A multi-condition adaptive task scheduling method and system for cutting and grinding processes

Abstract

This application relates to the field of embodied intelligence technology and discloses a multi-condition adaptive task scheduling method and system for cutting and grinding processes. The method includes: inputting collected multi-agent state information, workshop dust environment information, and standardized cutting and grinding task data into an initialized customized scheduling model to generate multiple feasible scheduling schemes, selecting the optimal scheme, and retaining multiple alternative schemes; each agent executes the cutting and grinding task according to the execution instructions, and the feedback information is processed hierarchically; if a trigger is detected, the adjusted scheduling scheme is transmitted to each agent for execution; a standardized training dataset is constructed, and a reinforcement learning training framework with the core requirements of the cutting and grinding process scheduling as the reward objective is built; the standardized training dataset is input into the framework for iterative training. This application improves the collaborative operation efficiency and stability of multiple embodied agents in the cutting and grinding process.

Description

Technical Field

[0001] This application relates to the field of embodied intelligence technology, and in particular to a multi-condition adaptive task scheduling method and system for cutting and grinding processes. Background Technology

[0002] Currently, embodied intelligence technology is increasingly widely used in the production of customized panel furniture. Multi-embodied intelligent agent collaborative scheduling has become a core supporting technology for promoting the intelligent and automated upgrading of the cutting and sanding processes in customized panel furniture. The cutting and sanding processes of customized panel furniture are characterized by high customization, diverse board specifications, high precision requirements, and significant interference from operating conditions (wood chips and dust). Existing technologies mainly focus on designing fixed task flows and preset scheduling rules. The core idea is to pre-plan the cutting and sanding task sequence and assign the tasks of the two processes to designated intelligent agents to achieve multi-agent collaborative operation. With the surge in demand for customized panel furniture, higher requirements are placed on the flexibility, adaptability, and collaborative precision of multi-embodied intelligent agent scheduling in the cutting and sanding processes. However, existing technologies still have significant limitations in their adaptability to these two core processes, making it difficult to meet the actual needs of customized board cutting and high-precision sanding.

[0003] Existing static scheduling methods based on preset rules involve manually pre-setting the priorities of customized panel furniture cutting and sanding tasks, as well as the assignment rules for intelligent agents. Cutting tasks are assigned to designated intelligent agents, and sanding tasks to another set of designated intelligent agents. The agents simply execute tasks according to preset instructions and lack the ability to adjust scheduling strategies based on changes in on-site conditions such as panel specifications, cutting accuracy deviations, substandard sanding results, or agent malfunctions. For example, if agent A is pre-set to be responsible for customized panel cutting and agent B for component sanding, the scheduling strategy remains unchanged regardless of on-site issues such as panel size deviations, substandard sanding accuracy, or agent malfunctions. This can easily lead to decreased cutting and sanding efficiency and lower product yield.

[0004] Existing semi-dynamic scheduling methods based on simple feedback add a simple feedback mechanism to static scheduling. When a cutting or grinding agent malfunctions and stops, its unfinished cutting or grinding tasks are assigned to other idle agents. However, these methods do not consider key factors such as the precision requirements of custom panel furniture cutting, the roughness requirements of grinding, the real-time load of agents, and interference from sawdust in the workshop. They can only handle single types of emergencies, have limited adaptability, and cannot adapt to the complex working conditions of custom panel furniture cutting and grinding. A related technology, CN121200016A, discloses a multi-machine collaborative interaction method and system for industrial robots based on artificial intelligence. This method relies on task transfer based on a preset threshold and does not achieve dynamic adaptive scheduling for the custom panel furniture cutting and grinding process.

[0005] The existing technical solutions have the following problems: Insufficient scheduling flexibility: Existing technologies are mostly based on preset rules or simple feedback, which cannot dynamically adjust the task allocation strategy according to real-time changes in customized panel furniture cutting and polishing tasks (such as changes in customized board specifications, adjustments in cutting precision requirements, and changes in polishing roughness requirements), changes in the status of intelligent agents (such as excessive load, minor faults, and insufficient power), and changes in workshop dust interference. This can easily lead to delays in the execution of cutting and polishing tasks, production line stagnation, and waste of resources.

[0006] Low resource utilization: The real-time load and skill matching degree (such as board cutting accuracy and grinding roughness control ability) of the intelligent agents for cutting and polishing customized panel furniture are not dynamically evaluated. This can easily lead to some intelligent agents being overloaded while others are idle. Especially in the scenario of batch processing of customized orders, the problem of uneven resource allocation is prominent, which restricts the capacity improvement of the cutting and polishing process.

[0007] Poor robustness: When faced with unexpected situations in the cutting and polishing of customized panel furniture (such as temporary failure of the intelligent agent, batch deviation of customized boards, and increased dust interference), the existing technology can only perform simple task transfer. It cannot quickly re-plan the cutting and polishing task sequence or adjust resource allocation, which can easily lead to production interruption and make it difficult to ensure the continuity of the cutting and polishing process and product precision.

[0008] The cutting and grinding process has poor adaptability: Most of the existing scheduling methods are general designs and have not been specifically optimized for the special characteristics of cutting and grinding custom panel furniture (such as diverse specifications of customized boards, high precision requirements, and large dust interference). When applying them, the scheduling rules need to be manually readjusted, which is costly and cannot meet the stringent requirements of cutting and grinding for scheduling accuracy and response speed. Summary of the Invention

[0009] To address the technical problems of insufficient scheduling flexibility, low resource utilization, poor robustness, and weak adaptability to the cutting and polishing process of customized panel furniture in existing technologies, this application provides a multi-condition adaptive task scheduling method and system for the cutting and polishing process. This method enables dynamic perception of the real-time status of the cutting and polishing task, the load and skill matching degree of multiple intelligent agents, and the interference of workshop dust environment. It automatically adjusts the task allocation strategy and execution sequence, balances the resource load of intelligent agents, improves the execution efficiency and continuity of the cutting and polishing task, adapts to the entire process scenario of cutting and polishing customized panel furniture, and helps the intelligent upgrade of this process.

[0010] This application discloses a multi-condition adaptive task scheduling method for cutting and grinding processes, which includes: Step 1: Collect cutting and polishing task information, multi-agent status information, and workshop dust environment information in real time at a preset frequency in the customized panel furniture cutting and polishing scenario. The preset frequency is the information collection frequency dynamically set according to the production rhythm of the cutting and polishing process. Quantify the workshop dust environment information to obtain the environmental interference coefficient, quantify the cutting and polishing task information to obtain the task difficulty coefficient, and analyze the multi-agent status information to obtain the skill matching degree. Step 2: Perform task splitting, priority sorting, and data standardization operations on the collected cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on task difficulty coefficient, skill matching degree, and environmental interference coefficient. Step 3: Input the collected multi-agent state information, workshop dust environment information, and standardized cutting and polishing task data into the initialized customized scheduling model. The customized scheduling model is an intelligent decision-making model built on a machine learning architecture and adapted to the scheduling needs of customized panel furniture cutting and polishing processes. Multiple feasible scheduling schemes are generated using the large model combined with a mind tree approach. The optimal scheduling scheme is selected after comprehensive scoring of the feasible scheduling schemes through a preset evaluation system, while multiple alternative scheduling schemes are retained. The preset evaluation system is a multi-dimensional scheme evaluation system set according to the production needs of the cutting and polishing process. The optimal scheduling scheme is broken down into execution instructions and transmitted to each agent to start task execution. The alternative scheduling schemes are stored in the scheduling system. Step 4: Each agent executes the cutting and grinding task according to the execution instructions. The agent's execution status, changes in cutting and grinding conditions, and changes in the workshop dust environment are collected in real time at a preset frequency. The feedback information is then processed in layers. Step 5: Determine whether the scheduling strategy adjustment is triggered based on the result of the hierarchical processing. If it is determined to be triggered, then based on the scene state information and feedback information, call the alternative scheduling scheme or re-optimize and generate the scheduling scheme for the sudden working conditions of the cutting and grinding process, and transmit the adjusted scheduling scheme to each intelligent agent for execution. Step 6: Collect the entire cutting and grinding process data generated in Steps 1 to 5. After preprocessing the data, construct a standardized training dataset. Build a reinforcement learning training framework with the core requirements of the cutting and grinding process scheduling as the reward objective. Input the standardized training dataset into the framework for iterative training. Correct the scheduling decision logic and adaptive adjustment rules of the large model through the reinforcement learning training framework. After the core indicators of the model tend to stabilize, update the optimized model parameters to the scheduling system. At the same time, establish an incremental training mechanism. Collect new data for incremental training after each batch of cutting and grinding tasks is completed.

[0011] Furthermore, the cutting and polishing task information includes task type, quantity, priority, completion time limit, required intelligent agent skills, and task association; the multi-agent status information includes the real-time load, operating status, current position and workable range, and work accuracy of each cutting and polishing intelligent agent; and the workshop dust environment information includes workshop dust concentration, sawdust accumulation, and obstacle distribution.

[0012] Furthermore, the real-time load includes the number of currently executing tasks, hardware utilization rate, remaining power, and operation duration; the operating status includes normal, minor fault, and shutdown fault; and the operation accuracy includes cutting size deviation and grinding roughness.

[0013] Furthermore, by using various sensors mounted on the intelligent agent as the core data acquisition carrier and combining edge computing technology, a full-dimensional state perception system is constructed to realize the real-time acquisition, quantification, and analysis of cutting and polishing scene information.

[0014] Furthermore, the task is broken down into multiple independently executable sub-tasks, specifying the completion time, skill requirements, difficulty level, and interrelationships of each sub-task. The priority ranking adopts the hierarchical analysis method, classifying tasks according to customized order requirements, task urgency, and cutting and polishing process cycle requirements. The normalization process unifies the task difficulty level, skill matching degree, and environmental interference coefficient to the same numerical range.

[0015] Furthermore, during the initialization of the customized scheduling model, a dedicated knowledge base is loaded, including collaborative rules for cutting and grinding processes, load balancing standards, dust environment adaptation parameters, and customized task processing rules. The mind tree is constructed with the core objectives of efficient execution of cutting and grinding, load balancing of intelligent agents, continuous production, and achievement of work accuracy standards, and multiple feasible scheduling schemes are generated. The scoring dimensions of the preset evaluation system include scheme feasibility, execution efficiency, load balancing, and environmental adaptability, with corresponding weight coefficients configured for each dimension.

[0016] Furthermore, the execution status of the intelligent agent includes task completion progress, execution accuracy, operation error, power consumption rate, and hardware occupancy rate; the changes in the cutting and sanding working conditions include the addition / cancellation of customized orders, changes in board specifications, and adjustments to the cutting and sanding process cycle time; the real-time changes in the workshop dust environment include dust concentration, sawdust accumulation, and obstacle position changes; the results of the layered processing include normal execution data, slightly abnormal data, and severely abnormal data; the normal execution data includes task progress meeting standards, intelligent agent status being normal, and dust concentration meeting requirements; the slightly abnormal data includes minor deviations in cutting dimensions and substandard sanding roughness; the severely abnormal data includes intelligent agent malfunction and shutdown, batch damage to boards, and excessive dust concentration.

[0017] Furthermore, the sudden working conditions include agent failure, production task change, dust environment abnormality, and agent load imbalance. The scheduling strategy adjustment process does not require manual intervention. Among them, dust environment abnormality is when the environmental interference coefficient reaches a preset threshold, and agent load imbalance is when the variance of the load of multiple agents reaches a preset threshold. The preset threshold is a judgment threshold set according to the process requirements of the customized panel furniture cutting and polishing process and the production environment.

[0018] Furthermore, the preprocessing of the entire process data includes data cleaning, noise reduction, and normalization. After removing invalid and redundant data, core feature parameters are extracted. These core feature parameters include task completion time, load balancing variance, anomaly handling response time, accuracy achievement rate, and product qualification rate. The reinforcement learning training framework is configured with an agent, environment, actions, and reward function. The agent is a scheduling decision system, the environment is the dynamic working conditions of cutting and polishing, and the actions are scheduling scheme generation, task allocation, and adaptive adjustment operations. The reward function is a multi-dimensional function that integrates task execution efficiency, agent load balancing, system robustness, anomaly response time, and product qualification rate. Each dimension is configured with corresponding weight coefficients, and the sum of the weight coefficients is 1. The iterative training is carried out for more than or equal to a preset number of rounds. The preset number of rounds is the number of iterations required to meet the model training convergence requirements. The criterion for the model's core indicators to stabilize is that the indicator fluctuations are within a preset range. The preset range is the range of indicator fluctuations that meets the scheduling accuracy requirements of the cutting and polishing process. The results of the incremental training are synchronously updated to the customized scheduling large model, the thinking tree branch logic, and the adaptive adjustment rules.

[0019] This application also discloses a multi-condition adaptive task scheduling system for cutting and grinding processes, used to implement the aforementioned multi-condition adaptive task scheduling method for cutting and grinding processes, comprising: The state perception module is used to collect cutting and grinding task information, multi-agent state information, and workshop dust environment information in real time at a preset frequency. It quantifies the workshop dust environment information to obtain the environmental interference coefficient, quantifies the cutting and grinding task information to obtain the task difficulty coefficient, and analyzes the multi-agent state information to obtain the skill matching degree. The task parsing module is used to perform task splitting, priority sorting, and data standardization operations on the cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on the task difficulty coefficient, skill matching degree, and environmental interference coefficient. The scheduling decision module has a built-in customized scheduling model. It receives the data collected by the state perception module and the preprocessed data by the task parsing module. It loads the knowledge base dedicated to the cutting and grinding process and generates a variety of feasible scheduling schemes by combining the large model with the mind tree. After comprehensive scoring by the preset evaluation system, the optimal scheduling scheme is selected and the execution instructions are output to each intelligent agent. At the same time, multiple alternative scheduling schemes are stored. The execution feedback module is used to collect three types of feedback information in real time during the task execution process: the execution status of the intelligent agent, the changes in the cutting and grinding conditions, and the real-time changes in the dust environment in the workshop. The feedback information is processed in layers, and the results of the layered processing and the execution log are transmitted to the scheduling decision module, the adaptive adjustment module, and the reinforcement learning training module. The adaptive adjustment module is used to determine whether to trigger the scheduling strategy adjustment based on the hierarchical processing results of the execution feedback module and the real-time information collected by the state perception module. In case of sudden working conditions, it calls the alternative scheduling scheme or re-optimizes and generates the scheduling scheme, and outputs the adjusted scheduling scheme to each intelligent agent. The reinforcement learning training module is used to collect data from the entire cutting and polishing process. After preprocessing the data, a standardized training dataset is constructed. A reinforcement learning training framework is built to carry out iterative and incremental training, correct the scheduling decision logic and adaptive adjustment rules of the customized scheduling model, and update the optimized model parameters to the scheduling decision module in a synchronous manner, forming a closed-loop iterative optimization of the scheduling model.

[0020] Due to the adoption of the above technical solution, this application has the following advantages: 1. To address the shortcomings of existing technologies in terms of scheduling flexibility and inability to adapt to the cutting and polishing processes of customized panel furniture, this application utilizes multi-dimensional real-time perception, large model + mind tree scheduling, and full-condition adaptive adjustment to adapt to dynamic scenarios such as changes in customized panel specifications and dust interference in real time. This requires no manual intervention, improves process adaptability by more than 40%, effectively solves the problem of rigid scheduling in existing technologies, and ensures the efficient progress of customized cutting and polishing tasks.

[0021] 2. To address the shortcomings of low resource utilization in existing technologies, this application ensures that the load balancing variance of the intelligent agent is ≤0.2 through refined task analysis and load balancing adjustment mechanism. In the specific implementation, the load balancing variance is only 0.16. Compared with the existing technology, the resource utilization rate is improved by 30%-35%, which solves the problem of uneven load and resource waste of the intelligent agent in the cutting and grinding process, and greatly improves the cutting and grinding capacity.

[0022] 3. To address the shortcomings of existing technologies, such as poor robustness and susceptibility to production interruptions, this application employs layered feedback and full-condition adaptive adjustment, achieving an anomaly response time of ≤0.8s and reducing production interruption time by more than 80%. Combined with alternative scheduling schemes, it effectively ensures the continuity of cutting and grinding processes while improving product accuracy and yield. Specifically, the cutting accuracy compliance rate is 99%, the grinding roughness compliance rate is 98.8%, and the product yield is 7% higher than that of existing technologies.

[0023] 4. To address the shortcomings of existing technologies in terms of weak process adaptability, this application focuses entirely on the cutting and polishing process of customized panel furniture. All technical features and scheduling strategies are specifically optimized for the customization, high precision, and high dust interference characteristics of this process, and can be directly applied without manual adjustment, resulting in low adaptation costs. At the same time, through reinforcement learning closed-loop iterative optimization, it can dynamically adapt to changes in customized orders and fluctuations in working conditions, significantly improving the intelligence level and production efficiency of the cutting and polishing process. In specific implementations, the scheduling efficiency is improved by 30% compared to existing technologies, and the response speed for handling anomalies is improved by 20%.

[0024] 5. The closed-loop mechanism of "data acquisition-analysis scheduling-execution feedback-adaptive adjustment-model optimization" constructed in this application relies on the data of the entire cutting and polishing process to continuously optimize the scheduling model. As the number of task executions increases, the scheduling accuracy and execution efficiency continue to improve. Long-term application can further reduce production costs and reduce manual intervention, helping to achieve full-process intelligent and automated upgrade of the cutting and polishing process of customized panel furniture. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings.

[0026] Figure 1 This is a flowchart illustrating a multi-condition adaptive task scheduling method for a cutting and grinding process according to an embodiment of this application. Figure 2 This is a flowchart of intelligent agent task processing and adaptive optimization for a cutting and grinding process according to an embodiment of this application. Figure 3 This is a schematic diagram of the intelligent agent task scheduling and execution system architecture based on a large model and mind tree, according to an embodiment of this application. Figure 4 This is a flowchart illustrating the task execution adaptive adjustment and model optimization method based on anomaly detection and reinforcement learning for the cutting and grinding scenario in this application embodiment. Detailed Implementation

[0027] The present application will be further described in conjunction with the accompanying drawings and embodiments. The described embodiments are only some, not all, of the embodiments of the present application. All other embodiments obtained by those skilled in the art should fall within the protection scope of the embodiments of the present application.

[0028] This application is specifically designed for the two core processes of panel cutting and polishing in the production of customized panel furniture, enabling dynamic collaborative scheduling and adaptive adjustment of multiple intelligent agents in these two processes. Focusing primarily on these two core processes, it is applicable to the entire cutting and polishing process of customized panel furniture, achieving dynamic task allocation and adaptive adjustment of multiple intelligent agents. This solves the technical challenges of insufficient flexibility, low resource utilization, and poor adaptability in the collaborative scheduling of multiple intelligent agents in the cutting and polishing process of customized panel furniture, thus facilitating the intelligent upgrade of the cutting and polishing process in customized panel furniture production. Cutting and polishing are core processes in the production of customized panel furniture, directly determining product precision and quality; their level of intelligence directly affects overall production efficiency. With the transformation of customized panel furniture towards personalization and large-scale production, the demand for collaborative scheduling of multiple intelligent agents in the cutting and polishing process is increasingly urgent. Existing technologies cannot meet actual production needs in terms of flexibility, resource utilization, robustness, and process adaptability. Therefore, there is an urgent need for a method that can perceive the cutting and polishing tasks of customized panel furniture, the status of intelligent agents and the dust environment in the workshop in real time, and realize dynamic adaptive scheduling to solve the shortcomings of existing technologies, improve the collaborative operation efficiency and stability of multiple intelligent agents in the cutting and polishing process, promote the intelligent upgrading of the cutting and polishing process of customized panel furniture, reduce production costs, and ensure production continuity.

[0029] Based on this, see Figure 1 This application provides an embodiment of a multi-condition adaptive task scheduling method for cutting and grinding processes, which includes: Step 1: Collect cutting and polishing task information, multi-agent status information, and workshop dust environment information in real time at a preset frequency in the customized panel furniture cutting and polishing scenario. The preset frequency is the information collection frequency dynamically set according to the production rhythm of the cutting and polishing process. Quantify the workshop dust environment information to obtain the environmental interference coefficient, quantify the cutting and polishing task information to obtain the task difficulty coefficient, and analyze the multi-agent status information to obtain the skill matching degree. Step 2: Perform task splitting, priority sorting, and data standardization operations on the collected cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on task difficulty coefficient, skill matching degree, and environmental interference coefficient. Step 3: Input the collected multi-agent state information, workshop dust environment information, and standardized cutting and polishing task data into the initialized customized scheduling model. The customized scheduling model is an intelligent decision-making model built on a machine learning architecture and adapted to the scheduling needs of customized panel furniture cutting and polishing processes. Multiple feasible scheduling schemes are generated using the large model combined with a mind tree approach. The optimal scheduling scheme is selected after comprehensive scoring of the feasible scheduling schemes through a preset evaluation system, while multiple alternative scheduling schemes are retained. The preset evaluation system is a multi-dimensional scheme evaluation system set according to the production needs of the cutting and polishing process. The optimal scheduling scheme is broken down into execution instructions and transmitted to each agent to start task execution. The alternative scheduling schemes are stored in the scheduling system. Step 4: Each agent executes the cutting and grinding task according to the execution instructions. The agent's execution status, changes in cutting and grinding conditions, and changes in the workshop dust environment are collected in real time at a preset frequency. The feedback information is then processed in layers. Step 5: Determine whether the scheduling strategy adjustment is triggered based on the result of the layered processing. If it is determined to be triggered, then based on the scene state information and feedback information, call the alternative scheduling scheme or re-optimize and generate the scheduling scheme for the sudden working conditions of the cutting and grinding process, and transmit the adjusted scheduling scheme to each intelligent agent for execution to ensure the continuous progress of the cutting and grinding process. Step 6: Collect the entire cutting and grinding process data generated in Steps 1 to 5. After preprocessing the data, construct a standardized training dataset. Build a reinforcement learning training framework with the core requirements of the cutting and grinding process scheduling as the reward objective. Input the standardized training dataset into the framework for iterative training. Correct the scheduling decision logic and adaptive adjustment rules of the large model through the reinforcement learning training framework. After the core indicators of the model tend to stabilize, update the optimized model parameters to the scheduling system. At the same time, establish an incremental training mechanism. After each batch of cutting and grinding tasks is completed, collect new data for incremental training, forming a closed-loop iterative optimization logic of the scheduling model: model optimization - scheduling execution - data collection - model re-optimization.

[0030] Optionally, the cutting and polishing task information includes task type, quantity, priority, completion time limit, required agent skills, and task association; the multi-agent status information includes the real-time load, operating status, current location and workable range, and work accuracy of each cutting and polishing agent; and the workshop dust environment information includes workshop dust concentration, sawdust accumulation, and obstacle distribution.

[0031] Optionally, the real-time load includes the number of currently executing tasks, hardware utilization, remaining power, and operation duration; the operating status includes normal, minor fault, and shutdown fault; and the operation accuracy includes cutting size deviation and grinding roughness.

[0032] Optionally, various sensors mounted on the intelligent agent can be used as the core data acquisition carriers, and edge computing technology can be combined to build a full-dimensional state perception system to realize the real-time acquisition, quantification and analysis of cutting and polishing scene information.

[0033] Optionally, the task is broken down into multiple independently executable sub-tasks, specifying the completion time, skill requirements, difficulty coefficient, and connection relationship between each sub-task; the priority ranking adopts the hierarchical analysis method, combining the customized order requirements, task urgency, and cutting and grinding process cycle requirements to classify tasks into levels; the normalization process unifies the task difficulty coefficient, skill matching degree, and environmental interference coefficient to the same numerical range.

[0034] Optionally, during the initialization of the customized scheduling model, a dedicated knowledge base is loaded, including collaborative rules for cutting and grinding processes, load balancing standards, dust environment adaptation parameters, and customized task processing rules. The mind tree is constructed with the core objectives of efficient execution of cutting and grinding, load balancing of intelligent agents, continuous production, and achievement of work accuracy standards, and multiple feasible scheduling schemes are generated. The scoring dimensions of the preset evaluation system include scheme feasibility, execution efficiency, load balancing, and environmental adaptability, with corresponding weight coefficients configured for each dimension.

[0035] Optionally, the execution status of the intelligent agent includes task completion progress, execution accuracy, operation error, power consumption rate, and hardware occupancy rate; the changes in cutting and sanding conditions include the addition / cancellation of customized orders, changes in board specifications, and adjustments to the cutting and sanding process cycle time; the real-time changes in the workshop dust environment include dust concentration, sawdust accumulation, and obstacle position changes; the results of the layered processing include normal execution data, slightly abnormal data, and severely abnormal data; the normal execution data includes task progress meeting standards, intelligent agent status being normal, and dust concentration meeting requirements; the slightly abnormal data includes minor deviations in cutting dimensions and substandard sanding roughness; and the severely abnormal data includes intelligent agent malfunction and shutdown, batch damage to boards, and excessive dust concentration.

[0036] Optionally, the sudden working conditions include agent failure, production task change, dust environment abnormality, and agent load imbalance. The scheduling strategy adjustment process does not require manual intervention. Among them, dust environment abnormality is when the environmental interference coefficient reaches a preset threshold, and agent load imbalance is when the variance of the load of multiple agents reaches a preset threshold. The preset threshold is a judgment threshold set according to the process requirements of the customized panel furniture cutting and polishing process and the production environment.

[0037] Optionally, the preprocessing of the entire process data includes data cleaning, noise reduction, normalization, and extraction of core feature parameters after removing invalid and redundant data. These core feature parameters include task completion time, load balancing variance, anomaly response time, accuracy achievement rate, and product qualification rate. The reinforcement learning training framework is configured with an agent, environment, actions, and a reward function. The agent is a scheduling decision system, the environment is the dynamic working conditions of cutting and polishing, and the actions are scheduling scheme generation, task allocation, and adaptive adjustment operations. The reward function is a multi-dimensional function that integrates task execution efficiency, agent load balancing, system robustness, anomaly response time, and product qualification rate. Each dimension is configured with corresponding weight coefficients, and the sum of these weight coefficients is 1. The iterative training is conducted for a number of iterations greater than or equal to a preset number of iterations, which is the number of iterations required to meet the model training convergence requirements. The criterion for stabilizing the model's core indicators is that the indicator fluctuations are within a preset range, which is the range of indicator fluctuations required to meet the scheduling accuracy requirements of the cutting and polishing process. The results of the incremental training are synchronously updated to the customized scheduling large model, the thought tree branch logic, and the adaptive adjustment rules.

[0038] This application also provides an embodiment of a multi-condition adaptive task scheduling system for a cutting and grinding process, used to implement the multi-condition adaptive task scheduling method for the cutting and grinding process described in the above embodiment, comprising: The state perception module is used to collect cutting and grinding task information, multi-agent state information, and workshop dust environment information in real time at a preset frequency. It quantifies the workshop dust environment information to obtain the environmental interference coefficient, quantifies the cutting and grinding task information to obtain the task difficulty coefficient, and analyzes the multi-agent state information to obtain the skill matching degree. The task parsing module is used to perform task splitting, priority sorting, and data standardization operations on the cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on the task difficulty coefficient, skill matching degree, and environmental interference coefficient. The scheduling decision module has a built-in customized scheduling model. It receives the data collected by the state perception module and the preprocessed data by the task parsing module. It loads the knowledge base dedicated to the cutting and grinding process and generates a variety of feasible scheduling schemes by combining the large model with the mind tree. After comprehensive scoring by the preset evaluation system, the optimal scheduling scheme is selected and the execution instructions are output to each intelligent agent. At the same time, multiple alternative scheduling schemes are stored. The execution feedback module is used to collect three types of feedback information in real time during the task execution process: the execution status of the intelligent agent, the changes in the cutting and grinding conditions, and the real-time changes in the dust environment in the workshop. The feedback information is processed in layers, and the results of the layered processing and the execution log are transmitted to the scheduling decision module, the adaptive adjustment module, and the reinforcement learning training module. The adaptive adjustment module is used to determine whether to trigger the scheduling strategy adjustment based on the hierarchical processing results of the execution feedback module and the real-time information collected by the state perception module. In case of sudden working conditions, it calls the alternative scheduling scheme or re-optimizes and generates the scheduling scheme, and outputs the adjusted scheduling scheme to each intelligent agent. The reinforcement learning training module is used to collect data from the entire cutting and polishing process. After preprocessing the data, a standardized training dataset is constructed. A reinforcement learning training framework is built to carry out iterative and incremental training, correct the scheduling decision logic and adaptive adjustment rules of the customized scheduling model, and update the optimized model parameters to the scheduling decision module in a synchronous manner, forming a closed-loop iterative optimization of the scheduling model.

[0039] For ease of understanding, this application provides a more specific embodiment: The overall technical solution of this application is as follows: A multi-condition adaptive task scheduling system for cutting and polishing processes is constructed. This system comprises six core modules: a state perception module, a task parsing module, a scheduling decision module, an execution feedback module, an adaptive adjustment module, and a reinforcement learning training module (focusing on the adaptation and optimization of the cutting and polishing processes). The state perception module collects real-time information on customized panel furniture cutting and polishing tasks, the operating status of multiple intelligent agents, and workshop dust environment information. The task parsing module decomposes, prioritizes, and standardizes the cutting and polishing tasks. Based on the preprocessed information, the scheduling decision module uses a large model combined with a mind tree approach to make scheduling decisions, generating various possible scheduling methods through the mind tree. The scheduling scheme, after optimization by a large model, achieves optimal matching of tasks with multiple agents and plans the sequence of cutting and polishing tasks. The execution feedback module collects real-time data on agent task execution status and changes in cutting and polishing conditions. When unexpected situations arise, such as task changes, agent malfunctions, or increased dust interference, the adaptive adjustment module triggers a re-optimization of the scheduling strategy, achieving dynamic adaptive scheduling of multiple agents collaborating in the cutting and polishing process. Finally, relying on a reinforcement learning training module, the large model and scheduling decision logic are continuously trained and optimized based on data from the entire cutting and polishing process, forming a closed-loop mechanism of "data acquisition - parsing and scheduling - execution feedback - adaptive adjustment - model optimization," ensuring efficient and continuous completion of cutting and polishing tasks. The entire technical solution requires no manual intervention, automatically adapts to various working conditions in customized panel furniture cutting and polishing, improves resource utilization and scheduling robustness, adapts to the entire cutting and polishing process, and facilitates the intelligent upgrading of this process.

[0040] The method described in this application includes the following steps, which are carried out one by one according to the task execution process of cutting and sanding customized panel furniture: Step 1: Collecting Scene Status Information for Cutting and Grinding The system collects three core types of information during the cutting and polishing of customized panel furniture using a state-aware module. This ensures the real-time nature and accuracy of information collection, providing reliable data support for scheduling decisions and adapting to the high precision and continuity requirements of cutting and polishing. Specifically, this includes: 1. Cutting and Grinding Task Information: Collect the type (e.g., custom irregular board cutting, regular board cutting, rough grinding, fine grinding), quantity, priority, completion deadline, and task difficulty coefficient (quantified based on board specification complexity, cutting precision requirements, and grinding roughness requirements, with a value range of 0-1, the larger the value, the higher the difficulty) of the custom panel furniture board cutting, grinding, and polishing tasks to be executed, as well as the intelligent agent skills required for task execution (e.g., high-precision board cutting skills, irregular board cutting skills, fine grinding skills), and task relationships (e.g., the order of cutting → grinding, the connection between rough grinding → fine grinding). 2. Status information of multiple intelligent agents: Real-time load (number of tasks currently being executed, CPU utilization, remaining battery power, and operation time) of each cutting and grinding tool intelligent agent, skill matching degree (the degree of compatibility between the cutting / grinding skills possessed by the agent and the skills required for the current task, with a value range of 0-1), operating status (normal, minor fault, shutdown fault), current position and workable range (cutting area, grinding area), and operation accuracy (cutting size deviation, grinding roughness); 3. Workshop dust environment information: Relying on dust sensors and obstacle sensors mounted on multiple intelligent agents, the dust concentration, sawdust accumulation, obstacle distribution and other interference factors in the customized panel furniture cutting and sanding workshop are collected to quantify the environmental interference coefficient (the value ranges from 0 to 0.5, the larger the value, the stronger the interference, which is directly related to the cutting accuracy and sanding roughness).

[0041] The information collection frequency can be adjusted according to the cutting and grinding cycle. By default, it collects data once every 50ms to ensure real-time tracking of cutting and grinding conditions, intelligent agent status, and changes in the dust environment, meeting high response requirements.

[0042] Step Two: Cutting and Grinding Task Analysis and Pre-processing The task parsing module preprocesses the collected cutting and polishing task information to provide standardized data for scheduling decisions, adapting to the multi-task and customized needs of cutting and polishing custom panel furniture. Specifically, this includes: 1. Task Breakdown: Complex cutting and polishing tasks (such as batch custom board cutting and polishing) are broken down into multiple independent sub-tasks (such as single custom board cutting, single board rough polishing, and single board fine polishing). Each sub-task has a clear and independent completion time limit, skill requirements, difficulty level, and relationship with other sub-tasks (such as cutting first and then polishing), ensuring that the sub-tasks can be executed independently without affecting the overall cutting and polishing rhythm. 2. Priority Ranking: Based on the requirements of customized orders, the urgency of tasks, and the cutting and sanding cycle time, the hierarchical analysis method is used to prioritize all tasks (including sub-tasks), dividing them into two levels: Level 1 (urgent and important, such as cutting and fine sanding of customized irregular-shaped boards) and Level 2 (important but not urgent, such as cutting and rough sanding of regular boards). Priority is given to ensuring the execution of Level 1 tasks, which is in line with the priority requirements of customized panel furniture production. 3. Data standardization: The quantitative indicators such as task difficulty coefficient, agent skill matching degree, and environmental interference coefficient are normalized and uniformly set to a value range of 0-1 to eliminate the influence of units, facilitate subsequent scheduling algorithm calculations, and ensure scheduling accuracy.

[0043] Step 3: Scheduling Decisions and Cutting & Grinding Task Allocation The scheduling decision module, based on preprocessed information, uses a combination of a large model and mind tree to make scheduling decisions and allocate tasks. The mind tree generates multiple feasible scheduling schemes, which are then selected by the large model to determine the optimal scheme and execute it. The specific process is as follows: 1. Large Model Initialization and Input: Preprocessed cutting and polishing task information, multi-agent state information, and workshop dust environment information (all multimodal standardized data) are input into a large model specifically optimized for customized panel furniture cutting and polishing. This large model adopts a Transformer architecture with 12 layers, and the training dataset contains over 100,000 data points related to customized panel furniture cutting and polishing scheduling. During model initialization, a dedicated knowledge base for cutting and polishing scheduling (including cutting and polishing process collaboration rules, load balancing standards, dust environment adaptation parameters, customized task processing rules, etc.) is loaded to ensure that decisions align with the needs of the cutting and polishing scenario. 2. Mind Tree Construction and Multiple Solution Generation: With "efficient execution of cutting and polishing, load balancing, continuous production, and accuracy achievement" as the core objectives, the scheduling decision logic is decomposed layer by layer using a mind tree approach to construct a three-level mind tree structure: the first-level nodes are the scheduling objectives, the second-level nodes are the decision dimensions (task allocation, sequence planning, path adaptation), and the third-level nodes are the specific scheduling solutions. Based on each branch of the mind tree, no fewer than three feasible scheduling solutions are generated. Each solution clearly defines the cutting / polishing task allocation results, the agent execution sequence, and the collaborative requirements, covering different load allocation and path planning scenarios, and adapting to customized cutting and polishing needs. 3. Large-scale model optimization and solution determination: The large-scale model constructs an evaluation system based on the core requirements of cutting and grinding (production efficiency, load balancing, dust environment adaptability, production continuity, and operation accuracy). It comprehensively scores the various solutions generated by the mind tree. The scoring dimensions include solution feasibility (weight 0.3), execution efficiency (weight 0.3), load balancing (weight 0.2), and environmental adaptability (weight 0.2). The optimal solution with the highest score is selected as the final scheduling solution. At the same time, the top two alternative solutions are retained for quick switching in case of emergencies. 4. Solution Output and Execution: The large model breaks down the optimal scheduling scheme into specific execution instructions, clearly defining the cutting / polishing task allocation list, execution order, estimated completion time, and coordination parameters for each embodied intelligent agent, and synchronously transmitting them to each intelligent agent to start task execution; alternative schemes are stored in the scheduling system and can be called at any time.

[0044] Step 4: Task execution and cutting / grinding status feedback Multiple intelligent agents synchronously receive specific execution instructions based on the optimal scheduling scheme output by the large model, initiating the customized panel furniture cutting and sanding task execution process. During execution, they strictly adhere to the cutting and sanding process requirements and coordination parameters to ensure that the actions of each agent are synchronized, smoothly connected, and in line with the cutting and sanding rhythm. The execution feedback module adopts a working mode of "real-time acquisition + layered feedback + anomaly warning" to track the entire task execution process. The specific feedback content and process are as follows: First, three types of core information are collected in real time, including the agent's own execution status (task completion progress, execution accuracy, operation error, power consumption rate, CPU utilization, etc.), dynamic changes in cutting and sanding conditions (customized order addition / cancellation, board specification change, cutting and sanding rhythm adjustment), and real-time changes in the workshop dust environment (dust concentration, sawdust accumulation, obstacle position changes, etc.). The acquisition frequency is maintained at once every 30ms to ensure data real-time performance and provide accurate data support for subsequent adaptive adjustments. Secondly, the collected information is processed in layers. Normal execution data (such as task progress meeting targets, agent status being normal, and dust concentration being compliant) is synchronized in real-time to the scheduling and decision-making module for record-keeping and optimization of the large model through self-learning. Minor anomalies (such as slight deviations in cutting dimensions or substandard grinding roughness) receive initial warnings and are simultaneously fed back to the agent, triggering minor adjustments (such as correcting the cutting angle or adjusting the grinding speed). Severe anomalies (such as agent malfunctions causing shutdowns, batch damage to materials, or excessive dust concentration) immediately trigger emergency feedback, quickly pushing the data to the scheduling and decision-making module to initiate the anomaly handling process. Finally, the feedback module records the entire task execution process data in real-time, forming an execution log. This log is used for real-time adjustments to the task and is also aggregated and transmitted to the large model to optimize the thought tree branch logic and solution evaluation system, improving the accuracy of subsequent scheduling decisions.

[0045] Step 5: Adaptive scheduling and adjustment of the cutting and grinding scene The adaptive adjustment module receives feedback information from the execution feedback module in real time. Based on the characteristics of the customized panel furniture cutting and sanding scenario, it determines whether the scheduling strategy needs adjustment. Specific adjustment scenarios and methods are as follows, fully adapting to various unexpected working conditions during cutting and sanding: 1. Agent Failure Scenarios: If the cutting / grinding agent experiences a minor failure (such as decreased cutting accuracy or substandard grinding roughness), adjust the agent's task allocation to reduce its workload. Assign some sub-tasks to agents with high skill matching and low workload, and adjust the task execution rhythm to ensure that the cutting and grinding cycle is not affected. If the agent experiences a shutdown failure, immediately transfer all its unfinished cutting / grinding tasks, re-plan the task sequence, and prioritize assign it to idle agents with suitable skills to ensure continuous production. 2. Production task change scenarios: If an emergency customized board cutting / grinding task is added, all tasks will be prioritized again, and the task allocation plan and execution sequence will be adjusted to ensure the execution of emergency customized tasks while taking into account the progress of the original production tasks to avoid affecting the overall production plan; if a task is canceled or its difficulty is adjusted (such as changes in board specifications or adjustments to grinding roughness requirements), the task information will be updated in a timely manner and the scheduling plan will be optimized again. 3. Dust environment change scenario: If the dust concentration in the workshop increases and the environmental interference coefficient is ≥0.4, adjust the execution path and operation parameters of the intelligent agent, give priority to assigning intelligent agents with strong dust resistance and high operation accuracy to perform tasks in this area, extend the operation detection time, ensure cutting accuracy and grinding roughness, and avoid product quality problems caused by dust interference; 4. Uneven load scenario: If the load variance of multiple agents is detected to be ≥0.2 (uneven load), the task allocation is adjusted, and some subtasks of agents with excessive load are transferred to idle agents or agents with lower load, so as to achieve agent load balancing and improve resource utilization.

[0046] Step Six: Training and Scheduling Optimization of Reinforcement Learning Model Based on Cutting and Grinding Data Based on the cutting and polishing process data generated in steps one through five, reinforcement learning is used to continuously train and optimize the large model and scheduling decision logic. The focus is on improving task scheduling efficiency, system robustness, and adaptability to cutting and polishing scenarios, forming a closed-loop iterative mechanism of "data acquisition - model training - scheduling optimization". The specific implementation process is as follows: 1. Production Data Aggregation and Preprocessing: Collect the full-process cutting and polishing data generated in steps one through five. The core data includes scene status acquisition data, task parsing and scheduling data, task execution feedback data, and adaptive adjustment data. Clean, denoise, and normalize the data, remove invalid and redundant data, and extract core feature parameters (such as task completion time, load balancing variance, anomaly handling response time, accuracy compliance rate, and product qualification rate) to build a standardized training dataset. 2. Reinforcement Learning Training Framework Construction: With "maximizing scheduling efficiency, optimizing load balancing, maximizing robustness, maximizing anomaly handling, and maximizing product qualification rate" as the core reward objectives, a reinforcement learning training framework is constructed. This framework defines the intelligent agent (i.e., the scheduling decision-making system), the environment (dynamic cutting and grinding conditions), actions (scheduling scheme generation, task allocation, and adaptive adjustment), and a reward function. The reward function is designed in conjunction with the core requirements of cutting and grinding, specifically as follows: In the formula, E represents task execution efficiency, L represents agent load balancing, Rb represents system robustness, T represents anomaly response time, and Q represents product qualification rate. The weighting coefficients (the sum of the five factors is 1) have specific values. , , , , ; 3. Reinforcement learning model training and iteration: Input the pre-processed and refined data into the reinforcement learning framework, use the scheduling and decision-making logic of the large model combined with the thinking tree as the initial strategy, conduct no less than 300 rounds of iterative training, and gradually correct the scheme evaluation weights of the large model, the branching logic of the thinking tree and the adaptive adjustment rules until the core indicators of the model tend to stabilize (indicator fluctuation ≤5%). 4. Implementation and continuous optimization of training results: The optimized model parameters after training are updated synchronously to the scheduling system. At the same time, an incremental training mechanism is established. After each batch of cutting and grinding tasks is completed, new data is automatically collected for training, so that the model can dynamically adapt to changes in cutting and grinding conditions and continuously improve scheduling efficiency and robustness.

[0047] This application focuses on the collaborative scheduling design of multiple intelligent agents in the cutting and polishing process of customized panel furniture. It specifically optimizes the process due to its high precision, customization, and significant dust interference. The various features work together in a coordinated and layered manner. The specific implementation method and its differences from existing technologies are as follows: 1. Multi-dimensional real-time status perception features of cutting and grinding scenarios, corresponding to step one.

[0048] Specific implementation: Using dust sensors, precision sensors, and position sensors mounted on the intelligent agents of cutting and grinding tools as core data acquisition carriers, and combining edge computing technology, a comprehensive state perception system for the cutting and grinding scenario is constructed. This system collects three core types of information in real time: cutting and grinding tasks, agent status, and dust environment, with a collection frequency of 50ms / time and a data error of ≤3%. It focuses on collecting specific information such as customized sheet material specifications, cutting precision requirements, and grinding roughness requirements, as well as interference parameters such as workshop dust concentration, providing accurate data support for scheduling decisions.

[0049] The difference from existing technologies is that this application focuses on the single process of cutting and polishing customized panel furniture, and the information collected is tailored to the core needs of this process. It specifically collects exclusive information such as the cutting accuracy, polishing roughness and dust interference of customized panels, so as to achieve seamless connection between information collection and scheduling decision-making. Existing technologies do not collect the above-mentioned exclusive information in a targeted manner, and the collection dimensions are not well adapted to the cutting and polishing process.

[0050] 2. Refined analysis and preprocessing features oriented towards cutting and grinding tasks, corresponding to step two.

[0051] Specific implementation methods: In view of the customized and multi-type characteristics of custom panel furniture cutting and sanding tasks, the complex task is broken down into independent sub-tasks of cutting and sanding, and the connection relationship between the sub-tasks of "cutting → sanding" is clarified; the hierarchical analysis method is adopted, and the tasks are divided into two levels of priority based on the urgency of the customized orders and the accuracy requirements of the operation, so as to give priority to the core customized tasks; the relevant quantitative indicators are standardized to provide standardized data for scheduling decisions.

[0052] The difference from the prior art is that the preprocessing of this application is in line with the characteristics of the cutting and grinding process, the task breakdown takes into account the process connection, and the priority ranking focuses on customized needs; the task breakdown of the prior art is rough, the priority ranking does not take into account the cutting and grinding precision requirements, and the preprocessing is poorly adapted to the process.

[0053] 3. The large model combines the segmentation and refinement of the mind tree with the scheduling and decision-making features, corresponding to step three.

[0054] Implementation details: A customized, optimized model for cutting and polishing panel furniture is used, loaded with a dedicated knowledge base for cutting and polishing, and combined with a three-level thinking tree to generate more than three feasible scheduling solutions. A dedicated evaluation system is used to select the best solution for execution, while two alternative solutions are retained. The large model's response time is ≤200ms, meeting real-time scheduling requirements. The solution is primarily adapted to the needs of customized panel cutting and high-precision polishing, while also taking into account environmental interference factors such as dust.

[0055] The difference from existing technologies is that the scheduling decision of this application focuses on the cutting and grinding process, loads a dedicated knowledge base into the large model, and adopts "large model + mind tree" to achieve the selection of multiple solutions, which significantly improves the decision accuracy and scenario adaptability. Existing technologies use general decision logic without the selection of multiple solutions, and cannot adapt to the customized and high-precision requirements of cutting and grinding.

[0056] 4. Layered feedback features for cutting and grinding conditions, corresponding to step four.

[0057] Specific implementation method: Construct a dedicated feedback system for cutting and grinding that consists of "real-time acquisition - hierarchical feedback - data retention". Core data is collected every 30ms. For minor anomalies such as cutting accuracy deviation or grinding roughness not meeting the standards, the agent is triggered to make minor adjustments. For serious anomalies such as agent failure or excessive dust, emergency feedback is triggered. At the same time, the entire process execution data is recorded to provide support for model training.

[0058] The difference from existing technologies is that the feedback system of this application is tailored to the cutting and grinding process and has a layered feedback mechanism, which can specifically handle abnormalities in cutting and grinding precision and dust interference; existing technologies do not have a layered feedback mechanism and cannot specifically handle the above-mentioned abnormalities.

[0059] 5. Adaptive adjustment features for all cutting and grinding conditions, corresponding to step five.

[0060] Specific implementation method: For common unexpected working conditions in the cutting and grinding process (smart agent failure, task change, dust interference, load imbalance), preset exclusive adjustment rules and judgment thresholds are set. The adjustment response time is ≤0.8s, no manual intervention is required, and priority is given to ensuring the customized task and operation accuracy, so as to ensure the continuous progress of the cutting and grinding process.

[0061] The difference from existing technologies is that this application covers all working conditions of cutting and grinding, and the adjustment rules are tailored to the characteristics of the process; existing technologies can only handle a single type of emergency, and the adjustment methods are not adapted to the characteristics of cutting and grinding, such as dust interference and high precision requirements.

[0062] 6. Reinforcement learning-driven closed-loop iterative optimization of features, corresponding to step six.

[0063] Specific implementation method: Based on the data of the entire cutting and polishing process, a reinforcement learning training framework is constructed. The core needs of cutting and polishing are used as the reward target. No less than 300 rounds of iterative training are carried out to establish an incremental training mechanism, realize the continuous optimization of the model, and adapt to the dynamic changes of the cutting and polishing working conditions of customized panel furniture.

[0064] The difference from existing technologies is that this application uses reinforcement learning combined with cutting and grinding-specific data to achieve continuous model iteration, which can adapt to the dynamic changes in customized cutting and grinding needs; existing technologies lack a model self-learning optimization mechanism and cannot adapt to the above dynamic changes.

[0065] The specific implementation method of this application is described in detail using a typical scenario of cutting and sanding customized panel furniture (batch customized wardrobe panel cutting and sanding process). This implementation method clearly defines the specific parameter settings, the entire process execution steps, the details of reinforcement learning training, and the effect verification, fully demonstrating the core advantages of this application and its adaptability and superiority in the cutting and sanding process; ensuring that those skilled in the art can fully implement this invention based on the description, and at the same time, it can be flexibly adapted to other customized panel furniture cutting and sanding scenarios.

[0066] Implementation method: Collaborative scenario of cutting and sanding panels for customized panel wardrobes This implementation uses the cutting and polishing process of customized panel wardrobe panels as an application scenario. Specifically, it applies the multi-agent collaborative adaptive scheduling method of this application to achieve efficient collaborative execution of the entire process of cutting (particleboard, MDF), rough polishing, and fine polishing of 30 sets of customized panel wardrobe panels. This process involves 6 agent robots (4 for cutting and 2 for polishing). All tasks are required to be completed within 6 hours, with priority given to the cutting and fine polishing of customized irregular-shaped panels. The cutting accuracy is guaranteed to be ≤0.5mm, the polishing roughness is ≤Ra1.6μm, the dust concentration in the workshop is controlled within 8mg / m³, and the production line operates continuously. At the same time, the scheduling model is continuously optimized by relying on reinforcement learning.

[0067] Step S1. Acquisition of cutting and grinding scene status information 1. Sensor Configuration: Each cutting agent is equipped with a dust sensor (measuring range...) The system includes a precision sensor (adapted for plate cutting precision detection), a position sensor (positioning accuracy ±0.5m), and a power sensor; each grinding intelligent unit is equipped with a dust sensor, a roughness sensor, and a power sensor; the scheduling system is equipped with an edge computing module for data preprocessing; and dust monitors are deployed in the workshop to assist in collecting dust concentration data.

[0068] 2. Parameter settings: Information acquisition frequency 50ms / time, environmental interference coefficient threshold 0.4 (dust concentration) (Judgment), agent skill matching threshold 0.7, remaining battery level threshold 25%, load balancing variance threshold 0.2, cutting accuracy threshold 0.5mm, and grinding roughness threshold Ra 1.6μm.

[0069] 3. Information Collection: Collect information on the cutting and sanding tasks of 30 sets of customized wardrobe panels (task types: irregular panel cutting, regular panel cutting, rough sanding, fine sanding; completion time limit: 6 hours, irregular panel cutting and fine sanding are the highest priority; difficulty coefficient: 0.5-0.8); collect status information of 6 intelligent agents (skill matching degree of 0.85-0.95 for 4 cutting agents, skill matching degree of 0.8-0.9 for 2 sanding agents); collect environmental information such as dust concentration in the workshop and quantify the environmental interference coefficient.

[0070] Step S2. Cutting and Grinding Task Analysis and Preprocessing 1. Task Breakdown: The task of cutting and polishing 30 sets of custom wardrobes is broken down into 180 sub-tasks (6 per set: 1 for irregular cutting, 2 for regular cutting, 2 for rough polishing, and 1 for fine polishing). The time limit for completing a single sub-task is ≤2 minutes, and the connection between "cutting → rough polishing → fine polishing" is clearly defined.

[0071] 2. Priority ranking: Using the analytic hierarchy process, the urgency of the order is assigned a weight of 0.4, the accuracy requirement of the operation is assigned a weight of 0.4, and the cycle time adaptability is assigned a weight of 0.2. Irregular cutting and fine grinding are assigned as first-level priority, and regular cutting and rough grinding are assigned as second-level priority.

[0072] 3. Data standardization: The task difficulty coefficient, agent skill matching degree, and environmental interference coefficient are normalized and uniformly ranged from 0 to 1.

[0073] Step S3. Scheduling Decision and Cutting & Grinding Task Allocation 1. Large Model Initialization and Input: The preprocessed multimodal data is input into the large model specifically optimized for cutting and polishing. This large model adopts the Transformer architecture with 12 layers. The training dataset contains more than 100,000 data related to the cutting and polishing scheduling of customized panel furniture. A dedicated knowledge base for cutting and polishing is loaded to ensure that the decisions are in line with the needs of the scenario.

[0074] 2. Mind Tree Construction and Multiple Solution Generation: Construct a three-level mind tree to generate three feasible scheduling solutions, clarify the task allocation results and execution sequence, and cover different load allocation scenarios.

[0075] 3. Large Model Optimization and Solution Execution: The large model scores the three solutions through a dedicated evaluation system, selects the optimal solution, breaks it down into execution instructions, assigns them to 6 intelligent agents, starts task execution, and retains 2 alternative solutions.

[0076] Step S4. Task execution and cutting / grinding status feedback Six intelligent agents execute cutting and grinding tasks according to instructions. The execution feedback module collects core data every 30ms. For minor anomalies such as slight deviations in cutting dimensions or slightly excessive grinding roughness, the agents are triggered to make adjustments. For serious anomalies such as malfunction of one cutting agent or excessive dust concentration, emergency feedback is immediately triggered and pushed to the scheduling decision module. At the same time, the entire process execution data is recorded to form an execution log.

[0077] Step S5. Adaptive scheduling and adjustment of the cutting and grinding scene. 1. Agent Failure Adjustment: The five irregular cutting sub-tasks that the faulty cutting agent failed to complete are assigned to another cutting agent with a lower load and a skill matching degree of 0.92, and the execution rhythm is adjusted to ensure that the progress is not affected.

[0078] 2. Dust Environment Adjustment: When the dust concentration reaches... (Interference coefficient 0.45) Adjust the execution path of the intelligent agent and extend the grinding detection time to ensure that the grinding roughness meets the standard.

[0079] 3. Load balancing adjustment: When a grinding agent was found to have a load variance of 0.23, its three coarse grinding subtasks were transferred to another idle grinding agent. After adjustment, the load variance was reduced to 0.16.

[0080] Step S6. Reinforcement Learning Model Training and Scheduling Optimization 1. Data aggregation and preprocessing: Collect the data from steps S1-S5, clean and denoise the data, extract the core feature parameters, and construct a standardized training dataset.

[0081] 2. Training Framework Setup: Construct a reinforcement learning training framework and define the reward function. The focus is on ensuring cutting accuracy, grinding roughness, and production efficiency.

[0082] 3. Model training and results implementation: Conduct 300 rounds of iterative training, optimize model parameters, update them synchronously to the scheduling system, and establish an incremental training mechanism.

[0083] In this embodiment, six intelligent agents completed the cutting and polishing of 30 sets of customized wardrobes within 6 hours, achieving a 100% task completion rate, a 99% cutting accuracy compliance rate, a 98.8% polishing roughness compliance rate, an agent load balancing variance of 0.16, and product errors caused by dust interference ≤0.3mm. After reinforcement learning optimization, scheduling efficiency improved by 14%, and anomaly handling response speed improved by 20%. Compared with existing static scheduling methods, task execution efficiency improved by 30%, resource utilization improved by 35%, production interruption time reduced by 80%, and product qualification rate improved by 7%, fully verifying the feasibility, superiority, and process adaptability of this application.

[0084] Figure 1The diagram shows the overall flowchart of the adaptive task scheduling method described in this application, clearly illustrating the closed-loop process of "data acquisition - task parsing - scheduling decision - execution feedback - adaptive adjustment - model optimization"; the arrows in the diagram indicate the data flow direction, and the rectangles indicate the core execution content of each step.

[0085] Figure 2 In this process, the workflow begins with the state awareness module (task / agent / environment), then proceeds sequentially through the task parsing module (splitting / sorting / standardization), the scheduling decision module (large model + mind tree), to multi-agent execution (segmentation / refinement), and finally forms a closed loop through the execution feedback module (layered feedback). The execution feedback module also connects to the adaptive adjustment module (anomalies / rescheduling) and the reinforcement learning training module (closed-loop optimization), both of which feed back to the scheduling decision module, enabling the system's adaptive adjustment and continuous optimization.

[0086] Figure 3 This system demonstrates the task allocation results, skill matching, and load balancing status of six embodied intelligent agents. The system includes: a standardized task pool for storing standardized task sets (Task_i set); a scheduling decision engine employing a large model and mind tree technology to receive tasks from the standardized task pool and make scheduling decisions; a slicing agent group (Agent_C1~C4) for executing slicing tasks; and a refining agent group (Agent_P1~P2) for executing refining tasks. The scheduling decision engine allocates tasks to the corresponding agent groups based on task requirements. The slicing agent group executes the slicing tasks, and the refining agent group executes the refining tasks, achieving parallel processing and efficient execution of tasks. This system improves the intelligence level and execution efficiency of task scheduling by combining a large model and mind trees.

[0087] Figure 4 In this process, the workflow begins with "Task execution starts," followed by the "Real-time status acquisition" step, which continuously monitors the task execution environment and status. The system evaluates the acquired status through an "Abnormality" judgment node: if no abnormality is detected, it directly proceeds to the "Continue task execution" step; if an abnormality is detected, it triggers the "Trigger scheduling adjustment" step, which then executes "Call alternative solutions / Replan" to generate a new execution strategy, and subsequently proceeds to "Continue task execution."

[0088] After the task continues, the system performs a "record execution data" step, storing key data from the current execution process. This data is then used for "reinforcement learning training" to optimize the decision model. Once training is complete, the system performs a "model update" step, updating the optimized model parameters to the decision engine. The updated model is then applied to the "real-time status acquisition" stage of subsequent tasks, forming a closed-loop optimization that improves the system's adaptability to anomalies and its long-term execution efficiency. This method, through the combination of anomaly detection, dynamic adjustment, and reinforcement learning, achieves intelligent and continuous optimization of the task execution process.

[0089] This application significantly improves the production efficiency of the cutting and sanding process for customized panel furniture. Compared with existing technologies, it increases task execution efficiency by 30%, resource utilization by 35%, reduces production downtime by 80%, and increases product qualification rate by 7%. It also reduces the cost of manual intervention, enables adaptive adjustment under multiple working conditions, and eliminates the need for manual intervention to handle emergencies. Furthermore, it improves product precision, achieving a 99% compliance rate for cutting accuracy and a 98.8% compliance rate for sanding roughness, thus meeting the needs of customized production. It is adaptable to various customized panel furniture cutting and sanding scenarios, with low adaptation costs, and can be quickly promoted and applied.

[0090] This application enriches the theoretical framework for the collaborative scheduling of multiple intelligent agents in specific processes, constructing a closed-loop scheduling theoretical framework of "perception-analysis-decision-feedback-adjustment-optimization," providing theoretical reference and practical paradigm for the research and development of intelligent scheduling technologies for similar intelligent manufacturing processes. Existing technologies are mostly static scheduling or simple semi-dynamic scheduling, and do not involve the integrated application of multi-dimensional working condition perception, large model + thinking tree decision-making, and reinforcement learning closed-loop optimization.

[0091] Alternative scenarios: For example, the sensor type can be replaced with other brands of sensors with the same accuracy and frequency according to actual needs; the basic architecture of the large model can be replaced with other Transformer-type architectures at the same level; the specific algorithm for data preprocessing can be replaced with a normalization algorithm of the same type, but the processing accuracy and efficiency must be consistent with the original solution.

[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and not to limit them. Although this application has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of this application. Any modifications or equivalent substitutions that do not depart from the spirit and scope of this application should be covered within the protection scope of the claims of this application.

Claims

1. A multi-condition adaptive task scheduling method for cutting and grinding processes, characterized in that, include: Step 1: Collect cutting and polishing task information, multi-agent status information, and workshop dust environment information in real time at a preset frequency in the customized panel furniture cutting and polishing scenario. The preset frequency is the information collection frequency dynamically set according to the production rhythm of the cutting and polishing process. Quantify the workshop dust environment information to obtain the environmental interference coefficient, quantify the cutting and polishing task information to obtain the task difficulty coefficient, and analyze the multi-agent status information to obtain the skill matching degree. Step 2: Perform task splitting, priority sorting, and data standardization operations on the collected cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on task difficulty coefficient, skill matching degree, and environmental interference coefficient. Step 3: Input the collected multi-agent state information, workshop dust environment information, and standardized cutting and polishing task data into the initialized customized scheduling model. The customized scheduling model is an intelligent decision-making model built on a machine learning architecture and adapted to the scheduling needs of customized panel furniture cutting and polishing processes. Multiple feasible scheduling schemes are generated using the large model combined with a mind tree approach. The optimal scheduling scheme is selected after comprehensive scoring of the feasible scheduling schemes through a preset evaluation system, while multiple alternative scheduling schemes are retained. The preset evaluation system is a multi-dimensional scheme evaluation system set according to the production needs of the cutting and polishing process. The optimal scheduling scheme is broken down into execution instructions and transmitted to each agent to start task execution. The alternative scheduling schemes are stored in the scheduling system. Step 4: Each agent executes the cutting and grinding task according to the execution instructions. The agent's execution status, changes in cutting and grinding conditions, and changes in the workshop dust environment are collected in real time at a preset frequency. The feedback information is then processed in layers. Step 5: Determine whether the scheduling strategy adjustment is triggered based on the result of the hierarchical processing. If it is determined to be triggered, then based on the scene state information and feedback information, call the alternative scheduling scheme or re-optimize and generate the scheduling scheme for the sudden working conditions of the cutting and grinding process, and transmit the adjusted scheduling scheme to each intelligent agent for execution. Step 6: Collect the entire cutting and grinding process data generated in Steps 1 to 5. After preprocessing the data, construct a standardized training dataset. Build a reinforcement learning training framework with the core requirements of the cutting and grinding process scheduling as the reward objective. Input the standardized training dataset into the framework for iterative training. Correct the scheduling decision logic and adaptive adjustment rules of the large model through the reinforcement learning training framework. After the core indicators of the model tend to stabilize, update the optimized model parameters to the scheduling system. At the same time, establish an incremental training mechanism. Collect new data for incremental training after each batch of cutting and grinding tasks is completed.

2. The multi-condition adaptive task scheduling method of cutting and grinding processes according to claim 1, characterized in that, The cutting and polishing task information includes task type, quantity, priority, completion time limit, required agent skills, and task association; the multi-agent status information includes the real-time load, operating status, current location and workable range, and work accuracy of each cutting and polishing agent; the workshop dust environment information includes workshop dust concentration, sawdust accumulation, and obstacle distribution.

3. The multi-condition adaptive task scheduling method of cutting and grinding processes according to claim 2, characterized in that, The real-time load includes the number of currently executing tasks, hardware utilization, remaining battery power, and operation duration; the operating status includes normal, minor fault, and shutdown fault; the operation accuracy includes cutting size deviation and grinding roughness.

4. The multi-condition adaptive task scheduling method of cutting and grinding processes according to claim 2, characterized in that, Using various sensors mounted on the intelligent agent as the core data acquisition carrier, and combining edge computing technology, a full-dimensional state perception system is constructed to realize the real-time acquisition, quantification, and analysis of cutting and polishing scene information.

5. The multi-condition adaptive task scheduling method for the cutting and grinding process according to claim 1, characterized in that, The task breakdown involves dividing the complex cutting and polishing task into multiple independently executable sub-tasks, clearly defining the completion deadline, skill requirements, difficulty coefficient, and interrelationships between each sub-task; the priority ranking adopts the hierarchical analysis method, combining customized order requirements, task urgency, and cutting and polishing process cycle requirements to classify tasks into levels; the normalization process unifies the task difficulty coefficient, skill matching degree, and environmental interference coefficient to the same numerical range.

6. The multi-condition adaptive task scheduling method for the cutting and grinding process according to claim 1, characterized in that, The customized scheduling model is initialized by loading a dedicated knowledge base containing collaborative rules for cutting and grinding processes, load balancing standards, dust environment adaptation parameters, and customized task processing rules. The thinking tree is constructed with the core objectives of efficient execution of cutting and grinding, load balancing of intelligent agents, continuous production, and achievement of work accuracy standards, and generates multiple feasible scheduling schemes. The preset evaluation system has scoring dimensions including scheme feasibility, execution efficiency, load balancing, and environmental adaptability, with corresponding weight coefficients configured for each dimension.

7. The multi-condition adaptive task scheduling method for the cutting and grinding process according to claim 1, characterized in that, The execution status of the intelligent agent includes task completion progress, execution accuracy, operation error, power consumption rate, and hardware occupancy rate; the changes in cutting and sanding conditions include the addition / cancellation of customized orders, changes in board specifications, and adjustments to the cutting and sanding process cycle time; the real-time changes in the workshop dust environment include dust concentration, sawdust accumulation, and obstacle position changes; the results of the layered processing include normal execution data, slightly abnormal data, and severely abnormal data; the normal execution data includes task progress meeting standards, intelligent agent status being normal, and dust concentration meeting requirements; the slightly abnormal data includes minor deviations in cutting dimensions and substandard sanding roughness; the severely abnormal data includes intelligent agent malfunction and shutdown, batch damage to boards, and excessive dust concentration.

8. The multi-condition adaptive task scheduling method for the cutting and grinding process according to claim 1, characterized in that, The sudden operating conditions include agent failure, production task change, dust environment abnormality, and agent load imbalance. The scheduling strategy adjustment process does not require manual intervention. Among them, dust environment abnormality is when the environmental interference coefficient reaches a preset threshold, and agent load imbalance is when the variance of the load of multiple agents reaches a preset threshold. The preset threshold is a judgment threshold set according to the process requirements of the customized panel furniture cutting and polishing process and the production environment.

9. The multi-condition adaptive task scheduling method for the cutting and grinding process according to claim 1, characterized in that, The preprocessing of the entire process data includes data cleaning, noise reduction, and normalization. After removing invalid and redundant data, core feature parameters are extracted. These core feature parameters include task completion time, load balancing variance, anomaly handling response time, accuracy achievement rate, and product qualification rate. The reinforcement learning training framework is configured with an agent, environment, actions, and reward function. The agent is a scheduling decision system, the environment is the dynamic working conditions of cutting and grinding, and the actions are scheduling scheme generation, task allocation, and adaptive adjustment operations. The reward function is a multi-dimensional function that integrates task execution efficiency, agent load balancing, system robustness, anomaly response time, and product qualification rate. Each dimension is configured with corresponding weight coefficients, and the sum of the weight coefficients is 1. The iterative training is carried out for more than or equal to a preset number of rounds. The preset number of rounds is the number of iterations required to meet the model training convergence requirements. The criterion for the model's core indicators to stabilize is that the indicator fluctuations are within a preset range, which is the range of indicator fluctuations that meets the scheduling accuracy requirements of the cutting and grinding process. The results of the incremental training are synchronously updated to the customized scheduling large model, the thinking tree branch logic, and the adaptive adjustment rules.

10. A multi-condition adaptive task scheduling system for a cutting and grinding process, used to implement the multi-condition adaptive task scheduling method for the cutting and grinding process as described in any one of claims 1 to 9, characterized in that, include: The state perception module is used to collect cutting and grinding task information, multi-agent state information, and workshop dust environment information in real time at a preset frequency. It quantifies the workshop dust environment information to obtain the environmental interference coefficient, quantifies the cutting and grinding task information to obtain the task difficulty coefficient, and analyzes the multi-agent state information to obtain the skill matching degree. The task parsing module is used to perform task splitting, priority sorting, and data standardization operations on the cutting and polishing task information in sequence, and output standardized cutting and polishing task data; the data standardization is to perform normalization processing on the task difficulty coefficient, skill matching degree, and environmental interference coefficient. The scheduling decision module has a built-in customized scheduling model. It receives the data collected by the state perception module and the preprocessed data by the task parsing module. It loads the knowledge base dedicated to the cutting and grinding process and generates a variety of feasible scheduling schemes by combining the large model with the mind tree. After comprehensive scoring by the preset evaluation system, the optimal scheduling scheme is selected and the execution instructions are output to each intelligent agent. At the same time, multiple alternative scheduling schemes are stored. The execution feedback module is used to collect three types of feedback information in real time during the task execution process: the execution status of the intelligent agent, the changes in the cutting and grinding conditions, and the real-time changes in the dust environment in the workshop. The feedback information is processed in layers, and the results of the layered processing and the execution log are transmitted to the scheduling decision module, the adaptive adjustment module, and the reinforcement learning training module. The adaptive adjustment module is used to determine whether to trigger the scheduling strategy adjustment based on the hierarchical processing results of the execution feedback module and the real-time information collected by the state perception module. In case of sudden working conditions, it calls the alternative scheduling scheme or re-optimizes and generates the scheduling scheme, and outputs the adjusted scheduling scheme to each intelligent agent. The reinforcement learning training module is used to collect data from the entire cutting and polishing process. After preprocessing the data, a standardized training dataset is constructed. A reinforcement learning training framework is built to carry out iterative and incremental training, correct the scheduling decision logic and adaptive adjustment rules of the customized scheduling model, and update the optimized model parameters to the scheduling decision module in a synchronous manner, forming a closed-loop iterative optimization of the scheduling model.

Images