Knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method and system

Abstract

The invention discloses a multi-factory collaborative manufacturing dynamic scheduling method and system based on knowledge driving, and belongs to the technical field of intelligent manufacturing and optimal scheduling. According to the method, a dynamic multi-target mathematical model which simultaneously considers the equipment fault, the performance degradation and the adjustable processing time is established, and the minimization of the maximum completion time and the minimization of the total energy consumption are taken as optimization targets; generating a high-quality initial population by adopting a hybrid initialization strategy fusing various heuristic rules and a random generation mode; in the optimization iteration process, environment change detection is executed, equipment abnormity and machining rate fluctuation are recognized in real time, a static optimization strategy or a dynamic response strategy is adaptively selected according to a detection result, and rapid adjustment of emergencies is achieved. The static optimization strategy comprises specific crossover and mutation operators, adaptive neighborhood search and energy-saving post-processing; the dynamic response strategy comprises elite demigration, neighborhood restoration and a diversity injection mechanism. Scheduling stability, energy efficiency and response speed are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and production scheduling technology, and in particular relates to a knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method and system, which is especially suitable for distributed production environments with complex dynamic characteristics such as adjustable processing time and frequent equipment failures. Background Technology

[0002] As global manufacturing moves towards distributed and flexible operations, multi-factory collaborative production has become the mainstream manufacturing model. The Distributed Flexible Job Shop Scheduling Problem (DFJSP) adds the decision dimension of factory allocation to the traditional Flexible Job Shop Scheduling Problem (FJSP), significantly increasing the complexity of the problem and making it a typical NP-hard combinatorial optimization problem.

[0003] In the context of current green manufacturing and dual-carbon strategies, production scheduling not only needs to optimize production cycles but also must take energy consumption into account. Research shows that by adjusting processing time increments, a flexible trade-off between energy consumption and time can be achieved while ensuring processing quality: higher time increments, although extending processing time, can significantly reduce energy consumption per unit time and reduce equipment stress; lower time increments can shorten the critical path but increase energy consumption. This production mode with adjustable processing rates provides new decision-making space and optimization perspectives for multi-objective optimization.

[0004] Equipment failures not only interrupt current processing tasks but also cause performance degradation after repair, extending subsequent processing times. This continuous dynamic disturbance severely impacts the effectiveness of pre-defined scheduling schemes, leading to production delays, energy waste, and decreased yield. Therefore, researching scheduling methods that can adapt to dynamic events such as equipment failures has significant theoretical and practical value.

[0005] For the dynamic scheduling problem of distributed flexible job shops considering machine failures, existing research has developed various technical approaches. Traditional heuristic methods such as SPT (Shortest Processing Time First), EDD (Earliest Due Date First), and CR (Critical Ratio, commonly used in scheduling) quickly generate solutions through predefined priority rules. However, rule design relies on expert experience, making it difficult to adapt to multiple scenarios and multi-objective optimization, and lacking flexibility in the face of dynamic environmental changes. Metaheuristic algorithms such as genetic algorithms and particle swarm optimization can perform global search, but in dynamic scenarios, they need to evolve from scratch after each event, failing to utilize historical information, resulting in excessively long response times. Furthermore, the general operators do not fully utilize the constraint structure characteristics of the scheduling problem, leading to low search efficiency. Additionally, the algorithm performance is highly dependent on parameter settings, resulting in weak generalization ability. In recent years, deep reinforcement learning has attracted attention, enabling rapid response through end-to-end learning. However, it faces the dilemma of extremely high training costs. The state and action spaces of large-scale problems grow exponentially, with training taking days or even weeks. Designing reward functions in multi-objective scenarios is difficult, and sample efficiency is low. More importantly, well-trained models exhibit significant performance degradation when faced with problem sizes, factory configurations, and failure modes outside the training set, demonstrating limited generalization ability. Furthermore, the black-box nature of neural networks makes the decision-making process difficult to interpret, hindering their trust in practical applications. While hybrid intelligent algorithms attempt to combine the advantages of multiple methods, the design of component coordination mechanisms is complex, computational overhead increases exponentially, and hybrid strategies are typically statically fixed, unable to be dynamically adjusted according to the optimization process. Prediction-based proactive scheduling methods are limited by failure uncertainty, resulting in low prediction accuracy. Robust schemes sacrifice normal efficiency, and reactive rescheduling is still required for sudden events, limiting their practicality.

[0006] Current technologies for dynamic scheduling problems have the following limitations:

[0007] 1. Single response mechanism to environmental changes: Traditional methods usually adopt a full rescheduling strategy, that is, whenever a dynamic event occurs, all historical information is discarded and a global re-search is performed. This is computationally expensive and slow, making it difficult to cope with frequently changing production environments.

[0008] 2. Lack of neighborhood structure tailored to dynamic problems: General neighborhood structures fail to fully utilize the structural characteristics of scheduling problems and do not effectively perturb decision schemes on critical paths, resulting in low local search efficiency and difficulty in quickly tracking the optimal solution set in dynamic environments.

[0009] 3. Lack of adaptability in neighborhood structure selection strategies: Existing methods mostly use random selection or fixed probability allocation of neighborhood operators, which cannot be dynamically adjusted according to the current optimization state. When facing dynamic time, they cannot track more promising neighborhood structures in time, resulting in unreasonable allocation of computing resources.

[0010] 4. Ignoring the optimization potential brought by processing rate adjustment: Most studies treat the processing time of the process as a fixed parameter and fail to take advantage of the fact that the processing time of the process in real processing scenarios is adjustable, thereby affecting machine stress and manufacturing power consumption, to achieve more refined and deeply coupled synergistic optimization of time and energy consumption.

[0011] The closest existing technology is: CN 120235404A (Comprehensive Processing Method for Dynamic Scheduling Data of Manufacturing Resources).

[0012] This paper mainly proposes a dynamic scheduling method that involves acquiring multi-source heterogeneous data, constructing a device fingerprint database, triggering abnormal event streams, building recommended production scheduling schemes based on knowledge graphs, and then combining edge-cloud joint local optimization and global optimization to form the final execution scheme.

[0013] Existing technical problems:

[0014] First, although the paper achieves multi-source data fusion and equipment status judgment, its optimization objective does not clearly consider the two typical multi-objectives of "minimizing the maximum completion time" and "minimizing the total energy consumption". At the same time, the model does not cover the cross-factory resource collaboration scenarios such as "adjustable processing time", "equipment performance degradation" and "multi-factory collaboration".

[0015] Secondly, it lacks systematic algorithm design in terms of initialization strategy, population optimization, elite solution migration and neighborhood repair mechanism; for example, it does not involve technical details such as heuristic rule fusion initialization, mixed population generation, and multi-neighborhood structure adaptive selection. Therefore, when equipment failure or environmental changes are frequent and the system is a multi-factory collaborative manufacturing environment, its real-time adjustment capability and response efficiency still have considerable room for improvement. Summary of the Invention

[0016] To address the problems existing in the prior art, this invention provides a knowledge-driven dynamic scheduling method and system for multi-factory collaborative manufacturing.

[0017] This invention is implemented as follows: A knowledge-driven dynamic scheduling method for multi-factory collaborative manufacturing includes:

[0018] S1. Establish a dynamic multi-objective mathematical model that comprehensively considers equipment failure, performance degradation, and adjustable processing rate;

[0019] S2. Before the optimization iteration begins, a hybrid initialization strategy that combines multiple heuristic rules with random generation methods is adopted to generate a high-quality initial population.

[0020] S3. During the optimization iteration process, execute the environmental change detection step to determine whether dynamic events such as equipment failure have occurred.

[0021] S4. Based on the detection results, adaptively decide and select one of the following strategies to execute:

[0022] S4-1) When no significant changes are detected in the environment, a static optimization strategy is executed, which further includes evolutionary operators, local search, and active energy-saving operators;

[0023] S4-2) When a significant change in the environment is detected, a dynamic response strategy based on environmental elite solution migration and repair is implemented.

[0024] Furthermore, the mathematical model aims to minimize the maximum completion time and the total energy consumption. The constraints include: workpiece allocation constraints, process processing constraints, machine capacity constraints, process sequence constraints, time slot allocation constraints, and equipment failure and maintenance constraints.

[0025] Furthermore, the hybrid initialization strategy includes eight heuristic initialization rules:

[0026] 1) Allocation rules based on factory load balancing;

[0027] 2) Machine selection rules based on minimum energy consumption;

[0028] 3) Machine selection rules based on the shortest processing time;

[0029] 4) Machine selection rules based on the earliest available time;

[0030] 5) Priority ranking rules based on the number of candidate machines;

[0031] 6) Selection rules based on critical machine avoidance;

[0032] 7) Generation rules based on random time-level perturbations;

[0033] 8) Completely random generation strategy;

[0034] In each of the sub-strategies in rules 1 to 7, a certain proportion of individuals are generated, and the remaining individuals are supplemented in a completely random manner to ensure the quality and diversity of the population.

[0035] Furthermore, the static optimization strategy includes the following steps:

[0036] 1) Employ specific crossover and mutation operators designed for the four-layer coding structure (factory allocation, machine selection, process sequencing, and time slots) to perform evolutionary operations;

[0037] 2) An adaptive neighborhood search strategy based on an unexpected popularity feedback mechanism is adopted to dynamically select the most effective neighborhood operator for local search;

[0038] 3) An energy-saving strategy based on active decoding and process pre-interpolation is adopted to optimize the scheduling solution through post-processing.

[0039] Furthermore, the adaptive neighborhood search strategy includes six problem-based neighborhood structures:

[0040] Random machine reassignment of neighborhood structure;

[0041] Factory exchange neighborhood structure for key processes;

[0042] : Inserting process into a neighborhood structure based on the key block structure;

[0043] Rearrangement of neighborhood structures within processes along the critical path;

[0044] : Improve the neighborhood structure by allocating time slots for non-critical processes;

[0045] : Reduce the neighborhood structure of the time frame for key processes;

[0046] The unexpected popularity model uses an adaptive response mechanism to calculate the unexpected popularity of each neighborhood structure based on the ratio of its historical success rate to its expected success rate, and dynamically adjusts the selection probability of each neighborhood structure accordingly.

[0047] Furthermore, the calculation steps of the unexpected popularity model include:

[0048] 1) Initialization: Apply all neighborhood structures to each high-quality solution, record the number of successes and failures, and calculate the initial selection probability;

[0049] 2) Historical record: Maintain a historical memory matrix of length HL to record the success and failure of each neighborhood structure in each iteration;

[0050] 3) Success rate update: Calculate the actual success rate SP for each neighborhood structure based on the historical memory matrix;

[0051] 4) Calculation of unexpected prevalence: Divide the actual success rate SP by the expected success rate EP (i.e. the selection probability of the previous generation) to obtain the unexpected prevalence index SPD;

[0052] 5) Rewards and Normalization: Add extra rewards to neighborhood structures with SPD greater than 1, and then normalize all selection probabilities.

[0053] Furthermore, the dynamic response strategy includes the following steps:

[0054] 1) Population reassessment and elite migration: When environmental changes are detected, the entire population is reassessed to identify and migrate the best-performing elite solutions.

[0055] 2) Solution repair based on neighborhood search: Apply variable neighborhood search or other perturbation methods to the migrated elite solutions to repair them and make them adapt to the new environment;

[0056] 3) Diversity Injection and Population Update: Retain the repaired elite solution and some current population individuals, and randomly generate new individuals to fill the remaining positions, forming a new generation of population.

[0057] Furthermore, the energy-saving strategy includes two levels:

[0058] 1) Processing energy consumption optimization: In the variable neighborhood search, try to increase the process time level of non-critical paths to a higher level, extend the processing time to reduce energy consumption per unit time;

[0059] 2) Comprehensive active scheduling: Using a forward and reverse bidirectional traversal method, the process is inserted to the left or right into the idle time period, so as to minimize the machine idle time while satisfying the process sequence constraints.

[0060] Another objective of this invention is to provide a knowledge-driven multi-factory collaborative manufacturing dynamic scheduling system, comprising:

[0061] The model building module is used to build a dynamic multi-objective mathematical model that comprehensively considers equipment failure, performance degradation, and adjustable processing rate.

[0062] The initialization module is used to generate a high-quality initial population by employing a hybrid initialization strategy that combines multiple heuristic rules with random generation methods.

[0063] The detection module is used to perform environmental change detection steps during the optimization iteration process to determine whether dynamic events such as equipment failure have occurred.

[0064] The adaptive optimization module is used to adaptively select and execute response strategies based on the detection results.

[0065] Another object of the present invention is to provide a computer device, the computer device including a memory and a processor; the memory stores computer program instructions, and when the computer program instructions are loaded and executed by the processor, the processor performs the steps of the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method of the present invention.

[0066] Another object of the present invention is to provide a computer-readable storage medium having computer program instructions stored thereon; when the computer program instructions are executed by a processor, the processor performs the steps of the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method of the present invention.

[0067] Another objective of this invention is to provide an information data processing terminal configured to run and implement the functions of the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling system described in this invention.

[0068] Based on the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by this invention are as follows:

[0069] (1) It realizes intelligent hierarchical response to dynamic environments.

[0070] Existing dynamic optimization algorithms typically employ a single, complete rescheduling strategy, performing a global research regardless of the severity of environmental changes, leading to wasted computational resources and sluggish response. This invention designs a dynamic response mechanism based on elite solution migration and repair. This mechanism can retain and migrate high-quality solutions available in the historical environment when the environment changes, and locally repair individuals with degraded quality through adaptive neighborhood search, enabling them to quickly adapt to the new environment. Simultaneously, by combining new environmental knowledge with an initialization strategy, a certain proportion of diverse new individuals is injected, balancing solution quality maintenance with population exploration capabilities. This mechanism avoids blind global search, significantly improving the algorithm's response speed and running efficiency, enabling the algorithm to maintain stable optimization performance in production environments with frequent dynamic events.

[0071] (2) Adaptive and intelligent neighborhood structure selection has been achieved.

[0072] Traditional variable neighborhood search methods typically employ random selection or fixed-probability neighborhood operator allocation, failing to adaptively adjust based on the current optimization state and problem characteristics. This invention innovatively introduces an adaptive operator selection mechanism based on Surprisingly Popularity. This mechanism maintains a historical memory matrix, recording the success and failure status of each neighborhood structure in real time, and calculating the ratio of its actual success rate to its expected success rate (i.e., the Surprisingly Popularity index). For neighborhood structures with a Surprisingly Popularity greater than 1, indicating that their current performance exceeds expectations, the algorithm automatically increases their selection probability and provides an additional reward. This mechanism can dynamically identify and prioritize the most effective neighborhood structures, achieving refined on-demand allocation of computational resources. Especially in dynamic environments, where historical experience may become invalid when the environment changes, this mechanism can quickly relearn and adapt to new environments, significantly improving the algorithm's robustness and optimization efficiency.

[0073] (3) It achieves deep coupling of time and energy consumption, improving optimization accuracy.

[0074] Most existing scheduling studies treat process time as a fixed parameter, failing to utilize the adjustable speed capabilities commonly found in modern manufacturing equipment. This invention innovatively introduces processing time increments as a decision variable into the scheduling model, establishing a mathematical relationship between processing time, energy consumption, and time increments. By intelligently adjusting the processing time increments during optimization, a flexible trade-off between time and energy consumption is achieved. Furthermore, a specific neighborhood structure based on processing time is designed: for processes on the critical path, higher increments can be used to extend processing time but reduce energy consumption; for processes on non-critical paths, lower increments can be used to shorten processing time but increase energy consumption. Through model building and strategy design, a more refined new perspective is provided for multi-objective optimization of time and energy consumption. In addition, this invention designs an energy-saving strategy based on active decoding and process pre-insertion, inserting processes into idle time periods through bidirectional traversal to maximize the reduction of machine idle energy consumption. This multi-layered energy-saving mechanism enables this invention to significantly reduce total energy consumption while ensuring production cycle time, providing an effective technical means for achieving green manufacturing.

[0075] (4) Achieving a deep integration between the scheduling model and physical reality.

[0076] Traditional scheduling models often treat equipment as idealized objects with constant performance, neglecting the performance degradation effect after equipment failure. This invention introduces an equipment degradation coefficient λ into the scheduling model, directly linking equipment health status with processing time: after each equipment failure and repair, its degradation coefficient increases within a set range, leading to an extension of subsequent processing time. This cumulative degradation effect more realistically reflects changes in the physical state of the equipment. By incorporating equipment health status into scheduling decisions, the scheduling scheme of this invention not only reflects the current state but also adaptively adjusts to performance changes caused by equipment failure, thus possessing high real-world adaptability and foresight, providing a solid foundation for achieving truly digital twin-driven intelligent manufacturing.

[0077] (5) The technical solution of the present invention fills the gap in related technologies at home and abroad.

[0078] This invention provides an adaptive scheduling technology solution that can collaboratively handle multiple optimization objectives (completion time and energy consumption), adapt to changes in physical equipment status (failure and degradation), and utilize the adjustable processing rate characteristics in dynamic and uncertain environments, filling a gap in related technical fields. Existing technologies in this area have significant limitations: one type of method focuses on multi-objective optimization in static environments, but its model assumptions are out of touch with real-world dynamic production environments; another type of solution considers dynamic events, but its response mechanism is usually fixed and non-adaptive, and fails to effectively account for the optimization space brought about by equipment performance degradation and time-based adjustments. This invention systematically and deeply couples the following four core technical elements: 1) Introducing a scheduling model for equipment failure and performance degradation; 2) A dynamic response mechanism based on elite solution migration and repair; 3) An adaptive neighborhood search strategy based on unexpected popularity; 4) A collaborative energy-saving mechanism based on time-based adjustments, constructing a closed-loop control system integrating environmental perception, intelligent decision-making, and scheduling execution. Through the synergistic effect of the above elements, this system solves the problem of the disconnect between traditional scheduling theory and complex physical reality, and overcomes the key technical problem of how to dynamically allocate computing resources according to the optimization process. It realizes the technological transformation from a passive and high-cost response mode to an active and refined management mode, demonstrating significant originality and progress in related technical fields.

[0079] (6) The technical solution of the present invention overcomes technical bias.

[0080] The technical solution of this invention successfully overcomes a deep-rooted technical bias in the field of dynamic scheduling, namely, that "neighborhood operator selection strategies should be based on long-term historical experience." This bias holds that neighborhood search strategies should rely on the experience accumulated from a large amount of historical data for greedy selection, or adopt a simple random selection method. This bias limits the adaptability of the algorithm in dynamic environments, because when the environment changes drastically, long-term historical experience often becomes invalid, and random selection cannot fully utilize the currently effective neighborhood structures. This invention breaks this bias, proving that the neighborhood operator selection strategy itself can also be dynamic and adaptive. By introducing an unexpected popularity model, this invention achieves online learning and dynamic adjustment capabilities for operator selection. This model does not rely on long-term historical experience, but is based on a finite-length historical memory window (HL), capturing the current performance of each neighborhood structure in real time, and identifying the most promising neighborhood structure by comparing the actual success rate with the expected success rate. Especially in dynamic environments, when the environment changes, the algorithm can quickly reset or update the historical memory and rapidly adapt to the new environment. This invention provides a short-term memory-driven adaptive operator selection framework, which overcomes traditional technical biases and significantly improves the intelligence level and adaptability of dynamic scheduling algorithms in complex and uncertain environments. Attached Figure Description

[0081] Figure 1 This invention provides a knowledge-driven dynamic scheduling method for multi-factory collaborative manufacturing.

[0082] Figure 2 This is a schematic diagram of a fault-based environmental response mechanism provided in an embodiment of the present invention.

[0083] Figure 3 This is a rendering of the dynamic scheduling method for flexible workshops that takes into account machine failures, provided in an embodiment of the present invention.

[0084] Figure 4 This is a schematic diagram of the cross-strategy for the process sequencing part in the multi-objective static optimization provided in the embodiment of the present invention.

[0085] Figure 5 This is a schematic diagram of the crossover strategy in the machine selection part of multi-objective static optimization provided in an embodiment of the present invention.

[0086] Figure 6 This is a flowchart of a flexible workshop dynamic scheduling method that takes into account machine failures, provided in an embodiment of the present invention.

[0087] Figure 7 This is a schematic diagram of the individual coding method of the flexible workshop dynamic scheduling method that takes into account machine failures provided in the embodiments of the present invention.

[0088] Figure 8This is a comparison of the Pareto front of each algorithm in the DP08 instance under environments 5-10 provided in this embodiment of the invention (with the maximum completion time vs. total energy consumption as the objective). Detailed Implementation

[0089] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0090] In multi-factory collaborative manufacturing systems, equipment operating states fluctuate randomly, processing times are constantly changing due to operating conditions, and energy prices and load curves fluctuate continuously. This makes it difficult for traditional static scheduling models to maintain the executability and energy economy of scheduling solutions in actual production. Existing methods generally rely on the assumptions of fixed processing times and static resource availability, but these assumptions do not hold true in discrete manufacturing scenarios with multiple workshops, production lines, and factories. Slight equipment performance degradation can cause batch-by-batch shifts in processing rates, and occasional downtime can disrupt already scheduled process links. In multi-factory cross-domain scheduling, these disturbances can be amplified in a cascading manner within the system, resulting in problems such as local congestion, peak energy loads, and delays in overall completion time, causing the scheduling model to become disconnected from the actual production state in the long term. This method addresses this industry pain point by constructing a dynamic scheduling framework that can be updated in real time according to changes in equipment status. This allows the scheduling strategy to make structural adjustments as soon as disturbances occur, achieving high adaptability to the complex environment of multiple factories.

[0091] This invention employs a dynamic reconfiguration mechanism driven by environmental knowledge in the optimization process. At the beginning of the algorithm, a mathematical model incorporating equipment fault constraints, performance degradation factors, and adjustable time settings characterizes the actual processing procedure, ensuring that the scheduling solution meets executability requirements under any environmental condition. To avoid getting trapped in a low-quality initial solution space under complex constraints, a hybrid initialization strategy is used to construct multiple initial solutions with varying distribution characteristics, enabling the algorithm to cover different process combinations and energy consumption levels from the outset. This initialization method provides sufficient structural diversity for subsequent dynamic searches, giving the algorithm stronger resilience to environmental changes.

[0092] During the iterative search process, the model continuously monitors equipment operating status, processing rate changes, and factory resource load, using a lightweight environmental change detection module to determine if the system is in a stable range. If the system remains undisturbed, the algorithm employs a static optimization strategy, using crossover and mutation operators specifically designed with a four-layer coding structure to drive the population towards a better solution. At this point, a neighborhood operator is selected through an unexpected popularity mechanism, extracting the actual contribution of various neighborhood structures to the scheduling solution from historical iterations, and dynamically adjusting its call frequency based on the success probability offset, making the neighborhood search more consistent with the true characteristics of the current production state. Simultaneously, active decoding and process pre-insertion strategies are used to fine-tune the scheduling solution at the energy consumption level, further reducing overall energy consumption.

[0093] When environmental changes are detected, the algorithm enters a dynamic response mode. All current solutions are recalculated for processing time and energy consumption based on the latest equipment status, and the optimal elite solutions undergo variable neighborhood repair to regain feasibility in the new environment. Subsequently, by retaining some repaired high-quality solutions and injecting random new solutions, the diversity of the population structure is maintained, enabling the scheduling system to quickly recover its global search capability after disturbances. Through the synergy of environmental knowledge and random generation, the algorithm achieves a unified approach to real-time scheduling, energy consumption optimization, and rapid disturbance recovery in dynamic multi-factory scenarios. This mechanism transforms the scheduling process from a static, single computation into a continuous computational behavior that evolves synchronously with the production process, effectively addressing the three major challenges of stability, responsiveness, and energy efficiency in actual multi-factory operations.

[0094] like Figure 1 , Figure 2 As shown, the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method provided by this embodiment of the invention includes the following steps:

[0095] S1. Establish a dynamic multi-objective mathematical model that comprehensively considers equipment failure, performance degradation, and adjustable processing rate.

[0096] This invention first establishes a dynamic, multi-objective, distributed, flexible workshop scheduling model based on the time-energy coupling problem, considering variable processing time and machine failure scenarios. Through simulations of equipment status, maintenance, and dynamic processing time adjustments in the processing flow, it accurately depicts the randomness caused by equipment status, environmental factors, etc., in actual production, especially machine failure scenarios caused by randomness and stress.

[0097] 1.1 Symbol Definition

[0098] - Maximum completion time of the scheduling scheme

[0099] - Total energy consumption

[0100] - Factory Index

[0101] - : Workpiece index

[0102] - Process Index

[0103] - Machine index,

[0104] - Time slot level

[0105] - Equipment No. Secondary fault

[0106] - : workpiece The Steps

[0107] - Process Completion time

[0108] - Process start time

[0109] - Process At the factory machine On the processing time

[0110] - Process At the factory machine Based on time slots In the Processing time after the fault

[0111] - Process At the factory machine Based on time slots Energy consumption per unit time

[0112] - Time slot Time scaling factor

[0113] - Time slot Energy scaling factor

[0114] - :factory machine In the Degradation coefficient after the second failure

[0115] - :factory machine free time

[0116] - :factory machine unit idle energy consumption

[0117] - : Binary variable, if the workpiece Assigned to the factory The value is 1 if it is 1, otherwise it is 0.

[0118] - : Binary variable, if process At the factory machine The value is 1 if the processing is performed, otherwise it is 0.

[0119] - : Binary variable, if process By time slot The value is 1 if executed, and 0 otherwise.

[0120] - : Binary variable, if factory machine At any moment The value is 0 if the system is under maintenance, and 1 otherwise.

[0121] 1.2 Time-based mechanism

[0122] The time-based scheduling mechanism allows processes to be executed at different energy efficiency levels:

[0123]

[0124]

[0125] in:

[0126] Time setting 0: Longer processing time, higher energy consumption; Time setting 1: Standard processing time, standard energy consumption; Time setting 2: Shorter processing time, lower energy consumption.

[0127] 1.3 Objective Function

[0128] 1.3-1) Objective Function 1: Total Energy Consumption (TEC)

[0129]

[0130] 1.3-2) Objective Function 2: Maximum Completion Time ( )

[0131]

[0132] 1.4 Constraints

[0133] Constraint 1: Workpiece Assignment Constraint: , Ensure that each workpiece can only be assigned to one factory.

[0134] Constraint 2: Process-Machine Compatibility Constraint: , Ensure that the process can only be carried out on compatible machines in the assigned factory.

[0135] Constraint 3: Time slot allocation constraint: , Ensure that each process can only be executed within one time slot.

[0136] Constraint 4: Process Completion Time Constraint , Define the completion time for each time bracket.

[0137] Constraint 5: Process Sequence Constraint: , Ensure the sequence constraints of processes within the same workpiece.

[0138] Constraint 6: Machine Capacity Constraint Ensure that there is at most one priority relationship on the same machine.

[0139] Constraint 7: Scheduling Separation Constraint Prevent scheduling conflicts.

[0140] Constraint 8: Equipment State Constraints Ensure that processing can only be carried out when the equipment is available.

[0141] S2. Employ a knowledge-driven hybrid initialization strategy.

[0142] To improve the quality and diversity of the initial population, this invention employs a knowledge-driven approach, specifically by designing seven heuristic initialization rules:

[0143] Initialization Rule IS1 (Factory Load Balancing-Based Allocation Rule): This rule uses a load balancing strategy to initialize the scheduling scheme. During the allocation process, the system iterates through all workpieces to be processed, calculates the current cumulative workload of each factory for each workpiece, and prioritizes allocating the workpiece to the factory with the lowest load, ensuring that the load distribution among factories is as even as possible. After the workpiece is allocated to a factory, for each process of that workpiece, the algorithm evaluates the current load of all available machines in that factory and selects the machine with the lowest load for processing, thereby achieving load balancing within the factory and avoiding uneven machine utilization.

[0144] Initialization rule IS2 (Machine Selection Based on Lowest Energy Consumption): This rule prioritizes energy consumption optimization in the production process. First, a random strategy is used to allocate workpieces to various factories, and the processing sequence of each workpiece is randomly determined to ensure diversity in the initial solutions. During the machine selection phase, the algorithm evaluates the energy consumption characteristics of all candidate machines for each process and selects the machine with the lowest energy consumption for processing. This strategy aims to reduce the total energy cost of the entire production process, achieving the goal of green manufacturing, and is particularly suitable for production scenarios with high energy costs or strict environmental protection requirements.

[0145] Initialization rule IS3 (Machine selection rule based on shortest processing time): This rule primarily aims to minimize processing time. Workpieces are randomly assigned to factories, and the sequence of processes is also randomly determined to ensure initial solution diversity. When selecting a processing machine for each process, the algorithm compares the processing times of each machine in the candidate machine set, always selecting the machine that can complete the process the fastest. While this greedy strategy may lead to overloading of some machines, it enables a rapid achievement of shorter completion times in the initial stage, providing a good starting point for subsequent optimization.

[0146] Initialization rule IS4 (Machine selection based on earliest available time): This rule focuses on the real-time availability of machines to reduce workpiece waiting time. After randomly assigning workpieces to factories and sequencing processes, the algorithm dynamically tracks the occupancy status and estimated available time of each machine, selecting the earliest machine that can start processing for each process, even if that machine's processing time may not be the shortest. This strategy helps improve machine utilization, reduce workpiece waiting time during production, and make the production process more efficient.

[0147] Initialization rule IS5 (priority sorting rule based on the number of candidate machines): This rule uses a bottleneck-first approach to construct the initial solution. Workpieces are first randomly assigned to various factories. Then, the algorithm analyzes the number of candidate machines for each process, prioritizing processes with the fewest candidate machines (these processes are more likely to become production bottlenecks). For these critical processes, the machine with the lowest current load is selected from its limited candidate set for processing, ensuring that bottleneck processes receive relatively good resource allocation and avoiding negative impacts on overall production efficiency due to improper scheduling of bottleneck processes.

[0148] Initialization rule IS6 (Critical Machine Avoidance Selection Rule): This rule employs a forward-looking strategy to avoid resource conflicts. After randomly assigning workpieces to the factory and determining the process sequence, the algorithm performs a pre-judgment check when selecting machines: if a machine in the candidate set is the only available machine for a subsequent process of that workpiece, then the current process should avoid occupying that machine, reserving resources for subsequent critical processes; if none of the candidate machines have such exclusive requirements, then a machine is randomly selected. This strategy effectively reduces scheduling conflicts caused by the premature occupation of critical resources.

[0149] Initialization rule IS7 (Generation rule based on random time-level perturbation): This rule introduces a time-level perturbation mechanism on top of conventional random initialization. First, the factory allocation of workpieces, process sequencing, and machine selection are completed randomly to generate a basic scheduling scheme. Then, with a suitable low probability, a subset of individuals (operations) are randomly selected, and their processing time levels are randomly adjusted to either a higher level (longer processing time but potentially lower cost) or a lower level (shorter processing time but potentially higher cost). This perturbation increases the diversity of the initial population in the time dimension, providing more possibilities for the algorithm to explore time-cost tradeoffs.

[0150] Initialization rule IS8 (completely random generation strategy): As a fallback strategy to ensure population diversity, this rule uses a completely random generation method for the remaining individuals. Each job is randomly assigned to any feasible factory. Under the premise of satisfying process constraints and dependencies between preceding and subsequent processes, all processes of the job are randomly ordered, and a machine is randomly selected from the set of available machines for each process. Although this completely random strategy may produce poor-quality initial solutions, it can maximize the diversity of the population, prevent the algorithm from converging to local optima too early, and provide sufficient genetic material for the evolutionary process. Each seed strategy (IS1-IS7) generates an appropriate number of individuals according to probability.

[0151] S3. Perform environmental change detection steps.

[0152] In the iterative optimization process of the Dynamic multi-objective optimization problem (DMOP), environmental monitoring is one of the most important components of the algorithm, and efficient monitoring methods are a necessary condition for algorithm efficiency.

[0153] Discrete-time index definition: in As an evolutionary iteration counter, when a dynamic event (such as equipment failure) occurs, the detection module monitors it in real time and provides real-time feedback to the scheduling console. It will record and transmit the information indicating that an environmental change has been triggered, and the algorithm will execute a dynamic response strategy.

[0154] S4. Execute static optimization strategy or dynamic response strategy:

[0155] Based on the environmental change detection results, the algorithm adaptively selects one of the following strategies:

[0156] S4-1 Static Optimization Strategy

[0157] When the environment does not change significantly, a static optimization strategy is implemented, including the following steps:

[0158] Step 1: Evolutionary Operation

[0159] Specific crossover and mutation operators are designed for the four-layer coding structure (Factory Assignment (FS), Machine Selection (MS), Operation Sequencing (OS), and Time Slots (TS): Crossover operations: Operation Sequencing (OS): A priority-based crossover operator (POX) is used to preserve operation priority constraints and ensure that the generated offspring are legal in operation sequencing. Factory Assignment (FS) and Machine Selection (MS): A uniform crossover operator (UX) is used to enhance the diversity and randomness of the search process. Time Slots (TS): A customized crossover method is designed. Identify identical operations processed on the same machine in two parent generations and swap their corresponding time slots to maintain scheduling feasibility. Mutation operations: A random two-point swap strategy is used for OS and MS; for TS, a random operation is selected, and its time slot is replaced with another feasible time slot. When the operation position changes during the mutation process, its time slot is adjusted accordingly to maintain scheduling feasibility.

[0160] Step 2: Adaptive Neighborhood Search Based on Unexpected Popularity

[0161] Design six problem-specific neighborhood structures:

[0162] N1 (Random Machine Reassignment Neighborhood Structure): Randomly selects an operation from the current scheduling scheme, removes it from its original machine, and reassigns it to another available machine with the same processing capacity. This reassignment aims to optimize the machine load distribution of the entire system, avoiding situations where some machines are overloaded while others are idle, thereby improving overall production efficiency.

[0163] N2 (Critical Process Factory Exchange Neighborhood Structure): First, identify critical operations located on the critical path (these operations directly affect the entire production cycle), and then adjust the factory allocation of the workpieces corresponding to these operations. By transferring workpieces from heavily loaded factories to lightly loaded factories, a balanced distribution of load is achieved among the factories, reducing production bottlenecks and improving the overall efficiency of multi-factory collaborative production.

[0164] N3 (Operation Insertion Neighborhood Structure Based on Critical Block Structure): This neighborhood structure specifically targets critical blocks within the critical path (a critical block is a sequence of operations processed consecutively on the same machine). For each identified critical block, the algorithm selects the first or last operation within the block as the move target and repositions it to another location within the block using an insertion move strategy. This targeted adjustment method, by changing the execution order of critical block boundary operations, effectively helps the algorithm escape local optima and explore better scheduling schemes.

[0165] N4 (Intra-Critical Path Operation Reordering Neighborhood Structure): As an extension of the N6 structure, N8 explores a larger neighborhood search space. In addition to including all the movement operations in N6, N8 also considers operations immediately before and after the critical block on the critical path (if these operations are also on the same machine). These adjacent operations can be swapped with operations at the critical block boundary or directly inserted into appropriate positions within the critical block. By providing a wider variety of operation sequence reordering possibilities, N8 can more flexibly adjust the critical path, thereby more effectively reducing the overall scheduling completion time.

[0166] N5 (Non-critical operation time tier enhancement neighborhood structure): Randomly select an operation with a relatively long time margin from the non-critical path. If the operation meets specific conditions (such as not causing new bottlenecks, having sufficient resource support, etc.), its processing time tier is adjusted upwards. This strategy frees up resources for critical operations by slowing down the execution of non-critical operations. Since the operation is not on the critical path, appropriately extending its processing time will not affect the total completion time, but can reduce processing costs or improve processing quality, achieving a balance between cost and time optimization.

[0167] N6 (Critical Operation Time Tier Reduction Neighborhood Structure): Randomly select a machine on the critical path and reduce its processing time tier to a lower level (i.e., use a faster processing speed, resulting in shorter processing time). By accelerating the execution of operations on the critical path, the overall scheduling completion time is directly shortened. Although reducing the time tier may affect processing costs or reduce processing accuracy, for time-sensitive production tasks, this trade-off can significantly improve overall scheduling performance and ensure on-time completion of production targets.

[0168] An adaptive operator selection mechanism based on Surprisingly Popular (SP) is employed to dynamically adjust the selection probability of each neighborhood structure. The specific steps are as follows:

[0169] (1) Initialization phase:

[0170] In the first generation, all neighborhood structures are applied to each high-quality solution, and the success count (SN) and failure count (FN) of each operator are recorded according to the dominance relationship. The initial selection probability of each neighborhood structure is determined through a normalization process.

[0171] (2) Maintenance of the historical memory matrix:

[0172] In subsequent iterations, a history memory matrix with HL rows is initialized, with all initial values ​​set to zero. In each generation, a roulette wheel selection mechanism is used to allocate high-quality solutions to neighborhood structures. Based on dominance relationships, the success and failure counts of each neighborhood structure are recorded and updated in the history memory matrix. When the matrix exceeds HL rows, the oldest row is deleted to maintain a fixed length.

[0173] (3) Success probability update:

[0174] The success probability of each neighborhood structure is updated according to the following formula:

[0175]

[0176] in Indicates the first Neighborhood Structure The probability of success, and They represent the first Neighborhood Structure The number of successes and failures. The expected probability of success is defined as:

[0177]

[0178] (4) Calculation of unexpected prevalence:

[0179] To adapt to environmental changes and fully tap the potential of various neighborhood structures, it is necessary to identify the most effective structures and allocate additional resources. Surprise Prevalence (SPD) is calculated using the following formula:

[0180]

[0181] in Measuring neighborhood structure In the The unexpected popularity of a generation reflects its current effectiveness relative to prior expectations.

[0182] (5) Rewards and Normalization:

[0183] like This indicates the structure This unexpectedly worked in this iteration. To reward this structure, [the following text is missing: "in its..."] Add an extra success probability increment (Typically, the value ranges from 0.1 to 0.2). To ensure that the sum of the selection probabilities of all neighborhood structures is 1, the success probability is normalized:

[0184]

[0185] in For applying reward-based neighborhood structure The success rate of the update. For the first Neighborhood Structure The normalized choice probability.

[0186] Step 3: Energy-saving strategy

[0187] Energy-saving strategies include two levels:

[0188] (1) Energy consumption optimization based on process processing time:

[0189] In variable neighborhood search, attempts are made to elevate the time increments of processes on non-critical paths to higher increments. This adjustment extends the overall processing time and reduces actual processing energy consumption, thereby minimizing idle time. Because these adjustments occur outside the non-critical paths, they do not affect the overall completion time.

[0190] (2) Fully active scheduling:

[0191] The fully proactive scheduling strategy involves two steps:

[0192] - Forward traversal: Traverse the processes from left to right. For a process on the same machine, if its processing time is less than the idle time on the left, then shift the process to the left and insert it into the idle time.

[0193] - Reverse traversal: Traverses the operations from right to left. For an operation on the same machine, if its processing time is less than the idle time on the right, then, provided the processing sequence constraint is satisfied, the operation is moved to the right and inserted into the idle time. The scheduling process terminates when no operation can be moved to the left or right for insertion. This bidirectional traversal ensures that idle time on the machine is minimized, thereby reducing energy waste while maintaining scheduling feasibility.

[0194] Step 4: Decomposition-based update strategy:

[0195] In the framework of multi-objective evolutionary algorithms, the environment selection or update strategy is a key factor determining the algorithm's performance. It directly affects the algorithm's ability to balance convergence and diversity. Traditional aggregation methods, such as weighted sums, have the drawback of failing to effectively approximate non-convex Pareto fronts; and the penalty-based boundary crossover method is highly sensitive to the setting of penalty parameters, and improper adjustment can affect the solution performance.

[0196] To overcome the limitations of the aforementioned methods, this invention employs an adaptive update strategy based on the Chebyshev aggregation function. This strategy aims to achieve efficient and robust optimization by guiding the search process directionally. The core idea of ​​the Chebyshev aggregation function is to decompose a multi-objective optimization problem into a set of single-objective scalar optimization subproblems, thereby transforming the complex Pareto front approximation problem into a series of more manageable single-objective optimization problems. Its mathematical form is defined as follows:

[0197]

[0198] In the formula, This is the current candidate solution to be evaluated; It is a weight vector that defines a specific search direction or preference, corresponding to a decomposed subproblem; It is an ideal reference point, consisting of the minimum value that each objective function in the current population can achieve; Solution The One objective function value; Representing the solution In the The distance between each target and the ideal reference point;

[0199] Calculate the maximum weighted distance across all targets; this value can be considered the solution. A comprehensive fitness on a subproblem defined by the weight vector λ.

[0200] From a geometric perspective, the Chebyshev aggregation function defines a function with an ideal point... Minimizing this function in the transformation space with the origin is equivalent to finding the point closest to the ideal point that intersects with the Pareto front. This geometric property allows the Chebyshev method to approximate Pareto fronts of arbitrary shapes, including concave, convex, and mixed shapes, while maintaining theoretical completeness and monotonicity guarantees.

[0201] This update strategy is executed in a neighborhood-based framework, and the specific steps are as follows:

[0202] Step 4-1) Dynamic update of ideal point: For newly generated solutions First, update the ideal reference point. Then, iterate through each target dimension. ,like Then let This ensures that the aggregation function is always evaluated based on the optimal reference point.

[0203] Step 4-2) Determine the update range: For the newly generated solution (Assuming it is generated from the i-th subproblem), a probabilistic update mechanism is used to generate random numbers. ,like ( (Usually taken as 0.9), then neighborhood update is selected, and the candidate update set is set as follows: ,in It is the neighborhood set of the i-th subproblem; otherwise, choose global update, let Neighborhood updates leverage the similarity of subproblems to efficiently propagate information, while global updates prevent getting trapped in local optima.

[0204] Step 4-3) Evaluation and Replacement: Evaluation of the candidate update set Each subproblem in Using its weight vector and the current ideal point Calculate the new solution respectively and original solution Chebyshev function value: , .like That is, the new interpretation If the performance is better, then perform the replacement. and will replace the counter Add 1. To maintain population diversity, set a maximum number of replacements. (Typical value is 2 or 3), when The update process will terminate at that time.

[0205] Through the above update mechanism, a newly generated excellent solution can simultaneously improve the solutions of multiple subproblems in its neighborhood, thereby accelerating the spread of high-quality information in the population and guiding the entire population to converge collaboratively toward the true Pareto front along multiple preset directions.

[0206] This strategy has the following significant advantages: (1) Theoretical completeness: It can approximate Pareto fronts of arbitrary shapes and is not limited by the concavity or convexity of the front; (2) Computational efficiency: The complexity of neighborhood update is O(n). Far lower than global comparison (3) Parameter robustness: The core parameters can maintain good performance over a wide range, reducing the difficulty of parameter tuning; (4) Co-evolution: Co-convergence is achieved through neighborhood information sharing, avoiding resource waste; (5) Diversity guarantee: Neighborhood structure and replacement quantity restrictions effectively maintain the uniformity of solution set distribution.

[0207] S4-2 Dynamic Response Strategy

[0208] When environmental changes are detected, a dynamic response strategy based on elite solution migration and repair is executed, including the following steps:

[0209] Step 1: Population Reassessment and Elite Migration

[0210] When environmental changes occur, the entire population is reassessed to identify the individuals (elite solutions) that perform best under the new conditions. These elite individuals are then migrated to preserve historically high-quality solutions and minimize the loss of solution quality caused by environmental changes.

[0211] Step 2: Solution Repair Based on Neighborhood Search

[0212] Variable Neighborhood Search (VNS) or other perturbation methods are applied to the migrating elite individuals to repair and adjust their structure. This repair process aims to adapt the solution to the new environment through minor adjustments while preserving its desirable properties.

[0213] Step 3: Diversity Injection and Population Renewal

[0214] The repaired elite solutions and a subset of individuals randomly selected from the current population are retained. New individuals are randomly generated to fill the remaining vacancies in the population. This diversity injection not only preserves high-quality solutions but also introduces randomness into the population, enhancing its exploratory capabilities and preventing premature convergence to local optima. Specifically, the new generation of the population consists of the following three parts (Q1, Q2, Q3):

[0215] Q1: Process the elite solution after re-evaluation using the dynamic transfer operator;

[0216] Q2: The diverse solutions that make up 20% of the population size include some high-quality solutions generated based on new environmental knowledge and other solutions generated by random strategies;

[0217] Q3: The solution obtained by performing a neighborhood search on the elite solution after re-evaluation.

[0218] Ultimately, the Chebyshev selection strategy was used to... Selected to form a new generation of population .

[0219] Step 4: Updating Historical Memory

[0220] When the environment changes, update the historical matrix by performing a full recalculation. , The method involves using an unexpected popularity model to evaluate and adjust the selection probability of neighborhood structures in the current environment. This allows the algorithm to learn and adapt to new environments quickly, significantly improving its robustness.

[0221] S5. Termination Condition Judgment

[0222] Repeat steps S3 to S4 until the termination condition is met (i.e., the maximum number of iterations or other preset stopping criteria are reached).

[0223] like Figure 4 As shown, an embodiment of the present invention provides a knowledge-driven multi-factory collaborative manufacturing dynamic scheduling system, comprising:

[0224] Model Building Module: This module is used to build a dynamic multi-objective mathematical model that comprehensively considers equipment failure, performance degradation, and adjustable processing rate. Based on real-time data from the production environment, this module constructs a mixed-integer linear programming model with the objectives of minimizing the maximum completion time and minimizing total energy consumption, and sets the corresponding constraints.

[0225] Initialization Module: This module employs a hybrid initialization strategy that combines multiple heuristics with random generation methods to generate a high-quality initial population. Through a combination of eight initialization sub-strategies, this module ensures that the initial population possesses both high quality and good diversity.

[0226] Detection Module: This module performs environmental change detection during optimization iterations to determine if dynamic events such as equipment malfunctions have occurred. It monitors whether machines in the machine matrix are in a faulty state, promptly identifying environmental changes and triggering corresponding response mechanisms.

[0227] Dynamic optimization module: This module adaptively selects between static optimization and dynamic response strategies based on the detection results. It intelligently switches optimization modes according to environmental conditions, ensuring the algorithm maintains high performance in various scenarios.

[0228] Neighborhood Search Module: This module dynamically selects the most effective neighborhood operator for local search based on an unexpected popularity feedback mechanism. It achieves online learning and adaptive adjustment of neighborhood operator selection by maintaining a historical memory matrix and calculating an unexpected popularity index.

[0229] Energy-saving optimization module: This module performs post-processing optimization on the scheduling solution to reduce total energy consumption. It maximizes the reduction of machine idle time and lowers actual processing energy consumption through two levels: processing energy consumption optimization and comprehensive active scheduling.

[0230] In this embodiment of the invention, the signal and data processing begins with comprehensive monitoring of the workshop production equipment. The system deploys various types of sensors on each processing device, assembly station, and material handling unit to continuously collect equipment operating status information, including operation duration, equipment utilization rate, power consumption mode, operating temperature range, and mechanical motion characteristics—all production process parameters. The collected signals are aggregated in real-time to the data processing center via an industrial communication network. Each data point is appended with a timestamp and equipment code to ensure data integrity and traceability. The system employs a data verification mechanism to identify and handle any data loss or anomalies that may occur during transmission, ensuring the reliability of information collection.

[0231] After the raw signal enters the data preprocessing stage, the system performs multi-step data optimization processing. First, noise filtering is applied to the acquired signal to remove data fluctuations caused by environmental interference and measurement errors. Second, data smoothing techniques are used to eliminate occasional abnormal peaks, improving data stability. Simultaneously, key feature parameters, such as average value, rate of change, and periodicity, are extracted from the raw signal to form a structured feature dataset. Finally, data from different sources and with different dimensions are standardized to ensure that all indicators have a uniform numerical range and comparability, providing high-quality input data for subsequent processing.

[0232] In the environmental change detection phase, the system compares and analyzes the real-time processed data with preset normal operating benchmarks. By establishing a multi-dimensional evaluation system, the degree of change in equipment status is quantitatively assessed. When equipment operating efficiency declines, failure frequency increases, or operating parameters deviate from normal ranges, the system can promptly identify these changes. The detection module accurately judges the health status and performance level of the equipment by setting reasonable judgment criteria and classification methods, and transforms the detection results into quantifiable status indicators, providing a basis for scheduling decisions.

[0233] After the data enters the adaptive scheduling module, the system dynamically adjusts the production scheduling plan based on the quantitative assessment results of environmental changes. The scheduling module comprehensively considers the current status of equipment, the urgency of tasks, and production constraints, rescheduling work plans while maintaining production continuity. When a piece of equipment malfunctions, the system automatically reassigns the affected production tasks to other available equipment or factories and adjusts the execution time and sequence of related processes. Through flexible task allocation and resource reorganization, the impact of abnormal events on the overall production schedule is minimized.

[0234] The optimized scheduling scheme is transmitted to each execution level through information feedback and visualization processing. The system issues new production instructions to the corresponding equipment control units, guiding the specific execution of on-site operations. Simultaneously, the monitoring center's display interface presents scheduling changes in an intuitive graphical manner, including key information such as the task allocation status of each piece of equipment, estimated completion time, and production progress. Operators can monitor the workshop's operational status and scheduling adjustment results in real time. New data generated during execution is re-entered into the acquisition system to participate in the next round of analysis and optimization, implementing a continuous cyclical processing mechanism.

[0235] To fully demonstrate the inventiveness and practical value of the technical solution of this invention, the following will take a multi-factory collaborative manufacturing enterprise with an annual output of 1.2 million automotive drive shaft parts as an example to illustrate the application of this invention in detail.

[0236] The technological background of this application scenario is that automotive parts manufacturers face the dual challenges of frequent equipment failures and cumulative performance degradation when implementing distributed production models. When performing high-intensity batch production tasks, the core processing equipment groups in the workshop not only experience sudden failures such as spindle bearing seizure and servo motor burnout, but also chronic degradation phenomena such as gradual deterioration of processing accuracy and continuous decline in positioning repeatability. This complex pattern of equipment health deterioration severely impacts the stable execution of production plans and the consistent control of product quality.

[0237] To address the aforementioned engineering challenges, the knowledge-driven dynamic scheduling system constructed in this invention can be seamlessly integrated into an enterprise's intelligent manufacturing platform, achieving deep integration with SCADA (Supervisory Control and Data Acquisition), equipment PHM (Prognostics and Health Management) modules, and ERP (Enterprise Resource Planning) planning layers. During system operation, it constructs a digital twin model of the equipment's health status by collecting multi-dimensional sensor data such as vibration spectrum, spindle current, and servo torque from various factory equipment in real time.

[0238] In actual production, when the spindle bearing of a CNC turning center in a factory experiences a sudden jamming failure due to fatigue wear, the dynamic event detection engine of this invention can identify the abnormal event within milliseconds and immediately activate the emergency response mechanism for dynamic events such as machine failures. The system first assesses the repair time of the faulty equipment and the degree of performance recovery after repair (based on historical data prediction), and then activates rescheduling based on knowledge transfer.

[0239] The algorithm performs the following intelligent decisions while retaining the core structure of the current effective scheduling scheme: (1) urgently transfer the work-in-process on the faulty equipment to the backup equipment in the same factory or allocate it to the cooperating factory across factories; (2) re-evaluate the processing capacity of the equipment after maintenance and adjust the subsequent task allocation density according to the expected performance degradation of the equipment; (3) optimize the processing time of the bottleneck process on the critical path, and ensure that the overall delivery node is not affected by sacrificing the processing time of some critical processes, while increasing the processing time of non-critical processes with sufficient time margin to reduce total energy consumption and improve processing stability.

[0240] More importantly, this invention can handle the ongoing performance degradation effects after equipment maintenance. For example, when a gear hobbing machine completes a cutter replacement, the system automatically updates the degradation coefficient λ of the equipment from 1.0 to the prediction range. This means that all subsequent processing tasks assigned to the equipment need to take into account additional time margins. Through this dynamic degradation modeling mechanism, the scheduling system can proactively avoid the risk of cascading delays caused by ignoring equipment performance degradation.

[0241] The implementation of this application fully validates the technical advantages of this invention in addressing the uncertainties of complex manufacturing environments. By organically combining the immediate response capability to equipment failures with a long-term tracking mechanism for performance degradation, this invention achieves a paradigm shift from passive fault handling to proactive health management, providing reliable technical support for building a resilient supply chain system in the automotive parts industry.

[0242] The performance of the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method of this application can be illustrated and analyzed through a specific embodiment.

[0243] Specifically, this application employs an improved MK and DP standard test set, comprising 20 test cases. To simulate the dynamic characteristics of real production, reasonable scenarios of continuous equipment failure and performance degradation are introduced. Processing time and operating power at different time levels are configured according to a preset ratio: the base processing power is... Idle power is The equipment maintenance is imperfect; after each repair, the equipment's processing time will increase due to degradation, with a degradation coefficient... From each failure It is randomly generated within the range, and the degree of degradation accumulates as the environment changes continuously.

[0244] To comprehensively measure algorithm performance, this experiment uses three improved, internationally recognized performance metrics (diversity, convergence) for dynamic multi-objective optimization (GD, MIGD, HV):

[0245] 1. Mean Generation Distance (MGD): The smaller the value, the better the convergence.

[0246] 2. Mean Inversion Generation Distance (MIGD): The smaller the value, the better the average performance;

[0247] 3. Mean supervolume (MHV): The larger the value, the better the convergence and diversity averaging performance;

[0248] To further demonstrate the effectiveness of KDMMA, it was compared with five state-of-the-art algorithms, including three classic evolutionary algorithms—NSGA-II, MOEA / D, and SPEA2—and two recent high-performance methods for energy-aware distributed flexible job shop scheduling: SPAMA and IMANS. To eliminate random variations and ensure experimental fairness, each algorithm was run independently 30 times on each test instance. The statistical results of the comparison algorithms with KDMMA for the MHV, MGD, and MIGD metrics are shown in Tables 1, 2, and 3 below.

[0249] Table 1

[0250]

[0251] Table 2

[0252]

[0253] Table 3

[0254]

[0255] Furthermore, this application ranks the results of the Frideman rank-sum test between KDMMA and the comparison algorithm (significance level). KDMMA ranked first in all three metrics: MHV, MGD, and MIGD. Overall, KDMMA achieved the best comprehensive ranking (Table 4), indicating that KDMMA has strong competitiveness in solving problems.

[0256] Table 4

[0257]

[0258] 1. Statistical results across the four metrics show that KDMMA significantly outperforms other algorithms on MGD and MHV. For MIGD, KDMMA performs comparably to SPAMA, IMANS, and MOEA / D on no more than three instances, and outperforms all compared algorithms on the remaining instances. On the MSpread metric, KDMMA is generally less competitive, but still achieves better or comparable results on some instances.

[0259] 2. For example Figure 8 As shown, the comparison of Pareto front solutions in environments 5 through 10 of the DP08 instance demonstrates that, intuitively, KDMMA outperforms other state-of-the-art algorithms in terms of both convergence and diversity, based on the distribution of solutions.

[0260] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0261] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A knowledge-driven dynamic scheduling method for multi-factory collaborative manufacturing, characterized in that, Includes the following steps: (1) Establish a dynamic multi-objective mathematical model that comprehensively considers equipment failure, performance degradation and adjustable processing time. The mathematical model takes minimizing the maximum completion time and minimizing the total energy consumption as optimization objectives. (2) Before optimization iteration, a hybrid initialization strategy that integrates multiple heuristic rules and random generation methods is adopted to generate a high-quality initial population; (3) Execute an environmental change detection step during the optimization iteration process to identify dynamic events such as equipment failure or processing rate fluctuations; (4) Based on the detection results, the static optimization strategy or the dynamic response strategy based on elite solution migration and repair is adaptively selected, and the diversity is injected based on environmental knowledge and random generation, thereby realizing real-time scheduling and energy consumption optimization of multiple plants.

2. The method according to claim 1, characterized in that, The constraints of the mathematical model include: Workpiece allocation constraints, process processing constraints, machine capacity constraints, process sequence constraints, time allocation constraints, and equipment failure and maintenance constraints. The time-level constraint ensures the executability of the scheduling solution by setting feasible start and end time intervals for each processing segment.

3. The method according to claim 1, characterized in that, The hybrid initialization strategy includes the following eight sub-strategies: The distribution rules are based on factory load balancing, machine selection rules are based on minimum energy consumption, machine selection rules are based on minimum processing time, machine selection rules are based on earliest available time, priority ranking rules are based on the number of candidate machines, selection rules are based on avoiding critical machines, generation rules are based on random time level perturbation, and completely random generation strategies are used.

4. The method according to claim 1, characterized in that, The static optimization strategy includes: Evolutionary operations are performed using specific crossover and mutation operators designed for a four-layer coding structure. An adaptive neighborhood search strategy based on an unexpected popularity feedback mechanism dynamically selects the most effective neighborhood operator for local search. An energy-saving optimization strategy of active decoding and process pre-interpolation is adopted to post-process the scheduling solution in order to reduce overall energy consumption.

5. The method according to claim 4, characterized in that, The adaptive neighborhood search strategy includes six neighborhood structures: Random machine reassignment neighborhood, factory exchange neighborhood for critical processes, process insertion neighborhood based on critical block structure, process rearrangement neighborhood within the critical path, time increment neighborhood for non-critical processes, and time increment neighborhood for critical processes.

6. The method according to claim 4, characterized in that, The unexpected popularity feedback mechanism includes: For each high-quality solution, apply the full neighborhood structure and count the number of successes and failures. A historical memory matrix of length HL is maintained to record the success and failure of each neighborhood structure in each iteration; The actual success rate is calculated based on the historical matrix, and the ratio of the actual success rate to the expected success rate is defined as the unexpected prevalence index. Add rewards to neighborhood structures with an unexpected popularity index greater than 1, and then normalize all selection probabilities.

7. The method according to claim 1, characterized in that, The dynamic response strategy includes: When environmental changes are detected, the entire population is reassessed. The best-performing elite solution is then repaired using a variable neighborhood search. The repaired elite solution and some current individuals are retained, and new individuals are randomly generated to maintain population diversity.

8. A knowledge-driven multi-factory collaborative manufacturing dynamic scheduling system, characterized in that, include: The model building module is used to construct a dynamic multi-objective mathematical model that takes into account equipment failure, performance degradation, and adjustable processing rate. The initialization module is used to generate the initial population using a hybrid initialization strategy that combines heuristic rules and random generation methods. The detection module is used to perform environmental change detection during the optimization iteration process; The adaptive optimization module is used to select between static optimization strategy and dynamic response strategy based on the detection results. The neighborhood search module is used to dynamically select the most effective neighborhood operator for local search based on the unexpected popularity feedback mechanism. The energy-saving optimization module is used to post-process the scheduling solution to reduce total energy consumption.

9. A dynamic scheduling device for multi-factory collaborative manufacturing to implement the method of claim 1, characterized in that, include: Processor, memory, and communication interface; The memory stores instructions that, when executed by the processor, implement the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the knowledge-driven multi-factory collaborative manufacturing dynamic scheduling method according to any one of claims 1 to 7.

Images