Method and system for predicting stability of integrated circuit test equipment, equipment and medium

Abstract

The invention discloses a method, a system and equipment for predicting the stability of integrated circuit test equipment and a medium, and belongs to the technical field of integrated circuit test. The method comprises the following steps: firstly, carrying out preprocessing and feature selection on historical data in an FT test stage, and adopting a Gaussian mixture model (GMM) to cluster and identify different operation condition clusters of a test machine; then, establishing a health state GMM reference for each working condition cluster, calculating a KL divergence value of a normal sample and the reference, and setting a dynamic anomaly detection threshold by 99.73% quantile of the KL divergence value; and finally, in real-time monitoring, calculating a KL divergence value of real-time data and a corresponding working condition cluster benchmark, and comparing the KL divergence value with a dynamic threshold value to realize accurate anomaly marking. The method effectively solves the problem of abnormal detection of the test data of the integrated circuit under complex and changeable working conditions, and improves the monitoring accuracy and working condition adaptability.

Description

Technical Field

[0001] This invention relates to the field of integrated circuit test machine performance research, specifically to methods, systems, equipment, and media for predicting the stability of integrated circuit test equipment. Background Technology

[0002] Automatic Test Equipment (ATE) is a core component for performing chip functional verification and performance evaluation, and its technological evolution is closely related to the development of integrated circuits. From the first-generation ATE equipment in the early 1960s, which only supported 16-pin DC parameter testing, to modern high-end test systems capable of handling more than 1024 pins and test rates exceeding 500MHz, ATE equipment has made significant progress in test accuracy, parallel processing capabilities, and functional coverage. However, the contradiction between test efficiency and test cost has become increasingly prominent. On the one hand, increased chip complexity leads to a surge in test data volume, with the test vector scale of a single high-end SoC chip reaching TB levels; on the other hand, extended test time directly drives up production costs. Statistics show that every 1 millisecond increase in test time results in tens of thousands of dollars in additional costs for a wafer fab with a monthly production capacity of millions of wafers. Against this backdrop, tester stability has become a core element in balancing test quality and test cost.

[0003] Stability analysis is a crucial process that comprehensively considers factors such as hardware, software, and environment. At the hardware level, in-depth and meticulous stability analysis and optimization of the test equipment's power system, clock circuit, and interface subsystem are required. This forms the physical foundation for ensuring the consistency and reliability of test results. At the software level, the focus is on the robustness of the test program and core algorithms, aiming to ensure that the test process maintains accurate, consistent, and repeatable logic cores even under various boundary conditions, noise interference, and minor fluctuations in the inherent parameters of the device under test. Furthermore, at the environmental level, given the potential diversity and dynamism of the actual operating environment of the test equipment, the stability of environmental factors (such as temperature, humidity, vibration spectrum, and electromagnetic compatibility) must also be included in the system considerations, as their impact permeates the entire process of maintaining test accuracy and ensuring equipment reliability.

[0004] Current mainstream ATE anomaly detection methods are based on the assumption of a static normal distribution: screening test data features that conform to a normal distribution and using the upper boundary value of the normal distribution as the anomaly threshold. However, integrated circuit testing conditions are complex and variable—differences in the process / type / testing stage of the chip under test, ATE hardware aging, and dynamic environmental changes all lead to dynamic shifts or changes in the variance of the test data distribution, making traditional methods difficult to adapt: ​​either normal data is misjudged as anomalies, increasing chip scrapping and testing costs; or actual anomaly data is missed, resulting in defective chips entering the market, causing repair and recall losses and brand risks.

[0005] Therefore, there is an urgent need to propose a prediction method to solve the above problems. Summary of the Invention

[0006] The technical problem to be solved by this invention is how to address the issue of false detection or missed detection caused by the difficulty of traditional anomaly detection methods in adapting to dynamic data distribution changes under complex and ever-changing integrated circuit testing conditions.

[0007] This invention solves the above-mentioned technical problems through the following technical means: a method for predicting the stability of integrated circuit testing equipment, comprising: Step S1, Data Preprocessing: Obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features. Step S2, Operating Condition Identification: Based on the preprocessed historical test dataset, unsupervised learning is performed using the Gaussian mixture model clustering algorithm to divide the test data into multiple operating condition clusters; Step S3, State Distribution Modeling: Establish a Gaussian Mixture Model (GMM) baseline for the healthy state for each working condition cluster. For each working condition cluster, calculate the KL divergence value of all its training samples relative to the healthy state GMM baseline of that working condition cluster. Based on the distribution of the KL divergence value, determine the dynamic anomaly detection threshold for that working condition cluster. Step S4: Real-time monitoring: Acquire real-time integrated circuit test data, determine its operating condition cluster after data preprocessing; calculate its KL divergence value relative to the corresponding operating condition cluster's health status GMM benchmark; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, mark the test data as abnormal.

[0008] This invention first optimizes the quality of input data through data cleaning, feature selection, and standardization. Then, it uses GMM clustering to accurately identify different operating conditions of the test machine, establishes a dedicated health model for each condition, and determines dynamic thresholds. Finally, it combines real-time monitoring to achieve targeted anomaly detection. This process is adaptable to the multi-condition characteristics of the equipment and can dynamically track stability changes, effectively reducing test deviations caused by misjudgments of equipment status, and lowering the risk of defective products leaving the equipment and maintenance cost losses.

[0009] Furthermore, the feature selection includes: Calculate the skewness, kurtosis, coefficient of variation, and interquartile range for each test feature. The KS test was used to evaluate the goodness of fit between the test features and 15 probability distribution models. The distribution model with the highest goodness of fit is selected as the optimal distribution model for that feature, and features that conform to or are close to a normal distribution are preferentially selected as the training set.

[0010] This invention analyzes feature distribution from multiple dimensions and prioritizes features that conform to or are close to a normal distribution as the training set, thereby reducing the error caused by feature mismatch with the model, providing high-quality input for subsequent predictions, and improving the accuracy of the results.

[0011] Furthermore, step S2 specifically includes: Initialization parameters: Specifies the number of clusters, and randomly initializes the weights for each Gaussian distribution. Mean vector Covariance Matrix ; Step E: Calculate each sample Belongs to the The posterior probability of a Gaussian distribution, i.e., the responsibility function ; M-step: Update the parameters of each Gaussian distribution according to the responsibility function; definition: For the first The number of effective samples from a Gaussian distribution; The weight update formula is as follows:

[0012] The mean update formula is:

[0013] The formula for updating the covariance matrix is:

[0014] Iteration and Termination: Repeat the E-step and M-step until the parameter change is less than the preset threshold, the log-likelihood change is less than the preset value, or the preset maximum number of iterations is reached. Select the optimal parameters of the model by using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Assign each sample to the cluster corresponding to the Gaussian distribution with the highest probability. Each cluster represents a working condition of the test machine. During the iteration process, log-likelihood estimation is used to evaluate the convergence of the parameters, and its calculation formula is as follows:

[0015] Where t represents the number of iterations. X Represents the entire observation dataset, for , Sample size This represents the total number of components in the Gaussian mixture. The first component of the Gaussian mixture One portion, It is the weight of the k-th Gaussian component. For the first The probability density function of a Gaussian distribution. For the first The mean vector of Gaussian components, For the first k The covariance matrix of Gaussian components.

[0016] This invention standardizes the GMM clustering iteration process and combines AIC / BIC to select the optimal parameters, ensuring accurate and stable working condition division and providing a reliable basis for health modeling and anomaly detection.

[0017] Furthermore, the optimal parameters of the model are selected using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). The calculation formulas for AIC and BIC are as follows:

[0018]

[0019] in, This represents the number of parameters in the model.

[0020] This invention selects the optimal parameters by combining the AIC / BIC criteria, avoiding overfitting or underfitting, ensuring accurate and concise division of operating conditions, and further improving the accuracy and reliability of equipment stability prediction.

[0021] Furthermore, step S3 specifically includes: For each working condition cluster divided in step S2, a GMM model is trained as a health status benchmark. The probability density function of the GMM model is:

[0022] in, It is the probability density function of a multivariate Gaussian distribution.

[0023] This invention constructs a dedicated GMM health benchmark for each operating condition cluster, accurately depicts the normal state of different operating conditions, provides a standardized basis for KL divergence calculation, and improves the accuracy of anomaly detection.

[0024] Furthermore, the formula for calculating the KL divergence value is as follows:

[0025] in, For the true distribution, For reference distribution, The probability density function of a healthy Gaussian mixture model at point The value at that location.

[0026] This invention uses KL divergence to quantitatively measure the difference between a sample and a healthy baseline, providing an objective judgment indicator, avoiding subjective errors, and providing reliable support for determining dynamic thresholds.

[0027] Furthermore, the dynamic anomaly detection threshold is taken as the 99.73% quantile of the KL divergence values ​​of all training samples in the corresponding working condition cluster.

[0028] This invention uses a dynamic threshold set at the 99.73 percentile to adapt to the characteristics of various operating conditions, significantly reducing the misjudgment rate of normal data and the missed judgment rate of abnormal data, thus ensuring the reliability of prediction.

[0029] This invention also provides a system for predicting the stability of integrated circuit testing equipment, comprising: Data preprocessing module: used to obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features. The working condition identification module is used to divide the test data into multiple working condition clusters based on the preprocessed historical test dataset using the Gaussian mixture model clustering algorithm for unsupervised learning. State distribution modeling module: used to establish a Gaussian mixture model benchmark for the healthy state for each working condition cluster. For each working condition cluster, calculate the KL divergence value of all its training samples relative to the healthy state GMM benchmark of that working condition cluster. Based on the distribution of the KL divergence value, determine the dynamic anomaly detection threshold of that working condition cluster. Real-time monitoring module: Used to acquire real-time integrated circuit test data, determine the operating condition cluster to which the data belongs after data preprocessing; calculate the KL divergence value of the data relative to the GMM health status benchmark of the corresponding operating condition cluster; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, the test data is marked as abnormal.

[0030] The present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above-described method steps by calling the program instructions.

[0031] The present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the above-described method steps.

[0032] The advantages of this invention are: This invention achieves precise multi-condition monitoring and early fault warning for the stability of integrated circuit test equipment by combining Gaussian Mixture Model (GMM) clustering with KL divergence anomaly detection. GMM learning automatically identifies various operating conditions of the test machine and establishes an independent health status benchmark model for each condition, effectively solving the false alarm and missed alarm problems caused by operating condition fluctuations in traditional methods. By calculating the KL divergence between real-time data and the corresponding operating condition benchmark and employing a dynamic threshold mechanism, abnormal deviations can be sensitively captured, significantly improving the accuracy and adaptability of equipment status monitoring. Furthermore, its feature selection process prioritizes normally distributed feature variables, enhancing the reliability of model statistical inference and providing a data-driven high-reliability and stability management solution for integrated circuit testing. Attached Figure Description

[0033] Figure 1 This is a flowchart of the method for predicting the stability of integrated circuit testing equipment according to Embodiment 1 of the present invention; Figure 2 This is a two-dimensional visualization of the clustering results of the test machine data in Embodiment 1 of the present invention; Figure 3 This is a confusion matrix diagram of the method for predicting the stability of integrated circuit testing equipment according to Embodiment 1 of the present invention; Figure 4 The ROC curve is the prediction method for the stability of integrated circuit testing equipment according to Embodiment 1 of the present invention. Detailed Implementation

[0034] 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 in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0035] Example 1 Figure 1 A flowchart illustrating a method for predicting the stability of integrated circuit test equipment. This method includes: Step S1, Data Preprocessing: Obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features.

[0036] Specifically, S11, Data Selection: Randomly select a set of test data (at least 5000 datasets) from the corresponding test data obtained after the integrated circuit FT stage test as the dataset for model training.

[0037] S12. Feature engineering includes: S121. Data Cleaning: Key steps in data cleaning include handling missing values ​​and outliers. For missing values, median imputation is used. For outliers, the LOF (Local Outlier Factor) algorithm is used for detection and removal. LOF is a density-based anomaly detection method suitable for identifying local outliers.

[0038] S122. Feature Selection: IC (Integrated Circuit) test data has high feature dimensionality. To reduce model training time, the test data is fitted with a distribution, and features that conform to or are close to a normal distribution are selected as the training set. Distribution fitting is performed based on the skewness, kurtosis, coefficient of variation, and interquartile range of the data distribution.

[0039] Skewness measures the asymmetry of a probability distribution, reflecting the degree and direction of data deviation from the center. Skewness is an important indicator in statistics describing the shape of a distribution. The calculation formula is:

[0040] in, Sample size It is the first One test sample, It is the mean of the test sample.

[0041] Kurtosis measures the thickness and flatness of the tails of a distribution, reflecting the probability of extreme values ​​occurring. It describes whether a distribution is steeply apex or flatly apex. The calculation formula is:

[0042] The coefficient of variation is a measure of relative dispersion, used to compare the degree of dispersion of datasets with different dimensions. The formula is:

[0043] in, It is the standard deviation of the test sample.

[0044] The interquartile range is a robust indicator that describes the dispersion of the middle 50% of observations in a dataset, and is not affected by extreme values.

[0045]

[0046] The calculation steps include: 1. Sort the data in ascending order:

[0047] 2. Calculate the first quartile. :

[0048] 3. Calculate the third quartile. :

[0049] Based on the above statistical characteristics, 15 common probability distribution models were selected to fit the test data features. These 15 probability distribution models include: beta distribution, t distribution, log-normal distribution, gamma distribution, normal distribution, uniform distribution, exponential distribution, Weibull distribution, chi-square distribution, F distribution, Laplace distribution, logistic distribution, Cauchy distribution, Pareto distribution, and Rayleigh distribution. The KS test was used to reflect the acceptance of the data by each probability model. The KS test is a non-parametric hypothesis testing method used to assess whether a sample dataset comes from a specific distribution. The calculation formula is:

[0050] in, It is the KS test statistic, and sup is the supremum. Experience distribution Theoretical distribution.

[0051] The empirical distribution of each feature was compared with various theoretical distributions using the KS test, and the distribution with the highest good fit for each feature was selected as its optimal distribution model. Furthermore, features with well-shaped distributions (including normal, beta, t, and gamma distributions) that are conducive to subsequent GMM modeling were prioritized as the final training set, and the selected features covered 95% of all features.

[0052] S123. Data Standardization: Since the selected features have different dimensions, in order to unify the dimensions and GMM clustering, and improve the convergence speed and prediction accuracy of subsequent algorithms, the selected feature data is standardized using the Z-Score standardization method.

[0053] Z-Score standardization is a common method in data processing. It allows data of different magnitudes to be converted into a unified Z-Score for comparison. The calculation formula is:

[0054] in, It is the standard deviation of the test sample.

[0055] After standardization, different test items in the training set data can be clustered, which ensures the feasibility of subsequent clustering.

[0056] Step S2, Operating Condition Identification: Based on the preprocessed historical test dataset, unsupervised learning is performed using the Gaussian mixture model clustering algorithm to divide the test data into multiple operating condition clusters.

[0057] Because the operating conditions of the test machine are complex and varied, a single model cannot accurately capture its operating characteristics. Therefore, in order to improve the accuracy of prediction, it is necessary to effectively identify the operating status of the test machine under different operating conditions.

[0058] Gaussian Mixture Model (GMM) is a probabilistic clustering algorithm that assumes data is a mixture of multiple Gaussian distributions. Each Gaussian distribution corresponds to a cluster, and data points belong to each cluster with a certain probability. Its core idea is: data points are generated from a mixture of multiple Gaussian distributions, and the EM algorithm alternately performs the E-step (calculating the responsibility value of a data point belonging to each Gaussian distribution) and the M-step (updating the model parameters with the calculated responsibility value) until convergence. Ultimately, each data point is assigned to the cluster corresponding to the Gaussian distribution with the highest probability. The GMM clustering algorithm can be divided into four steps: 1. Specify the number of clusters and randomly initialize the mean vector, covariance matrix, and Gaussian weights of each Gaussian distribution; 2. E-step: Calculate the posterior probability of each data point belonging to each Gaussian distribution; 3. M-step: Based on the posterior probability calculated in the E-step, re-estimate the parameters of each Gaussian distribution to maximize the probability of the data point under that distribution; 4. Repeat the E-step and M-step until the parameters converge. Specifically: (1) Initialize parameters Weights of the Gaussian distribution:

[0059] Mean vector:

[0060] Covariance matrix:

[0061] Where D represents the dimension of the data, i.e., the number of features in each dataset. This indicates that it is a D A real matrix of D.

[0062] (2) E-step: Calculate the posterior probability (responsibility function) For each sample Calculate its belonging to the first k Posterior probabilities of a Gaussian distribution:

[0063] in, For the first The probability density function of Gaussian components, For the first The mean vector of Gaussian components, For the first The covariance matrix of the nth Gaussian component. This value represents the... The sample "belongs to" the first The "responsibility" of a Gaussian distribution.

[0064] (3) M step: Update parameters definition: For the first The number of "effective samples" in a Gaussian distribution.

[0065] The updated formula is as follows: The weight update formula is:

[0066] The mean update formula is:

[0067] The formula for updating the covariance matrix is:

[0068] Where t represents the number of iterations.

[0069] (4) Iteration and Termination: Repeat steps E and M until the parameter change is less than a preset threshold, the log-likelihood change is less than a preset value, or the preset maximum number of iterations is reached. During the iteration process, log-likelihood estimation can be used to evaluate the convergence of the parameters. Its calculation formula is as follows:

[0070] in, This represents the total number of components in the Gaussian mixture. k The first component of the Gaussian mixture k One portion, Representing observation data In parameters The log-likelihood value is given below. It is the first k The weights of each Gaussian component, For the first k The probability density function of a Gaussian distribution. For the first k The mean vector of Gaussian components, For the first k The covariance matrix of Gaussian components. This value tends to stabilize with the number of iterations.

[0071] Akaike Information Criterion (AIC): Used to measure the goodness of a model fit. The calculation formula is:

[0072] in, This represents the number of parameters in the model, and its purpose is to prevent overfitting.

[0073] Bayesian Information Criterion (BIC): Used to measure the fit of a model, it avoids overfitting by incorporating a penalty term. The calculation formula is:

[0074] Finally, the optimal model parameters are selected based on the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). After multiple iterations, a set of optimal estimated parameters is obtained, and each sample is assigned to the most probable cluster, thus completing GMM clustering. Specifically, this method iterates through different candidate cluster numbers R, each for its training Gaussian mixture model, and calculates the corresponding AIC and BIC values. The R that minimizes both AIC and BIC values ​​is selected as the optimal number of clusters. Based on this optimal R value, the EM algorithm is iteratively converged to obtain the final model parameters (i.e., the weights, mean vector, and covariance matrix of each Gaussian component), and each sample is assigned to the cluster with the highest posterior probability, thus completing GMM clustering.

[0075] After clustering, each cluster represents a specific operating state during the actual operation of the test machine. The two-dimensional clustering result obtained through PCA dimensionality reduction is as follows: Figure 2 As shown in the results, the boundaries between clusters are clear, indicating that the GMM clustering algorithm can effectively capture the inherent distribution characteristics of the data.

[0076] Step S3, State Distribution Modeling: Establish a Gaussian Mixture Model (GMM) baseline for the healthy state for each operating condition cluster. For each operating condition cluster, calculate the KL divergence value of all its training samples relative to the healthy state GMM baseline of that operating condition cluster. Based on the distribution of the KL divergence value, determine the dynamic anomaly detection threshold for that operating condition cluster.

[0077] By incorporating the KL divergence into the pre-constructed working condition clusters, the state distribution is modeled to determine the probability density model of the "healthy" state, and the dynamic detection threshold is determined based on this model. The main steps can be divided into three steps: (1) training a Gaussian mixture model for each working condition cluster as a baseline for the healthy state; (2) calculating the KL divergence of all normal samples; and (3) determining the dynamic monitoring threshold based on the KL divergence distribution. The formula is: S31, GMM Model Construction The probability density function of GMM is:

[0078] in, It is the probability density function of a multivariate Gaussian distribution.

[0079] S32, KL divergence calculation KL divergence can quantify the degree of deviation of each sample from the healthy state. During the training phase of the healthy model, the KL divergence values ​​of normal samples are stored to determine the boundary values ​​of the healthy state.

[0080]

[0081] in, This represents the true distribution (i.e., the distribution to be evaluated). This is the reference distribution (i.e., the model distribution that has been constructed based on the sample data). and They are respectively and The probability distribution.

[0082] In this method, due to the requirement of real-time monitoring, the test samples collected at each moment need to be evaluated immediately, and are regarded as a highly concentrated distribution (mathematically approximated as a Dirac delta distribution). Therefore, the KL divergence can be simplified to the following form.

[0083]

[0084] in, This indicates the current sample point to be detected. Consider it as a definite point (mathematically, it can be modeled as a "Dirac delta distribution", the probability mass of which is completely concentrated in the center). point) The probability density function of the healthy Gaussian mixture model at point... The value at that location is calculated based on the sample point to be detected. The probability density of an occurrence in a healthy state. The larger the KL value, the greater the difference between the current sample and a healthy state, and the higher the probability of an anomaly. S33, Dynamic Threshold Determination Based on KL divergence, a boundary threshold for health status is determined for each cluster. A 99.73% quantile is introduced as a dynamic threshold, meaning that only 0.27% of the healthy samples in the training data will have KL divergence values ​​exceeding this threshold. This quantile approach is independent of the specific shape of the data distribution (such as a normal distribution) and therefore has stronger robustness. Specifically, for each cluster, the KL divergence values ​​of all its healthy training samples are calculated, and then the dynamic threshold is set to the 99.73% quantile of that KL divergence value sequence.

[0085] Specifically, for each operating condition cluster, we collect the KL divergence values ​​of all its healthy training samples, forming a sequence from smallest to largest:

[0086] in ，S This represents the number of samples in the cluster.

[0087] The dynamic threshold is directly selected from the 99.73% empirical quantile of the sequence.

[0088] in, Represents the empirical distribution function. This is the threshold for dynamic anomaly detection used for this cluster.

[0089] This means:

[0090] in, It is a KL divergence random variable obtained based on health data.

[0091] By using the 99.73% quantile of the KL divergence as the dynamic threshold, and then saving the GMM model and threshold for the cluster, the task of GMM-KL dynamic monitoring anomaly modeling can be completed.

[0092] Step S4: Real-time monitoring: Acquire real-time integrated circuit test data, determine its operating condition cluster after data preprocessing; calculate its KL divergence value relative to the corresponding operating condition cluster's health status GMM benchmark; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, mark the test data as abnormal.

[0093] First, 5% of outlier data was randomly injected into the test set and recorded. Then, the test samples were standardized using the standardization method described earlier. After standardization, the test set had characteristics consistent with the training set and could be directly applied to the monitoring model. The trained clustering model was used to predict the operating condition cluster to which the test set data belonged. For each test data point, the corresponding health GMM model was loaded according to its operating condition cluster, and the KL divergence value of the data point was calculated. The KL divergence was compared with the dynamic anomaly detection threshold of the operating condition cluster. If the KL divergence exceeded the threshold, the data point was marked as an anomaly. Finally, the performance of the method was verified using evaluation metrics such as confusion matrix, precision, and recall. Experimental verification was conducted. Figure 3 The confusion matrix results of this method are shown. Figure 4 The ROC curve is shown, and its precision, recall, and F1 score demonstrate the model's ability to distinguish between true and false data.

[0094] Performance metrics: Accuracy: 0.9924.

[0095] Accuracy: 0.9981.

[0096] Recall rate: 0.8502.

[0097] F1 score: 0.9182.

[0098] Example 2 Based on Embodiment 1, Embodiment 2 of the present invention also provides a system for predicting the stability of integrated circuit testing equipment, comprising: Data preprocessing module: used to obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features.

[0099] Specifically, the data preprocessing module includes a feature selection unit, which calculates the skewness, kurtosis, coefficient of variation, and interquartile range for each test feature. It uses the KS test to evaluate the goodness of fit between the test feature and 15 probability distribution models, selects the distribution model with the highest goodness of fit as the optimal distribution model for that feature, and prioritizes features that conform to or are close to a normal distribution as the training set.

[0100] Operating condition identification module: Based on the preprocessed historical test dataset, it uses Gaussian mixture model clustering algorithm for unsupervised learning to divide the test data into multiple operating condition clusters.

[0101] Specifically, the working condition recognition module includes: Initialization parameters: Specifies the number of clusters, and randomly initializes the weights for each Gaussian distribution. Mean vector Covariance Matrix .

[0102] Step E: Calculate each sample Belongs to the The posterior probability of a Gaussian distribution, i.e., the responsibility function .

[0103] M-step: Update the parameters of each Gaussian distribution based on the responsibility function.

[0104] definition: For the first The number of effective samples from a Gaussian distribution.

[0105] The weight update formula is as follows:

[0106] The mean update formula is:

[0107] The formula for updating the covariance matrix is:

[0108] Iteration and Termination: Repeat the E-step and M-step until the parameter change is less than the preset threshold, the log-likelihood change is less than the preset value, or the preset maximum number of iterations is reached. Select the optimal parameters of the model by using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Assign each sample to the cluster corresponding to the Gaussian distribution with the highest probability. Each cluster represents a working condition of the test machine. During the iteration process, log-likelihood estimation is used to evaluate the convergence of the parameters, and its calculation formula is as follows:

[0109] Where t represents the number of iterations. X Represents the entire observation dataset, for , Sample size This represents the total number of components in the Gaussian mixture. The first component of the Gaussian mixture One portion, It is the weight of the k-th Gaussian component. For the first The probability density function of a Gaussian distribution. For the first The mean vector of Gaussian components, For the first k The covariance matrix of Gaussian components.

[0110] The optimal parameters of the model are selected using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). The formulas for calculating AIC and BIC are as follows:

[0111]

[0112] in, This represents the number of parameters in the model.

[0113] State distribution modeling module: used to establish a Gaussian mixture model benchmark for the healthy state for each operating condition cluster. For each operating condition cluster, the KL divergence value of all its training samples relative to the healthy state GMM benchmark of that operating condition cluster is calculated. Based on the distribution of the KL divergence value, the dynamic anomaly detection threshold of that operating condition cluster is determined.

[0114] Specifically, a GMM model is trained for each working condition cluster divided by the working condition identification module as a health status benchmark. The probability density function of the GMM model is:

[0115] in, It is the probability density function of a multivariate Gaussian distribution.

[0116] Formula for calculating KL divergence:

[0117] in, For the true distribution, For reference distribution, The probability density function of a healthy Gaussian mixture model at point The value at that location.

[0118] The dynamic anomaly detection threshold is set at the 99.73% quantile of the KL divergence values ​​of all training samples in the corresponding working condition cluster.

[0119] Real-time monitoring module: Used to acquire real-time integrated circuit test data, determine the operating condition cluster to which the data belongs after data preprocessing; calculate the KL divergence value of the data relative to the GMM health status benchmark of the corresponding operating condition cluster; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, the test data is marked as abnormal.

[0120] Example 3 Based on Embodiment 1, Embodiment 3 of the present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method steps of Embodiment 1 by calling the program instructions.

[0121] Example 4 Based on Embodiment 1, Embodiment 4 of the present invention also provides a computer-readable storage medium storing computer instructions that cause the computer to perform the steps of the method described in Embodiment 1.

[0122] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. 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 of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method of predicting the stability of integrated circuit test equipment, characterised by, include: Step S1, Data Preprocessing: Obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features. Step S2, Operating Condition Identification: Based on the preprocessed historical test dataset, unsupervised learning is performed using the Gaussian mixture model clustering algorithm to divide the test data into multiple operating condition clusters; Step S3, State Distribution Modeling: Establish a Gaussian Mixture Model (GMM) baseline for the healthy state for each working condition cluster. For each working condition cluster, calculate the KL divergence value of all its training samples relative to the healthy state GMM baseline of that working condition cluster. Based on the distribution of the KL divergence value, determine the dynamic anomaly detection threshold for that working condition cluster. Step S4: Real-time monitoring: Acquire real-time integrated circuit test data, determine its operating condition cluster after data preprocessing; calculate its KL divergence value relative to the corresponding operating condition cluster's health status GMM benchmark; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, mark the test data as abnormal.

2. The method of claim 1, wherein the step of determining the stability of the integrated circuit test equipment is performed by a computer. The feature selection includes: Calculate the skewness, kurtosis, coefficient of variation, and interquartile range for each test feature. The KS test was used to evaluate the goodness of fit between the test features and 15 probability distribution models. The distribution model with the highest goodness of fit is selected as the optimal distribution model for that feature, and features that conform to or are close to a normal distribution are preferentially selected as the training set.

3. The method of claim 1, wherein the step of determining the stability of the integrated circuit test equipment is performed by a computer. Step S2 specifically includes: Initialization parameters: number of clusters, randomly initialize the weight of each Gaussian distribution , mean vector and covariance matrix ; Step E: Calculate each sample Belongs to the The posterior probability of a Gaussian distribution, i.e., the responsibility function ; M-step: Update the parameters of each Gaussian distribution according to the responsibility function; definition: For the first The number of effective samples from a Gaussian distribution; The weight update formula is as follows: The mean update formula is: The formula for updating the covariance matrix is: Iteration and Termination: Repeat the E-step and M-step until the parameter change is less than the preset threshold, the log-likelihood change is less than the preset value, or the preset maximum number of iterations is reached. Select the optimal parameters of the model by using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). Assign each sample to the cluster corresponding to the Gaussian distribution with the highest probability. Each cluster represents a working condition of the test machine. During the iteration process, log-likelihood estimation is used to evaluate the convergence of the parameters, and its calculation formula is as follows: Where t represents the number of iterations. X Represents the entire observation dataset, for , Sample size This represents the total number of components in the Gaussian mixture. The first component of the Gaussian mixture One portion, It is the weight of the k-th Gaussian component. For the first The probability density function of a Gaussian distribution. For the first The mean vector of Gaussian components, For the first k The covariance matrix of Gaussian components.

4. The method for predicting the stability of integrated circuit testing equipment according to claim 3, characterized in that, The optimal parameters of the model are selected using the Akaike Information Criterion (AIC) and the Bayesian Information Criterion (BIC). The formulas for calculating AIC and BIC are as follows: in, This represents the number of parameters in the model.

5. The method for predicting the stability of integrated circuit testing equipment according to claim 1, characterized in that, Step S3 specifically involves: For each working condition cluster divided in step S2, a GMM model is trained as a health status benchmark. The probability density function of the GMM model is: in, It is the probability density function of a multivariate Gaussian distribution.

6. The method for predicting the stability of integrated circuit testing equipment according to claim 1, characterized in that, The formula for calculating the KL divergence value is as follows: in, For the true distribution, For reference distribution, The probability density function of a healthy Gaussian mixture model at point The value at that location.

7. The method for predicting the stability of integrated circuit testing equipment according to claim 1, characterized in that, The dynamic anomaly detection threshold is set at the 99.73% quantile of the KL divergence values ​​of all training samples in the corresponding working condition cluster.

8. A system for predicting the stability of integrated circuit testing equipment, characterized in that, include: Data preprocessing module: used to obtain historical test datasets from the integrated circuit FT testing stage, perform data cleaning and feature selection on the historical test datasets, wherein feature selection includes distribution fitting for each test feature, selecting features that conform to or are close to a normal distribution; and standardizing the selected features. The working condition identification module is used to divide the test data into multiple working condition clusters based on the preprocessed historical test dataset using the Gaussian mixture model clustering algorithm for unsupervised learning. State distribution modeling module: used to establish a Gaussian mixture model benchmark for the healthy state for each working condition cluster. For each working condition cluster, calculate the KL divergence value of all its training samples relative to the healthy state GMM benchmark of that working condition cluster. Based on the distribution of the KL divergence value, determine the dynamic anomaly detection threshold of that working condition cluster. Real-time monitoring module: Used to acquire real-time integrated circuit test data, determine the operating condition cluster to which the data belongs after data preprocessing; calculate the KL divergence value of the data relative to the GMM health status benchmark of the corresponding operating condition cluster; if the KL divergence value exceeds the dynamic anomaly detection threshold of the operating condition cluster, the test data is marked as abnormal.

9. A processing device, characterized in that, The method includes at least one processor and at least one memory communicatively connected to the processor, wherein the memory stores program instructions executable by the processor, and the processor can execute the method as described in any one of claims 1 to 7 by invoking the program instructions.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions that cause the computer to perform the method as described in any one of claims 1 to 7.

Images