A polycrystalline silicon resistivity detection method based on principal component analysis and neural network model

Abstract

The application is a kind of polycrystalline silicon resistivity detection method based on principal component analysis and neural network model, comprising: obtaining historical rectification section heavy removal tower tank liquid daily full analysis detection data and reducing furnace out of furnace product resistivity data, removing most class samples, replacing undetected data with detection limit 1 / 2, obtaining data set X i and product resistivity data Y; data set X i standardization processing, principal component analysis, obtaining principal component function and data set F i ; according to data set F i , product resistivity data Y, obtaining data set F' i , product resistivity data Y'; establishing the corresponding relationship between product resistivity data Y' and data set F' i , constructing ANN neural network model, obtaining model NET; standardizing new rectification section heavy removal tower tank liquid daily full analysis detection data, substituting into principal component function, obtaining data set F new , inputting into model NET, obtaining out of furnace polycrystalline silicon resistivity ρ', correcting, obtaining daily out of furnace resistivity ρ. The application solves the existing problems such as hysteresis of prior art.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of polycrystalline silicon and particularly relates to a polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model. BACKGROUND

[0002] At present, the production process of electronic-grade polycrystalline silicon mainly adopts a modified Siemens method, and the main process is as follows: silicon powder reacts with hydrogen chloride and hydrogen in a cold hydrogenation synthesis furnace to produce trichlorosilane (TCS), silicon tetrachloride and other chlorosilane mixtures, and the chlorosilane mixtures form a material circulation together with the recovered chlorosilane in a reduction process tail gas. B and P impurities enter the rectification system together with the silicon powder and the material circulation. High-purity trichlorosilane obtained in the rectification section is gradually deposited into a silicon rod on the surface of a high-temperature silicon core by hydrogen reduction, and polycrystalline silicon products are obtained.

[0003] Because the B, P and metal impurity contents in the cold hydrogenation raw material silicon powder fluctuate in batches, the dynamic characteristics of the production process make the quality of high-purity TCS in the rectification section unable to remain stable, and whether PH3 is contained in the hydrogen in the tail gas recovery section will directly affect the quality of the polycrystalline silicon products. Because the finished polycrystalline silicon needs to undergo a gas phase deposition process of about 100 h, the physical and chemical indexes are also difficult to measure online, and therefore the prediction of the finished product resistivity is extremely important.

[0004] At present, the impurity content in the high-purity TCS buffer tank entering the reduction furnace is usually detected daily in production, and the impurity removal level of the rectification and tail gas is artificially judged in combination with the finished product resistivity data after the furnace is stopped, so that the rectification operation is adjusted. However, the prior art has the following problems: (1) the strategy adopted in production has obvious hysteresis, and the quality of the finished products out of the furnace is a fact, and there is no possibility to make up for it. (2) the detection of the key index monitoring point of the rectification section in production is to judge the operation level of the rectification according to experience combined with the finished product resistivity. This method seriously depends on the experience of the operator. Because the finished product resistivity and the numerous impurity detection items are not linearly related, the artificial judgment result is also prone to error, and quality accidents occur frequently. Therefore, there is an urgent need for a polycrystalline silicon resistivity detection method which can be dynamically monitored, does not depend on artificial judgment and can realize standardized polycrystalline silicon product resistivity prediction.

[0005] In view of this, the application provides a polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model, which does not depend on the experience of the operator and can realize a standardized polycrystalline silicon product resistivity prediction method. SUMMARY

[0006] The application aims to provide a polysilicon resistivity detection method based on principal component analysis and a neural network model.

[0007] To achieve the above-mentioned purpose, the technical scheme adopted is:

[0008] A polysilicon resistivity detection method based on principal component analysis and a neural network model comprises the following steps:

[0009] S10: Obtain historical rectification section heavy-removing tower kettle liquid daily full analysis detection data and product resistivity data of a reduction furnace;

[0010] S20: After removing most of the sample, the rectification section heavy-removing tower kettle liquid daily full analysis detection data and product resistivity data obtain full analysis detection data set and product resistivity data Y;

[0011] S30: The full analysis detection data set is replaced by 1 / 2 of the detection limit of the analyzer to obtain data set X i ;

[0012] S40: After the data set X i is standardized, principal component analysis is performed to obtain principal components, corresponding characteristic values, load coefficients, contribution rates, principal component functions and reduced data set F i ;

[0013] S50: According to the data set F i , the data set F' i of the average value of the same furnace is obtained, and then according to the product resistivity data Y, the product resistivity data Y' of the last day of the same furnace is obtained, and the corresponding relationship between the product resistivity data Y' and the data set F' i of the same day is established;

[0014] S60: Construct an ANN neural network model, set the data set F' i as the input layer and Y' as the output layer to obtain the model NET;

[0015] S70: Collect new rectification section heavy-removing tower kettle liquid daily full analysis detection data, and after standardization, substitute into the principal component function to obtain real-time data set F new ;

[0016] S80: Input the real-time data set F new into the model NET to obtain the resistivity ρ' of the out-of-furnace polysilicon;

[0017] S90: The tap-out resistivity p' is corrected to obtain the daily tap-out resistivity p.

[0018] Further, in the step S10, the daily analysis and detection data of the tower kettle liquid of the heavy-removing tower in the rectification section include the impurity contents of Ti, Ni, Cr, Zn, Mn, Cu, Ca, As, Al, Fe, P and B.

[0019] Further, in the step S20, the undersampling method is used to randomly remove most of the class samples.

[0020] In the step S40, the formula of the standardization processing is as follows:

[0021] The standardized data set In the formula, i=80, and k=12.

[0022] Further, in the step S20, 10 samples are taken from 0-8000 at intervals of 1000 Ω·cm.

[0023] Further, in the step S40, the principal components are extracted according to the principle that the eigenvalue is greater than 1, and the principal component function and the reduced data set F i are obtained.

[0024] The principal component feature vector Φ is the corresponding load coefficient divided by the square root of the eigenvalue, and the principal component feature vector Φ is multiplied by the standardized data set to obtain the reduced data set F i .

[0025] Further, in the step S40, the principal components are selected in the order of the contribution rate from large to small, and the cumulative contribution rate is not less than 0.9.

[0026] Further, in the step S50, the data of the current day and the data of the previous 3 days, a total of 4 days, are averaged, In the formula, 4≤i≤n.

[0027] In the step S60, the ANN neural network model includes an input layer, a training set, a test set, a hidden layer and an output layer.

[0028] Further, the step S60 further includes training the model with the training set data and testing the model with the test set data until the training and testing results are good, and the model NET is obtained.

[0029] Further, the training set data accounts for 60-70% of the total data, and the training method is the Levenberg-Marguardt method.

[0030] When the correlation coefficients of the training set, the test set and the validation set are all greater than 0.9, the training and the test result are good, that is, the model training is completed.

[0031] Further, in the step S90, the formula of the corrected daily tapping resistivity p is: p = θp' + 100.

[0032] In the formula, θ is a correction term, when PH3 in hydrogen is detected, θ = 0, otherwise θ = 1.

[0033] Compared with the prior art, the present application has the beneficial effects that:

[0034] 1. The polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model can predict the product resistivity according to the detection data of the tower kettle liquid of the heavy-removing tower in the rectification section, and avoids the delay caused by detecting the product resistivity after the silicon rod is tapped.

[0035] 2. The polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model can predict the product resistivity, correct the rectification parameters and tail gas parameters in time, improve the purity of high-purity trichlorosilane into the reduction furnace, and reduce the rate of defective products.

[0036] 3. The polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model does not rely on the experience of operators, and can realize standardized prediction of the product resistivity of polycrystalline silicon. BRIEF DESCRIPTION OF DRAWINGS

[0037] Figure 1 The flowchart of the present application is shown in the figure.

[0038] Figure 2 The experimental result graph is shown in the figure. DETAILED DESCRIPTION

[0039] In order to further illustrate the polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model, and achieve the intended purpose of the present application, the polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model according to the present application will be described in detail below in combination with preferred embodiments, the specific implementation, structure, features and effects of which are as follows. In the following description, different "an embodiment" or "embodiments" do not necessarily refer to the same embodiment. In addition, the specific features, structures or characteristics in one or more embodiments can be combined in any suitable form.

[0040] The polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model will be further described in detail below in combination with specific embodiments:

[0041] There is no polycrystalline silicon finished product resistivity prediction method in the prior art. There is a method of principal component analysis combined with a neural network model for predicting the quality of other chemical products or process product quality in the prior art. Principal component analysis is a dimension reduction method that extracts the main components from numerous influencing factors, and the neural network model is a nonlinear regression model based on big data, which has good convergence and small evaluation error. Based on this, the present application applies principal component analysis and a neural network model to polycrystalline silicon resistivity detection. However, the reason why the existing principal component analysis combined with a neural network model technology has not been applied in the polycrystalline silicon industry is related to the characteristics of the polycrystalline silicon industry. When the applicant applied it to polycrystalline silicon resistivity detection, the following problems were found:

[0042] (1) The high-purity TCS metal impurity content of the reduction furnace sample is low, close to or lower than the detection limit, the analysis result is inaccurate, and the model training cannot converge.

[0043] (2) In most cases, the polycrystalline silicon resistivity in the historical data is high, and the low resistivity sample is small, and the data is uneven, and the principal component analysis algorithm is sensitive to the distribution of the data, and if the data is uneven, the effect of the algorithm will be affected.

[0044] (3) The metal impurity content in high-purity TCS can be regarded as a continuous variable that changes, while PH3 in the tail gas recovery hydrogen is only marked as detected and not detected, which can be regarded as a discrete variable, and the principal component analysis method cannot simultaneously reduce the dimension of continuous variables and discrete variables.

[0045] The applicant has carried out further research and proposed a polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model. The technical scheme of the present application is as follows:

[0046] A polycrystalline silicon resistivity detection method based on principal component analysis and a neural network model, comprising the following steps:

[0047] S10: Obtain historical rectification section heavy-removing tower kettle liquid daily full analysis detection data and finished product resistivity data of the reduction furnace;

[0048] S20: The rectification section heavy-removing tower kettle liquid daily full analysis detection data and the finished product resistivity data are removed from the majority of samples to obtain the full analysis detection data set and the finished product resistivity data Y;

[0049] S30: The full analysis detection data set is replaced with 1 / 2 of the detection limit of the analyzer to obtain the data set X i ;

[0050] S40: After standardizing the data set X i , principal component analysis is performed to obtain the principal components, the corresponding characteristic values, the load coefficients, the contribution rates, the principal component functions and the reduced data set Fi ;

[0051] S50: According to the data set F i , the data set F' i is obtained, and then according to the finished product resistivity data Y, the finished product resistivity data Y' of the last day of the same furnace is obtained, and the correspondence between the finished product resistivity data Y' and the data set F i of the day is established;

[0052] S60: Constructing an ANN neural network model, setting the data set F' i as the input layer and Y' as the output layer to obtain the model NET;

[0053] S70: Collecting daily full analysis and detection data of the distillation section heavy-removing tower bottom liquid, performing standardization processing, and substituting into the principal component function to obtain the real-time data set F new ;

[0054] S80: Inputting the real-time data set F new into the model NET to obtain the resistivity p ’ of the furnace polysilicon;

[0055] S90: The furnace resistivity p' is corrected to obtain the daily furnace resistivity p.

[0056] In the above technical solution, the improvement points are:

[0057] (1) Since the impurity content of the reduction furnace sample is low, the process between the reduction furnace and the distillation section heavy-removing tower is stable. Therefore, the sampling point is moved to the distillation section heavy-removing tower bottom liquid, and effective impurity content data can be obtained.

[0058] (2) The sampling point is moved to the distillation section heavy-removing tower bottom liquid in the present application, although most of the effective data can be obtained, but there are still data close to or lower than the detection limit. The present application uses 1 / 2 of the detection limit of the analysis instrument to replace the "undetected" data, avoiding the influence of invalid data on the analysis results.

[0059] (3) The present application realizes sample balancing by under-sampling, randomly removing some majority class samples according to the finished product resistivity gradient.

[0060] Preferably, in the step S10, the daily full analysis and detection data of the distillation section heavy-removing tower bottom liquid includes Ti, Ni, Cr, Zn, Mn, Cu, Ca, As, Al, Fe, P, B and other impurity contents.

[0061] Preferably, in the step S20, under-sampling is used to randomly remove majority class samples;

[0062] The formula of the standardization processing in step S40 is as follows:

[0063] The data set after standardization In the formula, i=80, k=12.

[0064] Further preferably, in step S20, samples are taken at intervals of 1000 Ω·cm from 0-8000, with 10 samples in each interval.

[0065] Preferably, in step S40, principal components are extracted according to the principle that the eigenvalue is greater than 1, to obtain principal component functions and a data set F after dimension reduction i ;

[0066] The principal component feature vector Φ is the corresponding load coefficient divided by the square root of the eigenvalue, and the principal component feature vector Φ is multiplied by the data set after standardization to obtain the data set F after dimension reduction i .

[0067] Further preferably, in step S40, the principal components are selected in descending order of contribution rate, and the cumulative contribution rate is not less than 0.9.

[0068] Preferably, in step S50, the data of the day is averaged with the data of the previous 3 days, for a total of 4 days, In the formula, 4≤i≤n;

[0069] In step S60, the ANN neural network model comprises: an input layer, a training set, a test set, a hidden layer, and an output layer.

[0070] Preferably, step S60 further comprises training the model with training set data and testing the model with test set data until the training and testing results are good, obtaining the model NET.

[0071] Further preferably, the training set data accounts for 60-70% of the total data, and the training method is the Levenberg-Marguardt method.

[0072] When the correlation coefficients of the training set, the test set, and the validation set are all greater than 0.9, the training and testing results are good, and the model training is completed.

[0073] Preferably, in step S90, the formula of the corrected daily tapping resistivity ρ is: ρ=θρ'+100;

[0074] In the formula, θ is a correction term, PH3 is detected in hydrogen, and θ=0, otherwise θ=1.

[0075] In the above technical solution, the detection of PH3 in hydrogen gas is used as a correction. When PH3 is detected in hydrogen gas, the polysilicon product in this reduction furnace can be directly judged as unqualified. Therefore, by using the detection of PH3 in hydrogen gas as a correction, this invention can achieve the effect of accurately judging the quality of polysilicon.

[0076] Example 1.

[0077] Combination Figure 1 The specific operating steps are as follows:

[0078] S1: Obtain daily full analysis results of the bottom liquid from the de-heavy column in the distillation section of the polysilicon production process and historical data of the finished product from the reduction furnace. Specifically:

[0079] The bottom liquid test data of the de-heavy column in the distillation section for the past three years were obtained through the LIMS system interface. This included the content of 12 impurities: Ti, Ni, Cr, Zn, Mn, Cu, Ca, As, Al, Fe, P, and B, denoted as X. n =[X n (1),X n (2(,…X n [(m)]; where n is the batch and m represents the impurity element.

[0080] Collect the finished product resistivity data of the reduction furnace of the corresponding production line in the past three years.

[0081] S2: The daily full analysis results of the bottom liquid of the de-weighting tower in the distillation section of the polysilicon production process and the historical data of the finished product from the reduction furnace are obtained by undersampling. Some majority class samples are randomly removed according to the resistivity gradient of the finished product to achieve sample balance.

[0082] Samples were taken at intervals of 1000 Ω·cm from 0 to 8000, with 10 samples per interval, denoted as X. n ';The corresponding finished product resistivity data of the reduction furnace in the production line is Y=[Y(1),Y(2),…Y(k)].

[0083] S3: Replace dataset X with half of the instrument's detection limit. n Data that was "not detected" in the dataset is dataset X. i X i =[X i (1),X i (2),…X i [(k)], i = 80, k = 12.

[0084] S4: For dataset X i Standardize the dataset using the formula below to obtain the standardized dataset X'. i .

[0085]

[0086] S5: Based on the standardized data set X' obtained in S40 i , principal component analysis is performed to obtain principal components and their corresponding characteristic values, contribution rates, principal component functions and reduced data set F i . Specifically: the principal component analysis method is used to comprehensively evaluate the above samples and 12 detection items, and the principal component analysis is performed on the standardized data set X' i to obtain principal components and their corresponding characteristic values, loading coefficients, contribution rates and cumulative contribution rates. According to the principle of characteristic value > 1, the principal components are extracted; and the principal components are selected in the order of contribution rate from large to small, and the cumulative contribution rate is not less than 0.9.

[0087] The principal component feature vector Φ is equal to the corresponding loading coefficient divided by the square root of the characteristic value, that is,

[0088] Establishment of data set F i : Multiply the obtained feature vector Φ and the standardized variable X' i to obtain the reduced data set F i = ΦX' i .

[0089] S6: Since the reduction furnace gas phase deposition time is about 4 days, based on the data set F i , the average value data set F' i of the continuous 4 days is obtained. That is, according to the data set F i , the average value data set F' i of the same furnace is obtained. Then, the correspondence between the finished product resistance data set Y' and the average data set F' i of the previous 4 days is established. Specifically:

[0090] The average of the data of the current day and the data of the previous 3 days, a total of 4 days, is obtained, where 4 ≤ i ≤ n, to obtain a new data set F' i . The data of the previous 3 days in the finished product resistivity data set Y is deleted to obtain a data set Y', and the data set F' i and the data set Y' have the same number of data, so the correspondence between the finished product resistance data set Y' and the data set F' i is established.

[0091] S7: According to the correspondence between the finished product resistance data set Y' and the data set F' i , an ANN neural network model is constructed, including an input layer, a training set, a test set, a hidden layer and an output layer, wherein F' i is the input layer and Y' is the output layer. Specifically:

[0092] The ANN neural network model is constructed, and F i is set as the input layer, Y' is set as the output layer, the number of hidden layers is set to 10, the number of training sets is 50, the number of test sets and validation sets is 15 respectively, the loss function of the model is set to the root mean square error, and the training method is selected as Levenberg-Marguardt.

[0093] S8: The model is trained with the training data and tested with the test data, and if the training and test results are good, the model NET is saved; otherwise, the training and test are re-performed until the results are good.

[0094] The correlation coefficient > 0.9 is used as the model training result standard, and the finished product resistivity prediction model NET based on principal component analysis and ANN neural network is obtained.

[0095] S9: Collect daily full analysis results of the tower kettle liquid of the de-heavy tower in the rectification section, and standardize the results. For the newly collected samples in the process, the standardization needs to be performed, that is, the content of each component is divided by the average value in the modeling.

[0096] S10: The standardized new data obtained in S9 is substituted into the principal component function expression (data set F i =ΦX' i ) generated in step S5 to obtain a real-time data set, and the prediction input layer F new is obtained.

[0097] S11: The input layer F new data is imported into the NET model, and the resistivity p' of the polycrystalline silicon discharged on the day is obtained.

[0098] S12: A correction term θ is defined, and if PH3 is detected in the hydrogen gas within the day, θ = 0, otherwise θ = 1.

[0099] S13: The corrected resistivity of the polycrystalline silicon discharged on the day is p = θp' + 100.

[0100] The method of example 1 is used to predict the resistivity of the product in the reduction furnace of a certain production line in Inner Mongolia Daquan, and the cycle is 98 days. After prediction, the actual value is compared, and the results are shown in Figure 2 .

[0101] Figure 2 The model training prediction value and the actual value of the method are compared, and it can be seen that the method has a good effect on the finished product resistivity prediction, and can better reflect the change trend of the finished product resistivity.

[0102] Example 2.

[0103] The step of embodiment 2 is the same as that of embodiment 1, except that S11: after obtaining the resistivity p' of the polysilicon discharged on the day, the predicted resistivity data of the day is averaged with the predicted resistivity data of the previous 3 days to obtain the average resistivity data of the silicon rod stopped on the day The corrected resistivity of the polysilicon discharged on the day

[0104] By obtaining the resistivity data of the same furnace for 4 days, the quality of the finished polysilicon can be better monitored.

[0105] The above is only the preferred embodiment of the present embodiment, and does not limit the present embodiment in any form. Any simple modification, equivalent change and modification of the above embodiment according to the technical essence of the present embodiment still belongs to the scope of the technical solution of the present embodiment.

Claims

1. A method for detecting polysilicon resistivity based on principal component analysis and neural network model, characterized in that, The method comprises the following steps: S10: obtaining historical daily full analysis detection data of the distillation section heavy removal tower kettle liquid and the reduced furnace product resistivity data; S20: removing the majority class samples from the distillation section heavy removal tower kettle liquid daily full analysis detection data and the product resistivity data to obtain the full analysis detection data set and the product resistivity data Y; S30: The total analysis detection data set is replaced with 1 / 2 of the detection limit of the analysis instrument for the non-detected data to obtain data set X i ; S40: normalizing the data set X i After standardization, principal component analysis is performed to obtain principal components, corresponding characteristic values, loading coefficients, contribution rates, principal component functions and reduced data set F i ; S50: According to the data set F i , the data set F ' i After that, according to the finished product resistivity data Y, the finished product resistivity data Y' of the last day of the same furnace is obtained, and the correspondence between the finished product resistivity data Y' of the day and the data set F ' i ; S60: build ANN neural network model, set data set F ' i For the input layer, Y' is the output layer, and the model NET is obtained. S70: Collect daily full analysis test data of the new rectification section heavy-removal column column still liquid, after standardization processing, substitute into the principal component function, get real-time data set F new ; S80: inputting the real-time data set F new into the model NET to obtain the resistivity p' of the polysilicon. S90: correcting the resistivity p' of the reduced polycrystalline silicon to obtain the daily reduced resistivity p.

2. The polycrystalline silicon resistivity detection method according to claim 1, wherein the daily full analysis detection data of the distillation section heavy removal tower kettle liquid in step S10 comprises the impurity contents of Ti, Ni, Cr, Zn, Mn, Cu, Ca, As, Al, Fe, P and B.

3. The polycrystalline silicon resistivity detection method according to claim 1, wherein in step S20, the majority class samples are randomly removed by under-sampling; In step S40, the standardization processing formula is as follows:

4. The polycrystalline silicon resistivity detection method according to claim 3, wherein in step S20, 10 samples are taken from each interval at intervals of 1000 Ω·cm from 0-8000.

5. The polycrystalline silicon resistivity detection method according to claim 1, wherein in step S40, the principal components are selected in descending order of contribution rate, and the cumulative contribution rate is not less than 0.

9. Standardized dataset where i = 80, k = 12.

6. The polycrystalline silicon resistivity detection method according to claim 5, wherein in step S40, the principal components are selected in descending order of contribution rate, and the cumulative contribution rate is not less than 0.

9.

7. The polycrystalline silicon resistivity detection method according to claim 1, wherein in step S60, the ANN neural network model comprises an input layer, a training set, a test set, a hidden layer and an output layer.

8. The polycrystalline silicon resistivity detection method according to claim 1, wherein step S60 further comprises training the model with the training set data and testing the model with the test set data until the training and testing results are good, and obtaining the model NET. In step S40, according to the principle of eigenvalue > 1, the principal component is extracted, and the principal component function and the reduced dimension data set F are obtained i ; The principal component feature vector Φ is the corresponding load coefficient divided by the square root of the eigenvalue, the principal component feature vector Φ is multiplied by the normalized data set, and a reduced dimension data set F is obtained i .

9. The polycrystalline silicon resistivity detection method according to claim 8, wherein the training set data accounts for 60-70% of the total data, and the training method is the Levenberg-Marguardt method; When the correlation coefficients of the training set, the test set and the validation set are all greater than 0.9, the training and testing results are good, and the model training is completed.

10. The polycrystalline silicon resistivity detection method according to claim 1, wherein In step S50, the data of the day is averaged with the data of the previous 3 days, for a total of 4 days, where 4≤i≤n. ​ ​ ​ ​ ​ ​ ​ The formula of the corrected daily tapping resistivity p in step S90 is: p=0p ' +100; wherein, 0 is a correction term, PH3 is detected in hydrogen, and 0=0 if PH3 is detected, otherwise 0=1.

Images