Adhesive tape production line production scheduling system based on intelligent optimization

Abstract

The invention belongs to the technical field of production scheduling systems, and particularly relates to an adhesive tape production line production scheduling system based on intelligent optimization, which comprises a sensing layer for monitoring operation data of equipment through sensing equipment, sorting the equipment data into an equipment data packet and uploading the equipment data packet to a control center, the equipment data packet comprises a production state, an operation state, process parameters and material stock; static specifications, dynamic performance and plan constraints of equipment are converted into mathematical constraints which can be identified by a scheduling algorithm, and an equipment processing boundary is determined through an adaptive matrix, so that the problems that large equipment is allocated to small-specification orders or low-precision equipment is allocated to high-precision demands, the equipment specification mismatching rate is reduced, and the rework rate is reduced due to insufficient equipment capability are avoided; moreover, productivity expectation is dynamically corrected through an attenuation curve, progress delay caused by scheduling based on theoretical values is avoided, scheduling progress deviation is reduced through productivity attenuation correction, and the order on-time delivery rate is increased.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of production scheduling system, and in particular to a rubber belt production line production scheduling system based on intelligent optimization. BACKGROUND

[0002] The rubber belt production line production scheduling system is a comprehensive system integrating automation control, data acquisition, intelligent algorithm and production management logic, and the core goal is to realize optimal allocation of production resources, improve production line efficiency, reduce cost and guarantee delivery under the constraints of meeting order demand, equipment capacity, material supply, etc. The manufacturing system is essentially dynamic, and tasks and resources will change due to factors such as equipment failure, maintenance, demand fluctuations and changing market conditions. In addition, the rapid development of technological innovation and the increasing integration of equipment in the manufacturing process lead to the complexity and unpredictability of the production environment.

[0003] If customer demand shows characteristics of diversification and individualization, the production line needs to cope with multi-variety and small-batch production mode. In this mode, the amount of information data in the production process is huge, the data structure is complex and has strong randomness, making control difficult. Moreover, these factors require the production system to have high flexibility and real-time performance, making the gap between traditional scheduling theory and reality larger, and it is difficult to adapt to the scheduling needs of actual production lines. Moreover, the advanced production planning and scheduling system currently in use only deals with static scheduling problems. In the face of complex and dynamic environments, the usual solution is to split it into multiple static problems for rescheduling, which is slow in response and inefficient in actual application scenarios. Often, manual experience needs to be added to adaptively adjust, but manual adjustment depends on the experience and ability of the scheduling personnel, and as the complexity of the scheduling problem deepens, its quality and stability are difficult to guarantee. Moreover, manual adjustment often consumes a lot of time and manpower, and is low in intelligence and low in production efficiency. SUMMARY

[0004] In order to make up for the deficiencies of the prior art and solve the above technical problems, the application provides a rubber belt production line production scheduling system based on intelligent optimization.

[0005] The technical scheme adopted by the application to solve the technical problems is that the application provides a rubber belt production line production scheduling system based on intelligent optimization, which comprises:

[0006] The perception layer: monitor each device running data through the perception device, and upload the device data package including production state, running state, process parameters and material inventory to the control center.

[0007] Cloud edge collaborative layer: build a constraint condition library based on device data packets, which includes device constraint sub-library, material constraint sub-library and order constraint sub-library;

[0008] Scheduling layer: build a system for workshop scheduling based on deep reinforcement learning, including device unit, resource unit, scheduling unit and execution unit, device unit includes detection device of production equipment, resource unit includes past order scheduling records and scheduling logic, scheduling unit includes initial scheduling scheme transmitted by scheduling system, and execution unit includes transmission device and display device to notify and control the cooperation between production equipment;

[0009] Execution layer: implant a multi-target task scheduling system to evenly distribute order tasks into the scheduling system of the workshop;

[0010] Interaction layer: build a monitoring and early warning module, the detection data of the detection device in the scheduling layer is used as the feature extraction value, and the parameters of the adhesive tape product are used as the preset threshold value. When the deviation value exceeds the preset threshold value, the dynamic adjustment mechanism is triggered to update the scheduling scheme.

[0011] Preferably, the constraint condition library includes: the device constraint sub-library includes the equipment specification adaptation matrix, the capacity attenuation function curve and the downtime maintenance time window period, which is constructed based on historical operation data of the equipment; the material constraint sub-library is associated with the data of the material storage tray in the equipment, and the inventory, storage location and shelf life of the material are updated in real time, and multiple levels of inventory early warning threshold values are set; the order constraint sub-library uses the analytic hierarchy process to divide the priority, and the priority factors include the delivery time urgency, order amount and customer level.

[0012] Preferably, the system for workshop scheduling includes a workshop scheduling environment, an offline training module and an online application module:

[0013] Workshop scheduling environment: model the workshop scheduling problem as a Markov decision process, then use the disjunctive graph model to split the scheduling task into multiple nodes, and use the Gantt chart model to store the processing information matrix;

[0014] Offline training module: by continuously interacting the agent with the environment, the generated reinforcement learning four-tuple data is stored in the storage component, and for the production line scene, a deep reinforcement learning algorithm is selected from the algorithm pool to train the agent; the loss function is calculated continuously using the data of agent interaction, and the network weight of the agent is updated using the gradient descent algorithm until the network converges; finally, the trained network model and weight are saved;

[0015] Online application module: load the trained network model into the agent, then input the state of the actual workshop environment into the agent, the agent outputs the corresponding scheduling action through the network decision, then the workshop executes this scheduling action and updates to the next state, and the cycle continues until the scheduling task is completed.

[0016] Preferably, the disjunctive graph reflects the processing time of each process and the specific time scale of the entire scheduling scheme by means of the Gantt chart, and the scheduling problem is modeled into a discrete sequential decision process by means of the cooperation of the disjunctive graph and the Gantt chart, and the state transition function, the action space, the reward function and the agent network structure are designed respectively for the sequential decision process.

[0017] Preferably,

[0018] The state transition function contains three kinds of state information, processing time, processing mark and cumulative processing time;

[0019] The action space contains two parts of process order and machine selection, and is used for encoding the action space;

[0020] The reward function is a short-term return obtained by performing an action in the current state;

[0021] The agent network structure contains a feature extraction network and a decision network.

[0022] Preferably, the offline training module trains the agent by using a proximal policy optimization algorithm, uses two same agent networks in the algorithm architecture, one network is used for sampling, and the other network is used for repeated updating, so that the policy iteration gradually converges to the optimal policy.

[0023] Preferably, the multi-objective task scheduling system comprises a data collection and preprocessing module, a multi-objective optimization module, a cloud edge resource allocation module and a task scheduling algorithm.

[0024] A tape production line production scheduling method based on intelligent optimization, the specific steps are as follows:

[0025] Step 1: Collect the full-process data of the tape production through the industrial Internet of Things equipment, complete the multi-dimensional data collection and preprocessing work, including equipment data, process data, material data and order data;

[0026] Step 2: Build a dynamic constraint condition library based on the preprocessed data, including a device constraint sub-library, a material constraint sub-library and an order constraint sub-library;

[0027] Step 3: Define the core optimization target of scheduling, construct a mathematical model, and take minimizing the total completion time, reducing energy consumption and reducing production cost as the core target, determine the weight of each target by entropy weight method, and integrate the device, material and order constraints in step 2 into mathematical expressions, and convert the scheduling problem into a quantifiable mathematical optimization problem, balance the conflict between efficiency, cost and service quality;

[0028] Step 4: Generate an initial scheduling scheme by algorithm, select a comprehensive optimal solution from the optimization solution, and generate information including devices, processes and time;

[0029] Step 5: Implement the scheduling scheme through the industrial control system and track the execution status in real time.

[0030] Step 6: Extract features such as device load, material inventory, and order progress through a heterogeneous graph neural network, calculate the deviation between actual and planned values, and real-time feedback data to identify deviations and trigger a dynamic adjustment mechanism.

[0031] The beneficial effects of the present application are as follows:

[0032] 1. The intelligent optimization-based adhesive tape production line production scheduling system converts the static specifications, dynamic performance, and planning constraints of the equipment into mathematical constraints recognizable by the scheduling algorithm, clearly defines the equipment processing boundaries through an adaptation matrix, avoids assigning small-size orders to large equipment or high-precision requirements to low-precision equipment, reduces the mismatch rate of equipment specifications, and reduces the rework rate caused by insufficient equipment capacity; and, through a decay curve to dynamically correct the expected production capacity, avoid schedule delays based on theoretical values, and reduce the schedule progress deviation through production capacity decay correction, improving the on-time delivery rate of orders.

[0033] 2. The intelligent optimization-based adhesive tape production line production scheduling system solves the efficiency and cost balance problem through the multi-objective task allocation system of the execution layer, avoids resource waste caused by the single target of fastest completion; the interactive layer extracts features such as the deviation of actual coating thickness from the standard value and compares it with the threshold value, and when the deviation exceeds the limit, triggers the scheduling unit to recalculate: if the deviation is caused by material shortage, the material constraint sub-library is updated to warn of inventory, and the task order of the execution layer is adjusted; if the deviation is caused by equipment aging, the production capacity decay curve of the equipment constraint sub-library is called to correct the subsequent schedule; the monitoring and warning module of the interactive layer analyzes the real-time deviation, such as process parameter deviation, triggers dynamic adjustment, compensates for the rigidity of traditional scheduling, and ensures production stability. BRIEF DESCRIPTION OF DRAWINGS

[0034] The present application will be further described below with reference to the accompanying drawings.

[0035] Figure 1 is a system flowchart of the present application. DETAILED DESCRIPTION

[0036] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the accompanying drawings shown in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0037] Embodiment One:

[0038] To effectively solve the above problems, as shown in the accompanying drawings Figure 1 The scheduling system comprises:

[0039] The perception layer: monitor the operation data of each device through the perception equipment, and upload the device data package to the control center, the device data package including the production state, the running state, the process parameters and the material inventory;

[0040] The cloud edge collaborative layer: build a constraint condition library based on the device data package, the constraint condition library including a device constraint sub-library, a material constraint sub-library and an order constraint sub-library;

[0041] The scheduling layer: build a system for workshop scheduling based on deep reinforcement learning, including a device unit, a resource unit, a scheduling unit and an execution unit, the device unit including a detection device of the production device, the resource unit including the past order scheduling records and the scheduling logic, the scheduling unit including the initial scheduling scheme transmitted by the scheduling system, and the execution unit including a transmission device and a display device to notify and control the cooperation between the production devices;

[0042] The execution layer: implant a multi-target task scheduling system to evenly distribute the order tasks into the scheduling system of the workshop;

[0043] The interaction layer: build a monitoring and early warning module, the detection data of the detection device in the scheduling layer as the feature extraction value, and the parameters of the adhesive tape product as the preset threshold value, when the deviation value exceeds the preset threshold value, the dynamic adjustment mechanism is triggered to update the scheduling scheme;

[0044] The constraint condition library includes: the device constraint sub-library including the device specification adaptation matrix, the capacity attenuation function curve and the downtime maintenance time window period, which is constructed based on the historical operation data of the device; the material constraint sub-library is associated with the data of the material storage tray in the device, and the inventory, the storage location and the validity period of the material are updated in real time, and multi-level inventory early warning threshold is set; the order constraint sub-library adopts the analytic hierarchy process to divide the priority, and the priority factors include the delivery time urgency, the order amount and the customer level;

[0045] Specifically: the device, material, order data collected by the perception layer is processed by the cloud-edge collaboration layer and converted into dynamic parameters of the constraint condition library, such as real-time production capacity of the device, material inventory threshold, etc.; the deep reinforcement learning model of the scheduling layer generates an optimized solution through state perception, action decision, and reward feedback based on these constraints; the execution layer decomposes the solution into device instructions and monitors deviations in real time through the interaction layer to form a closed loop of data collection, modeling, decision-making, execution, and feedback; the perception layer solves the problem of data silos in traditional scheduling through multi-dimensional data collection, production status, operating parameters, and material inventory; because the continuity of tape production is strong, such as coating, slitting, and rewinding, which require consecutive operations, and process parameters, such as glue viscosity and coating thickness, have a great impact on quality, high-frequency data collection is required to ensure data timeliness, and data is integrated into standardized data packets to provide unified input for subsequent constraint modeling and intelligent decision-making, avoiding scheduling deviations caused by data format chaos;

[0046] Further, the multi-objective task allocation system of the execution layer solves the efficiency and cost balance problem and avoids resource waste caused by the single objective of the fastest completion; the interaction layer extracts features such as the deviation of the actual coating thickness from the standard value and the threshold value, and when the deviation exceeds the limit, it triggers the scheduling unit to recalculate: if the deviation is caused by material shortage, it updates the inventory warning of the material constraint sub-library and adjusts the task order of the execution layer; if it is caused by device aging, it calls the production capacity decay curve of the device constraint sub-library to correct the subsequent scheduling; the monitoring and warning module of the interaction layer analyzes the real-time deviation, such as process parameter deviation, triggers dynamic adjustment, and makes up for the rigidity of one-time scheduling in traditional scheduling to ensure production stability;

[0047] Through the cooperation of the perception layer, the constraint condition library, and the interaction layer, the production capacity decay state data curve is updated in real time, and the device constraint is changed from a static assumption to a dynamic adaptation, for example, after the coating machine runs for 1000 hours, the system automatically corrects its production speed through the running feature extraction of the interaction layer, avoiding scheduling deviations caused by production capacity misjudgment;

[0048] By setting the constraint condition library, the perception layer perceives the device through devices such as controllers and tension sensors to collect static specification data of the device, and through detection devices such as tape coating detectors to collect dynamic running data of the device, and based on historical device data packets to obtain historical performance data or production capacity change curves, including but not limited to:

[0049] Device constraint sub-library:

[0050] Data collection: use Python Pandas library to clean data, eliminate sensor outliers, normalize, and fill in missing maintenance data through time series analysis;

[0051] Constraint library construction: build a Boolean matching matrix with process, equipment, and product specifications as three-dimensional dimensions. Update based on historical equipment data packets through association rule mining;

[0052] Use linear regression or gradient boosting tree model to fit the decay curve with runtime as the independent variable and the ratio of actual capacity to theoretical capacity as the dependent variable , Run hours, recalibrate based on new data every quarter;

[0053] Integrate preventive maintenance plan and real-time fault warning data to build a time axis constraint. Mark the equipment unavailable period in the equipment constraint sub-library. When the equipment vibration value exceeds the threshold, automatically trigger the maintenance window in advance;

[0054] Cloud-edge collaborative update: edge computing nodes synchronize real-time data every preset period. If the capacity decay rate exceeds the limit or the maintenance plan changes, update the constraint library immediately;

[0055] In traditional scheduling, allocating equipment based on experience can lead to specification mismatch, capacity misjudgment, and plan conflict. Therefore, through the cooperation of constraint library and scheduling layer, based on data-driven equipment capacity, the static specifications, dynamic performance, and plan constraints of the equipment are converted into mathematical constraints recognizable by the scheduling algorithm. Through the adaptation matrix, the processing boundaries of the equipment are clearly defined, avoiding the allocation of small-size orders to large equipment or the allocation of high-precision requirements to low-precision equipment. The specification mismatch rate of the equipment is reduced, and the rework rate caused by insufficient equipment capacity is reduced. Moreover, through the decay curve, the expected capacity is dynamically corrected, avoiding schedule delays caused by theoretical values. The capacity decay correction reduces the schedule deviation, and the on-time delivery rate of orders is improved. Furthermore, by locking the unavailable period in advance, the time conflict between production and maintenance is avoided, and the maintenance window is pre-integrated to reduce unplanned downtime and improve equipment utilization, thereby improving the scheduling efficiency of the scheduling system of the tape production line;

[0056] Material constraint sub-library:

[0057] Data collection: use RFID readers of intelligent racks, weight sensors of material conveying lines, etc. to collect full-chain data of core materials for tape production; use MQTT protocol to push RFID data to cloud-edge collaboration layer every 10 seconds, and trigger inventory real-time update when materials are taken or stored;

[0058] Constraint library construction: based on material attributes and product specifications as dimensions, build a correlation table, update the adaptive relationship based on quality inspection data; based on production consumption rate / purchasing cycle / safety factor, set multi-level inventory thresholds of safety, early warning and emergency; quantify the time consumption and loss of different material switching to provide basis for order consolidation scheduling; when the inventory is lower than the corresponding threshold, the system automatically pushes the message, if the inventory of a certain material is lower than the emergency threshold, the scheduling system automatically queries the adaptive table, prioritizes production using another order, and at the same time, extends the start time of the material related order until the material is replenished;

[0059] When the material information lag, supply and demand dislocation problem occurs in traditional scheduling, based on real-time sensing, early warning linkage and demand matching, the inventory status, adaptability and supply cycle of the material are converted into scheduling constraints to avoid the situation of lack of materials caused by the difference between the account and the actual situation, and to realize the change from passive response to active prevention, to resolve supply risks in advance, and to ensure that the material quality meets the product requirements and avoid quality defects caused by material mismatch;

[0060] Order constraint sub-library:

[0061] Data collection: import order core information from ERP system, extract three priority factors including delivery time urgency, order amount and customer level, and normalize the factor value by range method;

[0062] Order priority weight allocation: compare the importance of factors according to the actual production needs, build a judgment matrix, and calculate the factor weight by eigenvalue method, for example: delivery time urgency score 40%, order amount score 30%, customer level score 30%, order sequence is formed according to the score;

[0063] Order association with scheduling layer: the scheduling system allocates resources to high-score orders first, when new orders are inserted, the priority is automatically recalculated, and if it is an urgent order, the local rearrangement of existing scheduling is triggered;

[0064] In traditional scheduling, the problem of unordered insertion of orders and unbalanced allocation of resources occurs, through the effect of multi-criteria quantization, scientific sorting and dynamic adaptation, the subjective order priority is converted into objective numerical constraint, realizing the optimal allocation of resources while adapting to the variability of multi-variety and small-batch production demand, and quickly responding to order changes;

[0065] Further, the device, material and order sub-libraries do not operate in isolation, but form a closed loop through data interconnection, constraint linkage and target cooperation, improve the effect of cloud edge cooperation layer assisting scheduling layer to optimize scheduling scheme, and can also cope with temporary and variable production scheduling requirements, thereby improving the practicality of the scheduling system;

[0066] The detection device comprises a reciprocating sliding mechanism, a bearing platform, a laser thickness gauge and a camera assembly. The bearing platform is installed on the reciprocating sliding mechanism, and the laser thickness gauge and the camera assembly are installed at the bottom of the bearing platform and face the adhesive tape. An air jet pipe is installed on the bearing platform, and the nozzle of the air jet pipe is located on one side of the camera assembly and faces the other side and the laser thickness gauge. The opening of the air jet pipe is flat. A mounting bracket is slidably connected to the bearing platform, and a sealing cover is slidably connected to the mounting bracket by a spring. An installation cover is slidably connected to the mounting bracket by a spring, and a protective film is arranged in the installation cover. An arc-shaped extruded rubber is arranged at the bottom of the installation cover. A sticky roller is rotatably connected to one end of the mounting bracket, and the sticky roller contacts the lens at the bottom of the camera assembly. A coating shaft is arranged at the top of the mounting bracket, and a smearing agent is injected into the coating shaft.

[0067] The wet coating thickness of the adhesive tape coating is directly related to the dry coating thickness after drying. Dry thickness = wet thickness The solid content of the glue is a known process parameter. The wet coating is detected before drying, and the coating amount problem can be found earlier than after drying, so as to reduce the loss from the source and detect the wet coating thickness before drying and correlate the dry coating thickness. Through solid content conversion, the core data support of front-end early warning, dynamic adjustment and resource optimization can be provided for the adhesive tape production line scheduling system. From passive response to quality problems to active control of production rhythm, the scheduling efficiency, production capacity utilization rate and order delivery quality are ultimately improved.

[0068] The wet coating detection can immediately identify the thickness deviation after coating. If the deviation is too large, the scheduling system can trigger immediate adjustment and local rejection through linkage with the control system, such as fine-tuning the scraper gap, cutting off the abnormal wet coating section, etc. The abnormal processing time is reduced, and the production schedule of subsequent orders is not affected, which is convenient for scheduling production quality of the scheduling system.

[0069] If the lens protection film is aged and residual old coating agent, it will cause the machine vision system to misjudge, false defect alarm caused by unplanned downtime, such as misjudgment after dispatching system emergency arrangement equipment maintenance, or due to missed real defects lead to unqualified products into the downstream process, need to follow-up rework production, disrupt the original order schedule; Therefore, when replacing the lens protection film on the camera component, the personnel push the installation frame close to the camera component until the adhesive roller contacts the lens protection film, the sealing cover is blocked by the camera component and stops moving, and the installation frame drives the adhesive roller to continue moving. The adhesive roller removes the old protection film while the lens is gradually separated from the lens. At the same time, part of the exposed lens immediately contacts the coating shaft of the sponge material. The coating shaft stores a conventional coating agent for promoting film sticking effect. As the old protection film is torn off, the coating shaft also gradually wipes the lens surface. The side of the coating shaft away from the installation cover wipes the old coating agent remaining on the old film, improving the cleanliness of the lens. The side of the coating shaft close to the installation cover wipes the new coating agent on the clean lens, so that the lens is replaced with a new coating agent while replacing the protection film, improving the film sticking effect, thereby improving the detection effect, and further improving the scheduling accuracy.

[0070] Further, when the installation frame drives the installation cover to approach the lens, the sealing cover extrudes the installation cover to block it from rising. When the sealing cover moves to the lens position below the film, the sealing cover moves away from the installation cover. The installation cover rises under the influence of the spring and sticks the new protection film on the lens, completing the replacement of the new and old films. This facilitates personnel operation, simplifies the work process, improves the replacement efficiency, and thus improves the detection efficiency. Moreover, the interval time between tearing the old film, applying the coating agent, and sticking the new film is short. While improving the film sticking efficiency, the time of lens exposure is shortened, dust and other impurities are avoided from staining the lens, the step of redundant lens cleaning is saved, and the lens is avoided from being scratched by multiple cleanings.

[0071] After the installation cover drives the new film to contact the lens, the spring continues to drive the installation cover to rise. The installation cover drives the extruded rubber to press the new film on the lens. The center of the extruded rubber first contacts the center of the lens. As the extruded rubber is flattened, the extruded rubber gradually sticks the surrounding part of the new film to the lens based on the center, so that the bubbles generated during the film sticking process can be excluded from the center to the periphery, avoiding bubble generation. Moreover, the extruded rubber can also press out the excess coating agent, avoiding the influence of too much coating agent on the shooting detection. Furthermore, the discharge of the coating agent can also drive the bubbles to be discharged together, improving the fluidity of the bubbles and accelerating the discharge efficiency of the bubbles, thereby improving the film sticking efficiency and reducing the detection maintenance time.

[0072] In addition, the film replacement data can be integrated into the device data of the scheduling system to establish a film replacement period and detection accuracy correlation function curve; the scheduling system can record the detection accuracy change after each film replacement, automatically generate an optimal film replacement period, and push a film replacement reminder to the dispatcher 24 hours in advance, so that the film replacement arrangement is included in the planned maintenance; and, the passive response to detection inaccuracy caused by film aging can be avoided, and the planned film replacement active control is realized, so that the scheduling plan is more in line with the equipment maintenance rhythm; and, the scheduling system can analyze the correlation between the lens detection accuracy and the effective production capacity of the production line, such as when the detection accuracy is greater than or equal to 99.5%, the effective production capacity is 30,000 square meters per day; when the accuracy is reduced to 98%, the effective production capacity is reduced to 28,000 square meters per day; when the detection accuracy shows a downward trend, the scheduling system can predict the production capacity fluctuation in advance, and actively adjust the subsequent order scheduling, so that the scheduling plan is upgraded from post-correction to pre-prediction, and the flexibility and risk resistance of the scheduling plan are improved.

[0073] Embodiment Two

[0074] On the basis of embodiment one, the system of workshop scheduling includes a workshop scheduling environment, an offline training module and an online application module:

[0075] Workshop scheduling environment: model the workshop scheduling problem as a Markov decision process, then use the disjunctive graph model to split the scheduling task into multiple nodes, and use the Gantt chart model to store the processing information matrix;

[0076] Offline training module: by continuously interacting the agent with the environment, store the generated reinforcement learning quadruple data into the storage component, select a deep reinforcement learning algorithm from the algorithm pool to train the agent for the production line scene; use the data of agent interaction to constantly calculate the loss function and update the network weights of the agent using the gradient descent algorithm until the network converges; finally, save the trained network model and weights;

[0077] Online application module: load the trained network model into the agent, then input the actual workshop environment state into the agent, the agent outputs the corresponding scheduling action through the network decision, then the workshop executes this scheduling action and updates to the next state, and the cycle continues until the scheduling task is completed;

[0078] Disjunctive graph uses Gantt chart to reflect the processing time of each process and the specific time scale of the entire scheduling scheme, and with the cooperation of disjunctive graph and Gantt chart, the scheduling problem is modeled as a discrete sequential decision process, and the state transition function, action space, reward function and agent network structure are designed for the sequential decision process;

[0079] The state transition function contains three kinds of state information, processing time, processing mark and cumulative processing time;

[0080] The action space includes a process sequence and a machine selection, and is used to encode the action space;

[0081] The reward function is a short-term return obtained by performing an action in the current state, and after the agent performs the action, the environment state is transferred and an immediate reward is given to the agent, since the maximum completion time cannot be known at the current time, but the end time of the previous state can be used to approximate the pros and cons of the current action by subtracting the end time of the current state;

[0082] The agent network structure includes a feature extraction network and a decision network, the feature extraction network is composed of a convolutional neural network and a graph neural network, the convolutional neural network extracts features from the basic state of the workshop, and embeds the state into a disjunctive graph, and then uses the graph neural network to obtain the association information of the nodes from the disjunctive graph, and finally the decision network inputs the features into a fully connected network, and outputs the probability distribution of the agent action and the action value estimation;

[0083] Specifically: when using a deep reinforcement learning algorithm to solve the tape production line scheduling problem, first, the workshop scheduling problem needs to be modeled as a Markov decision process; Then use the disjunctive graph model to split the scheduling task into multiple nodes, and use the Gantt chart model to store the processing information matrix; Because there are various dynamic events in actual workshop scheduling, human, weather and other disturbance events need to be added to the environment to form a dynamic production environment, and include state transition functions, reward functions and agent action decoders, etc.

[0084] Then, by continuously interacting the agent with the environment, the generated reinforcement learning quadruple data is stored in the storage component; That is, in the production line scene, select a deep reinforcement learning algorithm from the algorithm pool to train the agent, use the data of the agent interaction to continuously calculate the loss function and update the network weights of the agent using the gradient descent algorithm until the network converges; Finally, save the trained network model and weights;

[0085] Load the trained network model into the agent, then input the state of the actual workshop environment, such as the workshop scheduling environment, to the agent, the agent outputs the corresponding scheduling action through the network decision, then the workshop executes this scheduling action and updates to the next state, and so on until the scheduling task is completed;

[0086] Through the deep reinforcement learning technology, a three-layer architecture of environment modeling, offline training and online application is established. The workshop scheduling environment defines the boundary and interaction rules of the problem, provides a virtual training field for offline training, and provides a state translator for online application. Through interaction learning with the virtual environment, an optimized strategy suitable for the actual scene is generated. The decision-making ability of the offline model is applied to actual production, while real data is fed back to the offline training, forming a closed loop of virtual learning, real application and data backflow. The constraint library also shares the feature information of devices, materials and orders with the online application module. Before the scheduling decision is executed, the accuracy of the production scheduling decision is verified again through the real-time information of the constraint library to avoid deviations in the scheduling process caused by external factors.

[0087] Through the Markov decision process, the scheduling problem is converted into a sequential decision problem that can be iteratively optimized by the agent. The complex production constraints are converted into data states that can be understood by the agent. The logical constraints between processes are clearly expressed by the AND graph, reducing the difficulty of understanding constraints for the agent, reducing the modeling time of the scheduling problem, and improving the rapid response effect. Moreover, the Gantt chart intuitively stores the processing state, making high-dimensional information efficiently searchable and updateable, shortening the agent's query time, and meeting the real-time decision-making needs.

[0088] The dynamic modeling of state transition abstracts the production process into a quantifiable state flow through the combination of three types of time and marker information, meets the no-effect assumption of the Markov decision process, and enables the agent to make independent decisions based on the current state without tracing historical information, reducing the complexity of decision-making, avoiding complex decoding processes, and improving the efficiency of scheduling decision generation.

[0089] The structuring of the action space is based on the decision logic of selecting processes first and then assigning devices in the workshop scheduling. Two-dimensional coding is used to accurately map the actual scheduling process, avoiding action space redundancy and improving the decision-making efficiency of the agent.

[0090] The guidance of the reward function solves the pain point of sparse reward signals in reinforcement learning in scheduling problems. Traditional methods need to wait for the completion of the entire process to evaluate the action. Through immediate feedback, the learning process of the agent is accelerated, and the offline training cost is reduced under the same amount of training data.

[0091] The feature fusion of the network structure cooperates convolutional neural networks and graph neural networks to comprehensively extract numerical features and relationship features. The combination of the two enables the agent to grasp the state of production resources and logical constraints, making the decision-making more practical, reducing the deviation between the optimal solution output by the agent and the theoretical optimal solution, and improving the accuracy of the scheduling system.

[0092] Embodiment Three

[0093] Based on embodiment two, the offline training module trains the agent using the proximal policy optimization algorithm, uses two identical agent networks in the algorithm architecture, one network for sampling and the other network for repeated updating, so that the policy iteration gradually converges to the optimal policy;

[0094] The basic steps of the proximal policy optimization algorithm for solving the job shop scheduling problem are as follows:

[0095] Step 1: Initialize the parameters of the policy network and the value network, and the hyperparameters in the algorithm;

[0096] Step 2: Start iteration and record the number of iterations;

[0097] Step 3: Output the probability distribution of the action through the policy function, and put the data generated by the interaction between the agent and the job shop scheduling environment into the replay memory;

[0098] Step 4: Determine whether the current state is a terminal state of a round, if not, return to step 3, if yes, calculate the estimate of the backtracking reward function;

[0099] Step 5: Record the number of updates of the agent and update the agent using the offline policy, determine the number of updates, if the number of updates is less than the threshold, jump back to step 5, otherwise proceed to the next operation;

[0100] Step 6: Update the neural network weights to the old policy network, and compare the current scheduling scheme with the previous optimal scheduling scheme, if the current one is better than the previous one, save the current scheduling scheme and set it as the optimal scheduling scheme, otherwise discard it;

[0101] Step 7: Determine the number of iterations of the algorithm, if the number is less than the threshold, return to step 2 to continue iteration, otherwise the algorithm ends and outputs the optimal scheduling result and scheduling scheme;

[0102] Specifically: the proximal policy optimization algorithm realizes stable iteration of the agent policy through a double-network alternating update mechanism, and the core process is as follows:

[0103] Initialization configuration: set the initial parameters of the policy network and the value network, and configure the hyperparameters;

[0104] Interactive sampling: the policy network outputs the action probability distribution according to the current job shop state, and the agent interacts with the scheduling environment accordingly, generating four-tuple data, state, action, reward, next state, and storing them in the replay memory;

[0105] Round termination judgment: if the current state is a terminal state, such as all orders being completed, the total return of the round is calculated through the backtracking reward function, otherwise sampling continues;

[0106] Offline policy update: extract batch data from the memory bank, update the network offline; value network evaluates state value, calculates advantage function; policy network limits the deviation between new and old policies through proximal constraint, iteratively optimizes until the preset update number is reached;

[0107] Network synchronization and scheme saving: synchronize the updated network weight to the old policy network, compare the current scheduling scheme with the historical optimal scheme, and keep the better scheme;

[0108] Iteration termination judgment: if the total iteration number does not reach the threshold, return to step 2 to continue training; otherwise, output the optimal scheduling scheme;

[0109] In the traditional policy gradient algorithm, the policy update is strongly related to the sampling data, which leads to unstable training. By separating the sampling network and the update network, the sampling network remains relatively fixed, ensuring stable data distribution. The update network is optimized based on offline data, avoiding drastic fluctuations in policy and common performance shocks in traditional algorithms, ensuring stable convergence of the agent to a better policy and improving the accuracy of the scheduling system policy generation.

[0110] Moreover, the offline update mechanism allows a single interaction data to be reused multiple times, reducing the number of interactions with the environment and shortening the training period, thereby improving the speed of the scheduling system.

[0111] Furthermore, using historical data from the replay memory bank for batch processing updates reduces sample correlation. The advantage function quantifies the relative value of actions, allowing policy optimization to focus more on improving marginal returns, reducing the maximum completion time of proximal policy optimization algorithms, and further improving the accuracy of the scheduling system policy generation.

[0112] Embodiment Four

[0113] Based on embodiment three, the multi-objective task scheduling system includes a data collection and preprocessing module, a multi-objective optimization module, a cloud-edge resource allocation module, and a task scheduling algorithm.

[0114] Specifically, the core modules of the multi-objective task scheduling system based on cloud-edge collaboration and intelligent optimization technology include:

[0115] Multi-objective optimization module:

[0116] The core objectives are to minimize total completion time, minimize energy consumption, and minimize production cost, while incorporating a constraint library.

[0117] The adhesive tape production line scheduling problem is abstracted as a multi-objective optimization model. For example, for the coating process, correlation equations for device capacity, energy consumption, and processing efficiency are constructed to quantify multi-objective returns under different processing parameters.

[0118] Cloud edge collaboration layer distribution module:

[0119] The edge node is responsible for tasks with high real-time requirements and undertakes computationally intensive tasks.

[0120] Task scheduling algorithm module:

[0121] A non-dominated sorting genetic algorithm is used to solve multi-objective optimization problems. Through encoding, device allocation and process order conversion, population initialization, crossover mutation, non-dominated sorting, etc., a Pareto optimal solution group is generated, covering scheduling schemes with different target weights.

[0122] According to the priority requirements of the adhesive tape production line, the optimal scheme is selected from the Pareto frontier to generate a scheduling Gantt chart containing devices, processes, and time.

[0123] For the conflicting relationship of efficiency, cost, and energy consumption in the adhesive tape production line, such as increasing the coating speed to shorten the construction period but increase the energy consumption, the non-dominated sorting mechanism of the non-dominated sorting genetic algorithm is used to generate multiple optimal solutions under the premise of meeting the constraint conditions, so as to avoid the deterioration of other indicators caused by single objective optimization and provide flexible selection space for decision makers, such as shortening the construction period when expediting and reducing energy consumption when saving energy.

[0124] In addition, the comprehensive optimization rate of the adhesive tape production line in terms of completion time, energy consumption, and cost can be improved. The Pareto optimal solution group provides adaptive solutions for different scenarios, with short construction period solutions being preferred in emergency order scenarios and low energy consumption solutions being preferred in environmental management scenarios, thereby significantly improving the flexibility of the scheduling system.

[0125] Example five:

[0126] An adhesive tape production line production scheduling method based on intelligent optimization, the specific steps are as follows:

[0127] Step 1: Collect adhesive tape production full-process data through industrial Internet of Things devices, complete multi-dimensional data collection and preprocessing work, including device data, process data, material data, and order data.

[0128] Step 2: Build a dynamic constraint condition library based on preprocessed data, including device constraint sub-library, material constraint sub-library, and order constraint sub-library.

[0129] Step 3: Define the core optimization target of scheduling, construct a mathematical model, and take minimizing total completion time, reducing energy consumption, and reducing production cost as the core target. Determine the target weight through the entropy weight method, and integrate the device, material, and order constraints in step 2 into mathematical expressions to convert the scheduling problem into a quantifiable mathematical optimization problem, balancing the conflict between efficiency, cost, and service quality.

[0130] Step 4: Generate an initial scheduling scheme through an algorithm, select a comprehensive optimal scheme from the optimization solution, and generate information containing equipment, process, and time;

[0131] Step 5: Implement the scheduling scheme through the industrial control system and track the execution status in real time;

[0132] Step 6: Extract features such as equipment load, material inventory, and order progress through a heterogeneous graph neural network, calculate the deviation between actual and planned values, and real-time feedback data to identify deviations and trigger a dynamic adjustment mechanism.

[0133] The above shows and describes the basic principles, main features and advantages of the present application. Those skilled in the art should understand that the present application is not limited to the above embodiments, and the above embodiments and descriptions in the specification are only to illustrate the principles of the present application. Without departing from the spirit and scope of the present application, various changes and improvements can be made to the present application, and these changes and improvements all fall within the scope of the claimed present application. The scope of protection of the present application is defined by the appended claims and their equivalents.

Claims

1. A production scheduling system for a tape production line based on intelligent optimization, characterized in that, The scheduling system includes: The sensing layer monitors the operating data of each device through sensing devices, organizes the device data into device data packages, and uploads them to the control center. The device data packages include production status, operating status, process parameters, and material inventory. Cloud-edge collaboration layer: A constraint library is built based on device data packets. The constraint library includes a device constraint sub-library, a material constraint sub-library, and an order constraint sub-library. Scheduling Layer: Construct a workshop scheduling system based on deep reinforcement learning, including equipment units, resource units, scheduling units, and execution units. Equipment units include detection devices for production equipment, resource units include past order scheduling records and scheduling logic, scheduling units include the initial scheduling scheme transmitted by the scheduling system, and execution units include transmission and display devices to notify and control the coordination between production equipment. Execution layer: A multi-objective task scheduling system is embedded to evenly distribute order tasks into the workshop's scheduling system; Interaction Layer: A monitoring and early warning module is constructed. The detection data from the detection device in the scheduling layer is used as the feature extraction value, and the tape product parameters are used as the preset threshold. When the deviation value exceeds the preset threshold, a dynamic adjustment mechanism is triggered to update the scheduling scheme.

2. The intelligent optimization-based tape production line scheduling system according to claim 1, characterized in that: The constraint library includes: an equipment constraint sub-library containing equipment specification adaptation matrices, capacity decay function curves, and downtime maintenance time windows, built based on historical equipment operation data; a material constraint sub-library associated with material storage disks in the equipment, updating material inventory balances, storage locations, and expiration dates in real time, and setting multi-level inventory warning thresholds; and an order constraint sub-library using the analytic hierarchy process to prioritize orders, with priority factors including delivery urgency, order amount, and customer level.

3. The intelligent optimization-based tape production line scheduling system according to claim 1, characterized in that: The workshop scheduling system includes a workshop scheduling environment, an offline training module, and an online application module. Shop floor scheduling environment: The shop floor scheduling problem is modeled as a Markov decision process. Then, the scheduling task is broken down into multiple nodes using a disjunctive graph model, and a Gantt graph model is used to store the processing information matrix. Offline training module: By enabling the agent to continuously interact with the environment, the generated reinforcement learning quadruple data is stored in the storage component. For the production line scenario, a deep reinforcement learning algorithm is selected from the algorithm pool to train the agent. The loss function is continuously calculated using the data of the agent's interaction, and the network weights of the agent are updated using the gradient descent algorithm until the network converges. Finally, the trained network model and weights are saved. Online application module: The trained network model is loaded into the agent, and then the actual state of the workshop environment is input into the agent. The agent outputs the corresponding scheduling action through network decision, and then the workshop executes the scheduling action and updates to the next state. This cycle continues until all scheduling tasks are completed.

4. The intelligent optimization-based tape production line scheduling system according to claim 3, characterized in that: Disjunctive graphs utilize Gantt charts to reflect the processing time of each process and the specific time scale of the entire scheduling scheme. By combining disjunctive graphs and Gantt charts, the scheduling problem is modeled as a discrete sequential decision-making process. State transition functions, action spaces, reward functions, and agent network structures are designed for the sequential decision-making process.

5. The intelligent optimization-based tape production line scheduling system according to claim 4, characterized in that: The state transition function contains three types of state information: processing time, processing flag, and cumulative processing time; The motion space consists of two parts: process sequence and machine selection, and is used to encode the motion space; The reward function is the short-term reward obtained by performing an action in the current state; The intelligent agent network structure includes a feature extraction network and a decision network.

6. The intelligent optimization-based tape production line scheduling system according to claim 5, characterized in that: The offline training module uses a near-end policy optimization algorithm to train the agent. The algorithm architecture uses two identical agent networks, one of which samples and the other updates repeatedly, so that the policy iteration gradually converges to the optimal policy.

7. The intelligent optimization-based tape production line scheduling system according to claim 1, characterized in that: The multi-objective task scheduling system includes a data collection and preprocessing module, a multi-objective optimization module, a cloud-edge resource allocation module, and a task scheduling algorithm.

8. A production scheduling method for a tape production line based on intelligent optimization, wherein the scheduling method uses the scheduling system described in any one of claims 1-7 above, characterized in that: The specific steps are as follows: Step 1: Collect data from the entire tape production process using industrial IoT devices to complete multi-dimensional data collection and preprocessing, including equipment data, process data, material data, and order data; Step 2: Build a dynamic constraint library based on the preprocessed data, including equipment constraint sub-libraries, material constraint sub-libraries, and order constraint sub-libraries; Step 3: Define the core optimization objective of scheduling, construct a mathematical model with the core objectives of minimizing total completion time, reducing energy consumption, and reducing production costs. Determine the weight of each objective using the entropy weight method, and integrate the equipment, material, and order constraints from Step 2 into mathematical expressions. This transforms the scheduling problem into a quantifiable mathematical optimization problem, balancing the conflict between efficiency, cost, and service quality. Step 4: Generate an initial scheduling scheme through the algorithm, select the comprehensive optimal scheme from the optimized solutions, and generate information including equipment, processes, and time; Step 5: Implement the scheduling plan through the industrial control system and track the execution status in real time: Step 6: Extract features such as equipment load, material inventory, and order progress through heterogeneous graph neural networks, calculate the deviation between actual and planned values, provide real-time feedback data to identify deviations, and trigger a dynamic adjustment mechanism.

Images