Shoulder-releasing pulling speed control method, device, equipment, medium and program product

Abstract

The present invention discloses a method, device, equipment, medium, and program product for controlling shouldering pulling speed, relating to the field of crystal preparation technology. The method comprises: obtaining a crystal pulling dataset of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputting the crystal pulling dataset into a target shouldering initial state prediction model to obtain a shouldering initial state prediction result; integrating a static classification sub-model and a dynamic prediction sub-model into the target shouldering initial state prediction model; and controlling the pulling speed of the Czochralski single crystal during the shouldering process based on the shouldering initial state prediction result. The nonlinear relationship between the crystal pulling data and the shouldering initial state learned by the target shouldering initial state prediction model, which integrates the static classification sub-model and the dynamic prediction sub-model, is used to predict the shouldering initial state, thereby improving the model's accuracy in predicting the hot and cold states during the initial shouldering process.

Description

Technical Field

[0001] The present invention relates to the technical field of crystal preparation, and in particular to a shoulder-releasing pulling speed control method, device, equipment, medium and program product. Background Art

[0002] The Czochralski method is the primary method for producing single crystal silicon, which is used to refine polycrystalline silicon raw materials into single crystal silicon. The process of producing rod-shaped single crystal silicon crystals during the Czochralski process includes welding, seeding, shoulder release, shoulder rotation, equalization of diameters, and finishing.

[0003] Existing methods for judging and intervening in the initial hot / cold state of the shoulder during the early stages of the process are primarily categorized as "relying on human experience and manual intervention" and "judging by thresholds and adapting the process." The manual intervention method involves centralized control personnel determining whether the initial state of the shoulder is overcold / overheated based on the performance at the end of the seeding phase, and manually intervening in the shoulder shape using a higher / lower constant pull speed during the early stages of the process. This method is time-consuming and labor-intensive, and different personnel have different criteria for judging hot / cold, resulting in significant variability in shouldering results. The threshold judgment and process adaptation method typically utilizes the average seeding pull speed or the shoulder splitting time for threshold judgment. However, due to the lag in the effect of the heater on the furnace temperature, accurate and timely judgment of the initial hot / cold state of the shoulder based solely on the average seeding pull speed may be infeasible. Intervening in the pull speed to adjust the shoulder shape during the early stages of the process after the shoulder splitting time is known is already too late.

[0004] Therefore, the existing technology cannot quickly and accurately control the pulling speed in the initial state of shoulder release. Summary of the Invention

[0005] The present invention provides a shoulder-releasing drawing speed control method, device, equipment, medium and program product to solve the problem that the prior art cannot quickly and accurately control the drawing speed in the initial state of shoulder-releasing.

[0006] In a first aspect, an embodiment of the present invention provides a method for controlling a shoulder-releasing pulling speed, comprising:

[0007] Obtain a data set of Czochralski single crystals during the welding process, seeding process, and initial shoulder release process;

[0008] Inputting the crystal pulling data set into a target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model;

[0009] The pulling speed of the CZ single crystal during the shoulder releasing process is controlled according to the shoulder releasing initial state prediction result.

[0010] In a second aspect, an embodiment of the present invention provides a shoulder-releasing pulling speed control device, comprising:

[0011] The data acquisition module is used to obtain the crystal pulling data set of the CZ single crystal during the welding process, seeding process and the initial shoulder release process;

[0012] a shoulder release initial state prediction module, configured to input the crystal pulling data set into a target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model;

[0013] A pulling speed control module is used to control the pulling speed of the CZ single crystal during the shouldering process according to the shouldering initial state prediction result.

[0014] In a third aspect, an embodiment of the present invention provides an electronic device, comprising:

[0015] at least one processor;

[0016] and a memory communicatively coupled to the at least one processor;

[0017] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the shoulder release pulling speed control method described in any embodiment of the present invention.

[0018] In a fourth aspect, an embodiment of the present invention provides a computer-readable storage medium, wherein the computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the shoulder release pulling speed control method described in any embodiment of the present invention when executed.

[0019] In a fifth aspect, an embodiment of the present invention provides a computer program product including a computer program, which, when executed by a processor, implements the shoulder release pulling speed control method described in any embodiment of the present invention.

[0020] The technical solution of the embodiments of the present invention obtains a crystal pulling dataset of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputs the crystal pulling dataset into a target shouldering initial state prediction model to obtain a shouldering initial state prediction result; integrates the target shouldering initial state prediction model with a static classification sub-model and a dynamic prediction sub-model; and controls the pulling speed of the Czochralski single crystal during the shouldering process based on the shouldering initial state prediction result. By using the nonlinear relationship between the crystal pulling data and the initial shouldering state learned by the target shouldering initial state prediction model, which integrates the static classification sub-model and the dynamic prediction sub-model, the initial shouldering state is predicted. This improves the model's accuracy in predicting the hot and cold states during the initial shouldering process, resolving the problem of prior art in being unable to quickly and accurately control the pulling speed during the initial shouldering state.

[0021] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present invention, nor is it intended to limit the scope of the present invention. Other features of the present invention will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0023] Figure 1 This is a flow chart of a shoulder-releasing pulling speed control method provided in Example 1 of the present invention;

[0024] Figure 2 This is a flow chart of a shoulder-releasing pulling speed control method provided in the second embodiment of the present invention;

[0025] Figure 3 This is a flow chart of a shoulder-releasing pulling speed control method provided in Example 3 of the present invention;

[0026] Figure 4 Schematic diagram of the adaptive pulling speed and growth data when the initial state of the shoulder is relatively cold;

[0027] Figure 5 Schematic diagram of the adaptive pulling speed and growth data when the shoulder is initially hot;

[0028] Figure 6 A schematic structural diagram of a shoulder-releasing pulling speed control device provided in a fourth embodiment of the present invention;

[0029] Figure 7 A schematic structural diagram of an electronic device for implementing the shoulder-releasing pulling speed control method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts should fall within the scope of protection of the present invention.

[0031] It should be noted that the terms "first", "second", etc. in the description and claims of the present invention and the above-mentioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that the numbers used in this way can be interchanged where appropriate so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions. For example, a process, method, system, product or device that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0032] Example 1

[0033] Figure 1 This is a flowchart of a shouldering pulling speed control method provided in the first embodiment of the present invention. This embodiment is applicable to controlling the crystal pulling speed during the shouldering process of a CZ single crystal. The method can be executed by a shouldering pulling speed control device. The shouldering pulling speed control device can be implemented in the form of hardware and / or software. The shouldering pulling speed control device can be configured in an electronic device. Figure 1 As shown, the method includes:

[0034] S110, obtaining a crystal pulling data set of the Czochralski single crystal during the welding process, the seeding process, and the initial shoulder release process.

[0035] Czochralski (CZ) single crystals are single crystal silicon produced using the Czochralski method (abbreviated as CZ method). The Czochralski process primarily includes the welding process, seeding, shouldering, shoulder rotation, equalizing diameter, and finishing. This example primarily aims to provide timely control of the pulling speed during the shouldering process. Therefore, the CZ crystal pulling datasets are primarily collected for the initial and prior stages of shouldering, specifically the welding, seeding, and initial shouldering stages. Shouldering is the process of rapidly increasing the crystal diameter from the necking diameter to the desired equalizing diameter after seeding and necking by adjusting the crystal pulling speed and temperature. This stage primarily provides the appropriate crystal diameter and shape for subsequent equalizing diameter growth while ensuring crystal quality. The initial shouldering stage can be defined as a fixed-duration process starting with shouldering, or it can be defined as the time it takes for the CZ crystal to grow to a preset crystal length during the shouldering stage.

[0036] For example, data from the current welding process is obtained, including, for example, welding duration and average welding temperature; data from the current seeding process is obtained, including, for example, seeding duration, average seeding temperature, average seeding pull speed, and pull speed at the end of seeding; and data from the initial shouldering process is obtained, including, for example, actual diameter, set diameter, crystal pull speed, liquid level temperature, liquid inlet distance, and crucible position for shouldered crystals with a length less than 20 mm. By using multiple feature data such as pull speed, temperature, and time from the welding, seeding, and shouldering processes, the hot and cold conditions of the initial shouldering state are analyzed, reducing misjudgments caused by single features.

[0037] Optionally, the data collected during the welding and seeding processes of a Czochralski crystal are generally static statistical data that does not change over time, while the data collected during the initial shouldering process are generally time series data that changes dynamically over time. Therefore, the crystal pulling dataset for a Czochralski crystal during the welding, seeding, and initial shouldering processes may include: a static statistical data set consisting of the welding and seeding processes, and a dynamic time series data set during the initial shouldering process. The static statistical data set reflects the static statistical characteristics of the Czochralski crystal during the welding and seeding processes and can be used to analyze the impact of these static statistical characteristics on the initial cold and hot states of the shouldering process; the dynamic time series data set reflects the dynamic time series characteristics of the Czochralski crystal during the initial shouldering process and can be used to analyze the impact of these dynamic time series characteristics on the initial cold and hot states of the shouldering process.

[0038] S120, inputting the crystal pulling data set into the target shoulder release initial state prediction model to obtain the shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates the static classification sub-model and the dynamic prediction sub-model.

[0039] The target shoulder initial state prediction model can be understood as a model used to predict the hot and cold states of a CZ single crystal at the initial stage of shouldering, and is generally a trained or fully trained model. The target shoulder initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model. The static classification sub-model is used to analyze the mapping relationship between the static statistical data set in the crystal pulling data set and the shoulder initial state, and the dynamic prediction sub-model is used to analyze the mapping relationship between the dynamic time series data set in the crystal pulling data set and the shoulder initial state. The shoulder initial state prediction result can be understood as the prediction result for the hot and cold initial state of the shoulder output by the target shoulder initial state prediction model, which generally includes the hot and cold states and the predicted probability value. The shoulder initial state can include a cold initial state, a hot initial state, and a normal initial state.

[0040] Specifically, the crystal pulling data set collected during the welding process, seeding process and initial shoulder release process of the Czochralski single crystal is input into the target shoulder release initial state prediction model, and the target shoulder release initial state prediction model is used to predict the shoulder release initial state of the crystal pulling data set, thereby obtaining the shoulder release initial state prediction result.

[0041] For example, a target shoulder release initial state prediction model can be trained using a sample dataset of historical data collected from Czochralski crystals during the welding, seeding, and initial shoulder release processes. This sample dataset can include a static statistical sample dataset of the Czochralski crystal during the welding and seeding processes, as well as a dynamic time series sample dataset during the initial shoulder release process. The static statistical sample dataset is used to train the static classification submodel within the target shoulder release initial state prediction model, while the dynamic time series sample dataset is used to train the dynamic prediction submodel within the target shoulder release initial state prediction model.

[0042] S130. Control the pulling speed of the CZ single crystal during the shoulder-releasing process according to the prediction result of the initial state of the shoulder-releasing.

[0043] Specifically, the pulling speed of the CZ single crystal during the shoulder release process is controlled according to the hot and cold states at the initial stage of the shoulder release reflected by the prediction results of the initial state of the shoulder release.

[0044] The technical solution of the embodiments of the present invention provides a method for controlling the shouldering pulling speed. The method obtains a crystal pulling dataset of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputs the crystal pulling dataset into a target shouldering initial state prediction model to obtain a shouldering initial state prediction result; integrates the target shouldering initial state prediction model with a static classification sub-model and a dynamic prediction sub-model; and controls the pulling speed of the Czochralski single crystal during the shouldering process based on the shouldering initial state prediction result. The nonlinear relationship between the crystal pulling data and the shouldering initial state learned by the target shouldering initial state prediction model, which integrates the static classification sub-model and the dynamic prediction sub-model, is used to predict the shouldering initial state, thereby improving the model's accuracy in predicting the hot and cold states during the initial shouldering process.

[0045] Example 2

[0046] Figure 2 A flow chart of a shoulder release pulling speed control method provided in Example 2 of the present invention. This embodiment further refines the prediction process of the target shoulder release initial state prediction model on the basis of the above embodiments. Specifically, the crystal pulling data set is input into the target shoulder release initial state prediction model to obtain the shoulder release initial state prediction result, including: inputting the static statistical data set of the welding process and the crystal induction process in the crystal pulling data set into the static classification sub-model in the shoulder release initial state prediction model to obtain the first prediction result output by the static classification sub-model; inputting the dynamic time series data set of the initial shoulder release process in the crystal pulling data set into the dynamic prediction sub-model of the target shoulder release initial state prediction model to obtain the second prediction result output by the dynamic prediction sub-model; inputting the first prediction result and the second prediction result into the fusion sub-model in the shoulder release initial state prediction model to obtain the shoulder release initial state prediction result output by the fusion sub-model.

[0047] like Figure 2 As shown, the method includes:

[0048] S210 , obtaining a crystal pulling data set of the Czochralski single crystal during the welding process, the seeding process, and the initial shoulder release process.

[0049] S220 , inputting the static statistical data set of the welding process and the seeding process in the crystal pulling data set into the static classification sub-model in the target shoulder initial state prediction model to obtain a first prediction result output by the static classification sub-model.

[0050] The target shoulder release initial state prediction model may include a static classification sub-model, a dynamic prediction sub-model, and a fusion sub-model. The static classification sub-model is used to classify the shoulder release initial state of the static statistical data set in the crystal pulling data set. Optionally, the static classification sub-model may employ a decision tree model or a variant of a decision tree model, such as a gradient boosting decision tree (GBDT), XGBoost, or LightGBM. Other classification models may also be employed, and this is not limited in the present embodiment.

[0051] Specifically, the static statistical data set of the welding process and the seeding process in the crystal pulling data set is input into the static classification sub-model in the target shoulder release initial state prediction model; the static statistical data set is classified and predicted by the static classification sub-model to obtain a first prediction result, which may include the predicted shoulder release initial state type and predicted probability value of the CZO single crystal.

[0052] S230 , inputting the dynamic time series data set of the initial shoulder release process in the crystal pulling data set into the dynamic prediction sub-model of the target shoulder release initial state prediction model to obtain a second prediction result output by the dynamic prediction sub-model.

[0053] The dynamic prediction submodel is used to predict the initial state of the shoulder release for the dynamic time series dataset within the crystal pulling dataset. Optionally, the dynamic prediction submodel can be a Transformer model or a variant of the Transformer model, such as iTransformer, PatchFormer, Informer, or FEDformer. Other prediction models may also be used, and this is not a limitation in the present embodiment.

[0054] Specifically, the dynamic time series data set of the initial shoulder release process in the crystal pulling data set is input into the dynamic prediction sub-model of the target shoulder release initial state prediction model, and the dynamic prediction sub-model is used to predict the dynamic time series data set to obtain a second prediction result. The second prediction result may include the predicted shoulder release initial state type and predicted probability value of the CZO single crystal.

[0055] It is understandable that the execution order of S220 and S230 is not critical, and they can also be executed in parallel.

[0056] S240: Input the first prediction result and the second prediction result into a fusion sub-model in the target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result output by the fusion sub-model.

[0057] The fusion sub-model is used to fuse the first prediction result of the static classification sub-model and the second prediction result of the dynamic prediction sub-model. Optionally, the fusion sub-model can be a fully connected network.

[0058] Specifically, the first prediction result output by the static classification sub-model and the second prediction result output by the dynamic prediction sub-model are input into the fusion sub-model, and the first prediction result and the second prediction result are fused through the fusion sub-model to obtain the initial state prediction result of shoulder release, so that the initial state prediction result of shoulder release can integrate the influence of static statistical data and dynamic time series data on the initial state of shoulder release.

[0059] S250, controlling the pulling speed of the CZ single crystal during the shoulder releasing process according to the prediction result of the initial state of the shoulder releasing.

[0060] The technical solution of an embodiment of the present invention provides a method for controlling the shouldering pulling speed. The method comprises obtaining a crystal pulling data set of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputting a static statistical data set of the welding and seeding processes from the crystal pulling data set into a static classification sub-model of a target shouldering initial state prediction model to obtain a first prediction result output by the static classification sub-model; inputting a dynamic time series data set of the initial shouldering process from the crystal pulling data set into a dynamic prediction sub-model of the target shouldering initial state prediction model to obtain a second prediction result output by the dynamic prediction sub-model; inputting the first and second prediction results into a fusion sub-model of the target shouldering initial state prediction model to obtain a shouldering initial state prediction result output by the fusion sub-model; and controlling the pulling speed of the Czochralski single crystal during the shouldering process based on the shouldering initial state prediction result. The initial shouldering state is predicted using the relationship between the static statistical data and the initial shouldering state, as well as the relationship between the dynamic time series data and the initial shouldering state, learned by the target shouldering initial state prediction model fused with the static classification sub-model and the dynamic prediction sub-model, thereby improving the model's prediction accuracy for the hot and cold states of the initial shouldering process.

[0061] As an optional embodiment of the present application, the training method of the target shoulder initial state prediction model includes:

[0062] A1. Obtain a static statistical sample data set of the CZ single crystal during the welding process and the seeding process, as well as a dynamic time series sample data set at the initial stage of shoulder release; the sample data in the static statistical sample data set and the dynamic time series sample data set are marked with a priori label values ​​representing the initial state of shoulder release.

[0063] The static statistical sample dataset, consisting of historical static statistical data from multiple Czochralski single crystals during the welding and seeding processes, is used to train the static classification sub-model. The dynamic time series sample dataset, consisting of historical dynamic time series data from Czochralski single crystals during the initial shouldering phase, is used to train the dynamic prediction sub-model. Prior labels are used to represent the initial shouldering state represented by the sample data. For example, if the initial shouldering state is hot, the prior label value is set to 0; if the initial shouldering state is normal, the prior label value is set to 1; if the initial shouldering state is cold, the prior label value is set to 2.

[0064] For example, historical static data of a Czochralski crystal during the welding and seeding processes is collected. This data may include, for example, time, crystal length, diameter, crystal pulling speed, crucible lifting speed, liquid level temperature, and crucible position. This historical static data is then preprocessed, statistically analyzed, and labeled with priori labels to construct a static statistical sample dataset. This dataset may include, for example, welding time, seeding time, average welding temperature, average seeding temperature, average seeding pulling speed, and pulling speed at the end of seeding. Historical dynamic time series data of a Czochralski crystal during the initial shouldering process (e.g., data with a crystal length <20 mm during the shouldering process) is collected. This data may include, for example, actual diameter, set diameter, crystal pulling speed, liquid level temperature, liquid inlet distance, and crucible position. This sample dataset, after preprocessing and labeling with priori labels, is converted into a dynamic time series sample dataset. To further improve the quality of the sample dataset, we screened multiple single complete crystal pulling runs (including data on the splicing, seeding, shoulder release, shoulder rotation, and equalizing processes) from historical crystal pulling data from multiple single crystal furnaces over a period of time (e.g., one month). Based on this complete crystal pulling data, we determined a static statistical sample dataset and a dynamic time series sample dataset. It should be noted that both static statistical sample data and static statistical sample data come from the same single complete crystal pulling run.

[0065] The prior label values ​​of the static statistical sample dataset and the dynamic time series sample dataset can be determined based on the deviation between the actual diameter and the set diameter of the shoulder of the CZ single crystal, the deviation between the actual diameter slope and the set diameter slope, the average crystal pulling speed of the shoulder, and the coldness or hotness of the initial state of the shoulder corresponding to the current root dataset based on manual experience.

[0066] Optionally, the static statistical sample dataset and the dynamic time series sample dataset can be further divided into a training set, a test set, and a validation set. The training set is used to optimize model parameters and learn the relationship between the data features in the training set and the initial hot and cold states of the shoulder. The validation set is used to evaluate the model's performance in predicting the initial hot and cold states during training and to adjust the model's hyperparameters. The test set is used to evaluate the model's final prediction accuracy and generalization ability.

[0067] A2. Input the static statistical sample data set into the static classification sub-model in the initial shoulder release initial state prediction model to obtain a first output result output by the static classification sub-model, and perform model parameter training on the static classification sub-model according to the first output result.

[0068] The initial shoulder release prediction model can be understood as an untrained model. The initial shoulder release prediction model can include a static classification sub-model, a dynamic prediction sub-model, and a fusion sub-model. The first output result can be understood as the shoulder release prediction result output by the static classification sub-model during each iteration of the model training phase.

[0069] As an optional implementation of this embodiment, the static classification sub-model adopts a decision tree model. The decision tree model is a supervised learning algorithm that makes decisions based on a tree structure. By dividing the data set into multiple subsets, each subset corresponds to a decision rule, a tree-like decision model is constructed.

[0070] A20, inputting the static statistical sample data set into the static classification sub-model in the initial shoulder release initial state prediction model, obtaining a first output result output by the static classification sub-model, and performing model parameter training on the static classification sub-model according to the first output result, including:

[0071] A21. Select a feature from the static statistical sample data set as a currently selected feature, and select a partition threshold from a partition threshold set corresponding to the currently selected feature as the current partition threshold.

[0072] Among them, the features included in the static statistical sample data set may include welding time, seeding time, average welding temperature, average seeding temperature, average seeding pulling speed and seeding end pulling speed. Each feature may correspond to multiple selectable partitioning thresholds. Exemplarily, the partitioning threshold set corresponding to the average seeding pulling speed may include a single partitioning threshold and a partitioning threshold combination. A single partitioning threshold may include 200, 240, 280, 320, etc. A partitioning threshold of 240 means that it will be divided according to "average seeding pulling speed <240 mm / h" and "average seeding pulling speed >240 mm / h". It can also be a combination of two or more thresholds, such as 240 and 280, indicating that it will be divided according to "average seeding pulling speed <240 mm / h", "240 mm / h≤average seeding pulling speed <280 mm / h" and "average seeding pulling speed ≥280 mm / h". The currently selected feature can be understood as the feature selected from the dynamic time series sample data set during the current iterative training process. The current segmentation threshold can be understood as a segmentation threshold selected from a set of optional segmentation thresholds corresponding to the feature during the current iterative training process.

[0073] Specifically, a feature is selected from the features included in the static statistical sample data set as the current selected feature, and a partition threshold is selected from the optional partition threshold set corresponding to the current selected feature as the current partition threshold.

[0074] A22. Subsetting the static statistical sample data set according to the currently selected feature and the current partitioning threshold to obtain a partitioning result, and calculating the information gain of the partitioning result;

[0075] Specifically, the sample data contained in the static statistical sample dataset is subsetted according to the currently selected features and the current partitioning threshold. Static statistical sample data within the same threshold range is grouped together to obtain a partitioning result. The information gain corresponding to the partitioning result is then calculated. The information gain can be, for example, a gain ratio or a Gini index.

[0076] For example, for each partition result, the information gain of the partition result is equal to the original information entropy of the static statistical sample data set before subset partitioning Conditional entropy of static statistical sample data set after subtracting subset partitioning The original information entropy of the static statistical sample data set before subset division is:

[0077]

[0078] in, It belongs to The ratio of the number of static statistical samples of the class feature to the total number of samples in the static statistical sample dataset D, Indicates the initial state of shoulder release, such as It means that the initial state of shoulder release is hot; It means that the initial posture of shoulder placement is normal; Indicates that the initial state of the shoulder is relatively cold; n is the total number of features contained in the static statistical sample data set.

[0079] The conditional entropy of the static statistical sample data set after subset partitioning is:

[0080]

[0081] in, It is the subset divided according to the j-th partition threshold of the currently selected feature A. is a subset of static statistics The number of samples in is the number of samples in the static statistical sample data set D, and m is the total number of thresholds in the optional partition threshold set of the current selected feature A.

[0082] For example, using "average seeding speed" as the current partitioning feature, and using "average seeding speed < 240 mm / h," "240 mm / h ≤ average seeding speed < 280 mm / h," and "average seeding speed ≥ 280 mm / h" as the current partitioning thresholds, the sample data contained in the static statistical sample dataset is partitioned into three subsets. Assume that the ratio of the prior label values ​​in the dataset before partitioning is 1:1:1 for overheat:normal:undercooled. After partitioning, the ratio of the prior label values ​​in subset 1 is 5:1:2 for overheat:normal:undercooled; the ratio of the prior label values ​​in subset 2 is 1:6:1 for overheat:normal:undercooled; and the ratio of the prior label values ​​in subset 3 is 2:2:5 for overheat:normal:undercooled. Using the current partitioning thresholds, the data of different hot and cold initial states can be more purified in the three subsets.

[0083] A23. Select the next feature from the static statistical sample data set as the currently selected feature, and select the next partitioning threshold from the partitioning threshold set corresponding to the currently selected feature as the current partitioning threshold, and return the partitioning result obtained by subset partitioning the static statistical sample data set according to the current partitioning threshold of the currently selected feature, and calculate the information gain of the partitioning result until the iteration termination condition is reached.

[0084] Among them, the iteration termination condition can be that the prior label values ​​of the static statistical sample data in each division result are the same, or the static statistical sample data in at least one division result is less than the data volume threshold, or the depth of the decision tree is greater than the depth threshold.

[0085] Specifically, the features in the static statistical sample data set and the set of partitioning thresholds corresponding to the currently selected features are traversed, step A22 is repeated, the static statistical sample data set is subset-partitioned according to each currently selected feature and the current partitioning threshold to obtain the partitioning result, and the information gain of the partitioning result is calculated until any iteration termination condition is met, the iterative training process is terminated, thereby obtaining the information gain of all possible partitioning results.

[0086] A24. The feature corresponding to the partition result with the largest information gain is used as the optimal partition feature, and the partition threshold corresponding to the partition result with the largest information gain is used as the optimal partition threshold.

[0087] Specifically, according to the information gain of each segmentation result, the feature corresponding to the segmentation result with the largest information gain is selected as the optimal segmentation feature, and the corresponding segmentation threshold is selected as the optimal segmentation threshold.

[0088] A25. Adjust the model parameters of the static classification sub-model according to the optimal division feature and the optimal division threshold, and determine the division result with the maximum information gain as the first output result.

[0089] Specifically, the static classification sub-model parameters, such as the classification decision conditions, are adjusted based on the optimal partitioning features and the optimal partitioning threshold, and the static classification sub-model is output. The partitioning result with the maximum information gain is determined as the first output of the static classification sub-model, which is then input into the fusion sub-model.

[0090] This embodiment uses a decision tree class to build a static classification model, which can mine the mapping relationship between the data features of the welding and seeding processes and the initial hot and cold states of the shoulder release, and can tolerate the noise and missing values ​​that are widely present in industrial data. It analyzes the importance of input features to the prediction results, making the initial hot and cold state prediction results more interpretable.

[0091] A3. Perform model parameter training on the dynamic prediction sub-model in the initial shoulder release prediction model based on the dynamic time series sample data set to obtain a second output result output by the dynamic prediction sub-model, and perform model parameter training on the dynamic prediction sub-model based on the second output result and the prior label value.

[0092] Among them, the second output result can be understood as the initial state prediction result of the shoulder release output by the dynamic prediction sub-model during each iteration of the model training phase.

[0093] As an optional implementation of this embodiment, the dynamic prediction sub-model adopts a deep learning model based on an attention mechanism. A3, performing model parameter training on the dynamic prediction sub-model in the initial shoulder release prediction model based on the dynamic time series sample dataset, obtaining a second output result output by the dynamic prediction sub-model, and performing model parameter training on the dynamic prediction sub-model based on the second output result and the prior label value, includes:

[0094] A31. Input the dynamic time series sample data set into the embedding layer of the dynamic prediction sub-model in the initial shoulder-release initial state prediction model to obtain a query key value matrix.

[0095] Specifically, the dynamic prediction sub-model includes an embedding layer, a residual network, and a feedforward network. The initial dynamic continuous data is fed into the embedding layer to obtain a query key-value matrix consisting of query (Q), key (K), and value (V).

[0096] A32. Calculate the attention value matrix based on the query key value matrix, and process the data in the dynamic time series sample data set and the spliced ​​data of the corresponding attention values ​​in the attention value matrix through the residual network and the feedforward network to output a second output result.

[0097] Specifically, the attention value matrix is:

[0098]

[0099] Where Q is the query matrix, K is the key matrix, and V is the value matrix. is the dimension of the key vector, used to scale the dot product to prevent the gradient from disappearing, and T is the transposed sign.

[0100] The data in the dynamic time series sample dataset is concatenated with the corresponding attention values ​​in the attention value matrix. The concatenated data is processed by the residual network and the feedforward network to produce a second output result. The second output result can include the category and predicted probability value of each dynamic time series sample data, with the predicted probability value ranging from 0 to 1.

[0101] A33. Calculate a loss value based on the second output result and the priori label value, and iteratively adjust the parameters of the dynamic prediction sub-model based on the loss value.

[0102] Specifically, the loss value of the second output result and the prior label value is calculated, and the model with the smallest loss value in the validation set is selected as the trained dynamic prediction sub-model. For example, the loss value can be a cross entropy loss value (CrossEntropy Loss). The calculation formula of the cross entropy loss value is as follows:

[0103]

[0104] in, It is The dynamic time series sample data corresponds to the first The original probability of each class, It is samples The true category index of Is the output of the Softmax function, indicating that the model predicts samples Belongs to the real category The probability of is the negative log-likelihood loss, and N is the number of samples of dynamic time series sample data.

[0105] This example uses a deep learning model based on an attention mechanism to construct a dynamic prediction sub-model, effectively capturing the long-term nonlinear relationship between the dynamic time series data at the initial stage of shoulder release and the initial state of shoulder release. Compared to traditional recurrent neural networks (RNNs) or long short-term memory networks (LSTMs), the Transformer avoids the vanishing or exploding gradient issues when processing long sequences, improving model stability and prediction accuracy.

[0106] A4. Freeze the parameters of the static classification sub-model and the dynamic prediction sub-model, input the first output result and the second output result into the fusion sub-model in the initial shoulder release initial state prediction model, and obtain a third output result.

[0107] Among them, the third output result can be understood as the initial state prediction result of the shoulder release output by the fusion sub-model during each iteration of the model training stage.

[0108] Specifically, during each iterative training of the initial shoulder-dropping initial state prediction model, the model parameters of the static classification sub-model are trained based on the static statistical sample dataset, and the model parameters of the dynamic prediction sub-model are trained based on the dynamic time series sample dataset. The model parameters of the static classification sub-model and the dynamic prediction sub-model are then frozen. The first output result of the static classification sub-model and the second output result of the dynamic prediction sub-model are then input into the fusion sub-model to obtain a third output result. The third output result may include the shoulder-dropping initial state and predicted probability values ​​corresponding to the static statistical sample data and dynamic time series sample data of a single crystal pulling process, with the predicted probability value for each shoulder-dropping initial state ranging from 0 to 1.

[0109] For example, the fusion sub-model may use a fully connected network layer. The calculation formula of the fully connected neural network layer is as follows: in, is the output vector of the fully connected neural network layer, that is, the third output result, and its dimension is . is the input vector, i.e. the first output result and the second output result, whose dimension is ; is the weight matrix, whose dimension is ; is the bias vector, whose dimension is ; is the number of samples in the static statistical sample dataset (equal to the number of samples in the dynamic time series sample dataset), is the number of prior label values.

[0110] A5. Perform model parameter training on the fusion sub-model in the initial shoulder release initial state prediction model according to the loss function value between the third output result and the priori label value.

[0111] Specifically, the loss function value of the predicted value and the label is calculated, and the model with the smallest loss function value in the validation set is taken as the trained fusion sub-model.

[0112] A6. Construct the target shoulder initial state prediction model based on the trained static classification sub-model, dynamic prediction sub-model and fusion sub-model.

[0113] This embodiment adopts a method of fusing a static classification sub-model with a dynamic prediction sub-model to construct a shoulder initial state prediction model, which can learn the nonlinear relationship between the crystal pulling characteristics and the shoulder initial cold and hot states, thereby improving the prediction accuracy of the model.

[0114] Example 3

[0115] Figure 3 A flowchart of a method for controlling a shouldering process pulling speed is provided in accordance with a third embodiment of the present invention. This embodiment further refines the steps for controlling the pulling speed of a CZO crystal during the shouldering process based on a shouldering initial state prediction result based on the above-described embodiments. Controlling the pulling speed of a CZO crystal during the shouldering process based on the shouldering initial state prediction result includes: adjusting a reference pulling speed based on the shouldering initial state prediction result to obtain an adaptive reference pulling speed; obtaining current shouldering process data and performing proportional-integral-differential control on the current shouldering process data to obtain a pulling speed deviation; and determining the sum of the adaptive reference pulling speed and the pulling speed deviation as the pulling speed of the CZO crystal during the shouldering process.

[0116] like Figure 3 As shown, the method includes:

[0117] S310 , obtaining a crystal pulling data set of a Czochralski single crystal during a welding process, a seeding process, and an initial shoulder release process.

[0118] S320, inputting the crystal pulling data set into the target shoulder release initial state prediction model to obtain the shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates the static classification sub-model and the dynamic prediction sub-model.

[0119] S330: Determine an adaptive adjustment ratio according to the shoulder release initial state prediction result, and adjust the reference pulling speed according to the adaptive adjustment ratio to obtain an adaptive reference pulling speed.

[0120] The "base pulling speed" can be understood as the pulling speed calibrated during the shouldering process. The base pulling speed for different crystal lengths can be obtained from the Standard Operating Procedure (SOP). This SOP is typically provided by the manufacturer or production department of the single crystal growth equipment and contains standard operating parameters for the equipment. The "adaptive adjustment ratio" can be understood as the ratio by which the base pulling speed is adjusted. The adaptive base pulling speed can be understood as the base pulling speed after adaptive adjustment.

[0121] Specifically, the adaptive adjustment ratio is determined based on the shoulder initial state prediction result, and the reference casting speed is adjusted according to the adaptive adjustment ratio to obtain the adaptive reference casting speed. The adaptive adjustment ratio can scale the shoulder initial state prediction result output by the target shoulder initial state prediction model from [0,1] to the range from the minimum adaptive adjustment coefficient to the maximum adaptive adjustment coefficient. The specific calculation formula is:

[0122]

[0123] in, To adjust the ratio adaptively, is the initial state prediction result of the shoulder, and its value range is [0,1]. The model predicts the maximum value, The model predicts the minimum value, is the maximum value of the adaptive adjustment coefficient, is the minimum value of the adaptive adjustment coefficient.

[0124] This embodiment can adaptively adjust the reference pulling speed based on the shoulder release initial state prediction result based on linear scaling.

[0125] S340: Acquire current shoulder release process data, and perform proportional integral differential control on the current shoulder release process data to obtain a pulling speed deviation.

[0126] Current shouldering process data can be understood as data collected in real time during the shouldering process of a Czochralski crystal. Generally, the pulling speed during the initial waiting period is not adjusted. After the initial waiting period, current shouldering process data is collected and used to adjust the pulling speed. Current shouldering process data can include crystal length, actual diameter, set diameter, crystal pulling speed, and crucible pulling speed.

[0127] Specifically, during the shouldering process of the CZO single crystal, the current shouldering process data is collected in real time, and proportional-integral-derivative (PID) control is performed on the current shouldering process data to output the pulling speed deviation.

[0128] For example, PID control is performed based on the diameter deviation and diameter slope deviation corresponding to the current shoulder release process data. The calculation formula can be:

[0129] in: is the pulling speed deviation output by PID, is the diameter deviation and the diameter slope deviation, is the proportional gain, is the integral gain, is the differential gain.

[0130] Optionally, when the crystal length is greater than the preset maximum adjustable crystal length, proportional integral differential control is performed on the current shoulder release process data to obtain a pulling speed deviation to ensure that the preset crystal length diameter before shoulder release can grow slowly.

[0131] S350, determining the sum of the adaptive reference pulling speed and the pulling speed deviation as the adaptive output pulling speed, and controlling the pulling speed of the CZ single crystal during the shouldering process according to the adaptive output pulling speed.

[0132] Among them, the adaptive output pulling speed can be understood as the final output pulling speed, which is used to control the pulling speed of the CZ single crystal during the shoulder release process.

[0133] Specifically, Figure 4 This is a schematic diagram of the adaptive pulling speed and growth data when the shoulder is initially cold. Figure 5 The figure is a schematic diagram of the adaptive pulling speed and growth data under the condition of hot initial state of shoulder. Figure 4 As shown in the figure, when the target shoulder initial state prediction model determines that the initial state of shoulder is cold, the adaptive reference casting speed is increased compared with the reference casting speed, and the overall adaptive output casting speed is moved up to control the diameter growth in the early stage of shoulder not to be too fast. Figure 5 As shown in the figure, when the target shoulder release initial state prediction model determines that the initial state of shoulder release is too hot, the adaptive benchmark casting speed is lower than the benchmark casting speed, and the overall adaptive output casting speed decreases, so that the diameter growth in the early stage of shoulder release will not be too slow.

[0134] An embodiment of the present invention provides a method for controlling a shouldering pulling speed. The method obtains a crystal pulling dataset of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputs the crystal pulling dataset into a target shouldering initial state prediction model to obtain a shouldering initial state prediction result; the target shouldering initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model; determines an adaptive adjustment ratio based on the shouldering initial state prediction result, and adjusts a reference pulling speed based on the adaptive adjustment ratio to obtain an adaptive reference pulling speed; obtains current shouldering process data and performs proportional-integral-differential control on the current shouldering process data to obtain a pulling speed deviation; determines the sum of the adaptive reference pulling speed and the pulling speed deviation as an adaptive output pulling speed, and controls the pulling speed of the Czochralski single crystal during the shouldering process based on the adaptive output pulling speed. By adopting an adaptive pulling speed control strategy, the shouldering pulling speed is adaptively controlled according to the degree of hotness or coldness of the initial shouldering state, thereby improving the control accuracy of the shouldering pulling speed of the Czochralski crystal and avoiding overshoot.

[0135] Next, a specific example is given in combination with all the above embodiments. In the training process of the shoulder release initial state prediction model, 1,279 historical crystal pulling data from 80 vertical single crystal pulling furnaces for one month are used to screen the data of the welding process, the seeding process, and the initial shoulder release to train the initial shoulder release initial state prediction model. Table 1 is an evaluation table of the initial shoulder release initial state prediction model on the test set. As shown in Table 1, the static classification sub-model and the dynamic fusion sub-model are superior to the separate decision tree classification model or the Transformer prediction model in terms of accuracy, recall rate, and F1 score, which proves the effectiveness of the target shoulder release initial state prediction model provided by the embodiment of the present invention. The prediction accuracy of the target shoulder release initial state prediction model can reach 85.3%, which meets the application standard.

[0136] Table 1

[0137]

[0138] This embodiment conducts 715 actual shoulder-dropping pulling speed control operations on 40 vertical single crystal pulling furnaces over a period of one month. In the early stages of shoulder-dropping, the target shoulder-dropping initial state prediction model predicts the initial state of the shoulder-dropping, and adaptively adjusts the baseline pulling speed without manual intervention. Table 2 compares the results of the embodiment of the present invention with the experimental example and the comparative example. As shown in Table 2, the comparative example does not adjust the baseline pulling speed during the shoulder-dropping process. The embodiment and the comparative example have the same cooling power regulation and crucible lifting control during the shoulder-dropping process, and both use a PID controller with the same coefficient for pulling speed control. The experimental example and the comparative example have the same polysilicon raw material quality, quartz crucible batch, experimental environment, material disassembly and cleaning personnel, and process standards. The furnace types are all the same furnace types produced in the same batch.

[0139] Table 2

[0140]

[0141] A 10.5-inch single crystal silicon rod was produced by adopting the cold and hot initial state prediction and adaptive pulling speed control method of the shoulder release process of the experimental example of the present invention. The shoulder release survival rate and yield were higher than those of the control example, and the survival rate was improved by 1.35%.

[0142] Example 4

[0143] Figure 6 This is a schematic diagram of the structure of a shoulder-releasing pulling speed control device provided in the fourth embodiment of the present invention. Figure 6 As shown, the device includes: a data acquisition module 410, a shoulder initial state prediction module 420 and a pulling speed control module 430; wherein:

[0144] The data acquisition module 410 is used to acquire the crystal pulling data set of the CZ single crystal during the welding process, the seeding process and the initial shoulder release process;

[0145] The shoulder initial state prediction module 420 is used to input the crystal pulling data set into a target shoulder initial state prediction model to obtain a shoulder initial state prediction result; the target shoulder initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model;

[0146] The pulling speed control module 430 is used to control the pulling speed of the CZ single crystal during the shoulder release process according to the shoulder release initial state prediction result.

[0147] The technical solution of the embodiments of the present invention provides a Czochralski pulling speed control device. The device obtains a crystal pulling dataset of a Czochralski single crystal during the welding, seeding, and initial shouldering processes; inputs the crystal pulling dataset into a target shouldering initial state prediction model to obtain a shouldering initial state prediction result; and integrates the target shouldering initial state prediction model with a static classification sub-model and a dynamic prediction sub-model. The pulling speed of the Czochralski single crystal during the shouldering process is controlled based on the shouldering initial state prediction result. The nonlinear relationship between the crystal pulling data and the shouldering initial state learned by the target shouldering initial state prediction model, which integrates the static classification sub-model and the dynamic prediction sub-model, is used to predict the shouldering initial state, thereby improving the model's accuracy in predicting the hot and cold states during the initial shouldering process.

[0148] Optionally, the shoulder initial state prediction module 420 is specifically configured to:

[0149] Inputting the static statistical data set of the welding process and the seeding process in the crystal pulling data set into the static classification sub-model in the target shoulder initial state prediction model to obtain a first prediction result output by the static classification sub-model;

[0150] Inputting the dynamic time series data set of the initial shoulder release process in the crystal pulling data set into the dynamic prediction sub-model of the target shoulder release initial state prediction model to obtain a second prediction result output by the dynamic prediction sub-model;

[0151] The first prediction result and the second prediction result are input into a fusion sub-model in the target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result output by the fusion sub-model.

[0152] Optionally, the training method of the target shoulder initial state prediction model includes:

[0153] Obtaining a static statistical sample data set of the Czochralski single crystal during the welding process and the seeding process, as well as a dynamic time series sample data set at the initial stage of shoulder release; the sample data in the static statistical sample data set and the dynamic time series sample data set are marked with a priori label value indicating the initial state of shoulder release;

[0154] Inputting the static statistical sample data set into the static classification sub-model in the initial shoulder release initial state prediction model, obtaining a first output result output by the static classification sub-model, and performing model parameter training on the static classification sub-model according to the first output result;

[0155] Performing model parameter training on a dynamic prediction sub-model in the initial shoulder release initial state prediction model according to the dynamic time series sample data set, obtaining a second output result output by the dynamic prediction sub-model, and performing model parameter training on the dynamic prediction sub-model according to the second output result and the prior label value;

[0156] Freeze the parameters of the static classification sub-model and the dynamic prediction sub-model, input the first output result and the second output result into the fusion sub-model in the initial shoulder release initial state prediction model, and obtain a third output result;

[0157] Performing model parameter training on the fusion sub-model in the initial shoulder release initial state prediction model according to a loss function value between the third output result and the priori label value;

[0158] The target shoulder initial state prediction model is constructed based on the trained static classification sub-model, dynamic prediction sub-model and fusion sub-model.

[0159] Optionally, the static classification sub-model adopts a decision tree model; inputting the static statistical sample data set into the static classification sub-model in the initial shoulder release initial state prediction model, obtaining a first output result output by the static classification sub-model, and performing model parameter training on the static classification sub-model according to the first output result, includes:

[0160] Selecting a feature from the static statistical sample data set as a currently selected feature, and selecting a partition threshold from a set of optional partition thresholds corresponding to the currently selected feature as the current partition threshold;

[0161] Performing subset division on the static statistical sample data set according to the current selected feature and the current division threshold to obtain a division result, and calculating the information gain of the division result;

[0162] Selecting the next feature from the static statistical sample data set as the current selected feature, and selecting the next partitioning threshold from the set of optional partitioning thresholds corresponding to the current selected feature as the current partitioning threshold, and returning the partitioning result obtained by performing subset partitioning on the static statistical sample data set according to the current partitioning threshold of the current selected feature, and calculating the information gain of the partitioning result until the iteration termination condition is reached;

[0163] The feature corresponding to the partition result with the largest information gain is used as the optimal partition feature, and the partition threshold corresponding to the partition result with the largest information gain is used as the optimal partition threshold;

[0164] The model parameters of the static classification sub-model are adjusted according to the optimal division feature and the optimal division threshold, and the division result with the maximum information gain is determined as the first output result.

[0165] Optionally, the dynamic prediction sub-model adopts a deep learning model based on an attention mechanism; the performing model parameter training on the dynamic prediction sub-model in the initial shoulder release initial state prediction model according to the dynamic time series sample dataset, obtaining a second output result output by the dynamic prediction sub-model, and performing model parameter training on the dynamic prediction sub-model according to the second output result and the prior label value, includes:

[0166] Inputting the dynamic time series sample data set into the embedding layer of the dynamic prediction sub-model in the initial shoulder initial state prediction model to obtain a query key value matrix;

[0167] Calculating an attention value matrix according to the query key value matrix, and processing the data in the dynamic time series sample dataset and the concatenated data of the corresponding attention values ​​in the attention value matrix through a residual network and a feedforward network, and outputting a second output result;

[0168] A loss value is calculated according to the second output result and the priori label value, and parameters of the dynamic prediction sub-model are iteratively adjusted according to the loss value.

[0169] Optionally, the pulling speed control module 430 is specifically configured to:

[0170] determining an adaptive adjustment ratio according to the shoulder release initial state prediction result, and adjusting the reference pulling speed according to the adaptive adjustment ratio to obtain an adaptive reference pulling speed;

[0171] Obtain the current shoulder release process data, and perform proportional integral differential control on the current shoulder release process data to obtain the pulling speed deviation;

[0172] The sum of the adaptive reference pulling speed and the pulling speed deviation is determined as the adaptive output pulling speed, and the pulling speed of the CZ single crystal in the shouldering process is controlled according to the adaptive output pulling speed.

[0173] The shoulder-releasing pulling speed control device provided in the embodiment of the present invention can execute the shoulder-releasing pulling speed control method provided in any embodiment of the present invention, and has the corresponding functional sub-model and beneficial effects of the execution method.

[0174] Example 5

[0175] Figure 7A schematic diagram of an electronic device 10 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices (such as helmets, glasses, watches, etc.), and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or claimed herein.

[0176] like Figure 7 As shown, electronic device 10 includes at least one processor 11 and memory, such as read-only memory (ROM) 12 and random access memory (RAM) 13, communicatively connected to at least one processor 11. The memory stores computer programs executable by the at least one processor. Processor 11 can perform various appropriate actions and processes based on the computer programs stored in ROM 12 or loaded from storage unit 18 into RAM 13. RAM 13 can also store various programs and data required for the operation of electronic device 10. Processor 11, ROM 12, and RAM 13 are interconnected via bus 14. An input / output (I / O) interface 15 is also connected to bus 14.

[0177] Multiple components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc.; an output unit 17, such as various types of displays, speakers, etc.; a storage unit 18, such as a magnetic disk, an optical disk, etc.; and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0178] Processor 11 can be any general-purpose and / or specialized processing component with processing and computing capabilities. Some examples of processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any other suitable processor, controller, microcontroller, etc. Processor 11 executes the various methods and processes described above, such as the shoulder release and pull speed control method.

[0179] In some embodiments, the shoulder-release speed control method can be implemented as a computer program tangibly embodied in a computer-readable storage medium, such as storage unit 18. In some embodiments, part or all of the computer program can be loaded and / or installed on electronic device 10 via ROM 12 and / or communication unit 19. When the computer program is loaded into RAM 13 and executed by processor 11, one or more steps of the shoulder-release speed control method described above can be performed. Alternatively, in other embodiments, processor 11 can be configured to execute the shoulder-release speed control method via any other suitable means (e.g., via firmware).

[0180] Various embodiments of the systems and techniques described above can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0181] In some embodiments, the shoulder-release speed control method can be implemented as a computer program, which is invisibly included in a computer program product. When executed by a processor, the computer program implements the shoulder-release speed control method of the present invention. The computer program product can be understood as a software product whose solution is primarily implemented through a computer program. The computer program used to implement the method of the present invention can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flowchart and / or block diagram are implemented. The computer program can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0182] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in conjunction with an instruction execution system, device, or apparatus. A computer-readable storage medium may include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device, or apparatus, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. More specific examples of machine-readable storage media may include an electrical connection based on one or more wires, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0183] To provide interaction with a user, the systems and techniques described herein can be implemented on an electronic device that has: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0184] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer with a graphical user interface or web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0185] A computing system may include clients and servers. The clients and servers are typically remote from each other and typically interact via a communication network. This client-server relationship arises through computer programs running on the respective computers, creating a client-server relationship. The server may be a cloud server, also known as a cloud computing server or cloud host. This server is a hosting product within the cloud computing service ecosystem that addresses the management difficulties and limited scalability of traditional physical hosting and VPS services.

[0186] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present invention can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution of the present invention can be achieved. This is not limited herein.

[0187] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A shoulder-releasing pulling speed control method, characterized in that: include: Obtain a data set of Czochralski single crystals during the welding process, seeding process, and initial shoulder release process; The crystal pulling data set includes: a static statistical data set formed during the welding process and the seeding process, and a dynamic time series data set during the initial shoulder release process; Inputting the crystal pulling data set into a target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model; the static classification sub-model is used to analyze the mapping relationship between the static statistical data set and the shoulder release initial state, and the dynamic prediction sub-model is used to analyze the mapping relationship between the dynamic time series data set and the shoulder release initial state; The pulling speed of the CZ single crystal during the shoulder releasing process is controlled according to the shoulder releasing initial state prediction result.

2. The method according to claim 1, characterized in that Inputting the crystal pulling data set into a target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result includes: Inputting the static statistical data set of the welding process and the seeding process in the crystal pulling data set into the static classification sub-model in the target shoulder initial state prediction model to obtain a first prediction result output by the static classification sub-model; Inputting the dynamic time series data set of the initial shoulder release process in the crystal pulling data set into the dynamic prediction sub-model of the target shoulder release initial state prediction model to obtain a second prediction result output by the dynamic prediction sub-model; The first prediction result and the second prediction result are input into a fusion sub-model in the target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result output by the fusion sub-model.

3. The method according to any one of claims 1 to 2, characterized in that The training method of the target shoulder initial state prediction model includes: Obtaining a static statistical sample data set of the Czochralski single crystal during the welding process and the seeding process, as well as a dynamic time series sample data set at the initial stage of shoulder release; the sample data in the static statistical sample data set and the dynamic time series sample data set are marked with a priori label value indicating the initial state of shoulder release; Inputting the static statistical sample data set into the static classification sub-model in the initial shoulder release initial state prediction model, obtaining a first output result output by the static classification sub-model, and performing model parameter training on the static classification sub-model according to the first output result; Performing model parameter training on a dynamic prediction sub-model in the initial shoulder release initial state prediction model according to the dynamic time series sample data set, obtaining a second output result output by the dynamic prediction sub-model, and performing model parameter training on the dynamic prediction sub-model according to the second output result and the prior label value; Freeze the parameters of the static classification sub-model and the dynamic prediction sub-model, input the first output result and the second output result into the fusion sub-model in the initial shoulder release initial state prediction model, and obtain a third output result; Performing model parameter training on the fusion sub-model in the initial shoulder release initial state prediction model according to a loss function value between the third output result and the priori label value; The target shoulder initial state prediction model is constructed based on the trained static classification sub-model, dynamic prediction sub-model and fusion sub-model.

4. The method according to claim 3, characterized in that The static classification sub-model adopts a decision tree model; the static statistical sample data set is input into the static classification sub-model in the initial shoulder initial state prediction model, a first output result output by the static classification sub-model is obtained, and model parameter training of the static classification sub-model is performed according to the first output result, including: Selecting a feature from the static statistical sample data set as a currently selected feature, and selecting a partition threshold from a set of optional partition thresholds corresponding to the currently selected feature as the current partition threshold; Performing subset division on the static statistical sample data set according to the current selected feature and the current division threshold to obtain a division result, and calculating the information gain of the division result; Selecting the next feature from the static statistical sample data set as the current selected feature, and selecting the next partitioning threshold from the set of optional partitioning thresholds corresponding to the current selected feature as the current partitioning threshold, and returning the partitioning result obtained by performing subset partitioning on the static statistical sample data set according to the current partitioning threshold of the current selected feature, and calculating the information gain of the partitioning result until the iteration termination condition is reached; The feature corresponding to the partition result with the largest information gain is used as the optimal partition feature, and the partition threshold corresponding to the partition result with the largest information gain is used as the optimal partition threshold; The model parameters of the static classification sub-model are adjusted according to the optimal division feature and the optimal division threshold, and the division result with the maximum information gain is determined as the first output result.

5. The method according to claim 3, characterized in that The dynamic prediction sub-model adopts a deep learning model based on the attention mechanism; the dynamic prediction sub-model in the initial shoulder release prediction model is trained with model parameters according to the dynamic time series sample data set, a second output result of the dynamic prediction sub-model is obtained, and the model parameters of the dynamic prediction sub-model are trained according to the second output result and the prior label value, including: Inputting the dynamic time series sample data set into the embedding layer of the dynamic prediction sub-model in the initial shoulder initial state prediction model to obtain a query key value matrix; Calculating an attention value matrix according to the query key value matrix, and processing the data in the dynamic time series sample dataset and the concatenated data of the corresponding attention values ​​in the attention value matrix through a residual network and a feedforward network, and outputting a second output result; A loss value is calculated according to the second output result and the priori label value, and parameters of the dynamic prediction sub-model are iteratively adjusted according to the loss value.

6. The method according to any one of claims 1 to 2, characterized in that The controlling the pulling speed of the CZ single crystal during the shoulder releasing process according to the shoulder releasing initial state prediction result includes: determining an adaptive adjustment ratio according to the shoulder release initial state prediction result, and adjusting the reference pulling speed according to the adaptive adjustment ratio to obtain an adaptive reference pulling speed; Obtain the current shoulder release process data, and perform proportional integral differential control on the current shoulder release process data to obtain the pulling speed deviation; The sum of the adaptive reference pulling speed and the pulling speed deviation is determined as the adaptive output pulling speed, and the pulling speed of the CZ single crystal in the shouldering process is controlled according to the adaptive output pulling speed.

7. A shoulder-releasing pulling speed control device, characterized in that: include: The data acquisition module is used to obtain the crystal pulling data set of the CZ single crystal during the welding process, seeding process and the initial shoulder release process; The crystal pulling data set includes: a static statistical data set formed during the welding process and the seeding process, and a dynamic time series data set during the initial shoulder release process; a shoulder release initial state prediction module, configured to input the crystal pulling data set into a target shoulder release initial state prediction model to obtain a shoulder release initial state prediction result; the target shoulder release initial state prediction model integrates a static classification sub-model and a dynamic prediction sub-model; the static classification sub-model is configured to analyze the mapping relationship between the static statistical data set and the shoulder release initial state, and the dynamic prediction sub-model is configured to analyze the mapping relationship between the dynamic time series data set and the shoulder release initial state; A pulling speed control module is used to control the pulling speed of the CZ single crystal during the shouldering process according to the shouldering initial state prediction result.

8. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively coupled to the at least one processor; The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the shoulder-releasing pulling speed control method according to any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the shoulder-releasing pulling speed control method according to any one of claims 1 to 6 when executed.

10. A computer program product, characterized in that The computer program product includes a computer program, and when the computer program is executed by a processor, the computer program implements the shoulder-releasing pulling speed control method according to any one of claims 1 to 6.

Images