A human-computer co-conformable intelligent workshop scheduling method, system, medium and device
By dynamically modeling worker states and generating scheduling rules using large language models, combined with MAPPO agents for real-time scheduling, the problem of insufficient human factor modeling in existing methods is solved, and efficient and interpretable scheduling optimization in the workshop is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HOHAI UNIV
- Filing Date
- 2026-04-24
- Publication Date
- 2026-07-31
AI Technical Summary
Existing scheduling methods lack dynamic modeling of human factors, ignore the impact of worker skill level and actual condition on work efficiency, and traditional methods are insufficient in terms of transportation resource constraints and logical reasoning ability, resulting in a disconnect between production and logistics scheduling and making it difficult to adapt to actual production scenarios.
By collecting real-time production and order data from the workshop, dynamically modeling worker status, generating heuristic scheduling rules using a large language model, and combining this with the MAPPO agent for real-time scheduling, the production process is optimized by taking into account worker fatigue and skill experience.
It achieves highly efficient and interpretable scheduling in the workshop, proactively adapts to the workers' status, improves the collaborative efficiency and sustainability of the production process, and breaks through the limitations of fixed motion space in traditional methods.
Smart Images
Figure CN122491737A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to a human-machine collaborative intelligent workshop scheduling method, system, medium, and equipment, belonging to the field of intelligent manufacturing technology. Background Technology
[0002] In real-world hybrid assembly line workshops, scheduling problems not only involve matching machines and workpieces but also face complex constraints related to personnel status and logistics resources. Existing scheduling methods suffer from the following limitations: They lack dynamic modeling of human factors, and most existing methods assume constant worker efficiency, neglecting the impact of worker skill and actual condition on work efficiency. Furthermore, idealized assumptions about transportation resources ignore physical constraints such as the number and location of AGVs and transfer delays, leading to a disconnect between production and logistics scheduling, making them unsuitable for highly constrained real-world production scenarios. In addition, scheduling decision systems suffer from a lack of logical reasoning ability and rigid scheduling rules. While mainstream reinforcement learning methods offer good performance, their decisions lack interpretability and are typically limited to a pre-defined fixed action space, unable to create new strategies. Traditional genetic programming (GP), although capable of generating rules, has a huge search space and slow convergence speed. Summary of the Invention
[0003] The purpose of this invention is to provide a human-machine integrated intelligent workshop scheduling method, system, medium, and equipment. By dynamically modeling the worker's state, human factor information such as accumulated experience and worker fatigue is embedded into the scheduling agent. Combined with the scheduling rule library obtained by the evolution of a large language model, high-efficiency and interpretable scheduling of the workshop is achieved.
[0004] To achieve the above objectives, the present invention is implemented using the following technical solution.
[0005] On the one hand, the present invention provides a human-machine collaborative intelligent workshop scheduling method, comprising:
[0006] Real-time collection of workshop production and order data, and extraction of workshop status and worker status;
[0007] The workshop state and worker state are converted into state vectors and input into the pre-trained MAPPO agent to select a heuristic scheduling rule that is suitable for the current workshop from the preset action space.
[0008] The heuristic scheduling rules are executed to achieve real-time scheduling of the workshop;
[0009] The worker's condition is extracted through a pre-built experience accumulation model and a fatigue and recovery model;
[0010] The preset action space is constructed in the following way:
[0011] The semantic states of the workshop and workers are constructed, and they are converted into natural language prompts along with the basic composite scheduling rules, scheduling objectives and constraints. These prompts are then input into a large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
[0012] Optionally, the extraction workshop status includes:
[0013] Extract the workpiece's process index, remaining process set, earliest start time of the workpiece, and estimated remaining processing time required to complete the workpiece to form the workpiece status;
[0014] Extract the machine's earliest available time, load rate, buffer status, and maintenance window to form the machine status;
[0015] Extract order priority, urgent order flag, order arrival date, and order deadline to form the order status.
[0016] Optionally, the extraction of the worker's status includes:
[0017] Obtain workers' cumulative effective working time and use a pre-built experience accumulation model to calculate real-time skill experience values;
[0018] Using a pre-built fatigue and recovery model, the real-time fatigue value of workers is calculated;
[0019] Based on the worker's real-time skill experience value and real-time fatigue value, the standard processing time is corrected to obtain the actual processing time;
[0020] The ratio of standard processing time to actual processing time is calculated to obtain the worker's real-time comprehensive efficiency, which serves as the worker's status.
[0021] Optionally, the scheduling objective is to minimize the maximum completion time;
[0022] The constraints include at least: machine processing capacity constraints, process sequence constraints, buffer capacity constraints, AGV quantity constraints, transportation time and location constraints, personnel-machine matching constraints, and personnel status constraints.
[0023] Optionally, the large language model outputs heuristic scheduling rules after iterative evolution, including:
[0024] The basic composite scheduling rules are used as leaf nodes of the genetic programming tree, and mathematical and logical operators are used as connection nodes to randomly generate an initial population of rule expressions.
[0025] Using a large language model as the core operator for evolution, in each generation of evolution, the semantic states of workshops and workers are combined to perform semantically perceptual nonlinear combination and nesting of the rule expressions in the population, and output composite scheduling rule expressions that adapt to different semantic states of workshops and workers.
[0026] Optionally, before selecting a heuristic scheduling rule from the preset action space that fits the current workshop, the following steps are also included:
[0027] The production instructions in natural language form are parsed using a large language model. The workpiece scheduling priority in the production instructions is extracted and transformed into a structured guidance vector, which is then input into the MAPPO agent.
[0028] Optionally, the MAPPO agent is configured with a completion time reward function and a semantic execution reward function during the training phase;
[0029] Completion time bonus ,in, For the current reward point moment, This indicates that at the previous reward point, the agent receives the reward at each decision point. Its optimization objective is to minimize the time interval between adjacent reward points in order to minimize the maximum completion time.
[0030] Semantic execution reward This is used to quantify the alignment between the agent's decisions and the semantic guidance of the large language model; among which, and The preset instant reward value and end-game reward value, The variable is a phased completion indicator. It takes a value of 1 when a high-priority workpiece completes any phase of its task at the current decision point k, and 0 otherwise; n is the time when the scheduling ends. The total number of high-priority workpieces in the valid window.
[0031] Secondly, the present invention provides a human-machine collaborative intelligent workshop scheduling system, comprising:
[0032] The data acquisition module is used to collect workshop production and order data in real time, and extract workshop status and worker status.
[0033] The rule selection module is used to convert the workshop state and worker state into state vectors and input them into the pre-trained MAPPO agent to select heuristic scheduling rules that are suitable for the current workshop from the preset action space.
[0034] The execution module is used to execute the heuristic scheduling rules to achieve real-time scheduling of the workshop;
[0035] The worker's condition is extracted through a pre-built experience accumulation model and a fatigue and recovery model;
[0036] The preset action space is constructed in the following way:
[0037] The semantic states of the workshop and workers are constructed, and they are converted into natural language prompts along with the basic composite scheduling rules, scheduling objectives and constraints. These prompts are then input into a large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
[0038] Thirdly, the present invention provides a computer-readable storage medium having a computer program / instruction stored thereon, wherein when the computer program / instruction is executed by a processor, it implements the intelligent workshop scheduling method for human-machine collaboration as described in the first aspect.
[0039] Fourthly, the present invention provides a computer device, comprising:
[0040] Memory, used to store computer programs / instructions;
[0041] A processor for executing the computer program / instructions to implement the steps of the intelligent workshop scheduling method for human-machine collaboration as described in any one aspect.
[0042] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:
[0043] This invention dynamically models worker states and embeds them into the model's inference, enabling the workshop to proactively adapt to worker states during scheduling, thereby optimizing human-machine collaboration efficiency and ensuring sustainable production. By semanticizing workshop states and generating heuristic scheduling rules using a large language model, the interpretability of scheduling strategies is enhanced, overcoming the limitations of traditional reinforcement learning's fixed action space and demonstrating a certain degree of generalization performance during scheduling. Attached Figure Description
[0044] Figure 1 This is a schematic diagram of the intelligent workshop scheduling method for human-machine collaboration provided in an embodiment of the present invention;
[0045] Figure 2 This is a schematic diagram illustrating the functionality of the large language model provided in an embodiment of the present invention. Detailed Implementation
[0046] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.
[0047] Example 1
[0048] This embodiment provides a human-machine collaborative intelligent workshop scheduling method, such as... Figure 1 As shown, it includes the following steps:
[0049] S1. Real-time collection of workshop production and order data, and extraction of workshop status and worker status.
[0050] S1.1, Extraction of Workshop Status:
[0051] Workpiece Status: Extract the workpiece's process index, remaining process set, earliest start time of the workpiece, and estimated remaining processing time required to complete the workpiece, which are used to describe the current processing progress of the workpiece.
[0052] Machine Status: Extracts the earliest available time, current load rate, buffer status, and maintenance window of the machine to describe its load and availability.
[0053] Order status: Extract order priority, urgent order flag, order arrival date, and order deadline to describe the urgency and dynamic attributes of the task.
[0054] S1.2, Worker Status Extraction:
[0055] Existing scheduling methods lack dynamic modeling of human factors. Most existing methods assume constant worker efficiency and ignore the impact of worker proficiency and actual condition on work efficiency. Therefore, this embodiment extracts worker state through a pre-built experience accumulation model and a fatigue and recovery model.
[0056] S1.2.1 Obtain the worker's cumulative effective working time and calculate the real-time skill experience value using a pre-built experience accumulation model.
[0057] First, considering the positive impact of experience accumulation on efficiency, we no longer assume constant worker efficiency, but define the worker's real-time skill experience value based on the learning curve theory. The change curve describes the worker's skill experience value. The formula for calculating the growth over time is as follows:
[0058] ;
[0059] in, This represents the total effective working time of worker j. Real-time skill experience points; This represents the initial experience level of worker j. The maximum level of experience that worker j can achieve through learning; This represents the learning rate parameter. The time required for a worker to reach half of the maximum experience increment.
[0060] S1.2.2 Calculate the worker's real-time fatigue value using a pre-built fatigue and recovery model;
[0061] A fatigue and recovery model is introduced to consider the negative impact of fatigue on efficiency, quantifying the fatigue accumulation caused by continuous work and the physical recovery process during rest periods. The real-time fatigue state of workers is defined as increasing exponentially during continuous work and decreasing exponentially during rest. The system applies a penalty correction to the standard processing time based on the real-time fatigue value.
[0062] When worker j is in a continuous working state, the real-time fatigue value calculation formula is:
[0063]
[0064] When worker j is at rest, the formula for calculating his / her real-time fatigue value is:
[0065]
[0066] in, This represents the real-time fatigue value of worker j at the current time t. This represents the worker's fatigue level at the previous moment; Indicates from the previous moment The time interval up to the current time t; This represents the cumulative fatigue coefficient. This represents the fatigue recovery coefficient.
[0067] S1.2.3, Based on workers' real-time skill experience values and real-time fatigue value The standard processing time is adjusted to obtain the actual processing time;
[0068] Specifically, this involves: reducing standard processing time by accumulating experience, i.e., when real-time skill experience values are high; and increasing standard processing time by accumulating fatigue, i.e. when workers are in a state of high fatigue, thus obtaining the actual processing time.
[0069] S1.2.4. The ratio of the set standard processing time to the actual processing time is used as the worker's real-time comprehensive efficiency, which is also used as the worker's state. For example, if the standard processing time for a certain process is 100s, and the worker is in a state of high proficiency and low fatigue, the corrected actual processing time is 80s, then the real-time comprehensive efficiency filled into the state vector is 1.25; conversely, if the actual processing time is extended to 125s due to fatigue, the efficiency drops to 0.8.
[0070] S2. Convert the workshop state and worker state into state vectors and input them into the pre-trained MAPPO agent to select a heuristic scheduling rule that is suitable for the current workshop from the preset action space.
[0071] S2.1 Construct a workshop scheduling environment model using a Python simulation platform, and describe the problem of the environment model: After the workshop receives an order, each workpiece sequentially undergoes the same number of processing stages to complete the process. Each stage involves several processing machines and multiple workers, and the workers work in shifts. Furthermore, make reasonable assumptions about the environment model: one machine can only process one workpiece at a time, and each employee can operate at most one machine.
[0072] Under the above description and assumptions, the simulation environment can provide real-time feedback on the workshop status and support the input of external scheduling rules and instructions, returning the corresponding scheduling performance indicators, namely the maximum completion time.
[0073] S2.2 Constructing a preset action space:
[0074] like Figure 2 As shown, the semantic states of the workshop and workers are constructed (the workshop states are semanticized to fill the gap between numerical data and the cognition of the large language model), and they are converted into natural language prompts along with the basic composite scheduling rules, scheduling objectives and constraints. These prompts are then input into the large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
[0075] Specifically, the scheduling objective is to minimize the maximum completion time Makespan.
[0076] Specifically, the constraints include: machine processing capacity constraints, process sequence constraints, buffer capacity constraints, AGV quantity constraints, transportation time and location constraints, and personnel-machine matching constraints.
[0077] Due to machine processing capacity constraints, each machine can only process one workpiece at a time.
[0078] The process sequence is constrained; a workpiece must complete the previous process before it can proceed to the next stage.
[0079] The buffer capacity is constrained; the number of artifacts in each buffer is limited by the physical capacity.
[0080] AGV quantity constraint: The total number of AGVs available in the workshop is fixed, and the number of AGVs in active transport mode at any given time must not exceed the total limit.
[0081] Due to time and location constraints, the time for workpieces to move between different processing stages consists of three parts: waiting time for AGVs, empty travel time of AGVs to workpieces, and load transport time of AGVs carrying workpieces to the next machine.
[0082] Personnel-machine matching constraints: each processing stage is assisted by designated workers, and their efficiency affects the actual processing time.
[0083] Personnel status constraints mean that worker efficiency changes dynamically over time and is reflected in the processing time calculation.
[0084] Specifically, basic composite scheduling rules (Shortest Processing Time First, Earliest Delivery Date First, First-In-First-Out, FIFO, etc.) are used as leaf nodes of the genetic programming tree, and mathematical and logical operators are used as connection nodes. A large language model (LLM) infers and generates several structurally valid general programming (GP) rule expressions, forming the first generation population. Each rule expression is run in a simulation environment, and its maximum completion time (Makespan) is recorded. Using Makespan as the fitness value, the best-performing rules are retained as elite rules, and the rest are eliminated. The textual expressions of the elite rules are then combined with the semantic states of the workshop and workers to form mutation prompts. These prompts are input into the LLM for semantically perceptual nonlinear combination and nesting, outputting composite scheduling rule expressions adapted to different workshop and worker semantic states, forming the next generation population.
[0085] When the optimal Makespan no longer decreases significantly after several consecutive generations, the evolution stops, and the final set of elite rule expressions is the preset action space adapted to the semantic state of a specific workshop and worker.
[0086] S3. Execute the heuristic scheduling rules to achieve real-time scheduling of the workshop.
[0087] like Figure 2 As shown, in one embodiment, before selecting a heuristic scheduling rule that fits the current workshop from the preset action space, i.e. during the order status perception process, the parsing and feature extraction of human natural language instructions are also performed.
[0088] When receiving unstructured natural language instructions (e.g., "Prioritize high-priority jobs 10, 11, and 12 in order A"), the system invokes LLM for semantic parsing. Specifically, LLM extracts key scheduling features from the instructions, including job ID and priority. The system then transforms these features into a structured guidance vector. For example, urgent jobs are prioritized with a preset high weight, while ordinary jobs are assigned a standard weight by default. This structured guidance vector is then concatenated into a state space consisting of job, machine, order, and worker states and input into the MAPPO agent. In this way, human production intentions can be directly translated into state features that the agent can understand.
[0089] This agent employs an event-driven mechanism based on a task pool to trigger scheduling, mapping the semantic workshop state to a state vector and using the GP rule set generated by LLM as the discrete action space. A reward function is designed that includes completion time rewards and semantic execution rewards, guiding the agent to balance completion time and instruction fulfillment rate.
[0090] The completion time bonus Related to Makespan, Makespan minimization is achieved by minimizing the single-step time.
[0091] ;
[0092] in and These represent the time points corresponding to the current reward point and the previous reward point, respectively. During training, the agent obtains a single-step reward value based on the selected action. The total reward value is obtained by accumulating the single-step decision rewards, thereby effectively learning to select the optimal strategy at different scheduling points to obtain higher reward values and shorten Makespan.
[0093] Semantic execution reward It consists of two dimensions: process guidance and outcome evaluation. Real-time process guidance: Whenever a high-priority task successfully progresses to a certain process stage, an immediate reward is triggered, providing continuous gradient guidance for model training. Final goal incentive: At the end of the scheduling cycle, the system verifies the completion time of high-priority tasks. If the task is ultimately among the earliest completed K items in the sequence (Top-K window), a final reward is assigned based on the actual number of completed tasks, n.
[0094] The calculation is as follows:
[0095] ;
[0096] in, and These are the preset instant reward value and final reward value. The value is 1 when a high-priority workpiece completes any stage of the task at the current decision point t, and 0 otherwise; n is the total number of high-priority workpieces that enter the Top-K effective window at the end of the scheduling.
[0097] The MAPPO agent, consisting of multiple actor networks and one critic network, is trained using the following specific steps:
[0098] Environment initialization: The total number of training rounds is 15,000, the memory pool capacity is 1024, the batch size is 512, and there are 6 actor networks, which correspond to the 6 scheduling decision-makers in the workshop environment. Each actor network adopts a three-layer fully connected hidden layer structure, with the number of neurons in each layer being 512, 256, and 128 respectively. The activation function is ReLU, and the output layer outputs the action probability distribution.
[0099] A global commentator network is constructed using a three-layer fully connected hidden layer structure, with 1024, 512, and 256 neurons in each layer, respectively. The ReLU activation function is used, and the output layer outputs a scalar state value.
[0100] Experience collection: When the time point is a scheduling point, the agent obtains the workshop state of the scheduling point, then selects an appropriate scheduling rule based on the action probability distribution, executes the action, observes the state transition at the next scheduling point, obtains the new state and single-step reward, and then stores the decision experience of this round of training into the memory pool.
[0101] Network Update: Once the memory pool capacity reaches the update condition, historical decision-making experience is randomly selected from the memory pool. The loss functions of the actor network and the critic network are calculated, and the network parameters are updated through backpropagation. Specifically:
[0102] ;
[0103] Let the loss function be the actor network. B represents the parameters of the actor network; B represents the batch size. The probability ratio between the old and new strategies. For the current actor network in local observation The next output action probability, This represents the probability that the old policy would output the same action under the same observations before the update. For the dominant function, This is a cutoff factor used to constrain the magnitude of change in the probability ratio between the old and new strategies; This represents the trimming function, where min indicates taking the smaller value between the untrimmed term and the truncated term. The action entropy regularization coefficient, For strategy In observation The action distribution entropy is used to encourage exploration and prevent the strategy from converging to a local optimum too early.
[0104] ;
[0105] The loss function of the critic network, For the parameters of the critic network, For the commentator network on the global state The value of the predicted state, The target value is calculated based on the actual reward sequence and the discount factor.
[0106] In each round of parameter updates, empirical data of batch size B is sampled from the empirical memory pool, and the actor network loss is calculated according to the aforementioned objective function. and the loss of the critic network The network parameters were then adjusted using a backpropagation algorithm and an optimizer (such as Adam with a learning rate of 0.0005). and Perform gradient descent updates.
[0107] Model selection: Determine if the maximum number of training rounds has been reached. If not, proceed to a new round of training. If the maximum number of training rounds has been reached, training is complete. Select the optimal network parameters based on scheduling performance and deploy it as an online scheduling model.
[0108] Example 2
[0109] Based on the same inventive concept as Embodiment 1, this embodiment introduces a human-machine collaborative intelligent workshop scheduling system, including:
[0110] The data acquisition module is used to collect workshop production and order data in real time, and extract workshop status and worker status.
[0111] The rule selection module is used to convert the workshop state and worker state into state vectors and input them into the pre-trained MAPPO agent to select heuristic scheduling rules that are suitable for the current workshop from the preset action space.
[0112] The execution module is used to execute the heuristic scheduling rules to achieve real-time scheduling of the workshop.
[0113] Specifically, the worker's condition is extracted through a pre-built experience accumulation model and a fatigue and recovery model;
[0114] The preset action space is constructed in the following way:
[0115] The semantic states of the workshop and workers are constructed, and they are converted into natural language prompts along with the basic composite scheduling rules, scheduling objectives and constraints. These prompts are then input into a large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
[0116] The specific functions of each module described above are explained in the relevant content of the method in Embodiment 1, and will not be repeated here.
[0117] Example 3
[0118] Based on the same inventive concept as Embodiment 1, this embodiment introduces a computer-readable storage medium storing a computer program / instruction thereon, which, when executed by a processor, implements the intelligent workshop scheduling method for human-machine collaboration as described in Embodiment 1.
[0119] Example 4
[0120] Based on the same inventive concept as Embodiment 1, this embodiment introduces a computer device, including:
[0121] Memory, used to store computer programs / instructions;
[0122] A processor is used to execute the computer program / instructions to implement the steps of the intelligent workshop scheduling method for human-machine collaboration as described in any one of Embodiments 1.
[0123] In summary, this invention dynamically models worker states and embeds these states into the model's inference. It also embeds human factors information, such as accumulated experience and worker fatigue, into the scheduling agent, enabling the workshop to proactively adapt to worker states during scheduling. This optimizes human-machine collaboration efficiency and ensures sustainable production. By semanticizing workshop states and generating heuristic scheduling rules using a large language model, the interpretability of the scheduling strategy is enhanced, overcoming the limitations of traditional reinforcement learning's fixed action space and exhibiting generalization performance during scheduling. Furthermore, the evolution of scheduling rules considers constraints related to AGV quantity, transportation time and location, and personnel-machine matching, making the connection between production and logistics more efficient in actual workshop scheduling. The personnel-machine constraints also consider the impact of worker efficiency during the workpiece processing stage on the overall processing time.
[0124] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.
[0125] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.
[0126] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0127] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.
[0128] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A human-machine collaborative intelligent workshop scheduling method, characterized in that, include: Real-time collection of workshop production and order data, and extraction of workshop status and worker status; The workshop state and worker state are converted into state vectors and input into the pre-trained MAPPO agent to select a heuristic scheduling rule that is suitable for the current workshop from the preset action space. The heuristic scheduling rules are executed to achieve real-time scheduling of the workshop; The worker's condition is extracted through a pre-built experience accumulation model and a fatigue and recovery model; The preset action space is constructed in the following way: The semantic states of workshops and workers are constructed, and these, along with the basic composite scheduling rules, scheduling objectives, and constraints, are converted into natural language prompts. These prompts are then input into a large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
2. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, The status of the extraction workshop includes: Extract the workpiece's process index, remaining process set, earliest start time of the workpiece, and estimated remaining processing time required to complete the workpiece to form the workpiece status; Extract the machine's earliest available time, load rate, buffer status, and maintenance window to form the machine status; Extract order priority, urgent order flag, order arrival date, and order deadline to form the order status.
3. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, The extraction of the worker's status includes: Obtain workers' cumulative effective working time and use a pre-built experience accumulation model to calculate real-time skill experience values; Using a pre-built fatigue and recovery model, the real-time fatigue value of workers is calculated; Based on the worker's real-time skill experience value and real-time fatigue value, the standard processing time is corrected to obtain the actual processing time; The ratio of standard processing time to actual processing time is calculated to obtain the worker's real-time comprehensive efficiency, which serves as the worker's status.
4. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, The scheduling objective is to minimize the maximum completion time. The constraints include at least: machine processing capacity constraints, process sequence constraints, buffer capacity constraints, AGV quantity constraints, transportation time and location constraints, personnel-machine matching constraints, and personnel status constraints.
5. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, After iterative evolution, the large language model outputs heuristic scheduling rules, including: The basic composite scheduling rules are used as leaf nodes of the genetic programming tree, and mathematical and logical operators are used as connection nodes to randomly generate an initial population of rule expressions. Using a large language model as the core operator of evolution, in each generation of evolution, the semantic states of workshops and workers are combined to perform semantically perceptual nonlinear combination and nesting of the rule expressions in the population, and output composite scheduling rule expressions that adapt to different semantic states of workshops and workers.
6. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, Before selecting a heuristic scheduling rule that fits the current workshop from the preset action space, the following steps are also included: The production instructions in natural language form are parsed using a large language model. The workpiece scheduling priority in the production instructions is extracted and transformed into a structured guidance vector, which is then input into the MAPPO agent.
7. The intelligent workshop scheduling method for human-machine collaboration according to claim 1, characterized in that, During the training phase, the MAPPO agent is configured with a completion time reward function and a semantic execution reward function; Completion time bonus ,in, For the current reward point moment, This indicates that at the previous reward point, the agent receives the reward at each decision point. Its optimization objective is to minimize the time interval between adjacent reward points in order to minimize the maximum completion time. Semantic execution reward This is used to quantify the alignment between the agent's decisions and the semantic guidance of the large language model; among which, and The preset instant reward value and end-game reward value, The variable is a phased completion indicator. It takes a value of 1 when a high-priority workpiece completes any phase of its task at the current decision point k, and 0 otherwise; n is the time when the scheduling ends. The total number of high-priority workpieces in the effective window.
8. A human-machine collaborative intelligent workshop scheduling system, characterized in that, include: The data acquisition module is used to collect workshop production and order data in real time, and extract workshop status and worker status. The rule selection module is used to convert the workshop state and worker state into state vectors and input them into the pre-trained MAPPO agent to select heuristic scheduling rules that are suitable for the current workshop from the preset action space. The execution module is used to execute the heuristic scheduling rules to achieve real-time scheduling of the workshop; The worker's condition is extracted through a pre-built experience accumulation model and a fatigue and recovery model; The preset action space is constructed in the following way: The semantic states of workshops and workers are constructed, and these, along with the basic composite scheduling rules, scheduling objectives, and constraints, are converted into natural language prompts. These prompts are then input into a large language model, and after iterative evolution, heuristic scheduling rules are output to construct a preset action space.
9. A computer-readable storage medium having a computer program / instructions stored thereon, characterized in that, When the computer program / instruction is executed by the processor, it implements the intelligent workshop scheduling method for human-machine collaboration as described in any one of claims 1 to 7.
10. A computer device, characterized in that, include: Memory, used to store computer programs / instructions; A processor for executing the computer program / instructions to implement the steps of the intelligent workshop scheduling method for human-machine collaboration as described in any one of claims 1 to 7.