A production plan self-learning iterative optimization method and system

Abstract

The present application relates to the technical field of production and manufacturing, in particular to a production plan self-learning iterative optimization method and system; the method comprises the following steps: collecting production factor data and constraint condition data, and constructing a standardized production mathematical model; extracting a key feature vector from the standardized production mathematical model; inputting the key feature vector into a pre-trained AI inference model, completing an inference process through GPU parallel calculation, and outputting a preliminary scheduling scheme; performing multi-objective secondary optimization on the preliminary scheduling scheme according to a preset optimization target, and generating a final production plan; collecting actual production execution data, comparing the final production plan with the actual execution data, and calculating scheduling deviation; generating a labeled training sample based on the scheduling deviation, incrementally training the pre-trained AI inference model, and completing model self-learning iteration; and replacing a traditional commercial solver with AI inference, combining GPU parallel acceleration and closed-loop self-learning, and solving the pain points of high cost, low efficiency and poor adaptability of traditional scheduling.

Description

Technical Field

[0001] This invention relates to the field of manufacturing technology, and in particular to a self-learning iterative optimization method and system for production planning. Background Technology

[0002] Production planning optimization in manufacturing is a core element in improving enterprise production efficiency and reducing operating costs. Traditional advanced planning and scheduling systems primarily employ a technical approach combining mathematical modeling with commercial solvers. This involves establishing operations research models such as integer programming and linear programming, and then using CPU solvers to solve these models to obtain production scheduling solutions. Existing technologies have, after years of development, formed relatively mature theoretical systems and commercial products, capable of meeting the basic scheduling needs of some large manufacturing enterprises and demonstrating certain application effectiveness in high-volume, single-product production models.

[0003] However, in the aforementioned existing technologies, traditional commercial solvers have high licensing fees, long deployment and implementation cycles, and require professional operations researchers for maintenance and configuration, making them extremely difficult to use and unaffordable for most small and medium-sized enterprises. At the same time, traditional solvers are based on CPUs for serial calculations, which results in low computational efficiency when facing large-scale and complex production scenarios, failing to meet the real-time requirements of dynamic scheduling. Furthermore, the solvers use static models and cannot automatically optimize and adjust according to actual production conditions, causing the scheduling accuracy to gradually decrease as the production environment changes, making it difficult to adapt to the flexible production needs of the manufacturing industry with multiple varieties and small batches. Summary of the Invention

[0004] The purpose of this invention is to provide a self-learning iterative optimization method and system for production planning, aiming to solve the technical problems of high licensing fees for commercial solvers, long deployment and implementation cycles, the need for professional operations researchers for maintenance and configuration, extremely high barriers to entry, and the fact that most small and medium-sized enterprises cannot afford the related costs.

[0005] To achieve the above objectives, the present invention employs a production planning self-learning iterative optimization method, comprising the following steps: Collect production factor data and constraint data, and construct a standardized production mathematical model; Extracting key feature vectors from standardized production mathematical models; The key feature vectors are input into the pre-trained AI inference model, and the inference process is completed through GPU parallel computing to output a preliminary scheduling plan. Based on the preset optimization objectives, the initial scheduling plan is optimized in a multi-objective secondary manner to generate the final production plan; Collect actual production execution data and compare the final production plan with the actual execution data to calculate the scheduling deviation; Based on scheduling deviations, labeled training samples are generated to incrementally train the pre-trained AI inference model, completing the model's self-learning iteration.

[0006] Among the steps involved in collecting production factor data and constraint data to construct a standardized production mathematical model: Analyze, import, and integrate information on four basic production elements: equipment, personnel, materials, and processes. The integrated information undergoes unified encoding conversion and data format standardization processing. A four-category constraint logic system for generating process resource order exceptions; Based on a standardized information and constraint logic system, a standardized production mathematical model that conforms to the norms of operations research is generated. The standardized production mathematical model is verified for completeness and rationality, and the verification results are output along with error messages.

[0007] Among the steps in generating the four types of constraint logic systems for process resource order exceptions: Configure the logic for determining the sequence of processes, processing time, and preparation time; Verify the occupancy boundaries of equipment capacity, personnel skills, and material inventory; Match order delivery date priority batch scheduling rules; A response mechanism for equipment malfunctions, personnel leave requests, and material delays.

[0008] In the steps of inputting key feature vectors into a pre-trained AI inference model, completing the inference process through GPU parallel computing, and outputting a preliminary scheduling plan: The AI ​​inference computation task is broken down into multiple independent sub-tasks; The subtasks are distributed across multiple GPU cores for parallel execution. Merge the inference results from all GPUs to generate a complete preliminary scheduling plan; Monitor GPU resource usage in real time and dynamically adjust subtask allocation strategies.

[0009] In the step of performing multi-objective secondary optimization on the preliminary scheduling plan based on preset optimization objectives to generate the final production plan: Obtain the user-specified optimization target weights and calculate the comprehensive score of the preliminary scheduling plan; Select the candidate scheduling schemes with the highest overall scores; Resource conflict detection and correction are performed on the candidate scheduling schemes; The corrected optimal scheduling scheme is output as the final production plan.

[0010] In the step of collecting actual production execution data and comparing the final production plan with the actual execution data to calculate the scheduling deviation: Simultaneously acquire real-time data on all dimensions of process execution, equipment operation, material consumption, personnel attendance, and quality inspection. Compare execution time for each process, calculate resource consumption, and statistically analyze product quality differences. The overall scheduling deviation is obtained by summarizing the deviation data of all processes.

[0011] In the steps of generating labeled training samples based on scheduling bias, incrementally training the pre-trained AI inference model, and completing the model's self-learning iteration: Extract time dimension differences to calculate resource usage differences and statistical cost input differences; filter samples with deviation values ​​exceeding preset thresholds and add corresponding actual execution results as annotations; Count the number of labeled samples and start incremental training of the model when the trigger condition is met; The labeled training samples are used to incrementally train the pre-trained AI inference model; Evaluate the scheduling accuracy of the updated model. If the accuracy improves, enable the new model; otherwise, roll back to the previous version.

[0012] After generating labeled training samples based on scheduling bias, incrementally training the pre-trained AI inference model, and completing the model's self-learning iteration: The final production plan is displayed through a visual interface, and scheduling adjustment instructions input by the user are received. Execute adjustment instructions to generate an adjusted scheduling plan and record all adjustment operations; add the adjusted scheduling plan and corresponding constraints as labeled samples to the training sample library.

[0013] In the steps of inputting key feature vectors into a pre-trained AI inference model, completing the inference process through GPU parallel computing, and outputting a preliminary scheduling plan: The pre-trained AI inference model employs a deep neural network architecture; It is pre-trained based on massive historical production scheduling data; and adapted to the production scenarios of different types of manufacturing enterprises through transfer learning.

[0014] This invention also provides a production planning self-learning iterative optimization system, including a production planning modeling module, a feature extraction module, an AI inference scheduling module, a multi-objective optimization module, a deviation calculation module, and a self-learning iterative module, wherein: The production planning modeling module collects production factor data and constraint data, and constructs a standardized production mathematical model. The feature extraction module extracts key feature vectors from the standardized production mathematical model; The AI ​​inference scheduling module inputs key feature vectors into a pre-trained AI inference model, completes the inference process through GPU parallel computing, and outputs a preliminary scheduling plan. The multi-objective optimization module performs secondary multi-objective optimization on the preliminary scheduling plan based on preset optimization objectives to generate the final production plan. The deviation calculation module collects actual production execution data and compares the final production plan with the actual execution data to calculate the scheduling deviation. The self-learning iteration module generates labeled training samples based on scheduling deviations, performs incremental training on the pre-trained AI inference model, and completes the model's self-learning iteration.

[0015] This invention discloses a production planning self-learning iterative optimization method and system, which involves: collecting production factor data and constraint data to construct a standardized production mathematical model; extracting key feature vectors from the standardized production mathematical model; inputting the key feature vectors into a pre-trained AI inference model, completing the inference process through GPU parallel computing, and outputting a preliminary scheduling scheme; performing multi-objective secondary optimization on the preliminary scheduling scheme according to preset optimization objectives to generate a final production plan; collecting actual production execution data, comparing the final production plan with the actual execution data to calculate the scheduling deviation; generating labeled training samples based on the scheduling deviation, and incrementally training the pre-trained AI inference model to complete the model's self-learning iteration. By constructing a standardized production mathematical model, the input format of production factors and constraints is unified. The use of a pre-trained AI inference model to replace the traditional commercial solver significantly reduces the system deployment and usage costs. Relying on GPU parallel computing, the speed and efficiency of scheduling inference are significantly improved. Multi-objective secondary optimization is combined to achieve a balance of multi-dimensional scheduling requirements. At the same time, a closed-loop self-learning mechanism of execution data feedback and incremental model training is established, which can continuously optimize the scheduling accuracy. This effectively solves the core technical problems of high cost, low efficiency, static model, and poor adaptability of traditional scheduling systems. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the steps of the production planning self-learning iterative optimization method of the present invention.

[0018] Figure 2 This is a flowchart of steps S100 of the present invention.

[0019] Figure 3 This is a flowchart of step S103 of the present invention.

[0020] Figure 4 This is a flowchart of steps S300 of the present invention.

[0021] Figure 5 This is a flowchart of steps S400 of the present invention.

[0022] Figure 6 This is a flowchart of steps S500 of the present invention.

[0023] Figure 7 This is a flowchart of steps S600 of the present invention.

[0024] Figure 8 This is a schematic diagram of the production planning self-learning iterative optimization system of the present invention.

[0025] 701 - Production Planning Modeling Module, 702 - Feature Extraction Module, 703 - AI Inference Scheduling Module, 704 - Multi-Objective Optimization Module, 705 - Deviation Calculation Module, 706 - Self-Learning Iteration Module. Detailed Implementation

[0026] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0027] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more of the associated listed items.

[0028] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, such information should not be limited to these terms. These terms are only used to distinguish information of the same type from one another. For example, without departing from the scope of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the word "if" as used herein may be interpreted as "when," "when," or "in response to determination."

[0029] Please see Figures 1-7 This invention provides a self-learning iterative optimization method for production planning, comprising the following steps: S100: Collect production factor data and constraint data, and construct a standardized production mathematical model; In this implementation, production factor data and constraint data are collected to construct a standardized production mathematical model. Specifically, this includes: S101: Analyze, import, and integrate information on the four basic production elements: equipment, personnel, materials, and processes; S102: Perform unified encoding conversion and data format standardization on the integrated information; S103: Generate a logic system for four types of constraints on abnormal process resource orders; S1031: Logic for determining the sequence of processes, processing time, and preparation time; S1032: Verify the occupied boundaries of equipment capacity, personnel skills, material inventory; S1033: Scheduling rules for matching order delivery priority batches; S1034: Response mechanism for triggering equipment failure, personnel leave, and material delays.

[0030] S104: Based on a standardized information and constraint logic system, generate a standardized production mathematical model that conforms to the norms of operations research; S105: Perform a completeness and rationality check on the standardized production mathematical model, output the check results and prompt error messages.

[0031] This process involves connecting to ERP, MES, HR, and WMS systems via RESTful APIs, acquiring data changes since the last synchronization using incremental synchronization, converting data using the national unified manufacturing information coding standard to UTC timestamps and SI units, cleaning data using the 3σ principle to remove outliers, automatically deduplicating and filling in missing non-critical fields, establishing process dependency graphs to define predecessor and successor triggering conditions, representing processing time and preparation time as normal distribution parameters of mean ± standard deviation, calculating the maximum daily capacity of equipment and the effective daily working hours of personnel when verifying resource occupancy boundaries, calculating priority coefficients based on delivery urgency when matching order scheduling rules, defining the priority of resource reallocation in the exception response process, storing constraint logic in editable JSON format, converting it into an integer programming model containing objective function, decision variables, and constraints, performing three-level checks for data integrity, constraint consistency, and logical inconsistency, and outputting the specific location of error fields and conflict rules when checks fail. Among them, the four types of constraint logic systems refer to: the four types of constraint logic systems for generating process resource order anomalies; the process constraint logic system is used to configure the judgment logic of process sequence dependency, standard processing time, and preparation time; the resource constraint logic system is used to verify the occupancy boundaries of equipment capacity, personnel skills, and material inventory; the order constraint logic system is used to match the scheduling rules of order delivery date, priority, and batch; and the anomaly constraint logic system is used to trigger response mechanisms for equipment failure, personnel leave, and material delays. For the integer programming model, the objective function is to minimize the sum of the total production cycle and the order delay penalty, mathematically expressed as: minZ=Σ(T_i)+Σ(λ_j・max(0,F_j-D_j)), where T_i is the execution time of the i-th process, λ_j is the delay penalty coefficient of the j-th order, F_j is the final completion time of the j-th order, and D_j is the agreed delivery date of the j-th order; it includes at least three types of core constraints in its mathematical expression: first, the process sequence constraint, S_k≥F_i, where S_k is The constraints are: first, the start time of the successor process k, and F_i, the completion time of the predecessor process i; second, the equipment capacity constraint, Σ(x_{i,t})≤1, where x_{i,t} is a 0-1 decision variable indicating whether the equipment is assigned to process i in time period t; and third, the personnel skill constraint, y_{m,i}≤s_{m,e}, where y_{m,i} is a 0-1 decision variable indicating whether personnel m is assigned to process i, and s_{m,e} is a 0-1 attribute variable indicating whether personnel m has the skill to operate the equipment e required for process i.

[0032] S200: Extracting key feature vectors from standardized production mathematical models; Among them, a feature selection algorithm based on mutual information is adopted. The mutual information value between each parameter and the scheduling result is calculated by histogram estimation. The top 128 core parameters with the highest mutual information are selected. All features are mapped to the [0,1] interval by the min-max standardization formula X'=(X-Xmin) / (Xmax-Xmin). A feature cache pool is established to store the calculated feature vectors. The cache is set to expire in 1 hour. When the production factor data changes, the corresponding feature is automatically recalculated and the cache is updated.

[0033] S300: Input key feature vectors into the pre-trained AI inference model, complete the inference process through GPU parallel computing, and output a preliminary scheduling plan; In this implementation, key feature vectors are input into a pre-trained AI inference model, and the inference process is completed through GPU parallel computing to output a preliminary scheduling plan. Specifically, this includes: S301: Break down the AI ​​inference computation task into multiple independent sub-tasks; S302: Distribute the subtask to multiple GPU cores for parallel execution; S303: Merge the inference results of all GPUs to generate a complete preliminary scheduling plan; S304: Real-time monitoring of GPU resource usage and dynamic adjustment of subtask allocation strategy; S305: The pre-trained AI inference model adopts a deep neural network architecture; S306: It is obtained by pre-training based on massive historical production scheduling data; it is adapted to the production scenarios of different types of manufacturing enterprises through transfer learning.

[0034] The process involves splitting tasks into subtasks at the process group level, with each process group containing 5-10 consecutive processes with dependencies to ensure no data dependencies between subtasks. A weighted round-robin load balancing algorithm is used to distribute tasks to GPU cores. GPU utilization, memory usage, and temperature data are collected every 100 milliseconds. When the load on a single core exceeds 80%, unexecuted subtasks are migrated to cores with a load below 30%. Sub-schedule results are merged in chronological order and the process connection times are aligned. The model uses a 6-layer Transformer encoder and a 2-layer fully connected layer structure, employing a multi-head attention mechanism to capture the dependencies between production factors. It is pre-trained based on tens of millions of historical scheduling data from 12 industries, including machinery, electronics, and automobiles. During transfer learning, the bottom 6-layer encoder network is frozen, and only the top 2 fully connected layers are fine-tuned. Adaptation is completed using no less than 100 historical scheduling data from the target enterprise.

[0035] S400: Performs multi-objective secondary optimization on the preliminary scheduling plan based on preset optimization objectives to generate the final production plan; In this embodiment, the initial scheduling plan is subjected to multi-objective secondary optimization based on preset optimization objectives to generate the final production plan. Specifically, this includes: S401: Obtain the user-specified optimization target weights and calculate the comprehensive score of the preliminary scheduling plan; S402: Select the multiple candidate scheduling schemes with the highest overall scores; S403: Perform resource conflict detection and correction on the candidate scheduling scheme; S404: Output the corrected optimal scheduling scheme as the final production plan.

[0036] The system supports any combination of three objectives: shortest production cycle, highest equipment utilization, and lowest inventory cost. Users can set weights using sliders, and the sum of the weights is fixed at 1. The comprehensive scoring formula is S=w1T+w2(1-U)+w3*C, where T is the normalized value of production cycle, U is the normalized value of equipment utilization, and C is the normalized value of inventory cost. Ten preliminary scheduling schemes with different initial conditions are generated, and the top three candidate schemes with the highest comprehensive scores are selected. A one-hour sliding time window is used to detect resource conflicts and determine whether the same resource is allocated to multiple processes at the same time. Conflict correction is performed by adjusting the process time according to a 1-5 level quantitative priority system, prioritizing urgent orders, key equipment, and bottleneck processes. The comprehensive score of the corrected scheme is recalculated, and the highest-scoring scheme is selected to output three formats: Gantt chart, Excel table, and JSON data.

[0037] S500: Collects actual production execution data and compares the final production plan with the actual execution data to calculate scheduling deviation; In this implementation, actual production execution data is collected, and the scheduling deviation is calculated by comparing the final production plan with the actual execution data. Specifically, this includes: S501: Synchronously acquire real-time data on all dimensions of process execution, equipment operation, material consumption, personnel attendance, and quality inspection; S502: Compare the execution time of each process, calculate resource consumption, and statistically analyze product quality differences; S503: Summarize the deviation data of all processes to obtain the overall scheduling deviation.

[0038] Among them, vibration sensors, RFID readers and workstation touch terminals in the production site realize minute-level data collection, automatically record the actual start and end time of each process, equipment operating status, personnel clock-in information, material scanning consumption data and quality inspection results. Deviation calculation adopts the formula Δ=actual value-planned value to calculate time deviation, resource consumption deviation and quality deviation respectively. Process weight is determined by the critical path method. The process weight on the critical path is set to 2, and the process weight on the non-critical path is set to 1. All process deviations are weighted and summarized to obtain the overall scheduling deviation.

[0039] S600: Generates labeled training samples based on scheduling deviations, performs incremental training on the pre-trained AI inference model, and completes the model's self-learning iteration.

[0040] In this implementation, labeled training samples are generated based on scheduling deviations to incrementally train the pre-trained AI inference model, completing the model's self-learning iteration. Specifically, this includes: S601: Extract time dimension differences to calculate resource usage differences and statistical cost input differences; filter samples with deviation values ​​exceeding preset thresholds and add corresponding actual execution results as annotations; S602: Count the number of labeled samples and start incremental training of the model when the trigger condition is met; S603: Incrementally train the pre-trained AI inference model using the labeled training samples; S604: Evaluate the scheduling accuracy of the updated model. If the accuracy improves, enable the new model; otherwise, roll back to the previous version. S605: Display the final production plan through a visual interface and receive scheduling adjustment instructions input by the user; S606: Execute the adjustment command to generate the adjusted scheduling scheme and record all adjustment operations; add the adjusted scheduling scheme and corresponding constraints as labeled samples to the training sample library.

[0041] The process involves calculating the overall differences across time, resources, and cost, setting a 20% relative deviation threshold to filter samples, retaining only valid samples with deviations exceeding the threshold, generating labeled samples in the format of "128-dimensional feature vector + actual scheduling result," and initiating incremental training under dual trigger conditions of 500 samples or a 7-day interval. The Adam optimizer is used, with a learning rate set to 1 / 10 of the pre-training learning rate, a batch size of 32, a maximum training epoch of 10, and an early stopping mechanism that stops training when the mean absolute error on the validation set no longer decreases after 3 consecutive epochs. An independent test set is used to evaluate model accuracy; if the accuracy improves by more than 5%, the model is automatically switched to a new model; otherwise, it is rolled back to the previous stable version. Manually adjusted training weights are set to twice that of automatically collected samples. The model version management system automatically retains the most recent 10 versions and records training parameters and evaluation metrics.

[0042] Corresponding to the aforementioned embodiments of the production planning self-learning iterative optimization method, this application also provides embodiments of the production planning self-learning iterative optimization system.

[0043] Figure 8 This is a block diagram illustrating a production planning self-learning iterative optimization system according to an exemplary embodiment. (Refer to...) Figure 8 The system may include: a production planning modeling module 701, a feature extraction module 702, an AI inference scheduling module 70, a multi-objective optimization module 704, a deviation calculation module 705, and a self-learning iteration module 706, wherein: The production planning modeling module 701: collects production factor data and constraint data, and constructs a standardized production mathematical model; The feature extraction module 702 extracts key feature vectors from the standardized production mathematical model; The AI ​​inference scheduling module 703: inputs key feature vectors into the pre-trained AI inference model, completes the inference process through GPU parallel computing, and outputs a preliminary scheduling scheme; The multi-objective optimization module 704: performs multi-objective secondary optimization on the preliminary scheduling scheme according to preset optimization objectives to generate the final production plan; The deviation calculation module 705: collects actual production execution data and compares the final production plan with the actual execution data to calculate the scheduling deviation; The self-learning iteration module 706 generates labeled training samples based on scheduling deviations, performs incremental training on the pre-trained AI inference model, and completes the model's self-learning iteration.

[0044] In this embodiment, the production planning modeling module 701 collects production factor data and constraint data to construct a standardized production mathematical model; the feature extraction module 702 extracts key feature vectors from the standardized production mathematical model; the AI ​​inference scheduling module 703 inputs the key feature vectors into the pre-trained AI inference model, completes the inference process through GPU parallel computing, and outputs a preliminary scheduling plan; the multi-objective optimization module 704 performs multi-objective secondary optimization on the preliminary scheduling plan according to preset optimization objectives to generate the final production plan; the deviation calculation module 705 collects actual production execution data and calculates the scheduling deviation by comparing the final production plan with the actual execution data; and the self-learning iteration module 706 generates labeled training samples based on the scheduling deviation, performs incremental training on the pre-trained AI inference model, and completes the model's self-learning iteration.

[0045] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0046] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope.

Claims

1. A self-learning iterative optimization method for production planning, characterized in that, The steps include the following: Collect production factor data and constraint data, and construct a standardized production mathematical model; Extracting key feature vectors from standardized production mathematical models; The key feature vectors are input into the pre-trained AI inference model, and the inference process is completed through GPU parallel computing to output a preliminary scheduling plan. Based on the preset optimization objectives, the initial scheduling plan is optimized in a multi-objective secondary manner to generate the final production plan; Collect actual production execution data and compare the final production plan with the actual execution data to calculate the scheduling deviation; Based on scheduling deviations, labeled training samples are generated to incrementally train the pre-trained AI inference model, completing the model's self-learning iteration.

2. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the step of collecting production factor data and constraint data, and constructing a standardized production mathematical model: Analyze, import, and integrate information on four basic production elements: equipment, personnel, materials, and processes. The integrated information undergoes unified encoding conversion and data format standardization processing. A four-category constraint logic system for generating process resource order exceptions; Based on a standardized information and constraint logic system, a standardized production mathematical model that conforms to the norms of operations research is generated. The standardized production mathematical model is verified for completeness and rationality, and the verification results are output along with error messages.

3. The production planning self-learning iterative optimization method as described in claim 2, characterized in that, In the steps of generating the four types of constraint logic system for process resource order exceptions: Configure the logic for determining the sequence of processes, processing time, and preparation time; Verify the occupancy boundaries of equipment capacity, personnel skills, and material inventory; Match order delivery date priority batch scheduling rules; A response mechanism for equipment malfunctions, personnel leave requests, and material delays.

4. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the steps of inputting key feature vectors into a pre-trained AI inference model, completing the inference process through GPU parallel computing, and outputting a preliminary scheduling plan: The AI ​​inference computation task is broken down into multiple independent sub-tasks; The subtasks are distributed across multiple GPU cores for parallel execution. Merge the inference results from all GPUs to generate a complete preliminary scheduling plan; Monitor GPU resource usage in real time and dynamically adjust subtask allocation strategies.

5. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the step of performing multi-objective secondary optimization on the preliminary scheduling plan based on preset optimization objectives to generate the final production plan: Obtain the user-specified optimization target weights and calculate the comprehensive score of the preliminary scheduling plan; Select the candidate scheduling schemes with the highest overall scores; Resource conflict detection and correction are performed on the candidate scheduling schemes; The corrected optimal scheduling scheme is output as the final production plan.

6. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the step of collecting actual production execution data and comparing the final production plan with the actual execution data to calculate the scheduling deviation: Simultaneously acquire real-time data on all dimensions of process execution, equipment operation, material consumption, personnel attendance, and quality inspection. Compare execution time for each process, calculate resource consumption, and statistically analyze product quality differences. The overall scheduling deviation is obtained by summarizing the deviation data of all processes.

7. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the steps of generating labeled training samples based on scheduling bias, incrementally training the pre-trained AI inference model, and completing the model's self-learning iteration: Extract time dimension differences to calculate resource usage differences and statistical cost input differences; filter samples with deviation values ​​exceeding preset thresholds and add corresponding actual execution results as annotations; Count the number of labeled samples and start incremental training of the model when the trigger condition is met; The labeled training samples are used to incrementally train the pre-trained AI inference model; Evaluate the scheduling accuracy of the updated model. If the accuracy improves, enable the new model; otherwise, roll back to the previous version.

8. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, After generating labeled training samples based on scheduling bias, incrementally training the pre-trained AI inference model, and completing the model's self-learning iteration: The final production plan is displayed through a visual interface, and scheduling adjustment instructions input by the user are received. Execute adjustment instructions to generate an adjusted scheduling plan and record all adjustment operations; add the adjusted scheduling plan and corresponding constraints as labeled samples to the training sample library.

9. The production planning self-learning iterative optimization method as described in claim 1, characterized in that, In the steps of inputting key feature vectors into a pre-trained AI inference model, completing the inference process through GPU parallel computing, and outputting a preliminary scheduling plan: The pre-trained AI inference model employs a deep neural network architecture; It is pre-trained based on massive historical production scheduling data; and adapted to the production scenarios of different types of manufacturing enterprises through transfer learning.

10. A production planning self-learning iterative optimization system, employing the production planning self-learning iterative optimization method as described in claim 1, characterized in that, It includes a production planning modeling module, a feature extraction module, an AI inference scheduling module, a multi-objective optimization module, a deviation calculation module, and a self-learning iteration module, among which: The production planning modeling module collects production factor data and constraint data, and constructs a standardized production mathematical model. The feature extraction module extracts key feature vectors from the standardized production mathematical model; The AI ​​inference scheduling module inputs key feature vectors into a pre-trained AI inference model, completes the inference process through GPU parallel computing, and outputs a preliminary scheduling plan. The multi-objective optimization module performs secondary multi-objective optimization on the preliminary scheduling plan based on preset optimization objectives to generate the final production plan. The deviation calculation module collects actual production execution data and compares the final production plan with the actual execution data to calculate the scheduling deviation. The self-learning iteration module generates labeled training samples based on scheduling deviations, performs incremental training on the pre-trained AI inference model, and completes the model's self-learning iteration.

Images