Feed production equipment collaborative scheduling and operation optimization method based on deep learning

Abstract

The invention discloses a feed production equipment collaborative scheduling and operation optimization method based on deep learning. The method comprises the following steps: S1, constructing an equipment state sequence data set; s2, constructing a disturbance sequence data set; s3, inputting a disturbance resistance residual fusion network to generate a preliminary scheduling scheme; s4, constructing an equipment conflict reasoning graph; s5, embedding the equipment conflict reasoning graph into the scheduling network, and correcting the preliminary scheduling scheme; s6, inputting the corrected preliminary scheduling scheme into an improved NGBoost model, introducing a confidence factor estimation and dynamic distribution calibration mechanism, and outputting a predicted expected value and an uncertainty score of each scheduling behavior; s7, identifying a high-risk behavior according to the uncertainty score, and generating a scheduling correction candidate set; and S8, performing multi-target comprehensive evaluation on the scheduling correction candidate set, and screening the candidate set with the highest score as a final scheduling plan to be issued and executed. According to the method, deep learning and an improved NGBoost model are combined, and intelligent scheduling optimization of feed equipment is realized.

Description

Technical Field

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

[0002] In modern feed manufacturing, equipment collaborative scheduling and operation optimization have become the key links restricting the production efficiency, energy utilization rate and equipment safe operation level of enterprises. Feed production involves multiple process stages such as raw material crushing, batching, mixing, granulation, cooling, and packaging, and each stage relies on the collaborative operation of multiple different types of equipment. With the continuous improvement of factory automation level, more and more enterprises adopt 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 ability, and poor adaptability to real-time status and disturbances, and it is 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 production capacity models or standard time windows. Such methods cannot fully perceive the uncertain factors in the actual operation state, such as equipment performance decay, switching delay, sudden task insertion or raw material quality fluctuation, etc., resulting in the scheduling results being extremely sensitive to minor disturbances. Once a device experiences a slight delay or operation fluctuation, the system lacks an effective mechanism for adaptive adjustment, and often requires manual intervention to re-intervene or re-arrange tasks as a whole, resulting in frequent plan failures. In addition, most systems do not establish a unified dynamic disturbance modeling mechanism, nor can they structure historical abnormal behaviors for prediction and optimization. The scheduling strategy mainly stays at the rule execution level and cannot form a closed-loop scheduling system with learning ability.

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

[0005] The prior art has not fully considered the problem of multi-objective scheduling optimization. Most systems only consider the goal of maximizing production capacity during the optimization process, lacking multi-dimensional considerations of energy consumption costs, resource conflict levels, or task robustness. As a result, the scheduling plans finally output by the system are often difficult to maximize benefits in actual scenarios. At the same time, current systems generally lack an effective scheduling output evaluation mechanism and are unable to conduct structured quantitative comparisons and screenings for multiple generated scheduling alternative plans. They only rely on the scoring function within the model for decision-making, ignoring the actual differences between candidate plans in terms of disturbance adaptability, load balancing, etc.

[0006] Therefore, how to provide a collaborative scheduling and operation optimization method for feed production equipment based on deep learning 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 collaborative scheduling and operation optimization method for feed production equipment based on deep learning. The present invention combines deep learning modeling with a disturbance perception inference mechanism to construct a disturbance-resistant residual fusion network and an equipment conflict inference graph, realizing intelligent optimization of the feed production equipment scheduling process, having advantages such as high robustness to disturbances, strong adaptive correction ability, high production capacity achievement rate, and controllable energy consumption, effectively solving the problems of slow response to abnormal fluctuations, insufficient conflict logic processing, and single optimization goal in the existing scheduling system, and significantly improving the intelligent level and industrial practical value of the system.

[0008] The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to an embodiment of the present invention includes the following steps: S1. Collect the operation status data of each device and the corresponding task formula information during the feed production process to form a device status sequence data set; S2. Construct a disturbance sample set, where the disturbance sample set includes start-up delays, operation interruptions, and abnormal formula switches, and is structured into a disturbance sequence data set; S3. Input the device status sequence data set and the disturbance sequence data set into a disturbance-resistant residual fusion network. The disturbance-resistant residual fusion network includes a main path and a side path, and dynamically weights and fuses through a residual gate fusion unit to generate a preliminary scheduling plan; S4. Construct an equipment conflict inference graph based on the task formula information and spatial layout relationship. The nodes represent the device-batch operation status, and the edges represent non-parallel conflict conditions and are accompanied by conflict rule expressions; S5. Embed the equipment conflict inference graph as a structural constraint into the scheduling network to block the scheduling paths that violate the process logic and correct the preliminary scheduling plan; S6. Input the corrected preliminary scheduling plan into the improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism, and outputs the predicted expected value and uncertainty score for each scheduling behavior. S7. Identify high-risk behaviors based on the uncertainty scores and generate a candidate set for scheduling correction. S8. Conduct a multi-objective comprehensive evaluation on the candidate set for scheduling correction, calculate scores based on the production capacity achievement rate, energy consumption change, and disturbance adaptability, and select the scheduling correction candidate plan with the highest score as the final scheduling plan for execution.

[0009] Optionally, the operating state data includes the current production capacity of the device, energy consumption per unit time, operating load rate, working temperature, vibration amplitude, task execution time, and device topology connection information.

[0010] Optionally, the specific steps of S3 are as follows: S31. Perform time dimension alignment and feature channel normalization processing on the device state sequence dataset and the disturbance sequence dataset to 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 the main path network composed of a one-dimensional convolutional layer and a gated recurrent unit. Extract short-term dependence features through the one-dimensional convolutional layer and extract the operating trend across time steps through the gated recurrent unit to obtain the main path feature tensor. S33. Input the side path input tensor into the side path network composed of a convolutional structure, multi-scale residual connections, and an attention mechanism to extract the dynamic influence features of the disturbance at different time scales and output the disturbance path feature tensor. S34. Concatenate the main path feature tensor and the disturbance path feature tensor in the feature dimension and input them into the residual gated fusion unit. The residual gated fusion unit performs channel-wise weighted fusion on the main path feature tensor and the disturbance path feature tensor to obtain the fused feature tensor. S35. Compress the channels of the fused feature tensor and input it into the structured decoder, which includes a device number encoding layer, a time window encoding layer, and a load allocation encoding layer. The device number encoding layer performs a position mapping operation, embeds and encodes the spatial channels and device topology indices in the fused feature tensor, generates a vector representation of the number corresponding to each device, and constructs the device identification dimension in the scheduling instruction. The time window encoding layer performs convolutional expansion and sliding window offset processing on the time step indices in the fused feature tensor, extracts the corresponding start time segment and duration period in each scheduling path, and converts them into an integer interval identification vector. The load distribution coding layer performs a max-mean dual-path normalization operation based on the fusion feature tensor to form a load distribution ratio vector corresponding to the current task of each device; S36. Assemble the output result of the structural decoder into a preliminary scheduling plan, where the preliminary scheduling plan includes the start-stop sequence, number information, task execution time window, and load distribution ratio value of each device.

[0011] Optionally, step S4 specifically includes: S41. Analyze the task recipe information, extract the device usage sequence, process stage requirements, and device occupancy time period corresponding to each recipe, and form a recipe-structured task template; S42. Collect the spatial layout data of the feed workshop, including device numbers, installation locations, physical occupancy areas, and process channel relationships, and establish a spatial layout mapping table; S43. According to the occupancy of each device by each batch of tasks in the recipe-structured task template, set the state combination of each device under each batch of tasks as the node set of the device conflict inference graph. The nodes are uniquely identified by the device identifier-batch identifier combination, representing the device-batch operation state; S44. According to the spatial layout mapping table and the device combinations with parallel relationships in the recipe-structured task template, identify task pairs with resource overlap or time conflict, and define them as the conflict edge set; S45. Set a conflict type identifier and a conflict rule expression for each conflict edge. The conflict rule expression is composed of recipe differences, physical space overlap, operation stage incompatibility, and device sharing status; S46. Construct a complete device conflict inference graph, including the node set, edge set, and corresponding conflict rule expressions.

[0012] Optionally, step S5 specifically includes: S51. Extract the start-stop sequence, number information, task execution time window, and load distribution ratio value of each device task in the preliminary scheduling plan, and parse them into a scheduling path set; S52. Traverse the scheduling path set, and for each node combination involved in each scheduling path, find the corresponding nodes and adjacent conflict edges in the device conflict inference graph; S53. According to the conflict rule expressions recorded in the conflict edges, perform conflict verification on each current scheduling path item by item, and identify the conflict path branches that violate the spatial occupancy, recipe dependency, and resource mutual exclusion conditions; S54. Mark the identified conflict 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 conflict paths in the forward propagation stage; S55. Output the scheduling path filtered by the scheduling mask tensor as the revised preliminary scheduling plan, where the revised preliminary scheduling plan does not contain any scheduling path that violates the process logic of the device conflict inference graph.

[0013] Optionally, the improvement of the improved NGBoost model lies in introducing a confidence factor estimation and dynamic distribution calibration mechanism when predicting the expected value and uncertainty score of the scheduling behavior, specifically including: Based on the historical perturbation dataset, statistically calculate the prediction error amplitude of each scheduling behavior, and construct a confidence factor for dynamically adjusting the output distribution width: ; where, represents the confidence factor at time step , represents the exponential function, represents the adjustment parameter, represents the prediction error at time step , represents the mean prediction error; Introduce a dynamic distribution calibration mechanism in the natural gradient boosting process, and perform the following adjustments when updating the prediction distribution parameters each time: ; where, represents the calibrated natural gradient, represents the standard natural gradient, represents the distribution calibration intensity coefficient, represents the true perturbation distribution at time step , represents the prediction distribution at time step , represents the prediction distribution parameter; Decode the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into the predicted expected value and uncertainty score corresponding to each scheduling behavior, where the predicted expected value represents the performance expectation of the current scheduling behavior under perturbation; Use the confidence factor as a dynamic weight to scale and adjust the original variance value, and use the adjusted variance value as the uncertainty score corresponding to the current scheduling behavior.

[0014] Optionally, the specific content of S7 includes: S71. Set a scheduling risk identification threshold, mark the 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 high-risk behavior in the context structure of the revised preliminary scheduling plan according to the equipment number, task recipe requirements, and time window position of each high-risk behavior; S73. Generate a scheduling correction candidate plan for each high-risk behavior on the premise of not violating the current task recipe execution logic and equipment space conflict constraints; S74. Classify all the generated scheduling correction candidate plans into a scheduling correction candidate set and retain the mapping relationship in the revised preliminary scheduling plan.

[0015] Optionally, the specific steps of S8 are as follows: S81. Receive the generated scheduling correction candidate set and perform structured parsing on the equipment start-stop sequence, time window, and load distribution information associated with each scheduling correction candidate plan; S82. Calculate the production capacity achievement rate of each scheduling correction candidate plan according to the standard production capacity target of the current batch of tasks, where the production capacity achievement rate is the ratio of the actual executed production capacity to the theoretical recipe production capacity; S83. Estimate the energy consumption of the equipment operation plan corresponding to each scheduling correction candidate plan, and calculate the energy consumption change compared with the initial scheduling plan by combining the operating power per unit time and the load level; S84. Extract the average uncertainty score of each scheduling correction candidate plan under the improved NGBoost model prediction as a risk quantification index to measure the disturbance adaptability; S85. Construct a weighted comprehensive evaluation function based on the production capacity achievement rate, energy consumption change, and disturbance adaptability, and rank the scores of each scheduling correction candidate plan. The weights in the weighted comprehensive evaluation function are set according to the process objectives; S86. Screen the scheduling correction candidate plan with the highest score as the final scheduling plan for execution.

[0016] The beneficial effects of the present invention are as follows: First of all, by collecting equipment operation status data and recipe task information, the present invention constructs a multi-dimensional equipment status sequence and disturbance sample data set, and introduces a disturbance-resistant residual fusion network to jointly model the equipment status and historical disturbance behaviors, effectively enhancing the model's perception ability and robustness to minor fluctuations. Through the parallel residual structure design of the main path and the side path, and with the help of the residual gating fusion unit to achieve dynamic weighting of information, the core problems of instability and misjudgment of traditional models in the face of disturbance situations such as operation interruptions and switching anomalies are solved.

[0017] Secondly, the present invention introduces a device conflict inference graph structure, taking the operating status at the device-batch level as nodes, and jointly inferring the parallel conflict edges and their logical conditions between devices through scheduling logs and process rules, so as to realize the formal modeling of potential resource conflicts and process sequence conflicts between devices. After embedding this conflict inference graph as a structural constraint within the scheduling network, the model can automatically block scheduling paths that do not meet the production logic, ensuring that the output scheduling plan conforms to the actual operating rules and spatial layout limitations between devices, and significantly improving the executability of the plan.

[0018] In addition, the output of the scheduling behavior of the present invention is no longer limited to a single result, but realizes the distributed prediction of each scheduling behavior by introducing an improved NGBoost model, and outputs a probabilistic result including the predicted expected value and the uncertainty score. In particular, the improved part realizes the adaptive adjustment of the prediction credibility for regions with different disturbance degrees through the confidence factor estimation and dynamic distribution calibration mechanism, so as to accurately identify high-risk scheduling behaviors. Based on the uncertainty score, the system further constructs a scheduling correction candidate set, and introduces a multi-objective comprehensive evaluation mechanism to quantitatively score and rank all correction plans from three aspects: production capacity achievement rate, energy consumption change amount, and disturbance adaptability, and finally selects the optimal plan for execution, ensuring a dynamic balance between the efficiency and stability of the scheduling result.

[0019] Through the above technical path, the present invention not only breaks through the limitations of traditional scheduling algorithms in dealing with disturbances and conflicts at the modeling level, but also realizes the double improvement of interpretability and robustness at the output strategy level, providing an integrated intelligent scheduling solution for complex feed production systems with learning ability, structural constraints, self-correction, and multi-objective coordinated optimization capabilities. This method significantly reduces the scheduling failure rate, improves the production capacity stability and resource utilization efficiency, and has broad industrial application value and promotion prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings are used to provide a further understanding of the present invention, and constitute a part of the specification, and are used to explain the present invention together with the embodiments of the present invention, and do not constitute a limitation to the present invention. In the drawings: Figure 1 is the overall flowchart of the collaborative scheduling and operation optimization method for feed production equipment based on deep learning proposed by the present invention; Figure 2 is the structural schematic diagram of the disturbance-resistant residual fusion network of the collaborative scheduling and operation optimization method for feed production equipment based on deep learning proposed by the present invention; Figure 3Flowchart of the confidence factor estimation and dynamic distribution calibration mechanism in the improved NGBoost model for predicting the expected value and uncertainty score of scheduling behavior in the collaborative scheduling and operation optimization method of feed production equipment based on deep learning proposed by the present invention. Detailed implementation manners

[0021] The present invention will be further described in detail below with reference to the accompanying 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.

[0022] Refer to Figures 1-3 , the collaborative scheduling and operation optimization method of feed production equipment based on deep learning includes the following steps: S1. Collect the operation status data of each device and the corresponding task formula information in the feed production process to form a device status sequence data set; S2. Construct a perturbation sample set, which includes start-up delay, operation interruption, and abnormal formula switching, and is structured into a perturbation sequence data set; S3. Input the device status sequence data set and the perturbation sequence data set into a perturbation-resistant residual fusion network, which includes a main path and a bypass path, and dynamically weights and fuses through a residual gating fusion unit to generate a preliminary scheduling plan; S4. Construct a device conflict inference graph based on the task formula information and the spatial layout relationship, where the nodes represent the device-batch operation status, and the edges represent the non-parallel conflict conditions and are attached with conflict rule expressions; S5. Embed the device conflict inference graph as a structural constraint into the scheduling network, mask the scheduling paths that violate the process logic, and correct the preliminary scheduling plan; S6. Input the corrected preliminary scheduling plan into an improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism, and outputs the predicted expected value and uncertainty score of each scheduling behavior; S7. Identify high-risk behaviors according to the uncertainty score to generate a scheduling correction candidate set; S8. Conduct a multi-objective comprehensive evaluation on the scheduling correction candidate set, calculate the scores according to the production capacity achievement rate, energy consumption change amount, and perturbation adaptability, and select the scheduling correction candidate plan with the highest score as the final scheduling plan for distribution and execution.

[0023] This method realizes the collaborative optimization control of feed production equipment under complex operating scenarios by constructing an intelligent scheduling process that integrates multi-source data, disturbance modeling, structural constraints, and uncertainty prediction. The disturbance-resistant residual fusion network is used to enhance the robustness of the model to operating anomalies. The structural elimination of spatial and process conflicts is achieved by combining the equipment conflict inference graph. The risk scoring and adaptive correction of scheduling behaviors are realized by introducing an improved NGBoost model. The overall solution has characteristics such as automatic generation, dynamic adjustment, risk perception, and optimal screening. In the face of a production environment with frequent equipment state changes and complex task switching, it can significantly improve system stability, task completion rate, and energy efficiency level, meeting the multi-dimensional performance requirements of high-intensity industrial production for intelligent scheduling systems.

[0024] In this embodiment, the operating state data includes the current production capacity of the equipment, energy consumption per unit time, operating load rate, working temperature, vibration amplitude, task execution time, and topological connection information between devices.

[0025] In this embodiment, the specific steps of S3 are as follows: S31. Perform time dimension alignment and feature channel normalization processing on the equipment state sequence dataset and the disturbance sequence dataset to 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 the main path network composed of a one-dimensional convolutional layer and a gated recurrent unit. Extract short-term dependence features through the one-dimensional convolutional layer and extract the operating trend across time steps through the gated recurrent unit to obtain the main path feature tensor. S33. Input the side path input tensor into the side path network composed of a convolutional structure, multi-scale residual connections, and an attention mechanism to extract the dynamic influence features of disturbances at different time scales and output the disturbance path feature tensor. S34. Concatenate the main path feature tensor and the disturbance path feature tensor in the feature dimension and input them into the residual gated fusion unit. The residual gated fusion unit performs channel-wise weighted fusion on the main path feature tensor and the disturbance path feature tensor to obtain the fused feature tensor. S35. Compress the channels of the fused feature tensor and input it into the structure decoder, which includes an equipment number encoding layer, a time window encoding layer, and a load allocation encoding layer. The equipment number encoding layer performs a position mapping operation, embeds and encodes the spatial channels and equipment topology indices in the fused feature tensor, generates a numbered vector representation corresponding to each device, and constitutes the device identification dimension in the scheduling instruction. The time window encoding layer performs convolutional expansion and sliding window offset processing on the time step indices in the fused feature tensor, extracts the corresponding start time segments and durations in each scheduling path, and converts them into integer interval identification vectors; The load distribution encoding layer performs max-mean dual-path normalization operations based on the fused feature tensor to form a load distribution ratio vector corresponding to the current tasks of each device; S36. Assemble the output result of the structural decoder into a preliminary scheduling plan, where the preliminary scheduling plan includes the start and stop order, number information, task execution time window, and load distribution ratio value of each device.

[0026] Through the division of labor between the main path and the bypass path, the system can not only extract the normal operation mode, but also effectively identify disturbance features and dynamically fuse them, avoiding the failure of the scheduling plan due to minor fluctuations. In particular, the setting of the residual gating fusion unit enables the model to flexibly adjust the information flow intensity at different operation stages and dynamically optimize the output direction of the scheduling plan. The design of the structural decoder further strengthens the structural rationality and execution consistency of the output, ensuring the device matching degree, time consistency, and load balance in the actual deployment of the scheduling plan, and providing accurate and stable basic information output for the final scheduling plan. In this embodiment, the specific steps of S4 include: S41. Analyze the task recipe information, extract the device usage order, process stage requirements, and device occupation time periods corresponding to each recipe, and form a recipe-structured task template; S42. Collect the spatial layout data of the feed mill, including device numbers, installation locations, physical occupation areas, and process channel relationships, and establish a spatial layout mapping table; S43. According to the occupation of each device by each batch of tasks in the recipe-structured task template, set the state combination of each device under each batch of tasks as the node set of the device conflict inference graph. The nodes are uniquely identified by the device identifier-batch identifier combination, representing the device-batch operation state; S44. According to the spatial layout mapping table and the device combinations with parallel relationships in the recipe-structured task template, identify task pairs with resource overlap or time conflict, and define them as the conflict edge set; S45. Set a conflict type identifier and a conflict rule expression for each conflict edge. The conflict rule expression consists of recipe differences, physical space overlap, operation stage incompatibility, and device sharing status; S46. Construct a complete device conflict inference graph, including the node set, edge set, and corresponding conflict rule expressions.

[0027] 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, and can continuously learn and evolve in actual production. This mechanism improves the scheduling system's ability to adapt 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 safety of the scheduling system.

[0028] In this implementation manner, S5 specifically includes: S51, extracting the start and stop sequence, number information, task execution time window and load distribution ratio value of each equipment task in the preliminary scheduling plan, and parsing them into a scheduling path set; S52, traversing the scheduling path set, and for each node combination involved in the scheduling path, searching for corresponding nodes and adjacent conflicting edges in the device conflict reasoning graph; S53, based on the conflict rule expression recorded in the conflict edge, verify the conflicts one by one on the current scheduling path, and identify the conflict path branches that violate the space occupancy, recipe dependency and resource mutual exclusion conditions; S54, marking the identified conflict path branches, and constructing a scheduling mask tensor in the scheduling network, wherein the scheduling mask tensor is used to shield the output channel corresponding to the conflict path in the forward propagation stage; 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 reasoning graph rule.

[0029] By embedding structured graph constraints in the scheduling network, the model can explicitly shield non-compliant paths during the neural computing process, preventing invalid or conflicting scheduling schemes from participating in the final plan generation. Compared with the conventional soft constraint method based on loss function guidance, this scheme 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 operation efficiency is improved, and all output schemes are guaranteed to comply with production logic. This mechanism shows stronger robustness and generalization capabilities in scenarios with high-complexity task graphs, equipment parallel scheduling, and dynamic configuration changes, and effectively supports the complex requirements of synchronous execution of multi-batch and multi-stage tasks.

[0030] In this embodiment, the improvement of the improved NGBoost model lies in introducing a confidence factor estimation and dynamic distribution calibration mechanism when predicting the expected value and uncertainty score of scheduling behaviors, specifically including: Based on the historical perturbation dataset, statistically calculate the prediction error amplitude of each scheduling behavior, and construct a confidence factor for dynamically adjusting the output distribution width: ; Among them, represents the confidence factor at time step , represents the exponential function, represents the adjustment parameter, represents the time step when the prediction error, represents the mean prediction error; Introduce a dynamic distribution calibration mechanism during the natural gradient boosting process, and perform the following adjustments when updating the prediction distribution parameters each time: ; Among them, represents the calibrated natural gradient, represents the standard natural gradient, represents the distribution calibration intensity coefficient, represents the time step when the true perturbation distribution, represents the time step when the prediction distribution, represents the prediction distribution parameter; Decode the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into the predicted expected value and uncertainty score corresponding to each scheduling behavior. The predicted expected value represents the performance expectation of the current scheduling behavior under perturbations; Use the confidence factor as a dynamic weight to scale and adjust the original variance value, and use the adjusted variance value as the uncertainty score corresponding to the current scheduling behavior.

[0031] The improved NGBoost model realizes the uncertainty modeling of the scheduling behavior distribution under perturbations. By introducing a confidence factor estimation mechanism, the model can dynamically adjust the output confidence interval according to historical errors, and automatically expand the risk assessment range for sensitive areas of perturbations; at the same time, by introducing a dynamic distribution calibration mechanism, it corrects the offset of the prediction skewed area, making the prediction closer to the true distribution trend. This adaptive mechanism solves the problems that previous models cannot explain the prediction confidence and lack a judgment basis for high-risk scheduling, enabling the system to not only predict accurately but also know where it may be wrong, greatly enhancing the interpretability and decision-making robustness of scheduling behaviors in a high-uncertainty environment.

[0032] In this embodiment, 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 high-risk behaviors in the revised preliminary scheduling plan according to the equipment numbers, task recipe requirements, and time window positions of each high-risk behavior; S73. Generate scheduling correction candidate plans for each high-risk behavior on the premise of not violating the current task recipe execution logic and equipment space conflict constraints; S74. Classify all the generated scheduling correction candidate plans into a scheduling correction candidate set and retain the mapping relationship in the revised preliminary scheduling plan.

[0033] In the prediction output stage of the present invention, it no longer blindly trusts high-score results, but introduces a scheduling behavior risk identification and candidate plan generation process based on uncertainty scores. This mechanism can actively generate alternative plans when the model cannot clearly evaluate the execution results of certain behaviors, improving the fallback ability and safety redundancy of the system in the face of dynamic fluctuation scenarios. The generated scheduling correction candidate set retains context logic and structural consistency, ensuring that alternative plans can be seamlessly switched and quickly put into execution, greatly reducing the probability of production interruption caused by high-risk predictions. This design not only enhances the robustness of the scheduling system but also improves the risk control ability and fault response efficiency in industrial actual deployments.

[0034] In this embodiment, S8 specifically includes: S81. Receive the generated scheduling correction candidate set, and structurally analyze the equipment start-stop sequence, time window, and load distribution information associated with each scheduling correction candidate plan; S82. Calculate the production capacity achievement rate of each scheduling correction candidate plan according to the standard production capacity target of the current batch of tasks, where the production capacity achievement rate is the ratio of the actual executed production capacity to the theoretical recipe production capacity; S83. Estimate the energy consumption of the equipment operation plan corresponding to each scheduling correction candidate plan, and calculate the energy consumption change amount compared with the initial scheduling plan in combination with the unit time operation power and load level; S84. Extract the mean value of the uncertainty scores of each scheduling correction candidate plan under the improved NGBoost model prediction as a risk quantification index for measuring disturbance adaptability; S85. Construct a weighted comprehensive evaluation function based on the production capacity achievement rate, energy consumption change amount, and disturbance adaptability, and rank the scores of each scheduling correction candidate plan. The weights in the weighted comprehensive evaluation function are set according to the process objectives; S86. Screen the scheduling correction candidate plan with the highest score as the final scheduling plan for execution.

[0035] By introducing a multi-objective comprehensive evaluation mechanism, the screening of the scheduling plan no longer depends solely on a single performance index, but comprehensively considers three factors: production capacity achievement rate, energy consumption change, and disturbance adaptability, ensuring that the finally output plan achieves a balance among production efficiency, cost control, and system stability. The use of the weighted comprehensive scoring function provides a flexible and adjustable strategy control means, and enterprises can dynamically set the optimization focus according to the actual production goals. At the same time, the correction candidate plans are processed in a structured sorting manner to ensure the transparency and interpretability of the evaluation process and avoid the algorithm decision-making process becoming a black box. The overall plan significantly improves the quality of the scheduling output and the controllability of industrial deployment, providing strong support for the stable operation of the system.

[0036] Example 1: To verify the feasibility of the present invention in implementation, the present invention is applied to the intelligent production line of a modern feed processing factory with an annual output of 480,000 tons. The production line covers six major sections: crushing, batching, mixing, granulation, cooling, and packaging. A total of 24 automated devices are deployed, and the average daily task batches are about 180 groups. There are high dependencies and frequent switches among the devices, and the scheduling management is complex. Traditional rule engine methods are difficult to handle dynamic disturbances and equipment conflicts.

[0037] In this embodiment, the original system has long relied on fixed priority and equipment pairing rules for production scheduling. When the granulation equipment has operating fluctuations or the mixer delays in switching recipes, the downstream equipment fails to adjust in time, resulting in multiple chain blockages and batch misalignments. Especially in the case of multi-batch parallelism, there are physical overlaps in the spatial positions between some devices. For example, if two granulators sharing a belt are scheduled simultaneously, it will cause vibration conflicts and abnormal temperature rises. The traditional system lacks an effective early warning mechanism for this and often requires manual intervention by the dispatcher. To solve the above problems, the disturbance-resistant residual fusion network, equipment conflict inference graph, and uncertainty-driven scheduling correction mechanism in the present invention are completely deployed into the intelligent scheduling module of this factory.

[0038] First, by structuring the collection of all equipment operation logs, fault records, energy consumption data, and recipe task allocation information within three months, approximately 312,000 status sequence samples were constructed, and a perturbation sample set including situations such as startup delay, operation interruption, and recipe switching failure was constructed. The system jointly trains through the main path and the side path to model the behaviors of each equipment at different working periods, and has good contrast memory and perturbation memory capabilities. Subsequently, based on the equipment spatial layout diagram of the factory area and the equipment dependency chain between recipes, an equipment conflict inference graph with 2,412 equipment-batch nodes and 8,763 conflict edges was generated and embedded into the scheduling network as a legality filter for each scheduling path. The improved NGBoost model was introduced in the scheduling output process to score the expected performance and risk of each behavior, automatically identify unstable behaviors, and generate revised candidate solutions.

[0039] In the first month of deployment and operation, the system processed a total of 4,921 sets of scheduling tasks, identified 784 high-risk scheduling behaviors, automatically generated and replaced the revised solutions 412 times, and the stability of task execution was improved by approximately 23.7% on average each time. During the high-load test week (daily task batches > 200), the overall scheduling failure rate decreased from the original 4.6% to 1.1%, the production capacity achievement rate increased from 92.3% to 97.4%, the energy consumption per unit output decreased by 8.2%, and the collaborative scheduling efficiency between equipment was significantly optimized. Especially when equipment maintenance or sudden anomalies occur, the new system can complete the solution revision and rearrangement within 2 seconds, which is about 11 times faster than the manual scheduling response time of the original system, effectively ensuring the continuity and safety of the production line.

[0040] Before and after the deployment of the present invention, there are significant differences in the actual production line operation. Taking 30 consecutive days of operation as a cycle for statistics, the total number of scheduling behaviors is 4,921, the system identifies and processes approximately 7,300 perturbation samples, and replaces the scheduling behaviors 412 times. The scheduling failure rate decreased from 4.6% of the original system to 1.1%; the average production capacity achievement rate increased from 92.3% to 97.4%; the energy consumption per ton of feed decreased from 71.6 kWh to 65.7 kWh; the average delay of task switching decreased from 18.4 seconds to 7.1 seconds; the frequency of manual intervention decreased by approximately 69.2%. Among them, the prediction module introducing the confidence factor and the dynamic distribution calibration mechanism has an accuracy rate of 88.7% in identifying high-risk behaviors. Overall, it shows 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 rate of the feed processing industry.

[0041] The above are only the preferred specific embodiments 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 and inventive concept of the present invention, making equivalent substitutions or changes, shall be covered by the protection scope of the present invention.

Claims

1. A collaborative scheduling and operation optimization method for feed production equipment based on deep learning, characterized in that, It includes the following steps: S1. Collect the operation status data of each device and the corresponding task recipe information during the feed production process to form a device status sequence data set; S2. Construct a perturbation sample set, which includes start-up delay, operation interruption and abnormal recipe switching, and is structured into a perturbation sequence data set; S3. Input the device status sequence data set and the perturbation sequence data set into a perturbation-resistant residual fusion network, which includes a main path and a bypass path, and dynamically weights and fuses them through a residual gating fusion unit to generate a preliminary scheduling plan; S4. Construct a device conflict inference graph based on the task recipe information and the spatial layout relationship, where the nodes represent the device-batch operation status, and the edges represent the non-parallel conflict conditions and are attached with conflict rule expressions; S5. Embed the device conflict inference graph as a structural constraint into the scheduling network to block the scheduling paths that violate the process logic and correct the preliminary scheduling plan; S6. Input the corrected preliminary scheduling plan into an improved NGBoost model, which introduces a confidence factor estimation and dynamic distribution calibration mechanism, and outputs the predicted expected value and uncertainty score of each scheduling behavior; S7. Identify high-risk behaviors according to the uncertainty score to generate a scheduling correction candidate set; S8. Conduct a multi-objective comprehensive evaluation on the scheduling correction candidate set, calculate the scores according to the production capacity achievement rate, energy consumption change amount and perturbation adaptability, and select the scheduling correction candidate plan with the highest score as the final scheduling plan for execution.

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

3. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, wherein The specific content of S3 includes: S31. Perform time dimension alignment and feature channel normalization processing on the device status sequence data set and the perturbation sequence data set to construct a main path input tensor and a bypass path input tensor corresponding to the time axis; S32. Input the main path input tensor into a main path network composed of a one-dimensional convolutional layer and a gated recurrent unit, extract short-term dependence features through the one-dimensional convolutional layer, and extract the operation trend across time steps through the gated recurrent unit to obtain a main path feature tensor; S33. Input the bypass path input tensor into a bypass path network composed of a convolutional structure, multi-scale residual connections and an attention mechanism, extract the dynamic influence features of perturbations at different time scales, and output a perturbation path feature tensor; S34. Concatenate the main path feature tensor and the perturbation path feature tensor in the feature dimension and input them into a residual gating fusion unit, which performs per-channel weighted fusion on the main path feature tensor and the perturbation path feature tensor to obtain a fusion feature tensor; S35. Compress the channels of the fusion feature tensor and input it into a structural decoder, which includes a device number encoding layer, a time window encoding layer and a load allocation encoding layer; The device number encoding layer performs a location mapping operation to embed and encode the spatial channels in the fused feature tensor with the device topology index, generating a number vector representation for each device, which constitutes the device identification dimension in the scheduling instruction; The time window encoding layer performs convolutional expansion and sliding window offset processing on the time step index in the fused feature tensor, extracts the corresponding start time segment and duration period in each scheduling path, and converts them into an integer interval identification vector; The load allocation encoding layer performs a max-mean dual-path normalization operation based on the fused feature tensor to form a load allocation ratio vector for the current task of each device; S36. Assemble the output result of the structured decoder into a preliminary scheduling plan, where the preliminary scheduling plan includes the start-stop sequence, number information, task execution time window, and load allocation ratio value of each device.

4. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, characterized in that The specific steps of S4 are as follows: S41. Analyze the task recipe information, extract the device usage sequence, process stage requirements, and device occupancy time period corresponding to each recipe, and form a recipe structured task template; S42. Collect the spatial layout data of the feed mill, including device numbers, installation locations, physical occupancy areas, and process channel relationships, and establish a spatial layout mapping table; S43. According to the occupancy of each device by each batch of tasks in the recipe structured task template, set the state combination of each device under each batch of tasks as the node set of the device conflict inference graph. The nodes are uniquely identified by the device identification - batch identification combination, representing the device - batch operation state; S44. According to the spatial layout mapping table and the device combinations with parallel relationships in the recipe structured task template, identify task pairs with resource overlap or time conflict, and define them as the conflict edge set; S45. Set a conflict type identifier and a conflict rule expression for each conflict edge. The conflict rule expression is composed of recipe differences, physical space overlap, operation stage incompatibility, and device sharing status; S46. Construct a complete device conflict inference graph, including the node set, edge set, and corresponding conflict rule expressions.

5. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, wherein The specific steps of S5 are as follows: S51. Extract the start-stop sequence, number information, task execution time window, and load allocation ratio value of each device task in the preliminary scheduling plan, and parse them into a scheduling path set; S52. Traverse the scheduling path set, and for each node combination involved in each scheduling path, find the corresponding nodes and adjacent conflict edges in the device conflict inference graph; S53. According to the conflict rule expressions recorded in the conflict edges, perform conflict verification on each current scheduling path item by item, and identify the conflict path branches that violate the spatial occupancy, recipe dependency, and resource mutual exclusion conditions; S54. Mark the identified conflict 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 conflict paths in the forward propagation stage; S55. Take the output of the scheduling path filtered by the scheduling mask tensor as the corrected preliminary scheduling plan. The corrected preliminary scheduling plan does not contain any scheduling paths that violate the process logic of the device conflict inference graph rules.

6. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, characterized in that The improvements of the improved NGBoost model lie in introducing a 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 perturbation dataset, statistically calculate the prediction error amplitude of each scheduling behavior, and construct a confidence factor for dynamically adjusting the width of the output distribution: ； Among them, represents the confidence factor at time step , represents the exponential function, represents the adjustment parameter, represents the time step when the prediction error is represents the mean of the prediction errors; Introduce a dynamic distribution calibration mechanism during the natural gradient boosting process, and perform the following adjustments when updating the prediction distribution parameters each time: ； Among them, represents the calibrated natural gradient, represents the standard natural gradient, represents the distribution calibration intensity coefficient, represents the time step the true perturbation distribution at time step represents the time step the predicted distribution at time step represents the predicted distribution parameter; Decode the output prediction distribution parameters after confidence factor adjustment and dynamic distribution calibration into the predicted expected value and uncertainty score corresponding to each scheduling behavior. The predicted expected value represents the performance expectation of the current scheduling behavior under perturbations; Use the confidence factor as a dynamic weight to scale and adjust the original variance value, and the adjusted variance value obtained is used as the uncertainty score corresponding to the current scheduling behavior.

7. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, characterized in that The specific content of S7 includes: S71. Set a scheduling risk identification threshold, mark the scheduling behaviors with uncertainty scores exceeding the scheduling risk identification threshold as high-risk behaviors, and construct a high-risk behavior index list; S72. According to the device number, task recipe requirements, and time window position of each high-risk behavior, locate the context structure of the high-risk behavior in the revised preliminary scheduling plan; S73. Generate scheduling correction candidate plans for each high-risk behavior on the premise of not violating the current task recipe execution logic and device space conflict constraints; S74. Classify all the generated scheduling correction candidate plans into a scheduling correction candidate set and retain the mapping relationship in the revised preliminary scheduling plan.

8. The collaborative scheduling and operation optimization method for feed production equipment based on deep learning according to claim 1, wherein The specific content of S8 includes: S81. Receive the generated scheduling correction candidate set, and structurally analyze the device start-stop sequence, time window, and load allocation information associated with each scheduling correction candidate plan; S82. Calculate the production capacity achievement rate of each scheduling correction candidate plan according to the standard production capacity target of the current batch of tasks. The production capacity achievement rate is the ratio of the actual executed production capacity to the theoretical recipe production capacity; S83. Estimate the energy consumption of the device operation plan corresponding to each scheduling correction candidate plan, and calculate the energy consumption change compared with the initial scheduling plan in combination with the unit time operating power and load level; S84. Extract the mean value of the uncertainty scores of each scheduling correction candidate plan under the prediction of the improved NGBoost model as a risk quantification index for measuring perturbation fitness; S85. Construct a weighted comprehensive evaluation function based on the production capacity achievement rate, energy consumption change, and perturbation fitness, and rank the scores of each scheduling correction candidate plan. The weights in the weighted comprehensive evaluation function are set according to the process objectives; S86. Screen the scheduling correction candidate plan with the highest score as the final scheduling plan for issuance and execution.

Images