A method and system for predicting tbc removal efficiency

Abstract

The application provides a TBC removal efficiency prediction method and system, comprising collecting condition state parameters and operation control parameters in a TBC removal process, the condition state parameters comprising initial pollutant concentration and coexisting ion concentration, the operation control parameters comprising catalyst dosage, pH and temperature, and being respectively coded into corresponding feature vectors; after vector splicing, an initial feature interaction matrix is constructed through outer product, Hadamard product operation is performed in combination with interaction attention mask to obtain a weighted interaction tensor; the spliced vector is expanded into a main effect memory, the weighted interaction tensor is used as an interaction effect memory, tensor aggregation is performed, and the two types of memories are weighted and fused through a gate matrix to obtain a fusion representation; a query vector is generated relying on the spliced vector to calculate attention weight, the fusion representation is weighted to obtain a context vector, and the prediction of TBC removal efficiency is completed through a full connection network.

Description

Technical Field

[0001] This application belongs to the field of latex purification and processing, and in particular relates to a method and system for predicting TBC removal efficiency. Background Technology

[0002] With the continuous development of the synthetic rubber and polymer materials industry, the demand for latex products continues to increase. During latex production and storage, TBC (tert-butylcatechol) is typically added as a polymerization inhibitor to prevent monomer self-polymerization. However, residual TBC in latex can severely affect subsequent processing performance and the quality of the final product. Therefore, efficient removal of TBC from latex systems has become a crucial issue that urgently needs to be addressed in the fields of polymer synthesis and latex processing. In actual TBC removal processes, the removal efficiency is usually influenced by a variety of factors. These factors include both the conditional parameters of the latex system itself, such as the initial TBC concentration and the concentrations of various coexisting ions in the latex system, and the operational control parameters during the treatment process, such as the amount of catalyst or sodium hydroxide added, the pH value of the reaction system, and the ambient temperature. These parameters play different roles in the removal reaction system, exhibiting both independent influencing mechanisms and mutual coupling to jointly determine the final treatment effect. Cross-characteristic interaction effects exist in the TBC removal process. For example, coexisting ions may interfere with catalyst activity, and small fluctuations in pH and temperature may further amplify or weaken this synergistic or antagonistic effect. Under different actual operating conditions, the contribution ratio of the independent main effect of a single parameter and the interaction effect between multiple parameters to the removal result will also change. Traditional prediction methods often struggle to identify deep cross-dimensional feature interactions, cannot effectively measure the fusion correlation between different parameter combinations, and are difficult to dynamically assess and fuse main effects and interaction effects based on the current variable operating conditions. The aforementioned deficiencies in the ability to identify and fuse variable feature interactions can easily lead to significant deviations in model prediction results. Therefore, how to extract deep interactions between multi-source process parameters, achieve feature dimensional fusion and focus, and construct a high-precision TBC removal efficiency prediction model is a problem that urgently needs to be solved in this field. Summary of the Invention

[0003] According to a first aspect of the present invention, a method for predicting TBC removal efficiency is proposed to address the problem of how to extract the deep interaction relationships between multi-source process parameters, achieve feature dimension fusion and focus, and construct a high-precision TBC removal efficiency prediction model, comprising the following steps: The parameters of the TBC removal process are obtained, which are divided into conditional state parameters including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters including the catalyst dosage, pH and temperature. The conditional state parameters and operational control parameters are converted into conditional state vectors and operational parameter vectors, respectively. Based on the conditional state vector and the operation parameter vector, a concatenated vector is obtained. Based on the concatenated vector, an initial cross-feature interaction matrix is ​​generated through an outer product operation. An interaction probe query is generated using the concatenated vector and mapped to an interaction attention mask through a transformation network. The interaction attention mask is then subjected to a Hadamard product operation with the initial cross-feature interaction matrix to obtain a weighted interaction tensor. The concatenated vector is expanded by copying to generate a matrix with the same dimension as the weighted interaction tensor as the main effect memory. The weighted interaction tensor is used as the interaction effect memory. The interaction effect memory is summed and aggregated to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation. Based on the concatenated vector, a target query vector is obtained through a query network. The attention score between the target query vector and the fused memory representation is calculated, and the fused memory representation is weighted accordingly to generate a context vector. The context vector is then input into a fully connected layer to output a predicted TBC removal efficiency value.

[0004] Optionally, the acquisition of TBC removal process parameters is divided into conditional state parameters including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters including catalyst dosage, pH, and temperature. The conditional state parameters and operational control parameters are respectively converted into conditional state vectors and operational parameter vectors, including: The initial TBC concentration and coexisting ion concentration, as well as the catalyst dosage, pH and temperature, are subjected to range normalization. Using the range benchmark within the historical time window, the values ​​of each parameter are mapped to the range of 0 to 1. According to the preset bit width and feature order, the normalized conditional state parameters are concatenated and arranged into the conditional state vector, and the normalized operation control parameters are concatenated and arranged into the operation parameter vector.

[0005] Optionally, the step of obtaining a concatenated vector based on the conditional state vector and the operation parameter vector, and generating an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, includes: The conditional state vector and the operation parameter vector are concatenated end to end to obtain the concatenated vector. The concatenated vector is used as a column vector and transposed to obtain a row vector. The outer product operation is performed on the column vector and the row vector to obtain a two-dimensional grid-like parameter matrix. The matrix is ​​used as the initial cross-feature interaction matrix to identify the second-order cross features between the conditional state and the operation control.

[0006] Optionally, the step of generating an interactive probe query using the concatenated vector and mapping it to an interactive attention mask via a transformation network includes: The concatenated vector is input into a multilayer perceptron containing at least one hidden layer to extract a nonlinear feature sequence to obtain the interaction detection query. The interaction detection query is input into a fully connected mapping layer and subjected to numerical compression and normalization mapping through a Sigmoid activation function to output a weight matrix with the same dimension as the initial cross-feature interaction matrix. The weight matrix is ​​used as the interaction attention mask.

[0007] Optionally, generating a hybrid gated matrix using the aggregated tensor via a gated network includes: After flattening the aggregated tensor into a one-dimensional vector, it is input into a linear transformation neural network with specified weights to extract global interaction features. The global interaction features are then numerically compressed and normalized using the Sigmoid activation function to generate gating weight coefficients distributed between 0 and 1. Based on the gating weight coefficients, a matrix with the same dimensions as the main effect memory and the interaction effect memory is constructed through broadcast expansion, which serves as the hybrid gating matrix.

[0008] Optionally, the step of using the hybrid gating matrix to perform a weighted summation of the main effect memory and the interaction effect memory to obtain a fused memory representation includes: Each element in the hybrid gating matrix is ​​extracted as the main effect weight, and the residual difference between the all-one matrix and the hybrid gating matrix is ​​used as the interaction effect weight. The main effect weights are multiplied element-wise with the main effect memory to obtain a first weighted matrix. The interaction effect weights are multiplied element-wise with the interaction effect memory to obtain a second weighted matrix. The elements at corresponding positions of the first weighted matrix and the second weighted matrix are added and fused to obtain the fused memory representation.

[0009] Optionally, the step of obtaining the target query vector based on the concatenated vector through a query network, calculating the attention score between the target query vector and the fused memory representation, and weighting the fused memory representation accordingly to generate a context vector includes: The inner product of the target query vector, which has the same dimensions as each column vector in the fusion memory representation, and each column vector in the fusion memory representation is calculated to obtain the original score of local feature relevance. The original score is numerically transformed and dimensionlessly normalized by the Softmax function to obtain the attention score. Each column vector of the fusion memory representation is multiplied by the attention score corresponding to the vector, and all weighted column vectors are superimposed and aggregated to obtain the context vector for the final prediction target.

[0010] According to a second aspect of the present invention, a TBC removal efficiency prediction system is provided, comprising the following modules: The conversion module is used to obtain TBC removal process parameters, which are divided into conditional state parameters including initial TBC concentration and coexisting ion concentration, and operational control parameters including catalyst dosage, pH and temperature. The conditional state parameters and operational control parameters are converted into conditional state vectors and operational parameter vectors, respectively. The generation module is used to obtain a concatenated vector based on the conditional state vector and the operation parameter vector, generate an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, generate an interaction probe query using the concatenated vector, map it to an interaction attention mask through a transformation network, and perform a Hadamard product operation on the interaction attention mask and the initial cross-feature interaction matrix to obtain a weighted interaction tensor. The aggregation module is used to generate a matrix with the same dimension as the weighted interaction tensor by copying and expanding the spliced ​​vector as the main effect memory, use the weighted interaction tensor as the interaction effect memory, and sum and aggregate the interaction effect memories to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation. The output module is used to obtain the target query vector based on the concatenated vector through the query network, calculate the attention score between the target query vector and the fused memory representation, and weight the fused memory representation accordingly to generate a context vector. The context vector is then input into the fully connected layer to output the TBC removal efficiency prediction value.

[0011] Preferably, the acquisition of TBC removal process parameters is divided into conditional state parameters including initial TBC concentration and coexisting ion concentration, and operational control parameters including catalyst dosage, pH, and temperature. The conditional state parameters and operational control parameters are respectively converted into conditional state vectors and operational parameter vectors, including: The initial TBC concentration and coexisting ion concentration, as well as the catalyst dosage, pH and temperature, are subjected to range normalization. Using the range benchmark within the historical time window, the values ​​of each parameter are mapped to the range of 0 to 1. According to the preset bit width and feature order, the normalized conditional state parameters are concatenated and arranged into the conditional state vector, and the normalized operation control parameters are concatenated and arranged into the operation parameter vector.

[0012] Preferably, the step of obtaining a concatenated vector based on the conditional state vector and the operation parameter vector, and generating an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, includes: The conditional state vector and the operation parameter vector are concatenated end to end to obtain the concatenated vector. The concatenated vector is used as a column vector and transposed to obtain a row vector. The outer product operation is performed on the column vector and the row vector to obtain a two-dimensional grid-like parameter matrix. The matrix is ​​used as the initial cross-feature interaction matrix to identify the second-order cross features between the conditional state and the operation control.

[0013] Preferably, the step of generating an interactive probe query using the concatenated vector and mapping it to an interactive attention mask via a transformation network includes: The concatenated vector is input into a multilayer perceptron containing at least one hidden layer to extract a nonlinear feature sequence to obtain the interaction detection query. The interaction detection query is input into a fully connected mapping layer and subjected to numerical compression and normalization mapping through a Sigmoid activation function to output a weight matrix with the same dimension as the initial cross-feature interaction matrix. The weight matrix is ​​used as the interaction attention mask.

[0014] Preferably, the step of generating a hybrid gated matrix using the aggregated tensor via a gated network includes: After flattening the aggregated tensor into a one-dimensional vector, it is input into a linear transformation neural network with specified weights to extract global interaction features. The global interaction features are then numerically compressed and normalized using the Sigmoid activation function to generate gating weight coefficients distributed between 0 and 1. Based on the gating weight coefficients, a matrix with the same dimensions as the main effect memory and the interaction effect memory is constructed through broadcast expansion, which serves as the hybrid gating matrix.

[0015] Preferably, the step of using the hybrid gating matrix to perform a weighted summation of the main effect memory and the interaction effect memory to obtain a fused memory representation includes: Each element in the hybrid gating matrix is ​​extracted as the main effect weight, and the residual difference between the all-one matrix and the hybrid gating matrix is ​​used as the interaction effect weight. The main effect weights are multiplied element-wise with the main effect memory to obtain a first weighted matrix. The interaction effect weights are multiplied element-wise with the interaction effect memory to obtain a second weighted matrix. The elements at corresponding positions of the first weighted matrix and the second weighted matrix are added and fused to obtain the fused memory representation.

[0016] Preferably, the step of obtaining a target query vector based on the concatenated vector through a query network, calculating the attention score between the target query vector and the fused memory representation, and weighting the fused memory representation accordingly to generate a context vector includes: The inner product of the target query vector, which has the same dimensions as each column vector in the fusion memory representation, and each column vector in the fusion memory representation is calculated to obtain the original score of local feature relevance. The original score is numerically transformed and dimensionlessly normalized by the Softmax function to obtain the attention score. Each column vector of the fusion memory representation is multiplied by the attention score corresponding to the vector, and all weighted column vectors are superimposed and aggregated to obtain the context vector for the final prediction target.

[0017] This invention integrates multi-dimensional process information by acquiring and quantifying conditional state parameters and operational control parameters during the TBC removal process. It utilizes outer product operations to generate a cross-feature interaction matrix and performs Hadamard product operations in conjunction with an interaction attention mask to extract and enhance the nonlinear interaction features between key parameters, reducing the interference of redundant information on prediction results. By combining the main effect memory corresponding to a single feature with the interaction effect memory formed by multiple feature combinations, and using a hybrid gating matrix for adaptive weighted fusion, it achieves the collaborative expression of feature information at different levels, ensuring the integrity and richness of feature representation. Through an attention mechanism, it calculates the correlation score between the target query vector and the fused memory representation, focusing on core contextual features that have a more prominent impact on removal efficiency, thus improving the accuracy, stability, and reliability of TBC removal efficiency prediction. Simultaneously, the model is trained using historical TBC removal efficiency labels before prediction, enabling the network parameters to have effective predictive capabilities, thereby forming a complete prediction closed loop from sample training to online inference. Attached Figure Description

[0018] Figure 1 A flowchart of the first embodiment; Figure 2 Generate a network structure diagram for interactive probing queries and interactive attention masks; Figure 3 A schematic diagram of the feature interaction matrix after weighting interaction attention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] In the first embodiment, the present invention proposes a method for predicting TBC removal efficiency, such as... Figure 1 It includes the following steps: S1, obtain the TBC removal process parameters, which are divided into conditional state parameters including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters including the catalyst dosage, pH and temperature. Convert the conditional state parameters and operational control parameters into conditional state vectors and operational parameter vectors, respectively.

[0021] Real-time data and initial setpoints of the TBC removal reaction system were collected using sensors and laboratory testing instruments. Historical operation datasets were read, and conditional parameters including initial TBC concentration and coexisting ion concentration, as well as operational control parameters including catalyst dosage, pH, and temperature, were cleaned. TBC is the common abbreviation for the target polymerization inhibitor (p-tert-butylcatechol), a key component in the latex system that needs to be degraded or separated. The catalyst is Fe. 2+ The catalyst may contain iron, TiO2, Fe3O4, etc. In the alkaline washing method, the catalyst specifically refers to sodium hydroxide. For the sake of consistency, the term "catalyst" will be used, but those skilled in the art should know that it can also be sodium hydroxide, etc. Missing values ​​are filled with the mean. The above features are subjected to range standardization or zero-mean standardization to remove dimensional differences. In this embodiment, range standardization will be used as an example for explanation. After preprocessing, the conditional state parameters are converted into a one-dimensional conditional state vector, and the operation control parameters are also converted into a one-dimensional operation parameter vector. These two one-dimensional vectors are then converted into a tensor format that can circulate in the computation graph.

[0022] In an optional embodiment, the acquisition of TBC removal process parameters is divided into conditional state parameters including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters including catalyst dosage, pH, and temperature. The conditional state parameters and operational control parameters are respectively converted into conditional state vectors and operational parameter vectors, including: The initial TBC concentration and coexisting ion concentration, as well as the catalyst dosage, pH and temperature, are subjected to range normalization. Using the range benchmark within the historical time window, the values ​​of each parameter are mapped to the range of 0 to 1. According to the preset bit width and feature order, the normalized conditional state parameters are concatenated and arranged into the conditional state vector, and the normalized operation control parameters are concatenated and arranged into the operation parameter vector.

[0023] Variable parameters within the current reaction cycle are obtained through online monitoring equipment or offline testing. For example, the initial TBC concentration in the latex system is set to a range of 10–150 mg / L; the concentration range of coexisting ions such as chloride and sulfate ions is set to 5–50 mmol / L; the catalyst dosage is set to 0.1–2.0 g / L; the pH value is set between 3.0 and 9.0; and the operating temperature is set between 20 and 60°C. A preset historical time window is set to the continuous operation data records of the past 30 days. The maximum and minimum values ​​of each parameter during this historical window are extracted as the range benchmark. During the range standardization process, the normalization formula is called. Each input parameter is processed. When the maximum and minimum values ​​of any parameter within a historical time window are equal, the standardized value of that parameter is assigned a preset neutral value or zero. When the current parameter value exceeds the historical range benchmark, the standardized result is limited to remain within the 0 to 1 range. For example, if the initial TBC concentration measured in the current batch is 38 mg / L, and the historical maximum and minimum values ​​are 150 mg / L and 10 mg / L respectively, then this parameter is mapped to 0.200. For multiple coexisting ions such as chloride and sulfate ions, range standardization is performed according to their respective historical maximum and minimum values ​​to obtain the corresponding standardized ion concentrations. Then, a weighted sum or weighted average is performed based on the preset interference weights of each ion on the TBC removal reaction to obtain the comprehensive concentration of coexisting ions. The preset interference weights are predetermined based on the degree of influence of the changes in the concentration of each ion in historical samples on the measured removal efficiency, single-factor experimental calibration results, or expert experience. For example, let the standardized concentration of chloride ions be... The standardized concentration of sulfate ions is The corresponding weights are respectively and The overall concentration of coexisting ions It can be represented as When more coexisting ions are present, the calculation is expanded according to the same rules. All the standardized values ​​mentioned above are stored in a preset bit-width format of 32-bit single-precision floating-point numbers. Vector concatenation is performed according to the preset spatial arrangement order, concatenating the initial TBC concentration and the comprehensive concentration of coexisting ions in sequence to obtain a 2×1 conditional state vector; simultaneously, the three control parameters of catalyst dosage, pH, and temperature are concatenated in sequence to generate a 3×1 operational parameter vector, constructing a dimensionless data framework.

[0024] S2, a concatenated vector is obtained based on the conditional state vector and the operation parameter vector. An initial cross-feature interaction matrix is ​​generated based on the concatenated vector through an outer product operation. An interaction probe query is generated using the concatenated vector and mapped to an interaction attention mask through a transformation network. The interaction attention mask is then subjected to a Hadamard product operation with the initial cross-feature interaction matrix to obtain a weighted interaction tensor.

[0025] The conditional state vector and the operation parameter vector are concatenated along the feature dimension to obtain a complete concatenated vector. Tensor multiplication is then performed on this concatenated vector to generate a two-dimensional initial cross-feature interaction matrix. Diagonal elements represent the square of a single feature, and off-diagonal elements represent the interactions between different features. A query network is constructed, and the concatenated vector is input into this network for affine transformation to generate an interaction probe query vector. This query vector is then input into a transformation network containing a multilayer perceptron and a nonlinear activation function. At the network's end, an activation function normalizes the output value to the 0-1 range, generating an interaction attention mask matrix with the same dimension as the initial cross-feature interaction matrix. The interaction attention mask is then subjected to a Hadamard product with the initial cross-feature interaction matrix to enhance important feature interactions and control redundant interactions, outputting a two-dimensional weighted interaction tensor. The structure of the interaction probe query and interaction attention mask generation network is as follows: Figure 2 As shown.

[0026] In an optional embodiment, the step of obtaining a concatenated vector based on the conditional state vector and the operation parameter vector, and generating an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, includes: The conditional state vector and the operation parameter vector are concatenated end to end to obtain the concatenated vector. The concatenated vector is used as a column vector and transposed to obtain a row vector. The outer product operation is performed on the column vector and the row vector to obtain a two-dimensional grid-like parameter matrix. The matrix is ​​used as the initial cross-feature interaction matrix to identify the second-order cross features between the conditional state and the operation control.

[0027] A standardized 2×1 conditional state vector and a 3×1 operation parameter vector are concatenated along the first dimension to construct a 5×1 concatenated feature vector V containing 5 elements. This concatenated vector contains the system's operating condition constraints and control strategy state information, with each feature indexed sequentially to form global state information. The model performs an outer product operation on this 5×1 concatenated column vector and a 1×5 row vector obtained by matrix transposing the vector. In this embodiment, the outer product operation represents the tensor product form of the two-dimensional product matrix generated between the column vector and the row vector. This calculation process iterates through each element pair in the two vectors, generating a matrix corresponding to the coordinates. The value is equal to the scalar product of the i-th element of the column vector and the j-th element of the row vector. The model will output a 5×5 parameter matrix, which is the initial cross-feature interaction matrix. For example, if the temperature corresponds to a value of 0.6 in the vector and the catalyst dosage is 0.4, then this operation generates a combined feature value of 0.24 at the corresponding coordinate point in the interaction matrix, indicating a second-order cross-effect in the reaction system, such as the activation of a large dose of catalyst under specific high-temperature conditions.

[0028] In an optional embodiment, the step of generating an interaction probe query using the concatenated vector and mapping it to an interaction attention mask via a transform network includes: The concatenated vector is input into a multilayer perceptron containing at least one hidden layer to extract a nonlinear feature sequence to obtain the interaction detection query. The interaction detection query is input into a fully connected mapping layer and subjected to numerical compression and normalization mapping through a Sigmoid activation function to output a weight matrix with the same dimension as the initial cross-feature interaction matrix. The weight matrix is ​​used as the interaction attention mask.

[0029] The multilayer perceptron (MLP) has a feedforward neural network structure consisting of an input layer, one hidden layer, and a dimensionality-reduced output layer. The input layer has 5 neurons, the hidden layer has 16 or 32 neurons, and the dimensionality-reduced output layer has 5 neurons. The input to the MLP is a concatenated vector of length 5, and the output is a 5-dimensional one-dimensional sequence of hidden states, i.e., the interactive probe query vector. The fully connected mapping layer has a single-layer structure containing 25 linear output units. Its input is the 5-dimensional interactive probe query vector, and its output is a 25-dimensional sequence. The concatenated vector of length 5 is input into the MLP, which uses GELU or ReLU activation functions for forward propagation to extract features. After encoding in the hidden layer, the network uses the dimensionality-reduced layer to output a 5-dimensional sequence of hidden states, i.e., the interactive probe query vector Q. This interactive probe query vector Q is then input into the fully connected mapping layer, projecting the low-dimensional feature space into a 25-dimensional sequence. The 25 scalar values ​​were element-wise adjusted using the Sigmoid activation function, limiting them to the range of 0.0 to 1.0. Then, a dimension reshaping function was used to reshape these 25 normalized values ​​into a 5×5 two-dimensional tensor, serving as the final interaction attention mask. High weights, such as 0.88, are used to emphasize the second-order combined effect of key influencing factors like pH and coexisting ions, while low weights, such as 0.05, are used to suppress redundant interaction features. The interaction attention weighted interaction matrix is ​​shown below. Figure 3 As shown.

[0030] S3, the spliced ​​vector is expanded by copying to generate a matrix with the same dimension as the weighted interaction tensor as the main effect memory, the weighted interaction tensor is used as the interaction effect memory, and the interaction effect memory is summed and aggregated to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation.

[0031] A dimension is added to a specific dimension of the concatenated vector, and preferably, the concatenated vector is copied and expanded along the column direction. This ensures that each column of the expanded two-dimensional matrix retains the main effect information of the same concatenated vector. The concatenated vector is then deformed and aligned to form a two-dimensional matrix with the same number of rows and columns as the weighted interaction tensor, serving as the main effect memory that retains the original single feature information. The interaction effect memory matrix is ​​summed and aggregated along the feature dimension to reduce its dimensionality, extracting global interaction context information to obtain an aggregated one-dimensional tensor. This aggregated tensor is fed into a gate control network to calculate a hybrid gate control matrix with values ​​between 0 and 1 and matching dimensions. Based on the gate control mechanism principle, this hybrid gate control matrix is ​​used as weight coefficients and multiplied element-wise with the main effect memory. The hybrid gate control matrix is ​​subtracted from 1 to obtain complementary weights, which are then multiplied element-wise with the interaction effect memory. Finally, the two weighted results are summed element-wise to achieve the fusion of linear and nonlinear feature representations, outputting a fused memory representation matrix.

[0032] In an optional embodiment, generating a hybrid gating matrix using the aggregated tensor via a gating network includes: After flattening the aggregated tensor into a one-dimensional vector, it is input into a linear transformation neural network with specified weights to extract global interaction features. The global interaction features are then numerically compressed and normalized using the Sigmoid activation function to generate gating weight coefficients distributed between 0 and 1. Based on the gating weight coefficients, a matrix with the same dimensions as the main effect memory and the interaction effect memory is constructed through broadcast expansion, which serves as the hybrid gating matrix.

[0033] This includes a linear transformation neural network model. The network structure is a single-layer fully connected network containing an input layer and an output layer. The input layer contains 5 neurons, and the output layer contains 1 neuron. The input is a flattened one-dimensional feature vector from a 5-dimensional aggregation tensor sequence, and the output is a single global interaction feature value. For the 5×5 interaction effect memory tensor obtained by performing a Hadamard product between the interaction attention mask and the initial cross-feature interaction matrix, the model performs tensor accumulation and summation operations along the column directions to aggregate information from different dimensions, outputting a 5-dimensional aggregation tensor sequence. This sequence is flattened into a one-dimensional feature vector and input into a linear transformation neural network with preset initial weight parameters, such as Xavier uniformly distributed weights. The global interaction feature value representing the overall interaction level of the current working condition is extracted through a fully connected inner product. The Sigmoid activation function is used to numerically transform the output scalar, restricting it to the interval between 0 and 1, generating a gating weight coefficient. For example, the current network feedforward calculates this coefficient to be 0.68. After obtaining the gating scalar, it is expanded along the row and column dimensions of the two-dimensional space using a broadcast mechanism, constructing a 5×5 matrix with all 25 elements having a value of 0.68, which is the hybrid gating matrix. This hybrid gating matrix is ​​used to balance the main effects and interaction effects.

[0034] In an optional embodiment, the step of using the hybrid gating matrix to perform a weighted summation of the main effect memory and the interaction effect memory to obtain a fused memory representation includes: Each element in the hybrid gating matrix is ​​extracted as the main effect weight, and the residual difference between the all-one matrix and the hybrid gating matrix is ​​used as the interaction effect weight. The main effect weights are multiplied element-wise with the main effect memory to obtain a first weighted matrix. The interaction effect weights are multiplied element-wise with the interaction effect memory to obtain a second weighted matrix. The elements at corresponding positions of the first weighted matrix and the second weighted matrix are added and fused to obtain the fused memory representation.

[0035] Extract the previously generated 5×5 main effect memory matrix and the aforementioned 5×5 interaction effect memory matrix. Obtain the 5×5 mixing gate matrix with an element value of, for example, 0.68 generated in the previous step as the main effect weight matrix; generate a 5×5 matrix with all elements equal to 1.0, subtract the value of the mixing gate matrix from the corresponding element of the matrix with all elements equal to 0.32, and use this matrix as the interaction effect weight matrix for the interaction effects. Then, perform the fusion step, where the model performs Hadamard product (element-wise dot product) operations on the tensors: multiply the main effect weight matrix containing 0.68 elements with the main effect memory matrix to generate the first weighted matrix; multiply the interaction effect weight matrix containing 0.32 elements with the interaction effect memory matrix to generate the second weighted matrix. Add the first and second weighted matrices, both with dimensions of 5×5, at their corresponding coordinates to obtain the fused memory representation.

[0036] S4. Based on the concatenated vector, a target query vector is obtained through a query network. The attention score between the target query vector and the fused memory representation is calculated. The fused memory representation is then weighted accordingly to generate a context vector. The context vector is input into a fully connected layer to output a predicted TBC removal efficiency value.

[0037] Another fully connected query network is constructed, taking the concatenated vector as input and mapping it through a linear transformation to a query feature space with the same dimensions as the column vectors in the fused memory representation, generating a target query vector. The similarity between the target query vector and each column vector in the fused memory representation is calculated, i.e., matrix multiplication is performed to calculate the inner product, and the result is divided by the square root of the hidden layer dimension to suppress excessive softmax distribution saturation caused by excessively large inner product values. The resulting value is then normalized along the sequence dimension, transforming it into an attention score probability distribution with a sum of 1. Based on this attention score, a weighted summation operation is performed again on the fused memory representation matrix, compressing and fusing the two-dimensional representation into a one-dimensional context vector. This vector encapsulates the most critical feature information for TBC removal efficiency. A multilayer perceptron containing hidden layers and a single output node is constructed as a regression prediction model. The context vector is input into a stacked network consisting of multiple torch.nn.Linear layers and the nonlinear activation function ReLU. After passing through a linear layer without an activation function, the forward propagation algorithm calculates and outputs a continuous scalar value, which is the system's predicted TBC removal efficiency. When the removal efficiency label is represented in the range of 0 to 1, the continuous scalar value is subjected to Sigmoid mapping or amplitude limiting processing so that the output result is within the range of 0 to 1; when the removal efficiency is expressed as a percentage, the result in this range is converted into a removal efficiency prediction value of 0% to 100%.

[0038] In an optional embodiment, the step of obtaining a target query vector based on the concatenated vector via a query network, calculating the attention score between the target query vector and the fused memory representation, and weighting the fused memory representation accordingly to generate a context vector includes: The inner product of the target query vector, which has the same dimensions as each column vector in the fusion memory representation, and each column vector in the fusion memory representation is calculated to obtain the original score of local feature relevance. The original score is numerically transformed and dimensionlessly normalized by the Softmax function to obtain the attention score. Each column vector of the fusion memory representation is multiplied by the attention score corresponding to the vector, and all weighted column vectors are superimposed and aggregated to obtain the context vector for the final prediction target.

[0039] This includes a fully connected predictor network model. The network structure is a single-layer fully connected network containing an input layer and an output layer. The input layer contains 5 neurons, and the output layer contains at least 1 neuron. The input of this predictor is a 1×5 one-dimensional context vector, and the output is the final target prediction value. The 5×5 fused memory representation is divided into 5 column vectors, each containing 5 feature elements. For each column vector, a dot product operation is performed with the target query vector to calculate 5 raw scores representing feature relevance. For example, under certain conditions, the inner product operation yields 2.5, -0.8, 1.2, 3.1, and 0.4. The array of raw scores containing 5 elements is input into the Softmax function module, making the sum of the output attention scores equal to 1.0, resulting in a dimensionless attention distribution. For example, the numerical mapping output above is 0.32, 0.01, 0.09, 0.54, and 0.04. The obtained attention scores are weighted into five column vectors according to their corresponding positions, i.e., all features in the first column are multiplied by 0.32, and all features in the fourth column are multiplied by 0.54. After weighting, these five column vectors are merged and summed. A one-dimensional context vector of length 5 is generated and provided to the fully connected layer predictor for target prediction.

[0040] During the model training phase, multiple historical TBC removal samples were collected. Each sample included the initial TBC concentration, concentrations of multiple coexisting ions, catalyst dosage, pH, temperature, and the corresponding measured TBC removal efficiency. The measured TBC removal efficiency could be calculated based on the pre-reaction and post-reaction TBC concentrations, for example, using the pre-reaction TBC concentration... Based on, calculation The predicted removal efficiency is used as a label, or converted to a percentage. Historical samples are forward-propagated according to steps S1 to S4 to obtain the predicted removal efficiency. The mean squared error or mean absolute error between the predicted and measured removal efficiency labels is used as the loss function. Subsequently, the gradient of the loss function with respect to the trainable parameters in the multilayer perceptron, fully connected mapping layer, gated network, query network, and fully connected layer predictor is calculated, and gradient descent, Adam optimizer, or other optimization algorithms are used to update the network parameters. After multiple rounds of training, training stops when the validation set loss no longer decreases, the preset number of iterations is reached, or a preset error threshold is met, and the trained model parameters are saved. In the online prediction phase, the trained model parameters are loaded, forward propagation is performed on the current TBC removal process parameters, and the predicted TBC removal efficiency value is output.

[0041] In the second embodiment, the present invention also proposes a TBC removal efficiency prediction system, comprising the following modules: The conversion module is used to obtain TBC removal process parameters, which are divided into conditional state parameters including initial TBC concentration and coexisting ion concentration, and operational control parameters including catalyst dosage, pH and temperature. The conditional state parameters and operational control parameters are converted into conditional state vectors and operational parameter vectors, respectively. The generation module is used to obtain a concatenated vector based on the conditional state vector and the operation parameter vector, generate an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, generate an interaction probe query using the concatenated vector, map it to an interaction attention mask through a transformation network, and perform a Hadamard product operation on the interaction attention mask and the initial cross-feature interaction matrix to obtain a weighted interaction tensor. The aggregation module is used to generate a matrix with the same dimension as the weighted interaction tensor by copying and expanding the spliced ​​vector as the main effect memory, use the weighted interaction tensor as the interaction effect memory, and sum and aggregate the interaction effect memories to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation. The output module is used to obtain the target query vector based on the concatenated vector through the query network, calculate the attention score between the target query vector and the fused memory representation, and weight the fused memory representation accordingly to generate a context vector. The context vector is then input into the fully connected layer to output the TBC removal efficiency prediction value.

[0042] The above description represents the preferred embodiments of the present invention. It should be noted that, for those skilled in the art, various improvements and modifications can be made without departing from the principles of the present invention, and these improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A method for predicting TBC removal efficiency, characterized in that, Includes the following steps: The parameters of the TBC removal process are obtained, which are divided into conditional state parameters including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters including the catalyst dosage, pH and temperature. The conditional state parameters and operational control parameters are converted into conditional state vectors and operational parameter vectors, respectively. Based on the conditional state vector and the operation parameter vector, a concatenated vector is obtained. Based on the concatenated vector, an initial cross-feature interaction matrix is ​​generated through an outer product operation. An interaction probe query is generated using the concatenated vector and mapped to an interaction attention mask through a transformation network. The interaction attention mask is then subjected to a Hadamard product operation with the initial cross-feature interaction matrix to obtain a weighted interaction tensor. The concatenated vector is expanded by copying to generate a matrix with the same dimension as the weighted interaction tensor as the main effect memory. The weighted interaction tensor is used as the interaction effect memory. The interaction effect memory is summed and aggregated to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation. Based on the concatenated vector, a target query vector is obtained through a query network. The attention score between the target query vector and the fused memory representation is calculated, and the fused memory representation is weighted accordingly to generate a context vector. The context vector is then input into a fully connected layer to output a predicted TBC removal efficiency value.

2. The method according to claim 1, characterized in that, The process parameters for obtaining TBC removal are divided into conditional state parameters, including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters, including the catalyst dosage, pH, and temperature. The conditional state parameters and operational control parameters are respectively converted into conditional state vectors and operational parameter vectors, including: The initial TBC concentration and coexisting ion concentration, as well as the catalyst dosage, pH and temperature, are subjected to range normalization. Using the range benchmark within the historical time window, the values ​​of each parameter are mapped to the range of 0 to 1. According to the preset bit width and feature order, the normalized conditional state parameters are concatenated and arranged into the conditional state vector, and the normalized operation control parameters are concatenated and arranged into the operation parameter vector.

3. The method according to claim 1, characterized in that, The process of obtaining a concatenated vector based on the conditional state vector and the operation parameter vector, and generating an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, includes: The conditional state vector and the operation parameter vector are concatenated end to end to obtain the concatenated vector. The concatenated vector is used as a column vector and transposed to obtain a row vector. The outer product operation is performed on the column vector and the row vector to obtain a two-dimensional grid-like parameter matrix. The matrix is ​​used as the initial cross-feature interaction matrix to identify the second-order cross features between the conditional state and the operation control.

4. The method according to claim 3, characterized in that, The step of generating an interactive probe query using the concatenated vector and mapping it to an interactive attention mask via a transformation network includes: The concatenated vector is input into a multilayer perceptron containing at least one hidden layer to extract a nonlinear feature sequence to obtain the interaction detection query. The interaction detection query is input into a fully connected mapping layer and subjected to numerical compression and normalization mapping through a Sigmoid activation function to output a weight matrix with the same dimension as the initial cross-feature interaction matrix. The weight matrix is ​​used as the interaction attention mask.

5. The method according to claim 1, characterized in that, The step of generating a hybrid gated matrix using the aggregated tensor via a gated network includes: After flattening the aggregated tensor into a one-dimensional vector, it is input into a linear transformation neural network with specified weights to extract global interaction features. The global interaction features are then numerically compressed and normalized using the Sigmoid activation function to generate gating weight coefficients distributed between 0 and 1. Based on the gating weight coefficients, a matrix with the same dimensions as the main effect memory and the interaction effect memory is constructed through broadcast expansion, which serves as the hybrid gating matrix.

6. The method according to claim 1 or 5, characterized in that, The step of using the hybrid gating matrix to perform a weighted summation of the main effect memory and the interaction effect memory to obtain a fused memory representation includes: Each element in the hybrid gating matrix is ​​extracted as the main effect weight, and the residual difference between the all-one matrix and the hybrid gating matrix is ​​used as the interaction effect weight. The main effect weights are multiplied element-wise with the main effect memory to obtain a first weighted matrix. The interaction effect weights are multiplied element-wise with the interaction effect memory to obtain a second weighted matrix. The elements at corresponding positions of the first weighted matrix and the second weighted matrix are added and fused to obtain the fused memory representation.

7. The method according to claim 1, characterized in that, The process of obtaining a target query vector based on the concatenated vector via a query network, calculating the attention score between the target query vector and the fused memory representation, and weighting the fused memory representation accordingly to generate a context vector includes: The inner product of the target query vector, which has the same dimensions as each column vector in the fusion memory representation, and each column vector in the fusion memory representation is calculated to obtain the original score of local feature relevance. The original score is numerically transformed and dimensionlessly normalized by the Softmax function to obtain the attention score. Each column vector of the fusion memory representation is multiplied by the attention score corresponding to the vector, and all weighted column vectors are superimposed and aggregated to obtain the context vector for the final prediction target.

8. A TBC removal efficiency prediction system, characterized in that, Includes the following modules: The conversion module is used to obtain TBC removal process parameters, which are divided into conditional state parameters including initial TBC concentration and coexisting ion concentration, and operational control parameters including catalyst dosage, pH and temperature. The conditional state parameters and operational control parameters are converted into conditional state vectors and operational parameter vectors, respectively. The generation module is used to obtain a concatenated vector based on the conditional state vector and the operation parameter vector, generate an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, generate an interaction probe query using the concatenated vector, map it to an interaction attention mask through a transformation network, and perform a Hadamard product operation on the interaction attention mask and the initial cross-feature interaction matrix to obtain a weighted interaction tensor. The aggregation module is used to generate a matrix with the same dimension as the weighted interaction tensor by copying and expanding the spliced ​​vector as the main effect memory, use the weighted interaction tensor as the interaction effect memory, and sum and aggregate the interaction effect memories to obtain the aggregated tensor. The aggregated tensor is used to generate a hybrid gating matrix through a gating network. The hybrid gating matrix is ​​used to perform a weighted summation of the main effect memory and the interaction effect memory to obtain the fused memory representation. The output module is used to obtain the target query vector based on the concatenated vector through the query network, calculate the attention score between the target query vector and the fused memory representation, and weight the fused memory representation accordingly to generate a context vector. The context vector is then input into the fully connected layer to output the TBC removal efficiency prediction value.

9. The system according to claim 8, characterized in that, The process parameters for obtaining TBC removal are divided into conditional state parameters, including the initial TBC concentration and the concentration of coexisting ions, and operational control parameters, including the catalyst dosage, pH, and temperature. The conditional state parameters and operational control parameters are respectively converted into conditional state vectors and operational parameter vectors, including: The initial TBC concentration and coexisting ion concentration, as well as the catalyst dosage, pH and temperature, are subjected to range normalization. Using the range benchmark within the historical time window, the values ​​of each parameter are mapped to the range of 0 to 1. According to the preset bit width and feature order, the normalized conditional state parameters are concatenated and arranged into the conditional state vector, and the normalized operation control parameters are concatenated and arranged into the operation parameter vector.

10. The system according to claim 8, characterized in that, The process of obtaining a concatenated vector based on the conditional state vector and the operation parameter vector, and generating an initial cross-feature interaction matrix based on the concatenated vector through an outer product operation, includes: The conditional state vector and the operation parameter vector are concatenated end to end to obtain the concatenated vector. The concatenated vector is used as a column vector and transposed to obtain a row vector. The outer product operation is performed on the column vector and the row vector to obtain a two-dimensional grid-like parameter matrix. The matrix is ​​used as the initial cross-feature interaction matrix to identify the second-order cross features between the conditional state and the operation control.

Images