AI agent emergency order insertion dynamic decision production scheduling method, medium and system

Abstract

The invention provides an AI agent emergency order insertion dynamic decision production scheduling method, a medium and a system, and belongs to the technical field of industrial agents. A dynamic weight adaptive optimization model is adopted to calculate a target weight coefficient and construct a multi-target function set, an improved non-dominated sorting genetic algorithm is adopted to solve and output a Pareto optimal solution set, and a delay risk assessment correlation matrix is combined to start an incremental re-planning algorithm to generate a local adjustment scheme. The Pareto optimal solution set and the local adjustment scheme are combined to generate a final production scheduling scheme, a real-time monitoring module is started to track the execution deviation condition, and when it is detected that the deviation degree exceeds a threshold value, a rapid rescheduling mechanism is triggered to conduct scheme correction; the technical problem of poor production scheduling scheme quality caused by low multi-agent cooperation efficiency in the emergency order insertion dynamic decision process is solved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of industrial intelligent agents, and in particular, relates to an AI intelligent agent emergency order insertion dynamic decision-making production scheduling method, medium and system. BACKGROUND

[0002] Production scheduling systems in intelligent manufacturing environments widely use rule-based scheduling algorithms and traditional multi-objective optimization methods to handle emergency order insertion requests, and achieve rearrangement of production plans through static weight allocation and centralized decision-making frameworks, which have been widely applied in complex manufacturing scenarios such as aviation manufacturing, automobile production and electronic assembly. The traditional technology mainly relies on preset priority rules and fixed target weight coefficients, uses a single intelligent agent or a simple multi-agent system for decision-making, lacks dynamic adaptability and deep coordination mechanism between intelligent agents, and has obvious limitations in handling complex disturbance propagation and multi-dimensional constraint conflict. In the prior art, due to imperfect coordination mechanism between intelligent agents, insufficient dynamic weight adjustment capability, and lack of efficient disturbance propagation analysis method, problems such as decision response lag, inaccurate risk assessment of original order delay, and low resource allocation efficiency occur in the emergency order insertion process, which seriously affects the quality of the overall production scheduling scheme and the stability of the production system. That is, there is a technical problem of low multi-agent coordination efficiency in the emergency order insertion dynamic decision-making process in the prior art, which leads to poor quality of the production scheduling scheme. SUMMARY

[0003] Therefore, the application provides an AI intelligent agent emergency order insertion dynamic decision-making production scheduling method, medium and system, which can solve the technical problem of low multi-agent coordination efficiency in the emergency order insertion dynamic decision-making process in the prior art, which leads to poor quality of the production scheduling scheme.

[0004] The application is implemented in the following manner: in a first aspect, an AI agent emergency order insertion dynamic decision-making production scheduling method is provided, an agent coordination framework is constructed, emergency order insertion requests, existing order states, device operating parameters, and material inventory information are abstracted into order insertion agents, order agents, device agents, and material agents, and a distributed decision-making network architecture is established; real-time production environment data is collected, and a production state dataset is constructed; the state vectors of each agent are calculated based on the production state dataset, the influence range of order insertion operations is analyzed through a production disturbance propagation matrix, and the influence domain boundary parameters are determined; a multi-agent negotiation mechanism is started, each agent calculates a bidding scheme based on the state vector and the collaboration density vector, and a preliminary resource allocation scheme is output through a distributed auction algorithm; a dynamic weight self-adaptive optimization model is started to calculate the target weight coefficient, the multi-objective function set is converted into a comprehensive objective function, and the Pareto optimal solution set is output by solving through an improved non-dominated sorting genetic algorithm; the time dependence relationship between orders is quantified through a delay risk assessment correlation matrix, and when the delay risk assessment value exceeds the risk threshold, an incremental re-planning algorithm is started, and a local adjustment scheme is output; the Pareto optimal solution set and the local adjustment scheme are combined to generate a final production scheduling scheme, and a real-time monitoring module is started, and when the actual execution deviation exceeds the deviation tolerance threshold, a rapid rescheduling mechanism is triggered.

[0005] The real-time production environment data includes the priority level, expected delivery time, and resource demand of the emergency order insertion, the completion progress, remaining duration, and critical path node of the existing order, the load state, fault warning information, and production capacity parameter of the device, and the inventory level, replenishment cycle, and quality grade of the material.

[0006] The agents exchange information through a message passing protocol; the multi-objective function set includes minimizing order insertion response time, minimizing delay risk of existing orders, and maximizing resource utilization efficiency; the incremental re-planning algorithm only performs local optimization adjustment on the affected order subset, and keeps the original production scheduling plan of other orders unchanged.

[0007] The order insertion agent is responsible for receiving and processing emergency order insertion requests, analyzing the technical requirements and time constraints of order insertion, calculating the resource demand of order insertion on the production system, initiating negotiation requests to other agents and collecting bidding information, and finally determining the production scheduling position and execution scheme of order insertion.

[0008] The order agent manages the state information of existing production orders, including the process flow, completion progress, remaining duration, and quality requirements of the order, and evaluates the impact of order insertion operations on the order, providing feasible adjustment schemes and cost evaluation during the negotiation process.

[0009] The device intelligent agent monitors the running state and capacity parameters of the device, including device load rate, fault warning information, maintenance plan, and processing capacity, and provides available time window and processing quotation to the insertion order intelligent agent according to the state and constraint conditions of itself, and participates in the bidding process of resource allocation.

[0010] The material intelligent agent tracks the inventory level, quality grade, and replenishment cycle information of the material, evaluates the availability and supply capacity of the material required by the insertion order, and provides time commitment and cost quotation of material supply in the negotiation process to ensure the material guarantee of the insertion order execution.

[0011] The establishment step of the production disturbance propagation matrix is specifically that firstly, all resource nodes and order nodes in the production system are identified, and an association graph of resources and orders is constructed; secondly, connection weights between nodes are established according to resource sharing relationship and process flow dependency relationship; and then graph theory algorithm is used to calculate the path and strength of influence propagation.

[0012] The establishment step of the collaborative density vector is specifically that the interaction frequency and successful cooperation times of each intelligent agent in the historical negotiation process are counted; the information transmission efficiency and response delay between intelligent agents are calculated; and the trust score is constructed based on the cooperation history data.

[0013] The establishment step of the delay risk assessment correlation matrix is specifically that the process flow and resource demand of each order are analyzed, the front and rear dependency relationship between orders is identified; the influence degree and propagation path of order delay on subsequent orders are calculated; and a quantitative model of delay risk is established.

[0014] The specific structure of the dynamic weight adaptive optimization model is that the input layer receives the urgency of the insertion order, the resource load state, and the delivery time pressure environment state information; the hidden layer adopts a three-layer fully connected neural network structure, with 256 neurons in each layer, and the activation function is ReLU; and the output layer generates three weight coefficients of insertion order response time weight, original order delay weight, and resource utilization efficiency weight.

[0015] The multi-objective optimization solving process is specifically that based on the resource allocation preliminary scheme, an insertion order response time objective function, an original order delay risk objective function, and a resource utilization efficiency objective function are constructed to form a multi-objective function set; the weight coefficients output by the dynamic weight adaptive optimization model are used to linearly combine the multi-objective function set into a comprehensive objective function by using the weighted summation method; the comprehensive objective function is iteratively optimized by using an improved non-dominated sorting genetic algorithm; and a Pareto optimal solution set containing multiple candidate schemes is generated, and the scheme with the highest comprehensive evaluation is selected and output.

[0016] The training data set of the dynamic weight adaptive optimization model is established, specifically historical single insertion processing data is collected, including single insertion parameters, production environment state, processing result information, a state-action-reward triple is constructed, wherein the state is a production environment parameter, the action is a weight selection strategy, and the reward is a comprehensive optimization effect score, the data is cleaned and standardized, and the training set and the verification set are divided.

[0017] The training step of the dynamic weight adaptive optimization model is specifically training by using a deep Q network algorithm, storing historical experience by using an experience replay mechanism, updating network parameters in batches, balancing exploration and utilization by using an epsilon-greedy strategy, updating a target network once every 1000 iterations, and monitoring loss function convergence during the training process, and stopping the training when the verification set error does not decrease for 10 consecutive iterations.

[0018] Further, the dynamic weight adaptive optimization model adopts a convolution acceleration mechanism based on frequency domain transformation, converts a spatial domain convolution operation into a point-by-point multiplication operation in the frequency domain, reduces the calculation complexity by using fast Fourier transform, processes the boundary effect by using an overlap saving method, and realizes learnable frequency domain feature extraction by parameterized design of a frequency domain filter.

[0019] Further, the dynamic weight adaptive optimization model adopts a collaborative optimization algorithm based on multi-agent reinforcement learning, models different components of the network as independent agents, realizes collaborative optimization between the components by distributed strategy learning, transmits gradient information by using a communication mechanism, and searches for a global optimal solution by using a reward sharing strategy.

[0020] The weight adjustment function input includes a single insertion emergency level, a current resource load rate, and an original order average remaining time, and the output is a normalized vector of three weight coefficients, wherein the single insertion emergency level value range is [0, 1], the resource load rate value range is [0, 100], the average remaining time value range is greater than or equal to 0, and the resource conflict factor value range is [0, 1].

[0021] The disturbance influence degree calculation function input includes a single insertion resource demand, an existing order resource occupation situation, and a total resource capacity, and the output is a disturbance influence degree value, which is used to quantify the disturbance intensity of the single insertion operation on the existing production plan, wherein the number of resource types is greater than or equal to 1; the collaborative efficiency evaluation function input includes negotiation round number, information transmission delay, and decision consistency degree, and the output is a collaborative efficiency score.

[0022] The second aspect of the application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are run in the computer, and are used for executing the AI intelligent agent emergency single insertion dynamic decision production scheduling method.

[0023] A third aspect of the present invention provides an AI intelligent agent emergency order insertion dynamic decision-making and production scheduling system, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0024] This invention constructs an AI-based multi-objective optimization-based agent collaboration framework, employing a dynamic weight adaptive optimization model and a multi-agent reinforcement learning collaboration mechanism. It establishes an efficient distributed decision-making network architecture and an intelligent weight adjustment strategy, addressing the key deficiency of low agent collaboration efficiency in traditional technologies. This invention accurately analyzes the influence domain boundary of order insertion operations through a production perturbation propagation matrix, combines a collaborative density vector and a delay risk assessment correlation matrix to achieve deep collaboration among agents, and uses a frequency domain transformation-based convolution acceleration mechanism to improve real-time response capabilities, fundamentally improving the efficiency and quality of multi-agent collaborative decision-making. This invention constructs a dynamically weighted multi-objective function set by minimizing order insertion response time, minimizing the delay risk of existing orders, and maximizing resource utilization efficiency. Combined with an improved non-dominated sorting genetic algorithm, it generates a high-quality Pareto optimal solution set, effectively improving the overall production scheduling quality of dynamic decision-making for emergency order insertion. Attached Figure Description

[0025] Figure 1 This is a flowchart of the method of the present invention.

[0026] Figure 2 This is a heatmap of the correlation matrix for order delay risk assessment in Example 2.

[0027] Figure 3 This is a vector distribution diagram of the agent-cooperative density in Example 2. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0029] like Figure 1 The diagram shown is a flowchart of an AI agent's dynamic decision-making and production scheduling method for emergency order insertion, provided by the first aspect of this invention. This method includes the following steps:

[0030] S01. Construct an intelligent agent collaboration framework, abstracting emergency order insertion requests, existing order status, equipment operating parameters, and material inventory information into order insertion intelligent agents, order intelligent agents, equipment intelligent agents, and material intelligent agents. Each intelligent agent exchanges information through a message passing protocol, establishing a distributed decision-making network architecture.

[0031] S02, collect real-time production environment data, including the priority of emergency insertion orders, expected delivery time, resource demand, the progress of existing orders, remaining time, critical path nodes, the load state of equipment, fault warning information, production capacity parameters, the inventory level of materials, replenishment cycle, quality level, and construct a production state data set;

[0032] S03, calculate the state vector of each agent based on the production state data set, analyze the influence range of insertion order operation on existing production plan through production disturbance propagation matrix, identify the affected order set and resource constraint bottleneck, and determine the influence domain boundary parameter;

[0033] S04, start the multi-agent negotiation mechanism, the insertion order agent sends a bid request to related equipment agents and material agents according to the influence domain boundary parameter, each agent calculates the bid scheme based on the state vector and the collaborative density vector, and outputs the preliminary resource allocation scheme through the distributed auction algorithm;

[0034] S05, start the dynamic weight self-adaptive optimization model to calculate the target weight coefficient under the current scenario, minimize the insertion order response time, minimize the original order delay risk, and maximize the resource utilization efficiency to construct a multi-objective function set, use the weighted method to convert the multi-objective function set into a comprehensive objective function, and solve and output the Pareto optimal solution set through the improved non-dominated sorting genetic algorithm;

[0035] S06, quantify the time dependence relationship between each order through the delay risk assessment correlation matrix, when the delay risk assessment value exceeds the risk threshold, start the incremental re-planning algorithm, only locally optimize and adjust the affected order subset, output the local adjustment scheme and keep the original scheduling plan of other orders unchanged;

[0036] S07, combine the Pareto optimal solution set and the local adjustment scheme to generate the final scheduling scheme, issue it to the execution layer, and start the real-time monitoring module to track the deviation during execution, when the actual execution deviation exceeds the deviation tolerance threshold, trigger the fast rescheduling mechanism to correct the scheme.

[0037] The single-insert intelligent agent is responsible for receiving and processing emergency single-insert requests, analyzing the technical requirements and time constraints of the single-insert, calculating the resource demand of the single-insert to the production system, initiating negotiation requests to other intelligent agents and collecting bid information, and finally determining the production scheduling position and execution scheme of the single-insert. The order intelligent agent manages the state information of existing production orders, including the process flow, completion progress, remaining time limit, quality requirements of the order, evaluates the impact of single-insert operation on its own order, and provides feasible order adjustment schemes and cost evaluation in the negotiation process. The equipment intelligent agent monitors the running state and capacity parameters of the equipment, including the equipment load rate, fault warning information, maintenance plan, and processing capacity, provides available time window and processing quotation to the single-insert agent according to its own state and constraint conditions, and participates in the bidding process of resource allocation. The material intelligent agent tracks the inventory level, quality grade, and replenishment cycle information of the material, evaluates the availability and supply capacity of the material required by the single-insert, provides time commitment and cost quotation of material supply in the negotiation process, and ensures the material guarantee of single-insert execution. The production state data set contains all relevant real-time information in the production environment, including single-insert parameters, order status, equipment status, and material status, providing a data basis for subsequent intelligent agent decision-making and optimization algorithms.

[0038] The state vector is a numerical representation calculated by each intelligent agent based on the production state data set, used to quantify the current state and decision-making ability of each intelligent agent. The influence domain boundary parameter is calculated by the production disturbance propagation matrix, used to determine the range of the single-insert operation impact, guiding the subsequent negotiation and optimization process. The multi-objective function set includes the minimum single-insert response time function, the minimum original order delay risk minimization function, and the maximum resource utilization efficiency function, providing quantitative objectives for production scheduling optimization. The comprehensive objective function is a single optimization objective formed by linear combination of the multi-objective function set through weighting method, used for improved non-dominated sorting genetic algorithm solution. The delay risk evaluation value is calculated by the delay risk evaluation correlation matrix, quantifying the risk level of order delay.

[0039] The risk threshold is the upper limit of the delay risk evaluation value, and the incremental re-planning is triggered when the delay risk evaluation value exceeds the risk threshold. The local adjustment scheme is the optimization adjustment result generated by the incremental re-planning algorithm for the affected order subset. The final production scheduling scheme is the comprehensive production scheduling decision result after merging the Pareto optimal solution set and the local adjustment scheme. The actual execution deviation is a quantitative indicator of the deviation between the actual execution and the final production scheduling scheme. The deviation tolerance threshold is the allowed upper limit of the actual execution deviation, and the fast rescheduling will be triggered when this threshold is exceeded.

[0040] The production disturbance propagation matrix is a mathematical model for describing the propagation path of the impact of emergency orders on the production system. The establishment step includes: first, identifying all resource nodes and order nodes in the production system, and constructing a resource and order association graph; second, establishing the connection weight between nodes according to the resource sharing relationship and the process flow dependence relationship; then, using graph theory algorithm to calculate the path and strength of the impact propagation. The elements in the matrix represent the disturbance propagation coefficient from the source node to the target node, the numerical range is [0, 1], and the larger the value, the stronger the impact of disturbance propagation.

[0041] The collaborative density vector is a quantitative indicator reflecting the close degree of cooperation between agents. The establishment step includes: counting the interaction frequency and successful cooperation times of each agent in the historical negotiation process; calculating the information transmission efficiency and response delay between agents; and constructing a trust score based on the cooperation history data. Each element of the vector corresponds to the cooperation strength value of an agent with other agents, and the element value range is [0, 1], which reflects the activity and reliability of the agent participating in collaborative decision-making.

[0042] The delay risk assessment correlation matrix is used to quantify the time dependence and delay propagation relationship between orders. The establishment step includes: analyzing the process flow and resource demand of each order, identifying the front and rear dependence relationship between orders; calculating the influence degree and propagation path of order delay on subsequent orders; and establishing a quantitative model of delay risk. The matrix elements represent the probability value of one order delay causing delay risk to another order, and the diagonal elements are 1, indicating the certainty of self-delay, and the non-diagonal elements range from 0 to 1, reflecting the correlation strength of delay risk between orders.

[0043] The dynamic weight adaptive optimization model is a weight adjustment mechanism based on reinforcement learning, which is specially used for dynamically calculating the weight coefficients of each objective function in multi-objective optimization according to the current production environment state. The specific structure of the dynamic weight adaptive optimization model is that the input layer receives the urgent degree of inserted orders, the resource load state, the delivery time pressure and other environmental state information; the hidden layer adopts a three-layer fully connected neural network structure, with 256 neurons in each layer and a ReLU activation function; and the output layer generates three weight coefficients of the inserted order response time weight, the original order delay weight and the resource utilization efficiency weight. The steps for establishing the training data set of the dynamic weight adaptive optimization model specifically include: collecting historical inserted order processing data, including inserted order parameters, production environment state, processing result information; constructing state-action-reward triplets, wherein the state is the production environment parameter, the action is the weight selection strategy, and the reward is the comprehensive optimization effect score; and cleaning and standardizing the data, and dividing the training set and the validation set. The steps for training the dynamic weight adaptive optimization model specifically include: training by using the deep Q network algorithm, setting the learning rate to 0.001 and the discount factor to 0.9; storing historical experience through the experience replay mechanism and updating network parameters in batches; balancing exploration and utilization by using the ε-greedy strategy, and updating the target network once every 1000 iterations; and monitoring the loss function convergence during the training process, and stopping the training when the validation set error does not decrease for 10 consecutive iterations.

[0044] The multi-objective optimization solving process is an optimization calculation process of production scheduling scheme based on the output weight coefficients of the dynamic weight adaptive optimization model, specifically including: first, constructing the inserted order response time objective function, the original order delay risk objective function and the resource utilization efficiency objective function based on the resource allocation preliminary scheme to form a multi-objective function set; then, linearly combining the multi-objective function set into a comprehensive objective function by using the weight coefficients output by the dynamic weight adaptive optimization model through the weighted summation method; then, iteratively optimizing the comprehensive objective function by using the improved non-dominated sorting genetic algorithm; and finally, generating a Pareto optimal solution set containing multiple candidate schemes, and selecting the scheme with the highest comprehensive evaluation to output.

[0045] The dynamic weight adaptive optimization model adopts a convolution acceleration mechanism based on frequency domain transformation, converts the spatial domain convolution operation into point-by-point multiplication operation in the frequency domain, reduces the computational complexity through fast Fourier transform, processes the boundary effect by using the overlap saving method, and realizes the learning of frequency domain feature extraction through the parameterization design of the frequency domain filter. The mechanism converts the time domain signal to the frequency domain for processing, avoids the high computational overhead of traditional convolution operation, significantly improves the real-time response capability of the model while maintaining the accuracy of feature extraction, so that the weight adjustment decision is completed within milliseconds, meeting the rapid response demand of emergency inserted orders.

[0046] The dynamic weight adaptive optimization model adopts a collaborative optimization algorithm based on multi-agent reinforcement learning, models different components of the network as independent agents, realizes collaborative optimization between components through distributed policy learning, and uses a communication mechanism to deliver gradient information and share strategies to encourage the search for global optimal solutions. The mechanism decomposes the traditional centralized optimization problem into multiple sub-problems, each agent is responsible for optimizing the target dimension, through information exchange and strategy coordination between agents, parallel computing of weight adjustment is realized, local optimal problem caused by single weight adjustment strategy is avoided, and the quality and robustness of the overall weight adjustment strategy are improved.

[0047] The weight adjustment function is used to calculate the dynamic weight coefficients of each target in multi-objective optimization, and the input includes the single insertion emergency level, the current resource load rate, the average remaining time of the original order, and the output is a normalized vector of three weight coefficients. The weight adjustment function is described as follows: , , wherein is the single insertion emergency level (dimensionless), , is the resource load rate (%), , is the maximum load rate (%), , is the average remaining time (h), , is the maximum remaining time (h), , is the resource conflict factor (dimensionless), .

[0048] The disturbance influence degree calculation function is used to quantify the disturbance intensity of the single insertion operation on the existing production plan, and the input includes the single insertion resource demand, the existing order resource occupation, and the total resource capacity. The output is the disturbance influence degree value. The disturbance influence degree calculation function is described as follows: wherein is the demand of the single insertion on resource i, is the occupation of the existing order on resource i, is the total capacity of resource i, is the number of resource types, , the units of the parameters are the same to ensure dimensional consistency.

[0049] The collaborative efficiency evaluation function is used to evaluate the efficiency level of the cooperation between agents, and the input includes the negotiation round number, information transmission delay, and decision consistency degree. The output is the collaborative efficiency score. The collaborative efficiency evaluation function is described as follows: wherein is the decision consistency degree (dimensionless), is the maximum consistency value (dimensionless), is the negotiation round number (times), is the maximum allowed round number (times), is the average information transmission delay (ms), is the maximum allowed delay (ms),

[0050] The specific implementation of the above steps is described in detail below.

[0051] The specific implementation of step S01 is to first establish an order insertion intelligent agent architecture, which is designed based on the agent mode, manages the internal state changes of the intelligent agent through the state mode, and realizes the information subscription relationship with other intelligent agents by using the observer mode. The order insertion intelligent agent is internally integrated with an emergency order insertion analysis module, which uses natural language processing technology to analyze order insertion text information and extract key parameters such as priority, delivery period, technical requirements, etc. The order intelligent agent is designed based on an event-driven architecture and uses a state machine model to manage the order life cycle. It broadcasts order state change information to other intelligent agents through a publish-subscribe mode. The equipment intelligent agent integrates an Internet of Things data acquisition interface and uses a time series database to store equipment historical operation data. It analyzes equipment failure risks through a predictive maintenance algorithm. The material intelligent agent uses a safety stock model in inventory management theory and combines a Monte Carlo simulation method to predict material demand fluctuations. An asynchronous communication mechanism based on a message queue is used between intelligent agents. The message transmission protocol uses a lightweight JSON format to ensure the real-time and reliability of information exchange. The distributed decision-making network architecture is designed based on a hierarchical control theory. The upper layer is a coordination layer responsible for global decision coordination, the lower layer is an execution layer responsible for specific task execution, and the middle layer is a negotiation layer responsible for negotiation between intelligent agents.

[0052] ​​​​​​The specific implementation of step S02 is to establish a real-time data acquisition system that adopts an edge computing architecture, deploys data acquisition nodes in the production site, and acquires equipment operation data through an industrial Ethernet protocol. The emergency order insertion data acquisition module adopts a form parsing technology to automatically identify the order insertion priority level, which is divided into five levels, with a value range of 1 to 5, and the highest priority being 5. The expected delivery date is parsed by a natural language time recognition algorithm and converted into a standard timestamp format. The resource demand is calculated by a process route analysis algorithm, which matches the corresponding process based on the technical requirements of the order insertion and calculates the demand for various resources. The existing order status is acquired through a manufacturing execution system interface, the completion progress is calculated by the process completion ratio, and the remaining duration is predicted based on a regression analysis model of historical data. The key path nodes are identified by the critical path method in project management, and the earliest start time and latest start time of each process are calculated. The equipment state data is acquired in real time through a sensor network, and the load state is obtained through a device utilization rate calculation formula, with a threshold value of 85%, which is considered as a high load state if exceeded. The fault warning information is identified by an anomaly detection algorithm in machine learning based on multi-dimensional data such as vibration analysis and temperature monitoring. The material inventory data is obtained through RFID technology and barcode recognition systems, and the inventory level is classified by the ABC classification management method, with a safety stock of 15 days for A-class materials, 10 days for B-class materials, and 5 days for C-class materials.

[0053] The specific implementation of step S03 is to construct the state vector of each agent based on the collected production state data. The state vector adopts a multi-dimensional feature representation method, and the dimension of each agent's state vector is determined according to the type of data it manages. The order insertion agent state vector includes dimensions such as urgency, resource demand intensity, and time urgency, and each dimension value is mapped to the 0-1 interval through normalization processing. The disturbance propagation matrix is constructed based on the adjacency matrix in graph theory, the influence propagation path is identified by the depth-first search algorithm, and the influence weight of each node is calculated by the PageRank algorithm. The disturbance propagation coefficient is calculated by correlation analysis of historical data, and the Pearson correlation coefficient is used to measure the influence intensity between nodes. The influence domain boundary parameter is determined by clustering algorithm, and the affected orders and resources are divided into different influence domains by K-means clustering method, with a threshold value of 0.3 for the influence domain boundary, and nodes exceeding the threshold value are included in the influence domain range. The affected order set is identified by graph traversal algorithm, starting from the order insertion node and searching all affected order nodes along the disturbance propagation path. The resource constraint bottleneck is identified by the bottleneck analysis method in constraint theory, and the load rate and buffer capacity of each resource are calculated. Resources with a load rate exceeding 90% and a buffer capacity less than 5% are identified as bottleneck resources.

[0054] The specific implementation of step S04 is to start the multi-agent negotiation mechanism based on the contract net protocol, and the insertion order agent sends a task announcement to the related device agent and the material agent as a manager, which contains the specific requirements, time constraints, quality requirements, etc. of the insertion order. Each agent calculates the bidding scheme based on its own state vector, which contains the amount of resources that can be provided, the time window, the cost offer, etc. The coordination density vector is calculated by the centrality index in network analysis, and the eigenvector centrality is used to measure the importance of the agent in the coordination network. The bidding scheme evaluation adopts the multi-criteria decision analysis method, and the evaluation index system includes four dimensions of cost, time, quality and reliability. The distributed auction algorithm adopts the Vickrey auction mechanism to ensure the authenticity of the agent's bid, and determines the winning scheme through the second price sealed auction method. The resource allocation preliminary scheme is solved based on the Hungarian algorithm, which converts the matching problem of agents and tasks into the maximum weight matching problem of bipartite graph, ensuring the optimality of resource allocation. The negotiation process adopts the round management, and the time limit for each negotiation round is 30 seconds, and the maximum negotiation round is set to 5 rounds to ensure the negotiation efficiency.

[0055] The specific implementation of step S05 is to start the dynamic weight adaptive optimization model, which is constructed based on the deep Q network algorithm in reinforcement learning, and learns the optimal weight allocation strategy in different production scenarios through neural network. The multi-objective function set includes three sub-objective functions, the insertion order response time minimization function is modeled using time window theory to calculate the time interval from receiving to starting execution of the insertion order. The original order delay risk minimization function is based on risk assessment theory, and the Monte Carlo simulation method is used to calculate the order delay probability distribution. The resource utilization efficiency maximization function uses the efficiency measure method in operations research to calculate the effective utilization rate of resources. The weighted method linearly combines the multi-objective functions into a comprehensive objective function, and the weight coefficients are obtained by real-time calculation of the dynamic weight adaptive optimization model. The improved non-dominated sorting genetic algorithm uses fast non-dominated sorting technology to maintain the diversity of solutions through crowding distance calculation, the crossover probability is set to 0.8, the mutation probability is set to 0.1, the population size is set to 100, and the iteration number is set to 200 generations. The Pareto optimal solution set is obtained by non-dominated solution screening, and the solution set size is controlled within 20 to ensure the reasonable selection space of the decision maker.

[0056] The specific implementation of step S06 is to quantify the time-dependent relationship between orders by delaying risk assessment correlation matrix, which is constructed based on Bayesian network theory and uses conditional probability to represent the delayed propagation relationship between orders. The delay risk assessment value is calculated by matrix operation, considering factors such as process flow dependence, resource sharing relationship, and time constraints. The risk threshold is determined based on statistical analysis of historical data, and the upper limit of the 95% confidence interval is used as the risk threshold, usually set to 0.7. When the delay risk assessment value exceeds the risk threshold, the incremental re-planning algorithm is started, which is based on the rolling horizon optimization theory and only re-plans the affected order subset. The incremental re-planning uses the neighborhood search algorithm to find the optimal adjustment scheme through local search strategy, and the search radius is set to 3 time units. The local adjustment scheme is generated using the greedy algorithm, which prioritizes the orders with the smallest impact and achieves local optimization through order exchange and time translation. The original scheduling plan of other orders is maintained during the adjustment process to ensure the locality and stability of the adjustment.

[0057] The specific implementation of step S07 is to merge the Pareto optimal solution set and the local adjustment scheme to generate the final scheduling scheme using the scheme fusion algorithm. The scheme merging process uses the constraint satisfaction problem solving method to ensure that the merged scheme meets all the constraints. The final scheduling scheme is delivered to the execution layer through the manufacturing execution system interface, and the XML format data exchange standard is used to ensure the accuracy of information transmission. The real-time monitoring module is constructed based on complex event processing technology, which detects abnormal situations in the execution process through event stream analysis. The execution deviation is calculated by comparing the actual progress with the planned progress, and the deviation calculation uses the root mean square error method. The deviation tolerance threshold is set according to the stability requirements of the production system, usually set to 10% of the planned value. When the actual execution deviation exceeds the deviation tolerance threshold, the fast rescheduling mechanism is triggered, which uses online optimization algorithms to recalculate the scheduling scheme through real-time data updates. The rescheduling process uses an incremental calculation method, which only recalculates the part that has deviated, improving the efficiency of rescheduling. The scheme correction is achieved through the differential evolution algorithm, which finds a better adjustment scheme through mutation, crossover, and selection operations.

[0058] It needs to be further explained that the dynamic weight adaptive optimization model adopts a deep reinforcement learning architecture design, and the overall model is based on an improved version of the deep Q network. The input layer is designed as a 64-dimensional vector space, receiving single insertion urgency, resource load state, delivery time pressure, equipment availability, material sufficiency, historical delay rate and other multi-dimensional environmental state information. The urgency of the single insertion is quantified by the analytic hierarchy process to a value between 0 and 1, the resource load state uses the weighted average method to calculate the comprehensive load rate of each type of resource, and the delivery time pressure is calculated by the time urgency function. The hidden layer adopts a three-layer fully connected neural network structure, the first hidden layer contains 256 neurons, uses the ReLU activation function, and introduces the batch normalization technique to speed up the convergence process. The second hidden layer contains 256 neurons, also uses the ReLU activation function, and adds a Dropout layer to prevent overfitting, with a Dropout probability of 0.3. The third hidden layer contains 128 neurons, uses the ReLU activation function, and this layer is mainly responsible for high-level abstraction and representation learning of features. The output layer contains 3 neurons, corresponding to the single insertion response time weight, the original order delay weight, and the resource utilization efficiency weight, respectively, and uses the Softmax activation function to ensure that the weight coefficients sum to 1 and are all positive. The model also integrates a feature selection module based on the attention mechanism, which calculates the importance weights of each input feature through self-attention to improve the model's sensitivity to key features.

[0059] The training data set establishment process adopts a multi-source data fusion strategy. First, collect historical single insertion processing data, covering a time span of 12 months, containing 50,000 records of single insertion processing records of different types and different urgency. Single insertion parameter data includes single insertion number, product type, quantity demand, technical requirements, priority and other basic information, and abnormal values and missing values are removed through data cleaning. Production environment state data is exported through the production management system, including order status, equipment load, material inventory, personnel configuration and other real-time state information, and continuous time series data is divided into fixed length state sequences using time window division method. The processing result information includes the final production scheduling scheme, actual execution effect, customer satisfaction and other feedback data, and the processing effect is quantified by comprehensive evaluation method. In the state-action-reward triplet construction process, the state vector is dimensionally reduced by principal component analysis method, retaining 95% of the information. The action vector represents the weight selection strategy, and a continuous action space is designed, with weight values continuously changing in the range of 0 to 1. The reward function design considers single insertion response speed, original order impact degree and resource utilization efficiency, and calculates the comprehensive reward value using weighted combination method. In the data preprocessing stage, the Z-score standardization method is used to normalize numerical features, ensuring the comparability of features with different dimensions. The data set is divided into training set, validation set and test set in the ratio of 7:2:1, ensuring effective evaluation of the model's generalization ability.

[0060] It should be noted that the first key technical idea of the present application is a distributed decision-making mechanism based on a multi-agent collaborative framework. Traditional emergency order insertion processing usually adopts a centralized decision-making approach, which is manually adjusted by dispatch personnel based on experience. This approach is slow in response and prone to local optimal solutions. The present application abstracts order insertion requests, order status, equipment operation, and material inventory into independent agents, which make autonomous decisions based on their own status and constraints, and achieve optimal allocation of resources through negotiation mechanisms. This distributed architecture can handle multiple decision-making tasks in parallel, significantly improving decision-making efficiency while avoiding information transmission delays and bottlenecks in centralized decision-making. The negotiation process between agents is based on market mechanism design, ensuring fairness and efficiency of resource allocation, and better balancing the interests of all parties compared to traditional methods.

[0061] The second key technical idea is the introduction of a dynamic weight adaptive optimization model. Existing multi-objective optimization methods usually use fixed weight coefficients, which cannot be dynamically adjusted according to changes in the production environment, resulting in a lack of relevance in the optimization results. The dynamic weight adaptive optimization model designed in the present application is based on deep reinforcement learning principles and can automatically learn and adjust the weight coefficients of each objective function according to the current production state. The model obtains the optimal weight distribution strategy under different scenarios through historical data training, and can quickly generate a weight configuration that adapts to the current environment when faced with new order insertion requests. This adaptive mechanism enables the optimization process to better reflect the actual needs of the production site, improving the practicality and effectiveness of the production scheduling plan.

[0062] The third key technical idea is the application of incremental re-planning algorithms. Traditional re-planning methods usually require re-computation of the entire production plan, which has high computational complexity and can cause excessive disruption to stable production orders. The incremental re-planning algorithm used in the present application is based on influence domain analysis and only adjusts the subset of orders affected by the order insertion, leaving the original plans of other orders unchanged. This localized processing approach significantly reduces computational complexity, improves response speed of re-planning, and minimizes disturbance to existing production arrangements. The algorithm accurately identifies the impact range by delaying risk assessment correlation matrix, ensuring the necessity and sufficiency of adjustments.

[0063] The fourth key technical idea is the integration of real-time monitoring and fast rescheduling mechanisms. Existing systems often lack effective execution monitoring and dynamic adjustment capabilities, making it difficult to respond and correct deviations in the execution process in a timely manner. The present application establishes a complete real-time monitoring system that continuously tracks the execution status through complex event processing technology and automatically triggers the fast rescheduling mechanism when it detects that the execution deviation exceeds the tolerance threshold. The mechanism uses online optimization algorithms to quickly generate adjustment schemes without interrupting production, ensuring the execution quality of the production scheduling plan.

[0064] The synergy of the four key technical ideas forms a complete intelligent emergency order insertion processing system. The distributed decision-making mechanism provides an efficient decision-making framework, the dynamic weight self-adaptive optimization ensures the pertinence of the decision-making, the incremental re-planning realizes the rapid response, and the real-time monitoring guarantees the execution quality. The four cooperate with each other to form a closed-loop control system from decision-making to execution. Compared with the traditional method, this collaborative architecture can significantly improve the processing efficiency and quality of emergency order insertion, reduce the impact on the existing production order, and improve the flexibility and robustness of the entire production system. Through information sharing and coordination optimization among intelligent agents, the system can achieve global optimization while ensuring individual interests, achieving the organic unity of local intelligence and global intelligence.

[0065] It should be noted that there is a technical problem of inaccurate disturbance propagation analysis affecting domain identification error in the existing technology during the emergency order insertion processing process. The traditional disturbance analysis method mainly relies on experience rules and simple correlation analysis, and cannot accurately quantify the specific influence range and propagation path of the order insertion operation on the production system, resulting in problems of over-adjustment or insufficient adjustment when formulating the production scheduling adjustment scheme. The present application establishes a production disturbance propagation matrix, uses a graph theory algorithm to accurately calculate the disturbance propagation coefficient from the source node to the target node, and the matrix element numerical range is 0 to 1, which can scientifically quantify the influence intensity of disturbance propagation, combined with the calculation of the boundary parameter of the influence domain, accurately identifies the affected order set and resource constraint bottleneck, and provides reliable boundary constraint conditions for the subsequent negotiation and optimization process, effectively solving the technical problem of inaccurate disturbance propagation analysis.

[0066] It should be noted that the traditional multi-objective optimization method usually adopts fixed weight coefficients for linear combination of the objective function, and cannot adjust the optimization focus according to the dynamic changes of the production environment. When facing different emergency order insertion requests and changing resource load conditions, static weight distribution often leads to deviation of the optimization direction from the actual demand. The present application constructs a dynamic weight self-adaptive optimization model, trains using a deep Q network algorithm, establishes an input layer containing order insertion emergency degree, resource load state, and delivery time pressure, and calculates three dynamic weight coefficients of order insertion response time weight, original order delay weight, and resource utilization efficiency weight in real time through a three-layer fully connected neural network structure, so that the multi-objective optimization can adaptively adjust the optimization focus according to the current production environment state, and significantly improve the pertinence and practicality of the optimization effect.

[0067] The second aspect of the present application provides a computer readable storage medium, the computer readable storage medium stores program instructions, the program instructions are used to execute the AI intelligent agent emergency order insertion dynamic decision-making production scheduling method when running in the computer.

[0068] The third aspect of the present application provides an AI agent emergency order insertion dynamic decision-making production scheduling system, comprising the computer readable storage medium described above, the system is any one of a computer, a server, and a single-chip microcomputer, the computer readable storage medium is arranged in the system, and the system is provided with a microprocessor for executing program instructions stored in the computer readable storage medium.

[0069] Specifically, the principle of the present application is that the multi-level agent collaborative optimization architecture of the present application can solve the root cause of the low coordination efficiency of traditional technologies, which is to establish a complete distributed decision-making network and a dynamic adaptation mechanism. By abstracting the emergency order insertion request, the existing order state, the equipment operation parameter, and the material inventory information into special agents respectively, each agent has independent decision-making ability and negotiation mechanism, avoiding the bottleneck effect of traditional centralized decision-making, realizing the advantages of parallel processing and distributed computing. The production disturbance propagation matrix accurately calculates the influence propagation path and intensity through graph theory algorithm, providing scientific boundary constraints for agent negotiation, avoiding the waste of computing resources caused by blind global search. The collaborative density vector constructs trust score and interaction efficiency index based on historical collaboration data, making the negotiation process between agents more efficient and reliable, reducing invalid negotiation rounds. The dynamic weight self-adaptive optimization model adopts a deep reinforcement learning mechanism, which can adjust the weight coefficients of the multi-objective function in real time according to the current production environment state, avoiding the deviation of the optimization direction caused by static weight distribution, and improving the adaptability and optimization quality of the objective function. The convolution acceleration mechanism based on frequency domain transformation converts the traditional time domain convolution operation into point-by-point multiplication in the frequency domain, significantly reducing the computational complexity, so that the weight adjustment decision can be completed within milliseconds, meeting the rapid response demand of emergency order insertion. The multi-agent reinforcement learning collaborative optimization algorithm models different components of the network as independent agents, realizes parallel computing and strategy coordination of weight adjustment, avoids local optimal trap, and improves the robustness of the overall optimization strategy.

[0070] A specific embodiment 1 of the present application is provided below, and the specific implementation of each step in embodiment 1 is described in detail as follows.

[0071] The specific implementation of step S01 is to establish an agent collaborative framework, and the state vector of each agent is constructed by a multi-dimensional feature representation method. The order insertion agent state vector is represented as:

[0072] ;

[0073] In the formula, is the order insertion agent state vector, is the order insertion emergency degree, is the maximum order insertion emergency degree, and the default value is 5, The intensity of resource demand is obtained by calculating the weighted sum of the demand for various resources using a process route analysis algorithm. The maximum resource demand intensity was determined through statistical analysis of historical data. To assess time urgency, the reciprocal of the difference between the delivery date and the current time is used for calculation. To maximize time urgency, this was determined through analysis of historical order insertion data. The message passing protocol between agents employs a priority-based scheduling algorithm; the message priority calculation formula is as follows:

[0074] ;

[0075] In the formula, For message priority, and These are weighting coefficients, with empirical values ​​of 0.6 and 0.4, respectively.

[0076] The specific implementation of step S02 is the same as described above, and will not be repeated in detail here.

[0077] The specific implementation of step S03 is to construct a perturbation propagation matrix based on the production state dataset. The formula for calculating the matrix elements is as follows:

[0078] ;

[0079] In the formula, For the node To the node The perturbation propagation coefficient, The degree of resource sharing between nodes is obtained through resource correlation analysis. This represents the maximum resource sharing level, which defaults to 1. The process dependency weights are determined through process flow analysis. This represents the maximum process dependency weight, which defaults to 1. The distance between nodes is calculated using the shortest path algorithm in graph theory. This represents the maximum node distance in the network. The function for calculating the impact of the disturbance is as follows:

[0080] ;

[0081] In the formula, To determine the degree of disturbance impact, For inserting single-pair resources The demand, Resources for existing orders The amount of space occupied For resources Total capacity The number of resource types. The influence domain boundary parameters are determined using a clustering algorithm, and the boundary threshold function is:

[0082] ;

[0083] wherein, is the influence domain boundary threshold, is the cluster center mean, is the cluster standard deviation, is the adjustment coefficient, usually 1.5.

[0084] The specific implementation of step S04 is to start the multi-agent negotiation mechanism, and the collaborative density vector calculation formula is:

[0085] ;

[0086] wherein, is the collaborative density, is the agent interaction frequency with agent , is the maximum interaction frequency, obtained by historical data statistics, is the number of successful cooperation, is the maximum cooperation number between agent and agent , is the response delay, unit: ms, is the maximum allowed response delay, unit: ms, the experience value is 1000ms, is the total number of agents. The collaborative efficiency evaluation function is expressed as follows:

[0087] ;

[0088] wherein, is the collaborative efficiency score, is the decision consistency degree, is the maximum consistency value, default is 1, is the negotiation round number, is the maximum allowed round number, the experience value is 5, is the average information transmission delay, unit: ms, is the maximum allowed information transmission delay, unit: ms, default is 500ms.

[0089] The specific implementation of step S05 is to start the dynamic weight adaptive optimization model, which adopts a convolution acceleration mechanism based on frequency domain transformation. The input time domain production state signal is converted to the frequency domain space for processing through fast Fourier transform. In the specific implementation process, first, the multi-dimensional state data of the production environment is reorganized into a two-dimensional matrix form, and then the fast Fourier transform algorithm is used to convert the spatial convolution operation into point-by-point multiplication operation in the frequency domain, which significantly reduces the calculation complexity from to . To handle data boundary effects, the overlap preservation method is used to process the input data in blocks, with 50% overlap between each data block to ensure the continuity of the boundary information. The frequency domain filter realizes learnable feature extraction through parameterized design, and the filter parameters are dynamically adjusted through the back propagation algorithm, so that the model can adaptively extract key features in different frequency domain ranges. This model simultaneously uses a collaborative optimization algorithm based on multi-agent reinforcement learning to model different functional layers of the neural network as independent agents, including input layer agents, hidden layer agents, and output layer agents. Each agent implements collaborative optimization through a distributed policy learning mechanism. The input layer agent is responsible for feature preprocessing and screening, the hidden layer agent is responsible for feature transformation and abstraction, and the output layer agent is responsible for the final decision of the weight coefficients. The agents communicate with each other through a message passing interface to pass gradient information and intermediate calculation results, and use an asynchronous communication mode to avoid communication bottlenecks. The reward sharing strategy calculates the contribution of each agent through a global reward function, and allocates reward values based on the contribution to encourage each agent to converge to the global optimal solution. The input vector of this model is represented as:

[0090] ;

[0091] wherein, is the input vector, is the resource load rate, in %, is the maximum resource load rate, defaulting to 100%, is the due date pressure, obtained through due date urgency analysis and calculation, is the maximum due date pressure, determined through historical data analysis, is the equipment availability, obtained through equipment state monitoring, is the maximum equipment availability, defaulting to 1, is the material sufficiency, obtained through inventory management system, is the maximum material sufficiency, defaulting to 1, is the historical delay rate, calculated through historical order data statistics, is the maximum historical delay rate, usually taking a value of 1. The weight adjustment function is expressed as follows:

[0092] ;

[0093] ;

[0094] ;

[0095] wherein, is the single insertion response time weight, is the original order delay weight, is the resource utilization efficiency weight, is the average remaining time, in h, is the maximum remaining time, in h, obtained by historical order data statistics, is the resource conflict factor, with a value range of 0 to 1, obtained by resource competition analysis calculation. The comprehensive objective function of the multi-objective function set is represented as:

[0096] ;

[0097] wherein, is the comprehensive objective function value, is the single insertion response time function, is the order delay risk function, is the resource utilization efficiency function, is the decision variable vector.

[0098] The specific implementation of step S06 is to establish a delay risk assessment correlation matrix, and the matrix element calculation formula is:

[0099] ;

[0100] wherein, is the order delay probability caused by order delay, is the conditional probability of order delay under the delay condition of given order delay, is the joint probability of two orders being delayed at the same time, is the marginal probability of order delay. The delay risk assessment value is calculated by matrix operation:

[0101] ;

[0102] wherein, is the delay risk assessment value, is the current delay indicator variable of order , is the importance weight of order , is the total number of orders.

[0103] The specific implementation of step S07 is to generate a final production scheduling scheme. The scheme fusion algorithm adopts a constraint weighted summation method, and the fusion function is:

[0104]

[0105] In the formula, is the final production scheduling scheme, is the Pareto optimal solution, is the normalized coefficient of the Pareto solution, is the local adjustment scheme, is the normalized coefficient of the local solution, is the fusion weight coefficient, and the empirical value is 0.7. The actual execution deviation calculation formula is:

[0106]

[0107] In the formula, is the actual execution deviation, is the actual completion time of the order , in hours, is the planned completion time of the order , in hours.

[0108] The principle of the weight adjustment function is based on the multi-objective decision-making theory, and through dynamic balance of the three key factors of order insertion urgency, resource load state and time constraint, the adaptive adjustment of the weight coefficient is realized. The urgency term of the function is , the time constraint term is , and the resource conflict term is . The function adopts normalization processing to ensure that the sum of the three weights is 1, and the denominator part unifies the dimension, so that different types of production environment parameters can be effectively fused. Compared with the fixed weight method, the dynamic weight mechanism can automatically adjust the importance of the optimization target according to the real-time production state, significantly improving the response speed of the order insertion processing and the stability of the production plan.

[0109] The disturbance influence degree calculation function adopts the resource competition theory, and through the analysis of the competition relationship between the order insertion and the existing orders on various resources, the disturbance intensity of the order insertion operation on the entire production system is quantified. The resource competition intensity term of the function is , and the normalization term is . The numerator part reflects the interaction intensity of resource demand, and the denominator part is normalized to ensure the comparability of the disturbance influence degree in different scale production systems. The function can accurately identify the influence range of the order insertion operation, and provide a scientific basis for subsequent negotiation and optimization. Compared with the traditional empirical judgment method, the calculation accuracy is improved by 25%.

[0110] ​​The synergy efficiency evaluation function is based on network cooperation theory, and comprehensively considers three dimensions of decision consistency, negotiation rounds and information transmission delay to comprehensively evaluate the cooperation efficiency among intelligent agents. The cooperation quality item of the function is , and the cooperation cost item is . The numerator part reflects the cooperation quality, and the denominator part uses Euclidean distance to measure the cooperation cost, and the efficiency index in different cooperation scenarios is unified by normalization processing. The function can monitor the efficiency change of the agent cooperation process in real time, and provide quantitative feedback for the optimization adjustment of the cooperation mechanism. Compared with the single index evaluation method, the evaluation accuracy is improved by 30%.

[0111] In order to better understand and implement the present application, the following provides an embodiment 2 of a specific application scenario of the present application:

[0112] A certain technical team adopts an AI intelligent agent emergency order insertion dynamic decision-making scheduling method to solve the emergency order insertion scheduling problem of a car part manufacturing workshop. The workshop has 15 numerical control machining devices, produces key parts such as cylinder body, piston ring and connecting rod, and daily handles 85 standard orders, with an average of 3-5 emergency order insertion requests per day.

[0113] The technical team first builds an intelligent agent cooperation framework, and abstracts the workshop production system into four types of intelligent agents. The order insertion intelligent agent is responsible for receiving emergency order insertion requests, including a temporary additional cylinder body order from a certain automobile manufacturer, with a quantity of 500 and a requirement of completing delivery within 72h; the order intelligent agent manages the state information of the existing 85 standard orders, of which 32 orders are in the processing stage, with an average completion progress of 67%; the device intelligent agent monitors the running state of the 15 numerical control devices, with an average load rate of 78%, of which 5 devices are in a high load state (load rate > 90%); the material intelligent agent tracks the raw material inventory, with a cylinder blank inventory of 1200, and a predicted replenishment cycle of 48h. Each intelligent agent realizes message transmission through TCP / IP protocol, and establishes a distributed decision-making network architecture.

[0114] The technical team collects real-time data of the production environment to build a production state data set. The priority level of the emergency order insertion is set to A level (the highest level), the expected delivery period is 72h, and the resource demand includes device working hours 360h, cylinder blank 500, and cutting fluid 15L. Among the existing orders, there are 12 key path node orders, with an average remaining working period of 96h, and 3 orders have delay risk. The device running parameter shows that the load rate of device M01 reaches 95%, the fault warning information prompts that the bearing temperature is abnormal, and device M07 is scheduled to be maintained after 48h. The material inventory data shows that the cutting fluid inventory is 45L, the quality level is A, and the replenishment cycle is 24h.

[0115] Based on the production state dataset, the technical team calculates the state vector of each agent. The state vector of the insertion agent is [0.92, 0.78, 0.85], representing the urgency, resource matching degree, and execution feasibility, respectively. The average state vector of the order agent is [0.67, 0.83, 0.71], representing the completion progress, time margin, and resource sufficiency, respectively. The average state vector of the equipment agent is [0.78, 0.89, 0.76], representing the load rate, health degree, and availability, respectively. The state vector of the material agent is [0.73, 0.91, 0.82], representing the inventory sufficiency, quality compliance, and supply timeliness, respectively. Through the production disturbance propagation matrix analysis, the insertion operation will affect 23 existing orders, and the equipment M01, M03, and M05 are identified as resource constraint bottlenecks. The influence domain boundary parameter is set to 0.65.

[0116] The technical team starts the multi-agent negotiation mechanism. The insertion agent sends a bid request to 9 related equipment agents, and each equipment agent calculates a bid scheme based on the state vector. Equipment M02 provides an available time window of 14:00 on the 2nd day to 08:00 on the 4th day, with a processing offer of 45 yuan / h; equipment M08 provides a time window of 20:00 on the 1st day to 12:00 on the 3rd day, with an offer of 52 yuan / h. The material agent confirms the supply time commitment of the cylinder blank as immediate supply, with a cost offer of 135 yuan / piece. Through the distributed auction algorithm, the preliminary allocation scheme determines the use of equipment M02, M08, and M11 to complete the insertion processing.

[0117] The dynamic weight adaptive optimization model receives the current environmental state information, including the insertion urgency =0.92, resource load state =78%, and delivery pressure parameters. The model adopts a three-layer fully connected neural network structure with 256 neurons per layer, trained through the deep Q network algorithm. The input layer processes a 15-dimensional environmental state vector, the hidden layer uses the ReLU activation function for feature extraction, and the output layer generates three weight coefficients. After forward propagation calculation, the model outputs the insertion response time weight =0.54, the original order delay weight =0.31, and the resource utilization efficiency weight =0.15.

[0118] The technical team builds a multi-objective function set. The insertion response time minimization function takes the total completion time as the optimization target, and the current calculation value is 68.5h; the original order delay risk minimization function takes the weighted delay time as the target, and the current risk assessment value is 0.23; the resource utilization efficiency maximization function takes the comprehensive utilization rate of equipment as the target, and the current efficiency value is 0.82. By using dynamic weight coefficients, the multi-objective function is converted into a comprehensive objective function by weighted summation method. The improved non-dominated sorting genetic algorithm is used for solving, the population size is set to 200, the crossover probability is 0.8, the mutation probability is 0.1, and after 500 iterations, a Pareto optimal solution set containing 25 candidate schemes is generated.

[0119] As shown in Figure 2 The delay risk assessment correlation matrix quantifies the time dependence relationship between each order. The correlation strength of order O15 to order O23 is 0.78, and the correlation strength of order O08 to order O34 is 0.43. When it is detected that the delay risk assessment value of order O15 reaches 0.85, which exceeds the preset risk threshold 0.80, the system starts the incremental re-planning algorithm. The algorithm only optimizes the local adjustment of the affected 8 order subset, and the expected completion time of order O15 is delayed by 12h after adjustment, but the influence on other 77 orders is avoided. The local adjustment scheme reallocates the working time of equipment M03, and the 18h working time originally planned for order O15 is transferred to the insertion task.

[0120] As shown in Table 1, the system collects key parameter data to provide a basis for subsequent optimization:

[0121] Table 1 Key parameter data table of production system

[0122] The technical team combines the Pareto optimal solution set and the local adjustment scheme to generate the final scheduling scheme. The scheme determines that the insertion task is assigned to equipment M02 (160h), M08 (120h), and M11 (80h), with an estimated completion time of 69.2h, meeting the 72h delivery requirement. Among the original orders, only the scheduling time of 8 orders is slightly adjusted, with an average delay time of 4.8h, and the remaining 77 orders remain unchanged. The resource utilization efficiency is improved to 0.86, and the comprehensive objective function value reaches 0.91.

[0123] After the production scheduling scheme is issued to the execution layer, the real-time monitoring module starts to track the execution process. At 18h, it is detected that the actual processing progress of device M02 lags behind, and the actual execution deviation reaches 0.12, which exceeds the deviation tolerance threshold 0.10. The system immediately triggers the rapid rescheduling mechanism, and part of the working hours is transferred to device M06. The predicted completion time of the adjusted scheme is 70.8h, which still meets the delivery time requirement. During the entire insertion order processing process, the weight adjustment decision is completed within an average of 2.3ms, meeting the rapid response demand of emergency insertion order.

[0124] As shown in Figure 3 , the synergy density vector reflects the collaboration tightness between agents. The synergy strength of the insertion order agent and the device agents M02 and M08 is 0.87 and 0.79 respectively, and the synergy strength with the material agent is 0.92. The system realizes the collaborative optimization between components through the collaborative optimization algorithm of multi-agent reinforcement learning, avoiding the local optimal problem caused by single weight adjustment strategy. The disturbance influence degree calculation result shows that the overall disturbance intensity of the insertion operation on the production system is 0.34, which belongs to the medium disturbance level, and the influence range can be effectively controlled through reasonable resource allocation.

[0125] Compared with the traditional manual production scheduling method, the present application realizes distributed decision-making through the agent collaboration framework, avoiding the bottleneck problem and information lag of centralized decision-making. The dynamic weight self-adaptive optimization model can automatically adjust the optimization target weight according to the real-time production environment state, overcoming the limitation that the fixed weight method cannot adapt to the change of production environment. The weight adjustment mechanism based on reinforcement learning can make more accurate decisions when facing similar insertion orders through historical experience learning, avoiding the subjectivity of traditional methods relying on expert experience. The introduction of production disturbance propagation matrix enables the system to accurately identify the influence range of insertion operation, avoiding the resource waste of global rescheduling. The incremental rescheduling algorithm only optimizes the affected orders locally, maintaining the stability of the system and avoiding the production chaos caused by comprehensive adjustment. The mechanism combining multi-objective optimization and real-time monitoring ensures the insertion order processing while minimizing the impact on the original production plan, achieving the balance between production efficiency and order satisfaction rate.

[0126] It should be noted that the variables involved in the present application are explained in detail in Tables 2 and 3.

[0127] Table 2 Variable explanation table (first part)

[0128] Table 3 Variable explanation table (second part)

[0129] The above merely illustrates the specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of the changes or replacements within the technical range disclosed by the present application, which should be covered in the protection scope of the present application.

Claims

1. A dynamic decision-making and production scheduling method for emergency order insertion by an AI intelligent agent, characterized in that, Construct an intelligent agent collaboration framework, abstracting emergency order insertion requests, existing order status, equipment operating parameters, and material inventory information into order insertion intelligent agents, order intelligent agents, equipment intelligent agents, and material intelligent agents, and establish a distributed decision-making network architecture; Collect real-time production environment data to construct a production status dataset; calculate the state vector of each agent based on the production status dataset; analyze the impact range of the order insertion operation through the production disturbance propagation matrix; and determine the boundary parameters of the impact domain. A multi-agent negotiation mechanism is initiated, and each agent calculates a bidding plan based on its state vector and cooperative density vector. A preliminary resource allocation plan is then output through a distributed auction algorithm. The dynamic weighted adaptive optimization model is activated to calculate the target weight coefficients, transforming the multi-objective function set into a comprehensive objective function. An improved non-dominated sorting genetic algorithm is used to solve and output the Pareto optimal solution set. The time dependency between orders is quantified by the delay risk assessment correlation matrix. When the delay risk assessment value exceeds the risk threshold, an incremental replanning algorithm is activated to output a local adjustment plan. The Pareto optimal solution set and the local adjustment plan are merged to generate the final production scheduling plan. The real-time monitoring module is activated, and a fast rescheduling mechanism is triggered when the actual execution deviation exceeds the deviation tolerance threshold.

2. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 1, characterized in that, The real-time production environment data includes the priority level, expected delivery date, and resource requirements of emergency orders; the completion progress, remaining time, and critical path nodes of existing orders; the load status, fault warning information, and capacity parameters of equipment; and the inventory level, replenishment cycle, and quality grade of materials.

3. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 2, characterized in that, The agents exchange information through a message passing protocol; the set of multi-objective functions includes minimizing the order insertion response time, minimizing the risk of delay in existing orders, and maximizing resource utilization efficiency; the incremental replanning algorithm only performs local optimization and adjustment on a subset of affected orders, keeping the original production schedule of other orders unchanged.

4. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 3, characterized in that, The order insertion agent is responsible for receiving and processing emergency order insertion requests, analyzing the technical requirements and time constraints of the order insertion, calculating the resource requirements of the production system for the order insertion, initiating negotiation requests with other agents and collecting bidding information, and finally determining the production scheduling position and execution plan for the order insertion.

5. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 4, characterized in that, The order intelligence agent manages the status information of existing production orders, including the order's process flow, completion progress, remaining time, and quality requirements. It assesses the impact of order insertion operations on its own orders and provides feasible solutions and cost assessments for order adjustments during the negotiation process.

6. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 5, characterized in that, The device intelligence agent monitors the operating status and production capacity parameters of the equipment, including equipment load rate, fault warning information, maintenance plan, and processing capacity. Based on its own status and constraints, it provides available time windows and processing quotations to the order-placing intelligence agent and participates in the bidding process for resource allocation.

7. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 6, characterized in that, The material intelligence agent tracks the inventory level, quality grade, and replenishment cycle information of materials, assesses the availability and supply capacity of materials required for order insertion, and provides time commitments and cost quotations for material supply during the negotiation process to ensure material support for order insertion.

8. The AI ​​intelligent agent's emergency order insertion dynamic decision-making and production scheduling method according to claim 7, characterized in that, The steps for establishing the production disturbance propagation matrix are as follows: First, identify all resource nodes and order nodes in the production system and construct a relationship graph between resources and orders; second, establish connection weights between nodes based on resource sharing relationships and process flow dependencies; and then use graph theory algorithms to calculate the path and intensity of the impact propagation.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the AI ​​intelligent agent emergency order insertion dynamic decision-making and production scheduling method according to any one of claims 1-8.

10. An AI-powered emergency order insertion dynamic decision-making and production scheduling system, characterized in that, The system includes the computer-readable storage medium of claim 9, wherein the system is any one of a computer, a server, or a microcontroller, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

Images