the instantaneous frequency, and mark the sampling points that exceed the frequency threshold as electrical characteristic anomalies. The fourth unit is used to map the image features of the defect candidate region to the electrical characteristic parameter space through nonlinear transformation, convert the electrical characteristic anomaly points into the anomaly probability distribution in the image space through reverse mapping, calculate the feature mutual information value between the electrical characteristic parameter space and the image space, and determine that there is a defect point at the location when the feature mutual information value is greater than the correlation threshold. The fifth unit is used to calculate the diffusion field distribution with the defect location as the source point, and predict the diffusion path of the defect by combining the spatiotemporal evolution law of the electrical characteristic anomaly point. Calculate the correlation coefficient of image features at each sampling point along the diffusion path to determine the severity of the defect.

8. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory to execute the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the method described in any one of claims 1 to 6.

Images