A production scheduling method and system based on hybrid algorithm

Abstract

The application discloses a production scheduling method and system based on a hybrid algorithm, and the method comprises the following steps: generating a plurality of initial scheduling schemes corresponding to current basic production data according to a pre-constructed reinforcement learning model, and generating an initial solution based on the initial scheduling schemes; generating an initial population corresponding to the initial solution, obtaining the fitness value of each individual in the initial population, and performing non-dominated sorting and congestion calculation; performing iterative optimization on the initial population by using an evolutionary algorithm, and outputting an optimal scheduling scheme corresponding to a Pareto optimal solution pool when the iterative optimization of the evolutionary algorithm is completed. The application breaks through the limitation of a single algorithm, realizes multi-objective balanced optimization, adapts to different industry production characteristics, can quickly respond to production abnormalities and single insertion demands, significantly improves the global optimality, feasibility and dynamic response capability of the scheduling scheme, and is suitable for efficient production scheduling in various manufacturing scenarios.

Description

Technical Field

[0001] This invention relates to the field of production planning and scheduling, and in particular to a production scheduling method and system based on a hybrid algorithm. Background Technology

[0002] In the field of production planning and scheduling in manufacturing, production scheduling, as the core link connecting production planning and workshop execution, directly impacts a company's production efficiency, resource utilization, and customer delivery time achievement rate due to the scientific nature and feasibility of its solutions. As the manufacturing industry transforms towards multi-variety, small-batch, and customized production, the complexity of the production environment, the diversity of constraints, and the multidimensionality of production objectives place higher demands on scheduling technology.

[0003] Currently, the mainstream production scheduling solutions in the industry mainly include the following four categories: Rule-based scheduling systems: These systems use fixed priority rules to sequence work processes. Typical rules include First-In-First-Out (FIFO), Earliest Delivery Date (EDD), Shortest Processing Time (SPT), and other single rules. Their core logic is to quickly determine the execution order of processes and resource allocation schemes through preset fixed rules. They are characterized by simple logic and fast calculation speed, but the fixed nature of the rules makes them unable to flexibly adapt to the needs of different production scenarios. Heuristic scheduling schemes: These schemes use single optimization algorithms such as genetic algorithms and simulated annealing for global search, finding a better scheduling solution through algorithm iteration. However, they lack a dynamic adjustment mechanism, and once the algorithm parameters and search strategy are determined, they cannot be adapted in real time to changes in the production environment, resulting in limited optimization performance under complex constraints. Traditional Advanced Planning and Scheduling (APS) systems: These systems generate scheduling plans based on preset production models and static scheduling strategies. They can achieve a certain degree of multi-resource collaborative scheduling, but they cannot respond in real time to abnormal situations such as equipment failures, material shortages, and emergency orders during the production process. This results in a disconnect between the scheduling plan and actual production execution, and insufficient feasibility and practicality. Summary of the Invention

[0004] The purpose of this invention is to address the problems of poor algorithm adaptability, local optimum traps, and difficulties in multi-objective optimization faced by existing production planning and scheduling systems in complex manufacturing environments, and to propose a production scheduling method and system based on a hybrid algorithm.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution: Firstly, a production scheduling method based on a hybrid algorithm is provided, including the following steps: The pre-built reinforcement learning model generates several initial scheduling schemes based on the current basic production data, and generates an initial solution based on the obtained initial scheduling schemes; the input of the reinforcement learning model is the state vector corresponding to the current basic production data, and the output is the corresponding process selection probability distribution, equipment allocation probability distribution and time arrangement probability distribution; Based on the initial solution, a corresponding initial population is generated, the fitness value of each individual in the initial population is obtained, and non-dominated sorting and crowding calculation are performed. The initial population is iteratively optimized using an evolutionary algorithm. Upon completion of the iterations, the algorithm outputs the optimal scheduling scheme corresponding to the Pareto optimal solution pool. The iterative steps of the evolutionary algorithm are as follows: Based on fitness values, a binary tournament selection method is used to select parent individuals from the current population in the current iteration step. Based on the parent individuals, crossover and / or mutation operations are performed to obtain corresponding new individuals; Perform a heuristic local search on the newly obtained individuals to obtain optimized individuals; For each optimized individual, its fitness value is determined, and the Pareto front is identified by fast non-dominated sorting. Individuals are screened by combining crowding distance to generate a new generation of population and update the Pareto optimal solution pool.

[0006] As one possible implementation method, the fitness value is a vector composed of the delivery date achievement rate target, the resource utilization rate target, and the cost optimization target; in: The resource utilization target is calculated based on a weighted average of equipment utilization, personnel utilization, and material utilization. The cost optimization objective is calculated based on the total production cost and the cost budget.

[0007] As one possible implementation method: Based on the probability distributions of process selection, equipment selection, and time selection, sampling or greedy selection is performed to form a corresponding action sequence; The sequence of actions is mapped to a scheduling code that combines process sequence, equipment allocation, and start time; Based on the obtained scheduling code, several initial scheduling schemes are generated.

[0008] As one possible implementation method: The crossover operation includes a process sequence crossover operator that maintains process sequence constraints and an equipment allocation crossover operator that maintains equipment capacity constraints. The process sequence crossover operator uses process sequence crossover and performs process sequence constraint checks. The equipment allocation crossover operator determines the equipment allocation of the offspring to inherit from the parent according to a preset equipment crossover probability and performs equipment capacity constraint checks. If a violation is found, a replacement device is selected from the list of processable equipment.

[0009] As one possible implementation method: The mutation operations include bottleneck process forward mutation operators, batch merging mutation operators of the same process family, and low-load equipment reallocation mutation operators.

[0010] As one possible implementation method, a heuristic local search is performed on the obtained new individuals, and several sub-iterations are performed to obtain the optimized individuals. The heuristic local search algorithm adopts tabu search, step-counting hill climbing or big flood algorithm. A sub-iteration step includes the following steps; Calculate the current improvement rate, which indicates the optimization effect of the local search algorithm used in the current sliding window; When the current improvement rate is greater than or equal to a preset low improvement rate threshold, the currently used local search algorithm is used as the target search algorithm. When the current improvement rate is less than the preset low improvement rate, it indicates that the improvement rate is low. At this time, the number of consecutive low improvement rounds is accumulated by a counter. If the counter is greater than or equal to the switching threshold, the target local search algorithm is determined based on the preset dynamic switching rule and the counter is cleared. The dynamic switching rule is used to indicate the target local search algorithm corresponding to the currently used local search algorithm. The target individual is optimized based on the target local search algorithm, and the target individual is updated based on the optimization results.

[0011] As one possible implementation, a switching decision is made based on the number of constraint violations before calculating the current improvement rate: If the number of constraint violations in the current solution is greater than the constraint violation threshold, the tabu search algorithm is directly used as the target local search algorithm.

[0012] As one possible implementation, the current improvement rate is the sum of the improvement rates of each round within the sliding window; The improvement rate is a weighted combination of the improvement in the objective function and the decrease in constraint violations for the corresponding round: Specifically, the objective function values ​​are calculated based on the delivery date achievement rate, resource utilization rate, and cost ratio; the delivery date achievement rate corresponds to the delivery date achievement rate target, and the resource utilization rate corresponds to the resource utilization target; the improvement margin of the objective function is calculated based on the previous round's optimal objective function value and the current round's optimal objective function value.

[0013] As one possible implementation method, the dynamic switching rules include: When the current local search algorithm is the step-counting hill-climbing algorithm, switch to the big flood algorithm; Switch to the Big Flood algorithm when the current local search algorithm is the tabu search algorithm; When the current local search algorithm is the flood algorithm, switch to the step-counting hill climbing algorithm.

[0014] As one possible implementation method, the steps for generating the corresponding state vector based on the current basic production data are as follows: Based on the current basic production data, a sparse matrix transformation is performed to obtain the corresponding order-process association matrix, process-equipment processability matrix, time window-equipment capacity matrix, and material-process availability matrix. Based on pre-built rule constraints, the process-equipment matrices and time window-equipment capacity matrices are updated; at the same time, based on pre-built rule constraints, the changeover cleaning time is calculated, and the corresponding changeover time matrix is ​​determined. Feature extraction is performed based on the order-process association matrix, the updated process-equipment processability matrix, the updated time window-equipment capacity matrix, the material-process availability matrix, and the changeover time matrix to obtain the corresponding scheduling features.

[0015] Based on the obtained scheduling features, construct a state vector S=[S order ,S operation ,S equipment ,S time ,S material ], where S order S is the order layer state vector; operation S is the process layer state vector; equipment S is the device layer state vector; time S is the time-level state vector; material This is the state vector of the material layer.

[0016] As one possible implementation method, the scheduling problem is modeled as a Markov decision process, and a multi-objective reward function is constructed that includes rewards for meeting delivery deadlines, resource utilization, cost optimization, quality assurance, and penalties for constraint violations. An Actor-Critic network is used as the main policy network for the Markov decision process, where the Actor network is responsible for outputting action policies and the Critic network is responsible for evaluating state values ​​and guiding policy updates. The output layer of the Actor network includes a process selection branch, an equipment allocation branch, and a time scheduling branch, which output the process selection probability distribution, the equipment selection probability distribution, and the time window selection probability distribution, respectively. An action mask is applied to each of the three probability distributions, and the probability of inactive actions is set to zero and then normalized.

[0017] As one possible implementation method: Based on a multi-objective reward function, the PPO algorithm is used to train the main policy network, and the trained Actor network is used as a reinforcement learning model.

[0018] As one possible implementation method, the preset rule constraints include the free-floating rule of overlapping process lines, the rule of collaborative production between product categories, the three-dimensional cleaning and changeover rule of process / product / equipment, and the rule of centralized production in specific work areas; The process flow line overlap free float rule is used to determine the allowable overlap time, and the actual start time is calculated based on the allowable overlap time; The inter-category collaborative production rules are used to determine products belonging to the same process family; The process / product / equipment three-dimensional cleaning and replacement rules are used to calculate the replacement cleaning time based on the three dimensions of process type, product characteristics and equipment status, and generate the corresponding replacement time matrix. The specific workspace centralized production rules are used to indicate centralized production constraints and to determine the corresponding penalty coefficients based on the number of workspace switching times, environmental stabilization waiting time, and the number of cleanliness violations.

[0019] In a second aspect, a production scheduling system based on a hybrid algorithm is provided for executing the production scheduling method according to any one of claims 1-8.

[0020] This invention, by adopting the above technical solutions, has significant technical effects: A production scheduling method that uses a pre-trained reinforcement learning model to obtain an initial solution, uses an evolutionary algorithm to perform a global search, and combines a heuristic algorithm for local fine-tuning optimization, takes into account both global optimization and local feasibility, and obtains the optimal scheduling scheme. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1This is a flowchart illustrating a production scheduling method based on a hybrid algorithm according to the present invention. Figure 2 This is a flowchart illustrating the iterative steps of an evolutionary algorithm. Detailed Implementation

[0023] The exemplary embodiments of this disclosure are described below with reference to the accompanying drawings, including various details of the embodiments to aid understanding, and should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of this disclosure. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.

[0024] Existing scheduling systems often employ fixed, single-priority rules such as FIFO (First-In, First-Out), EDD (Earliest Delivery Date), and SPT (Shortest Processing Time) in the process sequencing stage. These rules only achieve local optimization and are difficult to adapt to complex and ever-changing production scenarios. At the optimization algorithm level, they mostly use single algorithms such as genetic algorithms and simulated annealing for global search, lacking dynamic adjustment mechanisms and easily getting trapped in local optima. In terms of scheduling strategies, they adopt a static scheduling mode, which cannot respond to production anomalies and order insertions in real time. At the same time, most scheduling systems only have a single optimization algorithm engine, making it difficult to achieve optimal balance in different production scenarios and different optimization stages.

[0025] To address the aforementioned issues, one embodiment of this application proposes a production scheduling method and system based on a hybrid algorithm, including a reinforcement learning model training step, and generating an initial solution based on the trained reinforcement learning model. Based on the initial solution, an evolutionary algorithm is used to perform a global search, and a heuristic algorithm is combined to perform local fine-tuning optimization, taking into account both global optimization and local feasibility, to obtain the optimal scheduling scheme. First, configure the rules: Pre-configure corresponding rule constraints according to industry characteristics. Each rule constraint is used to indicate the corresponding hard constraints, as well as the corresponding process prohibition / mandatory relationship, and to determine the constraint parameters and penalty parameters based on the corresponding basic production data. Those skilled in the art can set the rule constraints themselves according to actual needs. In practical applications, based on the input basic production data and various rule constraints, corresponding sets of hard constraints, soft constraint penalty coefficients, changeover time matrices, and process prohibition / mandatory relationship tables are generated. In this embodiment, the rule constraints include four core industry rules: overlapping and free-floating rules for process production lines, collaborative production rules between product categories, three-dimensional cleaning and replacement rules for processes / products / equipment, and centralized production rules for specific work areas. Specifically: A110, Overlapping Free Floating Rules of Process Production Lines: Applicable to industries such as pharmaceuticals, chemicals, and food processing, it sets up free-floating rules for overlapping process lines based on the maximum allowable overlap time (72 hours), allowing adjacent processes to overlap to a certain extent in time, with the overlap time not exceeding the process safety threshold.

[0026] The calculation is used to determine the allowable overlap time and, based on the allowable overlap time, to determine the actual start time; Formulas for calculating the allowed overlap time and the actual start time: Allowable overlap time = min(T) overlap max Processing time × α overlap ); Actual start time = Previous process end time - Allowable overlap time + T safety margin .

[0027] in: T overlap max : Maximum allowable overlap time, which is set to 72 hours in this embodiment; T safety margin : Safety interval time, which is set to 2 hours in this embodiment; α overlap : Overlap coefficient, with a value range of 0.0-1.0, and a value of 0.3 in this embodiment.

[0028] Constraints: The overlap time must not exceed T overlap max ; Process safety requirements must be guaranteed; There must be no material dependencies between overlapping processes.

[0029] A200, Inter-category Collaborative Production Rules: Suitable for enterprises producing multiple varieties in small batches, it allows similar products to be produced in the same time period, reducing the number of changeovers and cleaning time, and enabling collaborative production arrangements between different product categories.

[0030] Used to determine products belonging to the same process family; Formula for calculating category similarity: similarity(i,j) = w1 × process similarity + w2 × material similarity + w3 × equipment similarity; Wherein, process similarity = number of identical processes / total number of processes; material similarity = number of identical materials / total number of materials; equipment similarity = number of identical equipment / total number of equipment; w1, w2, and w3 are weighting coefficients, and w1 + w2 + w3 = 1. In this embodiment, w1 = 0.4, w2 = 0.3, and w3 = 0.3.

[0031] When similarity(i,j)≥similarity threshold If they are determined to be from the same process family, then they can be produced collaboratively, where similarity... threshold Product similarity threshold, with a value range of 0.0-1.0, and a value of 0.7 in this embodiment.

[0032] Those skilled in the art can also limit the minimum batch size, for example, the minimum batch size can be set to 10.

[0033] A300, Three-Dimensional Cleaning and Replacement Rules for Process / Product / Equipment: Applicable to industries such as chemical, pharmaceutical, and food, it establishes a three-dimensional cleaning and changeover constraint model, which is used to calculate the changeover cleaning time based on three dimensions: process type, product characteristics, and equipment status, and generate the corresponding changeover time matrix.

[0034] Transformation time matrix: T changeover =T base ×f process ×f product ×f equipment ; Among them, T base Basic replacement time (obtained from historical data statistics); f process This is the process change factor (>1.0 when the process type changes, =1.0 when they are the same); f product Product type change factor (>1.0 when product type changes, =1.0 when the product type is the same); f equipment This is the equipment condition coefficient (>1.0 when the equipment needs cleaning, =1.0 when it does not). Those skilled in the art can set the various coefficients according to actual needs. In this embodiment, the specific coefficient values ​​are as follows: Process change factor f process Same process = 1.0, similar process = 1.5, different process = 2.0, process conflict (such as switching from toxic products to food) = 5.0; similar processes are determined based on the above-mentioned inter-category collaborative production rules.

[0035] Product conversion factor f product Same product = 1.0, different specifications in the same series = 1.2, different series of products = 1.8, color switching (light color → dark color) = 1.0, color switching (dark color → light color) = 2.5; Equipment condition coefficient f equipment Equipment cleanliness is good = 1.0, equipment needs simple cleaning = 1.5, equipment needs deep cleaning = 3.0, equipment needs disinfection and sterilization = 5.0.

[0036] A400, Centralized Production Rules for Specific Workshops: Suitable for industries such as pharmaceuticals, electronics, and food (cleanrooms, constant temperature and humidity workshops, etc.), it allows for centralized production arrangements to meet the special requirements of cleanrooms, concentrating processes that need to be produced in specific work areas to reduce the number of work area changes.

[0037] It is used to indicate centralized production constraints, and also to determine the corresponding penalty coefficient based on the number of workspace switching, environmental stabilization waiting time, and number of cleanliness violations; Centralized production constraints: Constraint 1: Processes of the same cleanliness level should be arranged in consecutive time periods as much as possible, and the number of people entering at one time should be less than the maximum batch size for a single entry (due to the capacity of the work area). Constraint 2: After entering the cleanroom, continuous production should be maintained for at least N hours. min operations There are 10 processes, of which N min operations : Minimum number of continuous production steps, which is set to 3 in this embodiment.

[0038] Constraint 3: Switching between different cleanliness levels requires changing clothes and cleaning procedures. In this embodiment, the cleanliness levels include Class 100, Class 1000, and Class 10000. Penalty function: P cleanroom =w1 × number of workroom switching times + w2 × environmental stabilization waiting time + w3 × number of cleanliness violations; Where w1 is the workspace switching penalty weight; w2 is the environmental stability waiting penalty weight; w3 is the cleanliness violation penalty weight, which is a hard constraint; in this embodiment, w1=100, w2=10, w3=1000.

[0039] Environmental stabilization waiting time refers to the time required for the environment to stabilize after entering the cleanroom.

[0040] This application also includes a method for pre-training the reinforcement learning model, the training steps of which are as follows: B100, training data is constructed based on historical scheduling data; The specific steps are as follows: B110. Obtain historical scheduling data and preprocess it to obtain corresponding sample data; In this embodiment, corresponding historical scheduling data is generated based on pre-collected and recorded basic production data. As an example: Basic production data includes: Production plans, orders, BOMs (Bill of Materials), and master data from ERP (Enterprise Resource Planning) systems; among which, master data includes at least basic parameters such as equipment ledgers, process routes, standard working hours, shift calendars, personnel qualifications, changeover matrices, material master files, and supplier lead times; Production reporting and status information from the MES (Manufacturing Execution System); among which, equipment status includes at least the running / stopping / fault / maintenance status, current load, available start time, and last failure time; process parameters include at least industrial parameters such as process temperature range, pressure range, stirring speed, heat preservation time, cleanliness level, and line changeover cleaning time; Inventory and material data from the WMS (Warehouse Management System).

[0041] Those skilled in the art can set the basic production data according to actual needs. In this embodiment, the basic production sample data is multi-source full-volume production-related data, such as basic production data from ERP, MES, and WMS systems.

[0042] As an example, order information, equipment status, and process parameters are extracted from the basic production data as historical scheduling data; among them, equipment status parameters include at least utilization rate, failure frequency, mean time to repair (MTTR), mean time between failures (MTBF), maintenance plan window, and equipment changeover loss; process parameters include at least flow rate, rotation speed, processing time, cooling time, cleaning time, cleanliness level, and batch compatibility.

[0043] Those skilled in the art can set historical scheduling data according to actual needs, and this specification does not impose detailed limitations on it.

[0044] In this embodiment, the input historical scheduling data is cleaned to remove duplicate, missing, and abnormal data, thereby obtaining corresponding sample data.

[0045] B120. Convert the obtained sample data into a sparse matrix format to obtain the order-process correlation matrix M. op Process-equipment matrices M fe Time window - equipment capacity matrix M tc Material-process matrix M can be used mp ; The order-process association matrix is ​​used to represent the sequence of processes corresponding to each order.

[0046] The process-equipment machinability matrix is ​​used to characterize the candidate equipment corresponding to each process.

[0047] The time window-equipment capacity matrix is ​​used to characterize the availability of each piece of equipment in different time periods.

[0048] A material-process matrix can be used to characterize the material consumption requirements of each process.

[0049] Specifically: B121, Order-Process Association Matrix M op The definition and construction steps are as follows: Order-process correlation matrix M op Definition: M op ∈{0,1}^(n×m); Where n is the number of orders and m is the total number of processes.

[0050] Order-process correlation matrix M op Matrix element definition: M op [i,j]=1: This indicates that order i contains process j; M op [i,j]=0: This means that order i does not include process j.

[0051] In this embodiment, the order-process association matrix M op The construction steps are as follows: Initialize the all-zero matrix M op The dimension is n×m; Iterate through each order i (i=1,2,...,n); Obtain the process route for order i and get the process sequence {op i1 ,op i2 ,...,op ik}; For each operation op in the process sequence ij Set Mop[i,op] in the matrix ij ]=1; Repeat the above steps until all orders have been processed.

[0052] As an example: Suppose there are 3 orders and a total of 5 processes. Order 1 contains processes {1, 2, 3}, Order 2 contains processes {2, 4}, and Order 3 contains processes {1, 3, 5}. Then: M op =[[1,1,1,0,0],[0,1,0,1,0],[1,0,1,0,1]].

[0053] B122, Process-Equipment Machinability Matrix M fe The definition and construction steps are as follows: Process-Equipment Machinability Matrix M fe Definition: M fe ∈{0,1}^(m×e); Where m is the total number of processes and e is the total number of equipment.

[0054] Process-Equipment Machinability Matrix M fe Matrix element definition: Mfe[j,k]=1: This indicates that process j can be processed on equipment k; Mfe[j,k]=0: This means that process j cannot be processed on equipment k.

[0055] In this embodiment, the process-equipment processable matrix M fe The construction steps are as follows: Initialize the all-zero matrix M fe The dimension is m×e; Iterate through each process j (j=1,2,...,m); Query the process route of process j and obtain the list of available equipment {equip} j1 ,equip j2 ,...,equip jp}; For each device equip in the list of processable equipment jq Set M in the matrix fe [j,equip jq ]=1; Repeat the above steps until all processes are completed.

[0056] B123, Time Window - Equipment Capacity Matrix M tc The definition and construction steps are as follows: Time Window - Equipment Capacity Matrix M tc Definition: M tc ∈R^(t×e); Where t is the number of time windows (dividing the scheduling period into t time windows), and e is the total number of devices.

[0057] Time Window - Equipment Capacity Matrix M tc Matrix element definition: M tc [τ,k]: Represents the available capacity of device k within the time window τ (unit: minutes); The specific calculation formula is as follows: M tc[τ,k] = shift duration - planned maintenance duration - downtime due to fault - scheduled usage time.

[0058] In this embodiment, the time window-equipment capacity matrix M tc The construction steps are as follows: Determine the scheduling period [T] start ,T end The time windows are divided according to a fixed time interval Δt, resulting in the number of time windows t = [(T] end -T star [t) / Δt], in this embodiment, Δt = 60 minutes; Initialize matrix M tc The dimension is t×e; Iterate through each time window τ (τ=1,2,...,t) and each device k (k=1,2,...,e); Query the shift calendar of device k within the time window τ and calculate the available working hours (excluding planned maintenance, downtime due to malfunctions, etc.). Assign the calculated available time to M. tc [τ,k]; Repeat the above steps until all time windows and devices have been processed.

[0059] B124. The definition and construction steps of the material-process matrix Mmp are as follows: Material-process matrix M is available mp Definition: M mp ∈{0,1}^(p×m); Where p is the number of material types and m is the total number of processes.

[0060] Material-process matrix M is available mp Matrix element definition: M mp [μ,j]=1: indicates that material μ can be used in process j; M mp [μ,j]=0: indicates that material μ cannot be used in process j.

[0061] In this embodiment, the material-process matrix M can be used. mp The construction steps are as follows: Initialize the all-zero matrix M mp The dimension is p×m; Iterate through each process j (j=1,2,...,m); Query the BOM list for process j to obtain the required material list {mat j1 ,mat j2 ,...,mat jq}; For each material mat in the required material list jr Set M in the matrix mp [mat jr [j]=1; Repeat the above steps until all processes are completed.

[0062] B130. Based on historical scheduling data and various sparse matrices, output the corresponding hard constraint set, soft constraint penalty coefficient, changeover time matrix, and process prohibition / mandatory relationship table based on pre-configured rule constraints. At the same time, update the process-equipment matrices and the time window-equipment capacity matrix.

[0063] Those skilled in the art, knowing the above-mentioned rule constraints, can retrieve corresponding data from historical scheduling data, scheduling features, and / or sparse matrices based on each rule constraint to perform rule matching and related parameter calculation. Therefore, this embodiment will not describe it in detail.

[0064] Because of the design of the rule constraints, the capacity value of the time window-equipment capacity matrix and the process-equipment processability matrix will be modified based on the configured constraints. Those skilled in the art can easily update the process-equipment processability matrix and the time window-equipment capacity matrix given the knowledge of the rule constraints, so they will not be described in detail.

[0065] B140. Based on each sparse matrix and the transformation time matrix, extract features related to scheduling performance to obtain the corresponding scheduling features; The process-equipment processability matrix and the time window-equipment capacity matrix are the matrices updated in step B130. The scheduling characteristics include at least the following: order urgency, remaining processing path length, equipment availability window, equipment health, process changeover cost, material availability rate, work-in-process congestion, historical delay risk, energy intensity, and shift capacity.

[0066] In this embodiment, the data is standardized. All continuous features are normalized using min-max normalization and mapped to the [0,1] interval. Discrete features are encoded using one-hot encoding.

[0067] B150: Training data is constructed based on historical scheduling data, scheduling features, hard constraint set, soft constraint penalty coefficient, and transformation time matrix; In this embodiment, the scheduling problem is modeled as a Markov decision process, and the training data is used to indicate the state, action, transition, and reward corresponding to each sample point.

[0068] Those skilled in the art can determine the current state and transition status (the state of the next sample point) of each sample point based on the scheduling characteristics, and can determine the action corresponding to each sample point based on historical scheduling data. Given the reward function, the corresponding reward can be calculated based on the set of hard constraints and the penalty coefficient of soft constraints. Therefore, the construction of training data will not be described in detail here. Please refer to the following introduction to the Markov decision process model for details.

[0069] B200: Train the corresponding reinforcement learning model based on the training data; In this embodiment, the scheduling problem is modeled as a Markov Decision Process (MDP). A state space with multiple dimensions, including scheduling progress, resource utilization, and order priority, is defined. A scheduling action space, including actions such as process allocation, equipment selection, and time arrangement, is designed. A multi-objective reward function, including delivery date achievement rate, resource utilization, and production cost, is designed. An Actor network and a Critic network are constructed to obtain the master policy network. The PPO algorithm is used to train the master policy network to complete the training of the reinforcement learning model.

[0070] B210, MDP Definition: MDP is used to define states, actions, transitions, and rewards. The state space S includes multiple dimensions of production states such as scheduling progress, resource utilization, and order priority. The action space A includes scheduling decision actions such as process allocation, equipment selection, and time arrangement. The state transition probability P(s'|s,a) describes the state change after performing an action, i.e., the probability of transitioning to state s' after taking action a in state s. The reward function R(s,a,s') is designed based on multi-objective optimization requirements. In this embodiment, an Actor-Critic network is used as an approximate solution framework for this MDP. The Actor is responsible for outputting action policies, and the Critic is responsible for evaluating state value and guiding policy updates. As one possible implementation, during the training phase of the reinforcement learning model, a scheduling simulation environment or digital twin environment is constructed based on pre-configured rule constraints, a scheduling decoder, and a hard constraint checker. The reinforcement learning agent interacts in the environment to generate trajectory samples containing states, actions, rewards, and state transitions. Historical scheduling data is used to initialize the parameters of the main policy network, and is used for offline sample screening and value calibration of the DQN auxiliary module, as well as for calibrating the parameters of the reward function.

[0071] B211, State Space: In this embodiment, a multi-dimensional state space is constructed, including the order layer, process layer, equipment layer, time layer, and material layer. The actual workshop conditions are encoded as learnable inputs, so that the model can simultaneously perceive delivery pressure, equipment load, and material availability at each step of decision-making, avoiding making locally optimal decisions based on only a single objective. Detailed definition of the state space; The state vector S is composed of: S=[S order ,S operation ,S equipment ,S time ,S material ]; Among them, S order The order layer state vector has dimensions n×d. order , where n is the number of orders; S operation This is a process layer state vector with dimension m×d. op m is the total number of processes; S equipment The device layer state vector has dimensions e×d. eq e represents the total number of devices; S time The time-level state vector has a dimension of 1×d. time S material The material layer state vector has dimensions p×d. mat p represents the number of material types.

[0072] The specific definitions of the state vectors at each layer are as follows: Order layer state vector S order [i] (each order i): S order [i] = [priority weight, delivery urgency, number of remaining processes, number of completed processes, total order hours, processing hours, material availability rate, historical delay risk, order value weight, customer importance].

[0073] Process layer state vector S operation [j] (single process j): S operation [j] = [processing time, number of processable devices, completion status of preceding processes, number of subsequent processes, urgency of processes, complexity of processes, changeover cost of processes, quality risk of processes, energy consumption intensity of processes, cleanliness requirements of processes].

[0074] Device layer state vector S equipment [k] (single device k): S equipment[k] = [device availability status, current load rate, health status, utilization rate, most recent failure time interval, maintenance plan window, switching loss, energy efficiency, capability level, and specificity].

[0075] Time-layer state vector S time : S time = [Current scheduling time, scheduling cycle progress, shift progress, remaining available capacity, current time period type]; Material layer state vector S material [μ] (μ of a single material): S material [μ] = [Material inventory, material safety stock, material availability status, material supplier lead time, material in transit, material shortage risk].

[0076] All of the above dimensions are normalized to the interval [0,1].

[0077] In practical applications, orders, processes, equipment, and materials are each set according to a preset upper limit N. max M max E max P max Perform zero-padding or truncation, generate an effective bitmask mask, and finally concatenate them into a fixed-dimensional vector D. state For input to the Actor / Critic network.

[0078] Those skilled in the art can easily generate each state vector based on the obtained sparse matrix, scheduling features and corresponding basic production data, so they will not be described in detail in this embodiment.

[0079] B212, Action Space; This embodiment designs a scheduling action space that includes three core actions: process allocation, equipment selection, and time arrangement, covering all dimensions of scheduling decision-making. The composition of action vector A: A = (a op , a equip , a time ); Among them, a op Select an action for the process, and select the next process to be scheduled; a equip Assign actions to equipment and processing equipment to selected processes; a time Schedule actions for the time and determine the start time of each process.

[0080] Process selection action a op : Using a priority sorting method, output the selection probability of each process to be scheduled: a op = [p1, p2, ..., pm]; Where, p j Let be the probability that process j is selected, and summation is performed only on the processes to be scheduled to satisfy Σp. j = 1.

[0081] Process selection strategy: Get the current set of pending scheduling operations (OP) pending ; Calculate the priority score for each process; The selection probability distribution is output through the Actor network; Sampling based on probability distribution or greedy selection.

[0082] Equipment allocation action a equip : Output the probability of selecting the workable equipment for the selected process: a equip = [q1, q2, ..., q e ]; Where, q k Let k be the probability of being selected, summed only over processable equipment satisfying Σq k = 1.

[0083] Equipment selection strategy: Query Process - Equipment Processability Matrix M fe Get a list of processable equipment; Calculate the fit score for each processable device; The selection probability distribution is output through the Actor network; Sampling based on probability distribution or greedy selection.

[0084] Time scheduling action a time : Output the selection probability of the available time window for the selected process: a time = [t1, t2, ..., tT]; Among them, t τ Let τ be the probability of a time window being selected, summed only over the available time windows to satisfy Σt τ =1.

[0085] Time selection strategy: Query Time Window - Equipment Capacity Matrix M tc Get the list of available time windows; Calculate the fit score for each available time window; The selection probability distribution is output through the Actor network; Sampling based on probability distribution or greedy selection.

[0086] B213, Multi-objective reward function In this embodiment, a multi-objective reward function is constructed that includes delivery date achievement rate, resource utilization rate, and cost optimization, while simultaneously setting penalty items for constraint violations. In this embodiment, the formula for the multi-objective reward function R is: R=w delivery ×R delivery +w utilization ×R utilization +w cost ×R cost +w quality ×R quality -w penalty ×P constraint ; Among them, R delivery Rewards for meeting delivery deadlines; R utilization Incentives for resource utilization; R cost Incentives for cost optimization; R quality As a quality assurance reward; P constraint To constrain penalties for violations; w* represents the weighting coefficient of each objective; Those skilled in the art can set the weight coefficients corresponding to each objective themselves; this specification does not limit this. As an example, the weight coefficients for each objective are set as follows: Delivery date achievement weight w delivery =0.35, resource utilization rate weight w utilization =0.25, cost optimization weight w cost =0.20, Quality Assurance Weight w quality =0.15, constraint violation penalty weight w penalty =1.0.

[0087] Delivery date achievement reward R delivery The calculation formula is as follows: R delivery = (1 / N) ×Σ[max(0, 1 - delay i / T tolerance )]; Where N is the total number of orders, and delay i T represents the delay time for order i (a negative value if completed early). tolerance The tolerance threshold for delay is set to 24 hours in this embodiment; The calculation logic for the delivery date achievement reward is as follows: the reward for early completion is close to 1, the reward for on-time completion is 1, the reward for delay within the threshold is between 0 and 1, and the reward for delay exceeding the threshold is 0.

[0088] Resource utilization rate reward R utilization The calculation formula is as follows: R utilization =α×R equipment +β×R personnel +γ×R material ; Among them, R equipment For equipment utilization rate; R personnel For personnel utilization rate; R material α represents the material utilization rate; α represents the equipment utilization rate weight, β represents the personnel utilization rate weight, and γ represents the material utilization rate weight. Those skilled in the art can set each weight according to actual needs, as long as α + β + γ = 1.0. As an example, α = 0.5, β = 0.3, and γ = 0.2.

[0089] As an example, the formula for calculating equipment utilization is as follows: R equipment =(1 / E)×Σ(actual processing time / available capacity time); As an example, the formula for calculating personnel utilization rate is as follows: R personnel = (1 / W)×Σ(actual working hours / available working hours); As an example, the formula for calculating material utilization rate is as follows: R material =(1 / P)×Σ(actual consumed materials / input materials).

[0090] Cost optimization reward R cost The calculation formula is as follows: R cost =1-(C total / C max ); Among them, C total C represents the total production cost. max This represents the maximum allowable cost limit.

[0091] As an example, the formula for calculating total production cost is as follows: C total =C material +C labor +C energy +C setup +C inventory ; Among them, C material For material costs; C labor For labor costs; C energy For energy costs; C setup For replacement costs; C inventoryFor inventory costs.

[0092] Quality Assurance Bonus R quality The calculation formula is as follows: R quality =(1 / M)×ΣQ j ; Where M is the total number of processes, Q j The quality score for process j is given.

[0093] As an example, the quality score Q of process j is... j The calculation formula is as follows: Q j = w1 × Equipment compatibility + w2 × Process stability + w3 × Operator proficiency; Constraints and penalties P constraint The calculation formula is as follows: P constraint =Σ(violation type ×severity); Among them, violence type The type of constraint violated is 'severity', which represents the corresponding penalty weight.

[0094] Those skilled in the art can set corresponding penalty weights for each constraint type according to actual needs, for example: Penalties for violating hard constraints: 1000 times for process sequence violation, 1000 times for equipment capacity violation, 1000 times for material availability violation, and 10000 times for safety constraint violation; Soft constraint violation penalties: Delivery delay 100 / hour, equipment overload 50 / overload percentage, personnel overtime 30 / hour, inventory shortage 80 / shortage quantity.

[0095] B220. Construct an Actor network (Global Awareness Move Selector) for action decision-making and a Critic network (State Value Estimator) for state value evaluation. In this embodiment, a multilayer perceptron (MLP) is used as the basic network structure; PPO is used for Actor-Critic main policy learning, and DQN is used for auxiliary Q-value estimation and offline sample screening. The two are used in conjunction in the scheduling scenario of this invention.

[0096] B221, Actor network structure (Move Selector); Input layer: Fixed dimension D state The standardized state vector S fixed It is obtained from the state vector S through zero-padding / truncation and mask encoding; Hidden layers: 3 fully connected layers with 512, 256 and 128 neurons respectively, all using ReLU activation function, and a Dropout layer with a dropout rate of 0.3 to prevent overfitting; Output layer: 3 parallel branches, namely: Process selection branch: Fully connected layer, number of neurons = M max The activation function is Softmax, and the output process selection probability distribution is [p1, p2, ..., p]. m ]; Device-assigned branches: Fully connected layer, number of neurons = E max The activation function is Softmax, and the output device selection probability distribution is [q1, q2, ..., q]. e ]; Timing branch: Fully connected layer, number of neurons = Tmax, activation function Softmax, output time window selection probability distribution [t1, t2, ..., t] T ]; Output legality constraints: Apply action masks to the three branches respectively, set the probability of non-actions to zero and then normalize them to ensure that sampling is only performed within the set of actionable actions.

[0097] B222, Critic Network Structure (State Value Estimator): Input layer: Fixed dimension D state The standardized state vector S fixed It is obtained from the state vector S through zero-padding / truncation and mask encoding; Hidden layers: 3 fully connected layers with 512, 256 and 128 neurons respectively, all using ReLU activation function, and a Dropout layer with a dropout rate of 0.3; Output layer: Fully connected layer, number of neurons = 1, linear activation function, output state value estimate V(s)∈ℝ.

[0098] B230. Train the Critic network structure based on the training data to obtain the corresponding reinforcement learning model; In this embodiment, the deep reinforcement learning algorithm training mainly adopts the PPO (Proximal Policy Optimization) algorithm to train the Actor-Critic main policy network to ensure the stability of the training process; at the same time, DQN (Deep Q-Network) is used as an auxiliary Q-value evaluation module for offline sample screening and value calibration. Specifically as follows: B231. Network parameter initialization: The weights are initialized using Xavier, and the biases are initialized with zero. Xavier initialization formula: W~U[-sqrt (6 / (n in +n out )), sqrt (6 / (n in + n out ))]; Where, n in n is the number of input neurons. out U represents the number of output neurons, and U is a uniform distribution.

[0099] B232. Training Hyperparameter Settings: Those skilled in the art can set the training hyperparameters according to actual needs. In this embodiment, the training hyperparameters are set as follows: Learning rate parameter: Actor network learning rate α actor =3e-4, Critic network learning rate α critic =1e-3, discount factor γ= 0.99; PPO parameters: pruning parameter ε = 0.2, value function coefficient c1 = 0.5, entropy regularization coefficient c2 = 0.01; DQN parameters: Experience replay pool size 100,000, batch size 64, target network update frequency once every 1,000 steps, soft update coefficient τ = 0.005; Training parameters: maximum number of training rounds 1000, number of samples collected per round 2048, number of optimization iterations 10, early stopping patience value 50 (stop if there is no improvement after 50 consecutive rounds).

[0100] B233, Model Training; The training data is divided into a training set, a validation set, and a test set. In this embodiment, the training set, validation set, and test set are divided in a ratio of 7:2:1. Those skilled in the art can set the division ratio according to actual needs.

[0101] The Actor-Critic master policy network is trained using the training set; The generalization of the model is verified by using the validation set and the test set. Once the accuracy requirements are met, the model weights are fixed, and the trained Actor network is output as the corresponding reinforcement learning model.

[0102] Those skilled in the art can build an experience replay pool to store training samples during the training process, set appropriate learning rates and discount factors, and adopt a multi-step reward design to balance immediate rewards and long-term benefits, thereby improving training stability and sample utilization efficiency.

[0103] This application also includes a production scheduling method that uses a pre-trained reinforcement learning model to obtain the initial solution, performs a global search using an evolutionary algorithm, and combines a heuristic algorithm for local fine-tuning optimization, balancing global optimization and local feasibility, to obtain the optimal scheduling scheme; see reference Figure 1 The specific steps are as follows: C100: The pre-built reinforcement learning model generates several initial scheduling schemes based on the current basic production data, and the obtained initial scheduling schemes are used as initial solutions. The input to the reinforcement learning model is the state vector corresponding to the current basic production data, and the output is the corresponding process selection probability distribution, equipment allocation probability distribution, and time arrangement probability distribution. Several feasible solutions are generated as initial solutions based on the obtained process selection probability distribution, equipment selection probability distribution, and time selection probability distribution. The specific steps are as follows: C110. Generate the corresponding state vector based on the current basic production data; That is, based on the above steps B110 to B140 and the definition of the state space in step B211, the current basic production data is converted into the state vector of the on-site scheduling scenario. Specifically: C111. Based on the current basic production data, perform sparse matrix transformation to obtain the corresponding order-process correlation matrix M. op Process-equipment matrices M fe Time window - equipment capacity matrix M tc Material-process matrix M can be used mp ; C112. Based on pre-built rule constraints, the process-equipment matrices M are processed. fe Time window - equipment capacity matrix M tc Update the system; simultaneously, based on pre-built rule constraints, calculate the replacement cleaning time and determine the corresponding replacement time matrix. C113, Based on the order-process association matrix M op Updated process-equipment processable matrix M fe Updated Time Window - Equipment Capacity Matrix M tc Material-process matrix M can be used mp Feature extraction is performed on the time matrix of the transformation to obtain the corresponding scheduling features.

[0104] C114. Construct a state vector S=[S] based on the obtained scheduling features. order ,S operation ,S equipment ,Stime ,S material ].

[0105] The specific implementation steps have been described in detail above, so they will not be repeated here.

[0106] C120. The reinforcement learning model outputs the corresponding process selection probability distribution, equipment selection probability distribution, and time selection probability distribution based on the state vector. C130. Based on the process selection probability distribution, equipment selection probability distribution and time selection probability distribution, several initial scheduling schemes are produced; In this embodiment, the action sequence output by the reinforcement learning model is converted into a complete scheduling scheme through a scheduling decoder to obtain the corresponding initial scheduling scheme. Specifically, sampling or greedy selection is performed based on the action probability distribution output by the reinforcement learning model to form an action sequence. The action sequence is then mapped to a complete scheduling code of "process sequence + equipment allocation + start time" by the scheduling decoder to obtain the initial scheduling scheme. As an example, the steps are as follows: C131. Sample from the corresponding probability distribution and select the process, equipment and time window to be scheduled; C132. Check the preceding processes, equipment availability, and material completeness. If not feasible, resample or select the next best action. C133. Assign the selected process to the selected equipment, set the start time, calculate the end time, and update the scheduling plan. C134. Remove scheduled processes and update equipment availability, material inventory, and current time; C135. Return to the complete scheduling plan with all processes arranged, and obtain the corresponding initial scheduling plan.

[0107] C140. Select an initial solution from the obtained initial scheduling scheme; Those skilled in the art can combine hard constraints and the fitness of subsequent evolutionary algorithms to select an initial solution from the initial scheduling scheme; For example: Perform hard constraint checks on the initial scheduling scheme, eliminate initial scheduling schemes that violate the constraints, or modify initial scheduling schemes that violate the constraints. The fitness of the initial scheduling schemes (including the modified initial scheduling schemes) that have passed the hard constraint verification is pre-evaluated. The initial solution is determined based on the evaluation results, such as the N initial scheduling schemes with the highest fitness as the initial solution, or the initial scheduling schemes with fitness greater than the preset fitness threshold as the initial solution.

[0108] Those skilled in the art can set their own initial solution screening rules according to actual needs, and this specification does not impose detailed limitations on them.

[0109] C200. Based on the initial solution, generate the corresponding initial population, obtain the fitness value of each individual in the initial population, and perform non-dominated sorting and crowding calculation. The specific steps are as follows: C210, Population initialization; In this embodiment, the initial solution is used as a high-quality seed, and a heuristic method is used to generate the initial population to ensure the quality and diversity of solutions; Those skilled in the art can set an appropriate population size as needed to balance computational efficiency and search capability.

[0110] C220, Fitness Assessment: For all individuals in the initial population, calculate the fitness value based on a preset multi-objective fitness function; In this embodiment, the fitness value F = [Fdelivery, Futilization, Fcost] is a vector composed of the delivery date achievement rate target Fdelivery, the resource utilization rate target Futilization, and the cost optimization target Fcost. Specifically: The delivery date achievement rate target is calculated as follows: F delivery =(1 / N)×Σdelivery score [i]; Where N is the total number of orders; delivery score [i] represents the delivery date score for order i, specifically: When completed on time, delivery score [i]=1; The delay was no more than 24 hours, delivery score [i] = 1 - Delay duration / 24; Delayed by more than 24 hours, delivery score [i] = 0.

[0111] The resource utilization target is calculated based on a weighted average of equipment utilization, personnel utilization, and material utilization, as follows: F utilization = w eq ×F equipment +w per ×F personnel +w mat ×F material ; Where w* represents the weight coefficient of each objective, and w eq +w per+w mat =1, in this embodiment, w eq =0.5、w per =0.3、w mat =0.2; F equipment For equipment utilization rate; F personnel For personnel utilization rate; F material Specifically, regarding material utilization rate: Equipment utilization rate F equipment The calculation method is as follows: F equipment = (1 / E)×Σ(actual work time [k] / available time [k]); Where E represents the total number of devices; actual work time [k] represents the actual operating time of device k; available time [k] represents the available time of device k.

[0112] Personnel utilization rate F personnel The calculation method is as follows: F personne l = (1 / W)×Σ(actual work time [w] / available time [w]); Where W represents the total number of personnel; actual work time [w]: The actual working hours of personnel w; available time [w] represents the available time for person w.

[0113] Material utilization rate F material The calculation method is as follows: F material = (1 / P)×Σ(actual usage [μ] / planned usage [μ]); Where P is the number of material types; actual usage [μ] represents the actual usage of material μ; planned usage [μ] represents the planned usage of material μ.

[0114] Calculate the corresponding cost optimization objective F based on the total production cost and cost budget. cost The calculation method is as follows: F cost =1- (C total / C budget ), Among them, C totalC represents the total production cost. budget For cost budgeting.

[0115] In this embodiment, the total production cost includes material cost C. material Labor costs C labor Energy cost C energy Changeover cost C setup Inventory cost C inventory Delay penalty cost C penalty The calculation formula is as follows: C total =C material +C labor +C energy +C setup +C inventory +C penalty ; Specifically: C material =Σ(material unit price × material usage); C labor =Σ(labor hourly rate × number of working hours); C energy =Σ(Energy unit price × Energy consumption); C setup =Σ (Changeover cost × Number of changeovers); C inventory =Σ(Inventory holding cost × Inventory quantity × Holding time); C penalty =Σ(Delay penalty × Delay time).

[0116] C230. Based on the fitness value, the Pareto optimal solution pool maintenance algorithm is used to perform non-dominated sorting and crowding calculation. The specific steps are as follows: C231. Initialize the maximum capacity of the solution pool. In this embodiment, the maximum capacity is set to 100. C232. Non-dominance relationship judgment: If a new individual is dominated by any individual in the solution pool, it will not be added to the solution pool; if a new individual dominates some individuals in the solution pool, the dominated individuals will be removed. C233, Pool capacity control: If the pool capacity exceeds the maximum pool capacity, calculate the crowding distance of each solution, remove the individual with the smallest crowding distance, until the capacity meets the requirements; C234. Return to the updated solution pool to obtain the Pareto optimal solution.

[0117] The dominance relationship is defined as solution A dominating solution B if and only if the fitness values ​​A for all objectives are not inferior to B, and at least one objective has a fitness value A that is superior to B. The objectives are delivery date achievement rate (Fdelivery), resource utilization rate (Futilization), and cost optimization (Fcost).

[0118] The formula for calculating congestion distance is as follows: CD (S i ) = Σ (F k (S {i+1} ) - F k (S {i-1} )) / (F kmax - F k min ); Among them, S i S is the i-th solution in the Pareto solution pool (after non-dominated sorting); {i+1} , S {i-1} For adjacent solutions; F k For the k-th objective dimension (delivery time achievement rate objective, resource utilization rate objective, cost optimization objective), F k max For the corresponding target F k The maximum value, F k min For the corresponding target F k The minimum value; In this embodiment, the crowding distance of the boundary solutions is set to infinity. The larger the crowding distance, the sparser the solution is on the Pareto front, and it should be retained preferentially. Specifically, it is expressed as follows: CD(S1) = CD(S n ) = ∞.

[0119] C300: Perform iterative optimization of the initial population using an evolutionary algorithm. When the evolutionary algorithm iteration is complete, output the optimal scheduling scheme corresponding to the Pareto optimal solution pool. In this embodiment, as Figure 2 As shown, the diverse feasible initial solutions generated by reinforcement learning are used as high-quality seeds. A global search is performed through a dedicated crossover / mutation operator, followed by targeted repair through a heuristic local search. Then, the environment is selected based on Pareto order and crowding, and the constraint violation statistics are fed back to the next generation operator parameters to obtain the optimal scheduling scheme. Reference Figure 2 Each iteration step performs the following steps: C310, dual-championship selection; Based on fitness values, a binary tournament selection method is used to select parent individuals from the current population in the current iteration step. In this embodiment, diversity is maintained by combining crowding distance and selecting superior individuals as parents to balance selection pressure and population diversity. C320, crossover operation; Perform a crossover operation on the parent individuals to obtain the corresponding child individuals; Crossover operations can be performed using methods such as single-point crossover or multi-point crossover. In this preferred embodiment, a dedicated crossover operator for scheduling scenarios is used to perform crossover operations on the selected parent individuals to generate child individuals. In this embodiment, the crossover operation adopts order crossover (OX) and equipment allocation crossover; in order to maintain population diversity, the initial crossover probability is set to 0.85, and a higher crossover probability is maintained in the early stage of the population to expand the search range; In this embodiment, the term "dedicated" means that the constraint checker is called before and after the operator is executed to ensure that key constraints such as process sequence, equipment capacity and material availability are not violated, and the proportion of infeasible solutions is controlled through a penalty repair strategy. Among them, the special cross operators include the process sequence cross operator that maintains process sequence constraints and the equipment allocation cross operator that maintains equipment capacity constraints; C321, Process sequence crossover operator; This embodiment uses the existing publicly available process sequence crossover operator, applies it to the scheduling scenario, and adds process constraint checks; P1 represents the first parent individual, P2 represents the second parent individual, C1 represents the first child individual, and C2 represents the second child individual. The specific algorithm steps are as follows: In the process sequences of P1 and P2, two intersection points pos1 and pos2 (pos1 < pos2) are randomly selected to determine the intersection region [pos1, pos2]. Copy the intersection region, C1[pos1:pos2] = P1[pos1:pos2], C2[pos1:pos2] = P2[pos1:pos2]; Extract the processes from P2 that do not appear in the intersection area of ​​C1 in sequence, and fill them into the remaining positions of C1; extract the processes from P1 that do not appear in the intersection area of ​​C2 in sequence, and fill them into the remaining positions of C2. Check whether the offspring individuals meet the process sequence constraints. If they do not, identify the violating process pair (opi, opj), adjust the process sequence to ensure that the preceding process comes first, and repeat the check until all constraints are met. Return the offspring individuals C1 and C2.

[0120] C322, Equipment Allocation Cross Operator; The specific algorithm steps are as follows: Extract the device allocation sequence E1 = [e11, e12, ..., e1] from P1. mExtract the device allocation sequence E2 = [e21, e22, ..., e2] from P2. m ]; The equipment corresponding to each process j (j=1,2,...,m) in the child generation is determined based on the following steps: Get the random value corresponding to the current process. When the corresponding random value is less than the equipment crossover probability (0.5), C1 inherits the equipment allocation of P1 and C2 inherits the equipment allocation of P2. When the corresponding random value is greater than or equal to the equipment crossover probability, C1 inherits the equipment allocation of P2 and C2 inherits the equipment allocation of P1. Equipment capacity constraint check: Check whether the assigned equipment is in the list of processable equipment. If not, randomly select the equipment with the lowest load rate from the list of processable equipment. After all the equipment for each process is determined, the corresponding offspring individuals C1 and C2 are returned.

[0121] Example: Parent P1 device allocation: [1, 2, 1, 3, 2, 1, 3, 2, 1] Parent P2 device allocation: [2, 1, 2, 1, 3, 2, 1, 3, 2] Cross mask (randomly generated): [1, 0, 1, 0, 1, 0, 1, 0, 1] (1 indicates inheritance from P1, 0 indicates inheritance from P2) Child C1 device allocation: [1, 1, 2, 1, 3, 2, 1, 3, 1] Child C2 device allocation: [2, 2, 1, 3, 2, 1, 3, 2, 2] Equipment capacity constraint check: Assuming the processable equipment for process 3 is [1, 2], process 3 in C1 is assigned to equipment 2 (which satisfies the constraints). Assuming the workable equipment for process 4 is [1, 3], process 4 in C1 is assigned to equipment 1 (meeting the constraints). As one possible implementation, after completing the equipment capacity constraint check, the equipment load balance check step is also included: calculate the load rate of the equipment that passed the check (located in the list of processable equipment), and if the load rate of a certain equipment exceeds the preset load rate threshold (0.9), then select the equipment with the lowest load rate from the list of processable equipment for replacement; As one possible implementation method, the offspring individuals obtained from the crossover can be directly used as new individuals.

[0122] C330, Mutation Operation: Mutation operations can be performed using methods such as random mutation and Gaussian mutation. Preferably, this embodiment uses a dedicated mutation operator for scheduling scenarios to perform mutation operations on the offspring individuals after crossover to generate new individuals. In this embodiment, the mutation operation is designed to address the characteristics of scheduling problems, including process exchange and equipment reallocation. To maintain population diversity, the initial mutation probability is set to 0.10, and the mutation probability is adaptively increased to 0.15 in the later stages of convergence to avoid premature convergence. In this embodiment, the term "dedicated" means that the constraint checker is called before and after the operator is executed to ensure that key constraints such as process sequence, equipment capacity and material availability are not violated, and the proportion of infeasible solutions is controlled through a penalty repair strategy. Specialized mutation operators include at least: The bottleneck process shifting variation operator identifies and moves bottleneck processes forward to reduce waiting time and shorten the overall production cycle of orders. This is a batch merging variation operator for the same process family that combines processes into batches, reducing the number of changeovers and cleaning time. A low-load equipment reallocation mutation operator that identifies low-load equipment, reallocates processes on it, balances equipment load, and improves overall resource utilization. The specific algorithm steps for the bottleneck process shifting mutation operator (C331) are as follows: Based on the waiting time, downstream impact, and critical path criticality of each process in the individual to be mutated, the bottleneck index of the corresponding process is calculated, and the process with the largest bottleneck index is selected as the bottleneck process op. bottleneck ; The bottleneck index is calculated as follows: bottleneck index [j]=w1×wait time [j]+w2×downstream impact [j]+w3×criticality[j]; Among them, wait time [j] represents the waiting time for process j; downstream impact [j] represents the downstream impact (total waiting time of subsequent processes); criticality[j] represents the criticality (whether it is on the critical path; criticality=1 for processes on the critical path, otherwise=0); w* represents the weight coefficient of each objective, and w1+w2+w3=1. In this embodiment, w1=0.4, w2=0.4, and w3=0.2. Obtain the set of preceding processes for the bottleneck process and calculate the earliest possible start time. start = max(previous process end time), determines the forward time window.window = current start_time - earliest start , where current start_time This represents the start time of the current bottleneck process within the existing scheduling scheme.

[0123] Within the shiftable time window, a new start time new_start is randomly selected, where new_start ∈ [earliest_start, current_start_time]; Update the timing of the bottleneck process based on the new start time, and simultaneously adjust the timing of subsequent processes and their new start times. Check equipment availability, material availability, and process sequence constraints; if any are violated, roll back the variation operation. Returns the mutated individual.

[0124] Those skilled in the art can set equipment availability, material availability, and process sequence constraints according to actual needs.

[0125] The specific algorithm steps for the batch merging mutation operator of the same process family (C332) are as follows: Each process is clustered in advance using a process similarity calculation formula to form a process family set; The process similarity is calculated as follows: similarity (op i , op j = w1 × process parameter similarity + w2 × equipment similarity + w3 × material similarity; Wherein, process parameter similarity = number of identical process parameters / total number of process parameters; equipment similarity = number of identical processable equipment / total number of processable equipment; material similarity = number of identical materials / total number of materials; w* is the weight coefficient of each objective, and w1+w2+w3=1. In this embodiment, w1=0.5, w2=0.3, and w3=0.2. When similarity(op) i , op j )≥similarity threshold When determined to be from the same process family, in this embodiment, similarity threshold Take 0.7.

[0126] Check whether the processes within the process family are distributed across different time periods in the individual to be mutated, and calculate the corresponding combined benefit. i = Changeover time savings - Time window loss; if the benefit is greater than the preset benefit threshold, it is marked as mergeable. The time windows corresponding to the processes within the process family are statistically analyzed. Based on the statistical results, a benchmark time window (the most concentrated time window) is selected. Based on the benchmark time window, the processes within the same process family are arranged (as close as possible to the benchmark time window), and the process sequence is adjusted to ensure that process constraints are met. Check equipment, materials, and delivery deadlines; if any violations are found, partially roll back and merge the operations. Return the mutated individual mutated .

[0127] Those skilled in the art can set equipment, material, and delivery constraints according to actual needs.

[0128] The specific algorithm steps for the C333 low-load equipment reallocation mutation operator are as follows: Calculate the load rate of each device in the individual to be mutated, and identify the set of low-load devices whose load rate is lower than the preset low load rate threshold (0.3); The process of low-load equipment is reassigned. In this embodiment, if there are other processable equipment for the process, the process is marked as reassignable and the corresponding processable equipment is used as candidate equipment. For reassignable processes, calculate the fit degree of each candidate device, and select the candidate device with the highest fit degree as the target device for process migration. The device compatibility calculation method is as follows: score (k,j) = w1×(1-load rate [k])+w2×time window fitness (k,j)+w3×(1-setup cost (k,j)); Among them, load rate [k] represents the current load rate of device k (the lower the load rate, the higher the adaptability); time window fitness (k,j) represents the time window fit (whether device k has a suitable time window to accommodate process j); setup cost (k,j) represents the changeover cost (the changeover cost of processing step j on equipment k); w* represents the weight coefficient of each objective, and w1+w2+w3=1. In this embodiment, w1=0.4, w2=0.4, and w3=0.2. Among them, time window adaptation window fitness (k, j) is calculated as follows: Query and calculate the available capacity of equipment k within the time window of process j; Compare the available production capacity with the processing time of process j. If the available capacity is greater than or equal to the processing time of process j, the time window fit is 1; otherwise, the time window fit is the ratio of available capacity to process processing time. Check the target equipment's capacity, time window, and material constraints; if any are violated, cancel the reallocation of that process. Iterate until there are no low-load devices; Return the mutated individual mutated .

[0129] Those skilled in the art can set the target equipment's capabilities, time windows, and material constraints according to actual needs.

[0130] C340, Heuristic Local Search: For the newly obtained individuals, perform a heuristic local search; Those skilled in the art can select existing heuristic local search algorithms according to actual needs and perform local searches based on fixed heuristic local search algorithms; Prior to this embodiment, a heuristic local search is performed for each new individual. During the heuristic local search process, based on a preset improvement rate determination mechanism, the algorithm dynamically switches between tabu search, step-counting hill climbing, and big flood algorithm to iteratively obtain the optimal solution corresponding to each new individual, that is, to obtain the optimized new individual. The new individual to be heuristically searched is taken as the target individual. Each target individual performs several sub-iterations to obtain the corresponding optimal solution and updates the Pareto optimal solution pool. When the evolutionary algorithm completes the iteration, it outputs the optimal scheduling scheme corresponding to the Pareto optimal solution pool. Each sub-iteration step includes the following steps: C341. Calculate the current improvement rate; The improvement rate is the sum of the improvement rates for each round within the sliding window; The improvement rate is a weighted combination of the improvement magnitude of the objective function and the decrease magnitude of constraint violations in the corresponding round, used to quantitatively evaluate the optimization effect of the local search algorithm used within the sliding window; The calculation method is as follows: Improvement Rate window =(1 / W)×ΣImprovement Rate [ti]; Improvement Rate =α×Objective Improvement +β×Constraint Reduction ; Where W is the sliding window size (default W=5); t is the current round; and i is the index of the historical round (i=1,2,...,W).

[0131] ObjectiveImprovement Improvement magnitude of the objective function; Constraint Reduction To constrain the rate of decrease in violations; α is the weight of the improvement magnitude of the objective function, β is the weight of the rate of decrease in violations, and α+β= 1. In this embodiment, α=0.6, β=0.4.

[0132] Among them, the improvement magnitude of the objective function is calculated based on the optimal objective function value of the previous round and the optimal objective function value of the current round. Improvement The calculation method is as follows: Objective Improvement =(F best new -F best old ) / max(|F bestold |,ε); Among them, F best old F is the optimal objective function value from the previous round; best new The objective function value for the current round is ε; ε is a very small positive number to prevent the denominator from being zero, and in this embodiment, it is taken as 1e. -8 ; The objective function F is calculated based on the delivery date achievement rate, resource utilization rate, and cost ratio, as follows: F = w1 × delivery date achievement rate + w2 × resource utilization rate + w3 × (1 - cost percentage); Where F is the scalar target for monitoring the improvement rate, which is used in parallel with the Pareto vector target. The delivery date achievement rate corresponds to the delivery date achievement rate target, and the resource utilization rate corresponds to the resource utilization rate target.

[0133] Among them, the decrease in constraint violations Constraint Reduction The calculation method is as follows: Constraint Reduction =(Pold-P new ) / max(P old ,ε); Among them, P old P represents the total number of constraint violations in the previous round. new The total number of constraint violations in the current round; constraint violation P = Σ(number of violated constraints × constraint severity); ε is a very small positive number to prevent the denominator from being 0, and in this embodiment, it is taken as 1e. -8 ; C342. Make a switching decision based on the current improvement rate; If the improvement rate during the window period is greater than or equal to the high improvement rate threshold (5%), it indicates that the improvement rate is good. The current algorithm is maintained, that is, the current algorithm is used as the target local search algorithm to maintain the optimization rhythm. If the improvement rate during the window period is less than the low improvement rate threshold (1%), it indicates that the improvement rate is low. At this time, the counter of consecutive low improvement rounds is accumulated. If the counter is greater than or equal to the switching threshold, the local search algorithm used is switched based on the preset dynamic switching rules. The dynamic switching rule is used to indicate the target local search algorithm corresponding to the currently used local search algorithm.

[0134] The current algorithm is a step-counting hill-climbing algorithm; switch to the big flood algorithm to escape local optima. If the current algorithm is tabu search, switch to the Great Flood algorithm and reset the search direction. The current algorithm is a large flood, so we will switch to a step-counting hill-climbing algorithm for local fine-tuning. After switching algorithms, reset the counter to 0; If the improvement rate during the window period is between the low improvement rate threshold and the high improvement rate threshold, it indicates that the improvement rate is moderate, and the current algorithm is maintained, that is, the current algorithm is used as the target local search algorithm; The target individual is optimized based on the target local search algorithm, and the target individual is updated based on the optimization results.

[0135] As one possible implementation, the switching judgment is also based on the number of constraint violations. In this embodiment, if the number of constraint violations of the current solution is greater than the constraint violation threshold (set by those skilled in the art), it indicates that the constraint violation is serious, and the feasible repair is performed first (prior to the switching judgment based on the improvement rate), and the algorithm is switched to the tabu search algorithm. That is, the tabu search algorithm is directly used as the target local search algorithm.

[0136] As one possible implementation, the switching history is recorded: storing "iteration round, current algorithm, next algorithm, sliding window improvement rate, number of constraint violations, and switching reason" for tracking the algorithm optimization effect; C343. Termination condition judgment: The local search stops if any of the following conditions are met: The iterations reached the maximum number of iterations. The improvement rate is less than the low improvement rate threshold for three consecutive rounds and the number of constraint violations is 0; The number of constraint violations is 0 and the objective function value enters the Pareto front.

[0137] As one feasible implementation, the large flood algorithm is improved by introducing a water level reduction strategy; The improved flood algorithm is as follows: The Big Flood Algorithm: This algorithm resets the search direction by lowering the water level, preventing it from getting trapped in local optima and allowing it to escape local optima. The specific steps are as follows: From the process exchange neighborhood, equipment reallocation neighborhood, and time window adjustment neighborhood, randomly select one neighborhood type and generate one random neighborhood solution; Calculate the objective function value and the number of constraint violations for the neighborhood solution; If the objective function value of the neighborhood solution is greater than or equal to the current water level and the constraint violation count of the neighborhood solution is 0, then the neighborhood solution is accepted, and the current solution is updated to the neighborhood solution; if the objective function value of the current solution is better than the optimal solution, then the optimal solution is updated; in this embodiment, the initial water level is set to 90% of the objective function value of the initial solution. Each time the water level is updated, the current water level is multiplied by the rate of decline in each round to obtain the new current water level; in this embodiment, the rate of decline in each round is 0.5%; If the current water level is lower than the objective function value of the optimal solution, then the water level has dropped below the optimal solution, and the search is terminated prematurely. As one possible implementation method, the solution in the neighborhood is obtained based on the scheduling-specific neighborhood operation in tabu search, step-counting hill climbing, and big flood algorithms; The scheduling-specific neighborhood operations include process exchange neighborhood, equipment reallocation neighborhood, and time window adjustment neighborhood; Tabu search obtains a neighborhood solution set based on process exchange neighborhood, equipment reallocation neighborhood, and time window adjustment neighborhood; The step-counting hill-climbing and big flood algorithms randomly select a scheduling-specific neighborhood operation to generate the corresponding neighborhood solution; Specifically: The definition of the process exchange neighborhood is: for a scheduling scheme, the set of all scheduling schemes obtained by exchanging the positions of any two different processes in the scheme; the constraints are that the process sequence of the corresponding order must be satisfied after the exchange, and the processing time of the two processes on the allocated equipment does not overlap; the taboo attribute is the exchanged process pair; the neighborhood size is m×(m-1) / 2, where m is the total number of processes in the scheduling scheme; The definition of the equipment reallocation neighborhood is: for a scheduling scheme, the set of all scheduling schemes obtained after replacing the processing equipment of any operation in the scheme with other processing equipment of that operation; the constraints are that other processing equipment must come from the list of processing equipment corresponding to the operation-equipment processing matrix, and have available capacity during the planned processing period of the operation; the taboo property is the combination pair of operation and newly allocated equipment; the neighborhood size is Σ(number of processing equipment - 1).

[0138] The time window adjustment neighborhood is defined as: for a scheduling scheme, the set of all scheduling schemes obtained after adjusting the processing time window of any operation in the scheme from the original time window to other available time windows of that operation; its constraints are that other available time windows must come from the list of available time windows corresponding to the time window-equipment capacity matrix, and satisfy the following conditions: the completion time of the preceding operation is less than or equal to other available time windows, the operation processing time change time, and the start time of the subsequent operation; the taboo attribute is the combination pair of operation and the newly adjusted time window; the neighborhood size is m×T, where m is the total number of operations and T is the total number of time windows in the scheduling period.

[0139] C350, fitness assessment, and environment selection: For all optimized individuals, their fitness values ​​are determined, and Pareto fronts are identified through fast non-dominated sorting. Individuals are then selected by combining crowding distance to generate a new generation of population. At the same time, it maintains the Pareto optimal solution pool, provides multiple high-quality scheduling schemes, and supports planners to select the optimal scheme according to actual needs; In this embodiment, the optimal solution of the current Pareto front is directly retained to the next generation; C360. Termination Condition Judgment: If any of the following conditions are met, the evolutionary algorithm iteration is considered complete, and the optimal scheduling scheme in the Pareto solution pool is output: The iterations reached the maximum number of iterations. Determine whether the algorithm has converged based on the convergence threshold; The computation time limit is reached to ensure that the algorithm completes within a reasonable time.

[0140] C400: Based on the optimal scheduling scheme, perform full-dimensional hard constraint verification and violation repair to obtain the target scheduling scheme; For each optimal scheduling scheme, select it as a candidate scheduling scheme and perform the following steps: C410, Load candidate scheduling schemes and the corresponding hard constraint set and sparse matrix, and initialize the violation list; C420: Perform full-dimensional hard constraint verification and record all violation information; The comprehensive hard constraint verification includes process sequence constraint verification, equipment capacity constraint verification, equipment time window constraint verification, material availability constraint verification, personnel qualification constraint verification, and equipment maintenance constraint verification. Specifically: Process sequence constraint verification is a constraint verification that ensures that the processes are executed strictly in the order of the preset process route. The constraint condition is that if process A is the preceding process of process B, then the start time of process B is not earlier than the end time of process A. The verification method is to traverse the process route of each order and check the time logic of all preceding and following processes. The violation is handled by adjusting the start time of the subsequent process to ensure that the subsequent process is started only after the preceding process is completed. Equipment capacity constraint verification is a constraint verification to ensure that equipment can complete the assigned process within a specified time. The constraint condition is that the process can only be processed on equipment within its set of processable equipment. The verification method is to traverse all processes and check whether the assigned equipment is in the process-equipment processable list corresponding to the process-equipment processable list. Violation is handled by reassigning the equipment to the equipment with the lowest load rate in the processable equipment list. Equipment time window constraint verification is a constraint verification to ensure that the execution period of the process falls within the available time window of the equipment. The constraint condition is that adjacent processes on the same equipment must meet the requirement that the sum of the end time of the preceding process and the changeover time does not exceed the start time of the subsequent process. The verification method is to collect the processes by equipment and sort them by start time, and check whether the time interval between adjacent processes meets the requirements of changeover and non-overlap. Violation is handled by adjusting the start time of the subsequent process or moving it to other processable equipment. Material availability constraint verification is a constraint verification that ensures that all types of materials required when a process starts are available. The constraint condition is that the available quantity of materials required when a process starts is not less than the quantity required by the process. The verification method is to traverse all processes, calculate the available inventory of materials at the start time of the process and check whether the requirements are met. Violation is handled by postponing the start time of the process until the materials are available or adjusting the material procurement plan. Personnel qualification constraint verification is to ensure that the personnel performing the process have the corresponding qualifications, and that the scheduling time matches the process execution time. The constraint condition is that the personnel performing the process must have the qualifications required for the process and the scheduling time covers the process execution time. The verification method is to traverse all processes and check whether the qualifications of the assigned personnel meet the process requirements and whether the scheduling time matches. The violation is handled by reassigning personnel with the corresponding qualifications and matching scheduling time. Equipment maintenance constraint verification is a constraint verification that ensures the execution time of a process avoids the planned maintenance time period of the equipment and that the current maintenance status of the equipment meets the requirements of the process operation. The constraint condition is that the process must not be scheduled within the planned maintenance time period of the equipment. The verification method is to traverse the maintenance windows of all equipment and check whether the process time overlaps with the maintenance window. The violation is handled by adjusting the process time to avoid the maintenance window or reassigning the equipment.

[0141] C430. Verification Result Judgment: If all hard constraints are satisfied, the verification passes, and the scheduling scheme is output; if there are constraint violations, the constraint repair phase begins. After repair, the verification is re-executed. The specific steps are as follows: C431. Initialize violation information, create a violation list and set it to an empty list, and initialize the validation pass flag to indicate that all constraints are satisfied. C432. Perform six types of hard constraint checks in the following order: process sequence constraints, equipment capacity constraints, equipment time window constraints, material availability constraints, personnel qualification constraints, and equipment maintenance constraints. After each type of constraint check is completed, if a violation is detected, enter the process number, violation type, and violation details of the violation into the violation list, and set the check pass flag to "No". C433. Output the results of this full-dimensional hard constraint verification, including the verification pass flag, violation list, and statistics on the number of violations for each type of constraint; C434. Sort all violations in the violation list according to the severity of the constraint, with hard constraint violations taking precedence over soft constraint violations; C435. Perform the corresponding constraint violation repair operation according to the type of violation; C436. After completing the repair of all violations, re-execute the full-dimensional hard constraint verification to generate new verification results; C437. If the verification pass flag after re-verification is "Yes", then output the repaired final scheduling scheme; if the verification pass flag is still "No", then return the current scheduling scheme to the main engine of the evolutionary algorithm and re-execute the global iterative optimization.

[0142] C500, based on the final scheduling scheme, outputs an executable minute-level production scheduling plan and supporting decision support content; The minute-level production scheduling plan includes information such as process scheduling Gantt charts, supply chain network overview, equipment allocation, and time arrangements; The system outputs the formatted production schedule to the user interface or a file for use by production managers.

[0143] The supporting decision support content includes simulated scheduling comparison and what-if simulation analysis, scheduling result interpretation provided by AI assistant, and intelligent decision support; Among them, the simulated scheduling comparison and what-if simulation analysis are used to construct disturbance scenarios such as order insertion, equipment failure, raw material delays, and shift changes without affecting the online execution plan, and output the changes in indicators such as delivery rate, work-in-process, changeover frequency, and energy consumption under different solutions. The AI ​​assistant module is used to explain the reasons for the scheduling results (such as key bottleneck equipment, critical path processes, and root causes of delays) and provide actionable adjustment suggestions based on the Pareto solution pool (such as recommending outsourced orders, recommending changeover sequences, and recommending overtime windows).

[0144] As the device embodiment is basically similar to the method embodiment, the description is relatively simple, and relevant parts can be found in the description of the method embodiment.

[0145] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0146] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0147] This invention is described with reference to flowchart illustrations and / or block diagrams of the method, terminal device (system), and computer program product according to the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0148] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0149] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0150] It should be noted that: The phrase "an embodiment" or "an embodiment" used in this specification means that a particular feature, structure, or characteristic described in connection with the embodiment is included in at least one embodiment of the invention. Therefore, the phrase "an embodiment" or "an embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment.

[0151] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0152] Furthermore, it should be noted that the shapes and names of the parts and components described in the specific embodiments described in this specification may differ. All equivalent or simple variations made to the structure, features, and principles described in this patent concept are included within the protection scope of this patent. Those skilled in the art to which this invention pertains may make various modifications or additions to the described specific embodiments or use similar methods to replace them, as long as they do not depart from the structure of this invention or exceed the scope defined in these claims, they should all fall within the protection scope of this invention.

Claims

1. A production scheduling method based on a hybrid algorithm, characterized in that, Includes the following steps: The pre-built reinforcement learning model generates several initial scheduling schemes based on the current basic production data, and generates an initial solution based on the obtained initial scheduling schemes. The input to the reinforcement learning model is the state vector corresponding to the current basic production data, and the output is the corresponding process selection probability distribution, equipment allocation probability distribution, and time arrangement probability distribution. Based on the initial solution, a corresponding initial population is generated, the fitness value of each individual in the initial population is obtained, and non-dominated sorting and crowding calculation are performed. The initial population is iteratively optimized using an evolutionary algorithm. Upon completion of the iterations, the algorithm outputs the optimal scheduling scheme corresponding to the Pareto optimal solution pool. The iterative steps of the evolutionary algorithm are as follows: Based on fitness values, a binary tournament selection method is used to select parent individuals from the current population in the current iteration step. Based on the parent individuals, crossover and / or mutation operations are performed to obtain corresponding new individuals; Perform a heuristic local search on the newly obtained individuals to obtain optimized individuals; For each optimized individual, its fitness value is determined, and the Pareto front is identified by fast non-dominated sorting. Individuals are screened by combining crowding distance to generate a new generation of population and update the Pareto optimal solution pool.

2. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that: The fitness value is a vector composed of the delivery date achievement rate target, resource utilization rate target, and cost optimization target; in: The resource utilization target is calculated based on a weighted average of equipment utilization, personnel utilization, and material utilization. The cost optimization objective is calculated based on the total production cost and the cost budget.

3. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that: Based on the probability distributions of process selection, equipment selection, and time selection, sampling or greedy selection is performed to form a corresponding action sequence; The sequence of actions is mapped to a scheduling code that combines process sequence, equipment allocation, and start time; Based on the obtained scheduling code, several initial scheduling schemes are generated.

4. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that: The crossover operation includes a process sequence crossover operator that maintains process sequence constraints and an equipment allocation crossover operator that maintains equipment capacity constraints. The process sequence crossover operator uses process sequence crossover and performs process sequence constraint checks. The equipment allocation crossover operator determines the equipment allocation of the offspring to inherit from the parent according to a preset equipment crossover probability and performs equipment capacity constraint checks. If a violation is found, a replacement device is selected from the list of processable equipment.

5. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that, The mutation operations include bottleneck process forward mutation operators, batch merging mutation operators of the same process family, and low-load equipment reallocation mutation operators.

6. The production scheduling method based on a hybrid algorithm according to any one of claims 1-5, characterized in that, A heuristic local search is performed on the newly obtained individuals, and several sub-iterations are performed to obtain the optimized individuals. The heuristic local search algorithm adopts tabu search, step-counting hill climbing or big flood algorithm. A sub-iteration step includes the following steps; Calculate the current improvement rate, which indicates the optimization effect of the local search algorithm used in the current sliding window; When the current improvement rate is greater than or equal to a preset low improvement rate threshold, the currently used local search algorithm is used as the target search algorithm. When the current improvement rate is less than the preset low improvement rate, it indicates that the improvement rate is low. At this time, the number of consecutive low improvement rounds is accumulated by a counter. If the counter is greater than or equal to the switching threshold, the target local search algorithm is determined based on the preset dynamic switching rule and the counter is cleared. The dynamic switching rule is used to indicate the target local search algorithm corresponding to the currently used local search algorithm. The target individual is optimized based on the target local search algorithm, and the target individual is updated based on the optimization results.

7. The production scheduling method based on a hybrid algorithm according to claim 6, characterized in that, The dynamic switching rules include: When the current local search algorithm is the step-counting hill-climbing algorithm, switch to the big flood algorithm; Switch to the Big Flood algorithm when the current local search algorithm is the tabu search algorithm; When the current local search algorithm is the flood algorithm, switch to the step-counting hill climbing algorithm.

8. The production scheduling method based on a hybrid algorithm according to claim 6, characterized in that, Before calculating the current improvement rate, a switching decision is made based on the number of constraint violations: If the number of constraint violations in the current solution is greater than the constraint violation threshold, the tabu search algorithm is directly used as the target local search algorithm.

9. The production scheduling method based on a hybrid algorithm according to claim 6, characterized in that: The current improvement rate is the sum of the improvement rates of each round within the sliding window; The improvement rate is a weighted combination of the improvement in the objective function and the decrease in constraint violations for the corresponding round: Specifically, the objective function values ​​are calculated based on the delivery date achievement rate, resource utilization rate, and cost ratio; the delivery date achievement rate corresponds to the delivery date achievement rate target, and the resource utilization rate corresponds to the resource utilization target; the improvement margin of the objective function is calculated based on the previous round's optimal objective function value and the current round's optimal objective function value.

10. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that, The steps for generating the corresponding state vector based on the current basic production data are as follows: Based on the current basic production data, a sparse matrix transformation is performed to obtain the corresponding order-process association matrix, process-equipment processability matrix, time window-equipment capacity matrix, and material-process availability matrix. Based on pre-built rule constraints, the process-equipment matrices and time window-equipment capacity matrices are updated; at the same time, based on pre-built rule constraints, the changeover cleaning time is calculated, and the corresponding changeover time matrix is ​​determined. Feature extraction is performed based on the order-process association matrix, the updated process-equipment processability matrix, the updated time window-equipment capacity matrix, the material-process availability matrix, and the changeover time matrix to obtain the corresponding scheduling features; Based on the obtained scheduling features, construct a state vector S=[S order ,S operation ,S equipment ,S time ,S material ], where S order S is the order layer state vector; operation S is the process layer state vector; equipment S is the device layer state vector; time S is the time-level state vector; material This is the state vector of the material layer.

11. The production scheduling method based on a hybrid algorithm according to claim 1, characterized in that: The scheduling problem is modeled as a Markov decision process, and a multi-objective reward function is constructed that includes rewards for meeting delivery deadlines, resource utilization, cost optimization, quality assurance, and penalties for constraint violations. An Actor-Critic network is used as the main policy network for the Markov decision process, where the Actor network is responsible for outputting action policies and the Critic network is responsible for evaluating state values ​​and guiding policy updates. The output layer of the Actor network includes a process selection branch, an equipment allocation branch, and a time scheduling branch, which output the process selection probability distribution, equipment selection probability distribution, and time window selection probability distribution, respectively. Action masks are applied to the above three types of probability distributions, and the probability of inactive actions is set to zero and then normalized.

12. The production scheduling method based on a hybrid algorithm according to claim 11, characterized in that: Based on the multi-objective reward function, the PPO algorithm is used to train the main policy network, and the DQN algorithm is used as an auxiliary Q-value evaluation module for offline sample screening and value calibration. The DQN algorithm uses an experience replay mechanism to sample and update training samples, and a target network update mechanism is used to improve training stability. The trained Actor network is used as a reinforcement learning model.

13. The production scheduling method based on a hybrid algorithm according to claim 10, characterized in that: The preset rules and constraints include overlapping and free-floating rules for process lines, collaborative production rules between product categories, three-dimensional cleaning and changeover rules for processes / products / equipment, and centralized production rules for specific work areas; The process flow line overlap free float rule is used to determine the allowable overlap time, and the actual start time is calculated based on the allowable overlap time; The inter-category collaborative production rules are used to determine products belonging to the same process family; The process / product / equipment three-dimensional cleaning and replacement rules are used to calculate the replacement cleaning time based on the three dimensions of process type, product characteristics and equipment status, and generate the corresponding replacement time matrix. The specific workspace centralized production rules are used to indicate centralized production constraints and to determine the corresponding penalty coefficients based on the number of workspace switching times, environmental stabilization waiting time, and the number of cleanliness violations.

14. A production scheduling system based on a hybrid algorithm, characterized in that, Used to perform the production scheduling method according to any one of claims 1-13.

Images