Pesticide production control system and method

Abstract

The invention provides a pesticide production control system and method, and relates to the technical field of agricultural production control, and the system comprises a raw material feature modeling module which is configured to obtain multi-dimensional quality indexes of a to-be-produced pesticide raw material; utilizing a nonlinear embedding function to obtain a low-dimensional raw material feature vector; the dynamic prediction module is configured to obtain state parameters; generating a weighted state feature by using a weighted accumulation operation; utilizing normalization processing operation to construct a dynamic prediction model; the control optimization module is configured to determine a target reaction state parameter and an environmental regulation factor; constructing a control optimization model; using the control optimization model to generate control optimization parameters and control production equipment to control the operation of the optimization parameters; the problems that an existing pesticide production control system cannot predictively adjust technological parameters, and product quality deviation and even batch scrapping are easily caused are solved; and a mechanism for applying prediction error feedback to model correction is lacked, so that the prediction result cannot be corrected according to the prediction error.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of agricultural production control, and in particular to a pesticide production control system and method. BACKGROUND

[0002] As a key input in modern agricultural production, pesticide production involves complex chemical reactions, precise raw material proportioning, strict environmental control, and consistent product quality requirements. Traditional pesticide production control mostly adopts experience-driven or single-link feedback strategies, which are difficult to cope with the complex challenges of high coupling, nonlinear dynamics, and process disturbance in the reaction process. This results in insufficient reaction, large quality fluctuations, high production energy consumption, and other problems, which restrict the product quality stability and the intelligent level of process control.

[0003] In recent years, with the rapid development of industrial internet and intelligent manufacturing, digital twin technology has gradually become a new direction for industrial process optimization. Through real-time mapping and virtual modeling of physical objects, digital twin systems achieve high-precision modeling, dynamic simulation, and intelligent prediction of production processes, thereby providing real-time and interpretable decision-making basis for production process control. Currently, the digital twin concept has been applied to the process manufacturing field. One type of existing pesticide production control system is an open-loop control mode based on experience parameter setting, which drives equipment operation through pre-set process thresholds. Another type is a single-variable feedback regulation closed-loop control system that adjusts real-time for single process parameters such as temperature and pressure in pesticide production processes.

[0004] However, when the quality of raw materials fluctuates slightly, the system cannot predictively adjust the process parameters, which easily leads to product quality deviation or even batch rejection. Secondly, the prediction of pesticide production process status often relies on static experience models, which lack the ability to update synchronously with actual working conditions, causing prediction lag and decreased prediction accuracy. In addition, the system lacks a mechanism for feeding back prediction errors for model correction, and cannot correct prediction results based on prediction errors. SUMMARY

[0005] The present application provides a pesticide production control system and method to solve the technical problems that existing pesticide production control systems cannot predictively adjust process parameters, easily leading to product quality deviation or even batch rejection, and lack a mechanism for feeding back prediction errors for model correction, and cannot correct prediction results based on prediction errors.

[0006] The first aspect of the present application provides a pesticide production control system, comprising: a raw material feature modeling module, a dynamic prediction module, and a control optimization module connected in communication; The raw material feature modeling module is configured to: acquire a multi-dimensional quality index of a pesticide raw material to be produced; the multi-dimensional quality index comprises: effective component concentration, humidity, impurity proportion; based on the multi-dimensional quality index, obtain a low-dimensional raw material feature vector by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; The dynamic prediction module is configured to: acquire a state parameter of the pesticide raw material to be produced in a production device during production; the state parameter comprises: temperature, pH value, reaction speed, mixing degree; based on the state parameter, generate a weighted state feature by using a weighted cumulative operation; based on the weighted state feature and the low-dimensional raw material feature vector, construct a dynamic prediction model by using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector; The control optimization module is configured to: determine a target reaction state parameter and an environmental adjustment factor; the target reaction state parameter comprises: a target concentration of the pesticide to be produced, a yield; based on the twin prediction value, the target reaction state parameter, and the environmental adjustment factor, construct a control optimization model; generate a control optimization parameter by using the control optimization model and control the production device to operate at the control optimization parameter; the control optimization parameter comprises: heating power, feeding speed, stirring frequency.

[0007] In some embodiments, the raw material feature modeling module is further configured to: construct a high-dimensional quality index vector according to the multi-dimensional quality index; perform feature mapping on the high-dimensional quality index vector by using a nonlinear embedding function to obtain a low-dimensional raw material feature vector; the nonlinear embedding function is composed of a weight matrix and a bias term; the low-dimensional raw material feature vector is: ; wherein, the weight matrix represents a neural network embedding layer; the bias vector represents a bias vector; the transpose of the high-dimensional quality index vector of the batch of raw materials; the bias vector represents a bias vector.

[0008] In some embodiments, the dynamic prediction module is further configured to: construct a state observation value according to the state parameter; the state observation value is a state parameter combination vector at any time; The state observation value is subjected to a weighted cumulative operation to generate a weighted state feature; the weighted state feature is: ; In the formula, characterized as the state observation value of the historical time i in the batch reaction process; characterized as the current time; ; characterized as the time decay coefficient.

[0009] In some embodiments, the control optimization parameter is: ; In the formula, characterized as the target reaction state parameter; characterized as the control response network; characterized as the current control instruction vector to be optimized, including heating power, raw material adding speed, stirring frequency; characterized as the twin prediction value; characterized as the control instruction vector of the previous time; characterized as the weight vector of the target reaction state parameter components of the batch i at the time t ; characterized as the environmental adjustment factor of the batch i at the time t ; characterized as the weight coefficient of the control smoothing term; characterized as the weight coefficient of the control sparsity regularization term; characterized as the control optimization parameter of the batch i production equipment at time .

[0010] In some embodiments, the system further comprises: a correction module, which is communicatively connected to the control optimization module and the dynamic prediction module; the correction module is configured to: after the production equipment runs with the control optimization parameter, acquire the real-time state parameter of the pesticide raw material to be produced in the production equipment; calculate the difference between the real-time state parameter and the twin prediction value corresponding to the control optimization parameter; based on the difference, calculate a weight update term using a first-order gradient approximation method; input the weight update term into the dynamic prediction model to update the twin prediction value.

[0011] ​In some embodiments, the weight update term is: ; wherein, is a weight parameter matrix of the dynamic prediction model at the t-th time point; t is a learning rate; is a real-time state parameter; is a weight projection operation.

[0012] In some embodiments, the system further comprises: a warning module, which is communicatively connected to the raw material feature modeling module, the dynamic prediction module, and the control optimization module; the warning module is configured to: extract the low-dimensional raw material feature vector, the control optimization parameter, the twin prediction value, and the embedded feature vector of the weight update amplitude; the weight update amplitude is the difference value of the weight update term at adjacent time points; combine any two 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 tensor of the sub-tensor and perform a concatenation operation to obtain a fusion feature tensor; extract the composite features related to the quality index prediction from the fusion feature tensor; the quality index includes effective ingredient purity, impurity content, and concentration compliance rate; based on the composite features, use DropBlock and attention mechanism to obtain a confidence score; determine whether the confidence score is less than a confidence alarm threshold; if so, issue a warning.

[0013] In some embodiments, the warning module is further configured to: based on the composite features, use an activation function to obtain a quality index prediction value; determine whether the quality index prediction value is within a preset quality interval; if not, issue a warning.

[0014] In some embodiments, after the step of determining whether the quality index prediction value is within a preset quality interval, the method further comprises: determining a weight change amplitude; the weight change amplitude is: ; obtaining the gradient response distribution of the weight change amplitude on different input dimensions; ​According to the gradient response distribution, a prediction error factor report is obtained; the prediction error factor report is used to display a dominant factor causing the error of the control optimization parameter.

[0015] The second aspect of the application provides a pesticide production control method, applied to the pesticide production control system of any one of the first aspect, comprising: Obtaining a multi-dimensional quality index of the pesticide raw material to be produced; the multi-dimensional quality index includes: effective component concentration, humidity, impurity ratio; Based on the multi-dimensional quality index, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; Obtaining the state parameters of the pesticide raw material to be produced in the production equipment during the production process; the state parameters include: temperature, pH value, reaction speed, mixing degree; Based on the state parameters, a weighted state feature is generated by using a weighted cumulative operation; Based on the weighted state feature and the low-dimensional raw material feature vector, a dynamic prediction model is constructed by using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector; Determining a target reaction state parameter and an environmental adjustment factor; the target reaction state parameter includes: target concentration of the pesticide to be produced, yield; Based on the twin prediction value, the target reaction state parameter and the environmental adjustment factor, a control optimization model is constructed; Using the control optimization model, a control optimization parameter is generated and the production equipment is controlled to operate at the control optimization parameter; the control optimization parameter includes: heating power, feeding speed, stirring frequency.

[0016] The application provides a pesticide production control system and method, the system comprises: a raw material feature modeling module, a dynamic prediction module, and a control optimization module connected in communication; the raw material feature modeling module is configured to: obtain a multi-dimensional quality index of a pesticide raw material to be produced; the multi-dimensional quality index comprises: effective component concentration, humidity, and impurity proportion; based on the multi-dimensional quality index, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; the dynamic prediction module is configured to: obtain a state parameter of the pesticide raw material to be produced in a production device during production; the state parameter comprises: temperature, pH value, reaction speed, and mixing degree; based on the state parameter, a weighted state feature is generated by using a weighted cumulative operation; based on the weighted state feature and the low-dimensional raw material feature vector, a dynamic prediction model is constructed by using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector; the control optimization module is configured to: determine a target reaction state parameter and an environmental adjustment factor; the target reaction state parameter comprises: a target concentration of the pesticide to be produced and a yield; based on the twin prediction value, the target reaction state parameter, and the environmental adjustment factor, a control optimization model is constructed; the control optimization model is used to generate a control optimization parameter and control the production device to operate at the control optimization parameter; the control optimization parameter comprises: heating power, feeding speed, and stirring frequency; by monitoring the pesticide production process in real time, adjusting the production process parameter of the pesticide, and realizing precise perception, intelligent decision-making, and closed-loop optimization control of the pesticide production process, the pesticide product quality stability and process self-adaptive ability are improved. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions of the application, the drawings needed in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can also be obtained without creative labor based on these drawings.

[0018] Figure 1 A flow chart of the pesticide production control system in the application during operation; Figure 2 A flow chart of the early warning module in the application during operation.

[0019] Explanation of reference signs: 1-raw material feature modeling module; 2-dynamic prediction module; 3-control optimization module; 4-correction module; 5-early warning module. DETAILED DESCRIPTION

[0020] In order for those skilled in the art to better understand the technical solutions in the present application, the technical solutions in the embodiments of the present application will be clearly and completely described in the following with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor should fall within the scope of protection of the present application.

[0021] Since in some technologies, the pesticide production control system cannot predictively adjust the process parameters, it is easy to cause product quality deviation or even batch rejection; and there is no mechanism for feeding back the prediction error for model correction, and the prediction result cannot be corrected according to the prediction error, in order to solve the technical problem, the present application provides a pesticide production control system and method, which will be described below: As Figure 1 shown, it is a flow chart of the pesticide production control system in the present application when running.

[0022] The first aspect of the present application provides a pesticide production control system, comprising: a communication-connected raw material feature modeling module 1, a dynamic prediction module 2, and a control optimization module 3; The raw material feature modeling module 1 is configured to: obtain a multi-dimensional quality index of a raw material to be produced; the multi-dimensional quality index includes effective component concentration, humidity, and impurity ratio; based on the multi-dimensional quality index, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term.

[0023] For example, the raw material feature modeling module 1 collects the multi-dimensional quality index of the raw material through an online sensing device, constructs a high-dimensional quality index vector of the raw material, and converts it into a low-dimensional raw material feature vector through an embedding function, which is transmitted to the dynamic prediction module 2 and the control optimization module 3.

[0024] Specifically, the raw material feature modeling module 1 is further configured to: construct a high-dimensional quality index vector according to the multi-dimensional quality index; before the raw material to be produced enters the production process, the current batch of pesticide raw material is sampled for multi-dimensional quality index by an online sensing device (such as a near-infrared spectrometer and a component analyzer), including effective component concentration, humidity, impurity ratio, etc., to construct a high-dimensional quality index vector ; the high-dimensional quality index vector is: ; wherein, represents the high-dimensional quality index vector of the batch of pesticide raw material; represents batch number, is a positive integer, is the number of batches of pesticide raw materials before entering the production process, represents the first dimension number of high-dimensional quality indicators of the batch of pesticide raw materials.

[0025] The high-dimensional quality indicator vector is characterized by using a nonlinear embedding function to obtain a low-dimensional raw material feature vector; the nonlinear embedding function is composed of a weight matrix and a bias term; the high-dimensional quality indicator vector is characterized by using a nonlinear embedding function composed of a weight matrix and a bias term to obtain a low-dimensional raw material feature vector ; the low-dimensional raw material feature vector is: ; In the formula, represents a weight matrix of a neural network embedding layer, with a dimension of d x n, used to project input from n dimensions to d dimensions, and the initial value is usually in [-0.1, 0.1] or uses Xavier / He initialization; represents the transpose of the high-dimensional quality indicator vector of the first batch of raw materials, which is n x 1 after transposition; represents a bias vector with a dimension of d x 1, used to improve the nonlinear expression ability of the embedding function, and the initial value is generally set to 0, and is constructed as an end-to-end structure together with the network of the reaction dynamic prediction model, and is jointly optimized through a unified back propagation mechanism; represents the low-dimensional feature vector of the first batch of raw materials.

[0026] The dynamic prediction module 2 is configured to: obtain state parameters of the pesticide raw material to be produced in the production equipment during the production process; based on the state parameters, generate a weighted state feature using a weighted cumulative operation; based on the weighted state feature and the low-dimensional raw material feature vector, construct a dynamic prediction model using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector.

[0027] For example, the dynamic prediction module 2 uses a weighted state feature constructed by a time decay weighted integral mechanism based on the low-dimensional raw material feature vector and the state observation value, i.e. the state parameter, continuously collected by the sensor during the reaction process, and inputs it into the newly constructed dynamic prediction model with nonlinear modeling ability and online updating mechanism to generate a twin prediction value, which represents the dynamic state indicator of the current reaction process in real time.

[0028] Specifically, the dynamic prediction module 2 is further configured as follows: 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. .

[0029] 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: ; 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.

[0030] 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. With 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. The Xavier initialization is adopted, the training stage is optimized by back propagation, and the running stage is dynamically adjusted by combining feedback errors. The output layer linearly transforms the last hidden state through a set of trainable weight matrices and bias terms, and maps it to the prediction value of the key reaction indicator through an activation function, which is represented as the twin prediction value (normalized) for subsequent control instruction generation.

[0031] According to the process requirements, the target reaction state, i.e. the target reaction state parameter, is set as the ideal process indicator under the condition of product quality meeting the standard. Then, combined with the twin prediction value and considering the environmental adjustment factor , a control optimization model is constructed. The specific steps are as follows: The control optimization module 3 is configured to: determine the target reaction state parameter and the environmental adjustment factor; the target reaction state parameter includes the target concentration and yield of the pesticide to be produced.

[0032] The weighted Euclidean error between the target reaction state parameter and the expected reaction state vector output by the control response network is the main optimization objective, and the dynamic weighting of the error term is realized by introducing a time dynamic weight vector . At the same time, two constraints are combined: one is to realize the sparsity of the control instruction vector through L1 norm, and the other is to punish the sharp change of the control quantity by introducing the L2 norm square term between the current control instruction and the instruction at the last time, so as to improve the smoothness of the control instruction in time sequence, and guarantee the continuity and stability of the physical process.

[0033] Based on the twin prediction value, the target reaction state parameter and the environmental adjustment factor, a control optimization model is constructed.

[0034] Using the control optimization model, control optimization parameters are generated and the production equipment is controlled to run with the control optimization parameters; the control optimization parameters include heating power, feeding speed and stirring frequency. The control optimization parameters, i.e. the control optimization model solving formula, are as follows: ; In the formula, represents the target reaction state parameter; represents the control response network; represents the current control instruction vector to be optimized, including heating power, raw material adding speed and stirring frequency; represents the twin prediction value; represents the control instruction vector at the last time. characterized as the first i batch at t the weight vector of each component of the target reaction state parameter at the moment; characterized as the first i batch at t the environmental adjustment factor at the moment; characterized as the weight coefficient of the control smoothing term; characterized as the weight coefficient of the control sparsity regularization term; characterized as the first i batch production equipment at time the control optimization parameter.

[0035] wherein the control response network is constructed based on a multi-layer perceptron structure, the input of the control response network is composed of the control instruction vector , the current predicted state and the environmental adjustment factor three vectors spliced into joint input features into the control response network. Through the multi-layer weighted full connection structure and the nonlinear activation function, the nonlinear interaction feature between the control variable, the predicted state and the environmental adjustment factor is extracted layer by layer to enhance the expression and fitting ability of the control response behavior. All the learnable parameters of the control response network include the weight matrix and the bias vector of each layer, which are optimized and adjusted by gradient descent algorithm in the training stage. Finally, the control response network outputs a multidimensional dimensionless expected reaction state vector consistent with the target reaction state vector ; represents the optimal control instruction that the pesticide reaction control equipment of the first i batch should execute at time t , which represents the optimized control parameter combination (such as heating power, feeding speed, stirring frequency, etc.); represents the control instruction vector that makes achieve the minimum value; represents the current control instruction vector to be optimized, including heating power, raw material adding speed, stirring frequency, etc., which has been normalized and is dimensionless; represents the control instruction vector at the last moment, which is consistent with dimension, has been normalized and is dimensionless; represents the target reaction state, which is the ideal output index (such as target concentration, yield, etc.) set by the process, which has been normalized and is dimensionless; represents the optimal control instruction that the pesticide reaction control equipment of the first i batch should execute at time tThe weight vector of each component of the target reaction state at the time, the value of which is set according to expert experience to set the priority control index of different production stages, Each component of the weight vector has a value range of [0, 1], and the whole is normalized to make the weight sum equal to 1. The environmental adjustment factor of the batch at time t is represented by i The value of the environmental adjustment factor is derived from expert experience method, and the value range is set to [0, 1]. The weight coefficient of the control smoothing term is represented by -3 , and the value is set according to the actual device inertia characteristics, with a value range of [10 -4 , 1]. The weight coefficient of the control sparsity regularization term is represented by , and the value is derived from expert experience method, with a value range of [10 -5 , 1]. The control change smoothness term is represented by , which constrains the control variable from changing abruptly and ensures practical feasibility. The sparse regularization term is represented by , which encourages the use of minimum control to save energy. By solving the control optimization model, the optimal control instruction is obtained to drive the actual reaction equipment to achieve precise control.

[0036] For example, the control optimization module 3 constructs a control optimization model according to the set target reaction state parameters, and solves the optimal control parameters in combination with the output results of the dynamic prediction module 2 to generate control instructions to drive the production equipment and achieve process control.

[0037] The application provides a pesticide production control system, which comprises the following steps: acquiring a multi-dimensional quality index of pesticide raw materials to be produced; the multi-dimensional quality index comprises effective component concentration, humidity and impurity proportion; based on the multi-dimensional quality index, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; and the state parameters of the pesticide raw materials to be produced in a production device during production are acquired; the state parameters comprise temperature, pH value, reaction speed and mixing degree; based on the state parameters, a weighted state feature is generated by using a weighted cumulative operation; based on the weighted state feature and the low-dimensional raw material feature vector, a dynamic prediction model is constructed by using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector; target reaction state parameters and environmental adjustment factors are determined; the target reaction state parameters comprise target concentration and yield of the pesticide to be produced; based on the twin prediction value, the target reaction state parameters and the environmental adjustment factors, a control optimization model is constructed; the control optimization model is used to generate control optimization parameters and control the production device to operate at the control optimization parameters; the control optimization parameters comprise heating power, feeding speed and stirring frequency. The application acquires the multi-dimensional quality index of the pesticide raw materials to be produced in real time, and obtains control optimization parameters in real time by using the control optimization model, so that the production device is controlled to operate at the control optimization parameters. The real-time index of the pesticide raw materials to be produced is monitored in real time, so that the operating parameters of the production device are adjusted in real time, the pesticide product after production is close to the target reaction state parameters, and the technical problems that the reaction process is unobservable, the control parameters are difficult to optimize in real time, and the quality fluctuation is difficult to predict and correct in time in the existing pesticide production process are solved.

[0038] The system further comprises: a correction module 4, which is in communication connection with the control optimization module 3 and the dynamic prediction module 2; the correction module 4 is configured to: after the production device operates at the control optimization parameters, acquiring real-time state parameters of the pesticide raw materials to be produced in the production device; after the control instruction is executed, the actual state value of the current reaction process is collected in real time .

[0039] the difference between the real-time state parameters and the twin prediction value corresponding to the control optimization parameters is calculated; by comparing the twin prediction value output by the dynamic prediction model in the dynamic prediction module 2, a prediction error is calculated.

[0040] Based on the difference value, a first-order gradient approximation method is used to calculate a weight update term; a first-order gradient approximation method is used to calculate a weight adjustment increment based on the prediction error, and a parameter projection operation is introduced to limit the updated prediction model weight in the feasible region of the weight parameter, thereby realizing model weight correction with structural constraints and enhancing the fitting ability and prediction accuracy of the control-response relationship.

[0041] wherein the structural constraint refers to a set of explicit limiting structures applied in the updating process of the weight parameter, which can be combined in one or more of the following constraint methods according to actual implementation requirements: one is to limit the boundary of the value range of a single weight element; two is to limit the norm of the overall weight matrix; three is to introduce sparsity constraints to make some weight elements tend to zero; four is to use low-rank constraints to simplify the model structure and improve the generalization ability. The above structural constraints are used to ensure the numerical stability, physical realizability and control safety of model updating. Specifically, a first-order gradient approximation method is used to construct a weight update term, which is: ; wherein, represents the weight parameter matrix of the dynamic prediction model at the t time; represents the learning rate; represents the real-time state parameter; represents the weight projection operation.

[0042] wherein: represents the weight parameter matrix of the dynamic prediction model at the t +1 time, which is the updating result obtained by correcting based on the error between the twin prediction value t and the actual reaction state after the model executes the control instruction and collects the actual reaction state at the time; represents the learning rate, which is obtained by experimental tuning or adaptive optimization algorithm, and the value range is (10 -5 , 1); represents the actual process state vector of the i batch pesticide reaction at time t , which is obtained by collecting through field sensor equipment, contains multiple dimensions such as temperature, pH value, reaction rate, product concentration, etc., has been normalized, and the unit is dimensionless; represents the first-order gradient approximation term; represents the weight projection operation; represents the feasible region of the weight parameter in the reaction dynamic prediction model, which is used to constrain the weight a range of values to possess numerical stability, physical feasibility, and prediction reliability. The range of values is usually defined by limiting the element size of the weights or limiting the overall norm. The specific values are derived from expert experience, actual process constraints, or statistical characteristics trained based on historical operation data to adapt to the reaction behavior and system dynamic changes under different working conditions.

[0043] The weight update term is input into the dynamic prediction model to update the twin prediction value.

[0044] For example, the correction module 4 collects the actual reaction process state value, i.e., the real-time state parameter, after the control instruction is executed, and compares the error with the twin prediction value generated by the dynamic prediction module 2; based on the error result, the weight of the dynamic prediction model in the dynamic prediction module 2 is updated to enhance the adaptability of the dynamic prediction model to control intervention and environmental fluctuations, improve the prediction accuracy and model generalization ability, thereby realizing the correction of the model weight with structural constraints, enhancing the fitting ability and prediction accuracy, and further increasing the control progress of the pesticide production control system, so that the product concentration and yield of the pesticide production meet the user's requirements.

[0045] As shown in FIG. 6, it is a flow chart of the operation of the early warning module in the present application. Figure 2

[0046] The system further comprises: An early warning module 5, which is communicatively connected to the raw material feature modeling module 1, the dynamic prediction module 2, and the control optimization module 3; the early warning module 5 is configured to: extract the low-dimensional raw material feature vector, the control optimization parameter, the twin prediction value, and the embedded feature vector of the weight update amplitude; the weight update amplitude is the difference value of the weight update term at adjacent time points; to realize real-time prediction and early warning of the final product quality, four types of key input information are fused, which are the low-dimensional raw material feature vector, the optimal control instruction vector, the twin prediction value, and the weight update amplitude (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, which is used to predict the specific values of the key quality indicators of the current batch of pesticide products, including effective ingredient purity, impurity content, and concentration compliance rate, etc.

[0047] ​combining any two of the embedding feature vectors to obtain an embedding feature vector group; constructing a cross attention mechanism in the embedding feature vector group to obtain a composite interaction matrix; dividing a tensor of the composite interaction matrix into three sub-tensors through a channel splitting operation; extracting a feature tensor of the sub-tensor and performing a cascading operation to obtain a fused feature tensor.

[0048] Specifically, the HDIF-Net takes four inputs as the basis, and in the input decoupling module, the internal convolution attention block is used to extract four types of embedding feature vectors , , , , which correspond to the embedding feature vector of the low-dimensional raw material feature vector , the embedding feature vector of the optimal control instruction , the embedding feature vector of the twin prediction value , and the embedding feature vector of the weight update amplitude of the dynamic prediction module 2 . Then in the cross-modal interaction module, any two of the four types of embedding feature vectors are combined to obtain six pairs of embedding feature vectors, and a cross attention mechanism is constructed for each pair of embedding feature vectors to finally obtain six groups of cross attention outputs, which are used to construct a composite interaction matrix H to mine high-order coupling relationships. The composite interaction matrix H is sent to the multi-scale residual fusion module. First, the tensor of the interaction matrix H , that is, a data structure organized in the form of a multi-dimensional array, commonly used in deep learning to describe high-dimensional features with multiple channels, spatial positions or batches, is uniformly divided into three sub-tensors in the channel dimension, which are respectively transmitted into the local tower, the mesoscale tower and the global tower for parallel feature extraction. Each sub-tensor sequentially passes through a convolution layer (such as 1x1, 3x3 or dilated convolution), a normalization layer (BatchNorm) and an activation function (such as ReLU) in the corresponding tower, and combines cross-layer residual connection for multi-level nonlinear expression learning. The local tower focuses on capturing weak interaction relationships between local channels, the mesoscale tower extracts medium-distance coupling patterns, and the global tower integrates long-distance context dependencies by expanding the receptive field. The output feature tensors of the three channels are subjected to channel-level cascading operations in the fusion stage to form the fused fused feature tensor .

[0049] extracting a composite feature related to quality index prediction in the fusion feature tensor; the quality index includes: effective component purity, impurity content, concentration compliance rate; based on the composite feature, a confidence score is obtained by using DropBlock and attention mechanism; it is judged whether the confidence score is less than a confidence alarm threshold, and if so, a warning is issued.

[0050] The warning module 5 is further configured to: based on the composite feature, a quality index prediction value is obtained by using an activation function; it is judged whether the quality index prediction value is located in a preset quality interval; if not, a warning is issued.

[0051] Specifically, in the heterogeneous decoding output module (including two decoding subnetworks of main path and side path), the fusion feature tensor is sent into the main path decoding subnetwork as input, first, after being flattened into a one-dimensional vector, it is input into a fully connected layer in turn to compress and extract a composite feature related to key quality index prediction. The last layer of decoding is the output layer of the constructed hierarchical decoupling interaction fusion network, the number of neurons of which is consistent with the dimension of the key quality index to be predicted (for example: three dimensions of effective component purity, impurity content, and concentration compliance rate), and the output layer adopts a Sigmoid activation function, so that the output key quality index prediction value is normalized to the interval [0, 1], which is convenient for standard quality interval comparison and confidence alarm threshold judgment. The standard quality interval is set according to the process specification in the long-term production process and the historical quality data statistical result, and is usually constructed in the form of mean ± standard deviation, focusing on the "compliance" of the prediction value. The side path decoding subnetwork introduces DropBlock and attention mechanism, and outputs the confidence score of the current time, which is used to quantify the credibility of the key quality index prediction value generated by the main path decoding subnetwork. If it is detected that (where is a confidence alarm threshold, which is set by the lower quantile experience of the historical confidence score sample of , for example, the 10th percentile, focusing on the "credibility" of the prediction result), the abnormal quantity warning mechanism is triggered.

[0052] 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.

[0053] 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.

[0054] 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: ; 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.

[0055] Specifically, represents the prediction function response correction strength driven by the deviation between the current time model output and the true state. By analyzing the gradient response distribution on different input dimensions, it can be determined which factors mainly cause the prediction error: if the gradient is mainly concentrated in the raw material feature dimension, it means that raw material fluctuation is the main reason; if it is concentrated in the control variable dimension, it means that control deviation is the dominant factor; if the response of multiple dimensions is not significant, it may be caused by environmental disturbance or unmodeled external factors. This analysis process provides data support for subsequent control instruction optimization, and realizes more targeted strategy adjustment.

[0056] For example, the early warning module 5 fuses multi-source information such as low-dimensional raw material feature vectors, control instructions, twin prediction values, and the weight update amplitude of the dynamic prediction module 2, constructs a fusion type quality prediction model, outputs product key quality index prediction values, compares them with the standard quality interval to determine compliance, combines confidence score and threshold comparison to evaluate credibility, and triggers early warning and reverts control instruction vectors if any abnormality occurs. The deviation result is fed back to the control optimization module 3 to form a closed-loop feedback control.

[0057] The second aspect of the present application provides a pesticide production control method applied to the pesticide production control system of any of the above embodiments, comprising: Obtaining multi-dimensional quality indicators of raw materials to be produced; the multi-dimensional quality indicators include effective component concentration, humidity, and impurity ratio; Based on the multi-dimensional quality indicators, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; Obtaining state parameters of the raw materials to be produced in the production equipment during the production process; the state parameters include temperature, pH value, reaction speed, and mixing degree; Based on the state parameters, a weighted state feature is generated by using a weighted cumulative operation; Based on the weighted state feature and the low-dimensional raw material feature vector, a dynamic prediction model is constructed by using a normalization processing operation; the dynamic prediction model is used to generate twin prediction values according to the weighted state feature and the low-dimensional raw material feature vector; Determining target reaction state parameters and environmental adjustment factors; the target reaction state parameters include target concentration and yield of the pesticide to be produced; Based on the twin prediction values, target reaction state parameters, and environmental adjustment factors, a control optimization model is constructed; Using the control optimization model, control optimization parameters are generated and the production equipment is operated with the control optimization parameters; the control optimization parameters include heating power, feeding speed, and stirring frequency. ​

[0058] It should be noted that the effects of the above method embodiments can be seen from the effects of the above system embodiments, which will not be repeated here.

[0059] The above detailed description has further explained the purposes, technical solutions and beneficial effects of the embodiments of the present application. It should be understood that the above is only a specific implementation of the embodiments of the present application and is not intended to limit the protection scope of the embodiments of the present application. Any modification, equivalent replacement, improvement, etc. made on the basis of the technical solutions of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

1. A pesticide production control system characterized by comprising: The system comprises: a raw material feature modeling module (1), a dynamic prediction module (2), and a control optimization module (3) connected in communication; the raw material feature modeling module (1) is configured to: obtain a multi-dimensional quality index of a pesticide raw material to be produced; the multi-dimensional quality index comprises: effective component concentration, humidity, and impurity ratio; based on the multi-dimensional quality index, a low-dimensional raw material feature vector is obtained by using a nonlinear embedding function; the nonlinear embedding function is composed of a weight matrix and a bias term; the dynamic prediction module (2) is configured to: obtain a state parameter of the pesticide raw material to be produced in a production device during production; the state parameter comprises: temperature, pH value, reaction speed, and mixing degree; based on the state parameter, a weighted state feature is generated by using a weighted cumulative operation; based on the weighted state feature and the low-dimensional raw material feature vector, a dynamic prediction model is constructed by using a normalization processing operation; the dynamic prediction model is used to generate a twin prediction value according to the weighted state feature and the low-dimensional raw material feature vector; the control optimization module (3) is configured to: determine a target reaction state parameter and an environmental adjustment factor; the target reaction state parameter comprises: a target concentration of the pesticide to be produced and a yield; based on the twin prediction value, the target reaction state parameter, and the environmental adjustment factor, a control optimization model is constructed; by using the control optimization model, a control optimization parameter is generated and the production device is controlled to operate at the control optimization parameter; the control optimization parameter comprises: heating power, feeding speed, and stirring frequency.

2. The agricultural chemical production control system according to claim 1, characterized by The raw material feature modeling module (1) is further configured to: construct a high-dimensional quality index vector according to the multi-dimensional quality index; perform feature mapping on the high-dimensional quality index vector by using a nonlinear embedding function to obtain a low-dimensional raw material feature vector; the nonlinear embedding function is composed of a weight matrix and a bias term; the low-dimensional raw material feature vector is: ； wherein a weight matrix representing a neural network embedding layer; a bias vector representing a bias vector of the neural network embedding layer; a transpose of a high-dimensional quality indicator vector of the batch of raw materials; a bias vector representing a bias vector of the neural network embedding layer.

3. The agricultural chemical production control system according to claim 1, characterized by The dynamic prediction module (2) is further configured to: construct a state observation value according to the state parameter; the state observation value is a state parameter combination vector at any time; perform a weighted cumulative operation on the state observation value to generate a weighted state feature; the weighted state feature is: ； In the formula, characterized by the first i batch reaction process, at the historical time state observation value; characterized by the current time; ; characterized by the time decay coefficient.

4. The agricultural chemical production control system according to claim 1, characterized by The control optimization parameter is: ； In the formula, characterized as a target reaction state parameter; characterized as a control response network; characterized as a current control instruction vector to be optimized, including heating power, raw material adding speed, stirring frequency; characterized as a twin prediction value; characterized as a control instruction vector at the last time; characterized as the i weight vector of each component of the target reaction state parameter of the batch at time t ; characterized as the i environmental adjustment factor of the batch at time t ; characterized as a weight coefficient of a control smoothing term; characterized as a weight coefficient of a control sparsity regularization term; characterized as the i control optimization parameter of the production equipment at time .

5. The agricultural chemical production control system according to claim 1, characterized by The system further comprises: a correction module (4) connected in communication with the control optimization module (3) and the dynamic prediction module (2); the correction module (4) is configured to: after the production device operates at the control optimization parameter, real-time state parameters of the pesticide raw material to be produced in the production device are obtained; a difference between the real-time state parameters and the twin prediction value corresponding to the control optimization parameter is calculated; based on the difference, a weight update term is calculated by using a first-order gradient approximation method; the weight update term is input into the dynamic prediction model to update the twin prediction value.

6. The agricultural chemical production control system according to claim 5, characterized by The weight update term is: ； 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.

7. The agricultural chemical production control system according to claim 1, characterized by The system further comprises: An early warning module (5) is communicatively connected to the raw material feature modeling module (1), the dynamic prediction module (2), and the control optimization module (3). The early warning module (5) is configured to: extract the low-dimensional raw material feature vector, the control optimization parameter, the twin prediction value, and the embedded feature vector of the weight update amplitude; the weight update amplitude is the difference value of the weight update item at adjacent time points; combine any two embedded feature vectors to obtain an embedded feature vector group; build 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 tensor of the sub-tensor and perform a concatenation operation to obtain a fusion feature tensor; extract the composite features related to the quality index prediction from the fusion feature tensor; the quality index includes effective ingredient purity, impurity content, and concentration compliance rate; based on the composite features, use DropBlock and attention mechanism to obtain a confidence score; determine whether the confidence score is less than a confidence alarm threshold; if yes, issue a warning.

8. The agricultural chemical production control system according to claim 7, characterized by The early warning module (5) is also configured to: based on the composite features, use an activation function to obtain a quality index prediction value; determine whether the quality index prediction value is within a preset quality interval; if not, issue a warning.

9. A pesticidal production control system according to either one of claims 7 or 8, wherein, After the step of determining whether the quality index prediction value is within a preset quality interval, the method further includes: determine the weight change amplitude; the weight change amplitude is: ； obtain the gradient response distribution of the weight change amplitude on different input dimensions; obtain a prediction error factor report according to the gradient response distribution; the prediction error factor report is used to display the dominant factor causing the error of the control optimization parameter.

10. A method for controlling the production of a pesticide, applied to the production control system for a pesticide according to any one of claims 1 to 9, characterized by, includes: obtain the multi-dimensional quality index of the pesticide raw material to be produced; the multi-dimensional quality index includes effective ingredient concentration, humidity, and impurity ratio; based on the multi-dimensional quality index, use a nonlinear embedding function to obtain a low-dimensional raw material feature vector; the nonlinear embedding function is composed of a weight matrix and a bias term; obtain the state parameters of the pesticide raw material to be produced in the production equipment during production; the state parameters include temperature, pH value, reaction speed, and mixing degree; based on the state parameters, use a weighted cumulative operation to generate a weighted state feature; based on the weighted state feature and the low-dimensional raw material feature vector, use a normalization processing operation to build a dynamic prediction model; the dynamic prediction model is used to generate a twin prediction value based on the weighted state feature and the low-dimensional raw material feature vector; determine the target reaction state parameter and the environmental adjustment factor; the target reaction state parameter includes the target concentration and yield of the pesticide to be produced; based on the twin prediction value, the target reaction state parameter, and the environmental adjustment factor, build a control optimization model; use the control optimization model to generate control optimization parameters and control the production equipment to operate with the control optimization parameters; the control optimization parameters include heating power, feeding speed, and stirring frequency.

Images