Feed formula optimization and production line parameter adaptive control method based on machine learning

Abstract

The present invention discloses a method for feed formula optimization and production line parameter adaptive control based on machine learning, comprising the following steps: S1: constructing a hierarchical polymorphic graph structure model; S2: calculating the behavioral differences between nodes to generate a tension field tensor; S3: screening path combinations with tension values below a set tension threshold to generate optimized individuals; S4: inputting the optimized individuals into an improved jellyfish search algorithm, which includes an active motion phase and a passive drift phase; S5: selecting the optimized individual with the highest fitness value as the optimal individual; S6: updating the hierarchical polymorphic graph structure model based on the formula parameters and process parameters of the optimal individual; S7: determining whether the current raw material batch or production line operating conditions have changed. If so, the method returns to step S2 and repeats steps S3 to S6; otherwise, the method proceeds to step S8; S8: outputting the formula parameters and process parameters of the current optimal individual. The present invention integrates the hierarchical polymorphic graph structure model with the improved jellyfish algorithm to achieve collaborative optimization of feed formula and process parameters.

Description

Technical Field

[0001] The present invention relates to the technical field of feed production process optimization, and in particular to a feed formula optimization and production line parameter adaptive control method based on machine learning. Background Art

[0002] In the modern feed industry, formula optimization and production line process control are key links in achieving high-quality and high-efficiency production. Feed formula involves the combination of multiple raw materials. There are significant differences in nutritional value, palatability, stability and cost between different raw materials. At the same time, feed production process parameters such as steam temperature, extrusion pressure, water addition, oil injection amount, etc. will directly affect the final product's pellet hardness, molding rate and animal growth performance. Traditional feed design relies on the experience of nutrition experts and experimental verification to repeatedly adjust the formula. Process control also often uses static setting parameters and lacks refined modeling and dynamic optimization capabilities. As a result, the formula plan cannot fully adapt to fluctuations in production conditions and dynamic changes in target performance, affecting product consistency and resource utilization efficiency.

[0003] To improve the efficiency of feed formulation and the stability of the production process, the industry has gradually introduced computer-aided formulation design systems and production process control systems in recent years. Some companies use formulation recommendation methods based on linear programming or rule engines, and introduce expert systems to set rules for process parameters and trigger alarms. This type of system has improved the level of operational automation to a certain extent, but still lacks a collaborative modeling mechanism between the formulation and process levels. The mutual coupling effect between formulation and process cannot be modeled, resulting in the inability to accurately adjust the process response when the formulation changes; conversely, it is difficult to quickly match new formulation solutions when production conditions fluctuate. In addition, traditional optimization algorithms such as genetic algorithms, simulated annealing, and particle swarm optimization are limited in the scenario of collaborative optimization of feed formulation and process due to simple model structures, insufficient search efficiency, or weak path control capabilities. It is difficult to quickly find a formulation combination that meets multiple performance objectives and has a coordinated tension in a complex high-dimensional parameter space.

[0004] Existing applications of machine learning in feed optimization mostly focus on classification or regression prediction of historical recipes and production results, and a complete dynamic adaptive optimization mechanism has yet to be formed. While previous studies have attempted to use neural networks for performance prediction, these have primarily been used as black-box models to fit existing data. They lack structured modeling and path interpretability, and are unable to effectively describe the multi-level mapping relationships between raw materials, processes, and performance. It is also difficult to update the model in real time to cope with complex operating conditions such as changes in raw material properties, equipment fluctuations, or adjustments to performance indicators. Furthermore, most methods lack a response modeling mechanism driven by behavioral functions, making it impossible to establish dynamic behavioral expressions and learning paths for raw materials or process parameters within the system.

[0005] In terms of optimization methods, existing intelligent optimization algorithms are mostly based on single-objective or static multi-objective optimization, and fail to consider the key variable of path conflict. There are multiple combination paths between recipes and process parameters, and the tension of different paths in achieving performance goals varies significantly. Existing optimization methods are unable to evaluate or regulate the response conflicts between paths, resulting in the final selected parameter combinations often being unstable, uncoordinated, and difficult to replicate in actual implementation. In addition, most existing swarm intelligence algorithms, such as the standard jellyfish search algorithm, ant colony algorithm, and differential evolution, lack control mechanisms for path structure. Their jump strategies and movement processes are driven by global fitness, lack structural information guidance, and cannot effectively guide individuals to evolve towards path spaces with less tension.

[0006] Therefore, how to provide a feed formula optimization and production line parameter adaptive control method based on machine learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] One objective of the present invention is to propose a machine learning-based method for feed formulation optimization and adaptive control of production line parameters. This method integrates hierarchical polymorphic graph structure modeling with an improved jellyfish search algorithm to establish a tension-driven optimization mechanism for formulation and process parameters, enabling dynamic, coordinated optimization of multiple performance indicators in the feed system. Through behavioral function response modeling and path tension calculation, it effectively resolves conflicts and coordination issues between formulation and process. The introduction of a model self-learning mechanism enables the system to evolve continuously, significantly improving optimization efficiency, system stability, and parameter adaptability. This method is suitable for intelligent formulation design and production line control in complex production environments with multiple batches of raw materials.

[0008] The feed formula optimization and production line parameter adaptive control method based on machine learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collecting feed formula raw material attributes, process parameters, and performance indicators, and constructing a hierarchical polymorphic graph structure model. The hierarchical polymorphic graph structure model includes a formula layer, a process layer, and a performance layer. The three layers are connected by cross-layer edges to form a unified structure, and each node includes an attribute value and a behavior function.

[0010] S2. Based on the behavior functions of the nodes in the hierarchical polymorphic graph structure model, the behavior differences between the nodes are calculated to generate a tension field tensor for characterizing the tension state of the node combination path;

[0011] S3. Screening path combinations whose tension values are lower than a set tension threshold according to the tension field tensor to generate optimized individuals, each of which includes a set of recipe parameters and corresponding process parameters;

[0012] S4, inputting the optimized individuals into an improved jellyfish search algorithm, updating the recipe parameters and process parameters according to the tension gradient direction in the active motion phase, and performing a jump operation in the passive drift phase;

[0013] S5. Calculate the fitness value based on the updated recipe parameters and process parameters of each optimized individual, and select the optimized individual with the highest fitness value as the optimal individual;

[0014] S6. Based on the recipe parameters and process parameters of the optimal individual, updating the node behavior functions of the recipe layer and the process layer in the hierarchical polymorphic graph structure model;

[0015] S7. Determine whether the current raw material batch or production line operating conditions have changed. If so, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, proceed to step S8.

[0016] S8. Output the recipe parameters and process parameters of the current optimal individual, record the tension state and fitness value, and store them in the local database.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Collect the attributes of the feed formula raw materials, including nutrient content, cost information, and source stability indicators, as attribute values of the formula layer nodes. The formula layer nodes include energy raw materials, protein raw materials, and nutritional additive raw materials;

[0019] Collecting feed process parameters, including steam temperature, extrusion pressure, water addition amount, oil injection amount, and cooling temperature, as attribute values of process layer nodes, including steam temperature node, extrusion pressure node, water addition amount node, oil injection amount node, and cooling temperature node;

[0020] Collecting feed performance indicators, including palatability score, forming rate, pellet hardness, and growth performance indicators, as attribute values of performance layer nodes, wherein the performance layer nodes include palatability node, forming rate node, pellet hardness node, and growth performance node;

[0021] S12. Analyze the contribution relationship of each node to the feed performance index based on historical production data, and set the connection relationship and initial edge weight of the cross-layer edge based on the contribution relationship between the recipe layer node and the performance layer node, the process layer node and the performance layer node, and the recipe layer node and the process layer node;

[0022] S13. Define the behavior function of the recipe layer node:

[0023] ;

[0024] in, Indicates the The behavior function of each recipe layer node, Indicates the The addition ratio of each formula raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, Indicates the The response sensitivity of each formula raw material, Indicates the Onset shift of individual formulation ingredients;

[0025] S14. Define the behavior function of the process layer node:

[0026] ;

[0027] in, Indicates the The behavior function of each process layer node, Indicates the process parameter values, represents the adjustment coefficient, represents the exponential function, Represents regulatory factors, Indicates the The optimal theoretical value of the process parameter value;

[0028] S15. Define the behavior function of the performance layer node:

[0029] ;

[0030] in, Indicates the The behavior function of each performance layer node, represents the adjustment coefficient, Indicates the A collection of recipe layer nodes connected to performance layer nodes, Indicates the A collection of process layer nodes connected to performance layer nodes, Indicates the The recipe layer node and the The connection edge weight of each performance layer node, Indicates the The process layer node and the The connection edge weight of each performance layer node;

[0031] S16. Construct a hierarchical polymorphic graph structure model based on the attribute values of nodes, behavior functions and connection relationships between cross-layer edges.

[0032] Optionally, the S2 specifically includes:

[0033] S21, reading the behavior function output values of each recipe layer node and process layer node in the layered polymorphic graph structure model, arranging them in order of node numbers, and forming a behavior response vector;

[0034] S22. For any two nodes in the behavior response vector that are path-connected, calculate the difference between the behavior response values, and perform weighted processing based on the path edge weight to obtain a behavior difference value between the two nodes. A larger behavior difference value indicates a higher degree of response incoordination between the two nodes under the current conditions.

[0035] S23, accumulating the behavior difference values of all adjacent node pairs in the path and dividing the sum by the number of node pairs in the path to obtain an average behavior difference value of the path;

[0036] S24. For each path, the target indicator variation range between the starting node and the end node is calculated. Based on the consistency of the performance variation direction, a positive factor is assigned to correct the average behavior difference value to obtain the conflict value of the path.

[0037] S25. Store the conflict values of all paths in a three-dimensional structure array in the order of path numbers, where the first dimension is the path number, the second dimension is the node level position, and the third dimension is the conflict value. Level positions that do not exist in the path are filled with null values.

[0038] S26. Normalize the three-dimensional structure array and define it as a tension field tensor, which is used to characterize the tension state of the node combination path.

[0039] Optionally, the S3 specifically includes:

[0040] S31, traverse all path dimensions in the tension field tensor and read the conflict value corresponding to each path number;

[0041] S32, normalizing the maximum and minimum values of the conflict values of all paths, and constructing a standardized tension value sequence so that the tension values of all paths are within a unified range;

[0042] S33, setting a tension threshold, screening the standardized tension value sequence, retaining path combinations with tension values lower than the tension threshold, and constructing a candidate path set;

[0043] S34. Extract the recipe parameters and corresponding process parameters of each candidate path in the candidate path set to form an optimized individual.

[0044] Optionally, the improvements of the improved jellyfish search algorithm specifically include:

[0045] In the active movement phase, the tension value of the current candidate path of the optimization individual is obtained, and the tension difference with the upper and lower adjacent paths in the tension field tensor is calculated. If the tension value of the adjacent path is 5% lower than the tension value of the current path, the movement in the tension decreasing direction is triggered. The movement amplitude is no more than 2% of the recipe parameter adjustment ratio, and the process parameter adjustment step is no more than 3% of the corresponding parameter interval;

[0046] In the passive drift phase, the candidate paths with tension values in the top 20% are selected as the drift target set, and the jump target path is selected from the drift target set using the roulette wheel selection algorithm to perform the jump operation;

[0047] Reset the recipe parameters and process parameters of the current optimized individual according to the recipe parameters and process parameters bound to the jump target path.

[0048] In the improved jellyfish search algorithm employed in this invention, a tension difference threshold of 5% is set based on a sensitivity analysis of the impact of path tension changes on performance in actual feed formulation systems. A tension drop exceeding 5% typically indicates a substantial improvement in path behavior consistency, thus triggering individual movement. The movement amplitude is limited to a maximum of 2% for recipe parameter adjustments and a maximum of 3% for process parameter adjustments within the corresponding parameter interval. This ensures convergence stability during local search and prevents large-scale perturbations from causing behavioral function jumps or recipe imbalances. Path drift targets the top 20% of paths by tension, balancing search efficiency and transition quality. This ensures that individuals avoid being trapped in suboptimal paths and significantly improves overall tension coordination. Roulette wheel selection is used to execute path transitions within this set, facilitating the establishment of a gradient-driven global exploration mechanism in a low-tension space. These parameter designs effectively enhance the stability, adaptability, and convergence quality of the jellyfish search algorithm in high-dimensional recipe and process spaces, demonstrating clear technical benefits.

[0049] Optionally, the S5 specifically includes:

[0050] Based on the recipe parameters and process parameters of each optimized individual, the predicted output value of the performance layer node is calculated in the hierarchical polymorphic graph structure model. The predicted output value includes four indicators: palatability, forming rate, particle hardness, and growth performance. The four indicators are weighted according to preset weights to obtain a performance score;

[0051] Read the tension value of the path corresponding to the optimized individual in the tension field tensor, and convert the tension value into a tension score in a normalized manner;

[0052] The performance score and the tension score are weighted and combined according to a preset ratio to obtain the fitness value of the optimized individual;

[0053] Compare the fitness values among all optimized individuals and select the optimized individual with the highest fitness value as the optimal individual in the current round.

[0054] Optionally, the S6 specifically includes:

[0055] S61. Calculate the predicted output value of the performance layer node in the hierarchical polymorphic graph structure model based on the recipe parameters and process parameters in the optimal individual, compare it with each actual observed performance value, and generate a performance error vector;

[0056] S62. According to the performance error vector, trace the behavior function contribution path of the corresponding recipe layer node and process layer node, and determine the response impact of each node on the error;

[0057] S63, adjusting the behavior function based on the degree of influence of the node's response to the error, using a parameter adjustment strategy consistent with the error direction, and increasing or decreasing the adjustment coefficient in the behavior function within a set update rate range;

[0058] S64. Write the updated behavior function into the hierarchical polymorphic graph structure model to replace the behavior function of the original node, thereby completing the iterative evolution of the hierarchical polymorphic graph structure model.

[0059] Optionally, the parameter adjustment strategy for consistent error directions specifically includes:

[0060] When a performance indicator in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is increased to enhance the behavioral response strength of the node;

[0061] When a performance indicator in the performance error vector is positive, indicating that the predicted output value is higher than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is reduced to reduce the behavioral response intensity of the node;

[0062] The adjustment coefficient adjustment operation is limited by a set update rate range in each round of iteration, so as to prevent the behavior function from changing too much and causing the layered polymorphic graph structure model to become unstable.

[0063] The beneficial effects of the present invention are:

[0064] First, by constructing a hierarchical polymorphic graph model, this paper achieves structured modeling of the three core elements of a feed system: formula ingredients, process parameters, and performance indicators. Cross-layer edge connections between the formula layer, process layer, and performance layer not only clarify the dependencies between these elements but also provide a foundation for path-level dynamic behavior modeling. This structured representation overcomes the limitations of traditional methods that separate formula and process, and can accurately reflect the coupled impact of multiple factors on performance results.

[0065] Secondly, the introduction of behavior functions enables each node to no longer be a static variable, but rather have dynamic response capabilities. The behavior functions of recipe- and process-level nodes simulate their specific contributions to target performance under varying parameter conditions, while the behavior functions of the performance level measure overall performance by integrating upstream behavior outputs. This mechanism not only enhances the system's expressive power but also provides clear parameter entry for subsequent model learning and behavior adjustments.

[0066] Furthermore, this invention introduces the concept of a tension field tensor for the first time, quantifying the degree of conflict between different paths during behavioral responses. By calculating the behavioral differences between adjacent nodes in a path, a path tension metric is constructed and used to optimize individual selection and search direction guidance, transcending the limitations of traditional optimization methods that rely solely on fitness functions. The introduction of a tension mechanism stabilizes the optimization process, reduces invalid searches and local convergence issues, and enhances the physical interpretability and control precision of path selection.

[0067] Furthermore, the optimization process employs an improved jellyfish search algorithm. This algorithm not only incorporates tension gradients as a directional control during the active motion phase but also employs a tension distribution structure to construct a jump target space during the passive drift phase, thereby achieving a structure-guided global search capability. Furthermore, to avoid being trapped in local optima, a local perturbation mechanism and a method for resetting parameters after path jumps were designed, ensuring both diversity and adaptability in the search.

[0068] Finally, the present invention also establishes a dynamic model update mechanism. During each round of optimization, the optimal individual's recipe and process parameters are fed back into the hierarchical polymorphic graph structure model to adjust the behavior functions of relevant nodes. This enables the model to continuously learn and evolve, adapting to complex scenarios such as different raw material batches, process fluctuations, and changes in performance targets, truly achieving adaptive collaborative optimization of recipes and processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0069] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0070] Figure 1 This is the overall flow chart of the feed formula optimization and production line parameter adaptive control method based on machine learning proposed in the present invention;

[0071] Figure 2 This is an operational flow chart for constructing an optimization individual for the feed formula optimization and production line parameter adaptive control method based on machine learning proposed in the present invention. DETAILED DESCRIPTION

[0072] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0073] refer to Figure 1 and Figure 2 The feed formula optimization and production line parameter adaptive control method based on machine learning includes the following steps:

[0074] S1. Collecting feed formula raw material attributes, process parameters, and performance indicators, and constructing a hierarchical polymorphic graph structure model. The hierarchical polymorphic graph structure model includes a formula layer, a process layer, and a performance layer. The three layers are connected by cross-layer edges to form a unified structure, and each node includes an attribute value and a behavior function.

[0075] S2. Based on the behavior functions of the nodes in the hierarchical polymorphic graph structure model, the behavior differences between the nodes are calculated to generate a tension field tensor for characterizing the tension state of the node combination path;

[0076] S3. Screening path combinations whose tension values are lower than a set tension threshold according to the tension field tensor to generate optimized individuals, each of which includes a set of recipe parameters and corresponding process parameters;

[0077] S4, inputting the optimized individuals into an improved jellyfish search algorithm, updating the recipe parameters and process parameters according to the tension gradient direction in the active motion phase, and performing a jump operation in the passive drift phase;

[0078] S5. Calculate the fitness value based on the updated recipe parameters and process parameters of each optimized individual, and select the optimized individual with the highest fitness value as the optimal individual;

[0079] S6. Based on the recipe parameters and process parameters of the optimal individual, updating the node behavior functions of the recipe layer and the process layer in the hierarchical polymorphic graph structure model;

[0080] S7. Determine whether the current raw material batch or production line operating conditions have changed. If so, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, proceed to step S8.

[0081] S8. Output the recipe parameters and process parameters of the current optimal individual, record the tension state and fitness value, and store them in the local database.

[0082] This method realizes the structured modeling of the complex relationship between feed raw material properties, processing parameters and performance output by constructing a hierarchical polymorphic graph structure model consisting of a formula layer, a process layer and a performance layer. On this basis, the behavior function is used as the core attribute of the node to dynamically express the response process of each parameter to the target performance. The tension field tensor is constructed by driving the behavioral differences to further realize the path-level conflict evaluation and quantification, forming the response basis for the whole process optimization. The entire optimization system is based on model-driven logic rather than empirical parameter adjustment, so that the combination decision of formula and process has data dependence, behavioral feedback and structural evolution capabilities. Compared with traditional rule engines or single prediction models, this method not only significantly improves the clarity and adaptability of multi-parameter linkage modeling, but also has the self-learning ability of raw material fluctuations and changes in the production environment, thereby effectively improving the accuracy, consistency and engineering promotion value of feed production.

[0083] In this embodiment, S1 specifically includes:

[0084] S11. Collect the attributes of the feed formula raw materials, including nutrient content, cost information, and source stability indicators, as attribute values of the formula layer nodes. The formula layer nodes include energy raw materials, protein raw materials, and nutritional additive raw materials;

[0085] Collecting feed process parameters, including steam temperature, extrusion pressure, water addition amount, oil injection amount, and cooling temperature, as attribute values of process layer nodes, including steam temperature node, extrusion pressure node, water addition amount node, oil injection amount node, and cooling temperature node;

[0086] Collecting feed performance indicators, including palatability score, forming rate, pellet hardness, and growth performance indicators, as attribute values of performance layer nodes, wherein the performance layer nodes include palatability node, forming rate node, pellet hardness node, and growth performance node;

[0087] S12. Analyze the contribution relationship of each node to the feed performance index based on historical production data, and set the connection relationship and initial edge weight of the cross-layer edge based on the contribution relationship between the recipe layer node and the performance layer node, the process layer node and the performance layer node, and the recipe layer node and the process layer node;

[0088] S13. Define the behavior function of the recipe layer node:

[0089] ;

[0090] in, Indicates the The behavior function of each recipe layer node, Indicates the The addition ratio of each formula raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, Indicates the The response sensitivity of each formula raw material, Indicates the Onset shift of individual formulation ingredients;

[0091] S14. Define the behavior function of the process layer node:

[0092] ;

[0093] in, Indicates the The behavior function of each process layer node, Indicates the process parameter values, represents the adjustment coefficient, represents the exponential function, Represents regulatory factors, Indicates the The optimal theoretical value of the process parameter value;

[0094] S15. Define the behavior function of the performance layer node:

[0095] ;

[0096] in, Indicates the The behavior function of each performance layer node, represents the adjustment coefficient, Indicates A collection of recipe layer nodes connected to performance layer nodes, Indicates A collection of process layer nodes connected to performance layer nodes, Indicates the The recipe layer node and the The connection edge weight of each performance layer node, Indicates the The process layer node and the The connection edge weight of each performance layer node;

[0097] S16. Construct a hierarchical polymorphic graph structure model based on the attribute values of nodes, behavior functions and connection relationships between cross-layer edges.

[0098] This step introduces a structured collection mechanism for the three information sources of formula, process and performance, and classifies each element into the corresponding node layer, providing a traceable data foundation for the hierarchical modeling of the graph model. By distinguishing the raw material type, process node and performance dimension, the classification coding and standardized expression of the model input are realized, enhancing the generalization and scalability of the system. The setting of cross-layer edge weights combined with historical data accurately maps the actual contribution path and control sensitive points of raw materials to performance, so that the graph structure is no longer a static connection, but has behavioral meaning and quantitative basis. On this basis, it further provides contextual support for the definition of behavioral functions, and provides a stable source of structural input for subsequent response calculations, tension assessment and optimization mechanisms. The overall design constructs a logical closed loop between raw data and structural graph models, laying a solid foundation for system modeling.

[0099] In this embodiment, S2 specifically includes:

[0100] S21, reading the behavior function output values of each recipe layer node and process layer node in the layered polymorphic graph structure model, arranging them in order of node numbers, and forming a behavior response vector;

[0101] S22. For any two nodes in the behavior response vector that are path-connected, calculate the difference between the behavior response values, and perform weighted processing based on the path edge weight to obtain a behavior difference value between the two nodes. A larger behavior difference value indicates a higher degree of response incoordination between the two nodes under the current conditions.

[0102] S23, accumulating the behavior difference values of all adjacent node pairs in the path and dividing the sum by the number of node pairs in the path to obtain an average behavior difference value of the path;

[0103] S24. For each path, the target indicator variation range between the starting node and the end node is calculated. Based on the consistency of the performance variation direction, a positive factor is assigned to correct the average behavior difference value to obtain the conflict value of the path.

[0104] S25. Store the conflict values of all paths in a three-dimensional structure array in the order of path numbers, where the first dimension is the path number, the second dimension is the node level position, and the third dimension is the conflict value. Level positions that do not exist in the path are filled with null values.

[0105] S26. Normalize the three-dimensional structure array and define it as a tension field tensor, which is used to characterize the tension state of the node combination path.

[0106] This step establishes a path conflict evaluation system driven by behavioral responses by calculating the differences between the outputs of node behavioral functions. Unlike traditional performance difference or objective function methods, this method starts from the consistency of nodes within the path, quantifies the degree of behavioral coordination in each path, forms a path conflict value, and then constructs a tension field tensor. The normalization and hierarchical distribution representation of tension values give the model tension visibility and path structure control capabilities. With this mechanism, the system can not only dynamically screen low-tension path combinations, but also identify high-tension bottleneck nodes, thereby assisting in the subsequent tuning of behavioral functions. The overall mechanism strengthens the path structure's ability to guide the optimization direction, making the optimization behavior structurally aware, directional, and interpretable, significantly improving the convergence efficiency of the global optimal solution and the stability of the system response.

[0107] In this embodiment, S3 specifically includes:

[0108] S31, traverse all path dimensions in the tension field tensor and read the conflict value corresponding to each path number;

[0109] S32, normalizing the maximum and minimum values of the conflict values of all paths, and constructing a standardized tension value sequence so that the tension values of all paths are within a unified range;

[0110] S33, setting a tension threshold, screening the standardized tension value sequence, retaining path combinations with tension values lower than the tension threshold, and constructing a candidate path set;

[0111] S34. Extract the recipe parameters and corresponding process parameters of each candidate path in the candidate path set to form an optimized individual.

[0112] Based on the construction of the tension field tensor, this step describes in detail the process of screening the path combination with lower tension and generating optimized individuals. By normalizing the tension value and setting a threshold mechanism, structural screening of the entire path set is achieved, effectively controlling the scale of the search space, and ensuring that the paths entering the optimization stage have high tension coordination. Furthermore, each selected path is constructed into an optimized individual, and the structure is clearly a two-component set consisting of recipe parameters and process parameters, which improves the individual operability and the efficiency of objective function evaluation. This optimization individual generation logic provides a stable quality and clear structure initial population for the subsequent optimization algorithm, effectively avoiding the low-quality disturbance caused by random initialization, improving the overall optimization starting level, shortening the convergence cycle, and improving search efficiency and reliability.

[0113] In this embodiment, the improvements of the improved jellyfish search algorithm specifically include:

[0114] In the active movement phase, the tension value of the current candidate path of the optimization individual is obtained, and the tension difference with the upper and lower adjacent paths in the tension field tensor is calculated. If the tension value of the adjacent path is 5% lower than the tension value of the current path, the movement in the tension decreasing direction is triggered. The movement amplitude is no more than 2% of the recipe parameter adjustment ratio, and the process parameter adjustment step is no more than 3% of the corresponding parameter interval;

[0115] In the passive drift phase, the candidate paths with tension values in the top 20% are selected as the drift target set, and the jump target path is selected from the drift target set using the roulette wheel selection algorithm to perform the jump operation;

[0116] Reset the recipe parameters and process parameters of the current optimized individual according to the recipe parameters and process parameters bound to the jump target path.

[0117] By introducing an improved jellyfish search algorithm driven by tension gradients, this step solves the problems of unclear direction and inaccurate jumps in traditional optimization algorithms in high-dimensional path searches. During the active motion phase, the algorithm determines direction based on the path tension difference, triggering a jump only when the tension of adjacent paths decreases by more than a set ratio. It also implements boundary control over the adjustment range of recipes and process parameters, making the individual evolution process more stable and accurate. During the passive drift phase, a tension sorting strategy is used to limit the target path set, and a roulette wheel selection mechanism is used to achieve biased distribution jumps, enhancing the ability to explore low-tension areas. The entire optimization process is dominated by the tension value structure rather than the global fitness function, making the search more structured and directional, significantly improving global convergence ability and path structure matching.

[0118] In this embodiment, the S5 specifically includes:

[0119] Based on the recipe parameters and process parameters of each optimized individual, the predicted output value of the performance layer node is calculated in the hierarchical polymorphic graph structure model. The predicted output value includes four indicators: palatability, forming rate, particle hardness, and growth performance. The four indicators are weighted according to preset weights to obtain a performance score;

[0120] Read the tension value of the path corresponding to the optimized individual in the tension field tensor, and convert the tension value into a tension score in a normalized manner;

[0121] The performance score and the tension score are weighted and combined according to a preset ratio to obtain the fitness value of the optimized individual;

[0122] Compare the fitness values among all optimized individuals and select the optimized individual with the highest fitness value as the optimal individual in the current round.

[0123] After the optimization is completed, this step clarifies the specific calculation method for optimizing individual fitness, and incorporates the performance output value and the tension value into the fitness evaluation system. The performance part is standardized and weighted by combining four types of indicators to form a performance score, and the tension part is normalized to generate a tension score. The two are proportionally synthesized into a fitness value, taking into account both path coordination and target performance quality. This dual-dimensional fusion fitness mechanism avoids the drawback of ignoring path conflicts under traditional single-goal drive, so that the selected optimal solution is not only optimal in performance, but also has high tension coordination and feasibility of actual implementation. The optimal individual data structure finally selected is stable and logically clear, providing reliable input for behavior function updates and system feedback.

[0124] In this embodiment, S6 specifically includes:

[0125] S61. Calculate the predicted output value of the performance layer node in the hierarchical polymorphic graph structure model based on the recipe parameters and process parameters in the optimal individual, compare it with each actual observed performance value, and generate a performance error vector;

[0126] S62. According to the performance error vector, trace the behavior function contribution path of the corresponding recipe layer node and process layer node, and determine the response impact of each node on the error;

[0127] S63, adjusting the behavior function based on the degree of influence of the node's response to the error, using a parameter adjustment strategy consistent with the error direction, and increasing or decreasing the adjustment coefficient in the behavior function within a set update rate range;

[0128] S64. Write the updated behavior function into the hierarchical polymorphic graph structure model to replace the behavior function of the original node, thereby completing the iterative evolution of the hierarchical polymorphic graph structure model.

[0129] This step constructs a feedback loop from the optimal individual to the model behavior function, realizing the iterative self-evolution of the hierarchical polymorphic graph structure model. By inputting the recipe parameters and process parameters of the optimal individual into the model and comparing them with the target performance, an error vector is obtained, which is then traced back to the recipe and process nodes in the path. According to the degree of influence of the node's response to the error, a hierarchical update is performed, and a direction-consistent strategy is used to adjust the adjustment coefficient in the behavior function, and an update rate constraint is imposed to avoid model instability. This feedback mechanism enables the graph structure model to have online learning and dynamic adaptability, effectively responding to production disturbances such as changes in raw material batches and adjustments to performance targets, and significantly enhancing the long-term applicability and adaptive control capabilities of the model.

[0130] In this embodiment, the parameter adjustment strategy with consistent error direction specifically includes:

[0131] When a performance indicator in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is increased to enhance the behavioral response strength of the node;

[0132] When a performance indicator in the performance error vector is positive, indicating that the predicted output value is higher than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is reduced to reduce the behavioral response intensity of the node;

[0133] The adjustment coefficient adjustment operation is limited by a set update rate range in each round of iteration, so as to prevent the behavior function from changing too much and causing the layered polymorphic graph structure model to become unstable.

[0134] The parameter adjustment strategy with consistent error direction clarifies the strategic logic for updating behavior function parameters. It adjusts the adjustment coefficients of response nodes based on the direction of performance error, thereby increasing or decreasing the output response strength of the nodes. When performance predictions are insufficient, overall performance is improved by enhancing the response of key nodes; when predictions exceed the target, the response is appropriately weakened to avoid overfitting or resource waste. This strategy introduces update rate constraints to ensure smooth and controllable parameter changes, effectively preventing model oscillation or drift during the update process.

[0135] Example 1

[0136] To verify the feasibility of this invention, we applied it to the pig feed production line of a large feed processing company for actual deployment and testing. This company has long faced two typical problems: First, raw material ratio adjustments often rely on the experience of nutritionists, resulting in slow response, trial-and-error, and difficulty adapting to changes in raw material batches; second, production line process parameters are mostly set manually, resulting in low stability and large fluctuations in finished product performance, which seriously restricts feed product quality consistency and production efficiency.

[0137] In this embodiment, the system first deploys a sensor collection module to collect the nutritional content, source batches, and cost attributes of six types of raw materials entering the production line, including corn, soybean meal, fish meal, wheat, calcium hydrogen phosphate, and additives. At the same time, it monitors key process parameters in real time during production line operation, including steam temperature, extrusion pressure, water addition, oil injection volume, and cooling temperature. In addition, it also integrates terminal quality inspection equipment to transmit back data on palatability scores, molding rate, pellet hardness, and animal growth performance.

[0138] Using the method of the present invention, a hierarchical polymorphic graph structure model based on a three-layer structure of raw materials, processes, and performance is first established. Behavior functions for each node are constructed, and initial response relationships are fitted using historical data. Subsequently, a tension field tensor is constructed by calculating the differences between the behavior functions to quantify the degree of coordination between different path combinations. Paths with low tension are screened based on their tension values, and the corresponding recipe and process parameters are extracted. Optimized individuals are generated and input into an improved jellyfish search algorithm for global search. During the optimization process, active adjustment guided by tension gradients and a path drift mechanism under structural constraints are introduced to achieve efficient search for the optimal recipe path. Finally, the individual with the highest fitness is selected and used to reversely update the behavior function, achieving closed-loop model evolution.

[0139] After seven consecutive days of testing, the company's system output optimization results showed high stability and significant improvement. Taking the optimized parameters as an example, the system recommended a formula of 35% corn, 25% soybean meal, 10% fish meal, 20% wheat, 5% calcium hydrogen phosphate, and 5% additives. The corresponding process parameters were a steam temperature of 90°C, an extrusion pressure of 3.5MPa, a water addition of 20%, an oil injection of 4%, and a cooling temperature of 25°C. Under this parameter combination, the actual finished feed had a palatability score of 8.7 points, a molding rate of up to 96.2%, a pellet hardness maintained at 22.5N, and an average 11.4% improvement in pig growth performance. The overall fitness value was 0.925, a significant improvement compared to the 0.781 before optimization.

[0140] In actual operation, the system dynamically adjusted recipes and fine-tuned process parameters on an hourly basis, completing an average of one self-learning update cycle every 24 hours. It also achieved stable responses to two raw material batch changes and one equipment maintenance process, without experiencing performance drops or abnormal fluctuations. The results demonstrate that the system exhibits excellent adaptability, optimization efficiency, and system stability in industrial environments characterized by complex raw material variation, variable performance targets, and highly coupled production line parameters, demonstrating significant engineering value and potential for widespread adoption.

[0141] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for feed formula optimization and production line parameter adaptive control based on machine learning, characterized in that: The steps include: S1. Collecting feed formula raw material attributes, process parameters, and performance indicators, and constructing a hierarchical polymorphic graph structure model. The hierarchical polymorphic graph structure model includes a formula layer, a process layer, and a performance layer. The three layers are connected by cross-layer edges to form a unified structure, and each node includes an attribute value and a behavior function. S2. Based on the behavior functions of the nodes in the hierarchical polymorphic graph structure model, the behavior differences between the nodes are calculated to generate a tension field tensor for characterizing the tension state of the node combination path; S3. Screening path combinations whose tension values are lower than a set tension threshold according to the tension field tensor to generate optimized individuals, each of which includes a set of recipe parameters and corresponding process parameters; S4, inputting the optimized individuals into an improved jellyfish search algorithm, updating the recipe parameters and process parameters according to the tension gradient direction in the active motion phase, and performing a jump operation in the passive drift phase; S5. Calculate the fitness value based on the updated recipe parameters and process parameters of each optimized individual, and select the optimized individual with the highest fitness value as the optimal individual; S6. Based on the recipe parameters and process parameters of the optimal individual, updating the node behavior functions of the recipe layer and the process layer in the hierarchical polymorphic graph structure model; S7. Determine whether the current raw material batch or production line operating conditions have changed. If so, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, proceed to step S8. S8. Output the recipe parameters and process parameters of the current optimal individual, record the tension state and fitness value, and store them in the local database; The S3 specifically includes: S31, traverse all path dimensions in the tension field tensor and read the conflict value corresponding to each path number; S32, normalizing the maximum and minimum values of the conflict values of all paths, and constructing a standardized tension value sequence so that the tension values of all paths are within a unified range; S33, setting a tension threshold, screening the standardized tension value sequence, retaining path combinations with tension values lower than the tension threshold, and constructing a candidate path set; S34. Extract the recipe parameters and corresponding process parameters of each candidate path in the candidate path set to form an optimized individual.

2. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 1, characterized in that: Said S1 specifically includes: S11. Collect the attributes of the feed formula raw materials, including nutrient content, cost information, and source stability indicators, as attribute values of the formula layer nodes. The formula layer nodes include energy raw materials, protein raw materials, and nutritional additive raw materials; Collecting feed process parameters, including steam temperature, extrusion pressure, water addition amount, oil injection amount, and cooling temperature, as attribute values of process layer nodes, including steam temperature node, extrusion pressure node, water addition amount node, oil injection amount node, and cooling temperature node; Collecting feed performance indicators, including palatability score, forming rate, pellet hardness, and growth performance indicators, as attribute values of performance layer nodes, wherein the performance layer nodes include palatability node, forming rate node, pellet hardness node, and growth performance node; S12. Analyze the contribution relationship of each node to the feed performance index based on historical production data, and set the connection relationship and initial edge weight of the cross-layer edge based on the contribution relationship between the recipe layer node and the performance layer node, the process layer node and the performance layer node, and the recipe layer node and the process layer node; S13. Define the behavior function of the recipe layer node: ； in, Indicates the The behavior function of each recipe layer node, Indicates the The addition ratio of each formula raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, Indicates the The response sensitivity of each formula raw material, Indicates the Onset shift of individual formulation ingredients; S14. Define the behavior function of the process layer node: ； in, Indicates the The behavior function of each process layer node, Indicates the process parameter values, represents the adjustment coefficient, represents the exponential function, Represents regulatory factors, Indicates the The optimal theoretical value of the process parameter value; S15. Define the behavior function of the performance layer node: ； in, Indicates the The behavior function of each performance layer node, represents the adjustment coefficient, Indicates A collection of recipe layer nodes connected to performance layer nodes, Indicates the A collection of process layer nodes connected to performance layer nodes, Indicates the The recipe layer node and the The connection edge weight of each performance layer node, Indicates the The process layer node and the The connection edge weight of each performance layer node; S16. Construct a hierarchical polymorphic graph structure model based on the attribute values of nodes, behavior functions and connection relationships between cross-layer edges.

3. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 1 is characterized in that: The S2 specifically includes: S21, reading the behavior function output values of each recipe layer node and process layer node in the layered polymorphic graph structure model, arranging them in order of node numbers, and forming a behavior response vector; S22. For any two nodes in the behavior response vector that are path-connected, calculate the difference between the behavior response values, and perform weighted processing based on the path edge weight to obtain a behavior difference value between the two nodes. A larger behavior difference value indicates a higher degree of response incoordination between the two nodes under the current conditions. S23, accumulating the behavior difference values of all adjacent node pairs in the path and dividing the sum by the number of node pairs in the path to obtain an average behavior difference value of the path; S24. For each path, the target indicator variation range between the starting node and the end node is calculated. Based on the consistency of the performance variation direction, a positive factor is assigned to correct the average behavior difference value to obtain the conflict value of the path. S25. Store the conflict values of all paths in a three-dimensional structure array in the order of path numbers, where the first dimension is the path number, the second dimension is the node level position, and the third dimension is the conflict value. Level positions that do not exist in the path are filled with null values. S26. Normalize the three-dimensional structure array and define it as a tension field tensor, which is used to characterize the tension state of the node combination path.

4. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 1, characterized in that: The improvements of the improved jellyfish search algorithm specifically include: In the active movement phase, the tension value of the current candidate path of the optimization individual is obtained, and the tension difference with the upper and lower adjacent paths in the tension field tensor is calculated. If the tension value of the adjacent path is 5% lower than the tension value of the current path, the movement in the tension decreasing direction is triggered. The movement amplitude is no more than 2% of the recipe parameter adjustment ratio, and the process parameter adjustment step is no more than 3% of the corresponding parameter interval; In the passive drift phase, the candidate paths with tension values in the top 20% are selected as the drift target set, and the jump target path is selected from the drift target set using the roulette wheel selection algorithm to perform the jump operation; Reset the recipe parameters and process parameters of the current optimized individual according to the recipe parameters and process parameters bound to the jump target path.

5. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 1, characterized in that: The S5 specifically includes: Based on the recipe parameters and process parameters of each optimized individual, the predicted output value of the performance layer node is calculated in the hierarchical polymorphic graph structure model. The predicted output value includes four indicators: palatability, forming rate, particle hardness, and growth performance. The four indicators are weighted according to preset weights to obtain a performance score; Read the tension value of the path corresponding to the optimized individual in the tension field tensor, and convert the tension value into a tension score in a normalized manner; The performance score and the tension score are weighted and combined according to a preset ratio to obtain the fitness value of the optimized individual; Compare the fitness values among all optimized individuals and select the optimized individual with the highest fitness value as the optimal individual in the current round.

6. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 1, characterized in that: The S6 specifically includes: S61. Calculate the predicted output value of the performance layer node in the hierarchical polymorphic graph structure model based on the recipe parameters and process parameters in the optimal individual, compare it with each actual observed performance value, and generate a performance error vector; S62. According to the performance error vector, trace the behavior function contribution path of the corresponding recipe layer node and process layer node, and determine the response impact of each node on the error; S63, adjusting the behavior function based on the degree of influence of the node's response to the error, using a parameter adjustment strategy consistent with the error direction, and increasing or decreasing the adjustment coefficient in the behavior function within a set update rate range; S64. Write the updated behavior function into the hierarchical polymorphic graph structure model to replace the behavior function of the original node, thereby completing the iterative evolution of the hierarchical polymorphic graph structure model.

7. The feed formula optimization and production line parameter adaptive control method based on machine learning according to claim 6, characterized in that: The parameter adjustment strategy with consistent error direction specifically includes: When a performance indicator in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is increased to enhance the behavioral response strength of the node; When a performance indicator in the performance error vector is positive, indicating that the predicted output value is higher than the actual observed performance value, the adjustment coefficient in the behavior function of the corresponding recipe layer node and process layer node is reduced to reduce the behavioral response intensity of the node; The adjustment coefficient adjustment operation is limited by a set update rate range in each round of iteration, so as to prevent the behavior function from changing too much and causing the layered polymorphic graph structure model to become unstable.

Images