A pesticide preparation cross-contamination risk early warning system and storage medium

Abstract

This invention provides a pesticide formulation cross-contamination risk early warning system and storage medium, comprising: acquiring the physicochemical parameters, toxicological parameters, subsequent residue allowable limits, and cleanup procedure data of the current batch of formulations; retrieving similar historical model calibration coefficients; and calculating the cleanup residue adsorption index by combining molecular descriptors, formulations, and cleanup solvent parameters; constructing a cross-contamination feature vector and inputting it into a graph attention network, outputting scenario contribution weights and prediction calibration coefficients; using variational inference to obtain the posterior probability distribution of residues, and using the expected value of the posterior probability distribution as the model-predicted residue amount; after obtaining the actual residue test results, correcting the prediction calibration coefficients based on the deviation between the actual residue test results and the model-predicted residue amount, generating a two-dimensional risk signature, and triggering an early warning when the two-dimensional risk signature exceeds a preset boundary.

Description

Technical Field

[0001] This application belongs to the field of risk warning, and in particular relates to a risk warning system and storage medium for cross-contamination of pesticide formulations. Background Technology

[0002] In pesticide formulation production, cross-contamination is influenced by multiple factors, including the physicochemical properties and toxicological characteristics of preceding formulations, the residue limits of subsequent products, and cleaning procedures. Due to the complex interaction mechanisms between different formulations and active ingredients, current cleaning assessments largely rely on manual experience and periodic sampling and testing, making it difficult to quantify the cleaning difficulty and potential residue levels of preceding substances. Furthermore, facing a wide variety of interconnected risk factors, traditional methods struggle to establish a systematic correlation assessment framework. Operational deviations or parameter fluctuations during production can easily lead to delayed identification of cross-contamination risks, potentially resulting in product non-compliance or recalls. Applying graph attention networks and probabilistic inference algorithms to production risk prediction can extract implicit correlations between risk features through inter-node message passing and utilize probabilistic models to characterize uncertainties in the assessment process, thereby improving the intelligence level of cross-contamination risk early warning. Graph network models lack a calibration loop that links with historical real test data, making it difficult to update model calibration coefficients in multiple production iterations, which makes prediction results prone to deviating from actual production scenarios. At the same time, existing probabilistic inference methods cannot use the factor weights output by deep learning and measurement uncertainty together to construct the prior distribution, resulting in unstable generation of residual posterior probability distribution and two-dimensional risk signature, which in turn causes problems such as blurred warning boundaries, delayed warnings, and false alarms and missed alarms. Summary of the Invention

[0003] To address the problems in existing technologies, such as insufficient integration of risk factors, lack of historical test data for calibration and closed-loop, difficulty in quantifying measurement uncertainties and generating two-dimensional risk signatures, which lead to blurred warning boundaries and false alarms or missed alarms, this invention proposes a pesticide formulation cross-contamination risk early warning system and method.

[0004] In a first aspect, the present invention proposes a pesticide formulation cross-contamination risk early warning system, comprising: The calculation module is used to acquire the current batch data, including the physicochemical parameters of the preceding formulation, the toxicological parameters of the preceding formulation, the residue limits of the subsequent formulation, and the clearance procedure data; retrieve the historical model calibration coefficient with the highest similarity to the current batch from the database; and calculate the clearance residue adsorption index by combining the historical model calibration coefficient, the preceding molecule descriptor, the dosage form, and the Hansen solubility parameter of the clearance solvent. The output module integrates the preceding toxicological endpoint value, the subsequent residue allowable limit, and the cleanup residue adsorption index to form a cross-contamination feature vector; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through a message passing mechanism. The derivation module is used to establish a prior distribution by combining the scenario contribution weight vector and measurement uncertainty, derive the posterior probability distribution of cross-contamination residue based on the observed values ​​of each risk factor through a variational inference algorithm, and use the expected value of the posterior probability distribution as the model's predicted residue amount; after obtaining the actual residue test results after the current batch is cleared and verifying the data validity, the prediction calibration coefficient is corrected based on the deviation between the actual residue test results and the model's predicted residue amount, and the corrected prediction calibration coefficient is used as the new historical model calibration coefficient to update the database; The early warning module is used to calculate the cumulative probability of exceeding the residual allowable limit and the ratio of the residual posterior standard deviation to the residual allowable limit based on the posterior probability distribution, thus forming a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

[0005] On the other hand, the present invention also proposes a method for early warning of cross-contamination risk of pesticide formulations, including: Obtain current batch data, including physicochemical parameters of the preceding formulation, toxicological parameters of the preceding formulation, residue limits of the subsequent formulation, and clearance procedure data; retrieve the historical model calibration coefficient with the highest similarity to the current batch from the database; and calculate the clearance residue adsorption index by combining the historical model calibration coefficient, the preceding molecule descriptor, the dosage form, and the Hansen solubility parameter of the clearance solvent. The cross-contamination feature vector is constructed by integrating the preceding toxicological endpoint value, the subsequent residue allowable limit and the cleanup residue adsorption index; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through the message passing mechanism. A prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty. Based on the observed values ​​of each risk factor, the posterior probability distribution of cross-contamination residue is derived using a variational inference algorithm. The expected value of the posterior probability distribution is used as the model's predicted residue. After obtaining the actual residue test results after the current batch is cleared and verifying the data validity, the prediction calibration coefficient is corrected based on the deviation between the actual residue test results and the model's predicted residue. The corrected prediction calibration coefficient is then used as the new historical model calibration coefficient and updated in the database. Based on the posterior probability distribution, the cumulative probability of exceeding the residual allowable limit and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit are calculated to form a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

[0006] This invention integrates the physicochemical and toxicological parameters of preceding formulations, the allowable limits for subsequent residues, and data from cleanup procedures. It scientifically calculates the cleanup residue adsorption index by combining the molecular-level characteristics of Hansen's solubility parameters. Using a graph attention network with risk factors as graph nodes, it explores and processes the topological relationships between factors, outputting scenario contribution weights and predicted calibration coefficients. By using real test results to validate and update the database, a model calibration feedback mechanism based on real test results is formed. By combining measurement uncertainty with variational inference algorithms to derive the posterior probability distribution and construct a two-dimensional risk signature, the accuracy and stability of pesticide cross-contamination risk early warning are improved, and the risk of false alarms or missed alarms caused by single threshold judgments is reduced. Attached Figure Description

[0007] Figure 1 This is a posterior distribution map of the residual amount; Figure 2 This is a diagram illustrating the cumulative probability of exceeding the allowable residual limit. Figure 3 This is a diagram illustrating the performance comparison of the models. Detailed Implementation

[0008] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this specification. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this specification as detailed in the appended claims.

[0009] It should be understood that the terms “comprising” and “having”, and any variations thereof, in the embodiments of this specification are intended to cover but not exclude inclusion. For example, a product or device that includes a series of components is not necessarily limited to those components that are explicitly listed, but may include other components that are not explicitly listed or that are inherent to such product or device.

[0010] In this application, a pesticide formulation cross-contamination risk early warning system includes: The calculation module is used to acquire the current batch data, including the physicochemical parameters of the preceding formulation, the toxicological parameters of the preceding formulation, the residue limits of the subsequent formulation, and the clearance procedure data; retrieve the historical model calibration coefficient with the highest similarity to the current batch from the database; and calculate the clearance residue adsorption index by combining the historical model calibration coefficient, the preceding molecular descriptor, the dosage form, and the Hansen solubility parameter of the clearance solvent.

[0011] The system reads the current batch data table exported from the Manufacturing Execution System (MES), extracts the water solubility and octanol-water partition coefficient of the preceding formulation as physicochemical parameters, extracts the median lethal dose (LD50) and no significant adverse effect level (NOLAD) of the preceding formulation as toxicological parameters, extracts the residual allowable limit (LAD) of the subsequent formulation, and extracts the cleaning time, cleaning temperature, and cleaning agent flow rate from the cleaning procedure data. At least some parameters from the preceding formulation's water solubility, octanol-water partition coefficient, LD50, NOLAD, subsequent residual allowable limit, cleaning time, cleaning temperature, cleaning agent flow rate, preceding molecular descriptor, and uncalibrated Hansen solubility parameter distance are concatenated in a preset order to form the current batch feature vector. The system then reads the corresponding historical batch feature vectors of the same dimension from the database. After normalizing both the current batch feature vector and the historical batch feature vectors, the cosine similarity between them is calculated using the cosine similarity formula. All historical batch feature vectors are iterated through, and the historical batch with the highest cosine similarity is identified as the most similar historical batch. The corresponding historical model calibration coefficient is then retrieved.

[0012] The Descriptors module of the RDKit cheminformatics toolkit is used to calculate the topological polar surface area, octanol-water partition coefficient, and number of rotatable bonds of the active ingredient in the preceding formulation, which are then used as preceding molecular descriptors. Each preceding molecular descriptor is normalized according to the historical sample value range or preset maximum and minimum values, so that each descriptor is converted into a dimensionless value in the range of 0 to 1. Then, the normalized topological polar surface area, octanol-water partition coefficient, and number of rotatable bonds are weighted and summed according to preset weights to obtain the comprehensive value of the preceding molecular descriptor used to characterize the polarity, hydrophobicity, and structural flexibility of the preceding formulation molecule.

[0013] The three-dimensional Hansen solubility parameters of the preceding formulation and the clearing solvent, including dispersion forces, polar forces, and hydrogen bonding forces, were calculated using the HSPiP software interface based on the group contribution method. The solubility parameter distance was obtained by calculating the Euclidean distance between the preceding formulation and the clearing solvent in three-dimensional space. The difference in the solubility parameter was dimensionlessized. The historical model calibration coefficient, the comprehensive value of the preceding molecule descriptor, the dimensionless solubility parameter difference, and the empirical constant of the dosage form adhesion were weighted and fused to obtain the clearing residual adsorption index, which represents the removal of resistance. The clearing residual adsorption index is positively correlated with the dimensionless solubility parameter distance and the empirical constant of the dosage form adhesion.

[0014] The group contribution method refers to the method of dividing a molecule into several structural groups with known property contribution values, and estimating the overall physicochemical properties of the molecule by accumulating or weighted accumulating the contributions of each group.

[0015] In some implementations, obtaining the current batch of data includes: The system calls the production scheduling history work order database of the enterprise production control platform via the application programming interface. Use regular expressions to extract object codes containing the previous production completion identifier from work order records, and obtain the sequence of subsequent product codes to be produced; Based on the extracted object code and product code sequence, a query is initiated to the pesticide management cloud platform to extract the corresponding toxicological endpoint value, residue limit standard, median lethal dose, level of no visible harmful effect, as well as the cleaning solvent parameters and equipment volume parameters in the cleaning operation procedure, and store them in the system cache.

[0016] The backend calls the historical work order database of the enterprise manufacturing execution system based on MySQL architecture via HTTP protocol using a RESTful interface. When extracting text-level production records, it uses preset regular expressions to traverse and search the structured work order table to extract the unique pesticide object code with the completion identifier of the previous product, and associates it to obtain the sequence of object product codes that will be launched in the next batch, establishing a key-value pair structure in memory to represent the production variety switching sequence.

[0017] Using the aforementioned object codes and product code sequences as query parameters, a query request is sent to a remote agricultural pesticide management cloud platform via a secure encrypted communication protocol, returning the corresponding multi-dimensional physicochemical and toxicological JSON format data package. The preferred range and example values ​​for the extracted parameters include: toxicological median lethal dose, level of no visible adverse effects, and residue limits for subsequent formulations. Additionally, the type of clearing solvent and the corresponding reactor volume parameters are simultaneously obtained. The extracted full data is serialized and stored in a Redis high-concurrency memory cache cluster with a 24-hour expiration time, supporting rapid retrieval of cached data by the model.

[0018] In some implementations, retrieving the historical model calibration coefficients with the highest similarity to the current batch from the database includes: Extract the molecular polarity and solubility feature values ​​of the current batch and construct the target feature query vector; Traverse all historical batch records in the database that contain historical model calibration coefficients, and extract the corresponding historical feature vectors; After standardizing the target feature query vector and the historical feature vector, the feature deviation distance is calculated based on the multidimensional Euclidean distance formula, where the feature deviation distance is the square root of the sum of the squares of the differences of each dimension component of the standardized target feature query vector and the historical feature vector. Sort all calculated feature deviation distances from smallest to largest, extract the first historical batch record with the smallest distance, and extract and cache the historical model calibration coefficient corresponding to the record as the calibration coefficient of the current batch.

[0019] The system retrieves the attributes of preceding pesticide molecules in the current batch from the cache, quantifies and extracts the logarithms of their topological polar surface area and oil-water partition coefficient, and concatenates them according to a fixed dimension to obtain a two-dimensional target feature query vector such as [TPSA, logP]. Simultaneously, using a cursor, it scans the underlying PostgreSQL historical knowledge graph database row by row, reading all historical batch records from the past 5 years with manually calibrated or experimentally verified calibration coefficients into memory, and extracting historical feature vectors with consistent structures.

[0020] Before distance calculation, the Z-score zero-mean normalization method is applied to transform both the target query vector and the historical vector set into dimensionless vectors that follow a standard normal distribution. For example, after normalization, the current vector becomes [0.55, -0.21], and a certain historical vector is [0.50, -0.15]. Substituting them into the multidimensional Euclidean distance formula, the square root of the sum of squares of the corresponding component differences is calculated, yielding a feature deviation distance of 0.078. After the computing unit traverses and calculates the feature deviation distance array of all historical batches, it calls the quicksort algorithm to sort them in ascending order and extracts the historical batch with index 0, i.e., the smallest deviation distance. The historical model calibration coefficients bound to this matching record are extracted and cached in the process context as floating-point variables for the next stage of clearing residual adsorption index calibration calculation.

[0021] In some implementations, the calculation of the residual adsorption index for clearing, combining the historical model calibration coefficients, preceding molecule descriptors, dosage form, and Hansen's solubility parameter of the clearing solvent, includes: Hansen solubility parameters, including dispersion force parameters, polarity parameters, and hydrogen bond parameters, were obtained for the target compound of the preceding formulation and the clearing solvent, respectively. Calculate the solubility parameter distance between the preceding formulation and the clearing solvent in Hansen space. The solubility parameter distance is the square root of the sum of four times the square of the difference in dispersion force parameter, the square of the difference in polarity parameter, and the square of the difference in hydrogen bond parameter between the preceding formulation and the clearing solvent. The preceding molecular descriptor is normalized and a comprehensive value of the preceding molecular descriptor is generated. The solubility parameter distance is dimensionless. The historical model calibration coefficient, the comprehensive value of the preceding molecular descriptor, the dimensionless solubility parameter distance, and the empirical constant of the dosage form adhesion are weighted and fused to obtain the calibrated clearing residual adsorption index. The residual adsorption index after cleaning is used as a key parameter to represent removal resistance and stored in the inference calculation module.

[0022] The three-dimensional Hansen solubility parameters of the precursor formulation components and the cleaning solvent at standard atmospheric pressure and 25°C were extracted from the built-in thermodynamic property library. Specifically, the three core parameters extracted typically have the following numerical ranges: dispersion force parameter... Between 12 and 25 Between, polarity parameters From 0 to 20 Between, hydrogen bond parameters Between 0 and 30 Between. For example, the parameters for extracting the preceding avermectin are [ =18.5, =4.2, =8.5], the cleaning solvent parameters are [ =15.0, =12.5, =20.0].

[0023] In three-dimensional Hansen space, Euclidean space metric is performed based on the improved dissolution sphere equation, and the result is substituted into the formula. Calculate the compatibility distance. Substitute the example values ​​above to obtain the solubility parameter distance. 15.82 A smaller distance indicates better dissolution performance of the cleaning agent on the preceding substances, while a larger distance indicates that the preceding substances are relatively less easily dissolved and removed by the clearing solvent, resulting in higher removal resistance. First, the solubility parameter distance is dimensionlessly processed according to a preset reference distance or historical sample range to obtain the dimensionless solubility parameter distance. Then, the combined value of the preceding molecule descriptor, the dimensionless solubility parameter distance, and the empirical constant of the dosage form adhesion are weighted and summed according to preset weights, and multiplied by the historical model calibration coefficient to obtain the calibrated clearing residual adsorption index. For example, it can be expressed as... Where I represents the residual adsorption index, K represents the historical model calibration coefficient, D represents the normalized combined value of the preceding molecular descriptors, R represents the dimensionless solubility parameter distance, and F represents the empirical constant of dosage form adhesion. , , The weight coefficients are non-negative and =1. This yields the residual adsorption index for the cleanup process, which is then pushed into the message queue in a high-precision double-precision floating-point format.

[0024] The output module integrates the preceding toxicological endpoint value, the subsequent residue allowable limit, and the cleanup residue adsorption index to form a cross-contamination feature vector; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through a message passing mechanism.

[0025] The preceding toxicological endpoint, the subsequent residue tolerance limit, and the adsorption index of the cleanup residue calculated in the previous step are sequentially concatenated using a fully connected method and then normalized to zero mean to generate a one-dimensional cross-contamination feature vector. A graph attention network model is constructed, and the indicators of each dimension in the cross-contamination feature vector are set as node features of the graph. The adjacency matrix between nodes is constructed using the threshold truncation method of the Pearson correlation coefficient matrix to represent the interdependent topology of risk factors.

[0026] The constructed node feature matrix and adjacency matrix are input into the GATConv layer. In the message passing mechanism, the LeakyReLU activation function is used to calculate the self-attention coefficients of the center node and neighboring nodes. The self-attention coefficients are normalized by the softmax function. The features of the neighboring nodes are weighted and summed according to the normalized attention coefficients to update the features of the center node. After being fused by the multi-head attention mechanism, the feature is connected to the global average pooling layer for graph-level feature aggregation.

[0027] The aggregated graph-level features are input into two parallel fully connected multilayer perceptron networks. The first multilayer perceptron network uses softmax as the output layer activation function to generate the percentage weights corresponding to each risk factor to form the scenario contribution weight vector. The second multilayer perceptron network uses the scalar value output by the linear layer as the prediction calibration coefficient.

[0028] In some implementations, the step of inputting the cross-contamination feature vector into a graph attention network model with risk factors as graph nodes, and outputting a scenario contribution weight vector and prediction calibration coefficients through a message passing mechanism, includes: Each risk factor in the cross-contamination feature vector is used as a graph node, and the numerical and attribute features of each risk factor are used to construct the initial node feature vector. Based on the association rules between each risk factor, the connection edges between nodes are constructed to establish the graph topology. In the constructed interactive network, for each node and its neighboring nodes, the attention coefficients between nodes are calculated by sharing the weight matrix and the attention weight vector; Based on the calculated attention coefficients, the features of neighboring nodes are weighted and aggregated, and the updated node features are obtained through an activation function. After multi-layer network iteration and updates, the aggregated full-graph node features output a context contribution weight vector, and the aggregated full-graph node features are mapped to prediction calibration coefficients through a fully connected layer.

[0029] The graph attention network takes as input a graph node feature matrix and an adjacency matrix transformed from cross-contamination feature vectors, and outputs a context contribution weight vector and prediction calibration coefficients. The network structure includes a graph feature construction module, a multi-layer graph attention hidden layer, a global pooling layer, and a two-branch fully connected output module.

[0030] In the preprocessing stage of the graph data, the levels of no visible harmful effects of the preceding toxicology, the permissible limits of subsequent product residues, and the adsorption index of the cleaning residues in the cross-contamination feature vector are mapped to three discrete graph nodes. Each node is assigned an initial feature vector with dimensions such as 16. This vector integrates scalar values ​​normalized to the 0-1 range and variable classification labels, such as toxicity parameters or physical limits, represented using one-hot encoding.

[0031] Based on pre-defined pharmaceutical cross-contamination process knowledge, a Boolean adjacency matrix is ​​generated to construct undirected connecting edges, such as connecting adsorption nodes to residual allowable limit nodes, establishing a fully connected graph topology interaction network containing specific edge connection logic. After forward propagation into the graph attention network, in each graph attention hidden layer, a shared 16×32 linear learnable weight matrix is ​​applied to upscale the node features. For the target node and its first-order neighbor nodes, the upscaled features are concatenated and a dot product operation is performed with the attention weight vector of dimension 64×1.

[0032] The dot product result is then subjected to an exponential probability normalization using a modified linear unit activation function with leakage, with the negative half-axis slope parameter preferably set to 0.2. This results in inter-node attention coefficients of the form 0.2, 0.5, and 0.3.

[0033] The resulting tensor is split into two branches. One branch maps the context contribution weight vector representing the importance of each factor dimension through a softmax function layer, ensuring that each weight value is between 0 and 1 and the weight sum is 1. For example, the outputs are 0.45, 0.35, and 0.20. The other branch performs regression through a two-layer fully connected layer of a multilayer perceptron containing 64 and 32 hidden neurons, outputting a prediction calibration coefficient such as a continuous scalar value of 1.18, which is then simultaneously used in the next stage of calculation.

[0034] The derivation module is used to establish a prior distribution by combining the scenario contribution weight vector and measurement uncertainty, derive the posterior probability distribution of cross-contamination residue based on the observed values ​​of each risk factor through a variational inference algorithm, and use the expected value of the posterior probability distribution as the model's predicted residue amount; after obtaining the actual residue test results after the current batch is cleared and verifying the data validity, the prediction calibration coefficient is corrected based on the deviation between the actual residue test results and the model's predicted residue amount, and the corrected prediction calibration coefficient is updated to the database as a new historical model calibration coefficient.

[0035] The scenario contribution weight vector is weighted and summed with the normalized risk factor observations to obtain a standardized risk characterization value. This value is then combined with the residue allowable limit to calculate the prior mean of cross-contamination residue. The relative standard deviation of the high-performance liquid chromatography (HPLC) as specified in the equipment cleaning validation procedure is converted into a standard deviation with the same dimensions as the residue, and this standard deviation is assigned as the scaling parameter of the prior distribution, thus establishing a normal prior distribution of cross-contamination residue. Subsequently, the risk factor observations of each dimension are used as conditional observation data. A guided function based on a parameterized distribution is constructed as an approximate posterior distribution. The Adam optimizer is used, with the Trace_ELBO lower bound of evidence as the loss function. The guided function parameters are iteratively updated through backpropagation to maximize the lower bound of evidence until the loss function converges, thereby obtaining an approximate posterior probability distribution model of cross-contamination residue.

[0036] The current batch of equipment eluent is analyzed using high-performance liquid chromatography-tandem mass spectrometry (HPLC-MS / MS) to obtain the actual residue test results. When the actual residue test results meet the quality control and data validity requirements of the detection method, the test results are deemed valid. The expected value of the approximate posterior probability distribution of the cross-contamination residue is used as the model's predicted residue amount. Based on the deviation between the actual residue test results and the model's predicted residue amount, the prediction calibration coefficients output by the graph attention network are corrected. The corrected prediction calibration coefficients, along with the current batch identifier, are added as new records to the historical model calibration coefficient data table.

[0037] In some implementations, the prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty, and the posterior probability distribution of cross-contamination residue is derived based on the observed values ​​of each risk factor using a variational inference algorithm, including: The allowable deviation of the measuring equipment is read as the measurement uncertainty, and the prior distribution of the latent variable representing the implicit state of cross-contamination is constructed by combining it with the scenario contribution weight vector; The observation dataset is constructed by obtaining the variables representing production conditions from the observed values ​​of each risk factor, excluding the inherent biochemical toxicity constants and regulatory standards, and constructing a variational distribution to approximate the true posterior distribution. The variational distribution parameters are optimized by maximizing the lower bound of evidence. The lower bound loss function is set as the difference between the log-expected value of the joint distribution and the log-expected value of the variational distribution. The gradient descent algorithm is used to iteratively update the variational parameters, reduce the KL divergence between the variational distribution and the true posterior distribution, until the lower bound of evidence converges or the preset maximum number of iterations is reached, and the converged latent variable posterior distribution is obtained. The latent variables are transformed into non-negative cross-contamination residues through a pre-defined generation mapping function with non-negative activation functions at the ends, and the posterior probability distribution of the cross-contamination residues is output.

[0038] The variational autoencoder network takes an observation dataset representing production conditions as input and outputs a probability density function of the true amount of cross-contamination residue. The network structure includes an inference neural network as the encoder, a multivariate Gaussian distribution building block, and a decoder generator network containing multilayer networks and terminal activation functions.

[0039] The fixed permissible relative standard deviation of the high-performance liquid chromatograph used in the validation and testing phase is obtained from the experimental equipment library, for example, the preset measurement uncertainty is ±5%. Combining the scenario contribution weight vectors of 0.45, 0.35, and 0.20 output from the previous phase, the latent variable Z is initialized within the framework of the variational autoencoder, preferably in 8-dimensional space, so that its prior distribution follows a parameterized multivariate Gaussian distribution. The scaling of the variance matrix is ​​constrained by the aforementioned relative standard deviation error boundary of the equipment measurement, and the relative standard deviation is converted into the corresponding absolute standard deviation according to the residual prediction scale.

[0040] In constructing the observation dataset, constant labels with statically unchanged half-lethal doses or residual allowable limits are removed. Operating condition variables that can represent the instantaneous production state are selected, such as the actual flushing fluid volume of 800 liters, the in-situ cleaning and flushing duration of 45 minutes, and the pipeline flow velocity of 2.5 meters per second. These are used as the condition vector X and input into the inference neural network model to predict the Gaussian variational posterior distribution used to approximate the true unknown distribution. .

[0041] During network backpropagation optimization, the training process revolves around maximizing the lower bound of evidence. The objective loss function is constructed as the data reconstruction error term. The difference between the KL divergence regularization term KL(q(Z|X)||P(Z)) and the KL divergence regularization term. A moment estimation gradient descent optimizer is used, with the learning rate set to 0.0005 and the batch size to 32. The model weights are continuously adjusted using reparameterization techniques. After approximately 1500 to 2000 iterations, when the calculated change in the lower bound of evidence is below a set threshold for 10 consecutive training epochs... The convergence of the latent space is determined at that time.

[0042] The hidden tensor Z is extracted from the converged variational distribution and input into the decoder generator network containing a smooth, positive, non-negative activation function. This activation operation derives and outputs a non-negative cross-contamination true residue probability density function that follows a log-normal distribution and falls entirely within the range of 0 to positive infinity; for example, a statistical distribution parameter set with a mean of 0.02 mg / kg and a standard deviation of 0.005. The frequency statistics of this posterior distribution are as follows... Figure 1 As shown, the concentration trend and dispersion of residual amounts are presented.

[0043] The early warning module is used to calculate the cumulative probability of exceeding the residual allowable limit and the ratio of the residual posterior standard deviation to the residual allowable limit based on the posterior probability distribution, thus forming a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

[0044] The probability of being below the residual allowable limit is calculated by calling the cumulative distribution function of the posterior probability distribution model generated by variational inference. The cumulative probability of exceeding the residual allowable limit is obtained by subtracting this probability from 1. At the same time, the posterior standard deviation of the residual amount is calculated based on the posterior probability distribution, and the ratio of the posterior standard deviation to the statutory residual allowable limit is calculated to obtain the uncertainty limit ratio.

[0045] The calculated cumulative probability of exceeding the residual allowable limit and the uncertainty limit ratio are packaged and merged to generate a two-dimensional risk signature in binary format. This two-dimensional risk signature is monitored in real time. A preset boundary is set as follows: the cumulative probability of exceeding the residual allowable limit is greater than 5% or the uncertainty limit ratio is greater than a preset uncertainty upper limit, such as 0.2. When any dimension value in the two-dimensional risk signature meets the above logical judgment condition, an alarm message containing the current batch number and the excessive risk parameter is generated. The alarm command, encapsulated in MIME format, is sent to the designated email address of the quality control manager via SMTP email service to trigger the alarm.

[0046] In some implementations, the calculation of the cumulative probability of exceeding the residual allowable limit based on the posterior probability distribution and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit constitutes a two-dimensional risk signature, including: The Monte Carlo sampling mechanism is invoked to extract simulated residual sample data from the posterior probability distribution. When the posterior probability distribution has been constrained to be non-negative by a non-negative activation function or a log-normal distribution, non-negative samples are directly retained. When the sampling model still produces negative values, negative samples are removed by truncation to zero. The percentage of samples with residual levels exceeding the allowable limit can be obtained by Monte Carlo sampling, or by integrating the posterior probability distribution over the interval from the allowable limit to positive infinity to calculate the cumulative probability of residual levels exceeding the allowable limit. The posterior standard deviation of the residual amount is calculated based on the simulated residual amount sample data or the posterior probability distribution function, and the posterior standard deviation of the residual amount is divided by the residual allowable limit to obtain the uncertainty limit ratio. By combining the cumulative probability with the uncertainty limit ratio, a two-dimensional risk signature array representing the cross-contamination risk level of the current batch is obtained; The array is compared with the preset control boundary. If the cumulative probability is higher than the set upper limit or the uncertainty limit ratio is higher than the set upper limit, the alarm module is triggered to output a warning signal.

[0047] Internally, a Monte Carlo random sampler is instantiated, with the number of cyclic samplings set to be on the order of a large sample size. The sampler performs repeated sampling operations on the parameterized log-normal posterior probability distribution function output from the previous stage, generating a continuous empirical sample array of residual amounts consisting of ten thousand non-negative floating-point concentration values.

[0048] The extracted residue limits for subsequent varieties are used as the lower bound for integration. The area of ​​integration of the entire posterior probability density function over the interval [residue limit, +∞) is calculated. For example, when the residue limit is 0.05 mg / kg, the integral is performed over the interval [0.05, +∞) to obtain the cumulative probability representing the risk of violation. Simultaneously, the posterior standard deviation of the residue level is calculated based on the empirical sample array of continuous residue levels. This posterior standard deviation is then divided by the residue limit to obtain a second-dimensional risk index characterizing the dispersion and uncertainty level of the prediction. The relative positions of the probability density curve and the residue limit are shown below. Figure 2 As shown.

[0049] For example, when the cumulative probability of exceeding the residue allowable limit is 0.035, and the posterior standard deviation of the residue amount is 0.006 mg / kg and the residue allowable limit is 0.05 mg / kg, the uncertainty limit ratio is calculated to be 0.12, thus forming a risk signature in two-dimensional array format [0.035, 0.12], where 0.035 represents the cumulative probability of the residue amount exceeding the residue allowable limit, and 0.12 represents the ratio of the posterior standard deviation of the residue amount to the residue allowable limit. This two-dimensional risk signature is compared with the preset control boundaries using Boolean logic, for example, setting the upper limit of the cumulative probability to 0.05 and the upper limit of the uncertainty limit ratio to 0.2. When it is determined that any indicator in the real-time signature exceeds the preset control boundary, the underlying microservice gateway immediately pushes an alarm status frame to the central monitoring dashboard through the RabbitMQ message middleware mechanism, triggering the sound and light intervention action of the production line in the factory area, and issuing a warning blocking command with the out-of-bounds value in the form of SMS and enterprise work order.

[0050] The experiment used 5,000 historical cleaning residue batch data from an enterprise manufacturing execution system as the dataset, with 80% used as the training set and 20% as the test set. The hardware environment was an enterprise-level deep learning computing platform. Three comparative models were set up: Model 1 was a basic multilayer perceptron model used to predict residue points on the input feature vector; Model 2 was a feature extraction model containing only a graph attention network, without the variational inference probability distribution derivation process; Model 3 was the comprehensive model proposed in this application, combining a graph attention network and a variational inference algorithm, used to construct a variational posterior probability distribution and generate a two-dimensional risk warning result.

[0051] Under the same training epochs and hyperparameter configurations, the three models exhibited different prediction performances on the test set. Model 1 had a root mean square error (RMSE) of 0.045 mg / kg for predicting cross-contamination residues, with a risk exceedance warning accuracy of 75.2%. Model 2's RMSE decreased to 0.021 mg / kg, and its warning accuracy improved to 86.4%. Model 3 had a RMSE of 0.009 mg / kg, with an exceedance warning accuracy of 97.1% based on two-dimensional risk signatures. The comparison results of the RMSE and warning accuracy of the three models are as follows: Figure 3 As shown.

[0052] Analysis of the experimental results shows that Model 2 exhibits better prediction performance compared to Model 1. This is primarily because the graph attention network can construct a node interaction topology from biochemical indicators and operating parameters of different dimensions, thereby extracting the implicit correlation features between key risk factors. Model 3 further reduces prediction error and improves early warning accuracy compared to Model 2. This is mainly because the variational inference algorithm and sampling mechanism extend the deterministic residual quantity prediction to a non-negative posterior probability density distribution derivation process, enabling the model to consider the impact of measurement uncertainty in risk assessment. Therefore, the comprehensive model demonstrates good noise resistance and fault tolerance under the experimental conditions, helping to reduce the risk of false alarms and missed alarms.

[0053] In this application, a method for early warning of cross-contamination risk of pesticide formulations includes: Obtain current batch data, including physicochemical parameters of the preceding formulation, toxicological parameters of the preceding formulation, residue limits of the subsequent formulation, and clearance procedure data; retrieve the historical model calibration coefficient with the highest similarity to the current batch from the database; and calculate the clearance residue adsorption index by combining the historical model calibration coefficient, the preceding molecule descriptor, the dosage form, and the Hansen solubility parameter of the clearance solvent. The cross-contamination feature vector is constructed by integrating the preceding toxicological endpoint value, the subsequent residue allowable limit and the cleanup residue adsorption index; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through the message passing mechanism. A prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty. Based on the observed values ​​of each risk factor, the posterior probability distribution of cross-contamination residue is derived through variational inference algorithm, and the expected value of the posterior probability distribution is used as the model's predicted residue. After obtaining the actual residue test results after the current batch is cleared and verifying the data validity, based on the deviation between the actual residue test results and the model's predicted residue, the predicted calibration coefficients output by the graph attention network are incrementally corrected according to a preset learning rate or correction step size, and the corrected predicted calibration coefficients are used as new historical model calibration coefficients to update the database. Based on the posterior probability distribution, the cumulative probability of exceeding the residual allowable limit and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit are calculated to form a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

[0054] In some implementations, obtaining the current batch of data includes: The system calls the production scheduling history work order database of the enterprise production control platform via the application programming interface. Use regular expressions to extract object codes containing the previous production completion identifier from work order records, and obtain the sequence of subsequent product codes to be produced; Based on the extracted object code and product code sequence, a query is initiated to the pesticide management cloud platform to extract the corresponding toxicological endpoint value, residue limit standard, median lethal dose, level of no visible harmful effect, as well as the cleaning solvent parameters and equipment volume parameters in the cleaning operation procedure, and store them in the system cache.

[0055] In some implementations, retrieving the historical model calibration coefficients with the highest similarity to the current batch from the database includes: Extract the molecular polarity and solubility feature values ​​of the current batch and construct the target feature query vector; Traverse all historical batch records in the database that contain historical model calibration coefficients, and extract the corresponding historical feature vectors; After standardizing the target feature query vector and the historical feature vector, the feature deviation distance is calculated based on the multidimensional Euclidean distance formula, where the feature deviation distance is the square root of the sum of the squares of the differences of each dimension component of the standardized target feature query vector and the historical feature vector. Sort all calculated feature deviation distances from smallest to largest, extract the first historical batch record with the smallest distance, and extract and cache the historical model calibration coefficient corresponding to the record as the calibration coefficient of the current batch.

[0056] In some implementations, the calculation of the residual adsorption index for clearing, combining the historical model calibration coefficients, preceding molecule descriptors, dosage form, and Hansen's solubility parameter of the clearing solvent, includes: Hansen solubility parameters, including dispersion force parameters, polarity parameters, and hydrogen bond parameters, were obtained for the target compound of the preceding formulation and the clearing solvent, respectively. Calculate the solubility parameter distance between the preceding formulation and the clearing solvent in Hansen space. The solubility parameter distance is the square root of the sum of four times the square of the difference in dispersion force parameter, the square of the difference in polarity parameter, and the square of the difference in hydrogen bond parameter between the preceding formulation and the clearing solvent. The preceding molecular descriptor is normalized and a comprehensive value of the preceding molecular descriptor is generated. The solubility parameter distance is dimensionless. The historical model calibration coefficient, the comprehensive value of the preceding molecular descriptor, the dimensionless solubility parameter distance, and the empirical constant of the dosage form adhesion are weighted and fused to obtain the calibrated clearing residual adsorption index. The residual adsorption index after cleaning is used as a key parameter to represent removal resistance and stored in the inference calculation module.

[0057] In some implementations, the step of inputting the cross-contamination feature vector into a graph attention network model with risk factors as graph nodes, and outputting a scenario contribution weight vector and prediction calibration coefficients through a message passing mechanism, includes: Each risk factor in the cross-contamination feature vector is used as a graph node, and the numerical and attribute features of each risk factor are used to construct the initial node feature vector. Based on the association rules between each risk factor, the connection edges between nodes are constructed to establish the graph topology. In the constructed interactive network, for each node and its neighboring nodes, the attention coefficients between nodes are calculated by sharing the weight matrix and the attention weight vector; Based on the calculated attention coefficients, the features of neighboring nodes are weighted and aggregated, and the updated node features are obtained through an activation function. After multi-layer network iteration and updates, the aggregated full-graph node features output a context contribution weight vector, and the aggregated full-graph node features are mapped to prediction calibration coefficients through a fully connected layer.

[0058] In some implementations, the prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty, and the posterior probability distribution of cross-contamination residue is derived based on the observed values ​​of each risk factor using a variational inference algorithm, including: The allowable deviation of the measuring equipment is read as the measurement uncertainty, and the prior distribution of the latent variable representing the implicit state of cross-contamination is constructed by combining it with the scenario contribution weight vector; The observation dataset is constructed by obtaining the variables representing production conditions from the observed values ​​of each risk factor, excluding the inherent biochemical toxicity constants and regulatory standards, and constructing a variational distribution to approximate the true posterior distribution. The variational distribution parameters are optimized by maximizing the lower bound of evidence. The lower bound loss function is set as the difference between the log-expected value of the joint distribution and the log-expected value of the variational distribution. The gradient descent algorithm is used to iteratively update the variational parameters, reduce the KL divergence between the variational distribution and the true posterior distribution, until the lower bound of evidence converges or the preset maximum number of iterations is reached, and the converged latent variable posterior distribution is obtained. The latent variables are transformed into non-negative cross-contamination residues through a pre-defined generation mapping function with non-negative activation functions at the ends, and the posterior probability distribution of the cross-contamination residues is output.

[0059] In some implementations, the calculation of the cumulative probability of exceeding the residual allowable limit based on the posterior probability distribution and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit constitutes a two-dimensional risk signature, including: The Monte Carlo sampling mechanism is invoked to extract simulated residual sample data from the posterior probability distribution. When the posterior probability distribution has been constrained to be non-negative by a non-negative activation function or a log-normal distribution, non-negative samples are directly retained. When the sampling model still produces negative values, negative samples are removed by truncation to zero. The percentage of samples with residual levels exceeding the allowable limit can be obtained by Monte Carlo sampling, or by integrating the posterior probability distribution over the interval from the allowable limit to positive infinity to calculate the cumulative probability of residual levels exceeding the allowable limit. The posterior standard deviation of the residual amount is calculated based on the simulated residual amount sample data or the posterior probability distribution function, and the posterior standard deviation of the residual amount is divided by the residual allowable limit to obtain the uncertainty limit ratio. By combining the cumulative probability with the uncertainty limit ratio, a two-dimensional risk signature array representing the cross-contamination risk level of the current batch is obtained; The array is compared with the preset control boundary. If the cumulative probability is higher than the set upper limit or the uncertainty limit ratio is higher than the set upper limit, the alarm module is triggered to output a warning signal.

[0060] It should be understood that in the foregoing description of the embodiments in this specification, various features are combined in a single embodiment, drawing, or description for the purpose of simplifying the description and to aid in understanding a feature. However, this does not mean that the combination of these features is necessary, and those skilled in the art may understand some of the technical features as individual embodiments when reading this specification. That is, the embodiments in this specification can also be understood as an integration of multiple secondary embodiments. And the content of each secondary embodiment is also valid even if it contains fewer than all the features of a single foregoing disclosed embodiment.

Claims

1. A pesticide formulation cross-contamination risk early warning system, characterized by, Includes the following modules: The calculation module is used to acquire the current batch data, including the physicochemical parameters of the preceding formulation, the toxicological parameters of the preceding formulation, the residue limits of the subsequent formulation, and the clearance procedure data; retrieve the historical model calibration coefficient with the highest similarity to the current batch from the database; and calculate the clearance residue adsorption index by combining the historical model calibration coefficient, the preceding molecule descriptor, the dosage form, and the Hansen solubility parameter of the clearance solvent. The output module integrates the preceding toxicological endpoint value, the subsequent residue allowable limit, and the cleanup residue adsorption index to form a cross-contamination feature vector; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through a message passing mechanism. The derivation module is used to establish a prior distribution by combining the scenario contribution weight vector and measurement uncertainty, derive the posterior probability distribution of cross-contamination residue based on the observed values ​​of each risk factor through a variational inference algorithm, and use the expected value of the posterior probability distribution as the model's predicted residue amount; after obtaining the actual residue test results after the current batch is cleared and verifying the data validity, the prediction calibration coefficient is corrected based on the deviation between the actual residue test results and the model's predicted residue amount, and the corrected prediction calibration coefficient is used as the new historical model calibration coefficient to update the database; The early warning module is used to calculate the cumulative probability of exceeding the residual allowable limit and the ratio of the residual posterior standard deviation to the residual allowable limit based on the posterior probability distribution, thus forming a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

2. The system of claim 1, wherein, The process of obtaining the current batch of data includes: The system calls the production scheduling history work order database of the enterprise production control platform via the application programming interface. Use regular expressions to extract object codes containing the previous production completion identifier from work order records, and obtain the sequence of subsequent product codes to be produced; Based on the extracted object code and product code sequence, a query is initiated to the pesticide management cloud platform to extract the corresponding toxicological endpoint value, residue limit standard, median lethal dose, level of no visible harmful effect, as well as the cleaning solvent parameters and equipment volume parameters in the cleaning operation procedure, and store them in the system cache.

3. The system of claim 1, wherein, The step of retrieving the historical model calibration coefficients from the database that have the highest similarity to the current batch includes: Extract the molecular polarity and solubility feature values ​​of the current batch and construct the target feature query vector; Traverse all historical batch records in the database that contain historical model calibration coefficients, and extract the corresponding historical feature vectors; After standardizing the target feature query vector and the historical feature vector, the feature deviation distance is calculated based on the multidimensional Euclidean distance formula, where the feature deviation distance is the square root of the sum of the squares of the differences of each dimension component of the standardized target feature query vector and the historical feature vector. Sort all calculated feature deviation distances from smallest to largest, extract the first historical batch record with the smallest distance, and extract and cache the historical model calibration coefficient corresponding to the record as the calibration coefficient of the current batch.

4. The system of claim 1, wherein, The process of calculating the residual adsorption index for clearing, by combining the historical model calibration coefficients, preceding molecule descriptors, dosage forms, and Hansen's solubility parameters of the clearing solvent, includes: Hansen solubility parameters of the target compound in the preceding formulation and the clearing solvent were obtained, including dispersion force parameters, polarity parameters and hydrogen bond parameters. Calculate the solubility parameter distance between the preceding formulation and the clearing solvent in Hansen space. The solubility parameter distance is the square root of the sum of four times the square of the difference in dispersion force parameter, the square of the difference in polarity parameter, and the square of the difference in hydrogen bond parameter between the preceding formulation and the clearing solvent. The preceding molecular descriptor is normalized and a comprehensive value of the preceding molecular descriptor is generated. The solubility parameter distance is dimensionless. The historical model calibration coefficient, the comprehensive value of the preceding molecular descriptor, the dimensionless solubility parameter distance, and the empirical constant of the dosage form adhesion are weighted and fused to obtain the calibrated clearing residual adsorption index. The residual adsorption index after cleaning is used as a key parameter to represent removal resistance and stored in the inference calculation module.

5. The system of claim 2 or 3, wherein, The step of inputting the cross-contamination feature vector into a graph attention network model with risk factors as graph nodes, and outputting the scenario contribution weight vector and prediction calibration coefficients through a message passing mechanism, includes: Each risk factor in the cross-contamination feature vector is used as a graph node, and the numerical and attribute features of each risk factor are used to construct the initial node feature vector. Based on the association rules between each risk factor, the connection edges between nodes are constructed to establish the graph topology. In the constructed interactive network, for each node and its neighboring nodes, the attention coefficients between nodes are calculated by sharing the weight matrix and the attention weight vector; Based on the calculated attention coefficients, the features of neighboring nodes are weighted and aggregated, and the updated node features are obtained through an activation function. After multi-layer network iteration and updates, the aggregated full-graph node features output a context contribution weight vector, and the aggregated full-graph node features are mapped to prediction calibration coefficients through a fully connected layer.

6. The system of claim 1, wherein, The prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty. Based on the observed values ​​of each risk factor, the posterior probability distribution of cross-contamination residue is derived using a variational inference algorithm, including: The allowable deviation of the measuring equipment is read as the measurement uncertainty, and the prior distribution of the latent variable representing the implicit state of cross-contamination is constructed by combining it with the scenario contribution weight vector; The observed dataset is constructed by obtaining the variables representing production conditions from the observed values ​​of each risk factor, excluding the inherent biochemical toxicity constants and regulatory standards, and constructing a variational distribution to approximate the true posterior distribution. The variational distribution parameters are optimized by maximizing the lower bound of evidence. The lower bound loss function is set as the difference between the log-expected value of the joint distribution and the log-expected value of the variational distribution. The gradient descent algorithm is used to iteratively update the variational parameters, reduce the KL divergence between the variational distribution and the true posterior distribution, until the lower bound of evidence converges or the preset maximum number of iterations is reached, and the converged latent variable posterior distribution is obtained. The latent variables are transformed into non-negative cross-contamination residues through a pre-defined generation mapping function with non-negative activation functions at the ends, and the posterior probability distribution of the cross-contamination residues is output.

7. The system of claim 1, wherein, The calculation of the cumulative probability of exceeding the residual allowable limit based on the posterior probability distribution and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit constitutes a two-dimensional risk signature, including: The Monte Carlo sampling mechanism is invoked to extract simulated residual sample data from the posterior probability distribution. When the posterior probability distribution has been constrained to be non-negative by a non-negative activation function or a log-normal distribution, non-negative samples are directly retained. When the sampling model still produces negative values, negative samples are removed by truncation to zero. The percentage of samples with residual levels exceeding the allowable limit can be obtained by Monte Carlo sampling, or by integrating the posterior probability distribution over the interval from the allowable limit to positive infinity to calculate the cumulative probability of residual levels exceeding the allowable limit. The posterior standard deviation of the residual amount is calculated based on the simulated residual amount sample data or the posterior probability distribution function, and the posterior standard deviation of the residual amount is divided by the residual allowable limit to obtain the uncertainty limit ratio. By combining the cumulative probability with the uncertainty limit ratio, a two-dimensional risk signature array representing the cross-contamination risk level of the current batch is obtained; The array is compared with the preset control boundary. If the cumulative probability is higher than the set upper limit or the uncertainty limit ratio is higher than the set upper limit, the alarm module is triggered to output a warning signal.

8. A method for early warning of cross-contamination risk of a pesticide formulation, characterized by, include: Obtain current batch data, including physicochemical parameters of preceding formulations, toxicological parameters of preceding formulations, residue limits of subsequent formulations, and data on disposal procedures. The historical model calibration coefficient with the highest similarity to the current batch is retrieved from the database; the residual adsorption index for clearing is calculated by combining the historical model calibration coefficient, the preceding molecule descriptor, the dosage form, and the Hansen solubility parameter of the clearing solvent. The cross-contamination feature vector is constructed by integrating the preceding toxicological endpoint value, the subsequent residue allowable limit and the cleanup residue adsorption index; the cross-contamination feature vector is input into a graph attention network model with risk factors as graph nodes, and the scenario contribution weight vector and prediction calibration coefficient are output through the message passing mechanism. A prior distribution is established by combining the scenario contribution weight vector and measurement uncertainty. Based on the observed values ​​of each risk factor, the posterior probability distribution of cross-contamination residue is derived using a variational inference algorithm. The expected value of the posterior probability distribution is used as the model's predicted residue. After obtaining the actual residue test results after the current batch is cleared and verifying the data validity, the prediction calibration coefficient is corrected based on the deviation between the actual residue test results and the model's predicted residue. The corrected prediction calibration coefficient is then used as the new historical model calibration coefficient and updated in the database. Based on the posterior probability distribution, the cumulative probability of exceeding the residual allowable limit and the ratio of the posterior standard deviation of the residual amount to the residual allowable limit are calculated to form a two-dimensional risk signature; an early warning is triggered when any indicator exceeds the preset boundary.

9. The method of claim 8, wherein, The step of obtaining the current batch data includes: The system calls the production scheduling history work order database of the enterprise production control platform via the application programming interface. Use regular expressions to extract object codes containing the previous production completion identifier from work order records, and obtain the sequence of subsequent product codes to be produced; Based on the extracted object code and product code sequence, a query is initiated to the pesticide management cloud platform to extract the corresponding toxicological endpoint value, residue limit standard, median lethal dose, level of no visible harmful effect, as well as the cleaning solvent parameters and equipment volume parameters in the cleaning operation procedure, and store them in the system cache.

10. A computer readable storage medium having stored thereon a computer program, characterized in that The computer program, when executed by a processor, implements the method as described in any one of claims 8-9.

Images