WO2024045244A1 - 用于无水氟化氢生产的能源管理控制系统及其控制方法 - Google Patents

用于无水氟化氢生产的能源管理控制系统及其控制方法 Download PDF

Info

Publication number
WO2024045244A1
WO2024045244A1 PCT/CN2022/120885 CN2022120885W WO2024045244A1 WO 2024045244 A1 WO2024045244 A1 WO 2024045244A1 CN 2022120885 W CN2022120885 W CN 2022120885W WO 2024045244 A1 WO2024045244 A1 WO 2024045244A1
Authority
WO
WIPO (PCT)
Prior art keywords
vector
feature
product
matrix
temperature
Prior art date
Application number
PCT/CN2022/120885
Other languages
English (en)
French (fr)
Inventor
邱汉林
钟华麟
陈三凤
廖鸿辉
廖育能
罗丽华
蓝丽萍
Original Assignee
福建省龙氟新材料有限公司
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by 福建省龙氟新材料有限公司 filed Critical 福建省龙氟新材料有限公司
Publication of WO2024045244A1 publication Critical patent/WO2024045244A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B19/00—Programme-control systems
    • G05B19/02—Programme-control systems electric
    • G05B19/418—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM]
    • G05B19/41885—Total factory control, i.e. centrally controlling a plurality of machines, e.g. direct or distributed numerical control [DNC], flexible manufacturing systems [FMS], integrated manufacturing systems [IMS] or computer integrated manufacturing [CIM] characterised by modeling, simulation of the manufacturing system
    • G—PHYSICS
    • G05—CONTROLLING; REGULATING
    • G05B—CONTROL OR REGULATING SYSTEMS IN GENERAL; FUNCTIONAL ELEMENTS OF SUCH SYSTEMS; MONITORING OR TESTING ARRANGEMENTS FOR SUCH SYSTEMS OR ELEMENTS
    • G05B2219/00—Program-control systems
    • G05B2219/30—Nc systems
    • G05B2219/32—Operator till task planning
    • G05B2219/32339—Object oriented modeling, design, analysis, implementation, simulation language

Definitions

  • the present invention relates to the field of intelligent management of produced energy, and more specifically, to an energy management control system for anhydrous hydrogen fluoride production and a control method thereof.
  • Anhydrous hydrogen fluoride is widely used in atomic energy, chemical industry, petroleum and other industries. It is a strong oxidant and can be used to prepare elemental fluorine, various fluorine refrigerants, inorganic fluorides, organic fluorides, and can also be formulated into aqueous hydrogen for various purposes. Hydrofluoric acid.
  • the traditional preparation method of anhydrous hydrogen fluoride mainly involves mixing fluorspar powder and sulfuric acid, and directly producing hydrogen fluoride under heating conditions.
  • the traditional preparation method of anhydrous hydrogen fluoride directly heats the materials.
  • the materials are mixed unevenly, which not only increases unnecessary reaction energy consumption, but also the main reaction is insufficient and the product yield and purity are low.
  • the embodiment of the present application provides an energy management control system and a control method for anhydrous hydrogen fluoride production, which uses artificial intelligence control technology to control the rotation speed value of the converter reactor and the pressure in the furnace based on a deep neural network model. value, reaction temperature value and gas chromatogram of the product to extract implicit correlation features in the time series dimension, and dynamically adjust the converter reactor in real time based on the real-time generation state characteristics of the product and the change characteristics of the pressure in the furnace and the reaction temperature.
  • the rotation speed is adjusted to improve the efficiency of the reaction, thereby improving the utilization rate of reaction materials and the quality of the product.
  • an energy management control system for anhydrous hydrogen fluoride production which includes:
  • the data acquisition module is used to obtain the rotational speed value, furnace pressure value, reaction temperature value and product gas chromatogram of the converter reactor at multiple predetermined time points including the current time point within a predetermined time period; the structured module, Used to calculate the temperature input after arranging the furnace pressure values and reaction temperature values at multiple predetermined time points within the predetermined time period including the current time point into temperature input vectors and pressure input vectors according to the time dimension.
  • the rotational speed encoding module is used to pass the rotational speed values of the converter reactor at multiple predetermined time points within the predetermined time period including the current time point through a time series encoder containing a one-dimensional convolution layer to obtain the rotational speed feature vector; global control features A generation module, used to multiply the rotational speed characteristic vector and the temperature-pressure correlation characteristic matrix to obtain a parameter global control characteristic vector; a product data encoding module, used to encode the current time point within the predetermined time period.
  • the gas chromatogram of the product at multiple predetermined time points is passed through a second convolutional neural network using a three-dimensional convolution kernel to obtain a product feature vector; a correction module is used to perform eigenvalues at each position in the product feature vector. Modify to obtain the corrected product feature vector; an action representation module for calculating the transfer matrix of the parameter global control feature vector relative to the product feature vector as a classification feature matrix; and
  • An energy control result generation module is used to pass the classification feature matrix through a classifier to obtain a classification result.
  • the classification result is used to indicate that the rotation speed value of the converter reactor at the current point in time should be increased or decreased.
  • the structured module includes: a data-level correlation unit used to calculate all the parameters between the temperature input vector and the pressure input vector using the following formula: The temperature-pressure correlation matrix; wherein, the formula is:
  • V 1 represents the temperature input vector
  • V 2 represents the pressure input vector
  • M represents the temperature-pressure correlation matrix
  • the data-level correlation module includes: a shallow feature map extraction subunit for extracting shallow features from the Mth layer of the first convolutional neural network Feature matrix, M is an even number; the deep feature map extraction subunit is used to extract the deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number, and N is greater than 2 times of M; and, A feature map fusion subunit is used to fuse the shallow feature map and the deep feature map to generate the temperature-pressure correlation feature matrix.
  • the rotation speed encoding module includes: an input vector construction unit for converting the converter reactions at multiple predetermined time points within the predetermined time period including the current time point.
  • the rotational speed value of the encoder is arranged into a one-dimensional input vector according to the time dimension;
  • a fully connected encoding unit is used to use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector using the following formula to extract the input
  • the high-dimensional implicit features of the eigenvalues at each position in the vector where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, Represents matrix multiplication;
  • a one-dimensional convolution coding unit used to use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution coding on the input vector according to the following formula to extract the position of each position in the input vector
  • High-dimensional implicit correlation features between eigenvalues where the formula is
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel
  • X represents the input vector.
  • the product data encoding module is further used to: the second convolutional neural network using a three-dimensional convolution kernel encodes the input data in the forward transmission of the layer. Perform respectively: perform three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; perform mean pooling processing based on a local feature matrix on the convolution feature map to obtain a pooled feature map; And, perform nonlinear activation on the pooled feature map to obtain an activation feature map; wherein the output of the last layer of the second convolutional neural network is the product feature vector, and the second convolutional neural network
  • the input of the first layer is the gas chromatogram of the product at multiple predetermined time points including the current time point within the predetermined time period.
  • the correction module is further used to: correct the characteristic values of each position in the product characteristic vector with the following formula to obtain the corrected product characteristic vector ;wherein, the formula is:
  • V represents the product eigenvector
  • ⁇ represents the autocovariance matrix of the product eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of each two positions of the product eigenvector
  • ⁇ and ⁇ are the global mean and variance of the product feature vector respectively, and Represents the position-wise subtraction and addition of the eigenvector respectively
  • 2 represents the second norm of the eigenvector
  • exp( ⁇ ) represents the exponential operation of the vector, and the exponential operation using the vector as the power represents the position of the vector.
  • Eigenvalues are natural exponential function values raised as powers.
  • the function representation module is further used to: calculate the transfer matrix of the parameter global control feature vector relative to the product feature vector as the classification using the following formula Characteristic matrix; where the formula is:
  • V 1 M*V 2
  • V 1 represents the parameter global control feature vector
  • M represents the classification feature matrix
  • V 2 represents the product feature vector
  • the energy control result generation module processes the classification feature matrix according to the following formula to generate a classification result, where: The above formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • a control method of an energy management control system for anhydrous hydrogen fluoride production includes:
  • the classification feature matrix is passed through a classifier to obtain a classification result, which is used to indicate that the rotational speed value of the converter reactor at the current time point should be increased or decreased.
  • the furnace pressure values and reaction temperature values at multiple predetermined time points including the current time point within the predetermined time period are calculated according to the time dimension.
  • calculating the temperature-pressure correlation matrix between the temperature input vector and the pressure input vector includes: calculating the temperature input vector and the pressure input vector using the following formula The temperature-pressure correlation matrix between;
  • V 1 represents the temperature input vector
  • V 2 represents the pressure input vector
  • M represents the temperature-pressure correlation matrix
  • the temperature-pressure correlation matrix is passed through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain the temperature-pressure correlation matrix.
  • the pressure-related feature matrix includes: extracting a shallow feature matrix from the Mth layer of the first convolutional neural network, M is an even number; extracting a deep feature matrix from the Nth layer of the first convolutional neural network, where, N is an even number, and N is greater than 2 times of M; and, the shallow feature map and the deep feature map are fused to generate the temperature-pressure correlation feature matrix.
  • the rotation speed values of the converter reactor at multiple predetermined time points within the predetermined time period including the current time point are passed through a one-dimensional convolution layer.
  • the time series encoder is used to obtain the rotation speed feature vector, including: arranging the rotation speed values of the converter reactor at multiple predetermined time points within the predetermined time period including the current time point into a one-dimensional input vector according to the time dimension; using the time series
  • the fully connected layer of the encoder performs fully connected encoding on the input vector using the following formula to extract high-dimensional hidden features of the feature values of each position in the input vector, where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, represents matrix multiplication; use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution encoding on the input vector with the following formula to extract high-dimensional implicit correlation features between the eigenvalues of each position in the input vector,
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel
  • X represents the input vector.
  • the gas chromatograms of the products at multiple predetermined time points within the predetermined time period including the current time point are passed through using a three-dimensional convolution kernel.
  • the second convolutional neural network to obtain the product feature vector includes: the second convolutional neural network using a three-dimensional convolution kernel performs on the input data respectively in the forward pass of the layer: based on the three-dimensional convolution kernel, The input data is subjected to a three-dimensional convolution process to obtain a convolution feature map; the convolution feature map is subjected to a mean pooling process based on a local feature matrix to obtain a pooled feature map; and, the pooled feature map is subjected to nonlinear Activated to obtain an activation feature map; wherein, the output of the last layer of the second convolutional neural network is the product feature vector, and the input of the first layer of the second convolutional neural network is the predetermined time period Contains gas chromatograms of
  • the characteristic values of each position in the product characteristic vector are corrected to obtain the corrected product characteristic vector, including: using the following formula to modify the product characteristic
  • the eigenvalues of each position in the vector are corrected to obtain the corrected product eigenvector; where the formula is:
  • V represents the product eigenvector
  • ⁇ represents the autocovariance matrix of the product eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of each two positions of the product eigenvector
  • ⁇ and ⁇ are the global mean and variance of the product feature vector respectively, and Represents the position-wise subtraction and addition of the eigenvector respectively
  • 2 represents the second norm of the eigenvector
  • exp( ⁇ ) represents the exponential operation of the vector, and the exponential operation using the vector as the power represents the position of the vector.
  • Eigenvalues are natural exponential function values raised as powers.
  • calculating the transfer matrix of the parameter global control feature vector relative to the product feature vector as a classification feature matrix includes: calculating the parameters with the following formula The transfer matrix of the global control feature vector relative to the product feature vector is used as the classification feature matrix; wherein, the formula is:
  • V 1 M*V 2
  • V 1 represents the parameter global control feature vector
  • M represents the classification feature matrix
  • V 2 represents the product feature vector
  • passing the classification feature matrix through a classifier to obtain a classification result includes: using the classifier to process the classification feature matrix with the following formula To generate classification results, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the projection is a vector
  • W 1 to W n are the weight matrices of the fully connected layers of each layer
  • B 1 to B n represent the bias matrices of the fully connected layers of each layer.
  • the energy management control system and its control method provided by this application for anhydrous hydrogen fluoride production use artificial intelligence control technology to control the rotation speed value of the converter reactor and the furnace speed based on a deep neural network model.
  • the internal pressure value, the reaction temperature value and the gas chromatogram of the product are extracted with implicit correlation features in the time series dimension, and the converter is dynamically adjusted in real time based on the real-time generation state characteristics of the product and the change characteristics of the pressure and reaction temperature in the furnace.
  • the speed of the reactor is adjusted to improve the efficiency of the reaction, thereby improving the utilization rate of reaction materials and the quality of the product.
  • Figure 1 is an application scenario diagram of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • FIG. 2 is a block diagram of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • Figure 3 is a flow chart of a control method of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • Figure 4 is a schematic architectural diagram of a control method of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • anhydrous hydrogen fluoride is widely used in atomic energy, chemical industry, petroleum and other industries. It is a strong oxidant and can be used to prepare elemental fluorine, various fluorine refrigerants, inorganic fluorides, organic fluorides, and can also be formulated into various Uses include aqueous hydrofluoric acid.
  • the traditional preparation method of anhydrous hydrogen fluoride mainly involves mixing fluorspar powder and sulfuric acid, and directly producing hydrogen fluoride under heating conditions.
  • the traditional preparation method of anhydrous hydrogen fluoride directly heats the materials.
  • the materials are mixed unevenly, which not only increases unnecessary reaction energy consumption, but also the main reaction is insufficient and the product yield and purity are low. Therefore, an optimized scheme for producing anhydrous hydrogen fluoride is expected.
  • the preparation process is as follows: S1: After heating the mixed sulfuric acid and fluorspar powder respectively, mix them in a pre-reactor according to a certain weight ratio of the input materials for pre-reaction to obtain a mixed material; S2: Continuously feed the mixed material The material is fed into the converter reactor, and under the conditions of the converter reactor speed of 1.0r/min ⁇ 2.0r/min and the furnace pressure of -0.55 ⁇ -0.45KPa, two temperatures of 550 ⁇ 650°C and 700 ⁇ 800°C are used in sequence. The mixed material reacts at high temperature for 50 to 70 minutes to obtain crude hydrogen fluoride gas and solid material. Part of the obtained solid material is fed to the converter reactor through the return device for further reaction, and the remaining solid material enters the dihydrate gypsum production device. ;
  • S3 The crude hydrogen fluoride gas is processed through the scrubber, condenser, distillation tower and degassing tower to obtain pure hydrogen fluoride gas and remaining gas;
  • S4 Use sulfuric acid to recycle the remaining gas to absorb the hydrogen fluoride gas;
  • S5 Not used The remaining gas recovered in the cycle is discharged from the top of the tail gas tower after being treated in the water washing tower and the tail gas tower.
  • step S2 the speed control, furnace pressure control and reaction temperature control of the converter reactor are not only related to the reaction efficiency and reaction adequacy, but also related to the performance of the anhydrous hydrogen fluoride production line. Consumption. Therefore, the inventor of the present application hopes to be able to dynamically control the rotation speed, pressure in the furnace, and reaction temperature of the converter reactor in real time according to the characteristics of the production state of the product, so as to improve the efficiency of the reaction and thereby improve the utilization rate of the reaction materials. .
  • the rotation speed of the converter reactor at multiple predetermined time points including the current time point within a predetermined time period is collected through various sensors, such as a rotation speed sensor, a pressure sensor and a temperature sensor. value, furnace pressure value and reaction temperature value, and collect gas chromatograms of the product at the multiple predetermined time points through a gas chromatograph.
  • the temperature-pressure correlation matrix is processed through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain a temperature-pressure correlation feature matrix. It should be understood that processing using a convolutional neural network model in which adjacent convolutional layers are convolution kernels that are transposed to each other can simultaneously update network parameters and search for network parameter structures suitable for specific data structures during training. thereby improving the accuracy of classification.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation sum of the rotational speed value of the converter reactor in the temporal dimension through one-dimensional convolutional encoding.
  • the high-dimensional hidden features of the rotation speed value of the converter reactor are extracted through fully connected coding.
  • the global mean pooling along the channel dimension of the product feature map is a pixel-level forward global downsampling based on image semantics, which will lead to differences between the feature values of each position of the product feature vector.
  • the correlation is poor, thus affecting its ability to express the distribution of the product feature map along the channel dimension.
  • V represents the product eigenvector
  • ⁇ represents the autocovariance matrix of the product eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of each two positions of the product eigenvector
  • ⁇ and ⁇ are the global mean and variance of the product feature vector respectively, and Represents the position-wise subtraction and addition of the eigenvector respectively
  • 2 represents the second norm of the eigenvector
  • exp( ⁇ ) represents the exponential operation of the vector, and the exponential operation using the vector as the power represents the position of the vector.
  • Eigenvalues are natural exponential function values raised as powers.
  • this guided correction of forward propagation correlation guides feature engineering through a learnable normal sampling offset of the product eigenvector to effectively model the long-range dependencies between individual eigenvalues in its length direction to further
  • the correlation between feature values is repaired by considering the local and non-local neighborhoods of the feature values.
  • we can improve its ability to predict class probabilities as it continues to flow with depth in the model, thereby improving the accuracy of classification. accuracy.
  • the operation is a responsive feature to the associated change of the parameter. Therefore, in order to better integrate the parameter global control feature vector and the product feature vector, further calculate the parameter global control feature vector relative to the product feature vector.
  • the transfer matrix serves as a classification feature matrix. In this way, the classification feature matrix can be passed through a classifier to obtain a classification result indicating that the rotation speed value of the converter reactor at the current time point should be increased or decreased.
  • this application proposes an energy management control system for anhydrous hydrogen fluoride production, which includes: a data acquisition module used to obtain converter reactions at multiple predetermined time points within a predetermined time period, including the current time point. The speed value of the furnace, the pressure value in the furnace, the reaction temperature value and the gas chromatogram of the product; a structured module for converting the pressure values in the furnace at multiple predetermined time points within the predetermined time period, including the current time point.
  • a data-level correlation module is used to combine the The temperature-pressure correlation matrix passes through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain the temperature-pressure correlation feature matrix;
  • the rotation speed encoding module is used to convert the predetermined time period including the current time point into The rotational speed values of the converter reactor at multiple predetermined time points are passed through a time series encoder including a one-dimensional convolution layer to obtain a rotational speed feature vector;
  • a global control feature generation module is used to combine the rotational speed feature vector with the temperature- The pressure correlation feature matrix is multiplied to obtain the parameter global control feature vector;
  • the product data encoding module is used to convert the gas chromatogram of the product at multiple predetermined time points within the predetermined time period including the current time point by using a three-dimensional
  • the second convolutional neural network of the convolution kernel is used to obtain
  • FIG 1 illustrates an application scenario diagram of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • the rotation speed sensor T1 the pressure sensor shown in Figure 1 deployed in the converter reactor (for example, R shown in Figure 1)
  • Sensor T2 and temperature sensor T3 respectively obtain the rotation speed value, furnace pressure value and reaction temperature value of the converter reactor at multiple predetermined time points including the current time point within a predetermined time period, and use the gas chromatograph (for example, As illustrated in G) in Figure 1, gas chromatograms of the product (eg, P as illustrated in Figure 1) at multiple predetermined time points within the predetermined time period are collected.
  • the gas chromatograph for example, As illustrated in G
  • gas chromatograms of the product eg, P as illustrated in Figure 1
  • the obtained rotational speed values, furnace pressure values, reaction temperature values and product gas chromatograms of the converter reactor at the plurality of predetermined time points are input to a server deployed with an energy management control algorithm for anhydrous hydrogen fluoride production.
  • a server deployed with an energy management control algorithm for anhydrous hydrogen fluoride production for example, the cloud server S as shown in Figure 1
  • the server can use an energy management control algorithm for anhydrous hydrogen fluoride production to control the rotational speed values of the converter reactor at the plurality of predetermined time points
  • the pressure value in the furnace, the reaction temperature value and the gas chromatogram of the product are processed to generate a classification result indicating whether the rotational speed value of the converter reactor at the current point in time should be increased or decreased.
  • FIG. 2 illustrates a block diagram of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • the energy management control system 200 for anhydrous hydrogen fluoride production according to the embodiment of the present application includes: a data acquisition module 210, used to obtain multiple predetermined times including the current time point within a predetermined time period. The rotation speed value of the converter reactor at the point, the pressure value in the furnace, the reaction temperature value and the gas chromatogram of the product; the structured module 220 is used to combine multiple predetermined time points within the predetermined time period including the current time point.
  • the furnace pressure values and reaction temperature values are arranged into temperature input vectors and pressure input vectors according to the time dimension respectively, calculate the temperature-pressure correlation matrix between the temperature input vector and the pressure input vector;
  • the data level correlation module 230 used to pass the temperature-pressure correlation matrix through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain the temperature-pressure correlation feature matrix;
  • the rotation speed encoding module 240 is used to convert the predetermined time
  • the rotation speed values of the converter reactor at multiple predetermined time points in the segment including the current time point are passed through a time series encoder including a one-dimensional convolution layer to obtain the rotation speed feature vector;
  • the global control feature generation module 250 is used to generate the The rotational speed feature vector is multiplied by the temperature-pressure correlation feature matrix to obtain a parameter global control feature vector;
  • the product data encoding module 260 is used to encode multiple predetermined time points including the current time point within the predetermined time period.
  • the gas chromatogram of the product is passed through a second convolutional neural network using a three-dimensional convolution kernel to obtain the product feature vector; the correction module 270 is used to correct the feature values of each position in the product feature vector to obtain the corrected product Feature vector; the action representation module 280 is used to calculate the transfer matrix of the parameter global control feature vector relative to the product feature vector as a classification feature matrix; and the energy control result generation module 290 is used to convert the classification feature matrix
  • the classification result is obtained through the classifier, and the classification result is used to indicate that the rotation speed value of the converter reactor at the current point in time should be increased or decreased.
  • the data acquisition module 210 and the structuring module 220 are used to obtain the rotation speed values of the converter reactor at multiple predetermined time points within a predetermined time period, including the current time point. , the furnace pressure value, the reaction temperature value and the gas chromatogram of the product, and arrange the furnace pressure values and reaction temperature values at multiple predetermined time points within the predetermined time period including the current time point according to the time dimension. After inputting a temperature vector and a pressure input vector, calculate a temperature-pressure correlation matrix between the temperature input vector and the pressure input vector.
  • the speed control, furnace pressure control and reaction temperature control of the converter reactor are not only related to reaction efficiency and reaction adequacy, but also related to the energy consumption of the anhydrous hydrogen fluoride production line. Therefore, in the technical solution of the present application, it is expected that the rotation speed, pressure in the furnace and reaction temperature of the converter reactor can be controlled dynamically in real time according to the characteristics of the production state of the product, so as to improve the efficiency of the reaction and thereby improve the reaction efficiency. Material utilization rate.
  • the converter reactions at multiple predetermined time points including the current time point within a predetermined time period are collected through various sensors, such as a rotation speed sensor, a pressure sensor and a temperature sensor.
  • the rotation speed value of the furnace, the pressure value in the furnace and the reaction temperature value are collected, and the gas chromatograms of the products at the multiple predetermined time points are collected using a gas chromatograph.
  • the temperature is calculated.
  • a temperature-pressure correlation matrix between the input vector and the pressure input vector is used to integrate the correlation between the pressure value in the furnace and the reaction temperature value to facilitate subsequent feature extraction.
  • the structured module includes: a data-level correlation unit, used to calculate the temperature-pressure between the temperature input vector and the pressure input vector using the following formula Correlation matrix; where the formula is:
  • V 1 represents the temperature input vector
  • V 2 represents the pressure input vector
  • M represents the temperature-pressure correlation matrix
  • the data-level correlation module 230 is used to pass the temperature-pressure correlation matrix through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain Temperature-pressure correlation characteristic matrix. That is, in the technical solution of the present application, the temperature-pressure correlation matrix is further processed through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain the temperature-pressure correlation features. matrix. It should be understood that processing using a convolutional neural network model in which adjacent convolutional layers are convolution kernels that are transposed to each other can simultaneously update network parameters and search for network parameter structures suitable for specific data structures during training. thereby improving the accuracy of classification.
  • the data-level correlation module includes: a shallow feature map extraction subunit, used to extract a shallow feature matrix, M, from the Mth layer of the first convolutional neural network. is an even number; the deep feature map extraction subunit is used to extract the deep feature matrix from the Nth layer of the first convolutional neural network, where N is an even number, and N is greater than 2 times of M; and, the feature map fusion subunit A unit configured to fuse the shallow feature map and the deep feature map to generate the temperature-pressure correlation feature matrix.
  • the rotation speed encoding module 240 is used to encode the rotation speed values of the converter reactor at multiple predetermined time points within a predetermined time period including the current time point through a one-dimensional convolution layer. timing encoder to obtain the rotational speed feature vector.
  • a temporal encoder containing a one-dimensional convolution layer is used to encode it to extract the rotational speed value of the converter reactor in Dynamic change characteristics in the time dimension, thereby obtaining the rotational speed feature vector.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the rotational speed value of the converter reactor in the temporal dimension through one-dimensional convolutional encoding.
  • the correlation and high-dimensional hidden features of the rotation speed value of the converter reactor are extracted through fully connected encoding.
  • the rotation speed encoding module includes: an input vector construction unit for converting the rotation speed values of the converter reactor at multiple predetermined time points within a predetermined time period including the current time point.
  • a one-dimensional input vector arranged according to the time dimension;
  • a fully connected encoding unit used to use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector using the following formula to extract each position in the input vector High-dimensional latent features of eigenvalues, where the formula is: where X is the input vector, Y is the output vector, W is the weight matrix, and B is the bias vector, Represents matrix multiplication;
  • a one-dimensional convolution coding unit used to use the one-dimensional convolution layer of the temporal encoder to perform one-dimensional convolution coding on the input vector according to the following formula to extract the position of each position in the input vector High-dimensional implicit correlation features between eigenvalues, where the formula is:
  • a is the width of the convolution kernel in the x direction
  • F is the convolution kernel parameter vector
  • G is the local vector matrix that operates with the convolution kernel function
  • w is the size of the convolution kernel
  • X represents the input vector.
  • the global control feature generation module 250 is configured to multiply the rotational speed feature vector and the temperature-pressure correlation feature matrix to obtain a parameter global control feature vector. That is to say, in the technical solution of the present application, after obtaining the temperature-pressure correlation characteristic matrix with correlation characteristics between the pressure value in the furnace and the reaction temperature value and the dynamics of the rotation speed value of the converter reactor After changing the rotational speed characteristic vector of the characteristic, the rotational speed characteristic vector and the temperature-pressure correlation characteristic matrix are further multiplied to integrate the characteristic correlation information of the changing parameters, thereby obtaining a parameter global control characteristic vector.
  • the product data encoding module 260 is used to convert the gas chromatograms of the products at multiple predetermined time points within the predetermined time period including the current time point by using three-dimensional convolution.
  • the second convolutional neural network of the kernel is used to obtain the product feature vector.
  • the gas chromatogram of the product at the predetermined time point is processed in a second convolutional neural network using a three-dimensional convolution kernel to extract the product at multiple predetermined time points within the predetermined time period, including the current time point.
  • the local dynamics of the gas chromatogram imply the feature distribution, thereby obtaining the product feature vector.
  • the product data encoding module is further configured to: the second convolutional neural network using a three-dimensional convolution kernel separately performs on the input data in the forward pass of the layer: based on The three-dimensional convolution kernel performs three-dimensional convolution processing on the input data to obtain a convolution feature map; performs mean pooling processing based on a local feature matrix on the convolution feature map to obtain a pooling feature map; and The pooled feature map performs nonlinear activation to obtain an activation feature map; wherein, the output of the last layer of the second convolutional neural network is the product feature vector, and the first layer of the second convolutional neural network
  • the input is the gas chromatogram of the product at multiple predetermined time points including the current time point within the predetermined time period.
  • the correction module 270 is used to correct the feature values of each position in the product feature vector to obtain a corrected product feature vector.
  • the global mean pooling along the channel dimension of the product feature map is a pixel-level forward global downsampling based on image semantics, which will result in the feature values of each position of the product feature vector. The correlation between them is poor, thus affecting its ability to express the distribution of the product feature map along the channel dimension. Therefore, in the technical solution of the present application, it is further necessary to conduct guided correction of forward propagation correlation on the product feature vector.
  • the correction module is further configured to: correct the characteristic values of each position in the product characteristic vector with the following formula to obtain the corrected product characteristic vector; wherein, The formula is:
  • V represents the product eigenvector
  • ⁇ represents the autocovariance matrix of the product eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of each two positions of the product eigenvector
  • ⁇ and ⁇ are the global mean and variance of the product feature vector respectively, and Represents the position-wise subtraction and addition of the eigenvector respectively
  • 2 represents the second norm of the eigenvector
  • exp( ⁇ ) represents the exponential operation of the vector, and the exponential operation using the vector as the power represents the position of the vector.
  • Eigenvalues are natural exponential function values raised as powers.
  • the guided correction of the forward propagation correlation guides feature engineering through the learnable normal sampling offset of the product feature vector to effectively model the relationship between each feature value in its length direction. Long-range dependencies are used to further consider the local and non-local neighborhoods of feature values to repair the correlation between feature values. In this way, when we continue to calculate the transfer matrix with the parameter global control feature vector as the classification feature matrix and perform classification, we can improve its ability to predict class probabilities as it continues to flow with depth in the model, thereby improving the accuracy of classification. accuracy.
  • the action representation module 280 and the energy control result generation module 290 are used to calculate the transfer matrix of the parameter global control feature vector relative to the product feature vector as a classification feature matrix. , and pass the classification feature matrix through a classifier to obtain a classification result, which is used to indicate that the rotational speed value of the converter reactor at the current time point should be increased or decreased. It should be understood that due to the different feature scales between the dynamic change characteristics of the gas chromatogram of the product and the implicit features associated with the parameters, and the dynamic characteristics of the gas chromatogram of the product can be determined in a high-dimensional feature space is regarded as a responsive feature to the associated change of the parameter.
  • the parameter global control feature vector in order to better integrate the parameter global control feature vector and the product feature vector, further calculate the parameter global control feature vector relative to the product feature vector.
  • the transfer matrix is used as the classification feature matrix.
  • the classification feature matrix can be passed through a classifier to obtain a classification result indicating that the rotation speed value of the converter reactor at the current time point should be increased or decreased.
  • the classifier processes the classification feature matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the action representation module is further configured to: calculate the transfer matrix of the parameter global control feature vector relative to the product feature vector as the classification feature matrix using the following formula; where , the formula is:
  • V 1 M*V 2
  • V 1 represents the parameter global control feature vector
  • M represents the classification feature matrix
  • V 2 represents the product feature vector
  • the energy management control system 200 for anhydrous hydrogen fluoride production based on the embodiment of the present application has been clarified. It uses artificial intelligence control technology to control the rotation speed value of the converter reactor and the energy in the furnace based on the deep neural network model.
  • the pressure value, reaction temperature value and gas chromatogram of the product are extracted with implicit correlation features in the time series dimension, and the converter reaction is dynamically performed in real time based on the real-time generation state characteristics of the product and the change characteristics of the pressure and reaction temperature in the furnace.
  • the rotation speed of the reactor is adjusted to improve the efficiency of the reaction, thereby improving the utilization rate of reaction materials and the quality of the product.
  • the energy management control system 200 for anhydrous hydrogen fluoride production according to the embodiment of the present application can be implemented in various terminal devices, such as a server of an energy management control algorithm for anhydrous hydrogen fluoride production, etc.
  • the energy management control system 200 for anhydrous hydrogen fluoride production according to an embodiment of the present application can be integrated into a terminal device as a software module and/or a hardware module.
  • the energy management control system 200 for anhydrous hydrogen fluoride production can be a software module in the operating system of the terminal equipment, or can be an application program developed for the terminal equipment; of course, the energy management control system 200 for anhydrous hydrogen fluoride production
  • the energy management control system 200 for water hydrogen fluoride production can also be one of the many hardware modules of the terminal equipment.
  • the energy management control system 200 for anhydrous hydrogen fluoride production and the terminal device may also be separate devices, and the energy management control system 200 for anhydrous hydrogen fluoride production may be connected via a wired And/or a wireless network is connected to the terminal device, and the interactive information is transmitted according to the agreed data format.
  • Figure 3 illustrates a flow chart of a control method of an energy management control system for anhydrous hydrogen fluoride production.
  • the control method of the energy management control system for anhydrous hydrogen fluoride production includes step: S110, obtaining multiple predetermined time points including the current time point within the predetermined time period.
  • the rotation speed value is passed through a temporal encoder including a one-dimensional convolution layer to obtain a rotation speed feature vector;
  • S150 multiply the rotation speed feature vector and the temperature-pressure correlation feature matrix to obtain a parameter global control feature vector;
  • S160 Pass the gas chromatograms of the product at multiple predetermined time points within the predetermined time period, including the current time point, through a second convolutional neural network using a three-dimensional convolution kernel to obtain a product feature vector;
  • S170 for the product The eigenvalues of each position in the eigenvector are corrected to obtain the corrected product eigenvector;
  • S180 calculate the transfer matrix of the parameter global control eigenvector relative to the product eigenvector as a classification feature matrix; and, S190, convert the The classification feature matrix is passed through the classifier to obtain a classification result, which is used to indicate that the rotation speed value of the converter reactor at the current time point should be increased or decreased.
  • FIG. 4 illustrates an architectural schematic diagram of a control method of an energy management control system for anhydrous hydrogen fluoride production according to an embodiment of the present application.
  • the network architecture of the control method of the energy management control system for anhydrous hydrogen fluoride production first, multiple predetermined times including the current time point within the obtained predetermined time period are obtained.
  • the furnace pressure value (for example, P1 as shown in Figure 4) and the reaction temperature value (for example, P2 as shown in Figure 4) of the point are respectively arranged according to the time dimension as a temperature input vector (for example, as shown in Figure 4
  • V1 as shown the pressure input vector
  • the pressure input vector for example, V2 as shown in Figure 4
  • calculate the temperature-pressure correlation matrix between the temperature input vector and the pressure input vector for example, as in Figure 4 M1 as shown
  • the temperature-pressure correlation matrix is passed through the first convolutional neural network (for example, CNN1 as shown in Figure 4) using mutually transposed convolution kernels in adjacent layers to obtain Temperature-pressure correlation feature matrix (for
  • steps S110 and S120 the rotational speed value, furnace pressure value, reaction temperature value and gas chromatogram of the converter reactor at multiple predetermined time points including the current time point within the predetermined time period are obtained. diagram, and after arranging the furnace pressure values and reaction temperature values at multiple predetermined time points within the predetermined time period including the current time point into temperature input vectors and pressure input vectors according to the time dimension, calculate the temperature Temperature-pressure correlation matrix between the input vector and the pressure input vector.
  • the speed control of the converter reactor, the pressure control in the furnace, and the reaction temperature control are not only related to the reaction efficiency and reaction adequacy, but also related to the energy consumption of the anhydrous hydrogen fluoride production line.
  • the converter reactions at multiple predetermined time points including the current time point within a predetermined time period are collected through various sensors, such as a rotation speed sensor, a pressure sensor and a temperature sensor.
  • the rotation speed value of the furnace, the pressure value in the furnace and the reaction temperature value are collected, and the gas chromatograms of the products at the multiple predetermined time points are collected using a gas chromatograph.
  • the temperature is calculated.
  • a temperature-pressure correlation matrix between the input vector and the pressure input vector is used to integrate the correlation between the pressure value in the furnace and the reaction temperature value to facilitate subsequent feature extraction.
  • step S130 the temperature-pressure correlation matrix is passed through a first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain a temperature-pressure correlation feature matrix. That is, in the technical solution of the present application, the temperature-pressure correlation matrix is further processed through the first convolutional neural network using mutually transposed convolution kernels in adjacent layers to obtain the temperature-pressure correlation features. matrix. It should be understood that processing using a convolutional neural network model in which adjacent convolutional layers are convolution kernels that are transposed to each other can simultaneously update network parameters and search for network parameter structures suitable for specific data structures during training. thereby improving the accuracy of classification.
  • step S140 the rotational speed values of the converter reactor at multiple predetermined time points including the current time point within the predetermined time period are passed through a temporal encoder including a one-dimensional convolution layer to obtain a rotational speed feature vector.
  • a temporal encoder containing a one-dimensional convolution layer is used to encode it to extract the rotational speed value of the converter reactor in Dynamic change characteristics in the time dimension, thereby obtaining the rotational speed feature vector.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the rotational speed value of the converter reactor in the temporal dimension through one-dimensional convolutional encoding.
  • the correlation and high-dimensional hidden features of the rotation speed value of the converter reactor are extracted through fully connected encoding.
  • step S150 the rotation speed feature vector and the temperature-pressure correlation feature matrix are multiplied to obtain a parameter global control feature vector. That is to say, in the technical solution of the present application, after obtaining the temperature-pressure correlation characteristic matrix with correlation characteristics between the pressure value in the furnace and the reaction temperature value and the dynamics of the rotation speed value of the converter reactor After changing the rotational speed characteristic vector of the characteristic, the rotational speed characteristic vector and the temperature-pressure correlation characteristic matrix are further multiplied to integrate the characteristic correlation information of the changing parameters, thereby obtaining a parameter global control characteristic vector.
  • step S160 the gas chromatograms of the product at multiple predetermined time points including the current time point within the predetermined time period are passed through a second convolutional neural network using a three-dimensional convolution kernel to obtain the product. Feature vector.
  • the gas chromatogram of the product at the predetermined time point is processed in a second convolutional neural network using a three-dimensional convolution kernel to extract the product at multiple predetermined time points within the predetermined time period, including the current time point.
  • the local dynamics of the gas chromatogram imply the feature distribution, thereby obtaining the product feature vector.
  • step S170 the characteristic values of each position in the product characteristic vector are corrected to obtain a corrected product characteristic vector.
  • the global mean pooling along the channel dimension of the product feature map is a pixel-level forward global downsampling based on image semantics, which will result in the feature values of each position of the product feature vector. The correlation between them is poor, thus affecting its ability to express the distribution of the product feature map along the channel dimension. Therefore, in the technical solution of the present application, it is further necessary to conduct guided correction of forward propagation correlation on the product feature vector.
  • the transfer matrix of the parameter global control feature vector relative to the product feature vector is calculated as a classification feature matrix, and the classification feature matrix is passed through a classifier to obtain a classification result,
  • the classification result is used to indicate that the rotation speed value of the converter reactor at the current point in time should be increased or decreased. It should be understood that due to the different feature scales between the dynamic change characteristics of the gas chromatogram of the product and the implicit features associated with the parameters, and the dynamic characteristics of the gas chromatogram of the product can be determined in a high-dimensional feature space is regarded as a responsive feature to the associated change of the parameter. Therefore, in order to better integrate the parameter global control feature vector and the product feature vector, further calculate the parameter global control feature vector relative to the product feature vector.
  • the transfer matrix is used as the classification feature matrix. In this way, the classification feature matrix can be passed through a classifier to obtain a classification result indicating that the rotation speed value of the converter reactor at the current time point should be increased or decreased.
  • control method of the energy management control system for anhydrous hydrogen fluoride production based on the embodiments of the present application has been clarified, which uses artificial intelligence control technology to control the rotation speed value and speed of the converter reactor based on a deep neural network model.
  • the pressure value in the furnace, the reaction temperature value and the gas chromatogram of the product are used to extract implicit correlation features in the time series dimension, and based on the real-time generation state characteristics of the product and the change characteristics of the pressure in the furnace and the reaction temperature, the described information is dynamically analyzed in real time.
  • the rotation speed of the converter reactor is adjusted to improve the efficiency of the reaction, thereby improving the utilization rate of reaction materials and the quality of the product.
  • each component or each step can be decomposed and/or recombined. These decompositions and/or recombinations shall be considered equivalent versions of this application.

Landscapes

  • Engineering & Computer Science (AREA)
  • Manufacturing & Machinery (AREA)
  • General Engineering & Computer Science (AREA)
  • Quality & Reliability (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Feedback Control In General (AREA)

Abstract

一种用于无水氟化氢生产的能源管理控制系统及其控制方法，所述的控制系统通过基于深度神经网络模型来对于转炉反应器（R）的转速值、炉内压力值、反应温度值和产物的气相色谱图进行时序维度上的隐含关联特征提取，并根据产物的实时生成状态特征以及炉内压力和反应温度的变化特征来实时动态地对于所述转炉反应器（R）的转速进行调整以提高反应的效率，进而提高反应物料的利用率与产品的质量。

Description

用于无水氟化氢生产的能源管理控制系统及其控制方法 技术领域
本发明涉及生产能源的智能管理的领域，且更为具体地，涉及一种用于无水氟化氢生产的能源管理控制系统及其控制方法。
背景技术
无水氟化氢广泛应用于原子能、化工、石油等行业，是强氧化剂，可用来制取元素氟、各种氟制冷剂、无机氟化物、有机氟化物，也可配制成各种用途的有水氢氟酸。传统的无水氟化氢制备方法主要是将氟石粉和硫酸混合，在加热的条件下，直接产生氟化氢。
传统的无水氟化氢制备方法直接将物料进行加热处理，物料混合不均匀，不仅增加了不必要的反应能耗，而且主反应进行不充分，产品得率和纯度较低。
因此，期待一种优化的用于生产无水氟化氢的方案。
发明内容
为了解决上述技术问题，提出了本申请。本申请的实施例提供了一种用于无水氟化氢生产的能源管理控制系统及其控制方法，其采用人工智能控制技术，通过基于深度神经网络模型来对于转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图进行时序维度上的隐含关联特征提取，并根据产物的实时生成状态特征以及炉内压力和反应温度的变化特征来实时动态地对于所述转炉反应器的转速进行调整以提高反应的效率，进而提高反应物料的利用率与产品的质量。
根据本申请的一个方面，提供了一种用于无水氟化氢生产的能源管理控制系统，其包括：
数据采集模块，用于获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；结构化模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；数据级关联模块，用于将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；
转速编码模块，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；全局控制特征生成模块，用于将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；产物数据编码模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；校正模块，用于对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；作用表示模块，用于计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及
能源控制结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
在上述用于无水氟化氢生产的能源管理控制系统中，所述结构化模块，包括：数据级关联单元，用于以如下公式来计算所述温度输入向量和所述压力输入向量之间的所述温度-压力关联矩阵；其中，所述公式为：
Figure PCTCN2022120885-appb-000001
其中
Figure PCTCN2022120885-appb-000002
表示向量相乘，V 1表示所述温度输入向量，
Figure PCTCN2022120885-appb-000003
表示所述温度输入向量的转置向量，V 2表示所述压力输入向量，M表示所述温度-压力关联矩阵。
在上述用于无水氟化氢生产的能源管理控制系统中，所述数据级关联模块，包括：浅层特征图提取子单元，用于从所述第一卷积神经网络的第M层提取浅层特征矩阵，M是偶数；深层特征图提取子单元，用于从所述第一卷积神经网络的第N层提取深层特征矩阵，其中，N为偶数，且N大于M的2倍；以及，特征图融合子单元，用于融合所述浅层特征图和所述深层特征图以生成所述温度-压力关联特征矩阵。
在上述用于无水氟化氢生产的能源管理控制系统中，所述转速编码模块，包括：输入向量构造单元，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值按照时间维度排列为一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000004
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022120885-appb-000005
表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000006
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸，X表示所述输入向量。
在上述用于无水氟化氢生产的能源管理控制系统中，所述产物数据编码模块，进一步用于：所述使用三维卷积核的第二卷积神经网络在层的正向传递中对输入数据分别进行：基于所述三维卷积核对所述输入数据进行三维卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部特征矩阵的均值池化处理以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第二卷积神经网络的最后一层的输出为所述产物特征向量，所述第二卷积神经网络的第一层的输入为所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图。
在上述用于无水氟化氢生产的能源管理控制系统中，所述校正模块，进一步用于：以如下公式对所述产物特征向量中各个位置的特征值进行修正以得到所述校正后产物特征向量；其中，所述公式为：
Figure PCTCN2022120885-appb-000007
其中V表示所述产物特征向量，∑表示所述产物特征向量的自协方差矩阵，即矩阵的每个位置的值是所述产物特征向量的每两个位置的特征值之间的方差，μ和σ分别是所述产物特征向量的全局均值和方差，
Figure PCTCN2022120885-appb-000008
和
Figure PCTCN2022120885-appb-000009
分别表示特征向量的按位置减法和加法，||·|| 2表示特征向量的二范数，exp(·)表示向量的指数运算，以向量为幂的指数运算表示以向量的每个位置的特征值作为幂的自然指数函数值。
在上述用于无水氟化氢生产的能源管理控制系统中，所述作用表示模块，进一步用于：以如下公式计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为所述分类特征矩阵；其中，所述公式为：
V 1＝M*V 2
其中V 1表示所述参数全局控制特征向量，M表示所述分类特征矩阵，V 2表示所述产物特征向量。
在上述用于无水氟化氢生产的能源管理控制系统中，所述能源控制结果生成模块，进一步用于：所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
根据本申请的另一方面，一种用于无水氟化氢生产的能源管理控制系统的控制方法，其包括：
获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及
将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵，包括：以如下公式来计算所述温度输入向量和所述压力输入向量之间的所述温度-压力关联矩阵；
其中，所述公式为：
Figure PCTCN2022120885-appb-000010
其中
Figure PCTCN2022120885-appb-000011
表示向量相乘，V 1表示所述温度输入向量，
Figure PCTCN2022120885-appb-000012
表示所述温度输入向量的转置向量，V 2表示所述压力输入向量，M表示所述温度-压力关联矩阵。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵，包括：从所述第一卷积神经网络的第M层提取浅层特征矩阵，M是偶数；从所述第一卷积神经网络的第N层提取深层特征矩阵，其中，N为偶数，且N大于M的2倍；以及，融合所述浅层特征图和所述深层特征图以生成所述温度-压力关联特征矩阵。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量，包括：将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值按照时间维度排列为一维的输入向量；使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000013
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022120885-appb-000014
表示矩阵乘；使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000015
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸，X表示所述输入向量。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量，包括：所述使用三维卷积核的第二卷积神经网络在层的正向传递中对输入数据分别进行：基于所述三维卷积核对所述输入数据进行三维卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部特征矩阵的均值池化处理以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第二卷积神经网络的最后一层的输出为所述产物特征向量，所述第二卷积神经网络的第一层的输入为所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量，包括：以如下公式对所述产物特征向量中各个位置的特征值进行修正以得到所述校正后产物特征向量；其中，所述公式为：
Figure PCTCN2022120885-appb-000016
Figure PCTCN2022120885-appb-000017
其中V表示所述产物特征向量，∑表示所述产物特征向量的自协方差矩阵，即矩阵的每个位置的值是所述产物特征向量的每两个位置的特征值之间的方差，μ和σ分别是所述产物特征向量的全局均值和方差，
Figure PCTCN2022120885-appb-000018
和
Figure PCTCN2022120885-appb-000019
分别表示特征向量的按位置减法和加法，||·|| 2表示特征向量的二范数，exp(·)表示向量的指数运算，以向量为幂的指数运算表示以向量的每个位置的特征值作为幂的自然指数函数值。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵，包括：以如下公式计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为所述分类特征矩阵；其中，所述公式为：
V 1＝M*V 2
其中V 1表示所述参数全局控制特征向量，M表示所述分类特征矩阵，V 2表示所述产物特征向量。
在上述用于无水氟化氢生产的能源管理控制系统的控制方法中，将所述分类特征矩阵通过分类器以得到分类结果，包括：使用所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
与现有技术相比，本申请提供的用于无水氟化氢生产的能源管理控制系统及其控制方法，其采用人工智能控制技术，通过基于深度神经网络模型来对于转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图进行时序维度上的隐含关联特征提取，并根据产物的实时生成状态特征以及炉内压力和反应温度的变化特征来实时动态地对于所述转炉反应器的转速进行调整以提高反应的效率，进而提高反应物料的利用率与产品的质量。
附图说明
通过结合附图对本申请实施例进行更详细的描述，本申请的上述以及其他目的、特征和优势将变得更加明显。附图用来提供对本申请实施例的进一步理解，并且构成说明书的一部分，与本申请实施例一起用于解释本申请，并不构成对本申请的限制。在附图中，相同的参考标号通常代表相同部件或步骤。
图1为根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的应用场景图。
图2为根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的框图。
图3为根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的控制方法的流程图。
图4为根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的控制方法的架构示意图。
具体实施方式
下面，将参考附图详细地描述根据本申请的示例实施例。显然，所描述的实施例仅仅是本申请的一部分实施例，而不是本申请的全部实施例，应理解，本申请不受这里描述的示例实施例的限制。
场景概述
如前所述，无水氟化氢广泛应用于原子能、化工、石油等行业，是强氧化剂，可用来制取元素氟、各种氟制冷剂、无机氟化物、有机氟化物，也可配制成各种用途的有水氢氟酸。传统的无水氟化氢制备方法主要是将氟石粉和硫酸混合，在加热的条件下，直接产生氟化氢。
传统的无水氟化氢制备方法直接将物料进行加热处理，物料混合不均匀，不仅增加了不必要的反应能耗，而且主反应进行不充分，产品得率和纯度较低。因此，期待一种优化的用于生产无水氟化氢的方案。
在本申请的制备方案中，制备流程如下：S1：将混合硫酸和氟石粉分别加热后，按照一定投料重量比在预反应器中混合进行预反应，得到混合物料；S2：将混合物料连续进料到转炉反应器中，在转炉反应器转速为1.0r/min～2.0r/min和炉内压力为-0.55～-0.45KPa的条件下，依次采用550～650℃和700～800℃两种温度对混合物料进行反应，反应50～70分钟，得到粗氟化氢气体和固体物料，将得到的固体物料的一部分通过返料装置进料到转炉反应器再次反应，剩余固体物料进入二水石膏生产装置；
S3：粗氟化氢气体经洗涤塔、冷凝器、精馏塔和脱气塔处理，得到纯净的氟化氢气体及剩余气体；S4：使用硫酸对剩余气体进行循环回收，以吸收氟化氢气体；S5：未被循环回收的剩余气体经水洗塔处理和尾气塔处理后，从尾气塔顶部排出。
基于此，本申请发明人发现在步骤S2中，对于转炉反应器的转速控制、炉内压力控制以及反应温度控制不仅关乎于反应效率和反应充分度，还关乎于无水氟化氢生产产线的能耗。因此，本申请发明人期望能够根据产物的生成状态特征来实时动态地对于所述转炉反应器的转速、炉内压力以及反应温度进行控制，以提高反应的效率，进而能够提高反应物料的利用率。
具体地，在本申请的技术方案中，首先，通过各个传感器，例如转速传感器、压力传感器和温度传感器分别采集预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值和反应温度值，并且通过气相色谱仪采集所述多个预定时间点的产物的气相色谱图。
应可以理解，由于所述炉内压力值和所述反应温度值之间具有着关联关系，其会相互产生影响，例如当炉内压力值升高时，反应温度也会随着增大，因此，在本申请的技术方案中，需要进一步挖掘出这种关联性特征来进行实际转炉反应器的转速控制。具体地，将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵，以整合所述炉内压力值和所述反应温度值的关联性，便于后续对其特征进行提取。
然后，将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络中进行处理以得到温度-压力关联特征矩阵。应可以理解，使用邻卷积层为互为转置的卷积核的卷积神经网络模型进行处理，能够在训练时能够同时更新网络参数的更新和适合特定数据结构的网络参数结构的搜索，进而提高分类的准确性。
对于所述预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值，考虑到其在时序上具有着动态的隐含规律，因此，为了更为充分地提取出这种动态变化的隐含特征规律，使用包含一维卷积层的时序编码器对其进行编码，以提取出所述转炉反应器的转速值在时间维度上的动态变化特征，从而得到转速特征向量。在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出所述转炉反应器的转速值在时序维度上的关联和通过全连接编码提取所述转炉反应器的转速值的高维隐含特征。
在得到具有所述炉内压力值和所述反应温度值之间关联特征的所述温度-压力关联特征矩阵以及所述转炉反应器的转速值的动态变化特征的所述转速特征向量后，进一步将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以整合变化参数的特征关联信息，从而得到参数全局控制特征向量。
对于所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图，考虑到所述产物的气相色谱图在时间维度上也是动态变化的，因此若想基于产物的变化状态特征信息来对于所述转炉反应器的转速值进行实时动态地控制，还需要将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图的局部动态隐含特征分布，从而得到产物特征向量。
但是，注意到对所述产物特征图进行沿通道维度的全局均值池化是基于图像语义的像素级的前向全局下采样，这会导致所述产物特征向量的各个位置的特征值之间的相关性较差，从而影响其对所 述产物特征图沿通道维度上的分布的表达能力。
因此，在本申请的技术方案中，进一步还需对产物特征向量进行前向传播相关性的引导修正，具体为：
Figure PCTCN2022120885-appb-000020
其中V表示所述产物特征向量，∑表示所述产物特征向量的自协方差矩阵，即矩阵的每个位置的值是所述产物特征向量的每两个位置的特征值之间的方差，μ和σ分别是所述产物特征向量的全局均值和方差，
Figure PCTCN2022120885-appb-000021
和
Figure PCTCN2022120885-appb-000022
分别表示特征向量的按位置减法和加法，||·|| 2表示特征向量的二范数，exp(·)表示向量的指数运算，以向量为幂的指数运算表示以向量的每个位置的特征值作为幂的自然指数函数值。
这里，该前向传播相关性的引导修正通过对产物特征向量的可学习的正态采样偏移引导特征工程来有效地建模其长度方向上的各个特征值之间的长程依赖关系，以进一步考虑特征值的局部和非局部邻域来进行特征值之间的相关性的修复。这样，在继续计算其与所述参数全局控制特征向量的转移矩阵作为分类特征矩阵并进行分类时，可以提高其在模型中继续随深度流式传播时对于类概率的预测能力，进而提高分类的准确性。
进一步地，由于所述产物的气相色谱图的动态变化特征与所述参数关联的隐含特征之间的特征尺度不同，并且所述产物的气相色谱图的动态特征在高维特征空间中可以看作是对于所述参数关联变化的响应性特征，因此为了更好地融合所述参数全局控制特征向量和所述产物特征向量，进一步计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵。这样，就可以将所述分类特征矩阵通过分类器以得到用于表示当前时间点的转炉反应器的转速值应增大或应减小的分类结果。
基于此，本申请提出了一种用于无水氟化氢生产的能源管理控制系统，其包括：数据采集模块，用于获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；结构化模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；数据级关联模块，用于将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；转速编码模块，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；全局控制特征生成模块，用于将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；产物数据编码模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；校正模块，用于对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；作用表示模块，用于计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及，能源控制结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
图1图示了根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的应用场景图。如图1所示，在该应用场景中，首先，通过部署于转炉反应器(例如，如图1中所示意的R)的各个传感器(例如，如图1中所示意的转速传感器T1、压力传感器T2和温度传感器T3)分别获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值和反应温度值，并且通过气相色谱仪(例如，如图1中所示意的G)采集所述预定时间段内多个预定时间点的产物(例如，如图1中所示意的P)的气相色谱图。然后，将获取的所述多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图输入至部署有用于无水氟化氢生产的能源管理控制算法的服务器中(例如，如图1中所示意的云服务器S)，其中，所述服务器能够以用于无水氟化氢生产的能源管理控制算法对所述多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图进行处理，以生成用于表示当前时间点的转炉反应器的转速值应增大或应减小的分类结果。
在介绍了本申请的基本原理之后，下面将参考附图来具体介绍本申请的各种非限制性实施例。
示例性系统
图2图示了根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的框图。如图2所示，根据本申请实施例的用于无水氟化氢生产的能源管理控制系统200，包括：数据采集模块210，用于获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；结构化模块220，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；数据级关联模块230，用于将 所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；转速编码模块240，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；全局控制特征生成模块250，用于将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；产物数据编码模块260，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；校正模块270，用于对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；作用表示模块280，用于计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及，能源控制结果生成模块290，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
具体地，在本申请实施例中，所述数据采集模块210和所述结构化模块220，用于获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图，并将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵。如前所述，由于对于转炉反应器的转速控制、炉内压力控制以及反应温度控制不仅关乎于反应效率和反应充分度，还关乎于无水氟化氢生产产线的能耗。因此，在本申请的技术方案中，期望能够根据产物的生成状态特征来实时动态地对于所述转炉反应器的转速、炉内压力以及反应温度进行控制，以提高反应的效率，进而能够提高反应物料的利用率。
也就是，具体地，在本申请的技术方案中，首先，通过各个传感器，例如转速传感器、压力传感器和温度传感器分别采集预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值和反应温度值，并且通过气相色谱仪采集所述多个预定时间点的产物的气相色谱图。应可以理解，由于所述炉内压力值和所述反应温度值之间具有着关联关系，其会相互产生影响，例如当炉内压力值升高时，反应温度也会随着增大，因此，在本申请的技术方案中，需要进一步挖掘出这种关联性特征来进行实际转炉反应器的转速控制。具体地，将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵，以整合所述炉内压力值和所述反应温度值的关联性，便于后续对其特征进行提取。
更具体地，在本申请实施例中，所述结构化模块，包括：数据级关联单元，用于以如下公式来计算所述温度输入向量和所述压力输入向量之间的所述温度-压力关联矩阵；其中，所述公式为：
Figure PCTCN2022120885-appb-000023
Figure PCTCN2022120885-appb-000024
其中
Figure PCTCN2022120885-appb-000025
表示向量相乘，V 1表示所述温度输入向量，
Figure PCTCN2022120885-appb-000026
表示所述温度输入向量的转置向量，V 2表示所述压力输入向量，M表示所述温度-压力关联矩阵。
具体地，在本申请实施例中，所述数据级关联模块230，用于将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵。也就是，在本申请的技术方案中，进一步将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络中进行处理以得到温度-压力关联特征矩阵。应可以理解，使用邻卷积层为互为转置的卷积核的卷积神经网络模型进行处理，能够在训练时能够同时更新网络参数的更新和适合特定数据结构的网络参数结构的搜索，进而提高分类的准确性。
更具体地，在本申请实施例中，所述数据级关联模块，包括：浅层特征图提取子单元，用于从所述第一卷积神经网络的第M层提取浅层特征矩阵，M是偶数；深层特征图提取子单元，用于从所述第一卷积神经网络的第N层提取深层特征矩阵，其中，N为偶数，且N大于M的2倍；以及，特征图融合子单元，用于融合所述浅层特征图和所述深层特征图以生成所述温度-压力关联特征矩阵。
具体地，在本申请实施例中，所述转速编码模块240，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量。应可以理解，对于所述预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值，考虑到其在时序上具有着动态的隐含规律，因此，在本申请的技术方案中，为了更为充分地提取出这种动态变化的隐含特征规律，使用包含一维卷积层的时序编码器对其进行编码，以提取出所述转炉反应器的转速值在时间维度上的动态变化特征，从而得到转速特征向量。相应地，在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出所述转炉反应器的转速值在时序维度上的关联和通过全连接编码提取所述转炉反应器的转速值的高维隐含特征。
更具体地，在本申请实施例中，所述转速编码模块，包括：输入向量构造单元，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值按照时间维度排列为一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连 接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000027
Figure PCTCN2022120885-appb-000028
其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
Figure PCTCN2022120885-appb-000029
表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
Figure PCTCN2022120885-appb-000030
其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸，X表示所述输入向量。
具体地，在本申请实施例中，所述全局控制特征生成模块250，用于将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量。也就是，在本申请的技术方案中，在得到具有所述炉内压力值和所述反应温度值之间关联特征的所述温度-压力关联特征矩阵以及所述转炉反应器的转速值的动态变化特征的所述转速特征向量后，进一步将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以整合变化参数的特征关联信息，从而得到参数全局控制特征向量。
具体地，在本申请实施例中，所述产物数据编码模块260，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量。应可以理解，对于所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图，考虑到所述产物的气相色谱图在时间维度上也是动态变化的，因此若想基于产物的变化状态特征信息来对于所述转炉反应器的转速值进行实时动态地控制，在本申请的技术方案中，还需要将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图的局部动态隐含特征分布，从而得到产物特征向量。
更具体地，在本申请实施例中，所述产物数据编码模块，进一步用于：所述使用三维卷积核的第二卷积神经网络在层的正向传递中对输入数据分别进行：基于所述三维卷积核对所述输入数据进行三维卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部特征矩阵的均值池化处理以得到池化特征图；以及，对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第二卷积神经网络的最后一层的输出为所述产物特征向量，所述第二卷积神经网络的第一层的输入为所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图。
具体地，在本申请实施例中，所述校正模块270，用于对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量。应可以理解，注意到对所述产物特征图进行沿通道维度的全局均值池化是基于图像语义的像素级的前向全局下采样，这会导致所述产物特征向量的各个位置的特征值之间的相关性较差，从而影响其对所述产物特征图沿通道维度上的分布的表达能力。因此，在本申请的技术方案中，进一步还需对产物特征向量进行前向传播相关性的引导修正。
更具体地，在本申请实施例中，所述校正模块，进一步用于：以如下公式对所述产物特征向量中各个位置的特征值进行修正以得到所述校正后产物特征向量；其中，所述公式为：
Figure PCTCN2022120885-appb-000031
Figure PCTCN2022120885-appb-000032
其中V表示所述产物特征向量，∑表示所述产物特征向量的自协方差矩阵，即矩阵的每个位置的值是所述产物特征向量的每两个位置的特征值之间的方差，μ和σ分别是所述产物特征向量的全局均值和方差，
Figure PCTCN2022120885-appb-000033
和
Figure PCTCN2022120885-appb-000034
分别表示特征向量的按位置减法和加法，||·|| 2表示特征向量的二范数，exp(·)表示向量的指数运算，以向量为幂的指数运算表示以向量的每个位置的特征值作为幂的自然指数函数值。应可以理解，这里，该前向传播相关性的引导修正通过对所述产物特征向量的可学习的正态采样偏移引导特征工程来有效地建模其长度方向上的各个特征值之间的长程依赖关系，以进一步考虑特征值的局部和非局部邻域来进行特征值之间的相关性的修复。这样，在继续计算其与所述参数全局控制特征向量的转移矩阵作为分类特征矩阵并进行分类时，可以提高其在模型中继续随深度流式传播时对于类概率的预测能力，进而提高分类的准确性。
具体地，在本申请实施例中，所述作用表示模块280和所述能源控制结果生成模块290，用于计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵，并将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。应可以理解，由于所述产物的气相色谱图的动态变化特征与所述参数关联的隐含特征之间的特征尺度不同，并且所述产物的气相色谱图的动态特征在高维特征空间中可以看作是对于所述参数关联变化的响应性特征，因此为了更好地融合所述参数全局控制特征向量和所述产物特征向量，进一步计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵。这样， 就可以将所述分类特征矩阵通过分类器以得到用于表示当前时间点的转炉反应器的转速值应增大或应减小的分类结果。
相应地，在一个具体示例中，所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
更具体地，在本申请实施例中，所述作用表示模块，进一步用于：以如下公式计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为所述分类特征矩阵；其中，所述公式为：
V 1＝M*V 2
其中V 1表示所述参数全局控制特征向量，M表示所述分类特征矩阵，V 2表示所述产物特征向量。
综上，基于本申请实施例的所述用于无水氟化氢生产的能源管理控制系统200被阐明，其采用人工智能控制技术，通过基于深度神经网络模型来对于转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图进行时序维度上的隐含关联特征提取，并根据产物的实时生成状态特征以及炉内压力和反应温度的变化特征来实时动态地对于所述转炉反应器的转速进行调整以提高反应的效率，进而提高反应物料的利用率与产品的质量。
如上所述，根据本申请实施例的用于无水氟化氢生产的能源管理控制系统200可以实现在各种终端设备中，例如用于无水氟化氢生产的能源管理控制算法的服务器等。在一个示例中，根据本申请实施例的用于无水氟化氢生产的能源管理控制系统200可以作为一个软件模块和/或硬件模块而集成到终端设备中。例如，该用于无水氟化氢生产的能源管理控制系统200可以是该终端设备的操作系统中的一个软件模块，或者可以是针对于该终端设备所开发的一个应用程序；当然，该用于无水氟化氢生产的能源管理控制系统200同样可以是该终端设备的众多硬件模块之一。
替换地，在另一示例中，该用于无水氟化氢生产的能源管理控制系统200与该终端设备也可以是分立的设备，并且该用于无水氟化氢生产的能源管理控制系统200可以通过有线和/或无线网络连接到该终端设备，并且按照约定的数据格式来传输交互信息。
示例性方法
图3图示了用于无水氟化氢生产的能源管理控制系统的控制方法的流程图。如图3所示，根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的控制方法，包括步骤：S110，获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；S120，将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；S130，将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；S140，将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；S150，将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；S160，将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；S170，对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；S180，计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及，S190，将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
图4图示了根据本申请实施例的用于无水氟化氢生产的能源管理控制系统的控制方法的架构示意图。如图4所示，在所述用于无水氟化氢生产的能源管理控制系统的控制方法的网络架构中，首先，将获得的所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值(例如，如图4中所示意的P1)和反应温度值(例如，如图4中所示意的P2)分别按照时间维度排列为温度输入向量(例如，如图4中所示意的V1)和压力输入向量(例如，如图4中所示意的V2)后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵(例如，如图4中所示意的M1)；接着，将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络(例如，如图4中所示意的CNN1)以得到温度-压力关联特征矩阵(例如，如图4中所示意的MF1)；然后，将获得的预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值(例如，如图4中所示意的P3)通过包含一维卷积层的时序编码器(例如，如图4中所示意的E)以得到转速特征向量(例如，如图4中所示意的VF1)；接着，将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量(例如，如图4中所示意的VF2)；然后，将获得的所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图(例如，如图4中所示意的P4)通过使用三维卷积核的第二卷积神经网络(例如，如图4中所示意的CNN2)以得到产物特征向量(例如，如图4中所示意的VF3)；接着，对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量 (例如，如图4中所示意的VF4)；然后，计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵(例如，如图4中所示意的MF)；以及，最后，将所述分类特征矩阵通过分类器(例如，如图4中所示意的分类器)以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
更具体地，在步骤S110和步骤S120中，获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图，并将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵。应可以理解，由于对于转炉反应器的转速控制、炉内压力控制以及反应温度控制不仅关乎于反应效率和反应充分度，还关乎于无水氟化氢生产产线的能耗。因此，在本申请的技术方案中，期望能够根据产物的生成状态特征来实时动态地对于所述转炉反应器的转速、炉内压力以及反应温度进行控制，以提高反应的效率，进而能够提高反应物料的利用率。
也就是，具体地，在本申请的技术方案中，首先，通过各个传感器，例如转速传感器、压力传感器和温度传感器分别采集预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值和反应温度值，并且通过气相色谱仪采集所述多个预定时间点的产物的气相色谱图。应可以理解，由于所述炉内压力值和所述反应温度值之间具有着关联关系，其会相互产生影响，例如当炉内压力值升高时，反应温度也会随着增大，因此，在本申请的技术方案中，需要进一步挖掘出这种关联性特征来进行实际转炉反应器的转速控制。具体地，将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵，以整合所述炉内压力值和所述反应温度值的关联性，便于后续对其特征进行提取。
更具体地，在步骤S130中，将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵。也就是，在本申请的技术方案中，进一步将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络中进行处理以得到温度-压力关联特征矩阵。应可以理解，使用邻卷积层为互为转置的卷积核的卷积神经网络模型进行处理，能够在训练时能够同时更新网络参数的更新和适合特定数据结构的网络参数结构的搜索，进而提高分类的准确性。
更具体地，在步骤S140中，将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量。应可以理解，对于所述预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值，考虑到其在时序上具有着动态的隐含规律，因此，在本申请的技术方案中，为了更为充分地提取出这种动态变化的隐含特征规律，使用包含一维卷积层的时序编码器对其进行编码，以提取出所述转炉反应器的转速值在时间维度上的动态变化特征，从而得到转速特征向量。相应地，在一个具体示例中，所述时序编码器由交替设置的全连接层和一维卷积层组成，其通过一维卷积编码提取出所述转炉反应器的转速值在时序维度上的关联和通过全连接编码提取所述转炉反应器的转速值的高维隐含特征。
更具体地，在步骤S150中，将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量。也就是，在本申请的技术方案中，在得到具有所述炉内压力值和所述反应温度值之间关联特征的所述温度-压力关联特征矩阵以及所述转炉反应器的转速值的动态变化特征的所述转速特征向量后，进一步将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以整合变化参数的特征关联信息，从而得到参数全局控制特征向量。
更具体地，在步骤S160中，将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量。应可以理解，对于所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图，考虑到所述产物的气相色谱图在时间维度上也是动态变化的，因此若想基于产物的变化状态特征信息来对于所述转炉反应器的转速值进行实时动态地控制，在本申请的技术方案中，还需要将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图的局部动态隐含特征分布，从而得到产物特征向量。
更具体地，在步骤S170中，对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量。应可以理解，注意到对所述产物特征图进行沿通道维度的全局均值池化是基于图像语义的像素级的前向全局下采样，这会导致所述产物特征向量的各个位置的特征值之间的相关性较差，从而影响其对所述产物特征图沿通道维度上的分布的表达能力。因此，在本申请的技术方案中，进一步还需对产物特征向量进行前向传播相关性的引导修正。
更具体地，在步骤S180和步骤S190中，计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵，并将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。应可以理解，由于所述产物的气相色谱图的动态变化特征与所述参数关联的隐含特征之间的特征尺度不同，并且所述产物的气相色谱图的动态特征在高维特征空间中可以看作是对于所述参数关联变化的响应性特征，因此为了更好地融合所述参数全局控制特征向量和所述产物特征向量，进一步计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵。这样，就可以将所述分类特征矩阵通过分类器以得到用于表示当前时间点的转炉反应器的转速值应增大或应减小的分类结果。
综上，基于本申请实施例的所述用于无水氟化氢生产的能源管理控制系统的控制方法被阐明，其采用人工智能控制技术，通过基于深度神经网络模型来对于转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图进行时序维度上的隐含关联特征提取，并根据产物的实时生成状态特征以及炉内压力和反应温度的变化特征来实时动态地对于所述转炉反应器的转速进行调整以提高反应的效率，进而提高反应物料的利用率与产品的质量。
以上结合具体实施例描述了本申请的基本原理，但是，需要指出的是，在本申请中提及的优点、优势、效果等仅是示例而非限制，不能认为这些优点、优势、效果等是本申请的各个实施例必须具备的。另外，上述公开的具体细节仅是为了示例的作用和便于理解的作用，而非限制，上述细节并不限制本申请为必须采用上述具体的细节来实现。
本申请中涉及的器件、装置、设备、系统的方框图仅作为例示性的例子并且不意图要求或暗示必须按照方框图示出的方式进行连接、布置、配置。如本领域技术人员将认识到的，可以按任意方式连接、布置、配置这些器件、装置、设备、系统。诸如“包括”、“包含”、“具有”等等的词语是开放性词汇，指“包括但不限于”，且可与其互换使用。这里所使用的词汇“或”和“和”指词汇“和/或”，且可与其互换使用，除非上下文明确指示不是如此。这里所使用的词汇“诸如”指词组“诸如但不限于”，且可与其互换使用。
还需要指出的是，在本申请的装置、设备和方法中，各部件或各步骤是可以分解和/或重新组合的。这些分解和/或重新组合应视为本申请的等效方案。
提供所公开的方面的以上描述以使本领域的任何技术人员能够做出或者使用本申请。对这些方面的各种修改对于本领域技术人员而言是非常显而易见的，并且在此定义的一般原理可以应用于其他方面而不脱离本申请的范围。因此，本申请不意图被限制到在此示出的方面，而是按照与在此公开的原理和新颖的特征一致的最宽范围。
为了例示和描述的目的已经给出了以上描述。此外，此描述不意图将本申请的实施例限制到在此公开的形式。尽管以上已经讨论了多个示例方面和实施例，但是本领域技术人员将认识到其某些变型、修改、改变、添加和子组合。

Claims (10)

  1. 一种用于无水氟化氢生产的能源管理控制系统，其特征在于，包括：数据采集模块，用于获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；结构化模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；数据级关联模块，用于将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；转速编码模块，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；全局控制特征生成模块，用于将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；产物数据编码模块，用于将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；校正模块，用于对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；作用表示模块，用于计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及能源控制结果生成模块，用于将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
  2. 根据权利要求1所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述结构化模块，包括：数据级关联单元，用于以如下公式来计算所述温度输入向量和所述压力输入向量之间的所述温度-压力关联矩阵；其中，所述公式为：
    Figure PCTCN2022120885-appb-100001
    其中
    Figure PCTCN2022120885-appb-100002
    表示向量相乘，V 1表示所述温度输入向量，
    Figure PCTCN2022120885-appb-100003
    表示所述温度输入向量的转置向量，V 2表示所述压力输入向量，M表示所述温度-压力关联矩阵。
  3. 根据权利要求2所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述数据级关联模块，包括：浅层特征图提取子单元，用于从所述第一卷积神经网络的第M层提取浅层特征矩阵，M是偶数；深层特征图提取子单元，用于从所述第一卷积神经网络的第N层提取深层特征矩阵，其中，N为偶数，且N大于M的2倍；以及特征图融合子单元，用于融合所述浅层特征图和所述深层特征图以生成所述温度-压力关联特征矩阵。
  4. 根据权利要求3所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述转速编码模块，包括：输入向量构造单元，用于将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值按照时间维度排列为一维的输入向量；全连接编码单元，用于使用所述时序编码器的全连接层以如下公式对所述输入向量进行全连接编码以提取出所述输入向量中各个位置的特征值的高维隐含特征，其中，所述公式为：
    Figure PCTCN2022120885-appb-100004
    其中X是所述输入向量，Y是输出向量，W是权重矩阵，B是偏置向量，
    Figure PCTCN2022120885-appb-100005
    表示矩阵乘；一维卷积编码单元，用于使用所述时序编码器的一维卷积层以如下公式对所述输入向量进行一维卷积编码以提取出所述输入向量中各个位置的特征值间的高维隐含关联特征，其中，所述公式为：
    Figure PCTCN2022120885-appb-100006
    其中，a为卷积核在x方向上的宽度、F为卷积核参数向量、G为与卷积核函数运算的局部向量矩阵，w为卷积核的尺寸，X表示所述输入向量。
  5. 根据权利要求4所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述产物数据编码模块，进一步用于：所述使用三维卷积核的第二卷积神经网络在层的正向传递中对输入数据分别进行：基于所述三维卷积核对所述输入数据进行三维卷积处理以得到卷积特征图；对所述卷积特征图进行基于局部特征矩阵的均值池化处理以得到池化特征图；以及
    对所述池化特征图进行非线性激活以得到激活特征图；其中，所述第二卷积神经网络的最后一层的输出为所述产物特征向量，所述第二卷积神经网络的第一层的输入为所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图。
  6. 根据权利要求5所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述校正模块，进一步用于：以如下公式对所述产物特征向量中各个位置的特征值进行修正以得到所述校正后产物特征向量；其中，所述公式为：
    Figure PCTCN2022120885-appb-100007
    其中V表示所述产物特征向量，∑表示所述产物特征向量的自协方差矩阵，即矩阵的每个位置的值是所述产物特征向量的每两个位置的特征值之间的方差，μ和σ分别是所述产物特征向量的全局均值和方差，
    Figure PCTCN2022120885-appb-100008
    和
    Figure PCTCN2022120885-appb-100009
    分别表示特征向量的按位置减法和加法，||·|| 2表示特征向量的二范数，exp(·)表示向量的指数运算，以向量为幂的指数运算表示以向量的每个位置的特征值作为幂的自然指数函数值。
  7. 根据权利要求6所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述作用表 示模块，进一步用于：以如下公式计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为所述分类特征矩阵；其中，所述公式为：
    V 1＝M*V 2
    其中V 1表示所述参数全局控制特征向量，M表示所述分类特征矩阵，V 2表示所述产物特征向量。
  8. 根据权利要求7所述的用于无水氟化氢生产的能源管理控制系统，其特征在于，所述能源控制结果生成模块，进一步用于：所述分类器以如下公式对所述分类特征矩阵进行处理以生成分类结果，其中，所述公式为：softmax{(W n,B n):…:(W 1,B 1)|Project(F)}，其中Project(F)表示将所述分类特征矩阵投影为向量，W 1至W n为各层全连接层的权重矩阵，B 1至B n表示各层全连接层的偏置矩阵。
  9. 一种用于无水氟化氢生产的能源管理控制系统的控制方法，其特征在于，包括：
    获取预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值、炉内压力值、反应温度值和产物的气相色谱图；将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵；将所述温度-压力关联矩阵通过相邻层使用互为转置的卷积核的第一卷积神经网络以得到温度-压力关联特征矩阵；将预定时间段内包含当前时间点在内的多个预定时间点的转炉反应器的转速值通过包含一维卷积层的时序编码器以得到转速特征向量；将所述转速特征向量与所述温度-压力关联特征矩阵进行相乘以得到参数全局控制特征向量；将所述预定时间段内包含当前时间点在内的多个预定时间点的产物的气相色谱图通过使用三维卷积核的第二卷积神经网络以得到产物特征向量；对所述产物特征向量中各个位置的特征值进行修正以得到校正后产物特征向量；计算所述参数全局控制特征向量相对于所述产物特征向量的转移矩阵作为分类特征矩阵；以及
    将所述分类特征矩阵通过分类器以得到分类结果，所述分类结果用于表示当前时间点的转炉反应器的转速值应增大或应减小。
  10. 根据权利要求9所述的用于无水氟化氢生产的能源管理控制系统的控制方法，其特征在于，所述将所述预定时间段内包含当前时间点在内的多个预定时间点的炉内压力值和反应温度值分别按照时间维度排列为温度输入向量和压力输入向量后，计算所述温度输入向量和所述压力输入向量之间的温度-压力关联矩阵，包括：以如下公式来计算所述温度输入向量和所述压力输入向量之间的所述温度-压力关联矩阵；其中，所述公式为：
    Figure PCTCN2022120885-appb-100010
    其中
    Figure PCTCN2022120885-appb-100011
    表示向量相乘，V 1表示所述温度输入向量，
    Figure PCTCN2022120885-appb-100012
    表示所述温度输入向量的转置向量，V 2表示所述压力输入向量，M表示所述温度-压力关联矩阵。
PCT/CN2022/120885 2022-08-31 2022-09-23 用于无水氟化氢生产的能源管理控制系统及其控制方法 WO2024045244A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211055093.9A CN115599049B (zh) 2022-08-31 2022-08-31 用于无水氟化氢生产的能源管理控制系统及其控制方法
CN202211055093.9 2022-08-31

Publications (1)

Publication Number Publication Date
WO2024045244A1 true WO2024045244A1 (zh) 2024-03-07

Family

ID=84843304

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/120885 WO2024045244A1 (zh) 2022-08-31 2022-09-23 用于无水氟化氢生产的能源管理控制系统及其控制方法

Country Status (2)

Country Link
CN (1) CN115599049B (zh)
WO (1) WO2024045244A1 (zh)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118030040A (zh) * 2024-04-11 2024-05-14 克拉玛依市富城油气研究院有限公司 用于采油工程的生产动态监测系统及方法

Families Citing this family (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116110507B (zh) * 2023-02-15 2024-02-06 浙江宏丰炉料有限公司 镁碳砖智能化生产方法及系统
CN116284130B (zh) * 2023-02-22 2024-04-19 森淼(山东)药业有限公司 磷霉素钠的制备工艺
CN116825217B (zh) * 2023-03-15 2024-05-14 福建省德旭新材料有限公司 制备高纯五氟化磷的方法
CN116048031B (zh) * 2023-03-31 2023-08-04 克拉玛依市蓝润环保科技有限责任公司 用于石油助剂生产的控制系统及其方法
CN116820052B (zh) * 2023-07-13 2024-02-02 滁州优胜高分子材料有限公司 Pbt材料生产设备及其控制方法

Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4294816A (en) * 1979-10-20 1981-10-13 Vereinigte Aluminium-Werke Aktiengesellschaft Process for producing hydrogen fluoride from fluorine containing materials by continuous pyrohydrolysis
CN103848400A (zh) * 2012-12-03 2014-06-11 福建省邵武市永飞化工有限公司 一种无水氟化氢的制备方法
US20170316312A1 (en) * 2016-05-02 2017-11-02 Cavium, Inc. Systems and methods for deep learning processor
CN113321185A (zh) * 2021-05-20 2021-08-31 浙江瑞星氟化工业有限公司 一种无水氟化氢生产工艺
CN114768279A (zh) * 2022-04-29 2022-07-22 福建德尔科技股份有限公司 用于电子级二氟甲烷制备的精馏控制系统及其控制方法
CN114870416A (zh) * 2022-04-28 2022-08-09 福建德尔科技股份有限公司 用于电子级一氟甲烷制备的精馏控制系统及精馏控制方法
CN114913401A (zh) * 2022-07-13 2022-08-16 江苏烨明光电有限公司 Led灯芯柱与外壳的焊接设备及其焊接质量监控方法

Family Cites Families (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107703217A (zh) * 2016-08-09 2018-02-16 江苏康缘药业股份有限公司 一种过敏性鼻炎颗粒指纹图谱的制备方法及获得的指纹图谱
CN110187029A (zh) * 2019-06-06 2019-08-30 重庆医科大学 沉香化气片气相指纹图谱的建立方法及其指纹图谱
BR112021022593A2 (pt) * 2019-06-21 2022-01-04 Halliburton Energy Services Inc Método e sistema
CN112699919B (zh) * 2020-12-15 2024-02-27 广东工业大学 基于三维卷积神经网络模型的木材鉴别方法、装置
CN114815930B (zh) * 2022-06-30 2022-09-02 烟台黄金职业学院 煅烧器的温度控制系统及其温度控制方法

Patent Citations (7)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US4294816A (en) * 1979-10-20 1981-10-13 Vereinigte Aluminium-Werke Aktiengesellschaft Process for producing hydrogen fluoride from fluorine containing materials by continuous pyrohydrolysis
CN103848400A (zh) * 2012-12-03 2014-06-11 福建省邵武市永飞化工有限公司 一种无水氟化氢的制备方法
US20170316312A1 (en) * 2016-05-02 2017-11-02 Cavium, Inc. Systems and methods for deep learning processor
CN113321185A (zh) * 2021-05-20 2021-08-31 浙江瑞星氟化工业有限公司 一种无水氟化氢生产工艺
CN114870416A (zh) * 2022-04-28 2022-08-09 福建德尔科技股份有限公司 用于电子级一氟甲烷制备的精馏控制系统及精馏控制方法
CN114768279A (zh) * 2022-04-29 2022-07-22 福建德尔科技股份有限公司 用于电子级二氟甲烷制备的精馏控制系统及其控制方法
CN114913401A (zh) * 2022-07-13 2022-08-16 江苏烨明光电有限公司 Led灯芯柱与外壳的焊接设备及其焊接质量监控方法

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118030040A (zh) * 2024-04-11 2024-05-14 克拉玛依市富城油气研究院有限公司 用于采油工程的生产动态监测系统及方法

Also Published As

Publication number Publication date
CN115599049B (zh) 2023-04-07
CN115599049A (zh) 2023-01-13

Similar Documents

Publication Publication Date Title
WO2024045244A1 (zh) 用于无水氟化氢生产的能源管理控制系统及其控制方法
WO2024045247A1 (zh) 用于氟化铵生产的生产管理控制系统及其控制方法
CN115586755B (zh) 用于六氟磷酸锂制备的生产管理控制系统及其方法
CN114870416B (zh) 用于电子级一氟甲烷制备的精馏控制系统及精馏控制方法
WO2024026993A1 (zh) 氟化铵制备用的自动配料控制系统及其控制方法
WO2023206724A1 (zh) 用于电子级二氟甲烷制备的精馏控制系统及其控制方法
WO2024045243A1 (zh) 用于六氟磷酸锂制备的自动配料系统及其配料方法
WO2024021254A1 (zh) 电子级三氟化氯的智能分离纯化系统
WO2024021259A1 (zh) 用于缓冲氧化物蚀刻液生产的自动配料系统及其配料方法
WO2024000798A1 (zh) 用于电子级氢氟酸制备的生产管理控制系统及其控制方法
Sanderson et al. Simultaneous chemical and phase equilibrium calculation
WO2023226226A1 (zh) 用于电子级三氟甲烷制备的精馏控制系统及其控制方法
CN112102890B (zh) 一种基于机器学习模型的mcs合成装置反应温度预测方法
WO2024113599A1 (zh) 用于电子级六氟丁二烯制备的生产管理控制系统
CN115982736B (zh) 用于计算机网络信息的数据加密方法及系统
CN114048818A (zh) 一种基于加速Transformer模型的视频分类方法
WO2023226236A1 (zh) 用于电子级氢氟酸制备的能源管理控制系统及其控制方法
WO2024036690A1 (zh) 用于剥膜液生产的自动配料系统及其配料方法
CN115090200B (zh) 用于电子级氢氟酸制备的自动配料系统及其配料方法
Qi et al. Using stacked auto-encoder and bi-directional LSTM for batch process quality prediction
CN114330143A (zh) 一种基于多源时空信息的分布式参数系统状态预测方法
CN116920739B (zh) 用于六氟磷酸锂制备的液体循环雾化合成控制系统及方法
CN117046140B (zh) 用于电子级氢氟酸精馏的装置
WO2024138944A1 (zh) 用于电子级四氟化碳的废气回收再利用方法及其系统
CN116825217B (zh) 制备高纯五氟化磷的方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 22957063

Country of ref document: EP

Kind code of ref document: A1