A production line dynamic scheduling method based on graph representation learning

Abstract

The application discloses a production line dynamic scheduling method based on graph representation learning, constructs a production line dynamic scheduling model according to the resource type quantity of the production line, resource requirements of an order corresponding task, task duration and arrival time, extracts hidden features required by production line scheduling through state coding and feature embedding representation, forms a corresponding state graph representation, forms an initial solution, namely an initial state of scheduling, according to a first-in first-out rule, selects a region needing rescheduling each time in the state graph through a region selector module, selects a suitable exchange action to reschedule the region output by the region selector through a rule selector module, repeatedly inputs the scheduling solution into the scheduling selector and the rule selector module in sequence, and obtains a final solution after a termination condition is met, and the application can obtain an excellent production line scheduling scheme.

Description

Technical Field

[0001] This invention belongs to the field of production line scheduling technology, specifically relating to a dynamic production line scheduling method based on graph representation learning. Background Technology

[0002] Dynamic scheduling of production lines is a prevalent problem in manufacturing, particularly in industries such as automotive, electronics, steel, home appliances, and aerospace. It requires efficient scheduling under conditions of conflicting production goals and limited resources. Previous research has largely addressed these issues using carefully designed heuristics. A review of existing research reveals a typical process: first, a clever heuristic is proposed for a simplified model of the problem; then, meticulous testing and adjustments are made to achieve good performance in practice. Once aspects of the problem (such as workload or other metrics) change, this process must be repeated to accommodate these changes. Existing solutions to dynamic production line scheduling problems employ reactive priority rules or metaheuristic algorithms, which suffer from drawbacks such as insufficient intelligence (unable to intelligently respond to different operating conditions) or the need for extensive iterative optimization (leading to reduced efficiency and failing to meet the reaction time requirements of dynamic scheduling). Summary of the Invention

[0003] In view of this, the present invention provides a production line dynamic scheduling method based on graph representation learning, which can obtain a production line scheduling scheme with excellent results.

[0004] The technical solution for implementing the present invention is as follows:

[0005] A dynamic scheduling method for production lines based on graph representation learning is proposed. This method constructs a dynamic scheduling model for the production line based on the resource types and quantities of the production line, the resource requirements of the tasks corresponding to the orders, the task duration, and the arrival time. It extracts the implicit features required for production line scheduling through state encoding and feature embedding representation, forming a corresponding state graph representation. An initial solution, i.e., the initial state of scheduling, is formed according to a first-in, first-out (FIFO) rule. A region selector module selects regions in the state graph that need to be rescheduled each time. A rule selector module selects appropriate exchange actions to reschedule the regions output by the region selector. The scheduling solution is repeatedly input into the scheduling selector and rule selector modules, and the final solution is obtained after a termination condition is met.

[0006] Furthermore, the dynamic scheduling model for the production line is specifically as follows:

[0007] The production line has D resource types. m orders arrive online in discrete time steps. Any order j corresponds to a task with p... j =(pj,1 ,……,p j,D Resource requirements (0≤p) j,n ≤1, n=1,…,D), arrival time is A j The task duration is T. j And make the following assumptions:

[0008] 1) Throughout the entire production scheduling process, the resource requirements of each order's corresponding task are fixed and known upon arrival;

[0009] 2) Each order's corresponding task must run continuously until completion; preemption is not allowed.

[0010] For each order j, define the task start time as B. j The task ends at time C. j Task response ratio H j =C j -A j / T j .

[0011] Furthermore, the order node v j Corresponding order attribute p j A j ,T j B j Embedded into (D×(T) max +1)+1) dimension vector e j In the middle, T max This is the maximum duration of the task; this vector encodes the task attributes and resource occupancy status during the dynamic scheduling process of the entire production line, and the first D dimensions of the vector are p. j =(p j,1 ,p j,2 ,……,p j,D ), D×T j The dimensional vector describes the total amount of resources used by all order tasks during the execution of order j, using... Let p' be the total amount of resources used by all tasks at each time step t. t =(p′) t,1 ,p′ t,2 ,……,p′ t,D ), usually there is T j <T max The remaining D×(T) max -T j ) dimensional vectors are all 0, vector e j The last dimension represents the task response ratio H corresponding to order j in the current scheduling table. j ;

[0012] Calculate v for each order node jThe implicit feature h j For each order node v j C(j) is the set of all its child nodes, (h1,c1), (h2,c2), ..., (h... k ,c k If is the LSTM state of its child node, then its LSTM state is:

[0013]

[0014] Where h j For order node v j The implicit features of c j For order node v j The cell state, h k For child node v k The implicit features of c k For child node v k cellular state;

[0015] The final result is the scheduling state diagram and the order node v in the diagram. j The corresponding embedding vector e j With latent feature h j .

[0016] Furthermore, given state s t s t Includes any current order node v j The corresponding embedding vector e j With latent feature h j The region selector is for each region ω t ∈Ω(s t Calculate a fraction Q(s) t ,ω t ), Q(s) t ,ω t The current state s was measured. t Lower region ω t The benefits of rescheduling: a high score indicates the importance of s. t [ω t Rescheduling is preferable, Ω(s) t ) is a set of regions related to dynamic scheduling of the production line. It covers all order nodes that can be used for rescheduling, and is determined according to the score Q(s). t ,ω t Output a probability distribution π ω (ω t |s t )

[0017]

[0018] Select a region ω that needs to be rescheduled t0 Output to the rule selector; region ω t Covering order node v j0 There are W order nodes before and after the center (in order of arrival), meaning that each order node can exchange positions with a maximum of W order nodes before and after it during each rescheduling process.

[0019] Furthermore, given the state and region s t [ω t0 Rescheduling is required; the rule selector predicts a probability distribution π over the entire rule set u. u (u t |s t [ω t0 ]), and select a rule u t0 ∈u applied to s t [ω t The rescheduling action is equivalent to rescheduling the current order node v. j0 Move to region ω in the scheduling state diagram t0 Another order node v j0′ Or a child node of v0, that is, at the time A when order j0' is completed or when order j0 arrives. j0 Immediately assign the order j0; a new scheduling solution (scheduling state) s will then be obtained. t+1 .

[0020] Beneficial effects:

[0021] 1. This invention proposes a novel dynamic scheduling method for production lines based on graph representation learning, which demonstrates excellent performance in production line scheduling. This method employs directed acyclic graphs for state encoding and feature embedding representation, obtaining deeper features of the production line state. Based on this, the method iteratively improves the current scheduling solution through two components: a region selector and a rule selector, continuously enhancing the quality of the scheduling solution and ultimately obtaining a suboptimal solution for production line scheduling.

[0022] 2. Experiments demonstrate that the scheduling performance of the method of this invention is superior to various heuristic scheduling rules and the general operations research optimization tool Google OR-tools, proving the effectiveness of the method. Attached Figure Description

[0023] Figure 1 Examples of scheduling schemes and their corresponding graphical representations;

[0024] Figure 2 The process of embedding the order task;

[0025] Figure 3 This represents an instance of a region selector module.

[0026] Figure 4 The results show the comparative experiments between this method and various scheduling methods. Detailed Implementation

[0027] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0028] This invention proposes a dynamic scheduling method for production lines based on graph representation learning, the specific implementation of which is as follows:

[0029] 1. Construct a dynamic scheduling model for the production line.

[0030] Consider a production line with D types of resources. m orders arrive online in discrete time steps. The task corresponding to any order j has p... j =(p j,1 ,……,p j,D Resource requirements (0≤p) j,n ≤1, n=1,…,D), arrival time is A j The task duration is T. j And make the following assumptions:

[0031] 1) Throughout the entire production scheduling process, the resource requirements of each order's corresponding task are fixed and known upon arrival;

[0032] 2) Each order's corresponding task must run continuously until completion, and preemption is not allowed.

[0033] For each order j, define the task start time as B. j The task ends at time C. j Task response ratio H j =C j -A j / T j This invention uses a graph representation learning-based dynamic production line scheduling method to determine a scheduling timetable for all orders, minimizing the average task response ratio H. The average task response ratio is defined as follows:

[0034]

[0035] 2. State encoding and feature embedding representation

[0036] like Figure 1 As shown, the scheduling state is represented as a directed acyclic graph at each time step, which describes the dependencies between different order scheduling times (orders). Specifically, each order j is represented as an order node v in the directed acyclic graph. j At the same time, add a v0 node as the starting node, if order j arrives at time A jImmediately dispatched (i.e., B) j =A j If ), then add a directed edge to the graph. <v0,v j Otherwise, there must exist at least one order j′ such that C j′ =B j (That is, order j starts immediately after order j' ends), add an edge to such orders. <v j′ ,v j >See the diagram.

[0037] like Figure 2 As shown, the order node v j Corresponding order attribute p j A j ,T j B j Embedded into (D×(T) max +1)+1) dimension vector e j In the middle, T max This represents the maximum duration of the task. This vector encodes the task attributes and resource occupancy status during the dynamic scheduling process of the entire production line. The first D dimensions of the vector are p. j =(p j,1 ,p j,2 ,……,p j,D ), D×T j The dimensional vector describes the total amount of resources used by all order tasks during the execution of order j, using... Let p' be the total amount of resources used by all tasks at each time step t. t =(p′) t,1 ,p′ t,2 ,……,p′ t,D ), usually there is T j <T max The remaining D×(T) max -T j ) dimensional vectors are all 0, vector e j The last dimension represents the task response ratio H corresponding to order j in the current scheduling table. j .

[0038] Next, calculate v for each order node. j The implicit feature h j For each order node v j C(j) is the set of all its child nodes, (h1,c1), (h2,c2), ..., (h... k ,c k If is the LSTM state of its child node, then its LSTM state is:

[0039]

[0040] Where h j For order node v j The implicit features of c j For order node v j The cell state, h k For child node v k The implicit features of c k For child node v k The cellular state.

[0041] The final result is the scheduling state diagram and the order node v in the diagram. j The corresponding embedding vector e j With latent feature h j .

[0042] 3. Form an initial solution according to the first-in-first-out rule, which is the initial state of the scheduling.

[0043] The initial scheduling state s0 is obtained by following the first-in-first-out heuristic scheduling rule, which states that the first order to arrive is scheduled first.

[0044] 4. Region selector module: Selects the region in the state graph that needs to be rescheduled each time.

[0045] Given state s t s t Includes any current order node v j The corresponding embedding vector e j With latent feature h j ,like Figure 3 As shown, the region selector is for each region ω t ∈Ω(s t Calculate a fraction Q(s) t ,ω t It measures the current state s. t Lower region ω t The benefits of rescheduling: a high score indicates the importance of s. t [ω t Rescheduling is preferable, Ω(s) t ) is a set of regions related to dynamic scheduling of the production line. It covers all order nodes that can be used for rescheduling, and is determined according to the score Q(s). t ,ω t Output a probability distribution π ω (ω t |s t )

[0046]

[0047] And select a region ω that needs to be rescheduled.t0 The output is given to the rule selector. Region ω t Covering order node v j0 There are W order nodes before and after the center (in order of arrival), meaning that each order node can exchange positions with a maximum of W order nodes before and after it during each rescheduling process.

[0048] 5. The rule selector module selects the appropriate swap action to reschedule the regions output by the region selector.

[0049] Given a state and region s t [ω t0 Rescheduling is required; the rule selector predicts a probability distribution π over the entire rule set u. u (u t |s t [ω t0 ]), and select a rule u t0 ∈u applied to s t [ω t The rescheduling action is equivalent to moving the current order node v... j0 Move to region ω in the scheduling state diagram t0 Another order node v j0′ Or a child node of v0, that is, at the time A when order j0' is completed or when order j0 arrives. j0 Immediately assign order j0. This will yield a new scheduling solution (scheduling state) s. t+1 .

[0050] 6. Repeatedly input the scheduling solution into the region selector and rule selector modules to obtain a rescheduling sequence.

[0051] (s0,(ω0,u0)),(s1,(ω1,u1)),…,(s T-1 ,(ω T-1 ,u T-1 )),s T (4) When Q(s) t ,ω t )<∈ or rule u t If unavailable, the rescheduling process terminates immediately, and the final solution s is obtained by setting ∈ = 0. T .

[0052] The following implementation effects can be achieved using this method:

[0053] The proposed graph representation learning-based dynamic scheduling method for production lines is compared with various heuristic scheduling rules and the general operations research optimization tool Google OR-tools. The heuristic scheduling rules used are as follows:

[0054] (1) First In First Out (FIFO): Orders that arrive first are scheduled first;

[0055] (2) Shortest Processing Time First (SPTF): Orders with the shortest processing time are scheduled first.

[0056] (3) Longest Processing Time First (LPTF): Orders with the longest processing time are scheduled first.

[0057] (4) Least Resource Occupation First (LROF): Prioritize processing orders with the least resource consumption;

[0058] (5) Most Resource Occupation First (MROF): Prioritize processing orders with the highest resource usage;

[0059] (6) Least Waiting Time First (LWTF): Prioritize processing orders with the shortest remaining task waiting time;

[0060] (7) Random scheduling RAMDOM: Order priorities are randomly assigned.

[0061] Experimental results are as follows Figure 4 As shown, the results indicate that the production line dynamic scheduling method based on graph representation learning (represented by GRL in the graph) proposed in this invention has the smallest average task response ratio H, and its scheduling effect is better than other methods, proving the effectiveness of this method.

[0062] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A dynamic scheduling method for production lines based on graph representation learning, characterized in that, Based on the resource types and quantities of the production line, the resource requirements of the corresponding tasks in the order, the task duration, and the arrival time, a dynamic scheduling model for the production line is constructed. Implicit features required for production line scheduling are extracted through state encoding and feature embedding representation to form a corresponding state diagram representation. An initial solution, i.e., the initial state of scheduling, is formed according to the first-in, first-out (FIFO) rule. The region selector module selects the regions in the state diagram that need to be rescheduled each time. The rule selector module selects appropriate exchange actions to reschedule the regions output by the region selector. The scheduling solution is repeatedly input into the region selector and rule selector modules sequentially, and the final solution is obtained after the termination condition is met. The specific dynamic scheduling model for the production line is as follows: Setting up production lines has Types of resources, production lines Each order arrives online at discrete time steps; any order... The corresponding task has resource requirements, Arrival time is The task duration is And make the following assumptions: 1) Throughout the entire production scheduling process, the resource requirements of each order's corresponding task are fixed and known upon arrival; 2) Each order's corresponding task must run continuously until completion; preemption is not allowed. For each order Define the task start time as The mission ends at [time]. Task response ratio ; Order Node Corresponding order attributes Embedded into dimensional vector Among them This is the maximum duration of the task; This vector encodes the task attributes and resource occupancy status during the dynamic scheduling process of the entire production line. dimensional vector is , dimensional vectors describe orders The total amount of resources used by all order tasks during their execution period, using Let represent the time step, where each time step is denoted as . The total amount of resources used by all tasks is expressed as follows: , usually have The remaining All dimensional vectors are set to 0, and the vectors are... The last dimension represents the orders currently in the scheduling table. Corresponding task response ratio ; Calculate each order node Implicit features For each order node , It is the set of all its child nodes. If the LSTM state of its child node is given, then its LSTM state is: in For order nodes The implicit features, For order nodes The state of cells, child node Implicit features, child node cellular state; The final result is the scheduling state diagram and each order node in the diagram. corresponding embedding vector With implicit features .

2. The production line dynamic scheduling method based on graph representation learning as described in claim 1, characterized in that, Given state , Includes any current order node corresponding embedding vector With implicit features The region selector is for each region. Calculate a fraction , The current state was measured. Lower region The benefits of rescheduling: a high score indicates... Rescheduling is advisable. It is a set of regions related to dynamic scheduling of the production line, encompassing all order nodes that can be used for rescheduling, and is categorized by score. Output a probability distribution Select a region that needs to be rescheduled. Output to the rule selector; region Covering order nodes Before and after the center Each order node can, in each rescheduling process, interact with at most the preceding and following order nodes. The order nodes exchanged positions.

3. The production line dynamic scheduling method based on graph representation learning as described in claim 2, characterized in that, Given state and region Rescheduling is required; the rule selector is used throughout the entire rule set. Predict a probability distribution and select a rule Apply to ; The rescheduling action is equivalent to resetting the current order node. Move to region in the scheduling state diagram Another order node or The child node, that is, in the order Completed or in order Arrival time Assign the order immediately A new scheduling solution is obtained. .

Images