An efficient model changing scheduling method based on AI multi-agent dynamic negotiation, medium and system

Abstract

The application provides a kind of efficient change type production scheduling method based on AI multi-agent dynamic negotiation, medium and system, belongs to agent technology field, the application is by constructing multiscale time sequence perception model fusion equipment historical data, real-time state data and environmental parameter data, utilize residual connection mechanism to process the time sequence correlation characteristics under nonlinear working condition to establish change type time prediction basis, establish dynamic hierarchical negotiation framework to realize information interaction and decision transmission, equipment agent is based on multiscale time sequence perception model prediction change type time and generates bid scheme, coordination agent adopts Pareto front multi-objective optimization algorithm to process bid scheme and balance multiple targets, through mixed similarity evaluation mechanism dynamic adjustment model parameter guarantees prediction stability, when production disturbance occurs, automatically identify affected order subset triggers incremental re-negotiation process, solves the technical problem that change type time prediction is not accurate and leads to production scheduling plan frequent adjustment.

Description

Technical Field

[0001] This invention belongs to the field of intelligent agent technology, and more specifically, relates to an efficient changeover scheduling method, medium, and system based on AI multi-agent dynamic negotiation. Background Technology

[0002] In the field of production scheduling in manufacturing, changeover scheduling refers to the process scheduling of a production line switching from the current product type to another. Traditional changeover scheduling methods mainly rely on empirical models and heuristic algorithms based on historical statistical data for scheduling decisions. These methods establish changeover time estimation models by analyzing historical equipment changeover records, and then combine order demand and equipment constraints to formulate a production plan. Such methods can provide basic scheduling support under single-equipment, stable operating conditions. However, traditional methods have significant technical limitations. On the one hand, changeover time prediction models based on historical statistics cannot accurately reflect the impact of real-time equipment status changes on the changeover process. When equipment operating status fluctuates or environmental parameters change, prediction accuracy drops sharply. On the other hand, traditional heuristic scheduling algorithms lack dynamic negotiation mechanisms, and are prone to getting stuck in local optima when facing complex scheduling scenarios with multiple equipment and multiple orders, failing to achieve global optimization. In the current intelligent manufacturing environment, due to the increasing complexity of production systems, frequent changes in equipment status, and intensified fluctuations in order demand, the prediction error problem of traditional changeover scheduling methods becomes more prominent. When the changeover time prediction deviation exceeds a certain range, the original production plan becomes invalid and needs to be readjusted. Frequent plan adjustments not only increase system operating costs but also affect overall production efficiency. Summary of the Invention

[0003] In view of this, the present invention provides an efficient changeover scheduling method, medium and system based on AI multi-agent dynamic negotiation, which can solve the technical problem in the prior art that inaccurate changeover time prediction leads to frequent adjustments to the production schedule.

[0004] This invention is implemented as follows: The first aspect of this invention provides an efficient changeover scheduling method based on AI multi-agent dynamic negotiation. This method uses a multi-scale time-series perception model to fuse historical equipment changeover data, real-time operating status data, and environmental parameter data to construct a prediction foundation. It establishes equipment agents, order agents, process agents, material agents, and coordination agents to achieve dynamic hierarchical negotiation and decision-making. A hybrid similarity evaluation mechanism is used to dynamically adjust model parameters to ensure prediction stability. The method includes the following steps: constructing a multi-scale time-series perception model; fusing historical equipment changeover data, real-time operating status data, and environmental parameter data to establish a basic dataset for changeover time prediction; establishing equipment agents, order agents, process agents, and coordination agents; and achieving information interaction and coordination among agents through a dynamic hierarchical negotiation architecture. Decision transmission: The order agent sends a changeover scheduling request to the coordinating agent. The coordinating agent selects a set of equipment agents that meet the requirements and broadcasts a scheduling task invitation. The equipment agents predict the changeover time based on a multi-scale time-series perception model, generate bidding schemes, calculate the mixed similarity between the predicted confidence feature vector and the baseline confidence vector, and adjust the model parameters according to the range of mixed similarity values. The coordinating agent uses a Pareto front multi-objective optimization algorithm to process the collected bidding schemes, balancing multiple objectives such as changeover time, order delivery date, and equipment load. The process agent and material agent confirm the changeover preparation, and the equipment agents execute the changeover operation and provide real-time feedback on the execution status. When a disturbance event occurs during production, the system automatically identifies the affected order subset and triggers an incremental renegotiation process only for this subset of orders.

[0005] Specifically, the step of constructing the multi-scale temporal perception model is to process the temporal correlation features under nonlinear conditions through the residual connection mechanism. The specific structure of the multi-scale temporal perception model is an encoder-decoder architecture. The encoder contains three layers of long short-term memory network units, with 128 neurons in each layer. The decoder uses a multi-head attention mechanism to fuse features at multiple time scales. The output layer generates transformation time prediction values ​​and confidence evaluation values ​​through a fully connected network.

[0006] Specifically, the steps of establishing equipment intelligent agents, order intelligent agents, process intelligent agents, material intelligent agents, and coordination intelligent agents involve each intelligent agent storing its corresponding state parameters, historical data, and constraints. The steps of establishing the training dataset for the multi-scale time-series perception model involve collecting the model change history of all equipment within the past 6 months, including data such as product models before and after the model change, model change duration, equipment state parameters, ambient temperature, and humidity, and dividing the training set and test set in a 7:3 ratio.

[0007] The hybrid similarity is a similarity evaluation index that comprehensively considers Euclidean distance and mapped clustering distance. The spatial distance between the predicted confidence feature vector and the baseline confidence vector is calculated using the Euclidean distance calculation function. The clustering difference between the two vectors in high-dimensional space is calculated using the mapped clustering distance calculation function. Finally, the two distance information are fused using the hybrid similarity calculation function to obtain a comprehensive similarity evaluation result.

[0008] Specifically, the step of generating a bidding proposal by the device agent involves generating a bidding proposal based on its own equipment status and load conditions. During the training of the multi-scale time-series perception model, the loss function adopts a weighted combination of mean square error and confidence regularization term, and an early stopping mechanism is used to prevent overfitting. The dynamic hierarchical negotiation architecture is an organizational structure that divides the agent system into three layers: a core coordination layer, a functional processing layer, and a data execution layer.

[0009] Specifically, the step of balancing multiple objectives by the coordinating agent is to balance multiple objectives such as changeover time, order delivery time, and equipment load through an adaptive weight adjustment mechanism. In the dynamic hierarchical negotiation architecture, the core coordination layer is responsible for global decision-making and conflict arbitration, the functional processing layer includes equipment agents, order agents, and process agents responsible for corresponding domain processing, and the data execution layer is responsible for real-time data acquisition and instruction execution. Specifically, the model parameters for adjusting the numerical range of the mixed similarity are maintained if the mixed similarity ∈ [0.85, 1].

[0010] Specifically, the process agent and material agent execute steps to coordinate the agent to monitor the entire changeover process and record the actual changeover time for multi-scale temporal perception model parameter optimization. The mixed similarity adjustment of model parameters also includes adjusting the attention weight convergence factor if the mixed similarity ∈ [0.65, 0.85), and re-initializing the hidden layer parameters of the model if the mixed similarity < 0.65. The Pareto front multi-objective optimization algorithm is a multi-objective optimization method based on evolutionary computation.

[0011] Specifically, the disturbance events are equipment failures or emergency order insertions. The Pareto front multi-objective optimization algorithm approximates the Pareto optimal front by maintaining a non-dominated solution set, avoiding the subjectivity problem of weight setting in the traditional weighted summation method. At the same time, it optimizes conflicting objectives such as changeover time, order on-time rate, and equipment utilization. The adaptive weight adjustment mechanism is an algorithm that dynamically adjusts the weights of each objective in the multi-objective optimization according to the real-time production status.

[0012] Specifically, the step of coordinating the intelligent agent to select a set of equipment intelligent agents that meet the requirements involves selecting and broadcasting production scheduling tasks to invite the set of equipment intelligent agents based on process constraints. The step of training the multi-scale time-series perception model involves performing sliding window slicing on the time-series data to form sample pairs, using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 64, and the number of training rounds to 200.

[0013] The adaptive weight adjustment mechanism is determined by the following conditions: when the degree of target conflict variation is <5%, the strategy of maintaining stable weight allocation is used to ensure system reliability; if the degree of target conflict variation is ≥5%, the multi-target decomposition adaptive adjustment mechanism is activated to cope with the changes.

[0014] The prediction confidence feature vector is a numerical vector representing the reliability of prediction output by the multi-scale time-series-aware model. It includes dimensions such as prediction variance, model gradient rate of change, and historical prediction bias, and is used to evaluate the credibility of the current prediction result.

[0015] The benchmark confidence vector is a reference standard vector established based on the historical best prediction state. It includes dimensions such as the corresponding prediction variance benchmark value, the model gradient rate of change benchmark value, and the historical prediction deviation benchmark value, and serves as a reference benchmark for evaluating the prediction confidence feature vector.

[0016] The attention weight convergence factor is used to control the convergence speed of the multi-head attention mechanism in the multi-scale temporal perception model. The inputs include prediction error, historical convergence trend and current learning rate, and the output is the adjusted attention weight update magnitude.

[0017] The degree of target conflict variation is an indicator that measures the fluctuation of each objective function value during the Pareto front multi-objective optimization algorithm. It is obtained by calculating the ratio of the standard deviation to the mean of the objective function values ​​within a continuous time window, and is used to determine whether the multi-objective decomposition adaptive adjustment mechanism needs to be activated.

[0018] The incremental renegotiation process is a local optimization method in which the system only re-negotiates the production schedule for the part of the orders affected by the production disturbance when a production disturbance occurs. The scope of affected orders is identified through impact propagation analysis, thus avoiding a full recalculation of the entire production schedule.

[0019] Among them, the multi-scale temporal perception model embeds the Pareto front optimization algorithm into the forward propagation process of the neural network, so that the parameter update of each layer depends not only on the gradient information but also on the global search result of the Pareto front multi-objective optimization algorithm. At the same time, it uses a computational efficiency optimization algorithm based on separable attention to decompose the multi-head attention mechanism into two independent processes: time-dimensional attention calculation and feature-dimensional attention calculation.

[0020] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described efficient changeover scheduling method based on AI multi-agent dynamic negotiation.

[0021] A third aspect of the present invention provides an efficient changeover scheduling system based on AI multi-agent dynamic negotiation, 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.

[0022] This invention employs a multi-scale time-series perception model that fuses historical equipment changeover data, real-time operating status data, and environmental parameter data. Through a residual connection mechanism, it processes the time-series correlation characteristics under nonlinear operating conditions, accurately capturing the impact of equipment status changes on changeover time and significantly improving changeover time prediction accuracy. This invention establishes a dynamic hierarchical negotiation architecture comprising equipment agents, order agents, process agents, and coordination agents. Each agent achieves collaborative optimization through information interaction and decision transmission. The coordination agent uses a Pareto front multi-objective optimization algorithm to balance multiple objectives such as changeover time, order delivery date, and equipment load, avoiding the local optimum problem caused by traditional single-objective optimization. Simultaneously, a hybrid similarity evaluation mechanism dynamically adjusts model parameters, ensuring the long-term stability of the prediction model under complex operating conditions. This invention achieves accurate changeover time prediction and dynamic optimization of production scheduling through a multi-agent negotiation mechanism, effectively reducing the frequency of production scheduling adjustments due to prediction errors and solving the technical problem of frequent production scheduling adjustments caused by inaccurate changeover time prediction in existing technologies. Attached Figure Description

[0023] Figure 1 This is a flowchart of the method of the present invention.

[0024] Figure 2 The image shows the convergence curve of the multi-objective optimization in the example.

[0025] Figure 3 This is a schematic diagram illustrating the working principle of the hybrid similarity evaluation mechanism in the embodiment. Detailed Implementation

[0026] 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.

[0027] like Figure 1The diagram shown is a flowchart of an efficient changeover scheduling method based on AI multi-agent dynamic negotiation, provided by the first aspect of this invention. This method includes the following steps:

[0028] S01. Construct a multi-scale time-series perception model, integrate historical equipment replacement data, real-time operating status data and environmental parameter data, process the time-series correlation features under nonlinear operating conditions through the residual connection mechanism, and establish a basic dataset for replacement time prediction.

[0029] S02. Establish equipment intelligent agents, order intelligent agents, process intelligent agents and coordination intelligent agents. Each intelligent agent stores the corresponding state parameters, historical data and constraints. Information interaction and decision transmission between intelligent agents are realized through dynamic hierarchical negotiation architecture.

[0030] S03. The order agent sends a changeover scheduling request to the coordinating agent. The coordinating agent selects a set of equipment agents that meet the requirements based on the process constraints and broadcasts the scheduling task invitation to the set of equipment agents.

[0031] S04. The equipment agent predicts the replacement time based on the multi-scale time-series perception model, generates a bidding plan by combining its own equipment status and load conditions, and calculates the mixed similarity between the predicted confidence feature vector and the baseline confidence vector. If the mixed similarity ∈ [0.85, 1], the current prediction parameters are maintained. If the mixed similarity ∈ [0.65, 0.85), the attention weight convergence factor is adjusted. If the mixed similarity < 0.65, the hidden layer parameters of the model are re-initialized.

[0032] S05. The coordinating agent uses the Pareto front multi-objective optimization algorithm to process the collected bidding schemes. It balances multiple objectives such as changeover time, order delivery time and equipment load through an adaptive weight adjustment mechanism. When the degree of objective conflict variation is <5%, the weight allocation stable operation strategy is maintained to ensure system reliability. If the degree of objective conflict variation is ≥5%, the multi-objective decomposition adaptive adjustment mechanism is activated to cope with the changes.

[0033] S06. The process intelligence agent and the material intelligence agent confirm the changeover preparation. The equipment intelligence agent executes the changeover operation and provides real-time feedback on the execution status. The coordination intelligence agent monitors the entire changeover process and records the actual changeover time for multi-scale time-series perception model parameter optimization.

[0034] S07. When equipment failure or emergency order insertion occurs during the production process, the system automatically identifies the affected order subset and triggers an incremental renegotiation process only for the aforementioned order subset, while the remaining orders maintain their original production schedule.

[0035] The specific structure of the multi-scale temporal perception model is an encoder-decoder architecture. The encoder contains three layers of long short-term memory network units, with 128 neurons in each layer. The decoder uses a multi-head attention mechanism to fuse features from multiple time scales. The output layer generates transformation time prediction values ​​and confidence evaluation values ​​through a fully connected network.

[0036] The steps for establishing the training dataset of the multi-scale time-series perception model specifically include collecting the replacement history of all devices in the past 6 months, including data such as product models before and after replacement, replacement duration, device status parameters, ambient temperature and humidity, etc., dividing the training set and test set in a 7:3 ratio, and performing sliding window slicing on the time-series data to form sample pairs.

[0037] The training steps of the multi-scale time-series perception model specifically include using the Adam optimizer, setting the initial learning rate to 0.001, the batch size to 64, the number of training epochs to 200, using a weighted combination of mean squared error and confidence regularization term for the loss function, preventing overfitting through an early stopping mechanism, and finally controlling the prediction error of the multi-scale time-series perception model on the test set to within 5%.

[0038] The dynamic hierarchical negotiation architecture is an organizational structure that divides the intelligent agent system into three layers: the core coordination layer, the functional processing layer, and the data execution layer. The core coordination layer is responsible for global decision-making and conflict arbitration. The functional processing layer includes equipment intelligent agents, order intelligent agents, and process intelligent agents responsible for professional domain processing. The data execution layer is responsible for real-time data acquisition and instruction execution.

[0039] The adaptive weight adjustment mechanism is an algorithm that dynamically adjusts the weights of each objective in multi-objective optimization based on the real-time production status. When an urgent order is detected, the weight of the delivery order is automatically increased. When the equipment load is too high, the weight of the equipment utilization rate is increased. The weight adjustment strategy is continuously optimized through reinforcement learning.

[0040] The prediction confidence feature vector is a numerical vector representing the reliability of prediction output by the multi-scale time-series-aware model. It includes dimensions such as prediction variance, model gradient rate of change, and historical prediction bias, and is used to evaluate the credibility of the current prediction result.

[0041] The baseline confidence vector is a reference standard vector established based on the historical best prediction state. It includes dimensions such as the corresponding prediction variance baseline value, the model gradient rate of change baseline value, and the historical prediction deviation baseline value, and serves as a reference benchmark for evaluating the prediction confidence feature vector.

[0042] The hybrid similarity is a similarity evaluation index that comprehensively considers Euclidean distance and mapped clustering distance. The spatial distance between the predicted confidence feature vector and the baseline confidence vector is calculated using the Euclidean distance calculation function, and the clustering difference between the two vectors in high-dimensional space is calculated using the mapped clustering distance calculation function. Finally, the two distance information are fused using the hybrid similarity calculation function to obtain a comprehensive similarity evaluation result.

[0043] The Euclidean distance calculation function is used to calculate the spatial distance between two feature vectors. The input includes the predicted confidence feature vector and the baseline confidence vector, and the output is the standardized spatial distance between the two vectors. The formula is expressed as follows: ,in To predict the i-th component of the confidence feature vector, For the i-th component of the baseline confidence vector, and These are the maximum values ​​of the corresponding vectors.

[0044] The mapping clustering distance calculation function is used to calculate the clustering difference between two feature vectors in a high-dimensional mapping space. The input includes the predicted confidence feature vector and the baseline confidence vector, and the output is the numerical mapping distance between the two vectors in the clustering space. The formula is expressed as follows: ,in To predict the mapping value of the confidence feature vector at the j-th cluster center, The mapping value of the baseline confidence vector at the j-th cluster center. Let the weight of the j-th cluster be , Let be the standard deviation of the j-th cluster.

[0045] The hybrid similarity calculation function is used to fuse spatial distance and clustering distance information. The inputs include the standardized spatial distance value output by the Euclidean distance calculation function and the mapped distance value output by the mapped clustering distance calculation function. The output is the comprehensive similarity evaluation result. The formula is expressed as follows: ,in As a weighting balancing factor, This is the normalization coefficient for the maximum spatial distance.

[0046] The attention weight convergence factor is used to control the convergence speed of the multi-head attention mechanism in the multi-scale temporal perception model. The inputs include prediction error, historical convergence trend and current learning rate, and the output is the adjusted attention weight update magnitude.

[0047] The degree of target conflict variation is an indicator that measures the fluctuation of each objective function value during the Pareto front multi-objective optimization algorithm. It is obtained by calculating the ratio of the standard deviation to the mean of the objective function values ​​within a continuous time window, and is used to determine whether the multi-objective decomposition adaptive adjustment mechanism needs to be activated.

[0048] The incremental renegotiation process is a local optimization method in which the system only re-negotiates the production schedule for the part of the orders affected by the disturbance when a production disturbance occurs. The scope of affected orders is identified through impact propagation analysis, avoiding a full recalculation of the entire production schedule.

[0049] Furthermore, the multi-scale temporal perception model embeds the Pareto front optimization algorithm into the forward propagation process of the neural network, ensuring that the parameter updates of each layer depend not only on gradient information but also on the global search result of the Pareto front multi-objective optimization algorithm. Additionally, the multi-scale temporal perception model utilizes a computationally efficient optimization algorithm based on separable attention, decomposing the multi-head attention mechanism into two independent processes: temporal dimension attention computation and feature dimension attention computation. This significantly reduces computational complexity and memory consumption while maintaining attention modeling capabilities.

[0050] The Pareto embedding optimization function is used to integrate the results of the Pareto front multi-objective optimization algorithm into the neural network parameter update process. The inputs include the current network gradient, the Pareto solution set, and the target weight vector. The output is a parameter update direction vector that integrates multi-objective information. The formula is expressed as follows: ,in For the current network gradient, Let be the weight of the i-th objective. Let be the gradient of the i-th objective function. is the gradient fusion coefficient.

[0051] The separable attention calculation function is used to reduce the computational cost of multi-head attention mechanisms. The input includes a time-dimensional feature matrix and a feature-dimensional feature matrix, and the output is the decomposed attention weight matrix. The formula is expressed as follows: ,in Attention weights are based on the time dimension. Attention weights for feature dimensions This is the Kronecker product operation.

[0052] By introducing a hybrid similarity evaluation mechanism, the entire changeover scheduling scheme achieves significant improvements in prediction accuracy, system stability, and dynamic response capability. Hybrid similarity not only considers the direct distance between the predicted confidence feature vector and the baseline confidence vector in Euclidean space, but also deeply integrates clustering structure information in the high-dimensional mapping space, enabling the system to simultaneously evaluate the reliability of prediction results from multiple dimensions. Euclidean distance reflects the linear spatial relationship between vectors, quickly identifying the degree of deviation in predicted values, while the mapped clustering distance captures deep feature patterns of vectors in complex nonlinear data distributions through cluster affiliation analysis, revealing the relative position of the predicted state in the historical data distribution. The organic fusion of these two distance information allows hybrid similarity to more accurately identify abnormal prediction states, especially in complex production environments with nonlinear operating condition changes and multi-device coupling interference. When the predicted changeover state of a device deviates from its historical optimal operating mode, hybrid similarity can sensitively detect the dual changes in spatial distance and cluster affiliation, promptly triggering the parameter adaptive adjustment mechanism of the multi-scale time-series sensing model to prevent continuous deterioration of prediction accuracy. The mechanism described above enhances the anti-interference capability of the entire agent negotiation system. When faced with long-term, gradual changes such as equipment aging, process parameter drift, or environmental condition variations, hybrid similarity can continuously monitor the health of the prediction model, ensuring the long-term accuracy of changeover time predictions. Simultaneously, the multi-dimensional evaluation characteristics of hybrid similarity can effectively distinguish between temporary disturbances and systematic shifts, avoiding system oscillations caused by frequent adjustments to model parameters due to occasional anomalies, thus guaranteeing the stability and continuity of the entire dynamic negotiation process.

[0053] By embedding the Pareto front multi-objective optimization algorithm into the neural network training process, the multi-scale time-series sensing model achieves synergistic optimization of prediction accuracy and multi-objective balance, fundamentally solving the local optimum trap problem caused by traditional single-objective optimization. The Pareto embedding optimization function calculates the fusion vector of the current gradient and multi-objective gradients in each training batch of the neural network, ensuring that model parameter updates not only pursue minimizing prediction error but also simultaneously consider the balanced optimization of multiple production objectives such as changeover time, equipment utilization, and energy costs. This mechanism, through dynamic adjustment of the gradient fusion coefficient β, automatically balances the weight allocation of single-objective and multi-objective learning according to the current production state. When the production environment is relatively stable, a larger β value emphasizes improved prediction accuracy; when facing intensified multi-objective conflicts, a smaller β value strengthens multi-objective coordination capabilities. The Pareto embedding mechanism also establishes mutual constraints between objective functions, preventing over-optimization of a single objective at the expense of overall performance and ensuring the generalization ability and robustness of the multi-scale time-series sensing model under complex multi-constraint conditions. By maintaining the dynamic update of the Pareto solution set, the embedding mechanism can continuously track the frontier boundary of multi-objective optimization, providing diverse optimization direction guidance for neural networks and enhancing the model's adaptability in different production scenarios. The separable attention mechanism, through independent decomposition strategies in the time and feature dimensions, completely reconstructs the computational architecture of traditional multi-head attention, achieving a dual breakthrough in computational efficiency and modeling capability. The time-dimensional attention computation focuses on capturing the temporal dependencies during the transformation process, while the feature-dimensional attention computation deeply mines the correlation patterns between device state parameters. The independent processing of these two dimensions avoids redundant computations and cross-interference in traditional joint computation. The separable attention computation function reconstructs two low-dimensional attention matrices into a complete attention weight matrix through Kronecker product operations, reducing computational complexity from low to high while maintaining the original modeling accuracy. Reduce to This provides a feasible technical path for processing long-term series and high-dimensional feature data. The decomposition strategy is particularly suitable for large-scale time-series data processing in changeover and production scheduling scenarios, effectively alleviating the computational bottlenecks and memory limitations of traditional multi-head attention mechanisms when processing hundreds of time steps and dozens of device feature dimensions. The parallel computing characteristics of the separable attention mechanism are highly compatible with modern GPU architectures. Independent computation of the time and feature dimensions can fully utilize parallel processing units, significantly improving the training and inference efficiency of the multi-scale time-series awareness model. The deep integration of the two mechanisms enables the multi-scale time-series awareness model to have excellent computational performance while ensuring prediction accuracy, providing a solid technical foundation for the stable operation of the real-time dynamic negotiation system under strict latency constraints, and ensuring that the entire agent negotiation network can complete changeover decisions and production scheduling adjustments within millisecond response time.

[0054] The specific implementation methods of the above steps are described in detail below.

[0055] The specific implementation of step S01 involves constructing a multi-scale time-series sensing model to fuse multi-source data and establish a foundation for changeover time prediction. First, historical changeover data is collected, including changeover duration, frequency, and success rate, covering a nearly six-month production cycle to ensure the integrity of the seasonal and periodic characteristics of the data samples. Then, real-time operating status data is collected, encompassing key operating parameters such as equipment temperature, vibration frequency, current, voltage, and rotational speed, with a sampling frequency set to 1Hz to capture subtle changes in equipment status. Simultaneously, environmental parameter data, including workshop temperature, humidity, and air pressure, is acquired, with environmental data sampling intervals of 10 minutes to reflect slowly changing external influences. A residual connection mechanism is used to process the time-series correlation features under nonlinear operating conditions. Residual connections directly transmit feature information from different time levels to subsequent levels through skip connections, avoiding the gradient vanishing problem in deep networks, making it particularly suitable for handling complex nonlinear relationships during changeover processes. In the data preprocessing stage, a sliding window technique is used to slice the time-series data. The window size is set to 72 hours, and the sliding step size is 1 hour, forming overlapping time-series sample pairs to enhance the model's time-series learning ability. Outlier detection adopts a statistical distribution-based method, marking data points exceeding the mean plus or minus three standard deviations as outliers and smoothing them. Data normalization uses the min-maximum standardization method, mapping all feature values ​​to the interval between 0 and 1, eliminating the interference of numerical differences between features of different dimensions on model training.

[0056] The specific implementation of step S02 involves establishing multiple types of intelligent agents and constructing a dynamic hierarchical negotiation architecture to achieve information interaction between intelligent agents. The establishment process of the equipment intelligent agent includes creating an independent intelligent agent instance for each production device, storing key data such as the device's technical parameters, current status information, historical changeover records, and maintenance schedules. The intelligent agent integrates a decision engine to analyze its own changeover capabilities and load conditions. The construction of the order intelligent agent involves assigning an intelligent agent agent to each production order, recording constraints such as product specifications, quantity requirements, delivery deadlines, and priority levels. The intelligent agent possesses reasoning capabilities for demand analysis and resource matching. The design of the process intelligent agent focuses on storing and managing process flow information, including a knowledge base of changeover process steps, process parameter settings, quality inspection standards, and safe operating procedures, providing process feasibility analysis and constraint verification functions. The coordination intelligent agent, as the core of the system, assumes the responsibility of global coordination and decision arbitration, maintaining system-level data such as overall production scheduling status, resource allocation information, and conflict resolution strategies. The establishment process of the material intelligence agent involves creating an independent intelligence agent instance for each type of material resource, storing key data such as material specifications, inventory quantity, storage location, supplier information, quality certificates, expiration dates, and safety stock thresholds. The intelligence agent integrates a material demand analysis engine to assess material supply capacity and delivery timeliness. The material intelligence agent is responsible for monitoring material inventory status, predicting material consumption trends, coordinating material procurement plans, and verifying material quality compliance. It possesses reasoning capabilities for material availability judgment and supply matching. In the dynamic hierarchical negotiation architecture, the material intelligence agent, as an important component of the functional processing layer, collaborates with the equipment intelligence agent, order intelligence agent, and process intelligence agent. Through a message passing mechanism, it shares key data such as material preparation status, inventory changes, and supply constraints in real time, ensuring that changeover scheduling decisions fully consider the feasibility and timeliness of material supply, avoiding production delays and resource waste caused by material shortages, and providing material resource guarantees and constraint verification support for the entire intelligence agent negotiation network. The dynamic hierarchical negotiation architecture adopts a three-tier organizational structure. The core coordination layer is responsible for high-level management functions such as global decision-making, conflict arbitration, and resource optimization and allocation. The functional processing layer contains various intelligent agents responsible for domain-related analysis and processing tasks. The data execution layer is responsible for low-level operations such as real-time data acquisition, command issuance and execution, and status feedback reporting. Communication between intelligent agents adopts a message passing mechanism, defining standardized message formats and communication protocols to ensure the consistency and reliability of information exchange between different types of intelligent agents.

[0057] The specific implementation of step S03 involves sending production scheduling requests from the order agent and broadcasting tasks to the coordinating agent. When the order agent sends a changeover production scheduling request to the coordinating agent, it first provides a structured description of the order requirements, including key information such as the target product model, product specifications, production quantity, quality requirements, and delivery time window. Simultaneously, it calculates the order's urgency score and priority weight. Upon receiving the scheduling request, the coordinating agent immediately initiates a process constraint screening procedure. Based on product process requirements and equipment technical capabilities, it performs a matching analysis, excluding equipment lacking the corresponding changeover capabilities. Screening conditions include multi-dimensional constraints such as equipment precision level, processing capacity range, process compatibility, and equipment health status. The screening algorithm employs a rule-based reasoning expert system method, performing one-by-one matching verification through a pre-set process compatibility rule base. This rule base covers constraints such as material compatibility, size adaptability, and process path feasibility. After determining the set of qualified equipment agents, the coordinating agent constructs a standardized production task invitation message. The message content includes detailed descriptions such as basic order information, process requirements, time constraints, quality standards, and expected response time. Task broadcasting uses multicast communication to simultaneously send invitation messages to all members of the device agent set, ensuring the synchronicity and fairness of information transmission. A timeout mechanism is set for broadcast messages, with a timeout period of 60 seconds. Device agents that do not respond within the timeout period are considered currently unavailable and removed from the candidate set.

[0058] The specific implementation of step S04 involves the equipment agent generating a bidding proposal and performing prediction confidence assessment and parameter adaptive adjustment. After receiving a production scheduling task invitation, the equipment agent first calls a multi-scale time-series perception model to predict the changeover time. It inputs multi-dimensional information such as current equipment status parameters, historical changeover data, and environmental conditions. The model output includes the predicted changeover time and the corresponding confidence assessment value. The bidding proposal generation process comprehensively considers factors such as the current equipment load, predicted changeover time, equipment maintenance plan, and energy consumption costs to calculate the bid price and the promised completion time. Equipment load assessment uses a weighted average method, comprehensively calculating indicators such as the current number of tasks, equipment utilization rate, and remaining capacity. The load threshold is set at 85%. Equipment exceeding the threshold will adjust its bidding strategy to reduce its probability of winning the bid. The construction of the prediction confidence feature vector includes multiple dimensions such as prediction variance, model gradient rate of change, and historical prediction bias. Prediction variance reflects the uncertainty of the model output, gradient rate of change reflects the stability of the model parameters, and historical prediction bias measures the long-term accuracy performance of the model. The baseline confidence vector is established based on the historical best prediction state. By analyzing the confidence feature distribution during the period of highest prediction accuracy in the past, the corresponding variance baseline value, gradient baseline value, and bias baseline value are extracted as reference standards. The mixture similarity calculation first calculates the spatial distance between the predicted confidence feature vector and the baseline confidence vector using the Euclidean distance function, and uses standardization to eliminate the influence of dimensional differences. Then, it analyzes the clustering differences between the two vectors in high-dimensional space using a mapping clustering distance function. Finally, it uses a mixture similarity calculation function to weightedly fuse the spatial distance and clustering distance information. Different parameter adjustment strategies are implemented according to the mixture similarity value. When the mixture similarity is in the range of 0.85 to 1, it indicates a good prediction state, and the current parameter settings are maintained. When the mixture similarity is in the range of 0.65 to 0.85, the attention weight convergence factor adjustment mechanism is activated to optimize the model convergence performance. When the mixture similarity is below 0.65, the hidden layer parameters of the model are reinitialized to restore prediction capability.

[0059] The specific implementation of step S05 involves the coordinating agent using a Pareto front multi-objective optimization algorithm to process bidding proposals and achieve adaptive weight adjustment. After collecting bidding proposals submitted by all device agents, the coordinating agent transforms the bidding information into input parameters for a multi-objective optimization problem. The objective functions include conflicting optimization objectives such as minimizing changeover time, maximizing order on-time performance, balancing equipment utilization, and minimizing energy costs. The Pareto front multi-objective optimization algorithm finds the optimal balance point among multiple objectives by maintaining a set of non-dominated solutions. Each solution in the non-dominated solution set is superior to other solutions on at least one objective and not inferior to other solutions on other objectives. The algorithm employs an evolutionary computation framework, continuously updating the candidate solution population through genetic operations such as selection, crossover, and mutation. The population size is set to 100 individuals, and the number of generations is set to 50 to ensure convergence quality. Fitness evaluation uses a comprehensive ranking method combining crowding distance and dominance relation. Crowding distance is used to maintain solution diversity, while dominance relation is used to screen for high-quality solutions. The combination of both ensures the integrity and uniformity of the Pareto front. The adaptive weight adjustment mechanism dynamically modifies the weight allocation of each objective function based on real-time production status. When an urgent order is detected, the delivery weight is automatically increased to 0.6 to ensure priority processing of urgent orders. When equipment load is too high, the equipment utilization weight is increased to 0.5 to achieve load balancing. The degree of objective conflict variation is quantitatively evaluated by calculating the ratio of the standard deviation to the mean of each objective function value over five consecutive time windows. When the variation is less than 5%, the current weight allocation strategy is maintained to ensure stable system operation. When the variation is greater than or equal to 5%, the multi-objective decomposition adaptive adjustment mechanism is activated to cope with environmental changes. The multi-objective decomposition mechanism decomposes the complex multi-objective problem into multiple single-objective sub-problems that are solved in parallel. Each sub-problem focuses on a specific combination of objectives, and global optimization is achieved through information sharing between sub-problems.

[0060] The specific implementation of step S06 involves the process agent confirming the changeover preparation and monitoring the entire process of the equipment agent executing the changeover operation. Upon receiving the production scheduling plan, the process agent immediately initiates the changeover preparation confirmation procedure, verifying the correctness of process parameter settings, checking the feasibility of the process path, and verifying the completeness of quality inspection standards. Confirmation content includes multiple aspects such as the logical sequence of changeover process steps, the parameter range of key control points, and the execution requirements of safety operating procedures. Simultaneously, the material agent confirms the material preparation status, checking the adequacy of raw material inventory, verifying material specification matching, and ensuring material quality compliance. Material preparation confirmation involves detailed processes such as raw material batch traceability, quality certificate verification, and storage condition checks. After receiving the process confirmation and material confirmation signals, the equipment agent begins executing the changeover operation. The changeover process proceeds step-by-step according to the predetermined process flow, including standardized steps such as equipment shutdown, tooling replacement, parameter adjustment, trial operation, and quality inspection. A real-time feedback mechanism continuously monitors the changeover progress through a sensor network and data acquisition system. Key monitoring parameters include the completion time of each changeover stage, the stability of equipment operating parameters, and the compliance status of product quality indicators. The feedback frequency is set to update every 30 seconds to ensure real-time information. A coordinating agent monitors the entire changeover process, identifying potential delay risks by comparing actual progress with expected time points. When the actual progress lags behind the expected progress by more than 10%, an early warning mechanism is automatically triggered, and an emergency plan is activated. Actual changeover time records include detailed time information such as the changeover start timestamp, the completion timestamps of each stage, and the changeover end timestamp. This recorded data is automatically stored in a historical database for parameter optimization of the multi-scale time-series perception model. Model parameter optimization employs an online learning method, fusing new changeover time data with historical training data. An incremental learning algorithm updates the model weights, with a learning rate set to 0.001 to ensure the stability of parameter updates.

[0061] The specific implementation of step S07 involves the system automatically identifying disturbance events and performing incremental renegotiation processing on the affected order subset. The disturbance event detection mechanism continuously monitors the operating status of the production system, automatically detecting various disturbance events such as equipment failures, emergency order insertions, raw material shortages, and quality anomalies through an anomaly pattern recognition algorithm. The detection algorithm sets multi-level warning thresholds based on statistical process control principles. The first-level warning threshold is set at twice the standard deviation from the normal range, and the second-level warning threshold is set at three times the standard deviation from the normal range. The identification of the affected order subset employs an impact propagation analysis method, performing forward and backward impact analysis along the production process flow and resource dependencies from the source of the disturbance. Forward analysis identifies the scope of the disturbance's impact on subsequent production stages, while backward analysis determines the degree of impact of the disturbance on preparatory work. The quantification of impact degree uses an impact intensity scoring mechanism, calculating the impact intensity value based on factors such as the severity, duration, and scope of the disturbance event. Orders with an impact intensity exceeding 0.3 are included in the affected order subset. The incremental renegotiation process only re-negotiates production schedules for a subset of affected orders, keeping the original production schedules of the remaining unaffected orders unchanged. This avoids system turbulence and wasted computing resources caused by full recalculation. The renegotiation algorithm employs a local optimization strategy, reallocating resources and adjusting schedules within the affected order subset while considering interface constraints with unaffected orders to ensure consistency between local adjustments and the global plan. Renegotiation priority is set according to a dual standard of urgency and importance. Urgency is calculated based on order delivery deadlines and current progress, while importance is determined based on order value and customer level, with a weighting ratio of 6:4. Negotiation convergence is determined using a solution stability assessment method. When the change in the production schedule after three consecutive rounds of negotiation is less than 5%, the negotiation is considered converged, and the final adjustment plan is output and implemented.

[0062] Further explanation is needed regarding the detailed structure of the multi-scale temporal perception model, which employs an encoder-decoder architecture. The encoder consists of three layers of long short-term memory (LSM) network units, each containing 128 neurons to capture long-term dependencies in the temporal data. The first LSM network unit receives the raw temporal input, including multi-dimensional features such as device state parameters, environmental conditions, and historical data. It selectively remembers and forgets historical information through coordinated control of forget gates, input gates, and output gates. The second LSM network unit receives the hidden state output from the first layer and further extracts an abstract representation of the temporal features. The neuron activation function uses a hyperbolic tangent function to ensure that the output value is within the range of -1 to +1. The third LSM network unit generates the final encoded feature vector, which contains complete semantic information of the input temporal sequence. The decoder employs a multi-head attention mechanism to fuse features across multiple time scales. Eight attention heads are used, each focusing on different combinations of time scales and feature dimensions. Multi-head attention computation achieves weighted feature fusion through matrix operations on query vectors, key vectors, and value vectors. The attention weights reflect the importance of different time steps and feature dimensions to the current prediction task. The residual connection mechanism establishes a skip connection between the encoder and decoder, directly passing the encoder's raw input to the decoder's output layer, mitigating the gradient vanishing problem in deep networks and preserving low-level feature information. The output layer maps the attention mechanism's output to the final prediction result through a fully connected network, including two output nodes: the predicted replacement time and the confidence assessment value. The fully connected layer uses a linear activation function to ensure the continuity of the output values. The detailed steps for building the training dataset first involve collecting replacement history records for all equipment over the past six months. The data collection scope covers basic information such as the product model before replacement, the product model after replacement, the start time of replacement, the end time of replacement, and the duration of replacement. Equipment status parameters are collected, including dynamic parameters such as equipment temperature curves, vibration frequency changes, current and voltage fluctuations, and speed adjustment processes during the replacement process. The sampling frequency is set to 1Hz to capture subtle changes in equipment status. Environmental parameter data covers external environmental factors such as workshop temperature, relative humidity, atmospheric pressure, and air quality. The correlation between environmental data and replacement time is verified through correlation analysis, with a correlation coefficient threshold set to 0.3 to filter significantly correlated environmental factors. In the data preprocessing stage, the raw data is cleaned and standardized. Outlier detection uses an interquartile range (ICM)-based method to identify outliers, and missing value imputation uses temporal interpolation to maintain data continuity. The dataset is divided into training and test sets in a 7:3 ratio. The training set is used for model parameter learning, and the test set is used for model performance evaluation. The partitioning process ensures the uniform temporal distribution of the training and test sets.The sliding window slicing process for time series data divides continuous time series data into sample pairs of fixed length. The window size is set to 72 hours to cover the complete transformation cycle, and the sliding step size is set to 1 hour to form overlapping samples and enhance the generalization ability of the model.

[0063] Furthermore, this invention addresses the technical problem of biased optimization results caused by the strong subjectivity of objective weight settings in multi-objective optimization processes. Traditional production scheduling optimization methods typically use weighted summation to handle multiple conflicting objectives such as changeover time, order delivery time, and equipment utilization. The setting of weight parameters often relies on expert experience or experimental debugging, exhibiting significant subjectivity and limitations. When the production environment changes or priorities are adjusted, fixed weight settings are difficult to adapt to dynamically changing optimization needs, easily leading to optimization results biased towards a single objective at the expense of overall performance. This invention uses a Pareto front multi-objective optimization algorithm to maintain a non-dominated solution set that approximates the Pareto optimal front, avoiding the subjectivity problem of weight settings in traditional weighted summation methods. Combined with an adaptive weight adjustment mechanism, it dynamically adjusts the weights of each objective based on real-time production status. When an urgent order is detected, the delivery time weight is automatically increased; when equipment load is too high, the equipment utilization weight is increased. Through reinforcement learning methods, the weight adjustment strategy is continuously optimized, achieving objectivity and adaptability in the multi-objective optimization process.

[0064] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described efficient changeover scheduling method based on AI multi-agent dynamic negotiation.

[0065] A third aspect of the present invention provides an efficient changeover scheduling system based on AI multi-agent dynamic negotiation, 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.

[0066] Specifically, the principle of this invention is as follows: The reason why this invention can solve the core technical problem of inaccurate changeover time prediction lies in the establishment of a comprehensive technical architecture that integrates multi-dimensional data perception, multi-agent negotiation and decision-making, and adaptive parameter adjustment. The multi-scale temporal perception model integrates historical changeover data, real-time operating status data, and environmental parameter data through an encoder-decoder architecture. It uses long short-term memory network units to capture temporal dependencies and fuses feature information from different time scales through a multi-head attention mechanism. Compared with traditional prediction methods based on historical statistics, this model can dynamically perceive changes in equipment status and adjust prediction parameters in real time, significantly improving the accuracy and adaptability of changeover time prediction. The multi-agent dynamic negotiation mechanism establishes a well-defined and collaborative intelligent decision-making system. The equipment agent is responsible for generating bidding proposals based on its own status, the order agent proposes production scheduling requirements, the process agent ensures that process constraints are met, and the coordinating agent coordinates overall optimization decisions. This negotiation architecture can fully consider various constraints and target requirements, avoiding the information asymmetry and decision lag problems of traditional centralized production scheduling methods. The hybrid similarity evaluation mechanism, by fusing Euclidean distance and mapped clustering distance, can assess the reliability of prediction results from two dimensions: linear spatial relationship and nonlinear clustering structure. When a deviation from the normal range is detected, a parameter adjustment mechanism is triggered promptly, ensuring the model maintains long-term stability of prediction accuracy in complex and ever-changing production environments. The Pareto front multi-objective optimization algorithm is embedded in the neural network training process, enabling model parameter updates to not only minimize prediction error but also balance the optimization of multiple production objectives such as changeover time and equipment utilization. By maintaining a non-dominated solution set that approximates the Pareto optimal front, it avoids the overall performance imbalance that may result from single-objective optimization.

[0067] The following provides a specific embodiment 1 of the present invention, and the specific implementation of each step in this embodiment 1 is described in detail below.

[0068] The specific implementation of step S01 involves constructing a multi-scale time-series sensing model and establishing a basic dataset for transformation time prediction through data fusion. During data preprocessing, a sliding window normalization method is used to normalize the time-series data, as expressed in the following formula:

[0069] ;

[0070] In the formula, For a moment Standardized values; For a moment The original data values, with units determined based on specific parameters; The mean of the data within the sliding window is obtained by averaging all data points within the window; The standard deviation of the data within the sliding window is obtained by taking the square root of the average of the sum of squared deviations of the data points within the window. The sliding window size is set to 72 hours, and outlier detection within the window uses a method based on interquartile range. The formula for the outlier threshold is as follows:

[0071] ;

[0072] In the formula, The threshold for identifying outliers; The third quartile is obtained by sorting the data in ascending order and taking the value at the 75th percentile. The first quartile is obtained by sorting the data from smallest to largest and taking the value at the 25th percentile.

[0073] The specific implementation methods for steps S02-S03 are the same as those described above, and will not be repeated in detail here.

[0074] The specific implementation of step S04 involves the equipment agent generating a bidding plan based on a multi-scale time-series perception model and performing a prediction confidence assessment. The equipment load assessment employs a weighted comprehensive assessment method, expressed by the following formula:

[0075] ;

[0076] In the formula, The overall load index of the equipment; The current number of tasks, in units; The maximum task capacity is measured in units of 20; the experience value is 20. Current utilization rate, in percentages (%) To maximize utilization, a value of 100% is typically used. This represents remaining production capacity, expressed in units per hour. Total capacity, in units per hour; , , The weighting coefficients for equipment load assessment are empirically set to 0.4, 0.3, and 0.3. The prediction confidence feature vector construction includes three dimensions: prediction variance, model gradient rate of change, and historical prediction bias, represented as follows:

[0077] ;

[0078] In the formula, To predict the confidence feature vector; The current prediction variance is obtained by calculating the variance from multiple predictions output by the model. The variance of the prediction is obtained by using the variance statistics of the historical best prediction state as a reference, with an empirical value of 0.01. This represents the model gradient rate of change, expressed in units of 1 / time, and is obtained by calculating the Euclidean distance between the gradient vectors between two consecutive training iterations. The reference gradient change rate, in units of 1 / time, is obtained through the statistical value of the gradient change rate of the historical best training state, with an empirical value of 0.005. Historical forecast bias, expressed in hours, is obtained by calculating the average absolute error between recent forecasts and actual values. The reference prediction bias, in hours, is obtained through statistical values ​​of the deviation from the historical best prediction accuracy, with an empirical value of 0.5 hours. The baseline confidence vector is established based on the historical best state, and its vector representation is as follows:

[0079] ;

[0080] In the formula, The baseline confidence vector is defined as follows, with each component being a dimensionless, standardized value of 1.0. The Euclidean distance function is used to calculate the spatial distance between the predicted confidence feature vector and the baseline confidence vector, as expressed in the following formula:

[0081] ;

[0082] In the formula, To standardize Euclidean distance; To predict the confidence feature vector of the th One component; The baseline confidence vector is the first One component; To predict the maximum value of the confidence feature vector; This represents the maximum value of the baseline confidence vector; The vector dimension is 3. The mapping clustering distance calculation function is used to calculate the difference in cluster affiliation between two feature vectors in a high-dimensional mapping space. The formula is as follows:

[0083] ;

[0084] In the formula, To map cluster distance; To predict the confidence feature vector at the th The mapping value of each cluster center is obtained by calculating the mapping value from the vector to the cluster center using a clustering algorithm; The baseline confidence vector is at the th The mapping values ​​of each cluster center are calculated using the same clustering algorithm; For the first The weight of each cluster is determined by the proportion of the number of samples in each cluster to the total number of samples. For the first The standard deviation of each cluster is obtained by calculating the standard deviation of the distance from the sample to the cluster center within the cluster; This represents the number of clusters, with a default value of 5. The hybrid similarity calculation function combines spatial distance and cluster distance information, and its formula is as follows:

[0085] ;

[0086] In the formula, For mixed similarity; This is the weighting balancing factor, with an empirical value of 0.6; This is the normalization coefficient for the maximum spatial distance, which is usually set to 1.0.

[0087] The specific implementation of step S05 involves the coordinating agent using a Pareto front multi-objective optimization algorithm to process bidding schemes and achieve adaptive weight adjustment. The objective function vector representation of the multi-objective optimization problem is as follows:

[0088] ;

[0089] In the formula, A vector of multi-objective functions; The objective function for conversion time is given in hours. For reference, the changeover time is in hours, with an empirical value of 8 hours; The objective function is the order on-time rate. The objective function is the equipment load. This is a reference load value; the empirical value is 0.8. The objective function is energy consumption cost, with units of... ; For reference energy consumption values, the unit is... Experience value: 100 The Pareto embedding optimization function is used to incorporate the results of the Pareto front multi-objective optimization algorithm into the parameter update process of a neural network. The formula is as follows:

[0090] ;

[0091] In the formula, Update the direction vector for the parameters; This is the gradient fusion coefficient, with an empirical value of 0.7; This represents the current network gradient; For the first The weights of each objective are dynamically determined through an adaptive weight adjustment mechanism. For the first The gradient of each objective function; The number of objective functions is set to 4. The degree of variation in objective conflict is calculated using the coefficient of variation method for objective function values ​​within a continuous time window, as expressed in the following formula:

[0092] ;

[0093] In the formula, The degree of conflict variation of the target; For the first The standard deviation of each objective function within a time window is obtained by calculating the standard deviation of the objective function values ​​within five consecutive time windows. For the first The mean of each objective function within a time window is obtained by calculating the average of the objective function values ​​over five consecutive time windows. The adaptive weight adjustment mechanism dynamically adjusts the weight allocation based on real-time production status. The weight update formula is as follows:

[0094] ;

[0095] In the formula, For the updated number Each target weight; For the previous version Each target weight; This is the weighting adjustment factor, with an empirical value of 0.1; For the first The priority score of each target is determined by analyzing the current production status. When an urgent order is inserted, the priority score of the delivery date target is 1.5. When the equipment load is too high, the priority score of the load balancing target is 1.2. Under normal conditions, the priority score of each target is 1.0.

[0096] The specific implementation of step S06 involves the process intelligence agent confirming the changeover preparation and monitoring the equipment's execution of the changeover operation. The changeover execution progress deviation assessment uses a method to calculate the relative deviation between the actual progress and the expected progress, expressed by the following formula:

[0097] ;

[0098] In the formula, Schedule deviation rate; The actual changeover time is in hours. The expected changeover time is in hours. The early warning mechanism is triggered in a timely manner.

[0099] The specific implementation of step S07 involves the system identifying disturbance events and performing incremental renegotiation on the affected subset of orders. The impact intensity assessment employs a multi-dimensional comprehensive assessment method for disturbance events, expressed by the following formula:

[0100] ;

[0101] In the formula, To influence the intensity index; To score the severity of the disturbance, an expert evaluation system is used to score the severity of the disturbance event on a scale of 1 to 10. The maximum severity level is typically set to 10. The duration of the disturbance is in hours. This is the maximum duration, expressed in hours; the empirical value is 24 hours. The disturbance's impact range is expressed in units. To determine the maximum impact range, the unit is units; the empirical value is the total number of all devices. , , The empirical values ​​for the weighting coefficients affecting the intensity assessment are 0.5, 0.3, and 0.2, respectively.

[0102] Separable attention computation function is used to reduce the computational cost of multi-head attention mechanisms, and the formula is expressed as follows:

[0103] ;

[0104] In the formula, The attention weight matrix is ​​separable. Attention weights are assigned to the time dimension. Attention weights for feature dimensions; This refers to the Kronecker product operation. The formula for calculating attention in the time dimension is as follows:

[0105] ;

[0106] In the formula, Query the matrix based on the time dimension; The key matrix is ​​a time-dimension matrix; This represents the number of features in the time dimension; the default value is 64. The normalized exponential function transforms the input vector into a probability distribution. The formula for calculating the attention dimension in the feature dimension is as follows:

[0107] ;

[0108] In the formula, Query matrix for feature dimensions; The key matrix represents the feature dimension. This represents the number of feature dimensions, which defaults to 128.

[0109] The Euclidean distance calculation function eliminates the influence of numerical differences between features of different dimensions through standardization and uses a normalization method to map each component to a uniform scale for distance calculation.

[0110] ;

[0111] This function can quickly identify the degree of deviation between the predicted confidence vector and the baseline vector in linear space. Compared with the traditional unstandardized Euclidean distance calculation method, it avoids the calculation bias caused by inconsistency in dimensions, providing a reliable spatial distance basis for hybrid similarity assessment. The mapping clustering distance calculation function is based on the clustering similarity measurement principle of Gaussian kernel function. It evaluates the deep structural similarity between vectors by calculating the mapping response intensity of feature vectors at multiple cluster centers.

[0112] ;

[0113] The exponential decay term in this function It can effectively capture the relative positional relationships of vectors in nonlinear data distributions. Compared with simple spatial distance measurement methods, it can identify potential clustering patterns in the data, providing a more accurate similarity judgment for predictive state assessment under complex nonlinear conditions. The hybrid similarity calculation function integrates complementary information from spatial distance and clustering distance through linear weighting. The setting of the weight balancing factor allows the function to adjust the importance of the two distance information according to the application scenario.

[0114] ;

[0115] This function organically combines two different dimensions of similarity measurement, forming a more comprehensive and robust similarity evaluation index. Compared with single distance measurement methods, it significantly improves the accuracy and sensitivity of anomaly detection in predicted states, making it particularly suitable for complex production environments with multivariate coupling. The Pareto embedding optimization function directly integrates the global search capability of multi-objective optimization into the parameter update process of the neural network. Through gradient fusion, it ensures that network parameter updates not only consider minimizing prediction error but also simultaneously coordinate the optimization of multiple production objectives.

[0116] ;

[0117] Gradient fusion coefficients in this function It can dynamically balance the weight allocation between single-objective and multi-objective learning, avoiding the local optimum trap problem compared to traditional single-objective optimization methods. This achieves a synergistic improvement in prediction accuracy and multi-objective balance, providing an effective parameter optimization strategy for intelligent decision-making in complex, multi-constrained production environments. The separable attention calculation function reconstructs the independent attention calculation results in the time and feature dimensions into a complete attention weight matrix through Kronecker product operations.

[0118] ;

[0119] This decomposition strategy avoids redundant operations and cross-interference in traditional joint computation, significantly reducing computational complexity. arrive While maintaining the original attention modeling accuracy, it is particularly suitable for processing long-term series and high-dimensional feature data. Compared with traditional multi-head attention mechanisms, it achieves significant optimization in terms of computational efficiency and memory usage, providing an efficient computational foundation for real-time dynamic negotiation systems. The equipment load assessment function comprehensively quantifies the current working status of the equipment through multi-dimensional weighted synthesis.

[0120] ;

[0121] This function considers key load indicators such as task quantity, utilization rate, and remaining capacity. Normalization eliminates dimensional differences between different indicators, and the weighting coefficients reflect the relative importance of different load dimensions. Compared to single-load indicator evaluation methods, it more accurately reflects the true load status of equipment, providing a reliable load assessment basis for the bidding decisions of equipment agents. The impact intensity assessment function quantifies the impact intensity of disturbances on the production system by comprehensively considering multiple dimensions such as the severity, duration, and scope of the disturbance event.

[0122] ;

[0123] This function uses a normalized weighted summation method to integrate influencing factors from different dimensions into a unified intensity index. Compared with single-dimensional evaluation methods, it can more comprehensively reflect the actual impact of disturbance events, providing a scientific quantitative basis for triggering decisions for incremental renegotiation and ensuring that the system can accurately identify the range of orders that need to be renegotiated.

[0124] To better understand and implement this invention, the following is a specific application scenario of this invention, Example 2:

[0125] A manufacturing technology team faced a complex multi-product changeover scheduling problem. The production line needed to handle 15 different product specifications, involving 12 critical processing machines. Traditional static scheduling methods could not cope with frequent order changes and equipment status fluctuations, resulting in inaccurate changeover time predictions and low equipment utilization. To solve these problems, the technology team decided to adopt an efficient changeover scheduling method based on AI-powered multi-agent dynamic negotiation.

[0126] The technical team first constructed a multi-scale time-series awareness model. The system collected the replacement history of all devices over the past six months, including product models before and after the replacement, replacement duration, device status parameters, ambient temperature, and humidity, totaling 4280 records. The data was divided into training and testing sets in a 7:3 ratio, and the time-series data was processed using a sliding window to form sample pairs. The multi-scale time-series awareness model adopts an encoder-decoder architecture. The encoder contains three layers of long short-term memory network units, with 128 neurons per layer. The decoder uses a multi-head attention mechanism to fuse features from multiple time scales. Model training used the Adam optimizer with an initial learning rate of 0.001, a batch size of 64, and 200 training epochs. The loss function was a weighted combination of mean squared error and confidence regularization. An early stopping mechanism was used to prevent overfitting, and the final prediction error of the model on the test set was controlled within 5%.

[0127] The system establishes four types of intelligent agents: equipment intelligent agents, order intelligent agents, process intelligent agents, and coordination intelligent agents. Each intelligent agent stores its corresponding state parameters, historical data, and constraints, and achieves information interaction and decision transmission through a dynamic hierarchical negotiation architecture. The dynamic hierarchical negotiation architecture divides the intelligent agent system into three layers: a core coordination layer, a functional processing layer, and a data execution layer. The core coordination layer is responsible for global decision-making and conflict arbitration; the functional processing layer, including equipment intelligent agents, order intelligent agents, and process intelligent agents, is responsible for domain-specific processing; and the data execution layer is responsible for real-time data acquisition and instruction execution. (See Table 1 for details.)

[0128] Table 1. Agent Configuration Parameter Table

[0129]

[0130] In actual operation, the order agent sends a changeover scheduling request to the coordinating agent. One morning, the system received a batch of new orders, including production requirements for 180 units of product A, 220 units of product B, and 150 units of product C, with delivery times of 72 hours, 84 hours, and 96 hours respectively. The coordinating agent, based on process constraints, filters a set of equipment agents that meet the requirements and finds that equipment 1, equipment 3, equipment 7, equipment 9, and equipment 11 can handle the production of this batch of orders. It then broadcasts the scheduling task to these five equipment agents.

[0131] The intelligent device agent predicts the replacement time based on a multi-scale temporal perception model and generates a bidding plan by combining its own equipment status and load conditions. Device 1 currently has a load rate of 65%, and predicts a replacement time of 2.3 hours from its current product to product A, with a confidence level of 0.92. Device 3 currently has a load rate of 72%, and predicts a replacement time of 1.8 hours, with a confidence level of 0.89. Device 7 currently has a load rate of 58%, and predicts a replacement time of 2.1 hours, with a confidence level of 0.94. The system calculates the mixed similarity between the predicted confidence feature vector and the baseline confidence vector. The mixed similarity for device 1 is 0.88, for device 3 it is 0.82, and for device 7 it is 0.91. Based on the similarity threshold, the mixed similarity between device 1 and device 7 is within the range of [0.85, 1], so the current prediction parameters are maintained; the mixed similarity between device 3 is within the range of [0.65, 0.85), so the system adjusts the attention weight convergence factor from 0.15 to 0.18.

[0132] The coordinating agent uses a Pareto front multi-objective optimization algorithm to process the collected bidding proposals. The system balances multiple objectives, such as changeover time, order delivery time, and equipment load, through an adaptive weight adjustment mechanism. Currently, the weight for changeover time is 0.35, the weight for order delivery time is 0.40, and the weight for equipment load is 0.25. After 85 rounds of iterative optimization, the system obtains a Pareto optimal solution set containing 23 non-dominated solutions. The calculated objective conflict variation is 3.8%, less than the 5% threshold, and the system maintains a stable weight allocation strategy. The final production scheduling plan is as follows: Equipment 7 is responsible for the production of product A, with an estimated changeover time of 2.1 hours; Equipment 1 is responsible for the production of product B, with an estimated changeover time of 2.3 hours; and Equipment 9 is responsible for the production of product C, with an estimated changeover time of 1.9 hours.

[0133] The process agent and material agent confirm the changeover preparation. The process agent checks the process parameter settings for products A, B, and C, and confirms the readiness status of process equipment such as molds, cutting tools, and fixtures; the material agent verifies the adequacy of raw material inventory and delivery schedule. The equipment agent executes the changeover operation and provides real-time feedback on the execution status. The actual changeover time for equipment 7 is 2.15 hours, with a deviation of only 0.05 hours from the predicted value; the actual changeover time for equipment 1 is 2.28 hours, with a deviation of 0.02 hours; and the actual changeover time for equipment 9 is 1.92 hours, with a deviation of 0.02 hours. The coordination agent monitors the entire changeover process and records the actual changeover time for optimization of parameters in the multi-scale time-series perception model.

[0134] like Figure 2As shown, the multi-objective optimization convergence curve during system operation illustrates the evolution of three objectives: changeover time, order on-time rate, and equipment utilization. During production execution, a sudden equipment failure occurred. Equipment 3 experienced an abnormal spindle temperature alarm during its 48th hour of operation, affecting the production plans of two subsequent orders. The system automatically identified the affected order subset, including orders for 95 units of product model D and 110 units of product model E. Incremental renegotiation was triggered only for these two orders; the original production schedules for the remaining orders remained unchanged. The incremental renegotiation was completed within 8 minutes, reassigning the affected orders to equipment 5 and equipment 8, with the expected delay time controlled within 3.5 hours.

[0135] As shown in Table 2, the system exhibits excellent performance in key performance indicators during operation:

[0136] Table 2 System Performance Indicators Statistics

[0137]

[0138] The Pareto embedding optimization mechanism plays a crucial role in the multi-scale temporal awareness model. During neural network training, the gradient fusion coefficient β is dynamically adjusted from 0.3 to 0.8. When the production environment is relatively stable, the β value is set to 0.75 to focus on improving prediction accuracy. When facing increased multi-objective conflict, the β value is adjusted to 0.35 to enhance multi-objective coordination capabilities. The separable attention mechanism, through independent decomposition of the time and feature dimensions, reduces computational complexity from... Reduce to The time dimension attention process handles 168 time steps, and the feature dimension attention process handles 28 device feature dimensions, which significantly improves the training and inference efficiency of the model.

[0139] like Figure 3 As shown in the diagram, the working principle of the hybrid similarity assessment mechanism demonstrates the multi-dimensional comparison process between the predicted confidence feature vector and the baseline confidence vector. In the hybrid similarity calculation, the weight balancing factor α is set to 0.6, and the maximum spatial distance normalization coefficient... The value is 1.8. Euclidean distance calculation reflects the linear spatial relationship between vectors, while mapping clustering distance calculation captures deep feature patterns in complex nonlinear data distributions. When the predicted state of equipment replacement deviates from the historical optimal operating mode, hybrid similarity can sensitively detect the dual changes in spatial distance and cluster affiliation, triggering timely adaptive adjustment of model parameters.

[0140] The system excels in handling long-term, gradual changes. Faced with equipment aging, process parameter drift, or environmental condition variations, the hybrid similarity continuously monitors the health of the prediction model, ensuring the long-term accuracy of changeover time predictions. The multi-dimensional evaluation features effectively distinguish between temporary disturbances and systematic shifts, avoiding system oscillations caused by frequent adjustments to model parameters due to occasional anomalies, thus guaranteeing the stability and continuity of the entire dynamic negotiation process.

[0141] This invention represents a significant technological advancement over traditional methods. Traditional changeover scheduling methods rely on static rules and human experience, failing to handle dynamically changing production environments. This invention, however, achieves distributed decision-making and real-time response through a multi-agent negotiation mechanism. The hybrid similarity evaluation mechanism overcomes the limitations of a single distance metric by fusing spatial distance and clustering distance information, enabling a comprehensive assessment of prediction reliability and demonstrating stronger anomaly detection capabilities compared to traditional simple error statistics methods. The Pareto embedding optimization mechanism deeply integrates multi-objective optimization into the neural network training process, avoiding the subjectivity issues of traditional weight settings and achieving synergistic optimization of prediction accuracy and multi-objective balance. The separable attention mechanism completely reconstructs the computational architecture of traditional multi-head attention through a dimensionality decomposition strategy, significantly reducing computational complexity while maintaining modeling accuracy, providing efficient technical support for real-time dynamic negotiation. The incremental renegotiation mechanism employs a local optimization strategy for production disturbances, avoiding the efficiency problems of traditional full recalculation, and achieving accurate identification of the impact range and rapid adjustment response.

[0142] It should be noted that the variables involved in this invention are explained in detail in Tables 3 and 4 below.

[0143] Table 3. Variable Explanation Table (Part 1)

[0144]

[0145] Table 4. Variable Explanation Table (Part Two)

[0146]

[0147] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A highly efficient changeover scheduling method based on AI multi-agent dynamic negotiation, characterized in that, This paper constructs a prediction foundation by fusing historical equipment replacement data, real-time operating status data, and environmental parameter data using a multi-scale time-series perception model. It establishes equipment agents, order agents, process agents, material agents, and coordination agents to achieve dynamic hierarchical negotiation and decision-making. A hybrid similarity evaluation mechanism is used to dynamically adjust model parameters to ensure prediction stability. The steps include: constructing a multi-scale time-series perception model, fusing historical equipment replacement data, real-time operating status data, and environmental parameter data to establish a basic dataset for replacement time prediction; establishing equipment agents, order agents, process agents, and coordination agents, and achieving information interaction and decision transmission among agents through a dynamic hierarchical negotiation architecture; and the order agent sending replacement schedules to the coordination agent. The system receives production requests, coordinates the selection of a set of eligible equipment agents, and broadcasts production scheduling invitations. Equipment agents predict changeover times based on a multi-scale time-series perception model, generate bidding proposals, calculate the mixed similarity between the predicted confidence feature vector and the baseline confidence vector, and adjust model parameters according to the range of mixed similarity values. The coordination agent uses a Pareto front multi-objective optimization algorithm to process the collected bidding proposals, balancing multiple objectives such as changeover time, order delivery date, and equipment load. Process agents and material agents confirm changeover preparation, and equipment agents execute changeover operations and provide real-time feedback on execution status. When disturbances occur during production, the system automatically identifies the affected order subset and triggers an incremental renegotiation process only for that subset. The steps for constructing the multi-scale temporal perception model specifically involve processing temporal correlation features under nonlinear operating conditions through a residual connection mechanism. The specific structure of the multi-scale temporal perception model is an encoder-decoder architecture. The encoder contains three layers of long short-term memory network units, with 128 neurons in each layer. The decoder uses a multi-head attention mechanism to fuse features across multiple time scales. The output layer generates replacement time prediction values ​​and confidence evaluation values ​​through a fully connected network. The steps for the coordinating agent to process the collected bidding schemes using the Pareto front multi-objective optimization algorithm specifically involve balancing multiple objectives such as replacement time, order delivery date, and equipment load through an adaptive weight adjustment mechanism. The process intelligence agent and material intelligence agent perform changeover preparation confirmation, and the equipment intelligence agent executes the changeover operation and provides real-time feedback on the execution status. Specifically, after receiving the production scheduling plan, the process intelligence agent immediately initiates the changeover preparation confirmation program to verify the correctness of process parameter settings, check the feasibility of the process path, and verify the completeness of quality inspection standards. The confirmation content includes the logical sequence of changeover process steps, the parameter range of key control points, and the execution requirements of safety operating procedures. Simultaneously, the material intelligence agent confirms the material preparation status, checks the adequacy of raw material inventory, verifies the matching of material specifications, and ensures the qualification of material quality. Material preparation confirmation involves raw material batch traceability, quality certificate verification, and storage condition checks. After receiving the process confirmation and material confirmation signals, the equipment intelligence agent begins to execute the changeover operation. The changeover process proceeds step by step according to the predetermined process flow, including standardized steps such as equipment shutdown, tooling replacement, parameter adjustment, trial operation, and quality inspection. The real-time feedback mechanism for execution status continuously monitors the changeover progress through a sensor network and data acquisition system. Key monitoring parameters include the completion time of each changeover stage, the stability of equipment operating parameters, and the compliance status of product quality indicators. The feedback frequency is set to update once every 30 seconds to ensure the real-time nature of the information.

2. The method according to claim 1, characterized in that, The steps for establishing equipment intelligence agents, order intelligence agents, process intelligence agents, material intelligence agents, and coordination intelligence agents are as follows: each intelligence agent stores its corresponding state parameters, historical data, and constraints. The steps for establishing the training dataset for the multi-scale time-series perception model are as follows: all equipment replacement history records within the past 6 months are collected, including product models before and after replacement, replacement duration, equipment state parameters, and ambient temperature and humidity data, and the training set and test set are divided in a 7:3 ratio.

3. The method according to claim 2, characterized in that, The hybrid similarity is a similarity evaluation index that comprehensively considers Euclidean distance and mapped clustering distance. The spatial distance between the predicted confidence feature vector and the baseline confidence vector is calculated using the Euclidean distance calculation function, and the clustering difference between the two vectors in high-dimensional space is calculated using the mapped clustering distance calculation function. Finally, the two distance information are fused using the hybrid similarity calculation function to obtain a comprehensive similarity evaluation result.

4. The method according to claim 3, characterized in that, The specific steps for generating a bidding proposal by the device intelligent agent are as follows: the bidding proposal is generated by combining its own equipment status and load conditions; the loss function in the training process of the multi-scale time-series perception model adopts a weighted combination of mean square error and confidence regularization term; overfitting is prevented by an early stopping mechanism; and the dynamic hierarchical negotiation architecture is an organizational structure that divides the intelligent agent system into three layers: core coordination layer, functional processing layer, and data execution layer.

5. The method according to claim 4, characterized in that, In the dynamic hierarchical negotiation architecture, the core coordination layer is responsible for global decision-making and conflict arbitration, the functional processing layer includes equipment intelligence agents, order intelligence agents and process intelligence agents responsible for corresponding domain processing, and the data execution layer is responsible for real-time data acquisition and instruction execution. The numerical range of the mixed similarity is adjusted to adjust the model parameters. Specifically, if the mixed similarity ∈ [0.85, 1], the current prediction parameters are maintained.

6. The method according to claim 5, characterized in that, The disturbance events are specifically equipment failures or emergency order insertions. The Pareto front multi-objective optimization algorithm approximates the Pareto optimal front by maintaining a non-dominated solution set, avoiding the subjective problem of weight setting in the traditional weighted summation method. At the same time, it optimizes changeover time, order on-time rate and equipment utilization. The adaptive weight adjustment mechanism is an algorithm that dynamically adjusts the weights of each objective in the multi-objective optimization according to the real-time production status.

7. 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 efficient changeover scheduling method based on AI multi-agent dynamic negotiation as described in any one of claims 1-6.

8. A highly efficient changeover scheduling system based on AI multi-agent dynamic negotiation, characterized in that, The system includes the computer-readable storage medium of claim 7, 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