Failure area evaluation method applied to high Sigma scene in electronic engineering

Abstract

The invention belongs to the field of methods for evaluating rare events in electronic engineering, and particularly relates to a failure area evaluation method applied to a high Sigma scene in the electronic engineering. A failure area evaluation method applied to a high Sigma scene in electronic engineering comprises the following steps: establishing an initial sample, establishing a prediction model, screening a high-risk sample, simulating the high-risk sample and calculating an offset delta d, and determining an approximate failure area sample, a secondary approximate failure area sample and a common area sample. Spearman rank coefficients are calculated for the samples of the three areas through a formula, different weights are given to the Spearman rank coefficients, and a final weighted rank correlation evaluation result is calculated. According to the evaluation method, through a sequencing function of a weighted rank correlation evaluation substitution model, the accuracy of sequencing representation of a high Sigma failure area is remarkably improved; and unnecessary simulation times are greatly reduced, and computing resources are saved.

Description

Technical Field

[0001] The invention belongs to the field of methods for evaluating rare events in electronic engineering, and in particular relates to a failure area evaluation method applied to high Sigma scenarios in electronic engineering. Background Art

[0002] Integrated circuit (IC) component production in high Sigma scenarios can only tolerate a few defects in hundreds of millions or billions of instances. This is because these components are usually replicated in large arrays, so producing a properly functioning product requires a large number of replicated components to work correctly. Once fluctuations occur in the manufacturing process, this can lead to performance deviations or even failures. Therefore, analyzing IC yield has become a critical task.

[0003] The traditional Monte Carlo (MC) method is a common technique for evaluating IC yield. It collects a large amount of circuit data by simulating deviations in the circuit process, then simulates all circuits to determine the number of failed circuits, and finally obtains the circuit yield calculation. Its application efficiency is low in high Sigma scenarios because a large number of simulations are required to capture rare failure events, resulting in huge computing resources and time costs. To solve this problem, the industry has proposed a variety of accelerated Monte Carlo methods (such as importance sampling and model substitution methods), as well as high Sigma Monte Carlo (HSMC) methods. Although these methods have made some progress in reducing the number of samples and improving simulation efficiency, existing methods often ignore the importance of alternative models in the sorting of failure areas, resulting in inaccurate capture of failure events.

[0004] These methods are analyzed in detail as follows: 1. Traditional Monte Carlo method: Although the results are reliable and applicable to circuit designs of different dimensions, the number of samples required is extremely large in high Sigma scenarios, resulting in huge computing time and resource consumption; 2. Accelerated Monte Carlo method: Although the importance sampling method can reduce the number of simulation samples, it performs poorly when dealing with large-scale variables and can easily lead to failure probability estimation errors; 3. High Sigma Monte Carlo method (HSMC): HSMC reduces the number of simulations by giving priority to failed samples, but relies on global error minimization, ignores the sorting accuracy of the failure area, and cannot accurately capture rare events; 4. Maximum and minimum rank method: Although it improves the sorting accuracy of the model, it lacks flexibility in the allocation of sample weights, and it is difficult to accurately distinguish samples from failure areas and ordinary areas.

[0005] Weighted Rank is a ranking method that considers the weights of different factors or features in a ranking or scoring system. Unlike simple ranking, weighted rank allows certain factors to have a greater impact on the final ranking, while other factors have a smaller impact. The weighted rank algorithm is applied to the evaluation method to rank the different influencing factors in the rare event area according to their weights, and then evaluate the probability of occurrence of various rare events. The applicant believes that this can obviously overcome the limitations of the prior art. Summary of the invention

[0006] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a failure area assessment method applied to high Sigma scenarios in electronic engineering. This assessment method evaluates the sorting deviation of alternative models in rare event areas through a weighted rank correlation algorithm, thereby improving the accuracy of failure sample capture.

[0007] In order to solve the above technical problems, the present invention adopts the following technical solution: a failure area assessment method applied to high Sigma scenarios in electronic engineering, comprising the following steps:

[0008] Step 1: Create the initial sample Sam 初始 :Sam from the total sample set 总 Randomly select an initial sample Sam from 初始 , initial sample Sam 初始 The number of Qua 初始 Satisfying formula (1): Qua 初始 =max(Qua 总 / 1000,1000) (1), where Qua is the total sample size and the initial sample size Qua is 初始 is the total sample size Qua 总 One thousandth of the initial sample Sam, and no less than 1000. This quantity requirement ensures that 初始 Covering the total sample set Sam 总 The basic characteristics of , and can control the amount of calculation.

[0009] Step 2: Establish a prediction model: select the initial sample Sam 初始 Perform simulation and obtain the simulation result YS 初始 , using the simulation results YS 初始 Train machine learning equivalent surrogate models to generate predictive models.

[0010] Preferably, the simulation tool adopts a Spice circuit simulation tool.

[0011] Preferably, alternative models such as FFX algorithm, linear regression, Lasso regression, LightGBM, etc. can be used to establish a prediction model.

[0012] Step 3: Screening high-risk samples Sam 风险 :Use the prediction model to predict the remaining samples Sam 剩余 Make a prediction and get the prediction result YP 剩余 , where the remaining sample Sam 剩余 Sam is the total sample set 总 Remove the initial sample Sam 初始 The following sample set, the remaining sample Sam 剩余 The predicted results of YP 剩余 Arrange from large to small, and select high-risk samples Sam from large to small 风险 , high-risk sample Sam 风险 The number of Qua 风险 Satisfying formula (2): Qua 风险 =max(Qua 总 / 2000,500) (2), where Qua 总 is the total number of samples, high-risk sample Sam 风险 The number of Qua 风险 is the total sample size Qua 总 0.05% and no less than 500%, high-risk sample Sam 风险 The predicted value is YP 风险 .

[0013] Using the initial sample Sam 初始 Train the prediction model and then use it to predict the large number of remaining samples Sam 剩余 Make predictions, rank the prediction results, and select a certain number of remaining samples Sam with high rankings 剩余 As a high-risk sample Sam 风险 Further analysis and calculation are performed, and the remaining samples Sam ranked at the bottom 剩余 The failure rate is very low, so it is not studied.

[0014] Step 4: Simulate high-risk sample Sam 风险 :For high-risk sample Sam 风险 Perform simulation, obtain simulation results, and arrange the simulation results from large to small to obtain simulation results YS 风险 .

[0015] Step 5: Calculate the numerical offset Δd and the average numerical offset Δd 平均 : Arrange the simulation results YS from large to small 风险 Calculate the difference between two adjacent prediction values ​​one by one, that is, Δd i =(YS 风险 ) i -(YS 风险 )i+1 (3), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 Each natural number of -1; and the entire risk sample Sam 风险 The average numerical deviation of

[0016] Step 6: Determine the sample Sam close to the failure area 接近 and the next closest failure area sample Sam 次接近 :Set a simulation value threshold Thr 实际 , then the sample Sam close to the failure area 接近 Threshold Thr 接近 =Thr 实际 -3Δd 平均 (5), the sample Sam close to the failure area 次接近 Threshold Thr 次接近 =Thr 实际 -6Δd 平均 (6).

[0017] The simulation value is greater than or equal to Thr 接近 The sample is taken as the sample close to the failure area Sam 接近 , and the simulation value is greater than or equal to the sample threshold Thr of the next close to failure area 次接近 And it is smaller than the sample threshold Thr close to the failure area 接近 The sample of the next closest failure area is Sam 次接近 , then the sample Sam close to the failure area 接近 The number of 接近 , the sample Sam close to the failure area 次接近 The number of 次接近 .

[0018] Step 7: Determine the common area sample Sam 普通 :High-risk sample Sam 风险 Remove the sample Sam close to the failure area 接近 and the next closest failure area sample Sam 次接近 The following samples are samples from the normal area Sam 普通 , then the normal area sample Sam 普通 The number of Qua 普通 Satisfying formula (9), Qua 普通 =Qua 风险 -Qua 接近 -Qua 次接近 (7).

[0019] Step 8: Calculate the ranking offset Δrank of the prediction results: Compare the high-risk samples Sam风险 Each sample data simulation result YS 风险 And the predicted result YP 风险 The ranking difference is the ranking offset Δrank, which satisfies formula (8):

[0020] Δrank i =|rank(YS 风险 ) i -rank(YP 风险 ) i | (8), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 For each natural number, rank(YS 风险 ) is the simulation result YS 风险 The ranking number, rank(YP 风险 ) is the prediction result YP 风险 Rank number;

[0021] Step 9: Calculate the sample Sam close to the failure area 接近 The Spearman rank coefficient ρ1: as shown in formula (9), that is, Among them, the Δrank in the formula i 'The values ​​are as follows:

[0022] Among them, Thr1 and Thr2 are the artificially set first-level threshold Thr1 of samples close to the failure area and the second-level threshold Thr2 of samples close to the failure area.

[0023] When Δd≤Thr1, the sample ranking is considered to be very accurate and close to the failure area, so the relative ranking offset Δrank i ' is set to 0, indicating that these samples occupy the most important position in the evaluation and do not need to be sorted in the evaluation; and Thr1<Δd 平均 When ≤Thr2, the ranking error is small, and the ranking relative offset Δrank' is assigned to Δrank / 2, which means that these samples perform well in the failure area, but their importance is relatively reduced; when Δd 平均 >Thr2, Δrank′ is assigned Δrank, which means that the ranking deviation of the sample is large, but it is still within the extended range of the failure area, so its ranking offset is directly assigned.

[0024] Step 10: Calculate the next closest failure area sample Sam 次接近 The Spearman rank coefficient ρ2 and the common area sample Sam 普通 The Spearman rank coefficient ρ3: Then ρ2 is as in formula (10), that is, ρ3 is as in formula (11), that is, All the data in these two formulas are known or can be calculated, and ρ2 and ρ3 can be obtained by substituting them into the formulas.

[0025] Step 11: Comprehensive weighted sum ρ 总 :Define the weight of samples close to failure area as α1, the weight of samples next close to failure area as α2, and the weight of samples in common area as α3, then ρ 总 The calculation formula is as shown in formula (12), that is, ρ 总 =α1·ρ1+α2·ρ2+α3·ρ3 (12), where α3=1-α1-α2, and α1>α2>α3.

[0026] In the final evaluation, we weighted the samples in different areas and comprehensively considered the samples close to the failure area Sam. 接近 , Secondary failure area Sam 次接近 and general area sample Sam 普通 The contribution of each area is given different weights, among which the area close to the failure area Sam 接近 and the next closest failure area Sam 次接近 The samples in the normal area will be given greater weight, while the samples in the normal area Sam 普通 The weight of is usually set to a smaller value to reduce its impact on the overall evaluation results.

[0027] Preferably, we set α1 to 0.6-0.7 to ensure that the failure area sample Sam 接近 The dominant role in the overall evaluation; and when α2 is set to 0.2-0.3, the second closest area sample Sam 次接近 Although not as good as the sample Sam close to the failure area 接近 It is critical, but it is still in a relatively important position; α3 is usually set to a small value, such as 0.1-0.2, indicating that the sample Sam in the ordinary area 普通 The impact on the overall assessment is minimal.

[0028] Preferred, high-risk sample Sam 风险 Divide into n regions according to the situation, calculate the Spearman rank coefficient of each region, define the weight of each region, and calculate the final weighted rank correlation evaluation result. The calculation is as shown in formula (13): Among them, Pfina1 is the final weighted rank correlation evaluation result, ρi is the Spearman rank correlation coefficient of the i-th region sample, α i is the weight of the ith region and satisfies

[0029] More preferably, according to the actual situation, the high-risk sample Sam 风险It is divided into four areas according to the situation, namely, the area close to failure, the area close to failure, the extended area and the ordinary area, with weights of α1, α2, α3 and α4 respectively, and Spearman rank coefficients of ρ1, ρ2, ρ3 and ρ4 respectively. The final weighted rank result ρ final =α1·ρ1+α2·ρ2+α3·ρ3+α4·ρ3 (14), where α1=0.5, α2=0.2, α3=0.2, α4=0.1.

[0030] Preferably, the value ratio and the minimum value number in formula (1) and formula (2) are changed according to actual conditions.

[0031] Preferably, Δrank in formula (9) i The value of ′ changes according to the actual situation.

[0032] The beneficial effects of the present invention are:

[0033] This evaluation method evaluates the ranking function of the alternative model through weighted rank correlation, which significantly improves the accuracy, especially in the ranking performance in the high Sigma failure area; through more accurate model evaluation, it reduces the number of unnecessary simulations, improves efficiency, and saves time and computing resources.

[0034] By adopting the dynamic weighted formula, the regions can be freely divided and the weights can be adjusted dynamically. According to the needs of the actual engineering scenario, the evaluation model can be flexibly adjusted, which is more in line with the actual situation than the evaluation result. By introducing multi-region division and weighted calculation, it helps to improve the accuracy of rare event capture, especially in the evaluation of high Sigma failure areas, which can make better use of computing resources. The general formula can be applied to any number of divided regions and can be widely used in large-scale integrated circuits, high-reliability systems and other scenarios that require high Sigma failure areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic diagram of the overall process of the present invention;

[0036] Figure 2 is a flow chart of the calculation of the weighted rank correlation of the model evaluation in the present invention;

[0037] Figure 3 It is a division definition diagram of each area in the risk sample of the present invention;

[0038] Figure 4 It is a schematic diagram of the calculation of the number of risk samples in each area in the present invention;

[0039] Figure 5 A schematic diagram of Comparative Example 1 of the present invention;

[0040] Figure 6It is a schematic diagram of Comparative Example 2 of the present invention. DETAILED DESCRIPTION

[0041] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the structure of the present invention.

[0042] Embodiment 1: As shown in the attached specification Figure 1-2 As shown, the total sample set Sam 总 Contains 1,000,000 sample data, from which the initial sample Sam is randomly selected 初始 , according to the formula Qua 初始 =max(Qua 总 / 1000,1000) (1), the initial sample size Qua 初始 is 1000. For the selected initial sample Sam 初始 Use Spice circuit simulation tool to simulate and get the simulation result YS 初始 , the FFX algorithm is used to simulate the results YS 初始 Train machine learning equivalent surrogate models to generate predictive models.

[0043] Total sample setSam 总 Remove the initial sample Sam 初始 The sample set after that constitutes the remaining sample Sam 剩余 , then the remaining sample Sam 剩余 The number of samples is 999,000. For these 999,000 remaining samples Sam 剩余 Use the prediction model to make predictions and get the prediction results YP 剩余 , the predicted result YP 剩余 Arrange from large to small. We want to select high-risk samples Sam 风险 , high-risk sample Sam 风险 The number of Qua 风险 Satisfying formula (2): Qua 风险 =max(Qua 总 / 2000,500) (2), the top 500 samples are selected as high-risk samples Sam 风险 , high-risk sample Sam 风险 The predicted values ​​are arranged from large to small to get the predicted result YP 风险 This is the total sample set Sam 总 We will focus on analyzing the data with the highest risk.

[0044] In order to improve the accuracy of the evaluation method, we will continue to evaluate these 500 high-risk samples Sam 风险 Partitioning is performed and divided into samples Sam close to the failure area according to the actual situation接近 , Sample Sam of the area close to failure 次接近 and the general area sample Sam 普通 , and assign different weights to them for weighted calculation. Figure 3 For example, we have a high-risk sample Sam 风险 After the simulation, the simulation values ​​are sorted, and the samples Sam with higher risk rankings are close to the failure area. 接近 , and the risk ranking is the lowest for the general area sample Sam 普通 , the sample Sam located between the two is close to the failure area 次接近 .Depend on Figure 3 We can also see that the simulation results and prediction results for the same sample are not ranked the same.

[0045] The sample Sam close to the failure area 接近 , Sample Sam of the area close to failure 次接近 and the general area sample Sam 普通 The number of each includes how much, and the calculation method is as follows Figure 4 For simplicity, we only assume that 10 samples are risk samples Sam. 风险 .

[0046] We first determine the number of samples close to the failure region, where the numerical offset Δd and the average numerical offset Δd are used. 平均 : Arrange the simulation results YS from large to small 风险 Calculate the difference between two adjacent prediction values ​​one by one, that is, Δd i =(YS 风险 ) i -(YS 风险 ) i+1 (3), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 Each natural number of -1; and the entire risk sample Sam 风险 The average numerical deviation of

[0047] exist Figure 4 Among them, Δd1=68-58=10, Δd2=58-44=14. . . . . Δd9=6-4=2. The specific results are shown in the Δd column, and the average numerical offset Δd 平均 , the calculated value is 7.11.

[0048] We set a simulation value threshold Thr 实际 , then the sample Sam close to the failure area 接近 Threshold Thr 接近 =Thr 实际-3Δd 平均 (5), the sample Sam close to the failure area 次接近 Threshold Thr 次接近 =Thr 实际 -6Δd 平均 (6).

[0049] Here, we set the simulation value threshold Thr 实际 is 52, then the sample Sam is close to the failure area 接近 Threshold Thr 接近 =52-3*7.11=30.67, the next closest failure area sample Sam 次接近 Threshold Thr 次接近 =52-6*7.11=9.34, then all samples with simulation values ​​greater than or equal to 30.67 are selected as samples close to the failure area Sam 接近 ,Right now Figure 4 The red area in the middle; the samples with simulation values ​​greater than or equal to 9.34 and less than 30.67 are selected as the samples Sam close to the failure area 次接近 ,Right now Figure 4 The yellow area in the middle; the remaining samples are samples from the normal area Sam 普通 ,Right now Figure 4 The green area in the middle. That is, among these 10 samples, the samples ranked 1-4 are close to the failure area. 接近 , and the samples ranked 5-8 are Sam, which are close to the failure area. 次接近 , the remaining 9th and 10th place are common area samples Sam 普通 .

[0050] The high-risk sample Sam 风险 After being divided into three areas, we need to perform weighted summation on the three areas, where the sample Sam close to the failure area 接近 , Sample Sam of the area close to failure 次接近 and the general area sample Sam 普通 The weights α1, α2 and α3 are artificially defined. However, the Spearman rank coefficients ρ1, ρ2 and ρ3 of each region need to be calculated.

[0051] The sorting offset Δrank is used in the Spearman rank coefficient formula: the simulation result YS of each sample data is 风险 and its predicted result YP 风险 The ranking difference is obtained, that is, the ranking offset Δrank is obtained, and the ranking offset Δrank satisfies formula (8):

[0052] Δrank i =|rank(YS 风险 )i -rank(YP 风险 ) i | (8), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 For each natural number, rank(YS 风险 ) is the simulation result YS 风险 The ranking number, rank(YP 风险 ) is the prediction result YP 风险 's ranking number.

[0053] Calculate the sample Sam close to the failure area 接近 The Spearman rank coefficient ρ1: as shown in formula (9), that is, The relative ranking offset Δrank1' in the formula takes the following values: Among them, Thr1 and Thr2 are the artificially set first-level threshold Thr1 of samples close to the failure area and the second-level threshold Thr2 of samples close to the failure area.

[0054] When Δd≤Thr1, the sample ranking is considered to be very accurate and close to the failure area, so the relative ranking offset Δrank i ' is set to 0, indicating that these samples occupy the most important position in the evaluation and do not need to be sorted in the evaluation; and Thr1<Δd 平均 When ≤Thr2, the ranking error is small and Δrank' is assigned Δrank / 2, which means that these samples perform well in the failure area, but their importance is relatively reduced; when Δd 平均 >Thr2, Δrank' is assigned Δrank, which means that the ranking deviation of the sample is large, but it is still within the extended range of the failure area, so its ranking offset Δrank is directly assigned.

[0055] Still Figure 4 Taking the data in as an example, the calculation method of Δrank is: Δrank1=3-1=2, Δrank2=6-2=4. . . . . Δrank 10 =10-7=3, see Figure 4 The last column of data in the i Substituting into formula (11), we can obtain ρ1.

[0056] Sample Sam near failure area 次接近 The Spearman rank coefficient ρ2 and the common area sample Sam 普通 The Spearman rank coefficient ρ3: Then ρ2 is as in formula (10), that is, ρ3 is as in formula (11), that is, The calculation method only requires the sample data Δrank in each region i Just substitute the calculation.

[0057] In this way, we obtained these 500 high-risk samples Sam 风险 The weights and Spearman rank coefficients of the three regions in the equation are summed to obtain the final evaluation result. 总 The calculation formula is as shown in formula (12), that is, ρ 总 =α1·ρ1+α2·ρ2+α3·ρ3 (12), where α3=1-α1-α2, and α1>α2>α3.

[0058] Embodiment 2: Obviously, according to different actual situations, the high-risk sample Sam 风险 Divide into n regions according to the situation, calculate the Spearman rank coefficient of each region, define the weight of each region, and calculate the final weighted rank correlation evaluation result. The calculation is as shown in formula (13): Among them, ρ final is the final weighted rank correlation evaluation result, ρ i is the Spearman rank correlation coefficient of the samples in the ith region, α i is the weight of the ith region and satisfies

[0059] For example, the high-risk sample Sam 风险 It is divided into four areas according to the situation, namely, the area close to failure, the area close to failure, the extended area and the ordinary area, with weights of α1, α2, α3 and α4 respectively, and Spearman rank coefficients of ρ1, ρ2, ρ3 and ρ4 respectively. The final weighted rank result ρ final =α1·ρ1+α2·ρ2+α3·ρ3+α4·ρ3 (14), where α1=0.5, α2=0.2, α3=0.2, α4=0.1.

[0060] Obviously, the value ratio and the minimum value number in formula (1) and formula (2) change according to the actual situation, while Δrank in formula (9) i 'The value will change according to the actual situation.

[0061] Example 3: In actual applications, different samples have different characteristics. In addition to setting the threshold Thr 实际 The method is to determine the sample Sam close to the failure area 接近 , we can also directly set the number of samples close to the failure area Qua 接近 The method is to determine the sample Sam close to the failure area 接近 For example, setting thresholds S1 and S2,

[0062] Δd 平均 When ≤S1, the sample Sam is close to the failure area 接近 The number of Qua 接近 Satisfies formula (15):

[0063] Qua 接近 =max(Qua 总 / 20000,50) (15),

[0064] S1<Δd 平均 When ≤S2, the sample Sam is close to the failure area 接近 The number of Qua 接近 Satisfies formula (16):

[0065] Qua 接近 =max(Qua 总 / 5000,200) (16),

[0066] When Δd>S2, the sample Sam is close to the failure area 接近 The number of Qua 接近 Satisfies formula (17):

[0067] Qua 接近 =max(Qua 总 / 2000,500) (17),

[0068] Select simulation result YS 风险 Ranking from large to small Quad 接近 The samples are taken as the samples close to the failure area Sam 接近 , then the last sample Sam selected into the close failure area 接近 The simulation value is close to the failure area sample threshold Thr 接近 In this way, we determine the sample Sam close to the failure area 接近 The smaller the sorting error is, the smaller the failure area sample Sam is selected. 接近 The fewer the number of failure areas, and correspondingly, the larger the sorting error, the more failure area samples Sam is selected. 接近 The more.

[0069] Obviously, Thr2>Thr1, when Δd 平均 When ≤Thr1, it is considered that the ranking of the samples by simulation and prediction is very accurate and perfectly captures the tail failure area. Then we select a smaller proportion of samples as samples close to the failure area Sam 接近 It can better reflect the total sample set Sam 总 The minimum number is 50; and when Δd 平均>Thr1, we select the sample Sam close to the failure area 接近 The ratio should be larger, and the minimum number is 200; and when Δd 平均 When >Thr2, the minimum number is 500, which means that the processing results vary greatly and the probability of failure is high.

[0070] In this calculation method, Qua 次接近 =Qua 接近 -3Δd 平均 (18), determine the next closest failure area sample Sam in turn 次接近 and the general area sample Sam 普通 .

[0071] Comparative Example 1: Figure 5 As shown, it means that the high-risk sample Sam 风险 Necessity of partition setting. If two prediction models A and B are used to predict high-risk sample Sam 风险 After prediction, we get two groups of data, Y_PredictA and Y_PredictB. Among them, the ranking values ​​of the samples with high ranking in Y_PredictA have a huge gap with the simulation value ranking, while the ranking of the sample data in other dangerous areas is smaller than the simulation value ranking; while the ranking values ​​of the samples with high ranking in Y_PredictB have a small gap with the simulation value ranking, or even the same, and the ranking gap between the sample data in other dangerous areas and the simulation value is large. Through calculation, we get the Spearman coefficient of Y_PredictA to be 0.458, and the Spearman coefficient of Y_PredictB to be -0.458. Obviously, 0.458>-0.458. If only a single formula is used for calculation, the equivalent ranking accuracy of model A is higher than that of model B. But in fact, we are more concerned about the ranking of samples close to the failure area. The ranking of this area can best reflect the accuracy of the equivalent substitution model. So in order to avoid such a situation, we divide the top-ranked samples into samples close to the failure area Sam. 接近 , giving it a higher weight and setting more areas.

[0072] Comparative Example 2: Figure 6 As shown in the figure, it is necessary to set Δrank' when calculating the Spearman coefficient ρ1 of the sample close to the failure area. Assume that the two prediction models A and B are used to obtain the rankings of the two sets of prediction values ​​Y_PredictA and Y_PredictB, where the sample Sam close to the failure area of ​​Y_PredictA 接近 The predicted values ​​of Y_PredictB are ranked from 100 to 91, while the sample Sam close to the failure area 接近The predicted values ​​are ranked from 1 to 10. If the ratio between Δrank' and Δrank is not set, assuming that the value of α is 0.7, the Spearman coefficient rank ρ is calculated in both cases. A =ρ B =1.

[0073] However, the ranking effects of these two predictions are obviously not equivalent. Obviously, prediction result B is better. This is because when calculating the Spearman coefficient, we will calculate the Spearman coefficient of the two areas close to the failure area and other dangerous areas respectively. The sorting sequence number in the area is relative and cannot reflect the overall ranking quality. Therefore, we need to introduce a new indicator, the relative offset of the ranking Δrank'. When the numerical offset Δd is less than a certain value Thr1, generally 20% of the number of samples close to the failure area, we believe that the ranking is good, and the relative offset of the ranking Δrank' is set to 0 during calculation; when the numerical offset Δd is between Thr1 and Thr2, generally between 20%-50% of the number of samples close to the failure area, the relative offset of the ranking Δrank' is set to half of the original; in other cases, the sorting offset Δrank itself is taken. After calculation, the prediction result is consistent with the actual analysis result.

[0074] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A failure area assessment method for high Sigma scenarios in electronic engineering, characterized by: The following steps are involved: (S1) Establish the initial sample Sam 初始 :Sam from the total sample set 总 Randomly select an initial sample Sam from 初始 , initial sample Sam 初始 The number of Qua 初始 Satisfying formula (1): Qua 初始 =max(Qua 总 / 1000,1000) (1), where Qua 总 is the total sample size, the initial sample size Qua 初始 is the total sample size Qua 总 One thousandth of, and not less than 1000; (S2) Establish a prediction model: For the selected initial sample Sam 初始 Perform simulation and obtain the simulation result YS 初始 , using the simulation results YS 初始 Train machine learning equivalent surrogate models to generate predictive models; (S3) Screening high-risk samples Sam 风险 :Use the prediction model to predict the remaining samples Sam 剩余 Make a prediction and get the prediction result YP 剩余 , where the remaining sample Sam 剩余 Sam is the total sample set 总 Remove the initial sample Sam 初始 The following sample set, the remaining sample Sam 剩余 The predicted results of YP 剩余 Arrange from large to small (for the convenience of explanation, the indicators in this evaluation method are arranged in order from large to small by default. However, the specific sorting method can be adjusted according to the nature of the indicators. For example, the sorting of indicators such as power consumption and delay time from small to large will not affect the applicability and final results of the evaluation method. The same is true for S4 and S8). Select high-risk samples Sam from large to small 风险 , among which the high-risk sample Sam 风险 The number of Qua 风险 , satisfying formula (2): Qua 风险 =max(Qua 总 / 2000,500) (2), where Qua 总 is the total number of samples, high-risk sample Sam 风险 The number of Qua 风险 is the total sample size Qua 总 0.05% and no less than 500%, high-risk sample Sam 风险 The predicted value is YP 风险 ; (S4) Simulate high-risk sample Sam 风险 :For high-risk sample Sam 风险 Perform simulation, obtain simulation results, and arrange the simulation results from large to small to obtain simulation results YS 风险 ; (S5) Calculate the numerical offset Δd and the average numerical offset Δd 平均 : Arrange the simulation results YS from large to small 风险 Calculate the difference between two adjacent prediction values ​​one by one, that is, Δd i =(YS 风险 ) i -(YS 风险 ) i+1 (3), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 Each natural number of -1; and the entire risk sample Sam 风险 The average numerical deviation of (S6) Determine the sample Sam close to the failure area 接近 and the next closest failure area sample Sam 次接近 :Set a simulation value threshold Thr 实际 , then the sample Sam close to the failure area 接近 Threshold Thr 接近 =Thr 实际 -3Δd 平均 (5), the sample Sam close to the failure area 次接近 Threshold Thr 次接近 =Thr 实际 -6Δd 平均 (6) The simulation value is greater than or equal to Thr 接近 The sample is taken as the sample close to the failure area Sam 接近 , and the simulation value is greater than or equal to the sample threshold Thr of the next close to failure area 次接近 And it is smaller than the sample threshold Thr close to the failure area 接近 The sample of the next closest failure area is Sam 次接近 , then the sample Sam close to the failure area 接近 The number is Qua 接近 , the next closest failure area sample Sam 次接近 The number is Qua 次接近 ; (S7) Determine the normal area sample Sam 普通 :High-risk sample Sam 风险 Remove the sample Sam close to the failure area 接近 and the next closest failure area sample Sam 次接近 The following samples are normal area samples Sam 普通 , then the normal area sample Sam 普通 The number of Qua 普通 Satisfying formula (7), Qua 普通 =Qua 风险 -Qua 接近 -Qua 次接近 (7); (S8) Calculate the ranking offset Δrank: Compare the high-risk samples Sam 风险 Each sample data in step (S4) is simulated as result YS 风险 and its prediction result YP in step (S3) 风险 The ranking difference is the ranking offset Δrank, which satisfies formula (8): Δrank i =|rank(YS 风险 ) i -rank(YP 风险 ) i | (8), where i is ≤ high-risk sample Sam 风险 The number of Qua 风险 For each natural number, rank(YS 风险 ) is the simulation result YS 风险 The ranking number, rank(YP 风险 ) is the prediction result YP 风险 Rank number; (S9) Calculate the sample Sam close to the failure area 接近 The Spearman rank coefficient ρ1: as shown in formula (9), that is, Among them, the Δrank in the formula i 'The values ​​are as follows: Where Thr1 and Thr2 are the first-level threshold Thr1 and the second-level threshold Thr2 of samples close to the failure area, which are set manually; (S10) Calculate the next closest failure area sample Sam 次接近 The Spearman rank coefficient ρ2 and the common area sample Sam 普通 The Spearman rank coefficient ρ3: Then ρ2 is as in formula (10), that is, ρ3 is as in formula (11), that is, (S11) Comprehensive weighted sum ρ 总 :Define the weight of samples close to failure area as α1, the weight of samples next close to failure area as α2, and the weight of samples in common area as α3, then ρ 总 The calculation formula is as shown in formula (12), that is, ρ 总 =α1·ρ1+α2·ρ2+α3·ρ3 (12), where α3=1-α1-α2, and α1>α2>α3.

2. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: 0.6≤α1＜0.7，0.2≤α2＜0.3。 3. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: High-risk sample Sam 风险 Divide into n regions according to the situation, calculate the Spearman rank coefficient of each region, define the weight of each region, and calculate the final weighted rank correlation evaluation result. The calculation is as shown in formula (13): Among them, ρ final is the final weighted rank correlation evaluation result, ρ i is the Spearman rank correlation coefficient of the samples in the ith region, α i is the weight of the ith region, which is defined artificially and satisfies n≥2.

4. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 3, characterized in that: The high-risk sample Sam 风险 It is divided into four areas according to the situation, namely, the area close to failure, the area close to failure, the extended area and the ordinary area, with weights of α1, α2, α3 and α4 respectively, and Spearman rank coefficients of ρ1, ρ2, ρ3 and ρ4 respectively. The final weighted rank result ρ final =α1·ρ1+α2·ρ2+α3·ρ3+α4·ρ4 (14), where α1=0.5, α2=0.2, α3=0.2, α4=0.

1.

5. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: The simulation tool adopts Spice circuit simulation tool.

6. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: Alternative models such as FFX algorithm, linear regression, Lasso regression, LightGBM, etc. can be used to establish a prediction model.

7. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: The value ratio and the minimum value number in formula (1) and formula (2) will change according to the actual situation.

8. The failure area assessment method for high Sigma scenarios in electronic engineering according to claim 1, characterized in that: Δrank in formula (9) i 'The value changes according to the actual situation.

Images