A production line intelligent scheduling system and a multi-device collaborative control method

Abstract

The application relates to the technical field of scheduling control, and provides a production line intelligent scheduling system and a multi-device cooperative control method. Event data of a production line is acquired, a causal relationship between devices and tasks is identified, a causal diagram network of the production line is generated, a causal influence factor is broadcast, the performance of the devices is predicted and mapped to device capability constraints, it is judged whether a subpackage request is broadcast and a takeover device is determined, the causal diagram network is updated in a closed loop and the causal influence factor is rebroadcast, the production state change can be perceived in real time, potential risks can be accurately predicted, optimization decisions can be quickly made, efficient allocation of production resources and stable operation of the production process can be realized, and production cost can be effectively reduced, production efficiency and product quality can be improved.

Description

Technical Field

[0001] This application relates to the field of scheduling and control technology, and in particular to an intelligent scheduling system for a production line and a method for collaborative control of multiple devices. Background Technology

[0002] As the manufacturing industry moves towards intelligence and digitalization, intelligent production line scheduling technology has become the key to improving production efficiency and corporate competitiveness. Early production line scheduling mainly relied on human experience to formulate production plans and to allocate tasks through paper work orders or simple spreadsheets. This method was greatly affected by human factors, had low scheduling efficiency, was prone to errors, and was difficult to adapt to complex and ever-changing production environments.

[0003] With the development of computer technology, rule-based scheduling systems have emerged. These systems allocate tasks through preset production rules and priority strategies, such as formulating scheduling plans according to rules like order delivery dates and equipment capacity. This can improve the accuracy and efficiency of scheduling to some extent. However, such systems are not very flexible, and it is difficult to quickly adjust the scheduling plan when unexpected situations such as equipment failures or order changes occur during the production process.

[0004] To address these issues, researchers have proposed scheduling methods based on mathematical programming, such as intelligent optimization algorithms like genetic algorithms and particle swarm optimization. These algorithms solve production scheduling problems by constructing mathematical models and can search for optimal scheduling schemes within a certain range. They have achieved good application results in static production environments, but in actual production lines, factors such as equipment status and task requirements change frequently, making it difficult to meet the needs of dynamic scheduling.

[0005] In recent years, some enterprises have attempted to introduce IoT and big data technologies to achieve digital management of production lines, analyzing collected equipment operation data. However, these efforts are often limited to equipment status monitoring and simple statistical analysis, lacking in-depth exploration of the causal relationships between equipment and tasks. This fails to provide effective support for scheduling decisions. Regarding multi-device collaboration, existing technologies often treat equipment as independent entities for scheduling, ignoring the mutual influence and synergistic effects between devices. This results in underutilization of equipment resources and hinders further improvement in overall production line efficiency. Most of these approaches fail to address how to identify the causal relationships between various elements of the production line in complex production environments to achieve intelligent production line scheduling. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this application provides an intelligent scheduling system for production lines and a method for collaborative control of multiple devices.

[0007] In a first aspect, this application provides an intelligent scheduling system for a production line, the system comprising: acquiring event data of the production line, the event data including equipment status, task status and abnormal events, identifying causal relationships between equipment and tasks through a causal discovery algorithm to generate a causal graph network of the production line, and broadcasting causal influencing factors;

[0008] The system receives causal influencing factors, uses a time-series prediction algorithm to predict the device's performance based on these factors, and maps them to device capability constraints. When a task is detected to violate the device capability constraints, the system broadcasts a sub-packet request to neighboring devices. Devices receiving the sub-packet request determine the cooperation utility based on their own performance, bid on the cooperation utility to determine the takeover device, output scheduling instructions, and trigger an update of the takeover device's performance.

[0009] It receives scheduling instructions and controls the takeover equipment, locates downstream causal devices by combining the causal graph network, monitors the update changes in the capability performance of the takeover equipment to incrementally update the causal graph network, and rebroadcasts causal influencing factors.

[0010] As an optional implementation, the broadcast logic of the causal influencing factor includes:

[0011] Based on the edge weights and node attributes of the causal graph network, the causal influence factor of each causal relationship is determined by the fuzzy comprehensive evaluation method according to the importance of the equipment, the urgency of the task, and the severity of the anomaly.

[0012] The event data of the production line is analyzed by using a long short-term memory network to predict the importance of different causal factors in the scheduling process, and the broadcast priority and broadcast frequency of causal factors are dynamically adjusted according to the prediction results.

[0013] The causal graph network is divided into multiple subgraphs, each representing a functional area of ​​the production line, and causal influence factors are broadcast to each subgraph based on broadcast priority and broadcast frequency.

[0014] As an optional implementation, the generation sub-logic of the causal graph network of the production line includes:

[0015] Acquire event data from the production line, align event data of different frequencies using a dynamic time warping algorithm, and standardize the event data.

[0016] By combining the causal discovery algorithm with the temporal properties of the production line, the causal relationship between equipment and tasks is identified and analyzed to form a candidate set of causal relationships;

[0017] Using the production line's operating efficiency and scheduling optimization objectives as reward functions, a deep Q-network is used to iteratively filter and adjust the weights of the causal relationship candidate set to generate a causal graph network for the production line.

[0018] As an optional implementation, the output logic of the scheduling instruction includes:

[0019] The device that receives the subcontracting request formulates a bidding strategy based on the collaborative utility and dynamically adjusts its bidding price according to its own device status through the simulated annealing algorithm.

[0020] The equipment to be taken over is determined based on the bidding price, and initial scheduling instructions are generated. At the same time, the execution status of the tasks is monitored. When a task is detected to violate the equipment capacity constraints, a partial adjustment of the scheduling instructions is triggered.

[0021] As an optional implementation, the prediction sub-logic for the device's performance capabilities includes:

[0022] The causal influencing factors are sampled by sliding window through multi-scale time windows, and the spatiotemporal features of the causal graph network are extracted based on the spatiotemporal graph convolutional network. The spatiotemporal features include equipment status trends, task association features and anomaly propagation features.

[0023] By using Bayesian deep learning, the device's performance is probabilistically predicted based on spatiotemporal features. The output includes a predicted distribution with confidence intervals. Simultaneously, the perturbation of causal influencing factors is simulated during the prediction process, and the marginal effect of causal influencing factors on the device's performance is quantified.

[0024] The equipment's performance includes process set, task completion time, health index, and energy consumption. Mapping the equipment's performance to the constraint space forms the equipment's capability constraints.

[0025] As an optional implementation, the sub-logic for determining the cooperative utility includes:

[0026] Upon receiving a subcontracting request, the device extracts features of its own capabilities based on a self-attention mechanism to obtain device capability features.

[0027] The task attributes are parsed from the sub-packet request, and the device capability characteristics are dynamically fused with the task attributes to generate a fused feature vector.

[0028] The collaborative utility of each device that received the subpacket request is obtained by weighted summation of the fused feature vectors.

[0029] As an optional implementation, the rebroadcast logic of the causal influencing factor includes:

[0030] The incrementally updated causal graph network is processed using a convolutional neural network and an attention mechanism to output a weight matrix;

[0031] The reward function for reinforcement learning is based on the production line's operating efficiency, energy consumption cost, and task delay risk. Broadcast priority and broadcast frequency are assigned to causal influencing factors according to the weight matrix.

[0032] The causal influence factor is rebroadcast to each subgraph based on the assigned broadcast priority and broadcast frequency.

[0033] As an optional implementation, the positioning sub-logic of the causal downstream device includes:

[0034] Multi-level analysis of the causal graph network is performed using graph attention network to identify the device nodes associated with the takeover device node, so as to form a causal propagation path;

[0035] For each causal propagation path, the probability distribution and risk value of downstream devices being affected are calculated to form a risk propagation matrix, and devices with risk values ​​greater than the risk threshold are selected as causal downstream devices.

[0036] A secondary screening of downstream devices based on task attributes is performed to locate downstream devices.

[0037] As an optional implementation, the incremental update sub-logic of the causal graph network includes:

[0038] After the takenover equipment executes the scheduling command, the system monitors the changes in the capabilities of the takenover equipment in real time to generate event codes.

[0039] By comparing the differences between event codes and historical codes, changes in device state caused by the execution of scheduling instructions are identified. The causal relationship between devices and tasks is re-analyzed through a causal discovery algorithm. At the same time, affected causal edges and nodes are marked to form a local perturbation map.

[0040] The local perturbation map is fused with the causal graph network through a graph neural network to incrementally update the causal graph network.

[0041] Secondly, this application provides a method for collaborative control of multiple devices in a production line, the method comprising:

[0042] Acquire event data from the production line, including equipment status, task status, and abnormal events. Use causal discovery algorithms to identify causal relationships between equipment and tasks to generate a causal graph network for the production line and broadcast causal influencing factors.

[0043] Receive causal influencing factors, predict the device's performance based on the causal influencing factors using a time-series prediction algorithm, and map it to device capability constraints;

[0044] When a task is detected to violate the device's capability constraints, a sub-packet request is broadcast to neighboring devices. The devices that receive the sub-packet request determine the effectiveness of the collaboration based on their own capabilities.

[0045] The takeover device is determined based on collaborative utility bidding, scheduling instructions are output, and updates to the capabilities of the takeover device are triggered.

[0046] It receives scheduling instructions and controls the takeover equipment, locates downstream causal devices by combining the causal graph network, monitors the update changes in the capability performance of the takeover equipment to incrementally update the causal graph network, and rebroadcasts causal influencing factors.

[0047] Compared with existing technologies, the beneficial effects of this application are as follows: by acquiring event data from the production line, identifying the causal relationship between equipment and tasks and generating a causal graph network of the production line, and simultaneously broadcasting causal influencing factors, predicting the equipment's performance capabilities and mapping them to equipment capability constraints, it is possible to determine whether to broadcast subcontracting requests and identify the equipment to take over, thereby updating the causal graph network in a closed loop and rebroadcasting causal influencing factors. This enables real-time perception of changes in production status, accurate prediction of potential risks, and rapid optimization decisions, achieving efficient allocation of production resources and stable operation of the production process, effectively reducing production costs, improving production efficiency and product quality.

[0048] By acquiring multi-source data such as equipment status, task status, and abnormal events, and generating a causal graph network of the production line through a causal discovery algorithm and broadcasting causal influencing factors, it is possible to deeply analyze the causal relationships between various elements in the production line. This transforms complex production logic into a visualized and quantifiable causal relationship network, thereby avoiding the problems of traditional methods that fail to identify potential causal relationships, making it difficult to trace production anomalies and lacking a basis for scheduling decisions. Furthermore, by dynamically broadcasting causal influencing factors, it provides rich and accurate decision-making information, enabling the scheduling system to predict production changes in advance. This lays the foundation for intelligent and precise scheduling and effectively improves the transparency and predictability of the production line.

[0049] By receiving causal influencing factors, predicting equipment performance through time-series prediction algorithms and forming constraints, and combining collaborative utility bidding to determine the takeover equipment, scheduling instructions are output. This changes the traditional situation of coarse task allocation and uneven equipment load in scheduling, enabling fine-grained scheduling based on real-time equipment capabilities and task requirements. When a task is detected to violate equipment capability constraints, a subcontracting request and bidding mechanism are triggered in a timely manner, realizing the dynamic optimization allocation of production tasks among equipment. This not only fully leverages the performance advantages of each piece of equipment and avoids overuse or idleness, but also significantly improves the timeliness and accuracy of task execution, shortens the production cycle, enhances the adaptability of the production line to diversified orders and sudden production situations, and ensures efficient and stable operation of the production process.

[0050] Upon receiving scheduling instructions, the system locates downstream devices using a causal graph network. By monitoring device performance, it incrementally updates the causal graph network and rebroadcasts causal influencing factors, constructing a closed-loop feedback mechanism for production scheduling. This enables the scheduling system to track changes in production status in real time after the execution of scheduling instructions. By accurately locating downstream devices, it proactively prevents chain reactions caused by local equipment adjustments, reducing production risks. The incremental updates to the causal graph network and the rebroadcasting of causal influencing factors ensure that the production information held by the scheduling system is always synchronized with the actual production status, providing the latest and most accurate data support for subsequent scheduling decisions. This allows the production line to continuously self-adjust and improve in a complex and ever-changing production environment, maintaining efficient and stable operation, and further enhancing the production line's intelligence level and overall competitiveness. Attached Figure Description

[0051] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort. Wherein:

[0052] Figure 1 A flowchart of an intelligent production line scheduling system provided in this application embodiment;

[0053] Figure 2 A broadcast logic diagram of causal influencing factors for a production line intelligent scheduling system provided in this application embodiment;

[0054] Figure 3 An incremental update sub-logic diagram of a causal graph network for an intelligent production line scheduling system provided in this application embodiment;

[0055] Figure 4 This is a flowchart illustrating a multi-device collaborative control method for a production line, as provided in an embodiment of this application. Detailed Implementation

[0056] To make the objectives, technical solutions, and advantages of the embodiments of this application more apparent and understandable, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments.

[0057] Example 1

[0058] like Figure 1 The diagram shown is a flowchart of an intelligent production line scheduling system provided in this application embodiment. The system includes a causal perception module, an intelligent scheduling module, and a control feedback module.

[0059] The causal perception module is used to acquire event data from the production line, including equipment status, task status, and abnormal events. It uses a causal discovery algorithm to identify the causal relationships between equipment and tasks to generate a causal graph network for the production line and broadcasts causal influencing factors.

[0060] Furthermore, the generation sub-logic of the causal graph network of the production line includes:

[0061] Acquire event data from the production line, align event data of different frequencies using a dynamic time warping algorithm, and standardize the event data.

[0062] By combining the causal discovery algorithm with the temporal properties of the production line, the causal relationship between equipment and tasks is identified and analyzed to form a candidate set of causal relationships;

[0063] Using the production line's operating efficiency and scheduling optimization objectives as reward functions, a deep Q-network is used to iteratively filter and adjust the weights of the causal relationship candidate set to generate a causal graph network for the production line.

[0064] Taking a melamine tableware production line as an example, this line encompasses multiple processes including raw material proportioning, injection molding, pattern pressing, and polishing quality inspection. The sampling frequency and format of event data differ significantly across these processes. Injection molding machines monitor temperature and pressure data in real time, while quality inspection primarily relies on batch sampling data. Direct analysis can easily lead to misjudgments of causal relationships. Therefore, a dynamic time warping algorithm is needed to align the data and standardize it to unify data dimensions, laying the foundation for subsequent causal analysis. For the high-frequency temperature and pressure data sampled by the injection molding machine and the low-frequency pass rate data from the quality inspection process, a dynamic time warping algorithm is used to set the time window length to a single production batch cycle. Local path optimization is employed to match data points of different frequencies. Numerical data such as temperature and pressure are normalized using the Z-score method, while categorical data such as equipment start-up / shutdown and quality inspection results are converted using unique thermal encoding. This eliminates interference caused by differences in sampling frequency and dimensions, making the correlation between equipment fault warning signals and production anomalies clearer and avoiding causal mismatches due to data confusion.

[0065] Melamine tableware production has strict time-series characteristics. In actual production, insufficient holding time in the injection molding process will expose surface bubble defects in the subsequent polishing process. Here, a time-series-based causal discovery algorithm is used, with a maximum time lag of 3 production batches. By calculating the information transfer entropy between events in different processes, relationships with causal potential are screened out. Preferably, the correlation between injection molding temperature change and the drop resistance of the finished product after 2 batches is analyzed. For the initially screened relationships, the causal strength is calculated using a grey relational algorithm, and relationships with causal strength greater than the strength threshold are retained to form a candidate set of causal relationships. This effectively uncovers hidden causal relationships between processes, namely the correlation between raw material mixing ratio and injection molding stability, providing data support for optimizing the production process.

[0066] In melamine tableware production, factors such as order urgency and raw material supply stability dynamically change scheduling objectives. Urgent orders require priority to ensure delivery time, while ordinary orders focus on energy consumption control. A dual-objective reward function is designed, including production line operating efficiency and scheduling optimization objectives. When the scheduling optimization objective is to shorten delivery time, the reward function emphasizes the weight of task completion time; when the objective is to reduce costs, the weight of energy consumption is increased. The deep Q-network processes causal relationships through a two-layer long short-term memory network structure, and the experience replay buffer stores scheduling data. After every 20 scheduling executions, the parameters of the deep Q-network are updated based on reward feedback, and training fluctuations are suppressed through the target network. The deep Q-network can dynamically adjust the causal graph network according to actual production needs. In the case of urgent orders, it can automatically strengthen the causal relationships related to delivery time, improve response speed, and the optimized causal graph network accurately reflects the causal relationships between equipment and tasks under the current scheduling optimization objective, providing accurate network structure and weight information for the broadcast of causal influencing factors.

[0067] Specifically, such as Figure 2 As shown, the broadcast logic of causal influencing factors includes:

[0068] Based on the edge weights and node attributes of the causal graph network, the causal influence factor of each causal relationship is determined by the fuzzy comprehensive evaluation method according to the importance of the equipment, the urgency of the task, and the severity of the anomaly.

[0069] The event data of the production line is analyzed by using a long short-term memory network to predict the importance of different causal factors in the scheduling process, and the broadcast priority and broadcast frequency of causal factors are dynamically adjusted according to the prediction results.

[0070] The causal graph network is divided into multiple subgraphs, each representing a functional area of ​​the production line, and causal influence factors are broadcast to each subgraph based on broadcast priority and broadcast frequency.

[0071] In melamine tableware production lines, the impact of a single causal relationship on scheduling is influenced by multiple factors, including equipment importance, task urgency, and anomaly severity. The impact of injection molding machine malfunctions on production requires comprehensive evaluation, considering factors such as order delivery time and the equipment's core position on the production line; a single indicator cannot fully measure this. A three-layer fuzzy evaluation system is constructed by determining the causal influencing factors for each causal relationship based on equipment importance, task urgency, and anomaly severity. The factor set includes equipment importance, task urgency, and anomaly severity. Equipment importance includes injection molding machines and quality inspection equipment. Urgency includes regular orders and rush orders, while severity includes minor defects and batch scrap. The evaluation set is divided into three levels: high, medium, and low. The weight of each factor is determined through historical data, and specific indicators are mapped to the evaluation set through triangular membership functions. The causal influence factor of each causal relationship is calculated using a fuzzy synthesis operator. The fuzzy comprehensive evaluation method comprehensively considers the influence of multiple factors, accurately quantifies the actual impact of causal relationships on scheduling, avoids decision-making bias, and provides basic data for long short-term memory network prediction. Its accuracy directly affects the rationality of dynamic adjustment of broadcast priority.

[0072] In the production of melamine tableware, the importance of causal relationships changes dynamically with the production status. When raw material supply is insufficient, the importance of causal relationships related to raw material ratios increases significantly. Static broadcasting strategies cannot adapt to this change. Therefore, it is necessary to use Long Short-Term Memory (LSTM) networks to analyze event data and dynamically adjust broadcast priority and frequency. Historical causal influencing factor sequences are used as input to the LTM network to output predictions of the importance trends of each causal influencing factor in the future. The LTM network is trained through a dynamic learning rate optimizer, and the prediction results are updated periodically. Based on the predicted importance trends, the broadcast priority and frequency are adjusted. This allows the broadcasting strategy to respond to production changes in real time, ensuring the priority dissemination of key causal information, improving the timeliness of scheduling decisions, and guiding the information push strategy of the causal graph network subgraphs to achieve accurate information delivery.

[0073] The melamine tableware production line is functionally divided into raw material processing, molding, and post-processing areas. Each area focuses on different causal information; the molding area is particularly concerned with the causal relationship between injection molding and pressing processes. Broadcasting across the entire graph would cause information redundancy. Therefore, the causal graph network is divided into subgraphs, and causal influencing factors are broadcast to each subgraph in a targeted manner. Using a spectral clustering algorithm, the causal graph network is divided into three subgraphs—raw material, molding, and post-processing—based on the degree of correlation between processes between equipment. To ensure the transmission of key causal information across areas, edge nodes are set up as information relay stations. Each subgraph maintains an independent broadcast queue. High-priority causal influencing factors are transmitted quickly via multicast, while low-priority causal influencing factors are transmitted via unicast at a reduced frequency. Subgraph broadcasting optimization reduces invalid information transmission, lowers communication load, and enables equipment in each area to quickly obtain key causal information, improving scheduling efficiency. Precise information broadcasting provides effective input to the intelligent scheduling module, helping it to more accurately predict equipment performance and make reasonable scheduling decisions.

[0074] The intelligent scheduling module receives causal influencing factors, predicts the device's performance based on these factors using a time-series prediction algorithm, and maps them to device capability constraints. When a task violates the device capability constraints, it broadcasts a sub-packet request to neighboring devices. Devices receiving the sub-packet request determine the cooperation utility based on their own capability performance, bid on the cooperation utility to determine the takeover device, output scheduling instructions, and trigger an update of the takeover device's capability performance.

[0075] Furthermore, the predictive sub-logic for the device's performance capabilities includes:

[0076] The causal influencing factors are sampled by sliding window through multi-scale time windows, and the spatiotemporal features of the causal graph network are extracted based on the spatiotemporal graph convolutional network. The spatiotemporal features include equipment status trends, task association features and anomaly propagation features.

[0077] By using Bayesian deep learning, the device's performance is probabilistically predicted based on spatiotemporal features. The output includes a predicted distribution with confidence intervals. Simultaneously, the perturbation of causal influencing factors is simulated during the prediction process, and the marginal effect of causal influencing factors on the device's performance is quantified.

[0078] The equipment's performance includes process set, task completion time, health index, and energy consumption. Mapping the equipment's performance to the constraint space forms the equipment's capability constraints.

[0079] The melamine tableware production line involves multiple stages, including raw material injection molding, pattern pressing, high-temperature curing, and quality inspection. The changes in equipment status at each stage are complexly correlated across different time scales and spatial dimensions. Single-scale analysis cannot fully capture the operational patterns of the equipment. Therefore, multi-scale time window sampling and spatiotemporal graph convolutional networks are required to extract equipment status trends, task-related features, and anomaly propagation characteristics. Specifically, for the injection molding stage, short time windows are set to capture high-frequency fluctuations in injection pressure and temperature. For the curing stage, long-term windows are used to monitor temperature curve trends. Spatiotemporal feature extraction is performed on the causal graph network based on the spatiotemporal graph convolutional network. The production line equipment is abstracted as graph nodes, and the material transport and process connections between equipment are treated as edges. Multi-layer convolutional operations are used to extract the collaborative features of the equipment in the spatial dimension. Simultaneously, temporal convolution kernels are used to mine temporal features. Preferably, it identifies how abnormal injection molding machine pressure affects the molding quality of downstream pattern pressing equipment through the material transport chain. This comprehensive extraction of multi-scale spatiotemporal features allows for early prediction of potential equipment failures and process deviations, preventing chain reactions across the entire production line caused by local equipment anomalies and improving production stability.

[0080] The production process of melamine tableware is affected by uncertainties such as batch variations in raw materials and changes in environmental temperature and humidity. Bayesian deep learning, through probabilistic modeling, can output both predicted values ​​of equipment performance and confidence intervals for these predictions, quantifying the reliability of the prediction results and meeting the risk assessment needs of production scheduling. Extracted spatiotemporal features are input into a Bayesian neural network, where the parameters are random variables following a specific distribution. During training, the parameter distribution is sampled using the Markov chain Monte Carlo method to obtain the probability distribution of task completion time, health index, and energy consumption of the equipment under different process sets. Simultaneously, disturbances in causal influencing factors are simulated to analyze their marginal effects on equipment performance and assess potential risks. This provides a quantitative basis for scheduling decisions, allowing for advance adjustments to production plans based on the probability distribution of equipment energy consumption when raw material batches are unstable, thus reducing cost waste.

[0081] To ensure production quality and efficiency, equipment capabilities must be translated into executable constraints to limit task allocation and execution, preventing product defects or equipment damage due to overload. Based on the melamine tableware production process requirements, the equipment's process set, task completion time, health index, and energy consumption are used as performance indicators. The process set includes the injection pressure range and curing temperature range. These capabilities are then translated into a feasible domain within the constraint space. Specifically, the pressure fluctuation range of the injection molding machine is mapped to an acceptable pressure constraint for the task. When the required pressure exceeds this range, it is considered a violation of the equipment capability constraint. This effectively standardizes task allocation, reduces production anomalies caused by unreasonable scheduling, and ensures product qualification rate and equipment lifespan. Equipment capability constraints become a key basis for task execution monitoring. Once a task violates the constraint, a subcontracting request is immediately triggered, driving subsequent collaborative scheduling between equipment.

[0082] Furthermore, the judgment sub-logic for collaborative utility includes:

[0083] Upon receiving a subcontracting request, the device extracts features of its own capabilities based on a self-attention mechanism to obtain device capability features.

[0084] The task attributes are parsed from the sub-packet request, and the device capability characteristics are dynamically fused with the task attributes to generate a fused feature vector.

[0085] The collaborative utility of each device that received the subpacket request is obtained by weighted summation of the fused feature vectors.

[0086] The capabilities of various equipment on a melamine tableware production line vary at different production stages. For example, injection molding machines initially exhibit lower molding accuracy after mold changes. To accurately assess an equipment's ability to undertake subcontracted tasks, it's necessary to extract its core capability characteristics in its current state based on a self-attention mechanism. This mechanism extracts features from the equipment's process set, task completion time, health index, and energy consumption. By calculating attention weights between different feature dimensions, the mechanism focuses on key features. Preferably, when tableware patterns are complex, it emphasizes the precision-related features of the pattern-pressing equipment, filtering out redundant information to obtain the final capability characteristics. This allows for precise extraction of the equipment's core capability characteristics, ensuring that the equipment's self-assessment aligns more closely with actual production needs and avoids blindly participating in task bidding. The equipment capability characteristics provide effective information for integration with task attributes, ensuring that collaborative utility calculations accurately reflect the equipment's suitability for undertaking tasks.

[0087] Different tasks have different equipment capability requirements. For example, the production of large melamine plates and small tableware requires different injection molding machine pressures and mold specifications. Dynamically fusing equipment capability characteristics with task attributes can comprehensively consider the matching degree between the two parties and provide a comprehensive basis for calculating collaborative utility. Task attributes such as process requirements, delivery time, and quality standards are parsed from subcontracting requests and concatenated with equipment capability characteristics. The concatenated fused feature vector is processed through a gating loop unit. Based on the gating mechanism, the fusion ratio of equipment capability characteristics and task attributes is adaptively adjusted. For urgent orders, the fusion weight of the task delivery time attribute is increased to highlight the importance of the equipment's rapid response capability. The dynamic fusion of feature vectors fully considers the matching relationship between tasks and equipment, making the collaborative utility calculation more in line with the needs of actual production scenarios and improving the rationality of task allocation. The fused feature vector, as the input for collaborative utility calculation, directly determines the level of collaborative utility of each piece of equipment, thereby affecting the selection of takeover equipment.

[0088] By quantifying the value of subcontracted tasks undertaken by equipment, an objective basis is provided for determining which equipment to take over, achieving optimal resource allocation and ensuring overall production line efficiency. Based on melamine tableware production experience and production goals, weights are set for different dimensions of the fusion feature vector. In the stage of pursuing product quality, the weight of equipment precision-related features is increased. The collaborative utility of each equipment is obtained by multiplying each dimension of the fusion feature vector with its corresponding weight and summing them up through a weighted summation method. This allows for a quantifiable comparison of the competitiveness between equipment, ensuring that tasks are assigned to the most suitable equipment, improving the overall production efficiency of the production line. The collaborative utility results serve as the basis for equipment bidding, driving equipment to formulate reasonable bidding strategies based on its own capabilities to win the right to take over tasks.

[0089] Specifically, the output logic of the scheduling instruction includes:

[0090] The device that receives the subcontracting request formulates a bidding strategy based on the collaborative utility and dynamically adjusts its bidding price according to its own device status through the simulated annealing algorithm.

[0091] The equipment to be taken over is determined based on the bidding price, and initial scheduling instructions are generated. At the same time, the execution status of the tasks is monitored. When a task is detected to violate the equipment capacity constraints, a partial adjustment of the scheduling instructions is triggered.

[0092] During the production of melamine tableware, the equipment status changes dynamically. For example, after continuous operation of the injection molding machine, energy consumption increases and precision decreases. Equipment needs to adjust its bidding price based on its real-time status to reasonably undertake tasks while avoiding excessive competition that could lead to reduced production efficiency or equipment damage. Equipment receiving subcontracting requests formulates an initial bidding price based on its own collaborative utility. The bidding price can be linked to factors such as task completion time and energy costs. Through simulated annealing algorithms, the bidding price is dynamically adjusted based on status parameters such as the equipment's current health index and remaining capacity. When the equipment's health index is low, the bid is appropriately increased to reduce the amount of tasks undertaken and prevent overloading. This allows equipment to flexibly participate in task competition based on its own status, ensuring normal equipment operation and achieving reasonable task allocation. The bidding price becomes a key basis for determining which equipment to take over, directly affecting the generation of scheduling instructions and the result of task allocation.

[0093] The process involves selecting the most suitable takeover equipment from among numerous bidding devices, generating scheduling instructions, clarifying task allocation plans, and ensuring orderly production. It also involves comparing the bid prices of various devices, selecting the one with the best bid as the takeover equipment, generating initial scheduling instructions (including task content, execution time, and process parameters), and issuing these instructions to the takeover equipment. This process determines the optimal takeover equipment and generates scheduling instructions, ensuring reasonable task allocation, smooth production processes, and improved production efficiency. Simultaneously, it monitors the execution status of tasks, and immediately triggers local adjustments to the scheduling instructions if any violation of equipment capacity constraints is detected.

[0094] During the production of melamine tableware, unexpected situations such as raw material supply delays and sudden equipment failures may occur, causing task execution to violate equipment capacity constraints. To ensure production targets are met, scheduling instructions need to be partially adjusted, and task allocation re-optimized. When a task violates equipment capacity constraints, the cause of the violation is analyzed, including insufficient injection molding machine pressure leading to poor product molding. The collaborative effectiveness of adjacent equipment is reassessed, and the task allocation scheme is adjusted based on the new collaborative effectiveness results, generating partially adjusted scheduling instructions. Specifically, some injection molding tasks are subcontracted to other idle injection molding machines, or the time parameters of the curing stage are adjusted to compensate for delays in the injection molding stage. This effectively addresses unexpected situations during production, reduces production downtime, ensures on-time product delivery, and guarantees the smooth progress of the production plan. The adjusted scheduling instructions redistribute tasks, triggering equipment capacity performance updates and causal graph network correlation analysis again, driving the control feedback module to further optimize the system.

[0095] The control feedback module is used to receive scheduling instructions and control the takeover equipment, locate downstream causal devices by combining the causal graph network, monitor the update changes of the takeover equipment's capability performance to incrementally update the causal graph network, and rebroadcast causal influence factors.

[0096] Furthermore, the positioning sub-logic for downstream devices includes:

[0097] Multi-level analysis of the causal graph network is performed using graph attention network to identify the device nodes associated with the takeover device node, so as to form a causal propagation path;

[0098] For each causal propagation path, the probability distribution and risk value of downstream devices being affected are calculated to form a risk propagation matrix, and devices with risk values ​​greater than the risk threshold are selected as causal downstream devices.

[0099] A secondary screening of downstream devices based on task attributes is performed to locate downstream devices.

[0100] The production processes of melamine tableware are closely interconnected. An malfunction in the injection molding equipment can affect downstream pressing and polishing processes through the material transport chain. A single-dimensional analysis is insufficient to fully grasp the causal relationships between equipment. Therefore, a multi-level analysis of the causal graph network using graph attention networks is necessary to identify the complete causal propagation path associated with the receiving equipment, ensuring that no potentially affected equipment is overlooked. This multi-level analysis assigns attention weights to different equipment nodes and edges, prioritizing nodes closely related to the receiving equipment. Starting with directly adjacent equipment, it expands outwards to identify first-, second-, and third-level related equipment, forming causal propagation paths with varying degrees of impact. If the injection molding machine becomes the receiving equipment, the pressing equipment that directly receives the molded parts is identified first, followed by the polishing equipment affected by the pressing equipment, constructing a complete propagation chain. This comprehensive identification of causal propagation paths allows for the prediction of potential impacts of equipment scheduling on other parts of the production line, preventing chain reactions caused by local adjustments and ensuring the continuity of the production process.

[0101] Different equipment in a production line have varying degrees of importance and resilience. Simple screening cannot accurately pinpoint the truly critical downstream equipment requiring attention. By calculating the probability distribution and risk value of downstream equipment being affected, a risk propagation matrix is ​​constructed. This quantifies the risks faced by each piece of equipment, allowing for the identification of high-risk equipment and prioritizing appropriate countermeasures. Along each causal propagation path, the probability distribution of downstream equipment being affected is calculated by combining current equipment status, historical fault data, and the closeness of process connections. Factors such as production interruption losses and product quality degradation due to equipment failure are comprehensively considered to determine the risk value for each piece of equipment, forming a risk propagation matrix. A reasonable risk threshold is set, and equipment with risk values ​​exceeding the threshold is selected as initial causal downstream equipment. This allows for precise identification of high-risk equipment, enabling production managers to concentrate resources on monitoring and maintaining key equipment, improving resource utilization efficiency, reducing overall production risk, and providing data for secondary screening. Further analysis of task attributes helps determine the equipment requiring final focus.

[0102] Different tasks have varying degrees of dependence on equipment and quality requirements. Downstream causal devices located solely based on risk values ​​may not be closely related to the current task. By combining task attributes with secondary screening, we can ensure that the located equipment is indeed critical to the smooth execution of the current task, improving the accuracy of equipment location. Task attributes such as process requirements, delivery time, and quality standards are extracted from scheduling instructions. For initially screened downstream causal devices, their matching degree with task attributes is analyzed. For tasks requiring high-precision patterns, the focus is on pressing and grinding equipment, while for urgent orders, equipment that affects delivery time is prioritized. The devices are ranked according to their matching degree to ultimately determine the downstream causal devices. This secondary screening based on task attributes makes the location of downstream causal devices more aligned with actual production needs, avoiding ineffective monitoring and resource waste, ensuring timely and high-quality task completion. Accurately located downstream causal devices provide a clear direction for incremental updates of the causal graph network, identifying the parts that need to be analyzed and updated.

[0103] Furthermore, such as Figure 3 As shown, the incremental update sub-logic of the causal graph network includes:

[0104] After the takenover equipment executes the scheduling command, the system monitors the changes in the capabilities of the takenover equipment in real time to generate event codes.

[0105] By comparing the differences between event codes and historical codes, changes in device state caused by the execution of scheduling instructions are identified. The causal relationship between devices and tasks is re-analyzed through a causal discovery algorithm. At the same time, affected causal edges and nodes are marked to form a local perturbation map.

[0106] The local perturbation map is fused with the causal graph network through a graph neural network to incrementally update the causal graph network.

[0107] After a takenover device executes a scheduling command, its operating status changes. This change affects other devices through causal relationships. Real-time monitoring of the updated performance of the takenover device and generating event codes allows for the timely capture of dynamic information about the device, providing data support for subsequent analysis of changes in the causal relationship between the device and the task. Real-time acquisition of the takenover device's process set, task completion time, health index, and energy consumption, along with data encoding technology, converts the acquired event data into a unified format event code. This event code contains key information such as device identification, timestamp, and status parameters, facilitating subsequent comparison and analysis with historical codes. This enables timely reflection of changes in the device's performance, providing a basis for quickly identifying changes in device status caused by the execution of scheduling commands and improving the response speed to production changes.

[0108] Updates to equipment capabilities can alter the causal relationships between existing equipment and tasks. For example, adjusting injection molding machine pressure can affect product molding time, thus changing its time correlation with subsequent polishing processes. By reanalyzing causal relationships and marking affected causal edges and nodes, the impact of scheduling command execution on the production line's causal structure can be clearly presented, forming a local perturbation graph. By comparing the differences between current event codes and historical codes, changes in equipment state caused by scheduling command execution can be identified. Using causal discovery algorithms and combining time-series data before and after equipment state changes, the causal relationships between equipment and tasks are reanalyzed. For parts where causal relationships have changed, corresponding causal edges and nodes are marked, and the type and degree of causal relationship changes are recorded. These affected parts are integrated to form a local perturbation graph, visually displaying the causal relationship changes caused by scheduling command execution. The local perturbation graph clearly presents changes in causal relationships, helping production managers and the system understand changes in production state after scheduling command execution, and providing a clear direction for subsequent causal graph network updates.

[0109] The existing causal graph network can no longer accurately reflect the causal relationships of the current production line. It is necessary to fuse the local disturbance graph with the causal graph network in terms of topology, enabling the causal graph network to adapt to production changes in a timely manner and provide a more accurate basis for subsequent scheduling decisions. A graph neural network is used to integrate information such as new causal relationships and changed edge weights from the local disturbance graph into the causal graph network. During the fusion process, the attributes of nodes and edges in the causal graph network are adjusted according to the importance and magnitude of changes in causal relationships. Newly emerging strong causal relationships are assigned higher weights, while weakened causal relationships are weighted lower. Through iterative calculation, it is ensured that the fused causal graph network accurately reflects the causal structure of the current production line. The incrementally updated causal graph network reflects production changes in real time, supporting subsequent calculations of causal influence factors and equipment capacity prediction, improving the adaptability and decision accuracy of the scheduling system. The updated causal graph network is the basis for the rebroadcast of causal influence factors; its structure and weight changes determine the broadcast priority and frequency adjustment of each causal influence factor.

[0110] Specifically, the rebroadcast logic of causal influencing factors includes:

[0111] The incrementally updated causal graph network is processed using a convolutional neural network and an attention mechanism to output a weight matrix;

[0112] The reward function for reinforcement learning is based on the production line's operating efficiency, energy consumption cost, and task delay risk. Broadcast priority and broadcast frequency are assigned to causal influencing factors according to the weight matrix.

[0113] The causal influence factor is rebroadcast to each subgraph based on the assigned broadcast priority and broadcast frequency.

[0114] After incremental updates to the causal graph network, the causal relationships between devices and tasks change, and the degree of influence of each causal relationship on production scheduling also changes accordingly. The updated network is processed by a convolutional neural network and an attention mechanism to output a weight matrix, which quantifies the importance of each causal relationship and provides a basis for prioritizing causal influencing factors. The incrementally updated causal graph network is converted into a matrix form and input into a convolutional neural network for feature extraction, capturing local structural features in the network. Then, an attention mechanism is used to assign attention weights to different causal relationships, highlighting important causal relationships. Finally, a weight matrix is ​​output, where each element corresponds to a weight value for a causal relationship. The larger the weight value, the more important the causal relationship is to the scheduling under the current production state. The weight matrix accurately quantifies the importance of causal relationships, enabling the scheduling system to rationally allocate resources and priorities according to actual production needs, improving the scientific nature of scheduling decisions. The weight matrix is ​​a key input for assigning broadcast priorities and frequencies based on reinforcement learning, directly affecting the propagation strategy of causal influencing factors.

[0115] The operating status and scheduling objectives of the production line are constantly changing, including adjustments to order urgency and capacity changes due to equipment failures. Static broadcast priorities and frequencies cannot meet real-time scheduling needs. Using production line operating efficiency, energy costs, and task delay risks as reward functions, reinforcement learning dynamically allocates broadcast priorities and frequencies, enabling the propagation of causal factors to better adapt to production changes. If increasing the broadcast frequency of important causal factors improves production line efficiency, a positive reward is given; if excessive broadcasting leads to wasted communication resources without improving production efficiency, a negative reward is given. Through continuous interaction with the production environment, reinforcement learning can learn the optimal allocation of broadcast priorities and frequencies. Based on the learning results, a corresponding broadcast priority and frequency are assigned to each causal factor. This allows for the rational arrangement of the propagation of causal factors according to actual production conditions, ensuring timely transmission of important information and improving the response speed and resource utilization efficiency of production scheduling.

[0116] The production line for melamine tableware has different needs for causal influencing factors in different functional areas. The injection molding area focuses more on the causal relationship related to raw materials and molding. By rebroadcasting causal influencing factors to each subgraph according to the assigned broadcast priority and frequency, accurate information delivery can be achieved, avoiding interference from invalid information and improving information transmission efficiency. Based on the functional area division of the production line, the causal graph network is divided into multiple subgraphs. For each subgraph, the corresponding causal influencing factors are pushed to the equipment within the subgraph according to the assigned broadcast priority and frequency. For high-priority causal influencing factors, multicast or broadcast is used to ensure rapid propagation, while for low-priority factors, the propagation frequency is appropriately reduced to reduce communication resource consumption. At the same time, the reception and processing of causal influencing factors by the equipment within the subgraph is monitored to ensure effective information transmission. Hierarchical subgraph rebroadcasting achieves accurate propagation of causal influencing factors, improves the targeting and efficiency of information transmission, and enables equipment in each area to obtain key information relevant to itself in a timely manner, better supporting scheduling decisions and production execution. The rebroadcast causal influencing factors provide updated information to the causal perception module and intelligent scheduling module, promoting continuous optimization of scheduling strategies and forming a closed-loop feedback mechanism.

[0117] Example 2

[0118] like Figure 4 The diagram shown is a flowchart of a multi-device collaborative control method for a production line provided in this application embodiment. The method includes:

[0119] Acquire event data from the production line, including equipment status, task status, and abnormal events. Use causal discovery algorithms to identify causal relationships between equipment and tasks to generate a causal graph network for the production line and broadcast causal influencing factors.

[0120] Receive causal influencing factors, predict the device's performance based on the causal influencing factors using a time-series prediction algorithm, and map it to device capability constraints;

[0121] When a task is detected to violate the device's capability constraints, a sub-packet request is broadcast to neighboring devices. The devices that receive the sub-packet request determine the effectiveness of the collaboration based on their own capabilities.

[0122] The takeover device is determined based on collaborative utility bidding, scheduling instructions are output, and updates to the capabilities of the takeover device are triggered.

[0123] It receives scheduling instructions and controls the takeover equipment, locates downstream causal devices by combining the causal graph network, monitors the update changes in the capability performance of the takeover equipment to incrementally update the causal graph network, and rebroadcasts causal influencing factors.

[0124] Since the principle of the method in this application embodiment is similar to that of the system described in this application embodiment, the implementation of the method is the same as that of the system, and the repeated parts will not be described again.

Claims

1. A production line intelligent scheduling system, characterized in that, include: Acquire event data from the production line, including equipment status, task status, and abnormal events. Use causal discovery algorithms to identify causal relationships between equipment and tasks to generate a causal graph network for the production line and broadcast causal influencing factors. The broadcast logic of the causal influencing factors includes: Based on the edge weights and node attributes of the causal graph network, the causal influence factor of each causal relationship is determined by the fuzzy comprehensive evaluation method according to the importance of the equipment, the urgency of the task, and the severity of the anomaly. The event data of the production line is analyzed by using a long short-term memory network to predict the importance of different causal factors in the scheduling process, and the broadcast priority and broadcast frequency of causal factors are dynamically adjusted according to the prediction results. The causal graph network is divided into multiple subgraphs, each representing a functional area of ​​the production line, and causal influence factors are broadcast to each subgraph based on broadcast priority and broadcast frequency; The generation sub-logic of the cause-effect graph network of the production line includes: Acquire event data from the production line, align event data of different frequencies using a dynamic time warping algorithm, and standardize the event data. By combining the causal discovery algorithm with the temporal properties of the production line, the causal relationship between equipment and tasks is identified and analyzed to form a candidate set of causal relationships; Using the production line's operating efficiency and scheduling optimization objectives as reward functions, a deep Q-network is used to iteratively filter and adjust the weights of the causal relationship candidate set to generate a causal graph network for the production line. The system receives causal influencing factors, uses a time-series prediction algorithm to predict the device's performance based on these factors, and maps them to device capability constraints. When a task is detected to violate the device capability constraints, the system broadcasts a sub-packet request to neighboring devices. Devices receiving the sub-packet request determine the cooperation utility based on their own performance, bid on the cooperation utility to determine the takeover device, output scheduling instructions, and trigger an update of the takeover device's performance. It receives scheduling instructions and controls the takeover equipment, locates downstream causal devices by combining the causal graph network, monitors the update changes in the capability performance of the takeover equipment to incrementally update the causal graph network, and rebroadcasts causal influencing factors.

2. The intelligent production line scheduling system as described in claim 1, characterized in that, The output logic of the scheduling instruction includes: The device that receives the subcontracting request formulates a bidding strategy based on the collaborative utility and dynamically adjusts its bidding price according to its own device status through the simulated annealing algorithm. The equipment to be taken over is determined based on the bidding price, and initial scheduling instructions are generated. At the same time, the execution status of the tasks is monitored. When a task is detected to violate the equipment capacity constraints, a partial adjustment of the scheduling instructions is triggered.

3. The intelligent production line scheduling system as described in claim 2, characterized in that, The prediction sub-logic for the device's performance capabilities includes: The causal influencing factors are sampled by sliding window through multi-scale time windows, and the spatiotemporal features of the causal graph network are extracted based on the spatiotemporal graph convolutional network. The spatiotemporal features include equipment status trends, task association features and anomaly propagation features. By using Bayesian deep learning, the device's performance is probabilistically predicted based on spatiotemporal features. The output includes a predicted distribution with confidence intervals. Simultaneously, the perturbation of causal influencing factors is simulated during the prediction process, and the marginal effect of causal influencing factors on the device's performance is quantified. The equipment's performance includes process set, task completion time, health index, and energy consumption. Mapping the equipment's performance to the constraint space forms the equipment's capability constraints.

4. The intelligent production line scheduling system as described in claim 3, characterized in that, The sub-logic for determining the effectiveness of collaboration includes: Upon receiving a subcontracting request, the device extracts features of its own capabilities based on a self-attention mechanism to obtain device capability features. The task attributes are parsed from the sub-packet request, and the device capability characteristics are dynamically fused with the task attributes to generate a fused feature vector. The collaborative utility of each device that received the subpacket request is obtained by weighted summation of the fused feature vectors.

5. The intelligent production line scheduling system as described in claim 4, characterized in that, The rebroadcast logic of the causal influencing factors includes: The incrementally updated causal graph network is processed using a convolutional neural network and an attention mechanism to output a weight matrix; The reward function for reinforcement learning is based on the production line's operating efficiency, energy consumption cost, and task delay risk. Broadcast priority and broadcast frequency are assigned to causal influencing factors according to the weight matrix. The causal influence factor is rebroadcast to each subgraph based on the assigned broadcast priority and broadcast frequency.

6. The intelligent production line scheduling system as described in claim 5, characterized in that, The location sub-logic of the downstream device includes: Multi-level analysis of the causal graph network is performed using graph attention network to identify the device nodes associated with the takeover device node, so as to form a causal propagation path; For each causal propagation path, the probability distribution and risk value of downstream devices being affected are calculated to form a risk propagation matrix, and devices with risk values ​​greater than the risk threshold are selected as causal downstream devices. A secondary screening of downstream devices based on task attributes is performed to locate downstream devices.

7. The intelligent production line scheduling system as described in claim 6, characterized in that, The incremental update sub-logic of the causal graph network includes: After the takenover equipment executes the scheduling command, the system monitors the changes in the capabilities of the takenover equipment in real time to generate event codes. By comparing the differences between event codes and historical codes, changes in device state caused by the execution of scheduling instructions are identified. The causal relationship between devices and tasks is re-analyzed through a causal discovery algorithm. At the same time, affected causal edges and nodes are marked to form a local perturbation map. The local perturbation map is fused with the causal graph network through a graph neural network to incrementally update the causal graph network.

8. A method for collaborative control of multiple devices on a production line, implemented based on an intelligent scheduling system for a production line according to any one of claims 1-7, characterized in that, include: Acquire event data from the production line, including equipment status, task status, and abnormal events. Use causal discovery algorithms to identify causal relationships between equipment and tasks to generate a causal graph network for the production line and broadcast causal influencing factors. Receive causal influencing factors, predict the device's performance based on the causal influencing factors using a time-series prediction algorithm, and map it to device capability constraints; When a task is detected to violate the device's capability constraints, a sub-packet request is broadcast to neighboring devices. The devices that receive the sub-packet request determine the effectiveness of the collaboration based on their own capabilities. The takeover device is determined based on collaborative utility bidding, scheduling instructions are output, and updates to the capabilities of the takeover device are triggered. It receives scheduling instructions and controls the takeover equipment, locates downstream causal devices by combining the causal graph network, monitors the update changes in the capability performance of the takeover equipment to incrementally update the causal graph network, and rebroadcasts causal influencing factors.

Images