A method, system and device for intelligent scheduling and management of wood processing production

Abstract

The application discloses a kind of wood processing production intelligent scheduling and management method, system and device, it is related to intelligent production scheduling management technical field, comprising: step one: constructing basic scheduling task set;Step two: establish scheduling potential field tree;Step three: execute tree layer potential energy evolution processing, generate evolution scheduling potential field tree;Step four: evolution scheduling potential field tree is input into improved RetNet model, additional tree layer evolution reservation unit, carry out tree layer evolution state transmission to intermediate hidden state, generate scheduling evolution state chain;Step five: generate candidate scheduling path;Step six: based on execution state, candidate scheduling path is carried out rolling optimization processing, generates wood processing production intelligent scheduling scheme;Step seven: production management is carried out and writes back basic scheduling task set, updates wood processing production intelligent scheduling scheme.The application utilizes scheduling potential field tree and improved RetNet model, realizes wood processing production continuous stable scheduling and management.

Description

Technical Field

[0001] This invention relates to the field of intelligent production scheduling and management technology, and in particular to an intelligent scheduling and management method, system and device for timber processing production. Background Technology

[0002] With the increasing digitalization, flexibility, and shortening order delivery cycles in the timber processing industry, technologies for task scheduling, resource allocation, and dynamic management in timber processing production have received widespread attention. Existing timber processing production management methods primarily rely on manual scheduling, fixed-rule sequencing, or scheduling methods based on single equipment load for production arrangement. However, these methods commonly suffer from the following problems in practical applications:

[0003] In timber processing, tasks awaiting scheduling, available processing resources, delivery nodes, and execution status are typically scattered across different production stages. Existing scheduling methods struggle to organize tasks, resources, and delivery requirements into a hierarchical structure that can evolve, resulting in a lack of stable data foundation for the scheduling process. Traditional scheduling methods often allocate tasks based on order sequence, equipment availability, or human experience, failing to accurately express the traction relationship between delivery nodes and tasks, as well as the occupancy and compression relationship between task leaf layers and resource branch layers. This can easily lead to task accumulation on some available processing resources and idleness on others, affecting production continuity. Existing intelligent models, when processing scheduling data, typically treat task status as a normal sequence or static feature input, making it difficult to retain the hierarchical evolution and state transmission relationship between the time root layer, resource branch layer, and task leaf layer. This results in insufficient ability of the model to express changes in delivery traction, resource compression, and task attachment. Furthermore, when new tasks are added, resources change, or execution completion information appears, traditional scheduling schemes lack rolling optimization based on execution status, making it difficult to dynamically adjust unexecuted tasks while locking in executing tasks. This leads to low stability, disturbance resistance, and delivery reliability of the scheduling scheme.

[0004] Therefore, how to provide a method, system, and device for intelligent scheduling and management of wood processing production is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent scheduling and management method, system, and device for timber processing production. This invention unifies the scheduling tasks, available processing resources, delivery nodes, and execution status into a basic scheduling task set. A scheduling potential field tree is formed through a time root layer, resource branch layer, and task leaf layer, enabling the hierarchical expression of dispersed task, resource, and delivery information. Furthermore, the tree-layer potential energy evolution process transmits the delivery traction potential and returns the occupancy squeezing potential, dynamically adjusting the task leaf layer attachment position. An improved RetNet model with added tree-layer evolution retention units is then used to generate a scheduling evolution state chain, combined with rolling optimization processing to form an intelligent scheduling scheme for timber processing production, thereby improving the continuity of production scheduling, the balance of resource allocation, and the stability of management.

[0006] A method for intelligent scheduling and management of timber processing production according to an embodiment of the present invention includes the following steps:

[0007] Step 1: Obtain the tasks to be scheduled, available processing resources, delivery nodes, and execution status in the timber processing production, and construct a basic set of scheduling tasks;

[0008] Step 2: Establish a scheduling potential tree based on the basic scheduling task set. The scheduling potential tree includes a time root layer, a resource branch layer, and a task leaf layer.

[0009] Step 3: Perform tree-level potential evolution processing on the scheduling potential field tree, transfer the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfer the occupancy squeezing potential of the task leaf layer back to the resource branch layer, and adjust the attachment position of the task leaf layer to generate the evolution scheduling potential field tree.

[0010] Step 4: Input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive retention unit between the multi-scale retention unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate a scheduling recursive state chain.

[0011] Step 5: Based on the aforementioned scheduling evolution state chain, generate candidate scheduling paths;

[0012] Step 6: Based on the execution status, perform rolling optimization on the candidate scheduling paths to obtain the current execution path and generate an intelligent scheduling scheme for timber processing production;

[0013] Step 7: Perform production management according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and completion of tasks, write back to the basic scheduling task set, and update the intelligent scheduling scheme for timber processing and production.

[0014] Optionally, step one specifically includes:

[0015] Obtain the tasks to be scheduled in the timber processing production and assign a task number to each task to be scheduled.

[0016] Identify available processing resources in timber processing production and assign a resource number to each available processing resource.

[0017] Obtain delivery nodes in timber processing and production, wherein the delivery node includes the delivery time and the corresponding task number;

[0018] The execution status in wood processing production is obtained, including non-execution status, execution status, and execution completion status.

[0019] The tasks to be scheduled are matched with the delivery nodes according to the task number, and the tasks to be scheduled, available processing resources and execution status are grouped according to the same production time to build a basic scheduling task set.

[0020] Optionally, step two specifically involves:

[0021] Delivery nodes are obtained from the basic scheduling task set, and a delivery time sequence index is established according to the delivery time. Delivery nodes with the same delivery time are grouped into the same time root node to generate a time root layer.

[0022] Available processing resources are obtained from the basic scheduling task set, and a resource acceptance index is established according to the resource number. The available processing resources are converted into resource branch nodes to generate a resource branch layer.

[0023] Obtain the tasks to be scheduled from the basic scheduling task set, establish a task attachment index according to the task number, convert the tasks to be scheduled into task leaf nodes, and generate a task leaf layer.

[0024] Based on the resource acceptance index and the task attachment index, establish a resource attachment chain between the resource branch node and the corresponding task leaf node;

[0025] Based on the task number corresponding to the delivery node, determine the time root node corresponding to the task leaf node, and establish a delivery acceptance chain between the corresponding resource branch node and the time root node through the established resource attachment chain.

[0026] The delivery acceptance chain and resource attachment chain are hierarchically merged according to the time root layer, resource branch layer and task leaf layer to obtain the scheduling potential tree.

[0027] Optionally, step three specifically includes:

[0028] Arrange the time root nodes in the time root layer according to the delivery time sequence index, and generate the delivery traction corresponding to each time root node based on the arrangement position of each time root node in the delivery time sequence index.

[0029] The delivery traction corresponding to each time root node is passed along the delivery acceptance chain to the corresponding resource branch node, and then passed along the resource attachment chain to the corresponding task leaf node.

[0030] Based on the execution status corresponding to the task leaf node, generate the occupancy and squeezing potential of the task leaf layer;

[0031] The occupancy squeeze potential of the task leaf layer is transmitted back to the corresponding resource branch node along the resource link chain, and the number of task leaf nodes under the same resource branch node and the received occupancy squeeze potential are summarized to form the transmission squeeze result corresponding to each resource branch node.

[0032] At the same time, the resource branch nodes corresponding to the root node are arranged from smallest to largest according to the back-passing squeezing result, forming the resource branch node arrangement result.

[0033] At the same root node, the task leaf nodes are arranged according to the delivery traction received by the task leaf nodes, forming the task leaf node arrangement result.

[0034] Based on the arrangement of task leaf nodes, select task leaf nodes in sequence, and in the resource branch nodes corresponding to the root node at the same time, find the resource branch node corresponding to the selected task leaf node through the resource link chain.

[0035] When multiple resource branch nodes are found, the resource branch node that appears first in the list is selected as the target resource branch node.

[0036] When the target resource branch node is not the same as the resource branch node currently attached to the selected task leaf node, the selected task leaf node will be removed from the currently attached resource branch node and attached to the target resource branch node.

[0037] When the target resource branch node is the same as the resource branch node currently attached to the selected task leaf node, the current attachment position of the selected task leaf node is retained.

[0038] Adjust the attachment positions of each task leaf node under the root node at the same time, and generate a recursive scheduling potential tree.

[0039] Optionally, the improved RetNet model specifically includes a potential field node embedding unit, a multi-scale preservation unit, a tree layer evolution preservation unit, a feedforward mapping unit, and a state chain output unit;

[0040] The potential field node embedding unit encodes the time root layer, resource branch layer and task leaf layer as time root layer node embedding, resource branch layer node embedding and task leaf layer node embedding, respectively.

[0041] The multi-scale retention unit performs retention calculations on the embedding of time root layer nodes, resource branch layer nodes, and task leaf layer nodes based on the delivery acceptance chain and resource attachment chain, to obtain the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer.

[0042] The tree-layer evolution preservation unit is set between the multi-scale preservation unit and the feedforward mapping unit. It performs tree-layer evolution state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to obtain the tree-layer evolution hidden state.

[0043] The feedforward mapping unit sequentially performs linear mapping, GELU nonlinear activation, and LayerNorm normalization on the tree-layer recursive hidden state to obtain the scheduling mapping state features.

[0044] The state chain output unit inputs the scheduling mapping state characteristics into the task output branch, resource output branch, and risk output branch, respectively.

[0045] The task output branch generates task priority probabilities through a linear layer and a Softmax function, and arranges the tasks in descending order of priority probabilities to obtain the task priority order.

[0046] The resource output branches generate resource acceptance probabilities through a linear layer and a Softmax function, and arrange the resource acceptance order in descending order of resource acceptance probabilities.

[0047] The risk output branch generates delivery risk values ​​through a linear layer and a sigmoid function, and arranges the delivery risk values ​​from largest to smallest to obtain the delivery risk order;

[0048] Using task priority as the main chain, the resource acceptance order and delivery risk order of the corresponding tasks to be scheduled are written into the same chain node to generate a scheduling evolution state chain.

[0049] Optionally, the tree-level evolution preservation unit specifically includes a downlink evolution branch, an uplink backhaul branch, and a hierarchical state merging branch;

[0050] The downlink recursive branch receives the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer. According to the delivery acceptance chain, the intermediate hidden states of the time root layer are spliced ​​with the intermediate hidden states of the corresponding resource branch layer. After linear mapping and GELU nonlinear activation processing, the root-branch downlink state is obtained.

[0051] The downlink evolution branch splices the root-branch downlink state with the intermediate hidden state of the corresponding task leaf layer according to the resource attachment chain, and performs linear mapping to obtain the branch-leaf downlink state.

[0052] The uplink backhaul branch generates task occupancy backhaul features based on the occupancy squeeze potential of the task leaf layer, and splices the task occupancy backhaul features with the intermediate hidden state of the corresponding resource branch layer according to the resource attachment chain to obtain the leaf-branch backhaul input features.

[0053] The uplink backhaul branch concatenates the leaf-branch backhaul input features with the backhaul squeezing results and the resource branch node arrangement results, and then performs linear mapping and GELU nonlinear activation processing to obtain the leaf-branch backhaul state.

[0054] The hierarchical state converging branch splices the root-branch downlink state, branch-leaf downlink state, and leaf-branch backlink state in the hierarchical order of time root layer, resource branch layer, and task leaf layer, and performs linear mapping, GELU nonlinear activation, and LayerNorm normalization in sequence to generate the tree-level recursive hidden state.

[0055] Optionally, step five specifically includes:

[0056] The chain nodes in the scheduling evolution state chain are traversed sequentially according to the task priority, and the corresponding task to be scheduled, resource acceptance order, and delivery risk order are determined in each chain node.

[0057] In each chain node, the available processing resource that ranks first in the resource acceptance order is taken as the primary processing resource, and the available processing resource that ranks second in the resource acceptance order is taken as the alternative processing resource.

[0058] The delivery risk value corresponding to each chain node is determined according to the delivery risk order. When the delivery risk value is greater than or equal to the set delivery risk threshold, the corresponding task to be scheduled is marked as a risk-priority task. When the delivery risk value is less than the set delivery risk threshold, the corresponding task to be scheduled is marked as a regular deferred task.

[0059] First, connect the chain nodes corresponding to the risk-priority tasks according to the task priority order, and bind the risk-priority tasks, the corresponding main undertaking resources and the corresponding delivery nodes to generate risk-priority scheduling segments.

[0060] Continue connecting the chain nodes corresponding to the regular postponed tasks according to the task priority order, and bind the regular postponed tasks, the corresponding main receiving resources, and the corresponding delivery nodes. At the same time, write the corresponding alternative receiving resources as spatiotemporal adjustment margins into the corresponding nodes to generate regular postponed scheduling segments.

[0061] The risk-priority scheduling segment is arranged before the regular deferred scheduling segment, and the risk-priority scheduling segment and the regular deferred scheduling segment are connected according to the order of the chain nodes in the scheduling evolution state chain to generate candidate scheduling paths.

[0062] Optionally, step six specifically includes:

[0063] Determine the executing and non-executing chain nodes in the candidate scheduling path according to the execution status;

[0064] The executing chain node is rigidly retained in its original position in the candidate scheduling path to lock the current processing state, and the unexecuted chain node is used as the rolling selection node;

[0065] The first rolling selection node after the chain node in the execution is taken as the rolling starting point, and the rolling selection window is formed according to the order of the chain nodes in the candidate scheduling path;

[0066] Within the rolling selection window, the chain node positions in the risk-priority scheduling segment are reserved first, and the main and alternative resources corresponding to the regular delayed task are compared and accepted based on the spatiotemporal adjustment margin written by the corresponding node in the regular delayed scheduling segment.

[0067] When the primary resource corresponding to a regular deferred task is in an unoccupied state, the primary resource will be used as the current resource.

[0068] When the main receiving resource corresponding to the regular deferred task is in the execution state and the alternative receiving resource is in the unoccupied state, the alternative receiving resource will be used as the current receiving resource.

[0069] When both the primary and alternative resources corresponding to a regular deferred task are in the execution state, the primary resource is kept as the current resource, and the task deferred processing is executed.

[0070] According to the connection order of risk-priority scheduling segment first and regular deferred scheduling segment last within the rolling selection window, the tasks to be scheduled, the currently received resources and the delivery nodes corresponding to each chain node are bound to obtain the current execution path;

[0071] Write the current execution path into the scheduling execution list, and generate an intelligent scheduling scheme for timber processing and production based on the scheduling execution list.

[0072] An intelligent scheduling and management system for wood processing production according to an embodiment of the present invention includes the following modules:

[0073] The task aggregation module is used to obtain the tasks to be scheduled, available processing resources, delivery nodes and execution status in the timber processing and production, and to build a basic scheduling task set;

[0074] The potential field tree construction module is used to build a scheduling potential field tree based on the basic scheduling task set. The scheduling potential field tree includes a time root layer, a resource branch layer, and a task leaf layer.

[0075] The tree layer evolution module is used to perform tree layer potential energy evolution processing on the scheduling potential field tree, transfer the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfer the occupancy squeeze potential of the task leaf layer back to the resource branch layer, and adjust the attachment position of the task leaf layer to generate the evolution scheduling potential field tree.

[0076] The state chain solution module is used to input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive preservation unit between the multi-scale preservation unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate a scheduling recursive state chain.

[0077] The path orchestration module is used to generate candidate scheduling paths based on the scheduling evolution state chain;

[0078] The rolling execution module is used to perform rolling optimization of candidate scheduling paths based on the execution status, obtain the current execution path, and generate an intelligent scheduling scheme for timber processing and production.

[0079] The status write-back module is used to manage production according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and execution completion, and write back the basic scheduling task set to update the intelligent scheduling scheme for timber processing and production.

[0080] A smart scheduling and management device for timber processing production includes a processor and a memory. The memory stores computer programs, and the processor calls the computer programs stored in the memory to execute any smart scheduling and management method for timber processing production.

[0081] The beneficial effects of this invention are:

[0082] This invention gathers the tasks to be scheduled, available processing resources, delivery nodes, and execution status into a basic scheduling task set, and further establishes a scheduling potential tree including a time root layer, a resource branch layer, and a task leaf layer. This enables the task, resource, and delivery information in wood processing production to form a unified hierarchical scheduling structure, thereby improving the organizational stability and real-time expressive capability of scheduling data.

[0083] This invention uses tree-layer potential energy evolution processing to transmit the delivery traction potential downwards along the delivery acceptance chain and resource attachment chain, and to transmit the occupancy and compression potential of the task leaf layer upwards along the resource attachment chain, thereby adjusting the attachment position of the task leaf layer, thereby reducing task accumulation under a single available processing resource and improving resource acceptance balance.

[0084] This invention improves the RetNet model by adding a tree-layer evolution preservation unit, enabling the model to express the downlink and uplink propagation relationships between the time root layer, resource branch layer, and task leaf layer during intermediate hidden state stages. This generates a scheduling evolution state chain that includes task priority, resource acceptance order, and delivery risk order. Furthermore, by using rolling optimization to lock execution chain nodes and adjust unexecuted chain nodes, the intelligent scheduling scheme for timber processing production possesses better continuity, anti-disturbance capability, and delivery stability. Attached Figure Description

[0085] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0086] Figure 1 This is an overall flowchart of an intelligent scheduling and management method for timber processing production proposed in this invention;

[0087] Figure 2 This is a schematic diagram of the structure of an intelligent scheduling and management system for wood processing production proposed in this invention. Detailed Implementation

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0089] refer to Figure 1 A method for intelligent scheduling and management of timber processing production includes the following steps:

[0090] Step 1: Obtain the tasks to be scheduled, available processing resources, delivery nodes, and execution status in the timber processing production, and construct a basic set of scheduling tasks;

[0091] Step 2: Establish a scheduling potential tree based on the basic scheduling task set. The scheduling potential tree includes a time root layer, a resource branch layer, and a task leaf layer.

[0092] Step 3: Perform tree-level potential evolution processing on the scheduling potential field tree, transfer the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfer the occupancy squeeze potential of the task leaf layer back to the resource branch layer, and adjust the attachment position of the task leaf layer to generate the evolution scheduling potential field tree.

[0093] Step 4: Input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive retention unit between the multi-scale retention unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate the scheduling recursive state chain.

[0094] Step 5: Generate candidate scheduling paths based on the scheduling recursion state chain;

[0095] Step 6: Based on the execution status, perform rolling optimization on the candidate scheduling paths to obtain the current execution path and generate an intelligent scheduling scheme for timber processing production;

[0096] Step 7: Perform production management according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and completion of tasks, write back to the basic scheduling task set, and update the intelligent scheduling scheme for timber processing and production.

[0097] In this embodiment, step one specifically includes:

[0098] Obtain the tasks to be scheduled in the timber processing production and assign a task number to each task to be scheduled.

[0099] Identify available processing resources in timber processing production and assign a resource number to each available processing resource.

[0100] Obtain delivery nodes in the timber processing and production process. Each delivery node includes the delivery time and the corresponding task number.

[0101] Acquire the execution status in wood processing production, including non-execution status, execution status, and execution completion status;

[0102] The tasks to be scheduled are matched with the delivery nodes according to the task number, and the tasks to be scheduled, available processing resources and execution status are grouped according to the same production time to build a basic scheduling task set;

[0103] This invention establishes a clear correspondence between tasks and delivery nodes by assigning a task number to each task to be scheduled and including the delivery time and corresponding task number in the delivery node, thus avoiding mismatches in task delivery times during scheduling. By assigning a resource number to each available processing resource, the invention ensures that available processing resources have a clear source when building the scheduling potential tree, facilitating the writing of available processing resources into resource branches. The execution status is divided into unexecuted, in-process, and completed states, enabling the basic scheduling task set to distinguish the current production status of different tasks and preventing completed tasks from being repeatedly included in the scheduling process. Grouping tasks to be scheduled, available processing resources, and execution status according to the same production time ensures that the basic scheduling task set reflects the actual production situation at the same moment, providing a stable data foundation and improving the simplicity and real-time performance of scheduling data.

[0104] In this embodiment, step two specifically involves:

[0105] Delivery nodes are obtained from the basic scheduling task set, and a delivery time sequence index is established according to the delivery time. Delivery nodes with the same delivery time are grouped into the same time root node to generate the time root layer.

[0106] Obtain available processing resources from the basic scheduling task set, establish a resource acceptance index according to the resource number, convert the available processing resources into resource branch nodes, and generate resource branch layers;

[0107] Obtain the tasks to be scheduled from the basic scheduling task set, establish a task attachment index according to the task number, convert the tasks to be scheduled into task leaf nodes, and generate task leaf layers.

[0108] Based on the resource attachment index and the task attachment index, establish a resource attachment chain between the resource branch node and the corresponding task leaf node.

[0109] Based on the task number corresponding to the delivery node, determine the time root node corresponding to the task leaf node, and establish a delivery connection chain between the corresponding resource branch node and the time root node through the established resource connection chain.

[0110] The delivery acceptance chain and resource attachment chain are hierarchically merged according to the time root layer, resource branch layer and task leaf layer to obtain the scheduling potential field tree.

[0111] In this invention, the scheduling potential tree is used to organize the delivery nodes, available processing resources, and tasks to be scheduled in the basic scheduling task set into a hierarchical scheduling structure. Specifically, firstly, a delivery time sequence index is established based on the delivery time as the grouping criterion, grouping delivery nodes with the same delivery time into the same time root node, so that the time root layer can represent the scheduling pressure sources corresponding to different delivery times; then, a resource acceptance index is established based on the resource number as the distinguishing criterion, converting each available processing resource into an independent resource branch node, so that the resource branch layer can represent the processing resources that can participate in acceptance at the current production time; simultaneously, a task attachment index is established based on the task number, converting each task to be scheduled into a task leaf node, so that the task leaf layer can represent the specific task objects waiting to be arranged.

[0112] Resource linkage chains represent the connection between resource branch nodes and task leaf nodes, while delivery linkage chains represent the delivery traction relationship between time root nodes and resource branch nodes. By first establishing resource linkage chains and then determining the delivery linkage between corresponding resource branch nodes and time root nodes based on the time root nodes corresponding to task leaf nodes, the scheduling potential field tree can maintain a three-layer tree structure of time root layer, resource branch layer, and task leaf layer. By constructing the scheduling potential field, the originally scattered task, resource, and delivery information is transformed into a recursive hierarchical object, providing a clear transmission path for the tree-layer potential energy recursion processing, enabling the delivery traction potential to be transmitted downwards along the delivery linkage chains and resource linkage chains, and the occupancy squeezing potential to be transmitted upwards along the resource linkage chains.

[0113] In this embodiment, step three specifically includes:

[0114] Arrange the time root nodes in the time root layer according to the delivery time sequence index, and generate the delivery traction corresponding to each time root node based on the arrangement position of each time root node in the delivery time sequence index.

[0115] The delivery traction corresponding to each time root node is passed along the delivery acceptance chain to the corresponding resource branch node, and then passed along the resource attachment chain to the corresponding task leaf node.

[0116] Based on the execution status corresponding to the task leaf node, generate the occupancy and squeezing potential of the task leaf layer;

[0117] The occupancy squeeze potential of the task leaf layer is transmitted back to the corresponding resource branch node along the resource link chain, and the number of task leaf nodes under the same resource branch node and the received occupancy squeeze potential are summarized to form the transmission squeeze result corresponding to each resource branch node.

[0118] At the same time, the resource branch nodes corresponding to the root node are arranged from smallest to largest according to the back-passing squeezing result, forming the resource branch node arrangement result.

[0119] At the same root node, the task leaf nodes are arranged according to the delivery traction received by the task leaf nodes, forming the task leaf node arrangement result.

[0120] Based on the arrangement of task leaf nodes, select task leaf nodes in sequence, and in the resource branch nodes corresponding to the root node at the same time, find the resource branch node corresponding to the selected task leaf node through the resource link chain.

[0121] When multiple resource branch nodes are found, the resource branch node that appears first in the list is selected as the target resource branch node.

[0122] When the target resource branch node is not the same as the resource branch node currently attached to the selected task leaf node, the selected task leaf node will be removed from the currently attached resource branch node and attached to the target resource branch node.

[0123] When the target resource branch node is the same as the resource branch node currently attached to the selected task leaf node, the current attachment position of the selected task leaf node is retained.

[0124] Adjust the attachment positions of each task leaf node under the root node at the same time, and generate a recursive scheduling potential tree;

[0125] In the specific implementation process, the delivery time sequence index is arranged according to the delivery time within the production day. The starting point of the arrangement is set to the current production time. The time root node is numbered sequentially in the delivery time sequence index starting from 1. The smaller the number, the earlier the delivery node is, and the earlier the corresponding delivery traction potential is propagated downwards. The execution status corresponding to the task leaf node includes unexecuted status, executing status, and executed completed status. The unexecuted status is recorded as 1, the executing status as 2, and the executed completed status as 0. The executed completed status refers to the task leaf node that has just been completed within the current production scheduling cycle and has not yet been removed from the basic scheduling task set. Historically completed tasks that have been removed from the current production scheduling cycle are no longer involved in the construction of the scheduling potential field tree. The number of task leaf nodes attached to a corresponding resource branch node is the total number of task leaf nodes currently attached to that resource branch node. The occupancy pressure of a single task leaf node is determined by the execution status value of that task leaf node. The return pressure result for the same resource branch node is obtained by summing the current number of attachments to that resource branch node and the occupancy pressure of each task leaf node under that resource branch node. Specifically, if a resource branch node currently has 3 task leaf nodes attached, and the first task leaf node is in an unexecuted state, then the occupancy pressure of the first task leaf node is 1; the second task leaf node is in an executing state, then the occupancy pressure of the second task leaf node is 2; and the third task leaf node is in an executed state, then the occupancy pressure of the third task leaf node is 0. The return pressure result for this resource branch node is the sum of the number of attachments 3 and the occupancy pressure of the first task leaf node (1), the occupancy pressure of the second task leaf node (2), and the occupancy pressure of the third task leaf node (0), which equals 6. The value 6 is then written into the corresponding resource branch node in the resource branch layer.

[0126] After the resource branch node arrangement and task leaf node arrangement are formed, the task leaf nodes with higher receiving and delivery traction potentials are processed first, followed by those with lower receiving and delivery traction potentials. When a selected task leaf node corresponds to multiple resource branch nodes, the resource branch node with the highest ranking in the resource branch node arrangement is determined as the target resource branch node. When a selected task leaf node corresponds to only one resource branch node, the current attachment position of the selected task leaf node is retained, and no cross-resource branch node adjustment is performed. Through the above processing, the recursive scheduling potential field tree can adjust task leaf nodes from resource branch nodes with larger feedback squeeze results to resource branch nodes with smaller feedback squeeze results while keeping the root node constraints unchanged at the same time and the resource attachment chain correspondence intact. This reduces task backlog under a single available processing resource, and the improved RetNet model's received scheduling structure simultaneously includes delivery traction relationships and resource occupancy feedback relationships, thereby improving the continuity, stability, and executability of the intelligent scheduling scheme for wood processing production.

[0127] In this embodiment, the improved RetNet model specifically includes a potential field node embedding unit, a multi-scale preservation unit, a tree layer evolution preservation unit, a feedforward mapping unit, and a state chain output unit.

[0128] The potential field node embedding unit encodes the time root layer, resource branch layer, and task leaf layer as time root layer node embedding, resource branch layer node embedding, and task leaf layer node embedding, respectively.

[0129] Specifically, the potential field node embedding unit adopts a unified node encoding method for the time root layer, resource branch layer, and task leaf layer, with the node embedding dimension set to 128 dimensions. When encoding the time root layer, the delivery traction potential corresponding to each time root node is used as the input feature of the time root layer. The delivery traction potential is determined by the arrangement position of each time root node in the delivery time sequence index. The time root layer input features are then input into the linear mapping layer to obtain the time root layer node embedding.

[0130] When encoding the resource branch layer, the resource number, back-passing compression result, and resource branch node arrangement result corresponding to each resource branch node are vectorized and concatenated to construct the resource branch layer input feature vector. The resource branch layer input feature vector is then input into a linear mapping layer to obtain the resource branch layer node embedding. When encoding the task leaf layer, the task number, task leaf node arrangement result, and execution status corresponding to each task leaf node are vectorized and concatenated to construct the task leaf layer input feature vector. The task leaf layer input feature vector is then input into a linear mapping layer to obtain the task leaf layer node embedding. Among them, the unexecuted state is encoded as 1, the executing state is encoded as 2, and the execution completed state is encoded as 0.

[0131] The output dimensions of the time root layer node embedding, resource branch layer node embedding, and task leaf layer node embedding are all set to 128 dimensions. Through the above encoding method, the time root layer can express the position of the delivery node in the delivery time sequence index, the resource branch layer can express the pressure and arrangement position of available processing resources, and the task leaf layer can express the arrangement position and execution status of the tasks to be scheduled. This enables the improved RetNet model to obtain hierarchical scheduling features of a unified dimension before entering the multi-scale retention unit, and avoids the inability to perform unified retention calculations for nodes at different levels due to different feature types.

[0132] The multi-scale retention unit performs retention calculations on the embedding of time root layer nodes, resource branch layer nodes, and task leaf layer nodes based on the delivery acceptance chain and resource attachment chain, and obtains the intermediate hidden states of time root layer, resource branch layer and task leaf layer.

[0133] Specifically, the multi-scale retention unit receives 128-dimensional embeddings of time root layer nodes, resource branch layer nodes, and task leaf layer nodes, and establishes inter-layer retention paths along the delivery acceptance chain and resource attachment chain, respectively. The multi-scale retention unit sets four retention scales with scale lengths of 1, 2, 4, and 8, respectively. Scale length 1 is used to retain the scheduling state of a single node, scale length 2 is used to retain the local acceptance state between adjacent resource branch nodes or adjacent task leaf nodes, scale length 4 is used to retain the acceptance distribution state between multiple resource branch nodes under the same time root node, and scale length 8 is used to retain the delivery traction delay state across time root nodes.

[0134] During retention computation, the multi-scale retention unit maps the embedding of each level node into query vector, key vector, and value vector, respectively. Following the retention computation method of the RetNet model at each retention scale, it matches the query vector of the current node with the key vectors of associated nodes on the delivery acceptance chain or resource attachment chain. The matching result is then applied to the value vector of the associated node to obtain the retention response at the corresponding scale. For the time root layer, the retention response mainly comes from adjacent time root nodes under the same delivery time sequence index. For the resource branch layer, the retention response comes from resource branch nodes connected through the delivery acceptance chain under the same time root node. For the task leaf layer, the retention response comes from task leaf nodes connected through the resource attachment chain under the same resource branch node. The retention responses at the four retention scales are concatenated, and the concatenated response features are input into a linear mapping layer for dimension alignment. The output dimension is still set to 128 dimensions, yielding the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer, respectively.

[0135] The tree-layer evolution preservation unit is set between the multi-scale preservation unit and the feedforward mapping unit. It performs tree-layer evolution state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to obtain the tree-layer evolution hidden state.

[0136] The feedforward mapping unit sequentially performs linear mapping, GELU nonlinear activation, and LayerNorm normalization on the tree-layer recursive hidden state to obtain the scheduling mapping state features.

[0137] The state chain output unit inputs the scheduling mapping state characteristics into the task output branch, resource output branch, and risk output branch, respectively;

[0138] The task output branch generates task priority probabilities through a linear layer and a Softmax function, and then arranges the tasks in descending order of priority probabilities to obtain the task priority order.

[0139] The resource output branch generates resource acceptance probabilities through a linear layer and a Softmax function, and arranges the resources in descending order of their acceptance probabilities to obtain the resource acceptance order.

[0140] The risk output branch generates delivery risk values ​​through a linear layer and a sigmoid function, and arranges them in descending order of delivery risk value to obtain the delivery risk order;

[0141] Using task priority as the main chain, the resource acceptance order and delivery risk order of the corresponding tasks to be scheduled are written into the same chain node to generate a scheduling evolution state chain.

[0142] In this invention, the improved RetNet model first converts the time root layer, resource branch layer, and task leaf layer into a unified and computable node embedding through a potential field node embedding unit, enabling the delivery relationship, resource acceptance relationship, and task attachment relationship in the scheduling potential field tree to be processed by the deep model. The multi-scale retention unit performs retention calculations based on the delivery acceptance chain and resource attachment chain, allowing the model to not only focus on the state of a single node but also retain the acceptance relationship between different levels. The tree layer evolution retention unit is set between the multi-scale retention unit and the feedforward mapping unit, allowing the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer to continue to evolve according to the tree layer evolution state transmission method, enhancing the model's ability to express delivery traction, resource squeezing, and changes in task arrangement. The feedforward mapping unit further completes nonlinear mapping and normalization processing, making the scheduling features more stable. The state chain output unit generates the task priority order, resource acceptance order, and delivery risk order respectively, and forms a scheduling evolution state chain with the task priority order as the chain main line, thereby providing a clear correspondence between tasks, resources, and risks for generating candidate scheduling paths, improving the accuracy and executability of the intelligent scheduling scheme for wood processing production.

[0143] In this embodiment, the tree layer evolution preservation unit specifically includes a downlink evolution branch, an uplink backhaul branch, and a hierarchical state merging branch;

[0144] The downlink recursive branch receives the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer. According to the delivery acceptance chain, the intermediate hidden states of the time root layer are concatenated with the intermediate hidden states of the corresponding resource branch layer. After linear mapping and GELU nonlinear activation processing, the root-branch downlink state is obtained.

[0145] The downlink evolution branch splices the root branch downlink state with the intermediate hidden state of the corresponding task leaf layer according to the resource attachment chain, and performs linear mapping to obtain the branch leaf downlink state.

[0146] The uplink backhaul branch generates task occupancy backhaul features based on the occupancy squeeze potential of the task leaf layer, and splices the task occupancy backhaul features with the intermediate hidden state of the corresponding resource branch layer according to the resource attachment chain to obtain the leaf-branch backhaul input features.

[0147] The uplink backhaul branch concatenates the leaf-branch backhaul input features with the backhaul squeezing results and the resource branch node arrangement results, and then performs linear mapping and GELU nonlinear activation processing to obtain the leaf-branch backhaul state.

[0148] The hierarchical state merging branch splices the root-branch downlink state, branch-leaf downlink state, and leaf-branch backlink state in the hierarchical order of time root layer, resource branch layer, and task leaf layer, and performs linear mapping, GELU nonlinear activation, and LayerNorm normalization in sequence to generate tree-level recursive hidden states.

[0149] In the specific implementation process, the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer received by the tree layer recursion preservation unit are all set as 128-dimensional vectors. In the downlink recursion branch, the intermediate hidden states of the time root layer and the corresponding intermediate hidden states of the resource branch layer are concatenated to form a 256-dimensional root-branch concatenation vector. The linear mapping layer compresses the 256-dimensional root-branch concatenation vector into 128 dimensions, and after GELU nonlinear activation processing, the root-branch downlink state is obtained. This processing is used to make the delivery timing information and resource acceptance information in the delivery acceptance chain form a nonlinear association, so that the resource branch layer can receive the information from the time root layer. The delivery of the layer drives the change; subsequently, the root branch down state and the intermediate hidden state of the corresponding task leaf layer are concatenated again to form a 256-dimensional branch-leaf concatenation vector, which is then compressed into a 128-dimensional branch-leaf down state through a linear mapping layer. The branch-leaf down state is not subjected to GELU nonlinear activation processing because this state directly corresponds to the attachment order of the task leaf layer and the arrangement result of the task leaf nodes, and it is necessary to retain the linear ordering relationship between the task leaf nodes; if nonlinear activation is performed again at this position, it is easy to change the order interval between adjacent task leaf nodes, causing the task priority order generated by the state chain output unit to shift.

[0150] In the uplink backhaul branch, the occupancy compression potential of the task leaf layer is first converted into a 128-dimensional task occupancy backhaul feature through a linear mapping layer, and then concatenated with the intermediate hidden state of the corresponding resource branch layer to form a 256-dimensional leaf-branch backhaul input feature. Further, the backhaul compression result and the resource branch node arrangement result are encoded into 128-dimensional compression feature vectors and arrangement feature vectors respectively through their corresponding linear mapping layers. The 256-dimensional leaf-branch backhaul input feature is then concatenated with the 128-dimensional compression feature vector and the 128-dimensional arrangement feature vector to form... A 512-dimensional uplink composite backpropagation feature was constructed. This feature was then input into a linear mapping layer and compressed to 128 dimensions. After GELU nonlinear activation, a 128-dimensional leaf-branch backpropagation state was obtained. The hierarchical state merging branch concatenates the 128-dimensional root-branch downlink state, the 128-dimensional branch-leaf downlink state, and the 128-dimensional leaf-branch backpropagation state into a 384-dimensional hierarchical merging feature. This feature is then compressed to 128 dimensions via linear mapping, and after GELU nonlinear activation and LayerNorm normalization, the tree-layer evolutionary hidden state is generated. Through these processes, the tree-layer evolutionary preservation unit can simultaneously retain the downlink scheduling pull from the time root layer to the task leaf layer, the occupancy feedback from the task leaf layer to the resource branch layer, and the changes in the load-bearing pressure of the resource branch layer. This allows the improved RetNet model to form a state representation in the intermediate layers that better fits the scheduling potential tree structure.

[0151] In this invention, the improved RetNet model is formed by structural modification based on the existing RetNet model's retained computational framework. It retains the multi-scale preservation idea, linear mapping structure, GELU nonlinear activation structure, and LayerNorm normalization structure suitable for long sequence recursion expression in the existing RetNet model. However, the processing method of the existing RetNet model oriented towards ordinary sequence positions is adjusted to a hierarchical state processing method oriented towards the scheduling potential tree. In terms of specific structure, the improved RetNet model consists of potential node embedding units, multi-scale preservation units, tree layer recursion preservation units, feedforward mapping units, and state chain output units, forming a continuous network structure from the input of the scheduling potential tree to the output of the scheduling recursion state chain. The model training data comes from the historical scheduling records and simulation scheduling records of wood processing production. Compared to the existing RetNet model, the improved RetNet model adds a tree-layer recursion preservation unit between the multi-scale preservation unit and the feedforward mapping unit. This enables the model to complete the downlink recursion and uplink backpropagation between the time root layer, resource branch layer and task leaf layer in the intermediate hidden state stage, thereby enhancing the model's ability to express changes in delivery traction, resource squeezing and task attachment, and improving the stability and executability of the generated scheduling scheme.

[0152] In this embodiment, step five specifically includes:

[0153] The chain nodes in the scheduling evolution state chain are traversed sequentially according to the task priority, and the corresponding task to be scheduled, resource acceptance order, and delivery risk order are determined in each chain node.

[0154] In each chain node, the available processing resource that ranks first in the resource acceptance order is taken as the primary processing resource, and the available processing resource that ranks second in the resource acceptance order is taken as the alternative processing resource.

[0155] The delivery risk value corresponding to each chain node is determined according to the delivery risk order. When the delivery risk value is greater than or equal to the set delivery risk threshold, the corresponding task to be scheduled is marked as a risk-priority task. When the delivery risk value is less than the set delivery risk threshold, the corresponding task to be scheduled is marked as a regular deferred task.

[0156] First, connect the chain nodes corresponding to the risk-priority tasks according to the task priority order, and bind the risk-priority tasks, the corresponding main undertaking resources and the corresponding delivery nodes to generate risk-priority scheduling segments.

[0157] Continue connecting the chain nodes corresponding to the regular postponed tasks according to the task priority order, and bind the regular postponed tasks, the corresponding main receiving resources, and the corresponding delivery nodes. At the same time, write the corresponding alternative receiving resources as spatiotemporal adjustment margins into the corresponding nodes to generate regular postponed scheduling segments.

[0158] The risk-priority scheduling segment is arranged before the regular deferred scheduling segment, and the risk-priority scheduling segment and the regular deferred scheduling segment are connected according to the order of the chain nodes in the scheduling evolution state chain to generate candidate scheduling paths;

[0159] In the specific implementation process, each chain node in the scheduling recursion state chain includes the task to be scheduled, the resource acceptance order, and the delivery risk order. The delivery risk value is output by the risk output branch through the Sigmoid function, and the value range is set to 0 to 1. The delivery risk threshold is set by the median of the delivery risk values ​​of all chain nodes in the current scheduling recursion state chain. Specifically, all delivery risk values ​​are arranged from smallest to largest. When the number of chain nodes is odd, the value of the middle position is taken as the delivery risk threshold. When the number of chain nodes is even, the average of the two middle values ​​is taken as the delivery risk threshold. The available processing resource ranked first in the resource acceptance order is used as the primary acceptance resource, and the available processing resource ranked second is used as the alternative acceptance resource, so that each chain node has clear primary resource pointers and alternative resource information before generating candidate scheduling paths.

[0160] The risk-priority scheduling segment does not include alternative receiving resources; instead, it binds only the risk-priority task, its corresponding primary receiving resource, and its corresponding delivery node. This aims to rigidly lock the scheduling tasks with high delivery risk values, preventing high-risk tasks from repeatedly migrating between different resources during the rolling selection process due to frequent switching of alternative receiving resources, thus affecting the continuous execution order of the primary receiving resource. The regular deferred scheduling segment, on the other hand, writes alternative receiving resources as a spatiotemporal adjustment margin into the corresponding node, allowing scheduling tasks with lower delivery risk values ​​to have adjustment space when resources or execution status changes. Through this asymmetric path construction method—rigidly binding the risk-priority scheduling segment and flexibly retaining the regular deferred scheduling segment—the candidate scheduling path can prioritize the execution stability of high-delivery-risk tasks while retaining the dynamic adjustment capability of regular tasks, thereby improving the overall executability and anti-disturbance capability of the intelligent scheduling solution for timber processing and production.

[0161] In this embodiment, step six specifically includes:

[0162] Determine the executing and non-executing chain nodes in the candidate scheduling path according to the execution status;

[0163] The executing chain node is rigidly retained in its original position in the candidate scheduling path to lock the current processing state, and the unexecuted chain node is used as the rolling selection node;

[0164] The first rolling selection node after the chain node in the execution is taken as the rolling starting point, and the rolling selection window is formed according to the order of the chain nodes in the candidate scheduling path;

[0165] Within the rolling selection window, the chain node positions in the risk-priority scheduling segment are reserved first, and the main and alternative resources corresponding to the regular delayed task are compared and accepted based on the spatiotemporal adjustment margin written by the corresponding node in the regular delayed scheduling segment.

[0166] When the primary resource corresponding to a regular deferred task is in an unoccupied state, the primary resource will be used as the current resource.

[0167] When the main receiving resource corresponding to the regular deferred task is in the execution state and the alternative receiving resource is in the unoccupied state, the alternative receiving resource will be used as the current receiving resource.

[0168] When both the primary and alternative resources corresponding to a regular deferred task are in the execution state, the primary resource is kept as the current resource, and the task deferred processing is executed.

[0169] According to the connection order of risk-priority scheduling segment first and regular deferred scheduling segment last within the rolling selection window, the tasks to be scheduled, the currently received resources and the delivery nodes corresponding to each chain node are bound to obtain the current execution path;

[0170] Write the current execution path into the scheduling execution list, and generate an intelligent scheduling scheme for timber processing and production based on the scheduling execution list;

[0171] In practice, the rolling optimization window is set to a length of 8 chain nodes. When the number of unexecuted chain nodes after an executing chain node is less than 8, the actual number of unexecuted chain nodes is used as the length of the rolling optimization window. The rolling optimization window starts from the first rolling optimization node after the executing chain node and only adjusts unexecuted chain nodes that have not yet entered the processing state, so that tasks in progress are not interrupted. The risk-priority scheduling segment maintains the original chain node positions within the rolling optimization window to ensure that tasks with high delivery risk do not drift in order due to resource switching. The regular deferred scheduling segment compares the primary and alternative resources based on the spatiotemporal adjustment margin, allowing tasks with lower delivery risk to take on the role of flexible resource adjustment.

[0172] When performing task assignment comparison, if the primary task assignment resource corresponding to the regular task assignment is in an executing state and the alternative task assignment resource is in an unoccupied state, then the regular task assignment is assigned through the alternative task assignment resource; if both the primary task assignment resource and the alternative task assignment resource corresponding to the regular task assignment are in an executing state, then the production assignment of the regular task assignment on the primary task assignment resource is maintained, and the execution position is postponed one position along the chain node sequence relationship in the candidate scheduling path to generate a task assignment processing result; the scheduling execution list records the task to be scheduled, the current task assignment resource, the delivery node, the execution order, and the assignment execution mark corresponding to each chain node; when generating the intelligent scheduling scheme for timber processing production, the task issuance order is formed according to the execution order in the scheduling execution list, the current task assignment resource is mapped to the specific available processing resource, the delivery time index deadline corresponding to the delivery node is used as the task completion deadline, and the task assignment processing result is used as the assignment execution mark. Through the above processing, the intelligent scheduling solution for timber processing production can simultaneously reflect the locking of tasks in execution, the stable retention of risk-priority tasks, and the flexible adjustment of routinely postponed tasks. This avoids interference with ongoing processing tasks during the rolling selection process and maintains the integrity of the scheduling path through alternative acceptance or postponement when resources are occupied, thereby improving the continuity of production scheduling, its resistance to disturbances, and the stability of delivery.

[0173] refer to Figure 2 A smart scheduling and management system for timber processing production includes the following modules:

[0174] The task aggregation module is used to obtain the tasks to be scheduled, available processing resources, delivery nodes and execution status in the timber processing and production, and to build a basic scheduling task set;

[0175] The potential field tree construction module is used to build a scheduling potential field tree based on the basic scheduling task set. The scheduling potential field tree includes a time root layer, a resource branch layer, and a task leaf layer.

[0176] The tree-level evolution module is used to perform tree-level potential energy evolution processing on the scheduling potential field tree. It transfers the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfers the occupancy squeeze potential of the task leaf layer back to the resource branch layer, and adjusts the attachment position of the task leaf layer to generate the evolution scheduling potential field tree.

[0177] The state chain solution module is used to input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive preservation unit between the multi-scale preservation unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate the scheduling recursive state chain.

[0178] The path orchestration module is used to generate candidate scheduling paths based on the scheduling evolution state chain;

[0179] The rolling execution module is used to perform rolling optimization of candidate scheduling paths based on the execution status, obtain the current execution path, and generate an intelligent scheduling scheme for timber processing and production.

[0180] The status write-back module is used to manage production according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and execution completion, and write back the basic scheduling task set to update the intelligent scheduling scheme for timber processing and production.

[0181] A smart scheduling and management device for timber processing production includes a processor and a memory. The memory stores computer programs, and the processor calls the computer programs stored in the memory to execute any smart scheduling and management method for timber processing production.

[0182] Example 1: To verify the feasibility of this invention in practice, it was applied to the daily production scheduling management of a timber processing workshop. This workshop includes continuous processing areas for sawing, thickness setting, sanding, panel splicing, and packaging, with multiple orders being processed simultaneously each day. Before implementation, the workshop mainly relied on manual experience and equipment availability to schedule tasks. This often resulted in urgent tasks being taken over by regular tasks, some processing resources being continuously queued, and some processing resources being temporarily idle. Especially when new orders were inserted or equipment was temporarily occupied, the original production schedule needed to be repeatedly modified manually, leading to increased task waiting times.

[0183] In this embodiment, the system collects data on tasks to be scheduled, available processing resources, delivery nodes, and execution status every 30 minutes as a production time period, constructing a basic scheduling task set. During a given production day, a total of 126 tasks to be scheduled, 18 available processing resources, and 42 delivery nodes were collected, of which 31 tasks were in execution and 95 tasks were not yet executed. The system first establishes a time root layer based on the delivery time, a resource branch layer based on the resource number, and a task leaf layer based on the task number, forming a scheduling potential field tree. Then, through tree layer potential energy evolution processing, the delivery traction potential is transferred to the resource branch layer and the task leaf layer, and the occupancy squeeze potential of the task leaf layer is transferred back to the resource branch layer, resulting in an evolved scheduling potential field tree. After the evolved scheduling potential field tree is input into the improved RetNet model, the system outputs a scheduling evolution state chain and further generates candidate scheduling paths. The system locks the executing chain nodes in the rolling optimization window, only adjusts the non-executed chain nodes, maintains rigid execution for risk-priority scheduling segments, and uses spatiotemporal adjustment margin to compare the main receiving resources and alternative receiving resources for regular delayed scheduling segments, ultimately forming an intelligent scheduling scheme for timber processing and production.

[0184] To verify the practical effectiveness of this invention, three comparison schemes were set up. Comparison scheme one is a manual experience-based scheduling scheme, where team leaders manually sort tasks based on order delivery dates and equipment availability. Comparison scheme two is a rule-priority scheduling scheme, where tasks are sorted according to fixed rules based on delivery time and resource availability. Comparison scheme three is a conventional deep learning scheduling scheme, where tasks, resources, and delivery data are input as sequences into a regular RetNet model without building a scheduling potential tree or setting tree layer evolution retention units. The test data comes from actual scheduling records of 10 consecutive production days, including 1260 tasks to be scheduled, 418 delivery nodes, 96 resource usage changes, and 74 new task insertions. The comparison results are shown in Table 1.

[0185] Table 1. Comparison of Production Scheduling Effects of Different Scheduling Schemes

[0186] plan Number of test tasks / Average task wait time / h Average resource utilization rate / % On-time delivery rate / % Comparison Option 1 1260 5.8 73.4 86.7 Comparison Option 2 1260 4.6 78.9 90.8 Comparison Option 3 1260 3.7 82.6 93.1 Method of the present invention 1260 2.4 89.7 97.3

[0187] The data in Table 1 above shows that, under the same test task count of 1260, the overall scheduling effect of the method of the present invention is significantly better than the three comparative schemes. Comparative scheme one, which uses manual experience for scheduling, has an average task waiting time of 5.8 hours, an average resource utilization rate of 73.4%, and a delivery on-time rate of 86.7%. This indicates that this scheme relies heavily on the experience of on-site personnel and struggles to adjust the task acceptance order in a timely manner when faced with multiple parallel tasks and changes in resource usage. Comparative scheme two, which uses fixed-rule sorting, reduces the average task waiting time to 4.6 hours, increases the average resource utilization rate to 78.9%, and improves the delivery on-time rate to 90.8%. However, it still mainly relies on delivery time and resource idle status, lacking a dynamic expression of the evolutionary relationship between tasks, resources, and delivery. Compared to Scheme 3, which introduces a conventional deep learning model, the average task waiting time is further reduced to 3.7 hours, the average resource utilization rate reaches 82.6%, and the on-time delivery rate reaches 93.1%, indicating that the model prediction can improve the scheduling effect. However, since the scheduling potential field tree is not constructed and the tree layer potential energy evolution processing is not performed, the characterization of resource squeezing and task connection changes is still insufficient. The method of this invention reduces the average task waiting time to 2.4 hours, increases the average resource utilization rate to 89.7%, and achieves a delivery on-time rate of 97.3%, indicating that it can more effectively reduce task waiting, balance resource acceptance, and ensure delivery stability.

[0188] This embodiment verifies the practical application value of the present invention in the wood processing and production scenario. Faced with the production pressure of multiple tasks running in parallel, changing resource usage, and concentrated delivery nodes, the intelligent scheduling scheme for wood processing and production formed by the present invention can integrate task arrangement, resource acceptance, and delivery risks into the same scheduling chain for processing, which significantly shortens task waiting time, makes resource utilization more balanced, and further improves on-time delivery rate.

[0189] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for intelligent scheduling and management of timber processing production, characterized in that, Includes the following steps: Step 1: Obtain the tasks to be scheduled, available processing resources, delivery nodes, and execution status in the timber processing production, and construct a basic set of scheduling tasks; Step 2: Establish a scheduling potential tree based on the basic scheduling task set. The scheduling potential tree includes a time root layer, a resource branch layer, and a task leaf layer. Step 3: Perform tree-level potential evolution processing on the scheduling potential field tree, transfer the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfer the occupancy squeezing potential of the task leaf layer back to the resource branch layer, and adjust the attachment position of the task leaf layer to generate the evolution scheduling potential field tree. Step 4: Input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive retention unit between the multi-scale retention unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate a scheduling recursive state chain. Step 5: Based on the aforementioned scheduling evolution state chain, generate candidate scheduling paths; Step 6: Based on the execution status, perform rolling optimization on the candidate scheduling paths to obtain the current execution path and generate an intelligent scheduling scheme for timber processing production; Step 7: Perform production management according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and completion of tasks, write back to the basic scheduling task set, and update the intelligent scheduling scheme for timber processing and production. Step two specifically involves: Delivery nodes are obtained from the basic scheduling task set, and a delivery time sequence index is established according to the delivery time. Delivery nodes with the same delivery time are grouped into the same time root node to generate a time root layer. Available processing resources are obtained from the basic scheduling task set, and a resource acceptance index is established according to the resource number. The available processing resources are converted into resource branch nodes to generate a resource branch layer. Obtain the tasks to be scheduled from the basic scheduling task set, establish a task attachment index according to the task number, convert the tasks to be scheduled into task leaf nodes, and generate a task leaf layer. Based on the resource acceptance index and the task attachment index, establish a resource attachment chain between the resource branch node and the corresponding task leaf node; Based on the task number corresponding to the delivery node, determine the time root node corresponding to the task leaf node, and establish a delivery acceptance chain between the corresponding resource branch node and the time root node through the established resource attachment chain. The delivery acceptance chain and resource attachment chain are hierarchically merged according to the time root layer, resource branch layer and task leaf layer to obtain the scheduling potential tree.

2. The intelligent scheduling and management method for timber processing production according to claim 1, characterized in that, Step one specifically involves: Obtain the tasks to be scheduled in the timber processing production and assign a task number to each task to be scheduled. Identify available processing resources in timber processing production and assign a resource number to each available processing resource. Obtain delivery nodes in timber processing and production, wherein the delivery node includes the delivery time and the corresponding task number; The execution status in wood processing production is obtained, including non-execution status, execution status, and execution completion status. The tasks to be scheduled are matched with the delivery nodes according to the task number, and the tasks to be scheduled, available processing resources and execution status are grouped according to the same production time to build a basic scheduling task set.

3. The intelligent scheduling and management method for timber processing production according to claim 1, characterized in that, Step three specifically involves: Arrange the time root nodes in the time root layer according to the delivery time sequence index, and generate the delivery traction corresponding to each time root node based on the arrangement position of each time root node in the delivery time sequence index. The delivery traction corresponding to each time root node is passed along the delivery acceptance chain to the corresponding resource branch node, and then passed along the resource attachment chain to the corresponding task leaf node. Based on the execution status corresponding to the task leaf node, generate the occupancy and squeezing potential of the task leaf layer; The occupancy squeeze potential of the task leaf layer is transmitted back to the corresponding resource branch node along the resource link chain, and the number of task leaf nodes under the same resource branch node and the received occupancy squeeze potential are summarized to form the transmission squeeze result corresponding to each resource branch node. At the same time, the resource branch nodes corresponding to the root node are arranged from smallest to largest according to the back-passing squeezing result, forming the resource branch node arrangement result. At the same root node, the task leaf nodes are arranged according to the delivery traction received by the task leaf nodes, forming the task leaf node arrangement result. Based on the arrangement of task leaf nodes, select task leaf nodes in sequence, and in the resource branch nodes corresponding to the root node at the same time, find the resource branch node corresponding to the selected task leaf node through the resource link chain. When multiple resource branch nodes are found, the resource branch node that appears first in the list is selected as the target resource branch node. When the target resource branch node is not the same as the resource branch node currently attached to the selected task leaf node, the selected task leaf node will be removed from the currently attached resource branch node and attached to the target resource branch node. When the target resource branch node is the same as the resource branch node currently attached to the selected task leaf node, the current attachment position of the selected task leaf node is retained. Adjust the attachment positions of each task leaf node under the root node at the same time, and generate a recursive scheduling potential tree.

4. The intelligent scheduling and management method for timber processing production according to claim 1, characterized in that, The improved RetNet model specifically includes a potential field node embedding unit, a multi-scale preservation unit, a tree layer evolution preservation unit, a feedforward mapping unit, and a state chain output unit; The potential field node embedding unit encodes the time root layer, resource branch layer and task leaf layer as time root layer node embedding, resource branch layer node embedding and task leaf layer node embedding, respectively. The multi-scale retention unit performs retention calculations on the embedding of time root layer nodes, resource branch layer nodes, and task leaf layer nodes based on the delivery acceptance chain and resource attachment chain, to obtain the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer. The tree-layer evolution preservation unit is set between the multi-scale preservation unit and the feedforward mapping unit. It performs tree-layer evolution state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to obtain the tree-layer evolution hidden state. The feedforward mapping unit sequentially performs linear mapping, GELU nonlinear activation, and LayerNorm normalization on the tree-layer recursive hidden state to obtain the scheduling mapping state features. The state chain output unit inputs the scheduling mapping state characteristics into the task output branch, resource output branch, and risk output branch, respectively. The task output branch generates task priority probabilities through a linear layer and a Softmax function, and arranges the tasks in descending order of priority probabilities to obtain the task priority order. The resource output branches generate resource acceptance probabilities through a linear layer and a Softmax function, and arrange the resource acceptance order in descending order of resource acceptance probabilities. The risk output branch generates delivery risk values ​​through a linear layer and a sigmoid function, and arranges the delivery risk values ​​from largest to smallest to obtain the delivery risk order; Using task priority as the main chain, the resource acceptance order and delivery risk order of the corresponding tasks to be scheduled are written into the same chain node to generate a scheduling evolution state chain.

5. The intelligent scheduling and management method for timber processing production according to claim 4, characterized in that, The tree-level evolution preservation unit specifically includes a downlink evolution branch, an uplink backhaul branch, and a hierarchical state merging branch; The downlink recursive branch receives the intermediate hidden states of the time root layer, resource branch layer, and task leaf layer. According to the delivery acceptance chain, the intermediate hidden states of the time root layer are spliced ​​with the intermediate hidden states of the corresponding resource branch layer. After linear mapping and GELU nonlinear activation processing, the root-branch downlink state is obtained. The downlink evolution branch splices the root-branch downlink state with the intermediate hidden state of the corresponding task leaf layer according to the resource attachment chain, and performs linear mapping to obtain the branch-leaf downlink state. The uplink backhaul branch generates task occupancy backhaul features based on the occupancy squeeze potential of the task leaf layer, and splices the task occupancy backhaul features with the intermediate hidden state of the corresponding resource branch layer according to the resource attachment chain to obtain the leaf-branch backhaul input features. The uplink backhaul branch concatenates the leaf-branch backhaul input features with the backhaul squeezing results and the resource branch node arrangement results, and then performs linear mapping and GELU nonlinear activation processing to obtain the leaf-branch backhaul state. The hierarchical state converging branch splices the root-branch downlink state, branch-leaf downlink state, and leaf-branch backlink state in the hierarchical order of time root layer, resource branch layer, and task leaf layer, and performs linear mapping, GELU nonlinear activation, and LayerNorm normalization in sequence to generate the tree-level recursive hidden state.

6. The intelligent scheduling and management method for timber processing production according to claim 1, characterized in that, Step five specifically involves: The chain nodes in the scheduling evolution state chain are traversed sequentially according to the task priority, and the corresponding task to be scheduled, resource acceptance order, and delivery risk order are determined in each chain node. In each chain node, the available processing resource that ranks first in the resource acceptance order is taken as the primary processing resource, and the available processing resource that ranks second in the resource acceptance order is taken as the alternative processing resource. The delivery risk value corresponding to each chain node is determined according to the delivery risk order. When the delivery risk value is greater than or equal to the set delivery risk threshold, the corresponding task to be scheduled is marked as a risk-priority task. When the delivery risk value is less than the set delivery risk threshold, the corresponding task to be scheduled is marked as a regular deferred task. First, connect the chain nodes corresponding to the risk-priority tasks according to the task priority order, and bind the risk-priority tasks, the corresponding main undertaking resources and the corresponding delivery nodes to generate risk-priority scheduling segments. Continue connecting the chain nodes corresponding to the regular postponed tasks according to the task priority order, and bind the regular postponed tasks, the corresponding main receiving resources, and the corresponding delivery nodes. At the same time, write the corresponding alternative receiving resources as spatiotemporal adjustment margins into the corresponding nodes to generate regular postponed scheduling segments. The risk-priority scheduling segment is arranged before the regular deferred scheduling segment, and the risk-priority scheduling segment and the regular deferred scheduling segment are connected according to the order of the chain nodes in the scheduling evolution state chain to generate candidate scheduling paths.

7. The intelligent scheduling and management method for timber processing production according to claim 1, characterized in that, Step six specifically involves: Determine the executing and non-executing chain nodes in the candidate scheduling path according to the execution status; The executing chain node is rigidly retained in its original position in the candidate scheduling path to lock the current processing state, and the unexecuted chain node is used as the rolling selection node; The first rolling selection node after the chain node in the execution is taken as the rolling starting point, and the rolling selection window is formed according to the order of the chain nodes in the candidate scheduling path; Within the rolling selection window, the chain node positions in the risk-priority scheduling segment are reserved first, and the main and alternative resources corresponding to the regular delayed task are compared and accepted based on the spatiotemporal adjustment margin written by the corresponding node in the regular delayed scheduling segment. When the primary resource corresponding to a regular deferred task is in an unoccupied state, the primary resource will be used as the current resource. When the main receiving resource corresponding to the regular deferred task is in the execution state and the alternative receiving resource is in the unoccupied state, the alternative receiving resource will be used as the current receiving resource. When both the primary and alternative resources corresponding to a regular deferred task are in the execution state, the primary resource is kept as the current resource, and the task deferred processing is executed. According to the connection order of risk-priority scheduling segment first and regular deferred scheduling segment last within the rolling selection window, the tasks to be scheduled, the currently received resources and the delivery nodes corresponding to each chain node are bound to obtain the current execution path; Write the current execution path into the scheduling execution list, and generate an intelligent scheduling scheme for timber processing and production based on the scheduling execution list.

8. A timber processing production intelligent scheduling and management system, executing the timber processing production intelligent scheduling and management method according to any one of claims 1 to 7, characterized in that, Includes the following modules: The task aggregation module is used to obtain the tasks to be scheduled, available processing resources, delivery nodes and execution status in the timber processing and production, and to build a basic scheduling task set; The potential field tree construction module is used to build a scheduling potential field tree based on the basic scheduling task set. The scheduling potential field tree includes a time root layer, a resource branch layer, and a task leaf layer. The tree layer evolution module is used to perform tree layer potential energy evolution processing on the scheduling potential field tree, transfer the delivery traction potential of the time root layer to the resource branch layer and the task leaf layer, transfer the occupancy squeeze potential of the task leaf layer back to the resource branch layer, and adjust the attachment position of the task leaf layer to generate the evolution scheduling potential field tree. The state chain solution module is used to input the recursive scheduling potential tree into the improved RetNet model. The improved RetNet model adds a tree-layer recursive preservation unit between the multi-scale preservation unit and the feedforward mapping unit to perform tree-layer recursive state transfer on the intermediate hidden states of the time root layer, resource branch layer and task leaf layer to generate a scheduling recursive state chain. The path orchestration module is used to generate candidate scheduling paths based on the scheduling evolution state chain; The rolling execution module is used to perform rolling optimization of candidate scheduling paths based on the execution status, obtain the current execution path, and generate an intelligent scheduling scheme for timber processing and production. The status write-back module is used to manage production according to the intelligent scheduling scheme for timber processing and production, obtain information on new tasks, resource changes, and execution completion, and write back the basic scheduling task set to update the intelligent scheduling scheme for timber processing and production.

9. A smart scheduling and management device for timber processing production, characterized in that, It includes a processor and a memory, the memory being used to store computer programs, and the processor calling the computer programs stored in the memory to execute the intelligent scheduling and management method for wood processing production as described in any one of claims 1 to 7.

Images