A pesticide production control system and method

Abstract

The application provides a pesticide production control system and method, relates to the technical field of agricultural production control, and the system comprises: a raw material feature modeling module configured to: acquire multi-dimensional quality indexes of pesticide raw materials to be produced; and obtain a low-dimensional raw material feature vector by using a nonlinear embedding function; a dynamic prediction module configured to: acquire state parameters; generate weighted state features by using a weighted cumulative operation; and construct a dynamic prediction model by using a normalization processing operation; and a control optimization module configured to: determine target reaction state parameters and environmental adjustment factors; construct a control optimization model; generate control optimization parameters by using the control optimization model, and control production equipment to operate according to the control optimization parameters; so as to solve the problems that the current pesticide production control system cannot predictively adjust process parameters, is prone to product quality deviation or even batch rejection, and lacks a mechanism for feeding back prediction errors to model correction, and cannot correct prediction results according to prediction errors.

Description

Technical Field

[0001] This application relates to the field of agricultural production control technology, and in particular to a pesticide production control system and method. Background Technology

[0002] As an indispensable key input in modern agricultural production, pesticides involve complex chemical reactions, precise raw material ratios, strict environmental control, and consistent product quality requirements in their production process. Traditional pesticide production control often relies on experience-driven or single-stage feedback strategies, which struggle to address the complex challenges of high coupling, nonlinear dynamics, and operating condition disturbances in the reaction process. This results in problems such as incomplete reactions, large quality fluctuations, and high energy consumption, hindering product quality stability and the level of intelligent process control.

[0003] In recent years, with the rapid development of the Industrial Internet and intelligent manufacturing, digital twin technology has gradually become an emerging direction for industrial process optimization. Digital twin systems achieve high-precision modeling, dynamic simulation, and intelligent prediction of production processes through real-time mapping and virtual modeling of physical objects, thus providing real-time and interpretable decision-making support for production process control. Currently, attempts have been made to apply the digital twin concept to the process manufacturing field. Existing pesticide production control systems fall into two categories: one is an open-loop control mode primarily based on empirical parameter settings, driving equipment operation through preset process thresholds; the other is a closed-loop control system with single-variable feedback regulation, which performs real-time corrections for single process parameters such as temperature and pressure in the pesticide production process.

[0004] However, when raw material quality fluctuates slightly, the system cannot predictively adjust process parameters, which can easily lead to product quality deviations or even batch rejection. Secondly, pesticide production process status prediction often relies on static empirical models, lacking the ability to update synchronously with actual operating conditions, resulting in prediction lag and decreased accuracy. Furthermore, the system lacks a mechanism to feed back prediction errors for model correction, making it unable to adjust prediction results based on these errors. Summary of the Invention

[0005] This application provides a pesticide production control system and method to solve the technical problems of existing pesticide production control systems that cannot predictively adjust process parameters, which can easily lead to product quality deviations or even batch scrapping; and that they lack a mechanism to feed back prediction errors for model correction, making it impossible to correct prediction results based on prediction errors.

[0006] The first aspect of this application provides a pesticide production control system, comprising:

[0007] The communication-connected raw material characteristic modeling module, dynamic prediction module, and control optimization module;

[0008] The raw material feature modeling module is configured as follows:

[0009] Obtain multidimensional quality indicators of the raw materials for pesticide production; the multidimensional quality indicators include: active ingredient concentration, humidity, and impurity ratio.

[0010] Based on the aforementioned multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term.

[0011] The dynamic prediction module is configured as follows:

[0012] The state parameters of the pesticide raw materials to be produced within the production equipment during the production process are obtained; the state parameters include: temperature, pH value, reaction rate, and degree of mixing.

[0013] Based on the state parameters, a weighted state feature is generated using a weighted accumulation operation;

[0014] Based on the weighted state features and the low-dimensional raw material feature vector, a dynamic prediction model is constructed using a normalization operation; the dynamic prediction model is used to generate twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector.

[0015] The control optimization module is configured as follows:

[0016] Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced.

[0017] Based on the twin predictions, target response state parameters, and environmental regulation factors, a control optimization model is constructed.

[0018] Using the control optimization model, control optimization parameters are generated and the production equipment is controlled to operate according to the control optimization parameters; the control optimization parameters include: heating power, feeding speed, and stirring frequency.

[0019] In some embodiments, the raw material feature modeling module is further configured to:

[0020] Based on the multidimensional quality indicators, construct a high-dimensional quality indicator vector;

[0021] A low-dimensional raw material feature vector is obtained by feature mapping of the high-dimensional quality index vector using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term; the low-dimensional raw material feature vector is:

[0022] ;

[0023] In the formula, Represented as the weight matrix of the neural network embedding layer; Characterized as the first Transpose of the high-dimensional quality index vector of batch raw materials; It is represented as a bias vector.

[0024] In some embodiments, the dynamic prediction module is further configured to:

[0025] Based on the state parameters, state observations are constructed; the state observations are vectors of state parameters at any given time.

[0026] The state observations are weighted and accumulated to generate weighted state features; the weighted state features are:

[0027] ;

[0028] In the formula, Characterized as the first i During the batch reaction process, at a historical moment State observations; Represented as the current moment; ; It is characterized as a time decay coefficient.

[0029] In some embodiments, the control optimization parameters are:

[0030] ;

[0031] In the formula, Characterized as target reaction state parameters; Characterized as a control-response network; This is represented by the current control command vector to be optimized, including heating power, raw material addition rate, and stirring frequency; Characterized as twin predicted values; This is represented by the control command vector from the previous moment; Characterized as the first i Batch in t The weight vector of each component of the state parameter of the response to the target at any given time; Characterized as the first i Batch in t Environmental modulators at any given time; These are represented by the weighting coefficients of the control smoothing term; Characterized as the weight coefficients that control the sparsity regularization term; Characterized as the first i Batch production equipment in time The control optimization parameters are as follows.

[0032] In some embodiments, the system further includes:

[0033] A correction module, communicatively connected to the control optimization module and the dynamic prediction module, is configured to:

[0034] After the production equipment is operated with the control optimization parameters, the real-time status parameters of the pesticide raw materials to be produced in the production equipment are obtained.

[0035] Calculate the difference between the real-time state parameters and the twin prediction values ​​corresponding to the control optimization parameters;

[0036] Based on the difference, the weight update term is calculated using the first-order gradient approximation method;

[0037] The weight update term is input into the dynamic prediction model to update the twin prediction value.

[0038] In some embodiments, the weight update term is:

[0039] ;

[0040] In the formula, Characterized as a dynamic prediction model in the first t The weight parameter matrix at time step; Characterized by the learning rate; Characterized as real-time state parameters; This is represented as a weighted projection operation.

[0041] In some embodiments, the system further includes:

[0042] An early warning module is communicatively connected to the raw material characteristic modeling module, the dynamic prediction module, and the control optimization module; the early warning module is configured as follows:

[0043] The low-dimensional raw material feature vector, control optimization parameters, twin prediction values, and weight update magnitude are extracted into an embedded feature vector; the weight update magnitude is the difference between the weight update terms at adjacent time points.

[0044] Combine any two of the embedded feature vectors to obtain an embedded feature vector group;

[0045] A cross-attention mechanism is constructed in the embedded feature vector group to obtain a composite interaction matrix;

[0046] The tensor of the composite interaction matrix is ​​divided into three sub-tensors by channel splitting operation;

[0047] Extract the feature tensors of the sub-tensors and perform a concatenation operation to obtain the fused feature tensor;

[0048] Extract composite features related to quality index prediction from the fusion feature tensor; the quality index includes: purity of active ingredient, impurity content, and concentration compliance rate.

[0049] Based on the aforementioned composite features, a confidence score is obtained using DropBlock and attention mechanisms.

[0050] Determine whether the confidence score is less than the confidence alarm threshold; if so, issue an early warning.

[0051] In some embodiments, the warning module is further configured to:

[0052] Based on the aforementioned composite features, the predicted values ​​of the quality indicators are obtained using an activation function.

[0053] Determine whether the predicted value of the quality indicator is within the preset quality range; if not, issue an early warning.

[0054] In some embodiments, after the step of determining whether the predicted value of the quality indicator is within a preset quality range, the method further includes:

[0055] Determine the magnitude of the weight change; the magnitude of the weight change is:

[0056] ;

[0057] Obtain the gradient response distribution of the weight change magnitude across different input dimensions;

[0058] Based on the gradient response distribution, a prediction error factor report is obtained; the prediction error factor report is used to display the dominant factors that cause errors in the control optimization parameters.

[0059] The second aspect of this application provides a pesticide production control method, applied to a pesticide production control system as described in any one of the first aspects above, comprising:

[0060] Obtain multidimensional quality indicators of the raw materials for pesticide production; the multidimensional quality indicators include: active ingredient concentration, humidity, and impurity ratio.

[0061] Based on the aforementioned multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term.

[0062] The state parameters of the pesticide raw materials to be produced within the production equipment during the production process are obtained; the state parameters include: temperature, pH value, reaction rate, and degree of mixing.

[0063] Based on the state parameters, a weighted state feature is generated using a weighted accumulation operation;

[0064] Based on the weighted state features and the low-dimensional raw material feature vector, a dynamic prediction model is constructed using a normalization operation; the dynamic prediction model is used to generate twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector.

[0065] Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced.

[0066] Based on the twin predictions, target response state parameters, and environmental regulation factors, a control optimization model is constructed.

[0067] Using the control optimization model, control optimization parameters are generated and the production equipment is controlled to operate according to the control optimization parameters; the control optimization parameters include: heating power, feeding speed, and stirring frequency.

[0068] This application provides a pesticide production control system and method. The system includes: a raw material feature modeling module, a dynamic prediction module, and a control optimization module connected by communication. The raw material feature modeling module is configured to: acquire multidimensional quality indicators of the pesticide raw materials to be produced; the multidimensional quality indicators include: active ingredient concentration, humidity, and impurity ratio; based on the multidimensional quality indicators, obtain a low-dimensional raw material feature vector using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term. The dynamic prediction module is configured to: acquire state parameters of the pesticide raw materials to be produced within the production equipment during the production process; the state parameters include: temperature, pH value, reaction rate, and mixing degree; based on the state parameters, generate weighted state features using a weighted accumulation operation; and based on the weighted state features and the low-dimensional raw material feature vector, use normalization... The process involves constructing a dynamic prediction model. This model generates twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector. The control optimization module is configured to: determine target reaction state parameters and environmental adjustment factors; the target reaction state parameters include the target concentration and yield of the pesticide to be produced; construct a control optimization model based on the twin prediction values, target reaction state parameters, and environmental adjustment factors; use the control optimization model to generate control optimization parameters and control the production equipment to operate with these parameters; the control optimization parameters include heating power, feeding speed, and stirring frequency. This allows for real-time monitoring of the pesticide production process, adjustment of pesticide production process parameters, and the realization of precise perception, intelligent decision-making, and closed-loop optimization control of the pesticide production process, thereby improving the stability of pesticide product quality and the process adaptability. Attached Figure Description

[0069] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0070] Figure 1 This is a flowchart of the operation of the pesticide production control system in this application;

[0071] Figure 2 This is a flowchart of the early warning module in this application during operation.

[0072] Explanation of reference numerals in the attached figures:

[0073] 1-Raw material characteristic modeling module; 2-Dynamic prediction module; 3-Control optimization module; 4-Correction module; 5-Early warning module. Detailed Implementation

[0074] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in 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 in this application, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this application.

[0075] In some technologies, pesticide production control systems cannot predictively adjust process parameters, which can easily lead to product quality deviations or even batch scrapping. Furthermore, the lack of a mechanism to feed back prediction errors for model correction makes it impossible to correct prediction results based on these errors. To address this technical problem, this application provides a pesticide production control system and method, which are described below:

[0076] like Figure 1 The diagram shown is a flowchart of the operation of the pesticide production control system in this application.

[0077] The first aspect of this application provides a pesticide production control system, comprising:

[0078] The communication-connected raw material characteristic modeling module 1, dynamic prediction module 2, and control optimization module 3 are:

[0079] The raw material feature modeling module 1 is configured as follows:

[0080] Obtain multidimensional quality indicators of the pesticide raw materials to be produced; the multidimensional quality indicators include: active ingredient concentration, humidity, and impurity ratio; based on the multidimensional quality indicators, obtain a low-dimensional raw material feature vector using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term.

[0081] For example, the raw material feature modeling module 1 collects multi-dimensional quality indicators of raw materials through online sensing devices, constructs a high-dimensional quality indicator vector of raw materials, and converts it into a low-dimensional raw material feature vector through an embedding function, which is then passed to the dynamic prediction module 2 and the control optimization module 3.

[0082] Specifically, the raw material feature modeling module 1 is further configured as follows:

[0083] Based on the aforementioned multidimensional quality indicators, a high-dimensional quality indicator vector is constructed. Before the pesticide raw materials to be produced enter the production process, multidimensional quality indicators of the current batch of pesticide raw materials are sampled using online sensing devices (such as near-infrared spectrometers and component analyzers), including active ingredient concentration, humidity, impurity ratio, etc., to construct a high-dimensional quality indicator vector. The high-dimensional quality index vector for: ;

[0084] in, Indicates the first High-dimensional quality index vector of batch pesticide raw materials; Indicates the batch number. It is a positive integer. This refers to the batch quantity of pesticide raw materials before they enter the production process. Indicates the first The number of dimensions of high-dimensional quality indicators for batches of pesticide raw materials.

[0085] A low-dimensional raw material feature vector is obtained by feature mapping of the high-dimensional quality index vector using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term; the high-dimensional quality index vector is then mapped using a nonlinear embedding function consisting of a weight matrix and a bias term. Perform feature mapping to obtain low-dimensional raw material feature vectors. The low-dimensional raw material feature vector is:

[0086] ;

[0087] In the formula, The weight matrix, representing the embedding layer of a neural network, has dimensions d×n and is used to project the input from n dimensions to d dimensions. The initial values ​​are usually in [-0.1, 0.1] or initialized using Xavier / He. Characterized as the first The transpose of the high-dimensional quality index vector of the batch of raw materials is n×1. Represented as a bias vector with dimension d×1, it is used to enhance the nonlinear expressiveness of the embedding function, and its initial value is generally set to 0. and Together with the network of the reaction dynamics prediction model, it is constructed as an end-to-end structure and jointly optimized through a unified backpropagation mechanism; Indicates the first Low-dimensional feature vectors of batch raw materials.

[0088] The dynamic prediction module 2 is configured as follows:

[0089] The process involves acquiring the state parameters of the pesticide raw material to be produced within the production equipment during the production process; generating weighted state features based on the state parameters using a weighted accumulation operation; constructing a dynamic prediction model based on the weighted state features and the low-dimensional raw material feature vector using a normalization operation; and generating twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector.

[0090] For example, the dynamic prediction module 2 constructs weighted state features based on low-dimensional raw material feature vectors and state observation values ​​(i.e., state parameters) continuously collected by sensors during the reaction process. It uses an integral mechanism with time decay weights to construct weighted state features and inputs them into a newly constructed dynamic prediction model with nonlinear modeling capabilities and an online update mechanism to generate twin prediction values, which represent the dynamic state indicators of the current reaction process in real time.

[0091] Specifically, the dynamic prediction module 2 is further configured as follows:

[0092] Based on the state parameters, state observations are constructed; the state observations are a vector of state parameters at any given time; the state parameters include: temperature, pH value, reaction rate, and degree of mixing; after the reaction process begins, key process state parameters such as temperature, pH value, reaction rate, and degree of mixing are continuously collected by sensors to form state parameters that change over time. .

[0093] The state observations are weighted and accumulated to generate weighted state features. To model the state evolution trend, an integral mechanism with time decay weights is introduced to weight and accumulate the state parameters, generating the weighted state features at the current time. The weighted state characteristics are:

[0094] ;

[0095] In the formula, Characterized as the first i During the batch reaction process, at a historical moment State observations (such as a combined vector of temperature, pH, reaction rate, etc.); Represented as the current moment; ; Characterized by the time decay coefficient, a positive real number, it controls the rate of decay; the larger the value, the more emphasis is placed on the "recent" state. Indicates the first Batch at current time The time-weighted cumulative features are based on the time step. Represents a past moment, with a range of values. ; This represents the time decay function, which controls the weight of different historical time points; the closer the time is to the current moment... t The smaller the decay, the higher the weight.

[0096] Subsequently, a trainable dynamic prediction model is constructed based on a multilayer perceptron (MLP) architecture, consisting of an input layer, 2 to 4 hidden layers, and an output layer, supporting online weight updates. The input includes a low-dimensional raw material feature vector. Weighted state features The two are concatenated into a column vector after min-max normalization. Linear mapping and ReLU activation are performed sequentially at each layer, and residual connections are introduced to mitigate degradation issues. Dynamic prediction model weight parameters. Xavier initialization is used, with optimization via backpropagation during training and dynamic adjustment based on feedback error during runtime. The output layer linearly transforms the last hidden state using a set of trainable weight matrices and bias terms, then connects it to an activation function to map it as predicted values ​​for key response indicators, represented as Siamese predictions. (Normalized) for use in generating subsequent control commands.

[0097] The target reaction state, i.e., the target reaction state parameter, is set according to the process requirements. This serves as an ideal process indicator under conditions where product quality meets standards. Subsequently, it is combined with twin prediction values. And consider including environmental regulation factors. A control optimization model is constructed. The specific steps are as follows:

[0098] The control optimization module 3 is configured as follows:

[0099] Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced.

[0100] With target reaction state parameters With control response network The weighted Euclidean error between the output expected response state vectors is the primary optimization objective, which is achieved by introducing a time-dynamic weight vector. This achieves dynamic weighting of the error term. It also incorporates two constraints: firstly, it uses the L1 norm to implement the control command vector. The sparsity is addressed by two methods: firstly, by introducing the L2 norm squared term between the current control command and the command from the previous time step. This penalizes drastic changes in the control quantity, thereby improving the smoothness of the control commands over time and ensuring the continuity and stability of the physical process.

[0101] Based on the twin predictions, target response state parameters, and environmental adjustment factors, a control optimization model is constructed.

[0102] Using the aforementioned control optimization model, control optimization parameters are generated, and the production equipment is controlled to operate according to these parameters. The control optimization parameters include: heating power, feeding speed, and stirring frequency. The formulas for solving the control optimization model are as follows:

[0103] ;

[0104] In the formula, Characterized as target reaction state parameters; Characterized as a control-response network; This is represented by the current control command vector to be optimized, including heating power, raw material addition rate, and stirring frequency; Characterized as twin predicted values; This is represented by the control command vector from the previous moment; Characterized as the first i Batch in t The weight vector of each component of the state parameter of the response to the target at any given time; Characterized as the first i Batch in t Environmental modulators at any given time; These are represented by the weighting coefficients of the control smoothing term; Characterized as the weight coefficients that control the sparsity regularization term; Characterized as the first i Batch production equipment in time The control optimization parameters are as follows.

[0105] Among them, control response network It is built on a multilayer perceptron structure, and the input to the control response network is the control command vector. Current prediction status and environmental regulators Three vectors are concatenated and fed into the control response network as joint input features. Through a multi-layered weighted fully connected structure and a nonlinear activation function, the nonlinear interaction features between the control variables, predicted state, and environmental moderating factors are extracted layer by layer to enhance the expression and fitting ability of the control response behavior. All learnable parameters of the control response network, including the weight matrices and bias vectors of each layer, are optimized and adjusted during the training phase using a gradient descent algorithm. Finally, the control response network... Output a response state vector to the target A dimensionless, multidimensional expected response state vector with consistent dimensions; Indicates the first i Batch pesticide reaction control equipment in time t The optimal control command to be executed indicates the optimized combination of control parameters (e.g., heating power, feeding speed, stirring frequency, etc.). [] indicates that the [] indicates that the [] is found Control command vector to obtain the minimum value ; This represents the current control command vector to be optimized, including heating power, raw material addition rate, stirring frequency, etc. It has been normalized and the unit is dimensionless. This represents the control command vector from the previous moment, and... The dimensions are consistent, normalized, and the units are dimensionless. This represents the target reaction state, which is the ideal output indicator set by the process (such as target concentration, yield, etc.). It has been normalized and the unit is dimensionless. Indicates the first i Batch in t The weight vector of each component of the target response state at any given time is determined by the priority control indicators for different production stages based on expert experience. Each component takes values ​​in the range [0, 1], and the whole is normalized so that the sum of the weights is 1; Indicates the first i Batch in time The environmental adjustment factor at any given time is determined by expert experience and its range is set to [0, 1]. The weighting coefficients of the control smoothing term are set according to the actual inertial characteristics of the equipment, and the range of values ​​is

[10] . -3 ,10]; This represents the weighting coefficient for the sparsity regularization term, with values ​​derived from expert experience and ranging from

[10] . -4 ,1]; This term controls the smoothness of change, constraining the control variable to prevent abrupt changes and ensuring practical feasibility. The sparse regularization term encourages minimal control usage to save energy. The optimal control command is obtained by solving the control optimization model. It is used to drive actual reaction equipment to achieve precise control.

[0106] For example, the control optimization module 3 constructs a control optimization model based on the set target reaction state parameters, and combines the output results of the dynamic prediction module 2 to solve for the optimal control parameters, generate control commands to drive the production equipment, and realize process regulation.

[0107] This application provides a pesticide production control system that acquires multidimensional quality indicators of the pesticide raw materials to be produced, including: active ingredient concentration, humidity, and impurity ratio. Based on these multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function, which consists of a weight matrix and a bias term. The system also acquires state parameters of the pesticide raw materials within the production equipment during the production process, including: temperature, pH value, reaction rate, and mixing degree. Based on these state parameters, a weighted accumulation operation is used to generate weighted state features. Finally, the system uses these weighted state features and the low-dimensional raw material feature vector to generate a low-dimensional raw material feature vector. The material feature vector is normalized to construct a dynamic prediction model. This dynamic prediction model generates twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector. Target reaction state parameters and environmental adjustment factors are determined. These target reaction state parameters include the target concentration and yield of the pesticide to be produced. A control optimization model is constructed based on the twin prediction values, target reaction state parameters, and environmental adjustment factors. The control optimization model generates control optimization parameters and controls the production equipment to operate according to these parameters. These control optimization parameters include heating power, feeding speed, and stirring frequency. This application obtains multi-dimensional quality indicators of the pesticide raw materials in real time and uses the control optimization model to obtain control optimization parameters in real time. The production equipment is then controlled to operate according to these parameters. By monitoring the real-time indicators of the pesticide raw materials, the operating parameters of the production equipment are adjusted in real time to achieve a pesticide product that closely approximates the target reaction state parameters. This addresses the technical problems in existing pesticide production processes, such as unobservable reaction processes, difficulty in real-time optimization of control parameters, and difficulty in timely prediction and correction of quality fluctuations.

[0108] The system also includes:

[0109] Correction module 4, which is communicatively connected to control optimization module 3 and dynamic prediction module 2; correction module 4 is configured to:

[0110] After the production equipment operates with the aforementioned control optimization parameters, the real-time status parameters of the pesticide raw materials to be produced within the production equipment are obtained; under control commands... After execution, the actual state values ​​of the current reaction process are collected in real time. .

[0111] Calculate the difference between the twin prediction values ​​corresponding to the real-time state parameters and the control optimization parameters; calculate the prediction error by comparing it with the twin prediction values ​​output by the dynamic prediction model in the dynamic prediction module 2.

[0112] Based on the difference, the weight update term is calculated using the first-order gradient approximation method; based on the prediction error, the weight adjustment increment is calculated using the first-order gradient approximation method, and a parameter projection operation is introduced to restrict the updated prediction model weights within the feasible domain of the weight parameters, thereby realizing model weight correction with structural constraints and enhancing its fitting ability and prediction accuracy to the control-response relationship.

[0113] Structural constraints refer to a set of explicit restrictions imposed during the update of weight parameters. These constraints can be a combination of one or more of the following methods, depending on the actual implementation requirements: first, boundary restrictions on the value range of individual weight elements; second, restrictions on the norm of the overall weight matrix; third, the introduction of sparsity constraints to encourage some weight elements to approach zero; and fourth, the use of low-rank constraints to simplify the model structure and improve generalization ability. These structural constraints are used to ensure the numerical stability, physical realizability, and control safety of the model update. Specifically, a first-order gradient approximation method is used to construct the weight update term, which is:

[0114] ;

[0115] In the formula, Characterized as a dynamic prediction model in the first t The weight parameter matrix at time step; Characterized by the learning rate; Characterized as real-time state parameters; This is represented as a weighted projection operation.

[0116] in: Indicates the dynamic prediction model in the first... t The weight parameter matrix at time +1 is the one at the +1st time. t The time-lapse model executes control commands and acquires the actual reaction state. Then, based on the twin prediction values The error between them is corrected by the first-order gradient approximation method to obtain the updated result; This represents the learning rate, whose value is obtained through experimental tuning or adaptive optimization algorithms, and its range is (10...). -5 ,1); Indicates the first i Batch pesticide reaction time tThe actual process state vector at any given time is obtained through field sensor equipment and includes multiple dimensions, such as temperature, pH value, reaction rate, and product concentration. It has been normalized and the unit is dimensionless. This represents the first-order gradient approximation term; This represents the weight projection operation; This represents the feasible region of the weight parameters in the reaction dynamics prediction model, used to constrain the weights. The range of values ​​for the weights is determined to ensure numerical stability, physical feasibility, and predictive reliability. This range is typically defined by limiting the size of the elements constrained by the weights or by limiting the overall norm. Specific values ​​are derived from expert experience, actual process constraints, or statistical characteristics trained on historical operating data to adapt to different operating conditions and dynamic changes in the system.

[0117] The weight update term is input into the dynamic prediction model to update the twin prediction value.

[0118] For example, after the control command is executed, the correction module 4 collects the actual reaction process state value, i.e., the real-time state parameter, and compares the error with the twin prediction value generated by the dynamic prediction module 2; based on the error result, the weights of the dynamic prediction model in the dynamic prediction module 2 are updated to enhance the adaptability of the dynamic prediction model to control intervention and environmental fluctuations, improve prediction accuracy and model generalization ability, thereby realizing the correction of model weights with structural constraints, enhancing its fitting ability and prediction accuracy, further increasing the control progress of the pesticide production control system, and making the pesticide production product concentration and yield more in line with user requirements.

[0119] like Figure 2 The diagram shown is a flowchart of the early warning module in this application during operation.

[0120] The system also includes:

[0121] Early warning module 5, which is communicatively connected to raw material characteristic modeling module 1, dynamic prediction module 2, and control optimization module 3; the early warning module 5 is configured as follows:

[0122] The embedded feature vectors of the low-dimensional raw material feature vector, control optimization parameters, twin predicted values, and weight update magnitude are extracted; the weight update magnitude is the difference between the weight update terms at adjacent time points. To achieve real-time prediction and early warning of the final product quality, four types of key input information are integrated: low-dimensional raw material feature vector, optimal control command vector, twin predicted values, and weight update magnitude (as a measure of prediction confidence). Based on these four inputs, a Hierarchical Decoupled Interaction Fusion Network (HDIF-Net) is constructed, including an input decoupling module, a cross-modal interaction module, a multi-scale residual fusion module, and a heterogeneous decoding output module, used to predict the specific values ​​of key quality indicators of the current batch of pesticide products, including the purity of active ingredients, impurity content, and concentration compliance rate.

[0123] Combine any two of the embedded feature vectors to obtain an embedded feature vector group; construct a cross-attention mechanism in the embedded feature vector group to obtain a composite interaction matrix; divide the tensor of the composite interaction matrix into three sub-tensors through a channel splitting operation; extract the feature tensors of the sub-tensors and perform a cascade operation to obtain a fused feature tensor.

[0124] Specifically, HDIF-Net uses four inputs as its foundation. In the input decoupling module, it extracts four types of embedded feature vectors using internal convolutional attention blocks. , , , The corresponding features are low-dimensional raw material feature vectors. Embedded feature vectors, optimal control commands Embedded feature vectors, twin predictions The embedded feature vector and the weight update magnitude of dynamic prediction module 2. The embedded feature vectors are then used in the cross-modal interaction module. This involves arbitrarily pairwise combining the four types of embedded feature vectors to obtain six pairs. For each pair, a cross-attention mechanism is constructed, resulting in six sets of cross-attention outputs used to build the composite interaction matrix. H To uncover higher-order coupling relationships. Composite interaction matrix. H The data is fed into the multi-scale residual fusion module, where it first undergoes a channel splitting operation to separate the interaction matrix. H A tensor (i.e., a data structure organized as a multidimensional array, often used in deep learning to describe high-dimensional features with multiple channels, spatial locations, or batches) is uniformly divided into three sub-tensors along the channel dimension. The features are fed into local, mesoscale, and global towers for parallel feature extraction. Each sub-tensor passes through a convolutional layer (e.g., 1×1, 3×3, or dilated convolution), a normalization layer (BatchNorm), and an activation function (e.g., ReLU) within its corresponding tower, and is combined with cross-layer residual connections for multi-level nonlinear representation learning. The local tower focuses on capturing weak interactions between local channels, the mesoscale tower extracts mid-range coupling patterns, and the global tower integrates long-range contextual dependencies by expanding the receptive field. The output feature tensors of the three channels... During the fusion phase, channel-level cascading operations are performed, splicing them together to form the fused feature tensor. .

[0125] Extract composite features related to quality indicator prediction from the fused feature tensor; the quality indicators include: purity of active ingredients, impurity content, and concentration compliance rate; based on the composite features, obtain a confidence score using DropBlock and attention mechanisms; determine whether the confidence score is less than the confidence alarm threshold, and if so, issue an early warning.

[0126] The early warning module 5 is also configured to:

[0127] Based on the composite features, an activation function is used to obtain the predicted value of the quality indicator; it is then determined whether the predicted value of the quality indicator is within a preset quality range; if not, an early warning is issued.

[0128] Specifically, in the heterogeneous decoding output module (including two decoding sub-networks: the main path and the side path), the feature tensor is fused. As input, it is fed into the main path decoding subnetwork. First... After being flattened into a one-dimensional vector, the vector is sequentially fed into a fully connected layer to compress and extract composite features related to the prediction of key quality indicators. The final decoding layer is the output layer of the constructed hierarchical decoupled interactive fusion network, whose number of neurons matches the dimension of the key quality indicator to be predicted (e.g., three dimensions: purity of active ingredient, impurity content, and concentration compliance rate). The output layer uses the Sigmoid activation function to output the predicted values ​​of the key quality indicators. Normalized to the [0,1] interval, it facilitates comparison with the standard quality interval and the determination of the confidence alarm threshold. The standard quality interval is set based on the process specifications and historical quality data statistics of long-term production, and is usually constructed in the form of mean ± standard deviation, focusing on checking the "compliance" of the predicted value. The side path decoding subnetwork introduces DropBlock and attention mechanisms to output the confidence score at the current moment. Used to quantify the predicted values ​​of key quality indicators generated by the main path decoding subnetwork. The credibility level. If detected (in The confidence alarm threshold is determined by... If the lower quantile of the historical confidence score sample is empirically set, for example, the 10th percentile (the focus is on testing the "reliability" of the prediction results), then an abnormality warning mechanism will be triggered.

[0129] For example, during the training of the HDIF-Net network, all parameters, including the kernel weights and bias parameters of the convolutional attention blocks in the decoupling module, the cross-attention weights in the cross-modal interaction module, the convolutional layer parameters of each tower structure in the multi-scale residual fusion module, and the fully connected layers of the main path and side path decoding subnetworks in the heterogeneous decoding output module, as well as the parameters in the attention module and DropBlock, are initialized using He. The forward propagation process includes embedding, interaction modeling, residual fusion, and multi-path decoding operations. During the training phase, optimization is performed using joint loss: the main task is the numerical regression of key quality indicators to measure the numerical error between the predicted and true values; at the same time, a difference metric between the confidence output distribution and the stable sample target distribution is introduced to capture the predicted values ​​of key quality indicators output by the hierarchical decoupling and interactive fusion network. The uncertainty. If the hierarchical decoupled interaction fusion network continuously detects during training... The downward trend is addressed by introducing a learnable temperature parameter to adjust the Softmax weights in the cross-attention module to a temperature scaling effect, thereby concentrating the attention distribution and enhancing the focus on key interaction features. The temperature parameter controls the smoothness of the attention distribution; the lower the temperature, the more concentrated the attention, which strengthens the model's ability to focus on key interaction features. Combined with a gradient feedback mechanism, the elements of the attention scoring matrix are fine-tuned, thus adaptively enhancing the HDIF-Net network's ability to focus on high-confidence regions.

[0130] Among them, the predicted values ​​of key quality indicators in actual operation The system compares the predicted result with a preset standard quality range: if the result falls within the standard quality range, the current production process continues; if it exceeds the range, a quality warning mechanism is immediately triggered. The warning response includes: issuing a quality deviation alarm, reverting to the control command vector from the previous moment in the intelligent control command generation process, and re-introducing it into the intelligent control command generation step to generate the current optimal control command vector, thus forming a closed-loop feedback control. Simultaneously, it calls upon the model weight change magnitude... This serves as an important basis for analyzing the sources of deviation.

[0131] After the step of determining whether the predicted value of the quality indicator is within the preset quality range, the method further includes:

[0132] Determine the magnitude of the weight change; the magnitude of the weight change is:

[0133] ;

[0134] Obtain the gradient response distribution of the weight change magnitude on different input dimensions; based on the gradient response distribution, obtain a prediction error factor report; the prediction error factor report is used to display the dominant factors causing errors in the control optimization parameters.

[0135] Specifically, This represents the strength of the prediction function response correction driven by the deviation between the model output and the true state at the current moment. Through analysis... The distribution of gradient responses across different input dimensions can help determine the main factors causing prediction errors: if the gradients are mainly concentrated in the raw material characteristic dimension, it indicates that raw material fluctuations are the primary cause; if they are concentrated in the control variable dimension, it indicates that control deviation is the dominant factor; if the responses in multiple dimensions are insignificant, it may be due to environmental disturbances or unmodeled external factors. This analysis process provides data support for subsequent control command optimization, enabling more targeted strategy adjustments.

[0136] For example, the early warning module 5 integrates multi-source information such as low-dimensional raw material feature vectors, control commands, twin prediction values, and weight update magnitudes of the dynamic prediction module 2 to construct a fusion-type quality prediction model. It outputs predicted values ​​of key product quality indicators, compares them with standard quality ranges to determine compliance, and combines confidence scores with threshold comparisons to assess credibility. If any abnormality is detected, an early warning is triggered, the control command vector is rolled back, and the deviation result is fed back to the control optimization module 3 to form a closed-loop feedback control.

[0137] A second aspect of this application provides a pesticide production control method, applied to a pesticide production control system described in any of the above embodiments, comprising:

[0138] Obtain multidimensional quality indicators of the raw materials for pesticide production; the multidimensional quality indicators include: active ingredient concentration, humidity, and impurity ratio.

[0139] Based on the aforementioned multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term.

[0140] The state parameters of the pesticide raw materials to be produced within the production equipment during the production process are obtained; the state parameters include: temperature, pH value, reaction rate, and degree of mixing.

[0141] Based on the state parameters, a weighted state feature is generated using a weighted accumulation operation;

[0142] Based on the weighted state features and the low-dimensional raw material feature vector, a dynamic prediction model is constructed using a normalization operation; the dynamic prediction model is used to generate twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector.

[0143] Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced.

[0144] Based on the twin predictions, target response state parameters, and environmental regulation factors, a control optimization model is constructed.

[0145] Using the control optimization model, control optimization parameters are generated and the production equipment is controlled to operate according to the control optimization parameters; the control optimization parameters include: heating power, feeding speed, and stirring frequency.

[0146] It is worth noting that the effects of the above method embodiments can be found in the effects of the above system embodiments, and will not be repeated here.

[0147] The above detailed embodiments further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A pesticide production control system, characterized in that, include: The communication connection includes a raw material characteristic modeling module (1), a dynamic prediction module (2), and a control optimization module (3). The raw material feature modeling module (1) is configured as follows: Obtain multidimensional quality indicators of pesticide raw materials to be produced; The multidimensional quality indicators include: concentration of active ingredients, humidity, and proportion of impurities; Based on the aforementioned multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term. The dynamic prediction module (2) is configured as follows: The state parameters of the pesticide raw materials to be produced within the production equipment during the production process are obtained; the state parameters include: temperature, pH value, reaction rate, and degree of mixing. Based on the state parameters, a weighted state feature is generated using a weighted accumulation operation; Based on the weighted state features and the low-dimensional raw material feature vector, a dynamic prediction model is constructed using a normalization operation; the dynamic prediction model is used to generate twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector. The control optimization module (3) is configured as follows: Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced. Based on the twin predictions, target response state parameters, and environmental regulation factors, a control optimization model is constructed. Using the aforementioned control optimization model, control optimization parameters are generated, and the production equipment is controlled to operate according to these parameters. The control optimization parameters include: heating power, feeding speed, and stirring frequency. These parameters are determined by the following formula: ； In the formula, Characterized as target reaction state parameters; Characterized as a control-response network; This is represented by the current control command vector to be optimized, including heating power, raw material addition rate, and stirring frequency; Characterized as twin predicted values; This is represented by the control command vector from the previous moment; It is represented as the weight vector of each component of the target response state parameter of the i-th batch at time t; This is represented as the environmental adjustment factor for the i-th batch at time t; These are represented by the weighting coefficients of the control smoothing term; Characterized as the weight coefficients that control the sparsity regularization term; Characterized by the production equipment of the i-th batch at time The control optimization parameters are as follows.

2. The pesticide production control system according to claim 1, characterized in that, The raw material feature modeling module (1) is further configured as follows: Based on the multidimensional quality indicators, construct a high-dimensional quality indicator vector; A low-dimensional raw material feature vector is obtained by feature mapping of the high-dimensional quality index vector using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term; the low-dimensional raw material feature vector is: ； In the formula, Represented as the weight matrix of the neural network embedding layer; Characterized as the first Transpose of the high-dimensional quality index vector of batch raw materials; It is represented as a bias vector.

3. The pesticide production control system according to claim 1, characterized in that, The dynamic prediction module (2) is further configured as follows: Based on the state parameters, state observations are constructed; the state observations are vectors of state parameters at any given time. The state observations are weighted and accumulated to generate weighted state features; the weighted state features are: ； In the formula, Characterized as the reaction process of the i-th batch, at a historical moment The state observation value; t represents the current time; ; It is characterized as a time decay coefficient.

4. A pesticide production control system according to claim 1, characterized in that, The system also includes: The correction module (4) is communicatively connected to the control optimization module (3) and the dynamic prediction module (2); the correction module (4) is configured to: After the production equipment is operated with the control optimization parameters, the real-time status parameters of the pesticide raw materials to be produced in the production equipment are obtained. Calculate the difference between the real-time state parameters and the twin prediction values ​​corresponding to the control optimization parameters; Based on the difference, the weight update term is calculated using the first-order gradient approximation method; The weight update term is input into the dynamic prediction model to update the twin prediction value.

5. A pesticide production control system according to claim 4, characterized in that, The weight update term is: ； In the formula, It is represented as the weight parameter matrix of the dynamic prediction model at time t; Characterized by the learning rate; Characterized as real-time state parameters; This is represented as a weighted projection operation; Characterized as twin predicted values; It is characterized as a low-dimensional raw material feature vector.

6. A pesticide production control system according to claim 1, characterized in that, The system also includes: Early warning module (5), which is communicatively connected to raw material characteristic modeling module (1), dynamic prediction module (2), and control optimization module (3); the early warning module (5) is configured as follows: The low-dimensional raw material feature vector, control optimization parameters, twin prediction values, and weight update magnitude are extracted into an embedded feature vector; the weight update magnitude is the difference between the weight update terms at adjacent time points. Combine any two of the embedded feature vectors to obtain an embedded feature vector group; A cross-attention mechanism is constructed in the embedded feature vector group to obtain a composite interaction matrix; The tensor of the composite interaction matrix is ​​divided into three sub-tensors by channel splitting operation; Extract the feature tensors of the sub-tensors and perform a concatenation operation to obtain the fused feature tensor; Extract composite features related to quality index prediction from the fusion feature tensor; the quality index includes: purity of active ingredient, impurity content, and concentration compliance rate. Based on the aforementioned composite features, a confidence score is obtained using DropBlock and attention mechanisms. Determine whether the confidence score is less than the confidence alarm threshold; if so, issue an early warning.

7. A pesticide production control system according to claim 6, characterized in that, The early warning module (5) is also configured to: Based on the aforementioned composite features, the predicted values ​​of the quality indicators are obtained using an activation function. Determine whether the predicted value of the quality indicator is within the preset quality range; if not, issue an early warning.

8. A pesticide production control system according to any one of claims 6 or 7, characterized in that, After the step of determining whether the predicted value of the quality indicator is within the preset quality range, the method further includes: Determine the magnitude of the weight change; the magnitude of the weight change is: ； In the formula, Characterized as twin predicted values; Characterized as a low-dimensional raw material feature vector; Characterized by the learning rate; Characterized as the actual state value of the raw materials to be produced; Obtain the gradient response distribution of the weight change magnitude across different input dimensions; Based on the gradient response distribution, a prediction error factor report is obtained; the prediction error factor report is used to display the dominant factors that cause errors in the control optimization parameters.

9. A pesticide production control method, applied to a pesticide production control system according to any one of claims 1 to 8, characterized in that, include: Obtain multidimensional quality indicators of pesticide raw materials to be produced; The multidimensional quality indicators include: concentration of active ingredients, humidity, and proportion of impurities; Based on the aforementioned multidimensional quality indicators, a low-dimensional raw material feature vector is obtained using a nonlinear embedding function; the nonlinear embedding function consists of a weight matrix and a bias term. The state parameters of the pesticide raw materials to be produced within the production equipment during the production process are obtained; the state parameters include: temperature, pH value, reaction rate, and degree of mixing. Based on the state parameters, a weighted state feature is generated using a weighted accumulation operation; Based on the weighted state features and the low-dimensional raw material feature vector, a dynamic prediction model is constructed using a normalization operation; the dynamic prediction model is used to generate twin prediction values ​​based on the weighted state features and the low-dimensional raw material feature vector. Determine the target reaction state parameters and environmental regulation factors; the target reaction state parameters include: the target concentration and yield of the pesticide to be produced. Based on the twin predictions, target response state parameters, and environmental regulation factors, a control optimization model is constructed. Using the control optimization model, control optimization parameters are generated and the production equipment is controlled to operate according to the control optimization parameters; the control optimization parameters include: heating power, feeding speed, and stirring frequency.

Images