Distributed heterogeneous workshop scheduling method and equipment based on greedy heuristic enhancement

Abstract

The invention belongs to the related technical field of distributed workshop dynamic scheduling, and discloses a distributed heterogeneous workshop scheduling method and device based on greedy heuristic enhancement, and the method comprises the steps: inputting the state characteristics of a to-be-scheduled distributed heterogeneous workshop into a double-layer end-to-end strategy network; the double-layer end-to-end strategy network outputs an optimal workpiece to be processed, and then a final scheduling scheme is obtained; wherein the training environment of the double-layer end-to-end policy network is constructed based on a partial observable Markov decision process scheduling model; the state features are obtained based on a partial observable Markov decision process scheduling model; the partial observable Markov decision process scheduling model is obtained by modeling a dynamic distributed heterogeneous workshop scheduling problem by adopting a decentralized multi-agent architecture. According to the invention, the expansibility and the flexibility are improved.

Description

Technical Field

[0001] The present invention belongs to the technical field related to distributed workshop dynamic scheduling, and more specifically, relates to a distributed heterogeneous workshop scheduling method and device based on enhanced greedy heuristic. Background Art

[0002] In today's rapidly changing market environment, the manufacturing industry of 3C products (such as mobile phones) faces high production complexity and uncertainty. The distributed manufacturing mode, with its high responsiveness and flexibility, is becoming an important means to meet emergency production demands and achieve mass customization production. The distributed workshop scheduling problem is a typical application scenario for large-scale production in distributed manufacturing and has become one of the rapidly growing research hotspots in the scheduling field in recent years. However, the production facilities in distributed manufacturing are often not completely the same, and this heterogeneity is inevitable in the production process, stemming from differences in aspects such as process flexibility and equipment performance in the factory. Therefore, the research on the distributed heterogeneous workshop scheduling problem needs to comprehensively consider the differences in the workshops, which has important practical significance.

[0003] In the manufacturing process of 3C products such as mobile phones, heterogeneous production environments are particularly common. Taking the mobile phone manufacturer Xiaomi as an example, it has multiple production factories across China, including self-owned factories and subcontractors, such as the Changping factory in Beijing, the Foxconn subcontractor in Langfang, and the BYD subcontractor in Xi'an. The models and states of the machinery and equipment used in these factories are different, constituting heterogeneous production lines. At the same time, the uncertainty of customer demands also makes the release of production orders dynamic and random, further exacerbating the complexity of production scheduling. This randomness will lead to the random arrival of tasks, making the scheduling plan need to be adjusted in real time to cope with the sudden changes that occur at any time.

[0004] The distributed heterogeneous workshop scheduling problem is a generalization of the traditional workshop scheduling problem and belongs to a typical NP-hard problem. The heterogeneity between factories significantly increases the scheduling difficulty because its solution space grows exponentially with the increase in the number of factories. In this problem, assigning the same task sequence to different factories will result in different completion times, so task assignment requires more careful consideration. To solve such scheduling problems, the existing common methods mainly include exact algorithms, priority dispatching rules (PDR), and meta-heuristic algorithms. Exact algorithms can find the optimal solution for small-scale problems, but their efficiency is extremely low when the problem scale increases. PDR is suitable for dynamic environments and is easy to implement, but its generality in all scenarios is poor. Meta-heuristic algorithms achieve a balance between the quality and efficiency of solutions in medium and large-scale problems, but they perform poorly when dealing with frequent dynamic changes. And this distributed heterogeneous workshop scheduling problem requires real-time adjustment of the scheduling plan to cope with the uncertainty brought by the random arrival of tasks. Therefore, how to study a scheduling method while ensuring real-time response, decision-making quality, and scalability has important engineering value. Summary of the Invention

[0005] In view of the above defects or improvement requirements of the prior art, the present invention provides a distributed heterogeneous job shop scheduling method and device based on enhanced greedy heuristic, aiming to solve the problems that existing scheduling methods are difficult to solve dynamic events and scale expansion.

[0006] To achieve the above object, according to one aspect of the present invention, there is provided a distributed heterogeneous job shop scheduling method based on enhanced greedy heuristic, the method comprising the following steps:

[0007] Input the state characteristics of the distributed heterogeneous job shop to be scheduled into a two-layer end-to-end policy network, and the two-layer end-to-end policy network outputs the optimal workpiece to be processed, and then the final scheduling scheme is obtained;

[0008] Wherein, the training environment of the two-layer end-to-end policy network is constructed based on a partially observable Markov decision process scheduling model; the state characteristics are obtained based on the partially observable Markov decision process scheduling model; the partially observable Markov decision process scheduling model is obtained by modeling the dynamic distributed heterogeneous job shop scheduling problem using a decentralized multi-agent architecture.

[0009] Further, input the state characteristics of the distributed heterogeneous job shop to be scheduled into a two-layer end-to-end policy network, and the two-layer end-to-end policy network outputs the optimal workpiece to be processed to form an initial scheduling scheme; then, search the initial scheduling scheme based on the greedy heuristic enhanced reverse exchange search method, and further obtain the final scheduling scheme.

[0010] Further, the two-layer end-to-end policy network is constructed based on a reinforcement learning algorithm, which includes a feature extraction network and an action network. The feature extraction network is used to process the workpiece pool features, and the action network generates the selection probability of each workpiece based on the features obtained by the feature extraction network and the workshop state features.

[0011] Further, each workshop is regarded as an agent, and an agent scheduler is used to determine the execution order of multiple agents and share the state; the agent observes the environment and inputs the state characteristics into the two-layer end-to-end policy network to obtain a scheduling decision. After multiple agent scheduling decisions, it transitions to the next state.

[0012] Further, a two-layer end-to-end policy network is constructed based on the actor-critic reinforcement learning method; wherein, the actor network refers to the two-layer end-to-end policy network, and the critic network is used to evaluate the advantage of the decision taken.

[0013] Further, the scheduling model is set with a two-dimensional fusion state space that comprehensively considers heterogeneity and permutation characteristics, an end-to-end action space, and a reward function based on a single workshop dimension.

[0014] Furthermore, the two-dimensional fusion state space fuses the state characteristics of the work pool dimension and the workshop dimension, namely the workpiece pool characteristics and the workshop characteristics. The workpiece pool characteristics describe the workpiece information that can be selected by the current workshop agent, and the workshop characteristics describe the operating states of different heterogeneous workshops.

[0015] Furthermore, the action space contains all the workpieces in the current workpiece pool. All the workpieces that have arrived at the distributed workshop are sorted in ascending order of the total processing time, and the first w workpieces are selected to form the workpiece pool, that is, the action space; the reward function is used for the training of the double-layer end-to-end policy network.

[0016] The present invention also provides a distributed heterogeneous workshop scheduling system based on greedy heuristic enhancement. The system includes a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it executes the distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as described above.

[0017] The present invention also provides a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by the processor, the machine-executable instructions cause the processor to implement the distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as described above.

[0018] Generally speaking, compared with the prior art through the above technical solutions conceived by the present invention, the distributed heterogeneous workshop scheduling method and device provided by the present invention mainly have the following beneficial effects:

[0019] 1. A decentralized multi-agent architecture is adopted to model the dynamic distributed heterogeneous workshop scheduling problem to obtain a partially observable Markov decision process scheduling model, breaking through the limitations of traditional centralized modeling in terms of scalability. Moreover, the scheduling model takes into account the heterogeneous characteristics and the disturbances of uncertain events existing in distributed production, and has high fault tolerance and high scalability.

[0020] 2. Each workshop in distributed manufacturing is regarded as an agent, and each agent makes scheduling decisions based on its own information and limited communication, which well balances the effect and cost, has a high degree of flexibility and autonomy, is very suitable for large-scale and dynamic distributed manufacturing systems, and has high fault tolerance and high scalability.

[0021] 3. The initial scheduling scheme is searched based on the greedy heuristic enhanced reverse exchange search method. In this way, the advantages of the heuristic algorithm in terms of solution effect are utilized. While maintaining a second-level execution speed, the scheduling decisions made by the agents are further improved, greatly improving the production efficiency.

[0022] 4. The set reward function is used for the training of the double - layer end - to - end policy network, which solves the problem of low learning efficiency caused by sparse rewards, making the reinforcement learning method easier to learn the optimal scheduling strategy.

[0023] 5. The present invention can effectively solve the dynamic events and scale - expansion problems that are difficult to handle by traditional scheduling methods, and can output scheduling solutions for large - scale problems within seconds, significantly improving the scheduling efficiency and enterprise production benefits in the distributed manufacturing process. Brief Description of the Drawings

[0024] Figure 1 is a flowchart of a distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement provided by the present invention;

[0025] Figure 2 is a schematic diagram of the overall architecture of the double - layer end - to - end policy network provided by the present invention;

[0026] Figure 3 is a flowchart of the simulation environment operation involved in the double - layer end - to - end policy network provided by the present invention;

[0027] Figure 4 is a schematic diagram of the double - layer end - to - end policy network provided by the present invention. Detailed Embodiments

[0028] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0029] Decentralized scheduling is a way of distributing the decision - making power of scheduling tasks to multiple independent sub - nodes or subsystems, which has high flexibility and autonomy, and is particularly suitable for large - scale, dynamic distributed systems, improving the fault tolerance and scalability of the system. In recent years, deep reinforcement learning (DRL) has shown great potential in the workshop scheduling problem. Through interaction with the environment, the DRL agent can learn the optimal scheduling strategy, and the trained DRL agent can make scheduling decisions quickly, showing excellent response speed and stable performance in dynamic scenarios.

[0030] Please refer to Figure 1, the present invention provides a distributed heterogeneous job shop scheduling method enhanced based on greedy heuristics. The scheduling method can effectively solve the dynamic events and scale expansion problems that are difficult to handle by existing scheduling methods, and can output a scheduling solution for large-scale problems within seconds, which can significantly improve the scheduling efficiency and enterprise production benefits in the distributed manufacturing process.

[0031] The scheduling method mainly includes the following steps:

[0032] Step 1, use a decentralized multi-agent architecture to model the dynamic distributed heterogeneous job shop scheduling problem as a partially observable Markov decision process scheduling model. Among them, the scheduling model is set with a two-dimensional fusion state space that comprehensively considers heterogeneity and permutation characteristics, an end-to-end action space, and a reward function based on a single job shop dimension.

[0033] Among them, use a decentralized multi-agent architecture to model the dynamic distributed heterogeneous job shop scheduling problem as a partially observable Markov decision process scheduling model to break through the limitations of traditional centralized modeling in terms of scalability and optimize the makespan.

[0034] The two-dimensional fusion state space integrates the state characteristics of the job pool dimension and the job shop dimension, namely the job pool characteristics and the job shop characteristics. The job pool characteristics describe the job information that can be selected by the current job shop agent, and the job shop characteristics describe the operating states of different heterogeneous job shops.

[0035] The job pool characteristics adopt 9 priority feature indicators to describe the matching degree between the job and the processing capacity of the current job shop. Specifically:

[0036] 1) FT1 = n top_w , indicating the number of jobs in the job pool, which is used to measure the size of the action space at the current moment;

[0037] 2) FT2 = AVG cf,j , indicating the average total processing time of job j, which is used to measure the occupation degree of the machine when job j is processed in job shop cf; AVG cf,j The calculation formula of is as follows:

[0038]

[0039] 3) FT3 = r j , indicating the release time of job j, which is used to measure the arrival time of the job; if the arrival time of job 1 is 10, then FT3 = r j = 10.

[0040] 4) FT4 = AVG cf,j + STD cf,j, representing the sum of the average and variance of the total processing time of the workpiece, which takes into account the degree of deviation from the average processing time; STD cf,j The calculation formula is as follows:

[0041]

[0042] 5) FT5 = AVG cf,j + STD cf,j + abs(SKE cf,j ), representing the sum of the mean, deviation, and absolute value of skewness of the workpiece processing time, used to represent the degree of data deviation of the processing time; SKE cf,j The calculation formula is as follows:

[0043]

[0044] 6) FT6 = WWD cf,j , representing the weighted weight distribution index, used to measure the matching degree and priority between the workpiece and the workshop machine; WWD cf,j The calculation formula is as follows:

[0045] WWD cf,j = WWDB cf,j - WWDF cf,j

[0046]

[0047]

[0048] 7) FT7 = RFL cf,j , representing the ratio of the processing time of workpiece j on the first machine and the last machine, used to measure the difference between the first process and the last process; RFL cf,j The calculation formula is as follows:

[0049]

[0050] 8) FT8 = EC cf,j , representing the estimated completion time. Assuming that the workpiece has been arranged for processing in workshop cf, estimate its completion time; The calculation formula for the estimated completion time is as follows:

[0051]

[0052] 9) FT9 = EMU cf,j , representing the estimated machine utilization rate. Assuming that the workpiece has been arranged for processing in workshop cf, estimate the machine utilization rate of the workshop.

[0053] Among them, cf represents the workshop that needs to be scheduled at the current decision point; j represents the subscript of the workpiece in the workpiece pool; AVG cf,j represents the average total processing time of workpiece j in workshop cf; STD cf,j represents the standard deviation of the processing time of workpiece j in workshop cf; SKE cf,j represents the skewness of the processing time of workpiece j in workshop cf; WWD cf,j represents the weighted weight distribution index of workpiece j in workshop cf; RFL cf,j represents the ratio of the processing time of workpiece j on the first machine and the last machine; EC cf,j represents the estimated completion time of workpiece j processed in workshop cf; EMU cf,j represents the estimated machine utilization rate of workpiece j processed in workshop cf; n cf represents the number of workpieces that have been processed in workshop cf; w represents the maximum capacity of the workpiece pool; EC l represents the estimated completion time of workshop l; EMU l represents the estimated machine utilization rate of workshop l.

[0054] These 9 features are all dynamic features. The workpiece pool features are represented as a multi-dimensional vector, denoted as h JF ; Each row of the vector represents the features of a workpiece in the workpiece pool.

[0055] For the workshop features, 4 features are set to measure the operating state of the workshop to reflect the load situation of each workshop, including:

[0056] 1) FT 10 = 1 / f, which represents the reciprocal of the number of workshops and reflects the number of agents;

[0057] 2) FT 11 = n l , which represents the number of allocated workpieces in the workshop and indirectly reflects the load situation of the workshop;

[0058] 3) FT 12 = EC l , which represents the estimated completion time of workshop l and directly reflects the load situation of the workshop;

[0059] 4) FT 13 = EMU l , which represents the estimated machine utilization rate of workshop l and indicates the idle and occupied situation of the machines in the workshop;

[0060] Among them, FT 10 is a static feature, and the rest are dynamic features.

[0061] The action space contains all the workpieces in the current workpiece pool. All the workpieces that have arrived at the distributed workshop are sorted in ascending order of the total processing time, and the first w workpieces are selected to form the workpiece pool, which is the action space.

[0062] A reward function based on the single workshop dimension is set to solve the problems of sparse rewards and overly large differences in reward values. The reward function is used for the training of the double-layer end-to-end policy network, and its calculation formula is:

[0063]

[0064] where r(s t ,a t ) represents the reward obtained by taking action a t in state s; EC t represents the estimated completion time after the workshop cf executes the action; m cf,j(st+1) represents the number of machines in the workshop cf; p cf represents the processing time of workpiece j on machine i in the workshop cf; N cf,i,j represents the set of initially arrived workpieces. init

[0065] Step 2: Build a double-layer end-to-end policy network based on the reinforcement learning algorithm. The double-layer end-to-end policy network includes a feature extraction network and an action network. The feature extraction network is used to process the workpiece pool features, and the action network is used to generate the selection probability of each workpiece according to the features obtained by the feature extraction network and the workshop state features. The training environment of the double-layer end-to-end policy network is constructed based on the partially observable Markov decision process scheduling model.

[0066] A decentralized multi-agent architecture is adopted to achieve decentralized scheduling, enabling each workshop to act as an independent agent to complete the scheduling decision. Limited communication is allowed between agents to share partial state information, achieving a balance between scheduling performance and communication cost. Each workshop is regarded as an agent, and an agent scheduler is used to determine the execution order of multiple agents and share the state. The agent observes the environment and inputs the state features into the double-layer end-to-end policy network to output the scheduling decision. After the scheduling decisions of multiple agents, it transitions to the next state. The double-layer end-to-end policy network is used to output the scheduling decision, and the feature extraction network is used to capture the internal relationship between tasks and heterogeneous workshops, thereby enhancing the self-learning ability and policy evolution ability of the agent.

[0067] Build a double-layer end-to-end policy network based on the actor-critic reinforcement learning method. Among them, the actor network refers to the designed double-layer end-to-end policy network, including a feature extraction network and an action network. The feature extraction network is used to process the workpiece pool features, and the action network is used to generate the selection probability of each workpiece. The critic network is used to evaluate the advantage of the decisions taken.​

[0068] Step 3: Input the state features of the distributed heterogeneous workshop to be scheduled into the double-layer end-to-end policy network. The double-layer end-to-end policy network outputs the optimal workpiece to be processed to form an initial scheduling plan. Then, search the initial scheduling plan based on the reverse exchange search method enhanced by the greedy heuristic to obtain the final scheduling plan. Among them, the state features are obtained based on the partially observable Markov decision process scheduling model.

[0069] Please refer to Figure 2 、 Figure 3 and Figure 4 , which specifically includes the following steps:

[0070] 1) The scheduling simulation environment adopts a decentralized scheduling idea. Using a decentralized multi-agent architecture, it allows limited communication between agents to share partial state information to achieve a balance between scheduling performance and communication cost. Each workshop is regarded as an agent, and an agent scheduler is used to determine the execution order of multiple agents and share the state. The agent observes the environment and inputs the state features into the double-layer end-to-end policy network to output the scheduling decision. After multiple agent scheduling decisions, it transitions to the next state. The execution process of the multi-agent simulation environment is as Figure 3 shown, and the specific steps are as follows:

[0071] Step 1: Initialize the environment according to a specific DHPFSP instance, and set up the factory, tasks, machines, and agent scheduler;

[0072] Step 2: Check whether all tasks have been completed. If so, go to Step 10; otherwise, continue to execute Step 3;

[0073] Step 3: Move to the next decision point, and the agent scheduler selects a qualified factory and generates a queue;

[0074] Step 4: If the factory queue is not empty, go to Step 5; otherwise, continue to Step 8;

[0075] Step 5: Set the first factory in the queue as the current factory;

[0076] Step 6: Use a heuristic algorithm to construct a task pool to form the action space of the agent;

[0077] Step 7: The agent makes a decision based on the task pool and factory features and selects a task to process;

[0078] Step 8: Calculate the reward;

[0079] Step 9: State transition, output the state, reward, action, and the next state;

[0080] Step 10: End, output the final scheduling plan and the makespan.

[0081] 2) Construct a two - layer end - to - end policy network based on the actor - critic method. In the two - layer end - to - end policy network, the steps of a single decision - making process are as follows:

[0082] Step 1. The agent observes the environment, observes the job pool and the workshop to obtain the workpiece pool feature h JF ={h JF1 , h JF2 ,..., h JFw} and the factory feature h FF ;

[0083] Step 2. Flatten the workpiece pool feature h JF into a one - dimensional vector, and then input it into the feature extraction network to obtain the workpiece pool global feature vector h GJF , and the corresponding formula is:

[0084] h GJF = MLP ω [flatten(h JF )]

[0085] where MLP ω represents a multi - layer perceptron (MLP) with parameters ω.

[0086] Step 3. The action network generates a stochastic policy π θ (a t |s t ), generates a score for each optional action a t , and then uses the softmax function to output the probability distribution of each workpiece being selected; input the workpiece pool global feature vector h GJF and the factory feature vector h FF into the action network to obtain the stochastic policy π θ (a t |s t ):

[0087] μ(a t |s t ) = MLP θ [h GJF ‖h FF

[0088] where the operation [·||·] represents the concatenation operation of two vectors; the calculation formula for the selection probability of an action a t is as follows:

[0089]

[0090] ​Step 4. The critic network takes [h GJF ||h FF as the input and outputs a scalar υ φ (s t ) as the state estimate value.

[0091] 3) The decision is searched by using a greedy heuristic-based reverse exchange search. After the agent outputs the optimal action, the selected workpiece j is placed at the last position of the current shop scheduling sequence and its estimated completion time is calculated. Then, this workpiece j is exchanged with all possible previous positions, and it is necessary to ensure that the start processing time of the workpiece at the first machine at the exchange position is greater than the release time of workpiece j. If the completion time is improved after the exchange, the exchange position is recorded. Finally, workpiece j is exchanged with the workpiece at the optimal exchange position to improve the scheduling scheme.

[0092] The present invention also provides a distributed heterogeneous shop scheduling system enhanced based on greedy heuristic. The system includes a memory and a processor. The memory stores a computer program. When the processor executes the computer program, it executes the distributed heterogeneous shop scheduling method enhanced based on greedy heuristic as described above.

[0093] The present invention also provides a computer-readable storage medium. The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions cause the processor to implement the distributed heterogeneous shop scheduling method enhanced based on greedy heuristic as described above.

[0094] Those skilled in the art can easily understand that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement, characterized in that: The method comprises the following steps: The state characteristics of the distributed heterogeneous workshops to be scheduled are input into the double-layer end-to-end strategy network, which outputs the optimal workpiece to be processed, and then obtains the final scheduling plan; Among them, the training environment of the two-layer end-to-end policy network is constructed based on the partially observable Markov decision process scheduling model; the state characteristics are obtained based on the partially observable Markov decision process scheduling model; the partially observable Markov decision process scheduling model is obtained by modeling the dynamic distributed heterogeneous workshop scheduling problem using a decentralized multi-agent architecture.

2. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as claimed in claim 1, characterized in that: The state characteristics of the distributed heterogeneous workshops to be scheduled are input into the double-layer end-to-end strategy network, and the double-layer end-to-end strategy network outputs the optimal workpiece to be processed to form an initial scheduling plan; Afterwards, the initial scheduling solution is searched based on a greedy heuristic enhanced reverse exchange search method to obtain a final scheduling solution.

3. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as claimed in claim 1, characterized in that: The two-layer end-to-end policy network is built based on the reinforcement learning algorithm, which includes a feature extraction network and an action network. The feature extraction network is used to process the features of the workpiece pool, and the action network generates the selection probability of each workpiece based on the features obtained by the feature extraction network and the workshop status features.

4. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as claimed in claim 1, characterized in that: Each workshop is regarded as an intelligent agent, and an intelligent agent scheduler is used to determine the execution order of multiple agents and share the state; the agent observes the environment and inputs the state characteristics into a two-layer end-to-end strategy network to obtain scheduling decisions. After multiple agents make scheduling decisions, they transition to the next state.

5. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement according to any one of claims 1 to 4, characterized in that: A two-layer end-to-end policy network is constructed based on the actor-critic reinforcement learning method; the actor network refers to the two-layer end-to-end policy network, and the critic network is used to evaluate the advantages of the decisions taken.

6. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement according to any one of claims 1 to 4, characterized in that: The scheduling model is set up with a two-dimensional fusion state space that comprehensively considers heterogeneity and permutation characteristics, an end-to-end action space, and a reward function based on a single workshop dimension.

7. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as claimed in claim 6 is characterized by: The two-dimensional fusion state space integrates the state characteristics of the work pool dimension and the workshop dimension, which are the workpiece pool characteristics and the workshop characteristics. The workpiece pool characteristics describe the workpiece information that can be selected by the current workshop agent, and the workshop characteristics describe the operating status of each heterogeneous workshop.

8. The distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as claimed in claim 6, characterized in that: The action space contains all the workpieces in the current workpiece pool. All the workpieces that have arrived at the distributed workshop are sorted in ascending order according to the total processing time. The first w workpieces are taken to form the workpiece pool, namely the action space. The reward function is used for training the two-layer end-to-end policy network.

9. A distributed heterogeneous workshop scheduling system based on greedy heuristic enhancement, characterized by: The system includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement according to any one of claims 1 to 8 is executed.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores machine-executable instructions. When the machine-executable instructions are called and executed by a processor, the machine-executable instructions prompt the processor to implement the distributed heterogeneous workshop scheduling method based on greedy heuristic enhancement as described in any one of claims 1 to 8.

Images