Collaborative scheduling and operation optimization method of feed production equipment based on deep learning

Abstract

The present invention discloses a method for collaborative scheduling and operation optimization of feed production equipment based on deep learning, comprising the following steps: S1, constructing an equipment state sequence dataset; S2, constructing a disturbance sequence dataset; S3, inputting a disturbance-resistant residual fusion network to generate a preliminary scheduling plan; S4, constructing an equipment conflict reasoning graph; S5, embedding the equipment conflict reasoning graph into the scheduling network to revise the preliminary scheduling plan; S6, inputting the revised preliminary scheduling plan into an improved NGBoost model, introducing a confidence factor estimation and dynamic distribution calibration mechanism, and outputting a predicted expected value and uncertainty score for each scheduling behavior; S7, identifying high-risk behaviors based on the uncertainty score and generating a set of scheduling revision candidates; S8, performing a multi-objective comprehensive evaluation on the set of scheduling revision candidates, screening the one with the highest score and issuing it as the final scheduling plan for execution. The present invention combines deep learning with the improved NGBoost model to achieve intelligent scheduling optimization of feed equipment.

Description

Technical Field

[0001] The present invention relates to the field of deep learning and industrial automation technology, and in particular to a method for collaborative scheduling and operation optimization of feed production equipment based on deep learning. Background Art

[0002] In the modern feed manufacturing industry, equipment collaborative scheduling and operation optimization have become key links that restrict the company's production efficiency, energy utilization and equipment safety operation level. Feed production involves multiple process stages such as raw material crushing, batching, mixing, granulation, cooling, and packaging. Each stage relies on the collaborative operation of multiple different types of equipment. With the continuous improvement of factory automation levels, more and more companies are adopting scheduling methods based on rule engines, static scheduling systems or empirical formula control. However, traditional scheduling technologies usually have problems such as strong model rigidity, weak response capabilities, and poor adaptability to real-time conditions and disturbances. They are difficult to cope with high-frequency process fluctuations and equipment resource conflicts in complex dynamic environments, resulting in unreasonable production scheduling, energy waste, frequent production bottlenecks and other problems.

[0003] Existing feed equipment scheduling systems mainly rely on static scheduling tables or heuristic optimization algorithms to generate scheduling plans, and usually divide tasks based on average capacity models or standard time windows. This type of method cannot fully perceive the uncertainty factors in the actual operating status, such as equipment performance degradation, switching delays, sudden task insertions, or fluctuations in raw material quality, resulting in scheduling results that are extremely sensitive to small disturbances. Once a device experiences a slight delay or operational fluctuation, the system lacks an effective mechanism for adaptive adjustment, and often requires manual intervention to re-intervene or overall reschedule tasks, resulting in frequent plan failures. In addition, most systems have not established a unified dynamic disturbance modeling mechanism, nor have they structured historical abnormal behaviors for predictive optimization. Scheduling strategies remain more at the rule execution level, and are unable to form a closed-loop scheduling system with learning capabilities.

[0004] On the other hand, some studies have attempted to introduce intelligent methods such as neural networks or reinforcement learning to optimize production scheduling, but there are generally three significant problems: First, model construction is mostly based on the assumption of ideal working conditions, and lacks systematic modeling of real disturbance scenarios, resulting in a disconnect between training results and actual applications; second, most methods use a global single path as the core perspective when modeling scheduling, ignoring the process logic constraints and spatial layout conflicts between equipment, and cannot effectively avoid the implicit conflict relationships in the parallel execution of multiple devices; third, the scheduling strategy output is usually a hard instruction, lacking confidence assessment and risk perception of the prediction results, and cannot automatically guide the modification and replacement of the scheduling plan when the uncertainty of the results is high, resulting in insufficient robustness and increased failure rate in complex environments.

[0005] Existing technologies also fail to fully consider the issue of multi-objective scheduling optimization. Most systems only consider capacity maximization during the optimization process, lacking multi-dimensional considerations such as energy costs, resource conflicts, or task robustness. As a result, the scheduling solutions ultimately output by the system often fail to maximize benefits in real-world scenarios. Furthermore, current systems generally lack effective scheduling output evaluation mechanisms, making it impossible to perform structured quantitative comparisons and screening of the various generated scheduling alternatives. Decisions rely solely on the scoring function within the model, ignoring the actual differences between candidate solutions in terms of disturbance adaptability, load balancing, and other aspects.

[0006] Therefore, how to provide a method for collaborative scheduling and operation optimization of feed production equipment based on deep learning is an urgent problem that technicians in this field need to solve. Summary of the Invention

[0007] One purpose of the present invention is to propose a method for collaborative scheduling and operation optimization of feed production equipment based on deep learning. The present invention combines deep learning modeling with disturbance perception and reasoning mechanism to construct a disturbance-resistant residual fusion network and equipment conflict reasoning graph to realize intelligent optimization of the scheduling process of feed production equipment. It has the advantages of being highly robust to disturbances, strong adaptive correction capability, high production capacity achievement rate, and controllable energy consumption. It effectively solves the problems of slow response to abnormal fluctuations, insufficient conflict logic processing, and single optimization target in the existing scheduling system, and significantly improves the intelligence level and industrial practical value of the system.

[0008] The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to an embodiment of the present invention includes the following steps:

[0009] S1. Collect the operating status data of each device in the feed production process and the corresponding task recipe information to form an equipment status sequence data set;

[0010] S2. Construct a disturbance sample set, wherein the disturbance sample set includes startup delay, operation interruption and recipe switching anomaly, and is structured into a disturbance sequence data set;

[0011] S3. Inputting the device state sequence data set and the disturbance sequence data set into a disturbance-resistant residual fusion network, wherein the disturbance-resistant residual fusion network includes a main path and a side path, and dynamically weighted fusion is performed through a residual gated fusion unit to generate a preliminary scheduling plan;

[0012] S4. Constructing an equipment conflict reasoning graph based on the task recipe information and the spatial arrangement relationship, where nodes represent equipment-batch operation states, edges represent non-parallel conflict conditions and are accompanied by conflict rule expressions;

[0013] S5. Embed the equipment conflict reasoning diagram as a structural constraint into the scheduling network, shield the scheduling path that violates the process logic, and revise the preliminary scheduling plan;

[0014] S6. Input the revised preliminary scheduling plan into the improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism to output the predicted expected value and uncertainty score of each scheduling behavior;

[0015] S7. Identify high-risk behaviors based on the uncertainty score and generate a set of scheduling modification candidates;

[0016] S8. Perform a multi-objective comprehensive evaluation on the scheduling correction candidate set, calculate scores based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and select the scheduling correction candidate with the highest score to be issued as the final scheduling plan for execution.

[0017] Optionally, the operating status data includes the current production capacity of the equipment, energy consumption per unit time, operating load rate, operating temperature, vibration amplitude, task execution time and topological connection information between devices.

[0018] Optionally, the S3 specifically includes:

[0019] S31. Perform time dimension alignment and feature channel normalization processing on the device state sequence dataset and the disturbance sequence dataset, and construct a main path input tensor and a side path input tensor corresponding to the time axis;

[0020] S32. Input the main path input tensor into a main path network composed of a stack of one-dimensional convolutional layers and gated recurrent units, extract short-term dependency features through the one-dimensional convolutional layers, and extract running trends across time steps through the gated recurrent units to obtain a main path feature tensor;

[0021] S33, inputting the side path input tensor into a side path network composed of a convolutional structure, a multi-scale residual connection, and an attention mechanism, extracting dynamic impact features of the disturbance at different time scales, and outputting a disturbance path feature tensor;

[0022] S34, splicing the main path feature tensor and the perturbation path feature tensor in the feature dimension, and inputting the concatenated features into a residual gated fusion unit, wherein the residual gated fusion unit performs a channel-by-channel weighted fusion on the main path feature tensor and the perturbation path feature tensor to obtain a fused feature tensor;

[0023] S35, performing channel compression on the fused feature tensor and inputting the result into a structured decoder, wherein the structured decoder includes a device number encoding layer, a time window encoding layer, and a load distribution encoding layer;

[0024] The device number encoding layer performs a position mapping operation, embedding and encoding the spatial channels in the fused feature tensor and the device topology index, generating a number vector representation corresponding to each device, which constitutes the device identification dimension in the scheduling instruction;

[0025] The time window encoding layer performs convolution expansion and sliding window offset processing on the time step index in the fused feature tensor, extracts the corresponding starting time segment and duration period in each scheduling path, and converts them into an integer interval identification vector;

[0026] The load distribution encoding layer performs a maximum-mean two-pass normalization operation based on the fused feature tensor to form a load distribution ratio vector corresponding to the current task of each device;

[0027] S36. Assemble the output results of the structured decoder into a preliminary scheduling plan, wherein the preliminary scheduling plan includes the start and stop sequence, numbering information, task execution time window and load distribution ratio value of each device.

[0028] Optionally, the S4 specifically includes:

[0029] S41. Analyze task recipe information, extract the equipment usage sequence, process stage requirements, and equipment occupancy time period corresponding to each recipe, and form a recipe structured task template;

[0030] S42. Collect spatial layout data of the feed workshop, including equipment number, installation location, physical occupied area, and relationship between process channels, and establish a spatial layout mapping table;

[0031] S43. Based on the equipment occupancy of each batch task in the recipe structured task template, the state combination of each device under each batch task is set as a node set of the device conflict reasoning graph, where the node is uniquely identified by a device ID-batch ID combination, indicating the equipment-batch operation status;

[0032] S44. Based on the equipment combinations in the spatial arrangement mapping table and the recipe structured task templates that have a parallel relationship, identify task pairs with resource overlap or time conflict, and define them as conflict edge sets;

[0033] S45. Setting a conflict type identifier and a conflict rule expression for each conflicting edge, wherein the conflict rule expression is composed of recipe difference, physical space overlap, operation stage incompatibility, and device sharing status;

[0034] S46. Construct a complete device conflict reasoning graph, including a node set, an edge set, and corresponding conflict rule expressions.

[0035] Optionally, the S5 specifically includes:

[0036] S51, extracting the start and stop sequence, number information, task execution time window, and load distribution ratio of each equipment task in the preliminary scheduling plan, and parsing it into a scheduling path set;

[0037] S52, traversing the scheduling path set, and for each node combination involved in the scheduling path, searching for the corresponding node and adjacent conflicting edge in the device conflict reasoning graph;

[0038] S53. Based on the conflict rule expression recorded in the conflict edge, the current scheduling path is conflict-verified one by one, and conflict path branches that violate space occupancy, recipe dependency, and resource mutual exclusion conditions are identified;

[0039] S54. Mark the identified conflicting path branches and construct a scheduling mask tensor in the scheduling network. The scheduling mask tensor is used to mask the output channels corresponding to the conflicting paths in the forward propagation phase.

[0040] S55 , outputting the scheduling path filtered by the scheduling mask tensor as a revised preliminary scheduling plan, wherein the revised preliminary scheduling plan does not include any scheduling path that violates the process logic of the equipment conflict inference graph rule.

[0041] Optionally, the improved NGBoost model is improved by introducing a confidence factor estimation and dynamic distribution calibration mechanism when predicting the expected value and uncertainty score of the scheduling behavior, specifically including:

[0042] Based on the historical disturbance data set, the prediction error amplitude of each scheduling behavior is statistically analyzed to construct a confidence factor for dynamically adjusting the output distribution width:

[0043] ;

[0044] in, Represents the time step The confidence factor when represents the exponential function, represents the adjustment parameter, Represents the time step The prediction error when represents the mean of the prediction error;

[0045] A dynamic distribution calibration mechanism is introduced during the natural gradient boosting process, and the following adjustments are performed each time the prediction distribution parameters are updated:

[0046] ;

[0047] in, represents the natural gradient after calibration, represents the standard natural gradient, represents the distribution calibration intensity coefficient, Represents the time step The true disturbance distribution when Represents the time step The predicted distribution when represents the prediction distribution parameter;

[0048] Decoding the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into a prediction expectation value and uncertainty score corresponding to each scheduling behavior, wherein the prediction expectation value represents the performance expectation of the current scheduling behavior under disturbance;

[0049] The confidence factor is used as a dynamic weight to scale the original variance value, and the adjusted variance value is used as the uncertainty score corresponding to the current scheduling behavior.

[0050] Optionally, the S7 specifically includes:

[0051] S71. Set a scheduling risk identification threshold, mark scheduling behaviors with uncertainty scores exceeding the scheduling risk identification threshold as high-risk behaviors, and construct a high-risk behavior index list;

[0052] S72. Locate the context structure of the high-risk behavior in the revised preliminary scheduling plan based on the equipment number, task recipe requirements, and time window position of each high-risk behavior;

[0053] S73. Generate a scheduling correction candidate for each high-risk behavior without violating the current task recipe execution logic and equipment space conflict constraints;

[0054] S74: Classify all generated scheduling revision candidate solutions into a scheduling revision candidate set, and retain the mapping relationship in the revised preliminary scheduling solution.

[0055] Optionally, the S8 specifically includes:

[0056] S81: Receive the generated scheduling correction candidate set, and perform structured analysis on the equipment start / stop sequence, time window, and load distribution information associated with each scheduling correction candidate solution;

[0057] S82. Calculate the capacity achievement rate of each scheduling modification candidate plan based on the standard capacity target of the current batch task, where the capacity achievement rate is the ratio of the actual execution capacity to the theoretical formula capacity;

[0058] S83. Estimating the energy consumption of the equipment operation plan corresponding to each candidate scheduling revision plan, and calculating the energy consumption change compared to the initial scheduling plan by combining the unit time operation power and load level;

[0059] S84. Extract the mean uncertainty score of each scheduling correction candidate under the prediction of the improved NGBoost model as a risk quantification indicator for measuring the disturbance adaptability;

[0060] S85. Construct a weighted comprehensive evaluation function based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and rank each scheduling correction candidate plan by score, where the weights in the weighted comprehensive evaluation function are set according to the process objectives;

[0061] S86. Screen the scheduling correction candidate with the highest score and issue it as the final scheduling plan for execution.

[0062] The beneficial effects of the present invention are:

[0063] First, this invention collects equipment operating status data and recipe task information to construct a multidimensional equipment state sequence and disturbance sample dataset. It then introduces a disturbance-resistant residual fusion network to jointly model equipment status and historical disturbance behavior, effectively enhancing the model's sensitivity and robustness to small fluctuations. By designing a parallel residual structure for the main and side paths and implementing dynamic information weighting with a residual gated fusion unit, this approach addresses the core issue of traditional models, which are prone to instability and misjudgment in the face of disturbances such as operational interruptions and switching anomalies.

[0064] Secondly, this invention introduces a device conflict reasoning graph structure, using the operating status of the device-batch level as nodes. Using scheduling logs and process rules, the parallel conflict edges and their logical conditions between devices are inferred, thereby formally modeling potential resource conflicts and process sequence conflicts between devices. By embedding this conflict reasoning graph as a structural constraint within the scheduling network, the model automatically blocks scheduling paths that do not meet production logic, ensuring that the output scheduling plan conforms to the actual operating rules and spatial layout constraints between devices, significantly improving the plan's implementability.

[0065] In addition, the output of the scheduling behavior of the present invention is no longer limited to a single result, but by introducing an improved NGBoost model, a distributed prediction of each scheduling behavior is achieved, and a probabilistic result including a predicted expected value and an uncertainty score is output. In particular, the improved part realizes adaptive adjustment of the prediction credibility of areas with different disturbance levels through confidence factor estimation and dynamic distribution calibration mechanism, thereby accurately identifying scheduling behaviors with high risks. Based on the uncertainty score, the system further constructs a candidate set for scheduling corrections, and introduces a multi-objective comprehensive evaluation mechanism to quantitatively score and rank all correction plans from three aspects: capacity achievement rate, energy consumption change, and disturbance adaptability. Finally, the optimal plan is selected and issued for execution to ensure that the scheduling results achieve a dynamic balance between efficiency and stability.

[0066] Through this technical approach, the present invention not only overcomes the limitations of traditional scheduling algorithms in handling disturbances and conflicts at the modeling level, but also achieves enhanced interpretability and robustness at the output strategy level. This provides an integrated intelligent scheduling solution for complex feed production systems, equipped with learning capabilities, structural constraints, self-correction, and multi-objective coordinated optimization. This approach significantly reduces scheduling failure rates, improves production capacity stability, and enhances resource utilization efficiency, possessing broad industrial application value and promising prospects for widespread adoption. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] 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:

[0068] Figure 1 This is the overall flow chart of the deep learning-based collaborative scheduling and operation optimization method for feed production equipment proposed in the present invention;

[0069] Figure 2 This is a schematic diagram of the structure of the disturbance-resistant residual fusion network of the deep learning-based collaborative scheduling and operation optimization method for feed production equipment proposed in the present invention;

[0070] Figure 3 This is a flow chart of the confidence factor estimation and dynamic distribution calibration mechanism of the improved NGBoost model in the process of predicting the expected value and uncertainty scoring of scheduling behavior in the deep learning-based collaborative scheduling and operation optimization method of feed production equipment proposed in the present invention. DETAILED DESCRIPTION

[0071] 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.

[0072] refer to Figure 1-Figure 3 , a method for collaborative scheduling and operation optimization of feed production equipment based on deep learning, including the following steps:

[0073] S1. Collect the operating status data of each device in the feed production process and the corresponding task recipe information to form an equipment status sequence data set;

[0074] S2. Construct a disturbance sample set, wherein the disturbance sample set includes startup delay, operation interruption and recipe switching anomaly, and is structured into a disturbance sequence data set;

[0075] S3. Inputting the device state sequence data set and the disturbance sequence data set into a disturbance-resistant residual fusion network, wherein the disturbance-resistant residual fusion network includes a main path and a side path, and dynamically weighted fusion is performed through a residual gated fusion unit to generate a preliminary scheduling plan;

[0076] S4. Constructing an equipment conflict reasoning graph based on the task recipe information and the spatial arrangement relationship, where nodes represent equipment-batch operation states, edges represent non-parallel conflict conditions and are accompanied by conflict rule expressions;

[0077] S5. Embed the equipment conflict reasoning diagram as a structural constraint into the scheduling network, shield the scheduling path that violates the process logic, and revise the preliminary scheduling plan;

[0078] S6. Input the revised preliminary scheduling plan into the improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism to output the predicted expected value and uncertainty score of each scheduling behavior;

[0079] S7. Identify high-risk behaviors based on the uncertainty score and generate a set of scheduling modification candidates;

[0080] S8. Perform a multi-objective comprehensive evaluation on the scheduling correction candidate set, calculate scores based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and select the scheduling correction candidate with the highest score to be issued as the final scheduling plan for execution.

[0081] This method achieves collaborative optimization control of feed production equipment in complex operating scenarios by constructing an intelligent scheduling process that integrates multi-source data, disturbance modeling, structural constraints, and uncertainty prediction. The model's robustness to operational anomalies is enhanced by using a disturbance-resistant residual fusion network. The structural elimination of spatial and process conflicts is achieved by combining the equipment conflict reasoning graph, and the risk scoring and adaptive correction of scheduling behavior are achieved by introducing an improved NGBoost model. The overall solution has the characteristics of automatic generation, dynamic adjustment, risk perception, and optimal screening. In a production environment with frequent equipment status changes and complex task switching, it can significantly improve system stability, task achievement rate, and energy efficiency, meeting the multi-dimensional performance requirements of intelligent scheduling systems for high-intensity industrial production.

[0082] In this embodiment, the operating status data includes the current production capacity of the equipment, energy consumption per unit time, operating load rate, operating temperature, vibration amplitude, task execution time and topological connection information between devices.

[0083] In this embodiment, S3 specifically includes:

[0084] S31. Perform time dimension alignment and feature channel normalization processing on the device state sequence dataset and the disturbance sequence dataset, and construct a main path input tensor and a side path input tensor corresponding to the time axis;

[0085] S32. Input the main path input tensor into a main path network composed of a stack of one-dimensional convolutional layers and gated recurrent units, extract short-term dependency features through the one-dimensional convolutional layers, and extract running trends across time steps through the gated recurrent units to obtain a main path feature tensor;

[0086] S33, inputting the side path input tensor into a side path network composed of a convolutional structure, a multi-scale residual connection, and an attention mechanism, extracting dynamic impact features of the disturbance at different time scales, and outputting a disturbance path feature tensor;

[0087] S34, splicing the main path feature tensor and the perturbation path feature tensor in the feature dimension, and inputting the concatenated features into a residual gated fusion unit, wherein the residual gated fusion unit performs a channel-by-channel weighted fusion on the main path feature tensor and the perturbation path feature tensor to obtain a fused feature tensor;

[0088] S35, performing channel compression on the fused feature tensor and inputting the result into a structured decoder, wherein the structured decoder includes a device number encoding layer, a time window encoding layer, and a load distribution encoding layer;

[0089] The device number encoding layer performs a position mapping operation, embedding and encoding the spatial channels in the fused feature tensor and the device topology index, generating a number vector representation corresponding to each device, which constitutes the device identification dimension in the scheduling instruction;

[0090] The time window encoding layer performs convolution expansion and sliding window offset processing on the time step index in the fused feature tensor, extracts the corresponding starting time segment and duration period in each scheduling path, and converts them into an integer interval identification vector;

[0091] The load distribution encoding layer performs a maximum-mean two-pass normalization operation based on the fused feature tensor to form a load distribution ratio vector corresponding to the current task of each device;

[0092] S36. Assemble the output results of the structured decoder into a preliminary scheduling plan, wherein the preliminary scheduling plan includes the start and stop sequence, numbering information, task execution time window and load distribution ratio value of each device.

[0093] By separating the main and side paths, the system not only extracts regular operating patterns but also effectively identifies disturbance signatures and dynamically integrates them, preventing scheduling plans from failing due to minor fluctuations. In particular, the residual gated fusion unit allows the model to flexibly adjust information flow intensity during different operating phases, dynamically optimizing the output direction of the scheduling plan. The structured decoder design further enhances the structural rationality and execution consistency of the output, ensuring device matching, time consistency, and load balance in actual deployments, providing accurate and stable basic information output for the final scheduling plan.

[0094] In this embodiment, the S4 specifically includes:

[0095] S41. Analyze task recipe information, extract the equipment usage sequence, process stage requirements, and equipment occupancy time period corresponding to each recipe, and form a recipe structured task template;

[0096] S42. Collect spatial layout data of the feed workshop, including equipment number, installation location, physical occupied area, and relationship between process channels, and establish a spatial layout mapping table;

[0097] S43. Based on the equipment occupancy of each batch task in the recipe structured task template, the state combination of each device under each batch task is set as a node set of the device conflict reasoning graph, where the node is uniquely identified by a device ID-batch ID combination, indicating the equipment-batch operation status;

[0098] S44. Based on the equipment combinations in the spatial arrangement mapping table and the recipe structured task templates that have a parallel relationship, identify task pairs with resource overlap or time conflict, and define them as conflict edge sets;

[0099] S45. Setting a conflict type identifier and a conflict rule expression for each conflicting edge, wherein the conflict rule expression is composed of recipe difference, physical space overlap, operation stage incompatibility, and device sharing status;

[0100] S46. Construct a complete device conflict reasoning graph, including a node set, an edge set, and corresponding conflict rule expressions.

[0101] By constructing an equipment conflict reasoning graph, not only are the process dependencies and spatial exclusion relationships between multiple devices explicitly modeled, but these constraints are also incorporated into the model scheduling process through graph structure encoding. Compared with traditional linear scheduling methods, this graph structure is more suitable for representing complex, multi-dimensional resource conflict scenarios and can effectively identify potential conflicts across batches, processes, and equipment. The introduction of conflict rule expressions makes conflict judgments reasonable and updatable, allowing for sustainable learning and evolution in actual production. This mechanism improves the scheduling system's adaptability to changes in equipment combinations and production line expansions, while reducing the probability of logical violations or equipment interlocking errors in scheduling outputs, significantly enhancing the versatility and security of the scheduling system.

[0102] In this embodiment, the S5 specifically includes:

[0103] S51, extracting the start and stop sequence, number information, task execution time window, and load distribution ratio of each equipment task in the preliminary scheduling plan, and parsing it into a scheduling path set;

[0104] S52, traversing the scheduling path set, and for each node combination involved in the scheduling path, searching for the corresponding node and adjacent conflicting edge in the device conflict reasoning graph;

[0105] S53. Based on the conflict rule expression recorded in the conflict edge, the current scheduling path is conflict-verified one by one, and conflict path branches that violate space occupancy, recipe dependency, and resource mutual exclusion conditions are identified;

[0106] S54. Mark the identified conflicting path branches and construct a scheduling mask tensor in the scheduling network. The scheduling mask tensor is used to mask the output channels corresponding to the conflicting paths in the forward propagation phase.

[0107] S55 , outputting the scheduling path filtered by the scheduling mask tensor as a revised preliminary scheduling plan, wherein the revised preliminary scheduling plan does not include any scheduling path that violates the process logic of the equipment conflict inference graph rule.

[0108] By embedding structured graph constraints in the scheduling network, the model can explicitly block non-compliant paths during the neural computation process, preventing invalid or conflicting scheduling schemes from participating in the final plan generation. Compared with the conventional soft constraint method guided by the loss function, this solution provides a strong constraint form, which significantly improves the model's responsiveness to actual process rules. By constructing a mask tensor to perform forward path suppression, the computing space is effectively compressed, the model's operating efficiency is improved, and all output schemes are guaranteed to comply with production logic. This mechanism demonstrates stronger robustness and generalization capabilities in scenarios with high-complexity task graphs, parallel equipment scheduling, and dynamic configuration changes, effectively supporting the complex requirements of synchronous execution of multiple batches and multi-stage tasks.

[0109] In this embodiment, the improved NGBoost model is improved by introducing a confidence factor estimation and dynamic distribution calibration mechanism when predicting the expected value and uncertainty score of scheduling behavior, specifically including:

[0110] Based on the historical disturbance data set, the prediction error amplitude of each scheduling behavior is statistically analyzed to construct a confidence factor for dynamically adjusting the output distribution width:

[0111] ;

[0112] in, Represents the time step The confidence factor when represents the exponential function, represents the adjustment parameter, Represents the time step The prediction error when represents the mean of the prediction error;

[0113] A dynamic distribution calibration mechanism is introduced during the natural gradient boosting process, and the following adjustments are performed each time the prediction distribution parameters are updated:

[0114] ;

[0115] in, represents the natural gradient after calibration, represents the standard natural gradient, represents the distribution calibration intensity coefficient, Represents the time step The true disturbance distribution when Represents the time step The predicted distribution when represents the prediction distribution parameter;

[0116] Decoding the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into a prediction expectation value and uncertainty score corresponding to each scheduling behavior, wherein the prediction expectation value represents the performance expectation of the current scheduling behavior under disturbance;

[0117] The confidence factor is used as a dynamic weight to scale the original variance value, and the adjusted variance value is used as the uncertainty score corresponding to the current scheduling behavior.

[0118] The improved NGBoost model implements uncertainty modeling for the distribution of scheduling behavior under disturbance conditions. By introducing a confidence factor estimation mechanism, the model dynamically adjusts the output confidence interval based on historical errors, automatically expanding the risk assessment range for disturbance-sensitive areas. A dynamic distribution calibration mechanism is also introduced to offset areas of skewed predictions, bringing predictions closer to the true distribution trend. This adaptive mechanism addresses the inability of previous models to explain prediction confidence and the lack of a basis for judging high-risk scheduling. This enables the system to not only accurately predict but also understand where it may have made mistakes, significantly enhancing the interpretability and robustness of scheduling behavior in highly uncertain environments.

[0119] In this embodiment, the S7 specifically includes:

[0120] S71. Set a scheduling risk identification threshold, mark scheduling behaviors with uncertainty scores exceeding the scheduling risk identification threshold as high-risk behaviors, and construct a high-risk behavior index list;

[0121] S72. Locate the context structure of the high-risk behavior in the revised preliminary scheduling plan based on the equipment number, task recipe requirements, and time window position of each high-risk behavior;

[0122] S73. Generate a scheduling correction candidate for each high-risk behavior without violating the current task recipe execution logic and equipment space conflict constraints;

[0123] S74: Classify all generated scheduling revision candidate solutions into a scheduling revision candidate set, and retain the mapping relationship in the revised preliminary scheduling solution.

[0124] The present invention no longer blindly trusts high-scoring results in the prediction output stage, but instead introduces a scheduling behavior risk identification and candidate solution generation process based on uncertainty scoring. This mechanism can proactively generate alternative solutions when the model cannot clearly evaluate the execution results of certain behaviors, thereby improving the system's fallback capability and safety redundancy in the face of dynamic fluctuation scenarios. The generated scheduling correction candidate set retains contextual logic and structural consistency, ensuring that alternative solutions can be seamlessly switched and quickly put into execution, greatly reducing the probability of production interruptions caused by high-risk predictions. This design not only enhances the robustness of the scheduling system, but also improves risk control capabilities and fault response efficiency in actual industrial deployments.

[0125] In this embodiment, the S8 specifically includes:

[0126] S81: Receive the generated scheduling correction candidate set, and perform structured analysis on the equipment start / stop sequence, time window, and load distribution information associated with each scheduling correction candidate solution;

[0127] S82. Calculate the capacity achievement rate of each scheduling modification candidate plan based on the standard capacity target of the current batch task, where the capacity achievement rate is the ratio of the actual execution capacity to the theoretical formula capacity;

[0128] S83. Estimating the energy consumption of the equipment operation plan corresponding to each candidate scheduling revision plan, and calculating the energy consumption change compared to the initial scheduling plan by combining the unit time operation power and load level;

[0129] S84. Extract the mean uncertainty score of each scheduling correction candidate under the prediction of the improved NGBoost model as a risk quantification indicator for measuring the disturbance adaptability;

[0130] S85. Construct a weighted comprehensive evaluation function based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and rank each scheduling correction candidate plan by score, where the weights in the weighted comprehensive evaluation function are set according to the process objectives;

[0131] S86. Screen the scheduling correction candidate with the highest score and issue it as the final scheduling plan for execution.

[0132] By introducing a multi-objective comprehensive evaluation mechanism, the screening of scheduling solutions no longer relies solely on a single performance indicator. Instead, it comprehensively considers three factors: capacity achievement rate, energy consumption changes, and disturbance adaptability, ensuring that the final output solution strikes a balance between production efficiency, cost control, and system stability. The use of a weighted comprehensive scoring function provides a flexible and adjustable policy control method, allowing companies to dynamically set optimization priorities based on actual production goals. At the same time, a structured ranking process is performed on the revised candidate solutions to ensure the transparency and explainability of the evaluation process and prevent the algorithmic decision-making process from becoming a black box. The overall solution significantly improves the quality of scheduling output and the controllability of industrial deployment, providing strong support for the stable operation of the system.

[0133] Example 1:

[0134] To verify the feasibility of the present invention in practice, the present invention was applied to the intelligent production line of a modern feed processing plant with an annual output of 480,000 tons. The production line covers six major work sections: crushing, batching, mixing, granulation, cooling, and packaging. A total of 24 automated equipment were deployed, with an average daily task batch of approximately 180 groups. There is a high degree of dependence and frequent switching between equipment, and scheduling management is complex. Traditional rule engine methods have difficulty in dealing with dynamic disturbances and equipment conflicts.

[0135] In this embodiment, the original system has long relied on fixed priorities and equipment pairing rules for production scheduling. When the granulation equipment experiences operational fluctuations or the mixer switches recipes with delays, the downstream equipment fails to adjust in time, resulting in multiple chain blockages and batch misalignments. Especially in the case of multiple batches running in parallel, there is physical overlap in the spatial positions of certain equipment. For example, if two granulators sharing a belt are scheduled at the same time, vibration conflicts and abnormal temperature rise will occur. Traditional systems lack an effective early warning mechanism for this, and manual intervention by dispatchers is often required. To solve the above problems, the disturbance-resistant residual fusion network, equipment conflict reasoning graph, and uncertainty-driven scheduling correction mechanism in the present invention are fully deployed in the plant's intelligent scheduling module.

[0136] First, through structured collection of all equipment operation logs, fault records, energy consumption data, and recipe task assignment information over a three-month period, approximately 312,000 state sequence samples were constructed, along with a set of perturbation samples encompassing scenarios such as startup delays, operational interruptions, and recipe switching failures. The system, through joint training of main and side paths, models the behavior patterns of each device during different operating periods, demonstrating excellent comparative and perturbation memory capabilities. Subsequently, based on the plant's equipment spatial layout and the inter-recipe equipment dependency chains, an equipment conflict inference graph with 2,412 equipment-batch nodes and 8,763 conflict edges was generated. This graph was embedded into the scheduling network as a legitimacy filter for each scheduling path. The scheduling output process incorporates an improved NGBoost model to assign expected performance and risk scores to each behavior, automatically identifying unstable behaviors and generating candidate corrections.

[0137] During the first month of deployment, the system processed 4,921 scheduling tasks, identified 784 high-risk scheduling behaviors, and automatically generated and replaced correction plans 412 times. Each correction improved task execution stability by an average of approximately 23.7%. During a high-load test week (daily task batches >200), the overall scheduling failure rate dropped from 4.6% to 1.1%, the capacity achievement rate increased from 92.3% to 97.4%, and energy consumption per unit of output decreased by 8.2%, significantly improving the efficiency of coordinated scheduling between equipment. Especially during equipment maintenance or sudden anomalies, the new system can complete plan correction and rescheduling within 2 seconds, approximately 11 times faster than the manual scheduling response time of the original system, effectively ensuring the continuity and safety of the production line.

[0138] The present invention showed significant differences in actual production line operation before and after deployment. Based on statistics of 30 consecutive days of operation, the total number of scheduling behaviors was 4921, the system identified and processed approximately 7300 disturbance samples, and replaced scheduling behaviors 412 times. The scheduling failure rate dropped from 4.6% of the original system to 1.1%; the average capacity achievement rate increased from 92.3% to 97.4%; the energy consumption per ton of feed dropped from 71.6 kWh to 65.7 kWh; the average delay in task switching was shortened from 18.4 seconds to 7.1 seconds; and the frequency of manual intervention was reduced by approximately 69.2%. Among them, the prediction module that introduced the confidence factor and dynamic distribution calibration mechanism had an accuracy rate of 88.7% in identifying high-risk behaviors. Overall, it is shown that the scheduling optimization method provided by the present invention has strong robustness, flexible adaptability and practical implementation effects, can effectively replace the traditional static scheduling system, and significantly improve the automation operation efficiency and resource utilization of the feed processing industry.

[0139] 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 collaborative scheduling and operation optimization of feed production equipment based on deep learning, characterized in that: The steps include: S1. Collect the operating status data of each device in the feed production process and the corresponding task recipe information to form an equipment status sequence data set; S2. Construct a disturbance sample set, wherein the disturbance sample set includes startup delay, operation interruption and recipe switching anomaly, and is structured into a disturbance sequence data set; S3. Inputting the device state sequence data set and the disturbance sequence data set into a disturbance-resistant residual fusion network, wherein the disturbance-resistant residual fusion network includes a main path and a side path, and dynamically weighted fusion is performed through a residual gated fusion unit to generate a preliminary scheduling plan; S4. Constructing an equipment conflict reasoning graph based on the task recipe information and the spatial arrangement relationship, where nodes represent equipment-batch operation states, edges represent non-parallel conflict conditions and are accompanied by conflict rule expressions; S5. Embed the equipment conflict reasoning diagram as a structural constraint into the scheduling network, shield the scheduling path that violates the process logic, and revise the preliminary scheduling plan; S6. Input the revised preliminary scheduling plan into the improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism to output the predicted expected value and uncertainty score of each scheduling behavior; The improvement of the improved NGBoost model lies in the introduction of confidence factor estimation and dynamic distribution calibration mechanism when predicting the expected value and uncertainty score of scheduling behavior, specifically including: Based on the historical disturbance data set, the prediction error amplitude of each scheduling behavior is statistically analyzed to construct a confidence factor for dynamically adjusting the output distribution width: ； in, Represents the time step The confidence factor when represents the exponential function, represents the adjustment parameter, Represents the time step The prediction error when represents the mean of the prediction error; A dynamic distribution calibration mechanism is introduced during the natural gradient boosting process, and the following adjustments are performed each time the prediction distribution parameters are updated: ； in, represents the natural gradient after calibration, represents the standard natural gradient, represents the distribution calibration intensity coefficient, Represents the time step The true disturbance distribution when Represents the time step The predicted distribution when represents the prediction distribution parameter; Decoding the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into a prediction expectation value and uncertainty score corresponding to each scheduling behavior, wherein the prediction expectation value represents the performance expectation of the current scheduling behavior under disturbance; Using the confidence factor as a dynamic weight, the original variance value is scaled and adjusted, and the adjusted variance value is used as the uncertainty score corresponding to the current scheduling behavior; S7. Identify high-risk behaviors based on the uncertainty score and generate a set of scheduling modification candidates; S8. Perform a multi-objective comprehensive evaluation on the scheduling correction candidate set, calculate scores based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and select the scheduling correction candidate with the highest score to be issued as the final scheduling plan for execution.

2. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The operating status data includes the current production capacity of the equipment, energy consumption per unit time, operating load rate, operating temperature, vibration amplitude, task execution time and topological connection information between devices.

3. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The S3 specifically includes: S31. Perform time dimension alignment and feature channel normalization processing on the device state sequence dataset and the disturbance sequence dataset, and construct a main path input tensor and a side path input tensor corresponding to the time axis; S32. Input the main path input tensor into a main path network composed of a stack of one-dimensional convolutional layers and gated recurrent units, extract short-term dependency features through the one-dimensional convolutional layers, and extract running trends across time steps through the gated recurrent units to obtain a main path feature tensor; S33, inputting the side path input tensor into a side path network composed of a convolutional structure, a multi-scale residual connection, and an attention mechanism, extracting dynamic impact features of the disturbance at different time scales, and outputting a disturbance path feature tensor; S34, splicing the main path feature tensor and the perturbation path feature tensor in the feature dimension, and inputting the concatenated features into a residual gated fusion unit, wherein the residual gated fusion unit performs a channel-by-channel weighted fusion on the main path feature tensor and the perturbation path feature tensor to obtain a fused feature tensor; S35, performing channel compression on the fused feature tensor and inputting the result into a structured decoder, wherein the structured decoder includes a device number encoding layer, a time window encoding layer, and a load distribution encoding layer; The device number encoding layer performs a position mapping operation, embedding and encoding the spatial channels in the fused feature tensor and the device topology index, generating a number vector representation corresponding to each device, which constitutes the device identification dimension in the scheduling instruction; The time window encoding layer performs convolution expansion and sliding window offset processing on the time step index in the fused feature tensor, extracts the corresponding starting time segment and duration period in each scheduling path, and converts them into an integer interval identification vector; The load distribution encoding layer performs a maximum-mean two-pass normalization operation based on the fused feature tensor to form a load distribution ratio vector corresponding to the current task of each device; S36. Assemble the output results of the structured decoder into a preliminary scheduling plan, wherein the preliminary scheduling plan includes the start and stop sequence, numbering information, task execution time window and load distribution ratio value of each device.

4. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The S4 specifically includes: S41. Analyze task recipe information, extract the equipment usage sequence, process stage requirements, and equipment occupancy time period corresponding to each recipe, and form a recipe structured task template; S42. Collect spatial layout data of the feed workshop, including equipment number, installation location, physical occupied area, and relationship between process channels, and establish a spatial layout mapping table; S43. Based on the equipment occupancy of each batch task in the recipe structured task template, the state combination of each device under each batch task is set as a node set of the device conflict reasoning graph, where the node is uniquely identified by a device ID-batch ID combination, indicating the equipment-batch operation status; S44. Based on the equipment combinations in the spatial arrangement mapping table and the recipe structured task templates that have a parallel relationship, identify task pairs with resource overlap or time conflict, and define them as conflict edge sets; S45. Setting a conflict type identifier and a conflict rule expression for each conflicting edge, wherein the conflict rule expression is composed of recipe difference, physical space overlap, operation stage incompatibility, and device sharing status; S46. Construct a complete device conflict reasoning graph, including a node set, an edge set, and corresponding conflict rule expressions.

5. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The S5 specifically includes: S51, extracting the start and stop sequence, number information, task execution time window, and load distribution ratio of each equipment task in the preliminary scheduling plan, and parsing it into a scheduling path set; S52, traversing the scheduling path set, and for each node combination involved in the scheduling path, searching for the corresponding node and adjacent conflicting edge in the device conflict reasoning graph; S53. Based on the conflict rule expression recorded in the conflict edge, the current scheduling path is verified for conflicts one by one, and conflict path branches that violate space occupancy, recipe dependency, and resource mutual exclusion conditions are identified; S54. Mark the identified conflicting path branches and construct a scheduling mask tensor in the scheduling network. The scheduling mask tensor is used to mask the output channels corresponding to the conflicting paths in the forward propagation phase. S55 , outputting the scheduling path after filtering the scheduling mask tensor as a revised preliminary scheduling plan, wherein the revised preliminary scheduling plan does not include any scheduling path that violates the process logic of the equipment conflict inference graph rule.

6. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The S7 specifically includes: S71. Set a scheduling risk identification threshold, mark scheduling behaviors with uncertainty scores exceeding the scheduling risk identification threshold as high-risk behaviors, and construct a high-risk behavior index list; S72. Locate the context structure of the high-risk behavior in the revised preliminary scheduling plan based on the equipment number, task recipe requirements, and time window position of each high-risk behavior; S73. Generate a scheduling correction candidate for each high-risk behavior without violating the current task recipe execution logic and equipment space conflict constraints; S74: Classify all generated scheduling revision candidate solutions into a scheduling revision candidate set, and retain the mapping relationship in the revised preliminary scheduling solution.

7. The method for collaborative scheduling and operation optimization of feed production equipment based on deep learning according to claim 1, characterized in that: The S8 specifically includes: S81: Receive the generated scheduling correction candidate set, and perform structured analysis on the equipment start / stop sequence, time window, and load distribution information associated with each scheduling correction candidate solution; S82. Calculate the capacity achievement rate of each scheduling modification candidate plan based on the standard capacity target of the current batch task, where the capacity achievement rate is the ratio of the actual execution capacity to the theoretical formula capacity; S83. Estimating the energy consumption of the equipment operation plan corresponding to each candidate scheduling revision plan, and calculating the energy consumption change compared to the initial scheduling plan by combining the unit time operation power and load level; S84. Extract the mean uncertainty score of each scheduling correction candidate under the prediction of the improved NGBoost model as a risk quantification indicator for measuring the disturbance adaptability; S85. Construct a weighted comprehensive evaluation function based on the capacity achievement rate, energy consumption change, and disturbance adaptability, and rank each scheduling correction candidate plan by score, where the weights in the weighted comprehensive evaluation function are set according to the process objectives; S86. Screen the scheduling correction candidate with the highest score and issue it as the final scheduling plan for execution.

Images