Single crystal silicon edge breakage prediction method based on improved SSA and extreme gradient boosting

Abstract

The application provides a single crystal silicon edge break prediction method based on improved SSA and extreme gradient boosting, and relates to the technical field of data processing. The method comprises the following steps: calculating the difference between the current constant diameter growth stage data of the single crystal silicon to be predicted and the standard constant diameter growth stage data determined according to the weight coefficients formed by clustering the historical constant diameter growth stage data of a plurality of sample single crystal silicon, to obtain current difference constant diameter growth stage data; performing statistical feature extraction on the current difference constant diameter growth stage data to obtain current difference features; inputting the current difference features into a prediction model to obtain the edge break prediction result of the single crystal silicon to be predicted; and the prediction model is obtained by optimizing an extreme gradient boosting model based on the historical difference features and edge break labels of each sample single crystal silicon and an improved sparrow optimization algorithm (SSA). The application realizes efficient and accurate single crystal silicon edge break prediction by reducing the dependence on a mechanism model.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme value gradient enhancement. Background Technology

[0002] Dislocations are a major defect in the single-crystal pulling process. In the constant-diameter growth stage, they mainly manifest as edge breakage. Edge breakage refers to the fracture of the edges around the crystal rod. Dislocation lines reduce the resistivity of the crystal material, leading to a decline in the performance of crystalline silicon and devices, and reducing the photoelectric conversion efficiency of silicon solar cells. Therefore, how to accurately and effectively predict edge breakage in single-crystal silicon is a crucial issue that the industry urgently needs to address.

[0003] Currently, the prediction of edge breakage during the single-crystal silicon pulling process is mainly based on multi-physics coupling mechanism models or hybrid models that combine mechanism and data-driven approaches. While constructing a mechanism model from the perspective of multi-physics coupling can clarify the physical meaning of parameters and the dynamic behavioral relationship between explanatory variable parameters, the large amount of knowledge involved and the wide range of topics make the mechanism model too complex and difficult to implement in practice. Although the hybrid model uses a data-driven model to describe some complex mechanisms, it is still essentially based on the mechanism model, and its model complexity is still very high.

[0004] Therefore, the implementation of existing edge breakage prediction methods relies on mechanistic models. However, the complexity of mechanistic models makes it difficult to establish them efficiently and accurately, which in turn makes it difficult to guarantee the accuracy and efficiency of edge breakage prediction in monocrystalline silicon. Summary of the Invention

[0005] This invention provides a method for predicting edge breakage in monocrystalline silicon based on improved SSA and extreme value gradient enhancement, which solves the problem that existing technologies require reliance on mechanistic models to predict edge breakage, making it difficult to guarantee the accuracy and efficiency of monocrystalline silicon edge breakage prediction. This invention enables efficient and accurate prediction of monocrystalline silicon edge breakage.

[0006] This invention provides a method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme gradient enhancement, comprising:

[0007] Obtain the current constant diameter growth stage data of the monocrystalline silicon to be predicted within the current time period;

[0008] The difference between the current equal diameter growth stage data and the standard equal diameter growth stage data is calculated to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon.

[0009] Statistical features are extracted from the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted. The current differential features are then input into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted.

[0010] The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference characteristics and broken edge labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

[0011] According to the present invention, a method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme gradient enhancement is provided, wherein the standard constant diameter growth stage data is obtained through the following steps:

[0012] Obtain historical data on the constant diameter growth stages of each sample monocrystalline silicon within a historical time period;

[0013] The historical isodiameter growth stage data are preprocessed; the preprocessing includes data merging, missing value imputation, outlier correction, and redundant parameter filtering.

[0014] Statistical features were extracted from the preprocessed historical constant diameter growth stage data to obtain the historical statistical features of the single-crystal silicon samples.

[0015] Based on the historical statistical characteristics, the historical constant diameter growth stage data of each sample of monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs.

[0016] The weighting coefficients corresponding to the historical constant diameter growth stage data of each sample monocrystalline silicon are determined based on the distance between the historical constant diameter growth stage data of each sample monocrystalline silicon and the cluster center of the cluster to which it belongs.

[0017] The standard equal diameter growth stage data are obtained based on the weighting coefficients and the historical equal diameter growth stage data of each sample monocrystalline silicon.

[0018] According to the present invention, a method for predicting edge breakage in monocrystalline silicon based on improved SSA and extreme value gradient enhancement is provided. The method involves clustering the historical constant-diameter growth stage data of each sample monocrystalline silicon according to the historical statistical characteristics to obtain the cluster to which the historical constant-diameter growth stage data of each sample monocrystalline silicon belongs, including:

[0019] The historical statistical features are normalized.

[0020] Principal component dimensionality reduction is performed on the normalized historical statistical features;

[0021] Based on the historical statistical characteristics after dimensionality reduction, the historical constant diameter growth stage data of each sample monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample monocrystalline silicon belong.

[0022] According to the present invention, a method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme value gradient enhancement is provided. The prediction model is trained based on the following steps:

[0023] The difference between the historical constant diameter growth stage data and the standard constant diameter growth stage data of each sample monocrystalline silicon is calculated to obtain the historical difference constant diameter growth stage data corresponding to each sample monocrystalline silicon.

[0024] Statistical feature extraction was performed on the historical difference data of the equal diameter growth stage to obtain the historical difference characteristics of each sample of monocrystalline silicon.

[0025] Based on the improved SSA, the hyperparameters of the extreme gradient boosting model are optimized to obtain the initial model parameters of the extreme gradient boosting model.

[0026] The extreme value gradient boosting model is initialized according to the initial model parameters, and the historical difference features are input into the initialized extreme value gradient boosting model to obtain the edge breakage prediction results of each sample of single crystal silicon.

[0027] Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, and the complexity of the initialized extreme value gradient boosting model, the target loss function is obtained.

[0028] Based on the target loss function, the model parameters of the initialized extreme value gradient boosting model are iteratively optimized to obtain the prediction model.

[0029] According to the present invention, a method for predicting edge breakage in monocrystalline silicon based on improved SSA and extreme gradient boosting is provided. The step of obtaining a target loss function based on the deviation between the edge breakage prediction result of each sample monocrystalline silicon and the edge breakage label of each sample monocrystalline silicon, and the complexity of the initialized extreme gradient boosting model, includes:

[0030] Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, an error loss function is obtained;

[0031] The complexity of the initialized extreme gradient boosting model is obtained based on the number of leaf nodes and the leaf node weights of the initialized extreme gradient boosting model.

[0032] The target loss function is obtained by weighted fusion of the error loss function and the complexity.

[0033] According to the present invention, a method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme gradient enhancement is provided. The method involves optimizing the hyperparameters of the extreme gradient enhancement model according to the improved SSA to obtain the initial model parameters of the extreme gradient enhancement model, including:

[0034] According to the reverse learning strategy, the initial sparrow population generated by the original SSA is initialized to obtain the reverse sparrow population; the position of each individual in the initial sparrow population is determined based on various preset combinations of key model parameters of the extreme value gradient boosting model.

[0035] Calculate the fitness of each individual in the reverse sparrow population, and determine the discoverers, followers, and scouts in the reverse sparrow population based on the fitness.

[0036] According to the spiral search strategy, the spiral radius of the discoverer is updated, and the position of the discoverer is updated according to the updated spiral radius;

[0037] The positions of the follower and the scout are updated based on the original SSA;

[0038] The reverse sparrow population is iteratively updated based on the updated locations of the discoverers, followers, and scouts until a preset termination condition is met. The preset termination condition includes the number of iterations reaching the maximum number of iterations, or the fitness of individuals in the updated reverse sparrow population reaching a preset value.

[0039] The initial model parameters of the extreme value gradient boosting model are determined based on the position of the individual with the best fitness in the reverse sparrow population after the last update.

[0040] The present invention also provides a single-crystal silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement, comprising:

[0041] The acquisition unit is used to acquire data on the current constant diameter growth stage of the monocrystalline silicon to be predicted within the current time period;

[0042] The calculation unit is used to calculate the difference between the current equal diameter growth stage data and the standard equal diameter growth stage data to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon.

[0043] The prediction unit is used to extract statistical features from the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted, and input the current differential features into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted.

[0044] The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference characteristics and broken edge labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

[0045] The present invention also provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the single-crystal silicon edge breakage prediction method based on any of the above-described improved SSA and extreme gradient enhancement methods.

[0046] The present invention also provides a non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement as described above.

[0047] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the single-crystal silicon edge breakage prediction method based on any of the above-described improved SSA and extreme value gradient enhancement methods.

[0048] The present invention provides a method for predicting edge breakage in monocrystalline silicon based on improved SSA and extreme gradient enhancement. By integrating improved SSA and XGBoos for model optimization training, a high-performance prediction model is constructed from model methods, data methods, and fusion methods. The constructed prediction model can better capture the trend information of the constant diameter growth characteristics of monocrystalline silicon in various scenarios through data analysis, effectively reducing the dependence on the mechanism model, and thus more efficiently and accurately predicting edge breakage in monocrystalline silicon in various scenarios. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0050] Figure 1 This is one of the flowcharts of the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement provided by the present invention;

[0051] Figure 2 This is the second flowchart of the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement provided by the present invention.

[0052] Figure 3 This is a schematic diagram of the distribution of clusters formed by K-means clustering provided by the present invention;

[0053] Figure 4 This is a schematic diagram of the distribution of standard isodiameter growth stage data provided by the present invention;

[0054] Figure 5 This is a schematic diagram of the extreme value gradient boosting model provided by the present invention;

[0055] Figure 6 This is a schematic diagram of the single-crystal silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement provided by the present invention.

[0056] Figure 7 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0058] In monocrystalline silicon, Czochralski-grown monocrystalline silicon is widely used in solar cells. Its fabrication process is mature, its crystal lattice is intact, and its mechanical strength is high. It is a fundamental material for integrated circuits and one of the main materials for solar cells. Its growth process can be divided into five stages: crystal introduction, necking, shoulder formation, constant diameter growth, and tailing. Among these, constant diameter growth is the most critical stage in the monocrystalline silicon growth process.

[0059] During the constant-diameter growth stage, the main defect in monocrystalline silicon is edge breakage. Edge breakage refers to the fracture of the edges around the crystal rod. Dislocation lines reduce the resistivity of the crystal material, leading to a decline in the performance of both the crystalline silicon and the device, and ultimately reducing the photoelectric conversion efficiency of silicon solar cells. Therefore, accurately and effectively predicting edge breakage in monocrystalline silicon is a crucial issue that the industry urgently needs to address.

[0060] Currently, the prediction of edge breakage during the single-crystal silicon pulling process is mainly based on multi-physics coupling mechanism models or hybrid models that combine mechanism and data-driven approaches. While constructing a mechanism model from the perspective of multi-physics coupling can clarify the physical meaning of parameters and the dynamic behavioral relationship between explanatory variable parameters, the large amount of knowledge involved and the wide range of topics make the mechanism model too complex and difficult to implement in practice. Although the hybrid model uses a data-driven model to describe some complex mechanisms, it is still essentially based on the mechanism model, and its model complexity is still very high.

[0061] Data-driven methods, when applied to modeling complex industrial processes, can easily detect and predict the state of monocrystalline silicon pulling. Compared to mechanistic models based on multiphysics coupling, or hybrid models combining mechanisms and data-driven approaches, machine learning methods based on data-driven models do not rely on the Czochralski monocrystalline silicon growth mechanism and can improve prediction accuracy through continuous training. This is the main advantage of using data to build models.

[0062] Therefore, existing edge breakage prediction methods rely on mechanistic models. However, the complexity of these models makes them difficult to establish efficiently and accurately, leading to a lack of assurance in the accuracy and efficiency of edge breakage prediction for monocrystalline silicon. This embodiment uses a data-driven model formed by the fusion of an improved SSA and extreme gradient enhancement to predict edge breakage in monocrystalline silicon. By relying on data to determine the optimization of the model, the dependence on mechanistic models can be effectively reduced, thereby achieving efficient and accurate edge breakage prediction for monocrystalline silicon.

[0063] Figure 1 One of the flowcharts for the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement provided by this invention is shown below; Figure 1 As shown, the method includes the following steps:

[0064] Step 110: Obtain the current constant diameter growth stage data of the monocrystalline silicon to be predicted within the current time period;

[0065] Here, the monocrystalline silicon to be predicted refers to the monocrystalline silicon for which edge breakage prediction is required during the constant diameter growth stage. The time period can be set according to actual monitoring needs, such as dividing each period into 6 months.

[0066] Optionally, the current constant diameter growth stage data refers to the data generated during the constant diameter growth stage. It can be obtained by collecting data on the single crystal silicon produced by the same type of single crystal furnace associated with the single crystal silicon to be predicted during the current constant diameter growth stage in the current time period. The constant diameter growth stage data includes, but is not limited to, multiple items such as crystal rotation, crucible rotation, diameter, crystal lift, set growth rate, actual growth rate, main power, main current, actual liquid surface temperature, liquid outlet distance, crucible lift, main chamber pressure, dry pump frequency / throttle valve opening, argon flow rate, crucible position, seed crystal position, net weight, crystal weight, total weight, weight removed, remaining weight, crystal length, hot screen water temperature, and hot screen flow rate.

[0067] It should be noted that after obtaining the current constant diameter growth stage data of the monocrystalline silicon to be predicted within the current time period, the data can be preprocessed before performing subsequent prediction operations. This improves the data quality of the current constant diameter growth stage data, enabling more accurate and efficient acquisition of the monocrystalline silicon edge breakage prediction results. The preprocessing mentioned includes, but is not limited to, data merging, missing value imputation, outlier correction, and redundant parameter screening.

[0068] Step 120: Calculate the difference between the current equal diameter growth stage data and the standard equal diameter growth stage data to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon.

[0069] It should be noted that, since the task of producing monocrystalline silicon of the same size is the same each time under the same working conditions, and the expected monocrystalline silicon constant diameter growth trend curve can be fitted under normal circumstances, this embodiment adopts the standard curve feature construction method to obtain the standard constant diameter growth stage data.

[0070] Here, the steps for obtaining standard constant diameter growth stage data can be as follows: first, cluster the historical constant diameter growth stage data of multiple sample monocrystalline silicon, and then obtain the weight coefficients of the historical constant diameter growth stage data of each sample monocrystalline silicon based on the clustering results, so as to construct standard constant diameter growth stage data based on the weight coefficients.

[0071] The clustering referred to here can be either using the historical constant diameter growth stage data of the sample monocrystalline silicon as the feature parameters required for clustering, or using the historical constant diameter growth stage data of the sample monocrystalline silicon after preprocessing and feature extraction as the feature parameters required for clustering. No specific limitation is made on the clustering method here.

[0072] Optionally, after obtaining the current equal diameter growth stage data and the standard equal diameter growth stage data, trend features can be extracted from the current equal diameter growth stage data to obtain the current equal diameter growth stage trend features, and trend features can be extracted from the standard equal diameter growth stage data to obtain the standard equal diameter growth stage trend features.

[0073] Next, the difference between the current trend characteristics of the equal diameter growth stage and the trend characteristics of the standard equal diameter growth stage is calculated to form the current differential equal diameter growth stage data.

[0074] Step 130: Statistical feature extraction is performed on the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted. The current differential features are then input into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted. The prediction model is obtained by training an extreme value gradient boosting model based on the historical differential features and edge breakage labels of each sample monocrystalline silicon. The initial model parameters of the extreme value gradient boosting model are optimized based on the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on a back-learning strategy, a spiral search strategy, and the original SSA.

[0075] Here, the prediction model is built based on a data-driven model, which uses data to automatically detect whether the monocrystalline silicon has broken edges during the constant diameter growth stage. It combines the characteristics of high accuracy and real-time performance, and can quickly and accurately predict the real-time constant diameter growth stage data of monocrystalline silicon, so as to obtain the edge break prediction results of monocrystalline silicon efficiently and accurately.

[0076] The prediction model is trained using a fusion of an improved Sparrow Search Algorithm (SSA) and XGBoost (Extreme Gradient Boosting). This trained model better captures the edge breakage phenomenon in monocrystalline silicon, thus achieving more efficient and accurate prediction of edge breakage in monocrystalline silicon. The improved SSA is derived from the original SSA by incorporating a back-learning strategy and a spiral search strategy.

[0077] Optionally, before performing step 130, a prediction model capable of accurately predicting edge breakage in monocrystalline silicon needs to be pre-trained. This prediction model is trained based on the following steps:

[0078] First, to improve the prediction of edge breakage in monocrystalline silicon, it is necessary to collect historical constant-diameter growth stage data of sample monocrystalline silicon under various operating conditions. Following steps 120 and 130, differential constant-diameter growth stage data and statistical feature extraction are performed on the historical constant-diameter growth stage data of sample monocrystalline silicon under different operating conditions. This yields historical difference features of sample monocrystalline silicon under various operating conditions, ensuring that the sample dataset constructed based on these historical difference features has sufficient depth and breadth. This allows the trained prediction model to accurately predict edge breakage in monocrystalline silicon under different scenarios. The sample dataset can be historically collected, such as creating a sample dataset from historically collected sample monocrystalline silicon under different operating conditions; or it can be obtained by loading a sample dataset from an open-source database.

[0079] Secondly, the XGBoost iteration is optimized based on the sample dataset and the improved SSA to obtain a prediction model that can better capture the characteristics of single-crystal silicon edge breakage and can efficiently and accurately output the prediction results of single-crystal silicon edge breakage. This optimization can be achieved by using the improved SSA to perform a global search optimization on XGBoos based on the sample dataset to obtain the prediction model; or by first determining the initialization parameters of XGBoos based on the improved SSA to initialize XGBoos, and then further optimizing the XGBoos initialized based on the sample dataset to obtain the prediction model. The specific optimization steps of the prediction model are not specifically limited here.

[0080] Optionally, after obtaining the prediction model through iterative optimization, statistical features can be extracted from the current differential equal diameter growth stage data, including but not limited to maximum value, mean, variance, standard deviation, kurtosis, and minimum value, to obtain the current differential features of the monocrystalline silicon to be predicted; and the current differential features are input into the prediction model to predict the edge breakage of monocrystalline silicon based on the current differential features, thereby outputting the edge breakage prediction result used to characterize whether the monocrystalline silicon to be predicted has an edge breakage defect in the current time period.

[0081] The monocrystalline silicon edge breakage prediction method provided in this embodiment, based on improved SSA and extreme value gradient enhancement, optimizes the model training by fusing improved SSA and XGBoos. This approach constructs a high-performance prediction model from the perspectives of model method, data method, and fusion method. The constructed prediction model can better capture the trend information of the constant diameter growth characteristics of monocrystalline silicon under various scenarios through data analysis, effectively reducing the dependence on the mechanism model. As a result, it can more efficiently and accurately predict monocrystalline silicon edge breakage under various scenarios.

[0082] In some embodiments, the standard isodiameter growth stage data are obtained through the following steps:

[0083] Obtain historical data on the constant diameter growth stages of each sample monocrystalline silicon within a historical time period;

[0084] The historical isodiameter growth stage data are preprocessed; the preprocessing includes data merging, missing value imputation, outlier correction, and redundant parameter filtering.

[0085] Statistical features were extracted from the preprocessed historical constant diameter growth stage data to obtain the historical statistical features of the single-crystal silicon samples.

[0086] Based on the historical statistical characteristics, the historical constant diameter growth stage data of each sample of monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs.

[0087] The weighting coefficients corresponding to the historical constant diameter growth stage data of each sample monocrystalline silicon are determined based on the distance between the historical constant diameter growth stage data of each sample monocrystalline silicon and the cluster center of the cluster to which it belongs.

[0088] The standard equal diameter growth stage data are obtained based on the weighting coefficients and the historical equal diameter growth stage data of each sample monocrystalline silicon.

[0089] Figure 2 The second schematic diagram of the flowchart for the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement provided by this invention; as shown. Figure 2 As shown, the specific steps for obtaining standard isodiameter growth stage data include:

[0090] First, obtain the historical isodiameter growth stage data of each sample monocrystalline silicon in the historical time period from the sample dataset;

[0091] Next, the data from each historical constant diameter growth stage are preprocessed, including but not limited to merging the historical constant diameter growth stage data within a given period of monocrystalline silicon, filling in and correcting missing or outlier values ​​in the merged data, and performing parameter screening on the data from each historical constant diameter growth stage to remove redundant features. This parameter screening can be achieved through correlation system analysis, such as mutual information or Pearson correlation coefficients, to calculate parameter correlations before parameter screening.

[0092] For example, the specific process of preprocessing can be as follows:

[0093] Data on monocrystalline silicon produced during the historical constant-diameter growth stages of monocrystalline silicon samples from the same model of monocrystalline furnace within a historical period were collected and merged to obtain merged historical constant-diameter growth stage data for each sample. Secondly, since the data collected by the monocrystalline silicon parameter acquisition sensors may contain missing data, missing key parameters were filled in and outliers were corrected based on the merged historical constant-diameter growth stage data. Thirdly, since the purpose of feature selection is to extract a subset of feature attributes with a high impact on edge breakage before feature dimensionality reduction, and to perform feature dimensionality reduction on the sample point set containing this subset, and since the feature selection method is not directly related to model training, mutual information coefficients can be used to perform correlation analysis on the corrected data. This allows for redundant parameter filtering based on the correlation, ensuring that only the subset of feature attributes with a high impact on edge breakage is retained in the preprocessed historical constant-diameter growth stage data.

[0094] Mutual information is used to evaluate the amount of information contributed by the occurrence of one event to the occurrence of another event. The specific calculation formula is as follows:

[0095]

[0096] Wherein, p(d) i ,d j ) represents the parameter d in the corrected historical isodiameter growth stage data. i and parameter d j The joint probability between them; p(d i ) and p(d j ) are parameters d i and parameter d j The marginal probability.

[0097] The maximum mutual information coefficient (MPI) addresses the relationship between two variables by discretizing them into a two-dimensional space as scatter points and calculating the joint probability by examining the distribution of these points across different intervals. This solves the problem of finding the joint probability in mutual information analysis. MPI is not limited to a specific function type and can capture various linear or nonlinear relationships between variables. Therefore, using MPI to remove redundant parameters can completely capture and retain the subset of feature attributes that have a high impact on edge breakage.

[0098] Next, after the preprocessing of the historical constant diameter growth stage data is completed, statistical features can be extracted from the preprocessed historical constant diameter growth stage data to obtain effective information that can characterize the constant diameter growth characteristics of each sample of single crystal silicon, thereby obtaining the historical statistical characteristics of each sample of single crystal silicon.

[0099] As shown in Table 1, statistical feature extraction can be performed on the preprocessed historical isodiameter growth stage data to extract time-domain features, including but not limited to multiple features such as maximum value, minimum value, mean, peak-to-peak value, absolute mean, root mean square value, root mean square amplitude, variance, kurtosis, skewness, average amplitude, margin index, waveform index, impulse index, peak index, and kurtosis index. In Table 1, x represents the data points in the preprocessed historical isodiameter growth stage data; n represents the number of data points in the preprocessed historical isodiameter growth stage data.

[0100] Next, after obtaining the historical statistical characteristics of each sample of monocrystalline silicon, the historical constant diameter growth stage data of each sample of monocrystalline silicon can be clustered based on the historical statistical characteristics to divide the historical constant diameter growth stage data of each sample of monocrystalline silicon into the corresponding clusters, thereby obtaining the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs, as well as the cluster center; then, the weight coefficient of the sample monocrystalline silicon is constructed using the distance between the historical constant diameter growth stage data of each sample monocrystalline silicon and the cluster center, and then the standard monocrystalline silicon constant diameter growth stage data is constructed using the weight coefficient of the sample monocrystalline silicon and the historical constant diameter growth stage data of the sample monocrystalline silicon, thereby obtaining the standard constant diameter growth stage data.

[0101] Table 1 Statistical Characteristics Calculation Table

[0102]

[0103] Here, clustering can be performed by directly using the historical statistical characteristics of the sample monocrystalline silicon as the feature parameters required for clustering, or by performing data processing operations such as normalization and dimensionality reduction on the historical statistical characteristics of the sample monocrystalline silicon as the feature parameters required for clustering. No specific limitation is made on the clustering method here.

[0104] Here, the steps for constructing the standard constant diameter growth stage data can be based on the weighting coefficients of each sample of single-crystal silicon, and the historical constant diameter growth stage data of each sample of single-crystal silicon can be weighted and fused to obtain the standard constant diameter growth stage data.

[0105] The method provided in this embodiment preprocesses and performs statistical feature processing on historical constant diameter growth stage data to achieve clustering of historical constant diameter growth stage data of sample monocrystalline silicon, and obtains standard constant diameter growth stage data based on the clustering results, which can effectively improve the accuracy, flexibility and reliability of standard constant diameter growth stage data acquisition.

[0106] In some embodiments, the step of clustering the historical constant diameter growth stage data of each sample of monocrystalline silicon according to the historical statistical characteristics to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs includes:

[0107] The historical statistical features are normalized.

[0108] Principal component dimensionality reduction is performed on the normalized historical statistical features;

[0109] Based on the historical statistical characteristics after dimensionality reduction, the historical constant diameter growth stage data of each sample monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample monocrystalline silicon belong.

[0110] like Figure 2 As shown, in the field of machine learning, different evaluation metrics often have different dimensions and units. This can affect the subsequent feature engineering and model input. To eliminate the influence of dimensions between metrics, data standardization is required before clustering historical isochronous growth stage data to ensure comparability between data metrics. The normalization calculation formula is as follows:

[0111]

[0112] Among them, F * ={f1,f2,f3,…,f m} represents historical statistical features. By normalizing these historical statistical features, they can be made to have the same dimension, providing support for subsequent principal component dimensionality reduction.

[0113] Furthermore, in order to reduce the feature dimensionality and thus reduce computation, after normalizing the historical isochronous growth stage data, principal component dimensionality reduction is also required for the normalized historical statistical features.

[0114] Principal Component Analysis (PCA) transforms high-dimensional data containing redundant information into a small amount of low-dimensional data, namely principal components. Each principal component contains almost all the effective information of the original data. This process transforms complex data analysis problems into problems requiring only a few principal components, enabling more in-depth analysis and simplifying the process. The basic idea is to find a projection transformation matrix that best represents the main features of the original data under the constraint of minimum mean square error. The steps for PCA dimensionality reduction of the normalized historical statistical features F are as follows:

[0115] (1) Remove the average value, that is, subtract the average value of each feature dimension.

[0116] (2) Calculate the covariance matrix Where m is the number of statistical features;

[0117] (3) Calculate the covariance matrix using SVD (Singular Value Decomposition). eigenvalues ​​and eigenvectors.

[0118] (4) Sort the eigenvalues ​​from largest to smallest and select the k largest ones. Then, use the k corresponding eigenvectors as column vectors to form the eigenvector matrix P.

[0119] (5) Transform the normalized historical statistical features into a new space constructed by k feature vectors, i.e., F′=PF. F′ is the historical statistical feature after PCA dimensionality reduction, which retains at least 90% of the effective information of the original historical statistical features.

[0120] Subsequently, after obtaining the dimensionality-reduced historical statistical features, these features can be used as clustering parameters. K-means can be used to cluster the historical constant diameter growth stage data of each sample of monocrystalline silicon to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs.

[0121] The basic idea of ​​K-means clustering is to iteratively find a partitioning scheme of K clusters that minimizes the loss function corresponding to the clustering result.

[0122] Here, the number of clusters in K-means clustering can be configured according to actual needs. For example, configuring the number of clusters to 1 means that the historical constant diameter growth stage data of the sample monocrystalline silicon are clustered into one class. Correspondingly, the steps for constructing the standard constant diameter growth stage data for the dimensionality-reduced historical statistical features F′ include:

[0123] (1) First, the historical constant diameter growth stage data of the sample single crystal silicon were clustered by the historical statistical features after dimensionality reduction. K=1 was set, that is, the number of clusters was 1. Then, the cluster center of the cluster was determined according to the clustering results.

[0124] (2) Calculate the distance between the historical equal diameter growth stage data and the cluster center of each sample monocrystalline silicon, and construct the weighting coefficient of the historical equal diameter growth stage data of each sample monocrystalline silicon accordingly.

[0125] (3) Standard equal diameter growth stage data are obtained by using the weighting coefficients of the historical equal diameter growth stage data of each sample of monocrystalline silicon and the historical equal diameter growth stage data of each sample of monocrystalline silicon.

[0126] Figure 3 This is a schematic diagram illustrating the distribution of clusters formed by K-means clustering provided by the present invention; as shown. Figure 3 As shown, K-means clustering can accurately and effectively obtain the cluster to which the historical constant diameter growth stage data of each sample of single-crystal silicon belongs.

[0127] Figure 4 This is a schematic diagram illustrating the distribution of standard isodiameter growth stage data provided by the present invention; for example... Figure 4 As shown, the trend characteristics of the standard constant diameter growth stage constructed by clustering can effectively characterize the changing trend characteristics of standard monocrystalline silicon in the constant diameter growth stage.

[0128] The method in this embodiment normalizes historical statistical features and performs principal component dimensionality reduction before clustering historical isodiameter growth stage data. This eliminates the influence of dimensions between indicators while reducing the feature dimensionality, thereby ensuring the accuracy and efficiency of obtaining standard isodiameter growth stage data during the clustering process based on the dimensionality-reduced historical statistical features.

[0129] In some embodiments, the prediction model is trained based on the following steps:

[0130] The difference between the historical constant diameter growth stage data and the standard constant diameter growth stage data of each sample monocrystalline silicon is calculated to obtain the historical difference constant diameter growth stage data corresponding to each sample monocrystalline silicon.

[0131] Statistical feature extraction was performed on the historical difference data of the equal diameter growth stage to obtain the historical difference characteristics of each sample of monocrystalline silicon.

[0132] Based on the improved SSA, the hyperparameters of the extreme gradient boosting model are optimized to obtain the initial model parameters of the extreme gradient boosting model.

[0133] The extreme value gradient boosting model is initialized according to the initial model parameters, and the historical difference features are input into the initialized extreme value gradient boosting model to obtain the edge breakage prediction results of each sample of single crystal silicon.

[0134] Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, and the complexity of the initialized extreme value gradient boosting model, the target loss function is obtained.

[0135] Based on the target loss function, the model parameters of the initialized extreme value gradient boosting model are iteratively optimized to obtain the prediction model.

[0136] Figure 5 This is a schematic diagram of the structure of the extreme value gradient boosting model provided by the present invention; as shown below. Figure 5 As shown, XGBoost is an ensemble learning algorithm that combines the characteristics of gradient boosting and decision trees, and it exhibits excellent performance and generalization ability when dealing with nonlinear problems.

[0137] like Figure 2 As shown, the specific training steps for the prediction model include:

[0138] First, historical constant diameter growth stage data of each sample monocrystalline silicon were obtained from the sample dataset. Then, trend features were extracted from the historical constant diameter growth stage data and standard constant diameter growth stage data of each sample monocrystalline silicon to obtain the historical constant diameter growth trend features corresponding to the historical constant diameter growth stage data and the standard constant diameter growth trend features corresponding to the standard constant diameter growth stage data. The difference between the historical constant diameter growth trend features corresponding to the historical constant diameter growth stage data and the standard constant diameter growth trend features corresponding to the standard constant diameter growth stage data of each sample monocrystalline silicon was calculated to obtain the historical difference constant diameter growth stage data corresponding to each sample monocrystalline silicon.

[0139] Next, statistical features were extracted from the historical difference data of the same diameter growth stage corresponding to each sample of monocrystalline silicon, including but not limited to the maximum value, mean, variance, standard deviation, kurtosis, and minimum value, in order to obtain the historical difference features corresponding to the sample monocrystalline silicon.

[0140] Next, the key model parameters of the XGBoost model are initially optimized by introducing a sparrow search algorithm with an improved reverse learning strategy and a spiral search strategy, and finally the global optimal solution for the initialization of the XGBoost model is obtained; the key model parameters include, but are not limited to, the number of iterations, the learning rate, and the maximum depth of the tree.

[0141] Next, the XGBoost model is initialized according to the initial model parameters, and the historical difference features are input into the initialized extreme gradient boosting model. The historical difference features are then used to predict the edge breakage of monocrystalline silicon through the initialized extreme gradient boosting model, thereby obtaining the edge breakage prediction results of monocrystalline silicon for each sample.

[0142] Next, the deviation between the predicted edge breakage result of each sample of monocrystalline silicon and the edge breakage label of each sample of monocrystalline silicon, as well as the complexity of the initialized extreme value gradient boosting model, are calculated. Based on the deviation and complexity, the objective loss function of the XGBoost model is constructed. The model parameters of the initialized extreme value gradient boosting model are iteratively optimized based on this objective loss function as the optimization objective until the maximum number of iterations is reached, or the model performance index value of the optimized extreme value gradient boosting model reaches the set index value. Subsequently, the prediction model can be constructed based on the optimized extreme value gradient boosting model with the best model performance index obtained during the optimization process.

[0143] The method provided in this embodiment optimizes the XGBoost model by using an improved SSA and historical constant diameter growth stage data of sample monocrystalline silicon in various different scenarios during the training process. This can accelerate network convergence and efficiently obtain a prediction model that integrates multiple scenarios. This effectively improves the model training efficiency and also improves the prediction accuracy of monocrystalline silicon edge breakage in various complex scenarios.

[0144] In some embodiments, obtaining the target loss function based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, and the complexity of the initialized extreme value gradient boosting model, includes:

[0145] Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, an error loss function is obtained;

[0146] The complexity of the initialized extreme gradient boosting model is obtained based on the number of leaf nodes and the leaf node weights of the initialized extreme gradient boosting model.

[0147] The target loss function is obtained by weighted fusion of the error loss function and the complexity.

[0148] Optionally, the target loss function can be a loss function obtained by weighted fusion of the error loss function and the complexity, and the specific calculation formula is as follows:

[0149]

[0150] in, The error loss function can be a loss function obtained by fusing the loss values ​​constructed from the deviation between the edge breakage prediction results and the edge breakage labels of all sample monocrystalline silicon. l is the complexity; l1 and l2 are weight coefficients; i is the i-th sample monocrystalline silicon in the dataset; m represents the total amount of sample monocrystalline silicon imported into the k-th tree; r represents the total number of trees built in the XGBoost model; y i This indicates the broken edge label of the i-th sample monocrystalline silicon. This represents the edge breakage prediction result of the i-th sample monocrystalline silicon in the r-th prediction; g k Ω() represents the structural term of the tree model in the XGBoost model; L() represents the loss function, which describes the difference between the broken edge label and the broken edge prediction result; Ω() represents the complexity of the tree.

[0151] The complexity of the tree is calculated based on the number of leaf nodes and their weights, using the following formula:

[0152]

[0153] Where γ is a parameter used to control the proportion of leaves; T is the number of leaf nodes in the k-th tree; λ is the normalization coefficient; ω j Let be the leaf node weight of the j-th leaf node.

[0154] The method provided in this embodiment takes into account the influence of model complexity based on the loss function constructed by considering the deviation between the predicted edge breakage result of each sample of monocrystalline silicon and the edge breakage label of each sample of monocrystalline silicon. At the same time, it optimizes the generalization error of the model in terms of mathematical principles and seeks a balance point in the variance and bias dilemma of the model in order to generalize the model and improve the training efficiency of the model.

[0155] In some embodiments, the step of optimizing the hyperparameters of the extreme value gradient boosting model according to the improved SSA to obtain the initial model parameters of the extreme value gradient boosting model includes:

[0156] According to the reverse learning strategy, the initial sparrow population generated by the original SSA is initialized to obtain the reverse sparrow population; the position of each individual in the initial sparrow population is determined based on various preset combinations of key model parameters of the extreme value gradient boosting model.

[0157] Calculate the fitness of each individual in the reverse sparrow population, and determine the discoverers, followers, and scouts in the reverse sparrow population based on the fitness.

[0158] According to the spiral search strategy, the spiral radius of the discoverer is updated, and the position of the discoverer is updated according to the updated spiral radius;

[0159] The positions of the follower and the scout are updated based on the original SSA;

[0160] The reverse sparrow population is iteratively updated based on the updated locations of the discoverers, followers, and scouts until a preset termination condition is met. The preset termination condition includes the number of iterations reaching the maximum number of iterations, or the fitness of individuals in the updated reverse sparrow population reaching a preset value.

[0161] The initial model parameters of the extreme value gradient boosting model are determined based on the position of the individual with the best fitness in the reverse sparrow population after the last update.

[0162] Optionally, the steps for obtaining the initial model parameters of the extreme value gradient boosting model include:

[0163] First, the initial sparrow population generated by the original SSA is formed by initializing various preset key model parameters based on the extreme value gradient boosting model; the so-called key model parameters include, but are not limited to, the number of trees, the depth of the trees, and the learning rate.

[0164] Next, due to the uneven distribution and reduced diversity of the initial population randomly generated by the original sparrow optimization algorithm, the algorithm is prone to getting trapped in local optima in the later stages. Therefore, this embodiment initializes the initial sparrow population generated by the original SSA using a reverse learning strategy to obtain a reverse sparrow population. This effectively reduces the risk of the sparrow algorithm getting trapped in local optima and enhances the global search capability. Reverse learning is a strategy to expand the search area of ​​the algorithm. The reverse learning strategy selects closer individuals as the initial individuals of the population to improve the convergence speed of all individuals in the population. The specific operation is as follows: randomly generate the initial sparrow population and generate its reverse population; calculate the fitness value and select the better population as the next generation population.

[0165] Next, the fitness of each individual in the reverse sparrow population is calculated. Based on the fitness, the discoverers, followers, and scouts in the reverse sparrow population are determined. For example, individuals with higher fitness can be designated as discoverers, those with the second highest fitness as followers, and those with lower fitness as scouts.

[0166] Next, while the original sparrow search algorithm's discoverer position update formula is simple and efficient, it also significantly increases the probability of getting trapped in local optima. To address this, this embodiment introduces a spiral search strategy into the discoverer position update process of the original sparrow search algorithm. This algorithm draws inspiration from seagulls, using a random search approach to solve the problem. In seagull attack behavior, individuals capture their prey by spiraling their actions, constantly changing angles and speeds. This behavior balances the needs of local exploitation and global search, significantly increasing the algorithm's randomness and making it less prone to getting trapped in local optima.

[0167] The discoverer's position is updated by modifying the spiral radius of the discoverer based on the original SSA, taking into account the impact of population size. The formula for updating the discoverer's position is as follows:

[0168]

[0169] in, Let represent the position value of the i-th individual in the j-th dimension during the t-th iteration; Y is the warning value; AN is the preset safety value; M is a random number; and D is a matrix with a value of 1. x′, y′, and z′ represent the attack trajectory of the discoverer.

[0170] In this study, considering the influence of the sparrow population size F, the spiral radius r in the SOA (Seagull Optimization Algorithm) is improved, and the improved spiral radius r′ is used to determine the attacker's trajectory. The calculation formulas for variables such as spiral radius r′, attack trajectory x′, y′, and z′ are as follows:

[0171]

[0172] Among them, iter max α is the maximum number of iterations; α is a random number in (0,1]; λ is a random number in [0,2].

[0173] Furthermore, based on the original SSA, the positions of the followers and the scouts are updated. The reverse sparrow population is iteratively updated based on the updated positions of the discoverers, followers, and scouts until the maximum number of iterations is reached, or the fitness of individuals in the updated reverse sparrow population reaches a preset value.

[0174] Finally, the position of the individual with the best fitness is obtained from the reverse sparrow population after the last update. Based on the combined key model parameters corresponding to the position of the individual with the best fitness, the initial model parameters of the extreme value gradient boosting model are determined.

[0175] The method provided in this embodiment addresses the weakness of the original sparrow optimization algorithm in global search capability by introducing reverse learning and spiral search strategies to improve the original sparrow optimization algorithm. This improves the global search capability of the original sparrow optimization algorithm, and then optimizes the key parameters of the XGBoost model based on the improved sparrow optimization algorithm. It can quickly and accurately find the initial model parameters of the extreme gradient boosting model globally, thereby improving the convergence efficiency of the extreme gradient boosting model and thus improving the efficiency and accuracy of single-crystal silicon edge breakage prediction.

[0176] like Figure 2 As shown, the complete implementation steps for training the prediction model provided in this embodiment include:

[0177] First, the data tables of equal diameter growth stages for sample monocrystalline silicon over multiple months are merged to obtain historical equal diameter growth stage data for the sample monocrystalline silicon. Next, missing values ​​are imputed and corrected in the historical equal diameter growth stage data, and redundant parameters are removed using mutual information correlation coefficients. Then, time-domain feature extraction, normalization, and PCA dimensionality reduction are performed on the removed parameters. Next, the weight coefficients corresponding to the sample monocrystalline silicon are calculated using the K-means clustering algorithm to construct standard equal diameter growth stage data based on these weight coefficients. Then, the differences between the historical equal diameter growth stage data and the standard equal diameter growth stage data are calculated to form differential equal diameter growth stage data, and statistical features are extracted from this data to form differential features. Finally, based on the differential features and edge breakage labels of the sample monocrystalline silicon, and using improved SSA and XGBoost, model training is performed, and the trained prediction model is used to predict edge breakage in monocrystalline silicon. This achieves high efficiency in predicting edge breakage in monocrystalline silicon through data-driven methods and reduces the cost of establishing mechanistic models.

[0178] To verify the effectiveness of the method provided in this embodiment in predicting edge breakage during the constant diameter stage of monocrystalline silicon, experimental verification will be conducted using data from the constant diameter growth stage of monocrystalline silicon measured in a real-world scenario.

[0179] In this experiment, the difference between the constant diameter growth stage data of the sample monocrystalline silicon and the standard constant diameter growth stage data, as well as the edge breakage label of the sample monocrystalline silicon, were used to construct a sample set. The sample set was then divided proportionally, for example, the ratio between training data and test data was set to 7:3 to obtain the training dataset and the test dataset.

[0180] The XGBoost model was optimized using the training dataset and the improved SSA to obtain the prediction model;

[0181] During training, whether or not the single-crystal silicon experiences edge breakage is used as a label. The difference features of the differential equal diameter growth stage data are used as input to the XGBoost model, and whether or not edge breakage occurs in the end is used as the model output.

[0182] During testing, a test set is used to verify the accuracy and effectiveness of the trained prediction model.

[0183] For the binary classification problem between broken edge samples and normal samples, the sample points are classified into four cases based on the combination of their broken edge label and the broken edge prediction result of the prediction model trained based on improved SSA and XGBoost: true positive, false positive, true negative, and false negative, denoted by TP, FP, TN, and FN, respectively. TP + FP + TN + FN = number of sample points.

[0184] To effectively evaluate the prediction performance of edge breakage in the constant-diameter stage of monocrystalline silicon, this embodiment uses commonly used evaluation metrics of binary classification models for prediction performance assessment. Specific evaluation metrics include:

[0185] Accuracy is a commonly used metric for evaluating the performance of classification models. It measures the proportion of samples correctly classified by the model. The formula for calculating accuracy is as follows:

[0186]

[0187] Precision is a performance metric used to evaluate classification models. It measures the proportion of samples that the model predicts as positive, but which are actually positive. The specific formula is:

[0188]

[0189] Recall is a performance metric used in classification models. It measures the proportion of samples that are actually positive, but which are correctly predicted as positive by the model. The specific formula is:

[0190]

[0191] The F1-Score is a comprehensive metric that considers both precision and recall, used to evaluate the performance of a classification model. It is the harmonic mean of precision and recall, and helps to comprehensively assess the model's accuracy in predicting positive samples and its ability to capture positive samples. The formula for calculating the F1-Score is as follows:

[0192]

[0193] As shown in Table 2, comparing the performance of the prediction model trained based on the improved SSA and XGBoost provided in this embodiment with the prediction model obtained by supervised training of XGBoost based on the original sample set (hereinafter also referred to as the original XGBoost), and the prediction model trained based on the original SSA and XGBoost (hereinafter also referred to as SSA-XGBoost), it can be seen that the prediction model trained based on the improved SSA and XGBoost provided in this embodiment significantly outperforms the original XGBoost and SSA-XGBoost in all evaluation metrics.

[0194] Table 2 Comparison of Predictive Performance of Edge Breakage in Monocrystalline Silicon

[0195]

[0196] The following describes the monocrystalline silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement provided by the present invention. The monocrystalline silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement described below can be referred to in correspondence with the monocrystalline silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement described above.

[0197] Figure 6 A schematic diagram of the single-crystal silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement provided by the present invention; as shown. Figure 6 As shown, the device includes:

[0198] The acquisition unit 610 is used to acquire the current constant diameter growth stage data of the monocrystalline silicon to be predicted in the current time period;

[0199] The calculation unit 620 is used to calculate the difference between the current equal diameter growth stage data and the standard equal diameter growth stage data to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon.

[0200] The prediction unit 630 is used to extract statistical features from the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted, and input the current differential features into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted.

[0201] The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference characteristics and broken edge labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

[0202] The monocrystalline silicon edge breakage prediction device provided in this embodiment, based on improved SSA and extreme value gradient enhancement, optimizes the model training by fusing improved SSA and XGBoos. It constructs a high-performance prediction model from the perspectives of model method, data method, and fusion method. The constructed prediction model can better capture the trend information of the constant diameter growth characteristics of monocrystalline silicon in various scenarios through data analysis, effectively reducing the dependence on the mechanism model, and thus more efficiently and accurately predicting monocrystalline silicon edge breakage in various scenarios.

[0203] Figure 7 An example is a schematic diagram of the physical structure of an electronic device, such as... Figure 7As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730, and a communication bus 740, wherein the processor 710, the communication interface 720, and the memory 730 communicate with each other through the communication bus 740. The processor 710 can call logic instructions in the memory 730 to execute a single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient boosting. This method includes: acquiring the current equal-diameter growth stage data of the single-crystal silicon to be predicted within the current time period; calculating the difference between the current equal-diameter growth stage data and the standard equal-diameter growth stage data to obtain the current differential equal-diameter growth stage data corresponding to the single-crystal silicon to be predicted; the standard equal-diameter growth stage data is determined by weighting coefficients formed by clustering historical equal-diameter growth stage data of multiple sample single-crystal silicons; extracting statistical features from the current differential equal-diameter growth stage data to obtain the current difference features of the single-crystal silicon to be predicted; inputting the current difference features into the prediction model to obtain the edge breakage prediction result of the single-crystal silicon to be predicted; wherein the prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference features and edge breakage labels of each sample single-crystal silicon, and the improved Sparrow Optimization Algorithm (SSA); the improved SSA is constructed based on a back-learning strategy, a spiral search strategy, and the original SSA.

[0204] Furthermore, the logical instructions in the aforementioned memory 730 can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, essentially, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0205] On the other hand, the present invention also provides a computer program product, which includes a computer program that can be stored on a non-transitory computer-readable storage medium. When the computer program is executed by a processor, the computer is able to execute the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement provided by the above methods. The method includes: acquiring the current equal diameter growth stage data of the single-crystal silicon to be predicted in the current time period; calculating the difference between the current equal diameter growth stage data and the standard equal diameter growth stage data to obtain the current difference equal diameter growth stage data corresponding to the single-crystal silicon to be predicted; the standard equal diameter growth stage... The segment data is determined by weighting coefficients formed by clustering historical equal-diameter growth stage data of multiple sample monocrystalline silicon. Statistical features are extracted from the current difference equal-diameter growth stage data to obtain the current difference features of the monocrystalline silicon to be predicted. The current difference features are input into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted. The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference features and edge breakage labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the back-learning strategy, the spiral search strategy, and the original SSA.

[0206] In another aspect, the present invention also provides a non-transitory computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme gradient enhancement provided by the methods described above. This method includes: acquiring current constant diameter growth stage data of the single-crystal silicon to be predicted within the current time period; calculating the difference between the current constant diameter growth stage data and standard constant diameter growth stage data to obtain current difference constant diameter growth stage data corresponding to the single-crystal silicon to be predicted; the standard constant diameter growth stage data is based on multiple sample single-crystal silicon... The weighting coefficients are determined by clustering historical equal-diameter growth stage data; statistical features are extracted from the current differential equal-diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted, and the current differential features are input into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted; wherein, the prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical differential features and edge breakage labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA); the improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

[0207] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs. Those skilled in the art can understand and implement this without any creative effort.

[0208] Through the above description of the embodiments, those skilled in the art can clearly understand that each embodiment can be implemented by means of software plus necessary general-purpose hardware platforms, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0209] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for predicting edge breakage in single-crystal silicon based on improved SSA and extreme gradient enhancement, characterized in that, include: Obtain the current constant diameter growth stage data of the monocrystalline silicon to be predicted within the current time period; The difference between the current equal diameter growth stage data and the standard equal diameter growth stage data is calculated to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon. Statistical features are extracted from the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted. The current differential features are then input into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted. The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference characteristics and broken edge labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

2. The method for predicting single-crystal silicon edge breakage based on improved SSA and extreme value gradient enhancement according to claim 1, characterized in that, The standard isodiameter growth stage data were obtained through the following steps: Obtain historical data on the constant diameter growth stages of each sample monocrystalline silicon within a historical time period; The historical isodiameter growth stage data are preprocessed; the preprocessing includes data merging, missing value imputation, outlier correction, and redundant parameter screening. Statistical features were extracted from the preprocessed historical constant diameter growth stage data to obtain the historical statistical features of the single-crystal silicon samples. Based on the historical statistical characteristics, the historical constant diameter growth stage data of each sample of monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs. The weighting coefficients corresponding to the historical constant diameter growth stage data of each sample monocrystalline silicon are determined based on the distance between the historical constant diameter growth stage data of each sample monocrystalline silicon and the cluster center of the cluster to which it belongs. The standard equal diameter growth stage data are obtained based on the weighting coefficients and the historical equal diameter growth stage data of each sample monocrystalline silicon.

3. The method for predicting single-crystal silicon edge breakage based on improved SSA and extreme value gradient enhancement according to claim 2, characterized in that, The step of clustering the historical constant diameter growth stage data of each sample of monocrystalline silicon based on the historical statistical characteristics to obtain the cluster to which the historical constant diameter growth stage data of each sample of monocrystalline silicon belongs includes: The historical statistical features are normalized. Principal component dimensionality reduction is performed on the normalized historical statistical features; Based on the historical statistical characteristics after dimensionality reduction, the historical constant diameter growth stage data of each sample monocrystalline silicon are clustered to obtain the cluster to which the historical constant diameter growth stage data of each sample monocrystalline silicon belong.

4. The method for predicting single-crystal silicon edge breakage based on improved SSA and extreme value gradient enhancement according to any one of claims 1-3, characterized in that, The prediction model was trained based on the following steps: The difference between the historical constant diameter growth stage data and the standard constant diameter growth stage data of each sample monocrystalline silicon is calculated to obtain the historical difference constant diameter growth stage data corresponding to each sample monocrystalline silicon. Statistical feature extraction was performed on the historical difference data of the equal diameter growth stage to obtain the historical difference characteristics of each sample of monocrystalline silicon. Based on the improved SSA, the hyperparameters of the extreme gradient boosting model are optimized to obtain the initial model parameters of the extreme gradient boosting model. The extreme value gradient boosting model is initialized according to the initial model parameters, and the historical difference features are input into the initialized extreme value gradient boosting model to obtain the edge breakage prediction results of each sample of single crystal silicon. Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, and the complexity of the initialized extreme value gradient boosting model, the target loss function is obtained. Based on the target loss function, the model parameters of the initialized extreme value gradient boosting model are iteratively optimized to obtain the prediction model.

5. The method for predicting single-crystal silicon edge breakage based on improved SSA and extreme value gradient enhancement according to claim 4, characterized in that, The step of obtaining the target loss function based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, and the complexity of the initialized extreme value gradient boosting model, includes: Based on the deviation between the predicted edge breakage results of each sample monocrystalline silicon and the edge breakage labels of each sample monocrystalline silicon, an error loss function is obtained. The complexity of the initialized extreme gradient boosting model is obtained based on the number of leaf nodes and the leaf node weights of the initialized extreme gradient boosting model. The target loss function is obtained by weighted fusion of the error loss function and the complexity.

6. The method for predicting single-crystal silicon edge breakage based on improved SSA and extreme value gradient enhancement according to claim 4, characterized in that, The process of optimizing the hyperparameters of the extreme value gradient boosting model based on the improved SSA to obtain the initial model parameters of the extreme value gradient boosting model includes: According to the reverse learning strategy, the initial sparrow population generated by the original SSA is initialized to obtain the reverse sparrow population; the position of each individual in the initial sparrow population is determined based on various preset combinations of key model parameters of the extreme value gradient boosting model. Calculate the fitness of each individual in the reverse sparrow population, and determine the discoverers, followers, and scouts in the reverse sparrow population based on the fitness. According to the spiral search strategy, the spiral radius of the discoverer is updated, and the position of the discoverer is updated according to the updated spiral radius; The positions of the follower and the scout are updated based on the original SSA; The reverse sparrow population is iteratively updated based on the updated locations of the discoverers, followers, and scouts until a preset termination condition is met. The preset termination condition includes the number of iterations reaching the maximum number of iterations, or the fitness of individuals in the updated reverse sparrow population reaching a preset value. The initial model parameters of the extreme value gradient boosting model are determined based on the position of the individual with the best fitness in the reverse sparrow population after the last update.

7. A single-crystal silicon edge breakage prediction device based on improved SSA and extreme value gradient enhancement, characterized in that, include: The acquisition unit is used to acquire data on the current constant diameter growth stage of the monocrystalline silicon to be predicted within the current time period; The calculation unit is used to calculate the difference between the current equal diameter growth stage data and the standard equal diameter growth stage data to obtain the current difference equal diameter growth stage data corresponding to the monocrystalline silicon to be predicted; the standard equal diameter growth stage data is determined by weighting coefficients formed by clustering historical equal diameter growth stage data of multiple sample monocrystalline silicon. The prediction unit is used to extract statistical features from the current differential equal diameter growth stage data to obtain the current differential features of the monocrystalline silicon to be predicted, and input the current differential features into the prediction model to obtain the edge breakage prediction result of the monocrystalline silicon to be predicted. The prediction model is obtained by optimizing the extreme value gradient boosting model based on the historical difference characteristics and broken edge labels of each sample monocrystalline silicon, as well as the improved Sparrow Optimization Algorithm (SSA). The improved SSA is constructed based on the reverse learning strategy, the spiral search strategy, and the original SSA.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement as described in any one of claims 1 to 6.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement as described in any one of claims 1 to 6.

10. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the single-crystal silicon edge breakage prediction method based on improved SSA and extreme value gradient enhancement as described in any one of claims 1 to 6.

Images