A dynamic scheduling system for an eyeglass frame production line

Abstract

This invention relates to the field of intelligent manufacturing and scheduling optimization, specifically a dynamic scheduling system for an eyeglass frame production line. It includes modules for physical sensing, rheological control, disturbance prediction, deadlock assessment, and adaptive decision-making. The system calculates the rheological brittleness index and processing fluctuations based on process topology and raw material characteristics, assessing the deadlock risk under hard time windows. Its core is to dynamically generate entropy reduction or efficiency optimization instructions based on a comparison of risk probability and threshold. This invention quantifies the nonlinear attenuation caused by microscopic physical differences in materials, predicts delay risks in advance, and effectively eliminates scheduling deviations caused by raw material fluctuations.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production scheduling optimization, specifically to a dynamic scheduling system for an eyeglass frame production line. Background Technology

[0002] In the current high-end eyeglass frame precision manufacturing environment, the production process usually involves complex heterogeneous processes such as CNC machining, physical grinding and chemical surface treatment. Moreover, the production process is subject to the dual influence of strict hard time window constraints of chemical reactions and natural fluctuations in the physical properties of raw materials.

[0003] For the management and control of such production lines, existing scheduling schemes generally adopt a static strategy based on standard working hours or shortest processing time priority. The processing time of each process is usually preset as an ideal constant or a simple random variable, focusing on maximizing local output efficiency by optimizing the work sequence.

[0004] However, in actual operation, such solutions lack the depth of perception of implicit disturbances in the physical world. They cannot quantify the nonlinear decay of processing rates caused by physical differences in raw materials such as board hardness and moisture content, and they also struggle to identify potential rheological brittleness risks in the production flow topology. When microscopic fluctuations in working hours lead to congestion and encroach on the safety time window, existing systems often cannot react in advance, easily causing buffer overflows or system loop deadlocks. This forces work-in-process to remain in chemical processes such as electroplating or dyeing, violating the maximum dwell time limit, resulting in high-cost batch chemical scrap and production line downtime accidents. It is difficult to ensure the continuous and safe operation of the system under unpredictable disturbances.

[0005] Therefore, how to establish a defense mechanism that integrates the perception of the physical properties of raw materials with the prediction of hidden deadlock risks, so as to achieve a dynamic balance between survival safety and production efficiency while effectively avoiding catastrophic deadlock, has become an urgent technical problem to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a dynamic scheduling system for an eyeglass frame production line to solve the problems mentioned in the background art. Specifically, the technical solution of this invention includes:

[0007] The physical state perception module is used to collect operating status parameters in real time through a sensor network and convert them into physical feedback signals in a unified format. The parameters include: process topology, current buffer queue length, order sequence to be processed, physical properties of raw materials, and hard time window constraints for key processes.

[0008] A rheological state controller, connected to the physical state sensing module, is used to calculate, in real time, a rheological brittleness index reflecting the solidification tendency of the production flow as a first control parameter based on the process topology and buffer queue length; a processing disturbance predictor, connected to the physical state sensing module, is used to predict the processing time fluctuation value of random processing delay based on the physical properties of the raw materials as a second control parameter.

[0009] A deadlock risk assessor, connected to the rheological state controller and the processing disturbance predictor, is used to calculate the system deadlock risk probability under hard time window constraints based on the first and second control parameters and output a risk warning signal; an adaptive control decision-maker, connected to the deadlock risk assessor, is used to compare the risk probability with a safety threshold: when the probability is greater than the threshold, an entropy reduction control command is generated to reduce the system entropy value.

[0010] When the probability is less than or equal to the threshold, an optimized control instruction to improve efficiency is generated; the instruction issuance and execution module is connected to the adaptive control decision-maker and the physical execution mechanism, and is used to compile the instruction into a control signal sequence to drive the physical execution mechanism to adjust the workpiece flow path, processing sequence or equipment start and stop.

[0011] Preferably, the rheological state controller includes:

[0012] The graph structure processing unit is used to call the process topology and extract its features using a graph neural network to generate node embedding vectors that represent the dynamic relationship between workstations.

[0013] A queue state fusion unit is used to map the current buffer queue length to the state attribute of the node embedding vector in real time.

[0014] The congestion propagation calculation unit dynamically calculates the congestion propagation weight of the entire network traffic based on the fused node embedding vector through a graph attention mechanism.

[0015] The brittleness index calculation unit calculates the value of the system's ability to resist disturbances in real time based on the congestion transmission weight, and outputs the value as the rheological brittleness index.

[0016] Preferably, the processing disturbance predictor includes:

[0017] A characteristic data input interface is used to collect the physical characteristic data of the raw materials, wherein the data includes the board hardness parameter and moisture content parameter;

[0018] The association model calling unit is used to call the preset physical characteristics and processing rate association model;

[0019] The attenuation coefficient calculation unit is used to input the plate hardness parameter and the moisture content parameter into the correlation model and output the corresponding processing rate attenuation coefficient.

[0020] The fluctuation value determination unit is used to calculate the duration exceeding the standard processing time based on the processing rate attenuation coefficient and the standard processing time, and to determine the duration as the processing time fluctuation value.

[0021] Preferably, the deadlock risk assessor includes:

[0022] The constraint loading unit is used to invoke the hard time window constraints of the key process.

[0023] The state transition modeling unit, based on the time automata theory, constructs a system state transition model that includes the rheological brittleness index;

[0024] A random disturbance injection unit is used to inject the processing time fluctuation value as a random disturbance variable into the system state transition model.

[0025] The real-time risk analysis unit is used to perform reachability analysis, and to statistically analyze in real time the frequency at which the system state transitions to the final state that cannot meet the hard time window constraint under the influence of the random disturbance variable, and outputs the frequency as the deadlock risk probability.

[0026] Preferably, when the instruction issuing and execution module receives the entropy reduction control instruction, it specifically performs the following:

[0027] According to the entropy reduction control command, the forced load reduction logic is triggered;

[0028] Generate and send a sequence of control signals to the physical actuator containing at least one of the following operations: insert a preset time slot into the material flow corresponding to the order sequence to be processed, or break up the material flow corresponding to the same type of order with continuous attributes into non-continuous batches;

[0029] By executing the control signal sequence, the physical redundancy space of the current buffer queue length is directly increased.

[0030] Preferably, when the instruction issuing and execution module receives the optimization control instruction, it specifically performs the following:

[0031] Activate the deep reinforcement learning agent according to the optimized control instructions;

[0032] Using minimizing equipment switching costs and maximizing the response rate for urgent orders as objective functions, the sequence of orders to be processed is reordered to generate an optimized production sequence;

[0033] The optimized production sequence is compiled into a specific sequence of equipment action instructions and sent to the corresponding physical actuators for execution.

[0034] Preferably, the hard time window constraint for critical processes refers to:

[0035] After entering a specific chemical processing step, the target material must exit the step between a preset minimum residence time and a preset maximum residence time; otherwise, it is deemed scrap. This constraint is set by the production line's underlying process controller and fed back to the deadlock risk assessor.

[0036] Preferably, the graph neural network used in the graph structure processing unit is a heterogeneous graph structure, wherein:

[0037] The equipment in the process topology is defined as equipment nodes, and the orders in the order sequence to be processed are defined as task nodes;

[0038] The weight of the connection edge between the device node and the task node is dynamically determined by the current buffer queue length and is used to adjust the control logic in real time.

[0039] Compared with the prior art, the present invention has the following improvements and advantages:

[0040] 1. This invention uses a processing disturbance predictor to collect physical data such as the hardness and moisture content of the board material in real time, and uses a physical property-processing rate correlation model to convert these microscopic physical differences into specific and visible processing time fluctuation values. This enables the scheduling system to quantify the nonlinear decay of the processing rate caused by the hardness of the material or differences in moisture content, thereby predicting the risk of it crowding out the downstream safety time window before the actual delay occurs, and effectively eliminating the macroscopic scheduling deviation caused by the fluctuation of the physical properties of raw materials.

[0041] 2. This invention introduces a rheological state controller; by employing a heterogeneous graph neural network and an improved graph attention mechanism, the system not only monitors the length of a single buffer queue, but also takes into account the dynamic correlation between workstations and the congestion propagation weight, and calculates the rheological brittleness index that reflects the solidification tendency of the production flow in real time; this mechanism enables the system to go beyond the local perspective and accurately capture the cascading effect of congestion in the network, thereby identifying structural hidden dangers before micro-fluctuations evolve into system loop deadlock.

[0042] 3. This invention utilizes a deadlock risk assessor, based on time automata theory, to perform reachability analysis. It takes the rheological brittleness index and processing time fluctuation as inputs to predict in real time the probability of the system state migrating to a point where it violates the maximum dwell time limit for chemical processes. Unlike existing systems that passively respond only after congestion occurs, this system can provide early warnings based on risk probability, ensuring that work-in-process flows out of the electroplating or dyeing process strictly within a preset time window. This effectively prevents products from being forced to remain in the chemical bath for extended periods due to buffer overflow, thereby avoiding costly batch chemical scrap and production line downtime accidents.

[0043] 4. This invention uses an adaptive control decision-maker to flexibly switch control logic based on the deadlock risk probability. When the risk is high, the system executes entropy reduction control, increasing physical redundancy by forcibly inserting idle time slices or shuffling consecutive batches, sacrificing local efficiency to ensure the system does not experience catastrophic collapse. When the risk is controllable, a deep reinforcement learning agent is used to optimize the production sequence with the goal of minimizing switching costs. Simulation data shows that this strategy significantly reduces the frequency of deadlock shutdowns while effectively improving the overall output of the system, solving the technical problem that traditional scheduling schemes cannot guarantee continuous operation under unpredictable high-risk steady-state conditions. Attached Figure Description

[0044] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0045] Figure 1 This is a structural diagram of the system of the present invention; Detailed Implementation

[0046] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0047] Example 1:

[0048] Please see Figure 1 A dynamic scheduling system for an eyeglass frame production line includes: a physical state perception module, used to collect operating status parameters in real time through a sensor network and convert them into physical feedback signals in a unified format; the parameters include: process topology, current buffer queue length, order sequence to be processed, physical properties of raw materials, and hard time window constraints for key processes;

[0049] The rheological state controller, connected to the physical state sensing module, is used to calculate the rheological brittleness index, which reflects the solidification tendency of the production flow, as the first control parameter based on the process topology and buffer queue length. The processing disturbance predictor, connected to the physical state sensing module, is used to predict the processing time fluctuation value of random processing delay based on the physical properties of the raw materials as the second control parameter.

[0050] The deadlock risk assessor, connected to the rheological state controller and the processing disturbance predictor, is used to calculate the system deadlock risk probability under hard time window constraints based on the first and second control parameters and output a risk warning signal; the adaptive control decision-maker, connected to the deadlock risk assessor, is used to compare the risk probability with a safety threshold: when the probability is greater than the threshold, an entropy reduction control command is generated to reduce the system entropy value.

[0051] When the probability is less than or equal to the threshold, an optimized control instruction to improve efficiency is generated; the instruction issuance and execution module is connected to the adaptive control decision-maker and the physical execution mechanism, and is used to compile the instruction into a control signal sequence to drive the physical execution mechanism to adjust the workpiece flow path, processing sequence or equipment start and stop.

[0052] This embodiment details the core mechanism by which the system balances survival and efficiency under high-risk steady-state conditions; the physical state perception module, as a bridge connecting the digital world and the physical workshop, captures transient changes in the physical world in real time through a sensor network deployed on CNC machines, automatic polishing machines, and the entrances and exits of electroplating tanks.

[0053] Among them, process topology It refers to the dynamically changing logical connections between various pieces of equipment on the production line at the current moment.

[0054] The specific construction rule is as follows: the system iterates through the currently effective process route, and for any adjacent process... If the process equipment With process equipment If all elements are currently online (i.e., neither faulty nor offline), then a directed edge is created in the graph structure. If the equipment When the system is in a fault offline state, it will retrieve alternative equipment based on a pre-set redundancy plan. Establish directed edges Disconnect the original connection at the same time. Adjacency matrix Based on the dynamically generated edge set mentioned above Real-time updates are performed, where if edges exist... ,but ,otherwise ;

[0055] To enable those skilled in the art to more clearly understand the meaning of the symbols in the above topological definitions and subsequent formulas, a unified definition of general symbols and subscripts is provided here: Subscript The sequence index of the representative process in the process route, i.e. for The immediate preceding process; subscript Represents the index of any node in the graph structure; Represents the total number of nodes in the graph; superscript Represents the initial input layer of the neural network; Specifically refers to the standard time constants specified in the process documents, while For general time variables; Series symbols, such as Specifically refers to the hard time threshold of a chemical process;

[0056] Used to handle situations where equipment fails offline or is bypassed for maintenance; current buffer queue length. This refers to the quantity of work-in-process temporarily stored at each workstation. Under a zero intermediate inventory design, this value is usually extremely small, and any fluctuation can easily cause it to overflow.

[0057] Rheological state controller calculates rheological brittleness index This index does not simply reflect the degree of congestion, but rather quantifies the structural risk of production flow stagnation and loop deadlocks within the topological network. Simultaneously, the processing disturbance predictor transforms invisible physical differences in raw materials into visible time-dimensional delay risks, generating processing time fluctuation values. ;

[0058] Based on this, the deadlock risk assessor performs multi-dimensional data fusion and risk extrapolation under hard time window constraints. Calculate the probability of deadlock risk The adaptive control decision-maker makes a binary decision based on this probability: when When the safety threshold is reached, the system determines that a catastrophic deadlock is about to occur, and it no longer pursues output rate, but instead performs entropy reduction operation to forcibly reduce the disorder and density of the system; otherwise, it focuses on improving production efficiency.

[0059] Regarding the safety threshold mentioned in the instruction manual This embodiment discloses its specific setting and dynamic calibration logic; safety threshold It is not a fixed constant, but is determined based on a trade-off model of false positive cost versus false negative cost; specifically, the system presets the unit loss weight caused by deadlock. The corresponding high cost of chemical waste disposal and the unit loss weight caused by misjudging the rate reduction This corresponds to a low cost of capacity loss; by collecting historical operating data to construct an ROC curve, the true positive rate and false positive rate at different thresholds are calculated, and the function that maximizes the total expected loss is selected.

[0060] ;

[0061] Minimize the value as ;in, This refers to the prior probability of a deadlock event occurring. This value is derived from statistics on the frequency of deadlock downtime in historical production cycles; for example, it might be set to 0.05. During the initial calibration phase, this is set... Calculation The initial value is 0.75, which means that the system adopts a conservative strategy and prioritizes survival. As system operation data accumulates, this threshold can be adaptively fine-tuned periodically, such as weekly, based on the latest fault statistics distribution.

[0062] The instruction issuance and execution module compiles abstract decisions into PLC or SCADA signals, which drive the robotic arm or AGV to perform actions.

[0063] This embodiment introduces a rheological brittleness index and processing disturbance prediction to construct a defense mechanism that can detect hidden deadlock risks that traditional scheduling systems cannot detect; in particular, by taking into account the fluctuation of the physical properties of raw materials, the system can react in advance before the processing slows down due to the hardness of the sheet material and is about to squeeze the safe time window of the electroplating process.

[0064] To verify the above technical effects, the applicant constructed a digital twin simulation experiment containing 50 processing nodes and 200 work-in-process items; two groups were set up for comparative experiments: the control group adopted the traditional shortest processing time priority scheduling strategy, and the experimental group adopted the dynamic scheduling system of this embodiment; in the simulation test running continuously for 1000 hours, the same raw material hardness fluctuation data were input; the experimental results showed:

[0065] The control group experienced 18 deadlock shutdowns, resulting in an average fault recovery time of 45 minutes and a total output of 8,500 units. The experimental group, despite triggering 32 entropy reduction and speed-down operations, achieved 0 deadlock shutdowns and a total output of 9,800 units. The data confirms that the experimental group avoided mass chemical scrapping caused by deadlocks while increasing the effective output rate by 15.3%, thus validating at the data level that the system can effectively avoid deadlocks and achieve a dynamic balance between survival and efficiency.

[0066] It should be further explained that the simulation environment is not a deterministic ideal environment, but rather a dynamically injected environment that obeys random perturbation injection units. The distributed random variables are used to fully simulate the unpredictable physical disturbances in real production. The 32 entropy reduction operations triggered by the experimental group actually include the system's proactive defense behavior based on the principle of minimizing expected losses. That is, under the asymmetric weight constraint that the cost of misjudgment and deceleration is much lower than the cost of deadlock and scrapping, the system uses an adaptively calibrated safety threshold to tolerate moderate defensive intervention in exchange for the system's high probability of survival under extreme random disturbances. The experimental performance is consistent with the underlying risk control logic in high-risk industrial scenarios.

[0067] Example 2:

[0068] The graph structure processing unit is used to call the process topology and extract its features using a graph neural network to generate node embedding vectors that represent the dynamic relationship between workstations.

[0069] The queue state fusion unit is used to map the current buffer queue length to the state attributes of the node embedding vector in real time;

[0070] The congestion propagation calculation unit dynamically calculates the congestion propagation weight of the entire network traffic based on the fused node embedding vector through a graph attention mechanism.

[0071] The brittleness index calculation unit calculates the numerical value of the system's ability to resist disturbances in real time based on the congestion transmission weight, and outputs this value as the rheological brittleness index.

[0072] This embodiment details the specific implementation logic of the rheological state controller, focusing on utilizing graph neural networks to quantify the system's vulnerability; the graph structure processing unit constructs a heterogeneous graph containing device nodes and order task nodes. Prior to this, the system constructs an initial node feature matrix. As the input to the 0th layer of the graph neural network;

[0073] For any device node in the graph Its initial eigenvector It is composed of physical parameters in three dimensions: a one-hot encoding of the device's current operating state, with a vector dimension of 4, corresponding to the device's operating state; here. superscript Explicitly define the initial feature state before the graph neural network iteration, and then use... The vector generated by the layer iteration is denoted as The normalized rated processing rate of the equipment is taken as:

[0074] ;

[0075] Historical average failure rate of equipment For a task node, its feature vector includes: the number of remaining processing steps for the order; and the normalized remaining time before the hard time window ends.

[0076] It then uses GraphSAGE or GCN to extract features from the input topology and generate node embedding vectors. Next, the queue status fusion unit will collect the current buffer queue length in real time. Normalization is performed, and the resulting vector is then concatenated as a dynamic feature into the node embedding vector to form a fusion vector. The congestion propagation calculation unit uses a graph attention mechanism to calculate the mutual influence between nodes. To address the issue that simply using node features cannot reflect the physical obstruction of the queue, this embodiment explicitly introduces the weight of the connecting edges in the attention calculation. Define congestion propagation weights The revised calculation formula is as follows:

[0077] ;

[0078] in, Derived from model calculations, the physical meaning is nodes. The congestion status of nodes The extent of the impact; It originates from the queue state fusion unit, and its physical meaning is the node embedding vector that fuses the queue state; Derived from the graph structure processing unit, its physical meaning is based on neighbor nodes. The connection edge weights are calculated based on the current buffer queue length, for example... This is used as an attention injection mechanism for structural biases;

[0079] Specifically, neighbor nodes are used here. Queue length rather than target node The queue length is to ensure that the bias term is different for different neighbors in the summation operation, thereby avoiding the loss of physical information due to mathematical simplification; Derived from model training, its physical meaning is a learnable weight matrix; Derived from model training, its physical meaning is attention vector; Derived from a preset algorithm, its physical meaning is a non-linear activation function. Derived from the definition of graph structure, its physical meaning is nodes. The set of first-order neighbor nodes is used to limit the normalization computation range of the attention mechanism;

[0080] Among them, the formula uses The logarithmic form is based on the log-linear model assumption and aims to express the weights of connecting edges. The multiplicative probability effect is transformed into an additive bias term in the attention score calculation, thus maintaining numerical stability during gradient backpropagation; it should be clarified here that the weight matrix and attention vector The data is not randomly initialized, but obtained through offline pre-training on historical data; specifically, the system uses the prediction error of minimizing the next time step buffer queue length as the training objective; the following mean squared error loss function is constructed. :

[0081] ;

[0082] in, The total number of time steps for the training samples. For a set of nodes, Nodes recorded in historical data exist The actual buffer queue length at any given time. For graph neural networks based The predicted value of the state feature output at time step 1; the loss function is optimized using stochastic gradient descent or an optimizer, and the parameters are updated via backpropagation. and The training process continues until convergence. This process imbues the network with the physical meaning to identify congestion transmission characteristics, ensuring the effectiveness of subsequent calculations of the rheological brittleness index.

[0083] Based on this, the brittleness index calculation unit calculates the rheological brittleness index of the entire network using the aforementioned weights. The calculation formula is as follows:

[0084] ;

[0085] in, : Derived from calculation results, its physical meaning is the rheological brittleness index; Derived from network structure, its physical meaning is the total number of nodes in the network; : Originates from real-time data acquisition by sensors; its physical meaning is a node. Current buffer queue length; : Preset function, specifically defined as The physical meaning is a nonlinear activation function used to exponentially amplify the effects of a high-load queue;

[0086] coefficients in the function This is the queue sensitivity coefficient, a value determined experimentally. When it takes a value of [value missing], [the value is missing]. At that time, it can reduce the brittleness index In queue length Reaching the buffer capacity It exhibits a significant nonlinear jump;

[0087] This embodiment uses an improved graph attention mechanism to reduce the congestion level of queues in the physical world. This is directly mapped to the attention bias of the neural network. This rheological fragility index not only reflects the current congestion situation, but also captures the network topology correlation of congestion through graph structure, ensuring that the technical feature of dynamically determining edge weights using queue length is substantially applied in the core algorithm.

[0088] Example 3:

[0089] Processing disturbance predictor includes:

[0090] The characteristic data input interface is used to collect physical characteristic data of raw materials, including board hardness parameters and moisture content parameters;

[0091] The association model calling unit is used to call the preset physical characteristics and processing rate association model;

[0092] The attenuation coefficient calculation unit is used to input the plate hardness parameters and moisture content parameters into the correlation model and output the corresponding processing rate attenuation coefficient.

[0093] The fluctuation value determination unit is used to calculate the duration exceeding the standard processing time based on the processing rate attenuation coefficient and the standard processing time, and to determine this duration as the processing time fluctuation value.

[0094] This embodiment details the implementation logic of the processing disturbance predictor, focusing on predicting time fluctuations from a physical perspective; the characteristic data input interface collects physical characteristic data of the raw materials, specifically including the plate hardness parameters collected by sensors. Such as Vickers hardness value and moisture content parameters The system invokes a model relating physical properties to processing rate, which is fitted from experimental data in materials mechanics.

[0095] The attenuation coefficient calculation unit inputs the above parameters into the model to calculate the processing rate attenuation coefficient. The formula is as follows:

[0096] ;

[0097] in, : Derived from calculation results, its physical meaning is the processing rate attenuation coefficient, which is dimensionless; The data is collected by sensors and its physical meaning is the hardness parameter of the board material. The data is collected by sensors and its physical meaning is the moisture content parameter. : Derived from historical data regression analysis, its physical meaning is an empirical constant;

[0098] The sign of the aforementioned empirical constants depends on the physical properties of the specific material being processed. In this embodiment, cellulose acetate sheets are used as an example. An increase in their moisture content leads to material softening, which in turn counteracts the processing resistance caused by hardness. Therefore, the moisture content parameter... coefficient before Take a positive value or determine the sign direction based on the actual model fitting results;

[0099] Regarding the above formulas and empirical constants The specific method of obtaining the data is based on the model assumption of a variant of the Arrhenius equation: that is, the hindering effect of material hardness on processing rate decreases exponentially; in specific implementation, this is achieved by collecting no less than... Historical batch data of the group, including hardness Moisture content and actual rate attenuation ratio The regression fitting method is used to solve the problem; for example, in the case of acetate fiberboard processing, the typical calibrated parameter values ​​are: Basic cutting resistance coefficient Hardness-sensitive factor , water content softening compensation coefficient.

[0100] The fluctuation value determination unit calculates the expected processing time fluctuation value based on the attenuation coefficient. :

[0101] ;

[0102] in, : Derived from calculation results, its physical meaning is the processing time fluctuation value; : Derived from the preset process document, its physical meaning is the standard processing time for this process, and is represented here by the symbol. To distinguish it from other time concepts;

[0103] In actual computational logic, the system's output processing rate attenuation coefficient Set a lower safety threshold ,For example That is, the coefficients actually used in the calculation are To prevent extreme abnormalities in materials from causing At that time, the calculated processing time fluctuation value It tends towards infinity;

[0104] This embodiment establishes a quantitative relationship between physical properties and the time dimension, transforming processing time, which is traditionally considered a random variable in scheduling, into a predictable deterministic variable. By introducing physical parameters, the system can anticipate the time extension caused by material hardness, providing accurate disturbance input for subsequent deadlock assessment, thereby eliminating macroscopic scheduling deviations caused by microscopic physical differences. Furthermore, those skilled in the art should understand that the acquisition of processing time fluctuation values ​​in Embodiment 1 is not limited to the specific physical characteristics and processing rate correlation model in this embodiment, i.e., the exponential decay formula. In other alternative embodiments, this function can also be achieved through the following technical means:

[0105] Data-driven machine learning models, such as support vector regression or random forest models, are used to train the current batch by using the hardness and moisture content of the board material in historical production data as input features and the actual time deviation as labels.

[0106] Based on the lookup table method, the physical properties of raw materials are divided into several grade ranges, and the corresponding time fluctuation value can be directly queried through the preset property-time mapping table.

[0107] Rapid simulation based on finite element analysis establishes a simplified physical model of the material cutting process, and real-time simulation yields changes in processing resistance, thereby estimating processing time delays. Any technical means that can receive data on the physical properties of raw materials and predict processing time delays accordingly falls within the scope of this application.

[0108] Example 4:

[0109] Deadlock risk assessors include:

[0110] The constraint loading unit is used to invoke the hard time window constraints of critical processes;

[0111] The state transition modeling unit, based on time automata theory, constructs a system state transition model that includes a rheological brittleness index;

[0112] The random disturbance injection unit is used to inject processing time fluctuations as random disturbance variables in the system state transition model.

[0113] The real-time risk analysis unit is used to perform reachability analysis, and to statistically analyze in real time the frequency at which the system state transitions to a final state that cannot meet the hard time window constraints under the influence of random disturbance variables, and outputs this frequency as the deadlock risk probability.

[0114] This embodiment details the implementation logic of the deadlock risk assessor, focusing on using time automata theory for risk deduction; the constraint loading unit calls the hard time window constraints of key processes to parse out the preset minimum dwell time. and the preset maximum stay time The system is modeled based on the theory of time automata, defining position as the process state and clock as the workpiece dwell time. ;

[0115] The random perturbation injection unit performs a critical variable-distribution transformation operation to address the problem of simulation determinism caused by directly injecting scalars; specifically, this unit does not incorporate processing time fluctuations. Instead of treating it as a single, definite value, it is used as a location parameter of a probability distribution; based on the central limit theorem, the processing time fluctuation of a single process is caused by the accumulation of a large number of independent, minute random factors such as tool wear, material micro-inhomogeneity, and voltage fluctuations. Therefore, it is assumed that it follows a Gaussian distribution, which conforms to statistical laws; the system constructs a Gaussian probability density function:

[0116] ;

[0117] Among them, the mean , For standard working hours, The predicted volatility scalar; variance ; is a dimensionless scaling factor used to adjust the magnitude of the volatility uncertainty relative to the predicted volatility value, such that the right side of the equation... Dimensions Compared with the left variance Dimensions Maintaining strict consistency is crucial, as greater volatility indicates greater uncertainty; simultaneously, to avoid fluctuations in the predicted processing time value... When the probability distribution approaches 0, it collapses into a single-point value. The system sets a minimum baseline variance. The variance of the random disturbance in the actual injected model is:

[0118] ;

[0119] Wherein, scaling factor It is a dimensionless parameter, and its value range is... This coefficient is used to characterize the uncertainty gain of the processing system; the higher the degree of equipment aging, the larger the value of this coefficient.

[0120] In each Monte Carlo simulation iteration, the system moves from this distribution. Specific processing times are generated by random sampling. And inject it into the model; at the same time, combine the rheological brittleness index to correct the trigger probability of the migration condition; in order to solve the problem that the original linear correction formula is affected by the rheological brittleness index under high load conditions. The principle of exponential growth may lead to negative probability calculation values, so this embodiment uses exponential weighting instead of linear weighting.

[0121] Specifically, for any state node in the model, the first... The system identifies the state type that the path points to, such as: a path pointing to a waiting / blocked state is defined as a risky path, and a path pointing to a normal flow state is defined as a safe path; the system then calculates the corrected intermediate probability value. :

[0122] ;

[0123] in, The rheological brittleness index. The preset adjustment intensity coefficient; Similar to the inverse temperature parameter in the Boltzmann distribution, this parameter controls the degree to which the rheological brittleness index distorts the migration probability; a value of [value missing] is recommended. ; This is the binary path sensitivity coefficient, when the path... When pointing to a deadlock or blocked state, retrieve Otherwise take ; This is the path sensitivity coefficient; for example, for a risky path, it takes a value of [value missing]. For the safe path value or , : This stems from the initial settings of the system state transition model. The value is not an arbitrarily set probability, but a priori transition frequency derived from historical production log statistics; specifically defined as: within the statistical period, when the system is in the current state node, the actual transition frequency along the path... The percentage of times migrations occur; for example, for the normal workflow path from processing to processing completion, ;

[0124] in, This represents the historical average downtime probability of the equipment in this process; for risky paths pointing to deadlock / blockage, Set as the historical average blockage rate of this process Prior transition frequency The statistical period is the most recent Each production shift, and a sliding update is triggered at the end of each shift to accommodate the slow drift of equipment status;

[0125] The physical meaning is that under ideal conditions unaffected by the rheological brittleness index, the system along the first... The standard probability of path migration; by introducing an exponential function This ensured that no matter How big or Negative value, median probability The value is always positive, which conforms to the axiom of nonnegativity of probability; the system performs normalization processing and calculates the transition probability finally used for sampling. : ;

[0126] Where the denominator is all the values ​​in this state. The migration path corresponds to The sum; due to the monotonicity of the exponential function, the rheological brittleness index The numerical change of the brittleness index can non-linearly distort the probability space of state transitions; that is, the higher the brittleness index, the faster the probability weight of the safe path decays, and the greater the probability of the system transitioning to a deadlock or blocking state. After ensuring the legality of the probability distribution through a normalization step, the real-time risk analysis unit performs reachability analysis to calculate the deadlock risk probability. The calculation formula is as follows:

[0127] ;

[0128] in, : Derived from the calculation results, its physical meaning is the probability of deadlock risk; : Derived from preset values, the physical meaning is the total number of Monte Carlo simulations, for example, 10,000; : Derived from simulation statistics, the physical meaning is in In the random sampling simulation, the actual dwell time of the workpiece Violation of constraints or Or the number of times the system entered a deadlock state;

[0129] This embodiment corrects the risk of mathematical collapse caused by linear calculation in the original logic by introducing an exponential correction formula, ensuring that the rheological brittleness index can effectively and legally distort the probability space of state transitions, making the simulation results under high brittleness state more likely to expose potential deadlock risks; at the same time, it clarifies the dimensional balance relationship in variance calculation, restoring the randomness basis required for Monte Carlo simulation.

[0130] Example 5:

[0131] When the instruction issuance and execution module receives an entropy reduction control instruction, it specifically performs the following:

[0132] Based on the entropy reduction control command, the forced load reduction logic is triggered;

[0133] Generate and send a sequence of control signals to the physical actuator containing at least one of the following operations: insert a preset time slot into the material flow corresponding to the order sequence to be processed, or break up the material flow corresponding to the same type of order with continuous attributes into non-continuous batches;

[0134] By executing a sequence of control signals, the physical redundancy space of the current buffer queue length is directly increased.

[0135] This embodiment details the specific operations of the instruction issuance and execution module when it receives an entropy reduction control instruction; in response to risk probability. Greater than the safety threshold The system enters a survival mode; it triggers forced load reduction logic, shifting the system's focus from maximizing output to reducing entropy; it generates a sequence of control signals, including operation A: forcibly inserting a virtual empty order into the order sequence, i.e., controlling the upstream machine to pause feeding for a preset time, such as 30 seconds; and operation B: breaking up similar orders that were originally arranged consecutively to reduce mold changeover time, inserting different types of orders or idle orders in between; by executing the above sequence, the current buffer queue length is directly affected. This reduces the system's redundancy, thus physically increasing it.

[0136] This embodiment demonstrates a counterintuitive proactive shutdown strategy; under high-risk steady-state conditions, proactively performing entropy reduction operations is the only means to avoid systemic collapse; by sacrificing local explicit efficiency, such as 30 seconds of production capacity or increasing mold change costs, global survival safety is achieved, effectively cutting off the cascading effect of congestion and preventing the batch scrapping of high-value products due to forced prolonged soaking in chemical tanks caused by downstream blockages.

[0137] Example 6:

[0138] When the instruction issuance and execution module receives an optimization control instruction, it specifically performs the following:

[0139] Activate the deep reinforcement learning agent according to the optimized control instructions;

[0140] With the objective functions of minimizing equipment changeover costs and maximizing the response rate for urgent orders, the sequence of orders to be processed is reordered to generate an optimized production sequence.

[0141] The optimized production sequence is compiled into a specific sequence of equipment action instructions and then sent to the corresponding physical actuators for execution.

[0142] This embodiment details the specific operations of the instruction issuance and execution module when it receives an optimization control instruction; in response to risk probability. Less than or equal to the safety threshold The system activates a deep reinforcement learning agent; to ensure the reproducibility of the algorithm and the effectiveness of gradient updates, the input state and output action of the agent are explicitly defined here: 1. State space Defined as a vector:

[0143] ;

[0144] in, This is the normalized vector of the current buffer queue length for each process. A unique hot code for the operating status of critical equipment, used to characterize whether the equipment is in an idle, processing, or faulty state; This is the feature vector representing the remaining time until the hard time window expires for the current batch of orders;

[0145] Action space Defined as a discrete set of actions, corresponding to the reordering operation index of the order sequence to be processed; that is, actions. This indicates that the first item in the buffer queue will be... Orders in this location will be prioritized for processing.

[0146] Based on the above definition, in order to address the logical shortcomings of relying solely on historical statistical data, which leads to delayed reward signals and difficulties in credit allocation, this system introduces a prediction model based on digital twins as an environmental feedback mechanism; and sets the objective function of the deep reinforcement learning agent to maximize the cumulative reward. The formula is as follows:

[0147] ;

[0148] in, : Derived from the calculation results, its physical meaning is the optimization target reward value; : Derived from the output of the prediction model, its physical meaning is the execution of the current reordering action. Then, the estimated number of rush orders to be completed within a unit of time; this figure is not based on historical statistics, but rather on the number of actions... Input the future prediction values ​​derived from the built-in rapid simulation environment;

[0149] The built-in fast simulation environment runs on a discrete event simulation architecture, and the specific deduction logic steps are as follows:

[0150] Initialization: Set the simulation clock Create an event priority queue ; to the current action A defined order sequence is converted into a series of task release events. ;

[0151] Event Loop: From The event with the smallest timestamp is popped from the queue. If it is a processing start event, the device resources are checked. If available, a processing complete event is generated and inserted into the queue, with its occurrence time... ,in Call the fluctuation prediction value from Example 3; if the device is unavailable, add the task to the device's virtual waiting queue;

[0152] State update and termination: Update the virtual device after each event processing. When the simulation clock The simulation terminates when the preset simulation horizon is exceeded, or when all urgent orders are marked as completed.

[0153] Indicator Calculation: The number of urgent orders completed at the end of the simulation is used as the indicator. Record the timestamp of the last task as ;

[0154] : Derived from the output of the prediction model, its physical meaning is the expected production cycle time, which is also derived from the simulation environment based on the current action; : Derived from equipment parameters, its physical meaning is equipment switching cost, which is determined by actions. The determined process switching type is calculated directly; : Derived from preset values, its physical meaning is a weighting coefficient used to balance response speed and switching cost;

[0155] To correct the inconsistency of physical dimensions in the formula, this is clarified here. and Includes a dimension conversion factor to convert the output rate of the first term. The switching costs (time or amount) of the second term are converted into uniform dimensionless utility values ​​to ensure that the subtraction operation is valid in a physical sense; specifically, and These are the normalized weight coefficients determined using the analytic hierarchy process (AHP), and they satisfy... ;in, Set to 0.7. Setting it to 0.3 indicates that under the current strategy, the priority of the urgent order response rate is higher than the equipment switching cost;

[0156] The agent is based on the current state Output Action The order sequence to be processed is reordered to generate an optimized production sequence; this sequence is then compiled into specific equipment action instructions and issued for execution.

[0157] Under the premise of controllable safety risks, this embodiment utilizes the powerful search capabilities of reinforcement learning, combined with the predictive feedback of the built-in simulation environment, to find the optimal solution that balances urgent order queuing and batch production among a massive number of permutations and combinations. This mechanism ensures that there is a clear causal relationship between the reward signal and the agent action, enabling the policy network to converge correctly and maximizing production efficiency.

[0158] Example 7:

[0159] Hard time window constraints for critical processes refer to:

[0160] Once the target material enters a specific chemical processing step, it must exit the step between a preset minimum residence time and a preset maximum residence time; otherwise, it is deemed scrap. This constraint is set by the production line's underlying process controller and fed back to the deadlock risk assessor.

[0161] This embodiment specifically defines the hard time window constraint for critical processes; this constraint specifically refers to processes involving chemical reactions, such as the dyeing process of cellulose acetate sheets or the electroplating process of metal frames; specifically, it refers to the variable of the actual residence time of the workpiece after entering the tank. Must meet:

[0162] ;

[0163] In response to The defect was determined to be a quality defect caused by incomplete chemical reaction; in response to The condition was determined to be physical failure caused by excessive chemical corrosion; this was adopted. The series of symbols represents time thresholds and is defined as follows: This is a real-time variable, to distinguish it from the standard working time constant in the implementation of Example 3. This ensures strict adherence to the principle of uniqueness of mathematical symbols; the constraint parameter is set by the process controller at the bottom of the production line and fed back to the deadlock risk assessor in real time, serving as an invariant condition in the time automaton model.

[0164] This embodiment clearly defines the system's deadlock boundary; unlike general soft constraints such as delivery delays, the hard constraints of chemical processes will cause irreversible physical losses once violated; using this as the core criterion for deadlock assessment ensures that the scheduling algorithm always operates around the bottom line of product physical safety, fundamentally eliminating batch quality accidents caused by scheduling errors.

[0165] Example 8:

[0166] The graph neural network used in the graph structure processing unit is a heterogeneous graph structure, in which:

[0167] In the process topology, equipment is defined as equipment nodes, and orders in the order sequence to be processed are defined as task nodes;

[0168] The weight of the connection edge between the device node and the task node is dynamically determined by the current buffer queue length and is used to adjust the control logic in real time.

[0169] This embodiment further defines the graph neural network structure in the graph structure processing unit; it constructs a structure including device nodes. and task nodes Heterogeneous graph; define the weights of the connection edges between device nodes and task nodes. This is a dynamic variable; to ensure that this weight is not canceled out as a constant term in the Softmax normalization operation of the graph attention mechanism, this embodiment explicitly stipulates... The source node or neighbor node of the edge The state is determined, not by the target node. The state is uniquely determined; specifically, the weight of a connecting edge is determined by its neighboring nodes. Current buffer queue length The calculation formula is as follows:

[0170] ;

[0171] in, Derived from the calculation results, its physical meaning is the weight of the edge, reflecting the neighboring nodes. To the node The strength of information transmission or the smoothness of physical flow; : Originates from real-time sensor data; its physical meaning is neighboring nodes. , which is the source node in the attention mechanism, and the current length of the buffer queue; : Derived from a preset constant, its physical meaning is to prevent tiny values ​​where the denominator is zero; : Derived from preset values, its physical meaning is the adjustment coefficient; The range of values ​​is Its physical essence is signal amplification gain; when there is long-distance dependence in the network, it should be increased. Values ​​to prevent gradient vanishing;

[0172] This embodiment constructs a device-task heterogeneous graph and dynamically adjusts edge weights using the queue length of neighboring nodes, thus mathematically ensuring... targeting the same node different neighbors The values ​​are differentiated over time; this differentiation makes As an effective structural bias injection graph attention network, GNN can capture the congestion risk of specific tasks on specific devices in a more granular way, thereby providing high-dimensional feature input for subsequent precise control and improving the system's ability to analyze complex production scenarios.

[0173] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic scheduling system for an eyeglass frame production line, characterized in that, include: The physical state perception module is used to collect operating status parameters in real time through a sensor network and convert them into physical feedback signals in a unified format. The parameters include: process topology, current buffer queue length, order sequence to be processed, physical properties of raw materials, and hard time window constraints for key processes. A rheological state controller, connected to the physical state sensing module, is used to calculate, in real time, a rheological brittleness index reflecting the solidification tendency of the production flow as a first control parameter based on the process topology and buffer queue length; a processing disturbance predictor, connected to the physical state sensing module, is used to predict the processing time fluctuation value of random processing delay based on the physical properties of the raw materials as a second control parameter. A deadlock risk assessor, connected to the rheological state controller and the processing disturbance predictor, is used to calculate the system deadlock risk probability under hard time window constraints based on the first and second control parameters and output a risk warning signal; an adaptive control decision-maker, connected to the deadlock risk assessor, is used to compare the risk probability with a safety threshold: when the probability is greater than the threshold, an entropy reduction control command is generated to reduce the system entropy value. When the probability is less than or equal to the threshold, an optimized control instruction to improve efficiency is generated; the instruction issuance and execution module is connected to the adaptive control decision-maker and the physical execution mechanism, and is used to compile the instruction into a control signal sequence to drive the physical execution mechanism to adjust the workpiece flow path, processing sequence or equipment start and stop.

2. The dynamic scheduling system for an eyeglass frame production line according to claim 1, characterized in that, The rheological state controller includes: The graph structure processing unit is used to call the process topology and extract its features using a graph neural network to generate node embedding vectors that represent the dynamic relationship between workstations. A queue state fusion unit is used to map the current buffer queue length to the state attribute of the node embedding vector in real time. The congestion propagation calculation unit dynamically calculates the congestion propagation weight of the entire network traffic based on the fused node embedding vector through a graph attention mechanism. The brittleness index calculation unit calculates the value of the system's ability to resist disturbances in real time based on the congestion transmission weight, and outputs the value as the rheological brittleness index.

3. The dynamic scheduling system for an eyeglass frame production line according to claim 1, characterized in that, The processing disturbance predictor includes: A characteristic data input interface is used to collect the physical characteristic data of the raw materials, wherein the data includes the board hardness parameter and moisture content parameter; The association model calling unit is used to call the preset physical characteristics and processing rate association model; The attenuation coefficient calculation unit is used to input the plate hardness parameter and the moisture content parameter into the correlation model and output the corresponding processing rate attenuation coefficient. The fluctuation value determination unit is used to calculate the duration exceeding the standard processing time based on the processing rate attenuation coefficient and the standard processing time, and to determine the duration as the processing time fluctuation value.

4. The dynamic scheduling system for an eyeglass frame production line according to claim 1, characterized in that, The deadlock risk assessor includes: The constraint loading unit is used to invoke the hard time window constraints of the key process. The state transition modeling unit, based on the time automata theory, constructs a system state transition model that includes the rheological brittleness index; A random disturbance injection unit is used to inject the processing time fluctuation value as a random disturbance variable into the system state transition model. The real-time risk analysis unit is used to perform reachability analysis, and to statistically analyze in real time the frequency at which the system state transitions to the final state that cannot meet the hard time window constraint under the influence of the random disturbance variable, and to output the deadlock risk probability as the frequency.

5. The dynamic scheduling system for an eyeglass frame production line according to claim 1, characterized in that, When the instruction issuing and execution module receives the entropy reduction control instruction, it is specifically used for: According to the entropy reduction control command, the forced load reduction logic is triggered; Generate and send a sequence of control signals to the physical actuator containing at least one of the following operations: insert a preset idle time slice into the material flow corresponding to the order sequence to be processed, or break up the material flow corresponding to the same type of order with continuous attributes into non-continuous batches; By executing the control signal sequence, the physical redundancy space of the current buffer queue length is directly increased.

6. The dynamic scheduling system for an eyeglass frame production line according to claim 1, characterized in that, When the instruction issuing and execution module receives the optimization control instruction, it is specifically used for: Activate the deep reinforcement learning agent according to the optimized control instructions; Using minimizing equipment switching costs and maximizing the response rate for urgent orders as objective functions, the sequence of orders to be processed is reordered to generate an optimized production sequence; The optimized production sequence is compiled into a specific sequence of equipment action instructions and sent to the corresponding physical actuators for execution.

7. The dynamic scheduling system for an eyeglass frame production line according to claim 4, characterized in that, The hard time window constraint of the key process refers to: After entering a specific chemical processing step, the target material must exit the step between a preset minimum residence time and a preset maximum residence time; otherwise, it is deemed scrap. This constraint is set by the production line's underlying process controller and fed back to the deadlock risk assessor.

8. The dynamic scheduling system for an eyeglass frame production line according to claim 2, characterized in that, The graph neural network used in the graph structure processing unit is a heterogeneous graph structure, wherein: The equipment in the process topology is defined as equipment nodes, and the orders in the order sequence to be processed are defined as task nodes; The weight of the connection edge between the device node and the task node is dynamically determined by the current buffer queue length and is used to adjust the control logic in real time.

Images