A chip security mechanism evaluation method and computer device

Abstract

The present invention discloses a chip safety mechanism evaluation method and computer equipment, the method comprising: obtaining test data of historical model chips of the same series; extracting characteristic variables of the test data of historical model chips, establishing a prediction model based on machine learning, and evaluating the effectiveness of chip safety mechanisms; obtaining test data of new model chips with the same data structure as sample data; inputting the test data of new model chips into the established prediction model for prediction, and outputting the prediction result as an indicator of the effectiveness of chip safety mechanisms. The present invention can realize rapid functional safety detection and evaluation for new model chips of the same series, and significantly improve the functional safety detection efficiency and stability of new model chips in small sample scenarios.

Description

Technical Field

[0001] The present invention belongs to the technical field of chip functional safety, and in particular relates to a chip safety mechanism evaluation method and computer equipment. Background Art

[0002] Modern MCU chips are not only required to have powerful processing capabilities, but also need to ensure that their functional safety indicators meet strict industry standards. Functional safety has become a key consideration in the design and application of MCU chips, especially in areas such as automotive electronics and industrial control that have extremely high safety requirements. MCU chips of the same series maintain a high degree of similarity in design concepts and infrastructure. Different models of MCU chips in the same series integrate similar functional safety mechanisms to ensure stability and security in various application scenarios.

[0003] However, the functional safety testing of MCU chips in the same series still relies on independent testing of each model, using the traditional single chip testing method, that is, a large number of sample tests are conducted on each model of MCU chips to obtain their safety mechanism indicators. The similarity and historical test data of MCU chips in the same series are not fully utilized, resulting in a time-consuming and costly testing process. Summary of the invention

[0004] The purpose of the present invention is to provide a chip safety mechanism evaluation method, computer equipment, computer-readable storage medium and computer program product, which can significantly improve the functional safety detection efficiency and stability of new model chips in small sample scenarios.

[0005] In order to achieve the above object, one aspect of the present invention provides a chip security mechanism evaluation method, comprising:

[0006] Obtain test data of historical chips of the same series;

[0007] Extract characteristic variables from test data of historical model chips and establish a prediction model based on machine learning to evaluate the effectiveness of chip security mechanisms;

[0008] Obtain test data of a new model chip with the same data structure as the sample data;

[0009] The test data of the new model chip is input into the established prediction model for prediction, and the prediction result is output as an indicator of the effectiveness of the chip security mechanism.

[0010] Preferably, the prediction model based on machine learning constitutes a multi-level model architecture, including a basic model layer, a model-based model layer, an uncertainty model and a meta-model layer;

[0011] The basic model layer uses a gradient boosting regression model to build a universal feature extractor and establishes a basic model based on the statistical characteristics of the test data;

[0012] The model-based model layer uses a gradient boosting regression model to build specific model models for different model chips and capture model-specific features;

[0013] The uncertainty model uses a random forest model to evaluate the confidence interval of the prediction results according to the degree of dispersion of the prediction results of the basic model and the model-based model, and quantify the uncertainty of the prediction results;

[0014] The metamodel layer uses a gradient boosting regression model to establish a metamodel based on an adaptive learning strategy, integrates the prediction results of the basic model and the model-based model, and adaptively adjusts the weight of each prediction result in the final prediction result.

[0015] Preferably, the statistical features include the central tendency, dispersion and distribution form of the test data points, the model-specific features include statistical features, chip model identification and working condition characteristics, and the characteristics of the meta-model include the prediction results and uncertainty of the basic model, the prediction results and uncertainty of the model-based model, working condition characteristics and model similarity characteristics.

[0016] Preferably, the establishing of a prediction model based on machine learning comprises:

[0017] Extract statistical features and model features of test data, and standardize the statistical features and model features;

[0018] The standardized statistical features and model features are respectively input into the basic model and the model-based model for prediction. The prediction results output by each model are used as the input of the uncertainty model to perform uncertainty estimation and output the prediction results and uncertainty of each model.

[0019] According to the uncertainty of the prediction results of each model, the model similarity characteristics and the working condition characteristics, the weight of the prediction results of each model is determined, and the prediction results of each model are fused according to the weight, and the fused prediction results are used as the target variable;

[0020] A meta-model based on adaptive learning strategy is established, and the meta-model is trained with target variables to obtain the final prediction model.

[0021] Preferably, the establishing of a prediction model based on machine learning further comprises:

[0022] According to the weight of the prediction results of each model, the features corresponding to the target variable are dynamically weighted and adaptively selected, and the optimal feature combination is selected as the input of the gradient boosting regression algorithm of the meta-model.

[0023] Preferably, the indicators of the central trend include the mean and the median, the indicators of the degree of dispersion include the standard deviation, the maximum value, the minimum value and the quartiles, and the indicators of the distribution form include the skewness and the kurtosis; the working condition characteristics include the maximum voltage, the minimum voltage, the maximum temperature, the minimum temperature and the clock frequency; the uncertainty of the model prediction results based on the model includes the standard deviation of the specific model, the prediction lower bound of the specific model and the prediction upper bound of the specific model.

[0024] Preferably, the data format of the test data includes: chip identification, model identification, operating voltage range, operating temperature range and multiple test data points.

[0025] Another aspect of the present invention provides a computer device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above method.

[0026] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of the above method when executed by a processor.

[0027] Another aspect of the present invention provides a computer program product, comprising a computer program, which implements the steps of the above method when executed by a processor.

[0028] The chip safety mechanism evaluation method, computer device, computer-readable storage medium and computer program product according to the above aspects of the present invention can significantly improve the functional safety detection efficiency and stability of new chip models in small sample scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0029] In order to more clearly illustrate the technical solution of the present invention, the following briefly introduces the drawings used in the description of the embodiments of the present invention. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work:

[0030] Figure 1 is a flow chart of a chip security mechanism evaluation method according to an embodiment of the present invention;

[0031] Figure 2 is a schematic diagram of a multi-level model architecture of an embodiment of the present invention;

[0032] Figure 3 is a schematic diagram of a modeling process of a prediction model based on machine learning according to an embodiment of the present invention;

[0033] Figure 4is a comparison diagram of the prediction results and the real data of one embodiment of the present invention;

[0034] Figure 5 It is a structural diagram of a computer device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0036] An embodiment of the present invention provides a chip safety mechanism evaluation method, which is a chip functional safety mechanism evaluation method based on a machine learning prediction model. By comprehensively analyzing the similarities of MCUs in the same series, the functional safety detection efficiency of new chip models is improved, and safety indicators are predicted under small sample conditions, thereby shortening product verification time. Figure 1 As shown, the method of the embodiment of the present invention mainly includes steps S1 to S4.

[0037] Step S1: Obtain a large amount of test data of historical model chips of the same series, and prepare a large amount of historical test data of historical model chips of the same series for modeling in the subsequent step S2.

[0038] First, input the test results of the sample, taking the clock source failure (ClockSource Failures) caused by electromagnetic fault injection test as an example.

[0039] The formats of input data include:

[0040] ①Chip ID (chip unique identification)

[0041] ②Model ID (unique identifier of chip model)

[0042] ③Working voltage range

[0043] ④Operating temperature range

[0044] ⑤ Multiple test data points (under repeatable experimental conditions (such as 1000 electromagnetic fault injection tests, repeated 20 times), statistically analyze the probability of clock source failure caused by electromagnetic fault injection tests (obtained failure probability of a certain safety mechanism of the chip).

[0045] Use data with this structure as sample data input to prepare for subsequent modeling.

[0046] Step S2: Extract the feature variables ①-⑤ in step S1, establish a machine learning model, and obtain a prediction model that can be used to accurately evaluate the effectiveness of the chip safety mechanism. The prediction model can predict the clock source failure probability of a new model chip under the same electromagnetic fault injection test in a small sample case, thereby providing a stable and reliable prediction result.

[0047] This prediction model is based on supervised learning. It performs data preprocessing and feature extraction on the input data, and then inputs it into a multi-layer regression model for training to extract the potential patterns and relationships between the features.

[0048] It is worth noting that the data structure ⑤ described in step S1: In multiple repetitive experiments, by repeatedly testing a specific MCU chip in an electromagnetic fault injection test, multiple clock source failure probabilities are collected as data points. During the data preprocessing process, statistical features are extracted from these discrete data points for data processing, and a representative and stable clock source failure event probability index under the electromagnetic fault injection test is obtained through comprehensive calculation. This stable value, the "true value", not only serves as the target variable of the supervised learning model, but also provides a reliable benchmark reference for model performance evaluation. This method is used to transform scattered experimental data into highly generalizable quantitative indicators, so as to achieve accurate prediction and evaluation of the effectiveness of chip safety mechanisms.

[0049] Step S3: Get the test data of the new model chip, the number of tests and the number of repetitions are much less than the number of tests in the previous step S1, that is, small sample test data. Input the clock source failure probability data obtained in the electromagnetic fault injection test of the new model chip with the same data structure as in step S1 into the prediction model established in step S2.

[0050] Step S4: Input the new model chip test data of step S3 into the prediction model established in step S2 for prediction, and output the prediction result.

[0051] This prediction result is the prediction model established in step S2 with the test data of the new chip in step S3. The prediction model outputs the probability of the clock source failure event of the new chip in the electromagnetic fault injection test, or the numerical index of the safety mechanism, which is a representative and stable index. That is, it accurately predicts the effectiveness of the chip safety mechanism.

[0052] The prediction model based on machine learning established in step S2 includes four sub-models, forming a multi-level model architecture. Figure 2 The multi-level model architecture of this prediction model is shown, including:

[0053] (1) Basic model layer: A general feature extractor is constructed using the Gradient Boosting Regressor model, and modeling is performed based on the statistical characteristics of chip test data. The output is the prediction result based on the statistical characteristics of the test data.

[0054] (2) Model-based model layer: Use the gradient boosting regression model. Build model-specific models for different chip models to capture model-specific features. Output the prediction results based on the model-specific model for the test set to be tested.

[0055] (3) Uncertainty model: Use the random forest model. The uncertainty model evaluates the confidence interval of the prediction results based on the degree of dispersion of the prediction results of different models, and actually provides the uncertainty quantification of the prediction results of the basic model and the model-based model.

[0056] (4) Metamodel layer: Use the gradient boosting regression model. Based on the adaptive learning strategy, the prediction results of the basic model and the model-based model are integrated, and the weight of each prediction result in the final prediction result is adaptively adjusted. The final prediction result is output.

[0057] The purpose of building this multi-level model architecture is that the feature engineering of each sub-model will form a hierarchical feature engineering system. The main advantage is to achieve functional complementarity: the base model captures common patterns, the model-based model handles model specificity, and the meta-model integrates information from multiple dimensions. This hierarchical design can not only identify common failure modes, but also accurately model the characteristics of specific model chips, and intelligently fuse the prediction results of each layer through the meta-model layer to improve the prediction performance.

[0058] Table 1 below shows a hierarchical feature engineering system established by feature engineering of each sub-model based on the multi-level model architecture.

[0059] Table 1 Hierarchical feature engineering system

[0060]

[0061] (1) Basic model feature engineering: The basic model uses a pure statistical feature method to process the original data points to extract the following statistical features. The core statistical features include three aspects: central tendency, dispersion, and distribution shape. The central tendency indicators mainly include the mean, which reflects the average level of the data; the median, which reflects the central position of the data. The dispersion indicators include the standard deviation (std), which is used to measure the volatility of the data; the minimum value (min) and the maximum value (max), which represent the lower and upper limits of the data respectively; and the quartiles (q25, q75) reflect the distribution characteristics of the data. In terms of distribution shape, skew is used to measure the symmetry of the data distribution, while kurtosis reflects the sharpness of the distribution.

[0062] In terms of feature processing, StandardScaler is first used to standardize features to ensure that each feature has the same scale. Missing values ​​and outliers need to be properly processed to improve data quality. At the same time, the features should remain independent to avoid introducing model-related information to ensure the stability and versatility of the model.

[0063] (2) Model-based model feature engineering: By introducing features related to chip models and working conditions, the expressive power of basic statistical features is further enhanced. First, in the statistical feature layer, all statistical features of the basic model are inherited. Then, a model feature layer is added, which includes the chip model identifier (chip_version) and features related to working conditions, such as maximum voltage (voltage_max), minimum voltage (voltage_min), maximum temperature (temp_max), minimum temperature (temp_min), clock frequency, etc., but is not limited to this, and other working conditions may also be included. Moreover, in an embodiment of the present invention, model-specific features are not limited to statistical features, chip model identifiers, and working condition features, but may also include other features, which are just examples here. In terms of feature integration, the statistical features and model-specific features are spliced ​​in the axis=1 direction through the pd.concat method to ensure the integrity and independence of the features. Finally, modeling is performed separately for different models, and model-specific information is fully utilized to optimize model performance.

[0064] (3) Metamodel feature engineering: The metamodel uses the most complex feature engineering strategy to integrate multiple levels of prediction results and features to improve prediction performance. At the feature level, the metamodel integrates the prediction results (base_prediction) and uncertainty (base_std: standard deviation of the base model) of the base model, as well as the prediction output (version_{version}_pred) and uncertainty (version_{version}_std: standard deviation of a specific model; version_{version}_lower: prediction lower bound of a specific model; version_{version}_upper: prediction upper bound of a specific model) of each model model. It also introduces model identification features and generates corresponding identification columns for each model through one-hot encoding. At the same time, considering the working conditions, the model includes the minimum and maximum value features of temperature and voltage, and calculates the similarity with each known model through similarity calculation. By establishing a quantitative indicator of chip model similarity, it is possible to realize the correlation analysis of cross-model data features, and combined with statistical feature extraction methods, a complete feature expression framework is constructed.

[0065] To handle missing values, numerical features are filled with mean values, missing model identifiers are filled with 0, and the similarity of unknown models is set to the default value. All features are standardized, StandardScaler is used to ensure data consistency, and the order of feature columns is saved to facilitate consistency verification during prediction. In addition, the model ensures the consistency of feature lists during training and prediction, and can handle new and unknown models.

[0066] The method of the embodiment of the present invention combines the above-mentioned multi-level model architecture and feature engineering system to establish a prediction model based on machine learning. Figure 3 The comprehensive modeling process of the predictive model from data input to final prediction is demonstrated.

[0067] The entire process can be divided into three main stages:

[0068] Step S11, feature engineering of basic model and model-based model: input is the data structure in step S1. Feature extraction is performed through two main paths: basic model and model-based model.

[0069] (1) Basic characteristics: Extract key information from the original data points (the probability of the chip causing clock source failure in the electromagnetic fault injection test) through statistical methods (such as mean, standard deviation, skewness, kurtosis, etc.) to capture the overall distribution characteristics of the data;

[0070] (2) Model-based features: Combine the chip model with the specific operating environment, temperature, and voltage range to provide richer input for the model.

[0071] Both features need to be standardized and adjusted to the same scale as the output of this stage, and will subsequently be used as the input of the basic model and the model-based model in the prediction model integrated in step S12.

[0072] Step S12, integrated prediction system: consists of a basic model, an uncertainty model and a model-based model, and the input is the standardized features of the final output in step S11.

[0073] First, the output of step S11 is input into the basic model and the model-based model for prediction. The basic model provides the first prediction of the statistical characteristics of the data points of the overall data set, while the model-based model customizes the exclusive prediction model for different chip models. For the input data, different models are selected to make predictions respectively to obtain the prediction results of the chip based on different model models (used in the subsequent meta-model to adjust the proportion of the prediction results of the basic model and each model model to obtain the final prediction result).

[0074] Secondly, the prediction results output by each model are used as the input of the uncertainty model to estimate the uncertainty. The uncertainty model (random forest regressor) estimates the uncertainty of the prediction results of each model and quantifies the credibility of the prediction results of each model. The prediction results of the basic model and the prediction results based on the model model and the uncertainty of the corresponding prediction results are output, for example, in the form of [basic model prediction result X, confidence interval of the prediction result X].

[0075] Finally, the prediction results of the basic model and the model-based model are integrated. The integrated prediction results will serve as the output of this stage and as the input of the subsequent meta-model.

[0076] Step S13, meta-model based on adaptive learning strategy: fuse the multi-model prediction results outputted from step S12, and train the meta-model based on the fused prediction results (target variables) and optimized features. The purpose of training is to let the meta-model learn to automatically adjust the weight of each model prediction result in the final prediction result by extracting the model similarity, working conditions, and uncertainty of each model prediction result based on the input test data of the new model chip. This stage can be divided into two parts: the adaptive learning strategy core and the gradient boosting regression algorithm part.

[0077] (1) The core of the adaptive learning strategy. Its function is to optimize features and provide target variables for the subsequent (2) training process of the gradient boosting regression algorithm.

[0078] Providing target variables: The prediction results output in step S12 are input into the meta-model. The model does not simply average or directly combine the prediction results of each model, but adjusts the weight of the prediction results of each model through model similarity analysis (model similarity feature in meta-model feature engineering), model confidence assessment (uncertainty) and working environment adaptability assessment (working condition feature).

[0079] The weight of each model's prediction result = f (predicted value, uncertainty, model similarity, working environment)

[0080] Feature optimization: It is divided into dynamic adjustment of the weight of each feature and adaptive feature selection (the target variable is the output variable of the model, and the feature is the input variable of the model. The relationship is: the target variable is a function of the feature). When the target variable is the result of adjusting the weights of the prediction results of each model, the feature corresponding to the target variable should also be dynamically weighted. Then adaptive feature selection is performed. The role of adaptive feature selection is to select the most relevant and informative features from the original feature set, reduce the influence of irrelevant or redundant features, improve the generalization ability of the model and reduce the computational cost. This adaptive feature selection ensures that the model can always use the most representative and discriminative features for prediction, thereby improving the overall prediction accuracy. In fact, the optimal feature combination (the ones that are most helpful to the current prediction task and the features with large weights) is selected as the features to be input into the regression training in (2) the gradient boosting regression algorithm. At this point, feature optimization is completed.

[0081] (2) The meta-model’s gradient boosting regression algorithm first normalizes the optimal features outputted in the core of the adaptive learning strategy (1) and then inputs them into the gradient boosting regressor. The training is performed using the target variables provided in the core of the adaptive learning strategy (1). During the training process, the meta-model can adaptively adjust its learning strategy according to the changes in the input features. Through continuous iterative training, the meta-model can learn the optimal parameter settings and structural optimization strategies, thereby enhancing the model’s ability to adapt to new data in subsequent prediction tasks.

[0082] This machine learning-based prediction model can comprehensively utilize the prediction results of multiple models, consider the similarity between models, and integrate the working environment information, thereby improving the accuracy and stability of the prediction. The final output prediction results not only reflect the overall trend of the basic model, but also integrate the fine-grained insights based on the model model, representing the most accurate and comprehensive estimate of the probability of clock source errors caused by the model chip under electromagnetic fault injection testing. At the same time, the meta-model has a strong adaptive ability, can effectively handle new models of chips, realize knowledge transfer through similarity features, and dynamically adjust the weights of different models to ensure that efficient prediction performance can be maintained when facing different models and unknown situations.

[0083] Figure 4 The method of the embodiment of the present invention is used to predict the probability of a clock source failure under electromagnetic fault injection testing. The dotted line represents the true value, and the dot represents the final predicted value. By comparing the graphs of the true value and the predicted value, it can be observed that the predicted curve coincides well with the real data curve, which not only verifies the effectiveness of the model, but also highlights its strong potential for accurate modeling and prediction in complex nonlinear systems, fully proving the scientific value and practical application prospects of the method of the embodiment of the present invention in the field of chip reliability assessment.

[0084] In summary, the chip security mechanism evaluation method of the embodiment of the present invention has the following beneficial effects:

[0085] (1) Rapid functional safety testing and evaluation for new chip models in the same MCU series

[0086] The present invention develops an efficient data analysis model by making full use of the similarities in architecture and function between MCUs of the same series, as well as the accumulation of existing test data. The model can accurately predict the functional safety indicators of new MCU models based on the test data of existing models under small sample test conditions. By combining similarity analysis with data prediction technology, a rapid evaluation of functional safety testing of new chip models is achieved, greatly improving the detection efficiency.

[0087] (2) Reduce sample testing requirements and ensure the stability of test results

[0088] The present invention effectively solves the problem of poor test result stability in traditional detection methods under small sample testing conditions. Existing detection methods often lack a quantitative evaluation mechanism for the reliability of prediction results, making it difficult to provide credibility support for the detection results. The present invention ensures the reliability of the results in design. By establishing a scientific prediction model and utilizing existing test data and statistical analysis methods, the sample frequency required for physical testing is greatly reduced. This not only significantly reduces the testing cost, but also shortens the verification cycle of new products. On the premise of ensuring test reliability, redundant testing is reduced, the flexibility and efficiency of the test process are improved, and innovation and optimization of MCU functional safety testing are promoted.

[0089] An embodiment of the present invention further provides a computer device, which may be a server, and its internal structure diagram may be as shown in FIG. Figure 5 As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system, a computer program and a database. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The database of the computer device is used to store operating parameter data of each framework. The network interface of the computer device is used to communicate with an external terminal through a network connection. When the computer program is executed by the processor, the steps of the method of the embodiment of the present invention are implemented.

[0090] Those skilled in the art will understand that Figure 5 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0091] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the steps of the method of the embodiment of the present invention are implemented.

[0092] An embodiment of the present invention further provides a computer program product, including a computer program, which implements the steps of the method of the embodiment of the present invention when executed by a processor.

[0093] The above description is only by way of illustration of certain exemplary embodiments of the present invention. It is undoubted that those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the above drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. A chip security mechanism evaluation method, characterized in that: include: Obtain test data of historical model chips of the same series as sample data; Extract characteristic variables from test data of historical model chips and establish a prediction model based on machine learning to evaluate the effectiveness of chip security mechanisms; Obtain test data of a new model chip with the same data structure as the sample data; Input the test data of the new chip model into the established prediction model for prediction, and output the prediction result as an indicator of the effectiveness of the chip security mechanism; The prediction model based on machine learning constitutes a multi-level model architecture, including a basic model, a model-based model, an uncertainty model and a meta-model; The establishment of a prediction model based on machine learning includes: Extract statistical features and model-specific features of the test data, and perform standardization on the statistical features and model-specific features; The standardized statistical features and model-specific features are respectively input into the basic model and the model-based model for prediction. The prediction results output by each model are used as the input of the uncertainty model to perform uncertainty estimation and output the prediction results and uncertainty of each model. According to the uncertainty of the prediction results of each model, the model similarity characteristics and the working condition characteristics, the weight of the prediction results of each model is determined, and the prediction results of each model are fused according to the weight, and the fused prediction results are used as the target variable; A meta-model based on adaptive learning strategy is established, and the meta-model is trained with target variables to obtain the final prediction model.

2. The method according to claim 1, characterized in that Use the gradient boosting regression model to build a general feature extractor and establish a basic model based on the statistical characteristics of the test data; Use the gradient boosting regression model to build model-based models for different chip models to capture model-specific features; The uncertainty model was constructed using the random forest model. The confidence interval of the prediction results was evaluated according to the degree of dispersion of the prediction results of the basic model and the model-based model, and the uncertainty of the prediction results was quantified. Using the gradient boosting regression model, a meta-model based on an adaptive learning strategy is established to fuse the prediction results of the basic model and the model-based model, and adaptively adjust the weight of each prediction result in the final prediction result.

3. The method according to claim 2, characterized in that The statistical features include the central tendency, dispersion and distribution form of the test data points, the model-specific features include statistical features, chip model identification and working condition features, and the features of the meta-model include the prediction results and uncertainty of the basic model, the prediction results and uncertainty of the model-based model, working condition features and model similarity features.

4. The method according to claim 3, characterized in that The establishment of a prediction model based on machine learning also includes: According to the weight of the prediction results of each model, the features corresponding to the target variable are dynamically weighted and adaptively selected, and the optimal feature combination is selected as the input of the gradient boosting regression algorithm of the meta-model.

5. The method according to claim 3 or 4, characterized in that The indicators of central tendency include mean and median, the indicators of dispersion include standard deviation, maximum value, minimum value and quartiles, and the indicators of distribution form include skewness and kurtosis; the working condition characteristics include maximum voltage, minimum voltage, maximum temperature, minimum temperature and clock frequency; the uncertainty of the model prediction results based on the model includes the standard deviation of the specific model, the prediction lower bound of the specific model and the prediction upper bound of the specific model.

6. The method according to any one of claims 1 to 4, characterized in that The data format of the test data includes: chip identification, model identification, operating voltage range, operating temperature range, and multiple test data points.

7. A computer device comprising a memory, a processor and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A computer program product, comprising a computer program, characterized in that When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images