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

Abstract

The invention discloses a feed formula optimization and production line parameter adaptive control method based on machine learning. The method comprises the following steps: S1, constructing a hierarchical polymorphic graph structure model; s2, calculating a behavior difference between nodes to generate a tension field tensor; s3, screening path combinations of which the tension values are lower than a set tension threshold to generate optimized individuals; s4, inputting the optimized individuals into an improved jellyfish search algorithm, wherein the improved jellyfish search algorithm comprises an active motion stage and a passive drift stage; s5, selecting the optimized individual with the highest fitness value as an optimal individual; s6, updating the hierarchical polymorphic graph structure model based on the formula parameter and the process parameter of the optimal individual; s7, judging whether the current raw material batch or the working condition of the production line is changed or not, if so, returning to the step S2 and repeatedly executing the steps S3 to S6, otherwise, entering the step S8; and S8, outputting formula parameters and process parameters in the current optimal individual. According to the method, a layer polymorphic graph structure model and an improved jellyfish algorithm are fused, and collaborative optimization of a feed formula and process parameters is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of feed production process optimization, and particularly to a method for optimizing feed formulations and adaptively controlling production line parameters based on machine learning. Background Art

[0002] In the modern feed industry, formulation optimization and production line process control are key links to achieve high-quality and high-efficiency production. Feed formulations involve combinations of various raw materials, and there are significant differences in nutritional value, palatability, stability, and cost among different raw materials. At the same time, process parameters in feed production, such as steam temperature, extrusion pressure, water addition, oil spraying amount, etc., also directly affect the pellet hardness, forming rate, and animal growth performance of the final product. Traditional feed design relies on the experience of nutrition experts and repeated adjustments of the formulation through experimental verification. Process control also often uses static set parameters, lacking refined modeling and dynamic optimization capabilities, resulting in the formulation plan being unable to 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 development and the stability of the production process, computer-aided formulation design systems and production process control systems have been gradually introduced in the industry in recent years. Some enterprises adopt formulation recommendation methods based on linear programming or rule engines, and introduce expert systems to set rules and trigger alarms for process parameters. Such systems have improved the degree of operation automation to a certain extent, but there is still a lack of a collaborative modeling mechanism between the formulation layer and the process layer. The mutual coupling effect between the formulation and the process cannot be modeled, resulting in the inability to precisely adjust the process response when the formulation changes; conversely, it is also difficult to quickly match a new formulation plan 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 formulations and processes due to simple model structures, insufficient search efficiency, or weak path control capabilities, and it is difficult to quickly find a formulation combination that meets multiple performance objectives and has coordinated tensions in a complex high-dimensional parameter space.

[0004] The existing applications of machine learning in feed optimization mostly focus on classifying or regressing and predicting historical formulations and production results, and have not yet formed a complete dynamic adaptive optimization mechanism. Although existing research has tried to use neural networks for performance prediction, it mainly fits existing data as a black box model, lacking structured modeling and path interpretability capabilities, unable to effectively describe the multi-level mapping relationship between raw materials - processes - performance, and it is also difficult to update the model in real time to cope with complex working conditions such as changes in raw material properties, equipment fluctuations, or performance index adjustments. Moreover, most methods lack a response modeling mechanism driven by behavioral functions and are unable to establish dynamic behavioral expressions and learning paths for raw materials or process parameters in the system.

[0005] In terms of optimization methods, existing intelligent optimization algorithms mainly focus on single-objective or static multi-objective optimization, failing to consider the key variable of path conflict. There are multiple combination paths between formula and process parameters, and the tension differences in achieving performance goals among different paths are significant. Existing optimization methods cannot evaluate or regulate the response conflicts between paths, resulting in problems such as instability, incoordination, and difficulty in replication when the finally selected parameter combinations are actually implemented. In addition, most existing swarm intelligence algorithms such as the standard jellyfish search algorithm, ant colony algorithm, differential evolution, etc. lack a control mechanism for path structure. Their jump strategies and movement processes are driven by global fitness, lacking structural information guidance, and cannot effectively guide individuals to evolve towards the path space with less tension.

[0006] Therefore, how to provide a machine learning-based feed formula optimization and production line parameter adaptive control method is an urgent problem to be solved by those skilled in the art. Summary of the Invention

[0007] An object of the present invention is to propose a machine learning-based feed formula optimization and production line parameter adaptive control method. The present invention integrates a hierarchical polymorphic graph structure modeling and an improved jellyfish search algorithm, establishes a tension-driven optimization mechanism for formula parameters and process parameters, and realizes the dynamic collaborative optimization of multi-objective performance indicators in the feed system. Through behavior function response modeling and path tension calculation, the conflict coordination problem between formula and process is effectively solved, and a model self-learning mechanism is introduced to enable the system to have the ability of continuous evolution, significantly improving the optimization efficiency, system stability, and parameter adaptability, and being applicable to intelligent formula design and production line control under multi-batch raw materials and complex production environments.

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

[0009] S1. Collect the formula raw material attributes, process parameters, and performance indicators of the feed, and construct 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, calculate the behavior differences between the nodes, and generate a tension field tensor to characterize the tension state of the node combination path; S3. According to the tension field tensor, screen the path combinations with tension values lower than the set tension threshold to generate optimized individuals, and each optimized individual includes a set of formula parameters and corresponding process parameters; S4. Input the optimized individuals into the improved jellyfish search algorithm, and update the formula parameters and process parameters according to the tension gradient direction in the active movement stage, and perform jump operations in the passive drift stage; S5. Calculate the fitness value according to the updated formulation 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 formulation parameters and process parameters of the optimal individual, update the node behavior functions in the formulation layer and the process layer of the hierarchical polymorphic graph structure model; S7. Determine whether the current raw material batch or production line working condition has changed. If it has changed, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, enter step S8; S8. Output the formulation parameters and process parameters in the current optimal individual, record the tension state and fitness value, and store them in the local database.

[0010] Optionally, the specific steps of S1 include: S11. Collect the formulation raw material attributes of the feed, including nutrient content, cost information, and source stability index, as the attribute values of the nodes in the formulation layer. The nodes in the formulation layer include energy raw materials, protein raw materials, and nutritional additive raw materials; Collect the process parameters of the feed, including steam temperature, extrusion pressure, water addition amount, oil spraying amount, and cooling temperature, as the attribute values of the nodes in the process layer. The nodes in the process layer include steam temperature node, extrusion pressure node, water addition amount node, oil spraying amount node, and cooling temperature node; Collect the performance indicators of the feed, including palatability score, forming rate, pellet hardness, and growth performance index, as the attribute values of the nodes in the performance layer. The nodes in the performance layer 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 indicators based on historical production data, and set the connection relationship and initial edge weight of the cross-layer edges based on the contribution relationships between the nodes in the formulation layer and the performance layer, between the nodes in the process layer and the performance layer, and between the nodes in the formulation layer and the process layer; S13. Define the behavior function of the nodes in the formulation layer: ; where represents the behavior function of the th node in the formulation layer, represents the addition ratio of the th formulation raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, represents the response sensitivity of the th formulation raw material, represents the onset offset of the th formulation raw material; S14. Define the behavior function of the nodes in the process layer: ; wherein, represents the behavior function of the th process layer node, represents the th process parameter value, represents the adjustment coefficient, represents the exponential function, represents the regulation factor, represents the optimal theoretical value of the process for the th process parameter value; S15. Define the behavior function of the performance layer node: ; wherein, represents the behavior function of the th performance layer node, represents the adjustment coefficient, represents the set of formulation layer nodes connected to the th performance layer node, represents the set of process layer nodes connected to the th performance layer node, represents the connection edge weight between the th formulation layer node and the th performance layer node, represents the connection edge weight between the th process layer node and the th performance layer node; S16. Based on the attribute values, behavior functions of the nodes, and the connection relationships of the cross-layer edges, construct a hierarchical polymorphic graph structure model.

[0011] Optionally, the S2 specifically includes: S21. Read the output values of the behavior functions of each formulation layer node and process layer node in the hierarchical polymorphic graph structure model, and arrange them in the order of node numbers to form a behavior response vector; S22. For any two nodes in the behavior response vector that have a path connection, calculate the difference between the behavior response values, and perform weighted processing in combination with the path edge weights to obtain the behavior difference value between the two nodes. The greater the behavior difference value, the higher the degree of incoordination of the responses of the two nodes under the current conditions; S23. Accumulate the behavior difference values of all adjacent node pairs in the path, and divide by the number of node pairs in the path to obtain the average behavior difference value of the path; S24. Calculate the change range of the target index between the starting node and the ending node for each path, and according to the degree of consistency of the performance change direction, allocate a positive factor to correct the average behavior difference value to obtain the conflict value of the path; S25. Store the conflict values of all paths into a three-dimensional structure array in the order of path numbers. The first dimension is the path number, the second dimension is the node level position, and the third dimension is the conflict value. The level positions that do not exist in the path are filled with null values; S26. Define the normalized three-dimensional structure array as the tension field tensor, which is used to characterize the tension state of the node combination path.

[0012] Optionally, the specific steps of S3 include: S31. Traverse all path dimensions in the tension field tensor and read the conflict values corresponding to each path number; S32. Perform maximum and minimum normalization processing on the conflict values of all paths to construct a standardized tension value sequence, so that all path tension values are within a unified interval range; S33. Set a tension threshold, screen the standardized tension value sequence, and retain the path combinations with tension values lower than the tension threshold to construct 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 optimization individual.

[0013] Optionally, the improvements of the improved jellyfish search algorithm specifically include: In the active movement stage, obtain the current candidate path tension value of the optimization individual, and calculate the tension difference with the upper and lower adjacent paths in the tension field tensor. If the tension value of the adjacent path is lower than 5% of the current path tension value, trigger the movement along the tension decreasing direction, and the movement amplitude is that the adjustment ratio of the recipe parameters does not exceed 2%, and the adjustment step of the process parameters does not exceed 3% of the corresponding parameter interval; In the passive drift stage, select the candidate paths with the top 20% of the tension values as the drift target set, and use the roulette wheel selection algorithm to select the jump target path in the drift target set and perform the jump operation; Reset the recipe parameters and process parameters of the current optimization individual according to the recipe parameters and process parameters bound to the jump target path.

[0014] In the improved jellyfish search algorithm adopted by the present invention, the tension difference threshold is set to 5%, which is determined based on the sensitivity analysis of the impact of path tension changes on performance in the actual feed formulation system. When the tension decrease exceeds 5%, it usually represents a substantial improvement in the path behavior consistency, so individual movement can be triggered. The movement amplitude is that the adjustment ratio of the formulation parameters does not exceed 2%, and the adjustment step of the process parameters does not exceed 3% of the corresponding parameter interval, aiming to ensure the convergence stability of local search and avoid the jump of the behavior function or the imbalance of the formulation caused by large-scale perturbations. The path drift target selects the path set with the top 20% of the tension values, which is considered based on the balance of optimizing search efficiency and jump quality, ensuring that individuals do not fall into the sub-optimal path space and significantly improving the overall tension coordination; the roulette wheel selection algorithm is used to perform path jumps in this set, which helps to construct a global exploration mechanism dominated by gradients in the low-tension space. The above parameter design effectively enhances the stability, self-adaptability and convergence quality of the jellyfish search algorithm in the high-dimensional formulation and process space, and has clear technical effects.

[0015] Optionally, step S5 specifically includes: According to the formulation parameters and process parameters in each optimized individual, calculate the predicted output values of the performance layer nodes in the hierarchical polymorphic graph structure model. The predicted output values include four indicators: palatability, forming rate, particle hardness, and growth performance, and weight the four indicators 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; Weight and combine the performance score and the tension score 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.

[0016] Optionally, step S6 specifically includes: S61. According to the formulation parameters and process parameters in the optimal individual, calculate the predicted output values of the performance layer nodes in the hierarchical polymorphic graph structure model, compare them with the actual observed performance values, and generate a performance error vector; S62. According to the performance error vector, trace the behavior function contribution paths of the corresponding formulation layer nodes and process layer nodes, and determine the response influence degree of each node on the error; S63. Adjust the behavior function based on the response influence degree of the node on the error, adopt a parameter adjustment strategy with the same error direction, and increase or decrease the adjustment coefficient in the behavior function within the set update rate range; S64. Write the updated behavior function into the hierarchical polymorphic graph structure model, replace the behavior function of the original node, and complete the iterative evolution of the hierarchical polymorphic graph structure model.

[0017] Optionally, the parameter adjustment strategy with consistent error directions specifically includes: When a certain performance index in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, increase the adjustment coefficient in the behavior functions of the corresponding formulation layer nodes and process layer nodes to enhance the behavior response intensity of the nodes; When a certain performance index in the performance error vector is positive, indicating that the predicted output value is higher than the actual observed performance value, decrease the adjustment coefficient in the behavior functions of the corresponding formulation layer nodes and process layer nodes to reduce the behavior response intensity of the nodes; The adjustment operation of the adjustment coefficient is restricted by the set update rate range in each iteration, which is used to prevent the behavior function from changing too much and causing the instability of the hierarchical polymorphic graph structure model.

[0018] The beneficial effects of the present invention are: First of all, by constructing a hierarchical polymorphic graph structure model, the present invention realizes the structured modeling of three core elements in the feed system, namely formulation raw materials, process parameters and performance indicators. The formulation layer, process layer and performance layer are connected by cross-layer edges, which not only clarifies the dependence relationship between various elements, but also provides a basis for path-level dynamic behavior modeling. This structured expression method breaks through the limitation of the traditional method of treating formulations and processes separately, and can accurately reflect the coupling effect of multi-factor combinations on performance results.

[0019] Secondly, the introduction of the behavior function makes each node no longer a static variable, but has dynamic response capabilities. The behavior functions of the formulation layer and process layer nodes can simulate their specific contributions to the target performance under different parameter conditions, while the behavior function of the performance layer measures the overall performance effect by integrating the upstream behavior outputs. This mechanism not only enhances the expression ability of the system, but also provides a clear parameter entry for subsequent model learning and behavior adjustment.

[0020] In addition, the present invention first introduces the concept of the tension field tensor to quantify the conflict degree of different paths in the behavior response process. By calculating the behavior differences between adjacent nodes in the path, a path tension index is constructed and used to optimize individual screening and search direction guidance, breaking through the limitation of traditional optimization methods that only rely on fitness functions. The introduction of the tension mechanism makes the optimization process more stable, reduces the problems of invalid search and local convergence, and enhances the physical interpretability and control accuracy of path selection.

[0021] Furthermore, an improved jellyfish search algorithm is adopted in the optimization process. This algorithm not only introduces the tension gradient as the basis for direction control in the active movement stage, but also constructs a jump target space using the tension distribution structure in the passive drift stage, thereby achieving a globally searchable ability guided by the structure. Meanwhile, to avoid falling into local optima, a local perturbation mechanism and a method for resetting parameters after path jumping are designed to ensure the diversity and adaptability of the search.

[0022] Finally, the present invention also establishes a model dynamic update mechanism. In each round of the optimization process, the formula and process parameters of the optimal individual are fed back to the hierarchical polymorphic graph structure model to adjust the behavior functions of relevant nodes, enabling the model to have the ability of continuous learning and evolution, and to adapt to complex scenarios such as different raw material batches, process fluctuations, and changes in performance targets, truly realizing the adaptive collaborative optimization of the formula and process. Brief Description of the Drawings

[0023] The drawings are used to provide a further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention, but do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall flowchart of the method for optimizing feed formula and adaptively controlling production line parameters based on machine learning proposed by the present invention; Figure 2 is the operation flowchart of constructing an optimized individual of the method for optimizing feed formula and adaptively controlling production line parameters based on machine learning proposed by the present invention. Detailed Description of the Embodiments

[0024] Now, the present invention will be further described in detail with reference to the drawings. These drawings are all simplified schematic diagrams, only illustrating the basic structure of the present invention in a schematic manner, so they only show the components related to the present invention.

[0025] Refer to Figure 1 and Figure 2 , the method for optimizing feed formula and adaptively controlling production line parameters based on machine learning includes the following steps: S1. Collect the formula raw material attributes, process parameters, and performance indicators of the feed, and construct 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, calculate the behavior differences between the nodes, and generate a tension field tensor to characterize the tension state of the node combination path; S3. According to the tension field tensor, screen the path combinations with tension values lower than the set tension threshold to generate optimized individuals, and each optimized individual includes a set of formula parameters and corresponding process parameters; S4. Input the optimized individuals into the improved jellyfish search algorithm. In the active movement stage, update the formulation parameters and process parameters according to the direction of the tension gradient, and perform a jump operation in the passive drift stage; S5. Calculate the fitness value according to the updated formulation 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 formulation parameters and process parameters of the optimal individual, update the node behavior functions of the formulation layer and the process layer in the hierarchical polymorphic graph structure model; S7. Determine whether the current raw material batch or production line working condition has changed. If it has changed, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, enter step S8; S8. Output the formulation parameters and process parameters in the current optimal individual, record the tension state and fitness value, and store them in the local database.

[0026] This method realizes the structured modeling of the complex relationship among feed raw material attributes, processing parameters, and performance output by constructing a hierarchical polymorphic graph structure model composed of a formulation layer, a process layer, and a performance layer. On this basis, the behavior function, as the core attribute of the node, is used to dynamically express the response process of each parameter to the target performance. By constructing a tension field tensor driven by behavior differences, path-level conflict evaluation and quantification are further realized, constituting the response basis for the whole-process optimization. The entire optimization system is based on model-driven logic rather than empirical parameter tuning, making the combined decision-making of formulation and process data-dependent, behaviorally feedback, and structurally evolvable. 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 to raw material fluctuations and production environment changes, thus effectively improving the accuracy, consistency, and engineering promotion value of feed production.

[0027] In this embodiment, step S1 specifically includes: S11. Collect the formulation raw material attributes of the feed, including nutrient content, cost information, and source stability indicators, as the attribute values of the nodes in the formulation layer. The nodes in the formulation layer include energy raw materials, protein raw materials, and nutritional additive raw materials; Collect the process parameters of the feed, including steam temperature, extrusion pressure, water addition amount, oil spraying amount, and cooling temperature, as the attribute values of the nodes in the process layer. The nodes in the process layer include steam temperature node, extrusion pressure node, water addition amount node, oil spraying amount node, and cooling temperature node; Collect the performance indicators of the feed, including palatability score, forming rate, pellet hardness, and growth performance indicators, as the attribute values of the nodes in the performance layer. The nodes in the performance layer 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. Set the connection relationship and initial edge weights of the cross-layer edges according to the contribution relationships between the formula layer nodes and the performance layer nodes, the process layer nodes and the performance layer nodes, and the formula layer nodes and the process layer nodes. S13. Define the behavior function of the formula layer nodes: ; Among them, represents the behavior function of the th formula layer node, represents the addition ratio of the th formula raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, represents the response sensitivity of the th formula raw material, represents the onset offset of the th formula raw material; S14. Define the behavior function of the process layer nodes: ; Among them, represents the behavior function of the th process layer node, represents the th process parameter value, represents the adjustment coefficient, represents the exponential function, represents the regulation factor, represents the optimal theoretical value of the th process parameter value; S15. Define the behavior function of the performance layer nodes: ; Among them, represents the behavior function of the th performance layer node, represents the adjustment coefficient, represents the set of formula layer nodes connected to the th performance layer node, represents the set of process layer nodes connected to the th performance layer node, represents the connection edge weight between the th formula layer node and the th performance layer node, represents the connection edge weight between the th process layer node and the th performance layer node; S16. Construct a hierarchical polymorphic graph structure model based on the attribute values of nodes, behavior functions, and the connection relationships of cross-layer edges.

[0028] This step introduces a structured acquisition mechanism for three information sources: formula, process, and performance, and classifies each element into the corresponding node layer, providing a traceable data basis for the hierarchical modeling of the graph model. By distinguishing raw material types, process nodes, and performance dimensions, the classification coding and standardized expression of model inputs are realized, enhancing the generalization and expandability of the system. The setting of cross-layer edge weights combines historical data to accurately map the actual contribution path and control sensitive points of raw materials to performance, making the graph structure no longer a static connection but having behavioral significance and quantitative basis. On this basis, it further provides context support for the definition of behavior functions, providing a stable structural input source for subsequent response calculation, tension evaluation, and optimization mechanisms. The overall design constructs a logical closed-loop between the original data and the structure graph model, laying a solid foundation for system modeling.

[0029] In this embodiment, the specific steps of S2 are as follows: S21. Read the output values of the behavior functions of each formula layer node and process layer node in the hierarchical polymorphic graph structure model, arrange them in the order of node numbers, and form a behavior response vector. S22. For any two nodes in the behavior response vector that are connected by a path, calculate the difference between the behavior response values, and perform weighted processing in combination with the path edge weight to obtain the behavior difference value between the two nodes. The larger the behavior difference value, the higher the degree of response incoordination between the two nodes under the current conditions. S23. Accumulate the behavior difference values of all adjacent node pairs in the path and divide by the number of node pairs in the path to obtain the average behavior difference value of the path. S24. Calculate the change range of the target index between the starting node and the ending node for each path, and according to the degree of consistency of the performance change direction, assign a positive factor to correct the average behavior difference value to obtain the conflict value of the path. S25. Store the conflict values of all paths into a three-dimensional structure array in the order of path numbers. The first dimension is the path number, the second dimension is the node level position, and the third dimension is the conflict value. The 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 to characterize the tension state of the node combination path.

[0030] This step establishes a path conflict evaluation system based on behavior response driving through the calculation of differences between the outputs of node behavior functions. Different from traditional performance difference or objective function methods, this method starts from the consistency of internal nodes in the path, quantifies the degree of behavior coordination in each path, forms a path conflict value, and then constructs a tension field tensor. The normalization processing and hierarchical distribution representation of the tension value enable the model to have tension visibility and path structure control capabilities. With the help of 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 optimization of behavior functions. The overall mechanism strengthens the guiding ability of the path structure for the optimization direction, making the optimization behavior have structural awareness, directionality, and interpretability, and significantly improving the convergence efficiency of the global optimal solution and the system response stability.

[0031] In this embodiment, S3 specifically includes: S31. Traverse all path dimensions in the tension field tensor and read the conflict values corresponding to each path number; S32. Perform maximum and minimum normalization processing on the conflict values of all paths to construct a standardized tension value sequence, so that the tension values of all paths are within a unified interval range; S33. Set a tension threshold, screen the standardized tension value sequence, and retain the path combinations with tension values lower than the tension threshold to construct a candidate path set; S34. Extract the formulation parameters and corresponding process parameters of each candidate path in the candidate path set to form an optimization individual.

[0032] Based on the construction of the tension field tensor, this step details the process of screening path combinations with lower tension and generating optimization individuals. Through the normalization processing of the tension value and the setting of the threshold mechanism, the structural screening of the entire path set is realized, effectively controlling the scale of the search space, and ensuring that the paths entering the optimization stage have high tension coordination. Further, each selected path is constructed into an optimization individual, and the structure is clearly a two-component set composed of formulation parameters and process parameters, improving the operability of the individual and the evaluation efficiency of the objective function. The generation logic of this optimization individual provides a high-quality and clearly structured initial population for the subsequent optimization algorithm, effectively avoiding the low-quality perturbation caused by random initialization, improving the overall optimization starting level, shortening the convergence period, and enhancing the search efficiency and reliability.

[0033] In this embodiment, the improvements of the improved jellyfish search algorithm specifically include: In the active movement stage, obtain the current candidate path tension value of the optimized individual, and calculate the tension difference between the current path and the upper and lower adjacent paths in the tension field tensor. If the tension value of the adjacent path is lower than 5% of the current path tension value, trigger the movement along the direction of decreasing tension, and the movement amplitude is that the adjustment ratio of the formula parameters does not exceed 2%, and the adjustment step of the process parameters does not exceed 3% of the corresponding parameter interval. In the passive drift stage, select the candidate paths with the top 20% of the tension values as the drift target set, and use the roulette wheel selection algorithm to select the jump target path in the drift target set, and perform the jump operation. Reset the formula parameters and process parameters of the current optimized individual according to the formula parameters and process parameters bound to the jump target path.

[0034] By introducing an improved jellyfish search algorithm driven by tension gradient, this step solves the problems of unclear direction and inaccurate jumping in high-dimensional path search of traditional optimization algorithms. In the active movement stage, the algorithm determines the direction based on the path tension difference, triggers the jump only when the tension of the adjacent path decreases by more than the set ratio, and controls the adjustment range of the formula and process parameters, making the individual evolution process more stable and accurate. In the passive drift stage, by means of the tension sorting strategy to limit the target path set and using the roulette wheel selection mechanism, a biased distribution jump is realized, strengthening the exploration ability of the low-tension area. The entire optimization process is dominated by the tension value structure rather than the global fitness function, making the search more structural and directional, and significantly improving the global convergence ability and path structure matching degree.

[0035] In this embodiment, the specific steps of S5 include: According to the formula parameters and process parameters in each optimized individual, calculate the predicted output values of the performance layer nodes in the hierarchical polymorphic graph structure model. The predicted output values include four indicators: palatability, forming rate, particle hardness, and growth performance, and perform weighted processing on the four indicators according to the preset weights to obtain the 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. Weight and combine the performance score and the tension score according to the 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.

[0036] After optimization, this step clarifies the specific calculation method of the fitness of the optimized individuals, and jointly incorporates the performance output value and the tension value into the fitness evaluation system. The performance part is standardized and weighted combined through four types of indicators to form a performance score, and the tension part generates a tension score through normalization processing. The two are synthesized into a fitness value according to a certain proportion, taking into account the path coordination and the target performance quality. This fitness mechanism of two-dimensional fusion avoids the drawback that path conflicts are ignored under the traditional single-objective drive, making the selected optimal solution not only optimal in performance, but also having a high tension coordination degree and the feasibility of actual implementation. The finally selected optimal individual has a stable data structure and clear logic, providing reliable input for the update of the behavior function and system feedback.

[0037] In this embodiment, the S6 specifically includes: S61. According to the formula parameters and process parameters in the optimal individual, calculate the predicted output value of the performance layer nodes in the hierarchical polymorphic graph structure model, compare it with each actual observed performance value, and generate a performance error vector; S62. According to the performance error vector, trace the contribution paths of the corresponding formula layer nodes and process layer nodes of the behavior function, and determine the response influence degree of each node on the error; S63. Adjust the behavior function based on the response influence degree of the node on the error, adopt a parameter adjustment strategy with the same error direction, and increase or decrease the adjustment coefficient in the behavior function within the set update rate range; S64. Write the updated behavior function into the hierarchical polymorphic graph structure model, replace the behavior function of the original node, and complete the iterative evolution of the hierarchical polymorphic graph structure model.

[0038] This step constructs a feedback closed-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 formula parameters and process parameters of the optimal individual into the model and comparing them with the target performance, an error vector is obtained, and then the formula and process nodes in the path are traced back in reverse. According to the response influence degree of the node on the error, hierarchical updates are performed, the adjustment coefficient in the behavior function is adjusted using the same direction strategy, and an update rate constraint is imposed to avoid model instability. This feedback mechanism enables the graph structure model to have the ability of online learning and dynamic adaptation, effectively coping with production disturbances such as raw material batch changes and performance target adjustments, and significantly enhancing the long-term applicability and adaptive control ability of the model.

[0039] In this embodiment, the parameter adjustment strategy with the same error direction specifically includes: When a certain performance index in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, increase the adjustment coefficient in the behavior functions of the corresponding formula layer nodes and process layer nodes to enhance the behavior response intensity of the nodes; When a certain performance index 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 functions of the corresponding formulation layer nodes and process layer nodes is reduced to lower the behavior response intensity of the nodes; The adjustment operation of the adjustment coefficient is restricted by the set update rate range in each iteration, which is used to prevent the behavior function from changing too much and causing the instability of the hierarchical polymorphic graph structure model.

[0040] The parameter adjustment strategy with the same error direction clarifies the policy logic of updating the behavior function parameters, adjusts the adjustment coefficient of the response nodes based on the performance error direction, thereby increasing or decreasing the output response intensity of the nodes. When the performance prediction is insufficient, the overall performance is improved by enhancing the response of key nodes; when the prediction exceeds the target, the response is appropriately weakened to avoid overfitting or resource waste. This strategy introduces an update rate constraint, making the parameter changes have smoothness and controllability, and effectively preventing the model from oscillating or deviating during the update process.

[0041] Example 1

[0042] To verify the feasibility of the present invention in implementation, the present invention is applied to the pig feed production line of a large feed processing enterprise for actual deployment and testing. This enterprise has long faced two typical problems: First, the adjustment of raw material ratio often depends on the experience of nutritionists, with slow response and many trials and errors, and it is difficult to adapt to the changes in raw material batches; Second, most of the process parameters of the production line are set manually, with low stability, large fluctuations in the performance of the finished products, which seriously restricts the quality consistency and production efficiency of feed products.

[0043] In this embodiment, the system first deploys a sensing and acquisition module to collect the nutritional components, source batches, and cost attributes of six types of raw materials such as corn, soybean meal, fish meal, wheat, calcium hydrogen phosphate, and additives entering the production line. At the same time, it real-time monitors the key process parameters during the operation of the production line, including steam temperature, extrusion pressure, water addition amount, oil spraying amount, and cooling temperature. In addition, it integrates terminal quality detection equipment to transmit back the palatability score, forming rate, particle hardness, and animal growth performance data.

[0044] After adopting the method of the present invention, first, a hierarchical polymorphic graph structure model based on three layers of raw materials, processes, and performance is established, the behavior functions of each node are constructed, and the initial response relationship is fitted through historical data. Subsequently, a tension field tensor is constructed by calculating the differences between the behavior functions to quantify the coordination degree of different path combinations. According to the tension value, the paths with low tension are selected, the corresponding formulation parameters and process parameters of the paths are extracted to generate optimized individuals, and the improved jellyfish search algorithm is input for global search. During the optimization process, an active adjustment guided by the tension gradient and a path drift mechanism under structural constraints are introduced to achieve efficient search for the optimal formulation path. Finally, the individual with the highest fitness is selected and used for reverse update of the behavior function to realize the closed-loop evolution of the model.

[0045] After 7 consecutive days of operation testing, the optimized results output by the system showed high stability and significant improvement effects. Taking the parameters generated by optimization as an example, in the system-recommended formula, corn is 35%, soybean meal is 25%, fish meal is 10%, wheat is 20%, monocalcium phosphate is 5%, and additives are 5%. The corresponding process parameters are steam temperature of 90 °C, extrusion pressure of 3.5 MPa, water addition of 20%, oil spraying amount of 4%, and cooling temperature of 25 °C. Under this parameter combination, the palatability score of the actual finished feed reaches 8.7 points, the forming rate is as high as 96.2%, the particle hardness remains at 22.5 N, the average growth performance of pigs increases by 11.4%, and the overall fitness value is 0.925, showing a significant improvement compared with 0.781 before optimization.

[0046] During actual operation, the system achieved dynamic adjustment of the formula by the hour and fine-tuning of process parameters. On average, a self-learning update cycle was completed every 24 hours. Stable responses were achieved for both the switching of two raw material batches and one equipment maintenance process, without any performance cliff or abnormal fluctuations. The results show that the present invention exhibits good adaptability, optimization efficiency, and system stability in an industrial environment with complex raw material variations, variable performance targets, and highly coupled production line parameters, and has significant engineering value and promotion prospects.

[0047] The above is only the preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution of the present invention and its inventive concept, makes equivalent substitutions or changes, and all should be covered within the protection scope of the present invention.

Claims

1. A method for optimizing feed formulation and adaptively controlling production line parameters based on machine learning, characterized in that It includes the following steps: S1. Collect the formula raw material attributes, process parameters, and performance indicators of the feed, and construct 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, calculate the behavior differences between the nodes, and generate a tension field tensor to characterize the tension state of the node combination path; S3. According to the tension field tensor, screen the path combinations with tension values lower than the set tension threshold to generate optimized individuals. Each optimized individual includes a set of formula parameters and corresponding process parameters; S4. Input the optimized individuals into an improved jellyfish search algorithm. In the active movement stage, update the formula parameters and process parameters according to the tension gradient direction, and perform a jump operation in the passive drift stage; S5. Calculate the fitness value according to the updated formula 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 formula parameters and process parameters of the optimal individual, update the behavior functions of the nodes in the formula layer and the process layer of the hierarchical polymorphic graph structure model; S7. Judge whether the current raw material batch or production line working condition has changed. If it has changed, return to step S2 to regenerate the tension field tensor and repeat steps S3 to S6. Otherwise, enter step S8; S8. Output the formula parameters and process parameters in the current optimal individual, record the tension state and fitness value, and store them in the local database.

2. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, wherein The specific content of S1 includes: S11. Collect the formula raw material attributes of the feed, including nutrient content, cost information, and source stability indicators, as the attribute values of the nodes in the formula layer. The nodes in the formula layer include energy raw materials, protein raw materials, and nutritional additive raw materials; Collect the process parameters of the feed, including steam temperature, extrusion pressure, water addition amount, oil spraying amount, and cooling temperature, as the attribute values of the nodes in the process layer. The nodes in the process layer include steam temperature node, extrusion pressure node, water addition amount node, oil spraying amount node, and cooling temperature node; Collect the performance indicators of the feed, including palatability score, forming rate, particle hardness, and growth performance indicators, as the attribute values of the nodes in the performance layer. The nodes in the performance layer include palatability node, forming rate node, particle hardness node, and growth performance node; S12. Analyze the contribution relationship of each node to the feed performance indicators according to historical production data. Based on the contribution relationships between the nodes in the formula layer and the performance layer, between the nodes in the process layer and the performance layer, and between the nodes in the formula layer and the process layer, set the connection relationship and initial edge weight of the cross-layer edges; S13. Define the behavior function of the nodes in the formula layer: ； Among them, represents the behavior function of the th formulation layer node, represents the addition ratio of the th formulation raw material, represents the adjustment coefficient, represents the hyperbolic tangent function, represents the response sensitivity of the th formulation raw material, represents the onset offset of the th formulation raw material; S14. Define the behavior function of the nodes in the process layer: ； Among them, represents the behavior function of the th process layer node, represents the th process parameter value, represents the adjustment coefficient, represents the exponential function, represents the regulation factor, represents the optimal theoretical value of the process for the th process parameter value; S15. Define the behavior function of the nodes in the performance layer: ； Among them, represents the behavior function of the th performance layer node, represents the adjustment coefficient, represents the set of recipe layer nodes connected to the th performance layer node, represents the set of process layer nodes connected to the th performance layer node, represents the connection edge weight between the th recipe layer node and the th performance layer node, represents the connection edge weight between the th process layer node and the th performance layer node; S16. Based on the attribute values of the nodes, the behavior functions, and the connection relationships of the cross-layer edges, construct a hierarchical polymorphic graph structure model.

3. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, wherein The specific content of S2 includes: S21. Read the output values of the behavior functions of each formulation layer node and process layer node in the hierarchical polymorphic graph structure model, and arrange them in the order of node numbers to form a behavior response vector; S22. For any two nodes in the behavior response vector that are connected by a path, calculate the difference between the behavior response values, and perform weighted processing in combination with the path edge weights to obtain the behavior difference value between the two nodes. The larger the behavior difference value, the higher the degree of response incoordination between the two nodes under the current conditions; S23. Accumulate the behavior difference values of all adjacent node pairs in the path, and divide by the number of node pairs in the path to obtain the average behavior difference value of the path; S24. Calculate the change range of the target index between the starting node and the ending node for each path. According to the degree of consistency of the performance change direction, assign a positive factor 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. The first dimension is the path number, the second dimension is the node hierarchical position, and the third dimension is the conflict value. The hierarchical 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 to characterize the tension state of the node combination path.

4. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, characterized in that, The specific steps of S3 are as follows: S31. Traverse all path dimensions in the tension field tensor and read the conflict values corresponding to each path number; S32. Perform maximum and minimum normalization processing on the conflict values of all paths to construct a standardized tension value sequence, so that all path tension values are in a unified interval range; S33. Set a tension threshold, screen the standardized tension value sequence, and retain the path combinations with tension values lower than the tension threshold to construct a candidate path set; S34. Extract the formulation parameters and corresponding process parameters of each candidate path in the candidate path set to form an optimization individual.

5. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, wherein The improvements of the improved jellyfish search algorithm specifically include: In the active movement stage, obtain the current candidate path tension value of the optimization individual, and calculate the tension difference with the upper and lower adjacent paths in the tension field tensor. If the tension value of the adjacent path is lower than 5% of the current path tension value, trigger the movement along the tension decreasing direction, and the movement amplitude is that the adjustment ratio of the formulation parameters does not exceed 2%, and the adjustment step of the process parameters does not exceed 3% of the corresponding parameter interval; In the passive drift stage, select the candidate paths with the top 20% tension values as the drift target set, and use the roulette wheel selection algorithm to select the jump target path in the drift target set and perform the jump operation; Reset the formulation parameters and process parameters of the current optimization individual according to the formulation parameters and process parameters bound to the jump target path.

6. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, wherein The specific steps of S5 are as follows: According to the formulation parameters and process parameters in each optimization individual, calculate the predicted output values of the performance layer nodes in the hierarchical polymorphic graph structure model. The predicted output values include four indicators: palatability, forming rate, particle hardness, and growth performance, and perform weighted processing on the four indicators according to the preset weights to obtain the 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; Weight-combine the performance score and the tension score according to a preset ratio to obtain the fitness value of the optimized individual; Compare the fitness values among all the optimized individuals, and select the optimized individual with the highest fitness value as the optimal individual in the current round.

7. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 1, wherein The specific content of S6 is as follows: S61. According to the formulation parameters and process parameters in the optimal individual, calculate the predicted output value of the performance layer nodes in the hierarchical polymorphic graph structure model, compare it with each actual observed performance value, and generate a performance error vector; S62. According to the performance error vector, trace the contribution paths of the behavior functions of the corresponding formulation layer nodes and process layer nodes, and determine the response influence degree of each node on the error; S63. Adjust the behavior function based on the response influence degree of the node on the error, adopt a parameter adjustment strategy with the same error direction, and increase or decrease the adjustment coefficient in the behavior function within the 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, and complete the iterative evolution of the hierarchical polymorphic graph structure model.

8. The method for optimizing feed formula and adaptively controlling production line parameters based on machine learning according to claim 7, wherein The specific content of the parameter adjustment strategy with the same error direction is as follows: When a certain performance index in the performance error vector is negative, indicating that the predicted output value is lower than the actual observed performance value, increase the adjustment coefficient in the behavior functions of the corresponding formulation layer nodes and process layer nodes to enhance the behavior response intensity of the nodes; When a certain performance index in the performance error vector is positive, indicating that the predicted output value is higher than the actual observed performance value, decrease the adjustment coefficient in the behavior functions of the corresponding formulation layer nodes and process layer nodes to reduce the behavior response intensity of the nodes; The adjustment operation of the adjustment coefficient is restricted by the set update rate range in each round of iteration, which is used to prevent the behavior function from changing too much and causing the instability of the hierarchical polymorphic graph structure model.

Images