Method for intelligently predicting crystal pulling ending

Abstract

The invention provides a method for intelligently predicting crystal pulling ending. The method comprises the steps of collecting historical data and performing preprocessing; acquiring actual data; establishing a prediction model based on the preprocessed data and actual data; and correcting the prediction model to obtain the ending section drawing process. According to the method for intelligently predicting the crystal pulling ending, the ending section of the crystal pulling can be automatically simulated and subjected to de-noising judgment, so that the ending process matched with the existing crystal pulling process is obtained, the ending model of the drawn crystal bar is intelligently predicted, drawing of the ending section in the crystal bar can be automatically completed, and the drawing efficiency is improved. And the diameter size of the ending section is obtained in real time, the effectiveness is high, the accuracy is high, the risks of raw material waste and labor hour loss caused by human factors are reduced, the automatic control of ending section drawing is completely realized, and the working efficiency is improved.

Description

Technical Field

[0001] The present application belongs to the technical field of direct-pulling single crystal production, and in particular, relates to a method for intelligently predicting the end of crystal pulling. Background Art

[0002] The current finishing operation is monitored by the CCD on the single crystal furnace body. Since the diameter of the tail end is gradually reduced, it is blocked by the equal diameter segment in the CCD projection, and its projection cannot be captured, making it impossible to intuitively and accurately judge its state, and it is also impossible to obtain the real-time diameter value of the measured finishing segment. When the real-time diameter of the finishing segment cannot be directly obtained in real time, it is necessary to rely on the process parameters set by manual experience to complete the finishing drawing within a fixed time, that is, the entire finishing process belongs to the passive execution of process parameters, that is, the set finishing must be completed within a fixed time. This method of manually controlling the finishing needs to be determined based on the operator's work experience and operating techniques, has poor stability, and is very likely to cause waste of raw materials and ineffective loss of working hours. Summary of the invention

[0003] The present application provides an intelligent prediction method for crystal pulling tailing, which solves the technical problems in the prior art of relying on manual experience to perform the tail pulling stage, resulting in poor tail quality consistency and serious waste of labor and materials.

[0004] In order to solve at least one of the above technical problems, the technical solution adopted in this application is:

[0005] An intelligent prediction method for crystal pulling completion, comprising the following steps:

[0006] Collect historical data and pre-process it;

[0007] Get the actual data;

[0008] A prediction model is established based on fitting the preprocessed data with the actual data.

[0009] Furthermore, the collecting of historical data and preprocessing thereof include:

[0010] Filter out the end sections of crystal ingots of the same specification from the data recorded on the furnace;

[0011] Obtain the parameter data required for the final stage of each crystal ingot;

[0012] The parameter data is calculated to obtain the required diameter of the corresponding crystal rod tail section.

[0013] Furthermore, the parameter data obtained in the finishing section mainly include the height of the finishing section, the total weight of silicon material used to draw the finishing section, the total time used for the finishing, and the diameter size corresponding to each moment.

[0014] Furthermore, the parameter data required for the final stage of each crystal rod is obtained, specifically:

[0015] First, obtain the parameter data of the entire time period involved in the final stage of each crystal ingot;

[0016] Selecting parameter data corresponding to a preset time period from the parameter data in the full time period;

[0017] Then, the parameter data within the preset time period is cleaned to obtain the diameter data of the tail section of the crystal rod when the preset time period ends;

[0018] Among them, the preset time period is the first 200s of the final stage of drawing;

[0019] Clean the parameter data, including removing duplicate values, missing values ​​and outliers in all parameters in the data.

[0020] Furthermore, the parameter data is calculated using the formula:

[0021]

[0022] Calculate it, where

[0023] ΔM is the weight of the end section of the pulled crystal rod within the preset time;

[0024] ΔV is the volume of the end section of the pulled crystal rod within the preset time;

[0025] ΔL is the length of the end section of the pulled crystal rod within the preset time;

[0026] ρ is the density of monovalent silicon, 2.4 g / cm 3 ;

[0027] d1 is the diameter of the end section of the pulled crystal rod at the start of the preset time;

[0028] d2 is the diameter of the end section of the pullable crystal rod at the end of the preset time.

[0029] Furthermore, the actual data obtained are measured parameters obtained after measuring the end section of the crystal rod, specifically:

[0030] Obtain the end section of the ingot and measure its height;

[0031] Based on the measured height, a number of points are equally obtained along the generatrix of the contour;

[0032] The diameter size corresponding to each point is obtained, which is the measured parameter.

[0033] The end section of each crystal rod is divided into at least 10-25 points;

[0034] The sample size of the collected crystal rods is at least 200;

[0035] The crystal rods collected when obtaining actual data are the same as the crystal rods collected when collecting historical data.

[0036] Furthermore, the prediction model is established based on fitting the preprocessed data with the actual data, specifically:

[0037] Based on the pre-processed data, simulation is performed to obtain a prediction model for the diameter curve of each crystal rod end section;

[0038] Based on the actual data, the simulation is performed to obtain the actual model of the diameter curve of each crystal rod end section;

[0039] Classify the data in the prediction model;

[0040] Then train the prediction model on the divided data set;

[0041] The prediction model is continuously modified to finally obtain a prediction model that meets the composite requirements.

[0042] Furthermore, the data in the prediction model is divided into a training set and a test set;

[0043] Compare some data in the actual model with the data in the test set to obtain a diameter deviation set;

[0044] Continuously adjust other data in the actual input parameters to the test set and obtain the weight coefficient of the diameter until the deviation between the predicted diameter and the actual diameter is minimized.

[0045] Furthermore, based on the features inputted in the training set, the diameter error with the actual model is calculated by comparing it with other data in the actual model;

[0046] Based on the trained structure, the model parameters are updated again until the gap between the output diameter and the fitted diameter is small enough, and then the training is stopped.

[0047] Furthermore, when trimming the prediction model, when the error value exceeds the threshold, it is necessary to adjust the prediction model and retrain it, or to improve the preprocessing solution, or to expand the historical data again; when the error values ​​of all prediction models are less than the threshold, it means that the final prediction model can be applied to actual production.

[0048] An intelligent prediction method for crystal pulling ending designed in the present application can automatically simulate the ending section of crystal pulling and perform denoising judgment on it to obtain a ending process that matches the existing crystal pulling process, and intelligently predict the ending model of the pulled crystal rod, thereby automatically completing the pulling of the ending section in the crystal rod and obtaining the diameter size of the ending section in real time. It is highly effective and accurate, reduces the risk of raw material waste and work time loss due to human factors, fully realizes the automated control of the pulling of the ending section, and improves work efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a flow chart of a method for intelligently predicting the end of crystal pulling according to an embodiment of the present application;

[0050] Figure 2 It is a schematic diagram of the truncated cone-shaped closing section corresponding to the preset time of this application;

[0051] Figure 3 It is a flowchart of the prediction model processing of one embodiment of the present application. DETAILED DESCRIPTION

[0052] The present application is described in detail below with reference to the accompanying drawings and specific embodiments.

[0053] This embodiment proposes an intelligent prediction method for crystal pulling end, such as Figure 1-2 As shown, the steps include:

[0054] First collect historical data and pre-process it, that is, first filter it from the system data that has been drawn, and then pre-process the filtered data.

[0055] Then, actual data is obtained, that is, actual measurement is performed based on the end section of the pulled solid crystal rod to obtain actual data.

[0056] A prediction model is established based on fitting the preprocessed data with the actual data.

[0057] Specifically, historical data is collected and preprocessed, including: selecting the end of the crystal rod of the same specification from the data recorded in the furnace, that is, when selecting the data, all the crystal rods of the same model are selected, that is, all are N-type and have the same diameter and the same specification. Such data is not only convenient for statistics, but also convenient for traceability, so that it can be comparable. These data are collected based on the data records of the finished crystal rods that have been pulled in all single crystal furnaces, and are not limited to the data of a single furnace. For multiple furnaces of the same or multiple batches, as long as they pull crystal rods of the same specification, they can be used as the source of historical data.

[0058] Obtain the parameter data required for the final stage of each crystal rod, and derive the corresponding data from the furnace data. These parameter data mainly include the height of the final stage, the total weight of silicon material used to pull the final stage, the total time used for the final stage, and the diameter size corresponding to each moment.

[0059] The parameter data required in the final stage of each crystal ingot are as follows:

[0060] First, obtain the parameter data of the entire time period involved in the tail section of each crystal rod. Since the length of the tail section of each crystal rod is different, its pulling time, overall crystal length, and corresponding weight are not exactly the same, so several groups of historical data of crystal rods are needed as source data.

[0061] Then select the parameter data corresponding to the preset time period from the parameter data in the full time period, that is, select the data that meets the standard parameter requirements from these source data. The general outline of the end section is an inverted cone structure. In order to ensure the accuracy and reproducibility of the calculation simulation, the data within the preset time is selected and the structure corresponding to the preset time is a flat truncated cone structure, such as Figure 2 As shown. The preset time period is the first 200s of the drawing end section. If the selected time exceeds the preset time of 200s, the shape of the selected end section will be irregular, deviating from the appearance of the end section being a truncated cone structure. If the selected time is less than the preset time of 200s, ΔM is greatly affected by the weighing accuracy and the measurement is inaccurate.

[0062] The parameter data within the preset time period is then cleaned to obtain the diameter data of the tail section of the crystal rod at the end of the preset time period. The cleaning data includes removing duplicate values, missing values ​​and abnormal values ​​in all parameters in the data. This is because in the collected data, there are many duplicate data, or data with missing parameters, or abnormal data that obviously exceeds the standard size range. Such parameter data must be eliminated to avoid too much impact on the final prefabricated model.

[0063] Based on the historical parameter data after cleaning, the parameter data is calculated to obtain the estimated diameter value required in the corresponding crystal rod tail section.

[0064] Among them, the parameter data is calculated using the formula:

[0065]

[0066] Calculate it, where

[0067] ΔM is the weight of the end section of the pulled crystal rod within the preset time;

[0068] ΔV is the volume of the end section of the pulled crystal rod within the preset time;

[0069] ΔL is the length of the end section of the pulled crystal rod within the preset time;

[0070] ρ is the density of monovalent silicon, 2.4 g / cm 3 ;

[0071] d1 is the diameter of the end section of the pulled crystal rod at the start of the preset time;

[0072] d2 is the diameter of the end section of the pullable crystal rod at the end of the preset time.

[0073] Based on the above formula and the collected historical parameter data, the parameter data is calculated to obtain the estimated diameter required in the corresponding crystal rod tail section, that is, the diameter size at all times corresponding to the preset time being the first 200 seconds of the tail section is calculated.

[0074] Furthermore, the actual data obtained are the measured parameters obtained after measuring the end section of the crystal rod, specifically:

[0075] First, obtain the end sections of the crystal rods in the workshop and measure them to obtain the total height of the end sections. The crystal rods collected when measuring the end sections of the crystal rods are one-to-one corresponding to the crystal rods collected when collecting historical data, that is, they are all crystal rods of the same type and equal diameter, and the sample size of the collected crystal rods is at least 200.

[0076] Based on the measured height, several points are equally divided along its contour generatrix, and the end section of each crystal rod is equally divided into at least 10-25 points, so that the data obtained is referenceable.

[0077] Then, at each point, a projection line is drawn along its periphery to obtain the diameter size corresponding to the point, and then the parameters of several measured diameter sizes in each finishing section can be obtained, and all these measured parameters are true values.

[0078] like Figure 3 As shown, a prediction model is established based on the fitting of the preprocessed data and the actual data, specifically:

[0079] Based on the preprocessed data, simulation is performed to obtain a prediction model for the diameter curve of the tail section of each crystal rod. Since the preprocessed data includes the diameter size corresponding to each moment in the first 200 seconds of the tail section and the crystal length size, the diameter fitting curve corresponding to different crystal length sizes can be obtained through data processing.

[0080] Then, based on the actual data, the diameter sizes corresponding to several points of the tail section are obtained in the actual data. Based on the same processing method, the diameter fitting curve corresponding to the tail section of each crystal rod for different crystal lengths is simulated, which is the actual model.

[0081] Then the data in the prediction model is classified and divided into training set and test set. The test set is used for interactive comparison with the fitting curve in the actual model to gradually correct the prediction model in the simulation. The training set is used as a training model, and the output is compared with the threshold value many times, and the training is repeated to finally obtain a prediction model that meets the composite requirements.

[0082] Specifically, part of the data in the actual model is compared with the data in the test set to obtain the diameter deviation set; and other data in the input actual parameters are continuously adjusted to the test set, and the weight coefficient of the diameter is obtained until the deviation between the predicted diameter and the actual diameter is minimized.

[0083] Then the prediction model is trained on the divided data set, and the prediction model is continuously adjusted to finally obtain a prediction model that meets the composite requirements. That is, based on the characteristics of the training set input, it is compared with other data in the actual model to calculate the diameter error with the actual model. At the same time, based on the training structure, the model parameters are updated again until the difference between the output diameter and the fitted diameter is small enough, and then the training is stopped.

[0084] Furthermore, when trimming the prediction model, when the error value exceeds the threshold, where the threshold range is 5-10mm, it is necessary to adjust the prediction model for retraining, or to improve the preprocessing scheme, or to expand the historical data again. The error refers to the difference between the predicted value and the true value. When the error is large, it means that the performance of the prediction model is poor and is not enough to deduce the diameter trend. First, the amount of data should be expanded, and more sticks should be fed with more data to at least cover the regular closing situation. When the amount of data is sufficient, the prediction effect is still poor, then start from two aspects. One is to replace the prediction model or adjust the model structure so that it can better analyze the laws of weight, length and diameter changes and adapt to the prediction task. The second is data preprocessing, such as digital filtering of the weight sequence to make the input data fluctuate less.

[0085] When the error values ​​of all prediction models are less than the threshold, it means that the final prediction model can be applied to actual production.

[0086] An intelligent prediction method for crystal pulling ending designed in the present application can automatically simulate the ending section of crystal pulling and perform denoising judgment on it to obtain a ending process that matches the existing crystal pulling process, and intelligently predict the ending model of the pulled crystal rod, thereby automatically completing the pulling of the ending section in the crystal rod and obtaining the diameter size of the ending section in real time. It is highly effective and accurate, reduces the risk of raw material waste and work time loss due to human factors, fully realizes the automated control of the pulling of the ending section, and improves work efficiency.

[0087] The above detailed description of the embodiments of the present application is only a preferred embodiment of the present application and cannot be considered to limit the scope of implementation of the present application. All equivalent changes and improvements made within the scope of the present application should still fall within the scope of the patent coverage of the present application.

Claims

1. An intelligent prediction method for crystal pulling, characterized in that the steps include: Collect historical data and pre-process it; Get the actual data; A prediction model is established based on fitting the preprocessed data with the actual data.

2. The intelligent prediction method for crystal pulling completion according to claim 1, characterized in that: The collecting of historical data and preprocessing thereof include: Filter out the end sections of crystal ingots of the same specification from the data recorded on the furnace; Obtain the parameter data required for the final stage of each crystal ingot; The parameter data is calculated to obtain the required diameter of the corresponding crystal rod tail section.

3. The intelligent prediction method for crystal pulling completion according to claim 2, characterized in that: The parameter data obtained in the finishing section mainly include the height of the finishing section, the total weight of silicon materials used to draw the finishing section, the total time used for the finishing, and the diameter size corresponding to each moment.

4. An intelligent prediction method for crystal pulling completion according to claim 2 or 3, characterized in that: The parameter data required for the final stage of each crystal ingot is obtained as follows: First, obtain the parameter data of the entire time period involved in the final stage of each crystal ingot; Selecting parameter data corresponding to a preset time period from the parameter data in the full time period; Then, the parameter data within the preset time period is cleaned to obtain the diameter data of the tail section of the crystal rod when the preset time period ends; Among them, the preset time period is the first 200s of the final stage of drawing; Clean the parameter data, including removing duplicate values, missing values ​​and outliers in all parameters in the data.

5. The intelligent prediction method for crystal pulling completion according to claim 4, characterized in that: The parameter data is calculated using the formula: Calculate it, where ΔM is the weight of the end section of the pulled crystal rod within the preset time; ΔV is the volume of the end section of the pulled crystal rod within the preset time; ΔL is the length of the end section of the pulled crystal rod within the preset time; ρ is the density of monovalent silicon, 2.4 g / cm 3 ; d1 is the diameter of the end section of the pulled crystal rod at the start of the preset time; d2 is the diameter of the end section of the pullable crystal rod at the end of the preset time.

6. The intelligent prediction method for crystal pulling completion according to claim 1, characterized in that: The actual data obtained are the measured parameters obtained after measuring the end section of the crystal rod, specifically: Obtain the end section of the ingot and measure its height; Based on the measured height, a number of points are equally obtained along the generatrix of the contour; The diameter size corresponding to each point is obtained, which is the measured parameter. The end section of each crystal rod is divided into at least 10-25 points; The sample size of the collected crystal rods is at least 200; The crystal rods collected when obtaining actual data are the same as the crystal rods collected when collecting historical data.

7. An intelligent prediction method for crystal pulling completion according to any one of claims 1-3, 5-6, characterized in that: The establishment of a prediction model based on fitting the preprocessed data with the actual data is specifically as follows: simulation is performed based on the preprocessed data to obtain a prediction model for the diameter curve of each crystal rod tail section; Based on the actual data, the simulation is performed to obtain the actual model of the diameter curve of each crystal rod end section; Classify the data in the prediction model; Then train the prediction model on the divided data set; The prediction model is continuously modified to finally obtain a prediction model that meets the composite requirements.

8. The intelligent prediction method for crystal pulling completion according to claim 7, characterized in that: The data in the prediction model is divided into a training set and a test set; Compare part of the data in the actual model with the data in the test set to obtain the diameter deviation set; continuously adjust other data in the input actual parameters to the test set, and obtain the diameter weight coefficient until the deviation between the predicted diameter and the actual diameter is minimized.

9. The intelligent prediction method for crystal pulling completion according to claim 8, characterized in that: Based on the features input from the training set, the diameter error with the actual model is calculated by comparing it with other data in the actual model; Based on the trained structure, the model parameters are updated again until the gap between the output diameter and the fitted diameter is small enough, and then the training is stopped.

10. An intelligent prediction method for crystal pulling completion according to claim 8 or 9, characterized in that: When trimming the prediction model, when the error value exceeds the threshold, it is necessary to adjust the prediction model and retrain it, or to improve the preprocessing plan, or to expand the historical data again; when the error values ​​of all prediction models are less than the threshold, it means that the final prediction model can be applied to actual production.

Images