Polycrystalline silicon reduction furnace tail temperature prediction method based on convolutional network

Abstract

The invention relates to a polycrystalline silicon reduction furnace tail temperature prediction method based on a convolutional network, and the method comprises the steps: reading the issuing data and feedback data of a historical heat, carrying out the preprocessing of an issuing material table, the water temperature and flow of a furnace cylinder, and the voltage and tail temperature data; global information extraction is carried out on the lower charging table, the water temperature and flow of the furnace barrel and the voltage data through a pyramid pooling technology; according to the method, a convolutional network model used for tail temperature prediction is established, an original data set is constructed into a sample set conforming to input and prediction of the established convolutional network model, the sample set is further divided into a training set, a test set and a verification set, and then training and tuning of the convolutional network model are completed. According to the method, the periodicity and tendency characteristics reflected in the change process of the tail temperature of the reduction furnace are comprehensively considered, meanwhile, the change trends of the tail temperature in the short term and the long term are considered, and the accuracy and robustness of prediction of the tail temperature of the reduction furnace can be improved.

Description

Technical Field

[0001] The invention relates to a method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network, and belongs to the technical field of intelligent detection and prediction. Background Art

[0002] The polysilicon reduction furnace is a crucial piece of equipment for producing high-purity polysilicon. Controlling the reaction state within the furnace plays a key role in its production. Currently, accurate data on the furnace's internal conditions cannot be directly obtained through sensors. On-site engineers primarily rely on direct observation of furnace images and analysis of tail temperature data to determine and control the reaction state within the furnace.

[0003] However, this method of regulating the furnace status based on manual judgment has a large lag. On-site personnel can only identify the problem through images or data after it occurs, and thus often miss the best time to regulate. Therefore, whether it is possible to make advance predictions about the future furnace status based on system history and real-time data is of great practical significance.

[0004] Currently, the data acquisition of furnace images is costly due to the lack of appropriate acquisition equipment and the need to modify hardware equipment. However, the tail temperature data of the reduction furnace is relatively easy to obtain. The change of the tail temperature of the reduction furnace with various inputs is a highly nonlinear and complex dynamic process. The intermediate variables that affect the change of tail temperature are numerous and complex, making it difficult to fit by directly establishing a mathematical convolutional network model. By utilizing the self-learning, self-organizing, and adaptive characteristics of artificial neural network technology, the input samples are automatically calculated to obtain the time-varying patterns between various inputs and tail temperature, thereby making accurate predictions of tail temperature in the future and providing a technical basis for judging the furnace status and subsequent regulation. Summary of the Invention

[0005] In order to solve the above problems, the present invention discloses a method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network. The specific technical solution is as follows:

[0006] A method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network comprises the following steps:

[0007] Step (1) reads the historical heat delivery data and feedback data from the DCS system, specifically including response time, delivery table, furnace water temperature and flow rate, voltage and tail temperature data;

[0008] Step (2) preprocessing the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data, including processing of missing values ​​and abnormal values ​​and normalization of data;

[0009] Step (3) extracts global information from the sent data and feedback data in step (1) by using pyramid pooling technology to obtain new feature data;

[0010] Step (4) building a convolutional network model for tail temperature prediction, the convolutional network model includes a global feature extraction layer, a local feature extraction layer, a time embedding layer and an output layer;

[0011] Step (5) constructs the new feature data obtained in step (3) into a sample set that meets the input and prediction of the convolutional network model constructed in step (4), and further splits the sample set into a training set, a test set, and a validation set, thereby completing the training and tuning of the convolutional network model.

[0012] Furthermore, the frequency of sending and feeding back data during the reaction time in step (1) is 1 minute / time. In a complete reaction cycle, the reaction time starts from the first minute and accumulates to the end of the entire furnace.

[0013] Furthermore, the missing values ​​in step (2) are processed by filling up the missing values. The specific process is as follows:

[0014] m t =x t-p

[0015]

[0016] The formula represents the position of the nearest non-missing sample before time t. For example, the missing sample at the 10th minute is recorded as m 10 , all the non-missing samples in history are x0, x1, ... x9, then the non-missing sample value with p = 9 is taken as the filling value.

[0017] p'∈[0,t-1]

[0018] where m t is the missing sample at time t, x t-p is the sample at time tp, p is a moment before t, and p' is the entire time range from the beginning to time t, from 0 minutes to t-1 minutes, that is, the missing values ​​are filled with the non-null values ​​of the most recent moment before it.

[0019] Furthermore, the processing of outliers in step (2) is based on the 3σ principle in a limited time domain, and the specific process is as follows:

[0020] Assume x t represents the sample at time t, m is the length of the statistical time domain, u t and σ t For the calculated mean and standard deviation statistics, for x t Do the following:

[0021] If x t ≥u t +3σ t , then x t =u t +3σ t ;

[0022] If x t ≤u t +3σ t , then x t =u t -3σ t ;

[0023] in

[0024]

[0025] Where m is the total sample size, and i is the time index when traversing each sample.

[0026] Furthermore, in step (2), the original feature variable x is normalized, and the specific process is as follows:

[0027]

[0028] Where x is the original feature variable, x max and x min are the maximum and minimum values ​​of the original feature variable x, is the normalized result.

[0029] Furthermore, in step (3), global information is extracted based on the pyramid pooling technology, and the specific calculation is as follows:

[0030]

[0031] where x i is the sample at the i-th moment, represents the mean of the samples from the 1st to the tth moment; T represents transpose, and vectors are generally displayed in column form; z t,k It represents the extraction result obtained when k features are extracted from the historical sequence at time t;

[0032] t / / k refers to the value obtained by dividing k by t, which means rounding operation;

[0033] Z t,{k1,k2…kn}It means that n extraction strategies are used for feature extraction at time t, where {k1, k2…kn} represents the parameters of n different extraction strategies. After obtaining the corresponding pyramid features, the extracted results will be spliced ​​again by column to obtain new feature data, which will be used as the input of the subsequent convolutional network model.

[0034] Furthermore, the convolutional network model building process in step (4) is as follows:

[0035] 4.1) Use pyramid pooling technology to complete the extraction of global features and build a global feature extraction layer.

[0036] 4.2) Based on the two-dimensional convolution technology, the local features are extracted and the local feature extraction layer is constructed.

[0037] Two-dimensional convolution technology extracts specific feature information by sliding a small matrix called a convolution kernel or filter on the input. In the two-dimensional convolution layer, the convolution kernel is a two-dimensional matrix that is multiplied and summed element-by-element with the input features to obtain the output feature map.

[0038] 4.3) Add embedding of time features and build time embedding layer.

[0039] 4.4) Concatenate the data results of the global feature extraction layer, local feature extraction layer and time embedding layer constructed in steps 4.1)-4.3) by column.

[0040] 4.5) The data concatenation result obtained in step 4.4) is introduced into the fully connected layer, and the final prediction result is output through the output layer.

[0041] Furthermore, the convolutional network model in step (4) includes a global feature extraction layer, a local feature extraction layer, a time embedding layer and an output layer, wherein:

[0042] The input of the global feature extraction layer is the information extracted in step (3), which is composed of two fully connected layers, and its purpose is to learn the information of sequence history;

[0043] The input of the local feature extraction layer is a fixed-length sequence of W times the input sample time, where W is an integer multiple of hours. The convolutional network structure is built using convolutional layers. Based on the two-dimensional convolutional layer, feature extraction is performed along the time dimension to learn the recent tail temperature trend information of the sequence.

[0044] The time embedding layer embeds the reaction time to obtain the hidden features of time, with the aim of obtaining the implicit information of the tail temperature trend and relative size at the current time node;

[0045] The output layer obtains the final prediction result by splicing the output results of the global feature extraction layer, local feature extraction layer and embedding layer together for re-learning.

[0046] Furthermore, the step (5) constructs the new feature data obtained in step (3) into a sample set in minutes, wherein the sample set includes global features, local features, time features and data labels;

[0047] The global feature is the new feature data constructed in step (3); the time feature is the moment t of the current reaction;

[0048] The local features and data labels are constructed as follows:

[0049] Assume that the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data samples at time t are x t , the actual tail temperature data is y t , then:

[0050] X t =[x t-s ,x t-s+1 ,…x t ] T

[0051] s≥1

[0052] Y t =y t+1

[0053] where X t and Y t are the local features and data labels at time t, and s is the window size for local feature extraction.

[0054] Furthermore, the training and evaluation of the convolutional network model uses MSE as its evaluation indicator, and the specific form is:

[0055]

[0056] where f(x j ) and y j Represent the predicted value and actual value of the jth sample respectively, and N is the total number of samples.

[0057] The beneficial effects of the present invention are:

[0058] The present invention comprehensively considers the periodicity and trend characteristics reflected in the change process of the reduction furnace tail temperature. Based on convolutional networks and embedding technology, historical information is put into three sub-networks: the global feature extraction layer, the local feature extraction layer, and the time embedding layer for learning respectively. At the same time, the short-term and long-term change trends of the tail temperature are taken into account, which can improve the accuracy and robustness of the reduction furnace tail temperature prediction. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of the reduction furnace tail temperature prediction method of the present invention;

[0060] Figure 2 A schematic diagram of a convolutional network according to an embodiment of the present invention;

[0061] Figure 3 This is a schematic diagram of data set construction according to an embodiment of the present invention. DETAILED DESCRIPTION

[0062] The present invention will be further described below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not intended to limit the scope of the present invention.

[0063] like Figure 1 As shown, the tail temperature prediction method of the present invention includes the following steps:

[0064] S1. Read the historical heat delivery and feedback data from the DCS system, including response time, delivery table, furnace water temperature and flow rate, voltage and tail temperature data;

[0065] S2. Preprocess the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data, mainly including the processing of missing values ​​and outliers and data normalization;

[0066] The method for handling missing values ​​is to fill in the values ​​upwards, as follows:

[0067] x t =x t-p

[0068]

[0069] p∈[0,t-1

[0070] where x t is the sample value at time t, and p is the time before t.

[0071] The processing of outliers is based on the 3σ principle within a limited time domain, as follows:

[0072] Assume x t represents the sample at time t, m is the length of the statistical time domain, u t and σ t For the calculated mean and standard deviation statistics, for x t Do the following:

[0073] When x t ≥u t +3σ t When xt =u t +3σ t ;

[0074] When x t ≤u t +3σ t When x t =u t -3σ t ;

[0075] in

[0076]

[0077] The data is normalized as follows:

[0078]

[0079] Where x is the original feature variable, x max and x min are the maximum and minimum values ​​of the variable respectively, and x is the normalized result.

[0080] S3, using pyramid pooling technology to extract global information from the material distribution table, furnace water temperature and flow rate, and voltage data;

[0081] The specific calculation of extracting global information based on pyramid pooling technology is as follows:

[0082]

[0083] s=t / / k

[0084] Z t，k =[z t，k1 , z t，k2 ,...z t，kn ] T

[0085] where x i is the sample at the i-th moment, X 1:t Represents the mean of samples from time 1 to time t.

[0086] z t,k It represents the extraction result obtained when k features are extracted from the historical sequence at time t.

[0087] Z t,k It means that multiple extraction strategies are used for feature extraction at time t, and the extraction results are spliced.

[0088] S4. Build a convolutional network model for tail temperature prediction. The network diagram is as follows: Figure 2 As shown in the figure, the process of building a convolutional network model is:

[0089] 4.1) Use pyramid pooling technology to complete the extraction of global features and build a global feature extraction layer.

[0090] 4.2) Based on the two-dimensional convolution technology, the local features are extracted and the local feature extraction layer is constructed.

[0091] Two-dimensional convolution technology extracts specific feature information by sliding a small matrix called a convolution kernel or filter on the input. In the two-dimensional convolution layer, the convolution kernel is a two-dimensional matrix that is multiplied and summed element-by-element with the input features to obtain the output feature map.

[0092] 4.3) Add embedding of time features and build time embedding layer.

[0093] 4.4) Concatenate the data results of the global feature extraction layer, local feature extraction layer and time embedding layer constructed in steps 4.1)-4.3) by column.

[0094] 4.5) The data concatenation result obtained in step 4.4) is introduced into the fully connected layer and the final prediction result is output.

[0095] The input of the convolutional network model consists of three parts: global features, local features, and temporal features; see Figure 2 After the global features are processed by the convolution layer, all historical information from the beginning to the present is learned; after the local features are processed by the fully connected layer 1, recent trend information is learned; after the time features are processed by the embedding layer, the hidden information at that moment is learned; the network data of the above three parts will be spliced ​​together next, and then passed through the fully connected layer 2 to obtain the final prediction output.

[0096] S5. The new feature data obtained in step S3 is constructed into a sample set that meets the input and prediction of the convolutional network model constructed in step S4, and is further divided into a training set, a test set, and a validation set according to a certain ratio; the training set and the test set of the convolutional network model are used to complete the training and parameter tuning of the convolutional network model, and the validation set is used to evaluate the final effect of the convolutional network model.

[0097] The sample set input and prediction of the convolutional network model includes global features, local features, time features and data labels;

[0098] The global feature is the new feature data constructed in step S3; the time feature is the moment t of the current reaction;

[0099] The local features and data labels are constructed as follows:

[0100] Assume that the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data samples at time t are xt , the actual tail temperature data is y t , then:

[0101] X t =[x t-s ,x t-s+1 ,…x t ] T

[0102] s≥1

[0103] Y t =y t+1

[0104] where X t and Y t are the local features and data labels at time t, and s is the window size for local feature extraction.

[0105] like Figure 3 As shown in the figure, a specific example is shown: assuming a sliding window size of 2, at the 7th minute, the time feature is 7 and the data label is Y8. Using the pyramid pooling technique, all features except time from minutes 1 to 7 are extracted to obtain global features. Information is directly extracted from features from minutes 6 to 7 to obtain local features. The same operation is performed on all data with a time of 2 minutes or more to obtain the entire sample set.

[0106] The obtained sample set is further split into training set, validation set and test set in a ratio of 7:2:1. The training, tuning and evaluation of the convolutional network model use MSE as its evaluation indicator. The specific form is:

[0107]

[0108] where f(x i ) and y i Represent the predicted value and actual value of the i-th sample respectively, and n is the total number of samples.

[0109] The present invention is based on convolutional networks and embedding technology, and puts historical information into three sub-networks: the global feature extraction layer, the local feature extraction layer, and the time embedding layer for learning respectively. It comprehensively considers the periodicity and trend characteristics reflected in the change process of the reduction furnace tail temperature, and can improve the accuracy and robustness of the reduction furnace tail temperature prediction.

[0110] With the above-described preferred embodiments of the present invention as a guide, and with reference to the above description, relevant personnel are fully capable of making various changes and modifications without departing from the technical scope of this invention. The technical scope of this invention is not limited to the contents of the specification and must be determined according to the scope of the claims.

Claims

1. A method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network, characterized in that: The following steps are involved: Step (1) reading the historical furnace delivery data and feedback data from the DCS system, including reaction time, delivery table, furnace water temperature and flow rate, voltage and tail temperature data; Step (2) preprocessing the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data, the preprocessing includes processing of missing values ​​and abnormal values ​​and normalization of data; Step (3) extracts global information from the sent data and feedback data in step (1) by using pyramid pooling technology to obtain new feature data; Step (4) building a convolutional network model for tail temperature prediction, the convolutional network model comprising a global feature extraction layer, a local feature extraction layer, a time embedding layer and an output layer; Step (5) constructs the new feature data obtained in step (3) into a sample set that conforms to the input and prediction of the convolutional network model constructed in step (4), and further splits the sample set into a training set, a test set, and a validation set, thereby completing the training and tuning of the convolutional network model.

2. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1 is characterized in that: The frequency of sending and feeding back data during the reaction time in step (1) is 1 minute / time. In a complete reaction cycle, the reaction time starts from the first minute and accumulates until the end of the entire furnace.

3. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: The method for handling missing values ​​in step (2) is the upward filling method, and the specific process is as follows: m t =x t-p The formula represents the position of the nearest non-missing sample before time t. p′∈[0,t-1] Where m t is the missing sample at time t, x t-p is the sample at time tp, p is a moment before t, and p' is the entire time range from the beginning to time t, from 0 minutes to t-1 minutes, that is, the missing values ​​are filled with the non-null values ​​of the most recent moment before them.

4. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: The processing of outliers in step (2) is based on the 3σ principle in a limited time domain, and the specific process is as follows: Assume x t represents the sample at time t, m is the length of the statistical time domain, u t and σ t is the calculated mean and standard deviation statistics for x t Do the following: If x t ≥u t +3σ t , then x t =u t +3σ t ; If x t ≤u t +3σ t , then x t =u t -3σ t ; in Where m is the total sample size, and i is the time index when traversing each sample.

5. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: In step (2), the original feature variable x is normalized, and the specific process is as follows: Where x is the original feature variable, x max and x min are the maximum and minimum values ​​of the original feature variable x, respectively. is the normalized result.

6. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: In step (3), global information is extracted based on the pyramid pooling technology, and the specific calculation is as follows: s=t / / k WITH t,{k1,k2…kn} =[of t,k1 ,With t,k2 ,…With t,kn ] T where x i is the sample at the i-th moment, represents the mean of samples from the 1st to the tth moment; T represents transposition, and vectors are generally displayed in column form; z t,k It represents the extraction result obtained when k features are extracted from the historical sequence at time t; t / / k refers to the value obtained by dividing k by t, which means rounding operation; Z t,{k1,k2…kn} It means that n extraction strategies are used for feature extraction at the tth time, where {k1, k2…kn} represents the parameters of n different extraction strategies. After obtaining the corresponding pyramid features, the extracted results will be spliced ​​again by column to obtain new feature data, which will be used as the input of the subsequent convolutional network model.

7. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: The convolutional network model building process in step (4) is as follows: 4.1) Use pyramid pooling technology to extract global features and build a global feature extraction layer; 4.2) Based on the two-dimensional convolution technology, the local features are extracted and the local feature extraction layer is constructed. The two-dimensional convolution technique is to extract specific feature information by sliding a small matrix called a convolution kernel or filter on the input. In the two-dimensional convolution layer, the convolution kernel is a two-dimensional matrix, which is multiplied and summed element by element with the input features to obtain the output feature map. 4.3) Add embedding of temporal features and construct a temporal embedding layer; 4.4) splicing the data results of the global feature extraction layer, local feature extraction layer and time embedding layer constructed in steps 4.1)-4.3) by column; 4.5) The data concatenation result obtained in step 4.4) is introduced into the fully connected layer, and the final prediction result is output through the output layer.

8. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 7, characterized in that: The convolutional network model in step (4) includes a global feature extraction layer, a local feature extraction layer, a time embedding layer and an output layer. in: The input of the global feature extraction layer is the information extracted in step (3), which consists of two fully connected layers, and its purpose is to learn the information of sequence history; The input part of the local feature extraction layer is a fixed-length sequence of W lengths pushed forward from the input sample time, where W is an integer multiple of hours. The convolutional network structure is built using convolutional layers, and feature extraction is performed along the time dimension based on the two-dimensional convolutional layer. The purpose is to learn the tail temperature trend information of the sequence in the near future; The time embedding layer embeds the reaction time to obtain the hidden features of time, with the aim of obtaining the implicit information of the tail temperature trend and relative size at the current time node; The output layer concatenates the output results of the global feature extraction layer, the local feature extraction layer, and the embedding layer for relearning to obtain the final prediction result.

9. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: The step (5) constructs the new feature data obtained in step (3) into a sample set in minutes, where the sample set includes global features, local features, time features and data labels; The global feature is the new feature data constructed in step (3); the time feature is the moment t of the current reaction; The local features and data labels are constructed as follows: Assume that the material delivery table, furnace water temperature and flow rate, voltage and tail temperature data samples at time t are x t , the actual tail temperature data is y t , then: X t =[x t-s ,x t-s+1 ,…x t ] T s≥1 AND t =and t+1 Where X t and Y t are the local features and data labels at time t, and s is the window size for local feature extraction.

10. The method for predicting the tail temperature of a polysilicon reduction furnace based on a convolutional network according to claim 1, characterized in that: The training and evaluation of the convolutional network model uses MSE as its evaluation indicator, and the specific form is: Where f(x j ) and y j They represent the predicted value and actual value of the jth sample respectively, and N is the total number of samples.

Images