Enterprise order collaborative production scheduling intelligent dispatching service system based on deep intelligent optimization algorithm combining ai model and improved genetic algorithm

Abstract

The application discloses an enterprise order collaborative production scheduling intelligent dispatching service system based on a deep intelligent optimization algorithm of AI model and improved genetic algorithm, relates to the technical field of intelligent manufacturing and production planning and scheduling, and comprises the following modules: an order acquisition module, a multi-agent collaborative analysis module, a production scheduling scheme optimization module, a large language model evaluation module, a dynamic rearrangement module and an intelligent interpretation output module.The large language model is embedded in the middle link of the production scheduling optimization process, is intervened after the improved genetic algorithm completes numerical level iterative optimization and before the final scheme is output, and performs semantic level comprehensive evaluation on the production scheduling scheme.Based on deep understanding of the knowledge in the field of production scheduling, the large language model can identify deep risks such as excessive concentration of resources, fragile key path and disconnection of material supply which cannot be reflected by a pure numerical scoring system, so that the optimization dimension of the production scheduling scheme is expanded to semantic level scheme rationality verification.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and production planning and scheduling technology, specifically to an intelligent scheduling service system for collaborative production scheduling of enterprise orders based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms. Background Technology

[0002] As the global manufacturing industry accelerates its transformation towards digitalization and intelligence, production planning and scheduling, as the core hub connecting enterprise order commitments and shop floor execution, directly determine an enterprise's delivery capabilities and operational efficiency. Currently, three mainstream technical paths have emerged regarding production scheduling: the first is rule-based scheduling based on Enterprise Resource Planning (ERP) and Manufacturing Execution Systems (MES), relying on human experience to set priority rules (such as FIFO, shortest processing time priority) and sorting orders through fixed rules; the second is Advanced Planning and Scheduling (APS) based on mathematical programming, using mixed integer programming or constrained programming to model and solve scheduling problems; and the third is intelligent scheduling methods based on metaheuristic algorithms, using genetic algorithms, particle swarm optimization, or simulated annealing algorithms to perform global search optimization in the feasible solution space, gradually becoming a research hotspot in academia and industry.

[0003] However, the aforementioned existing technologies still have the following shortcomings: the application of Large Language Models (LLMs) in the field of production scheduling is still in its early stages. In all currently disclosed technical solutions, LLMs are only used as a tool for post-production scheduling interpretation. That is, after the production scheduling algorithm completes all calculations and outputs the final solution, LLMs then provide a natural language interpretation of the solution. This serial architecture of "calculate first, then interpret" means that the deep semantic understanding and domain reasoning capabilities of LLMs cannot feed back into the production scheduling optimization process itself. The quality of the production scheduling solution depends entirely on the numerical computation capabilities of the underlying algorithm. The cross-modal understanding capabilities of LLMs regarding production constraints, resource matching rules, and scheduling common sense are severely wasted. Production scheduling optimization remains in the single dimension of numerical computation and cannot achieve a qualitative leap from "numerical optimization" to "semantic verification and intelligent decision-making."

[0004] In response, this application proposes an intelligent scheduling service system for collaborative production scheduling of enterprise orders based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms, in order to solve the above problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent scheduling service system for collaborative production scheduling of enterprise orders based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms. This system addresses the problem in existing technologies where the sequential architecture of computation followed by interpretation prevents the deep semantic understanding and domain reasoning capabilities of LLM from feeding back into the production scheduling optimization process itself. The quality of the production scheduling plan depends entirely on the numerical computation capabilities of the underlying algorithm, and the cross-modal understanding capabilities of LLM regarding production constraints, resource matching rules, and scheduling common sense are severely wasted. Production scheduling optimization remains at the single dimension of numerical computation, failing to achieve a qualitative leap from numerical optimization to semantic verification and intelligent decision-making.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] An intelligent scheduling service system for collaborative production scheduling of enterprise orders, based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms, includes:

[0008] The order acquisition module is used to acquire the enterprise's pending production order data and its real-time production constraints, and obtain a set of order requirements with constraint tags.

[0009] The multi-agent collaborative parsing module is used to perform collaborative parsing and resource matching on the order demand set carrying constraint tags through multiple professional intelligent agents, and generate an initial collaborative production scheduling plan.

[0010] The production scheduling optimization module is used to iteratively optimize the initial collaborative production scheduling scheme through an improved genetic algorithm, and to accurately calculate the time position of each process in a shift-aware manner during the decoding process, so as to obtain the production scheduling scheme optimized by the algorithm.

[0011] The large language model evaluation module is used to perform a semantic-level comprehensive evaluation of the algorithm-optimized production scheduling scheme, identify potential risk nodes, and generate corresponding optimization and adjustment suggestions.

[0012] The dynamic rescheduling module is used to lock the processes that have started in the current production schedule when a sudden event is detected, and only reschedule the processes that have not started. It also generates a dynamically adapted production schedule by combining the dual-track constraints of the resource calendar.

[0013] The intelligent interpretation and output module is used to intelligently interpret and output the dynamic adaptation scheduling scheme in natural language.

[0014] Furthermore, the order acquisition module concurrently collects order data and bill of materials data from the enterprise resource planning system and the manufacturing execution system through a standardized interface;

[0015] A depth-first recursive algorithm is used to expand the multi-level bill of materials for each order. For each sub-assembly, the corresponding finished or semi-finished product inventory quantity is queried. When the inventory quantity is greater than or equal to the required quantity, the sub-assembly is marked as not requiring production. When the inventory quantity is less than the required quantity, only the difference in production order is generated to participate in production scheduling.

[0016] The real-time production constraints include at least equipment capacity constraints, worker skill constraints, material availability time constraints, and process sequence dependency constraints.

[0017] Furthermore, the multi-agent collaborative parsing module includes an order agent, a resource agent, and a process agent;

[0018] The order agent generates an order processing sequence based on the priority level and delivery requirements of each order in the order demand set carrying constraint tags;

[0019] The resource agent acquires the real-time availability status of each production resource, performs initial matching of equipment and workers based on the process resource requirements of each order in the order processing sequence, and generates a resource allocation mapping table.

[0020] The process intelligence agent determines the complete process chain of each order based on the expanded multi-level bill of materials, and allocates each process to the corresponding production resources according to the resource allocation mapping table;

[0021] The order agent, resource agent, and process agent iteratively negotiate through a collaborative game mechanism until all process allocations meet all hard constraints, generating the initial collaborative production scheduling plan.

[0022] Furthermore, the cooperative game mechanism is specifically as follows:

[0023] The resource agent calculates the set of processes that can be undertaken based on the current load and available time of each production resource, and feeds back the set of processes that can be undertaken to the order agent.

[0024] The order agent selects the solution with the highest delivery rate from multiple candidate resource allocation schemes based on the set of processes that can be undertaken and the delivery date requirements of the order itself, and sends the selected resource allocation scheme to the process agent;

[0025] The process intelligence agent verifies whether the sequential dependencies and sub-assembly timing relationships between each process under the selected resource allocation scheme are satisfied. If there are dependency conflicts, it sends an adjustment request to the order intelligence agent. The order intelligence agent adjusts the order processing sequence accordingly and re-initiates the allocation process until all hard constraints are satisfied, at which point it outputs the initial collaborative scheduling scheme.

[0026] Furthermore, the chromosome of the improved genetic algorithm is composed of a three-dimensional hybrid encoding of the order sequence, the equipment allocation preference value sequence, and the worker allocation preference value sequence. The equipment allocation preference value and the worker allocation preference value are continuous floating-point numbers from 0 to 1, which are converted into corresponding equipment indices and worker indices through a rounding mapping function.

[0027] The improved genetic algorithm dynamically adjusts the mutation rate based on the number of consecutive generations without improvement. When the number of consecutive generations without improvement exceeds a preset threshold, the mutation rate is increased; when a new breakthrough in fitness occurs, the mutation rate is decreased.

[0028] The improved genetic algorithm calculates the MD5 hash value of the chromosome encoding of each generation of individuals as a cache key, and directly returns the stored fitness score when the cache is hit.

[0029] Elite retention employs a strict greater than decision strategy, replacing elite individuals only when the fitness score of the current generation's best individual is strictly greater than the historical best score;

[0030] The optimized scheduling scheme is the scheduling scheme corresponding to the individual with a hard constraint violation rate of zero and the highest fitness score.

[0031] Furthermore, when decoding the chromosomes of the improved genetic algorithm, the production scheduling optimization module determines the earliest start time of each process in the available time slots of each production resource through binary search, and accumulates the process duration in a shift-aware manner, automatically skipping non-shift periods, holidays, and resource unavailability periods.

[0032] Before determining the start time of the process, the dynamic query sequence depends on the mold change matrix to obtain the actual mold change time from the previous process product to the current process product and adds it to the process time;

[0033] If there is no corresponding product combination record in the mold change matrix, the process will revert to the default mold change time.

[0034] In the algorithm-optimized production scheduling scheme, each process has a planned start time and a planned end time after shift calibration.

[0035] Furthermore, the resource calendar includes a first track of planned constraints and a second track of contingency constraints;

[0036] The first track of planned constraints includes regular equipment maintenance, statutory inspections, and fixed leave, which cannot be cleared by resource recovery operations;

[0037] The second track of contingency constraints includes equipment failure and temporary leave, which can be cleared by resource recovery operations;

[0038] During production scheduling, the first track's planned constraints and the second track's contingency constraints take effect together, forming the aforementioned resource unavailability period.

[0039] The dynamic rescheduling module merges and processes multiple sudden events accumulated within a preset time window, triggering only one local rescheduling. The resource allocation and time arrangement of the already started processes in the dynamically adapted production scheduling scheme generated after the rescheduling remain unchanged.

[0040] Furthermore, the large language model evaluation module is pre-set with a structured evaluation template that includes basic feature fields of the production scheduling plan and risk identification rules;

[0041] The time nodes, resource allocation relationships, and critical path information of each process in the algorithm-optimized production scheduling plan are filled into the structured evaluation template to form a plan evaluation prompt.

[0042] The large language model is invoked to perform semantic reasoning on the scheme evaluation prompts, and the identifiers of three types of potential risk nodes in the algorithm-optimized production scheduling scheme, namely, excessive resource concentration, critical path vulnerability, and material supply disconnection, as well as their corresponding natural language optimization and adjustment suggestions are output.

[0043] The optimization and adjustment suggestions include specific process identifiers, suggested adjustment directions, and explanations of expected impacts.

[0044] Furthermore, the intelligent interpretation and output module collects the planned start time, planned end time, allocated equipment and worker identifiers, delivery date achievement status of each order, and total utilization rate of all production resources for each process in the dynamic adaptation production scheduling scheme;

[0045] The collected data is used to fill the interpretation prompt template containing the premise statement, and the large language model is called to generate interpretation content in natural language form;

[0046] The interpretation content includes at least the delivery status forecast for each order, identification of key bottleneck resources, reminders of key time nodes that need to be focused on, and production risk warnings;

[0047] The intelligent interpretation output module uses a streaming transmission method to push the interpretation content to the user terminal word by word.

[0048] Furthermore, the system also includes a production scheduling strategy self-evolution module and a collaborative service layer;

[0049] The production scheduling strategy self-evolution module collects actual execution feedback data of the dynamically adapted production scheduling scheme from the manufacturing execution system, compares the actual execution data with the planned data for deviation at each process, generates delivery date deviation value and resource utilization deviation value, and automatically adjusts the sorting weight parameters of the order agent and the resource allocation preference parameters of the resource agent in the multi-agent collaborative parsing module according to the deviation value.

[0050] The collaborative service layer is used to connect with order data and capacity data from multiple factories or external enterprises, allocate production orders from different sources to the corresponding factories or production lines, and summarize and output the dynamic adaptation production scheduling schemes generated by each factory or production line.

[0051] Compared to existing technologies, the enterprise order collaborative scheduling intelligent scheduling service system provided by this invention, based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms, embeds a large language model into the intermediate stage of the scheduling optimization process. This allows the model to intervene after the genetic algorithm completes numerical-level iterative optimization but before the final solution output, performing a semantic-level comprehensive evaluation of the scheduling plan. Based on a deep understanding of production scheduling knowledge, the large language model can identify deep-seated risks that are difficult to reflect in purely numerical scoring systems, such as excessive resource concentration, critical path fragility, and material supply disruptions. It generates natural language optimization suggestions containing specific process identifiers and adjustment directions, expanding the optimization dimension of the scheduling plan from numerical "score improvement" to semantic "solution rationality verification." Simultaneously, by analyzing historical scheduling plan execution feedback data, the large language model automatically generates weight configuration adjustment suggestions, enabling the system to form a closed-loop self-evolutionary capability of "scheduling → execution → feedback → optimization" during continuous use. This mechanism, for the first time, transforms the large language model from a "passive interpretation of results" role in traditional scheduling systems to an "active participation in optimization decision-making," opening up a new technical path for the deep integration of semantic intelligence and computational intelligence. Attached Figure Description

[0052] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0053] Figure 1 This is a block diagram of an intelligent scheduling service system for collaborative production scheduling of enterprise orders based on deep intelligent optimization algorithms using AI models, provided in an embodiment of the present invention.

[0054] Figure 2 The system overall architecture diagram provided for embodiments of the present invention;

[0055] Figure 3 This is a diagram of the system main interface / dashboard provided in an embodiment of the present invention;

[0056] Figure 4 This is an overall flowchart of the improved genetic algorithm provided in an embodiment of the present invention;

[0057] Figure 5This is a schematic diagram of a multidimensional hybrid coding chromosome structure provided in an embodiment of the present invention;

[0058] Figure 6 The adaptive mutation rate adjustment flowchart provided in the embodiments of the present invention;

[0059] Figure 7 This is a diagram of the MD5 fitness caching mechanism provided in an embodiment of the present invention;

[0060] Figure 8 A preference value function graph provided for embodiments of the present invention;

[0061] Figure 9 An 11-dimensional scoring radar chart provided for an embodiment of the present invention;

[0062] Figure 10 A diagram illustrating the hard constraint veto mechanism provided in this embodiment of the invention;

[0063] Figure 11 This is a diagram of the production scheduling parameter configuration page provided in an embodiment of the present invention;

[0064] Figure 12 This is a real-time status panel diagram of production scheduling provided in an embodiment of the present invention;

[0065] Figure 13 The binary search constraint satisfaction decoding flowchart provided in the embodiments of the present invention;

[0066] Figure 14 This is a schematic diagram of shift perception time calculation provided in an embodiment of the present invention;

[0067] Figure 15 A diagram of the shift management configuration page provided in an embodiment of the present invention;

[0068] Figure 16 This is a diagram of the equipment / worker shift configuration page provided in an embodiment of the present invention;

[0069] Figure 17 A flowchart of BOM expansion and inventory deduction provided for embodiments of the present invention;

[0070] Figure 18 A diagram of the product BOM management interface provided in an embodiment of the present invention;

[0071] Figure 19 This is a diagram of the BOM configuration management interface provided in an embodiment of the present invention;

[0072] Figure 20 This is a diagram of a multi-level BOM configuration management interface provided in an embodiment of the present invention;

[0073] Figure 21 This is a schematic diagram of the sequential dependency mold-changing matrix and intelligent group scheduling provided in an embodiment of the present invention;

[0074] Figure 22 This is a diagram of the mold change matrix viewing interface provided in an embodiment of the present invention;

[0075] Figure 23 This is a diagram of the model change matrix maintenance page provided in an embodiment of the present invention. Detailed Implementation

[0076] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.

[0077] As attached Figure 1 To be continued Figure 23 As shown:

[0078] Example 1:

[0079] like Figure 1 and Figure 2 As shown, this invention relates to an intelligent scheduling service system for collaborative production scheduling of enterprise orders, based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms. The system comprises an algorithm engine layer, a fitness evaluation module, a decoder, and an LLM interpretation module. The algorithm engine layer is the core of this invention, employing an improved genetic algorithm and including an encoder, decoder, fitness evaluation module, genetic evolution module, and LLM interpretation module. The service layer provides a production scheduling routing module (new production scheduling / dynamic production scheduling / production status query), a data management routing module (CRUD operations for orders / products / materials / BOM / processes / equipment / workers / shifts / mold change matrices), a resource calendar routing module (dual-track management of planned constraints and emergencies), an event-driven routing module (emergency event monitoring and batch rescheduling triggering), an LLM interpretation routing module (large language model API calls and streaming proxy), and an SSE push module (real-time progress / log / status push). The interaction layer provides Gantt chart visualization, production scheduling parameter configuration, production scheduling result display, and an AI interpretation panel. The system supports independent operation without external systems and reserves standard interface adaptation capabilities.

[0080] like Figure 3 As shown, the dashboard serves as the system's overall overview entry point. It aggregates multi-source data such as resource status, production scheduling results, and inventory information in real time via backend APIs, providing production managers with a one-stop global view. The dashboard's data sources include: the service layer's resource status API (for real-time querying of equipment malfunctions and personnel leave records), the production scheduling result API (for obtaining the latest production scheduling plan's total score and process status distribution), the inventory API (for querying material shortages and finished / semi-finished product inventory status), and the order API (for statistical analysis of order priority distribution and the number of outsourced orders).

[0081] Specifically, overall layout

[0082] Left navigation bar: System function menu, including core modules such as dashboard, resource management, all production scheduling, new production scheduling, dynamic production scheduling, order management, production scheduling, and system settings.

[0083] Top control bar: Displays username, system message notifications, and quick action buttons.

[0084] Main content area: various data statistics cards, early warning information, and chart visualizations.

[0085] Core functional modules

[0086] Resource status alert (red zone)

[0087] Displays information such as resource unavailability and estimated recovery time, including equipment failure warnings and staff leave requests.

[0088] Operation button: "Go to Resource Calendar" allows for one-click navigation, facilitating quick and easy troubleshooting.

[0089] Key Indicator Statistics Card

[0090] Displays real-time quantities of various indicators.

[0091] Current production schedule overview

[0092] Displays the current production schedule's total score, number of processes, status distribution, etc.

[0093] 3. System status monitoring

[0094] Production scheduling status: Displays the current running status of production scheduling tasks (running / completed, etc.).

[0095] Material shortage: Displays the material shortage warning level (high / medium / low priority).

[0096] Staffing shortages for each shift: Displays staffing status for each shift.

[0097] Priority distribution: Statistics on the number of high / medium / low priority orders

[0098] The data visualization charts include a resource statistics pie chart (showing the proportion of four types of resources: equipment, workers, products, and materials) and an order priority distribution bar chart (showing the comparison of the number of orders with different priorities). The information status is conveyed intuitively through color coding (red for warning, green for normal).

[0099] The dashboard's innovation is reflected in three aspects: First, real-time aggregation of multi-source data—through concurrent queries of resource status, production scheduling results, inventory information, and other multi-source data via backend APIs, global visualization is achieved on a single interface, avoiding the problem of switching between multiple modules in traditional systems; Second, linkage between alerts and resource calendar—the dashboard's red alert area is linked to the dual-track management mechanism of the resource calendar, automatically pushing records of emergencies such as equipment failures and personnel leave to the dashboard, and clicking the "Go to Resource Calendar" button allows for one-click jump to processing, realizing a closed-loop process from alert discovery to anomaly handling; Third, real-time synchronization of production scheduling status with SSE—the dashboard's production plan overview area is linked to the SSE real-time push mechanism, synchronizing the running status of production tasks (waiting / running / completed / stopped) to the dashboard in real time, ensuring that production managers can grasp the production scheduling progress immediately.

[0100] 4. Data visualization charts

[0101] Resource statistics pie chart: showing the percentage distribution of four types of resources: equipment, workers, products, and materials.

[0102] Order Priority Distribution Chart: The bar chart shows the comparison of the number of orders with different priorities.

[0103] Example 2:

[0104] Improved genetic algorithm:

[0105] like Figure 4 As shown, the overall process of the improved genetic algorithm of this invention includes: initializing parameters → generating an initial population → decoding and fitness evaluation → determining whether to terminate → selection → crossover → adaptive mutation → strictly superior determination of elite retention → next generation, and repeating until the termination condition is met.

[0106] Multidimensional hybrid encoding: such as Figure 5 As shown, the chromosome of this invention is composed of three dimensions: order priority and process sequence genes (determining the order of order processing and process execution, constrained by the process route), equipment allocation genes (0~1 continuous preference values, mapped to equipment indices through rounding), and worker allocation genes (0~1 continuous preference values, mapped to worker indices through rounding). Compared with traditional single-dimensional encoding, multi-dimensional hybrid encoding greatly expands the search space and can find better production scheduling schemes.

[0107] Adaptive variability rate: such as Figure 6As shown, this invention dynamically adjusts the mutation rate based on the evolutionary state. When there is no fitness improvement for N consecutive generations (default 20 generations), the mutation rate is increased (e.g., from 0.1 to 0.3) to escape local optima; when a new breakthrough occurs (the current generation's best score is strictly greater than the historical highest score), the mutation rate is decreased (e.g., from 0.3 to 0.1) to accelerate convergence. The GA evolutionary state log format is: "Generation {current generation} / {total generations}↑ New breakthrough! Best: {best score} points Average: {average score} points", where "↑ New breakthrough!" is only displayed when it is strictly greater than.

[0108] MD5 fitness caching: such as Figure 7 As shown, this invention uses the MD5 hash of chromosome codes as a cache key and stores the fitness results in the cache. When the same code is encountered again, the cached value is returned directly, avoiding duplicate calculations. Since there are many duplicate individuals in the genetic algorithm, the cache hit rate can reach 30%~60%, significantly accelerating the computation.

[0109] Strict is superior to judgment: Combination Figure 4 In the elite retention stage of the overall flowchart shown, the elite retention strategy of this invention adopts a strict greater than decision mechanism to replace elite individuals. Specifically, the decision mechanism is as follows: zero violations of hard constraints are achieved through a 500-point veto penalty mechanism—each hard constraint violation deducts 500 points (far exceeding the maximum score of 21 points for a single dimension), ensuring that the total score of any individual with hard constraint violations is lower than that of individuals with zero violations, thus ensuring that the genetic algorithm prioritizes individuals with zero violations. Elite replacement uses a strict greater than decision mechanism, comparing the total fitness score of the current generation's best individual with the historical best score in each generation of evolution. Only when the current score is strictly greater than the latter (i.e., current_best > best_fitness, not current_best >= best_fitness) is the individual replaced with a new elite individual, and the continuous no-improve counter _no_improve_count is reset to zero to trigger the adaptive mutation rate recovery logic; otherwise, the original elite individual is retained, and the counter is incremented to trigger the adaptive mutation rate increase logic. After all evolution is complete, a hard constraint zero-violation check is performed on the final output optimal solution. If hard constraint violations still exist, an error is directly reported, prompting an increase in the number of generations or population size. This mechanism avoids the problem of retaining suboptimal individuals or even causing elite regression due to equal scores in traditional methods, ensuring the monotony and non-decreasing nature of the evolutionary direction.

[0110] (5) 11-dimensional weighted scoring: such as Figure 8 , 9As shown, the fitness function of this invention uses an 11-dimensional weighted scoring system: delivery date achievement (weight 20), material availability (weight 10), cost (weight 15), load balancing (weight 15), mold changeover time (weight 5), resource utilization (weight 10), quality (weight 5), energy cost (weight 5), work-in-process inventory (weight 5), sub-assembly collaboration (weight 5), and group optimization (weight 5). The sum of the weights is 100, and the weight value reflects the key dimensions that production scheduling focuses on; a higher weight indicates a greater focus on that dimension during production scheduling. Figure 8 As shown, the continuous 0-1 preference values ​​of the equipment allocation gene and the worker allocation gene are converted into resource indices through a rounding mapping function, transforming the discrete resource selection problem into a continuous optimization problem, enabling the genetic algorithm to simultaneously optimize resource allocation schemes. The production scheduling scoring template displays the ratio of the actual score to the full weight for each item on the results page.

[0111] Hard constraints can lead to veto power: such as Figure 10 As shown, this invention sets three hard constraints: process sequence constraints (a subsequent process cannot begin until the preceding process is completed), resource mutual exclusion constraints (only one process can be processed by the same resource at the same time), and sub-assembly timing constraints (a sub-assembly must be completed before the parent order). Each violation of a hard constraint deducts 500 points, achieving a veto-level penalty—500 points is 25 times the maximum score for a single dimension (maximum weight 20 points), ensuring that any violation of a hard constraint will cause the total score to be far lower than a zero-violation solution, and the genetic algorithm will inevitably prioritize the zero-violation solution. The final production scheduling plan must have zero hard constraint violations; if a violation occurs, an error will be directly reported, prompting the user to increase the generation or population size.

[0112] like Figure 11 As shown, the production scheduling parameter configuration page displays a visual configuration interface for all parameters of the genetic algorithm, including population size (default 100), number of generations (default 100), crossover rate (default 0.85), initial mutation rate (default 0.15), number of elites (default 5), tournament selection size (default 3), and slider controls for adjusting the weights of 11 dimensions of the scoring algorithm. Users can adjust the weights of each dimension in real time through the sliders to achieve flexible production scheduling strategies such as prioritizing delivery time (increasing the weight of delivery time achievement), prioritizing cost (increasing the weight of labor costs and mold change time), and prioritizing resources (increasing the weight of resource utilization and load balancing). The innovation of this page lies in making all parameters of the genetic algorithm available to users for visual adjustment, enabling production scheduling strategies to be flexibly customized according to the actual needs of enterprises, rather than the fixed parameter black box mode of traditional systems.

[0113] like Figure 12As shown, the real-time production scheduling status panel achieves real-time transparency of the production scheduling process through SSE (Server-Sent Events) technology. The panel consists of five parts: a production scheduling parameter configuration area, a convergence curve area, a running log area, a scoring details area, and an LLM scheme evaluation area. The production scheduling parameter configuration area is used to maintain production scheduling-related parameters and initialize multiple agents—solving the "losing at the start" problem. Multiple agents, such as order agents, resource agents, and process agents, collaborate in a game to provide a reliable initial answer. The convergence curve area plots the evolutionary trajectory of the optimal and average scores for each generation in real time, allowing users to intuitively observe the convergence speed of the genetic algorithm, whether it gets stuck in local optima, and when new breakthroughs occur. The running log area continuously outputs three types of logs: GA evolution status logs (e.g., "Generation {current generation} / {total generation}↑ New breakthrough! Optimal: {optimal score} points, Average: {average score} points", continuous no improvement logs, and logs in progress), and ENCODER inventory deduction logs (format: "[ENCODER] Order / Sub-assembly {number} ({name})"). Finished / semi-finished inventory is partially deducted ({inventory quantity} / {demand quantity}), requiring only the production of the {difference quantity}. A decoding log (recording process allocation and time calculation) is used, with log colors differentiated by level (success green / warning yellow / error red / information gray). The scoring details area displays the actual score and weighted full score ratio for each of the 11 dimensions. The AI ​​solution evaluation area addresses the issue of "not understanding just by looking at scores," helping scheduling personnel understand which indicators are dragging down performance; whether there are hidden risks; and how to improve. For example, it assesses over-concentration of resources, critical path vulnerability, and material supply disruptions. Besides risk warning, it also provides scheduling personnel with adjustment strategies and decision-making basis, achieving a closed-loop scheduling management system where the scheduling algorithm calculates, AI evaluates, and scheduling personnel make decisions and reschedules. This achieves "first round of scheduling completed → view AI evaluation → identify risks → modify configuration on the corresponding page → reschedule → view evaluation again →..." The closed-loop scheduling mode, which "repeates until a satisfactory result is achieved," uses AI evaluation as a "translator" between scheduling personnel and the scheduling algorithm. This helps scheduling personnel understand the deeper issues within the scheduling plan and then adjust configuration conditions and algorithm parameters based on this information to improve future scheduling. The innovation of this panel lies in its complete transparency of the algorithm's operation. Users can observe the evolutionary status of the genetic algorithm in real time after optimization, including when new breakthroughs occur, how many consecutive generations have seen no improvement, and whether the mutation rate is dynamically adjusted. This solves the problem of traditional scheduling systems operating like a black box, where users cannot judge the quality of scheduling. Simultaneously, through LLM (Limited Ledger Management), users can understand which indicators are dragging down the current scheduling results, whether there are hidden risks, how to improve, and adjust indicators and rearrange according to suggestions until a satisfactory result is achieved.

[0114] Example 3:

[0115] Decoder, constraint satisfaction and shift perception:

[0116] like Figure 13 As shown, the decoder of this invention employs a binary search constraint satisfaction algorithm. For each process, the decoder determines the earliest possible start time through binary search within the available resource period, ensuring that the process pre- and post-process constraints and resource mutual exclusion constraints are satisfied. When determining the process start time, the decoder also needs to dynamically query the mode change matrix:

[0117] If the last product processed on the equipment is different from the product of the current process, then find the switching time between the two products in the mold change matrix and add it to the time before the start of the process;

[0118] If no corresponding record is found in the model change matrix, the default model change time is used. This mechanism enables the decoder to accurately calculate the actual time cost of each product switch, providing an accurate time basis for the intelligent group scheduling optimization of the genetic algorithm.

[0119] like Figure 14 As shown, the shift-sensing time calculation of this invention supports time accumulation across days, shifts, and holidays. For each process, the decoder automatically skips: non-shift periods (such as night shifts), holidays, planned constraint periods (such as regular maintenance), and periods of unexpected events (such as equipment failure). Time format parsing supports the special case of '24:00', which is achieved by converting it to '23:59' and then adding 1 minute after parsing to ensure compatibility.

[0120] like Figure 15 , Figure 16 As shown, the shift management configuration page displays a visual configuration interface for shift parameters, including shift name (e.g., morning shift, afternoon shift, evening shift), start time, end time, rest period, applicable dates, and other parameters. The system supports flexible configuration of multiple shifts, allowing the definition of different shift systems such as morning shift (08:00~16:00), afternoon shift (16:00~24:00), and evening shift (00:00~08:00), and automatically identifies shifts spanned by the production process during scheduling to calculate time. The innovation of this page lies in parameterizing the multi-shift production mode of discrete manufacturing, enabling the scheduling time calculation to accurately match the actual production cycle, avoiding the distortion problem of traditional systems that estimate based on 24-hour continuous production.

[0121] Example 4:

[0122] BOM Deployment and Inventory Deduction

[0123] like Figure 17As shown, this invention combines finished / semi-finished product inventory for deduction during BOM expansion. For each order and sub-assembly, its finished / semi-finished product inventory quantity is queried: when the inventory quantity is greater than or equal to the required quantity, production of that order / sub-assembly is skipped; when the inventory quantity is less than the required quantity, only the difference needs to be produced. The production scheduling log records inventory deduction information of type ENCODER, in the format: "[ENCODER] Order / Sub-assembly {Number} ({Name}) Finished / Semi-finished Product Inventory Partial Deduction ({Inventory Quantity} / {Required Quantity}), Only {Difference Quantity} Needs to be Produced". Product data includes a stock field to record finished / semi-finished product inventory; if the stock field is absent, it defaults to 0.

[0124] like Figure 18 As shown, the product BOM management interface displays a visual distinction design between materials and subassemblies: material usage icons, white background, gray border, and blue serial number badges; subassemblies use... The interface features icons, a light purple background, a purple border, and purple serial number badges. Through this four-fold visual distinction—icon, background color, border color, and serial number badge color—users can intuitively identify which components in the BOM structure are purchased materials and which are sub-assemblies requiring in-house manufacturing. The interface's innovation lies in its intuitive and visual management of complex product structures, reducing the cognitive burden of multi-level BOM configurations and improving data entry accuracy.

[0125] like Figure 19 As shown, the BOM configuration management interface displays editing functions for material quantity, unit, and type (self-made / purchased), as well as mixed configuration management of sub-assemblies and materials. The system supports adding materials and sub-assemblies simultaneously in the same BOM level, and adjusting the BOM hierarchy structure through drag-and-drop operations. The material type field (self-made / purchased) directly affects the production scheduling logic: self-made materials generate production orders and participate in production scheduling, while purchased materials generate procurement requirements and set expected delivery times.

[0126] like Figure 20 As shown, the multi-level BOM configuration management interface displays a tree-structured BOM with three or more levels, supporting nested configuration of sub-assemblies at any level. The system employs a recursive algorithm to process multi-level BOM expansion, with each sub-assembly generating an independent sub-assembly order and establishing a timing constraint relationship with the parent order. The innovation of this interface lies in its recursive processing capability for complex multi-level BOMs, enabling unified management and scheduling of everything from simple single-level BOMs to complex product structures containing dozens of levels.

[0127] Example 5:

[0128] 11-Dimensional Scoring and Hard Constraint Details:

[0129] like Figure 9 and Figure 11As shown, the 11-dimensional scoring system and hard constraint veto mechanism of this invention are the core guarantees of production scheduling quality. The 11-dimensional scoring covers delivery date achievement, material availability, labor cost, load balancing, mold changeover time, resource utilization, quality, energy consumption cost, work-in-process inventory, sub-assembly collaboration and group optimization, and each dimension has a clear calculation method and weight.

[0130] A hard constraint veto mechanism ensures that the final solution has zero violations. The hard constraint penalty value is set at 500 points (veto level). Compared to the maximum score for a single dimension (maximum 20 points), a hard constraint violation will significantly lower the total score, ensuring that the genetic algorithm prioritizes solutions with zero violations. The final solution must have zero hard constraint violations; if a violation exists, an error will be reported directly, prompting an increase in the number of generations or the population size.

[0131] Example 6:

[0132] Sequential dependency on mold change matrix and intelligent group scheduling:

[0133] like Figure 21 As shown, the mold change matrix module of this invention adopts a sequence-dependent modeling of mold change time, solving the core pain point of fixed estimation of mold change time in traditional production scheduling systems. Traditional methods use a fixed mold change time (e.g., 0.5 hours) for all product changes, ignoring the actual mold change time differences caused by product similarity differences—changing similar products only requires minor adjustments (0.1 hours), while changing different types of products requires tool changes, fixture changes, cleaning, etc. (0.5~1.0 hours), with errors reaching 300%~500%, resulting in a serious disconnect between production scheduling plans and actual production.

[0134] The mold-changing matrix is ​​maintained separately for each equipment type or individual equipment. Matrix element M[i][j] represents the mold-changing time (in hours) required to switch from product i to product j. The matrix supports asymmetry (the mold-changing time from A to B is not equal to the time from B to A), perfectly matching actual production scenarios. When calculating the process start time, the decoder dynamically queries the mold-changing matrix using the _get_dynamic_setup_time method. Based on the ID of the last processed product on the equipment and the current product ID, it accurately obtains the actual mold-changing time, rather than using a fixed value.

[0135] This invention further integrates the mold-change matrix with a genetic algorithm to achieve intelligent group production scheduling optimization. In the fitness function, mold-change time is used as an independent scoring dimension in the calculation. The algorithm adjusts the relative positions of similar products in the order priority genes to ensure that similar products are produced as continuously as possible, thereby reducing the number of mold changes and the total mold-change time. Compared to traditional methods that sort by order arrival order or delivery date, intelligent group optimization can reduce the total mold-change time by 60%–75% and increase equipment utilization by 15%–25%, with particularly significant effects in discrete manufacturing scenarios involving multiple product types and small batches.

[0136] The mold-changing matrix supports flexible configuration: it can be configured in batches according to equipment type (e.g., all CNC machines share one matrix), or it can be configured individually for a single machine (e.g., a dedicated machine has unique mold-changing characteristics). When a corresponding product combination is not found in the matrix, the system automatically reverts to the default mold-changing time (read from `scheduling_defaults.default_setup_time` in `config.json`), ensuring a balance between flexible configuration and stable operation. Mold-changing matrix data can be synchronized from ERP / MES systems via an adapter, supporting independent maintenance and batch import.

[0137] like Figure 22 As shown, the mold change matrix viewing interface displays the mold change time matrix between products in tabular form. Rows represent the source product (product before switching), columns represent the target product (product after switching), and the cross cell M[i][j] represents the mold change time (in hours) required to switch from product i to product j. The matrix supports asymmetric display—the mold change time from A to B can differ from the mold change time from B to A, perfectly reflecting actual production scenarios (such as the difference in labor time between disassembly and installation). The interface uses color to intuitively indicate the length of mold change times (light color indicates short mold change, dark color indicates long mold change), helping users quickly identify product combinations with high mold change costs. The innovation of this interface lies in the visualization of sequentially dependent mold change times—compared to the fixed mold change times of traditional systems, the matrix view makes mold change time data readily apparent, providing a data foundation for intelligent group scheduling optimization.

[0138] like Figure 23As shown, the mold change matrix maintenance page displays the editing interface for mold change time data. It supports two modes: batch configuration by equipment type (e.g., all CNC machines sharing one matrix) and individual configuration for a single machine. Users can directly edit the mold change time values ​​in the cells of the table, or synchronize mold change data from the ERP / MES system via the batch import function. When the corresponding product combination is not found in the matrix, the system automatically reverts to the default mold change time (read from `scheduling_defaults.default_setup_time` in `config.json`), ensuring a balance between flexible configuration and stable operation. The innovation of this page lies in its flexible configuration capabilities for mold change data—supporting a combination of batch and single-machine configuration, and a combination of manual maintenance and system synchronization, adapting to the data management needs of manufacturing enterprises at different levels of maturity.

[0139] (1) Equipment manufacturing industry: including CNC machine tools, engineering machinery, agricultural machinery and other complete machine assembly manufacturing. It is characterized by multiple processes, multi-level BOM, strict delivery time and many outsourced processes. The multi-level BOM expansion and inventory deduction, outsourced process handling and delivery time priority scheduling strategy of this invention can effectively solve the scheduling problem of this industry.

[0140] (2) Consumer electronics manufacturing industry: including the manufacturing of 3C products such as mobile phones, computers, and home appliances. It is characterized by a wide variety of products, small batches, frequent mold changes, and cost sensitivity. The sequential-dependent mold change matrix and intelligent group scheduling optimization of the present invention can reduce the total mold change time by 60% to 75%, significantly reducing production costs.

[0141] (3) Heavy machinery manufacturing industry: including large equipment manufacturing such as mining equipment, metallurgical equipment, and petroleum equipment. It is characterized by expensive equipment resources, tight capacity, and long process cycle. The resource balance scheduling strategy and 11-dimensional weighted scoring of this invention can improve equipment utilization by 15% to 30%.

[0142] (4) Automotive parts manufacturing: including engine parts, chassis parts, body parts and other parts manufacturing. It is characterized by large batches, high quality requirements and strict supply chain coordination. The ERP / MES / WMS integration interface and quality inspection rework process of this invention can effectively support the supply chain coordination and production scheduling of this industry.

[0143] (5) Medical device manufacturing industry: including the manufacturing of medical equipment, surgical instruments, in vitro diagnostic reagents, etc. It is characterized by strict regulatory supervision, high traceability requirements and frequent process changes. The process template cascading synchronization and event-driven dynamic reordering of the present invention can effectively cope with process change scenarios.

[0144] (6) General discrete manufacturing industry: covering various multi-variety small-batch production scenarios such as mold manufacturing, hardware processing, injection molding, sheet metal processing, etc. The comprehensive balanced production scheduling strategy and AI production scheduling assistant of this invention can reduce the threshold of system use and are suitable for various small and medium-sized manufacturing enterprises.

[0145] As shown above, this invention incorporates equipment allocation preferences and worker allocation preferences into the chromosome structure in the form of continuous floating-point numbers through multi-dimensional hybrid encoding. It utilizes rounding mapping to convert from continuous space to discrete resources, expanding the search scope of the genetic algorithm from the single process ordering dimension of traditional schemes to the joint solution space of the three factors of production scheduling. This fundamentally overcomes the structural defect of fixed heuristic rules that prevent resource allocation from participating in evolution. Simultaneously, this invention uses a strict greater-than-or-equal-than strategy instead of the traditional greater-than-equal-than decision in the elite retention stage, completely eliminating elite degeneration and fitness regression caused by replacing the best individual with a suboptimal individual due to equal fitness scores. Combined with a dynamic adjustment mechanism that adaptively increases the mutation rate when there is no continuous improvement and adaptively decreases the mutation rate when a new breakthrough occurs, a self-balance is established between escaping local optima and accelerating convergence, achieving the dual technical effect of maintaining a monotonic evolutionary process and balancing search efficiency. These mechanisms work synergistically, enabling the genetic algorithm to stably approach the global optimum in the expanded solution space.

[0146] This invention embeds a large language model into the intermediate stage of the production scheduling optimization process, intervening after the genetic algorithm completes numerical-level iterative optimization and before the final solution output, to perform a semantic-level comprehensive evaluation of the production scheduling plan. Based on a deep understanding of production scheduling knowledge, the large language model can identify deep-seated risks that are difficult to reflect by purely numerical scoring systems, such as excessive resource concentration, critical path fragility, and material supply disruptions. It generates natural language optimization suggestions containing specific process identifiers and adjustment directions, expanding the optimization dimension of the production scheduling plan from numerical-level "score improvement" to semantic-level "solution rationality verification." Simultaneously, by analyzing feedback data from the actual execution of historical production scheduling plans, the large language model automatically generates weight configuration adjustment suggestions, enabling the system to form a closed-loop self-evolutionary capability of "scheduling → execution → feedback → optimization" during continuous use. This mechanism, for the first time, transforms the large language model from a "passive interpretation of results" role in traditional production scheduling systems to an "active participation in optimization decision-making," opening up a new technical path for the deep integration of semantic intelligence and computational intelligence.

[0147] The foregoing has only described certain exemplary embodiments of the present invention by way of illustration. Undoubtedly, those skilled in the art can modify the described embodiments in various ways without departing from the spirit and scope of the present invention. Therefore, the foregoing drawings and descriptions are illustrative in nature and should not be construed as limiting the scope of protection of the claims of the present invention.

Claims

1. An intelligent scheduling service system for collaborative production scheduling of enterprise orders based on a deep intelligent optimization algorithm that integrates AI models and improved genetic algorithms, characterized in that: include: The order acquisition module is used to acquire the enterprise's pending production order data and its real-time production constraints, and obtain a set of order requirements with constraint tags. The multi-agent collaborative parsing module is used to perform collaborative parsing and resource matching on the order demand set carrying constraint tags through multiple professional intelligent agents, and generate an initial collaborative production scheduling plan. The production scheduling optimization module is used to iteratively optimize the initial collaborative production scheduling scheme through an improved genetic algorithm, and to accurately calculate the time position of each process in a shift-aware manner during the decoding process, so as to obtain the production scheduling scheme optimized by the algorithm. The large language model evaluation module is used to perform a semantic-level comprehensive evaluation of the algorithm-optimized scheduling scheme, identify potential risk nodes, and generate corresponding optimization and adjustment suggestions. The dynamic rescheduling module is used to lock the processes that have started in the current production schedule when a sudden event is detected, and only reschedule the processes that have not started. It also generates a dynamically adapted production schedule by combining the dual-track constraints of the resource calendar. The intelligent interpretation and output module is used to intelligently interpret and output the dynamic adaptation scheduling scheme in natural language.

2. The enterprise order collaborative production scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The order acquisition module collects order data and bill of materials data concurrently from the enterprise resource planning system and the manufacturing execution system through a standardized interface; A depth-first recursive algorithm is used to expand the multi-level bill of materials for each order. For each sub-assembly, the corresponding finished or semi-finished product inventory quantity is queried. When the inventory quantity is greater than or equal to the required quantity, the sub-assembly is marked as not requiring production. When the inventory quantity is less than the required quantity, only the difference in production order is generated to participate in production scheduling. The real-time production constraints include at least equipment capacity constraints, worker skill constraints, material availability time constraints, and process sequence dependency constraints.

3. The enterprise order collaborative production scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The multi-agent collaborative parsing module includes an order agent, a resource agent, and a process agent; The order agent generates an order processing sequence based on the priority level and delivery requirements of each order in the order demand set carrying constraint tags; The resource agent acquires the real-time availability status of each production resource, performs initial matching of equipment and workers based on the process resource requirements of each order in the order processing sequence, and generates a resource allocation mapping table. The process intelligence agent determines the complete process chain of each order based on the expanded multi-level bill of materials, and allocates each process to the corresponding production resources according to the resource allocation mapping table; The order agent, resource agent, and process agent iteratively negotiate through a collaborative game mechanism until all process allocations meet all hard constraints, generating the initial collaborative production scheduling plan.

4. The enterprise order collaborative scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 3, characterized in that, The cooperative game mechanism is specifically as follows: The resource agent calculates the set of processes that can be undertaken based on the current load and available time of each production resource, and feeds back the set of processes that can be undertaken to the order agent. The order agent selects the solution with the highest delivery rate from multiple candidate resource allocation schemes based on the set of processes that can be undertaken and the delivery date requirements of the order itself, and sends the selected resource allocation scheme to the process agent; The process intelligence agent verifies whether the sequential dependencies and sub-assembly timing relationships between each process under the selected resource allocation scheme are satisfied. If there are dependency conflicts, it sends an adjustment request to the order intelligence agent. The order intelligence agent adjusts the order processing sequence accordingly and re-initiates the allocation process until all hard constraints are satisfied, at which point it outputs the initial collaborative scheduling scheme.

5. The enterprise order collaborative scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The chromosome of the improved genetic algorithm is composed of a hybrid encoding of three dimensions: order arrangement sequence, equipment allocation preference value sequence, and worker allocation preference value sequence. The equipment allocation preference value and worker allocation preference value are continuous floating-point numbers from 0 to 1, which are converted into corresponding equipment index and worker index through a rounding mapping function. The improved genetic algorithm dynamically adjusts the mutation rate based on the number of consecutive generations without improvement. When the number of consecutive generations without improvement exceeds a preset threshold, the mutation rate is increased; when a new breakthrough in fitness occurs, the mutation rate is decreased. The improved genetic algorithm calculates the MD5 hash value of the chromosome encoding of each generation of individuals as a cache key, and directly returns the stored fitness score when the cache is hit. Elite retention employs a strict greater than decision strategy, replacing elite individuals only when the fitness score of the current generation's best individual is strictly greater than the historical best score; The optimized scheduling scheme is the scheduling scheme corresponding to the individual with a hard constraint violation rate of zero and the highest fitness score.

6. The enterprise order collaborative scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 5, characterized in that, When decoding the chromosomes of the improved genetic algorithm, the production scheduling optimization module uses binary search to determine the earliest start time of each process in the available time slots of each production resource, and accumulates the process duration in a shift-aware manner, automatically skipping non-shift periods, holidays, and periods when resources are unavailable. Before determining the start time of the process, the dynamic query sequence depends on the mold change matrix to obtain the actual mold change time from the previous process product to the current process product and adds it to the process time; If there is no corresponding product combination record in the mold change matrix, the process will revert to the default mold change time. In the algorithm-optimized production scheduling scheme, each process has a planned start time and a planned end time after shift calibration.

7. The enterprise order collaborative production scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The resource calendar includes a first track of planned constraints and a second track of contingency constraints. The first track of planned constraints includes regular equipment maintenance, statutory inspections, and fixed leave, which cannot be cleared by resource recovery operations; The second track of contingency constraints includes equipment failure and temporary leave, which can be cleared by resource recovery operations; During production scheduling, the first track's planned constraints and the second track's contingency constraints take effect together, forming the aforementioned resource unavailability period. The dynamic rescheduling module merges and processes multiple sudden events accumulated within a preset time window, triggering only one local rescheduling. The resource allocation and time arrangement of the already started processes in the dynamically adapted production scheduling scheme generated after the rescheduling remain unchanged.

8. The enterprise order collaborative scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The large language model evaluation module is pre-set with a structured evaluation template that includes basic feature fields of the production scheduling plan and risk identification rules; The time nodes, resource allocation relationships, and critical path information of each process in the algorithm-optimized production scheduling plan are filled into the structured evaluation template to form a plan evaluation prompt. The large language model is invoked to perform semantic reasoning on the scheme evaluation prompts, and the identifiers of three types of potential risk nodes in the algorithm-optimized production scheduling scheme, namely, excessive resource concentration, critical path vulnerability, and material supply disconnection, as well as their corresponding natural language optimization and adjustment suggestions are output. The optimization and adjustment suggestions include specific process identifiers, suggested adjustment directions, and explanations of expected impacts.

9. The enterprise order collaborative production scheduling intelligent dispatch service system based on a deep intelligent optimization algorithm integrating AI model and improved genetic algorithm as described in claim 1, characterized in that, The intelligent interpretation and output module collects the planned start time, planned end time, allocated equipment and worker identifiers, delivery date achievement status of each order, and total utilization rate of all production resources for each process in the dynamic adaptation and scheduling scheme. The collected data is used to fill the interpretation prompt template containing the premise statement, and the large language model is called to generate interpretation content in natural language form; The interpretation content includes at least the delivery status forecast for each order, identification of key bottleneck resources, reminders of key time nodes that need to be focused on, and production risk warnings; The intelligent interpretation output module uses a streaming transmission method to push the interpretation content to the user terminal word by word.

10. The enterprise order collaborative scheduling intelligent dispatching service system based on a deep intelligent optimization algorithm integrating AI models and improved genetic algorithms as described in any one of claims 1-9, characterized in that, The system also includes a production scheduling strategy self-evolution module and a collaborative service layer; The production scheduling strategy self-evolution module collects actual execution feedback data of the dynamically adapted production scheduling scheme from the manufacturing execution system, compares the actual execution data with the planned data for deviation at each process, generates delivery date deviation value and resource utilization deviation value, and automatically adjusts the sorting weight parameters of the order agent and the resource allocation preference parameters of the resource agent in the multi-agent collaborative parsing module according to the deviation value. The collaborative service layer is used to connect with order data and capacity data from multiple factories or external enterprises, allocate production orders from different sources to the corresponding factories or production lines, and summarize and output the dynamic adaptation production scheduling schemes generated by each factory or production line.

Images