Crystal furnace control parameter optimization method and system based on deep learning

Abstract

The invention relates to the technical field of crystal furnace control parameter optimization, and discloses a crystal furnace control parameter optimization method and system based on deep learning, and the method comprises the following steps: 1, obtaining temperature value data in a crystal furnace heating process, and analyzing the temperature value data to obtain a crystal growth state characterization value; and 2, comparing the crystal growth state characterization value with a crystal growth state characterization threshold value to generate a crystal growth state abnormal signal. The temperature value data of the crystal furnace is obtained and compared with the temperature threshold range, whether the temperature is abnormal or not is judged, the temperature abnormal data is obtained, and the crystal state characterization value is calculated based on the obtained temperature abnormal data, so that the crystal growth state in the crystal furnace can be indirectly judged through the temperature of the crystal furnace. The crystal growth state condition is judged through the crystal growth state characterization value, and a crystal growth state normal signal or a crystal growth state abnormal signal is generated.

Description

Technical Field

[0001] The present invention relates to the technical field of crystal furnace control parameter optimization, and in particular to a crystal furnace control parameter optimization method and system based on deep learning. Background Art

[0002] Crystal furnace is a special equipment used for crystal growth and preparation. It uses high temperature and specific environmental conditions to melt raw materials and crystallize them into required crystals through physical or chemical methods. Crystal furnace is widely used in semiconductor, optoelectronics, laser, material science and other fields.

[0003] However, most of the existing crystal furnace control systems operate based on preset fixed parameters or simplified physical models, lacking real-time feedback and adaptive adjustment mechanisms. During the crystal growth process, fluctuations in key parameters such as temperature and gas flow may have a significant impact on the crystal quality of the crystal. In addition, the existing crystal furnaces are unable to establish a prediction model based on the direct relationship between temperature, gas flow and crystal growth state to optimize parameter control, which may affect the quality and yield of the crystal in the actual production process.

[0004] In view of this, we proposed a crystal furnace control parameter optimization method and system based on deep learning. Summary of the invention

[0005] The purpose of the present invention is to provide a crystal furnace control parameter optimization method and system based on deep learning to solve the problem that manual adjustment has certain shortcomings in the above background.

[0006] The purpose of the present invention can be achieved through the following technical solutions:

[0007] A method and system for optimizing crystal furnace control parameters based on deep learning, comprising the following steps:

[0008] Step 1: Obtain the temperature value data during the crystal furnace heating process, and analyze the temperature value data to obtain the crystal growth state characterization value; Step 2: Compare the crystal growth state characterization value with the crystal growth state characterization threshold to generate a crystal growth state abnormal signal; Step 3: Based on the generated crystal growth state abnormal signal, obtain the gas flow data during the crystal furnace heating process, and analyze the gas flow data and the temperature value data to calculate the overlap judgment value; Step 4: Compare the overlap judgment value with the overlap judgment threshold to determine whether the gas flow has an impact on the temperature and generate a related impact signal; Step 5: Based on the related impact signal, obtain a crystal growth state prediction model.

[0009] As a further solution of the present invention: the process of generating the temperature abnormality signal is:

[0010] Obtain heating temperature value data within a crystal furnace detection cycle, and compare the heating temperature value with a temperature threshold range;

[0011] If the heating temperature value is not within the temperature threshold range, a temperature abnormality signal is generated.

[0012] As a further solution of the present invention: the process of obtaining the crystal growth state characterization value is:

[0013] Obtain the duration of generating the abnormal temperature signal, calculate the ratio of the duration to the detection period, and obtain the duration ratio CX;

[0014] Obtain the temperature extreme value within the duration of generating the abnormal temperature signal to obtain the abnormal temperature extreme value, obtain the temperature value data of generating the normal temperature signal, and sum and average them to obtain the normal temperature average value;

[0015] The difference between the abnormal temperature extreme value and the normal temperature mean is calculated, and the absolute value of the difference is calculated as a ratio with the normal temperature mean to obtain the abnormal temperature deviation ratio PC;

[0016] Substitute the duration ratio CX and the abnormal temperature deviation ratio PC into the formula SZ=a1*CX+a2*PC to calculate the crystal growth state characterization value SZ, where a1 and a2 are preset proportional coefficients.

[0017] As a further solution of the present invention: the process of generating the abnormal crystal growth state signal is as follows:

[0018] Obtaining a crystal growth state characterization value, and comparing the crystal growth state characterization value with a crystal growth state characterization threshold;

[0019] If the crystal growth state characterization value is greater than the crystal growth state characterization threshold, a crystal growth state abnormal signal is generated.

[0020] As a further solution of the present invention: the process of obtaining the overlap judgment value is:

[0021] Obtaining the time period for generating the gas flow abnormality signal and the time period for generating the temperature abnormality signal, and obtaining the overlapping time period by calculation;

[0022] Obtain the temperature value data and gas flow data of the overlapping time period, divide the overlapping time into several time detection points, and mark two adjacent time detection points as detection periods;

[0023] The temperature values ​​of two adjacent time detection points are calculated to obtain the absolute value of the difference, which is recorded as the temperature change value during the detection period;

[0024] The difference of the gas flow rates at two adjacent time detection points is calculated to obtain the absolute value of the difference, which is recorded as the gas flow rate change value during the detection period;

[0025] Calculate the ratio of the temperature change value and the gas flow change value in the same detection period to obtain the change ratio of the relevant value in the detection period;

[0026] The relevant numerical value change ratios of all detection periods are summed and averaged to obtain the mean of the relevant numerical value change ratios;

[0027] The difference between the change ratio of the relevant value in each detection period and the mean value of the change ratio of the relevant value is calculated to obtain the deviation value of the change ratio of the relevant value;

[0028] Compare the relevant numerical value change ratio deviation value with the relevant numerical value change ratio deviation threshold;

[0029] If the relevant numerical value change ratio deviation value is less than or equal to the relevant numerical value change ratio deviation threshold, it is marked as a normal change ratio;

[0030] The number of values ​​marked as normal change ratio is obtained, and the normal change ratio value is compared with the total value of the number of detection periods to obtain the overlap judgment value.

[0031] As a further solution of the present invention: the process of generating the gas flow abnormality signal is:

[0032] Obtaining gas flow data during the crystal furnace heating process, the gas flow data is obtained through a gas flow meter;

[0033] comparing the gas flow data with a gas flow data standard range;

[0034] If the gas flow data is not within the gas flow data standard range, a gas flow abnormality signal is generated.

[0035] As a further solution of the present invention: the process of generating the relevant influence signal is:

[0036] Obtaining a coincidence judgment value, and comparing the coincidence judgment value with a coincidence judgment threshold;

[0037] If the coincidence judgment value is greater than the coincidence judgment threshold, a relevant influence signal is generated.

[0038] As a further solution of the present invention: the process of obtaining the crystal growth state prediction model is:

[0039] Acquire crystal growth state characterization values, gas flow abnormality time ratio and gas flow abnormality degree value in the gas flow abnormality period for multiple times;

[0040] The gas flow abnormality time ratio, the gas flow abnormality degree value and the crystal growth state characterization value obtained in each gas flow abnormality period are taken as a data group to obtain N data groups;

[0041] Randomly select 0.7*N data groups from N data groups and mark them as calculation data groups, and mark the remaining 0.3*N data groups as verification data groups;

[0042] Randomly selecting three data groups from a plurality of calculation data groups as a solution data group to obtain a plurality of substitution data groups;

[0043] The preset crystal growth state prediction model is Y=A1*ZB+A2*CD+B, where Y is the crystal growth state characterization value, ZB represents the gas flow abnormal time ratio, CD represents the gas flow abnormal degree value, A1 represents the first coefficient, A2 represents the second coefficient, and B represents the crystal growth state error coefficient;

[0044] Substituting the input data set into the crystal growth state prediction model function equation, respectively calculating multiple sets of first coefficients, second coefficients and crystal growth state error coefficients;

[0045] The A1 value in the crystal growth model is obtained by summing up all the first coefficients obtained and taking the average value, the A2 value in the crystal growth model is obtained by summing up all the second coefficients obtained and taking the average value, and the B value in the crystal growth model is obtained by summing up all the crystal growth state error coefficients obtained and taking the average value, so as to obtain the crystal growth state prediction model;

[0046] Substituting the volume flow abnormal time ratio and the gas flow abnormal degree value in the verification data group into the crystal growth state prediction model obtained above to obtain a theoretical crystal growth state characterization value, performing a difference calculation between the obtained theoretical crystal growth state characterization value and the actual crystal growth state characterization value in the group, and marking the absolute value of the difference as a calculation deviation value;

[0047] comparing the calculated deviation value to a calculated deviation threshold;

[0048] If the calculated deviation value is less than or equal to the calculated deviation threshold, a valid array is generated;

[0049] If the calculated deviation value is greater than the calculated deviation threshold, an invalid array is generated;

[0050] Get the number of valid data and the number of invalid arrays;

[0051] Calculate the ratio of the number of valid arrays to the number of verification data groups to obtain the ratio of valid arrays;

[0052] Calculate the ratio of the number of invalid arrays to the number of verification data groups to obtain the invalid array ratio;

[0053] Compare the valid array proportion value with the invalid array proportion value;

[0054] If the percentage of valid arrays is greater than the percentage of invalid arrays, a model qualification signal is generated;

[0055] If the valid array ratio is less than or equal to the invalid array ratio, a model failure signal is generated.

[0056] As a further solution of the present invention: based on the model failure signal, all gas flow abnormal time proportions are integrated into a gas flow abnormal time proportion data group, all gas flow abnormal degree values ​​are integrated into a gas flow abnormal degree value data group, and all crystal growth state characterization values ​​are integrated into a crystal growth state characterization data group;

[0057] Obtain the standard deviations corresponding to the gas flow abnormality time proportion data group, the gas flow abnormality degree value data group, and the crystal growth state characterization data group respectively;

[0058] Performing ratio processing on the standard deviation of the gas flow abnormal time proportion data group and the standard deviation corresponding to the crystal growth state characterization data group to obtain a first standard deviation ratio;

[0059] Performing ratio processing on the standard deviation corresponding to the gas flow abnormality value data group and the standard deviation corresponding to the crystal growth state characterization value data group to obtain a second standard deviation ratio;

[0060] comparing the first standard deviation ratio to the second standard deviation ratio;

[0061] If the first standard deviation ratio is greater than the second standard deviation ratio, the gas flow abnormal time proportion data group is selected as the target data group;

[0062] If the first standard deviation ratio is smaller than the second standard deviation ratio, the gas flow abnormality degree value data group is selected as the target data group;

[0063] If the first standard deviation ratio is equal to the second standard deviation ratio, the gas flow abnormal time proportion data group or the gas flow abnormal degree value data group is selected as the target data group;

[0064] Based on the target data set, a crystal growth state prediction model is obtained.

[0065] As a further solution of the present invention: specifically comprising:

[0066] Data acquisition and calculation module: obtains the temperature value data during the crystal furnace heating process, and analyzes the temperature value data to obtain the crystal growth state characterization value;

[0067] State judgment module: compares the crystal growth state characterization value with the crystal growth state characterization threshold, and generates an abnormal crystal growth state signal;

[0068] Coincidence judgment analysis module: based on the generation of abnormal crystal growth state signals, the gas flow data during the crystal furnace heating process is obtained, and the coincidence judgment value is calculated based on the analysis of the gas flow data and the temperature value data;

[0069] Impact judgment module: compares the coincidence judgment value with the coincidence judgment threshold, determines whether the gas flow has an impact on the temperature, and generates a related impact signal;

[0070] Prediction model acquisition module: based on relevant influencing signals, obtain the crystal growth state prediction model.

[0071] Beneficial effects of the present invention:

[0072] (1) by obtaining the temperature value data of the crystal furnace and comparing it with the temperature threshold range, determining whether the temperature is abnormal, obtaining temperature abnormality data, and calculating the crystal state characterization value based on the obtained temperature abnormality data, so that the crystal growth state in the crystal furnace can be indirectly determined by the temperature of the crystal furnace, and the crystal growth state can be determined by the crystal growth state characterization value, and a crystal growth state normal signal or a crystal growth state abnormal signal can be generated;

[0073] (2) By obtaining the gas flow data of the crystal furnace, determining whether the gas flow data is normal, and calculating the overlap time period of the gas flow abnormality time and the temperature abnormality, obtaining the temperature value data and gas flow data of the overlap time period, calculating the overlap judgment value, and comparing the overlap judgment value with the overlap judgment threshold, it can be determined whether the gas flow affects the temperature;

[0074] (3) By obtaining the crystal growth state characterization value, the gas flow abnormality time ratio within the gas flow abnormality period, and the gas flow abnormality degree value, firstly, a crystal growth state prediction model is preset, and the obtained data is used for calculation and verification to determine whether the model is qualified, and a model qualified signal or a model unqualified signal is generated. When the model qualified signal is generated, the preset model is the desired model. When the model unqualified signal is generated, the obtained data is processed to obtain the desired model, so that the crystal growth state can be predicted based on the prediction. In the process of crystal growth, the appearance time of gas flow abnormality will be earlier than the appearance time of temperature abnormality. Predicting the crystal growth state through gas flow data is conducive to earlier discovery of abnormal crystal growth state problems, and then the control parameters can be optimized to improve the quality of the crystal growth state. BRIEF DESCRIPTION OF THE DRAWINGS

[0075] The present invention will be further described below in conjunction with the accompanying drawings.

[0076] Figure 1 It is a flow chart of a method for optimizing crystal furnace control parameters based on deep learning according to the present invention;

[0077] Figure 2 It is a schematic diagram of a process for obtaining a crystal growth state characterization value in a crystal furnace control parameter optimization method based on deep learning of the present invention;

[0078] Figure 3 It is a schematic diagram of the process of obtaining the overlap judgment value in a crystal furnace control parameter optimization method based on deep learning of the present invention. DETAILED DESCRIPTION

[0079] 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.

[0080] Example 1: Please refer to Figure 1-Figure 3 As shown, the present invention is a crystal furnace control parameter optimization method based on deep learning, comprising the following steps:

[0081] Step 1: Obtain temperature data during the crystal furnace heating process, and analyze the temperature data to obtain a crystal growth state characterization value;

[0082] In some embodiments, the heating temperature value data within the crystal furnace detection period (the detection period is set by the staff in this field according to work experience) is obtained, and the heating temperature value is compared with the temperature threshold range;

[0083] Among them, the temperature threshold range is set by the staff in this field according to the production specifications;

[0084] If the heating temperature value is within the temperature threshold range, it means that the heating temperature meets the production specified temperature and a normal temperature signal is generated;

[0085] If the heating temperature value is not within the temperature threshold range, it means that the heating temperature does not meet the production requirements, and a temperature abnormality signal is generated;

[0086] Obtain the duration of generating the abnormal temperature signal, calculate the ratio of the duration to the detection period, and obtain the duration ratio CX;

[0087] Obtain the temperature extreme value within the duration of generating the abnormal temperature signal to obtain the abnormal temperature extreme value, obtain the temperature value data of generating the normal temperature signal, and sum and average them to obtain the normal temperature average value;

[0088] The difference between the abnormal temperature extreme value and the normal temperature mean is calculated, and the absolute value of the difference is calculated as a ratio with the normal temperature mean to obtain the abnormal temperature deviation ratio PC;

[0089] Substitute the duration ratio CX and the abnormal temperature deviation ratio PC into the formula SZ=a1*CX+a2*PC to calculate the crystal growth state characterization value SZ, where a1 and a2 are preset proportional coefficients;

[0090] It should be explained that the duration ratio value represents the ratio of the generated temperature anomaly signal to the detection period, that is, the larger the duration ratio value is, the longer the temperature anomaly signal lasts, the larger the crystal growth state value is, and the worse the crystal growth state is;

[0091] The abnormal temperature deviation ratio indicates the deviation ratio between the abnormal temperature extreme value (maximum temperature or minimum temperature) and the normal temperature mean value, that is, the larger the abnormal temperature deviation ratio is, the larger the abnormal temperature extreme value is, the larger the crystal growth state value is, and the worse the crystal formation state is;

[0092] Step 2: comparing the crystal growth state characterization value with the crystal growth state characterization threshold to generate a crystal growth state normal signal or a crystal growth state abnormal signal;

[0093] In some embodiments, a crystal growth state characterization value is obtained, and the crystal growth state characterization value is compared with a crystal growth state characterization threshold value (the crystal growth state characterization threshold value is set by a person skilled in the art based on historical experimental data summary);

[0094] If the crystal growth state characterization value is less than or equal to the crystal growth state characterization threshold, it means that the current heating temperature, abnormal duration and temperature deviation are within the compliance range, and a crystal growth state normal signal is generated;

[0095] If the crystal growth state characterization value is greater than the crystal growth state characterization threshold, it means that the current heating temperature, abnormal duration and temperature deviation are not within the compliance range, and a crystal growth state abnormal signal is generated;

[0096] The core technology of this embodiment is: by obtaining the temperature value data of the crystal furnace and comparing it with the temperature threshold range, determining whether the temperature is abnormal, obtaining temperature anomaly data (duration ratio and abnormal temperature deviation ratio), and calculating the crystal state characterization value based on the obtained temperature anomaly data, so that the crystal growth state in the crystal furnace can be indirectly determined by the temperature of the crystal furnace, and the crystal growth state can be determined by the crystal growth state characterization value, thereby generating a normal crystal growth state signal or an abnormal crystal growth state signal.

[0097] Embodiment 2: Based on the above embodiment, the change of gas flow rate will directly affect the convective heat transfer effect in the crystal furnace, thereby affecting the temperature distribution. A larger gas flow rate can enhance the convective heat transfer, which may cause the heat to dissipate too quickly, while a smaller gas flow rate may cause the heat to accumulate locally and form a temperature gradient. The present invention is a crystal furnace control parameter optimization method based on deep learning, which also includes the following steps:

[0098] Step 3: Based on the generated crystal growth state abnormal signal, the gas flow data in the crystal furnace heating process is obtained, and the coincidence judgment value is calculated based on the analysis of the gas flow data and the temperature value data;

[0099] In some embodiments, gas flow data is obtained during the crystal furnace heating process, and the gas flow data is obtained by a gas flow meter;

[0100] Compare the gas flow data with the standard range of gas flow data (the standard range of gas flow data is set by the staff in this field based on the historical experimental data summary);

[0101] If the gas flow data is within the gas flow data standard range, it means that the gas flow data meets the production specifications and a normal gas flow signal is generated;

[0102] If the gas flow data is not within the gas flow data standard range, it means that the gas flow data does not meet the production specifications and a gas flow abnormality signal is generated;

[0103] Obtaining the time period for generating the gas flow abnormality signal and the time period for generating the temperature abnormality signal, and obtaining the overlapping time period by calculation;

[0104] Obtain the temperature value data and gas flow data of the overlapping time period, divide the overlapping time into several time detection points, and mark two adjacent time detection points as detection periods;

[0105] The temperature values ​​of two adjacent time detection points are calculated to obtain the absolute value of the difference, which is recorded as the temperature change value during the detection period;

[0106] The difference of the gas flow rates at two adjacent time detection points is calculated to obtain the absolute value of the difference, which is recorded as the gas flow rate change value during the detection period;

[0107] Calculate the ratio of the temperature change value and the gas flow change value in the same detection period to obtain the change ratio of the relevant value in the detection period;

[0108] The relevant numerical value change ratios of all detection periods are summed and averaged to obtain the mean of the relevant numerical value change ratios;

[0109] The difference between the change ratio of the relevant value in each detection period and the mean value of the change ratio of the relevant value is calculated to obtain the deviation value of the change ratio of the relevant value;

[0110] Compare the relevant numerical value change ratio deviation value with the relevant numerical value change ratio deviation threshold (the relevant numerical value change ratio deviation threshold is set by the staff in this field based on the historical test data summary);

[0111] If the deviation value of the relevant numerical value change ratio is less than or equal to the relevant numerical value change ratio deviation threshold, it means that the relevant numerical value change ratio during the detection period is not much different from the mean value of the relevant numerical value change ratio, and it is marked as a normal change ratio;

[0112] If the deviation value of the correlation value change ratio is greater than the deviation threshold of the correlation value change ratio, it means that the correlation value change ratio in the detection period is significantly different from the mean value of the correlation value change ratio, and it is marked as an abnormal change ratio;

[0113] Obtain a value marked as a normal change ratio, perform ratio processing on the normal change ratio value and the total value of the number of detection periods to obtain a coincidence judgment value;

[0114] Step 4: Compare the coincidence judgment value with the coincidence judgment threshold value to determine whether the gas flow rate has an impact on the temperature, and generate a relevant impact signal or an irrelevant impact signal;

[0115] In some embodiments, the coincidence judgment value is obtained, and the coincidence judgment value is compared with the coincidence judgment threshold (the coincidence judgment threshold is set by the staff in this field based on the historical test data summary);

[0116] If the coincidence judgment value is greater than the coincidence judgment threshold, it means that the normal change ratio has more values, and the higher the coincidence degree, the gas flow rate has an impact on the temperature, that is, it has an impact on the crystal growth state, and a related impact signal is generated;

[0117] If the coincidence judgment value is less than or equal to the coincidence judgment threshold, it means that the normal change ratio has a smaller value, and the lower the coincidence, the more irrelevant influence signals are generated;

[0118] The core technology of this embodiment is: by obtaining the gas flow data of the crystal furnace, judging whether the gas flow data is normal, and calculating the overlapping time period of the gas flow abnormality time and the temperature abnormality, obtaining the temperature value data and the gas flow data of the overlapping time period, and calculating the temperature change value and the gas flow change value of the detection period, calculating the relevant numerical value change ratio of the detection period, and based on the relevant numerical value change ratio, calculating the relevant numerical value change ratio deviation value, judging and marking the normal change ratio and the abnormal change ratio, calculating the proportion of the normal change ratio as the overlap judgment value, comparing the overlap judgment value with the overlap judgment threshold, generating a relevant influence signal or an irrelevant influence signal, so as to judge whether the gas flow affects the temperature.

[0119] Embodiment 3: Based on the above embodiments, it is found that during the crystal growth process, the gas flow rate and temperature are interrelated, the change of the gas flow rate will affect the temperature distribution, and the change of the temperature will directly affect the growth state of the crystal. The present invention is a crystal furnace control parameter optimization method based on deep learning, which also includes the following steps:

[0120] Step 5: Based on the relevant influencing signals, a crystal growth state prediction model is obtained;

[0121] In some embodiments, the crystal growth state characterization value, the gas flow abnormality time ratio within the gas flow abnormality period, and the gas flow abnormality degree value are obtained multiple times;

[0122] It should be explained that the gas flow abnormal time ratio is the same as the method for calculating the duration ratio CX in step 1, and the gas flow abnormal degree value is the same as the method for calculating the abnormal temperature deviation degree ratio PC in step 1;

[0123] The gas flow abnormality time ratio, the gas flow abnormality degree value and the crystal growth state characterization value obtained in each gas flow abnormality period are taken as a data group to obtain N data groups;

[0124] Randomly select 0.7*N data groups from N data groups and mark them as calculation data groups, and mark the remaining 0.3*N data groups as verification data groups;

[0125] Randomly selecting three data groups from a plurality of calculation data groups as a solution data group to obtain a plurality of substitution data groups;

[0126] The preset crystal growth state prediction model is Y=A1*ZB+A2*CD+B, where Y is the crystal growth state characterization value, ZB represents the gas flow abnormal time ratio, CD represents the gas flow abnormal degree value, A1 represents the first coefficient, A2 represents the second coefficient, and B represents the crystal growth state error coefficient;

[0127] Substituting the input data set into the crystal growth state prediction model function equation, respectively calculating multiple sets of first coefficients, second coefficients and crystal growth state error coefficients;

[0128] It should be noted that the reason for selecting three data groups as a substitution data group is that the crystal growth state prediction model has three unknowns, so it is necessary to substitute three data groups to solve the first coefficient, the second coefficient and the crystal growth state error coefficient;

[0129] The A1 value in the crystal growth model is obtained by summing up all the first coefficients obtained and taking the average value, the A2 value in the crystal growth model is obtained by summing up all the second coefficients obtained and taking the average value, and the B value in the crystal growth model is obtained by summing up all the crystal growth state error coefficients obtained and taking the average value, so as to obtain the crystal growth state prediction model;

[0130] Substituting the volume flow abnormal time ratio and the gas flow abnormal degree value in the verification data group into the crystal growth state prediction model obtained above to obtain a theoretical crystal growth state characterization value, performing a difference calculation between the obtained theoretical crystal growth state characterization value and the actual crystal growth state characterization value in the group, and marking the absolute value of the difference as a calculation deviation value;

[0131] Compare the calculated deviation value with the calculated deviation threshold (the calculated deviation threshold is set by the staff in this field based on the historical test data summary);

[0132] If the calculated deviation value is less than or equal to the calculated deviation threshold, it means that the theoretical crystal growth state characterization value is not much different from the actual crystal growth state characterization value, and a valid array is generated;

[0133] If the calculated deviation value is greater than the calculated deviation threshold, it means that the theoretical crystal growth state characterization value and the actual crystal growth state characterization value are greatly different, and an invalid array is generated;

[0134] Get the number of valid data and the number of invalid arrays;

[0135] Calculate the ratio of the number of valid arrays to the number of verification data groups to obtain the ratio of valid arrays;

[0136] Calculate the ratio of the number of invalid arrays to the number of verification data groups to obtain the invalid array ratio;

[0137] Compare the valid array proportion value with the invalid array proportion value;

[0138] If the percentage of valid arrays is greater than the percentage of invalid arrays, a model qualification signal is generated;

[0139] Based on the qualified signals of the above model, the crystal growth state prediction model is Y = A1*ZB+A2*CD+B;

[0140] If the valid array ratio is less than or equal to the invalid array ratio, a model failure signal is generated;

[0141] Based on the model failure signal, all gas flow abnormality time proportions are integrated into a gas flow abnormality time proportion data group, all gas flow abnormality degree values ​​are integrated into a gas flow abnormality degree value data group, and all crystal growth state characterization values ​​are integrated into a crystal growth state characterization data group;

[0142] Obtain the standard deviations corresponding to the gas flow abnormality time proportion data group, the gas flow abnormality degree value data group, and the crystal growth state characterization data group respectively;

[0143] Performing ratio processing on the standard deviation of the gas flow abnormal time proportion data group and the standard deviation corresponding to the crystal growth state characterization data group to obtain a first standard deviation ratio;

[0144] Performing ratio processing on the standard deviation corresponding to the gas flow abnormality value data group and the standard deviation corresponding to the crystal growth state characterization value data group to obtain a second standard deviation ratio;

[0145] comparing the first standard deviation ratio to the second standard deviation ratio;

[0146] If the first standard deviation ratio is greater than the second standard deviation ratio, the gas flow abnormal time proportion data group is selected as the target data group;

[0147] If the first standard deviation ratio is smaller than the second standard deviation ratio, the gas flow abnormality degree value data group is selected as the target data group;

[0148] If the first standard deviation ratio is equal to the second standard deviation ratio, the gas flow abnormal time proportion data group or the gas flow abnormal degree value data group is selected as the target data group;

[0149] Based on the target data set, a crystal growth state prediction model is obtained;

[0150] The method of obtaining the crystal growth state prediction model includes but is not limited to: constructing a linear regression equation according to the target data set and the crystal growth state characterization value, and then solving the regression coefficient by the least square method or other optimization algorithm to obtain the crystal growth state prediction model;

[0151] The core technology of this embodiment is: by obtaining the crystal growth state characterization value, the gas flow abnormality time ratio within the gas flow abnormality period and the gas flow abnormality degree value, firstly, by presetting the crystal growth state prediction model, calculating and verifying through the acquired data, judging whether the model is qualified, generating a model qualified signal or a model unqualified signal, when generating the model qualified signal, the preset model is the required model, when generating the model unqualified signal, the acquired data is processed to obtain the required model, so that the crystal growth state can be predicted according to the prediction, and in the process of crystal growth, the appearance time of gas flow abnormality will be earlier than the appearance time of temperature abnormality, predicting the crystal growth state through gas flow data is conducive to earlier discovery of the problem of abnormal crystal growth state, and then the control parameters can be optimized to improve the quality of the crystal growth state.

[0152] Embodiment 4: Based on the above embodiments, the present invention is a crystal furnace control parameter optimization system based on deep learning, which specifically includes:

[0153] Data acquisition and calculation module: obtains the temperature value data during the crystal furnace heating process, and analyzes the temperature value data to obtain the crystal growth state characterization value;

[0154] State judgment module: compares the crystal growth state characterization value with the crystal growth state characterization threshold value to generate a crystal growth state normal signal or a crystal growth state abnormal signal;

[0155] Coincidence judgment analysis module: obtains the gas flow data during the crystal furnace heating process, and analyzes the gas flow data and temperature value data to calculate the coincidence judgment value;

[0156] Impact judgment module: compares the coincidence judgment value with the coincidence judgment threshold, determines whether the gas flow has an impact on the temperature, and generates a relevant impact signal or an irrelevant impact signal;

[0157] Prediction model acquisition module: based on relevant influencing signals, obtain the crystal growth state prediction model.

[0158] The above is a detailed description of an embodiment of the present invention, but the content is only a preferred embodiment of the present invention and cannot be considered to limit the scope of implementation of the present invention. All equivalent changes and improvements made within the scope of the present invention should still fall within the scope of the patent coverage of the present invention.

Claims

1. A crystal furnace control parameter optimization method based on deep learning, characterized in that: The following steps are involved: Step 1: Obtain temperature data during the crystal furnace heating process, and analyze the temperature data to obtain a crystal growth state characterization value; Step 2: comparing the crystal growth state characterization value with the crystal growth state characterization threshold to generate a crystal growth state abnormality signal; Step 3: Based on the generated crystal growth state abnormal signal, the gas flow data in the crystal furnace heating process is obtained, and the coincidence judgment value is calculated based on the analysis of the gas flow data and the temperature value data; Step 4: Compare the coincidence judgment value with the coincidence judgment threshold to determine whether the gas flow rate has an impact on the temperature and generate a related impact signal; Step 5: Based on relevant influencing signals, obtain a crystal growth state prediction model.

2. The method for optimizing crystal furnace control parameters based on deep learning according to claim 1, characterized in that: The process of generating temperature anomaly signal is: Obtain heating temperature value data within a crystal furnace detection cycle, and compare the heating temperature value with a temperature threshold range; If the heating temperature value is not within the temperature threshold range, a temperature abnormality signal is generated.

3. The method for optimizing crystal furnace control parameters based on deep learning according to claim 2, characterized in that: The process of obtaining the crystal growth state characterization value is as follows: Obtain the duration of generating the abnormal temperature signal, calculate the ratio of the duration to the detection period, and obtain the duration ratio CX; Obtain the temperature extreme value within the duration of generating the abnormal temperature signal to obtain the abnormal temperature extreme value, obtain the temperature value data of generating the normal temperature signal, and sum and average them to obtain the normal temperature average value; The difference between the abnormal temperature extreme value and the normal temperature mean is calculated, and the absolute value of the difference is calculated as a ratio with the normal temperature mean to obtain the abnormal temperature deviation ratio PC; Substitute the duration ratio CX and the abnormal temperature deviation ratio PC into the formula SZ=a1*CX+a2*PC to calculate the crystal growth state characterization value SZ, where a1 and a2 are preset proportional coefficients.

4. The method for optimizing crystal furnace control parameters based on deep learning according to claim 3, characterized in that: The process of generating abnormal crystal growth state signal is as follows: Obtaining a crystal growth state characterization value, and comparing the crystal growth state characterization value with a crystal growth state characterization threshold; If the crystal growth state characterization value is greater than the crystal growth state characterization threshold, a crystal growth state abnormal signal is generated.

5. The method for optimizing crystal furnace control parameters based on deep learning according to claim 1, characterized in that: The process of obtaining the coincidence judgment value is as follows: Obtaining the time period for generating the gas flow abnormality signal and the time period for generating the temperature abnormality signal, and obtaining the overlapping time period by calculation; Obtain the temperature value data and gas flow data of the overlapping time period, divide the overlapping time into several time detection points, and mark two adjacent time detection points as detection periods; The temperature values ​​of two adjacent time detection points are calculated to obtain the absolute value of the difference, which is recorded as the temperature change value during the detection period; The difference of the gas flow rates at two adjacent time detection points is calculated to obtain the absolute value of the difference, which is recorded as the gas flow rate change value during the detection period; Calculate the ratio of the temperature change value and the gas flow change value in the same detection period to obtain the change ratio of the relevant value in the detection period; The relevant numerical value change ratios of all detection periods are summed and averaged to obtain the mean of the relevant numerical value change ratios; The difference between the change ratio of the relevant value in each detection period and the mean value of the change ratio of the relevant value is calculated to obtain the deviation value of the change ratio of the relevant value; Compare the relevant numerical value change ratio deviation value with the relevant numerical value change ratio deviation threshold; If the relevant numerical value change ratio deviation value is less than or equal to the relevant numerical value change ratio deviation threshold, it is marked as a normal change ratio; The number of values ​​marked as normal change ratio is obtained, and the normal change ratio value is compared with the total value of the number of detection periods to obtain the overlap judgment value.

6. The method for optimizing crystal furnace control parameters based on deep learning according to claim 5, characterized in that: The process of generating gas flow abnormality signal is: Obtaining gas flow data during the crystal furnace heating process, the gas flow data is obtained through a gas flow meter; comparing the gas flow data with a gas flow data standard range; If the gas flow data is not within the gas flow data standard range, a gas flow abnormality signal is generated.

7. The method for optimizing crystal furnace control parameters based on deep learning according to claim 5, characterized in that: The process of generating relevant impact signals is: Obtaining a coincidence judgment value, and comparing the coincidence judgment value with a coincidence judgment threshold; If the coincidence judgment value is greater than the coincidence judgment threshold, a relevant influence signal is generated.

8. The method for optimizing crystal furnace control parameters based on deep learning according to claim 1, characterized in that: The process of obtaining the crystal growth state prediction model is as follows: Acquire crystal growth state characterization values, gas flow abnormality time ratio and gas flow abnormality degree value in the gas flow abnormality period for multiple times; The gas flow abnormality time ratio, the gas flow abnormality degree value and the crystal growth state characterization value obtained in each gas flow abnormality period are taken as a data group to obtain N data groups; Randomly select 0.7*N data groups from N data groups and mark them as calculation data groups, and mark the remaining 0.3*N data groups as verification data groups; Randomly selecting three data groups from a plurality of calculation data groups as a solution data group to obtain a plurality of substitution data groups; The preset crystal growth state prediction model is Y=A1*ZB+A2*CD+B, where Y is the crystal growth state characterization value, ZB represents the gas flow abnormal time ratio, CD represents the gas flow abnormal degree value, A1 represents the first coefficient, A2 represents the second coefficient, and B represents the crystal growth state error coefficient; Substituting the input data set into the crystal growth state prediction model function equation, respectively calculating multiple sets of first coefficients, second coefficients and crystal growth state error coefficients; The A1 value in the crystal growth model is obtained by summing up all the first coefficients obtained and taking the average value, the A2 value in the crystal growth model is obtained by summing up all the second coefficients obtained and taking the average value, and the B value in the crystal growth model is obtained by summing up all the crystal growth state error coefficients obtained and taking the average value, so as to obtain the crystal growth state prediction model; Substituting the volume flow abnormal time ratio and the gas flow abnormal degree value in the verification data group into the crystal growth state prediction model obtained above to obtain a theoretical crystal growth state characterization value, performing a difference calculation between the obtained theoretical crystal growth state characterization value and the actual crystal growth state characterization value in the group, and marking the absolute value of the difference as a calculation deviation value; comparing the calculated deviation value to a calculated deviation threshold; If the calculated deviation value is less than or equal to the calculated deviation threshold, a valid array is generated; If the calculated deviation value is greater than the calculated deviation threshold, an invalid array is generated; Get the number of valid data and the number of invalid arrays; Calculate the ratio of the number of valid arrays to the number of verification data groups to obtain the ratio of valid arrays; Calculate the ratio of the number of invalid arrays to the number of verification data groups to obtain the invalid array ratio; Compare the valid array proportion value with the invalid array proportion value; If the percentage of valid arrays is greater than the percentage of invalid arrays, a model qualification signal is generated; If the valid array ratio is less than or equal to the invalid array ratio, a model failure signal is generated.

9. The method for optimizing crystal furnace control parameters based on deep learning according to claim 8, characterized in that: Based on the model failure signal, all gas flow abnormality time proportions are integrated into a gas flow abnormality time proportion data group, all gas flow abnormality degree values ​​are integrated into a gas flow abnormality degree value data group, and all crystal growth state characterization values ​​are integrated into a crystal growth state characterization data group; Obtain the standard deviations corresponding to the gas flow abnormality time proportion data group, the gas flow abnormality degree value data group, and the crystal growth state characterization data group respectively; The standard deviation of the gas flow abnormal time proportion data group is processed by ratio with the standard deviation corresponding to the crystal growth state characterization data group to obtain a first standard deviation ratio; Performing ratio processing on the standard deviation corresponding to the gas flow abnormality value data group and the standard deviation corresponding to the crystal growth state characterization value data group to obtain a second standard deviation ratio; comparing the first standard deviation ratio to the second standard deviation ratio; If the first standard deviation ratio is greater than the second standard deviation ratio, the gas flow abnormal time proportion data group is selected as the target data group; If the first standard deviation ratio is smaller than the second standard deviation ratio, the gas flow abnormality degree value data group is selected as the target data group; If the first standard deviation ratio is equal to the second standard deviation ratio, the gas flow abnormal time proportion data group or the gas flow abnormal degree value data group is selected as the target data group; Based on the target data set, a crystal growth state prediction model is obtained.

10. A system for optimizing crystal furnace control parameters based on deep learning, the system implementing the method according to any one of claims 1 to 9, characterized in that: Specifically include: Data acquisition and calculation module: obtains the temperature value data during the crystal furnace heating process, and analyzes the temperature value data to obtain the crystal growth state characterization value; State judgment module: compares the crystal growth state characterization value with the crystal growth state characterization threshold, and generates an abnormal crystal growth state signal; Coincidence judgment analysis module: based on the generation of abnormal crystal growth state signals, the gas flow data during the crystal furnace heating process is obtained, and the coincidence judgment value is calculated based on the analysis of the gas flow data and the temperature value data; Impact judgment module: compares the coincidence judgment value with the coincidence judgment threshold, determines whether the gas flow has an impact on the temperature, and generates a related impact signal; Prediction model acquisition module: based on relevant influencing signals, obtain the crystal growth state prediction model.

Images