A semiconductor data detection system and method based on artificial intelligence

Through the semiconductor data detection system based on artificial intelligence, using neural network algorithms and fault feature data, a fault warning model is built and judgment standards are optimized, which solves the problems of high error rate and high detection cost in semiconductor fault detection, and achieves a higher level of judgment accuracy and efficiency.

CN118625756BActive Publication Date: 2024-12-06JIANGSU MANWANG SEMICON TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410643498.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-23
Publication Date
2024-12-06
Estimated Expiration
2044-05-23

Smart Images

  • Figure CN118625756B_ABST
    Figure CN118625756B_ABST
Patent Text Reader

Abstract

The present invention discloses a semiconductor data detection system and method based on artificial intelligence, which belongs to the field of model optimization technology. The present invention collects all data in the semiconductor production process, extracts characteristic data reflecting semiconductor failure after preprocessing the collected data; collects environmental data when characteristic data changes in the history of semiconductor production, extracts control parameters that affect characteristic data in the semiconductor production process; calculates the influence function of control parameters on characteristic data; uses a neural network algorithm to build a fault warning model based on the calculated characteristic data; collects historical data of semiconductor production, calculates the fault threshold of semiconductors based on the characteristic data, measures the characteristic data of semiconductors in real time, determines whether the produced semiconductors fail based on the fault threshold and issues a warning; collects control parameters in semiconductor production in real time, and optimizes the fault warning model based on real-time control parameters.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of model optimization, and in particular to a semiconductor data detection system and method based on artificial intelligence. Background Art

[0002] In the early 20th century, the development of solid-state physics provided the foundation for the rise of semiconductor technology. People began to conduct in-depth research on the electronic structure, conduction mechanism and band structure of solid materials, and these studies laid the foundation for the subsequent semiconductor device theory. In the 1960s, the concept of integrated circuits was proposed and realized, marking a leap forward in the manufacturing technology of electronic devices. Integrated circuits integrate multiple electronic devices on a chip, making electronic devices more compact, high-performance and low-cost. With the continuous advancement of microelectronics technology, including photolithography, ion implantation, thin film technology, etc., the manufacturing process of semiconductor devices has become more precise and controllable, the size of devices has continued to shrink, and the performance has continued to improve. From the late 20th century to the early 21st century, the rapid development of the digital revolution and information technology has promoted the rapid advancement of semiconductor technology. Computers, communications, consumer electronics and other fields have put forward higher performance and functional requirements for semiconductor devices, prompting the continuous innovation and evolution of semiconductor technology. However, with the development of semiconductor technology, semiconductors have become more precise and small, and the detection of semiconductor failures has become more difficult; among them, the rise of big data and Internet of Things technology has provided new opportunities for semiconductor failure warning. By real-time monitoring and analysis of the working status and performance data of the equipment, potential signs of failure can be discovered and early warnings can be given to reduce the losses caused by failures. However, when early warning monitoring of semiconductor failures is performed, certain data is analyzed, and the data may change with changes in the environment, resulting in inaccurate judgments on semiconductor failures, greatly increasing the possibility of misjudgment. Summary of the invention

[0003] The object of the present invention is to provide a semiconductor data detection system and method based on artificial intelligence to solve the problems raised in the above background technology.

[0004] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0005] A semiconductor data detection method based on artificial intelligence, the method comprising the following steps:

[0006] S100, collecting all data in the semiconductor production process, pre-processing the collected data, and extracting characteristic data reflecting semiconductor failure;

[0007] Furthermore, the specific steps of extracting characteristic data reflecting semiconductor failure are as follows:

[0008] S101, collecting the records of semiconductor failures in use in history, extracting all data before and after the semiconductor failure, pre-processing all the collected data, selecting the nearest data value to fill the missing data value according to the proximity principle, searching the data for data that exceeds the normal data value by more than 10 times as much as the normal data value as variant data, clearing the variant data and re-filling it according to the proximity principle;

[0009] S102. Calculate the change in data before and after the semiconductor fails. The formula is:

[0010]

[0011] In the formula, S_c represents the data difference before and after the semiconductor fails, S after Indicates the data value after the semiconductor fails, S before It is represented by the data value before the semiconductor fails, and t represents the time when the semiconductor fails. After calculation, the data difference set before and after the semiconductor fails is {S_c1, S_c2, S_c3...S_c n}, S_c1, S_c2, S_c3...S_c n It is represented by the calculated 1st, 2nd, 3rd...nth data difference before and after the semiconductor fails, where n represents the data type in the semiconductor and is a positive integer;

[0012] S103, comparing the calculated data differences before and after the n semiconductors fail, and selecting the data difference with the largest absolute value as the characteristic data F_ch for determining whether the semiconductor fails.

[0013] According to the change range of other data when semiconductor failure occurs in history, the data with the largest change range is selected as the characteristic data reflecting the semiconductor failure. After finding the characteristic data, when judging the semiconductor failure, only the real-time characteristic data needs to be collected. According to the characteristic data, it can be judged whether the semiconductor fails, which greatly reduces the detection cost of the semiconductor.

[0014] S200, collecting environmental data when characteristic data in the semiconductor production process changes in history, and extracting control parameters that affect the characteristic data in the semiconductor production process;

[0015] Furthermore, the specific steps of extracting the control parameters that affect the characteristic data in the semiconductor production process are:

[0016] S201, collecting characteristic data and environmental data when semiconductor failure occurs in history, assuming that the value of the characteristic data collected when the semiconductor failure occurs is {F_ch1, F_ch2, ​​F_ch...F_ch m}, F_ch1, F_ch2, ​​F_ch...F_ch m It is represented by the value of the 1st, 2nd, 3rd...mth characteristic data when the semiconductor fails in the collected history, m is the number of times the semiconductor fails, and m is a positive integer; suppose the types of collected environmental data are {E1, E2, E3...E G}, E1, E2, E3...E G Indicates 1, 2, 3, .. G types of environmental data collected when a semiconductor fails, where G is a positive integer;

[0017] S202, sorting the collected characteristic data when the semiconductor fails, drawing a fault environment curve chart using the collected characteristic data as an independent variable and the environmental data as a variable, and obtaining a curve in which each environmental data changes with the characteristic data when the semiconductor fails;

[0018] S203, collecting characteristic data and environmental data when the semiconductor does not fail, and drawing a normal environment curve graph for each environmental data, to obtain a curve showing how each environmental data changes with the characteristic data when the semiconductor fails; comparing the two drawn curve graphs, analyzing the curves of each environmental data and characteristic data, and selecting environmental data with the same curve change rule in the two curve graphs as the control parameter F_co of the semiconductor characteristic data;

[0019] S204, after comparison, the control parameter types of the semiconductor characteristic data are obtained as {F_co1, F_co2, F_co3...F_co k}, F_co1, F_co2, F_co3...F_co k Indicates the 1st, 2nd, 3rd, ... kth control parameters obtained by judgment.

[0020] When judging whether a semiconductor has failed based on characteristic data, a judgment standard needs to be determined. However, due to environmental factors, the judgment standard may change in different environments. Therefore, it is necessary to find environmental data that affects the characteristic data and set it as a control parameter. By optimizing the judgment standard based on the real-time control parameters, the misjudgment of semiconductor failures can be reduced and the accuracy of semiconductor failure judgment can be increased.

[0021] S300, calculating the influence function of the control parameter on the characteristic data according to the historical characteristic data and the change data of the control parameter;

[0022] Furthermore, the specific steps for calculating the influence function of the control parameters on the characteristic data are:

[0023] S301, extract the characteristic data and k control parameters collected in the history in S200, take each control parameter as an independent variable, and the characteristic data as a variable to draw a scatter plot corresponding to the k control parameters, perform regression analysis on each scatter plot, and assume that the linear regression curve of each scatter plot is F_ch=a×F_co+b, where F_ch represents the value of the characteristic data, F_co represents the value of each control parameter, a represents the slope of the linear regression equation, and b represents the intercept of the linear regression equation; calculate the values ​​of a and b in the equation, and the formula is:

[0024]

[0025] b=F_ch pi -a×F_co pi

[0026] In the formula, P represents the number of points in the scatter plot, F_ch i Indicates the value of the characteristic data, F_co i Indicates the value of each control parameter, F_ch pi represents the average value of characteristic data, F_co pi represents the average value of the control parameters, i ranges from 1 to P; the linear regression equation corresponding to the k control parameters is finally calculated as {f1, f2, f3...f k}, f1, f2, f3...f k It is represented by the regression equation corresponding to the 1st, 2nd, 3rd, ..., control parameters obtained by calculation, where k is a positive integer;

[0027] S302, analyzing the linear regression equations of k control parameters, extracting a from the k regression equations, and obtaining the slopes {a1, a2, a3...a k}, a1, a2, a3...a k It represents the slope values ​​of the 1st, 2nd, 3rd...kth regression equations calculated. The slopes of k regression equations are analyzed and the proportion of the slope of each regression curve is calculated. The formula is:

[0028]

[0029] In the formula, a_z is the proportion of the slope of each regression equation, a r is the slope of each regression equation, r ranges from 1 to k; the slopes of the k regression equations calculated by design account for {a_z1, a_z2, a_z3...a_z k};

[0030] S303, comprehensively analyzing the influence of the slope proportion of each regression equation on the control parameters and the characteristic data, and calculating the comprehensive influence function of the control parameters on the characteristic data, the formula is:

[0031]

[0032] In the formula, a_z r represents the proportion of the slope in each regression equation, and a r represents the slope of each regression equation, and r takes values from 1 to k.

[0033] Construct an influence function of each control parameter on the characteristic data based on historical data. The linear regression equation can represent the influence relationship of the control parameter on the characteristic data, and the difference in the slope represents the different influence degrees of different control parameters on the characteristic data. According to the calculated proportion of the slope in each control parameter regression equation, comprehensively analyze the control parameter using the proportion to obtain the comprehensive influence function of the control parameter on the characteristic data;

[0034] S400. Based on the calculated characteristic data, use the neural network algorithm to construct a fault warning model; collect the historical data of semiconductor production, calculate the fault threshold of the semiconductor according to the characteristic data, measure the characteristic data of the semiconductor in real time, and judge whether the produced semiconductor fails and issue a warning according to the fault threshold;

[0035] Furthermore, the specific steps for judging whether the produced semiconductor fails and issuing a warning according to the fault threshold are as follows:

[0036] S401. Extract the values of the characteristic data when the semiconductor fails in the historical data collected in S200, and calculate the fault threshold according to the values of the characteristic data when the semiconductor fails in the history. The formula is:

[0037]

[0038] In the formula, y_f represents the fault threshold for judging whether the semiconductor fails, and F_ch e represents the characteristic data when the semiconductor fails in the history, and m represents the number of historical characteristic data collected;

[0039] S402. The characteristic data of the semiconductor measured in real time is F_ch_s. When F_ch_s ≥ y_f, it is judged that the semiconductor fails and a warning is issued; when F_ch_s < y_f, it is judged that the semiconductor does not fail and works normally;

[0040] S403. Use the neural network algorithm to construct a semiconductor fault warning model. Take the judgment mechanism in S402 as the network layer of the fault warning model, set the input layer and the output layer. After inputting the real-time measured characteristic data into the input layer and passing through the judgment of the network layer, a warning signal is output at the output layer.

[0041] S500, collecting control parameters in semiconductor production in real time, and optimizing the fault warning model according to the real-time control parameters.

[0042] Furthermore, the specific steps of optimizing the fault warning model according to the real-time control parameters are as follows:

[0043] S501, when measuring the characteristic data of the semiconductor in real time, measuring the control parameters in the real-time environment, and obtaining the real-time control parameters in k as {F_co_s1, F_co_s2, F_co_s3...F_co_s k}, F_co_s1, F_co_s2, F_co_s3...F_co_s k It is represented as the first, second, third, ... control parameters in the real-time measured environment. The predicted value of the characteristic data is calculated based on the real-time measured control parameters. The formula is:

[0044]

[0045] In the formula, F_y is the predicted value of real-time semiconductor characteristic data, F_co_s r represents the control parameter in the real-time measurement environment, and r ranges from 1 to k;

[0046] S502, extracting the characteristic data F_ch_s of the semiconductor measured in real time in S402 as the true value of the characteristic data, and calculating the error between the true value of the characteristic data and the predicted value, the formula is:

[0047] W_α=F_y-F_ch_s

[0048] In the formula, W_α represents the error between the true value and the predicted value of the characteristic data, F_y represents the predicted value of the real-time semiconductor characteristic data, and F_ch_s represents the true value of the characteristic data;

[0049] S503. When W_α≠0, it is determined that the fault warning model needs to be optimized, and the fault warning model is optimized using the calculated error value. The formula is:

[0050] y_f_op=y_f-W_α

[0051] In the formula, y_f_op represents the optimized fault threshold, y_f represents the fault threshold before optimization, and W_α represents the error between the predicted value and the true value of the characteristic data; the semiconductor fault is re-judged after the fault is optimized;

[0052] When W_α=0, there is no need to optimize the fault warning model and the system continues to work.

[0053] The real-time characteristic data is calculated using the control parameters to obtain the predicted value of the characteristic data. After measuring the true value of the characteristic data, the error between the predicted value and the true value is calculated. The fault threshold in the fault warning model is optimized according to the error. When y_f_op is positive, it means that the predicted value is greater than the true value, the true value becomes smaller, and the corresponding fault threshold should be reduced; when y_f_op is negative, it means that the predicted value is less than the true value, which means that the true value becomes larger, and the corresponding fault threshold should also be increased;

[0054] A semiconductor data detection system based on artificial intelligence, the semiconductor data detection system includes a data collection module, a feature search module, a control parameter search module, a function relationship calculation module, a fault warning model construction module and a model optimization module;

[0055] The data collection module is used to collect data on the semiconductor failure and non-failure in history;

[0056] The feature search module is used to calculate the difference between the data before and after the semiconductor failure in the collected history, and search for the feature data reflecting the semiconductor failure after judgment;

[0057] The control parameter search module is used to draw a curve chart of the change law of environmental data when the semiconductor characteristic data changes in history, and to find the control parameters that affect the characteristic data after analyzing the curve chart;

[0058] The function relationship calculation module is used to draw a scatter plot according to the characteristic data and control parameters in the history, construct a regression curve equation for each scatter plot, and then calculate the comprehensive influence function of the characteristic data and the control parameters after comprehensive analysis of k regression curve equations;

[0059] The fault warning model building module sets the input layer, the network layer and the output layer according to the neural network algorithm, calculates the fault threshold according to the historical data, measures the characteristic data in real time, judges the fault of the semiconductor according to the fault threshold, and issues a warning;

[0060] The model optimization module is used to compare the real value and the predicted value of the characteristic data, calculate the error value, and optimize the fault warning model according to the error value.

[0061] The function relationship calculation module is used to draw scatter plots corresponding to k control parameters with each control parameter as an independent variable and the characteristic data as a variable, perform regression analysis on each scatter plot, calculate the regression equations corresponding to the k control parameters, and then comprehensively analyze and calculate the comprehensive influence function of the control parameters on the characteristic data based on the proportion of the slope of each regression equation.

[0062] The fault warning model building module includes a fault threshold calculation unit and a model building unit;

[0063] The fault threshold calculation unit is used to calculate the fault threshold for determining whether a semiconductor fails based on the characteristic data collected in the history;

[0064] The model building unit is used to set the input layer, network layer and output layer of the historical neural network algorithm respectively, set the calculated fault threshold value at the network layer, input the real-time measurement feature data into the input layer, and output the warning signal after judgment by the network layer.

[0065] The model optimization module is used to measure the control parameters in the environment in real time when performing real-time fault judgment on semiconductors, calculate the predicted value of the characteristic data using the comprehensive influence function of the control parameters and the characteristic data, measure the real-time characteristic data as the true value, compare the error between the predicted value and the true value, and optimize the fault warning model according to the error; specifically, use the error value to optimize the fault threshold of the network layer in the fault warning model.

[0066] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0067] 1. The present invention analyzes the data of semiconductor failures in history and extracts characteristic data reflecting semiconductor failures. When judging semiconductor failures, it is only necessary to measure the characteristic data to determine whether the semiconductor has failed, which greatly reduces the time and resource consumption when judging semiconductor failures.

[0068] 2. The present invention extracts control parameters that affect characteristic data through historical data. When judging semiconductor faults, the error between the true value of the characteristic data and the measured value is calculated by measuring the real-time control parameters. Since the fault threshold for judging semiconductor faults is calculated from the characteristic data, the fault threshold can be optimized in real time through the error, thereby increasing the accuracy of semiconductor fault judgment and reducing the probability of misjudgment of semiconductor faults. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 It is a module distribution diagram of a semiconductor data detection system based on artificial intelligence of the present invention;

[0071] Figure 2 It is a schematic diagram of the steps of a semiconductor data detection method based on artificial intelligence of the present invention;

[0072] Figure 3 It is a fault curve diagram of a semiconductor data detection method based on artificial intelligence of the present invention;

[0073] Figure 4 It is a non-fault curve diagram of a semiconductor data detection method based on artificial intelligence of the present invention;

[0074] Figure 5 It is an ambient temperature-characteristic data scatter diagram of a semiconductor data detection method based on artificial intelligence of the present invention;

[0075] Figure 6 It is a dust density-characteristic data scatter diagram of a semiconductor data detection method based on artificial intelligence of the present invention. DETAILED DESCRIPTION

[0076] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. 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.

[0077] See also Figure 1-Figure 6 , the present invention provides a technical solution:

[0078] A semiconductor data detection method based on artificial intelligence, the method comprising the following steps:

[0079] S100, collecting all data in the semiconductor production process, pre-processing the collected data, and extracting characteristic data reflecting semiconductor failure;

[0080] The specific steps to extract the characteristic data of semiconductor failure are as follows:

[0081] S101. Collect the records of semiconductor failures in history, extract all data before and after the semiconductor failure, pre-process all the collected data, select the nearest data value to fill the missing data value according to the proximity principle, find the data that exceeds the normal data value by more than 10 times as the variant data, clear the variant data and then re-fill it according to the proximity principle;

[0082] S102. Calculate the change in data before and after the semiconductor fails. The formula is:

[0083]

[0084] In the formula, S_c represents the data difference before and after the semiconductor fails, S after Indicates the data value after the semiconductor fails, Sbefore It is represented by the data value before the semiconductor fails, and t represents the time when the semiconductor fails. After calculation, the data difference set before and after the semiconductor fails is {S_c1, S_c2, S_c3...S_c n}, S_c1, S_c2, S_c3...S_c n It is represented by the calculated 1st, 2nd, 3rd...nth data difference before and after the semiconductor fails, where n represents the data type in the semiconductor and is a positive integer;

[0085] S103, comparing the calculated data differences before and after the n semiconductors fail, and selecting the data difference with the largest absolute value as the characteristic data F_ch for determining whether the semiconductor fails.

[0086] According to the change range of other data when semiconductor failure occurs in history, the data with the largest change range is selected as the characteristic data reflecting the semiconductor failure. After finding the characteristic data, when judging the semiconductor failure, only the real-time characteristic data needs to be collected. According to the characteristic data, it can be judged whether the semiconductor fails, which greatly reduces the detection cost of the semiconductor.

[0087] S200, collecting environmental data when characteristic data in the semiconductor production process changes in history, and extracting control parameters that affect the characteristic data in the semiconductor production process;

[0088] The specific steps for extracting the control parameters that affect the characteristic data in the semiconductor production process are:

[0089] S201, collecting characteristic data and environmental data when semiconductor failure occurs in history, assuming that the value of the characteristic data collected when the semiconductor failure occurs is {F_ch1, F_ch2, ​​F_ch...F_ch m}, F_ch1, F_ch2, ​​F_ch...F_ch m It is represented by the value of the 1st, 2nd, 3rd...mth characteristic data when the semiconductor fails in the collected history, m is the number of times the semiconductor fails, and m is a positive integer; suppose the types of collected environmental data are {E1, E2, E3...E G}, E1, E2, E3...E G Indicates 1, 2, 3, .. G types of environmental data collected when a semiconductor fails, where G is a positive integer;

[0090] S202, sorting the collected characteristic data when the semiconductor fails, drawing a fault environment curve chart using the collected characteristic data as an independent variable and the environmental data as a variable, and obtaining a curve in which each environmental data changes with the characteristic data when the semiconductor fails;

[0091] S203, collecting characteristic data and environmental data when the semiconductor does not fail, and drawing a normal environment curve graph for each environmental data, to obtain a curve showing how each environmental data changes with the characteristic data when the semiconductor fails; comparing the two drawn curve graphs, analyzing the curves of each environmental data and characteristic data, and selecting environmental data with the same curve change rule in the two curve graphs as the control parameter F_co of the semiconductor characteristic data;

[0092] S204, after comparison, the control parameter types of the semiconductor characteristic data are obtained as {F_co1, F_co2, F_co3...F_co k}, F_co1, F_co2, F_co3...F_co k Indicates the 1st, 2nd, 3rd, ... kth control parameters obtained by judgment.

[0093] When judging whether a semiconductor has failed based on characteristic data, a judgment standard needs to be determined. However, due to environmental factors, the judgment standard may change in different environments. Therefore, it is necessary to find environmental data that affects the characteristic data and set it as a control parameter. By optimizing the judgment standard based on the real-time control parameters, the misjudgment of semiconductor failures can be reduced and the accuracy of semiconductor failure judgment can be increased.

[0094] S300, calculating the influence function of the control parameter on the characteristic data according to the historical characteristic data and the change data of the control parameter;

[0095] The specific steps for calculating the influence function of the control parameters on the characteristic data are:

[0096] S301, extract the characteristic data and k control parameters collected in the history in S200, take each control parameter as an independent variable, and the characteristic data as a variable to draw a scatter plot corresponding to the k control parameters, perform regression analysis on each scatter plot, and assume that the linear regression curve of each scatter plot is F_ch=a×F_co+b, where F_ch represents the value of the characteristic data, F_co represents the value of each control parameter, a represents the slope of the linear regression equation, and b represents the intercept of the linear regression equation; calculate the values ​​of a and b in the equation, and the formula is:

[0097]

[0098] b=F_ch pi -a×F_co pi

[0099] In the formula, P represents the number of points in the scatter plot, F_ch i Indicates the value of the characteristic data, F_co iIndicates the value of each control parameter, F_ch pi represents the average value of characteristic data, F_co pi represents the average value of the control parameters, i ranges from 1 to P; the linear regression equation corresponding to the k control parameters is finally calculated as {f1, f2, f3...f k}, f1, f2, f3...f k It is represented by the regression equation corresponding to the 1st, 2nd, 3rd, ..., control parameters obtained by calculation, where k is a positive integer;

[0100] S302, analyzing the linear regression equations of k control parameters, extracting a from the k regression equations, and obtaining the slopes {a1, a2, a3...a k}, a1, a2, a3...a k It represents the slope values ​​of the 1st, 2nd, 3rd...kth regression equations calculated. The slopes of k regression equations are analyzed and the proportion of the slope of each regression curve is calculated. The formula is:

[0101]

[0102] In the formula, a_z is the proportion of the slope of each regression equation, a r is the slope of each regression equation, r ranges from 1 to k; the slopes of the k regression equations calculated by design account for {a_z1, a_z2, a_z3...a_z k};

[0103] S303, comprehensively analyzing the influence of the slope proportion of each regression equation on the control parameters and the characteristic data, and calculating the comprehensive influence function of the control parameters on the characteristic data, the formula is:

[0104]

[0105] In the formula, a_z r Indicates the proportion of slope in each regression equation, a r It represents the slope of each regression equation, and r takes values ​​from 1 to k.

[0106] The influence function of each control parameter on the characteristic data is constructed based on historical data. The linear regression equation can represent the influence relationship of the control parameter on the characteristic data, and the different slopes represent the different influence degrees of different control parameters on the characteristic data. According to the calculated proportion of the slope in the regression equation of each control parameter, the control parameters are comprehensively analyzed using the proportion to obtain the comprehensive influence function of the control parameters on the characteristic data.

[0107] S400. Based on the calculated characteristic data, construct a fault warning model using a neural network algorithm; collect historical data of semiconductor production, calculate the fault threshold of the semiconductor according to the characteristic data, measure the characteristic data of the semiconductor in real time, and determine whether the produced semiconductor has a fault and issue a warning according to the fault threshold;

[0108] The specific steps for determining whether the produced semiconductor has a fault and issuing a warning according to the fault threshold are as follows:

[0109] S401. Extract the values of the characteristic data when the semiconductor fails in the historical data collected in S200, and calculate the fault threshold according to the values of the characteristic data when the semiconductor fails in the history. The formula is:

[0110]

[0111] In the formula, y_f represents the fault threshold for judging whether the semiconductor has a fault, F_ch e represents the characteristic data when the semiconductor fails in the history, and m represents the number of historical characteristic data collected;

[0112] S402. The characteristic data of the semiconductor measured in real time is F_ch_s. When F_ch_s ≥ y_f, it is judged that the semiconductor has a fault and a warning is issued; when F_ch_s < y_f, it is judged that the semiconductor will not have a fault and works normally;

[0113] S403. Use a neural network algorithm to construct a semiconductor fault warning model. Take the judgment mechanism in S402 as the network layer of the fault warning model, set the input layer and the output layer. After inputting the characteristic data measured in real time into the input layer and passing through the judgment of the network layer, a warning signal is output at the output layer.

[0114] S500. Collect the control parameters in semiconductor production in real time, and optimize the fault warning model according to the real-time control parameters.

[0115] The specific steps for optimizing the fault warning model according to the real-time control parameters are as follows:

[0116] S501. When measuring the characteristic data of the semiconductor in real time, measure the control parameters in the real-time environment to obtain k real-time control parameters as {F_co_s1, F_co_s2, F_co_s3...F_co_s k}, F_co_s1, F_co_s2, F_co_s3...F_co_s k represent the 1st, 2nd, 3rd... types of control parameters in the measured real-time environment. Calculate the predicted value of the characteristic data according to the measured real-time control parameters. The formula is:

[0117]

[0118] In the formula, F_y is the predicted value of real-time semiconductor characteristic data, F_co_s r represents the control parameter in the real-time measurement environment, and r ranges from 1 to k;

[0119] S502, extracting the characteristic data F_ch_s of the semiconductor measured in real time in S402 as the true value of the characteristic data, and calculating the error between the true value of the characteristic data and the predicted value, the formula is:

[0120] W_α=F_y-F_ch_s

[0121] In the formula, W_α represents the error between the true value and the predicted value of the characteristic data, F_y represents the predicted value of the real-time semiconductor characteristic data, and F_ch_s represents the true value of the characteristic data;

[0122] S503. When W_α≠0, it is determined that the fault warning model needs to be optimized, and the fault warning model is optimized using the calculated error value. The formula is:

[0123] y_f_op=y_f-W_α

[0124] In the formula, y_f_op represents the optimized fault threshold, y_f represents the fault threshold before optimization, and W_α represents the error between the predicted value and the true value of the characteristic data; the semiconductor fault is re-judged after the fault is optimized;

[0125] When W_α=0, there is no need to optimize the fault warning model and the system continues to work.

[0126] The real-time characteristic data is calculated using the control parameters to obtain the predicted value of the characteristic data. After measuring the true value of the characteristic data, the error between the predicted value and the true value is calculated. The fault threshold in the fault warning model is optimized according to the error. When y_f_op is positive, it means that the predicted value is greater than the true value, the true value becomes smaller, and the corresponding fault threshold should be reduced; when y_f_op is negative, it means that the predicted value is less than the true value, which means that the true value becomes larger, and the corresponding fault threshold should also be increased;

[0127] A semiconductor data detection system based on artificial intelligence, the semiconductor data detection system includes a data collection module, a feature search module, a control parameter search module, a function relationship calculation module, a fault warning model construction module and a model optimization module;

[0128] The data collection module is used to collect data on the semiconductor failure and non-failure in history;

[0129] The feature search module is used to calculate the difference between the data before and after the semiconductor failure in the collected history, and search for the feature data reflecting the semiconductor failure after judgment;

[0130] The control parameter search module is used to draw a curve chart of the change law of environmental data when the semiconductor characteristic data changes in history, and to find the control parameters that affect the characteristic data after analyzing the curve chart;

[0131] The function relationship calculation module is used to draw a scatter plot according to the characteristic data and control parameters in the history, construct a regression curve equation for each scatter plot, and then calculate the comprehensive influence function of the characteristic data and the control parameters after comprehensive analysis of k regression curve equations;

[0132] The fault warning model building module sets the input layer, the network layer and the output layer according to the neural network algorithm, calculates the fault threshold according to the historical data, measures the characteristic data in real time, judges the fault of the semiconductor according to the fault threshold, and issues a warning;

[0133] The model optimization module is used to compare the real value and the predicted value of the characteristic data, calculate the error value, and optimize the fault warning model according to the error value.

[0134] The function relationship calculation module is used to draw scatter plots corresponding to k control parameters with each control parameter as an independent variable and the characteristic data as a variable, perform regression analysis on each scatter plot, calculate the regression equations corresponding to the k control parameters, and then comprehensively analyze and calculate the comprehensive influence function of the control parameters on the characteristic data based on the proportion of the slope of each regression equation.

[0135] The fault warning model building module includes a fault threshold calculation unit and a model building unit;

[0136] The fault threshold calculation unit is used to calculate the fault threshold for determining whether a semiconductor fails based on the characteristic data collected in the history;

[0137] The model building unit is used to set the input layer, network layer and output layer of the historical neural network algorithm respectively, set the calculated fault threshold value at the network layer, input the real-time measurement feature data into the input layer, and output the warning signal after judgment by the network layer.

[0138] The model optimization module is used to measure the control parameters in the environment in real time when performing real-time fault judgment on semiconductors, calculate the predicted value of the characteristic data using the comprehensive influence function of the control parameters and the characteristic data, measure the real-time characteristic data as the true value, compare the error between the predicted value and the true value, and optimize the fault warning model according to the error; specifically, use the error value to optimize the fault threshold of the network layer in the fault warning model.

[0139] Example 1

[0140] Collect three kinds of data before and after the failure of semiconductors in history, namely temperature, power consumption and current, and calculate the change of each data by formula: After calculation, the changes of the three data are 23, 52, and 102, and the current is finally judged to be the characteristic data of semiconductor failure.

[0141] Collect the characteristic data and environmental data values ​​of the semiconductor when it fails and does not fail in history, and draw a curve chart such as Figure 3 and Figure 4 As shown, after judgment, the environmental data with the same curve change law are the environmental temperature and dust density data, that is, the environmental temperature and dust density are the control parameters of the characteristic data;

[0142] Draw a scatter plot based on the collected feature data and control parameters, such as Figure 5 and Figure 6 As shown, the regression equations of the two control parameters are calculated as F_ch=0.4737×F_co+175.06 and F_ch=1.5946×F_co+160.62. After comprehensive analysis of the two regression equations, the comprehensive influence function is: F_ch=1.33×F_co+167.84.

[0143] The fault threshold calculated through historical data is 209, and the real-time collected feature data is 200. It is preliminarily judged that the semiconductor will not fail. Then the real-time collected ambient temperature and dust density are 45 and 33. The predicted value of the feature data is 215 after calculation. The error between the predicted value and the true value is 15. After optimizing the fault threshold, the fault threshold is 194. After optimization, it is judged that the real-time semiconductor may fail and an early warning is issued.

[0144] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, article or device.

[0145] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art can still modify the technical solutions described in the aforementioned embodiments or replace some of the technical features therein by equivalents. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A semiconductor data detection method based on artificial intelligence, characterized in that: The method comprises the following steps: S100, collecting all data in the semiconductor production process, pre-processing the collected data, and extracting characteristic data reflecting semiconductor failure; S200, collecting environmental data when characteristic data in the semiconductor production process changes in history, and extracting control parameters that affect the characteristic data in the semiconductor production process; S300, calculating a comprehensive influence function of the control parameters on the characteristic data according to the historical characteristic data and the change data of the control parameters; S400, constructing a fault warning model using a neural network algorithm based on the calculated characteristic data; collecting historical data of semiconductor production, calculating a semiconductor fault threshold based on the characteristic data, measuring the semiconductor characteristic data in real time, determining whether the produced semiconductor has a fault based on the fault threshold and issuing a warning; S500, collecting control parameters in semiconductor production in real time, and optimizing the fault warning model according to the real-time control parameters; The specific steps for optimizing the fault warning model according to the real-time control parameters are as follows: S501, when measuring the characteristic data of the semiconductor in real time, measuring the control parameters in the real-time environment, and obtaining the real-time control parameters in k as {F_co_s1, F_co_s2, F_co_s3...F_co_s k }, F_co_s1, F_co_s2, F_co_s3...F_co_s k It is represented as the first, second, third, ... control parameters in the real-time measured environment. The predicted value of the characteristic data is calculated based on the real-time measured control parameters. The formula is: In the formula, F_y is the predicted value of real-time semiconductor characteristic data, F_co_s r represents the control parameter in the real-time measurement environment, and r ranges from 1 to k; S502, extracting the characteristic data F_ch_s of the semiconductor measured in real time in S402 as the true value of the characteristic data, and calculating the error between the true value of the characteristic data and the predicted value, the formula is: W_α=F_y-F_ch_s In the formula, W_α represents the error between the true value and the predicted value of the characteristic data, F_y represents the predicted value of the real-time semiconductor characteristic data, and F_ch_s represents the true value of the characteristic data; S503. When W_α≠0, it is determined that the fault warning model needs to be optimized, and the fault warning model is optimized using the calculated error value. The formula is: y_f_op=y_f-W_α In the formula, y_f_op represents the optimized fault threshold, y_f represents the fault threshold before optimization, and W_α represents the error between the predicted value and the true value of the characteristic data; the semiconductor fault is re-judged after the fault is optimized; When W_α=0, there is no need to optimize the fault warning model and the system continues to work.

2. The semiconductor data detection method based on artificial intelligence according to claim 1, characterized in that: The specific steps of extracting characteristic data of reaction semiconductor failure in S100 are: S101. Collect the records of semiconductor failures in history, extract all data before and after the semiconductor failure, pre-process all the collected data, select the nearest data value to fill the missing data value according to the proximity principle, find the data that exceeds the normal data value by more than 10 times as the variant data, clear the variant data and then re-fill it according to the proximity principle; S102. Calculate the change in data before and after the semiconductor fails. The formula is: In the formula, S_c represents the data difference before and after the semiconductor fails, S after Indicates the data value after the semiconductor fails, S before It is represented by the data value before the semiconductor fails, and t represents the time when the semiconductor fails. After calculation, the data difference set before and after the semiconductor fails is {S_c1, S_c2, S_c3...S_c n }, S_c1, S_c2, S_c3...S_c n It is represented by the calculated 1st, 2nd, 3rd...nth data difference before and after the semiconductor fails, where n represents the data type in the semiconductor and is a positive integer; S103, comparing the calculated data differences before and after the n semiconductors fail, and selecting the data difference with the largest absolute value as the characteristic data F_ch for determining whether the semiconductor fails.

3. The semiconductor data detection method based on artificial intelligence according to claim 1, characterized in that: The specific steps of extracting the control parameters that affect the characteristic data in the semiconductor production process in S200 are: S201, collecting characteristic data and environmental data when semiconductor failure occurs in history, assuming that the value of the characteristic data collected when the semiconductor failure occurs is {F_ch1, F_ch2, ​​F_ch...F_ch m }, F_ch1, F_ch2, ​​F_ch...F_ch m It is represented by the value of the 1st, 2nd, 3rd...mth characteristic data when the semiconductor fails in the collected history, m is the number of times the semiconductor fails, and m is a positive integer; suppose the types of collected environmental data are {E1, E2, E3...E G }, E1, E2, E3...E G Indicates 1, 2, 3, .. G types of environmental data collected when a semiconductor fails, where G is a positive integer; S202. Sort the characteristic data when the semiconductor fails. Take the collected characteristic data as the independent variable and the environmental data as the variable to plot a fault environment curve graph, and obtain the curve of each environmental data changing with the characteristic data when the semiconductor fails. S203. Collect the characteristic data and environmental data when the semiconductor does not fail, and plot a normal environment curve graph for each environmental data, and obtain the curve of each environmental data changing with the characteristic data when the semiconductor fails. Compare the two plotted curve graphs, analyze the curves of each environmental data and characteristic data, and select the environmental data with the same curve change rule in the two curve graphs as the control parameter F_co of the semiconductor characteristic data. S204, after comparison, the control parameter types of the semiconductor characteristic data are obtained as {F_co1, F_co2, F_co3...F_co k }, F_co1, F_co2, F_co3...F_co k Indicates the 1st, 2nd, 3rd, ... kth control parameters obtained by judgment.

4. The semiconductor data detection method based on artificial intelligence according to claim 1, characterized in that: The specific steps for calculating the comprehensive influence function of the control parameter on the characteristic data in S300 are as follows: S301. Extract the historical characteristic data and k control parameters collected in S200. Take each control parameter as the independent variable and the characteristic data as the variable to plot a scatter diagram corresponding to the k control parameters. Conduct a regression analysis on each scatter diagram. Assume that the linear regression curve of each scatter diagram is F_ch = a×F_co + b. In the function, F_ch represents the value of the characteristic data, F_co represents the value of each control parameter, a represents the slope of the linear regression equation, and b represents the intercept of the linear regression equation. Calculate the values of a and b in the equation. The formula is: b=F_ch pi -a×F_co pi In the formula, P represents the number of points in the scatter plot, F_ch i Indicates the value of characteristic data, F_co i Indicates the value of each control parameter, F_ch pi represents the average value of characteristic data, F_co pi represents the average value of the control parameters, i ranges from 1 to P; the linear regression equation corresponding to the k control parameters is finally calculated as {f1, f2, f3...f k }, f1, f2, f3...f k It is represented by the regression equation corresponding to the 1st, 2nd, 3rd, ..., control parameters obtained by calculation, where k is a positive integer; S302, analyzing the linear regression equations of k control parameters, extracting a from the k regression equations, and obtaining the slopes {a1, a2, a3...a k }, a1, a2, a3...a k It represents the slope values ​​of the 1st, 2nd, 3rd...kth regression equations calculated. The slopes of k regression equations are analyzed and the proportion of the slope of each regression curve is calculated. The formula is: In the formula, a_z is the proportion of the slope of each regression equation, a r is the slope of each regression equation, r ranges from 1 to k; the slopes of the k regression equations calculated by design account for {a_z1, a_z2, a_z3...a_z k }; S303. Conduct a comprehensive analysis on the influence relationship between the control parameter and the characteristic data according to the slope ratio of each regression equation, and calculate the comprehensive influence function of the control parameter on the characteristic data. The formula is: In the formula, a_z r Indicates the proportion of slope in each regression equation, a r It represents the slope of each regression equation, and r takes values ​​from 1 to k.

5. The semiconductor data detection method based on artificial intelligence according to claim 1, characterized in that: The specific steps for judging whether the produced semiconductor fails and issuing a warning according to the fault threshold in S400 are as follows: S401. Extract the value of the characteristic data when the semiconductor fails in the history collected in S200, and calculate the fault threshold according to the value of the characteristic data when the semiconductor fails in the history. The formula is: In the formula, y_f represents the fault threshold for determining whether a semiconductor fails, and F_ch e represents the characteristic data when the semiconductor fails in history, and m represents the number of historical characteristic data collected; S402. Measure the characteristic data of the semiconductor in real time as F_ch_s. When F_ch_s ≥ y_f, judge that the semiconductor fails and issue a warning; when F_ch_s < y_f, judge that the semiconductor does not fail and works normally. S403. Use the neural network algorithm to construct a semiconductor fault warning model. Take the judgment mechanism in S402 as the network layer of the fault warning model. Set the input layer and output layer. After inputting the real-time measured characteristic data into the input layer, after being judged by the network layer, generate a warning signal and output it in the output layer.

6. A semiconductor data detection system based on artificial intelligence, characterized in that: The semiconductor data detection system includes a data collection module, a characteristic search module, a control parameter search module, a function relationship calculation module, a fault warning model construction module, and a model optimization module. The data collection module is used to collect the data when the semiconductor fails and does not fail in the history. The characteristic search module is used to calculate the difference between the data before and after the semiconductor fails in the collected history, and search for the characteristic data reflecting the semiconductor fault after judgment. The control parameter search module is used to draw a curve chart of the change law of environmental data when the semiconductor characteristic data changes in history, and to find the control parameters that affect the characteristic data after analyzing the curve chart; The function relationship calculation module is used to draw a scatter plot according to the characteristic data and control parameters in the history, construct a regression curve equation for each scatter plot, and then calculate the comprehensive influence function of the characteristic data and the control parameters after comprehensive analysis of k regression curve equations; The fault warning model building module sets the input layer, the network layer and the output layer according to the neural network algorithm, calculates the fault threshold according to the historical data, measures the characteristic data in real time, judges the fault of the semiconductor according to the fault threshold, and issues a warning; The model optimization module is used to compare the real value and the predicted value of the characteristic data, calculate the error value, and optimize the fault warning model according to the error value.

7. The semiconductor data detection system based on artificial intelligence according to claim 6, characterized in that: The functional relationship calculation module is used to draw scatter plots corresponding to k control parameters with each control parameter as an independent variable and the characteristic data as a variable, perform regression analysis on each scatter plot, calculate the regression equations corresponding to the k control parameters, and then comprehensively analyze and calculate the comprehensive influence function of the control parameters on the characteristic data based on the proportion of the slope of each regression equation.

8. The semiconductor data detection system based on artificial intelligence according to claim 6, characterized in that: The fault warning model building module includes a fault threshold calculation unit and a model building unit; The fault threshold calculation unit is used to calculate the fault threshold for determining whether a semiconductor fails based on the characteristic data collected in the history; The model building unit is used to set the input layer, network layer and output layer of the historical neural network algorithm respectively, set the calculated fault threshold value at the network layer, input the real-time measurement feature data into the input layer, and output the warning signal after judgment by the network layer.

9. The semiconductor data detection system based on artificial intelligence according to claim 6, characterized in that: The model optimization module is used to measure the control parameters in the environment in real time when performing real-time fault judgment on semiconductors, calculate the predicted value of the characteristic data using the comprehensive influence function of the control parameters and the characteristic data, measure the real-time characteristic data as the true value, compare the error between the predicted value and the true value, and optimize the fault warning model according to the error; specifically, optimize the fault threshold of the network layer in the fault warning model using the error value.