Pet feed formula precision optimization method based on industrial big data

Abstract

The present application belongs to the technical field of industrial big data, and particularly relates to a pet feed formula precision optimization method based on industrial big data. The method collects multi-source data to construct a heterogeneous data warehouse, constructs a feature correlation model based on original chemical index volatility and production process parameter sensitivity, and generates a multi-dimensional dynamic index structure; according to the target constraint condition, the historical data subset is positioned, and a local weighted regression model is constructed to calculate the optimal formula parameter. The scheme realizes fast data positioning by using the multi-dimensional dynamic index structure, reduces the data reading delay in the calculation process, and improves the precision of the formula optimization result and the production feasibility.

Description

Technical Field

[0001] This invention belongs to the field of industrial big data technology, specifically relating to a method for precise optimization of pet food formulations based on industrial big data. Background Technology

[0002] Currently, in the field of pet food formulation optimization, industrial big data is primarily stored in relational databases. Existing technical solutions typically store raw material data from the supply chain and process data from the production line in separate data tables. When retrieving historical data, the system uses a traversal query method or is based on simple keyword matching. This storage architecture mainly aims to meet basic data read and write requirements and ensure data entry and backup. The formulation calculation model runs on a static dataset, assuming that the various physicochemical properties of the raw materials remain constant. Existing data processing workflows only perform basic cleaning of abnormal data and do not delve into the potential connections between different data sources.

[0003] When performing formulation calculations, existing technologies typically extract current raw material index data, combine it with preset nutritional standards, and use linear programming or standard regression algorithms to solve for the optimal ratio. This process treats process parameters such as temperature and pressure on the production line as fixed constraints, ignoring the sensitivity of raw material characteristic fluctuations to these parameters. The system does not establish a specific index structure to filter historical samples highly relevant to the current production status during optimization. The optimization model performs calculations directly on the full dataset, failing to distinguish the differences in reference value of different historical data samples for current formulation optimization. This calculation method based on full dataset leads to a waste of computational resources and increased computation time.

[0004] With the continuous growth of industrial data, the core problem with existing technologies lies in the lack of a multi-dimensional dynamic index structure based on feature correlation. This makes it impossible to quickly locate effective samples that meet the requirements of current raw material volatility and production parameter sensitivity from massive historical data. Because the indexing mechanism fails to establish a mapping between the variability rate of raw material nutrients and the threshold values ​​for temperature and pressure parameter changes in the production line, the system is inefficient during retrieval and easily introduces a large amount of non-correlated data as noise interference. This reduces the computational accuracy and adaptability of the formulation optimization model, making the generated formulation schemes unable to cope with changes in the production environment caused by raw material fluctuations. Summary of the Invention

[0005] The purpose of this invention is to provide a method for precise optimization of pet food formulation based on industrial big data, which can solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A method for precise optimization of pet food formulation based on industrial big data includes the following steps: collecting raw material supply chain data, production line sensor time-series data, and finished product quality inspection data to construct a heterogeneous data warehouse; constructing a feature association model based on the volatility of raw material physicochemical indicators and the sensitivity of production process parameters; establishing a mapping relationship between the variation rate of raw material nutrient components and the temperature and pressure parameter change threshold of the production line using the feature association model to generate a multi-dimensional dynamic index structure; locating a subset of historical data that meets production feasibility based on the target nutritional constraints and cost range using the multi-dimensional dynamic index structure; and constructing a local weighted regression model based on the historical data subset to calculate and output the optimal formulation parameters.

[0008] Preferably, the steps for constructing a heterogeneous data warehouse specifically include: cleaning and denoising the collected raw material supply chain data, production line sensor time-series data, and finished product quality inspection data; identifying and removing abnormal records containing missing fields or exceeding physical extreme value ranges; converting the cleaned unstructured text data into structured numerical vectors; mapping data from different sources to a unified time axis for time-series alignment; and storing the aligned data in a distributed database according to the dimensions of raw material identification, production batch, and inspection time, forming a standardized multi-source heterogeneous dataset with a unified time benchmark.

[0009] Preferably, the steps of constructing the feature association model specifically include: calculating the mean and variance of the physicochemical indicators of each batch of raw materials within a predetermined time window, and using the variance as the volatility of the raw material physicochemical indicators; simultaneously analyzing the correlation coefficient between the changes in temperature and pressure parameters of the production line and the changes in finished product indicators, using the correlation coefficient as the sensitivity of the production process parameters; using the Pearson correlation coefficient algorithm to calculate the correlation strength between the volatility of the raw material physicochemical indicators and the sensitivity of the production process parameters; and constructing a feature association matrix.

[0010] Preferably, the steps for generating a multidimensional dynamic index structure specifically include: discretizing the correlation strength values ​​in the feature correlation matrix to generate multi-level index keys; using the multi-level index keys to establish an R-tree index structure in the database; the leaf nodes of the R-tree index structure store historical production records that conform to a specific correlation strength, and the non-leaf nodes store pointers to child nodes and the boundaries of the correlation strength range, thereby realizing the rapid location of the mapping relationship between the variation rate of raw material nutrient components and the temperature and pressure parameter change threshold of the production line.

[0011] Preferably, the steps for calculating and outputting the optimal formula parameters specifically include: obtaining the target nutrient constraint vector and cost range value in the current formula calculation request; traversing the R-tree index structure through the multi-dimensional dynamic index structure; selecting historical data with raw material index fluctuations within a preset tolerance range and matching production parameters as sample points; assigning weights based on the Euclidean distance between the sample points and the current query conditions; and using the weighted least squares method to solve the coefficients of the local weighted regression model to obtain the optimal formula parameters.

[0012] Preferably, during the cleaning and denoising process of the collected data, the isolated forest algorithm is used to identify abnormal values ​​in the raw material supply chain data and mark the identified abnormal values ​​as invalid data. At the same time, for missing values ​​in the time series data of the production line sensor, cubic spline interpolation is used to fit and fill in the missing values ​​based on the data values ​​of the adjacent preset number of time points, so as to ensure the continuity of the data input into the heterogeneous data warehouse in the time series and the smoothness of the values.

[0013] Preferably, the multidimensional dynamic index structure also includes a dynamic update mechanism. The dynamic update mechanism monitors the data volume threshold of newly added data in the heterogeneous data warehouse. When the number of newly added data reaches the preset update threshold, the fluctuation of raw material physicochemical indicators and the sensitivity of production process parameters, including the newly added data, are recalculated, and the feature association matrix is ​​updated. The node splitting and merging strategy of the R-tree index structure is adjusted according to the updated feature association matrix to maintain the real-time consistency between the multidimensional dynamic index structure and the current industrial big data distribution characteristics.

[0014] Preferably, after solving the coefficients of the local weighted regression model using the weighted least squares method, the obtained optimal formula parameters are substituted into a preset production constraint function for verification. The production constraint function includes the maximum capacity limit of the equipment and the available inventory of raw materials. If the verification result does not meet the production constraint function, the weight allocation ratio of the sample points is adjusted and iterative calculation is performed again until the optimal formula parameters that meet all preset conditions are output.

[0015] Preferably, the method further includes a step of visualizing the formula optimization results, constructing a visualization interface, obtaining the deviation values ​​between the optimal formula parameters and the standard nutritional formula parameters, generating a radar chart in the visualization interface, wherein the axes of the radar chart represent different nutritional index dimensions, and the deviation values ​​are displayed by highlighting points at the corresponding positions of the axes, and the trend curve of the optimal formula parameters changing over historical time is displayed in the visualization interface by a line graph.

[0016] Preferably, the method further includes an automatic production parameter distribution step, establishing a communication connection with the production line control system, converting the calculated optimal formula parameters into a control instruction set that conforms to the communication protocol of the production line control system, wherein the control instruction set includes raw material ratio setting values ​​and mixer mixing time setting values, and sending the control instruction set to the production line control system through a preset application programming interface to control the automatic batching equipment to perform pet food production and processing actions according to the optimal formula parameters.

[0017] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0018] 1. This invention achieves a mapping between the variation rate of raw material nutrient components and the threshold values ​​of temperature and pressure parameter changes in the production line by constructing a feature association model and generating a multi-dimensional dynamic index structure. This index structure enables the system to quickly locate a subset of historical data that meets production feasibility, skipping a large amount of irrelevant data based on target nutrient constraints and cost ranges. This feature association-based indexing method shortens the response time in the data retrieval process. Simultaneously, the solution utilizes the located historical data subset to construct a locally weighted regression model, allocating weights based on the Euclidean distance between sample points and query conditions, making the calculation process focus more on historical data similar to the current production state. This approach reduces the interference of irrelevant data on the calculation results, improves the accuracy of optimal formula parameter calculation, and solves the problem of insufficient optimization accuracy caused by low data retrieval efficiency and weak correlation in existing technologies.

[0019] 2. This invention employs the isolated forest algorithm to identify outliers and utilizes cubic spline interpolation to fill in missing values, ensuring the continuity and smoothness of the data input into the data warehouse over time, and improving the stability of subsequent calculations. The multi-dimensional dynamic index structure has a dynamic update mechanism; when the amount of new data reaches a threshold, the node splitting and merging strategy of the R-tree index structure is readjusted to ensure that the index structure always remains consistent with the current distribution characteristics of industrial big data, maintaining long-term stability of retrieval performance. Verifying the calculation results through preset production constraint functions avoids outputting formula parameters that exceed the maximum capacity limits of equipment or the available raw material inventory, enhancing the executability of the optimization scheme on the production line. Furthermore, converting the formula parameters into a control instruction set conforming to the communication protocol and sending it to the production line control system realizes automated flow from formula calculation to production execution. Attached Figure Description

[0020] Figure 1 This is a flowchart of the industrial big data acquisition and heterogeneous data warehouse construction process of the present invention;

[0021] Figure 2 This is a flowchart illustrating the construction of the raw material-process feature association model of the present invention;

[0022] Figure 3 This is a flowchart illustrating the generation and dynamic updating of the multidimensional dynamic index structure of the present invention.

[0023] Figure 4 This is a flowchart illustrating the process of locating a subset of historical data that meets production feasibility requirements according to the present invention.

[0024] Figure 5 This is a flowchart of the optimal formula parameter calculation and production constraint verification process for this invention.

[0025] Figure 6 This is a flowchart illustrating the visualization of formula optimization results and the automatic distribution of production parameters for this invention. Detailed Implementation

[0026] refer to Figure 1 In one embodiment, the pet food formulation precision optimization method based on industrial big data first collects raw material supply chain data, production line sensor time-series data, and finished product quality inspection data to construct a heterogeneous data warehouse. The raw material supply chain data includes information such as raw material name, origin, supplier, purchase batch, warehousing time, protein content, fat content, crude fiber content, ash content, moisture content, amino acid composition, and mineral content. The production line sensor time-series data includes real-time acquisition values ​​of parameters such as mixer speed, mixing time, conditioning temperature, conditioning pressure, pelleting temperature, pelleting pressure, cooling temperature, cooling time, puffing temperature, and puffing pressure, with an acquisition frequency of once per second. The finished product quality inspection data includes test results such as finished product batch number, production date, protein content, fat content, crude fiber content, ash content, moisture content, pellet hardness, pellet uniformity, water stability, and microbial indicators.

[0027] refer to Figure 2 Based on the volatility of raw material physicochemical indicators and the sensitivity of production process parameters, a feature correlation model is constructed. The volatility of raw material physicochemical indicators reflects the degree of variation in nutritional composition among different batches of raw materials, while the sensitivity of production process parameters reflects the degree of impact of changes in production process parameters on the quality indicators of the finished product. By analyzing the correlation between the volatility of raw material physicochemical indicators and the sensitivity of production process parameters, the feature correlation model reveals the intrinsic link between raw material characteristic fluctuations and production process adjustments.

[0028] refer to Figure 3A feature association model is used to establish a mapping relationship between the variability rate of raw material nutrients and the temperature and pressure parameter change thresholds of the production line, generating a multi-dimensional dynamic index structure. The variability rate of raw material nutrients refers to the relative deviation between the actual nutrient composition of the raw material and the standard nutrient composition. The temperature and pressure parameter change thresholds of the production line refer to the maximum allowable adjustment range of the temperature and pressure parameters of the production line while ensuring the quality of the finished product. The multi-dimensional dynamic index structure organizes and stores the mapping relationship between the variability rate of raw material nutrients and the temperature and pressure parameter change thresholds of the production line, enabling rapid retrieval of historical production data.

[0029] refer to Figure 4 Based on target nutritional constraints and cost ranges, a multi-dimensional dynamic index structure is used to locate historical data subsets that meet production feasibility requirements. Target nutritional constraints include the minimum and maximum values ​​of various nutritional indicators that the finished product needs to achieve, while the cost range includes the upper and lower limits of raw material costs per unit of finished product. The historical data subset that meets production feasibility requirements refers to the set of production records in historical production that used raw materials with similar characteristics, adopted similar production process parameters, and produced finished products that met target nutritional constraints and whose costs were within the specified range.

[0030] refer to Figure 5 A locally weighted regression model is constructed based on a subset of historical data to calculate and output the optimal formulation parameters. The locally weighted regression model is a non-parametric regression method that builds a locally weighted linear regression model for each query point, assigning greater weight to sample points closer to the query point. The optimal formulation parameters include the proportions of various raw materials, mixing time, conditioning temperature, conditioning pressure, granulation temperature, and granulation pressure.

[0031] In this embodiment, the process of collecting raw material supply chain data is achieved through integration with the enterprise's ERP system, which stores information on the procurement, warehousing, and inspection of all raw materials. The process of collecting time-series data from production line sensors is achieved through integration with an Industrial Internet of Things (IIoT) platform. This platform connects to all sensor devices on the production line, collecting and storing various parameters during the production process in real time. The process of collecting finished product quality inspection data is achieved through integration with a laboratory information management system, which stores the quality inspection results of all finished products.

[0032] The process of building a heterogeneous data warehouse begins with a unified format conversion of the collected multi-source data, transforming data from different systems with varying formats into a standardized JSON format. Next, metadata management is applied to the converted data, adding metadata information such as data source, collection time, data type, and data precision to each data item. Finally, the processed data is stored in a distributed database using the HBase architecture, which offers high availability, high scalability, and high read / write performance.

[0033] The process of constructing the feature association model first involves standardizing the physicochemical indicators of raw materials and production process parameters, converting indicators of different dimensions and orders of magnitude into dimensionless standard scores. Then, the mean and variance of each physicochemical indicator for each batch of raw materials are calculated, with the variance representing the variability of the physicochemical indicators for that batch. Next, the correlation coefficient between each production process parameter and the finished product quality indicator is calculated, and this correlation coefficient is used as the sensitivity of that production process parameter. Finally, the correlation strength between the variability of raw material physicochemical indicators and the sensitivity of production process parameters is calculated, constructing a feature association matrix.

[0034] The process of generating a multidimensional dynamic index structure first discretizes the association strength values ​​in the feature association matrix, dividing the continuous association strength values ​​into several discrete intervals, each interval corresponding to an index key value. Then, these index keys are used to build an R-tree index structure in the distributed database. Each node in the R-tree index structure corresponds to a multidimensional spatial region. Leaf nodes store the primary keys of historical production records, while non-leaf nodes store pointers to child nodes and the boundaries of the multidimensional spatial regions covered by the child nodes.

[0035] The process of locating a subset of historical data that meets production feasibility first transforms the target nutrient constraints and cost range into multi-dimensional query conditions, with each query condition corresponding to a value range for one dimension. Then, a multi-dimensional dynamic index structure is used to traverse the R-tree index structure to quickly locate the leaf nodes that intersect with the query conditions. Finally, the primary keys of all historical production records are extracted from these leaf nodes, and complete historical production data is retrieved from the distributed database based on these primary keys, forming the historical data subset.

[0036] The process of constructing a locally weighted regression model first calculates the Euclidean distance between each sample point in the historical data subset and the current query condition. A smaller Euclidean distance indicates a more similar sample point to the current production state. Then, weights are assigned to each sample point based on the Euclidean distance, with closer sample points receiving larger weights. Finally, the coefficients of the locally weighted regression model are solved using the weighted least squares method to obtain the optimal formula parameters.

[0037] Table 1. Mapping Relationship Between Raw Material Nutrient Composition Variation Rate and Production Line Temperature and Pressure Parameter Change Thresholds

[0038] 0%-2% ±2 ±0.05 ±1 ±0.03 2%-5% ±4 ±0.10 ±2 ±0.06 5%-8% ±6 ±0.15 ±3 ±0.09 8%-12% ±8 ±0.20 ±4 ±0.12 >12% ±10 ±0.25 ±5 ±0.15

[0039] In this embodiment, Table 1 illustrates the mapping relationship between the variation rate of raw material nutrient composition and the threshold values ​​for temperature and pressure parameter changes in the production line. When the variation rate of raw material nutrient composition is in the range of 0%-2%, the temperature and pressure parameters of the production line can be adjusted within a small range; when the variation rate of raw material nutrient composition exceeds 12%, the temperature and pressure parameters of the production line need to be adjusted within a larger range to ensure the stability of the finished product quality. This mapping relationship was obtained by analyzing a large amount of historical production data, reflecting the objective law between raw material characteristic fluctuations and production process adjustments.

[0040] In this embodiment, the formula for calculating the variation rate of raw material nutrient composition is:

[0041]

[0042] in, Indicates the first Variation rate of nutrient composition of raw materials Indicates the first The actual measured values ​​of the nutritional components of the raw materials. Indicates the first Standard values ​​for the nutritional components of each raw material.

[0043] The formula for calculating the threshold of temperature and pressure parameter changes in the production line is:

[0044]

[0045] in, Indicates the corresponding to the first The first item of the variation rate of raw material nutrient composition The threshold for the change of temperature and pressure parameters Indicates the first Nutritional components of the raw materials and the first The proportionality coefficient between the temperature and pressure parameters is calculated using a feature correlation model.

[0046] The prediction formula for the locally weighted regression model is:

[0047]

[0048] in, Indicates at the query point The predicted value at that location, The feature matrix representing the sample points, Represents the target value vector of the sample points. This represents the weight matrix, which is a diagonal matrix where the elements on the diagonal are the weights of each sample point.

[0049] The formula for calculating the sample point weights is:

[0050]

[0051] in, Indicates the first The weight of each sample point Indicates the first The feature vector of each sample point The feature vector representing the query point. This represents the bandwidth parameter, which controls the rate at which the weights decay with distance.

[0052] In this embodiment, bandwidth parameters The bandwidth parameter is determined using cross-validation. A subset of historical data is divided into a training set and a validation set. A locally weighted regression model is constructed on the training set using different bandwidth parameters. The prediction error of the model is calculated on the validation set, and the bandwidth parameter that minimizes the prediction error is selected as the final bandwidth parameter.

[0053] This embodiment integrates multi-source industrial data by constructing a heterogeneous data warehouse, reveals the intrinsic relationship between raw material characteristic fluctuations and production process adjustments by building a feature association model, enables rapid retrieval of historical production data by generating a multi-dimensional dynamic index structure, and calculates optimal formulation parameters by constructing a local weighted regression model. This method can effectively address changes in the production environment caused by raw material fluctuations, improving the efficiency and accuracy of formulation optimization.

[0054] In a preferred embodiment, the steps of constructing a heterogeneous data warehouse specifically include cleaning and denoising the collected raw material supply chain data, production line sensor time-series data, and finished product quality inspection data; identifying and removing abnormal records containing missing fields or exceeding physical extreme value ranges; converting the cleaned unstructured text data into structured numerical vectors; mapping data from different sources to a unified time axis for time-series alignment; and storing the aligned data in a distributed database according to the dimensions of raw material identification, production batch, and inspection time, thereby forming a standardized multi-source heterogeneous dataset with a unified time benchmark.

[0055] In this embodiment, the data cleaning and denoising process first performs an integrity check on the collected raw data to identify records containing missing fields. For records with missing fields, if the missing field is a critical field, such as raw material identifier, production batch number, or testing time, the record is directly marked as invalid data and removed; if the missing field is a non-critical field, the record is retained, and the missing value is filled in in subsequent steps.

[0056] Then, the data undergoes a reasonableness check to identify abnormal records that exceed the physical extreme value range. The physical extreme value range is determined based on the physicochemical properties of the raw materials and the technical parameters of the production equipment. For example, the moisture content of the raw materials cannot exceed 100%, and the speed of the mixer cannot exceed its rated speed. Records that exceed the physical extreme value range are directly marked as invalid data and discarded.

[0057] Next, the unstructured text data undergoes structuring processing, transforming unstructured text descriptions into structured numerical vectors. For example, the origin information of raw materials is converted into geographic coordinate vectors, the supplier information of raw materials is converted into supplier reputation vectors, and the sensory evaluation information of finished products is converted into sensory quality score vectors. The structuring processing of unstructured text data is achieved using a pre-trained language model, which can convert text information into high-dimensional numerical vectors while preserving the semantic information of the text.

[0058] Then, a time-series alignment operation is performed to map data from different sources to a unified timeline. The time granularity of raw material supply chain data is days, the time granularity of production line sensor time-series data is seconds, and the time granularity of finished product quality inspection data is hours. The time-series alignment operation first converts the timestamps of all data to a unified UTC time format, then associates data from different sources according to the production batch number, and aggregates the raw material supply chain data and production line sensor time-series data according to the time granularity of the finished product quality inspection data to obtain aggregated data with the same time granularity as the finished product quality inspection data.

[0059] Finally, the aligned data is stored in a distributed database according to raw material identification, production batch, and testing time. The distributed database employs a columnar storage architecture, storing data of different dimensions in different column families to improve data retrieval efficiency. Simultaneously, a version number is added to each data item to record the data's modification history, ensuring data traceability.

[0060] During the cleaning and denoising process of the collected data, the isolated forest algorithm is used to identify abnormal values ​​in the raw material supply chain data and mark the identified abnormal values ​​as invalid data. At the same time, for missing values ​​in the time series data of the production line sensor, cubic spline interpolation is used to fit and fill in the missing values ​​based on the data values ​​of the adjacent preset number of time points, so as to ensure the continuity of the data input into the heterogeneous data warehouse in the time series and the smoothness of the values.

[0061] The Isolation Forest algorithm is an anomaly detection algorithm based on isolation. It constructs multiple decision trees by randomly selecting features and randomly splitting feature values. Anomalies are isolated more quickly within these decision trees, resulting in shorter average path lengths. In this embodiment, the input to the Isolation Forest algorithm is the numerical values ​​of various physicochemical indicators in the raw material supply chain data, and the output is the anomaly score for each data point. A higher anomaly score indicates that the data point is more likely to be an anomaly. When the anomaly score of a data point exceeds a preset threshold, the data point is marked as invalid data.

[0062] The formula for calculating the outlier score in the Isolation Forest algorithm is as follows:

[0063]

[0064] in, Representing data points Abnormal scores, This indicates the number of samples in the training set. Representing data points Average path length across all decision trees Indicates a given The average path length of the binary tree for each sample is calculated using the following formula:

[0065]

[0066] in, Indicates the first A harmonic number, , is Euler's constant, approximately equal to 0.5772.

[0067] Cubic spline interpolation is a piecewise interpolation method that constructs a cubic polynomial between each adjacent interpolation node, ensuring that the entire interpolation function is continuous at all nodes, and that its first and second derivatives are also continuous. In this embodiment, for missing values ​​in the time-series data of the production line sensors, data from five time points before and after the missing value are selected as interpolation nodes, and the cubic spline interpolation method is used to calculate the estimated value of the missing value.

[0068] The expression for the cubic spline interpolation function is:

[0069]

[0070] in, This represents the cubic spline interpolation function. Indicates the first The time of each interpolation node Indicates the first The coefficients of a cubic polynomial in each interval.

[0071] The coefficients of the cubic polynomial are obtained by solving the following system of equations:

[0072]

[0073] in, Indicates the first The values ​​of the interpolation nodes, express The first derivative, express The second derivative of .

[0074] Table 2 Comparison of data before and after cleaning

[0075] Raw material supply chain data 12568 324 156 12244 99.2 Production line sensor data 8640000 125600 89200 8514400 98.5 Finished Product Inspection Data 9876 156 89 9720 99.1 total 8662444 126080 89445 8536364 98.6

[0076] In this embodiment, Table 2 shows the comparison before and after data cleaning. After cleaning and denoising, a large number of invalid records were removed, missing values ​​were filled, and the data integrity rate was significantly improved. The cleaned data is more continuous in time series and the values ​​are smoother, providing a high-quality data foundation for subsequent feature association analysis and model construction.

[0077] In this embodiment, the specific process of time-series alignment is as follows: First, the raw material supply chain data and finished product quality inspection data are associated according to the production batch number, with each production batch corresponding to one raw material supply chain record and one finished product quality inspection record. Then, the production line sensor time-series data is divided according to production batches, with each production batch corresponding to a continuous segment of sensor time-series data. Next, the sensor time-series data of each production batch is aggregated according to time windows, and statistical characteristics such as the mean, variance, maximum value, and minimum value within each time window are calculated. Finally, the aggregated sensor time-series data is merged with the corresponding raw material supply chain data and finished product quality inspection data to form a complete production record.

[0078] The distributed database storage architecture employs a master-slave replication model. The master node is responsible for writing and updating data, while the slave nodes are responsible for reading and querying data. Data consistency is maintained between the master and slave nodes through asynchronous replication. When the master node fails, the system automatically promotes a slave node to become the new master, ensuring high availability.

[0079] This embodiment improves data quality and completeness by employing the Isolation Forest algorithm to identify outliers and cubic spline interpolation to fill in missing values. Time-series alignment links data from different sources, forming a complete production record. A distributed database storage architecture enables efficient storage and retrieval of massive amounts of industrial data.

[0080] In a preferred embodiment, the step of constructing a feature association model specifically includes calculating the mean and variance of the physicochemical indicators of each batch of raw materials within a predetermined time window, and using the variance as the volatility of the raw material physicochemical indicators. At the same time, the correlation coefficient between the changes in temperature and pressure parameters of the production line and the changes in finished product indicators is analyzed, and the correlation coefficient is used as the sensitivity of the production process parameters. The Pearson correlation coefficient algorithm is used to calculate the correlation strength between the volatility of the raw material physicochemical indicators and the sensitivity of the production process parameters, and a feature association matrix is ​​constructed.

[0081] In this embodiment, the predetermined time window size is 30 days, which involves calculating the mean and variance of various physicochemical indicators of all batches of raw materials over the past 30 days. A 30-day time window is chosen because the procurement cycle for raw materials is typically around one month, and a 30-day window can reflect recent fluctuations in raw material characteristics. The size of the time window can be adjusted according to actual circumstances. For example, when the raw material market fluctuates significantly, the time window size can be appropriately reduced to more promptly reflect changes in raw material characteristics.

[0082] The formula for calculating the average physicochemical properties of raw materials is:

[0083]

[0084] in, Indicates the first Within the first time window The average values ​​of the physicochemical properties of the raw materials. Indicates the first The number of raw material batches within a time window Indicates the first Within the first time window The first batch of raw materials The measured values ​​of the physicochemical indicators.

[0085] The formula for calculating the variance of the physicochemical properties of raw materials is:

[0086]

[0087] in, Indicates the first Within the first time window The variance of the physicochemical properties of raw materials, i.e. the volatility of the physicochemical properties of raw materials.

[0088] The calculation process for the sensitivity of production process parameters is as follows: First, calculate the deviation between each temperature and pressure parameter of each production batch and the standard temperature and pressure parameter. Then, calculate the deviation between each finished product quality index of each production batch and the standard finished product quality index. Finally, calculate the Pearson correlation coefficient between the temperature and pressure parameter deviation and the finished product quality index deviation, and use this correlation coefficient as the sensitivity of production process parameters.

[0089] The formula for calculating the Pearson correlation coefficient is:

[0090]

[0091] in, Indicates the first The temperature and pressure parameters of the first item are related to the first item. Pearson correlation coefficients among the finished product quality indicators Indicates the production batch quantity. Indicates the first The first production batch Deviation of temperature and pressure parameters This indicates the first of all production batches. The mean deviation of the temperature and pressure parameters. Indicates the first The first production batch Deviation of finished product quality indicators This indicates the first of all production batches. The mean of the deviations of the quality indicators of the finished products.

[0092] The Pearson correlation coefficient ranges from -1 to 1. A closer absolute value to 1 indicates a stronger linear correlation between the two variables; a positive correlation indicates a positive correlation, and a negative correlation indicates a negative correlation. In this embodiment, the absolute value of the Pearson correlation coefficient is used as the sensitivity of the production process parameter. A larger absolute value indicates a greater impact of the production process parameter on the finished product quality indicators.

[0093] The Pearson correlation coefficient algorithm was used to calculate the correlation strength between the volatility of raw material physicochemical indicators and the sensitivity of production process parameters, and a characteristic correlation matrix was constructed. The rows of the characteristic correlation matrix correspond to the raw material physicochemical indicators, the columns correspond to the production process parameters, and the elements in the matrix represent the Pearson correlation coefficients between the volatility of the corresponding raw material physicochemical indicators and the sensitivity of the production process parameters.

[0094] The expression for the feature correlation matrix is:

[0095]

[0096] in, Represents the characteristic correlation matrix. This indicates the quantity of the physicochemical properties of the raw materials. Indicates the number of production process parameters. Indicates the first The fluctuation of the physicochemical properties of the raw materials and the first Pearson correlation coefficient between the sensitivity of production process parameters.

[0097] The steps for generating a multidimensional dynamic index structure specifically include discretizing the correlation strength values ​​in the feature correlation matrix to generate multi-level index keys, and using the multi-level index keys to build an R-tree index structure in the database. The leaf nodes of the R-tree index structure store historical production records that meet specific correlation strengths, while the non-leaf nodes store pointers to child nodes and the boundaries of the correlation strength range, thereby enabling rapid location of the mapping relationship between the variation rate of raw material nutrient components and the temperature and pressure parameter change threshold of the production line.

[0098] In this embodiment, the discretization of association strength values ​​employs an equal-frequency discretization method, dividing continuous association strength values ​​into several intervals, each containing approximately the same number of samples. The equal-frequency discretization method avoids the problem of insufficient sample numbers in certain intervals due to uneven data distribution, thus improving the query efficiency of the index structure.

[0099] The generation process of multi-level index keys is as follows: First, the correlation strength values ​​of each row in the feature correlation matrix are discretized to generate the first-level index key value; then, the first-level index key values ​​are weighted and combined according to the importance of the raw material physicochemical indicators to generate the second-level index key value; finally, the second-level index key value is combined with the raw material nutrient variation rate and the production line temperature and pressure parameter change threshold to generate the third-level index key value. Multi-level index keys enable rapid multi-dimensional and multi-level retrieval of historical production data.

[0100] The construction process of the R-tree index structure is as follows: First, all historical production records are sorted according to the third-level index key value; then, records are selected sequentially from the sorted records to construct the leaf nodes of the R-tree, each leaf node containing a fixed number of records; next, the leaf nodes are grouped according to the range of their index key values ​​to construct the non-leaf nodes of the R-tree, each non-leaf node containing a fixed number of child nodes; finally, the upper-level nodes of the R-tree are recursively constructed until the root node.

[0101] The query process of the R-tree index structure is as follows: First, the query conditions are converted into the corresponding index key value range; then, starting from the root node, the R-tree is traversed to determine whether the index key value range covered by the child node of the current node intersects with the index key value range of the query conditions; if they intersect, the child node is traversed; if they do not intersect, the child node is skipped; finally, the leaf node is reached to extract all historical production records that meet the query conditions.

[0102] The multidimensional dynamic index structure also includes a dynamic update mechanism. This mechanism monitors the data volume threshold of newly added data in the heterogeneous data warehouse. When the number of newly added data reaches the preset update threshold, it recalculates the volatility of raw material physicochemical indicators and the sensitivity of production process parameters, including the newly added data, and updates the feature association matrix. Based on the updated feature association matrix, it adjusts the node splitting and merging strategy of the R-tree index structure to maintain the real-time consistency between the multidimensional dynamic index structure and the current industrial big data distribution characteristics.

[0103] In this embodiment, the preset update threshold is 1000 new production records. When the number of new production records in the heterogeneous data warehouse reaches 1000, the dynamic update mechanism is triggered. The dynamic update mechanism first recalculates the mean and variance of the physicochemical indicators of all batches of raw materials in the past 30 days, updating the volatility of the raw material physicochemical indicators; then it recalculates the Pearson correlation coefficient between production process parameters and finished product quality indicators, updating the sensitivity of production process parameters; next, it recalculates the correlation strength between the volatility of raw material physicochemical indicators and the sensitivity of production process parameters, updating the feature association matrix; finally, it regenerates multi-level index key values ​​based on the updated feature association matrix and adjusts the node splitting and merging strategy of the R-tree index structure.

[0104] The R-tree index structure employs a quadratic splitting algorithm. When the number of records in a node exceeds its maximum capacity, the node is split into two. The quadratic splitting algorithm first selects the two farthest records as seeds for the two new nodes, then assigns the remaining records sequentially to the closer seed nodes, until all records have been assigned.

[0105] The node merging strategy for an R-tree index structure is as follows: when the number of records in a node is less than the minimum capacity, the node is merged with its sibling nodes. If the merged node exceeds the maximum capacity, it is split according to the node splitting strategy.

[0106] Table 3 Example of Feature Correlation Matrix

[0107] Physicochemical properties of raw materials Temperature sensitivity of tempering Conditioning pressure sensitivity Granulation temperature sensitivity Granulation pressure sensitivity Protein content 0.78 0.65 0.82 0.71 Fat content 0.62 0.74 0.58 0.69 Crude fiber content 0.55 0.48 0.63 0.52 Ash content 0.42 0.39 0.45 0.37 Moisture content 0.85 0.79 0.88 0.81

[0108] In this embodiment, Table 3 shows an example of a feature correlation matrix. As can be seen from the table, the correlation strength between the fluctuation of moisture content and the sensitivity of various temperature and pressure parameters is the strongest, indicating that fluctuations in moisture content have the greatest impact on the adjustment of production process parameters. Conversely, the correlation strength between the fluctuation of ash content and the sensitivity of various temperature and pressure parameters is the weakest, indicating that fluctuations in ash content have the least impact on the adjustment of production process parameters. This difference in correlation strength provides a basis for adjusting production process parameters. When the moisture content of the raw materials fluctuates significantly, parameters such as conditioning temperature and granulation temperature should be adjusted first.

[0109] In this embodiment, the importance of the raw material physicochemical indicators is determined using principal component analysis (PCA). PCA can transform multiple related raw material physicochemical indicators into a few unrelated principal components, each of which can explain a portion of the variance of the original data. The larger the variance contribution rate of a principal component, the more important that principal component is. The variance contribution rate of each principal component is used as the weight of the corresponding raw material physicochemical indicator to generate the second-level index key value.

[0110] The calculation process of principal component analysis is as follows: First, the physicochemical index data of the raw materials are standardized so that their mean is 0 and their variance is 1; then, the covariance matrix of the standardized data is calculated; next, the eigenvalues ​​and eigenvectors of the covariance matrix are calculated; finally, the eigenvalues ​​are sorted in descending order, and the top eigenvalues ​​are selected. The eigenvectors corresponding to the largest eigenvalues ​​are taken as principal components.

[0111] The formula for calculating the variance contribution rate of principal components is:

[0112]

[0113] in, Indicates the first The variance contribution rate of each principal component Indicates the first 1 eigenvalue, This indicates the quantity of the physicochemical properties of the raw materials.

[0114] This embodiment reveals the correlation between the volatility of raw material physicochemical indicators and the sensitivity of production process parameters by constructing a feature correlation matrix. It achieves rapid retrieval of historical production data by generating multi-level index keys and an R-tree index structure, and maintains consistency between the index structure and the current data distribution characteristics through a dynamic update mechanism. This method effectively improves data retrieval efficiency and reduces the interference of non-related data on formula optimization.

[0115] In a preferred embodiment, the step of calculating and outputting the optimal formula parameters specifically includes obtaining the target nutrient constraint vector and cost range value in the current formula calculation request, traversing the R-tree index structure through a multi-dimensional dynamic index structure, selecting historical data with raw material index volatility within a preset tolerance range and matching production parameters as sample points, assigning weights according to the Euclidean distance between the sample points and the current query conditions, and using the weighted least squares method to solve the coefficients of the local weighted regression model to obtain the optimal formula parameters.

[0116] In this embodiment, the target nutrient constraint vector is a A dimensional vector, where This represents the number of nutritional indicators, where each element in the vector represents the target value or range of values ​​for the corresponding nutritional indicator. For example, the target nutritional constraint vector can be represented as... ,in Indicates the first The target value for each nutritional indicator. If there are limitations on the range of values ​​for a nutritional indicator, then... It can be represented as an interval .

[0117] The cost range includes the upper limit of the raw material cost per unit of finished product. and lower limit The formula for calculating the raw material cost per unit of finished product is:

[0118]

[0119] in, This represents the raw material cost per unit of finished product. Indicates the types and quantities of raw materials. Indicates the first The proportions of the raw materials Indicates the first The unit price of the raw materials.

[0120] The process of traversing the R-tree index structure using a multidimensional dynamic index structure first converts the target nutrient constraint vector and cost interval values ​​into corresponding query conditions. Then, it traverses from the root node of the R-tree, gradually locating the leaf nodes that meet the query conditions. Finally, it extracts all historical production records from the leaf nodes as candidate sample points.

[0121] The process of selecting sample points first calculates the volatility of raw material indicators for each candidate sample point to determine if it falls within a preset tolerance range. The preset tolerance range is determined based on the actual measured values ​​and historical volatility of the current raw materials; for example, it can be set to ±20% of the current raw material indicator volatility. Next, the matching degree between the production parameters of each candidate sample point and the current production equipment parameters is calculated. The matching degree is obtained by calculating the Euclidean distance between the production parameters; a smaller Euclidean distance indicates a higher matching degree. Finally, the sample points with the highest matching degree are selected. The candidate sample points are used as the final sample points. The value is usually between 50 and 200.

[0122] The formula for calculating the Euclidean distance between a sample point and the current query condition is:

[0123]

[0124] in, Indicates the first The Euclidean distance between each sample point and the current query condition Indicates the number of feature dimensions. Indicates the first The first sample point 1 eigenvalue, This indicates the first element of the current query condition. Each feature value.

[0125] Weights are assigned to each sample point based on Euclidean distance, with the weights inversely proportional to the Euclidean distance; sample points that are closer in distance have higher weights. The weights are calculated using a Gaussian kernel function, as shown in the formula for calculating the Euclidean distance between a sample point and the current query condition.

[0126] The coefficients of a locally weighted regression model are solved using weighted least squares. Weighted least squares is an improved least squares method that assigns different weights to each sample point, making the model focus more on sample points with larger weights. The goal of weighted least squares is to minimize the weighted sum of squared residuals.

[0127]

[0128] in, This represents the weighted sum of squared residuals. This represents the coefficient vector of the locally weighted regression model. Indicates the first The target value for each sample point Indicates the first The feature vector of each sample point.

[0129] right Differentiating and setting it to 0, we obtain the normal equation for the weighted least squares method:

[0130]

[0131] in, The feature matrix representing the sample points, Represents the weight matrix. This represents the target value vector of the sample points.

[0132] Solving the normal equation yields the coefficient vector of the locally weighted regression model:

[0133]

[0134] Substituting the feature vector of the current query conditions into the local weighted regression model, we obtain the optimal formula parameters:

[0135]

[0136] After solving the coefficients of the local weighted regression model using the weighted least squares method, the obtained optimal formula parameters are substituted into the preset production constraint function for verification. The production constraint function includes the maximum capacity limit of the equipment and the available inventory of raw materials. If the verification result does not meet the production constraint function, the weight allocation ratio of the sample points is adjusted and the iterative calculation is performed again until the optimal formula parameters that meet all preset conditions are output.

[0137] The constraint function for limiting the maximum capacity of the equipment is:

[0138]

[0139] in, This indicates the equipment's maximum capacity, which is the number of finished products it can produce per unit of time.

[0140] The constraint function for limiting the availability of raw material inventory is:

[0141]

[0142] in, Indicates the planned quantity of finished products. Indicates the first Available inventory of various raw materials.

[0143] If the optimal formulation parameters do not satisfy the production constraint function, the weight allocation of the sample points is adjusted. The adjustment method is as follows: increase the weight of sample points that satisfy the production constraint function and decrease the weight of sample points that do not. Then, the coefficients of the local weighted regression model are solved again using the weighted least squares method to obtain new optimal formulation parameters. This process is repeated until the optimal formulation parameters satisfy all production constraint functions.

[0144] refer to Figure 6 The method for precise optimization of pet food formulation based on industrial big data also includes a visualization step for the formulation optimization results. A visualization interface is constructed to obtain the deviation values ​​between the optimal formulation parameters and the standard nutritional formulation parameters. A radar chart is generated in the visualization interface. The axes of the radar chart represent different nutritional index dimensions, and the deviation values ​​are displayed by highlighting the corresponding points on the axes. At the same time, the trend curve of the optimal formulation parameters changing over historical time is displayed in the visualization interface through a line graph.

[0145] In this embodiment, the visualization interface is developed using Web technology, supporting access on various devices such as computers, tablets, and mobile phones. The visualization interface includes three parts: a formula parameter display area, a nutritional deviation display area, and a historical trend display area. The formula parameter display area presents the optimal formula parameters in tabular form, including the proportions of various raw materials, mixing time, conditioning temperature, conditioning pressure, granulation temperature, and granulation pressure. The nutritional deviation display area uses a radar chart to show the deviation between the optimal formula parameters and the standard nutritional formula parameters. Each axis of the radar chart represents a nutritional indicator, the length of the axis indicates the numerical value of the nutritional indicator, and the highlighted points represent the nutritional indicator values ​​of the optimal formula parameters, which are compared with the values ​​of the standard nutritional formula parameters. The historical trend display area uses a line graph to show the changing trend of the optimal formula parameters over a period of time, helping production managers understand the changing patterns of the formula parameters.

[0146] The method for precise optimization of pet food formulation based on industrial big data also includes an automatic production parameter distribution step. This involves establishing a communication connection with the production line control system, converting the calculated optimal formulation parameters into a control instruction set that conforms to the communication protocol of the production line control system. The control instruction set includes raw material ratio settings and mixer mixing time settings. The control instruction set is sent to the production line control system through a preset application programming interface to control the automatic batching equipment to perform pet food production and processing actions according to the optimal formulation parameters.

[0147] In this embodiment, the production line control system adopts a PLC control system, and the communication protocol uses the Modbus TCP protocol. The process of establishing a communication connection with the production line control system first configures communication parameters, including IP address, port number, slave address, etc.; then establishes a TCP connection; and finally performs a communication handshake to confirm that the connection is normal.

[0148] The process of converting optimal formula parameters into a control instruction set first involves converting the optimal formula parameters into a numerical format that the PLC control system can recognize. For example, the raw material ratio is converted into the corresponding inverter frequency setting value, and the mixing time is converted into the corresponding timer setting value. Then, the control instructions are encapsulated according to the Modbus TCP protocol format, including function code, register address, data length, and data content. Finally, the control instruction set is generated.

[0149] The control command set is sent to the production line control system through a pre-defined application programming interface (API). The API uses a RESTful API style and supports the HTTP protocol. The process of sending the control command set involves first constructing an HTTP request with the control command set as the request body; then sending the HTTP request to the production line control system's API; and finally receiving the response from the production line control system to confirm that the control command has been successfully received and executed.

[0150] Table 4 Comparison of Optimal Formulation Parameters and Standard Formulation

[0151] Corn ratio (%) 60.0 58.5 -1.5 Soybean meal ratio (%) 25.0 26.2 +1.2 Fish meal ratio (%) 8.0 7.6 -0.4 Oil ratio (%) 4.0 4.5 +0.5 Mineral ratio (%) 2.0 2.1 +0.1 Vitamin ratio (%) 1.0 1.1 +0.1 Mixing time (min) 5.0 5.2 +0.2 Conditioning temperature (°C) 80.0 78.5 -1.5 Conditioning pressure (MPa) 0.30 0.28 -0.02 Granulation temperature (°C) 85.0 83.0 -2.0 Granulation pressure (MPa) 0.40 0.38 -0.02 Unit cost (yuan / kg) 3.20 3.12 -0.08

[0152] In this embodiment, Table 4 shows a comparison between the optimal formula parameters and the standard formula parameters. As can be seen from the table, the optimal formula parameters were slightly adjusted from the standard formula parameters, reducing the proportion of corn and fishmeal while increasing the proportions of soybean meal, oil, minerals, and vitamins, and appropriately adjusting the production process parameters. The unit cost of the optimal formula parameters is 0.08 yuan / kg lower than that of the standard formula parameters, while ensuring that the nutritional indicators of the finished product meet the requirements.

[0153] In this embodiment, the visualization step of the formula optimization results further includes generating a formula optimization report. The formula optimization report is generated in PDF format and includes a formula parameter comparison table, a nutrient deviation radar chart, a historical trend line graph, and a production feasibility analysis. The formula optimization report can be printed or exported for reference by production managers and technicians.

[0154] The automatic production parameter distribution process also includes production process monitoring. During production, various parameters of the production line are collected in real time and compared with the optimal formula parameters. When the deviation between the actual parameters and the optimal formula parameters exceeds a preset threshold, an alarm message is issued, and the production parameters are automatically adjusted to restore them to the range of the optimal formula parameters.

[0155] This embodiment calculates the optimal formulation parameters by constructing a local weighted regression model, verifies the production feasibility of the optimal formulation parameters through a production constraint function, displays the formulation optimization results through a visual interface, and achieves automated control of the production process by automatically issuing control commands. This method can improve the accuracy of formulation optimization and the level of automation in the production process, reduce production costs, and ensure the stability of finished product quality.

Claims

1. A method for precise optimization of pet food formulation based on industrial big data, characterized in that, Includes the following steps: Collect raw material supply chain data, production line sensor time-series data, and finished product quality inspection data to build a heterogeneous data warehouse; A feature correlation model is constructed based on the fluctuation of raw material physicochemical properties and the sensitivity of production process parameters; The feature association model is used to establish a mapping relationship between the variation rate of raw material nutrient components and the temperature and pressure parameter change threshold of the production line, and a multi-dimensional dynamic index structure is generated. Based on the target nutritional constraints and cost range, the multidimensional dynamic index structure is used to locate a subset of historical data that meets production feasibility requirements. A local weighted regression model is constructed based on the aforementioned subset of historical data to calculate and output the optimal formula parameters.

2. The method according to claim 1, characterized in that, The specific steps for building a heterogeneous data warehouse include: cleaning and denoising the collected raw material supply chain data, production line sensor time-series data, and finished product quality inspection data; identifying and removing abnormal records containing missing fields or exceeding physical extreme value ranges; converting the cleaned unstructured text data into structured numerical vectors; mapping data from different sources to a unified time axis for time-series alignment; and storing the aligned data in a distributed database according to the dimensions of raw material identification, production batch, and inspection time, forming a standardized multi-source heterogeneous dataset with a unified time benchmark.

3. The method according to claim 1, characterized in that, The specific steps for constructing a feature association model include: calculating the mean and variance of the physicochemical indicators of each batch of raw materials within a predetermined time window, and using the variance as the volatility of the raw material physicochemical indicators; simultaneously analyzing the correlation coefficient between the changes in temperature and pressure parameters of the production line and the changes in finished product indicators, using the correlation coefficient as the sensitivity of the production process parameters; using the Pearson correlation coefficient algorithm to calculate the correlation strength between the volatility of the raw material physicochemical indicators and the sensitivity of the production process parameters; and constructing a feature association matrix.

4. The method according to claim 3, characterized in that, The steps for generating a multidimensional dynamic index structure specifically include: discretizing the correlation strength values ​​in the feature correlation matrix to generate multi-level index keys; using the multi-level index keys to establish an R-tree index structure in the database; storing historical production records that meet specific correlation strengths in the leaf nodes of the R-tree index structure; and storing pointers to child nodes and the boundaries of the correlation strength range in the non-leaf nodes, thereby enabling rapid positioning of the mapping relationship between the variation rate of raw material nutrient components and the temperature and pressure parameter change threshold of the production line.

5. The method according to claim 4, characterized in that, The steps for calculating and outputting the optimal formula parameters specifically include: obtaining the target nutrient constraint vector and cost range values ​​in the current formula calculation request; traversing the R-tree index structure through the multi-dimensional dynamic index structure; selecting historical data with raw material index fluctuations within a preset tolerance range and matching production parameters as sample points; assigning weights based on the Euclidean distance between the sample points and the current query conditions; and using the weighted least squares method to solve the coefficients of the local weighted regression model to obtain the optimal formula parameters.

6. The method according to claim 2, characterized in that, During the cleaning and denoising process of the collected data, the isolated forest algorithm is used to identify abnormal values ​​in the raw material supply chain data and mark the identified abnormal values ​​as invalid data. At the same time, for missing values ​​in the time series data of the production line sensor, cubic spline interpolation is used to fit and fill in the missing values ​​based on the data values ​​of the adjacent preset number of time points, so as to ensure the continuity of the data input into the heterogeneous data warehouse in the time series and the smoothness of the values.

7. The method according to claim 4, characterized in that, The multidimensional dynamic index structure also includes a dynamic update mechanism. The dynamic update mechanism monitors the data volume threshold of newly added data in the heterogeneous data warehouse. When the number of newly added data reaches the preset update threshold, the fluctuation of raw material physicochemical indicators and the sensitivity of production process parameters, including the newly added data, are recalculated, and the feature association matrix is ​​updated. The node splitting and merging strategy of the R-tree index structure is adjusted according to the updated feature association matrix to maintain the real-time consistency between the multidimensional dynamic index structure and the current industrial big data distribution characteristics.

8. The method according to claim 5, characterized in that, After solving the coefficients of the local weighted regression model using the weighted least squares method, the obtained optimal formula parameters are substituted into the preset production constraint function for verification. The production constraint function includes the maximum capacity limit of the equipment and the available inventory of raw materials. If the verification result does not meet the production constraint function, the weight allocation ratio of the sample points is adjusted and the iterative calculation is performed again until the optimal formula parameters that meet all preset conditions are output.

9. The method according to claim 1, characterized in that, It also includes a step to visualize the results of formula optimization, constructing a visualization interface, obtaining the deviation values ​​between the optimal formula parameters and the standard nutritional formula parameters, generating a radar chart in the visualization interface, wherein the axes of the radar chart represent different nutritional index dimensions, and the deviation values ​​are displayed by highlighting the corresponding positions of the axes. At the same time, the trend curve of the optimal formula parameters changing over historical time is displayed in the visualization interface by a line graph.

10. The method according to claim 1, characterized in that, It also includes an automatic production parameter distribution step, establishing a communication connection with the production line control system, converting the calculated optimal formula parameters into a control instruction set that conforms to the communication protocol of the production line control system. The control instruction set includes raw material ratio setting values ​​and mixer mixing time setting values. The control instruction set is sent to the production line control system through a preset application programming interface to control the automatic batching equipment to perform pet food production and processing actions according to the optimal formula parameters.

Images