An integrated circuit intelligent testing method and system

Abstract

The application discloses an integrated circuit intelligent testing method and system, which firstly uses a positive predictor and a negative predictor to synchronously predict all integrated circuits needing to be tested, compares the prediction results of the two predictions, identifies the prediction result greater than a reliability threshold T stab and converts it into traditional testing, effectively improving the testing reliability. Meanwhile, the application performs weighted summation on the prediction results of the positive and negative predictors as the final prediction result, so that the final result prediction error is less than 1 / 2 of the single model prediction error, the prediction accuracy is significantly improved, and in addition, the variance of the final prediction error is less than 1 / 2 of the single model prediction error variance, which effectively improves the testing stability.

Description

Technical Field

[0001] This invention belongs to the field of integrated circuit testing technology, and more specifically, relates to an intelligent testing method and system for integrated circuits. Background Technology

[0002] Testing is a crucial step in integrated circuit manufacturing, ensuring the reliability of products delivered to customers. During integrated circuit production, every chip on the production line undergoes thorough testing, and only chips with all specifications within acceptable ranges are delivered to customers.

[0003] Traditional integrated circuit testing methods require expensive testing equipment, complex testing procedures, and lengthy testing times. With the development of machine learning technology, machine learning-based intelligent testing techniques have attracted attention. Machine learning-based intelligent testing is being used to replace traditional testing methods to reduce testing costs and time.

[0004] However, current technical personnel lack confidence in such testing technologies, which limits the application and development of intelligent testing technologies. On the one hand, current intelligent testing technologies cannot identify prediction results with large errors, which reduces the reliability of the test. Even if the prediction model has a good average prediction error, it may still produce large-error prediction results, which are unacceptable in actual testing. On the other hand, the current intelligent testing technology lacks accuracy, resulting in an excessively high probability of misclassification of samples near the specification boundary, with qualified products being classified as unqualified and unqualified products as qualified. At the same time, existing technologies lack stability. The prediction error of the prediction model fluctuates greatly across all samples, making the indicators characterizing the prediction model's performance, such as mean square error, root mean square error, and mean error, unrepresentative, preventing testers from effectively evaluating model performance. Therefore, there is an urgent need to propose an intelligent testing technology with sufficient reliability, testing accuracy, and testing stability. Summary of the Invention

[0005] The purpose of this invention is to overcome the shortcomings of the prior art and provide an intelligent testing method and system for integrated circuits to improve the reliability, accuracy and stability of testing.

[0006] To achieve the above-mentioned objectives, the intelligent testing method for integrated circuits of the present invention is characterized by comprising the following steps:

[0007] (1) First, use positive predictor and negative predictor to predict all integrated circuits that need to be tested simultaneously: input excitation signal to integrated circuit, collect response signal of integrated circuit and extract features and input them into positive predictor and negative predictor respectively for prediction, and obtain their respective prediction results, where the prediction results are the performance indicators of integrated circuit;

[0008] (2) Then, calculate the difference P between the prediction results of the positive predictor and the negative predictor for the i-th integrated circuit. diff_i ,i=1,2,…,n, where n is the total number of integrated circuits that need to be tested;

[0009] (3) Then, by comparing the difference P diff_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the result is positive, the prediction is considered reliable; otherwise, the prediction is considered unreliable.

[0010] (4) Finally, for prediction results judged as unreliable, the testing of the corresponding integrated circuits is changed to traditional testing; for prediction results judged as reliable, a weighted sum is performed to obtain the final prediction result P. final_i :

[0011] P final_i =1 / 2×P pos_i +1 / 2×P neg_i

[0012] Among them, P pos_i It is the prediction result of the positive predictor, P neg_i It is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable;

[0013] The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0014] Before using the positive predictor for prediction, its loss function during training is: loss pos =W a ×f1(P pre ,P lab )+W pnel_pos Among them, W s For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value.

[0015] The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0016] Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lab )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

[0017] The present invention relates to an intelligent testing method and system for integrated circuits, characterized in that it comprises: a feature extraction module, a prediction module, a comparison module, and a fusion module;

[0018] The feature extraction module is used to acquire the response signal of the integrated circuit and extract features when the input excitation signal is input to the integrated circuit, and then input the features into the prediction module;

[0019] The prediction module has built-in positive and negative predictors to predict the input features respectively, obtain their respective prediction results, and input them into the comparison module. The prediction results are the performance indicators of the integrated circuit.

[0020] The comparison module is used to calculate the positive predictor prediction result P for the i-th integrated circuit. pos_i The negative predictor's prediction result P neg_i The difference P diff_i Let i = 1, 2, ..., n, where n is the total number of integrated circuits to be tested. Then, the difference P is compared. difd_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the prediction is positive, the prediction result is considered reliable; otherwise, it is considered unreliable. For prediction results judged as unreliable, the testing of the corresponding integrated circuit is changed to traditional testing. For prediction results judged as reliable, the prediction result P of the positive predictor is... pos_i The prediction result P of the negative predictor neg_i The input is fed into the fusion module, where i is the serial number of the integrated circuit whose prediction result is reliable;

[0021] The fusion module affects the prediction result P pos_i P neg_iWe perform a weighted summation to obtain the final prediction result P. final_i :

[0022] P final_i =1 / 2×P pos_i +1 / 2×P neg_i

[0023] Among them, P pos_i It is the prediction result of the positive predictor, P neg_i It is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable;

[0024] The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0025] Before using the positive predictor for prediction, its loss function during training is: loss pos =W a ×f1(P pre ,P lab )+W pnel_pos Among them, W a For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value.

[0026] The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0027] Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lab )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

[0028] The objective of this invention is achieved as follows:

[0029] The present invention relates to an intelligent testing method and system for integrated circuits. First, a positive predictor and a negative predictor are used simultaneously to predict the performance of all integrated circuits requiring testing. By comparing the prediction results of the two predictions, circuits with reliability thresholds greater than T are identified. stab The prediction results are converted into traditional tests, effectively improving test reliability. Simultaneously, this invention performs a weighted summation of the prediction results from both positive and negative predictors to obtain the final prediction result, making the final prediction error less than half of the single-model prediction error, significantly improving prediction accuracy. Furthermore, the variance of the final prediction error is less than half of the variance of the single-model prediction error, also effectively improving test stability. Attached Figure Description

[0030] Figure 1 This is a flowchart of a specific implementation of the intelligent testing method for integrated circuits of the present invention;

[0031] Figure 2 It is a comparison chart of the error before and after fusing the prediction results of the positive and negative predictors, where (a) is the prediction error curve of the positive and negative predictors, and (b) is the prediction error curve of the result after fusion.

[0032] Figure 3 This is a schematic diagram of the training of the positive and negative predictor, where (a) is the training process of the positive and negative predictor, and (b) is a schematic diagram of the training principle of the positive and negative predictor.

[0033] Figure 4 This is a schematic diagram illustrating the principle of the integrated circuit intelligent testing system of the present invention during testing;

[0034] Figure 5 This is a schematic diagram illustrating the principle of a specific implementation of the intelligent integrated circuit testing system of the present invention. Detailed Implementation

[0035] The specific embodiments of the present invention will now be described with reference to the accompanying drawings to enable those skilled in the art to better understand the invention. It should be particularly noted that in the following description, detailed descriptions of known functions and designs that might obscure the main content of the invention will be omitted here.

[0036] Figure 1 This is a flowchart of a specific implementation of the intelligent testing method for integrated circuits of the present invention.

[0037] In this embodiment, as Figure 1 As shown, the intelligent testing method for integrated circuits of the present invention includes the following steps:

[0038] Step S1: Perform predictions on the integrated circuit using positive and negative predictors.

[0039] First, positive and negative predictors are used simultaneously to predict all integrated circuits that need to be tested: the excitation signal is input to the integrated circuit, the response signal of the integrated circuit is collected and the features are extracted and input into the positive and negative predictors respectively for prediction, and the prediction results are obtained. The prediction results are the performance indicators of the integrated circuit.

[0040] Step S2: Calculate the difference between the prediction results

[0041] Then, calculate the difference P between the prediction results of the positive predictor and the negative predictor for the i-th integrated circuit. diff_i ,i=1,2,…,n, where n is the total number of integrated circuits that need to be tested.

[0042] Step S3: Compare with a reliability threshold

[0043] Then, by comparing the difference P diff_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the prediction is true, the prediction is considered reliable; otherwise, the prediction is considered unreliable.

[0044] Step S4: Weighted summation of reliable prediction results

[0045] Finally, for predictions deemed unreliable, the testing of their corresponding integrated circuits is converted to traditional testing; for predictions deemed reliable, a weighted sum is performed to obtain the final prediction result P. final_i :

[0046] P final_i =1 / 2×P pos_i +1 / 2×P neg_i

[0047] Among them, P pos_i It is the prediction result of the positive predictor, P neg_i is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable.

[0048] The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0049] Before using the positive predictor for prediction, its loss function during training is: loss pos =Wa ×f1(P pre ,P lab )+W pnel_pos Among them, W a For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value.

[0050] The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0051] Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lab )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

[0052] Figure 2 This is a comparison chart of the errors before and after fusing the prediction results of the positive and negative predictors.

[0053] from Figure 2 (a) As we can see, this invention identifies large error prediction results by comparing the prediction results of positive and negative predictors; the rectangle represents the difference P. diff_i Greater than or equal to the reliability threshold T stab The prediction results are unreliable. Figure 2 (b) We can see that fusing the results of positive and negative predictors reduces the prediction error and the variance of the prediction error;

[0054] Figure 3 This is a schematic diagram of the training of the positive and negative predictor, where (a) is the training process of the positive and negative predictor and (b) is a schematic diagram of the training principle of the positive and negative predictor.

[0055] 1. Feature Extraction

[0056] according to Figure 3 (a) The process measures the responses of all integrated circuits under test, and after feature extraction, uses these responses as features for the training / test dataset of the positive and negative predictors. Preferably, principal component analysis is used as the feature extraction method.

[0057] 2. Dataset Construction

[0058] The known metrics of all integrated circuits are used as labels, each corresponding to a feature of a sample, forming a complete dataset. A portion of this dataset is used as test data to evaluate the intelligent system, while the remaining data is used as training / test data for the positive and negative predictors.

[0059] 3. Predictor Model Building

[0060] The predictor structure is preferably constructed using a deep neural network structure, which mainly includes a fully connected layer with the same number of neurons as the number of features as the input layer, multiple fully connected layers as hidden layers, and a fully connected layer with one neuron as the output layer.

[0061] 4. Loss Function Design

[0062] like Figure 3 (b) Design loss functions for the positive and negative predictors respectively. pos With loss neg The two loss functions mainly consist of a comparison function and a penalty term. Preferably, the mean square function or the root mean square function is used as the comparison function to measure the model's prediction result P. pre With label value P lab The difference between them. Positive model penalty term W pnel_pos The value of W is positively correlated with the number of samples in the prediction results where the predicted value is less than the label value. pnel_neg The value is positively correlated with the number of samples whose predicted value is greater than the label value in the prediction results.

[0063] 5. Predictor Training

[0064] like Figure 3 (b) The dataset is divided into training and testing sets, and the predictor is trained using an iterative optimization method. Preferably, as follows: Figure 3 (b) A training method using nested folding, rounds, and batches is employed. The data is divided into K folds. In each iteration, one fold is selected as the test data, and the rest are used as training data. This process is repeated K times until all folds in the dataset have been used as the test set. For each partition condition, a total of m rounds of training are performed. In each round of training, batch training is used, where all training data is fed into the model in batches. The model parameters are optimized sequentially in each batch. One round is considered complete when all data has been trained in batches.

[0065] Figure 4 This is a schematic diagram illustrating the principle of the integrated circuit intelligent testing system of the present invention.

[0066] In this embodiment, as Figure 4As shown, an input excitation signal is fed into the integrated circuit, and the integrated circuit outputs a response signal to the integrated circuit intelligent test system. The integrated circuit intelligent test system performs predictions to obtain the final prediction result. For unreliable prediction results, the test is converted to traditional testing.

[0067] Figure 5 This is a schematic diagram illustrating the principle of a specific implementation of the intelligent integrated circuit testing system of the present invention.

[0068] In this embodiment, as Figure 5 As shown, the integrated circuit intelligent testing method and system of the present invention includes a feature extraction module 1, a prediction module 2, a comparison module 3, and a fusion module 4.

[0069] Feature extraction module 1 is used to acquire the response signal of the integrated circuit and extract features when the input excitation signal is input to the integrated circuit, and input them to prediction module 2.

[0070] The prediction module 2 has built-in positive and negative predictors to predict the input features respectively, obtain their respective prediction results, and input them into the comparison module 3. The prediction results are the performance indicators of the integrated circuit.

[0071] Comparison module 3 is used to calculate the positive predictor prediction result P for the i-th integrated circuit. pos_i The negative predictor's prediction result P neg_i The difference P diff_i =P pos_i -P neg_i Let i = 1, 2, ..., n, where n is the total number of integrated circuits to be tested. Then, the difference P is compared. diff_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the prediction is positive, the prediction result is considered reliable; otherwise, it is considered unreliable. For prediction results judged as unreliable, the testing of the corresponding integrated circuit is changed to traditional testing. For prediction results judged as reliable, the prediction result P of the positive predictor is... pos_i The prediction result P of the negative predictor neg_i The input is fed into the fusion module 4, where i is the serial number of the integrated circuit whose prediction result is reliable.

[0072] Fusion module 4 on the prediction result P pos_i P neg_i We perform a weighted summation to obtain the final prediction result P. fina_i :

[0073] P final_i =1 / 2×P pos_i+1 / 2×P neg_i

[0074] Among them, P pos_i It is the prediction result of the positive predictor, P neg_i It is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable;

[0075] The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0076] Before using the positive predictor for prediction, its loss function during training is: loss pos =W a ×f1(P pre ,P lab )+W pnel_pos Among them, W a For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value.

[0077] The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit.

[0078] Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lan )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

[0079] Example

[0080] A test system was built using the intelligent test method proposed in this invention to test the common-mode rejection ratio of the high-precision analog operational amplifier integrated circuit ADA4927-1.

[0081] 1. Feature Extraction

[0082] For 1200 ADA4927-1s with known common-mode rejection ratios, the circuit response is acquired, and principal component analysis is used to process the circuit response. The first three features of the analysis results are retained as the input features of the predictor.

[0083] 2. Dataset Construction

[0084] Of the 1200 ADA4927-1 chips used to build the dataset for testing, 200 integrated circuits were allocated for system testing, and the remaining 1000 integrated circuits were used for model training / testing.

[0085] 3. Predictor Model Building

[0086] A predictor model was constructed using a deep neural network structure, specifically consisting of one fully connected layer as the input layer, five fully connected layers as hidden layers, and one fully connected layer as the output layer. The output of each layer was non-linearly mapped using the ReLU function.

[0087] 4. Loss Function Design

[0088] The loss function is designed, preferably using the following method to design the penalty term: In each batch of training, the number of samples whose predicted results are less than the label value is counted, and the remaining number of samples is subtracted as the penalty term W for the positive prediction model. pnel_pos Conversely, the number of samples in the batch whose predicted value is greater than the label value is subtracted from the number of remaining samples, which is used as the negative model penalty term W. pnel_neg .

[0089] 5. Predictor Model Training

[0090] The positive and negative predictor models are trained separately. Preferably, the 1000 sets of data used for training are divided into 5 folds, and each fold is trained for 200 rounds. In each round of training, the 1000 sets of data are divided into batches of 32 samples.

[0091] 6. Construction of Intelligent Testing System

[0092] To build a testing system, preferably, the reliability threshold T is set. stab Set the metric to be predicted as 3% of the mean common-mode suppression ratio. This means that if the difference between the predicted common-mode suppression ratios of the positive and negative predictors for the same sample is greater than 3% of the mean of all sample labels, the result is considered unreliable.

[0093] 7. System Testing

[0094] Using 200 divided samples, the present invention was analyzed and tested to verify that the prediction results of the present invention have high reliability, prediction accuracy and prediction stability.

[0095] Although the illustrative specific embodiments of the present invention have been described above to enable those skilled in the art to understand the invention, it should be understood that the invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the invention as defined and determined by the appended claims, and all inventions utilizing the concept of the present invention are protected.

Claims

1. A smart testing method for integrated circuits, characterized in that, Includes the following steps: (1) First, use positive predictor and negative predictor to predict all integrated circuits that need to be tested simultaneously: input excitation signal to integrated circuit, collect response signal of integrated circuit and extract features and input them into positive predictor and negative predictor respectively for prediction, and obtain their respective prediction results, where the prediction results are the performance indicators of integrated circuit; (2) Then, calculate the difference P between the prediction results of the positive predictor and the negative predictor for the i-th integrated circuit. diff_i ,i=1,2,…,n, where n is the total number of integrated circuits that need to be tested; (3) Then, by comparing the difference P diff_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the result is positive, the prediction is considered reliable; otherwise, the prediction is considered unreliable. (4) Finally, for prediction results judged as unreliable, the testing of the corresponding integrated circuits is changed to traditional testing; for prediction results judged as reliable, a weighted sum is performed to obtain the final prediction result P. final_i : P final_i =1 / 2×P pos_i +1 / 2×P neg_i Among them, P pos_i It is the prediction result of the positive predictor, P neg_i It is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable; The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit. Before using the positive predictor for prediction, its loss function during training is: loss pos =W a ×f1(P pre ,P lab )+W pnel_pos Among them, W a For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value. The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit. Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lab )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

2. The integrated circuit intelligent testing method according to claim 1, characterized in that, In each training batch, the number of samples whose predicted value is less than the label value is counted, and the remaining number of samples is subtracted as the penalty term W for the positive prediction model. pnel_pos Conversely, the number of samples in the batch whose predicted value is greater than the label value minus the number of remaining samples is used as the negative model penalty term W. pnel_neg .

3. The integrated circuit intelligent testing method according to claim 1, characterized in that, The performance metric of the integrated circuit is the common-mode rejection ratio.

4. A method and system for intelligent testing of integrated circuits, characterized in that, include: Feature extraction module, prediction module, comparison module, and fusion module; The feature extraction module is used to acquire the response signal of the integrated circuit and extract features when the input excitation signal is input to the integrated circuit, and then input the features into the prediction module; The prediction module has built-in positive and negative predictors to predict the input features respectively, obtain their respective prediction results, and input them into the comparison module. The prediction results are the performance indicators of the integrated circuit. The comparison module is used to calculate the positive predictor prediction result P for the i-th integrated circuit. pos_i The negative predictor's prediction result P neg_i The difference P diff_i Let i = 1, 2, ..., n, where n is the total number of integrated circuits to be tested. Then, the difference P is compared. diff_i Compared with the artificially set reliability threshold T stab The magnitude relationship between the positive and negative predictors is used to determine whether their predictions for the i-th integrated circuit are reliable. If the difference P... diff_i Less than the reliability threshold T stab If the prediction is positive, the prediction result is considered reliable; otherwise, it is considered unreliable. For prediction results judged as unreliable, the testing of the corresponding integrated circuit is changed to traditional testing. For prediction results judged as reliable, the prediction result P of the positive predictor is... pos_i The prediction result P of the negative predictor neg_i The input is fed into the fusion module, where i is the serial number of the integrated circuit whose prediction result is reliable; The fusion module affects the prediction result P pos_i P neg_i We perform a weighted summation to obtain the final prediction result P. final_i : P final_i =1 / 2×P pos_i +1 / 2×P neg_i Among them, P pos_i It is the prediction result of the positive predictor, P neg_i It is the prediction result of the negative predictor, and i is the serial number of the integrated circuit whose prediction result is reliable; The positive predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit. Before using the positive predictor for prediction, its loss function during training is: loss pos =W a ×f1(P pre ,P lab )+W pnel_pos Among them, W s For the set coefficient, P pre For the prediction result, i.e., the predicted value, P lab Here, f1 is the label value, and W is the comparison function. pnel_pos This is a penalty term, and its value is positively correlated with the number of samples in the training batch whose predicted value is less than the label value. The negative predictor is a regression prediction model based on machine learning. When using the trained positive predictor to predict the integrated circuit sample set, the prediction error of the predictor for more than 80% of the integrated circuits in the integrated circuit sample set is greater than or equal to 0. The prediction error is calculated as prediction result minus label value, where the label value is the true performance indicator of the integrated circuit. Before using the negative predictor for prediction, its loss function during training is: loss neg =W a ×f1(P pre ,P lab )+W pnel_neg Penalty item W pnel_neg The value is positively correlated with the number of samples in the training batch whose predicted value is greater than the label value.

Images