Method and device for controlling atomization in reduction furnace and method and device for training target model

Abstract

The invention discloses an atomization control method and device in a reduction furnace and a training method and device of a target model, and relates to the technical field of new energy. The atomization control method in the reduction furnace comprises the steps that real-time parameters in a target reduction furnace are obtained, the real-time parameters comprise at least one of N first-class real-time parameters in one-to-one correspondence with N different moments in a current polycrystalline silicon growth cycle and second-class real-time parameters at the current moment, and N is a positive integer; a target model corresponding to the real-time parameters is utilized to predict the real-time parameters, a first prediction result is obtained, and the target model comprises at least one of a target atomization trend prediction model and a target atomization state prediction model; and determining a target control strategy according to the first prediction result, and executing the target control strategy. According to the embodiment of the invention, the method can improve the quality of the polycrystalline silicon in the process of preparing the polycrystalline silicon.

Description

Technical Field

[0001] The present application belongs to the field of new energy technology, and specifically relates to a method and device for controlling atomization in a reduction furnace, and a method and device for training a target model. Background Art

[0002] In the process of preparing polysilicon by the Siemens method, atomization is a common challenge. A large amount of fine silicon powder is generated during the reaction. These fine silicon powders not only affect the purity and performance of the product, but may also have a negative impact on production equipment and process stability.

[0003] In the prior art, the atomization condition is usually judged manually. Manual judgment has a lag, that is, when the abnormality is judged, atomization often occurs in the furnace, and low-quality polysilicon has been deposited on the silicon rods. Although normal production can be restored through timely adjustments, some low-quality polysilicon has already been deposited, which will seriously affect the subsequent silicon participation reaction, and ultimately lead to a decrease in the quality of this batch of polysilicon. That is, the quality of polysilicon in the process of preparing polysilicon in the prior art needs to be improved. Summary of the invention

[0004] The technical problem to be solved by the present application is to provide a method and device for controlling atomization in a reduction furnace, and a method and device for training a target model in view of the above-mentioned deficiencies in the prior art. By using the method for controlling atomization in a reduction furnace, the occurrence of atomization can be predicted in advance or the atomization state can be determined in time, thereby improving the quality of polysilicon in the process of preparing polysilicon.

[0005] In a first aspect, an embodiment of the present application provides a method for controlling atomization in a reduction furnace, comprising:

[0006] Acquire real-time parameters in the target reduction furnace, the real-time parameters including N first-category real-time parameters corresponding to N different moments in the current polysilicon growth cycle, and at least one of the second-category real-time parameters at the current moment, where N is a positive integer;

[0007] Predicting the real-time parameter using a target model corresponding to the real-time parameter to obtain a first prediction result, wherein the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model;

[0008] According to the first prediction result, a target control strategy is determined and executed, and the target control strategy is used to reduce the atomization degree in the target reduction furnace.

[0009] Based on the same inventive concept, in a second aspect, an embodiment of the present application further provides a method for training a target model, wherein the target model adopts any one of the target models in the first aspect, and the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model, and the method includes:

[0010] Obtaining an initial polynomial order and a first training set, wherein the first training set includes M first training samples, each of the M first training samples including N first-class historical parameters corresponding to N different moments in a historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-class historical parameters; N, M, and K are all positive integers;

[0011] Using M first training samples and initial polynomial orders for training, a target atomization trend prediction model is obtained;

[0012] and / or, obtaining a second training set, the second training set comprising K second training samples, each of the K second training samples comprising a second type of historical parameter at a historical moment, and a historical atomization state corresponding to the second type of historical parameter;

[0013] The initial atomization state prediction model is trained using K second training samples to obtain a target atomization state prediction model. The initial atomization state prediction model is trained using a random forest ensemble learning algorithm.

[0014] Based on the same inventive concept, in a third aspect, an embodiment of the present application provides an atomization control device in a reduction furnace, comprising:

[0015] A first acquisition module is used to acquire real-time parameters in a target reduction furnace, wherein the real-time parameters include N first-category real-time parameters corresponding to N different moments in a current polysilicon growth cycle, and at least one of the second-category real-time parameters at the current moment, where N is a positive integer;

[0016] A first prediction module, connected to the first acquisition module, is used to predict the real-time parameter using a target model corresponding to the real-time parameter to obtain a first prediction result, wherein the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model;

[0017] The first determination module is connected to the first prediction module and is used to determine a target control strategy according to the first prediction result, and the target control strategy is used to reduce the atomization degree in the target reduction furnace.

[0018] Based on the same inventive concept, an embodiment of the present application further provides a target model training device, wherein the target model adopts any target model of the first aspect, and the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model, and the device includes:

[0019] A second acquisition module is used to acquire an initial polynomial order and a first training set, wherein the first training set includes M first training samples, each of the M first training samples includes N first-category historical parameters corresponding to N different moments in a historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-category historical parameters; N, M and K are all positive integers;

[0020] A first training module, connected to the second acquisition module, is used to perform training using M first training samples and an initial polynomial order to obtain a target atomization trend prediction model;

[0021] and / or,

[0022] A second acquisition module is used to acquire a second training set, where the second training set includes K second training samples, and each of the K second training samples includes a second type of historical parameter at a historical moment, and a historical atomization state corresponding to the second type of historical parameter;

[0023] The first training module is connected to the second acquisition module and is used to train the initial atomization state prediction model using K second training samples to obtain a target atomization state prediction model. The initial atomization state prediction model is obtained by training with a random forest ensemble learning algorithm.

[0024] According to the atomization control method and device in the reduction furnace and the training method and device of the target model provided in the embodiments of the present application, the real-time parameters are predicted by the target model corresponding to the real-time parameters, and the first prediction result can be automatically predicted without human intervention. The atomization situation in the reaction process in the polysilicon reduction furnace can be comprehensively monitored and predicted in time, and then according to the first prediction result, the target control strategy can be quickly determined, so that by executing the target control strategy, the potential atomization risk can be dealt with in time, the atomization degree can be reduced, and the quality of polysilicon in the process of preparing polysilicon can be improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 A schematic flow chart of a method for controlling atomization in a reduction furnace provided in an embodiment of the present application is shown;

[0026] Figure 2 A schematic diagram showing the corresponding relationship between time and predicted atomization value provided in an embodiment of the present application;

[0027] Figure 3 A schematic diagram showing a flow chart of a method for training a target model provided in an embodiment of the present application;

[0028] Figure 4a and Figure 4b A decision tree diagram of key factors of a target atomization state prediction model provided by an embodiment of the present application is shown;

[0029] Figure 5a and Figure 5b Another key factor decision tree diagram of the target atomization state prediction model provided by an embodiment of the present application is shown;

[0030] Figure 6a and Figure 6b Another key factor decision tree diagram of the target atomization state prediction model provided by an embodiment of the present application is shown;

[0031] Figure 7 Another schematic flow chart of the atomization control method in the reduction furnace provided in the embodiment of the present application is shown;

[0032] Figure 8 A schematic diagram showing a structure of an atomization control device in a reduction furnace provided in an embodiment of the present application;

[0033] Fig. 9 A structural schematic diagram of a target model training device provided in an embodiment of the present application is shown. DETAILED DESCRIPTION

[0034] In order to enable those skilled in the art to better understand the technical solution of the present application, the present application is further described in detail below with reference to the accompanying drawings and embodiments.

[0035] The features and exemplary embodiments of various aspects of the present application will be described in detail below. In order to make the purpose, technical solutions and advantages of the present application clearer, the present application will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only configured to explain the present application and are not configured to limit the present application. For those skilled in the art, the present application can be implemented without the need for some of these specific details. The following description of the embodiments is only to provide a better understanding of the present application by illustrating the examples of the present application.

[0036] 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. In the absence of further restrictions, the elements defined by the statement "include..." do not exclude the presence of other identical elements in the process, method, article or device including the elements.

[0037] It should be understood that the term "and / or" used in this article is only a description of the association relationship of associated objects, indicating that there can be three relationships. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. In addition, the character " / " in this article generally indicates that the associated objects before and after are in an "or" relationship.

[0038] Example 1

[0039] The atomization control method in a reduction furnace provided in the embodiment of the present application can be executed by an atomization control device in a reduction furnace and electronic equipment, etc. The following is explained by taking the atomization control method in a reduction furnace executed by an electronic equipment as an example.

[0040] like Figure 1 As shown, the atomization control method in the reduction furnace provided in the embodiment of the present application may include steps S110 to S130.

[0041] S110, obtaining real-time parameters in the target reduction furnace, the real-time parameters including N first-category real-time parameters corresponding to N different moments in the current polysilicon growth cycle, and at least one of the second-category real-time parameters at the current moment, where N is a positive integer.

[0042] S120, predicting the real-time parameters using a target model corresponding to the real-time parameters to obtain a first prediction result, where the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model.

[0043] S130. Determine a target control strategy based on the first prediction result, and execute the target control strategy, where the target control strategy is used to reduce the atomization degree in the target reduction furnace.

[0044] According to the atomization control method in a reduction furnace provided in an embodiment of the present application, the real-time parameters are predicted by a target model corresponding to the real-time parameters, and the first prediction result can be automatically predicted without human intervention. The atomization situation in the reaction process in the polysilicon reduction furnace can be comprehensively monitored and predicted in a timely manner, and then according to the first prediction result, the target control strategy can be quickly determined, so that by executing the target control strategy, the potential atomization risk can be dealt with in a timely manner, the degree of atomization can be reduced, and the quality of polysilicon in the process of preparing polysilicon can be improved.

[0045] The specific implementation methods of the above steps are introduced below.

[0046] In step S110, the target reduction furnace can be any reduction furnace that needs to be atomized. The chassis layout of the target reduction furnace, as well as the number, height and caliber of the nozzles on the chassis layout can be set according to actual conditions and are not limited here.

[0047] As an example, the real-time parameters include first-category real-time parameters; as another example, the real-time parameters include second-category real-time parameters; as yet another example, the real-time parameters include first-category real-time parameters and second-category real-time parameters.

[0048] The first category of real-time parameters includes at least one of the first real-time flow rate of hydrogen, the first real-time flow rate of trichlorosilane, the first real-time value of multiphase current, and the first real-time value of total voltage. The first real-time value of multiphase current may be the actual value of multiphase current used to heat the silicon rod. As an example, the first category of real-time parameters includes the first real-time flow rate of hydrogen; as another example, the first category of real-time parameters includes the first real-time flow rate of trichlorosilane and the first real-time value of multiphase current; as another example, the first category of real-time parameters includes the first real-time flow rate of trichlorosilane, the first real-time value of multiphase current, and the first real-time value of total voltage; as another example, the first category of real-time parameters includes the first real-time flow rate of hydrogen, the first real-time flow rate of trichlorosilane, the first real-time value of multiphase current, and the first real-time value of total voltage.

[0049] The second category of real-time parameters includes at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace chassis water return real-time flow rate, the first exhaust gas real-time temperature, the second exhaust gas real-time temperature, the first total power, the second multi-phase current real-time value, and the second total voltage real-time value.

[0050] The first tail gas real-time temperature may be the real-time tail gas temperature after the jacket water heat exchange. The second tail gas real-time temperature may be the real-time tail gas temperature in the furnace after the existing control system estimates. The second multi-phase current real-time value may be the actual value of the multi-phase current used to heat the silicon rod.

[0051] As an example, the second category of real-time parameters includes the second hydrogen real-time flow rate; as another example, the second category of real-time parameters includes the second trichlorosilane real-time flow rate, the furnace chassis water return real-time flow rate and the first exhaust gas real-time temperature; as yet another example, the second category of real-time parameters includes the first total power, the second multi-phase current real-time value and the second total voltage real-time value; as yet another example, the second category of real-time parameters includes the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace chassis water return real-time flow rate, the first exhaust gas real-time temperature, the second exhaust gas real-time temperature, the first total power, the second multi-phase current real-time value and the second total voltage real-time value.

[0052] For example, the electronic device may pre-store the first category real-time parameters and the second category real-time parameters for subsequent direct call. Alternatively, the electronic device may select and download the first category real-time parameters and the second category real-time parameters in response to a user's selection and download operation based on the electronic device operation interface. The selection and download operation may be an operation capable of selecting and downloading the first category real-time parameters and the second category real-time parameters.

[0053] It should be noted that the value of N can be set according to actual conditions and is not limited here. For example, N can be 2, 3, 5, 10, etc.

[0054] In some implementations, before predicting the real-time parameter using the target model corresponding to the real-time parameter to obtain the first prediction result, the method further includes:

[0055] Preprocessing the real-time parameters to obtain preprocessed real-time parameters, the preprocessing including at least one of supplementation processing, removal processing and outlier processing;

[0056] The real-time parameter is predicted using a target model corresponding to the real-time parameter to obtain a first prediction result, including:

[0057] The preprocessed real-time parameters are predicted using the target model corresponding to the preprocessed real-time parameters to obtain a first prediction result.

[0058] In this embodiment, by performing at least one of supplementary processing, removal processing and outlier processing on the real-time parameters, since the preprocessed real-time parameters are more accurate than the real-time parameters, the preprocessed real-time parameters are predicted using the target model corresponding to the preprocessed real-time parameters to obtain a first prediction result, which can improve the accuracy of the first prediction result and thereby improve the accuracy of atomization control in the target reduction furnace.

[0059] As one example, the preprocessing includes supplementation processing; as another example, the preprocessing includes removal processing and outlier processing; as yet another example, the preprocessing includes supplementation processing, removal processing, and outlier processing.

[0060] In some examples, preprocessing includes supplementation processing, removal processing, and outlier processing.

[0061] Preprocess the real-time parameters to obtain preprocessed real-time parameters, including:

[0062] Using the first target software to supplement the missing data in the real-time parameters to obtain supplemented real-time parameters, the first target software uses the mean, median, mode or preset value to supplement the missing data in the real-time parameters;

[0063] Using the second target software to remove the data row where the missing value of the supplemented real-time parameter is located, to obtain the real-time parameter after removal;

[0064] The outliers of the real-time parameters after removal are removed by using the third target software, or the outliers of the real-time parameters are supplemented by using the mean, median, mode or preset value to obtain the preprocessed real-time parameters.

[0065] Exemplarily, the first target software may include Impute Missing Values ​​in the knime software; the second target software may include Filter Rows; and the third target software may include Statistics. That is, by using Impute Missing Values ​​in the knime software to supplement the missing data in the real-time data, the mean, median, mode or specific value may be used for supplementation; Filter Rows may be used to remove the data rows with missing values; and Statistics may be used to identify the outliers, that is, to remove the outliers, or to supplement them with the mean, median, mode or specific value (i.e., preset value).

[0066] It should be noted that the preset value can be set according to actual conditions and is not limited here.

[0067] In step S120, after acquiring the real-time parameters in the target reduction furnace, the electronic device may also predict the real-time parameters using the target model corresponding to the real-time parameters to obtain a first prediction result.

[0068] As an example, the target model includes a target atomization trend prediction model; as another example, the target model includes a target atomization state prediction model; as yet another example, the target model includes a target atomization trend prediction model and a target atomization state prediction model.

[0069] In some embodiments, the first type of real-time parameters includes at least one of a first real-time hydrogen flow rate, a first real-time trichlorosilane flow rate, a first real-time multiphase current value, and a first real-time total voltage value, the target model includes a target atomization trend prediction model, and the first prediction result includes a first predicted atomization trend;

[0070] The real-time parameter is predicted using a target model corresponding to the real-time parameter to obtain a first prediction result, including:

[0071] The target atomization trend prediction model is used to predict at least one of the first hydrogen real-time flow rate, the first trichlorosilane real-time flow rate, the first multi-phase current real-time value, and the first total voltage real-time value to obtain a first predicted atomization trend.

[0072] In this embodiment, the target atomization trend prediction model is used to predict at least one of the first hydrogen real-time flow rate, the first trichlorosilane real-time flow rate, the first multi-phase current real-time value and the first total voltage real-time value, so that the first predicted atomization trend can be obtained automatically, quickly and accurately.

[0073] In some examples, the target atomization trend prediction model includes:

[0074]

[0075] Among them, z is the atomization prediction value; w is the first independent variable; x is the first second independent variable; y is the first third independent variable; w i is the i+1th first independent variable; x i is the i+1th second independent variable; y i is the i+1th third independent variable; α i is the coefficient of the i-th first variable; β i is the coefficient of the i-th second variable; θ i is the i-th third variable coefficient; i is greater than or equal to 0 and less than or equal to 5.

[0076] Specifically, w1 is the second first independent variable, w2 is the third first independent variable, w3 is the fourth first independent variable, w4 is the fifth first independent variable, and w5 is the sixth first independent variable; x1 is the second second independent variable, x2 is the third second independent variable, x3 is the fourth second independent variable, x4 is the fifth second independent variable, and x5 is the sixth second independent variable; y1 is the second third independent variable, y2 is the third third independent variable, y3 is the fourth third independent variable, y4 is the fifth third independent variable, and y5 is the sixth third independent variable; α1 is the first The first variable coefficient, α2 is the second first variable coefficient, α3 is the third first variable coefficient, α4 is the fourth first variable coefficient, and α5 is the fifth first variable coefficient; β1 is the first second variable coefficient, β2 is the second second variable coefficient, β3 is the third second variable coefficient, β4 is the fourth second variable coefficient, and β5 is the fifth second variable coefficient; θ1 is the first third variable coefficient, θ2 is the second third variable coefficient, θ3 is the third third variable coefficient, θ4 is the fourth third variable coefficient, and θ5 is the fifth third variable coefficient.

[0077] For example, the value ranges of α1 to α5, β1 to β5, and θ1 to θ5 can all be [-1000, 1000]. The value of each variable coefficient can be determined based on experience.

[0078] As an example, the first independent variable may be the first real-time flow rate of hydrogen, the second independent variable may be the first real-time flow rate of trichlorosilane, and the third independent variable may be the first real-time value of multiphase current. As another example, the first independent variable may be the first real-time flow rate of trichlorosilane, the second independent variable may be the first real-time value of multiphase current, and the third independent variable may be the first real-time value of total voltage.

[0079] Of course, the above target atomization trend prediction model is only used for example and is used to limit the present application. The target atomization trend prediction model provided in the embodiment of the present application can be a model that can predict the first predicted atomization trend according to the first type of real-time parameters.

[0080] The first predicted atomization trend may include an upward trend and a downward trend, etc.

[0081] It should be noted that the output of the target atomization trend prediction model may be a first predicted atomization trend drawn according to N predicted atomization values ​​corresponding one to one at N different moments in the current polysilicon growth cycle.

[0082] In other embodiments, the second type of real-time parameters includes at least one of a second hydrogen real-time flow rate, a second trichlorosilane real-time flow rate, a furnace bottom pan water return water real-time flow rate, a first tail gas real-time temperature, a second tail gas real-time temperature, a first total power, a second multiphase current real-time value, and a second total voltage real-time value, the target model includes a target atomization state prediction model, and the first prediction result includes a first predicted atomization state;

[0083] The real-time parameter is predicted using a target model corresponding to the real-time parameter to obtain a first prediction result, including:

[0084] The target atomization state prediction model is used to predict at least one of the second hydrogen real-time flow rate, the second chlorosilane real-time flow rate, the furnace chassis water return real-time flow rate, the first exhaust gas real-time temperature, the second exhaust gas real-time temperature, the first total power, the second multi-phase current real-time value, and the second total voltage real-time value to obtain a first predicted atomization state.

[0085] The first predicted atomization state may include a clear atomization state, a normal atomization state, a light atomization state, a moderate atomization state, a heavy atomization state, and the like.

[0086] The target atomization state prediction model provided in the embodiment of the present application may be a model capable of determining a first predicted atomization state according to the second type of real-time parameters.

[0087] In step S130, after the electronic device predicts the real-time parameters using the target model corresponding to the real-time parameters and obtains a first prediction result, it can also determine a target control strategy based on the first prediction result and execute the target control strategy, which is used to reduce the atomization degree in the target reduction furnace.

[0088] In some embodiments, the first prediction result includes a first predicted atomization trend, and determining a target control strategy according to the first prediction result includes:

[0089] In the case where the first predicted atomization trend includes an upward trend, determining the target control strategy to be a first type of control strategy;

[0090] Among them, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate, increasing the first hydrogen real-time flow rate, reducing the first trichlorosilane real-time flow rate, reducing the temperature of the target area, reducing the increase value of the first multi-phase current real-time value, and increasing at least one of the voltage baseline.

[0091] The target area may be an area where multiple predicted atomization values ​​are greater than the preset atomization value, that is, an area with a high atomization tendency. Figure 2 As shown, the atomization trend in the target area A gradually increases, that is, the area with a high trend, or the area where the predicted atomization value is greater than 10. The preset atomization value can be set according to the actual situation and is not limited here.

[0092] As an example, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate; as another example, the first type of control strategy includes lowering the temperature of the target area and reducing the increase value of the first multi-phase current real-time value; as yet another example, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate, increasing the first hydrogen real-time flow rate, reducing the first trichlorosilane real-time flow rate, lowering the temperature of the target area, reducing the increase value of the first multi-phase current real-time value, and increasing the voltage baseline.

[0093] It should be noted that, in this embodiment, the specific adjustment range can be set according to actual conditions and is not limited here.

[0094] It is understandable that, in the case where the first predicted atomization trend includes a downward trend, no operation may be performed. After determining that the target control strategy is the first type of control strategy, the first type of control strategy may be executed.

[0095] In some other embodiments, the first prediction result includes a first predicted atomization state, and determining a target control strategy according to the first prediction result includes:

[0096] In a case where the first predicted atomization state includes one of a light atomization state, a moderate atomization state, and a heavy atomization state, determining the target control strategy to be a second type of control strategy;

[0097] Among them, the second type of control strategy includes regulating at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace bottom pan water return real-time flow rate, the second multi-phase current real-time value and the second total voltage real-time value.

[0098] It should be noted that, in this embodiment, the specific control range can be set according to actual conditions and is not limited here.

[0099] Exemplarily, the second type of control strategy may include at least one of increasing the second hydrogen real-time flow rate, decreasing the second trichlorosilane real-time flow rate, increasing the furnace bottom pan water return real-time flow rate, decreasing the second multi-phase current real-time value, and decreasing the second total voltage real-time value.

[0100] More specifically, as an example, the second type of control strategy may include increasing the second hydrogen real-time flow rate; as another example, the second type of control strategy may include decreasing the second trichlorosilane real-time flow rate and increasing the furnace bottom pan water return real-time flow rate; as yet another example, the second type of control strategy may include increasing the second hydrogen real-time flow rate, decreasing the second trichlorosilane real-time flow rate, increasing the furnace bottom pan water return real-time flow rate, decreasing the second multi-phase current real-time value and decreasing the second total voltage real-time value.

[0101] It can be understood that, in the case where the first predicted atomization state includes a clear atomization state or a normal atomization state, no operation may be performed. In the case where the target control strategy is determined to be the second type of control strategy, the second type of control strategy may be executed.

[0102] It should be noted that in the embodiment of the present application, before the polysilicon target reduction furnace is put into production and operation, the first type of real-time parameters can be predicted by the target atomization trend prediction model to obtain the first prediction result, and the target control strategy corresponding to the first prediction result can be executed to reduce the degree of atomization in the production and operation of the polysilicon target reduction furnace; during the production and operation of the polysilicon target reduction furnace, the second type of real-time parameters can be predicted by the target atomization state prediction model to obtain the first prediction result, and the target control strategy corresponding to the first prediction result can be executed to enable the control system to more accurately control the control parameters during the operation. Among them, the control parameters include but are not limited to at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the second multi-phase current rise value for heating the silicon rod, the second total voltage real-time value, and the furnace bottom pan water return water real-time flow rate.

[0103] Example 2

[0104] The target model training method provided in the embodiment of the present application can be executed by a target model training device and an electronic device. The following is an example of the target model training method being executed by an electronic device.

[0105] In the embodiment of the present application, the target model adopts any one of the target models in Embodiment 1, and the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model.

[0106] like Figure 3 As shown, the target model training method provided in the embodiment of the present application includes steps S310 to S320.

[0107] S310, obtaining an initial polynomial order and a first training set, and / or obtaining a second training set, wherein the first training set includes M first training samples, each of the M first training samples includes N first-category historical parameters corresponding to N different moments in a historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-category historical parameters; the second training set includes K second training samples, each of the K second training samples includes a second-category historical parameter at a historical moment, and a historical atomization state corresponding to the second-category historical parameter; N, M, and K are all positive integers;

[0108] S320, using M first training samples and initial polynomial order for training to obtain a target atomization trend prediction model, and / or using K second training samples to train an initial atomization state prediction model to obtain a target atomization state prediction model, the initial atomization state prediction model is trained by a random forest ensemble learning algorithm.

[0109] According to the target model training method provided in the embodiment of the present application, M first training samples and the initial polynomial order are used for training to obtain a target atomization trend prediction model, and / or, the initial atomization state prediction model is trained using K second training samples to obtain a target atomization state prediction model, thereby providing a basis for subsequent prediction of real-time parameters through the target model corresponding to the real-time parameters.

[0110] The specific implementation of the above steps is introduced below.

[0111] In step S310, at least some of the M first training samples in the first training set may be nozzles of different sizes and heights, and / or first training samples under different chassis layouts; at least some of the K second training samples in the second training set may be nozzles of different sizes and heights, and / or second training samples under different chassis layouts.

[0112] Each of the N first-category historical parameters includes at least one of a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multiphase current historical value, and a first total voltage historical value. The first multiphase current historical value may be a multiphase current historical value used to heat the silicon rod. As an example, the first-category historical parameter includes a first hydrogen historical flow rate; as another example, the first-category historical parameter includes a first hydrogen historical flow rate and a first trichlorosilane historical flow rate; as yet another example, the first-category historical parameter includes a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, and a first multiphase current historical value; as yet another example, the first-category historical parameter includes a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multiphase current historical value, and a first total voltage historical value.

[0113] The second category of historical parameters includes at least one of the second hydrogen historical flow rate, the second trichlorosilane historical flow rate, the furnace bottom pan water return historical flow rate, the first exhaust gas historical temperature, the second exhaust gas historical temperature, the second total power, the second multi-phase current historical value, and the second total voltage historical value.

[0114] The first exhaust gas historical temperature may be the historical exhaust gas temperature after jacket water heat exchange. The second exhaust gas historical temperature may be the furnace exhaust gas temperature after estimation by the existing control system. The second multi-phase current historical value may be the multi-phase current historical value used to heat the silicon rod.

[0115] As an example, the second category of historical parameters includes the second hydrogen historical flow rate; as another example, the second category of historical parameters includes the second hydrogen historical flow rate and the second trichlorosilane historical flow rate; as yet another example, the second category of historical parameters includes the second exhaust gas historical temperature, the second total power, the second multi-phase current historical value and the second total voltage historical value; as yet another example, the second category of historical parameters includes the second hydrogen historical flow rate, the second trichlorosilane historical flow rate, the furnace bottom pan water return water historical flow rate, the first exhaust gas historical temperature, the second exhaust gas historical temperature, the second total power, the second multi-phase current historical value and the second total voltage historical value.

[0116] Exemplarily, the initial polynomial order, the first training set and the second training set may be pre-stored in the electronic device for subsequent direct calling.

[0117] In step S320, in some embodiments, each of the N first-category historical parameters includes at least one of a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multiphase current historical value, and a first total voltage historical value.

[0118] The target atomization trend prediction model is obtained by training using the M first training samples and the initial polynomial order, including:

[0119] A1. For the first target training sample among the M first training samples, based on the initial polynomial order, N first-category historical parameters are used as input, and N historical atomization values ​​are used as output, to obtain an initial polynomial regression model by fitting;

[0120] A2. For the second target training samples other than the first target training samples in the M first training samples, an initial polynomial regression model is used to perform atomization prediction on the N first-category historical parameters to obtain N second predicted atomization values ​​corresponding to N different moments in the historical polysilicon growth cycle;

[0121] A3. For each second target training sample other than the first target training sample in the M first training samples, determine a determination coefficient according to the N historical atomization values ​​and the N second predicted atomization values;

[0122] A4. When the determination coefficient is less than the preset determination coefficient, adjust the initial polynomial regression model to obtain an intermediate polynomial regression model, and update the initial polynomial regression model to the intermediate polynomial regression model, and update the second target training sample to the third target training sample among the M first training samples excluding the first target training sample and the second target training sample, and repeat steps A2 to A3 until the iterative determination coefficient is greater than or equal to the preset determination coefficient, and determine the iterative initial polynomial regression model as the target atomization trend prediction model.

[0123] In this embodiment, the preset determination coefficient may be in the range of [-1, 1]. The preset determination coefficient may be set according to actual conditions and is not limited here. For example, the preset determination coefficient may be 0.513.

[0124] The closer the value of the determination coefficient is to 1, the stronger the explanatory power of the target atomization trend prediction model is, which can be used to directly reflect the accuracy and fit of the intermediate polynomial regression model in predicting continuous variables.

[0125] In some examples, using M first training samples and initial polynomial orders for training to obtain a target atomization trend prediction model may also include: preprocessing the M first training samples to obtain M preprocessed first training samples, and using the M preprocessed first training samples and initial polynomial orders for training to obtain a target atomization trend prediction model. The preprocessing process is similar to the process of preprocessing the real-time parameters in Example 1, and will not be repeated here.

[0126] For example, when training is performed using the first training set, the selected variables and the initial polynomial order may be optimized and adjusted according to the second predicted atomization value.

[0127] Exemplarily, when training with the first training set, a portion of the first training samples can be selected from the first training set as a test set, the intermediate polynomial regression model is evaluated through the test set, and the model parameters of the intermediate polynomial regression model are iteratively updated to obtain a target atomization trend prediction model.

[0128] For example, the target atomization trend prediction model can be performed using a polynomial regression machine learning algorithm, which captures nonlinear relationships in the data by adding polynomial terms to the input variables. This method is very useful in predicting continuous value tasks, especially when the data exhibits nonlinear trends.

[0129] For example, before using training data to train a polynomial regression model for atomization trend prediction, the following steps need to be performed to configure the model and process the data:

[0130] 1. Handling missing values ​​(i.e. preprocessing): Use appropriate methods such as interpolation, deletion, or filling to handle missing values ​​in the dataset.

[0131] 2. Feature selection: Determine the input variables used for prediction, such as hydrogen flow, trichlorosilane flow, multiphase current value, total voltage real-time value, etc.

[0132] 3. Feature transformation: Although polynomial regression inherently involves nonlinear transformation of features, it may also be necessary to standardize or normalize the data to improve the training efficiency and performance of the model.

[0133] For the target atomization trend prediction model, the main parameters include 1) Variable Coeff., which is used to reflect the direction and intensity of the influence of the independent variable on the dependent variable. 2) Std.Err. is used to calculate the confidence interval and conduct hypothesis testing. 3) t-value, which is used to reflect whether the difference between the variable coefficient and zero is significant. It is the ratio of the variable coefficient to its standard error, that is, t = coefficient / standard error. 4) P>|t|, which is used to judge whether the variable coefficient is significant. In a two-tailed test, if the value of P>|t| is less than the preset significance level (such as 0.05), the null hypothesis is rejected and the coefficient is considered to be significantly not zero; if the value of P>|t| is greater than or equal to the preset significance level, the null hypothesis is not rejected and the coefficient is considered to be zero or not significant.

[0134] Exemplarily, the main steps of training the initial polynomial regression model using the first training set are as follows:

[0135] 1. Choose the order of the polynomial: Determine the order of the polynomial (such as quadratic, cubic, etc.), which determines the upper limit of the nonlinear complexity that the model can capture.

[0136] 2. Model fitting: Use the first training set (including hydrogen flow (i.e., the first hydrogen historical flow), trichlorosilane flow (i.e., the first trichlorosilane historical flow), multiphase current value (i.e., the first multiphase current historical value), total voltage real-time value (i.e., the first total voltage historical value), etc. as input, and atomization real-time value (i.e., historical atomization value) as output) to fit the initial polynomial regression model. The atomization real-time value can be obtained by data export through the existing control system.

[0137] 3. Parameter estimation: By calculating the coefficient of determination for the machine learning model for atomization trend prediction, the closer the R2 value is to 1, the stronger the explanatory power of the model. These indicators can directly reflect the accuracy and fit of the polynomial regression model in predicting continuous variables.

[0138] 4. Model prediction: Use the trained polynomial regression model to predict the new input data (i.e., the second target training samples other than the first target training samples in the M first training samples) to obtain the predicted value of the atomization trend (i.e., the second predicted atomization value).

[0139] 5. Result analysis: Evaluate the prediction results (a complete cycle of polysilicon growth atomization value prediction, corresponding to the time, the atomization trend curve can be drawn), including calculating the prediction error and drawing a comparison chart between the predicted value and the actual value. By performing trend analysis on the prediction results (such as trend line drawing), the change of atomization trend can be more clearly understood.

[0140] 6. Model optimization and adjustment: According to the test results, the initial polynomial regression model is optimized and adjusted, which may include: Adjusting the polynomial order: increasing or decreasing the order of the polynomial to find the balance between the best fit and overfitting; Feature engineering: trying different feature combinations or transformations to improve the predictive performance of the model, that is, first selecting three of the first category historical parameters for fitting. If the effect is not good, continue to add features, or reselect the other three first category historical parameters for fitting.

[0141] Specifically, an atomization trend prediction model was constructed with a 40-pair rod reduction furnace and a chassis layout of 0+0+12+8. The parameters such as Variable, Exponent, Coeff., Std. Err., t-value, and P>|t| are shown in Table 1.

[0142] Table 1

[0143]

[0144]

[0145] Among them, Table 1 shows the power of the function variables, coefficients, intercept and other parameters of the intermediate polynomial regression model, and R2 is the result obtained by measuring the original data and the predicted data. The values ​​in Table 1 are only used for example and are not used to limit this application.

[0146] In some examples, determining a determination coefficient based on N historical atomization values ​​and N second predicted atomization values ​​includes:

[0147] Determine the historical atomization average value according to N historical atomization values;

[0148] Substitute the historical atomization mean, N historical atomization values ​​and N second predicted atomization values ​​into formula (1) to calculate and obtain the determination coefficient;

[0149]

[0150] Among them, R 2 represents the coefficient of determination; y i represents the i-th historical atomization value; represents the i-th second predicted atomization value; Represents the historical fog average.

[0151] Exemplarily, the ratio of the sum of N historical atomization values ​​to N is determined as the historical atomization mean value.

[0152] For example, N historical atomization values ​​are [3, 4, 5, 6, 7]; N second predicted atomization values ​​are [2, 4, 6, 5, 7], then the historical atomization mean is =5, substitute the above data into formula (1) to calculate and get R 2 =0.7.

[0153] In other embodiments, the second type of historical parameters includes at least one of a second hydrogen historical flow rate, a second trichlorosilane historical flow rate, a furnace bottom pan water return historical flow rate, a first tail gas historical temperature, a second tail gas historical temperature, a second total power, a second multiphase current historical value, and a second total voltage historical value.

[0154] The initial atomization state prediction model is trained using K second training samples to obtain a target atomization state prediction model, including:

[0155] B1. For a fourth target training sample among the K second training samples, use the initial atomization state prediction model to predict the fourth target training sample to obtain a second predicted atomization state;

[0156] B2. determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample;

[0157] B3. When the correlation coefficient is less than the preset correlation coefficient, the fourth target training sample is updated to any second training sample among the K second training samples except the fourth target training sample, and steps B1 to B2 are repeated until the correlation coefficient after iteration is greater than or equal to the preset correlation coefficient, and the initial atomization state prediction model after iteration is determined as the target atomization state prediction model.

[0158] In this embodiment, the preset correlation coefficient can be set according to actual conditions and is not limited here. For example, the preset correlation coefficient can be 0.999.

[0159] In some examples, using K second training samples to train the initial atomization state prediction model to obtain a target atomization state prediction model may also include: preprocessing the K second training samples to obtain K preprocessed second training samples, and using the K preprocessed second training samples to train the initial atomization state prediction model to obtain a target atomization state prediction model. The preprocessing process is similar to the process of preprocessing the real-time parameters in Example 1, and will not be repeated here.

[0160] In some examples, determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample includes:

[0161] Using a preset correlation analysis method, a correlation analysis is performed on the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample to obtain a correlation coefficient;

[0162] Among them, the preset correlation analysis method includes at least one of the chart method, covariance method, correlation coefficient method, regression analysis method, information entropy and mutual information method, chi-square test method and correlation coefficient method.

[0163] Exemplarily, the correlation coefficient method includes the Pearson product moment correlation coefficient r, which is a dimensionless index ranging from -1.0 to 1. (including -1.0 and 1.0) that reflects the degree of linear correlation between two data sets. The function has two parameters, namely the independent variable set AVERAGE (array1) and the dependent variable set AVERAGE (array2); wherein the parameters can be numbers, or names, array constants or references containing numbers. The specific formula is as follows:

[0164]

[0165] Among them, x and y represent the test results; i and i represents the original data (i.e., the second training sample); and Represents the average value of the training samples in the training set, that is, is the sample average AVERAGE(array1), is the sample average AVERAGE(array2).

[0166] In some examples, before the initial atomization state prediction model is trained using the second training set, the model parameters of the target atomization state prediction model may be configured, and the model parameters may include the maximum depth of the tree and the branch parameters. More specifically, configuring the model parameters of the target atomization state prediction model may include: selecting the model parameters of the target atomization state prediction model, and the model parameters may include at least one of the operation time, atomization state, the ratio of hydrogen flow rate to trichlorosilane flow rate, atomization real-time value, furnace exhaust gas temperature, total power, total water power, total water power ratio, exhaust gas temperature change rate, voltage change rate, etc.; clearly predicting the target parameter as atomization state; selecting the maximum depth of the tree; the atomization real-time value is used as the first judgment parameter, and the rest are used as branch parameters.

[0167] In some examples, when training the initial atomization state prediction model, some second training samples can be selected from K second training samples as a test set (i.e., the fourth target training samples), and the initial atomization state prediction model is evaluated through the test set, and the model parameters of the initial atomization state prediction model are iteratively updated to obtain the target atomization state prediction model.

[0168] In some examples, the target atomization state prediction model is tested using a test set that has parameters consistent with the second training set, and a correlation analysis is performed between the test result (i.e., the second predicted atomization state) and its corresponding second training sample (i.e., the fourth target training sample). When the correlation coefficient between the test result and the second training sample reaches a preset correlation coefficient, the target atomization state prediction model is trained to reach the target training state.

[0169] In some examples, the model parameters of the target atomization state prediction model include: 1) n_estimators, which specifies the number of trees in the random forest. Increasing the number of trees can improve the stability and accuracy of the model, but too many trees will also increase the computational cost. 2) max_depth, which specifies the maximum depth of each decision tree, controls the growth of the tree, and limits the depth of the tree to prevent overfitting and improve the generalization ability of the model. 3) min_samples_split, the maximum depth of each decision tree limits the number of subtrees that continue to be divided. If the number of samples of a node is less than min_samples_split, it will not continue to try to select the optimal feature for division. This helps to avoid overfitting. 4) min_samples_leaf, the minimum number of samples required for leaf nodes, limits the minimum number of samples for leaf nodes. If the number of leaf nodes is less than the number of samples, it will be pruned together with the sibling nodes. This also helps to prevent overfitting. 5) max_features, the maximum number of features to be considered at each split. Increasing max_features generally improves the performance of the model because there are more features to choose from at each node. However, this will also reduce the diversity of individual trees, thus affecting the performance of random forests. Therefore, a proper balance is needed. 6) bootstrap, whether to use randomly sampled training data with replacement to build each tree. Using sampling with replacement can increase the randomness of the tree, thereby improving the generalization ability of the model.

[0170] In some examples, more specifically, the main steps of the training process of the target atomization state prediction model are as follows:

[0171] S1. Randomly select features and samples: During the training process, Random Forest will randomly select features and samples (i.e., the second training samples) to build multiple decision trees. Specifically, for each decision tree, a portion of samples will be randomly selected with replacement from the original data set (i.e., the second training set) as the training set, and a portion of features will be randomly selected from all features (i.e., the second type of historical parameters) to build the decision tree.

[0172] S2. Constructing a decision tree: For each trained sample set (i.e., the second training set) and feature set (including at least one of the second hydrogen historical flow rate, the second trichlorosilane historical flow rate, the furnace bottom tray water return historical flow rate, the first exhaust gas historical temperature, the second exhaust gas historical temperature, the second total power, the second multiphase current historical value, and the second total voltage historical value), a decision tree will be constructed. The decision tree construction process usually includes selecting the optimal splitting attribute and splitting point until the stopping condition is met (such as too few samples in the node or too small information gain).

[0173] S3. Model tuning, adjusting hyperparameters: During the training process, some hyperparameters may need to be adjusted, such as the number of decision trees, the number of features used in each decision tree, the depth of the decision tree, etc., to improve the performance and generalization ability of the model. When the correlation coefficient does not reach the expected target value, adjust the number of decision trees. For example, if the expected goal is achieved, but the usage rate of some feature parameters is not high, or the number of decision trees is too large, reduce the number and check again whether the result meets the expected goal. If it can be met, repeat the steps further until the model content is streamlined and the correlation coefficient meets the expected goal.

[0174] S4. Optimize model structure: In addition to adjusting hyperparameters, the performance of the model can also be improved by optimizing the model structure. For example, pruning can be used to reduce the complexity of the decision tree and avoid overfitting.

[0175] S5. Model evaluation uses a test set to evaluate the model: the above model is tested through a test set (i.e., the fourth target training sample) that is consistent with the training set parameter settings, and the test result (i.e., the second predicted atomization state) is compared and analyzed with the original data (i.e., the fourth target training sample). When the correlation coefficient between the test result and the original data reaches a preset value (i.e., the preset correlation coefficient), the training of the atomization state prediction machine learning model reaches the target training state.

[0176] Among them, the above steps S1 to S4 may be the training process of the initial atomization state prediction model.

[0177] Correlation analysis can adopt graphical method, covariance method, correlation coefficient method, regression analysis method, information entropy and mutual information method, chi-square test method, etc.; the embodiment of the present application can adopt the correlation coefficient method to measure the strength and direction of the linear relationship between variables by calculating the correlation coefficient.

[0178] In an embodiment of the present application, the atomization states before and after the prediction are collected, and the correlation coefficient is calculated using the Pearson correlation coefficient. When the correlation coefficient r≥0.999, it can be considered that the predicted atomization state (i.e., the second predicted atomization state) is positively correlated with the original atomization state data (i.e., the historical atomization state corresponding to the fourth target training sample), and the correlation is strong. The atomization state prediction model (i.e., the initial atomization state prediction model) has met the target training state.

[0179] Specifically, taking a 40-pair rod target reduction furnace as an example, a target atomization state prediction model with three different nozzles and chassis layouts was constructed. The key factors are as follows: Figures 4a to 6b shown.

[0180] Figure 4a and Figure 4bThe decision tree diagram of key factors when the atomization state is mild, moderate, and severe for the atomization state prediction model (i.e., the target atomization state prediction model) trained under the first nozzle and chassis layout mode (arranged in concentric circles, exhaust hole: inlet hole = 4:29) of 40 pairs of rod reduction furnaces is shown. Figure 4a and Figure 4b The double dashed line S1 indicates Figure 4a and Figure 4b , where the first key parameter is the real-time value of atomization (i.e. the predicted atomization value corresponding to the first predicted atomization state), followed by the total water power, the total water power ratio, and the real-time value of the exhaust gas temperature, the exhaust gas temperature change rate, etc.; a correlation analysis was performed on them, and the correlation coefficient r=0.99996989, which can be considered to meet the current atomization state prediction and achieve the target training state.

[0181] Figure 5a and Figure 5b The decision tree diagram of key factors when the atomization state is mild, moderate and severe for the atomization state prediction model (i.e., the target atomization state prediction model) trained under the first nozzle and chassis layout mode (arranged in a regular hexagonal form, exhaust hole: intake hole = 8:25) of 40 pairs of rod reduction furnaces is shown. Figure 5a and Figure 5b The double dashed line S2 indicates Figure 5a and Figure 5b The first key parameter is the exhaust temperature change rate, followed by the atomization real-time value, the voltage change rate, and then the total water power, the total water power ratio, the ratio, etc.; a correlation analysis was performed on them, and the correlation coefficient r=0.99998907, which can be considered to meet the current atomization state prediction and achieve the target training state.

[0182] Figure 6a and Figure 6b The decision tree diagram of key factors when the atomization state is mild, moderate, and severe for the atomization state prediction model (i.e., the target atomization state prediction model) trained under the first nozzle and chassis layout mode (arranged in concentric circles, exhaust hole: inlet hole = 4:20) of the 40-pair rod reduction furnace is shown in FIG. Figure 6a and Figure 6b The double dashed line S3 indicates Figure 6a and Figure 6b , where the first key parameter is the exhaust gas temperature change rate, followed by the real-time value of atomization, and then the running time, exhaust gas temperature change rate, total power, etc.; a correlation analysis was performed on them, and the correlation coefficient r=0.99996301, which can be considered to meet the current atomization state prediction and achieve the target training state.

[0183] Through the trained model, the atomization trend prediction machine learning model is used to predict the atomization trend of the initial operating parameters of different nozzles and chassis layouts at different times before the production and operation of the polysilicon reduction furnace; adjustments are made when the atomization trend rises rapidly to reduce the degree of atomization during operation;

[0184] During operation, an atomization state prediction machine learning model is configured in the control system of the existing polysilicon reduction furnace. After the initial parameter data determined above is put into use, the operation is started. The atomization state prediction model is used to predict the atomization state. Potential atomization risks can be discovered and dealt with in a timely manner, so that the control system can more accurately control the control parameters during operation, reduce the degree of atomization, and improve product quality and output.

[0185] In order to better understand the atomization control method in the reduction furnace provided in the embodiment of the present application, a more specific implementation method is described below.

[0186] like Figure 7 As shown, the atomization control method in the reduction furnace provided in the embodiment of the present application may include:

[0187] First, for the reduction furnace system used to produce polysilicon, an atomization prediction model based on the nozzle and chassis layout combined with multiple parameters is constructed; the reduction furnace control system is used to obtain the initial parameter setting information (i.e., the first type of historical parameters) and the operating data information during the patrol process (i.e., the second type of historical parameters); then, before the production and operation of the polysilicon reduction furnace, the atomization trend prediction machine learning model (i.e., the target atomization trend prediction model) is used to predict the atomization trend of the initial operating parameters at different nozzles and chassis layouts at different times, and it is found that the atomization trend is adjusted during the rapid rising stage (i.e., the first atomization trend includes an upward trend) to reduce the degree of atomization during operation; then, during the operation, an atomization state prediction machine learning model (i.e., the target atomization state prediction model) is configured in the existing polysilicon reduction furnace control system, and the operation is started after the above-mentioned determined initial parameter data (i.e., real-time parameters) are put into use, and the atomization state is predicted by the atomization state prediction model (i.e., the target atomization state prediction model), so that the control system can more accurately control the control parameters during the operation.

[0188] The nozzle and chassis layout used on different types of reduction furnaces such as 24 pairs of rods, 40 pairs of rods, and 60 pairs of rods, and the nozzle designs with different numbers, heights, and calibers on different chassis layouts, and the subsequent model parameters are collected and summarized based on the different characteristics of the internal parts of the polysilicon reduction furnace equipment;

[0189] The initial parameter setting information of the system is input into the control system of the existing polysilicon reduction furnace before production. The initial parameter setting information includes the size and height of the nozzle used, the chassis layout, the trichlorosilane flow setting value at different operating times, the hydrogen flow setting value, the multi-phase current setting value and the rise value for heating the silicon rod, the voltage setting value, etc.

[0190] Using the initial parameter setting information, before the polysilicon reduction furnace is operated based on the information, the atomization prediction model is used to predict the atomization trend during the batch operation, and before starting the operation, the initial parameter setting information is found to have an upward trend in atomization, and the atomization trend is re-optimized to reduce the possibility of severe atomization;

[0191] Furthermore, when the existing polysilicon reduction furnace control system is used for production, the operation data of the polysilicon reduction furnace is obtained, and the operation data information includes but is not limited to the size and height of the nozzle used, the chassis layout, the actual flow rate of trichlorosilane, the actual flow rate of hydrogen, the return water flow rate of the furnace chassis, the tail gas temperature after the jacket water heat exchange, the real-time tail gas temperature in the furnace after the existing control system estimates, the total power, and the actual value of the multi-phase current and the actual value of the voltage used to heat the silicon rods;

[0192] The atomization prediction model is used to generate the atomization state prediction information at the next moment for the above-mentioned polysilicon reduction furnace operation data, and the control parameters in the reduction furnace are adjusted by the existing polysilicon reduction furnace control system, so that the atomization state in the reduction furnace reaches a normal state;

[0193] The control parameters include, but are not limited to, hydrogen flow, trichlorosilane flow, multiphase current rise value for heating silicon rods, voltage value, and furnace drum water bottom tray water return flow.

[0194] Build a machine learning model for atomization trend prediction and a machine learning model for atomization status prediction, including:

[0195] Acquire initial and runtime parameter data corresponding to the polysilicon reduction furnace producing polysilicon, and prepare a training set based on the acquired corresponding parameter data, wherein the training data set includes parameter data information of multiple different batches under different nozzle and chassis layouts;

[0196] The training sets generated above are used to train the atomization trend prediction machine learning model and the atomization state prediction machine learning model respectively, so as to generate an atomization prediction model after reaching the target training state;

[0197] During operation, the atomization trend prediction machine learning model is used to generate the overall atomization trend before operation. At the same time, the atomization state prediction machine model is used to generate the atomization state prediction information during operation. Combined with the existing control system, the next moment hydrogen, trichlorosilane set value, multi-phase current rise value, and voltage set value are output.

[0198] Before using the training data to train the polynomial regression model for fogging trend prediction, you need to perform the following steps to configure the model and process the data:

[0199] 1. Handling missing values: Use appropriate methods such as interpolation, deletion, or filling to handle missing values ​​in the dataset.

[0200] 2. Feature selection: Determine the input variables used for prediction, such as hydrogen flow, trichlorosilane flow, multiphase current value, total voltage real-time value, etc.

[0201] 3. Feature transformation: Although polynomial regression inherently involves nonlinear transformation of features, it may also be necessary to standardize or normalize the data to improve the training efficiency and performance of the model.

[0202] Furthermore, the atomization trend prediction machine learning model can be expressed as follows:

[0203]

[0204] Among them, z is the real-time value of atomization prediction, w, x, y are independent variables, and β, α, θ are variable coefficients.

[0205] Furthermore, when using the training data and training the atomization trend prediction machine learning model, the trained model is evaluated through the test set, and the parameters of the atomization state prediction machine learning model are iteratively updated;

[0206] During the training process, the coefficient of determination, R 2 The closer the value is to 1, the stronger the explanatory power of the model. These indicators can directly reflect the accuracy and fit of the polynomial regression model in predicting continuous variables. Specifically, R 2 The calculation formula is as follows:

[0207]

[0208] in, It represents the residual sum of squares, that is, the sum of squares of the model predicted value (i.e., the second predicted atomization value) and the actual observed value (i.e., the historical atomization value). represents the total sum of squares, that is, the sum of the squares of the differences between all observations and their mean (i.e., the historical fog mean).

[0209] Before using the training data and training the atomization state prediction machine learning model, configure the model parameters of the atomization state prediction machine learning model, including the maximum depth of the tree and the branch parameters; 1. Select the required model parameters. In the embodiment of the present application, the parameters selected are the running time, atomization state, ratio, atomization real-time value, furnace exhaust gas temperature, total power, total water power, total water power ratio, exhaust gas temperature change rate, and voltage change rate; 2. Clearly predict the target parameter as the atomization state; 3. Select the maximum depth of the tree: 11; the atomization real-time value is used as the first judgment parameter, and the rest are used as branch parameters;

[0210] Furthermore, when using the training data and training the atomization state prediction machine learning model, the trained model is evaluated through the test set, and the parameters of the atomization state prediction machine learning model are iteratively updated;

[0211] During the training process, the above model is tested by a test set with consistent parameters with the training set, and a correlation analysis is performed between the test results and the original data. When the correlation coefficient between the test results and the original data reaches a preset value, the training of the atomization state prediction machine learning model reaches the target training state.

[0212] Correlation analysis can be performed using graphical methods, covariance methods, correlation coefficient methods, regression analysis methods, information entropy and mutual information methods, chi-square test methods, etc. The embodiment of the present application uses a correlation coefficient method to measure the strength and direction of the linear relationship between variables by calculating the correlation coefficient.

[0213] Specifically, the Pearson product-moment correlation coefficient r is a dimensionless index ranging from -1.0 to 1.0 (including -1.0 and 1.0) that reflects the degree of linear correlation between two data sets. The function has two parameters, including the independent variable set, AVERAGE(array1) and the dependent variable set, AVERAGE(array2); the parameters can be numbers, or names, array constants, or references containing numbers. The specific formula is as follows:

[0214]

[0215] in, and are the sample averages AVERAGE(array1) and AVERAGE(array2).

[0216] For the reduction furnace control system used in the production of polysilicon, an atomization prediction model based on nozzle and chassis layout combined with multiple parameters is constructed. The atomization trend prediction model is used to output the atomization trend curve after the initial parameter data at different times are run. The atomization trend curve is used to further adjust and optimize the initial parameter data, reduce or eliminate the rapid atomization and rapid increase trend, and discover potential atomization risks in advance.

[0217] Furthermore, the constructed atomization prediction model outputs real-time atomization state prediction information, and the atomization state prediction information is used in combination with the existing control system to regulate the reduction furnace during operation, thereby reducing the atomization degree and improving product quality and output.

[0218] The nozzle and chassis layout includes 24 pairs of rods, 40 pairs of rods, 60 pairs of rods, etc., and different numbers, heights, and calibers of nozzles are designed on different chassis layouts. Subsequent model parameter collection and aggregation are performed based on the different characteristics of the internal parts of the polysilicon reduction furnace equipment;

[0219] The initial parameter setting information of the system is input into the control system of the existing polysilicon reduction furnace before production. The initial parameter setting information includes the size and height of the nozzle used, the chassis layout, the trichlorosilane flow setting value at different operating times, the hydrogen flow setting value, the multi-phase current setting value and the rise value for heating the silicon rod, the voltage setting value, etc.

[0220] Before the polysilicon reduction furnace is operated based on the information, the atomization prediction model is used to predict the atomization trend during the batch operation using the initial parameter setting information, and the initial parameter setting information is found to have an upward trend in atomization before the operation starts. Specific methods include: 1. Increase the ratio (i.e., increase the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate), increase the hydrogen flow rate (i.e., increase the first hydrogen real-time flow rate), and reduce the trichlorosilane flow rate (i.e., reduce the first trichlorosilane real-time flow rate); 2. Reduce the temperature in places with high atomization trends (i.e., reduce the temperature in the target area), reduce the current increase value (i.e., reduce the increase value of the first multi-phase current real-time value), or increase the voltage baseline, etc., and reduce the possibility of atomization becoming serious by the above methods;

[0221] During operation, the initial parameter setting information after previous optimization needs to be input into the existing control system of the reduction furnace. Among them, trichlorosilane and hydrogen are first mixed and then sent into the reduction furnace through a regulating valve. At the same time, the current increase value and voltage setting value are input to heat the silicon core. The specific values ​​can be adjusted through the existing control system.

[0222] Furthermore, when using the existing polysilicon reduction furnace control system for production, the operation data of the polysilicon reduction furnace is obtained, and the operation data information includes but is not limited to the size and height of the nozzle used, the chassis layout, the actual flow rate of trichlorosilane, the actual flow rate of hydrogen, the return water flow rate of the furnace chassis, the tail gas temperature, the tail gas temperature in the furnace, the total power, and the actual value of the multi-phase current and the actual value of the voltage used to heat the silicon rods;

[0223] The atomization prediction model (i.e., the target atomization state prediction model) is used to generate the atomization state prediction information at the next moment for the above-mentioned polysilicon reduction furnace operation data, and the control parameters in the reduction furnace are adjusted by the existing polysilicon reduction furnace control system, so that the atomization state in the reduction furnace reaches a normal state;

[0224] The control parameters include, but are not limited to, hydrogen flow rate, trichlorosilane flow rate, multi-phase current rise value for heating silicon rods, voltage value, and furnace water flow rate.

[0225] Build a machine learning model for atomization trend prediction and a machine learning model for atomization status prediction, including:

[0226] Furthermore, the running data collected from the above training set is preprocessed, and the missing data is supplemented by using Impute Missing Values ​​in the knime software, which can be supplemented by the mean, median, mode or specific value; optionally, Filter Rows is used to remove data rows with missing values; and Statistics can be used to identify outliers, and outliers can be optionally removed or supplemented by the mean, median, mode or specific value;

[0227] The training sets generated above are used to train the atomization trend prediction machine learning model and the atomization state prediction machine learning model respectively, so as to generate an atomization prediction model after reaching the target training state;

[0228] During operation, the atomization trend prediction machine learning model is used to generate the overall atomization trend before operation. At the same time, the atomization state prediction machine model is used to generate the atomization state prediction information during operation. Combined with the existing control system, the next moment hydrogen, trichlorosilane set value, multi-phase current rise value, and voltage set value are output.

[0229] Acquire initial and runtime parameter data corresponding to the polysilicon reduction furnace producing polysilicon, and prepare a training set based on the acquired corresponding parameter data, wherein the training data set includes parameter data information of multiple different batches under different nozzle and chassis layouts;

[0230] The training sets generated above are used to train the atomization trend prediction machine learning model and the atomization state prediction machine learning model respectively, so as to generate an atomization prediction model after reaching the target training state;

[0231] During operation, the atomization trend prediction machine learning model is used to generate the overall atomization trend before operation. At the same time, the atomization state prediction machine model is used to generate the atomization state prediction information during operation. Combined with the existing control system, the next moment hydrogen, trichlorosilane set value, multi-phase current rise value, and voltage set value are output.

[0232] When using training data and training the machine learning model for atomization trend prediction, the selected independent variables and polynomial order are optimized and adjusted according to the test results;

[0233] Furthermore, when training the atomization trend prediction machine learning model using the training data, the trained model is evaluated through the test set, and the parameters of the atomization state prediction machine learning model are iteratively updated;

[0234] The fogging trend prediction machine learning model is performed using the Polynomial Regression machine learning algorithm, which captures nonlinear relationships in the data by adding polynomial terms to the input variables. This method is very useful in tasks that predict continuous values, especially when the data exhibits nonlinear trends.

[0235] Before using the training data to train the polynomial regression model for fogging trend prediction, you need to perform the following steps to configure the model and process the data:

[0236] 1. Handling missing values: Use appropriate methods such as interpolation, deletion, or filling to handle missing values ​​in the dataset.

[0237] 2. Feature selection: Determine the input variables used for prediction, such as hydrogen flow, trichlorosilane flow, multiphase current value, total voltage real-time value, etc.

[0238] 3. Feature transformation: Although polynomial regression inherently involves nonlinear transformation of features, it may also be necessary to standardize or normalize the data to improve the training efficiency and performance of the model.

[0239] For the machine learning model for atomization trend prediction, the main parameters include 1) Variable Coeff., which reflects the direction and intensity of the influence of the independent variable on the dependent variable. 2) Std.Err. is used to calculate the confidence interval and conduct hypothesis testing. 3) t-value, which reflects whether the difference between the variable coefficient and zero is significant. It is the ratio of the variable coefficient to its standard error, that is, t = coefficient / standard error. 4) P>|t|, which is used to judge whether the variable coefficient is significant. In a two-tailed test, if the value of P>|t| is less than the preset significance level (such as 0.05), the null hypothesis is rejected and the coefficient is considered to be significantly not zero; if the value of P>|t| is greater than or equal to the preset significance level, the null hypothesis is not rejected and the coefficient is considered to be zero or not significant.

[0240] The main steps for training a polynomial regression model using training data are as follows:

[0241] 1. Choose the order of the polynomial: Determine the order of the polynomial (such as quadratic, cubic, etc.), which determines the upper limit of the nonlinear complexity that the model can capture.

[0242] 2. Model fitting: Use the training data set (i.e., the first training set, including hydrogen flow, trichlorosilane flow, multiphase current value, total voltage real-time value, etc. as input, and atomization real-time value as output) to fit the polynomial regression model.

[0243] 3. Parameter estimation: By calculating the coefficient of determination for the machine learning model for atomization trend prediction, the closer the R2 value is to 1, the stronger the explanatory power of the model. These indicators can directly reflect the accuracy and fit of the polynomial regression model in predicting continuous variables.

[0244] 4. Model prediction: Use the trained polynomial regression model to predict the new input data and obtain the predicted value of the atomization trend.

[0245] 5. Result analysis: Evaluate the prediction results, including calculating the prediction error and drawing a comparison chart between the predicted value and the actual value. By performing trend analysis on the prediction results (such as drawing trend lines), you can have a clearer understanding of the changes in the atomization trend.

[0246] 6. Model optimization and adjustment: Based on the test results, the polynomial regression model is optimized and adjusted, which may include: Adjusting the polynomial order: increasing or decreasing the order of the polynomial to find the balance between the best fit and overfitting; Feature engineering: trying different feature combinations or transformations to improve the predictive performance of the model.

[0247] The machine learning model for atomization state prediction uses the Random Forest ensemble learning algorithm, which is an ensemble learning algorithm composed of multiple decision trees. These decision trees introduce randomness during the construction process to ensure that the prediction ability of each tree is relatively independent and different. By combining the prediction results of multiple trees, the overfitting phenomenon that may occur in a single decision tree can be significantly reduced, and the generalization ability of the model can be improved.

[0248] Before training the atomization state prediction machine learning model using the training data, configuring the model parameters of the atomization state prediction machine learning model, including the maximum depth of the tree and branch parameters;

[0249] Furthermore, when using the training data and training the atomization state prediction machine learning model, the trained model is evaluated through the test set, and the parameters of the atomization state prediction machine learning model are iteratively updated;

[0250] For the machine learning model for atomization state prediction, the main parameters include 1) n_estimators, which specifies the number of trees in the random forest. Increasing the number of trees can improve the stability and accuracy of the model, but too many trees will also increase the computational cost. 2) max_depth, which specifies the maximum depth of each decision tree, controls the growth of the tree, and limits the depth of the tree to prevent overfitting and improve the generalization ability of the model. 3) min_samples_split, the maximum depth of each decision tree limits the number of subtrees that continue to be divided. If the number of samples of a node is less than min_samples_split, it will not continue to try to select the optimal feature for division. This helps to avoid overfitting. 4) min_samples_leaf, the minimum number of samples required for leaf nodes, limits the minimum number of samples for leaf nodes. If the number of leaf nodes is less than the number of samples, it will be pruned together with the sibling nodes. This also helps to prevent overfitting. 5) max_features, the maximum number of features to be considered at each split. Increasing max_features generally improves the performance of the model because there are more features to choose from at each node. However, this will also reduce the diversity of individual trees, thus affecting the performance of random forests. Therefore, a proper balance is needed. 6) bootstrap, whether to use randomly sampled training data with replacement to build each tree. Using sampling with replacement can increase the randomness of the tree, thereby improving the generalization ability of the model.

[0251] The main steps of the training process of the atomization state prediction machine learning model (i.e., the target atomization state prediction model) are as follows:

[0252] 1. Randomly select features and samples: During the training process, random forests will randomly select features and samples to build multiple decision trees. Specifically, for each decision tree, a portion of samples will be randomly selected with replacement from the original data set as the training set, and a portion of features will be randomly selected from all features to build the decision tree.

[0253] 2. Build a decision tree: For each trained sample set and feature set, a decision tree is built. The decision tree construction process usually includes selecting the optimal splitting attributes and splitting points until the stopping condition is met (such as too few samples in the node or too small information gain).

[0254] 3. Model tuning and adjusting hyperparameters: During the training process, you may need to adjust some hyperparameters, such as the number of decision trees, the number of features used in each decision tree, the depth of the decision tree, etc., to improve the performance and generalization ability of the model.

[0255] 4. Optimize model structure: In addition to adjusting hyperparameters, you can also improve the performance of the model by optimizing the model structure. For example, you can reduce the complexity of the decision tree by pruning to avoid overfitting.

[0256] 5. Model evaluation uses a test set to evaluate the model: the above model is tested using a test set with parameters consistent with the training set, and the test results are compared and analyzed with the original data. When the correlation coefficient between the test results and the original data reaches a preset value, the training of the atomization state prediction machine learning model reaches the target training state.

[0257] Correlation analysis can be performed using graphical methods, covariance methods, correlation coefficient methods, regression analysis methods, information entropy and mutual information methods, chi-square test methods, etc. The embodiment of the present application uses a correlation coefficient method to measure the strength and direction of the linear relationship between variables by calculating the correlation coefficient.

[0258] In an embodiment of the present application, atomization state data before and after prediction are collected, and the correlation coefficient is calculated using the Pearson correlation coefficient. When the correlation coefficient r≥0.999, it can be considered that the predicted atomization state is positively correlated with the original atomization state data, and the correlation is strong. The atomization state prediction model has met the target training state.

[0259] It should be noted that compared with the prior art, the embodiments of the present application have at least the following beneficial effects:

[0260] 1. By combining different nozzles and chassis layouts to optimize the prediction model parameters, the model can be used more widely and the results more accurately;

[0261] 2. The embodiment of the present application can predict the atomization trend of the initial parameter data before operation through the atomization trend prediction model, and make adjustments when the atomization trend rises rapidly, thereby discovering potential atomization risks and reducing the atomization degree during operation;

[0262] 3. In the embodiment of the present application, the atomization state is predicted by an atomization state prediction model so that the control system can more accurately control the control parameters during operation, reduce the degree of atomization, and improve product quality and output.

[0263] Example 3

[0264] like Figure 8 As shown, the atomization control device in the reduction furnace provided in the embodiment of the present application includes:

[0265] A first acquisition module 810 is used to acquire real-time parameters in a target reduction furnace, wherein the real-time parameters include N first-category real-time parameters corresponding to N different moments in the current polysilicon growth cycle, and at least one item of the second-category real-time parameters at the current moment, where N is a positive integer;

[0266] A first prediction module 820, connected to the first acquisition module 810, is used to predict the real-time parameter using a target model corresponding to the real-time parameter to obtain a first prediction result, wherein the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model;

[0267] The first determination module 830 is connected to the first prediction module 820, and is used to determine a target control strategy according to the first prediction result, and the target control strategy is used to reduce the atomization degree in the target reduction furnace.

[0268] In some embodiments, the first type of real-time parameters includes at least one of a first real-time hydrogen flow rate, a first real-time trichlorosilane flow rate, a first real-time multiphase current value, and a first real-time total voltage value, the target model includes a target atomization trend prediction model, and the first prediction result includes a first predicted atomization trend;

[0269] The first prediction module 820 is specifically used for:

[0270] The target atomization trend prediction model is used to predict at least one of the first hydrogen real-time flow rate, the first trichlorosilane real-time flow rate, the first multi-phase current real-time value, and the first total voltage real-time value to obtain a first predicted atomization trend.

[0271] In some embodiments, the first prediction result includes a first predicted atomization trend, and the first determination module 830 is specifically configured to:

[0272] In the case where the first predicted atomization trend includes an upward trend, determining the target control strategy to be a first type of control strategy;

[0273] Among them, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate, increasing the first hydrogen real-time flow rate, reducing the first trichlorosilane real-time flow rate, reducing the temperature of the target area, reducing the first multi-phase current real-time value, and increasing at least one of the voltage baseline.

[0274] In some embodiments, the first prediction result includes a first predicted atomization state, and the first determination module 830 is specifically configured to:

[0275] In a case where the first predicted atomization state includes one of a light atomization state, a moderate atomization state, and a heavy atomization state, determining the target control strategy to be a second type of control strategy;

[0276] Among them, the second type of control strategy includes regulating at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace bottom pan water return real-time flow rate, the second multi-phase current real-time value and the second total voltage real-time value.

[0277] The atomization control device in the reduction furnace provided in the embodiment of the present application can be used to execute the atomization control method in the reduction furnace, that is, it has the beneficial effects and implementation methods of the atomization control method in the reduction furnace provided in Example 1 of the present application. For details, please refer to the specific description of the atomization control method in the reduction furnace in the above Example 1, and this embodiment will not be repeated here.

[0278] Example 4

[0279] In the embodiment of the present application, the target model adopts the target model of any one of Embodiment 1, and the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model.

[0280] like Fig. 9 As shown, the target model training device provided in the embodiment of the present application includes:

[0281] The second acquisition module 910 is used to obtain the initial polynomial order and the first training set, wherein the first training set includes M first training samples, each of the M first training samples includes N first-class historical parameters corresponding to N different moments in the historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-class historical parameters; N, M and K are all positive integers;

[0282] A first training module 920, connected to the second acquisition module 910, is used to perform training using M first training samples and an initial polynomial order to obtain a target atomization trend prediction model;

[0283] and / or,

[0284] A second acquisition module 910 is used to acquire a second training set, where the second training set includes K second training samples, each of the K second training samples includes a second type of historical parameter at a historical moment, and a historical atomization state corresponding to the second type of historical parameter;

[0285] The first training module 920 is connected to the second acquisition module 910 and is used to train the initial atomization state prediction model using K second training samples to obtain a target atomization state prediction model. The initial atomization state prediction model is obtained by training with a random forest ensemble learning algorithm.

[0286] In some embodiments, each of the N first-category historical parameters includes at least one of a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multi-phase current historical value, and a first total voltage historical value.

[0287] In the case where the target model includes a target atomization trend prediction model:

[0288] The first training module 920 is specifically used for:

[0289] A1. For the first target training sample among the M first training samples, based on the initial polynomial order, N first-category historical parameters are used as input, and N historical atomization values ​​are used as output, to obtain an initial polynomial regression model by fitting;

[0290] A2. For the second target training samples other than the first target training samples in the M first training samples, an initial polynomial regression model is used to perform atomization prediction on the N first-category historical parameters to obtain N second predicted atomization values ​​corresponding to N different moments in the historical polysilicon growth cycle;

[0291] A3. For each second target training sample other than the first target training sample in the M first training samples, determine a determination coefficient according to the N historical atomization values ​​and the N second predicted atomization values;

[0292] A4. When the determination coefficient is less than the preset determination coefficient, adjust the initial polynomial regression model to obtain an intermediate polynomial regression model, and update the initial polynomial regression model to the intermediate polynomial regression model, and update the second target training sample to the third target training sample among the M first training samples excluding the first target training sample and the second target training sample, and repeat steps A2 to A3 until the iterative determination coefficient is greater than or equal to the preset determination coefficient, and determine the iterative initial polynomial regression model as the target atomization trend prediction model.

[0293] In some embodiments, the second type of historical parameters includes at least one of a second hydrogen historical flow rate, a second trichlorosilane historical flow rate, a furnace bottom pan water return historical flow rate, a first tail gas historical temperature, a second tail gas historical temperature, a second total power, a second multiphase current historical value, and a second total voltage historical value.

[0294] In the case where the target model includes a target atomization state prediction model:

[0295] The first training module 920 is specifically used for:

[0296] B1. For a fourth target training sample among the K second training samples, use the initial atomization state prediction model to predict the fourth target training sample to obtain a second predicted atomization state;

[0297] B2. determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample;

[0298] B3. When the correlation coefficient is less than the preset correlation coefficient, the fourth target training sample is updated to any second training sample among the K second training samples except the fourth target training sample, and steps B1 to B2 are repeated until the correlation coefficient after iteration is greater than or equal to the preset correlation coefficient, and the initial atomization state prediction model after iteration is determined as the target atomization state prediction model.

[0299] The target model training device provided in the embodiment of the present application can be used to execute the target model training method, that is, it has the beneficial effects and implementation methods of the target model training method provided in Example 2 of the present application. For details, please refer to the specific description of the target model training method in the above Example 2, which will not be repeated in this embodiment.

[0300] It is to be understood that the above embodiments are merely exemplary embodiments used to illustrate the principles of the present application, but the present application is not limited thereto. For those skilled in the art, various modifications and improvements can be made without departing from the spirit and substance of the present application, and these modifications and improvements are also considered to be within the scope of protection of the present application.

Claims

1. A method for controlling atomization in a reduction furnace, characterized in that: include: Acquire real-time parameters in the target reduction furnace, wherein the real-time parameters include N first-category real-time parameters corresponding to N different moments in the current polysilicon growth cycle, and at least one of the second-category real-time parameters at the current moment, where N is a positive integer; Predicting the real-time parameter using a target model corresponding to the real-time parameter to obtain a first prediction result, wherein the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model; According to the first prediction result, a target control strategy is determined and executed, wherein the target control strategy is used to reduce the atomization degree in the target reduction furnace.

2. The method according to claim 1, characterized in that Before predicting the real-time parameter using the target model corresponding to the real-time parameter to obtain the first prediction result, the method further includes: Preprocessing the real-time parameters to obtain preprocessed real-time parameters, wherein the preprocessing includes at least one of supplementation processing, removal processing and outlier processing; The using the target model corresponding to the real-time parameter to predict the real-time parameter to obtain a first prediction result includes: The preprocessed real-time parameters are predicted using the target model corresponding to the preprocessed real-time parameters to obtain a first prediction result.

3. The method according to claim 2, characterized in that The preprocessing includes supplementation processing, removal processing and outlier processing. Preprocessing the real-time parameters to obtain preprocessed real-time parameters includes: Using the first target software to supplement the missing data in the real-time parameters to obtain supplemented real-time parameters, wherein the first target software supplements the missing data in the real-time parameters using a mean, a median, a mode or a preset value; Using the second target software to remove the data row where the missing value of the supplemented real-time parameter is located, to obtain the real-time parameter after removal; The outliers of the real-time parameters after removal are removed by using the third target software, or the outliers of the real-time parameters are supplemented by using the mean, median, mode or preset value to obtain the preprocessed real-time parameters.

4. The method according to claim 1, characterized in that: The first type of real-time parameters includes at least one of a first real-time hydrogen flow rate, a first real-time trichlorosilane flow rate, a first real-time multiphase current value, and a first real-time total voltage value; the target model includes a target atomization trend prediction model; and the first prediction result includes a first predicted atomization trend; Predicting the real-time parameter using the target model corresponding to the real-time parameter to obtain a first prediction result includes: The target atomization trend prediction model is used to predict at least one of the first hydrogen real-time flow rate, the first trichlorosilane real-time flow rate, the first multi-phase current real-time value and the first total voltage real-time value to obtain a first predicted atomization trend.

5. The method according to any one of claims 1 to 4, characterized in that: The second type of real-time parameters includes at least one of a second hydrogen real-time flow rate, a second trichlorosilane real-time flow rate, a furnace bottom pan water return real-time flow rate, a first tail gas real-time temperature, a second tail gas real-time temperature, a first total power, a second multiphase current real-time value, and a second total voltage real-time value, the target model includes a target atomization state prediction model, and the first prediction result includes a first predicted atomization state; Predicting the real-time parameter using the target model corresponding to the real-time parameter to obtain a first prediction result includes: The target atomization state prediction model is used to predict at least one of the second hydrogen real-time flow rate, the second chlorosilane real-time flow rate, the furnace chassis water return real-time flow rate, the first tail gas real-time temperature, the second tail gas real-time temperature, the first total power, the second multi-phase current real-time value and the second total voltage real-time value to obtain a first predicted atomization state.

6. The method according to claim 1, characterized in that The first prediction result includes a first predicted atomization trend, and determining a target control strategy according to the first prediction result includes: In the case where the first predicted atomization trend includes an upward trend, determining the target control strategy to be a first type of control strategy; Among them, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate, increasing the first hydrogen real-time flow rate, reducing the first trichlorosilane real-time flow rate, reducing the temperature of the target area, reducing the increase value of the first multi-phase current real-time value, and increasing at least one of the voltage baseline.

7. The method according to claim 1, characterized in that The first prediction result includes a first predicted atomization state, and determining a target control strategy according to the first prediction result includes: In the case where the first predicted atomization state includes one of a light atomization state, a moderate atomization state, and a heavy atomization state, determining the target control strategy to be a second type of control strategy; Among them, the second type of control strategy includes regulating at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace bottom pan water return real-time flow rate, the second multi-phase current real-time value and the second total voltage real-time value.

8. The method according to claim 1, characterized in that The target atomization trend prediction model includes: Among them, z is the atomization prediction value; w is the first independent variable; x is the first second independent variable; y is the first third independent variable; w i is the i+1th first independent variable; x i is the i+1th second independent variable; y i is the i+1th third independent variable; α i is the coefficient of the i-th first variable; β i is the coefficient of the i-th second variable; θ i is the i-th third variable coefficient; i is greater than or equal to 0 and less than or equal to 5.

9. A method for training a target model, characterized in that: The target model adopts the target model of any one of claims 1 to 8, the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model, and the method includes: Obtaining an initial polynomial order and a first training set, wherein the first training set includes M first training samples, each of the M first training samples including N first-class historical parameters corresponding to N different moments in a historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-class historical parameters; N, M, and K are all positive integers; Using M first training samples and initial polynomial orders for training, a target atomization trend prediction model is obtained; and / or, obtaining a second training set, the second training set comprising K second training samples, each of the K second training samples comprising a second type of historical parameter at a historical moment, and a historical atomization state corresponding to the second type of historical parameter; The initial atomization state prediction model is trained using K second training samples to obtain a target atomization state prediction model. The initial atomization state prediction model is trained using a random forest ensemble learning algorithm.

10. The method according to claim 9, characterized in that Each of the N first-category historical parameters includes at least one of a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multi-phase current historical value, and a first total voltage historical value. In the case where the target model includes a target atomization trend prediction model: The target atomization trend prediction model is obtained by training using the M first training samples and the initial polynomial order, including: A1. For a first target training sample among the M first training samples, based on the initial polynomial order, the N first-category historical parameters are used as input, and the N historical atomization values ​​are used as output, to obtain an initial polynomial regression model by fitting; A2. For the second target training samples other than the first target training samples in the M first training samples, using the initial polynomial regression model to perform atomization prediction on the N first-category historical parameters, and obtain N second predicted atomization values ​​corresponding to N different moments in the historical polysilicon growth cycle; A3. For each second target training sample other than the first target training sample in the M first training samples, determine a determination coefficient according to the N historical atomization values ​​and the N second predicted atomization values; A4. When the determination coefficient is less than the preset determination coefficient, adjust the initial polynomial regression model to obtain an intermediate polynomial regression model, update the initial polynomial regression model to the intermediate polynomial regression model, and update the second target training sample to the third target training sample among the M first training samples excluding the first target training sample and the second target training sample, repeat steps A2 to A3 until the iterative determination coefficient is greater than or equal to the preset determination coefficient, and determine the iterative initial polynomial regression model as the target atomization trend prediction model.

11. The method according to claim 10, characterized in that Determining a determination coefficient according to the N historical atomization values ​​and the N second predicted atomization values ​​includes: Determine the historical atomization average value according to N historical atomization values; Substitute the historical atomization mean, N historical atomization values ​​and N second predicted atomization values ​​into formula (1) to calculate and obtain the determination coefficient; Among them, R 2 represents the coefficient of determination; y represents the historical atomization value; represents the second predicted atomization value; Represents the historical fog average.

12. The method according to claim 9, characterized in that The second type of historical parameters includes at least one of the second hydrogen historical flow rate, the second trichlorosilane historical flow rate, the furnace bottom pan water return historical flow rate, the first tail gas historical temperature, the second tail gas historical temperature, the second total power, the second multiphase current historical value, and the second total voltage historical value. In the case where the target model includes a target atomization state prediction model: The initial atomization state prediction model is trained using K second training samples to obtain a target atomization state prediction model, including: B1. For a fourth target training sample among the K second training samples, use the initial atomization state prediction model to predict the fourth target training sample to obtain a second predicted atomization state; B2. determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample; B3. When the correlation coefficient is less than the preset correlation coefficient, the fourth target training sample is updated to any second training sample among the K second training samples except the fourth target training sample, and steps B1 to B2 are repeated until the correlation coefficient after iteration is greater than or equal to the preset correlation coefficient, and the initial atomization state prediction model after iteration is determined as the target atomization state prediction model.

13. The method according to claim 12, characterized in that Determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample includes: Using a preset correlation analysis method, a correlation analysis is performed on the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample to obtain a correlation coefficient; Among them, the preset correlation analysis method includes at least one of the chart method, covariance method, correlation coefficient method, regression analysis method, information entropy and mutual information method, chi-square test method and correlation coefficient method.

14. A reduction furnace atomization control device, characterized in that: include: A first acquisition module is used to acquire real-time parameters in a target reduction furnace, wherein the real-time parameters include N first-category real-time parameters corresponding to N different moments in a current polysilicon growth cycle, and at least one item of the second-category real-time parameters at the current moment, where N is a positive integer; A first prediction module, connected to the first acquisition module, is used to predict the real-time parameter using a target model corresponding to the real-time parameter to obtain a first prediction result, wherein the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model; The first determination module is connected to the first prediction module and is used to determine a target control strategy based on the first prediction result, wherein the target control strategy is used to reduce the atomization degree in the target reduction furnace.

15. The device according to claim 14, characterized in that The first type of real-time parameters includes at least one of a first real-time hydrogen flow rate, a first real-time trichlorosilane flow rate, a first real-time multiphase current value, and a first real-time total voltage value; the target model includes a target atomization trend prediction model; and the first prediction result includes a first predicted atomization trend; The first prediction module is specifically used for: The target atomization trend prediction model is used to predict at least one of the first hydrogen real-time flow rate, the first trichlorosilane real-time flow rate, the first multi-phase current real-time value and the first total voltage real-time value to obtain a first predicted atomization trend.

16. The device according to claim 14, characterized in that The first prediction result includes a first predicted atomization trend and a first determination module, which is specifically used to: In the case where the first predicted atomization trend includes an upward trend, determining the target control strategy to be a first type of control strategy; Among them, the first type of control strategy includes increasing the ratio of the first hydrogen real-time flow rate to the first trichlorosilane real-time flow rate, increasing the first hydrogen real-time flow rate, reducing the first trichlorosilane real-time flow rate, reducing the temperature of the target area, reducing the increase value of the first multi-phase current real-time value, and increasing at least one of the voltage baseline.

17. The device according to claim 14, characterized in that The first prediction result includes a first predicted atomization state and a first determination module, which is specifically used to: In the case where the first predicted atomization state includes one of a light atomization state, a moderate atomization state, and a heavy atomization state, determining the target control strategy to be a second type of control strategy; Among them, the second type of control strategy includes regulating at least one of the second hydrogen real-time flow rate, the second trichlorosilane real-time flow rate, the furnace bottom pan water return real-time flow rate, the second multi-phase current real-time value and the second total voltage real-time value.

18. A target model training device, characterized in that: The target model adopts the target model of any one of claims 1 to 8, the target model includes at least one of a target atomization trend prediction model and a target atomization state prediction model, and the device includes: A second acquisition module is used to acquire an initial polynomial order and a first training set, wherein the first training set includes M first training samples, each of the M first training samples includes N first-category historical parameters corresponding to N different moments in a historical polysilicon growth cycle, and N historical atomization values ​​corresponding to the N first-category historical parameters; N, M and K are all positive integers; A first training module, connected to the second acquisition module, is used to perform training using M first training samples and an initial polynomial order to obtain a target atomization trend prediction model; and / or, A second acquisition module is used to acquire a second training set, where the second training set includes K second training samples, and each of the K second training samples includes a second type of historical parameter at a historical moment, and a historical atomization state corresponding to the second type of historical parameter; The first training module is connected to the second acquisition module and is used to train the initial atomization state prediction model using K second training samples to obtain a target atomization state prediction model. The initial atomization state prediction model is obtained by training with a random forest ensemble learning algorithm.

19. The device according to claim 18, characterized in that Each of the N first-category historical parameters includes at least one of a first hydrogen historical flow rate, a first trichlorosilane historical flow rate, a first multi-phase current historical value, and a first total voltage historical value. In the case where the target model includes a target atomization trend prediction model: The first training module is specifically used for: A1. For a first target training sample among the M first training samples, based on the initial polynomial order, the N first-category historical parameters are used as input, and the N historical atomization values ​​are used as output, to obtain an initial polynomial regression model by fitting; A2. For the second target training samples other than the first target training samples in the M first training samples, using the initial polynomial regression model to perform atomization prediction on the N first-category historical parameters, and obtain N second predicted atomization values ​​corresponding to N different moments in the historical polysilicon growth cycle; A3. For each second target training sample other than the first target training sample in the M first training samples, determine a determination coefficient according to the N historical atomization values ​​and the N second predicted atomization values; A4. When the determination coefficient is less than the preset determination coefficient, adjust the initial polynomial regression model to obtain an intermediate polynomial regression model, update the initial polynomial regression model to the intermediate polynomial regression model, and update the second target training sample to the third target training sample among the M first training samples excluding the first target training sample and the second target training sample, repeat steps A2 to A3 until the iterative determination coefficient is greater than or equal to the preset determination coefficient, and determine the iterative initial polynomial regression model as the target atomization trend prediction model.

20. The device according to claim 18, characterized in that The second type of historical parameters includes at least one of the second hydrogen historical flow rate, the second trichlorosilane historical flow rate, the furnace bottom pan water return historical flow rate, the first tail gas historical temperature, the second tail gas historical temperature, the second total power, the second multiphase current historical value, and the second total voltage historical value. In the case where the target model includes a target atomization state prediction model: The first training module is specifically used for: B1. For a fourth target training sample among the K second training samples, use the initial atomization state prediction model to predict the fourth target training sample to obtain a second predicted atomization state; B2. determining a correlation coefficient according to the second predicted atomization state and the historical atomization state corresponding to the fourth target training sample; B3. When the correlation coefficient is less than the preset correlation coefficient, the fourth target training sample is updated to any second training sample among the K second training samples except the fourth target training sample, and steps B1 to B2 are repeated until the correlation coefficient after iteration is greater than or equal to the preset correlation coefficient, and the initial atomization state prediction model after iteration is determined as the target atomization state prediction model.

Images