Compound fertilizer production control method and system based on data processing

Abstract

The present application belongs to the technical field of production control, and particularly relates to a compound fertilizer production control method and system based on data processing, comprising: obtaining historical production data; constructing a quality index space as a Riemann manifold, setting a metric tensor in combination with a process influence degree and sample time proximity, and taking an absolute value of a neighborhood Ricci curvature scalar as a production state criticality index; forming a basic hierarchical unit through unsupervised clustering, constructing a time series data subset according to a global time sequence, and fine-tuning adjacent subset boundaries to reduce hierarchical proportion deviation and criticality difference; and then training and fusing a prediction model in a time forward increasing manner, and inversely solving optimal process parameters by using a particle swarm optimization to realize closed-loop control of compound fertilizer production. The present application can take into account time series forward and quality space distribution to improve prediction accuracy and stably guide closed-loop control.

Description

Technical Field

[0001] This invention relates to the field of production control technology. More specifically, this invention relates to a data processing-based method and system for controlling the production of compound fertilizer. Background Technology

[0002] Compound fertilizer production involves multiple continuous processes, including raw material batching, granulation, drying, cooling, and screening, constituting a typical multi-stage, multi-variable industrial process. Product quality is comprehensively characterized by multi-dimensional indicators such as the content of effective nitrogen, phosphorus, and potassium components, moisture content, and particle size distribution. Process parameters such as granulation temperature, rotation speed, batch ratio, and drying intensity have cross-coupled influences on these quality indicators at different production stages. Furthermore, time-varying disturbances in equipment operating status, raw material batch characteristics, and environmental conditions result in production data exhibiting not only significant time dependence but also complex characteristics such as uneven distribution, local abrupt changes, and boundary sensitivity within a multi-dimensional quality state space. Therefore, relying solely on empirical parameter tuning or conventional data-driven models makes it difficult to identify operating conditions on the edge of quality fluctuations in real time, and also fails to reliably obtain suitable combinations of process parameters for subsequent production control.

[0003] Currently, most existing compound fertilizer production quality prediction models use conventional cross-validation or stratified sampling to divide the training and validation sets. For production data with continuous time-series attributes, random partitioning can easily disrupt the sequential order of samples, leading to the leakage of future information during model training and failing to accurately reproduce the historical evolution of operating conditions. On the other hand, ordinary stratification methods usually only use sample categories or quality interval proportions as the basis for partitioning, lacking a detailed characterization of the spatial geometric structure of multidimensional quality indicators and critical sensitive regions, which can easily lead to uneven distribution of extreme operating conditions and boundary samples among the subsets.

[0004] However, the predictive model trained in this way lacks generalization ability in actual production environments, and its prediction bias is further amplified in the parameter back-optimization stage, making it difficult for the solved process parameters to stably guide closed-loop control. Therefore, it is urgent to construct a data processing and predictive control method that can take into account time sequence, quality spatial distribution, and critical equilibrium of production status. Summary of the Invention

[0005] To address the technical problem of insufficient predictive generalization caused by temporal distortion and uneven mass spatial distribution in existing models, this invention proposes a compound fertilizer production control method and system based on data processing, which can take into account both temporal progression and mass spatial distribution to improve prediction accuracy and stably guide closed-loop control.

[0006] In a first aspect, the present invention provides a data processing-based compound fertilizer production control method, comprising: S1, acquiring historical production data containing multidimensional quality indicators, constructing a Riemannian manifold from the vector space composed of the multidimensional quality indicators, setting a metric tensor for the sample points, calculating the Ricci curvature scalar of the neighborhood of each sample point based on the metric tensor, and using the absolute value of the curvature scalar as an index representing the criticality of the production state; S2, dividing all samples into several data clusters in the quality indicator space as basic hierarchical units, and dividing all samples into K non-overlapping time series according to the global time order. Data subsets are used, and without disrupting the time order, the boundaries of adjacent subsets are fine-tuned by combining the distribution of basic hierarchical units and criticality indices to reduce the hierarchical ratio deviation and criticality index difference of each subset; S3, the production quality prediction model is trained by using the first k time series data subsets in turn according to the time-incrementing principle, and verified by the adjacent k+1 time series data subset. k increases from 1 to K-1 to obtain and fuse the integrated prediction model. Based on the quality indicators of compound fertilizer, the optimal combination of process parameters is solved in reverse by the particle swarm optimization algorithm to perform closed-loop control of the compound fertilizer production process.

[0007] By employing the aforementioned technical solution, the multidimensional quality index space is constructed as a Riemannian manifold, and the absolute value of the Ricci curvature scalar is calculated as a criticality index. This allows for the quantitative characterization of the fluctuation boundaries and sensitivity of production states within the quality space. Based on this, unsupervised clustering is used to form basic hierarchical units, and the boundaries of adjacent temporal subsets are fine-tuned while maintaining the global temporal order, making the distribution of each subset more balanced in terms of hierarchical proportion and criticality index. Furthermore, a time-increasing ensemble training strategy is combined with particle swarm optimization for inverse solution, forming a closed-loop control. This technical solution helps alleviate the problem of insufficient predictive generalization ability caused by temporal disruption and uneven distribution of the quality space, thereby improving the model's predictive stability and the reliability of process parameter optimization in actual production environments.

[0008] Preferably, the step of constructing the vector space composed of multidimensional quality indicators as a Riemannian manifold and setting a metric tensor for the sample points includes: extracting a set of neighboring sample points within a preset distance for each sample point; using the process parameters and multidimensional quality indicator data in the set of neighboring sample points, fitting a mapping function through a local polynomial and taking the partial derivative of the mapping function to obtain the Jacobian matrix of the multidimensional quality indicators, calculating the matrix product of the Jacobian matrix and the transpose of the Jacobian matrix as the spatial influence matrix; calculating the absolute value of the timestamp difference between the sample point and each neighboring sample point in the set of neighboring sample points, converting the absolute value of the timestamp difference into a time weight, and summing them to obtain the aggregated time weight; multiplying the spatial influence matrix by the aggregated time weight and adding a preset dimensionless diagonal regularization term to ensure that the matrix is ​​positive definite, and using the result as the metric tensor of the sample point.

[0009] By employing the above technical solution, neighboring sample points are extracted, and the Jacobian matrix is ​​obtained through local polynomial fitting between process parameters and quality indicators. Combined with timestamp differences, aggregated time weights are generated, ultimately constructing a metric tensor that incorporates both spatial influence and temporal proximity information. This tensor can more accurately reflect the degree of coupling influence of process parameter changes on multidimensional quality indicators. Furthermore, the introduction of a regularization term ensures the positive definiteness of the metric tensor, providing a geometric basis for subsequent criticality index calculations and contributing to improved accuracy in characterizing the local structure of the quality space.

[0010] Preferably, the index representing the criticality of the production state includes: parameterizing the metric tensor of each sample point with continuous differentiability and positive definite constraints to obtain an interpolable parameter vector corresponding to the metric tensor; continuously interpolating the interpolable parameter vector using a radial basis function network, and reconstructing the continuous metric tensor field in the multidimensional quality index space, so that the reconstructed continuous metric tensor field remains symmetric, positive definite, and invertible; calculating the Ricci curvature scalar of the neighborhood of each sample point based on the continuous metric tensor field, and using the absolute value of the curvature scalar as the index representing the criticality of the production state.

[0011] By employing the above technical solution, the metric tensor of discrete sample points is parameterized to be continuously differentiable while maintaining positive definite constraints. Combined with a radial basis function network, a continuous metric tensor field is reconstructed in the quality index space, ensuring that the tensor field remains symmetric, positive definite, and invertible throughout the entire space. This processing method extends the metric information of discrete samples into a continuously differentiable geometric field, thus providing the necessary data conditions for subsequent curvature calculations based on differential geometry. This helps improve the continuity and computational accuracy of the criticality index across the entire operating condition space.

[0012] Preferably, the index representing the criticality of the production state includes: calculating the first and second type of Christofel symbols by analytical differentiation based on the continuous metric tensor field and the corresponding inverse matrix field; obtaining the corresponding Riemann curvature tensor by taking the partial derivatives of the second type of Christofel symbols with respect to the corresponding spatial coordinate variables and combining them with the product terms of the second type of Christofel symbols; performing tensor contraction on the spatial coordinate indices of the Riemann curvature tensor and calculating the Ricci tensor; further contracting the Ricci tensor with the inverse matrix of the continuous metric tensor field to obtain the trace corresponding to the operation matrix, and extracting the value of the trace as the Ricci curvature scalar; and assigning the absolute value of the Ricci curvature scalar at the sample point as the index of the criticality of the production state at the sample point.

[0013] By employing the aforementioned technical solution, the Christofel symbol, Riemann curvature tensor, and Ricci tensor are calculated sequentially. The Ricci curvature scalar is then extracted through a contraction operation, and its absolute value is ultimately taken as the criticality index for the production state. This computational path clarifies the complete analytical derivation relationship from the metric tensor field to the scalar index, enabling the quantification of the local geometric curvature of the mass space into a comparable numerical index. This facilitates the objective identification of operating condition samples located at the mass fluctuation boundary or in the phase transition sensitive region, providing a calculable quantitative basis for subsequent hierarchical division and boundary fine-tuning.

[0014] Preferably, the step of dividing all samples into several data clusters in the quality index space as basic hierarchical units includes: calling a clustering algorithm and randomly assigning the same number of initial centroid coordinates as the preset clusters; calculating the Euclidean distance of all samples relative to each initial centroid in the quality index space, and assigning each sample to the cluster containing the centroid with the smallest distance value to the sample; calculating the arithmetic mean of the multidimensional quality index vectors of all samples within each cluster, and using the arithmetic mean to replace the original centroid coordinates of the corresponding clusters; repeating the above distance calculation and centroid replacement steps until the centroid coordinates no longer shift position after iteration and the update stops, at which point all samples existing within the same cluster are integrated into a basic hierarchical unit.

[0015] By employing the above technical solution and using a clustering algorithm, the samples in the quality indicator space are divided into several data clusters through iterative optimization of the centroid position, and samples within the same cluster are integrated into basic hierarchical units. This approach can automatically identify the inherent clustering patterns of data based on the spatial distribution structure of the quality indicators themselves, without prior labels. This helps to group samples with similar quality characteristics into homogeneous units, thereby providing a structured data foundation for subsequent temporal subset partitioning and hierarchical ratio control, and reducing the interference caused by the heterogeneity of the quality indicator distribution on model training.

[0016] Preferably, reducing the stratification ratio deviation and criticality index difference of each subset includes: dividing all samples into K consecutive and non-overlapping time intervals according to timestamp order; determining the target sample quota of each subset in each basic stratification unit based on the total number of baseline target samples and the sample ratio of each basic stratification unit; allocating the remaining samples after rounding according to the decimal remainder, the current stratification ratio deviation, the cumulative criticality index, and the deviation; if the number of samples in any stratification unit is less than K, allowing some subsets not to be allocated samples in that stratification unit, with the allocation completed under the constraints of non-overlapping global time order and minimizing stratification ratio deviation; within the boundary window of adjacent time intervals, performing boundary fine-tuning based on the criticality index of the samples to ensure that the time order of adjacent subsets does not overlap after fine-tuning, thereby reducing the basic stratification ratio deviation of each subset and making the cumulative criticality index of each subset tend to be balanced.

[0017] Preferably, obtaining and fusing the integrated prediction model includes: constructing a multilayer feedforward neural network model with a set number of hidden layers and neuron node parameters; in the k-th iteration of the K-1 iterations, retaining the (k+1)-th time series data subset as an independent data subset for verification, feeding the first k time series data subsets into the multilayer feedforward neural network model, and adjusting and optimizing the weight nodes of each layer; substituting the currently retained independent data subset into the multilayer feedforward neural network with adjusted weights to calculate the predicted index, calculating the mean absolute error by subtracting it from the actual index, and after ending the K-1 iterations, saving K-1 multilayer feedforward neural networks with mutually independent training weights; synchronously inputting the candidate process parameters into the K-1 multilayer feedforward neural networks with independent training weights to perform feedforward inference, collecting multiple sets of quality prediction results, calculating the arithmetic mean of each index dimension, and integrating them into the output answer of the integrated prediction model.

[0018] Preferably, the closed-loop control of the compound fertilizer production process based on the optimal combination of process parameters includes: limiting the upper and lower limits of the hardware processing threshold of the physical manufacturing equipment; randomly generating a specific number of particle positions within the vector space defined by the upper and lower limits of the threshold, and randomly assigning corresponding velocity vectors, so that any particle position serves as a set of candidate process parameter vectors; calling the integrated prediction model to sequentially read in each set of candidate process parameter vectors, and calculating the corresponding multidimensional quality expectation index; calculating the distance difference between each set of expectation indexes and the quality index of the predetermined target compound fertilizer after dimensionless processing; using the reciprocal of the difference to map the fitness score for performing the selection operation on each set of particle positions; determining the global optimal fitness space position based on the fitness score, and performing closed-loop control of the compound fertilizer production process based on the global optimal fitness space position.

[0019] Preferably, the closed-loop control of the compound fertilizer production process based on the global optimal fitness spatial position includes: recording and updating the historical optimal fitness spatial position with the highest fitness score obtained by each individual particle during the search process, and synchronously updating the global optimal fitness spatial position with the highest fitness score generated during the search process of the group of particles; updating the velocity vector and position vector of each particle by using step-size iterative relationship calculation and learning and attenuation damping coefficient, so as to promote the convergence of the particle search process to the optimal region; once the number of steps for terminating the loop iteration is reached, the globally optimal fitness spatial position locked at this time is converted into a set execution parameter signal, and an instruction is transmitted to require the programmable logic controller to execute the updated control action to the underlying quantitative feeding execution component and regulating valve.

[0020] Secondly, the present invention provides a compound fertilizer production control system based on data processing, including a processor and a memory, wherein the memory stores computer program instructions, and when the computer program instructions are executed by the processor, the above-mentioned compound fertilizer production control method based on data processing is implemented.

[0021] By adopting the above technical solution, a computer program is generated from the above-mentioned compound fertilizer production control method based on data processing and stored in a memory so that it can be loaded and executed by a processor. In this way, a terminal device can be made based on the memory and the processor for convenient use.

[0022] The present invention has the following beneficial effects:

[0023] This invention constructs the quality index space as a Riemannian manifold and calculates the Ricci curvature scalar as a criticality index by combining the degree of process influence and temporal proximity, which can accurately characterize the fluctuation boundary of the operating condition. At the same time, it uses unsupervised clustering to form basic hierarchical units and performs equalization fine-tuning on the boundaries of adjacent subsets without destroying the temporal order, thus constructing a scientific partitioning mechanism that takes into account the distribution of sample categories, temporal continuity and critical operating condition distribution characteristics, effectively reducing the probability of extreme operating conditions and uneven distribution of boundary samples.

[0024] Furthermore, this invention combines time-incremental training with fusion to obtain an integrated prediction model, which greatly improves the model's generalization ability and prediction stability in actual production environments. In the reverse optimization stage, it combines particle swarm optimization algorithm to accurately solve the optimal combination of process parameters to implement closed-loop control, thereby significantly improving the overall compliance rate of nitrogen, phosphorus and potassium content, moisture and particle size distribution, and comprehensively enhancing the stability and intelligent management level of compound fertilizer production process. Attached Figure Description

[0025] Figure 1 This is a flowchart of a compound fertilizer production control method based on data processing; Figure 2 This is a comparison of the error convergence curves of the baseline method and the method proposed in this paper. Detailed Implementation

[0026] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0027] In this application, a data processing-based method for controlling compound fertilizer production is provided, such as... Figure 1 As shown, the method includes: S1. Obtain historical production data including timestamps, process parameters, and multidimensional quality indicators consisting of nitrogen, phosphorus, and potassium content, moisture, and particle size distribution. Construct a Riemannian manifold from the vector space of the multidimensional quality indicators. Based on the degree of influence of process parameter changes on quality indicator changes at each sample point and the temporal proximity of the sample point to other sample points, set a metric tensor for the sample point. Calculate the Ricci curvature scalar of the neighborhood of each sample point based on the metric tensor and use the absolute value of the curvature scalar as an index representing the criticality of the production state.

[0028] The system connects to the distributed control system database of the compound fertilizer production line through an industrial communication interface or database access interface. Historical production data is read according to the production batch number, collection time range, and equipment section identifier. The read data is parsed and missing values ​​are removed. Process parameters such as timestamp, granulation temperature, reaction pH value, and spindle speed are extracted, as well as multi-dimensional quality indicators such as nitrogen, phosphorus and potassium content, moisture and particle size distribution, to form a historical production sample dataset.

[0029] A process parameter feature matrix and a quality index feature matrix are defined, and both are dimensionlessly normalized to obtain normalized process parameter feature matrices and normalized quality index feature matrices. The first-order partial derivatives of the normalized quality indexes with respect to each normalized process parameter are calculated to construct the Jacobian matrix. Simultaneously, the time difference between samples is calculated based on timestamps, and a Gaussian kernel function is input to obtain the time weights corresponding to each neighboring sample. These time weights are summed to obtain the aggregated time weights. The Jacobian matrix is ​​multiplied by its transpose to obtain the spatial influence matrix. This spatial influence matrix is ​​then multiplied by the aggregated time weights and a dimensionless diagonal regularization term is added to obtain the Riemannian metric tensor for each sample point. On the continuous metric tensor field, the first-order partial derivatives of the metric tensor with respect to the multidimensional quality index coordinate space are solved. Substituting these derivatives into the Christoffel notation formula, the first and second types of Christoffel notations are obtained.

[0030] Based on the spatial partial derivatives of the second kind of Christofel symbols and combined with the Christofel symbol product terms, the Riemann curvature tensor is calculated according to the standard definition of the Riemann curvature tensor. The Riemann curvature tensor is then subjected to index-wise contraction to obtain the Ricci tensor. Tensor contraction is then performed between the inverse matrix of the metric tensor and the Ricci tensor to obtain the Ricci curvature scalar, and its absolute value is extracted and assigned to the critical exponential characteristic column.

[0031] In some embodiments, constructing the vector space composed of multidimensional quality indicators as a Riemannian manifold, and setting a metric tensor for the sample point based on the degree of influence of process parameter changes on quality indicator changes for each sample point and the temporal proximity of the sample point to other sample points, includes: Extract a set of neighboring sample points within a preset distance for each sample point; By using dimensionless process parameters and multidimensional quality index data from the set of neighboring sample points, a mapping function is fitted by a local polynomial and the partial derivative of the mapping function is obtained to obtain the Jacobian matrix of the multidimensional quality index with respect to the process parameters. Calculate the matrix product of the Jacobian matrix and the transpose of the Jacobian matrix as the spatial influence matrix; Calculate the absolute value of the timestamp difference between the sample point and each neighboring sample point in the neighboring sample point set, convert the absolute value of the timestamp difference into a time weight using a Gaussian kernel function, and sum them to obtain the aggregated time weight; The spatial influence matrix is ​​multiplied by the aggregated time weight and a preset dimensionless diagonal regularization term is added to ensure that the matrix is ​​positive definite, and the result is used as the metric tensor of the sample points.

[0032] When extracting neighboring samples from a sample point, a set of neighboring sample points within a preset distance range can be extracted, or a set of M nearest neighbors can be extracted, where the number of nearest neighbors M is preferably 15 to 30 points, for example, M=20. Process parameters include, but are not limited to, granulation temperature (e.g., 70 to 90℃), reaction pH (e.g., 5.5 to 6.5), and spindle speed (e.g., 30 to 50 revolutions per minute). Multidimensional quality indicators include the effective content of nitrogen, phosphorus, and potassium (e.g., fluctuating around 15%), moisture content (1.5% to 2.5%), and a particle size distribution of 2 to 4 mm accounting for more than 90%.

[0033] Range standardization is used to uniformly scale the above parameters to the 0-1 range to achieve dimensionlessness. A second-order Taylor expansion or a multivariate quadratic polynomial is used, combined with the least squares method, to fit the nearest neighbor data, obtaining a mapping function representing the process parameters to quality indicators. The first-order partial derivative of this mapping function is calculated to construct a 5×3 Jacobian matrix (representing 5 quality indicators with respect to 3 process parameters). Multiplying the Jacobian matrix by its transpose yields a 5×5 spatial influence matrix, which quantitatively represents the spatial structure of the multidimensional quality fluctuation covariance caused by minute process variations.

[0034] A Gaussian kernel function with a bandwidth parameter range of 1 to 10 hours, preferably 4 hours, is used to process the timestamp difference. Assuming the absolute value of the time difference between the current sample point and a certain nearest neighbor point is 2 hours, the converted single-point time weight is approximately 0.88. The aggregated time weight is generated by summing all weights within the nearest neighbor set.

[0035] The aforementioned 5×5 spatial influence matrix is ​​multiplied by the aggregated time weight, and then a unit diagonal matrix with a coefficient preferably of 0.001 is superimposed as a dimensionless diagonal regularization term. This operation not only increases the influence of temporally closer operating conditions on the local geometry, but also ensures that the output 5×5 metric tensor has positive definite properties through regularization.

[0036] In some embodiments, the step of calculating the Ricci curvature scalar of the neighborhood of each sample point based on the metric tensor and using the absolute value of the curvature scalar as an index representing the criticality of the production state includes: The metric tensor of each sample point is parameterized while maintaining a continuously differentiable and positive definite constraint to obtain an interpolable parameter vector corresponding to the metric tensor. The interpolable parameter vector is continuously interpolated using a radial basis function network, and a continuous metric tensor field is reconstructed in a multidimensional quality index space, so that the reconstructed continuous metric tensor field remains symmetric, positive definite and invertible. Based on the continuous metric tensor field and the corresponding inverse matrix field, the first and second type of Christofel symbols are calculated by analytical differentiation. By taking the partial derivatives of the corresponding spatial coordinate variables using the Christofel symbol of the second kind, and combining them with the product terms of the Christofel symbol of the second kind, the corresponding Riemann curvature tensor is obtained. Tensor shrinkage is performed on the spatial coordinate index of the Riemann curvature tensor, and the Ricci tensor is obtained by computation; The inverse matrix of the Ricci tensor and the continuous metric tensor field is shrunk again to obtain the trace corresponding to the operation matrix, and the value of the trace is extracted as the Ricci curvature scalar. The absolute value of the Ricci curvature scalar at the sample point is assigned as an index of the criticality of the production state at the sample point.

[0037] A radial basis function (RBF) network with a Gaussian kernel is used to perform continuous interpolation of the metric tensor. The Gaussian kernel shape parameter is preferably set to between 0.1 and 1, for example, 0.5. The RBF network takes a multidimensional quality index vector as input and an interpolable parameter vector obtained after positive definite parameterization of the metric tensor as output. The network structure includes an input layer, hidden layers, and an output layer. Specifically, the input layer receives the multidimensional quality index coordinates; the hidden layers calculate the distance activation values ​​between the input coordinates and the center nodes of each hidden layer based on the Gaussian kernel function; the output layer linearly combines the distance activation values ​​to output the interpolable parameter vector at the corresponding coordinate points. To ensure that the interpolated metric tensor maintains positive definiteness, the positive definite metric tensor at each discrete sample point is first decomposed using Cholesky decomposition to obtain a lower triangular matrix; then, the off-diagonal elements and the logarithm of the diagonal elements in the lower triangular matrix are combined to form a parameter vector, which is used as the training target of the RBF network.

[0038] At the target quality index coordinate point, the corresponding lower triangular matrix is ​​reconstructed based on the parameter vector output by the radial basis function network, and the continuous metric tensor is recovered by multiplying the lower triangular matrix with its transpose matrix; if necessary, a preset positive definite lower limit term is superimposed to ensure that the obtained continuous metric tensor field remains symmetric, positive definite and invertible in the multidimensional quality index space.

[0039] This continuous metric tensor field is at least second-order continuously differentiable, allowing partial derivative calculations to be performed analytically or numerically by differentiating the Gaussian kernel function and reconstruction parameters. On this continuous field, the first-type Christofel symbols are calculated for each component according to the rules of standard differential geometry, and then an inner product is performed with the inverse metric tensor field to obtain the second-type Christofel symbols used to describe the degree of spatial coordinate distortion.

[0040] By performing a partial differential operation on the second type of Christoferne symbols in each dimension and combining it with the Christoferne symbol product term, a corresponding fourth-order Riemann curvature tensor is constructed. Tensor shrinking and order reduction are then performed on the first and third indices of this fourth-order tensor to generate a second-order symmetric Ricci tensor. This tensor is then shrunk to a single real number, the Ricci curvature scalar, by performing a dot product trace operation with the inverse matrix of the continuous metric tensor field.

[0041] The absolute value of this curvature scalar is extracted as a criticality index. In practical applications, this index typically remains at a low level below 10 in stable production zones. However, when the formulation process is at the phase transition boundary where even slight parameter drift leads to agglomeration or pulverization, the absolute value of the curvature scalar may rapidly rise to 50 or even above 100.

[0042] S2 employs an unsupervised clustering algorithm to divide all samples into several data clusters in the quality index space as basic hierarchical units. All samples are divided into K non-overlapping time-series data subsets according to the global time order. Without disrupting the time order, the boundaries of adjacent subsets are fine-tuned by combining the distribution of basic hierarchical units and criticality index, so as to reduce the hierarchical ratio deviation and criticality index difference of each subset.

[0043] The K-means clustering algorithm was used, the number of cluster centers was set, and unsupervised clustering was performed on the multidimensional quality index data to generate cluster labels for each sample, thereby dividing the basic hierarchical units.

[0044] All samples are sorted globally in ascending order according to their timestamp feature column, and initially divided into K consecutive and non-overlapping time series intervals, forming K initial time series data subsets. Subsequently, the sample proportion of each basic hierarchical unit and the cumulative sum of criticality indices in each initial time series data subset are calculated.

[0045] A boundary fine-tuning compensation algorithm is defined, which iteratively compares the cumulative sum of critical exponents and the basic stratification ratio deviation of adjacent i-th and (i+1)-th time-series data subsets. If the sum of the former minus the sum of the latter is greater than a set tolerance threshold, and the temporal order non-overlap constraint is still satisfied after adjacent boundary adjustments, then the tail sample with the latest timestamp of the i-th subset is extracted and moved to the head of the (i+1)-th subset. If it is less than a negative tolerance threshold, and the temporal order non-overlap constraint is still satisfied after adjacent boundary adjustments, then the head sample of the (i+1)-th subset is extracted and moved to the tail of the i-th subset. After each move, the variance of the cumulative sum of all subsets and the stratification ratio deviation are updated and calculated in real time. When the variance no longer decreases after ten consecutive iterations or reaches the preset maximum number of iterations, the loop is exited, and the subset equalization fine-tuning is completed.

[0046] In some embodiments, the unsupervised clustering algorithm is used to divide all samples into several data clusters in the quality index space as basic hierarchical units, including: Call the K-means clustering algorithm and randomly assign the same number of initial centroid coordinates as the preset number of clusters; Calculate the Euclidean distance of all samples relative to each initial centroid in the quality index space, and assign each sample to the cluster containing the centroid with the smallest distance value from the sample. Calculate the arithmetic mean of the multidimensional quality index vector of all samples within each cluster, and use the arithmetic mean to replace the original centroid coordinates of the corresponding cluster. Repeat the above distance calculation and centroid substitution steps until the centroid coordinates no longer shift after iteration and stop updating. At this point, all samples within the same cluster are integrated into a basic hierarchical unit.

[0047] During the algorithm initialization phase, the preset number of clusters, i.e., the C value, is preferably determined in advance between 3 and 8 using the silhouette coefficient method. For example, setting C=5 is used to represent the five typical production phases that objectively exist in the system's historical operation: the high-quality product zone, the moisture exceeding the standard zone, the fine particle size zone, the nitrogen content fluctuation zone, and the agglomerated substandard product zone.

[0048] For a sample set of 5-dimensional quality indicators totaling tens of thousands, the K-means++ strategy is used for spatial uniform distribution initialization. In the iterative allocation phase, the algorithm rapidly calculates the multidimensional Euclidean distance between each sample and the five initial centroids, and assigns the sample to the cluster with the shortest distance. After allocation, for all samples within each cluster, the multidimensional quality indicators are extracted dimension by dimension, summed, and divided by the total number of samples. For example, if a cluster contains 2000 samples, the arithmetic mean of the nitrogen content of each sample, such as 15.15%, can be used to generate the new centroid coordinates of that cluster in the nitrogen dimension. The same logic applies to other dimensions such as moisture (e.g., 2.2%) and granularity (e.g., 92%).

[0049] Set the centroid offset convergence tolerance threshold to 0.0001 or set the maximum number of iterations to 500. When the centroid Euclidean distance variation generated by two consecutive iterations is all below this threshold, the iteration stop mechanism is triggered. At this point, each cluster is converted into the basic hierarchical unit for subsequent processes, eliminating the implicit heterogeneity between mixed production batches and improving the consistency of feature distribution within the basic data unit.

[0050] In some embodiments, dividing all samples into K non-overlapping time-series data subsets according to global time order, and fine-tuning the boundaries of adjacent subsets based on the basic stratification unit distribution and criticality index without disrupting the time order, so as to reduce the stratification ratio deviation and criticality index difference of each subset, includes: Based on the chronological order of the timestamps of all samples, all samples are initially divided into K consecutive and non-overlapping time intervals, such that the maximum timestamp of the samples in the previous time interval is no later than the minimum timestamp of the samples in the next time interval. Based on the baseline target sample total number of a single time series data subset and the sample ratio of each basic hierarchical unit, determine the target sample quota of each time series data subset within each basic hierarchical unit. The remaining samples after rounding the target sample amount are supplemented and allocated according to the decimal remainder, the current stratification ratio deviation, and the cumulative deviation of the criticality index. When the number of samples in any basic stratification unit is less than K, some time series data subsets are allowed not to be allocated samples in that basic stratification unit, and the allocation is completed with the constraints of global time order non-overlap and minimizing the stratification ratio deviation. Within the boundary window of adjacent time series intervals, boundary fine-tuning is performed based on the criticality index corresponding to the sample, ensuring that the time order of adjacent time series data subsets does not overlap after fine-tuning, thereby reducing the deviation of the basic stratification ratio of each time series data subset and making the accumulation and equilibrium of the criticality index of each time series data subset more or less equal.

[0051] Assuming the historical production database stores 10,000 samples, and the fold number K for cross-validation is set to 10, if a basic stratified unit contains 3,000 samples, then the percentage ratio constant for that unit is calculated to be 30%. For each subset of time-series data, the theoretical total number of baseline target samples is 1,000. By multiplying the aforementioned ratio constant and combining it with rounding and residual sample allocation rules, the reference target sample quota for each time-series data subset within that basic stratified unit is determined to be approximately 300. This results in 10 quantitative quota tasks that are as close as possible in terms of the distribution structure of quality indicators.

[0052] In the specific partitioning phase, the basic sample sorting is based on the global timestamp forward principle. During the cutting boundary window period, such as the period from the 295th to the 305th sample, a fine-tuning mechanism based on the balance of the critical index sum is triggered. Within the boundary window, multiple candidate split points are set according to the timestamp order. The stratification ratio deviation and the cumulative deviation of the critical index between two adjacent time series data subsets under different candidate split points are calculated. The candidate split point with the smaller comprehensive deviation and satisfying that the maximum timestamp of the previous subset is not later than the minimum timestamp of the next subset is selected as the new boundary.

[0053] This fine-tuning operation is repeated iteratively until the relative standard deviation between the final critical sums of the 10 subsets decreases to a preset range, preferably no more than 5%, or until the relative standard deviation no longer decreases after multiple consecutive iterations, at which point the boundary fine-tuning stops.

[0054] S3: The production quality prediction model is trained by taking turns using the first k time series data subsets in a forward increasing order of time and then validated by the next k+1 time series data subset. k increases sequentially from 1 to K-1 to obtain and merge the integrated prediction model. Based on the quality indicators of the target compound fertilizer, the optimal combination of process parameters is solved in reverse using the particle swarm optimization algorithm through the integrated prediction model. The compound fertilizer production process is then controlled in a closed loop based on the optimal combination of process parameters.

[0055] A multi-layer feedforward neural network model is constructed as the basic production quality prediction model. Time series cross-validation is set up, with the control variable k incrementing from 1 to K-1. In each loop, the first k time series data subsets are concatenated row-wise as the training set, and the training interface is called to complete the model training. The (k+1)th time series data subset is used as the validation set, and the prediction interface is called to output the predicted values ​​and calculate the mean squared error. After the loop ends, K-1 trained regressors are obtained. The reciprocals of the mean squared errors of all regressors are extracted as weights, normalized using the softmax algorithm, and summed to construct a weighted ensemble prediction model.

[0056] Using the particle swarm optimization algorithm, the input process parameter particles are fed into the integrated prediction model to obtain the predicted nitrogen, phosphorus, and potassium content, moisture, and particle size distribution, and to calculate the dimensionless comprehensive distance between them and the input target compound fertilizer quality index; when the target compound fertilizer quality index is subject to upper and lower limit constraints, the dimensionless comprehensive distance is calculated based on the violation amount of the prediction results relative to the upper and lower limit constraints.

[0057] An initial population of 200 particles is established, with inertia weights and learning factors set. Within the maximum number of iterations, the optimization objective is to minimize the dimensionless comprehensive distance or maximize the corresponding fitness score, continuously updating the historical best position and the global best position of each particle. After the algorithm converges, the global best position is output as the optimal combination of process parameters. The underlying socket communication library is invoked to establish a TCP connection with the workshop programmable logic controller via the OPC unified architecture protocol. The optimal combination of process parameters is packaged into a hexadecimal control message and sent to drive the batching regulating valve and the granulator frequency converter, achieving closed-loop control.

[0058] In some embodiments, the step of training the production quality prediction model using the first k time-series data subsets in a sequentially increasing manner and validating it with the next adjacent (k+1)th time-series data subset, where k increases sequentially from 1 to K-1 to obtain and fuse the ensemble prediction model, includes: Construct a multilayer feedforward neural network model with a set number of hidden layers and neuron node parameters; The process involves a total of K-1 iterations. In the k-th iteration, the k+1th time series data subset is stripped and retained as an independent data subset for verification. The k time series data subsets preceding this subset are fed into the multi-layer feedforward neural network model. The weight nodes of each layer in the neural network are adjusted and optimized based on the gradient descent method. The independent subset of data retained at this time is called and substituted into the multilayer feedforward neural network that has just completed weight adjustment. The predicted index is calculated, and the difference between the index and the actual index is calculated to solve the mean absolute error and monitor the convergence trend of the loss. After completing all K-1 iterations, K-1 multilayer feedforward neural networks with independent training weights are extracted and preserved. After receiving candidate process parameters from the outside, the same parameter is synchronously distributed to K-1 multilayer feedforward neural networks with independent weights, and feedforward inference is performed concurrently. After collecting multiple sets of quality prediction results from each network, the arithmetic mean of the multiple sets of results on each index dimension is calculated, and the average value is integrated into the unique output answer of the integrated prediction model.

[0059] The initial multi-layer feedforward neural network optimization architecture includes one input layer with an adaptive parameter dimension, three hidden layers, and one output layer containing a 5-dimensional quality index. The number of neurons in the hidden layers is configured in a decreasing structure of 64, 32, and 16, and ReLU activation functions are embedded between layers to prevent gradient vanishing.

[0060] Assuming the cross-fold number K is 10, the entire system is started in a total of 9 time-incrementing loops. In the first loop with k=1, the first subset, such as the data from hour 1 to hour 1000, is trained using the Adam optimizer with a batch size of 64 and an initial learning rate of 0.001, and the connection weights are updated. At the same time, the second subset is used as an independent validation set for testing.

[0061] Within a single loop, the quality metric obtained from the forward computation, such as the predicted moisture content of 2.2%, is immediately subtracted from the true label in the validation set, such as the actual moisture value of 2%, to obtain the absolute residual. This error is then integrated to calculate the multidimensional mean absolute error (MAR). If this MAR does not show a decreasing trend over 15 consecutive training generations, an early stopping module is triggered. After all nine incremental expansions are completed, for example, the ninth iteration uses the first nine subsets for joint training and the tenth subset for validation, nine models with a shared network structure but different weight mapping matrices in the high-dimensional space are saved.

[0062] When entering the online inference phase, the central control system inputs a set of real-time operating parameters, such as a spindle speed of 45 revolutions per minute and a temperature of 82°C. This data is immediately and synchronously input to nine networks for parallel processing, resulting in nine sets of predicted value vectors. The intradimensional arithmetic mean of these nine sets of data is calculated. This integration strategy reduces the prediction variance of individual models in time-varying production systems.

[0063] In some embodiments, the step of using an integrated prediction model to solve for the optimal combination of process parameters using a particle swarm optimization algorithm based on the quality indicators of the target compound fertilizer, and then performing closed-loop control of the compound fertilizer production process based on the optimal combination of process parameters, includes: The upper and lower limits of the hardware processing threshold of the physical manufacturing equipment are defined, and within the vector space defined by the upper and lower limits, a certain number of particle positions are randomly generated and corresponding velocity vectors are randomly assigned, so that any particle position can serve as a set of candidate process parameter vectors. The integrated prediction model is called to read in each set of candidate process parameter vectors one by one, and the corresponding multidimensional quality expectation index is calculated. The distance difference between each set of expected indexes and the quality index of the predetermined target compound fertilizer after dimensionless processing is calculated. The reciprocal of the difference is used as the fitness score for performing the selection operation on each set of particle positions. The entire process records and updates the historical best fitness space position with the highest fitness score obtained by each individual particle during the search, and simultaneously updates the global best fitness space position with the highest fitness score generated during the search of the swarm of particles. By using step-size iterative relationship calculations and learning and attenuation damping coefficients, the velocity vector and position vector of each particle are updated, prompting the particle search process to converge toward the optimal region. Once the predetermined number of steps for terminating the loop iteration is reached, the globally optimal fitness space position locked at this point is converted into a set execution parameter signal, which transmits instructions to the programmable logic controller to execute the updated control actions to the underlying quantitative feeding execution components and regulating valves.

[0064] For actual compound fertilizer production lines, safety boundaries are defined: for example, the upper limit of granulator speed is 60 revolutions per minute and the lower limit is 20 revolutions per minute; the upper and lower limits of spraying temperature are set at 100℃ and 60℃, respectively. Eighty individual particles with random initial velocity vectors are uniformly generated in this constrained three-dimensional parameter space using a Sobol sequence.

[0065] In each iteration, the aforementioned solidified multidimensional integrated prediction model acts as a virtual production line, reading 80 sets of test parameters in parallel and providing the predicted indicator distribution. When the target order provides a clear target value, the dimensionless distance between the predicted indicator and the target quality indicator vector is calculated. When the target order provides upper and lower limit standards, the constraint violation amount is calculated; for example, the moisture violation amount is max(0, predicted moisture - 1.5%), and the total nutrient violation amount is max(0, 45% - predicted total nutrient), and a dimensionless comprehensive distance is constructed based on each constraint violation amount. The reciprocal of the sum of this comprehensive distance and the positive small offset value of 1e-6 is taken as the fitness score of the current parameter combination, and the direction of particle swarm optimization is taken as the direction of maximum fitness score.

[0066] As the search process progresses, the particle swarm system employs a decreasing inertia weight strategy, where the inertia weight linearly decreases from 0.9 to 0.4 with each iteration, and both the individual cognitive learning factor and the social learning factor are configured to be 2. This ensures a broad global search capability in the early stages and enhances the ability to shrink and approach extreme regions for local optimization in the later stages.

[0067] The system continuously tracks and updates the optimal coordinates of each particle and the historical global optimal coordinates generated by all 80 particles. The loop stops when the number of iterations reaches a preset limit of 150 steps. The optimal spatial point information locked at this point, such as the optimal operating condition calculation result of 42.5 revolutions per minute and 86.2°C, is compiled into standard industrial control protocol data frames such as Profinet or Modbus-TCP control words, and control commands are issued to the programmable logic controller (PLC). Upon receiving the update signal, the PLC synchronously converts the digital quantity into a 4-20mA current signal to drive the frequency converter and steam pneumatic regulating valve, achieving the closed-loop control objective of online calculation and minute-level process control.

[0068] This ablation experiment used 10,000 continuous operational data points collected over one year from a real compound fertilizer production line as the sample library. The target production formula was set as total nutrients greater than 45% and moisture not exceeding 1.5%. The experimental hardware employed an industrial server platform equipped with a high-performance computing module. The validation set consisted of 800 batch operation data points collected over a subsequent month of continuous operation from the production line. Four comparative schemes were set up: the baseline model used a conventional feedforward neural network combined with a basic particle swarm optimization algorithm; group two introduced a Riemannian-based Ricci curvature criticality assessment mechanism on top of the baseline model; group three further incorporated clustering hierarchical and critical index balancing partitioning strategies on top of group two; and group four employed a complete closed-loop control scheme including a time-incrementing integrated neural network mechanism.

[0069] After multiple rounds of testing, the baseline model's mean absolute error (MAO) was 1.24%, with three instances of gradient divergence during training. The actual closed-loop control yield was 82.5%. Group 2 identified 42 high-critical samples near phase transition or mass fluctuation boundaries using Riemannian manifold metric tensors, reducing the model's MAO to 0.85% and increasing the yield to 89.2%. Group 3, employing clustering hierarchies and critical quota fine-tuning, reduced the relative standard deviation of the sum of critical indices for each validation subset to below 4.2%, improving training stability and decreasing the MAO to 0.56%. Group 4, after running the complete scheme and implementing inverse solution control, further reduced the MAO to 0.31%, stabilized the moisture content of the finished product at around 1.2%, and increased the overall yield to 98.8%. A comparison of the error convergence curves of the baseline method and the proposed method is provided. Figure 2 As shown.

[0070] The experimental data from each group show that conventional prediction methods lack the ability to characterize local geometric changes in high-dimensional mass space, and are prone to prediction bias when there are many process boundary samples. By introducing a continuous metric tensor field and a Ricci curvature scalar, the system can quantitatively identify critical conditions in the mass space. Furthermore, by combining a hierarchical basic data clustering and boundary fine-tuning allocation strategy, the distribution of hazardous condition samples in the training and validation sets becomes more balanced, reducing the impact of concentrated critical sample distribution on model training stability. The combination of a time-forward ensemble inference architecture and particle swarm optimization improves the stability of prediction results while maintaining temporal verification logic, and provides more reliable model support for the reverse solution of process parameters, thereby enhancing the prediction accuracy and regulation stability of the compound fertilizer production process.

[0071] This invention also discloses a data processing-based compound fertilizer production control system, including a processor and a memory. The memory stores computer program instructions, which, when executed by the processor, implement a data processing-based compound fertilizer production control method according to the present invention.

[0072] The system also includes other components well known to those skilled in the art, such as communication buses and communication interfaces, the settings and functions of which are known in the art and will not be described in detail here.

[0073] In the description of this specification, "multiple" or "several" means at least two, such as two, three or more, unless otherwise expressly and specifically defined.

Claims

1. A method for controlling compound fertilizer production based on data processing, characterized in that, Includes the following steps: S1. Obtain historical production data containing multidimensional quality indicators, construct the vector space composed of multidimensional quality indicators into a Riemannian manifold, set a metric tensor for the sample points, calculate the Ricci curvature scalar of the neighborhood of each sample point based on the metric tensor, and use the absolute value of the curvature scalar as an index representing the criticality of the production state. S2 divides all samples into several data clusters in the quality index space as basic hierarchical units. According to the global time order, all samples are divided into K non-overlapping time series data subsets. Without destroying the time order, the boundaries of adjacent subsets are fine-tuned in combination with the distribution of basic hierarchical units and criticality index, so as to reduce the hierarchical ratio deviation and criticality index difference of each subset. S3: The production quality prediction model is trained by taking turns using the first k time series data subsets according to the time-incrementing principle, and verified by the next k+1 time series data subset. k increases from 1 to K-1 to obtain and merge the integrated prediction model. Based on the quality indicators of compound fertilizer, the optimal combination of process parameters is solved in reverse by the particle swarm optimization algorithm to achieve closed-loop control of the compound fertilizer production process.

2. The method according to claim 1, characterized in that, The process of constructing a Riemannian manifold from the vector space of multidimensional quality indicators and setting a metric tensor for the sample points includes: Extract a set of neighboring sample points within a preset distance for each sample point; By utilizing process parameters and multidimensional quality index data from the set of neighboring sample points, a mapping function is fitted using a local polynomial and the partial derivative of the mapping function is obtained to acquire the Jacobian matrix of the multidimensional quality index. Calculate the matrix product of the Jacobian matrix and the transpose of the Jacobian matrix as the spatial influence matrix; Calculate the absolute value of the timestamp difference between the sample point and each neighboring sample point in the neighboring sample point set, convert the absolute value of the timestamp difference into a time weight, and sum them to obtain the aggregated time weight; The spatial influence matrix is ​​multiplied by the aggregated time weight and a preset dimensionless diagonal regularization term is added to ensure that the matrix is ​​positive definite, and the result is used as the metric tensor of the sample points.

3. The method according to claim 1, characterized in that, The index used to represent the criticality of the production state includes: The metric tensor of each sample point is parameterized while maintaining a continuously differentiable and positive definite constraint to obtain an interpolable parameter vector corresponding to the metric tensor. The interpolable parameter vector is continuously interpolated using a radial basis function network, and a continuous metric tensor field is reconstructed in a multidimensional quality index space, so that the reconstructed continuous metric tensor field remains symmetric, positive definite and invertible. The Ricci curvature scalar of the neighborhood of each sample point is calculated based on the continuous metric tensor field, and the absolute value of the curvature scalar is used as an index representing the criticality of the production state.

4. The method according to claim 3, characterized in that, The index representing the criticality of the production state includes: calculating the first type of Christofel symbol and the second type of Christofel symbol by analytical differentiation based on the continuous metric tensor field and the corresponding inverse matrix field; By taking the partial derivatives of the corresponding spatial coordinate variables using the Christofel symbol of the second kind, and combining them with the product terms of the Christofel symbol of the second kind, the corresponding Riemann curvature tensor is obtained. Tensor shrinkage is performed on the spatial coordinate index of the Riemann curvature tensor, and the Ricci tensor is obtained by computation; The inverse matrix of the Ricci tensor and the continuous metric tensor field is shrunk again to obtain the trace corresponding to the operation matrix, and the value of the trace is extracted as the Ricci curvature scalar. The absolute value of the Ricci curvature scalar at the sample point is assigned as an index of the criticality of the production state at the sample point.

5. The method according to claim 1, characterized in that, The step of dividing all samples into several data clusters in the quality index space as basic hierarchical units includes: Call the clustering algorithm and randomly assign the same number of initial centroid coordinates as the preset clusters; Calculate the Euclidean distance of all samples relative to each initial centroid in the quality index space, and assign each sample to the cluster containing the centroid with the smallest distance value from the sample. Calculate the arithmetic mean of the multidimensional quality index vector of all samples within each cluster, and use the arithmetic mean to replace the original centroid coordinates of the corresponding cluster. Repeat the above distance calculation and centroid substitution steps until the centroid coordinates no longer shift after iteration and stop updating. At this point, all samples within the same cluster are integrated into a basic hierarchical unit.

6. The method according to claim 1, characterized in that, The method to reduce the stratification ratio deviation and criticality index difference of each subset includes: dividing all samples into K consecutive and non-overlapping time intervals according to timestamp order; determining the target sample quota of each subset in each basic stratification unit based on the total number of baseline target samples and the sample ratio of each basic stratification unit; allocating the remaining samples after rounding according to the decimal remainder, the current stratification ratio deviation, the cumulative criticality index, and the deviation; if the number of samples in any stratification unit is less than K, allowing some subsets to not be allocated samples in that stratification unit, with the allocation completed under the constraints of non-overlapping global time order and minimizing stratification ratio deviation; within the boundary window of adjacent time intervals, performing boundary fine-tuning based on the criticality index of the samples to ensure that the time order of adjacent subsets does not overlap after fine-tuning, thereby reducing the basic stratification ratio deviation of each subset and making the cumulative criticality index of each subset tend to be balanced.

7. The method according to claim 1, characterized in that, The process of obtaining and fusing the integrated prediction model includes: constructing a multilayer feedforward neural network model with a set number of hidden layers and neuron node parameters; in the k-th iteration of the K-1 iterations, retaining the (k+1)-th time series data subset as an independent data subset for verification, feeding the first k time series data subsets into the multilayer feedforward neural network model, and adjusting and optimizing the weight nodes of each layer; substituting the independent data subset retained in this iteration into the multilayer feedforward neural network with adjusted weights to calculate the predicted index, calculating the mean absolute error by subtracting it from the actual index, and after the K-1 iterations are completed, saving K-1 multilayer feedforward neural networks with mutually independent training weights; synchronously inputting the candidate process parameters into the K-1 multilayer feedforward neural networks with independent training weights to perform feedforward inference, collecting multiple sets of quality prediction results, calculating the arithmetic mean of each index dimension, and integrating them into the output answer of the integrated prediction model.

8. The method according to claim 1, characterized in that, The closed-loop control of the compound fertilizer production process based on the optimal combination of process parameters includes: limiting the upper and lower limits of the hardware processing threshold of the physical manufacturing equipment; randomly generating a specific number of particle positions within the vector space defined by the upper and lower limits of the threshold, and randomly assigning corresponding velocity vectors, so that any particle position can serve as a set of candidate process parameter vectors; calling the integrated prediction model to read in each set of candidate process parameter vectors one by one, and calculating the corresponding multidimensional quality expectation index; calculating the distance difference between each set of expectation indexes and the quality index of the predetermined target compound fertilizer, and using the reciprocal of the difference to map the fitness score for performing the selection operation on each set of particle positions; The optimal fitness spatial location is determined based on the fitness score, and the compound fertilizer production process is then controlled in a closed loop based on this optimal fitness spatial location.

9. The method according to claim 8, characterized in that, The closed-loop control of the compound fertilizer production process based on the global optimal fitness spatial location includes: The entire process records and updates the historical best fitness space position with the highest fitness score obtained by each individual particle during the search, and simultaneously updates the global best fitness space position with the highest fitness score generated during the search of the swarm of particles. By using step-size iterative relationship calculations and learning and attenuation damping coefficients, the velocity vector and position vector of each particle are updated, prompting the particle search process to converge toward the optimal region. Once the predetermined number of steps for terminating the loop iteration is reached, the globally optimal fitness space position locked at this point is converted into a set execution parameter signal, which transmits instructions to the programmable logic controller to execute the updated control actions to the underlying quantitative feeding execution components and regulating valves.

10. A compound fertilizer production control system based on data processing, characterized in that, include: A processor and a memory, the memory storing computer program instructions that, when executed by the processor, implement a data processing-based compound fertilizer production control method according to any one of claims 1-9.

Images