WO2024036690A1 - 用于剥膜液生产的自动配料系统及其配料方法 - Google Patents

用于剥膜液生产的自动配料系统及其配料方法 Download PDF

Info

Publication number
WO2024036690A1
WO2024036690A1 PCT/CN2022/119832 CN2022119832W WO2024036690A1 WO 2024036690 A1 WO2024036690 A1 WO 2024036690A1 CN 2022119832 W CN2022119832 W CN 2022119832W WO 2024036690 A1 WO2024036690 A1 WO 2024036690A1
Authority
WO
WIPO (PCT)
Prior art keywords
matrix
feature
component
adjacency
feature matrix
Prior art date
Application number
PCT/CN2022/119832
Other languages
English (en)
French (fr)
Inventor
罗永春
袁瑞明
谢荣周
林金华
罗霜
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 WO2024036690A1 publication Critical patent/WO2024036690A1/zh

Links

Images

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16C—COMPUTATIONAL CHEMISTRY; CHEMOINFORMATICS; COMPUTATIONAL MATERIALS SCIENCE
    • G16C20/00—Chemoinformatics, i.e. ICT specially adapted for the handling of physicochemical or structural data of chemical particles, elements, compounds or mixtures
    • G16C20/60—In silico combinatorial chemistry
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
    • Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
    • Y02P90/30—Computing systems specially adapted for manufacturing

Definitions

  • the present invention relates to intelligent batching in the field of intelligent production, and more specifically, to an automatic batching system for the production of film stripping liquid and a batching method thereof.
  • PCB Printed Circuit Board
  • NaOH sodium hydroxide
  • COD chemical oxygen demand
  • Patent No. CN110331048B discloses an environmentally friendly film stripping liquid, which includes the components of the environmentally friendly film stripping liquid in terms of mass percentage: 35% to 45% of complex alkali, 3% to 5% of polyether surfactant, and water. 50% to 62%; wherein, the main components of the composite base include Ca(OH) 2 , activated white mud, diatomite and activated carbon; the polyether surfactant is fatty alcohol polyoxyethylene ether.
  • the automatic batching system for liquid production takes into account the surface state characteristics of the object to be peeled off when batching to obtain an optimal formula adapted to the specific peeling object.
  • Embodiments of the present application provide an automatic batching system and batching method for the production of peeling liquid, which uses a convolutional neural network model based on deep learning to mine the surface state characteristics of the image of the object to be peeled, and based on The global correlation characteristics of the mass percentage of each component of the peeling liquid in the ingredients and the logical correlation characteristics between the various components of the peeling liquid are used to assist in the adjustment of the mass percentage of the polyether surfactant, thereby obtaining The optimal formula adapted to specific peeling objects to obtain better peeling effects and ensure the quality of PCB boards.
  • an automatic batching system for the production of peeling liquid which includes: a batching data acquisition module, used to obtain the mass percentage of each component of the peeling liquid in the batching, wherein,
  • the components of the peeling liquid include complex alkali, polyether surfactant and water.
  • the complex alkali includes Ca(OH)2, activated white mud, diatomite and activated carbon;
  • the component mass percentage coding module is used The mass percentage of each component of the peeling liquid in the ingredients is passed through the context encoder based on the converter to obtain multiple component feature vectors; the component mass percentage associated coding module is used to combine the multiple component feature vectors.
  • the component feature vectors are arranged into a two-dimensional matrix and then passed through the first convolutional neural network as a feature extractor to obtain the component feature matrix;
  • a rule operation module is used to combine the components of the peeling liquid based on and disjunctive logical operation rules to construct a conjunctive adjacency matrix and a disjunctive adjacency matrix between the component feature vectors corresponding to each component, wherein the conjunctive logical operation rules represent the parallel relationship between the rules,
  • the disjunctive logical operation represents the substitution relationship between rules;
  • a logical matrix encoding module is used to pass the conjunctive adjacency matrix and the disjunctive adjacency matrix through the second convolutional neural network to obtain the first feature matrix and the third feature matrix.
  • a class adjacency matrix building module used to calculate the position-weighted sum between the first feature matrix and the second feature matrix to obtain a class adjacency matrix
  • a fusion module used to fuse the component features matrix and the class adjacency matrix to obtain the recipe feature matrix
  • the object to be peeled encoding module is used to pass the acquired image of the object to be peeled through the third convolutional neural network as a feature extractor to obtain the surface state feature matrix
  • the response module used to calculate the transfer matrix of the formula feature matrix relative to the surface state feature matrix as a classification feature matrix
  • an ingredient control module used to pass the classification feature matrix through a classifier to obtain a classification result, the classification result Used to indicate whether the mass percentage of polyether surfactant should be increased or decreased.
  • the component mass percentage associated encoding module includes: an input vector conversion unit for using the embedding layer of the converter-based context encoder to respectively convert the The mass percentage of each component of the peeling liquid in the ingredients is converted into an input vector to obtain a sequence of input vectors; and, a context encoding unit for using the converter of the converter-based context encoder to encode the input
  • the sequence of vectors is globally-based contextual semantic encoding to obtain the plurality of component feature vectors.
  • the rule operation module includes: a conjunction adjacency matrix construction unit for logical operation rules based on the conjunction between the various components of the peeling liquid.
  • the conjunctive adjacency matrix between the component feature vectors corresponding to each component is constructed using the following formula; wherein the formula is:
  • a conjunctive matrix used to represent that the matrix position is 1 when the corresponding pair of rules form a conjunctive normal form, and is 0 when it is not a conjunctive normal form
  • a disjunctive adjacency matrix construction unit used to based on the peeling
  • the logical operation rules of the disjunction between the various components of the liquid are used to construct the disjunction adjacency matrix between the component feature vectors corresponding to the respective components using the following formula; wherein, the formula is: where, is a disjunctive adjacency matrix, used to indicate that the matrix position is 1 when the corresponding pair of rules form a disjunctive normal form, and the matrix position is 0 when it is not a disjunctive normal form.
  • the logical matrix encoding module is further used to perform: each layer of the second convolutional neural network in the forward transmission of the layer: perform on the input data Convolution processing to obtain a convolution feature map; performing mean pooling based on local channel dimensions on the convolution feature map to obtain a pooled feature map; and performing nonlinear activation on the pooled feature map to obtain activation features Figure; wherein, the output of the last layer of the second convolutional neural network is the first feature matrix and the second feature matrix, and the input of the first layer of the second convolutional neural network is the conjunction adjacency matrix and the disjunctive adjacency matrix.
  • M s the class adjacency matrix
  • M 1 is the first characteristic matrix
  • M 2 is the first characteristic matrix.
  • Two feature matrices, "+” means the addition of elements at corresponding positions of the first feature matrix and the second feature matrix, ⁇ and ⁇ are used to control the sum of the first feature matrix in the class adjacency matrix
  • the encoding module of the object to be peeled is further used to: each layer of the third convolutional neural network performs respectively: in the forward transmission of the layer: input data Perform convolution processing to obtain a convolution feature map; perform mean pooling based on local channel dimensions on the convolution feature map to obtain a pooled feature map; and perform nonlinear activation on the pooled feature map to obtain activation Feature map; wherein, the output of the last layer of the third convolutional neural network is the surface state feature matrix, and the input of the first layer of the third convolutional neural network is the image of the object to be peeled off.
  • the correction unit is further used to: correct the characteristic values of each position in the formula characteristic matrix based on the surface state characteristic matrix with the following formula to obtain the The corrected formula characteristic matrix; wherein, the formula is: f i ⁇ max-norm (M 1 ) and f j ⁇ max-norm (M 2 )
  • M 1 represents the formula characteristic matrix
  • M 2 represents the surface state characteristic matrix
  • fi represents the characteristic value of the corresponding position of the formula characteristic matrix normalized to the [0,1] interval
  • f j represents The corresponding position of the surface state characteristic matrix is normalized to the characteristic value in the [0,1] interval
  • d( fi ,f j ) represents the characteristic value of each position in the formula characteristic matrix and the surface state
  • max-norm( ⁇ ) represents maximum value normalization
  • ⁇ is the control hyperparameter.
  • the batching control module is further configured to: use the classifier to process the classification feature matrix with the following formula to generate a classification result, wherein the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • an automatic batching method for the production of peeling liquid includes: obtaining the mass percentage of each component of the peeling liquid in the batching, wherein the components of the peeling liquid Including complex alkali, polyether surfactant and water, the complex alkali includes Ca (OH) 2, activated white mud, diatomaceous earth and activated carbon; the quality of each component of the film stripping liquid in the ingredients is The percentage is passed through the context encoder based on the converter to obtain multiple component feature vectors; the multiple component feature vectors are arranged into a two-dimensional matrix and then passed through the first convolutional neural network as a feature extractor to obtain the component features.
  • Matrix construct a conjunctive adjacency matrix and a disjunctive adjacency matrix between the component feature vectors corresponding to each component based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid,
  • the conjunctive logic operation rules represent the parallel relationship between the rules
  • the disjunctive logic operation represents the substitution relationship between the rules
  • the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through the second convolution
  • the neural network obtains the first feature matrix and the second feature matrix; calculates the position-weighted sum between the first feature matrix and the second feature matrix to obtain the class adjacency matrix; fuses the component feature matrix and the
  • the above-mentioned adjacency matrix is used to obtain the formula feature matrix;
  • the acquired image of the object to be peeled is passed through the third convolutional neural network as a feature extractor to obtain the surface state feature matrix;
  • the formula feature matrix is calculated relative to the surface state feature
  • the mass percentage of each component of the peeling liquid in the batch is passed through the context encoder based on the converter to obtain multiple component feature vectors, including: Use the embedding layer of the converter-based context encoder to respectively convert the mass percentage of each component of the peeling fluid in the ingredients into an input vector to obtain a sequence of input vectors; and, use the converter-based context encoder
  • the converter of the context encoder performs global-based context semantic encoding on the sequence of input vectors to obtain the plurality of component feature vectors.
  • one of the component feature vectors corresponding to each component of the peeling liquid is constructed based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid.
  • the conjunctive adjacency matrix and the disjunctive adjacency matrix between the film peeling liquid include: based on the conjunctive logical operation rules between the various components of the peeling liquid, using the following formula to construct one of the component feature vectors corresponding to each component.
  • disjunctive adjacency matrix used to indicate that the matrix position is 1 when the corresponding pair of rules form a disjunctive normal form, and the matrix position is 0 when it is not a disjunctive normal form.
  • the conjunctive adjacency matrix and the disjunctive adjacency matrix are passed through the second convolutional neural network to obtain the first characteristic matrix and the second characteristic matrix, including:
  • Each layer of the second convolutional neural network is performed separately in the forward pass of the layer: the input data is convolved to obtain a convolution feature map; the convolution feature map is subjected to mean pooling based on the local channel dimension.
  • the output of the last layer of the second convolutional neural network is the first feature matrix and The second feature matrix
  • the input of the first layer of the second convolutional neural network is the conjunctive adjacency matrix and the disjunctive adjacency matrix.
  • calculating the position-weighted sum between the first feature matrix and the second feature matrix to obtain a class adjacency matrix includes: calculating the first feature matrix with the following formula The position-weighted sum between a feature matrix and the second feature matrix is used to obtain the class adjacency matrix; wherein, the formula is:
  • M s is the class adjacency matrix
  • M 1 is the first feature matrix
  • M 2 is the second feature matrix
  • "+" means that the first feature matrix and the second feature matrix correspond to each other.
  • the elements at the position are added
  • ⁇ and ⁇ are weighting parameters used to control the balance between the first feature matrix and the second feature matrix in the class adjacency matrix.
  • the acquired image of the object to be peeled is passed through the third convolutional neural network as a feature extractor to obtain the surface state feature matrix, including: the third convolutional neural network
  • Each layer of the network is performed separately in the forward pass of the layer: convolution processing is performed on the input data to obtain a convolution feature map; mean pooling based on the local channel dimension is performed on the convolution feature map to obtain a pooled feature map.
  • the output of the last layer of the third convolutional neural network is the surface state feature matrix, and the third convolution The input of the first layer of the neural network is the image of the object to be peeled off.
  • the characteristic values of each position in the formula characteristic matrix are corrected based on the surface state characteristic matrix to obtain the corrected formula characteristic matrix, including: based on the surface state
  • the characteristic matrix uses the following formula to correct the characteristic values of each position in the formula characteristic matrix to obtain the corrected formula characteristic matrix; wherein the formula is: f i ⁇ max-norm (M 1 ) and f j ⁇ max-norm(M 2 )
  • M 1 represents the formula characteristic matrix
  • M 2 represents the surface state characteristic matrix
  • fi represents the characteristic value of the corresponding position of the formula characteristic matrix normalized to the [0,1] interval
  • f j represents The corresponding position of the surface state characteristic matrix is normalized to the characteristic value in the [0,1] interval
  • d( fi ,f j ) represents the characteristic value of each position in the formula characteristic matrix and the surface state
  • max-norm( ⁇ ) represents maximum value normalization
  • ⁇ is the control hyperparameter.
  • 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 a classification result , where the formula is: softmax ⁇ (W n ,B n ):...:(W 1 ,B 1 )
  • the automatic batching system and batching method provided by this application for the production of peeling liquid uses a convolutional neural network model based on deep learning to mine the surface state characteristics of the image of the object to be peeled, and Based on the global correlation characteristics of the mass percentage of each component of the peeling liquid in the ingredients and the logical correlation characteristics between the various components of the peeling liquid, the adjustment of the mass percentage of the polyether surfactant is assisted, and then To obtain the optimal formula suitable for specific peeling objects, in order to obtain better peeling effects and ensure the quality of PCB boards.
  • Figure 1 is an application scenario diagram of an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • FIG. 2 is a block diagram of an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • Figure 3 is a block diagram of a response module in an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • Figure 4 is a flow chart of an automatic batching method for film stripping liquid production according to an embodiment of the present application.
  • Figure 5 is a schematic structural diagram of an automatic batching method for film stripping liquid production according to an embodiment of the present application.
  • the film stripping process is the first step in the PCB stripping process. A key step in the preparation process.
  • NaOH sodium hydroxide
  • COD chemical oxygen demand
  • Patent No. CN110331048B discloses an environmentally friendly film stripping liquid, which includes the components of the environmentally friendly film stripping liquid in terms of mass percentage: 35% to 45% of complex alkali, 3% to 5% of polyether surfactant, and water. 50% to 62%; wherein, the main components of the composite base include Ca(OH) 2 , activated white mud, diatomite and activated carbon; the polyether surfactant is fatty alcohol polyoxyethylene ether.
  • the automatic batching system for liquid production takes into account the surface state characteristics of the object to be peeled off when batching to obtain an optimal formula adapted to the specific peeling object.
  • the inventor of the present application considered that in the actual peeling process, the surface states of PCB boards of different electrical equipment have different characteristics, and the polyether surfactant can significantly reduce the surface tension. Therefore, if Better peeling effects can be obtained for different peeling objects. It is expected that the surface state characteristics of the peeling objects will be taken into consideration when formulating to produce a better formula adapted to different peeling objects. This is essentially a classification problem, that is, based on the image of the object to be stripped to represent its surface state features, and assisted by the mass percentage correlation features and logical correlation features of each component of the peeling liquid in the ingredients. The mass percentage of the polyether surfactant is classified and adjusted to ensure the quality of the PCB board.
  • the mass percentage of each component of the film stripping liquid in the ingredients is obtained.
  • the components of the film stripping liquid include complex alkali, polyether surfactant and water.
  • the composite base includes Ca(OH) 2 , activated white mud, diatomaceous earth and activated carbon. It should be understood that considering that there is a correlation between the mass percentage data of each component of the peeling liquid in the ingredients, a converter-based context encoder is used to encode it to extract the The global high-dimensional semantic features between the mass percentages of each component of the peeling liquid in the ingredients are more suitable for characterizing the essential proportion characteristics of the peeling liquid, thereby obtaining multiple component feature vectors.
  • the multiple component feature vectors are arranged into a two-dimensional matrix to integrate the global feature information of the mass percentage of each component of the film stripping liquid in the ingredients, and use the method with the ability to extract implicit correlation features Use the convolutional neural network model with excellent performance to perform feature mining on it, thereby obtaining the component feature matrix.
  • logical operations between rules usually include conjunction and disjunction, represented by symbols ⁇ and ⁇ respectively, they are used to express the parallel or substitution relationship between rules, which is the meaning of "and” and "or”.
  • This relationship also exists between the various components of the peeling liquid.
  • the components of the peeling liquid that need to cooperate with each other and react with each other are in an "and” relationship, and the peeling liquid has an "and” relationship.
  • the components of each component of the membrane fluid that can be replaced by other components are in an "or” relationship. Therefore, in the technical solution of the present application, it is also necessary to perform auxiliary characterization of the implicit correlation characteristics between the component percentages of the stripping liquid based on this logical operation rule relationship.
  • the conjunctive adjacency matrix and disjunction between the component feature vectors corresponding to the respective components are constructed.
  • Adjacency matrix here, the conjunctive logic operation rules represent the parallel relationship between the rules, and the disjunctive logic operation represents the substitution relationship between the rules.
  • feature extraction is performed on the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to dig out the hidden conjunctions and disjunctions between the various components of the peeling fluid.
  • Logical operation rule characteristics thereby obtaining the first characteristic matrix and the second characteristic matrix.
  • the position-weighted sum between the first feature matrix and the second feature matrix can be calculated to fuse the conjunctive logical association features and the disjunctive logical association between the various components of the peeling fluid.
  • features to obtain a class adjacency matrix are fused to obtain a recipe feature matrix, and then the global correlation features of the mass percentage of each component of the peeling liquid in the ingredients and the peeling liquid are merged.
  • the implicit logical association characteristics between the various components of the liquid are used to improve the accuracy of subsequent classification.
  • the image of the object to be stripped is passed through a third convolutional neural network as a feature extractor for feature mining to extract local high-dimensional hidden features of the image of the object to be stripped, thereby obtaining a surface state feature matrix.
  • the mass percentage representing the polyether surfactant can be obtained. Classification results that should be increased or decreased.
  • the recipe feature matrix since the recipe feature matrix incorporates the logical adjacency and disjunction relationships of the adjacency-like matrix, the recipe feature matrix Departure from parameter correlation semantics in feature distribution.
  • the transfer matrix of the recipe feature matrix relative to the surface state feature matrix is directly calculated, it may be that the recipe feature matrix and the surface state feature matrix respectively reside in the entire high dimension due to anisotropy.
  • a narrow subset of the feature space causes the solution space of the transfer matrix to degenerate and lack continuity.
  • it increases the difficulty of calculating the transfer matrix (for example, the iterative process of matrix inversion is difficult to converge), and on the other hand, it affects the classification performance of the transfer matrix. .
  • each eigenvalue of the formula characteristic matrix is first compared to the search space isotropic based on the surface state characteristic matrix, that is: f i ⁇ max-norm (M 1 ) and f j ⁇ max -norm(M 2 )
  • M 1 represents the formula characteristic matrix
  • M 2 represents the surface state characteristic matrix
  • fi represents the characteristic value of the corresponding position of the formula characteristic matrix normalized to the [0,1] interval
  • f j represents The corresponding position of the surface state characteristic matrix is normalized to the characteristic value in the [0,1] interval
  • d( fi ,f j ) represents the characteristic value of each position in the formula characteristic matrix and the surface state
  • max-norm( ⁇ ) represents maximum value normalization
  • ⁇ is the control hyperparameter, for example, the initial setting is the distance between the matrices M 1 and M 2 .
  • the feature distribution of the formula feature matrix can be transferred to a representation space that is isotropic and differentiated from the surface state feature matrix, thereby simplifying the calculation of the transfer matrix. , enhances the distribution continuity of the feature representation of the transfer matrix, improves its classification performance, and thereby improves the accuracy of classification.
  • this application proposes an automatic batching system for the production of peeling liquid, which includes: a batching data collection module for obtaining the mass percentage of each component of the peeling liquid in the batching, wherein, the The components of the stripping liquid include complex alkali, polyether surfactant and water.
  • the complex alkali includes Ca(OH)2, activated white mud, diatomaceous earth and activated carbon; the component mass percentage coding module is used to The mass percentage of each component of the film stripping liquid in the ingredients is passed through the context encoder based on the converter to obtain multiple component feature vectors; the component mass percentage associated coding module is used to combine the multiple components The feature vectors are arranged into a two-dimensional matrix and then passed through the first convolutional neural network as a feature extractor to obtain the component feature matrix; the rule operation module is used to combine and analyze based on the various components of the peeling liquid.
  • the logical operation rules are taken to construct the conjunctive adjacency matrix and the disjunctive adjacency matrix between the component feature vectors corresponding to each component, wherein the conjunctive logical operation rules represent the parallel relationship between the rules, and the The disjunctive logical operation represents the substitution relationship between the rules;
  • the logical matrix encoding module is used to pass the conjunctive adjacency matrix and the disjunctive adjacency matrix through the second convolutional neural network to obtain the first feature matrix and the second feature matrix;
  • a class adjacency matrix building module for calculating a position-weighted sum between the first feature matrix and the second feature matrix to obtain a class adjacency matrix;
  • a fusion module for fusing the component feature matrix and
  • the class adjacency matrix is used to obtain the recipe feature matrix;
  • the object to be peeled encoding module is used to pass the acquired image of the object to be peeled through the third convolutional neural network as a feature extractor
  • FIG. 1 illustrates an application scenario diagram of an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • the peeling liquid in the ingredients for example, T as shown in Figure 1
  • the cloud storage terminal for example, D as shown in Figure 1
  • the mass percentage of each component, here, the components of the film stripping liquid include complex alkali, polyether surfactant and water, the complex alkali includes Ca(OH) 2 , activated white mud, diatomaceous earth and activated carbon, and based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid in the cloud storage terminal, a conjunctive adjacency matrix sum between the component feature vectors corresponding to the respective components is constructed.
  • the adjacency matrix is extracted, and an image of the object to be peeled off (eg, P illustrated in FIG. 1 ) is acquired through a camera (eg, C illustrated in FIG. 1 ).
  • the object to be stripped may be a PCB circuit.
  • the obtained mass percentage of each component of the peeling liquid, the conjunctive adjacency matrix and the disjunctive adjacency matrix, and the image of the object to be peeled are input to an automatic machine deployed for peeling liquid production.
  • the server of the batching algorithm for example, the cloud server S as shown in Figure 1
  • the server can calculate the mass percentage of each component of the stripping liquid with an automatic batching algorithm for the production of the stripping liquid.
  • the conjunctive adjacency matrix, the disjunctive adjacency matrix, and the image of the object to be stripped are processed to generate a classification result indicating that the mass percentage of the polyether surfactant should be increased or decreased.
  • FIG. 2 illustrates a block diagram of an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • the automatic batching system 200 for the production of peeling liquid includes: a batching data acquisition module 210, used to obtain the mass percentage of each component of the peeling liquid in the batching, wherein, the components of the peeling liquid include complex alkali, polyether surfactant and water, the complex alkali includes Ca(OH)2, activated white mud, diatomite and activated carbon; the component mass percentage coding module 220, used to pass the mass percentage of each component of the peeling liquid in the ingredients through the context encoder based on the converter to obtain multiple component feature vectors; the component mass percentage associated encoding module 230, used to The plurality of component feature vectors are arranged into a two-dimensional matrix and then passed through the first convolutional neural network as a feature extractor to obtain a component feature matrix; the rule operation module 240 is used to calculate each component of the peeling liquid
  • the logical matrix encoding module 250 is used to pass the conjunctive adjacency matrix and the disjunctive adjacency matrix through the second convolutional neural network to obtain The first feature matrix and the second feature matrix;
  • the class adjacency matrix building module 260 used to calculate the position-weighted sum between the first feature matrix and the second feature matrix to obtain the class adjacency matrix;
  • the fusion module 270 Used to fuse the component feature matrix and the class adjacency matrix to obtain a recipe feature matrix;
  • the object to be stripped encoding module 280 is used to pass the acquired image of the object to be stripped through a third convolutional neural network as a feature extractor To obtain the surface state feature matrix;
  • the ingredient data acquisition module 210 and the component mass percentage encoding module 220 are used to obtain the mass percentage of each component of the peeling liquid in the ingredients, wherein, the The components of the film stripping liquid include complex alkali, polyether surfactant and water.
  • the complex alkali includes Ca(OH) 2 , activated white mud, diatomaceous earth and activated carbon, and the film stripping liquid in the ingredients is The mass percentage of each component of the liquid is passed through the converter-based context encoder to obtain multiple component feature vectors.
  • the components of the film stripping liquid include complex alkali, polyether surface activity agent and water, and the composite alkali includes Ca(OH) 2 , activated white mud, diatomaceous earth and activated carbon.
  • the component mass percentage associated encoding module includes: an input vector conversion unit for using the embedding layer of the converter-based context encoder to respectively convert the ingredients in the ingredients.
  • the mass percentage of each component of the peeling fluid is converted into an input vector to obtain a sequence of input vectors; and, a context encoding unit for using the converter of the converter-based context encoder to encode the sequence of the input vectors
  • Global-based contextual semantic encoding is performed to obtain the plurality of component feature vectors.
  • the component mass percentage correlation encoding module 230 is used to arrange the multiple component feature vectors into a two-dimensional matrix and then pass it through the first convolutional neural network as a feature extractor. to obtain the component characteristic matrix. That is, in the technical solution of the present application, further, the plurality of component feature vectors are arranged into a two-dimensional matrix to integrate the global feature information of the mass percentage of each component of the peeling liquid in the ingredients. , and use the convolutional neural network model with excellent performance in implicit correlation feature extraction to perform feature mining on it, thereby obtaining the component feature matrix.
  • the rule operation module 240 and the logic matrix encoding module 250 are used to calculate the result based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid. Construct a conjunctive adjacency matrix and a disjunctive adjacency matrix between the component feature vectors corresponding to each component, wherein the conjunctive logic operation rule represents the parallel relationship between the rules, and the disjunctive logic operation represents the rule substitution relationship between them, and pass the conjunctive adjacency matrix and the disjunctive adjacency matrix through the second convolutional neural network to obtain the first feature matrix and the second feature matrix.
  • the conjunctive adjacency matrix and disjunction between the component feature vectors corresponding to the respective components are constructed.
  • Adjacency matrix here, the conjunctive logic operation rules represent the parallel relationship between the rules, and the disjunctive logic operation represents the substitution relationship between the rules.
  • feature extraction is performed on the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to dig out the hidden conjunctions and disjunctions between the various components of the peeling fluid.
  • Logical operation rule characteristics thereby obtaining the first characteristic matrix and the second characteristic matrix.
  • the logical matrix encoding module is further configured to perform convolution processing on the input data in each layer of the second convolutional neural network in the forward pass of the layer to obtain Convolution feature map; perform mean pooling based on local channel dimensions 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 first feature matrix and the second feature matrix, and the input of the first layer of the second convolutional neural network is the conjunction adjacency matrix and the Disjunctive adjacency matrix.
  • the rule operation module includes: a conjunctive adjacency matrix construction unit, which is used to calculate the following formula based on the conjunctive logical operation rules between the various components of the peeling liquid. To construct the conjunctive adjacency matrix between the component feature vectors corresponding to each component; wherein, the formula is:
  • a conjunctive matrix used to represent that the matrix position is 1 when the corresponding pair of rules form a conjunctive normal form, and is 0 when it is not a conjunctive normal form
  • a disjunctive adjacency matrix construction unit used to based on the peeling
  • the logical operation rules of the disjunction between the various components of the liquid are used to construct the disjunction adjacency matrix between the component feature vectors corresponding to the respective components using the following formula; wherein, the formula is: where, is a disjunctive adjacency matrix, used to indicate that the matrix position is 1 when the corresponding pair of rules form a disjunctive normal form, and the matrix position is 0 when it is not a disjunctive normal form.
  • the class adjacency matrix building module 260 and the fusion module 270 are used to calculate the position-weighted sum between the first feature matrix and the second feature matrix to obtain class adjacency matrix, and fuse the component feature matrix and the class adjacency matrix to obtain a recipe feature matrix. That is to say, in the technical solution of the present application, it is further possible to calculate the position-weighted sum between the first characteristic matrix and the second characteristic matrix to fuse the differences between the various components of the peeling liquid. Conjunctive logic associates features and disjunctive logic associates features to obtain adjacency-like matrices.
  • the component feature matrix and the class adjacency matrix are fused to obtain a recipe feature matrix, and then the global correlation features of the mass percentage of each component of the peeling liquid in the ingredients and the peeling liquid are merged
  • the implicit logical association characteristics between the various components of the liquid are used to improve the accuracy of subsequent classification.
  • the class adjacency matrix building module is further configured to: calculate the position-weighted sum between the first feature matrix and the second feature matrix using the following formula to obtain the The above-mentioned adjacency matrix; wherein, the formula is:
  • M s is the class adjacency matrix
  • M 1 is the first feature matrix
  • M 2 is the second feature matrix
  • "+" means that the first feature matrix and the second feature matrix correspond to each other.
  • the elements at the position are added
  • ⁇ and ⁇ are weighting parameters used to control the balance between the first feature matrix and the second feature matrix in the class adjacency matrix.
  • the object to be stripped encoding module 280 is used to pass the acquired image of the object to be stripped through a third convolutional neural network as a feature extractor to obtain a surface state feature matrix. It should be understood that if you want to adjust the mass percentage of the polyether surfactant according to the actual situation, you also need to obtain an image of the object to be stripped through a camera. Then, the image of the object to be stripped is passed through a third convolutional neural network as a feature extractor for feature mining to extract local high-dimensional hidden features of the image of the object to be stripped, thereby obtaining a surface state feature matrix.
  • the object encoding module to be stripped is further used to perform convolution on the input data in each layer of the third convolutional neural network in the forward pass of the layer. Processing to obtain a convolution feature map; performing mean pooling based on local channel dimensions on the convolution feature map to obtain a pooled feature map; and performing nonlinear activation on the pooled feature map to obtain an activation feature map;
  • the output of the last layer of the third convolutional neural network is the surface state feature matrix
  • the input of the first layer of the third convolutional neural network is the image of the object to be peeled off.
  • the response module 290 is used to calculate the transfer matrix of the formula feature matrix relative to the surface state feature matrix as a classification feature matrix. It should be understood that, further, by calculating the transfer matrix of the formula characteristic matrix relative to the surface state characteristic matrix, so as to fuse the characteristic information of the two for classification, the mass percentage representing the polyether surfactant can be obtained. Classification results that should be increased or decreased. However, when calculating the transfer matrix of the recipe feature matrix relative to the surface state feature matrix, since the recipe feature matrix incorporates the logical adjacency and disjunction relationships of the adjacency-like matrix, the recipe feature matrix Departure from parameter correlation semantics in feature distribution.
  • the transfer matrix of the recipe feature matrix relative to the surface state feature matrix is directly calculated, it may be that the recipe feature matrix and the surface state feature matrix respectively reside in the entire high dimension due to anisotropy.
  • a narrow subset of the feature space causes the solution space of the transfer matrix to degenerate and lack continuity.
  • it increases the difficulty of calculating the transfer matrix (for example, the iterative process of matrix inversion is difficult to converge), and on the other hand, it affects the classification performance of the transfer matrix. . Therefore, in the technical solution of the present application, preferably before calculating the transfer matrix, each eigenvalue of the formula characteristic matrix is first compared to the search space isotropic based on the surface state characteristic matrix.
  • the feature distribution of the recipe feature matrix can be transferred to a representation space that is isotropic and differentiated from the surface state feature matrix, thereby simplifying the transfer. While calculating the matrix, the distribution continuity of the feature representation of the transfer matrix is enhanced, its classification performance is improved, and the accuracy of the classification is improved.
  • the response module includes: first, correcting the characteristic values of each position in the formula characteristic matrix based on the surface state characteristic matrix to obtain a corrected formula characteristic matrix.
  • the characteristic values of each position in the formula characteristic matrix are corrected with the following formula based on the surface state characteristic matrix to obtain the corrected formula characteristic matrix;
  • the formula is: f i ⁇ max-norm(M 1 )and f j ⁇ max-norm(M 2 )
  • M 1 represents the formula characteristic matrix
  • M 2 represents the surface state characteristic matrix
  • fi represents the characteristic value of the corresponding position of the formula characteristic matrix normalized to the [0,1] interval
  • f j represents The corresponding position of the surface state characteristic matrix is normalized to the characteristic value in the [0,1] interval
  • d( fi ,f j ) represents the characteristic value of each position in the formula characteristic matrix and the surface state
  • max-norm( ⁇ ) represents maximum value normalization
  • ⁇ is the control hyperparameter, for example, the initial setting is the distance between the matrices M 1 and M 2 .
  • Figure 3 illustrates a block diagram of a response module in an automatic batching system for film stripping liquid production according to an embodiment of the present application.
  • the batching control module 300 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 mass percentage of polyether surfactant should be increased. larger or should be reduced.
  • the classifier is used to process the classification feature matrix with the following formula to generate a classification result, where the formula is: softmax ⁇ (W n ,Bn):...:(W 1 ,B 1 )[Project(F) ⁇ , where Project(F) represents projecting the classification feature matrix into a vector, W 1 to W n are the weight matrices of the fully connected layers of each layer, and B 1 to B n represent each The bias matrix of the fully connected layer.
  • the automatic batching system 200 for peeling liquid production based on the embodiment of the present application is clarified, which uses a convolutional neural network model based on deep learning to mine the surface state characteristics of the object image to be peeled, and based on The global correlation characteristics of the mass percentage of each component of the peeling liquid in the ingredients and the logical correlation characteristics between the various components of the peeling liquid assist in the adjustment of the mass percentage of the polyether surfactant, and then Obtain the optimal formula suitable for specific peeling objects to obtain better peeling effects and ensure the quality of PCB boards.
  • the automatic batching system 200 for the production of peeling liquid according to the embodiment of the present application can be implemented in various terminal devices, such as a server of the automatic batching algorithm for the production of peeling liquid, etc.
  • the automatic dispensing system 200 for film stripping liquid production according to the embodiment of the present application can be integrated into the terminal device as a software module and/or hardware module.
  • the automatic batching system 200 for film peeling liquid production 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 automatic batching system 200 for film peeling liquid
  • the automatic batching system 200 for liquid production can also be one of the many hardware modules of the terminal equipment.
  • the automatic batching system 200 for peeling liquid production and the terminal device can also be separate devices, and the automatic batching system 200 for peeling liquid production can be connected via wired and/or Or a wireless network is connected to the terminal device, and interactive information is transmitted according to the agreed data format.
  • Figure 4 illustrates a flow chart of an automated batching method for stripping liquid production.
  • the automatic batching method for the production of peeling liquid includes the step: S110, obtaining the mass percentage of each component of the peeling liquid in the batching, wherein the peeling liquid
  • the components of the liquid include complex alkali, polyether surfactant and water.
  • the complex alkali includes Ca(OH)2, activated white mud, diatomaceous earth and activated carbon; S120, add the film stripping liquid in the ingredients.
  • the mass percentage of each component is passed through the context encoder based on the converter to obtain multiple component feature vectors; S130, the multiple component feature vectors are arranged into a two-dimensional matrix and then passed through the first volume as a feature extractor. Accumulate the neural network to obtain the component feature matrix; S140, construct a relationship between the component feature vectors corresponding to each component based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid.
  • Conjunctive adjacency matrix and disjunctive adjacency matrix wherein the conjunctive logic operation rule represents the parallel relationship between rules, and the disjunctive logic operation represents the substitution relationship between rules;
  • S150 convert the conjunctive adjacency matrix and the disjunctive adjacency matrix through the second convolutional neural network to obtain the first feature matrix and the second feature matrix;
  • S160 calculate the weighted sum by position between the first feature matrix and the second feature matrix to Obtain the class adjacency matrix;
  • S170 fuse the component feature matrix and the class adjacency matrix to obtain the recipe feature matrix;
  • S180 pass the acquired image of the object to be stripped through the third convolutional neural network as a feature extractor to obtain Surface state feature matrix;
  • S190 calculate the transfer matrix of the formula feature matrix relative to the surface state feature matrix as a classification feature matrix;
  • S200 pass the classification feature matrix through a classifier to obtain a classification result, the classification
  • the results are used to indicate whether the mass percentage of poly
  • FIG. 5 illustrates a schematic architectural diagram of an automatic batching method for film stripping liquid production according to an embodiment of the present application.
  • the obtained mass percentage of each component of the peeling liquid in the batching (for example, as shown in Figure P shown in Figure 5) is passed through a transformer-based context encoder (e.g., E shown in Figure 5) to obtain multiple component feature vectors (e.g., VF1 shown in Figure 5); then, The plurality of component feature vectors are arranged into a two-dimensional matrix (for example, M as shown in Figure 5) and then passed through the first convolutional neural network as a feature extractor (for example, CNN1 as shown in Figure 5 ) to obtain the component characteristic matrix (for example, MC as shown in Figure 5); then, each group is constructed based on the logical operation rules of conjunction and disjunction between the various components of the peeling liquid.
  • a transformer-based context encoder e.g., E shown in Figure 5
  • component feature vectors e.g., VF1 shown in
  • step S110 and step S120 the mass percentage of each component of the film stripping liquid in the ingredients is obtained, wherein the components of the film stripping liquid include complex alkali, polyether surfactant and water,
  • the composite alkali includes Ca(OH)2, activated white mud, diatomaceous earth and activated carbon
  • the mass percentage of each component of the film stripping liquid in the ingredients is passed through a converter-based context encoder to obtain Multiple component eigenvectors.
  • the components of the film stripping liquid include complex alkali, polyether surface activity agent and water, and the composite alkali includes Ca(OH) 2 , activated white mud, diatomaceous earth and activated carbon.
  • step S130 the plurality of component feature vectors are arranged into a two-dimensional matrix and then passed through the first convolutional neural network as a feature extractor to obtain a component feature matrix. That is, in the technical solution of the present application, further, the plurality of component feature vectors are arranged into a two-dimensional matrix to integrate the global feature information of the mass percentage of each component of the peeling liquid in the ingredients. , and use the convolutional neural network model with excellent performance in implicit correlation feature extraction to perform feature mining on it, thereby obtaining the component feature matrix.
  • the relationship between the component feature vectors corresponding to the respective components is constructed based on the logical operation rules of conjunction and disjunction between the respective components of the film stripping liquid.
  • Conjunctive adjacency matrix and disjunctive adjacency matrix wherein the conjunctive logic operation rule represents the parallel relationship between rules, the disjunctive logic operation represents the substitution relationship between rules, and the conjunctive adjacency matrix and
  • the disjunctive adjacency matrix is passed through a second convolutional neural network to obtain a first feature matrix and a second feature matrix.
  • the conjunctive adjacency matrix and disjunction between the component feature vectors corresponding to the respective components are constructed.
  • Adjacency matrix here, the conjunctive logic operation rules represent the parallel relationship between the rules, and the disjunctive logic operation represents the substitution relationship between the rules.
  • feature extraction is performed on the conjunctive adjacency matrix and the disjunctive adjacency matrix through a second convolutional neural network to dig out the hidden conjunctions and disjunctions between the various components of the peeling fluid.
  • Logical operation rule characteristics thereby obtaining the first characteristic matrix and the second characteristic matrix.
  • a position-weighted sum between the first feature matrix and the second feature matrix is calculated to obtain a class adjacency matrix, and the component feature matrix and the Class adjacency matrix to obtain the recipe feature matrix. That is to say, in the technical solution of the present application, it is further possible to calculate the position-weighted sum between the first characteristic matrix and the second characteristic matrix to fuse the differences between the various components of the peeling liquid. Conjunctive logic associates features and disjunctive logic associates features to obtain adjacency-like matrices.
  • the component feature matrix and the class adjacency matrix are fused to obtain a recipe feature matrix, and then the global correlation features of the mass percentage of each component of the peeling liquid in the ingredients and the peeling liquid are merged
  • the implicit logical association characteristics between the various components of the liquid are used to improve the accuracy of subsequent classification.
  • step S180 the acquired image of the object to be peeled is passed through a third convolutional neural network as a feature extractor to obtain a surface state feature matrix.
  • a third convolutional neural network as a feature extractor for feature mining to extract local high-dimensional hidden features of the image of the object to be stripped, thereby obtaining a surface state feature matrix.
  • a transfer matrix of the formula feature matrix relative to the surface state feature matrix is calculated as a classification feature matrix. It should be understood that, further, by calculating the transfer matrix of the formula characteristic matrix relative to the surface state characteristic matrix, so as to fuse the characteristic information of the two for classification, the mass percentage representing the polyether surfactant can be obtained. Classification results that should be increased or decreased. However, when calculating the transfer matrix of the recipe feature matrix relative to the surface state feature matrix, since the recipe feature matrix incorporates the logical adjacency and disjunction relationships of the adjacency-like matrix, the recipe feature matrix Departure from parameter correlation semantics in feature distribution.
  • the transfer matrix of the recipe feature matrix relative to the surface state feature matrix is directly calculated, it may be that the recipe feature matrix and the surface state feature matrix respectively reside in the entire high dimension due to anisotropy.
  • a narrow subset of the feature space causes the solution space of the transfer matrix to degenerate and lack continuity.
  • it increases the difficulty of calculating the transfer matrix (for example, the iterative process of matrix inversion is difficult to converge), and on the other hand, it affects the classification performance of the transfer matrix. . Therefore, in the technical solution of the present application, preferably before calculating the transfer matrix, each eigenvalue of the formula characteristic matrix is first compared to the search space isotropic based on the surface state characteristic matrix.
  • the feature distribution of the recipe feature matrix can be transferred to a representation space that is isotropic and differentiated from the surface state feature matrix, thereby simplifying the transfer. While calculating the matrix, the distribution continuity of the feature representation of the transfer matrix is enhanced, its classification performance is improved, and the accuracy of the classification is improved.
  • the classification feature matrix is passed through a classifier to obtain a classification result, which is used to indicate that the mass percentage of the polyether surfactant should be increased or decreased.
  • the classifier is used to process 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 automatic batching method for peeling liquid production based on the embodiment of the present application has been clarified, which uses a convolutional neural network model based on deep learning to mine the surface state characteristics of the image of the object to be peeled, and is based on The global correlation characteristics of the mass percentage of each component of the peeling liquid in the ingredients and the logical correlation characteristics between the various components of the peeling liquid are used to assist in the adjustment of the mass percentage of the polyether surfactant, thereby obtaining The optimal formula adapted to specific peeling objects to obtain better peeling effects and ensure the quality of PCB boards.
  • 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)
  • General Health & Medical Sciences (AREA)
  • Medicinal Chemistry (AREA)
  • Crystallography & Structural Chemistry (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Health & Medical Sciences (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computing Systems (AREA)
  • Theoretical Computer Science (AREA)
  • Image Analysis (AREA)
  • Manufacture Of Macromolecular Shaped Articles (AREA)

Abstract

本申请涉及智能生产领域中的智能配料，其具体地公开了一种用于剥膜液生产的自动配料系统及其配料方法，通过基于深度学习的卷积神经网络模型来挖掘出待剥离对象图像的表面状态特征，并且基于在配料中的剥膜液的各个组分的质量百分比的全局关联特征以及所述剥膜液的各个组分之间的逻辑关联特征来辅助进行聚醚表面活性剂的质量百分比的调整，进而来获得适配于特定剥离对象的较优配方，以得到更好的剥膜效果，保证PCB板的质量。

Description

用于剥膜液生产的自动配料系统及其配料方法 技术领域
本发明涉及智能生产领域中的智能配料，且更为具体地，涉及一种用于剥膜液生产的自动配料系统及其配料方法。
背景技术
印刷电路板(Printed circuit board，PCB)制备过程中，蚀刻后的抗蚀刻膜是否能够完全去除，直接影响着后工序的进行和PCB板的质量，因此，剥膜工序是PCB制备过程中的一道关键工序。
目前，一般使用氢氧化钠(NaOH)溶液作为剥膜液，剥膜处理后产生的废液的化学需氧量(Chemical Oxygen Demand，COD)较高，废液后处理成本较高。
专利号CN110331048B揭露了一种环保型剥膜液，其包括所述环保剥膜液的组分以质量百分比计为：复合碱35％～45％，聚醚表面活性剂3％～5％，水50％～62％；其中，所述复合碱的主要成分包括Ca(OH) 2、活性白泥、硅藻土和活性碳；所述聚醚表面活性剂为脂肪醇聚氧乙烯醚。
但是，在实际产业中，由于剥膜液所要剥离的对象不同，即便是都是PCB线路，但用于不同电气设备的PCB板的表面状态拥有不同的特性，因此，期待一种用于剥膜液生产的自动配料系统，以在配料时考虑待剥离对象的表面状态特征来获得适配于特定剥离对象的较优配方。
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剥膜液生产的自动配料系统及其配料方法，其通过基于深度学习的卷积神经网络模型来挖掘出待剥离对象图像的表面状态特征，并且基于在配料中的剥膜液的各个组分的质量百分比的全局关联特征以及所述剥膜液的各个组分之间的逻辑关联特征来辅助进行聚醚表面活性剂的质量百分比的调整，进而来获得适配于特定剥离对象的较优配方，以得到更好的剥膜效果，保证PCB板的质量。
根据本申请的一个方面，提供了一种用于剥膜液生产的自动配料系统，其包括：配料数据采集模块，用于获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；组分质量百分比编码模块，用于将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；组分质量百分比关联编码模块，用于将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；规则运算模块，用于基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；逻辑矩阵编码模块，用于将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；类邻接矩阵构建模块，用于计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合模块，用于融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；待剥离对象编码模块，用于将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；响应模块，用于计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵；以及配料控制模块，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聚醚表面活性剂的质量百分比应增大或应减小。
在上述用于剥膜液生产的自动配料系统中，所述组分质量百分比关联编码模块，包括：输入向量转化单元，用于使用所述基于转换器的上下文编码器的嵌入层分别将所述在配料中的剥膜液的各个组分的质量百分比转化为输入向量以获得输入向量的序列；以及，上下文编码单元，用于使用所述基于转换器的上下文编码器的转换器对所述输入向量的序列进行基于全局的上下文语义编码以获得所述多个组分特征向量。
在上述用于剥膜液生产的自动配料系统中，所述规则运算模块，包括：合取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的合取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述合取邻接矩阵；其中，所述公式为：
Figure PCTCN2022119832-appb-000001
其中，
Figure PCTCN2022119832-appb-000002
是合取矩阵，用于表示相应的一对规则构成合取范式时矩阵位置取1，而非合取范式时矩阵位置取0；以及，析取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的析取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述析取邻接矩阵；其中，所述 公式为：其中，
Figure PCTCN2022119832-appb-000003
是析取邻接矩阵，用于表示相应的一对规则构成析取范式时矩阵位置取1，而非析取范式时矩阵位置取0。
在上述用于剥膜液生产的自动配料系统中，所述逻辑矩阵编码模块，进一步用于：所述第二卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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第一特征矩阵和第二特征矩阵，所述第二卷积神经网络的第一层的输入为所述合取邻接矩阵和所述析取邻接矩阵。
在上述用于剥膜液生产的自动配料系统中，所述类邻接矩阵构建模块，进一步用于：以如下公式计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得所述类邻接矩阵；其中，所述公式为：M s＝αM 1+βM 2其中，M s为所述类邻接矩阵，M 1为所述第一特征矩阵，M 2为所述第二特征矩阵，“+”表示所述第一特征矩阵和所述第二特征矩阵相对应位置处的元素相加，α和β为用于控制所述类邻接矩阵中所述第一特征矩阵和所述第二特征矩阵之间的平衡的加权参数。
在上述用于剥膜液生产的自动配料系统中，所述待剥离对象编码模块，进一步用于：所述第三卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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，所述第三卷积神经网络的最后一层的输出为所述表面状态特征矩阵，所述第三卷积神经网络的第一层的输入为所述待剥离对象的图像。
在上述用于剥膜液生产的自动配料系统中，所述响应模块，包括：校正单元，用于基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵；以及，转移单元，用于以如下公式计算所述校正后配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为所述分类特征矩阵，其中，所述公式为：S＝T*F，其中F表示所述校正后配方特征矩阵，T表示所述分类特征矩阵，S表示所述表面状态特征矩阵。
在上述用于剥膜液生产的自动配料系统中，所述校正单元，进一步用于：基于所述表面状态特征矩阵以如下公式对所述配方特征矩阵中各个位置的特征值进行校正以得到所述校正后配方特征矩阵；其中，所述公式为：f i∈max-norm(M 1)and f j∈max-norm(M 2)
其中M 1表示所述配方特征矩阵，M 2表示所述表面状态特征矩阵，f i表示所述配方特征矩阵的相应位置的归一化到[0,1]区间内的特征值，f j表示所述表面状态特征矩阵的相应位置的归一化到[0,1]区间内的特征值，d(f i,f j)表示所述配方特征矩阵中各个位置的特征值与所述表面状态特征矩阵中各个位置的特征值之间的距离，max-norm(·)表示最大值归一化，且ρ是控制超参数。
在上述用于剥膜液生产的自动配料系统中，所述配料控制模块，进一步用于：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。
根据本申请的另一方面，一种用于剥膜液生产的自动配料方法，其包括：获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；计算所述配方特征矩阵相对于所述表面状态特征矩阵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聚醚表面活性剂的质量百分比应增大或应减小。
在上述用于剥膜液生产的自动配料方法中，将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量，包括：使用所述基于转换器的上下文编码器的嵌入层分别将所述在配料中的剥膜液的各个组分的质量百分比转化为输入向量以获得输入向量的序列；以及，使用所述基于转换器的上下文编码器的转换器对所述输入向量的序列进行基于全局的上下文语义编码以获得所述多个组分特征向量。
在上述用于剥膜液生产的自动配料方法中，基于所述剥膜液的各个组分之间的合取和析取的逻辑 运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，包括：基于所述剥膜液的各个组分之间的合取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述合取邻接矩阵；其中，所述公式为：
Figure PCTCN2022119832-appb-000004
其中，
Figure PCTCN2022119832-appb-000005
是合取矩阵，用于表示相应的一对规则构成合取范式时矩阵位置取1，而非合取范式时矩阵位置取0；以及，基于所述剥膜液的各个组分之间的析取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述析取邻接矩阵；其中，所述公式为：
其中，
Figure PCTCN2022119832-appb-000006
是析取邻接矩阵，用于表示相应的一对规则构成析取范式时矩阵位置取1，而非析取范式时矩阵位置取0。
在上述用于剥膜液生产的自动配料方法中，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵，包括：所述第二卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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第一特征矩阵和第二特征矩阵，所述第二卷积神经网络的第一层的输入为所述合取邻接矩阵和所述析取邻接矩阵。
在上述用于剥膜液生产的自动配料方法中，计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵，包括：以如下公式计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得所述类邻接矩阵；其中，所述公式为：
M s＝αM 1+βM 2
其中，M s为所述类邻接矩阵，M 1为所述第一特征矩阵，M 2为所述第二特征矩阵，“+”表示所述第一特征矩阵和所述第二特征矩阵相对应位置处的元素相加，α和β为用于控制所述类邻接矩阵中所述第一特征矩阵和所述第二特征矩阵之间的平衡的加权参数。
在上述用于剥膜液生产的自动配料方法中，将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵，包括：所述第三卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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，所述第三卷积神经网络的最后一层的输出为所述表面状态特征矩阵，所述第三卷积神经网络的第一层的输入为所述待剥离对象的图像。
在上述用于剥膜液生产的自动配料方法中，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵，包括：基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵；以及，以如下公式计算所述校正后配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为所述分类特征矩阵，其中，所述公式为：S＝T*F，其中F表示所述校正后配方特征矩阵，T表示所述分类特征矩阵，S表示所述表面状态特征矩阵。
在上述用于剥膜液生产的自动配料方法中，基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵，包括：基于所述表面状态特征矩阵以如下公式对所述配方特征矩阵中各个位置的特征值进行校正以得到所述校正后配方特征矩阵；其中，所述公式为：f i∈max-norm(M 1)and f j∈max-norm(M 2)
其中M 1表示所述配方特征矩阵，M 2表示所述表面状态特征矩阵，f i表示所述配方特征矩阵的相应位置的归一化到[0,1]区间内的特征值，f j表示所述表面状态特征矩阵的相应位置的归一化到[0,1]区间内的特征值，d(f i,f j)表示所述配方特征矩阵中各个位置的特征值与所述表面状态特征矩阵中各个位置的特征值之间的距离，max-norm(·)表示最大值归一化，且ρ是控制超参数。
在上述用于剥膜液生产的自动配料方法中，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，本申请提供的用于剥膜液生产的自动配料系统及其配料方法，其通过基于深度学习的卷积神经网络模型来挖掘出待剥离对象图像的表面状态特征，并且基于在配料中的剥膜液的各个组分的质量百分比的全局关联特征以及所述剥膜液的各个组分之间的逻辑关联特征来辅助进行聚醚表面活性剂的质量百分比的调整，进而来获得适配于特定剥离对象的较优配方，以得到更好的剥膜效果，保证PCB板的质量。
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剥膜液生产的自动配料系统的应用场景图。
图2为根据本申请实施例的用于剥膜液生产的自动配料系统的框图。
图3为根据本申请实施例的用于剥膜液生产的自动配料系统中响应模块的框图。
图4为根据本申请实施例的用于剥膜液生产的自动配料方法的流程图。
图5为根据本申请实施例的用于剥膜液生产的自动配料方法的架构示意图。
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，印刷电路板(Printed circuit board，PCB)制备过程中，蚀刻后的抗蚀刻膜是否能够完全去除，直接影响着后工序的进行和PCB板的质量，因此，剥膜工序是PCB制备过程中的一道关键工序。
目前，一般使用氢氧化钠(NaOH)溶液作为剥膜液，剥膜处理后产生的废液的化学需氧量(Chemical Oxygen Demand，COD)较高，废液后处理成本较高。
专利号CN110331048B揭露了一种环保型剥膜液，其包括所述环保剥膜液的组分以质量百分比计为：复合碱35％～45％，聚醚表面活性剂3％～5％，水50％～62％；其中，所述复合碱的主要成分包括Ca(OH) 2、活性白泥、硅藻土和活性碳；所述聚醚表面活性剂为脂肪醇聚氧乙烯醚。
但是，在实际产业中，由于剥膜液所要剥离的对象不同，即便是都是PCB线路，但用于不同电气设备的PCB板的表面状态拥有不同的特性，因此，期待一种用于剥膜液生产的自动配料系统，以在配料时考虑待剥离对象的表面状态特征来获得适配于特定剥离对象的较优配方。
相应地，本申请发明人考虑到由于在实际的剥膜工序中，不同电气设备的PCB板的表面状态拥有不同的特性，而所述聚醚表面活性剂可以显著降低表面的张力，因此若想对于不同的剥膜对象都能够得到更好的剥膜效果，期望在配料时就考虑到剥膜对象的表面状态特征来进行适配于不同剥离对象的较优配方。这本质上是一个分类的问题，也就是，基于待剥离对象的图像来表示其表面的状态特征，并利用在配料中的剥膜液的各个组分的质量百分比关联特征和逻辑关联特征来辅助进行所述聚醚表面活性剂的质量百分比的分类调整，进而保证PCB板的质量。
具体地，在本申请的技术方案中，首先，获取在配料中的剥膜液的各个组分的质量百分比，这里，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH) 2、活性白泥、硅藻土和活性碳。应可以理解，考虑到所述在配料中的剥膜液的各个组分的质量百分比数据之间存在着关联，因此使用基于转换器的上下文编码器来对其进行编码，以提取出所述在配料中的剥膜液的各个组分的质量百分比之间的基于全局的高维语义特征以更适于表征所述剥膜液的本质配比特征，从而得到多个组分特征向量。
进一步地，将所述多个组分特征向量排列为二维矩阵以整合所述在配料中的剥膜液的各个组分的质量百分比的全局特征信息，并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来对其进行特征挖掘，从而得到组分特征矩阵。
考虑到规则之间的逻辑运算通常包括合取和析取，分别以符号∧和∨表示，用于表示规则之间的并列或者替代关系，也就是“并且”和“或者”的含义。而针对所述剥膜液的各个组分之间也存在这种关系，例如所述剥膜液的各个组分之间需要相互配合反应使用的组分就是“并且”的关系，而所述剥膜液的各个组分中可以使用别的成分替代的组分就是“或者”的关系。因此，在本申请的技术方案中，还需要基于这种逻辑运算规则关系来进行所述剥离液的各个组分百分比之间的隐含关联特征的辅助表征。也就是，首先，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，这里，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系。然后，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络中进行特征提取，以挖掘出所述剥膜液的各个组分之间的隐藏的合取和析取的逻辑运算规则特征，从而获得第一特征矩阵和第二特征矩阵。
这样，就可以计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和，来融合所述剥膜液的各个组分之间的合取逻辑关联特征和析取逻辑关联特征，以获得类邻接矩阵。然后，再融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵，进而融合了所述在配料中的剥膜液的各个组分 的质量百分比的全局关联特征和所述剥膜液的各个组分之间的隐含逻辑关联特征，以提高后续分类的准确性。
若想根据实际的情况来进行所述聚醚表面活性剂的质量百分比的调整，还需要通过摄像头获取待剥离对象的图像。然后，将所述待剥离对象的图像通过作为特征提取器的第三卷积神经网络中进行特征挖掘，以提取出所述待剥离对象的图像的局部高维隐含特征，从而得到表面状态特征矩阵。
应可以理解，进一步地，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵，以融合两者的特征信息来进行分类，就可以获得用于表示聚醚表面活性剂的质量百分比应增大或应减小的分类结果。但是，在计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵时，由于所述配方特征矩阵融合了所述类邻接矩阵的逻辑邻取和析取关系，使得所述配方特征矩阵在特征分布上偏离了参数关联语义。
由此，如果直接计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵，可能会由于所述配方特征矩阵与所述表面状态特征矩阵由于各向异性而分别驻留在整个高维特征空间的一个狭窄子集中，导致转移矩阵的解空间退化而缺乏连续性，一方面提高转移矩阵的计算难度(例如，矩阵求逆的迭代过程难以收敛)，另一方面影响转移矩阵的分类性能。
因此，优选地在计算转移矩阵之前，首先基于所述表面状态特征矩阵对所述配方特征矩阵的各个特征值进行对比搜索空间同向化，即：f i∈max-norm(M 1)and f j∈max-norm(M 2)
其中M 1表示所述配方特征矩阵，M 2表示所述表面状态特征矩阵，f i表示所述配方特征矩阵的相应位置的归一化到[0,1]区间内的特征值，f j表示所述表面状态特征矩阵的相应位置的归一化到[0,1]区间内的特征值，d(f i,f j)表示所述配方特征矩阵中各个位置的特征值与所述表面状态特征矩阵中各个位置的特征值之间的距离，max-norm(·)表示最大值归一化，且ρ是控制超参数，例如初始设置为矩阵M 1和M 2之间的距离。
这样，通过以上对比搜索空间同向化，可以将所述配方特征矩阵的特征分布转移到与所述表面状态特征矩阵各向同性且有区分度的表示空间，从而在简化转移矩阵的计算的同时，增强了转移矩阵的特征表示的分布连续性，提高了其分类性能，进而提高了分类的准确性。
基于此，本申请提出了一种用于剥膜液生产的自动配料系统，其包括：配料数据采集模块，用于获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；组分质量百分比编码模块，用于将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；组分质量百分比关联编码模块，用于将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；规则运算模块，用于基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；逻辑矩阵编码模块，用于将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；类邻接矩阵构建模块，用于计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合模块，用于融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；待剥离对象编码模块，用于将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；响应模块，用于计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵；以及，配料控制模块，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聚醚表面活性剂的质量百分比应增大或应减小。
图1图示了根据本申请实施例的用于剥膜液生产的自动配料系统的应用场景图。如图1所示，在该应用场景中，首先，通过云存储端(例如，如图1中所示意的D)获取在配料中的剥膜液(例如，如图1中所示意的T)的各个组分的质量百分比，这里，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH) 2、活性白泥、硅藻土和活性碳，并且基于云存储端中所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，以及通过摄像头(例如，如图1中所示意的C)获取待剥离对象(例如，如图1中所示意的P)的图像。特别地，在该应用场景中，所述待剥离对象可以为PCB线路。然后，将获得的所述剥膜液的各个组分的质量百分比、所述合取邻接矩阵和所述析取邻接矩阵以及所述待剥离对象的图像输入至部署有用于剥膜液生产的自动配料算法的服务器中(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剥膜液生产的自动配料算法对所述剥膜液的各个组分的质量百分比、所述合取邻接矩阵和所述析取邻接矩阵以及所述待剥离对象的图像进行处理，以生成用于表示聚醚表面活性剂的质量百分比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剥膜液生产的自动配料系统的框图。如图2所示，根据本申请实施例的用于剥膜液生产的自动配料系统200，包括：配料数据采集模块210，用于获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；组分质量百分比编码模块220，用于将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；组分质量百分比关联编码模块230，用于将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；规则运算模块240，用于基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；逻辑矩阵编码模块250，用于将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；类邻接矩阵构建模块260，用于计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合模块270，用于融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；待剥离对象编码模块280，用于将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；响应模块290，用于计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵；以及，配料控制模块300，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聚醚表面活性剂的质量百分比应增大或应减小。
具体地，在本申请实施例中，所述配料数据采集模块210和所述组分质量百分比编码模块220，用于获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH) 2、活性白泥、硅藻土和活性碳，并将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量。如前所述，由于在实际的剥膜工序中，不同电气设备的PCB板的表面状态拥有不同的特性，而所述聚醚表面活性剂可以显著降低表面的张力，因此若想对于不同的剥膜对象都能够得到更好的剥膜效果，在本申请的技术方案中，期望在配料时就考虑到剥膜对象的表面状态特征来进行适配于不同剥离对象的较优配方。这本质上是一个分类的问题，也就是，基于待剥离对象的图像来表示其表面的状态特征，并利用在配料中的剥膜液的各个组分的质量百分比关联特征和逻辑关联特征来辅助进行所述聚醚表面活性剂的质量百分比的分类调整，进而保证PCB板的质量。
也就是，具体地，在本申请的技术方案中，首先，获取在配料中的剥膜液的各个组分的质量百分比，这里，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH) 2、活性白泥、硅藻土和活性碳。应可以理解，考虑到所述在配料中的剥膜液的各个组分的质量百分比数据之间存在着关联，因此，在本申请的技术方案中，进一步使用基于转换器的上下文编码器来对其进行编码，以提取出所述在配料中的剥膜液的各个组分的质量百分比之间的基于全局的高维语义特征以更适于表征所述剥膜液的本质配比特征，从而得到多个组分特征向量。
更具体地，在本申请实施例中，所述组分质量百分比关联编码模块，包括：输入向量转化单元，用于使用所述基于转换器的上下文编码器的嵌入层分别将所述在配料中的剥膜液的各个组分的质量百分比转化为输入向量以获得输入向量的序列；以及，上下文编码单元，用于使用所述基于转换器的上下文编码器的转换器对所述输入向量的序列进行基于全局的上下文语义编码以获得所述多个组分特征向量。
具体地，在本申请实施例中，所述组分质量百分比关联编码模块230，用于将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵。也就是，在本申请的技术方案中，进一步地，将所述多个组分特征向量排列为二维矩阵以整合所述在配料中的剥膜液的各个组分的质量百分比的全局特征信息，并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来对其进行特征挖掘，从而得到组分特征矩阵。
具体地，在本申请实施例中，所述规则运算模块240和所述逻辑矩阵编码模块250，用于基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系，并将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵。应可以理解，考虑到规则之间的逻辑运算通常包括合取和析取，分别以符号∧和∨表示，用于表示规则之间的并列或者替代关系，也就是“并且”和“或者”的含义。而针对所述剥膜液的各个组分之间也存在这种关系，例如所述剥膜液的各个组分之间需要相互配合反应使用的组分就是“并且”的关系，而所述剥膜液的各个组分中可以使用别的成分替代的组 分就是“或者”的关系。因此，在本申请的技术方案中，还需要基于这种逻辑运算规则关系来进行所述剥离液的各个组分百分比之间的隐含关联特征的辅助表征。也就是，首先，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，这里，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系。然后，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络中进行特征提取，以挖掘出所述剥膜液的各个组分之间的隐藏的合取和析取的逻辑运算规则特征，从而获得第一特征矩阵和第二特征矩阵。
相应地，在一个具体示例中，所述逻辑矩阵编码模块，进一步用于：所述第二卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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第一特征矩阵和第二特征矩阵，所述第二卷积神经网络的第一层的输入为所述合取邻接矩阵和所述析取邻接矩阵。
更具体地，在本申请实施例中，所述规则运算模块，包括：合取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的合取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述合取邻接矩阵；其中，所述公式为：
Figure PCTCN2022119832-appb-000007
其中，
Figure PCTCN2022119832-appb-000008
是合取矩阵，用于表示相应的一对规则构成合取范式时矩阵位置取1，而非合取范式时矩阵位置取0；以及，析取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的析取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述析取邻接矩阵；其中，所述公式为：其中，
Figure PCTCN2022119832-appb-000009
是析取邻接矩阵，用于表示相应的一对规则构成析取范式时矩阵位置取1，而非析取范式时矩阵位置取0。
具体地，在本申请实施例中，所述类邻接矩阵构建模块260和所述融合模块270，用于计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵，并融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵。也就是，在本申请的技术方案中，进一步就可以计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和，来融合所述剥膜液的各个组分之间的合取逻辑关联特征和析取逻辑关联特征，以获得类邻接矩阵。然后，再融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵，进而融合了所述在配料中的剥膜液的各个组分的质量百分比的全局关联特征和所述剥膜液的各个组分之间的隐含逻辑关联特征，以提高后续分类的准确性。
更具体地，在本申请实施例中，所述类邻接矩阵构建模块，进一步用于：以如下公式计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得所述类邻接矩阵；其中，所述公式为：
M s＝αM 1+βM 2
其中，M s为所述类邻接矩阵，M 1为所述第一特征矩阵，M 2为所述第二特征矩阵，“+”表示所述第一特征矩阵和所述第二特征矩阵相对应位置处的元素相加，α和β为用于控制所述类邻接矩阵中所述第一特征矩阵和所述第二特征矩阵之间的平衡的加权参数。
具体地，在本申请实施例中，所述待剥离对象编码模块280，用于将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵。应可以理解，若想根据实际的情况来进行所述聚醚表面活性剂的质量百分比的调整，还需要通过摄像头获取待剥离对象的图像。然后，将所述待剥离对象的图像通过作为特征提取器的第三卷积神经网络中进行特征挖掘，以提取出所述待剥离对象的图像的局部高维隐含特征，从而得到表面状态特征矩阵。
更具体地，在本申请实施例中，所述待剥离对象编码模块，进一步用于：所述第三卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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，所述第三卷积神经网络的最后一层的输出为所述表面状态特征矩阵，所述第三卷积神经网络的第一层的输入为所述待剥离对象的图像。
具体地，在本申请实施例中，所述响应模块290，用于计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵。应可以理解，进一步地，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵，以融合两者的特征信息来进行分类，就可以获得用于表示聚醚表面活性剂的质量百分比应增大或应减小的分类结果。但是，在计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵时，由于所述配方特征矩阵融合了所述类邻接矩阵的逻辑邻取和析取关系，使得所述配方特征矩阵在特征分布上偏离了参数关联语义。由此，如果直接计算所述配方特征矩阵相对 于所述表面状态特征矩阵的转移矩阵，可能会由于所述配方特征矩阵与所述表面状态特征矩阵由于各向异性而分别驻留在整个高维特征空间的一个狭窄子集中，导致转移矩阵的解空间退化而缺乏连续性，一方面提高转移矩阵的计算难度(例如，矩阵求逆的迭代过程难以收敛)，另一方面影响转移矩阵的分类性能。因此，在本申请的技术方案中，优选地在计算所述转移矩阵之前，首先基于所述表面状态特征矩阵对所述配方特征矩阵的各个特征值进行对比搜索空间同向化。也就是，通过以上所述对比搜索空间同向化，可以将所述配方特征矩阵的特征分布转移到与所述表面状态特征矩阵各向同性且有区分度的表示空间，从而在简化所述转移矩阵的计算的同时，增强了所述转移矩阵的特征表示的分布连续性，提高了其分类性能，进而提高了分类的准确性。
更具体地，在本申请的实施例中，所述响应模块，包括：首先，基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵。相应地，在一个具体示例中，基于所述表面状态特征矩阵以如下公式对所述配方特征矩阵中各个位置的特征值进行校正以得到所述校正后配方特征矩阵；
其中，所述公式为：f i∈max-norm(M 1)and f j∈max-norm(M 2)
其中M 1表示所述配方特征矩阵，M 2表示所述表面状态特征矩阵，f i表示所述配方特征矩阵的相应位置的归一化到[0,1]区间内的特征值，f j表示所述表面状态特征矩阵的相应位置的归一化到[0,1]区间内的特征值，d(f i,f j)表示所述配方特征矩阵中各个位置的特征值与所述表面状态特征矩阵中各个位置的特征值之间的距离，max-norm(·)表示最大值归一化，且ρ是控制超参数，例如初始设置为矩阵M 1和M 2之间的距离。
然后，以如下公式计算所述校正后配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为所述分类特征矩阵，其中，所述公式为：S＝T*F，其中F表示所述校正后配方特征矩阵，T表示所述分类特征矩阵，S表示所述表面状态特征矩阵。
图3图示了根据本申请实施例的用于剥膜液生产的自动配料系统中响应模块的框图。如图3所示，所述响应模块290，包括：校正单元291，用于基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵；以及，转移单元292，用于以如下公式计算所述校正后配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为所述分类特征矩阵，其中，所述公式为：S＝T*F，其中F表示所述校正后配方特征矩阵，T表示所述分类特征矩阵，S表示所述表面状态特征矩阵。
具体地，在本申请实施例中，所述配料控制模块300，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聚醚表面活性剂的质量百分比应增大或应减小。相应地，在一个具体示例中，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,Bn):…:(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。
综上，基于本申请实施例的所述用于剥膜液生产的自动配料系统200被阐明，其通过基于深度学习的卷积神经网络模型来挖掘出待剥离对象图像的表面状态特征，并且基于在配料中的剥膜液的各个组分的质量百分比的全局关联特征以及所述剥膜液的各个组分之间的逻辑关联特征来辅助进行聚醚表面活性剂的质量百分比的调整，进而来获得适配于特定剥离对象的较优配方，以得到更好的剥膜效果，保证PCB板的质量。
如上所述，根据本申请实施例的用于剥膜液生产的自动配料系统200可以实现在各种终端设备中，例如用于剥膜液生产的自动配料算法的服务器等。在一个示例中，根据本申请实施例的用于剥膜液生产的自动配料系统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，该用于剥膜液生产的自动配料系统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，该用于剥膜液生产的自动配料系统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，该用于剥膜液生产的自动配料系统200与该终端设备也可以是分立的设备，并且该用于剥膜液生产的自动配料系统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4图示了用于剥膜液生产的自动配料方法的流程图。如图4所示，根据本申请实施例的用于剥膜液生产的自动配料方法，包括步骤：S110，获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；S120，将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；S130，将所述多个组分特征向量排列为二维矩阵后通过作为特征提取 器的第一卷积神经网络以得到组分特征矩阵；S140，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；S150，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；S160，计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；S170，融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；S180，将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；S190，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵；以及，S200，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聚醚表面活性剂的质量百分比应增大或应减小。
图5图示了根据本申请实施例的用于剥膜液生产的自动配料方法的架构示意图。如图5所示，在所述用于剥膜液生产的自动配料方法的网络架构中，首先，将获得的所述在配料中的剥膜液的各个组分的质量百分比(例如，如图5中所示意的P)通过基于转换器的上下文编码器(例如，如图5中所示意的E)以得到多个组分特征向量(例如，如图5中所示意的VF1)；接着，将所述多个组分特征向量排列为二维矩阵(例如，如图5中所示意的M)后通过作为特征提取器的第一卷积神经网络(例如，如图5中所示意的CNN1)以得到组分特征矩阵(例如，如图5中所示意的MC)；然后，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵(例如，如图5中所示意的M1)和析取邻接矩阵(例如，如图5中所示意的M2)；接着，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络(例如，如图5中所示意的CNN2)以获得第一特征矩阵(例如，如图5中所示意的MF1)和第二特征矩阵(例如，如图5中所示意的MF2)；然后，计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵(例如，如图5中所示意的MF)；接着，融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵(例如，如图5中所示意的MA)；然后，将获取的待剥离对象的图像(例如，如图5中所示意的Q)通过作为特征提取器的第三卷积神经网络(例如，如图5中所示意的CNN3)以得到表面状态特征矩阵(例如，如图5中所示意的MS)；接着，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵(例如，如图5中所示意的ML)；以及，最后，将所述分类特征矩阵通过分类器(例如，如图5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聚醚表面活性剂的质量百分比应增大或应减小。
更具体地，在步骤S110和步骤S120中，获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳，并将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量。应可以理解，由于在实际的剥膜工序中，不同电气设备的PCB板的表面状态拥有不同的特性，而所述聚醚表面活性剂可以显著降低表面的张力，因此若想对于不同的剥膜对象都能够得到更好的剥膜效果，在本申请的技术方案中，期望在配料时就考虑到剥膜对象的表面状态特征来进行适配于不同剥离对象的较优配方。这本质上是一个分类的问题，也就是，基于待剥离对象的图像来表示其表面的状态特征，并利用在配料中的剥膜液的各个组分的质量百分比关联特征和逻辑关联特征来辅助进行所述聚醚表面活性剂的质量百分比的分类调整，进而保证PCB板的质量。
也就是，具体地，在本申请的技术方案中，首先，获取在配料中的剥膜液的各个组分的质量百分比，这里，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH) 2、活性白泥、硅藻土和活性碳。应可以理解，考虑到所述在配料中的剥膜液的各个组分的质量百分比数据之间存在着关联，因此，在本申请的技术方案中，进一步使用基于转换器的上下文编码器来对其进行编码，以提取出所述在配料中的剥膜液的各个组分的质量百分比之间的基于全局的高维语义特征以更适于表征所述剥膜液的本质配比特征，从而得到多个组分特征向量。
更具体地，在步骤S130中，将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵。也就是，在本申请的技术方案中，进一步地，将所述多个组分特征向量排列为二维矩阵以整合所述在配料中的剥膜液的各个组分的质量百分比的全局特征信息，并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来对其进行特征挖掘，从而得到组分特征矩阵。
更具体地，在步骤S140和步骤S150中，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系，并将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵。应可以理解，考虑到规则之间的逻辑运算通常包括合取和析取，分别以符号∧和∨表示，用于表示规则之间的并列或者替代关系，也就是“并且”和“或者”的含义。而针对所述剥膜液的各个组分之间也存在 这种关系，例如所述剥膜液的各个组分之间需要相互配合反应使用的组分就是“并且”的关系，而所述剥膜液的各个组分中可以使用别的成分替代的组分就是“或者”的关系。因此，在本申请的技术方案中，还需要基于这种逻辑运算规则关系来进行所述剥离液的各个组分百分比之间的隐含关联特征的辅助表征。也就是，首先，基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，这里，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系。然后，将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络中进行特征提取，以挖掘出所述剥膜液的各个组分之间的隐藏的合取和析取的逻辑运算规则特征，从而获得第一特征矩阵和第二特征矩阵。
更具体地，在步骤S160和步骤S170中，计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵，并融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵。也就是，在本申请的技术方案中，进一步就可以计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和，来融合所述剥膜液的各个组分之间的合取逻辑关联特征和析取逻辑关联特征，以获得类邻接矩阵。然后，再融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵，进而融合了所述在配料中的剥膜液的各个组分的质量百分比的全局关联特征和所述剥膜液的各个组分之间的隐含逻辑关联特征，以提高后续分类的准确性。
更具体地，在步骤S180中，将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵。应可以理解，若想根据实际的情况来进行所述聚醚表面活性剂的质量百分比的调整，还需要通过摄像头获取待剥离对象的图像。然后，将所述待剥离对象的图像通过作为特征提取器的第三卷积神经网络中进行特征挖掘，以提取出所述待剥离对象的图像的局部高维隐含特征，从而得到表面状态特征矩阵。
更具体地，在步骤S190中，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵。应可以理解，进一步地，计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵，以融合两者的特征信息来进行分类，就可以获得用于表示聚醚表面活性剂的质量百分比应增大或应减小的分类结果。但是，在计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵时，由于所述配方特征矩阵融合了所述类邻接矩阵的逻辑邻取和析取关系，使得所述配方特征矩阵在特征分布上偏离了参数关联语义。由此，如果直接计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵，可能会由于所述配方特征矩阵与所述表面状态特征矩阵由于各向异性而分别驻留在整个高维特征空间的一个狭窄子集中，导致转移矩阵的解空间退化而缺乏连续性，一方面提高转移矩阵的计算难度(例如，矩阵求逆的迭代过程难以收敛)，另一方面影响转移矩阵的分类性能。因此，在本申请的技术方案中，优选地在计算所述转移矩阵之前，首先基于所述表面状态特征矩阵对所述配方特征矩阵的各个特征值进行对比搜索空间同向化。也就是，通过以上所述对比搜索空间同向化，可以将所述配方特征矩阵的特征分布转移到与所述表面状态特征矩阵各向同性且有区分度的表示空间，从而在简化所述转移矩阵的计算的同时，增强了所述转移矩阵的特征表示的分布连续性，提高了其分类性能，进而提高了分类的准确性。
更具体地，在步骤S200中，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聚醚表面活性剂的质量百分比应增大或应减小。相应地，在一个具体示例中，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。
综上，基于本申请实施例的所述用于剥膜液生产的自动配料方法被阐明，其通过基于深度学习的卷积神经网络模型来挖掘出待剥离对象图像的表面状态特征，并且基于在配料中的剥膜液的各个组分的质量百分比的全局关联特征以及所述剥膜液的各个组分之间的逻辑关联特征来辅助进行聚醚表面活性剂的质量百分比的调整，进而来获得适配于特定剥离对象的较优配方，以得到更好的剥膜效果，保证PCB板的质量。
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. 一种用于剥膜液生产的自动配料系统，其特征在于，包括：配料数据采集模块，用于获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；组分质量百分比编码模块，用于将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；组分质量百分比关联编码模块，用于将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；规则运算模块，用于基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；逻辑矩阵编码模块，用于将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；类邻接矩阵构建模块，用于计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合模块，用于融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；待剥离对象编码模块，用于将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；响应模块，用于计算所述配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为分类特征矩阵；以及配料控制模块，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聚醚表面活性剂的质量百分比应增大或应减小。
  2. 根据权利要求1所述的用于剥膜液生产的自动配料系统，其特征在于，所述组分质量百分比关联编码模块，包括：输入向量转化单元，用于使用所述基于转换器的上下文编码器的嵌入层分别将所述在配料中的剥膜液的各个组分的质量百分比转化为输入向量以获得输入向量的序列；以及上下文编码单元，用于使用所述基于转换器的上下文编码器的转换器对所述输入向量的序列进行基于全局的上下文语义编码以获得所述多个组分特征向量。
  3. 根据权利要求2所述的用于剥膜液生产的自动配料系统，其特征在于，所述规则运算模块，包括：合取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的合取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述合取邻接矩阵；其中，所述公式为：
    Figure PCTCN2022119832-appb-100001
    其中，
    Figure PCTCN2022119832-appb-100002
    是合取矩阵，用于表示相应的一对规则构成合取范式时矩阵位置取1，而非合取范式时矩阵位置取0；以及析取邻接矩阵构造单元，用于基于所述剥膜液的各个组分之间的析取的逻辑运算规则以如下公式来构建所述各个组分对应的组分特征向量之间的所述析取邻接矩阵；其中，所述公式为：其中，
    Figure PCTCN2022119832-appb-100003
    是析取邻接矩阵，用于表示相应的一对规则构成析取范式时矩阵位置取1，而非析取范式时矩阵位置取0。
  4. 根据权利要求3所述的用于剥膜液生产的自动配料系统，其特征在于，所述逻辑矩阵编码模块，进一步用于：所述第二卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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第一特征矩阵和第二特征矩阵，所述第二卷积神经网络的第一层的输入为所述合取邻接矩阵和所述析取邻接矩阵。
  5. 根据权利要求4所述的用于剥膜液生产的自动配料系统，其特征在于，所述类邻接矩阵构建模块，进一步用于：以如下公式计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得所述类邻接矩阵；其中，所述公式为：M s＝αM 1+βM 2其中，M s为所述类邻接矩阵，M 1为所述第一特征矩阵，M 2为所述第二特征矩阵，“+”表示所述第一特征矩阵和所述第二特征矩阵相对应位置处的元素相加，α和β为用于控制所述类邻接矩阵中所述第一特征矩阵和所述第二特征矩阵之间的平衡的加权参数。
  6. 根据权利要求5所述的用于剥膜液生产的自动配料系统，其特征在于，所述待剥离对象编码模块，进一步用于：所述第三卷积神经网络的各层在层的正向传递中分别进行：对输入数据进行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通道维度的均值池化以得到池化特征图；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，所述第三卷积神经网络的最后一层的输出为所述表面状态特征矩阵，所述第三卷积神经网络的第一层的输入为所述待剥离对象的图像。
  7. 根据权利要求6所述的用于剥膜液生产的自动配料系统，其特征在于，所述响应模块，包括：校正单元，用于基于所述表面状态特征矩阵对所述配方特征矩阵中各个位置的特征值进行校正以得到校正后配方特征矩阵；以及转移单元，用于以如下公式计算所述校正后配方特征矩阵相对于所述表面状态特征矩阵的转移矩阵作为所述分类特征矩阵，其中，所述公式为：S＝T*F，其中F表示所述校正 后配方特征矩阵，T表示所述分类特征矩阵，S表示所述表面状态特征矩阵。
  8. 根据权利要求7所述的用于剥膜液生产的自动配料系统，其特征在于，所述校正单元，进一步用于：基于所述表面状态特征矩阵以如下公式对所述配方特征矩阵中各个位置的特征值进行校正以得到所述校正后配方特征矩阵；其中，所述公式为：f i∈max-norm(M 1)and f j∈max-norm(M 2)其中M 1表示所述配方特征矩阵，M 2表示所述表面状态特征矩阵，f i表示所述配方特征矩阵的相应位置的归一化到[0,1]区间内的特征值，f j表示所述表面状态特征矩阵的相应位置的归一化到[0,1]区间内的特征值，d(f i,f j)表示所述配方特征矩阵中各个位置的特征值与所述表面状态特征矩阵中各个位置的特征值之间的距离，max-norm(·)表示最大值归一化，且ρ是控制超参数。
  9. 根据权利要求8所述的用于剥膜液生产的自动配料系统，其特征在于，所述配料控制模块，进一步用于：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。
  10. 一种用于剥膜液生产的自动配料方法，其特征在于，包括：获取在配料中的剥膜液的各个组分的质量百分比，其中，所述剥膜液的组分包括复合碱、聚醚表面活性剂和水，所述复合碱包括Ca(OH)2、活性白泥、硅藻土和活性碳；将所述在配料中的剥膜液的各个组分的质量百分比通过基于转换器的上下文编码器以得到多个组分特征向量；将所述多个组分特征向量排列为二维矩阵后通过作为特征提取器的第一卷积神经网络以得到组分特征矩阵；基于所述剥膜液的各个组分之间的合取和析取的逻辑运算规则来构建所述各个组分对应的组分特征向量之间的合取邻接矩阵和析取邻接矩阵，其中，所述合取逻辑运算规则表示规则之间的并列关系，所述析取逻辑运算表示规则之间的替换关系；将所述合取邻接矩阵和所述析取邻接矩阵通过第二卷积神经网络以获得第一特征矩阵和第二特征矩阵；计算所述第一特征矩阵和所述第二特征矩阵之间的按位置加权和以获得类邻接矩阵；融合所述组分特征矩阵和所述类邻接矩阵以得到配方特征矩阵；将获取的待剥离对象的图像通过作为特征提取器的第三卷积神经网络以得到表面状态特征矩阵；计算所述配方特征矩阵相对于所述表面状态特征矩阵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聚醚表面活性剂的质量百分比应增大或应减小。
PCT/CN2022/119832 2022-08-15 2022-09-20 用于剥膜液生产的自动配料系统及其配料方法 WO2024036690A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202210971924.0A CN115472242B (zh) 2022-08-15 2022-08-15 用于剥膜液生产的自动配料系统及其配料方法
CN202210971924.0 2022-08-15

Publications (1)

Publication Number Publication Date
WO2024036690A1 true WO2024036690A1 (zh) 2024-02-22

Family

ID=84367702

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2022/119832 WO2024036690A1 (zh) 2022-08-15 2022-09-20 用于剥膜液生产的自动配料系统及其配料方法

Country Status (2)

Country Link
CN (1) CN115472242B (zh)
WO (1) WO2024036690A1 (zh)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116004105A (zh) * 2023-02-17 2023-04-25 河南新旭体育设施有限公司 水性自修复硅pu面漆及其制备方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112488241A (zh) * 2020-12-18 2021-03-12 贵州大学 一种基于多粒度融合网络的零样本图片识别方法
CN114141320A (zh) * 2021-12-02 2022-03-04 鞍山浦项特种耐火材料有限公司 镁碳滑板的制备方法、系统和电子设备
CN114149254A (zh) * 2021-12-10 2022-03-08 鞍山浦项特种耐火材料有限公司 不烧滑板及其制备方法
CN114529085A (zh) * 2022-02-21 2022-05-24 杭州邬萍科技有限公司 基于大数据的居民收入预测系统及其预测方法
CN114610488A (zh) * 2022-03-11 2022-06-10 杭州青芷科技有限公司 基于负载均衡的数据采集方法、系统和电子设备
CN114866427A (zh) * 2022-04-22 2022-08-05 杭州雅深科技有限公司 全屋路由组网方法及其系统

Family Cites Families (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JP5544291B2 (ja) * 2007-09-06 2014-07-09 ザ コカ・コーラ カンパニー 飲料ディスペンサー及び方法
CN108549332B (zh) * 2017-12-19 2020-08-25 中南大学 一种基于钴酸锂配料系统的生产状态预测方法
CN110331048B (zh) * 2019-07-02 2021-07-20 深圳市大正瑞地科技有限公司 环保剥膜液及其制备方法

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112488241A (zh) * 2020-12-18 2021-03-12 贵州大学 一种基于多粒度融合网络的零样本图片识别方法
CN114141320A (zh) * 2021-12-02 2022-03-04 鞍山浦项特种耐火材料有限公司 镁碳滑板的制备方法、系统和电子设备
CN114149254A (zh) * 2021-12-10 2022-03-08 鞍山浦项特种耐火材料有限公司 不烧滑板及其制备方法
CN114529085A (zh) * 2022-02-21 2022-05-24 杭州邬萍科技有限公司 基于大数据的居民收入预测系统及其预测方法
CN114610488A (zh) * 2022-03-11 2022-06-10 杭州青芷科技有限公司 基于负载均衡的数据采集方法、系统和电子设备
CN114866427A (zh) * 2022-04-22 2022-08-05 杭州雅深科技有限公司 全屋路由组网方法及其系统

Also Published As

Publication number Publication date
CN115472242A (zh) 2022-12-13
CN115472242B (zh) 2023-04-18

Similar Documents

Publication Publication Date Title
WO2024036690A1 (zh) 用于剥膜液生产的自动配料系统及其配料方法
CN111259945B (zh) 引入注意力图谱的双目视差估计方法
CN114475350A (zh) 智能充电系统及其工作方法
WO2023206914A1 (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
CN115599049A (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
CN115693918A (zh) 一种建筑综合智能用电系统及其方法
CN115235612B (zh) 伺服电机的故障智能诊断系统及其诊断方法
CN115982736B (zh) 用于计算机网络信息的数据加密方法及系统
CN115212790B (zh) 用于光阻剥离液生产的自动配料系统及其配料方法
CN114202740A (zh) 一种基于多尺度特征融合的行人重识别方法
CN115601318B (zh) 快吸收低反渗纸尿裤智能生产方法及其系统
CN111242181B (zh) 基于图像语义和细节的rgb-d显著性物体检测器
CN115580564B (zh) 一种物联网通信网关动态调用装置
CN114979267B (zh) 面向多服务需求的语义通信方法及装置
CN116051853A (zh) 一种自控加水的和面罐及其使用方法
CN115938505A (zh) 基于图神经网络块结构与多头注意力机制融合的药物分子筛选方法及系统
CN113868451B (zh) 基于上下文级联感知的社交网络跨模态对话方法及装置
CN117409431B (zh) 一种多模态大型语言模型训练方法、电子设备及存储介质
CN105718070A (zh) 一种拼音的长句连打输入方法及其系统
CN110826500A (zh) 基于运动链接空间的对抗性网络估计3d人体姿态的方法
CN112395974A (zh) 一种基于对象间依赖关系的目标置信度矫正方法
CN117152431A (zh) 基于多模态特征融合的图像语义分割算法及应用
CN115097884A (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
CN115723280A (zh) 厚度可调节的聚酰亚胺薄膜的生产设备
CN116027874A (zh) 笔记本电脑功耗控制方法及其系统

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: 22955466

Country of ref document: EP

Kind code of ref document: A1