Production scheduling system of gene genetic algorithm based on MOSO coding

Abstract

The invention relates to the field of intelligent manufacturing, and discloses a production scheduling system of a gene genetic algorithm based on MOSO coding, comprising: a computing power execution module used for splitting a production scheduling task into a subtask set based on a machine / Lot dimension, processing the subtask set through a multi-thread parallel computing architecture, and outputting a chromosome population containing a local optimal solution; and the MOSO coding module is used for adopting a multi-structure optimization coding mechanism to map the chromosome population into a structured gene sequence containing a machine code and an operation code, the machine code represents a Lot-machine distribution relation, and the operation code represents a Lot processing sequence. A large-scale production scheduling task is divided into subtask sets of a machine / Lot dimension, distributed processing of the task is achieved through a multi-thread parallel computing architecture, the computing time consumption of the complex production scheduling problem is shortened, and the Lot-machine distribution relation and the processing sequence are expressed through structured mapping of machine codes and operation codes.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a production scheduling system based on a MOSO-coded genetic algorithm. Background Art

[0002] In modern intelligent manufacturing, semiconductor manufacturing processes, due to their high precision, multiple processes, and high complexity, place extremely high demands on production scheduling systems. Semiconductor production lines encompass dozens of processes, including etching, deposition, and packaging, and involve the coordinated scheduling of hundreds of machines and thousands of product lots. Scheduling is essentially an NP-hard combinatorial optimization problem. With the continuous reduction in semiconductor process size and the expansion of production line scale, traditional rule-based scheduling methods have become inadequate for large-scale, dynamically changing production scenarios. Scheduling technologies based on intelligent optimization algorithms have gradually become a research hotspot in the industry.

[0003] At present, traditional production scheduling systems mostly use heuristic algorithms, rule-driven engines or traditional scheduling models to allocate lots and machines. However, with the increase in the number of lots and the scale of equipment, the production scheduling problem presents the characteristics of a high combinatorial explosion. Most of its scheduling strategies are heuristic or simplified rule-based models, lacking global optimization capabilities, making it difficult to cope with the optimal scheduling needs in large-scale and complex scenarios, and unable to meet the real-time scheduling requirements. Summary of the Invention

[0004] In view of the shortcomings of the existing technology, the present invention provides a production scheduling system based on a MOSO-coded genetic algorithm to solve the problem that the traditional production scheduling system cannot meet the real-time requirements of production scheduling.

[0005] To achieve the above objectives, the present invention is implemented through the following technical solutions: a production scheduling system based on a MOSO-encoded genetic algorithm, comprising:

[0006] Computing execution module: used to split the production scheduling task into a set of subtasks based on the machine / lot dimension, process them through a multi-threaded parallel computing architecture, and output a chromosome population containing the local optimal solution;

[0007] MOSO encoding module: used to use the multi-structure optimization encoding mechanism to map the chromosome population into a structured gene sequence containing machine code and operation code. The machine code represents the lot-machine allocation relationship, and the operation code represents the lot processing order.

[0008] Algorithm optimization module: It is used to iteratively evolve structured gene sequences through selection, crossover, and mutation operators controlled by adaptive parameters. The mutation probability is dynamically adjusted according to the population fitness distribution to generate a production scheduling plan that meets multiple objective constraints.

[0009] Causal reasoning module: This module is used to build a task-resource causal graph based on historical production data. By simulating the impact of the current production schedule on the production line load in future preset time slices, it outputs optimization suggestions including bottleneck predictions.

[0010] Dynamic production scheduling module: used to perform incremental recoding of the affected lot subsets based on optimization recommendations, and rapidly evolve the population using the historical optimal solution as the initial population to generate a real-time adjusted production scheduling plan.

[0011] By adopting the above technical solution, large-scale production scheduling tasks are split into a set of subtasks in the machine / lot dimension. A multi-threaded parallel computing architecture is used to achieve distributed processing of tasks, shortening the computational time of complex production scheduling problems. A structured mapping of machine code and operation code is used to express the lot-machine allocation relationship and processing sequence. A coding verification mechanism is combined to ensure that the solution meets process constraints and avoid invalid calculations. Adaptive parameter-controlled genetic operations and dynamic mutation probability adjustment are used to accelerate the population's convergence to the optimal solution and reduce the number of iterations. A causal relationship diagram is constructed based on historical data. The impact of the production scheduling plan on the production line load is simulated in advance and bottlenecks are predicted, shifting from passive response to active avoidance. Incremental recoding is triggered for abnormal events, and only the affected lot subset is optimized locally. Compared with global recalculation, the computational complexity is significantly reduced, achieving a transition from "offline batch optimization" to "real-time dynamic scheduling". The system can quickly generate optimal solutions in large-scale production scheduling scenarios and respond to production anomalies within minutes. This effectively solves the problem that traditional production scheduling systems cannot meet the real-time scheduling requirements and meets the dual requirements of production scheduling response speed and optimization accuracy for highly complex production lines such as semiconductors.

[0012] Preferably, the computing power execution module includes a task splitting unit, a load balancing unit, an executor computing unit, a result aggregation unit, and a data communication unit. The task splitting unit is used to split the production scheduling task into independent subtask sets according to the machine processing capacity and the Lot process path. Each subtask corresponds to a combination of a part of the Lot and the machine. The load balancing unit is used to dynamically distribute the subtask set to each executor node based on the real-time computing power index of the executor. The computing power index includes CPU utilization and memory occupancy. The executor computing unit is used to process the subtask set in parallel to generate chromosome fragments of the local optimal solution. What does the parallel processing include? The result aggregation unit is used to merge the chromosome fragments into a chromosome population containing the local optimal solution through a non-dominated sorting algorithm. The data communication unit is used to transmit various data based on the Redis middleware. The Redis middleware supports million-level QPS data interaction.

[0013] Preferably, the MOSO encoding module includes an MS code generation unit, an OS code generation unit, an encoding verification unit, and a structure mapping unit. The MS code generation unit is used to map the chromosome population to generate a machine code sequence with a length equal to the number of Lots as a machine code according to a multi-structure optimization encoding mechanism, and each element corresponds to a machine number. The mapping rule is based on the Lot process path and the machine processing capability matrix. The OS code generation unit is used to generate a Lot full permutation sequence as an operation code from the chromosome population according to the multi-structure optimization encoding mechanism. The full permutation sequence defines the processing order of all Lots, wherein each element is a unique identifier of the Lot. The encoding verification unit is used to verify the validity of the machine code and the operation code according to preset conditions, and the preset conditions include Equipment compatibility constraints, sequence dependency constraints, and time window constraints. The equipment compatibility constraints are used to verify whether the machine assigned to each Lot meets its process requirements. The sequence dependency constraints are used to verify whether the Lot processing sequence in the operation code meets the process path requirements. The time window constraints are used to verify whether the combination of machine code and operation code meets the quality retention time limit of the Lot. The structure mapping unit is used to map the verified machine code and operation code into a binary-coded structured gene sequence, where: the machine code represents a machine number per 4 binary bits, and the operation code is converted into a binary sequence through ordinal coding. Each Lot identifier corresponds to a unique binary string, and the gene sequence adopts a segmented structure, with the first n bits being the machine code and the last m bits being the operation code, and the middle being identified by a separator.

[0014] Preferably, the algorithm optimization module includes a fitness calculation unit, a genetic operation unit, and an evolution control unit. The fitness calculation unit is used to calculate the chromosome fitness based on a multi-objective function, parse the structured gene sequence, extract the lot-machine allocation relationship and processing order, and output an individual quality evaluation vector containing the value of each objective function. The genetic operation unit is used to use a selection operator, a crossover operator, and a mutation operator to iteratively evolve the individual quality evaluation vector. The selection operator adopts a tournament selection method, randomly selecting 5 individuals from the population each time, and retaining the individual with the highest fitness to enter the next generation. The crossover operator includes performing a single-point crossover on the machine code, randomly selecting a crossover point to exchange the parent chromosome segment, and performing a partial mapping crossover on the operation code so that the generated offspring is still a valid arrangement. The mutation operator includes performing adaptive site mutation on the machine code, and the mutation probability is dynamically adjusted by the evolution control unit. The operation code performs exchange mutation and randomly exchanges the lot numbers of two positions. The evolution control unit is used to monitor the iterative evolution, dynamically adjust the algorithm parameters and control the termination of the iteration to generate a production scheduling plan that meets the multi-objective constraints.

[0015] Preferably, the multi-objective function Fitness = ω1·F machine +ω2·F lot+ω3·F constraint , where ω1, ω2, ω3 are weight coefficients, satisfying ω1+ω2+ω3=1, and the machine factor function in, is the total recipe conversion time, t switch (m i ,m i+1 ) is the machine m i Switch to m i+1 The conversion time, is the load balancing degree of the machine, L m is the cumulative working time of machine m, α, β are normalization coefficients, satisfying α+β=1, Lot factor function in, is the total delay time, is the priority-weighted completion rate, w i Lot i The priority coefficient, p i is a binary variable that is completed on time, γ and δ are normalization coefficients, satisfying γ + δ = 1, and the constraint penalty function Among them, v j is the number of violations of the jth constraint, λ is the penalty coefficient, which is dynamically adjusted according to the importance of the constraint, and the mutation probability Among them, p mmin =0.1,p mmax =0.65, f is the current individual fitness, f max and f avg are the maximum and average fitness of the population, respectively.

[0016] Preferably, the causal reasoning module includes a causal graph construction unit, an impact simulation unit, and a bottleneck prediction unit. The causal graph construction unit is used to construct a causal relationship graph of "machine allocation-load change-bottleneck formation" in the form of a directed acyclic graph based on historical production data, wherein the nodes represent production variables and the edges represent the intensity of the causal effect between the variables. The impact simulation unit is used to input the production scheduling plan into the causal relationship graph, simulate its impact on the production line load in the future preset time slice through the Bayesian network inference algorithm, and output a simulation result containing the load prediction value of each machine. The bottleneck prediction unit is used to identify machines whose load rate exceeds a preset threshold as potential bottlenecks according to the simulation results, and generate optimization suggestions for task allocation.

[0017] Preferably, the dynamic production scheduling module includes an impact assessment unit, an incremental optimization unit, and a solution verification unit. The impact assessment unit is used to generate an event trigger signal based on a production abnormality event and determine the affected Lot subset. The production abnormality event includes an order change, equipment failure, and material shortage. The affected Lot subset includes the Lot currently being processed by the faulty equipment, the first N Lots that are about to enter the faulty equipment according to the original production scheduling plan, and all Lots directly related to the shortage of materials. The incremental optimization unit is used to perform optimization operations on the affected Lot subset. The optimization operation includes extracting gene fragments related to the affected Lots in the historical optimal solution, and performing local recoding on the extracted gene fragments, keeping the unaffected part unchanged, using the recoded genes as the initial population, and performing rapid evolution through a genetic algorithm to obtain an optimized production scheduling plan. The solution verification unit is used to verify the effectiveness of the optimized production scheduling plan, including equipment compatibility, Lot priority, and quality retention time limit, and output a real-time adjusted production scheduling plan.

[0018] A production scheduling method based on a MOSO-encoded genetic algorithm is applied to the above-mentioned production scheduling system based on a MOSO-encoded genetic algorithm, comprising the following steps:

[0019] Computing execution: Split the production scheduling task into a set of subtasks based on the machine / lot dimension, process them through a multi-threaded parallel computing architecture, and output a chromosome population containing the local optimal solution;

[0020] MOSO encoding: Using a multi-structure optimization encoding mechanism, the chromosome population is mapped into a structured gene sequence containing machine codes and operation codes. The machine code represents the lot-machine allocation relationship, and the operation code represents the lot processing order.

[0021] Algorithm optimization: Through selection, crossover, and mutation operators controlled by adaptive parameters, the structured gene sequence is iteratively evolved, and the mutation probability is dynamically adjusted according to the population fitness distribution to generate a production scheduling plan that meets multiple objective constraints;

[0022] Causal reasoning: Builds a task-resource causal graph based on historical production data. By simulating the impact of the current production schedule on the production line load in future preset time slices, it outputs optimization suggestions including bottleneck predictions.

[0023] Dynamic production scheduling: Incrementally recode the affected lot subsets based on optimization recommendations, and rapidly evolve the population using the historical optimal solution as the initial population to generate a real-time adjusted production scheduling plan.

[0024] The present invention provides a production scheduling system based on a MOSO-encoded genetic algorithm. It has the following beneficial effects:

[0025] 1. The present invention splits large-scale production scheduling tasks into a set of subtasks in the machine / Lot dimension, uses a multi-threaded parallel computing architecture to realize distributed processing of tasks, shortens the calculation time of complex production scheduling problems, expresses the Lot-machine allocation relationship and processing sequence through structured mapping of machine code and operation code, combines the coding verification mechanism to ensure that the solution meets the process constraints and avoids invalid calculations, accelerates the convergence of the population to the optimal solution through genetic operations controlled by adaptive parameters and dynamic mutation probability adjustment, reduces the number of iterations, constructs a causal relationship diagram based on historical data, simulates the impact of the production scheduling plan on the production line load in advance and predicts bottlenecks, changes passive response to active avoidance, triggers incremental recoding for abnormal events, and only performs local optimization on the affected Lot subset, which greatly reduces the amount of calculation compared to global recalculation, and solves the problem that traditional production scheduling systems cannot meet the real-time requirements of production scheduling.

[0026] 2. The present invention adopts a distributed multi-threaded parallel computing architecture to split the production scheduling tasks into a set of subtasks according to the machine and lot dimensions, and uses an executor cluster to parallelly process local optimization problems, thereby greatly shortening the calculation time in large-scale production scheduling scenarios and providing a computing power foundation for real-time scheduling.

[0027] 3. The present invention converts the lot-machine allocation relationship and processing sequence into a gene sequence through a structured mapping mechanism between machine code and operation code, and combines it with constraint verification mechanisms such as equipment compatibility and process sequence dependency to filter out invalid solutions at the coding stage, thus avoiding the iterative calculation of a large number of invalid solutions in traditional algorithms, improving the search efficiency of valid solutions, ensuring that the production scheduling plan meets the actual production process requirements from the coding layer, and reducing redundant calculations in the optimization process.

[0028] 4. The present invention introduces genetic operations controlled by adaptive parameters and a dynamic mutation probability adjustment mechanism to regulate the evolutionary strategy in real time according to the population fitness distribution: when the population converges to the local optimum, the mutation probability is automatically increased to enhance the search diversity; when the individual fitness is close to the global optimum, the mutation probability is reduced to accelerate convergence and reduce the number of iterations, so that the production scheduling plan can converge to the multi-objective optimal solution in a shorter time, thereby improving the optimization efficiency and robustness of the algorithm.

[0029] 5. The present invention constructs a task-resource causal relationship diagram based on historical production data, uses a Bayesian network to simulate the impact of production scheduling plans on future production line loads, identifies potential bottleneck machines in advance and generates task reallocation suggestions, and transforms the "post-adjustment" mode of the traditional production scheduling system into a "pre-prediction" mode, thereby reducing the bottleneck risk caused by load imbalance on the production line and improving the foresight of the production scheduling plan and the balance of production line utilization.

[0030] 6. The present invention uses the historical optimal solution as the initial population for local evolution. When faced with abnormal events such as equipment failure and order changes, it only performs incremental recoding on the affected Lot subset, so that the production scheduling plan under abnormal events can be adjusted in a short time, maintaining the continuity and real-time performance of production line scheduling, reducing the impact of production abnormalities on the overall production scheduling rhythm, and improving the dynamic adaptability of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a system architecture diagram of a production scheduling system based on a MOSO-encoded genetic algorithm proposed in the present invention;

[0032] Figure 2 This is a flow chart of a production scheduling method based on MOSO-encoded genetic algorithm proposed by the present invention. DETAILED DESCRIPTION

[0033] The following will clearly and completely describe the technical solution of the present invention in conjunction with the accompanying drawings. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0034] Please see the attached Figure 1 The embodiment of the present invention provides a production scheduling system based on a MOSO-encoded genetic algorithm, comprising:

[0035] Computing power execution module: used to split the production scheduling task into a set of subtasks based on the machine / Lot dimension, and process them through a multi-threaded parallel computing architecture, and output a chromosome population containing a local optimal solution; the computing power execution module includes a task splitting unit, a load balancing unit, an executor computing unit, a result aggregation unit, and a data communication unit. The task splitting unit is used to split the production scheduling task into an independent subtask set according to the machine processing capacity and the Lot process path. Each subtask corresponds to a combination of a part of the Lot and the machine. The load balancing unit is used to dynamically distribute the subtask set to each executor node based on the executor's real-time computing power indicators. The computing power indicators include CPU utilization and memory usage. The executor computing unit is used to process the subtask set in parallel and generate chromosome fragments of the local optimal solution. What does parallel processing include? The result aggregation unit is used to merge the chromosome fragments into a chromosome population containing the local optimal solution through a non-dominated sorting algorithm. The data communication unit is used to transmit various data based on the Redis middleware. The Redis middleware supports million-level QPS data interaction.

[0036] Specifically, the computing power execution module implements parallel processing of production scheduling tasks based on a distributed computing architecture. The task splitting unit decomposes production scheduling tasks into multiple dimensions based on machine processing capabilities and Lot process paths. Each subtask corresponds to a specific combination of a machine group and a Lot subset, thereby ensuring data locality within the subtask and reducing cross-node communication overhead. For example, 10 machines are divided into different groups based on process type, such as etching groups and deposition groups. Each group processes a Lot subset with similar process requirements, so that each subtask only involves a specific combination of a machine group and a Lot, thereby reducing the complexity of cross-task data dependencies.

[0037] The load balancing unit dynamically distributes subtasks based on the real-time computing power indicators of the executors, and regularly collects the CPU utilization, memory usage and task queue length of each node through the heartbeat mechanism. When the CPU utilization of a node exceeds the preset threshold, the system automatically reduces its task allocation weight and prioritizes allocating computationally intensive subtasks to high-performance computing nodes, thereby improving cluster resource utilization and shortening the overall computing time compared to static allocation schemes. For example, when the CPU load of an executor node exceeds the preset threshold, the system automatically prioritizes new tasks to other idle nodes, and prioritizes scheduling computationally intensive tasks (such as large-scale chromosome mutation operations) to nodes equipped with high-performance processors to ensure balanced utilization of cluster resources.

[0038] The executor computing unit uses a multi-threaded parallel architecture to process subtasks. Each executor node launches multiple computing threads to form a two-level parallel structure: thread-level parallelism processes multiple chromosomes within the same subtask, while instruction-level parallelism accelerates fitness calculations using SIMD technology, thereby improving the efficiency of single-chromosome fitness calculations. For example, each thread is responsible for processing several chromosomes within a subtask. The SIMD instruction set vectorizes and accelerates the numerical operations in the chromosome fitness calculation, significantly improving the efficiency of batch processing of similar data.

[0039] The results aggregation unit merges the chromosome segments output by each executor using a non-dominated sorting algorithm. First, chromosomes are divided into different levels based on Pareto dominance relationships. Crowding distances are then calculated for individuals within the same level to maintain population diversity. This is achieved through a distributed merge sort, ensuring efficient merging of large chromosomes. For example, each executor first performs a Pareto sort on the chromosomes locally, and then sends the sorted results to the master node for global merging. During the merging process, crowding distances are calculated to maintain population diversity and avoid falling into local optimal solutions.

[0040] The data communication unit achieves efficient data flow based on the Redis cluster, deploying multiple master-slave nodes using a read-write separation architecture and evenly distributing data through hash slot sharding. For frequently accessed hot data, local caching is enabled to reduce Redis access, thereby ensuring that concurrent data requests from a large number of actuator nodes receive timely responses. For example, frequently accessed hot data such as process constraints and machine status is cached in the actuator's local memory, reducing access pressure on the Redis cluster. For task status synchronization information with high real-time requirements, Redis's publish-subscribe mechanism is used to achieve rapid cross-node transmission, ensuring data consistency during distributed computing.

[0041] MOSO encoding module: used to adopt a multi-structure optimization coding mechanism to map the chromosome population into a structured gene sequence containing machine code and operation code. The machine code represents the Lot-machine allocation relationship, and the operation code represents the Lot processing order; the MOSO encoding module includes an MS code generation unit, an OS code generation unit, a coding verification unit, and a structure mapping unit. The MS code generation unit is used to map the chromosome population into a machine code sequence with a length equal to the number of Lots as the machine code according to the multi-structure optimization coding mechanism, and each element corresponds to the machine number. The mapping rule is based on the Lot process path and the machine processing capability matrix. The OS code generation unit is used to generate a Lot full permutation sequence as the operation code from the chromosome population according to the multi-structure optimization coding mechanism. The full permutation sequence defines the processing order of all Lots, where each element is a unique identifier of the Lot The encoding verification unit is used to verify the validity of the machine code and the operation code according to the preset conditions. The preset conditions include equipment compatibility constraints, sequence dependency constraints, and time window constraints. The equipment compatibility constraint is to verify whether the machine assigned to each Lot meets its process requirements. The sequence dependency constraint is to verify whether the Lot processing sequence in the operation code meets the process path requirements. The time window constraint is to verify whether the combination of the machine code and the operation code meets the quality retention time limit of the Lot. The structure mapping unit is used to map the verified machine code and operation code into a binary-coded structured gene sequence, where: the machine code represents a machine number per 4 binary bits, and the operation code is converted into a binary sequence through ordinal coding. Each Lot identifier corresponds to a unique binary string, and the gene sequence adopts a segmented structure, with the first n bits being the machine code and the last m bits being the operation code, and the middle being identified by a separator.

[0042] Specifically, the MOSO encoding module realizes the gene sequence expression of the production scheduling plan through a multi-structure optimization encoding mechanism, and each unit works together to ensure the effectiveness and structure of the encoding. The MS code generation unit constructs mapping rules based on the Lot (production batch in semiconductor manufacturing) process path and the machine processing capability matrix. For example, for a certain type of Lot that requires an etching process, the unit will screen out a set of machines with corresponding etching functions from the machine capability matrix, and generate a machine code sequence corresponding to the machine number for each Lot, so that each element in the sequence points to a machine that can execute its process. The OS code generation unit uses a full permutation algorithm (such as the Johnson-Trotter algorithm) to generate a Lot processing order sequence. Each element in the sequence corresponds to a unique identifier of the Lot, and the order of any two elements defines the processing priority of the Lot. For example, for a scenario containing three Lots A, B, and C, the generated full permutation sequence may be [A, B, C] or [B, A, C], etc., to represent different processing order combinations.

[0043] The code verification unit verifies the validity of machine codes and opcodes from multiple perspectives. Equipment compatibility constraint verification compares lot process requirements with the tool capability list to ensure that each lot is assigned to the corresponding tool. Sequential dependency constraint verification relies on the process path database, for example, determining whether a lot in the deposition process is processed after a lot in the photolithography process. Time window constraint verification combines the quality retention time requirements of a lot to check whether the combination of machine code and opcode causes the lot to wait at the tool for longer than the permitted threshold. The structure mapping unit converts verified machine codes and opcodes into binary gene sequences. The machine code portion determines the number of bits per segment based on the tool number range (e.g., 4 bits are used to represent each number when supporting 10 tools). The opcode uses ordinal encoding to map the lot identifier to a unique binary string. The resulting gene sequence adopts a segmented structure, with the first half consisting of the machine code binary sequence and the second half consisting of the opcode binary sequence, separated by a specific delimiter to facilitate the genetic algorithm's subsequent targeted operations on different code segments. This encoding method improves the efficiency and directionality of genetic operations through structured design, enabling the algorithm to converge to the optimal solution more quickly when dealing with large-scale production scheduling problems.

[0044] Algorithm optimization module: It is used to iteratively evolve the structured gene sequence through selection, crossover and mutation operators controlled by adaptive parameters, and the mutation probability is dynamically adjusted with the population fitness distribution to generate a production scheduling plan that meets multi-objective constraints; the algorithm optimization module includes a fitness calculation unit, a genetic operation unit, and an evolution control unit. The fitness calculation unit is used to calculate the chromosome fitness based on the multi-objective function, parse the structured gene sequence, extract the lot-machine allocation relationship and processing sequence, and output an individual quality evaluation vector containing the value of each objective function. The genetic operation unit is used to use the selection operator, crossover operator, and mutation operator to evaluate the individual quality evaluation vector. Iterative evolution is performed, and the selection operator adopts the tournament selection method. Five individuals are randomly selected from the population each time, and the individual with the highest fitness is retained to enter the next generation. The crossover operator includes performing single-point crossover on the machine code, randomly selecting the crossover point to exchange the parent chromosome fragment, and performing partial mapping crossover on the operation code so that the generated offspring is still a valid arrangement. The mutation operator includes performing adaptive site mutation on the machine code. The mutation probability is dynamically adjusted by the evolution control unit, performing exchange mutation on the operation code, and randomly exchanging the lot numbers of two positions. The evolution control unit is used to monitor the iterative evolution, dynamically adjust the algorithm parameters and control the termination of the iteration to generate a production scheduling plan that meets multiple objective constraints.

[0045] Multi-objective function Fitness = ω1·F machine +ω2·F lot +ω3·F constraint , where ω1, ω2, ω3 are weight coefficients, satisfying ω1+ω2+ω3=1, and the machine factor function in, is the total recipe conversion time, t switch (m i ,m i+1 ) is the machine m i Switch to m i+1 The conversion time, is the load balancing degree of the machine, L m is the cumulative working time of machine m, α, β are normalization coefficients, satisfying α+β=1, Lot factor function in, is the total delay time, is the priority-weighted completion rate, w i Lot i The priority coefficient, p i is a binary variable that is completed on time, γ and δ are normalization coefficients, satisfying γ + δ = 1, and the constraint penalty function Among them, v j is the number of violations of the jth constraint, λ is the penalty coefficient, which is dynamically adjusted according to the importance of the constraint, and the mutation probability Among them, pmmin =0.1,p mmax =0.65, f is the current individual fitness, f max and f avg are the maximum and average fitness of the population, respectively.

[0046] Specifically, the algorithm optimization module optimizes production scheduling through the collaborative work of multiple objective functions and adaptive genetic operations. The fitness calculation unit first decodes the structured gene sequence. For example, it extracts the MS code from a chromosome to determine the lot-machine assignment relationship and the OS code to clarify the processing sequence, and then calculates the values ​​of each objective function. For example, in the machine factor function, if the machine assignment corresponding to a chromosome results in two recipe conversions, the system obtains the corresponding conversion times based on the machine switching matrix and accumulates them to obtain the total recipe conversion time. The system also calculates the load balance degree based on the accumulated working time of each machine, and then normalizes the load balance and incorporates it into the multi-objective function calculation.

[0047] The genetic operation unit performs evolutionary operations based on fitness calculations. For example, in tournament selection, several individuals are randomly selected from the population, and the individuals with the highest fitness are retained for the next generation, ensuring an elite strategy. During crossover operations, when performing single-point crossover on the machine code, such as swapping the machine allocation segments of the two parent chromosomes at a certain crossover point, a partial mapping crossover is used on the operation code. This establishes a gene mapping relationship to avoid sequence conflicts and ensure that the offspring still has a valid arrangement. During mutation operations, the mutation probability of the machine code is dynamically adjusted based on the difference between the current individual's fitness and the population average fitness. For example, when an individual's fitness falls below the population average, the system automatically increases its mutation probability to enhance search diversity. Operation codes are mutated by randomly swapping two lot numbers.

[0048] The evolution control unit monitors the population's evolutionary status in real time and triggers a termination mechanism when the fitness improvement over multiple generations falls below a preset threshold. For example, assuming a scheduling plan has a machine factor weight of 0.4, a lot factor weight of 0.5, and a constraint penalty weight of 0.1, the system calculates the overall fitness through weighted summation. If a chromosome performs well in terms of machine load balancing but has poor lead time, the evolution control unit dynamically adjusts the mutation probability based on the fitness distribution, guiding the algorithm toward a multi-objective balance, ultimately generating an optimized scheduling plan that meets production capacity, delivery date, and equipment constraints.

[0049] The multi-objective function takes the machine factor function, the lot factor function, and the constraint penalty function as input. When the machine allocation corresponding to a chromosome results in two recipe transitions with a small difference in machine load, while the lot delivery delay is low and there are no constraint violations, a comprehensive fitness value is output after weighted summation using the weight coefficients. This value is used to measure the quality of the chromosome's corresponding production scheduling plan, achieving multi-objective balanced optimization. In the machine factor function, the total recipe transition time is calculated using the machine switching matrix as input. If the machine allocation sequence for a chromosome is machine 1 → machine 2 → machine 1, the system extracts the corresponding transition time from the switching matrix and accumulates it, outputting the total transition time to evaluate machine switching efficiency. The machine load balance calculation uses the accumulated working time of each machine as input and, through normalization, outputs a load balance indicator for optimizing machine resource allocation.

[0050] In the Lot factor function, the calculation of the total delay time takes the Lot completion time and delivery date as inputs, determines whether each Lot is delayed and accumulates the time, and outputs the total delay time to evaluate the delivery cycle; the priority weighted completion rate takes the Lot priority coefficient and completion status as inputs, and outputs the priority satisfaction through weighted summation to ensure that high-priority Lots are processed first. The constraint penalty function takes the number of constraint violations and the penalty coefficient as inputs. When a chromosome has an equipment compatibility violation, the output penalty value is calculated through an exponential function to suppress the violation plan and ensure that the production schedule meets the process constraints. The calculation of the mutation probability takes the individual fitness, the maximum fitness of the population, and the average fitness as inputs. When the fitness of a chromosome is lower than the average level of the population, the mutation probability is dynamically adjusted to improve the search diversity. For example, the mutation probability is increased after calculation through a formula, making it easier for the algorithm to jump out of the local optimal solution and achieve adaptive evolution.

[0051] Causal reasoning module: used to construct a task-resource causal graph based on historical production data, and output optimization suggestions including bottleneck predictions by simulating the impact of the current production scheduling plan on the production line load in the future preset time slice; the causal reasoning module includes a causal graph construction unit, an impact simulation unit, and a bottleneck prediction unit. The causal graph construction unit is used to construct a causal relationship graph of "machine allocation-load change-bottleneck formation" in the form of a directed acyclic graph based on historical production data, where the nodes represent production variables and the edges represent the intensity of the causal effect between variables. The impact simulation unit is used to input the production scheduling plan into the causal relationship graph, simulate its impact on the production line load in the future preset time slice through the Bayesian network inference algorithm, and output simulation results including the load prediction value of each machine. The bottleneck prediction unit is used to identify machines whose load rate exceeds the preset threshold as potential bottlenecks based on the simulation results, and generate optimization suggestions for task allocation.

[0052] Specifically, the causal reasoning module predicts and optimizes production scheduling bottlenecks by constructing a causal graph and conducting dynamic simulations. The causal graph construction unit uses the Peter-Clark (PC) algorithm based on historical production data to identify causal relationships between variables. For example, it can mine a causal path between "the lot allocation of lithography machine A" and "the load rate of etching machine B" from thousands of production records. During construction, nodes contain over 50 production variables, including machine status, lot attributes, and process parameters. Edge weights are determined using Granger causality tests. For example, for every 1-unit increase in "the lot allocation of lithography machine A," "the load rate of etching machine B" increases by an average of 0.3 units, forming a directed acyclic graph. The impact simulation unit maps the current production schedule to the causal graph and performs inference using the Markov Chain Monte Carlo (MCMC) sampling algorithm of a Bayesian network. For example, after inputting a production schedule, the system simulates the load for three future time slots (each 15 minutes) and calculates the load probability distribution for each machine in each time slot. For the process path "lithography machine A → etching machine B → deposition machine C," the system considers the cascading effects of output delays on subsequent machines. For example, if the load rate of lithography machine A exceeds 80%, the probability of overload on etching machine B in the subsequent time slice increases to 65%. The bottleneck prediction unit performs dual-threshold detection based on simulation results. It first identifies machines with a load rate exceeding 75% as warning nodes and further calculates their "bottleneck propagation index" (BPI), which is the cumulative probability that an overload at this node will cause overloads in the three subsequent process machines. For example, when the BPI value of deposition machine C reaches 0.82, the system identifies it as a critical bottleneck and generates an optimization recommendation to "shift 50% of the high-priority lots from deposition machine C to deposition machine D." This recommendation is verified through counterfactual reasoning, simulating the production line status after the recommendation is implemented to ensure a more balanced load distribution. This improves the response time of production scheduling adjustments and overall capacity utilization.

[0053] The dynamic scheduling module is used to incrementally recode the affected lot subsets based on optimization recommendations and rapidly evolve them using the historical optimal solution as the initial population to generate a real-time adjusted scheduling plan. The dynamic scheduling module includes an impact assessment unit, an incremental optimization unit, and a solution verification unit. The impact assessment unit generates event trigger signals based on production anomalies and determines the affected lot subsets. Production anomalies include order changes, equipment failures, and material shortages. The affected lot subsets include the lot currently being processed by the faulty equipment, the first N lots about to enter the faulty equipment according to the original production schedule, and all lots directly related to the material shortage. The incremental optimization unit performs optimization operations on the affected lot subsets. This optimization operation involves extracting gene fragments related to the affected lots from the historical optimal solution and performing partial recoding on these extracted gene fragments, while keeping the unaffected portions unchanged. Using the recoded genes as the initial population, rapid evolution is performed using a genetic algorithm to obtain the optimized scheduling plan. The solution verification unit verifies the effectiveness of the optimized scheduling plan, including equipment compatibility, lot priority, and quality retention time limits, and outputs the real-time adjusted scheduling plan.

[0054] Specifically, the dynamic scheduling module achieves rapid response to production anomalies through a three-level linkage mechanism. Each unit forms a closed-loop optimization loop based on the characteristics of MOSO coding and historical optimization data. The impact assessment unit monitors production anomalies in real time. When a fault signal is detected, it immediately retrieves the current processing queue and process path database of the faulty equipment to locate the affected lot subset. Using a process correlation analysis algorithm, lots with high dependencies on subsequent processes are identified and prioritized. The incremental optimization unit performs local optimization at the gene segment level for the affected subset. First, relevant MS and OS code segments are retrieved from the historical optimal solution database. Using a "hot restart" mechanism, these segments serve as the initial population. Single-point crossover and adaptive mutation are performed only on the coding bits of the affected lots, while the coding bits of unaffected lots remain unchanged. For example, if a lot was originally assigned to a faulty machine, the mutation operation will select an available machine from the machine capacity matrix and recode it, while maintaining its relative position in the processing sequence in the OS code. The solution verification unit employs a three-tiered verification system: first, the re-encoded MS code is verified using the equipment compatibility matrix; second, the OS code sequence is checked based on the priority queue; and finally, the Q-time constraint engine is invoked to simulate whether the waiting time of a lot on the new machine exceeds the quality maintenance threshold. For example, if a lot has a Q-time constraint, the system calculates its waiting time on the machine in the new production plan to determine whether it meets the constraint.

[0055] This module uses an incremental recoding strategy to maintain historical optimization results for most chromosome segments and only perform local evolution on the affected areas. Compared with global rescheduling, it reduces a lot of computational effort and ensures that in emergency scenarios such as equipment failure, new production scheduling plans can be generated and verified in a relatively short time, effectively reducing the impact of abnormal events on the overall production rhythm and maintaining the stability of production line utilization.

[0056] The computational execution module splits production scheduling tasks into independent subtasks based on machine / lot dimensions, utilizing a multi-threaded parallel computing architecture for distributed processing, significantly reducing computational time for large-scale production scheduling problems. The MOSO encoding module utilizes a structured encoding mechanism combining machine code and opcodes to map lot-to-machine assignments and processing sequences into genetic sequences. Combined with a constraint-checking mechanism, this avoids the numerous invalid solution iterations common in traditional algorithms and improves the efficiency of searching for valid solutions. The algorithm optimization module intelligently controls the evolutionary process through adaptive parameter-controlled genetic operations and dynamic mutation probability adjustment, accelerating population convergence to the optimal solution and reducing the number of iterations. The causal reasoning module constructs a causal relationship graph based on historical production data and simulates the impact of scheduling plans on future production line loads using a Bayesian network. This allows for proactive prediction of bottleneck risks and generates optimization recommendations, avoiding the lag in adjustments required after bottlenecks occur in traditional systems. When an anomaly is detected, the dynamic scheduling module performs local recoding on only the affected lot subset, using the historically optimal solution as the initial population for rapid evolution. This significantly reduces computational effort compared to traditional global recalculation and ensures that scheduling can be adjusted quickly in the event of an anomaly. This has achieved a technological leap from "post-adjustment" to "pre-judgment" and from "global recalculation" to "local optimization", and has controlled the calculation time and abnormal response time of large-scale production scheduling problems within an acceptable time window for production, solving the real-time defects of traditional production scheduling systems caused by insufficient computing power, low algorithm efficiency, and poor dynamic adaptability.

[0057] Please see the attached Figure 2 A production scheduling method based on a MOSO-encoded genetic algorithm is applied to the above-mentioned production scheduling system based on a MOSO-encoded genetic algorithm, comprising the following steps:

[0058] Computing execution: Split the production scheduling task into a set of subtasks based on the machine / lot dimension, process them through a multi-threaded parallel computing architecture, and output a chromosome population containing the local optimal solution;

[0059] MOSO encoding: It uses a multi-structure optimization encoding mechanism to map the chromosome population into a structured gene sequence containing machine codes and operation codes. The machine code represents the lot-machine allocation relationship, and the operation code represents the lot processing order.

[0060] Algorithm optimization: Through selection, crossover, and mutation operators controlled by adaptive parameters, the structured gene sequence is iteratively evolved, and the mutation probability is dynamically adjusted according to the population fitness distribution to generate a production scheduling plan that meets multiple objective constraints;

[0061] Causal reasoning: Builds a task-resource causal graph based on historical production data. By simulating the impact of the current production schedule on the production line load in future preset time slices, it outputs optimization suggestions including bottleneck predictions.

[0062] Dynamic production scheduling: Incrementally recode the affected lot subsets based on optimization recommendations, and rapidly evolve the population using the historical optimal solution as the initial population to generate a real-time adjusted production scheduling plan.

[0063] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A production scheduling system based on a MOSO-encoded genetic algorithm, characterized in that: include: Computing execution module: used to split the production scheduling task into a set of subtasks based on the machine / lot dimension, process them through a multi-threaded parallel computing architecture, and output a chromosome population containing the local optimal solution; MOSO encoding module: used to use the multi-structure optimization encoding mechanism to map the chromosome population into a structured gene sequence containing machine code and operation code. The machine code represents the lot-machine allocation relationship, and the operation code represents the lot processing order. Algorithm optimization module: It is used to iteratively evolve structured gene sequences through selection, crossover, and mutation operators controlled by adaptive parameters. The mutation probability is dynamically adjusted according to the population fitness distribution to generate a production scheduling plan that meets multiple objective constraints. Causal reasoning module: This module is used to build a task-resource causal graph based on historical production data. By simulating the impact of the current production schedule on the production line load in future preset time slices, it outputs optimization suggestions including bottleneck predictions. Dynamic production scheduling module: used to perform incremental recoding of the affected lot subsets based on optimization recommendations, and rapidly evolve the population using the historical optimal solution as the initial population to generate a real-time adjusted production scheduling plan.

2. The production scheduling system based on the MOSO-encoded genetic algorithm according to claim 1, characterized in that: The computing power execution module includes a task splitting unit, a load balancing unit, an executor computing unit, a result aggregation unit, and a data communication unit. The task splitting unit is used to split the production scheduling task into independent subtask sets according to the machine processing capacity and the Lot process path. Each subtask corresponds to a combination of a part of the Lot and the machine. The load balancing unit is used to dynamically distribute the subtask set to each executor node based on the real-time computing power index of the executor. The computing power index includes CPU utilization and memory occupancy. The executor computing unit is used to process the subtask set in parallel to generate chromosome fragments of the local optimal solution. What does the parallel processing include? The result aggregation unit is used to merge the chromosome fragments into a chromosome population containing the local optimal solution through a non-dominated sorting algorithm. The data communication unit is used to transmit various data based on the Redis middleware. The Redis middleware supports million-level QPS data interaction.

3. The production scheduling system based on the MOSO-encoded genetic algorithm according to claim 1, characterized in that: The MOSO encoding module includes an MS code generation unit, an OS code generation unit, an encoding verification unit, and a structure mapping unit. The MS code generation unit is used to map the chromosome population to a machine code sequence with a length equal to the number of Lots as the machine code according to the multi-structure optimization encoding mechanism, and each element corresponds to the machine number. The mapping rule is based on the Lot process path and the machine processing capability matrix. The OS code generation unit is used to generate a Lot full arrangement sequence as the operation code according to the multi-structure optimization encoding mechanism. The full arrangement sequence defines the processing order of all Lots, where each element is a unique identifier of the Lot. The encoding verification unit is used to verify the validity of the machine code and the operation code according to preset conditions, and the preset conditions include the equipment Compatibility constraints, sequence dependency constraints, and time window constraints. The equipment compatibility constraint is to verify whether the machine assigned to each Lot meets its process requirements. The sequence dependency constraint is to verify whether the Lot processing sequence in the operation code meets the process path requirements. The time window constraint is to verify whether the combination of machine code and operation code meets the quality retention time limit of the Lot. The structure mapping unit is used to map the verified machine code and operation code into a binary-coded structured gene sequence, wherein: the machine code represents a machine number per 4 binary bits, and the operation code is converted into a binary sequence through ordinal coding. Each Lot identifier corresponds to a unique binary string, and the gene sequence adopts a segmented structure, with the first n bits being the machine code and the last m bits being the operation code, and the middle being identified by a separator.

4. The production scheduling system based on the MOSO-encoded genetic algorithm according to claim 1, characterized in that: The algorithm optimization module includes a fitness calculation unit, a genetic operation unit, and an evolution control unit. The fitness calculation unit is used to calculate the chromosome fitness based on a multi-objective function, parse the structured gene sequence, extract the lot-machine allocation relationship and processing order, and output an individual quality evaluation vector containing the value of each objective function. The genetic operation unit is used to iteratively evolve the individual quality evaluation vector using a selection operator, a crossover operator, and a mutation operator. The selection operator adopts a tournament selection method, randomly selecting 5 individuals from the population each time, and retaining the individual with the highest fitness to enter the next generation. The crossover operator includes performing a single-point crossover on the machine code, randomly selecting a crossover point to exchange the parent chromosome fragment, and performing a partial mapping crossover on the operation code so that the generated offspring is still a valid arrangement. The mutation operator includes performing adaptive site mutation on the machine code, and the mutation probability is dynamically adjusted by the evolution control unit. The operation code performs exchange mutation and randomly exchanges the lot numbers of two positions. The evolution control unit is used to monitor the iterative evolution, dynamically adjust the algorithm parameters and control the termination of the iteration to generate a production scheduling plan that meets the multi-objective constraints.

5. The production scheduling system based on the MOSO-encoded genetic algorithm according to claim 4, characterized in that: The multi-objective function Fitness = ω1·F machine +ω2·F lot +ω3·F constraint , where ω1, ω2, ω3 are weight coefficients, satisfying ω1+ω2+ω3=1, and the machine factor function in, is the total recipe conversion time, t switch (m i ,m i+1 ) is the machine m i Switch to m i+1 The conversion time, is the load balancing degree of the machine, L m is the cumulative working time of machine m, α, β are normalization coefficients, satisfying α+β=1, Lot factor function in, is the total delay time, is the priority-weighted completion rate, w i is the priority coefficient of Loti, p i is a binary variable that is completed on time, γ and δ are normalization coefficients, satisfying γ + δ = 1, and the constraint penalty function Among them, v j is the number of violations of the jth constraint, λ is the penalty coefficient, which is dynamically adjusted according to the importance of the constraint, and the mutation probability Among them, p mmin =0.1,p mmax =0.65, f is the current individual fitness, f max and f avg are the maximum and average fitness of the population, respectively.

6. The production scheduling system based on the MOSO-encoded genetic algorithm according to claim 1, characterized in that: The causal reasoning module includes a causal graph construction unit, an impact simulation unit, and a bottleneck prediction unit. The causal graph construction unit is used to construct a causal relationship graph of "machine allocation-load change-bottleneck formation" in the form of a directed acyclic graph based on historical production data, wherein the nodes represent production variables and the edges represent the intensity of the causal effect between the variables. The impact simulation unit is used to input the production scheduling plan into the causal relationship graph, simulate its impact on the production line load in the future preset time slice through the Bayesian network inference algorithm, and output a simulation result containing the load prediction value of each machine. The bottleneck prediction unit is used to identify machines whose load rate exceeds a preset threshold as potential bottlenecks based on the simulation results, and generate optimization suggestions for task allocation.

7. The production scheduling system based on MOSO coding genetic algorithm according to claim 1, characterized in that: The dynamic production scheduling module includes an impact assessment unit, an incremental optimization unit, and a solution verification unit. The impact assessment unit is used to generate an event trigger signal based on a production abnormality event and determine the affected Lot subset. The production abnormality event includes an order change, equipment failure, and material shortage. The affected Lot subset includes the Lot currently being processed by the faulty equipment, the first N Lots that are about to enter the faulty equipment according to the original production scheduling plan, and all Lots directly related to the shortage material. The incremental optimization unit is used to perform optimization operations on the affected Lot subset. The optimization operation includes extracting gene fragments related to the affected Lots in the historical optimal solution, and performing local recoding on the extracted gene fragments, keeping the unaffected part unchanged, using the recoded genes as the initial population, and performing rapid evolution through a genetic algorithm to obtain an optimized production scheduling plan. The solution verification unit is used to verify the effectiveness of the optimized production scheduling plan, including equipment compatibility, Lot priority, and quality retention time limit, and output a real-time adjusted production scheduling plan.

8. A production scheduling method based on a MOSO-encoded genetic algorithm, characterized by: A production scheduling system based on a MOSO-encoded genetic algorithm as claimed in any one of claims 1 to 7, comprising the following steps: Computing execution: Split the production scheduling task into a set of subtasks based on the machine / lot dimension, process them through a multi-threaded parallel computing architecture, and output a chromosome population containing the local optimal solution; MOSO encoding: Using a multi-structure optimization encoding mechanism, the chromosome population is mapped into a structured gene sequence containing machine codes and operation codes. The machine code represents the lot-machine allocation relationship, and the operation code represents the lot processing order. Algorithm optimization: Through selection, crossover, and mutation operators controlled by adaptive parameters, the structured gene sequence is iteratively evolved, and the mutation probability is dynamically adjusted according to the population fitness distribution to generate a production scheduling plan that meets multiple objective constraints; Causal reasoning: Builds a task-resource causal graph based on historical production data. By simulating the impact of the current production schedule on the production line load in future preset time slices, it outputs optimization suggestions including bottleneck predictions. Dynamic production scheduling: Incrementally recode the affected lot subsets based on optimization recommendations, and rapidly evolve the population using the historical optimal solution as the initial population to generate a real-time adjusted production scheduling plan.

Images