WO2024021258A1 - 电子级氢氧化钾的智慧产线的控制系统及其控制方法 - Google Patents

电子级氢氧化钾的智慧产线的控制系统及其控制方法 Download PDF

Info

Publication number
WO2024021258A1
WO2024021258A1 PCT/CN2022/119555 CN2022119555W WO2024021258A1 WO 2024021258 A1 WO2024021258 A1 WO 2024021258A1 CN 2022119555 W CN2022119555 W CN 2022119555W WO 2024021258 A1 WO2024021258 A1 WO 2024021258A1
Authority
WO
WIPO (PCT)
Prior art keywords
feature
vector
dynamic
feature map
eigenvector
Prior art date
Application number
PCT/CN2022/119555
Other languages
English (en)
French (fr)
Inventor
罗霜
林金华
石凌斌
张永彪
袁瑞明
Original Assignee
福建天甫电子材料有限公司
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 福建天甫电子材料有限公司 filed Critical 福建天甫电子材料有限公司
Publication of WO2024021258A1 publication Critical patent/WO2024021258A1/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
    • G05B19/04—Programme control other than numerical control, i.e. in sequence controllers or logic controllers
    • G05B19/042—Programme control other than numerical control, i.e. in sequence controllers or logic controllers using digital processors
    • G05B19/0428—Safety, monitoring
    • C—CHEMISTRY; METALLURGY
    • C01—INORGANIC CHEMISTRY
    • C01D—COMPOUNDS OF ALKALI METALS, i.e. LITHIUM, SODIUM, POTASSIUM, RUBIDIUM, CAESIUM, OR FRANCIUM
    • C01D1/00—Oxides or hydroxides of sodium, potassium or alkali metals in general
    • C01D1/04—Hydroxides
    • 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
    • G05B2219/20—Pc systems
    • G05B2219/24—Pc safety
    • G05B2219/24024—Safety, surveillance

Definitions

  • the present invention relates to intelligent control in the field of intelligent manufacturing, and more specifically, to a control system and a control method for a smart production line of electronic grade potassium hydroxide.
  • Patent 201010576437.1 discloses a method of dissolving potassium metaborate into an aqueous sodium hydroxide solution, heating to reflux, and solid-liquid separation to obtain solid potassium hydroxide. This method has a complicated process flow. , high energy consumption, and does not involve the purity of potassium hydroxide.
  • Patent 201410290942.8 discloses a continuous production method of high-purity potassium hydroxide aqueous solution. In this method, the potassium hydroxide solution needs to pass through chelating resin, cation exchange resin, anion exchange resin, electrodialysis and multi-stage filter circulation system to remove impurities in sequence. This process The method steps are cumbersome and the process cost is high.
  • Patent CN202010747995.3 uses purified potassium chloride for electrolysis to produce electronic-grade potassium hydroxide. There are problems such as complex process routes and difficulty in handling acidic waste brine.
  • Embodiments of the present application provide a control system and control method for a smart production line of electronic grade potassium hydroxide, which uses artificial intelligence control technology and is based on a deep neural network model to control the power dynamic change characteristics of the sprayer and spray impurity removal.
  • the dynamic characteristics of the image frames are deeply mined to dynamically control the power of the sprayer in real-time and dynamically in the smart production line of electronic grade potassium hydroxide, thereby ensuring the effect of crystal yield and the efficiency of impurity transfer.
  • a control system for a smart production line of electronic grade potassium hydroxide which includes: a spray data acquisition module used to obtain the power values of the sprayer at multiple predetermined time points within a predetermined time period; Monitoring video of spray impurities collected by the camera during the predetermined time period; a spatial encoding module for extracting multiple key frames from the monitoring video, and separately encoding each of the multiple key frames using spatial
  • the first convolutional neural network of the attention mechanism is used to obtain multiple spatially focused feature maps; the difference module is used to calculate the difference between each adjacent two spatially focused feature maps in the multiple spatially focused feature maps to obtain multiple spatially focused feature maps.
  • a differential feature map for passing the plurality of differential feature maps through a second convolutional neural network using a three-dimensional convolution kernel to obtain a de-impurity dynamic feature map; a dimensionality reduction module for decontaminating the The impurity dynamic feature map performs global mean pooling along the channel dimension to obtain the impurity removal dynamic feature vector; the correction module is used to calculate each of the impurity removal dynamic feature vectors based on the autocovariance matrix of the impurity removal dynamic feature vector.
  • the eigenvalues of the positions are corrected to obtain the corrected impurity-removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity-removal dynamic eigenvector are the corresponding two positions of the impurity-removal dynamic eigenvector.
  • a power timing encoding module used to pass the power values of the atomizer at multiple predetermined time points within the predetermined time period through a timing encoder including a one-dimensional convolution layer to obtain a power control feature vector; response a linear estimation module for calculating the transfer matrix of the power control eigenvector relative to the corrected impurity removal dynamic eigenvector; and a control result generation module for passing the transfer matrix through a classifier to obtain a classification result, so
  • the above classification results are used to indicate that the power value of the nebulizer at the current point in time should be increased or decreased.
  • the spatial encoding module is further used to: each layer of the first convolutional neural network performs on the input data respectively in the forward transmission of the layer: Perform convolution processing based on a two-dimensional convolution kernel on the input data to generate a convolution feature map; perform pooling processing on the convolution feature map to generate a pooled feature map; activate the pooled feature map Process to generate an activation feature map; perform global average pooling along the channel dimension on the activation feature map to obtain a spatial feature matrix; perform convolution processing and activation processing on the spatial feature matrix to generate a weight vector; and, so The weight value of each position in the weight vector weights each feature matrix of the activation feature map to obtain a generated feature map; wherein, the generated feature map output by the last layer of the first convolutional neural network is For the plurality of spatially focused feature maps, the input of the first layer of the first convolutional neural network is each key frame among the plurality of key frames
  • the differential module is further used to calculate the difference between each two adjacent spatial focus feature maps in the plurality of spatial focus feature maps using the following formula: Difference to obtain the multiple differential feature maps; wherein, the formula is:
  • F i represents the spatial focus feature map at the i-th position in the plurality of spatial focus feature maps
  • F i+1 represents the spatial focus feature map at the i+1th position in the plurality of spatial focus feature maps
  • F represents the differential feature map
  • the dynamic encoding module is further used to: the second convolutional neural network using a three-dimensional convolution kernel processes 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 on the convolution feature map to obtain a pooling feature map; and, perform 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 impurity-free dynamic feature map, and the first layer of the second convolutional neural network
  • the input of the layer is the plurality of differential feature maps.
  • the correction module is further used to: based on the auto-covariance matrix of the impurity removal dynamic eigenvector, calculate the impurity removal dynamic eigenvector according to the following formula The eigenvalues of each position in are corrected to obtain the corrected impurity removal dynamic eigenvector; wherein, the formula is:
  • V represents the impurity removal dynamic eigenvector
  • ⁇ is the autocovariance matrix of the impurity removal dynamic eigenvector
  • ⁇ and ⁇ are the global mean and variance of the impurity removal dynamic eigenvector respectively
  • exp( ⁇ ) represents Exponential operation using a vector as a power
  • the exponential operation using a vector as a power means using the value of each position of the vector as a power to find the exponent, and then filling the result into each position of the vector to obtain the vector operation result, and Represent position-wise subtraction and addition of vectors respectively
  • 2 represents the second norm of the feature vector.
  • the power timing encoding module is further used to: arrange the power values of the sprayers at multiple predetermined time points within the predetermined time period into one according to the time dimension.
  • dimensional input vector use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector with the following formula to extract the high-dimensional hidden features of the eigenvalues of each position in the input vector, where, the formula for: 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, Among them, 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 responsiveness estimation module is further used to calculate all the parameters of the power control eigenvector relative to the corrected impurity removal dynamic eigenvector according to the following formula:
  • the transfer matrix wherein, the formula is:
  • V 2 M*V 1
  • V 1 represents the power control eigenvector
  • M represents the transfer matrix
  • V 2 represents the corrected impurity removal dynamic eigenvector
  • the control result generation module processes the transfer matrix according to the following formula to generate a classification result, wherein, The formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • a method for controlling a smart production line of electronic grade potassium hydroxide which includes: obtaining the power values of the sprayer at multiple predetermined time points within a predetermined time period and the predetermined time collected by a camera. A segment of surveillance video with spray removal of impurities; extract multiple key frames from the surveillance video, and pass each of the multiple key frames through the first convolutional neural network using the spatial attention mechanism to obtain multiple key frames.
  • Spatially focused feature maps calculate the difference between each adjacent two spatial focused feature maps in the plurality of spatially focused feature maps to obtain multiple differential feature maps; convert the multiple differential feature maps by using a three-dimensional convolution kernel
  • the second convolutional neural network is used to obtain the impurity-removal dynamic feature map; the global mean pooling along the channel dimension is performed on the impurity-removal dynamic feature map to obtain the impurity-removal dynamic feature vector; the automatic algorithm based on the impurity-removal dynamic feature vector
  • the covariance matrix corrects the eigenvalues of each position in the impurity removal dynamic eigenvector to obtain the corrected impurity removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity removal dynamic eigenvector are is the variance between the eigenvalues of the corresponding two positions in the impurity removal dynamic eigenvector; the power values of the sprayers at multiple predetermined time points within the predetermined time period are
  • Convolutional neural network to obtain multiple spatially focused feature maps, including: each layer of the first convolutional neural network performs on the input data respectively in the forward pass of the layer: performs on the input data based on a two-dimensional convolution kernel Convolution processing to generate a convolution feature map; pooling processing on the convolution feature map to generate a pooling feature map; activation processing on the pooling feature map to generate an activation feature map; performing activation processing on the activation feature map Perform global average pooling along the channel dimension to obtain a spatial feature matrix; perform convolution processing and activation processing on the spatial feature matrix to generate a weight vector; and use the weight values of each position in the weight vector to Each feature matrix of the activation feature map is weighted to obtain a generated feature map; wherein the generated feature map output by the last layer of the first convolutional neural network is the pluralit
  • the difference between each two adjacent spatial focus feature maps in the multiple spatial focus feature maps is calculated to obtain multiple differential feature maps, including:
  • the following formula calculates the difference between each adjacent two spatial focus feature maps in the plurality of spatial focus feature maps to obtain the plurality of differential feature maps; wherein the formula is:
  • F i represents the spatial focus feature map at the i-th position in the plurality of spatial focus feature maps
  • F i+1 represents the spatial focus feature map at the i+1th position in the plurality of spatial focus feature maps
  • F represents the differential feature map
  • the plurality of differential feature maps are passed through a second convolutional neural network using a three-dimensional convolution kernel to obtain an impurity removal dynamic feature map, including: using The second convolutional neural network of the three-dimensional convolution kernel separately performs three-dimensional convolution processing on the input data based on the three-dimensional convolution kernel to obtain a convolution feature map; The convolution feature map is subjected to mean pooling processing to obtain a pooled feature map; and, the pooled feature map is nonlinearly activated to obtain an activation feature map; wherein, the last layer of the second convolutional neural network The output is the impurity-removing dynamic feature map, and the input of the first layer of the second convolutional neural network is the plurality of differential feature maps.
  • the eigenvalues of each position in the impurity removal dynamic eigenvector are corrected to obtain the corrected
  • the impurity removal dynamic eigenvector includes: based on the autocovariance matrix of the impurity removal dynamic eigenvector, the eigenvalues of each position in the impurity removal dynamic eigenvector are corrected with the following formula to obtain the corrected impurity removal dynamic eigenvector.
  • Characteristic vector where, the formula is:
  • V represents the impurity removal dynamic eigenvector
  • ⁇ is the autocovariance matrix of the impurity removal dynamic eigenvector
  • ⁇ and ⁇ are the global mean and variance of the impurity removal dynamic eigenvector respectively
  • exp( ⁇ ) represents Exponential operation using a vector as a power
  • the exponential operation using a vector as a power means using the value of each position of the vector as a power to find the exponent, and then filling the result into each position of the vector to obtain the vector operation result, and Represent position-wise subtraction and addition of vectors respectively
  • 2 represents the second norm of the feature vector.
  • the power values of the sprayers at multiple predetermined time points within the predetermined time period are passed through a timing encoder including a one-dimensional convolution layer to obtain the power control feature vector.
  • a timing encoder including a one-dimensional convolution layer including: arranging the power values of the sprayers at multiple predetermined time points within the predetermined time period into a one-dimensional input vector according to the time dimension; using the fully connected layer of the temporal encoder to perform the following formula on the input vector: Fully connected encoding is used to extract high-dimensional hidden features of the eigenvalues at 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, Among them, 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.
  • calculating the transfer matrix of the power control eigenvector relative to the corrected impurity removal dynamic eigenvector includes: calculating the power control eigenvector with the following formula The transfer matrix relative to the corrected impurity removal dynamic eigenvector; wherein, the formula is:
  • V 2 M*V 1
  • V 1 represents the power control eigenvector
  • M represents the transfer matrix
  • V 2 represents the corrected impurity removal dynamic eigenvector
  • the transfer matrix is passed through a classifier to obtain a classification result.
  • the classification result is used to indicate that the power value of the sprayer at the current time point should be increased or decreased.
  • the classifier processes the transfer matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • control system and control method of the smart production line of electronic grade potassium hydroxide adopts artificial intelligence control technology and is based on the deep neural network model to control the dynamic change characteristics of the power of the sprayer and the spray
  • the dynamic characteristics of impurity removal image frames are deeply explored to dynamically control the power of the sprayer in real-time in the smart production line of electronic grade potassium hydroxide, thereby ensuring the effect of crystal yield and the efficiency of impurity transfer.
  • Figure 1 is an application scenario diagram of a control system for a smart production line of electronic grade potassium hydroxide according to an embodiment of the present application.
  • FIG. 2 is a block diagram of a control system of a smart production line for electronic grade potassium hydroxide according to an embodiment of the present application.
  • Figure 3 is a flow chart of a control method for a smart production line of electronic grade potassium hydroxide according to an embodiment of the present application.
  • Figure 4 is a schematic structural diagram of a control method for a smart production line of electronic grade potassium hydroxide according to an embodiment of the present application.
  • Patent 201010576437.1 discloses a method of dissolving potassium metaborate into an aqueous sodium hydroxide solution, heating to reflux, and solid-liquid separation to obtain solid potassium hydroxide. This method has a complicated process flow. , high energy consumption, and does not involve the purity of potassium hydroxide.
  • Patent 201410290942.8 discloses a continuous production method of high-purity potassium hydroxide aqueous solution. In this method, the potassium hydroxide solution needs to pass through chelating resin, cation exchange resin, anion exchange resin, electrodialysis and multi-stage filter circulation system to remove impurities in sequence. This process The method steps are cumbersome and the process cost is high.
  • Patent CN202010747995.3 uses purified potassium chloride for electrolysis to produce electronic-grade potassium hydroxide. There are problems such as complex process routes and difficulty in handling acidic waste brine.
  • step S3 Centrifuge the material obtained in step S2;
  • the controlled parameters are as follows:
  • step S2 circulating water is used for cooling.
  • the temperature of the circulating water is 5°C ⁇ 25°C, and the preferred choice is 10°C ⁇ 20°C;
  • step S3 ultrapure water is used to atomize and spray potassium hydroxide while centrifuging; in step S3, the purpose of using ultrapure water to atomize and spray potassium hydroxide while centrifuging is Because the spray and crystals are in a state of hypergravity, it can speed up the mass transfer process and quickly take away impurities. At the same time, the yield loss of the crystals will also be reduced, ensuring good washing and yield effects.
  • step S3 the power control of the sprayer is the key. It should be understood that if the power is too high, the crystals will be partially transferred away, resulting in a decrease in the purification rate. Decrease, and if the power is too small, it will lead to low impurity transfer efficiency and low lifting efficiency. Therefore, the inventor of the present application expects to intelligently control the power of the sprayer in the smart production line of electronic grade potassium hydroxide to ensure the effect of crystal yield and the efficiency of impurity transfer.
  • the first convolutional neural network of the spatial attention mechanism processes each of the multiple key frames to obtain multiple spatially focused feature maps.
  • the difference between each two adjacent spatial focus feature maps among the plurality of spatial focus feature maps is calculated to obtain multiple differential feature maps.
  • the multiple differential feature maps are processed through a second convolutional neural network using a three-dimensional convolution kernel to extract the dynamic implications of the spray impurity removal in the multiple differential feature maps.
  • the global mean pooling along the channel dimension is further performed on the impurity removal dynamic feature map to obtain the decontamination dynamic feature map.
  • the spatial attention mechanism is used to focus on the local features in the spatial dimension of the feature matrix, and the global mean pooling along the channel dimension is further based on the spatial features of the feature matrix.
  • the global feature description is used, which weakens the correlation between the feature values at each position of the impurity-free dynamic feature vector, and the prediction pressure for the class probability occurs during the streaming propagation of the convolutional neural network with depth.
  • the impurity removal dynamic feature vector V is optimized and expressed as:
  • V represents the impurity-removal dynamic eigenvector
  • ⁇ is the autocovariance matrix of the impurity-removal dynamic eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of every two positions of the vector V
  • ⁇ and ⁇ are the global mean and variance of the impurity-removing dynamic feature vector respectively.
  • exp( ⁇ ) represents the exponential operation with the vector as the power, where the exponential operation with the vector as the power represents the value of each position of the vector as Find the exponent of the power, and then fill in the result into each position of the vector to get the vector operation result, and Represent position-wise subtraction and addition of vectors respectively
  • 2 represents the second norm of the feature vector.
  • the above optimization can guide the correction of the forward propagation correlation.
  • the characteristics of the forward propagation based on downsampling of the features based on the global mean pooling along the channel dimension through the learnable normal sampling offset Guide feature engineering of convolutional neural networks to effectively model long-range dependencies in the spatial dimension within the feature matrix and in the channel dimension between feature matrices of the impurity-free dynamic feature map, and consider the local and non-local aspects of the feature matrix
  • the neighborhood is used to repair the correlation between the eigenvalues of the impurity-free dynamic feature vector, thereby improving the class probability of the impurity-free dynamic feature vector during the streaming propagation process of the convolutional neural network with depth. predictive ability.
  • the temporal encoder of the layer encodes the power values of the atomizer at multiple predetermined time points within the predetermined time period to obtain a power control feature vector.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the power value of the sprayer in the temporal dimension through one-dimensional convolutional coding. and extract high-dimensional latent features of the power value of the sprayer through fully connected encoding.
  • the ratio of the power control eigenvector relative to the corrected impurity removal dynamic eigenvector is further calculated. transfer matrix. Furthermore, a classifier is used to perform classification processing on the transfer matrix to obtain a classification result indicating that the power value of the sprayer at the current time point should be increased or decreased.
  • this application proposes a control system for a smart production line of electronic grade potassium hydroxide, which includes: a spray data acquisition module, used to obtain the power values of the sprayer at multiple predetermined time points within a predetermined time period and the power values generated by the camera.
  • the collected monitoring video of spray impurities in the predetermined time period ; a spatial encoding module for extracting multiple key frames from the monitoring video, and separately converting each of the multiple key frames by using spatial attention.
  • the first convolutional neural network of the mechanism is used to obtain multiple spatially focused feature maps; a difference module is used to calculate the difference between each two adjacent spatially focused feature maps in the multiple spatially focused feature maps to obtain multiple differences.
  • Feature map Feature map
  • a dynamic encoding module used to pass the plurality of differential feature maps through a second convolutional neural network using a three-dimensional convolution kernel to obtain an impurity-free dynamic feature map
  • a dimensionality reduction module used to perform the impurity-free dynamic feature map
  • the feature map performs global mean pooling along the channel dimension to obtain the impurity-removal dynamic feature vector
  • the correction module is used to calculate the impurity-removal dynamic feature vector at each position based on the autocovariance matrix of the impurity-removal dynamic feature vector.
  • the eigenvalues are corrected to obtain a corrected impurity-removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity-removal dynamic eigenvector are the eigenvalues of the corresponding two positions in the impurity-removal dynamic eigenvector.
  • FIG. 1 illustrates an application scenario diagram of a control system for a smart production line of electronic grade potassium hydroxide according to an embodiment of the present application.
  • a sensor for example, the power measuring instrument T as shown in Figure 1
  • the sprayer for example, as shown in Figure 1
  • the power value of P is shown and the monitoring video of the spray impurity removal in the predetermined period of time is collected by the camera (for example, C as shown in FIG. 1 ).
  • the obtained power values of the sprayer at multiple predetermined time points within the predetermined time period and the monitoring video of the spray impurity removal during the predetermined time period are input to the control algorithm of the smart production line where electronic grade potassium hydroxide is deployed.
  • a server for example, the cloud server S as shown in Figure 1
  • the server can use the control algorithm of the smart production line of electronic grade potassium hydroxide to control multiple predetermined time points within the predetermined time period.
  • the power value of the atomizer and the monitoring video of the spray impurity removal in the predetermined time period are processed to generate a classification result indicating that the power value of the atomizer at the current time point should be increased or decreased.
  • FIG. 2 illustrates a block diagram of a control system of a smart production line for electronic grade potassium hydroxide according to an embodiment of the present application.
  • the control system 200 of a smart production line for electronic grade potassium hydroxide according to an embodiment of the present application includes: a spray data acquisition module 210, used to obtain the power of the sprayer at multiple predetermined time points within a predetermined time period.
  • the spatial encoding module 220 is used to extract multiple key frames from the surveillance video, and convert each of the multiple key frames into A plurality of spatially focused feature maps are obtained by using the first convolutional neural network of the spatial attention mechanism; the difference module 230 is used to calculate the difference between each two adjacent spatially focused feature maps in the plurality of spatially focused feature maps.
  • the dynamic encoding module 240 is used to pass the plurality of differential feature maps through a second convolutional neural network using a three-dimensional convolution kernel to obtain a de-impurity dynamic feature map;
  • the dimensionality reduction module 250 is used to perform global mean pooling along the channel dimension on the impurity removal dynamic feature map to obtain the impurity removal dynamic feature vector;
  • the correction module 260 is used to perform the impurity removal dynamic feature vector based on the autocovariance matrix of the impurity removal dynamic feature vector.
  • the eigenvalues of each position in the impurity removal dynamic eigenvector are corrected to obtain the corrected impurity removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity removal dynamic eigenvector are the impurity removal dynamic eigenvectors.
  • the spray data acquisition module 210 and the spatial encoding module 220 are used to obtain the power values of the sprayer at multiple predetermined time points within a predetermined time period and the predetermined values collected by the camera.
  • the surveillance video of the time period is sprayed to remove impurities, and multiple key frames are extracted from the surveillance video, and each of the multiple key frames is passed through the first convolutional neural network using the spatial attention mechanism to obtain Multiple spatially focused feature maps.
  • the power control of the sprayer is the key, it should be understood that if the power is too high, the crystals will be partially transferred away, resulting in a decrease in the purification rate.
  • the power of the sprayer is intelligently controlled in the smart production line of electronic grade potassium hydroxide, the effect of spraying impurities must be monitored dynamically in real time. Therefore, first, the power value of the sprayer at multiple predetermined time points within a predetermined time period is obtained through a power measuring instrument, and the monitoring video of the spray impurity removal during the predetermined time period is collected by a camera. Then, in order to facilitate subsequent feature mining and reduce the amount of calculation, multiple key frames are extracted from the surveillance video, and then the multiple key frames are processed by using convolutional neural networks that have excellent performance in implicit correlation feature extraction.
  • the network model performs feature mining on it.
  • the first convolutional neural network of the spatial attention mechanism processes each of the multiple key frames to obtain multiple spatially focused feature maps.
  • the spatial coding module is further configured to perform: each layer of the first convolutional neural network on the input data in the forward transfer of the layer: Perform convolution processing based on a two-dimensional convolution kernel to generate a convolution feature map; perform pooling processing on the convolution feature map to generate a pooled feature map; perform activation processing on the pooled feature map to generate an activation feature Figure; perform global average pooling along the channel dimension on the activation feature map to obtain a spatial feature matrix; perform convolution processing and activation processing on the spatial feature matrix to generate a weight vector; and, with each of the weight vectors The weight value of the position weights each feature matrix of the activation feature map to obtain a generated feature map; wherein, the generated feature map output by the last layer of the first convolutional neural network is the multiple spaces Focusing on the feature map, the input of the first layer of the first convolutional neural network is each key frame in the plurality of key frames.
  • the difference module 230 and the dynamic encoding module 240 are used to calculate the difference between each two adjacent spatial focus feature maps in the plurality of spatial focus feature maps to obtain A plurality of differential feature maps are passed through a second convolutional neural network using a three-dimensional convolution kernel to obtain an impurity-free dynamic feature map. That is to say, in the technical solution of the present application, further, in order to express the dynamic effect of the spray impurity removal, it is necessary to pay more attention to the comparison of the focus feature maps of each space, especially the two adjacent spaces. Dynamic comparison of focused feature maps.
  • the difference between each adjacent two spatial focus feature maps among the plurality of spatial focus feature maps is further calculated to obtain multiple differential feature maps.
  • the multiple differential feature maps are processed through a second convolutional neural network using a three-dimensional convolution kernel to extract the dynamic implications of the spray impurity removal in the multiple differential feature maps.
  • the second convolutional neural network using a three-dimensional convolution kernel performs a three-dimensional convolution on the input data in the forward pass of the layer: based on the three-dimensional convolution kernel, the input data is three-dimensionally convoluted.
  • Convolution processing is performed to obtain a convolution feature map; mean pooling is performed on the convolution feature map to obtain a pooling feature map; and nonlinear activation is performed on the pooling feature map to obtain an activation feature map; where,
  • the output of the last layer of the second convolutional neural network is the impurity-free dynamic feature map, and the input of the first layer of the second convolutional neural network is the plurality of differential feature maps.
  • the difference module is further configured to: calculate the difference between each adjacent two spatial focus feature maps in the plurality of spatial focus feature maps using the following formula to obtain the Multiple differential feature maps; wherein, the formula is:
  • F i represents the spatial focus feature map at the i-th position in the plurality of spatial focus feature maps
  • F i+1 represents the spatial focus feature map at the i+1th position in the plurality of spatial focus feature maps
  • F represents the differential feature map
  • the dimensionality reduction module 250 is used to perform global mean pooling along the channel dimension on the impurity-removing dynamic feature map to obtain an impurity-removing dynamic feature vector. It should be understood that considering that the number of parameters in the impurity removal dynamic feature map is large, which will make subsequent processing more complicated, therefore, in the technical solution of the present application, the impurity removal dynamic feature map is further processed along the Global mean pooling in the channel dimension is used to obtain impurity-free dynamic feature vectors to reduce the number of parameters, prevent over-fitting, and thereby improve the accuracy of subsequent classification.
  • the correction module 260 is used to correct the eigenvalues of each position in the impurity removal dynamic eigenvector based on the autocovariance matrix of the impurity removal dynamic eigenvector to obtain Corrected impurity-removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity-removal dynamic eigenvector are the variances between the eigenvalues of corresponding two positions in the impurity-removal dynamic eigenvector.
  • the spatial attention mechanism is used to focus on the local features in the spatial dimension of the feature matrix, and the global mean pooling along the channel dimension is further based on
  • the spatial features of the feature matrix are globally characterized, which weakens the correlation between the feature values at each position of the impurity-free dynamic feature vector, which occurs during the streaming process of the convolutional neural network with depth. Probabilistic predictive pressure. Therefore, in the technical solution of the present application, the impurity removal dynamic feature vector V needs to be optimized.
  • the correction module is further configured to: based on the autocovariance matrix of the impurity-removing dynamic eigenvector, use the following formula to calculate the eigenvalues of each position in the impurity-removing dynamic eigenvector. Calibration is performed to obtain the corrected impurity removal dynamic feature vector; wherein, the formula is:
  • V represents the impurity-removal dynamic eigenvector
  • ⁇ is the autocovariance matrix of the impurity-removal dynamic eigenvector, that is, the value of each position of the matrix is the variance between the eigenvalues of every two positions of the vector V
  • ⁇ and ⁇ are the global mean and variance of the impurity-removing dynamic feature vector respectively.
  • exp( ⁇ ) represents the exponential operation with the vector as the power, where the exponential operation with the vector as the power represents the value of each position of the vector as Find the exponent of the power, and then fill in the result into each position of the vector to get the vector operation result, and Represent position-wise subtraction and addition of vectors respectively
  • 2 represents the second norm of the feature vector.
  • the optimization can guide the correction of the forward propagation correlation.
  • the learned normal sampling offset guides the feature engineering of convolutional neural networks to effectively model the long-range dependencies within the spatial dimension of the feature matrix and the channel dimension between the feature matrices of the decontamination dynamic feature map, and consider The local and non-local neighborhoods of the feature matrix are used to repair the correlation between the eigenvalues of the impurity-removing dynamic eigenvector, thereby improving the flow of the impurity-removing dynamic eigenvector with depth in the convolutional neural network.
  • the power timing encoding module 270 is used to pass the power values of the sprayer at multiple predetermined time points within the predetermined time period through a timing encoder including a one-dimensional convolution layer to obtain Power control feature vector.
  • a timing encoder including a one-dimensional convolutional layer is used to encode the power values of the sprayer at multiple predetermined time points within the predetermined time period to obtain a power control feature vector.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the power value of the sprayer in the temporal dimension through one-dimensional convolutional coding. and extract high-dimensional latent features of the power value of the sprayer through fully connected encoding.
  • the power timing encoding module is further configured to: arrange the power values of the sprayers at multiple predetermined time points within the predetermined time period into a one-dimensional input vector according to the time dimension; Use the fully connected layer of the temporal encoder to perform fully connected encoding on the input vector with the following formula to extract the high-dimensional hidden features of the eigenvalues 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, Among them, 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 responsiveness estimation module 280 and the control result generation module 290 are used to calculate the transfer matrix of the power control feature vector relative to the corrected impurity removal dynamic feature vector,
  • the transfer matrix is passed through a classifier to obtain a classification result, which is used to indicate that the power value of the sprayer at the current time point should be increased or decreased.
  • the power control feature vector is further calculated relative to The transfer matrix of the corrected impurity-removing dynamic eigenvector. Furthermore, a classifier is used to perform classification processing on the transfer matrix to obtain a classification result indicating that the power value of the sprayer at the current time point should be increased or decreased.
  • the classifier processes the transfer matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the bias matrix of the connection layer is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the bias matrix of the connection layer is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • V 1 represents the power control eigenvector
  • M represents the transfer matrix
  • V 2 represents the corrected impurity removal dynamic eigenvector
  • control system 200 of the smart production line of electronic grade potassium hydroxide based on the embodiment of the present application is clarified, which uses artificial intelligence control technology and is based on a deep neural network model to control the dynamic change characteristics of the power of the sprayer and the spray removal.
  • the dynamic characteristics of the impurity image frames are deeply mined to dynamically control the power of the sprayer in real-time and dynamically in the smart production line of electronic grade potassium hydroxide, thereby ensuring the effect of crystal yield and the efficiency of impurity transfer.
  • control system 200 of the intelligent production line of electronic grade potassium hydroxide according to the embodiment of the present application can be implemented in various terminal devices, such as the server of the control algorithm of the intelligent production line of electronic grade potassium hydroxide, etc.
  • the control system 200 of the intelligent production line of electronic grade potassium hydroxide according to the embodiment of the present application can be integrated into the terminal device as a software module and/or a hardware module.
  • the control system 200 of the smart production line of electronic grade potassium hydroxide can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the electronic grade
  • the control system 200 of the smart production line of potassium hydroxide can also be one of the many hardware modules of the terminal equipment.
  • control system 200 of the smart production line of electronic grade potassium hydroxide and the terminal device can also be separate devices, and the control system 200 of the smart production line of electronic grade potassium hydroxide can Connect to the terminal device through a wired and/or wireless network, and transmit interactive information according to the agreed data format.
  • Figure 3 illustrates a flow chart of a control method for a smart production line of electronic grade potassium hydroxide.
  • the control method of a smart production line for electronic grade potassium hydroxide includes steps: S110, obtaining the power values of the sprayers at multiple predetermined time points within a predetermined time period and the power values collected by the camera. Monitoring video of spray impurities in the predetermined period; S120, extract multiple key frames from the monitoring video, and pass each of the multiple key frames through the first convolution using the spatial attention mechanism.
  • Neural network to obtain multiple spatially focused feature maps S130, calculate the difference between each adjacent two spatially focused feature maps in the multiple spatially focused feature maps to obtain multiple differential feature maps; S140, convert the multiple spatially focused feature maps The differential feature maps are passed through the second convolutional neural network using a three-dimensional convolution kernel to obtain the impurity-removal dynamic feature map; S150, perform global mean pooling along the channel dimension on the impurity-removal dynamic feature map to obtain the impurity-removal dynamic feature map Vector; S160, based on the autocovariance matrix of the impurity removal dynamic eigenvector, correct the eigenvalues of each position in the impurity removal dynamic eigenvector to obtain a corrected impurity removal dynamic eigenvector, wherein, the impurity removal dynamic eigenvector The eigenvalues of each position in the autocovariance matrix of the dynamic eigenvector are the variances between the eigenvalues of the corresponding two positions in the impurity-free dynamic
  • the power value of the sprayer is passed through a temporal encoder including a one-dimensional convolutional layer to obtain a power control feature vector; S180, calculate the transfer matrix of the power control feature vector relative to the corrected impurity removal dynamic feature vector; and, S190, The transfer matrix is passed through a classifier to obtain a classification result, which is used to indicate that the power value of the nebulizer at the current time point should be increased or decreased.
  • Figure 4 illustrates a schematic architectural diagram of a control method for a smart production line of electronic grade potassium hydroxide according to an embodiment of the present application.
  • a first convolutional neural network using a spatial attention mechanism (for example, as shown in Figure 4 CNN1) to obtain multiple spatially focused feature maps (for example, F1 as shown in Figure 4); then, calculate the difference between each adjacent two spatially focused feature maps in the multiple spatially focused feature maps to A plurality of differential feature maps (for example, F2 as shown in Figure 4) are obtained; then, the multiple differential feature maps are passed through a second convolutional neural network using a three-dimensional convolution kernel (for example, as shown in Figure 4 Schematic CNN2) to obtain the impurity-removal dynamic feature map (for example, F3 as shown in Figure 4); then, perform global mean
  • Impurity dynamic feature vector for example, VF2 as shown in Figure 4
  • the power values of the atomizer at multiple predetermined time points within the predetermined time period for example, Q as shown in Figure 4
  • the temporal encoder of the one-dimensional convolution layer for example, E as shown in Figure 4
  • the power control feature vector for example, VF as shown in Figure 4
  • the transfer matrix of the impurity dynamic feature vector is removed (for example, the MF as illustrated in Figure 4); and, finally, the transfer matrix is passed through a classifier (for example, the classifier as illustrated in Figure 4 ) to obtain a classification result, which is used to indicate that the power value of the nebulizer at the current point in time should be increased or decreased.
  • a classifier for example, the classifier as illustrated in Figure 4
  • steps S110 and S120 the power values of the sprayer at multiple predetermined time points within a predetermined time period and the monitoring video of spraying impurities in the predetermined time period collected by the camera are obtained, and the monitoring video is obtained from the monitoring
  • Multiple key frames are extracted from the video, and each of the multiple key frames is passed through a first convolutional neural network using a spatial attention mechanism to obtain multiple spatially focused feature maps.
  • the power control of the sprayer is the key, it should be understood that if the power is too high, the crystals will also be partially transferred away, resulting in a decrease in the purification rate, and if the power is too much, If it is small, it will lead to low impurity transfer efficiency and low lifting efficiency. Therefore, in the technical solution of this application, it is expected to intelligently control the power of the sprayer in the smart production line of electronic grade potassium hydroxide to ensure the effect of crystal yield and the efficiency of impurity transfer.
  • the power of the sprayer is intelligently controlled in the smart production line of electronic grade potassium hydroxide, the effect of spraying impurities must be monitored dynamically in real time. Therefore, first, the power value of the sprayer at multiple predetermined time points within a predetermined time period is obtained through a power measuring instrument, and the monitoring video of the spray impurity removal during the predetermined time period is collected by a camera. Then, in order to facilitate subsequent feature mining and reduce the amount of calculation, multiple key frames are extracted from the surveillance video, and then the multiple key frames are processed by using convolutional neural networks that have excellent performance in implicit correlation feature extraction.
  • the network model performs feature mining on it.
  • the first convolutional neural network of the spatial attention mechanism processes each of the multiple key frames to obtain multiple spatially focused feature maps.
  • the difference between each adjacent two spatially focused feature maps in the plurality of spatially focused feature maps is calculated to obtain a plurality of differential feature maps, and the multiple differential feature maps are The feature map is passed through the second convolutional neural network using a three-dimensional convolution kernel to obtain the impurity-free dynamic feature map. That is to say, in the technical solution of the present application, further, in order to express the dynamic effect of the spray impurity removal, it is necessary to pay more attention to the comparison of the focus feature maps of each space, especially the two adjacent spaces. Dynamic comparison of focused feature maps. Therefore, in the technical solution of the present application, the difference between each adjacent two spatial focus feature maps among the plurality of spatial focus feature maps is further calculated to obtain multiple differential feature maps.
  • the multiple differential feature maps are processed through a second convolutional neural network using a three-dimensional convolution kernel to extract the dynamic implications of the spray impurity removal in the multiple differential feature maps.
  • the second convolutional neural network using a three-dimensional convolution kernel performs a three-dimensional convolution on the input data in the forward pass of the layer: based on the three-dimensional convolution kernel, the input data is three-dimensionally convoluted.
  • Convolution processing is performed to obtain a convolution feature map; mean pooling is performed on the convolution feature map to obtain a pooling feature map; and nonlinear activation is performed on the pooling feature map to obtain an activation feature map; where,
  • the output of the last layer of the second convolutional neural network is the impurity-free dynamic feature map, and the input of the first layer of the second convolutional neural network is the plurality of differential feature maps.
  • steps S150 and S160 global mean pooling along the channel dimension is performed on the impurity removal dynamic feature map to obtain an impurity removal dynamic feature vector, and based on the autocovariance of the impurity removal dynamic feature vector matrix, correct the eigenvalues of each position in the impurity removal dynamic eigenvector to obtain the corrected impurity removal dynamic eigenvector, wherein the eigenvalues of each position in the autocovariance matrix of the impurity removal dynamic eigenvector are The variance between the eigenvalues of the corresponding two positions in the impurity removal dynamic eigenvector.
  • the impurity removal dynamic feature map is further processed along the Global mean pooling in the channel dimension is used to obtain impurity-free dynamic feature vectors to reduce the number of parameters, prevent over-fitting, and thereby improve the accuracy of subsequent classification.
  • the spatial attention mechanism is used to focus on local features in the spatial dimension of the feature matrix, and the global mean pooling along the channel dimension is further based on the feature matrix
  • the spatial features are globally characterized, which weakens the correlation between the feature values of each position of the impurity-free dynamic feature vector, which occurs during the streaming propagation of the convolutional neural network with depth for the class probability forecast pressure. Therefore, in the technical solution of the present application, the impurity removal dynamic feature vector V needs to be optimized.
  • step S170 the power values of the atomizer at multiple predetermined time points within the predetermined time period are passed through a temporal encoder including a one-dimensional convolution layer to obtain a power control feature vector.
  • a temporal encoder including a one-dimensional convolutional layer is used to encode the power values of the sprayer at multiple predetermined time points within the predetermined time period to obtain a power control feature vector.
  • the temporal encoder consists of alternately arranged fully connected layers and one-dimensional convolutional layers, which extract the correlation of the power value of the sprayer in the temporal dimension through one-dimensional convolutional coding. and extract high-dimensional latent features of the power value of the sprayer through fully connected encoding.
  • a transfer matrix of the power control feature vector relative to the corrected impurity removal dynamic feature vector is calculated, and the transfer matrix is passed through a classifier to obtain a classification result, the The classification result is used to indicate that the power value of the nebulizer at the current point in time should be increased or decreased.
  • the power control feature vector is further calculated relative to The transfer matrix of the corrected impurity-removing dynamic eigenvector. Furthermore, a classifier is used to perform classification processing on the transfer matrix to obtain a classification result indicating that the power value of the sprayer at the current time point should be increased or decreased.
  • control method of the smart production line of electronic grade potassium hydroxide based on the embodiment of the present application is clarified, which uses artificial intelligence control technology and is based on a deep neural network model to control the dynamic change characteristics of the power of the sprayer and the spray removal of impurities.
  • the dynamic characteristics of image frames are deeply mined to dynamically control the power of the sprayer in real-time and dynamically in the smart production line of electronic grade potassium hydroxide, thereby ensuring the effect of crystal yield and the efficiency of impurity transfer.
  • 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

  • Chemical & Material Sciences (AREA)
  • Engineering & Computer Science (AREA)
  • Organic Chemistry (AREA)
  • Materials Engineering (AREA)
  • Inorganic Chemistry (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Automation & Control Theory (AREA)
  • Image Analysis (AREA)

Abstract

公开了一种电子级氢氧化钾的智慧产线的控制系统及其控制方法，涉及智能制造领域下的智能控制。该控制系统采用人工智能控制技术，基于深度神经网络模型来对于喷雾器的功率动态变化特征和喷雾去杂质图像帧的动态特征进行深层挖掘，以在电子级氢氧化钾的智慧产线中对喷雾器的功率进行实时动态地控制，进而保证晶体收率的效果和杂质传递的效率。

Description

电子级氢氧化钾的智慧产线的控制系统及其控制方法 技术领域
本发明涉及智能制造领域下的智能控制，且更为具体地，涉及一种电子级氢氧化钾的智慧产线的控制系统及其控制方法。
背景技术
随着全球高新电子线路板的发展，全球对高纯电子级氢氧化钾的需求不断上升。工业级氢氧化钾已经满足不了要求。因此，有必要对工业级氢氧化钾进行纯化，提升品质。
现有诸多用于氢氧化钾的纯化方案，例如，专利201010576437.1公开了一种将偏硼酸钾溶解到氢氧化钠水溶液中，通过加热回流，固液分离得到固体氢氧化钾，该法工艺流程复杂，能耗大，未涉及到氢氧化钾纯度问题。专利201410290942.8公开了一种高纯氢氧化钾水溶液的连续生产方法，该方法中氢氧化钾溶液需依次通过螯合树脂、阳离子交换树脂、阴离子交换树脂、电渗析及多级滤芯循环系统除去杂质，该工艺方法步骤烦琐且工艺成本高。专利CN202010747995.3中采用纯化氯化钾进行电解制得电子级氢氧化钾工艺，存在工艺路线复杂，酸性废盐水不好处理等问题。
因此，期待一种优化的电子级氢氧化钾的制备方案。
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电子级氢氧化钾的智慧产线的控制系统及其控制方法，其采用人工智能控制技术，基于深度神经网络模型来对于喷雾器的功率动态变化特征和喷雾去杂质图像帧的动态特征进行深层挖掘，以在电子级氢氧化钾的智慧产线中对喷雾器的功率进行实时动态地控制，进而保证晶体收率的效果和杂质传递的效率。
根据本申请的一个方面，提供了一种电子级氢氧化钾的智慧产线的控制系统，其包括：喷雾数据采集模块，用于获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；空间编码模块，用于从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；差分模块，用于计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；动态编码模块，用于将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；降维模块，用于对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；校正模块，用于基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；功率时序编码模块，用于将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；响应性估计模块，用于计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及控制结果生成模块，用于将所述转移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喷雾器的功率值应增大或应减小。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述空间编码模块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对所述输入数据进行基于二维卷积核的卷积处理以生成卷积特征图；对所述卷积特征图进行池化处理以生成池化特征图；对所述池化特征图进行激活处理以生成激活特征图；对所述激活特征图进行沿通道维度的全局平均池化以获得空间特征矩阵；对所述空间特征矩阵进行卷积处理和激活处理以生成权重向量；以及，以所述权重向量中各个位置的权重值分别对所述激活特征图的各个特征矩阵进行加权以获得生成特征图；其中，所述第一卷积神经网络的最后一层输出的所述生成特征图为所述多个空间聚焦特征图，所述第一卷积神经网络的第一层的输入为所述多个关键帧中各个关键帧。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述差分模块，进一步用于：以如下公式计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到所述多个差分特征图；其中，所述公式为：
Figure PCTCN2022119555-appb-000001
其中，F i表示所述多个空间聚焦特征图中第i个位置的空间聚焦特征图，F i+1表示所述多个空间聚焦特征图中第i+1个位置的空间聚焦特征图，F表示所述差分特征图，
Figure PCTCN2022119555-appb-000002
表示按位置差分。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述动态编码模块，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去杂质动态特征图，所述第二卷积神经网络的第一层的输入为所述多个差分特征 图。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述校正模块，进一步用于：基于所述去杂质动态特征向量的自协方差矩阵，以如下公式对所述去杂质动态特征向量中各个位置的特征值进行校正以得到所述校正后去杂质动态特征向量；其中，所述公式为：
Figure PCTCN2022119555-appb-000003
其中V表示所述去杂质动态特征向量，∑是所述去杂质动态特征向量的自协方差矩阵，μ和σ分别是所述去杂质动态特征向量的全局均值和方差，exp(·)表示以向量为幂的指数运算，其中，以向量为幂的指数运算表示以向量的每个位置的值作为幂求指数，再将结果填入向量的各个位置以得到向量运算结果，
Figure PCTCN2022119555-appb-000004
和
Figure PCTCN2022119555-appb-000005
分别表示向量的按位置减法和加法，||·|| 2表示特征向量的二范数。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述功率时序编码模块，进一步用于：将所述预定时间段内多个预定时间点的喷雾器的功率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 PCTCN2022119555-appb-000006
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 PCTCN2022119555-appb-000007
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 PCTCN2022119555-appb-000008
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。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述响应性估计模块，进一步用于：以如下公式计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的所述转移矩阵；其中，所述公式为：
V 2＝M*V 1
其中V 1表示所述功率控制特征向量，M表示所述转移矩阵，V 2表示所述校正后去杂质动态特征向量。
在上述电子级氢氧化钾的智慧产线的控制系统中，所述控制结果生成模块，进一步用于：所述分类器以如下公式对所述转移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)表示将所述转移矩阵投影为向量，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，一种电子级氢氧化钾的智慧产线的控制方法，其包括：获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及将所述转移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喷雾器的功率值应增大或应减小。
在上述电子级氢氧化钾的智慧产线的控制方法中，从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图，包括：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对所述输入数据进行基于二维卷积核的卷积处理以生成卷积特征图；对所述卷积特征图进行池化处理以生成池化特征图；对所述池化特征图进行激活处理以生成激活特征图；对所述激活特征图进行沿通道维度的全局平均池化以获得空间特征矩阵；对所述空间特征矩阵进行卷积处理和激活处理以生成权重向量；以及，以所述权重向量中各个位置的权重值分别对所述激活特征图的各个特征矩阵进行加权以获得生成特征图；其中，所述第一卷积神经网络的最后一层输出的所述生成特征图为所述多个空间聚焦特征图，所述第 一卷积神经网络的第一层的输入为所述多个关键帧中各个关键帧。
在上述电子级氢氧化钾的智慧产线的控制方法中，计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图，包括：以如下公式计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到所述多个差分特征图；其中，所述公式为：
Figure PCTCN2022119555-appb-000009
其中，F i表示所述多个空间聚焦特征图中第i个位置的空间聚焦特征图，F i+1表示所述多个空间聚焦特征图中第i+1个位置的空间聚焦特征图，F表示所述差分特征图，
Figure PCTCN2022119555-appb-000010
表示按位置差分。
在上述电子级氢氧化钾的智慧产线的控制方法中，将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图，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去杂质动态特征图，所述第二卷积神经网络的第一层的输入为所述多个差分特征图。
在上述电子级氢氧化钾的智慧产线的控制方法中，基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，包括：基于所述去杂质动态特征向量的自协方差矩阵，以如下公式对所述去杂质动态特征向量中各个位置的特征值进行校正以得到所述校正后去杂质动态特征向量；其中，所述公式为：
Figure PCTCN2022119555-appb-000011
其中V表示所述去杂质动态特征向量，∑是所述去杂质动态特征向量的自协方差矩阵，μ和σ分别是所述去杂质动态特征向量的全局均值和方差，exp(·)表示以向量为幂的指数运算，其中，以向量为幂的指数运算表示以向量的每个位置的值作为幂求指数，再将结果填入向量的各个位置以得到向量运算结果，
Figure PCTCN2022119555-appb-000012
和
Figure PCTCN2022119555-appb-000013
分别表示向量的按位置减法和加法，||·|| 2表示特征向量的二范数。
在上述电子级氢氧化钾的智慧产线的控制方法中，将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量，包括：将所述预定时间段内多个预定时间点的喷雾器的功率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 PCTCN2022119555-appb-000014
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 PCTCN2022119555-appb-000015
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 PCTCN2022119555-appb-000016
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。
在上述电子级氢氧化钾的智慧产线的控制方法中，计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵，包括：以如下公式计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的所述转移矩阵；其中，所述公式为：
V 2＝M*V 1
其中V 1表示所述功率控制特征向量，M表示所述转移矩阵，V 2表示所述校正后去杂质动态特征向量。
在上述电子级氢氧化钾的智慧产线的控制方法中，将所述转移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喷雾器的功率值应增大或应减小，包括：所述分类器以如下公式对所述转移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)表示将所述转移矩阵投影为向量，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，本申请提供的电子级氢氧化钾的智慧产线的控制系统及其控制方法，其采用人工智能控制技术，基于深度神经网络模型来对于喷雾器的功率动态变化特征和喷雾去杂质图像帧的动态特征进行深层挖掘，以在电子级氢氧化钾的智慧产线中对喷雾器的功率进行实时动态地控制，进而保证晶体收率的效果和杂质传递的效率。
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电子级氢氧化钾的智慧产线的控制系统的应用场景图。
图2为根据本申请实施例的电子级氢氧化钾的智慧产线的控制系统的框图。
图3为根据本申请实施例的电子级氢氧化钾的智慧产线的控制方法的流程图。
图4为根据本申请实施例的电子级氢氧化钾的智慧产线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，随着全球高新电子线路板的发展，全球对高纯电子级氢氧化钾的需求不断上升。工业级氢氧化钾已经满足不了要求。因此，有必要对工业级氢氧化钾进行纯化，提升品质。
现有诸多用于氢氧化钾的纯化方案，例如，专利201010576437.1公开了一种将偏硼酸钾溶解到氢氧化钠水溶液中，通过加热回流，固液分离得到固体氢氧化钾，该法工艺流程复杂，能耗大，未涉及到氢氧化钾纯度问题。专利201410290942.8公开了一种高纯氢氧化钾水溶液的连续生产方法，该方法中氢氧化钾溶液需依次通过螯合树脂、阳离子交换树脂、阴离子交换树脂、电渗析及多级滤芯循环系统除去杂质，该工艺方法步骤烦琐且工艺成本高。专利CN202010747995.3中采用纯化氯化钾进行电解制得电子级氢氧化钾工艺，存在工艺路线复杂，酸性废盐水不好处理等问题。
因此，期待一种优化的电子级氢氧化钾的制备方案。
在专利CN 113860336A中，其技术原理和方案如下：
S1：取工业级氢氧化钾溶液和氢氧化钾片碱，搅拌溶解，配制为50％～75％的氢氧化钾溶液；
S2：将氢氧化钾溶液降温至30℃～45℃，得到氢氧化钾晶体；
S3：对步骤S2中得到的物料进行离心分离；
S4：离心后的氢氧化钾晶体溶于纯水，配制为不同浓度氢氧化钾溶液，通过精密过滤器过滤，得到电子级氢氧化钾溶液。
其中，控制的参数如下：
1.结晶温度；
2.步骤S2中降温过程采用循环水降温，循环水的温度为5℃～25℃，较优选择为10℃～20℃；
3.步骤S3所述的离心分离中，在离心的同时采用超纯水对氢氧化钾进行雾化喷淋；步骤S3中在离心的同时采用超纯水对氢氧化钾雾化喷淋的目的是喷雾和晶体处于超重力状态能够加快传质过程，快速带走杂质，同时对于晶体的收率损失也会减少，能够保证好的洗涤效果和收率效果。
相应地，本申请发明人发现在该制备方案中，在步骤S3中，喷淋的喷雾器的功率控制是关键，应可以理解，如果功率过大会导致晶体也被部分地传递走，造成提纯率的下降，而如果功率多小则会导致杂质传递效率低，提升效率低下。因此，本申请发明人期待在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制以保证晶体收率的效果和杂质传递的效率。
具体地，在本申请的技术方案中，考虑到在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制时，还要实时动态地对于喷雾去杂质的效果进行监测，因此，首先，通过功率测量仪获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集所述预定时间段的喷雾去杂质的监控视频。然后，为了便于后续的特征挖掘以及降低计算量，从所述监控视频中提取出多个关键帧，再将所述多个关键帧通过使用在隐含关联特征提取方面具有优异表现的卷积神经网络模型对其进行特征挖掘。应可以理解，在实际的所述喷雾去杂质的过程中，需要更加关注于所述各个关键帧中的晶体和杂质在空间上的关联关系，并更加聚焦于所述杂质的传递，因此，使用空间注意力机制的第一卷积神经网络对所述多个关键帧中各个关键帧进行处理，以得到多个空间聚焦特征图。
进一步地，为了表达出所述喷雾去杂的动态效果，需要更加关注于所述各个空间聚焦特征图的对比，尤其是相邻的两个所述空间聚焦特征图的动态性比较，因此，在本申请的技术方案中，计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图。然后，将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述多个差分特征图中的关于所述喷雾去杂质的动态性隐含特征，从而得到去杂质动态特征图。
这样，考虑到所述去杂质动态特征图中的参数数量较大，这会使得后续的处理较为复杂，因此，进一步对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量，以降低参数的数量，防止过拟合，进而提高后续分类的准确性。
应可以理解，在获得去杂质动态特征向量的过程当中，使用空间注意力机制聚焦了特征矩阵的 空间维度上的局部特征，且沿通道维度的全局均值池化又进一步基于特征矩阵的空间特征进行了全局性特征描述，这使得所述去杂质动态特征向量的每个位置的特征值之间的相关性减弱，在卷积神经网络随深度的流式传播过程中发生对于类概率的预测压力。
因此，对所述去杂质动态特征向量V进行优化，表示为：
Figure PCTCN2022119555-appb-000017
其中V表示所述去杂质动态特征向量，∑是所述去杂质动态特征向量的自协方差矩阵，即矩阵的每个位置的值是向量V的每两个位置的特征值之间的方差，μ和σ分别是所述去杂质动态特征向量的全局均值和方差，exp(·)表示以向量为幂的指数运算，其中，以向量为幂的指数运算表示以向量的每个位置的值作为幂求指数，再将结果填入向量的各个位置以得到向量运算结果，
Figure PCTCN2022119555-appb-000018
和
Figure PCTCN2022119555-appb-000019
分别表示向量的按位置减法和加法，||·|| 2表示特征向量的二范数。
这样，上述优化可以对前向传播相关性进行引导修正，具体地，基于沿通道维度的全局均值池化对于特征进行的基于下采样的前向传播的特点，通过可学习的正态采样偏移引导卷积神经网络的特征工程来有效地建模所述去杂质动态特征图的特征矩阵内的空间维度和特征矩阵之间的通道维度上的长程依赖关系，并考虑特征矩阵的局部和非局部邻域来进行所述去杂质动态特征向量的各特征值间的相关性的修复，从而提高了所述所述去杂质动态特征向量在卷积神经网络随深度的流式传播过程中对于类概率的预测能力。
对于所述喷雾器的功率值，考虑到其在时序维度上具有着隐含的动态性的关联特征信息，因此，为了能够充分地提取出这种隐含的动态关联特征，使用包含一维卷积层的时序编码器对所述预定时间段内多个预定时间点的喷雾器的功率值进行编码以得到功率控制特征向量。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喷雾器的功率值在时序维度上的关联和通过全连接编码提取所述喷雾器的功率值的高维隐含特征。
然后，考虑到由于所述喷雾器的功率数据和所述喷雾去杂质的监控视频中的各个关键帧数据的特征尺度不同，并且所述喷雾去杂质的动态图像特征在高维空间中可以看作是针对所述喷雾器的功率动态变化特征的响应性特征，因此为了更好地融合这两者的特征信息来进行分类，进一步计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵。进而，再使用分类器对所述转移矩阵进行分类处理，以获得用于表示当前时间点的喷雾器的功率值应增大或应减小的分类结果。
基于此，本申请提出了一种电子级氢氧化钾的智慧产线的控制系统，其包括：喷雾数据采集模块，用于获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；空间编码模块，用于从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；差分模块，用于计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；动态编码模块，用于将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；降维模块，用于对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；校正模块，用于基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；功率时序编码模块，用于将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；响应性估计模块，用于计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及，控制结果生成模块，用于将所述转移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喷雾器的功率值应增大或应减小。
图1图示了根据本申请实施例的电子级氢氧化钾的智慧产线的控制系统的应用场景图。如图1所示，在该应用场景中，首先，通过传感器(例如，如图1中所示意的功率测量仪T)获取预定时间段内多个预定时间点的喷雾器(例如，如图1中所示意的P)的功率值以及由摄像头(例如，如图1中所示意的C)采集所述预定时间段的喷雾去杂质的监控视频。然后，将获取的所述预定时间段内多个预定时间点的喷雾器的功率值以及所述预定时间段的喷雾去杂质的监控视频输入至部署有电子级氢氧化钾的智慧产线的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，所述服务器能够以电子级氢氧化钾的智慧产线的控制算法对所述预定时间段内多个预定时间点的喷雾器的功率值以及所述预定时间段的喷雾去杂质的监控视频进行处理，以生成用于表示当前时间点的喷雾器的功率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电子级氢氧化钾的智慧产线的控制系统的框图。如图2所示，根 据本申请实施例的电子级氢氧化钾的智慧产线的控制系统200，包括：喷雾数据采集模块210，用于获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；空间编码模块220，用于从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；差分模块230，用于计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；动态编码模块240，用于将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；降维模块250，用于对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；校正模块260，用于基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；功率时序编码模块270，用于将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；响应性估计模块280，用于计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及，控制结果生成模块290，用于将所述转移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喷雾器的功率值应增大或应减小。
具体地，在本申请实施例中，所述喷雾数据采集模块210和所述空间编码模块220，用于获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频，并从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图。如前所述，考虑到由于在该制备方案中，喷淋的喷雾器的功率控制是关键，应可以理解，如果功率过大会导致晶体也被部分地传递走，造成提纯率的下降，而如果功率多小则会导致杂质传递效率低，提升效率低下。因此，在本申请的技术方案中，期待在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制以保证晶体收率的效果和杂质传递的效率。
也就是，具体地，在本申请的技术方案中，考虑到在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制时，还要实时动态地对于喷雾去杂质的效果进行监测，因此，首先，通过功率测量仪获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集所述预定时间段的喷雾去杂质的监控视频。然后，为了便于后续的特征挖掘以及降低计算量，从所述监控视频中提取出多个关键帧，再将所述多个关键帧通过使用在隐含关联特征提取方面具有优异表现的卷积神经网络模型对其进行特征挖掘。应可以理解，在实际的所述喷雾去杂质的过程中，需要更加关注于所述各个关键帧中的晶体和杂质在空间上的关联关系，并更加聚焦于所述杂质的传递，因此，使用空间注意力机制的第一卷积神经网络对所述多个关键帧中各个关键帧进行处理，以得到多个空间聚焦特征图。
更具体地，在本申请实施例中，所述空间编码模块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对所述输入数据进行基于二维卷积核的卷积处理以生成卷积特征图；对所述卷积特征图进行池化处理以生成池化特征图；对所述池化特征图进行激活处理以生成激活特征图；对所述激活特征图进行沿通道维度的全局平均池化以获得空间特征矩阵；对所述空间特征矩阵进行卷积处理和激活处理以生成权重向量；以及，以所述权重向量中各个位置的权重值分别对所述激活特征图的各个特征矩阵进行加权以获得生成特征图；其中，所述第一卷积神经网络的最后一层输出的所述生成特征图为所述多个空间聚焦特征图，所述第一卷积神经网络的第一层的输入为所述多个关键帧中各个关键帧。
具体地，在本申请实施例中，所述差分模块230和所述动态编码模块240，用于计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图，并将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图。也就是，在本申请的技术方案中，进一步地，为了表达出所述喷雾去杂的动态效果，需要更加关注于所述各个空间聚焦特征图的对比，尤其是相邻的两个所述空间聚焦特征图的动态性比较。因此，在本申请的技术方案中，进一步计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图。然后，将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述多个差分特征图中的关于所述喷雾去杂质的动态性隐含特征，从而得到去杂质动态特征图。相应地，在一个具体示例中，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去杂质动态特征图，所述第二卷积神经网络的第一层的输入为所述多个差分特征图。
更具体地，在本申请实施例中，所述差分模块，进一步用于：以如下公式计算所述多个空间聚 焦特征图中每相邻两个空间聚焦特征图之间的差分以得到所述多个差分特征图；其中，所述公式为：
Figure PCTCN2022119555-appb-000020
其中，F i表示所述多个空间聚焦特征图中第i个位置的空间聚焦特征图，F i+1表示所述多个空间聚焦特征图中第i+1个位置的空间聚焦特征图，F表示所述差分特征图，
Figure PCTCN2022119555-appb-000021
表示按位置差分。
具体地，在本申请实施例中，所述降维模块250，用于对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量。应可以理解，考虑到所述去杂质动态特征图中的参数数量较大，这会使得后续的处理较为复杂，因此，在本申请的技术方案中，进一步对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量，以降低参数的数量，防止过拟合，进而提高后续分类的准确性。
具体地，在本申请实施例中，所述校正模块260，用于基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差。应可以理解，在获得所述去杂质动态特征向量的过程当中，使用所述空间注意力机制聚焦了特征矩阵的空间维度上的局部特征，且所述沿通道维度的全局均值池化又进一步基于特征矩阵的空间特征进行了全局性特征描述，这使得所述去杂质动态特征向量的每个位置的特征值之间的相关性减弱，在卷积神经网络随深度的流式传播过程中发生对于类概率的预测压力。因此，在本申请的技术方案中，需要对所述去杂质动态特征向量V进行优化。
更具体地，在一个具体示例中，所述校正模块，进一步用于：基于所述去杂质动态特征向量的自协方差矩阵，以如下公式对所述去杂质动态特征向量中各个位置的特征值进行校正以得到所述校正后去杂质动态特征向量；其中，所述公式为：
Figure PCTCN2022119555-appb-000022
其中V表示所述去杂质动态特征向量，∑是所述去杂质动态特征向量的自协方差矩阵，即矩阵的每个位置的值是向量V的每两个位置的特征值之间的方差，μ和σ分别是所述去杂质动态特征向量的全局均值和方差，exp(·)表示以向量为幂的指数运算，其中，以向量为幂的指数运算表示以向量的每个位置的值作为幂求指数，再将结果填入向量的各个位置以得到向量运算结果，
Figure PCTCN2022119555-appb-000023
和
Figure PCTCN2022119555-appb-000024
分别表示向量的按位置减法和加法，||·|| 2表示特征向量的二范数。应可以理解，这样，所述优化可以对前向传播相关性进行引导修正，具体地，基于所述沿通道维度的全局均值池化对于特征进行的基于下采样的前向传播的特点，通过可学习的正态采样偏移引导卷积神经网络的特征工程来有效地建模所述去杂质动态特征图的特征矩阵内的空间维度和特征矩阵之间的通道维度上的长程依赖关系，并考虑特征矩阵的局部和非局部邻域来进行所述去杂质动态特征向量的各特征值间的相关性的修复，从而提高了所述所述去杂质动态特征向量在卷积神经网络随深度的流式传播过程中对于类概率的预测能力。
具体地，在本申请实施例中，所述功率时序编码模块270，用于将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量。应可以理解，对于所述喷雾器的功率值，考虑到其在时序维度上具有着隐含的动态性的关联特征信息，因此，在本申请的技术方案中，为了能够充分地提取出这种隐含的动态关联特征，使用包含一维卷积层的时序编码器对所述预定时间段内多个预定时间点的喷雾器的功率值进行编码以得到功率控制特征向量。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喷雾器的功率值在时序维度上的关联和通过全连接编码提取所述喷雾器的功率值的高维隐含特征。
更具体地，在本申请实施例中，所述功率时序编码模块，进一步用于：将所述预定时间段内多个预定时间点的喷雾器的功率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 PCTCN2022119555-appb-000025
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 PCTCN2022119555-appb-000026
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 PCTCN2022119555-appb-000027
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，所述响应性估计模块280和所述控制结果生成模块290，用于计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵，并将所述转移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喷雾器的功率值应增大或应减小。应可以理解，考虑到由于所述喷雾器的功率数据和所述喷雾去杂质的监控视频中的各个关键帧数据的特征 尺度不同，并且所述喷雾去杂质的动态图像特征在高维空间中可以看作是针对所述喷雾器的功率动态变化特征的响应性特征，因此为了更好地融合这两者的特征信息来进行分类，在本申请的技术方案中，进一步计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵。进而，再使用分类器对所述转移矩阵进行分类处理，以获得用于表示当前时间点的喷雾器的功率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转移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)表示将所述转移矩阵投影为向量，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，所述响应性估计模块，进一步用于：以如下公式计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的所述转移矩阵；其中，所述公式为：V 2＝M*V 1
其中V 1表示所述功率控制特征向量，M表示所述转移矩阵，V 2表示所述校正后去杂质动态特征向量。
综上，基于本申请实施例的所述电子级氢氧化钾的智慧产线的控制系统200被阐明，其采用人工智能控制技术，基于深度神经网络模型来对于喷雾器的功率动态变化特征和喷雾去杂质图像帧的动态特征进行深层挖掘，以在电子级氢氧化钾的智慧产线中对喷雾器的功率进行实时动态地控制，进而保证晶体收率的效果和杂质传递的效率。
如上所述，根据本申请实施例的电子级氢氧化钾的智慧产线的控制系统200可以实现在各种终端设备中，例如电子级氢氧化钾的智慧产线的控制算法的服务器等。在一个示例中，根据本申请实施例的电子级氢氧化钾的智慧产线的控制系统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，该电子级氢氧化钾的智慧产线的控制系统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，该电子级氢氧化钾的智慧产线的控制系统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，该电子级氢氧化钾的智慧产线的控制系统200与该终端设备也可以是分立的设备，并且该电子级氢氧化钾的智慧产线的控制系统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图示了电子级氢氧化钾的智慧产线的控制方法的流程图。如图3所示，根据本申请实施例的电子级氢氧化钾的智慧产线的控制方法，包括步骤：S110，获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；S120，从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；S130，计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；S140，将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；S150，对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；S160，基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；S170，将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；S180，计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及，S190，将所述转移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喷雾器的功率值应增大或应减小。
图4图示了根据本申请实施例的电子级氢氧化钾的智慧产线的控制方法的架构示意图。如图4所示，在所述电子级氢氧化钾的智慧产线的控制方法的网络架构中，首先，从获得的所述监控视频(例如，如图4中所示意的P1)提取多个关键帧(例如，如图4中所示意的P2)，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络(例如，如图4中所示意的CNN1)以得到多个空间聚焦特征图(例如，如图4中所示意的F1)；接着，计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图(例如，如图4中所示意的F2)；然后，将所述多个差分特征图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)以得到去杂质动态特征图(例如，如图4中所示意的F3)；接着，对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量(例如，如图4中所示意的VF1)；然后，基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量(例如，如图4中所示意的VF2)；接着，将所述预定时间段内多个预定时间点的喷雾器的功率值(例如，如图4中所示意的Q)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)以得到功率控制特征向量(例如，如图4中所示意的VF)；然后，计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵(例如，如图4中所示 意的MF)；以及，最后，将所述转移矩阵通过分类器(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喷雾器的功率值应增大或应减小。
更具体地，在步骤S110和步骤S120中，获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频，并从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图。应可以理解，考虑到由于在该制备方案中，喷淋的喷雾器的功率控制是关键，应可以理解，如果功率过大会导致晶体也被部分地传递走，造成提纯率的下降，而如果功率多小则会导致杂质传递效率低，提升效率低下。因此，在本申请的技术方案中，期待在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制以保证晶体收率的效果和杂质传递的效率。
也就是，具体地，在本申请的技术方案中，考虑到在电子级氢氧化钾的智慧产线中对喷雾器的功率进行智能控制时，还要实时动态地对于喷雾去杂质的效果进行监测，因此，首先，通过功率测量仪获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集所述预定时间段的喷雾去杂质的监控视频。然后，为了便于后续的特征挖掘以及降低计算量，从所述监控视频中提取出多个关键帧，再将所述多个关键帧通过使用在隐含关联特征提取方面具有优异表现的卷积神经网络模型对其进行特征挖掘。应可以理解，在实际的所述喷雾去杂质的过程中，需要更加关注于所述各个关键帧中的晶体和杂质在空间上的关联关系，并更加聚焦于所述杂质的传递，因此，使用空间注意力机制的第一卷积神经网络对所述多个关键帧中各个关键帧进行处理，以得到多个空间聚焦特征图。
更具体地，在步骤S130和步骤S140中，计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图，并将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图。也就是，在本申请的技术方案中，进一步地，为了表达出所述喷雾去杂的动态效果，需要更加关注于所述各个空间聚焦特征图的对比，尤其是相邻的两个所述空间聚焦特征图的动态性比较。因此，在本申请的技术方案中，进一步计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图。然后，将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络中进行处理，以提取出所述多个差分特征图中的关于所述喷雾去杂质的动态性隐含特征，从而得到去杂质动态特征图。相应地，在一个具体示例中，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去杂质动态特征图，所述第二卷积神经网络的第一层的输入为所述多个差分特征图。
更具体地，在步骤S150和步骤S160中，对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量，并基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差。应可以理解，考虑到所述去杂质动态特征图中的参数数量较大，这会使得后续的处理较为复杂，因此，在本申请的技术方案中，进一步对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量，以降低参数的数量，防止过拟合，进而提高后续分类的准确性。但是，在获得所述去杂质动态特征向量的过程当中，使用所述空间注意力机制聚焦了特征矩阵的空间维度上的局部特征，且所述沿通道维度的全局均值池化又进一步基于特征矩阵的空间特征进行了全局性特征描述，这使得所述去杂质动态特征向量的每个位置的特征值之间的相关性减弱，在卷积神经网络随深度的流式传播过程中发生对于类概率的预测压力。因此，在本申请的技术方案中，需要对所述去杂质动态特征向量V进行优化。
更具体地，在步骤S170中，将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量。应可以理解，对于所述喷雾器的功率值，考虑到其在时序维度上具有着隐含的动态性的关联特征信息，因此，在本申请的技术方案中，为了能够充分地提取出这种隐含的动态关联特征，使用包含一维卷积层的时序编码器对所述预定时间段内多个预定时间点的喷雾器的功率值进行编码以得到功率控制特征向量。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喷雾器的功率值在时序维度上的关联和通过全连接编码提取所述喷雾器的功率值的高维隐含特征。
更具体地，在步骤S180和步骤S190中，计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵，并将所述转移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喷雾器的功率值应增大或应减小。应可以理解，考虑到由于所述喷雾器的功率数据和所述喷雾去杂质的监控视频中的各个关键帧数据的特征尺度不同，并且所述喷雾去杂质的动态图像特征在高维空间中可以看作是针对所述喷雾器的功率动态变化特征的响应性特征，因此为了更好地融合这两 者的特征信息来进行分类，在本申请的技术方案中，进一步计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵。进而，再使用分类器对所述转移矩阵进行分类处理，以获得用于表示当前时间点的喷雾器的功率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电子级氢氧化钾的智慧产线的控制方法被阐明，其采用人工智能控制技术，基于深度神经网络模型来对于喷雾器的功率动态变化特征和喷雾去杂质图像帧的动态特征进行深层挖掘，以在电子级氢氧化钾的智慧产线中对喷雾器的功率进行实时动态地控制，进而保证晶体收率的效果和杂质传递的效率。
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. 一种电子级氢氧化钾的智慧产线的控制系统，其特征在于，包括：喷雾数据采集模块，用于获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；空间编码模块，用于从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；差分模块，用于计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；动态编码模块，用于将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；降维模块，用于对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；校正模块，用于基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；功率时序编码模块，用于将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；响应性估计模块，用于计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及控制结果生成模块，用于将所述转移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喷雾器的功率值应增大或应减小。
  2. 根据权利要求1所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述空间编码模块，进一步用于：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对所述输入数据进行基于二维卷积核的卷积处理以生成卷积特征图；对所述卷积特征图进行池化处理以生成池化特征图；对所述池化特征图进行激活处理以生成激活特征图；对所述激活特征图进行沿通道维度的全局平均池化以获得空间特征矩阵；对所述空间特征矩阵进行卷积处理和激活处理以生成权重向量；以及以所述权重向量中各个位置的权重值分别对所述激活特征图的各个特征矩阵进行加权以获得生成特征图；其中，所述第一卷积神经网络的最后一层输出的所述生成特征图为所述多个空间聚焦特征图，所述第一卷积神经网络的第一层的输入为所述多个关键帧中各个关键帧。
  3. 根据权利要求2所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述差分模块，进一步用于：以如下公式计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到所述多个差分特征图；其中，所述公式为：
    Figure PCTCN2022119555-appb-100001
    其中，F i表示所述多个空间聚焦特征图中第i个位置的空间聚焦特征图，F i+1表示所述多个空间聚焦特征图中第i+1个位置的空间聚焦特征图，F表示所述差分特征图，
    Figure PCTCN2022119555-appb-100002
    表示按位置差分。
  4. 根据权利要求3所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述动态编码模块，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去杂质动态特征图，所述第二卷积神经网络的第一层的输入为所述多个差分特征图。
  5. 根据权利要求4所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述校正模块，进一步用于：基于所述去杂质动态特征向量的自协方差矩阵，以如下公式对所述去杂质动态特征向量中各个位置的特征值进行校正以得到所述校正后去杂质动态特征向量；其中，所述公式为：
    Figure PCTCN2022119555-appb-100003
    其中V表示所述去杂质动态特征向量，∑是所述去杂质动态特征向量的自协方差矩阵，μ和σ分别是所述去杂质动态特征向量的全局均值和方差，exp(·)表示以向量为幂的指数运算，其中，以向量为幂的指数运算表示以向量的每个位置的值作为幂求指数，再将结果填入向量的各个位置以得到向量运算结果，
    Figure PCTCN2022119555-appb-100004
    和
    Figure PCTCN2022119555-appb-100005
    分别表示向量的按位置减法和加法，||·|| 2表示特征向量的二范数。
  6. 根据权利要求5所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述功率时序编码模块，进一步用于：将所述预定时间段内多个预定时间点的喷雾器的功率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 PCTCN2022119555-appb-100006
    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 PCTCN2022119555-appb-100007
    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 PCTCN2022119555-appb-100008
    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。
  7. 根据权利要求6所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述响应性估计模块，进一步用于：以如下公式计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的所述转移矩阵；其中，所述公式为：V 2＝M*V 1
    其中V 1表示所述功率控制特征向量，M表示所述转移矩阵，V 2表示所述校正后去杂质动态特征向量。
  8. 根据权利要求7所述的电子级氢氧化钾的智慧产线的控制系统，其特征在于，所述控制结果生成模块，进一步用于：所述分类器以如下公式对所述转移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)表示将所述转移矩阵投影为向量，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. 一种电子级氢氧化钾的智慧产线的控制方法，其特征在于，包括：获取预定时间段内多个预定时间点的喷雾器的功率值以及由摄像头采集的所述预定时间段的喷雾去杂质的监控视频；从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图；计算所述多个空间聚焦特征图中每相邻两个空间聚焦特征图之间的差分以得到多个差分特征图；将所述多个差分特征图通过使用三维卷积核的第二卷积神经网络以得到去杂质动态特征图；对所述去杂质动态特征图进行沿通道维度的全局均值池化以得到去杂质动态特征向量；基于所述去杂质动态特征向量的自协方差矩阵，对所述去杂质动态特征向量中各个位置的特征值进行校正以得到校正后去杂质动态特征向量，其中，所述去杂质动态特征向量的自协方差矩阵中各个位置的特征值为所述去杂质动态特征向量中相应两个位置的特征值之间的方差；将所述预定时间段内多个预定时间点的喷雾器的功率值通过包含一维卷积层的时序编码器以得到功率控制特征向量；计算所述功率控制特征向量相对于所述校正后去杂质动态特征向量的转移矩阵；以及将所述转移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喷雾器的功率值应增大或应减小。
  10. 根据权利要求9所述的电子级氢氧化钾的智慧产线的控制方法，其特征在于，所述从所述监控视频提取多个关键帧，并分别将所述多个关键帧中各个关键帧通过使用空间注意力机制的第一卷积神经网络以得到多个空间聚焦特征图，包括：所述第一卷积神经网络的各层在层的正向传递中对输入数据分别进行：对所述输入数据进行基于二维卷积核的卷积处理以生成卷积特征图；对所述卷积特征图进行池化处理以生成池化特征图；对所述池化特征图进行激活处理以生成激活特征图；对所述激活特征图进行沿通道维度的全局平均池化以获得空间特征矩阵；对所述空间特征矩阵进行卷积处理和激活处理以生成权重向量；以及以所述权重向量中各个位置的权重值分别对所述激活特征图的各个特征矩阵进行加权以获得生成特征图；其中，所述第一卷积神经网络的最后一层输出的所述生成特征图为所述多个空间聚焦特征图，所述第一卷积神经网络的第一层的输入为所述多个关键帧中各个关键帧。
PCT/CN2022/119555 2022-07-29 2022-09-19 电子级氢氧化钾的智慧产线的控制系统及其控制方法 WO2024021258A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210905300.9A CN115202265A (zh) 2022-07-29 2022-07-29 电子级氢氧化钾的智慧产线的控制系统及其控制方法
CN202210905300.9 2022-07-29

Publications (1)

Publication Number Publication Date
WO2024021258A1 true WO2024021258A1 (zh) 2024-02-01

Family

ID=83584785

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/119555 WO2024021258A1 (zh) 2022-07-29 2022-09-19 电子级氢氧化钾的智慧产线的控制系统及其控制方法

Country Status (2)

Country Link
CN (1) CN115202265A (zh)
WO (1) WO2024021258A1 (zh)

Families Citing this family (8)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115350640B (zh) * 2022-10-20 2023-03-14 济宁九德半导体科技有限公司 基于半导体扩散炉用氮氧混合设备及其控制方法
CN115376075B (zh) * 2022-10-25 2023-02-14 中节能绿建环保科技有限公司 智能楼宇的新风节能系统及其控制方法
CN116000297A (zh) * 2023-01-03 2023-04-25 赣州市光华有色金属有限公司 高强度钨镧丝的制备装置及其方法
CN116047987B (zh) * 2023-03-31 2023-06-16 福建天甫电子材料有限公司 用于电子级缓冲氧化物蚀刻液生产的智能控制系统
CN116149280B (zh) * 2023-04-04 2023-07-07 福建德尔科技股份有限公司 电子级氢氧化钾的智能生产系统
CN116821745B (zh) * 2023-04-10 2024-04-09 浙江万能弹簧机械有限公司 智能线切割慢走丝设备的控制方法及其系统
CN116652681A (zh) * 2023-05-24 2023-08-29 浙江沃沃阀门有限公司 锻钢阀阀盖自动化多工位机加工设备及方法
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 (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5852817A (en) * 1991-08-14 1998-12-22 Kabushiki Kaisha Toshiba Intelligent control apparatus
CN107168255A (zh) * 2017-05-16 2017-09-15 浙江工业大学 一种基于集成神经网络的聚丙烯熔融指数混合建模方法
CN114371649A (zh) * 2022-01-10 2022-04-19 江苏大学 一种基于卷积神经网络的喷雾流量调控系统与方法
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

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US5852817A (en) * 1991-08-14 1998-12-22 Kabushiki Kaisha Toshiba Intelligent control apparatus
CN107168255A (zh) * 2017-05-16 2017-09-15 浙江工业大学 一种基于集成神经网络的聚丙烯熔融指数混合建模方法
CN114371649A (zh) * 2022-01-10 2022-04-19 江苏大学 一种基于卷积神经网络的喷雾流量调控系统与方法
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

Also Published As

Publication number Publication date
CN115202265A (zh) 2022-10-18

Similar Documents

Publication Publication Date Title
WO2024021258A1 (zh) 电子级氢氧化钾的智慧产线的控制系统及其控制方法
CN115079572B (zh) 用于六氟磷酸锂制备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
Eldesokey et al. Propagating confidences through cnns for sparse data regression
WO2024045243A1 (zh) 用于六氟磷酸锂制备的自动配料系统及其配料方法
WO2024026993A1 (zh) 氟化铵制备用的自动配料控制系统及其控制方法
CN115599049B (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
Lin et al. Sparse4d: Multi-view 3d object detection with sparse spatial-temporal fusion
CN115688592B (zh) 用于电子级四氟化碳制备的精馏控制系统及其方法
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
CN116149280B (zh) 电子级氢氧化钾的智能生产系统
WO2024007445A1 (zh) 用于氟化锂制备的能源管理控制系统及其控制方法
CN115903705B (zh) 用于电子级六氟丁二烯制备的生产管理控制系统
CN115231525B (zh) 电子级三氟化氯的智能分离纯化系统
CN117523498A (zh) 危险废物的处理回收系统及方法
Lin et al. Sparse4d v3: Advancing end-to-end 3d detection and tracking
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
CN117315585A (zh) 电解槽智能控制系统及其方法
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
Zhang et al. Dyna-depthformer: Multi-frame transformer for self-supervised depth estimation in dynamic scenes
CN117046140B (zh) 用于电子级氢氟酸精馏的装置
CN116920739A (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
CN116825217B (zh) 制备高纯五氟化磷的方法
CN116825215B (zh) 用于六氟磷酸锂制备的流体循环反应控制系统及其方法

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: 22952699

Country of ref document: EP

Kind code of ref document: A1