Clothing intelligent manufacturing process dynamic optimization management system based on large model

Abstract

The invention relates to the technical field of industrial data processing and resource scheduling, and discloses a garment intelligent manufacturing process dynamic optimization management system based on a large model, comprising an unstructured load waveform semantic analysis module used for receiving a load demand data stream and analyzing to generate a discrete load pulse sequence, calculating a semantic cognitive entropy representing the certainty of the load request; the resource capacity space-time constraint dynamic reconstruction module is used for generating a capacity slack variable based on the semantic cognitive entropy and embedding the capacity slack variable into the resource network load distribution model as a hard physical safety margin parameter; according to the invention, a direct mapping channel from a probability semantic space to a physical resource space-time environment is established, so that the problem of resource allocation mismatch caused by unstructured load fluctuation is effectively solved; and the stability and the self-adaptive balancing capability of the production resource network in dealing with the high-uncertainty load are improved.

Description

Dynamic Optimization Management System for Intelligent Garment Manufacturing Process Based on Large Model Technical Field

[0001] This invention relates to a dynamic optimization management system for intelligent garment manufacturing processes based on a large model, belonging to the field of industrial data processing and resource scheduling technology. Background Technology

[0002] In the current process of modernization and transformation of discrete manufacturing, enterprise resource planning (ERP) systems and advanced planning and scheduling (APS) systems have become the core hubs for allocating production factors. These systems typically use linear programming, mixed integer programming, or heuristic algorithms to construct deterministic mathematical models based on defined process routes and standard working hour data, thereby achieving equipment load balancing and accurate calculation of order delivery dates. With the popularization of flexible manufacturing models, the source of production data has gradually shifted from manual input to automatic parsing of unstructured data based on artificial intelligence. Large language models or computer vision technology are used to automatically extract process parameters from original documents such as design drawings and process sheets. This technical approach has improved the throughput of data collection, but it has introduced new systemic contradictions in the underlying logic of data interaction.

[0003] Regarding the software control dimension of production processes, AI-based scheduling strategies have logical flaws. For example, Chinese invention patent CN119962938B discloses an intelligent production scheduling and anomaly management method based on a large language model. This method uses a semantic parser to convert user requests into structured data, utilizes an attention-based graph neural network to model production data, and combines a retrieval-enhanced generation framework with reinforcement learning to generate scheduling instructions. While this technology improves scheduling adaptability through multimodal data integration and dynamic strategy adjustment, it suffers from a fundamental technical blind spot: the solution essentially tends to directly accept the results generated by the large language model, focusing on using the model's capabilities to generate instructions and relying on backend reinforcement learning for correction. The processing logic lacks a pre-emptive quantification mechanism for the inherent uncertainties in the large model's parsing process, fails to establish a mathematical buffer mapping between probabilistic semantic output and rigid physical time windows, and cannot distinguish the difference in physical resource occupation risks between model-sure instructions and model-doubted instructions. This leads to the inability to pre-determine necessary safety margins when facing highly ambiguous unstructured process descriptions, relying solely on reactive adjustments afterward. This makes it difficult to fundamentally eliminate scheduling oscillations caused by model illusions or misunderstandings. However, when applying the aforementioned unstructured data parsing techniques to high-precision scheduling scenarios, existing technical solutions exhibit fundamental limitations in both logical architecture and engineering practice. The core issue lies in the dimensional mismatch between probabilistic data generation mechanisms and deterministic scheduling solution logic. Traditional scheduling models presuppose that the input parameters are physical truth values, possessing single and definite numerical attributes. However, when large neural network-based models parse fuzzy instructions or complex process descriptions, their output is essentially a predicted value based on statistical probability, carrying non-negligible cognitive uncertainty or confidence distribution characteristics. The defuzzification strategy commonly adopted by existing technologies is to directly extract the single value with the highest probability as the standard parameter input to the scheduling system. This forced dimensionality reduction process directly strips away the inherent risk characteristics of the data.

[0004] Therefore, the technical problem to be solved by this invention is how to establish a method to quantify the cognitive uncertainty in the process of unstructured data parsing and transform it into a dynamically variable elastic constraint boundary in the scheduling model, thereby solving the logical disconnect between probabilistic input and deterministic solution. Summary of the Invention

[0005] To address the problems raised in the background art, the technical solution of this invention is as follows: A dynamic optimization management system for intelligent garment manufacturing processes based on a large model, comprising: an unstructured load waveform semantic parsing module, used to receive unstructured load demand data streams, use a large language model to parse and generate discrete load pulse sequences, and simultaneously extract the probability distribution data of the large language model when generating each discrete load pulse unit, so as to calculate the semantic cognitive entropy characterizing the determinism of the load request; and a resource capacity spatiotemporal constraint dynamic reconstruction module, connected to the unstructured load waveform semantic parsing module, used to perform nonlinear mapping operations based on the semantic cognitive entropy to generate an independent capacity for each discrete load pulse unit. The system measures and incorporates capacity slack variables as hard physical safety margin parameters, embedding them into the occupancy slot inequality constraints of the resource network's load allocation model. This constructs a robust allocation model with load fluctuation suppression capabilities. A laminar load balancing solution module, connected to a resource capacity spatiotemporal constraint dynamic reconstruction module, is used to solve the robust allocation model and outputs resource scheduling instructions that reserve non-uniform capacity buffers for high-entropy load pulse units. By directly transforming semantic cognitive entropy, an information-theoretic index, into capacity slack variables as physical scheduling boundaries, the system establishes a direct mapping channel from the probabilistic semantic space to the deterministic physical resource spatiotemporal environment within the robust allocation model.

[0006] Preferably, the resource capacity spatiotemporal constraint dynamic reconstruction module includes: an entropy-capacity margin mapping submodule, used to store and execute a preset mapping rule, which defines a positively correlated monotonic function relationship between the numerical range of semantic cognitive entropy and the numerical value of capacity slack variables, so that the higher the semantic cognitive entropy, the larger the value of the generated capacity slack variable used to offset the risk of load mutation; and a constraint boundary expansion submodule, used to multiply the preset standard physical baseline time using the capacity slack variable to form the minimum resource occupancy threshold in the occupancy time slot inequality constraint condition.

[0007] Preferably, the system further includes: a load feature vector distance calculation module, used to convert multiple load access requests to be allocated into high-dimensional physical feature vectors and calculate the feature vector distance between any two load access requests; the laminar load balancing solution module further optimizes by minimizing the sum of feature vector distances between adjacent discrete load pulse units, thereby generating a laminar load sequence with continuously changing physical medium properties.

[0008] Preferably, when constructing the robust assignment model, the constraint boundary expansion submodule targets the first... For each discrete load pulse unit, the time slot inequality constraint is constructed as follows: ,in, The planned resource allocation time slot allocated by the system to the i-th discrete load pulse unit. Let μ be the physical reference time for the i-th discrete load pulse unit, and μ be the preset global resource conversion efficiency coefficient. The capacity relaxation variable is generated by the entropy-capacity margin mapping submodule based on the semantic cognitive entropy of the i-th discrete load pulse unit.

[0009] Preferably, the system further includes: a real-time load response feedback and deviation calibration module, used to collect execution phase deviation data during the actual operation of the resource network; and an unstructured load waveform semantic parsing module, which is also used to: correct the output confidence parameters of the large language model based on the execution phase deviation data, and dynamically update the calculation benchmark of semantic cognitive entropy in subsequent parsing, thereby forming a closed-loop self-evolution mechanism for load prediction accuracy.

[0010] Preferably, the load feature vector distance calculation module includes: a physical medium attribute extraction submodule, used to extract key load physical attributes, including medium impedance characteristics, connection topology type, and auxiliary resource specifications, from the unstructured load demand data stream; a switching impedance matrix generation submodule, used to quantify and generate a state switching impedance matrix based on the feature vector distance, representing the resource reset time data required for switching between different load access requests; and a laminar load balancing solution module that directly calls the state switching impedance matrix as the impedance penalty term parameter for the full sequence of resource load allocation.

[0011] Preferably, the system further includes: a nonlinear coupling correlation analysis module, used to identify the physical coupling relationship between adjacent units in a discrete load pulse sequence; and a resource capacity spatiotemporal constraint dynamic reconstruction module, which, when generating capacity relaxation variables, also introduces a transfer superposition factor based on the physical coupling relationship, used to cascade and amplify the capacity relaxation variables of continuous high-entropy load units on critical resource paths.

[0012] Preferably, the unstructured load waveform semantic parsing module is specifically used to: accumulate or weight the logarithmic probability of each token output by the large language model to obtain the sequence-level semantic cognitive entropy; and perform sensitivity marking on the fuzzy qualifiers contained in the unstructured load demand data stream, and when a fuzzy qualifier is detected, forcibly increase the numerical level of the semantic cognitive entropy of the corresponding discrete load pulse unit.

[0013] Preferably, the system further includes: a multimodal data verification interface module for receiving image stream data or video stream data from the trial production verification stage; and an unstructured load waveform semantic parsing module for cross-comparing the image stream data or video stream data with the unstructured load demand data stream, and adjusting the final output value of the semantic cognitive entropy based on the consistency of the comparison results.

[0014] Preferably, the laminar load balancing solution module includes: a mixed integer programming solution submodule, used to find the solution vector with the shortest total job cycle or the highest resource utilization rate under the premise of satisfying the occupancy time slot inequality constraint; and a dynamic reallocation triggering submodule, used to trigger the mixed integer programming solution submodule to perform local resource reallocation calculation based on the remaining job load when the actual occupied time slot exceeds the range covered by the capacity slack variable is detected during the operation of the resource network.

[0015] Compared with existing technologies, the beneficial effects of this invention are: 1. In the intelligent manufacturing process of clothing, it realizes adaptive compensation for scheduling risks based on semantic cognitive entropy. This invention changes the traditional scheduling system's processing logic of treating the output of a large language model as an absolute truth value. By extracting the labeled log probability sequence in the semantic parsing process and calculating the semantic cognitive entropy, the inherent cognitive uncertainty in unstructured data processing is numerically represented. Based on this entropy value, the system dynamically generates time slack variables in a mixed integer programming model and constructs inequality constraints with elastic buffers, so that the generated production scheduling plan naturally has the ability to resist data noise: for mature processes with high model confidence, the system automatically generates compact time windows to ensure efficiency; for new or fuzzy processes with low model confidence, the system automatically reserves slack space to absorb potential execution fluctuations. This resource allocation method based on information quality solves the mismatch problem between probabilistic generated data and deterministic scheduling logic at the mathematical model level, avoiding the disconnect between planning and execution caused by initial parameter deviations, thereby ensuring the executability and stability of the scheduling scheme in complex workshop environments.

[0016] 2. Establishing a resource-precise adaptation logic based on micro-behavioral spectrum: This invention analyzes the timing distribution data of motor actions during production equipment operation, calculates the interaction density characteristics of effective working time and auxiliary operation time, and maps them to skill friction loss labels with management significance. This breaks down the originally general attribution of time delays into specific resource adaptation problems. When the system detects high-frequency intermittent distribution characteristics, it no longer simply increases the time budget, but adjusts the affinity weight in the resource compatibility matrix to force the solver to transfer high-difficulty processes to resource nodes with corresponding skill characteristics in subsequent rescheduling. This logical mapping from physical waveforms to management weights realizes the transformation from passive time correction to proactive resource optimization, effectively reducing the implicit capacity loss caused by human-machine skill mismatch and improving the effective output rate per unit time.

[0017] 3. Constructing a laminar flow production sequence based on semantic distance: This invention utilizes the parsed process attribute feature set to calculate the semantic vector distance between different orders in the scheduling queue and transform it into a process switching cost matrix. When constructing the scheduling objective function, the system introduces a collaborative optimization objective that minimizes the full sequence switching time penalty value, enabling the solver to generate a production sequence with gradually changing process attributes. By utilizing the clustering effect in mathematical programming, fragmented heterogeneous orders are reorganized into a quasi-laminar flow operation mode with logical continuity. Without changing the physical speed of single-piece processing, the marginal connection loss caused by frequent changes in line color, tension adjustment, or auxiliary material replacement is reduced by decreasing the process span between adjacent tasks, thereby uncovering the potential throughput of the production line in a multi-variety mixed flow mode. Attached Figure Description

[0018] Figure 1 is a diagram of the core module architecture and logical interaction principle of the system of the present invention; Figure 2 is a diagram of the full-link data processing flow and resource scheduling execution of the present invention. Detailed Implementation

[0019] The technical solutions of the embodiments of this application will be clearly described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of this application are within the scope of protection of this application.

[0020] A dynamic optimization management system for intelligent garment manufacturing processes based on a large model includes: an unstructured load waveform semantic parsing module, used to receive unstructured load demand data streams, use a large language model to parse and generate discrete load pulse sequences, and simultaneously extract the probability distribution data of the large language model when generating each discrete load pulse unit to calculate the semantic cognitive entropy representing the determinism of the load request; and a resource capacity spatiotemporal constraint dynamic reconstruction module, connected to the unstructured load waveform semantic parsing module, used to perform nonlinear mapping operations based on semantic cognitive entropy, generate independent capacity relaxation variables for each discrete load pulse unit, and apply the capacity relaxation... The variable, as a hard physical safety margin parameter, is embedded in the occupancy slot inequality constraint of the load allocation model of the resource network, thereby constructing a robust allocation model with load fluctuation suppression capability. The laminar load balancing solution module is connected to the resource capacity spatiotemporal constraint dynamic reconstruction module to solve the robust allocation model and output resource scheduling instructions containing a non-uniform capacity buffer band reserved for high-entropy load pulse units. The system establishes a direct mapping channel from the probabilistic semantic space to the deterministic physical resource spatiotemporal environment within the robust allocation model by directly transforming the information theory index of semantic cognitive entropy into the physical scheduling boundary of capacity relaxation variable.

[0021] Preferably, the resource capacity spatiotemporal constraint dynamic reconstruction module includes: an entropy-capacity margin mapping submodule, used to store and execute a preset mapping rule, which defines a positively correlated monotonic function relationship between the numerical range of semantic cognitive entropy and the numerical value of capacity slack variables, so that the higher the semantic cognitive entropy, the larger the value of the generated capacity slack variable used to offset the risk of load mutation; and a constraint boundary expansion submodule, used to multiply the preset standard physical baseline time using the capacity slack variable to form the minimum resource occupancy threshold in the occupancy time slot inequality constraint condition.

[0022] Preferably, the system further includes: a load feature vector distance calculation module, used to convert multiple load access requests to be allocated into high-dimensional physical feature vectors and calculate the feature vector distance between any two load access requests; the laminar load balancing solution module further optimizes by minimizing the sum of feature vector distances between adjacent discrete load pulse units, thereby generating a laminar load sequence with continuously changing physical medium properties.

[0023] Preferably, when constructing the robust allocation model, the constraint boundary expansion submodule constructs the occupancy time slot inequality constraint condition for the i-th discrete load pulse unit as follows: ,in, The planned resource allocation time slot allocated by the system to the i-th discrete load pulse unit. Let μ be the physical reference time for the i-th discrete load pulse unit, and μ be the preset global resource conversion efficiency coefficient. The capacity relaxation variable is generated by the entropy-capacity margin mapping submodule based on the semantic cognitive entropy of the i-th discrete load pulse unit.

[0024] Preferably, the system further includes: a real-time load response feedback and deviation calibration module, used to collect execution phase deviation data during the actual operation of the resource network; and an unstructured load waveform semantic parsing module, which is also used to: correct the output confidence parameters of the large language model based on the execution phase deviation data, and dynamically update the calculation benchmark of semantic cognitive entropy in subsequent parsing, thereby forming a closed-loop self-evolution mechanism for load prediction accuracy.

[0025] Preferably, the load feature vector distance calculation module includes: a physical medium attribute extraction submodule, used to extract key load physical attributes, including medium impedance characteristics, connection topology type, and auxiliary resource specifications, from the unstructured load demand data stream; a switching impedance matrix generation submodule, used to quantify and generate a state switching impedance matrix based on the feature vector distance, representing the resource reset time data required for switching between different load access requests; and a laminar load balancing solution module that directly calls the state switching impedance matrix as the impedance penalty term parameter for the full sequence of resource load allocation.

[0026] Preferably, the system further includes: a nonlinear coupling correlation analysis module, used to identify the physical coupling relationship between adjacent units in a discrete load pulse sequence; and a resource capacity spatiotemporal constraint dynamic reconstruction module, which, when generating capacity relaxation variables, also introduces a transfer superposition factor based on the physical coupling relationship, used to cascade and amplify the capacity relaxation variables of continuous high-entropy load units on critical resource paths.

[0027] Preferably, the unstructured load waveform semantic parsing module is specifically used to: accumulate or weight the logarithmic probability of each token output by the large language model to obtain the sequence-level semantic cognitive entropy; and perform sensitivity marking on the fuzzy qualifiers contained in the unstructured load demand data stream, and when a fuzzy qualifier is detected, forcibly increase the numerical level of the semantic cognitive entropy of the corresponding discrete load pulse unit.

[0028] Preferably, the system further includes: a multimodal data verification interface module for receiving image stream data or video stream data from the trial production verification stage; and an unstructured load waveform semantic parsing module for cross-comparing the image stream data or video stream data with the unstructured load demand data stream, and adjusting the final output value of the semantic cognitive entropy based on the consistency of the comparison results.

[0029] Preferably, the laminar load balancing solution module includes: a mixed integer programming solution submodule, used to find the solution vector with the shortest total job cycle or the highest resource utilization rate under the premise of satisfying the occupancy time slot inequality constraint; and a dynamic reallocation triggering submodule, used to trigger the mixed integer programming solution submodule to perform local resource reallocation calculation based on the remaining job load when the actual occupied time slot exceeds the range covered by the capacity slack variable is detected during the operation of the resource network.

[0030] Example 1: In a flexible production workshop scenario for smart garment manufacturing, where the daily order volume exceeds 2000 styles and the complexity of fabric and process combinations increases exponentially, the system faces the challenge of processing order data from the design end that includes an unstructured process description of using high-elastic velvet fabric for bias cutting and splicing. The physical characteristics implied in this description can easily cause fabric slippage and deformation during actual sewing, making standard operating time highly unstable. If traditional deterministic scheduling logic is used, the system only performs tight scheduling based on a single average estimated working time. Once a small process delay occurs on-site, it will trigger a cascading blockage and delivery default across the entire production line. To address this situation, the unstructured load waveform semantic parsing module receives the unstructured load demand data stream containing the unstructured process description. Using a large language model, it maps the continuous process text into a discrete load pulse sequence that can be recognized by the computer. The system simultaneously extracts the large language... The model generates probability distribution data for discrete load pulse units such as velvet bias stitching. When the model outputs the difficulty parameter corresponding to the process, its output labeled log probability sequence exhibits discrete characteristics. The system truncates and extracts the top 5 highest probability candidates from the model output, sums the original probability values ​​of these 5 candidates, and divides the independent probability value of each candidate by the summation value to complete the local probability quality renormalization, ensuring that the sum of probabilities within the calculation window equals 1. This indicates that the model has a high cognitive uncertainty regarding the execution difficulty of this specific process based on historical data. Based on this statistical characteristic, the module calculates that the process node has a high semantic cognitive entropy value. The resource capacity spatiotemporal constraint dynamic reconstruction module receives the semantic signal with this high entropy value. Its internal entropy-capacity margin mapping submodule executes a preset nonlinear mapping rule to convert the high level of semantic cognitive entropy into a numerical capacity slack variable. For example, if set to 0.35, the system prioritizes the physical feasibility of scheduling. Its internal constraint boundary expansion submodule uses this variable to reconstruct the mathematical model, constructing an occupancy slot inequality constraint condition containing a hard physical safety margin for the i-th discrete load pulse unit. ,in The planned resource allocation time slot allocated by the system to the i-th discrete load pulse unit. The physical baseline time for the i-th discrete load pulse unit is given by μ, which is the preset global resource conversion efficiency coefficient. Through this step, the system forcibly reserves an additional 35% time buffer for this high-risk process at the mathematical programming level, making the originally implicit cognitive risk explicit into a calculable resource occupation cost.

[0031] Finally, the laminar load balancing solution module solves this robust allocation model with elastic constraints. Because the model incorporates capacity slack variables for high-entropy nodes, the solver automatically allocates the complex order to a resource slot with high redundancy and postpones the start time of subsequent tasks when searching for the global optimum. This generates resource scheduling instructions with continuously changing physical medium properties. When the production instruction is issued to the workshop for execution, even if the actual sewing time of the velvet fabric exceeds the standard working time by 20% due to slippage, the delay still falls entirely within the system's preset capacity slack variables. Within the covered area, actual system operation data shows that the high-difficulty orders in this batch achieved smooth flow without interfering with other parallel tasks. The overall cycle time of the production line remained above 98%. By quantitatively transforming the semantic cognitive entropy at the data parsing level into the capacity relaxation variable at the management decision level, the logical contradiction between probabilistic AI reasoning and deterministic industrial control was resolved within a single architecture. This enabled the manufacturing management system to shift from passively responding to deviations to proactively managing risks. When constructing resource scheduling instructions, the laminar load balancing solution module maps the process feature vectors of the orders to be scheduled to a high-dimensional Euclidean space. By calculating the feature vector distance between adjacent discrete load pulse units, the process switching cost is quantified. A full-sequence switching cost penalty term is introduced into the objective function of the mixed integer programming model, so that the generated production sequence presents a continuous and gradual state in physical attributes. This reduces the auxiliary time caused by frequent changes in line color, tension adjustment, or replacement of auxiliary materials. The laminar flow operation mode is achieved by utilizing the clustering effect of mathematical programming, which increases the overall throughput of the production line without changing the processing speed of individual equipment.

[0032] Example 2: In a high-fidelity intelligent garment manufacturing process simulation verification environment built based on digital twin technology, this experiment aims to objectively verify the engineering effectiveness and boundary stability of the resource capacity spatiotemporal constraint dynamic reconstruction module in handling high uncertainty loads, addressing the technical challenge of production scheduling failures and resource utilization fluctuations caused by unstructured process descriptions. The experimental platform is deployed on a high-performance computing cluster equipped with 128 processors and 512GB of memory, running a discrete event simulation model that accurately maps 50 sewing production lines and their supporting logistics system in a physical workshop. The experimental data comes from anonymized historical production records of a garment manufacturing company, covering 2000 real order data, including approximately 400 orders containing fuzzy semantic descriptions such as high-elastic fabric splicing and special dart treatment. To ensure the test results reflect the complexity of the industrial environment, for highly complex orders, Gaussian white noise with a signal-to-noise ratio of 20dB was superimposed at the data input end to simulate manual data entry errors. A random disturbance factor with a mean of 0 and a variance of 0.15 was introduced during the simulation to simulate common engineering environmental noises such as fluctuations in worker efficiency and minor equipment downtime. The test design followed a multi-dimensional control system, establishing three independent test sample groups for comparative verification. The control sample group adopted traditional APS scheduling logic, performing deterministic scheduling based on standard working hours without introducing any slack variables. The sample group of this invention adopted the technical solution described in this specification, activating the unstructured load waveform semantic parsing module and the resource capacity spatiotemporal constraint dynamic reconstruction module, dynamically generating capacity slack variables based on real-time calculated semantic cognitive entropy. The out-of-range control group was set to a global fixed relaxation mode, that is, no semantic parsing was performed, and a fixed time redundancy of 50% was uniformly applied to all processes. The setting of the key parameter global resource conversion efficiency coefficient μ followed the calibration procedure based on historical capacity data. By backtracking the ratio of actual output to theoretical working hours in the past three months, the weighted average value of 0.85 was taken as the test benchmark to balance the systematic deviation between theoretical capacity and actual loss.

[0033] In the first phase of the experiment, a mixed load stream containing a process with 20% high semantic cognitive entropy was synchronously injected into each sample group. For the control group, the system read the standard time of the velvet bias cutting process as 120 seconds and arranged a tight sequence of preceding and following tasks accordingly. However, simulation data showed that due to random disturbance factors, the actual execution time of this process fluctuated between 135 and 150 seconds, causing a cascading delay in subsequent task chains. Ultimately, monitoring data showed that the on-time delivery rate of orders in the control group was only 68.5%, and the resource utilization rate of the production line was significantly reduced due to frequent scheduling adjustments. The data exhibited severe fluctuations, with a mean squared error of 0.24, confirming that deterministic models cannot adapt to the internal fluctuations of unstructured loads in the absence of an uncertainty measurement mechanism. In the second phase of the experiment, the operational data of the sample group showed differences. The unstructured load waveform semantic parsing module analyzed the input unstructured load demand data stream. For the aforementioned velvet bias cutting process, the probability distribution output by the large language model showed high dispersion, and the calculated semantic cognitive entropy H reached 2.8 bits. The entropy-capacity margin mapping submodule, based on a preset nonlinear mapping rule, transformed it into a capacity relaxation variable. Its value is 0.32, and the system constructs the corrected occupancy slot inequality constraint, namely... This reserved approximately 135 seconds of planned time slot for the process. During simulation, although the actual physical time of the process was still affected by disturbances, reaching 142 seconds, the fluctuation remained within a certain range. The constructed elastic buffer zone edge allows the system to fine-tune the start times of subsequent tasks without reconstructing the entire scheduling graph. Final data shows that the on-time delivery rate of the sample group improved to 96.2%, and while maintaining a high delivery rate, the root mean square error of production line resource utilization decreased to 0.08, exhibiting laminar flow stability characteristics. In the third stage of the experiment, data analysis of the out-of-range control group revealed the rationality of the parameter boundaries. Data showed that although the on-time delivery rate of this group reached 98.5%, slightly higher than that of the sample group of the present invention, its overall resource utilization rate dropped significantly to below 55%, far lower than the 88% of the sample group of the present invention. This nonlinear inflection point... The results show that while blindly increasing slack variables can suppress delays, it can lead to resource idleness and soaring costs. Comparative results demonstrate that the on-demand slack mechanism achieved by semantic cognitive entropy in this invention is a key technical path to achieve the optimal balance between delivery risk and resource cost. Furthermore, gradient testing of the semantic cognitive entropy threshold reveals that when the threshold is set too low, the system degenerates into an inefficient mode similar to the out-of-range control group; while when the threshold is too high, the system degenerates into a high-risk mode similar to the control group. Experimental data confirms that when the semantic cognitive entropy threshold is set in the range of 1.5 to 3.0, the system can achieve a synergistic effect with both high delivery rate and utilization rate.

[0034] Example 3: Regarding the specific calculation path of semantic cognitive entropy in the unstructured load waveform semantic parsing module and the parameter calibration procedure of the nonlinear mapping rule in the entropy-capacity margin mapping submodule, this example transparently constructs the core algorithm logic. When performing deep parsing, the unstructured load waveform semantic parsing module establishes a probability distribution acquisition channel based on a sliding window mechanism. It extracts the original Top-K Logits values ​​generated from the decoding layer of the large language model for the text describing a specific process. The system does not calculate the entropy value for all output texts, but uses a preset industrial entity dictionary to anchor key feature words and extracts the probability distribution data of the feature word positions involving process actions and material attributes. For the key feature word w in the i-th discrete load pulse unit, the module normalizes the original Logits values ​​into a probability distribution set using the Softmax function. Based on Shannon's definition of information entropy, the local uncertainty index of the feature word is calculated, and the indices of all key feature words within the unit are weighted and summed to obtain the semantic cognitive entropy of the process node. Its calculation formula is Where M is the number of key feature words contained in the process description. The weighting coefficients are determined by the reciprocal of the frequency of feature words in the process ontology library. K is the size of the candidate word sampling window output by the model. The system defaults to K=5 to cover the main high-probability candidate regions. The unstructured load waveform semantic parsing module obtains the log probability data sequence corresponding to the output tokens through the large language model inference interface. It uses a statistical window to extract the probability distribution of feature word positions related to fabric characteristics and process difficulty. The log probability data is converted into a normalized probability distribution to calculate the Shannon entropy, thus obtaining the semantic cognitive entropy that represents the cognitive determinism of the model. , The confidence metric is predicted for the i-th process node. The entropy value is constrained by the frequency of process distribution in the training corpus, reflecting the confidence bias of the model in predicting the execution time of fabric characteristics and material combinations. A statistical correlation is established between semantic uncertainty and physical workstation execution risk.

[0035] The entropy-capacity margin mapping submodule receives the above quantization. Instead of using a simple linear scaling transformation, a sigmoid logistic response function with truncation characteristics is used to generate capacity slack variables. To simulate the diminishing marginal returns of risk in industrial scenarios, the mapping logic is constructed as follows: ,in The maximum relaxation boundary allowed by the physical system is set to 0.5, which means that a maximum of 50% additional buffer time is allowed. To determine the risk trigger threshold for cognitive entropy, the system analyzes the entropy distribution of historical normal work orders and selects its 95th percentile value as the threshold. The initial setpoint is, for example, 2.5 bits; α is the risk sensitivity coefficient, used to adjust the response rate of the slack variable to changes in entropy, initially set to 1.5. This function logic ensures that when the semantic cognitive entropy... Below the threshold hour, Approaching zero, the system maintains a compact schedule; while when After exceeding the threshold, Rising rapidly and approaching The system tends to saturate, thus preventing unreasonable resource consumption caused by excessive model hallucination.

[0036] To ensure the adaptability of the aforementioned imaging model during long-term operation, the system integrates a real-time load response feedback and deviation calibration module, executes a periodic parameter post-hoc correction process, and collects actual execution phase deviation data for each process at the end of each production cycle. And construct the target loss function. Using the gradient descent algorithm, the module calculates the loss function with respect to parameters α and β. The partial derivatives are calculated, and these two core parameters are updated in reverse according to the preset learning rate η. When a specific type of process is detected... When the value continuously exceeds the system's reserved buffer, the algorithm automatically reduces the buffer size. Alternatively, increasing α allows for the application of more aggressive relaxation strategies for this type of process in subsequent scheduling, transforming the originally static parameter settings into an adaptive control process that dynamically evolves with the physical production environment. This ensures that the resource allocation model remains mathematically isomorphic to the actual entropy reduction trend in the workshop. The entropy-capacity margin mapping submodule executes a nonlinear transformation procedure using the logistic function, calculates the average entropy value by selecting a set of historical steady-state work orders with a time deviation rate of less than 5%, and sets the quantile value as the risk trigger threshold. Capacity slack variables are generated in conjunction with the risk sensitivity coefficient α. The actual execution phase deviation ΔT collected at the production site consistently exceeds the capacity slack variable. When constructing an elastic buffer band, the gradient descent algorithm is used to calculate the loss function for parameters α and β. The partial derivatives are updated in reverse, causing the minimum resource occupancy threshold in the occupancy time slot inequality constraint to converge to the optimal configuration range as the physical production environment fluctuates.

[0037] Example 4: Addressing the cold-start recognition accuracy deviation and parameter adaptability issues of the entropy-capacity margin mapping submodule in different manufacturing scenarios faced by the unstructured load waveform semantic parsing module during the initial system deployment phase, this example transparently constructs the pre-deployment calibration and debugging procedures for the system. Before the system is formally connected to the workshop production network, an offline calibration process based on a standard process corpus is executed to construct an initial semantic cognitive entropy calculation benchmark. This process selects a desensitized dataset containing no fewer than 5000 historical process description texts, covering the entire spectrum of process types from basic flat stitching to complex three-dimensional cutting. Senior industrial engineers manually annotate each text to establish a baseline ground truth value for its physical operation difficulty. A large language model is used to perform a full extrapolation of this dataset, collecting the logarithmic probability distribution data output by the model and calculating the corresponding semantic cognitive entropy. The system fits the difficulty value of manual annotation with the calculated value using the least squares method. The correlation curve between them was used to determine the initial normalization coefficient β of semantic cognitive entropy, ensuring that the entropy value output by the model can be linearly mapped to the fluctuation range of physical operation time, eliminating the systematic recognition error caused by the bias of the training data of the model itself, and laying a quantifiable benchmark for subsequent real-time analysis.

[0038] After completing offline calibration, the system entered the dynamic parameter fine-tuning procedure during the on-site trial operation phase. This eliminated the influence of environmental specificity on the capacity slack variable generation mechanism. During the first production cycle, the system ran in parallel in shadow mode to calculate and record the recommended capacity slack variables for each process. Instead of actually issuing the schedule to the execution end, the system simultaneously collects the actual process time deviation of the physical production line in real time. And calculate the theoretical buffer time corresponding to the recommended slack variables. Coverage metrics between When detected When the value falls below 95% of the preset safety threshold, it indicates that the uncontrollable disturbances in the current environment exceed the model's expectations. The system automatically triggers the parameter adaptive correction logic, gradually increasing the risk sensitivity coefficient α in steps of 0.05 until... By returning to the safe range, the system can automatically converge to the optimal combination of relaxation strategy parameters when faced with differences in the proficiency of personnel and the aging of equipment in different factories, thereby achieving engineering implementation without human intervention.

[0039] Example 5: Weighting coefficients of key feature words w in the semantic parsing module of unstructured load waveforms Risk trigger threshold in the entropy-capacity margin mapping submodule The initial calibration and adaptive evolution procedure, in this embodiment, transparently constructs the systematic determination method for the core parameters; during the system cold start phase, a static weight calibration process for feature words based on inverse text frequency is executed. The system accesses a historical corpus containing no less than 100,000 process orders in the garment manufacturing field, extracts all process feature words through word segmentation and entity recognition algorithms, and counts the document occurrence frequency df(w) of each feature word w in the corpus, using the formula The initial weights are calculated, where N is the total number of documents in the corpus. To correct for the bias of purely statistical methods in specific vertical domains, the system introduces an expert knowledge calibration mechanism. For highly sensitive words involving fabric properties and special processes, an artificial gain coefficient γ is set to adjust their weights. This calibration process ensures that the system can still accurately capture key semantic features that affect working hours even in the absence of real-time feedback data.

[0040] Risk trigger thresholds for the entropy-capacity margin mapping submodule The system employs a statistical calibration procedure based on historical work hour distribution, retrospectively analyzing the actual work hour deviation rate data of all completed work orders within the past year. It then filters out a set of standard steady-state work orders with an absolute deviation rate of less than 5%. For each work order in this set, the system calculates the semantic cognitive entropy of the corresponding process description text, constructs a steady-state entropy value distribution histogram, and selects the 95th percentile of this distribution as the initial... This means that the semantic uncertainty of 95% of normal work orders should be below this threshold. During system online operation, to cope with dynamic changes in the production environment, a sliding window adaptive update logic is integrated. The system maintains a queue of recently completed work orders of length L, and monitors the entropy distribution changes of on-time delivered work orders in this queue in real time. If the 95th quantile of this distribution is detected to be different from the current... If the deviation exceeds the preset drift threshold ϵ, the system will automatically trigger a threshold reset operation. The statistical quantiles of the current window are updated to ensure that the system can automatically adapt to the baseline of process complexity under different seasons and fabric batches, thus achieving dynamic stability of parameter settings.

[0041] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A dynamic optimization management system for intelligent garment manufacturing processes based on a large model, characterized in that, include: The unstructured load waveform semantic parsing module receives unstructured load demand data streams, uses a large language model to parse and generate discrete load pulse sequences, and simultaneously extracts the probability distribution data of the large language model when generating each discrete load pulse unit to calculate the semantic cognitive entropy representing the determinism of load requests. The resource capacity spatiotemporal constraint dynamic reconstruction module, connected to the unstructured load waveform semantic parsing module, performs nonlinear mapping operations based on semantic cognitive entropy to generate independent capacity relaxation variables for each discrete load pulse unit. The capacity relaxation variables are then used as hard physical safety margin parameters and embedded into the occupancy slot inequality constraint conditions of the load allocation model of the resource network, thereby constructing a robust allocation model with load fluctuation suppression capabilities. The laminar load balancing solution module, connected to the resource capacity spatiotemporal constraint dynamic reconstruction module, solves the robust allocation model and outputs resource scheduling instructions containing non-uniform capacity buffer bands reserved for high-entropy load pulse units. By directly transforming the information-theoretic index of semantic cognitive entropy into the physical scheduling boundary of capacity relaxation variables, the system establishes a direct mapping channel from the probabilistic semantic space to the deterministic physical resource spatiotemporal environment within the robust allocation model.

2. The garment intelligent manufacturing process dynamic optimization management system based on a large model as described in claim 1, characterized in that, The resource capacity spatiotemporal constraint dynamic reconstruction module includes: an entropy-capacity margin mapping submodule, which stores and executes a preset mapping rule. This rule defines a positively correlated monotonic function relationship between the numerical range of semantic cognitive entropy and the numerical value of the capacity slack variable, so that the higher the semantic cognitive entropy, the larger the value of the generated capacity slack variable used to offset the risk of load mutation; and a constraint boundary expansion submodule, which uses the capacity slack variable to multiply expand the preset standard physical baseline time.

3. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 2, characterized in that, The system also includes: a load feature vector distance calculation module, which converts multiple load access requests to be allocated into high-dimensional physical feature vectors and calculates the feature vector distance between any two load access requests; and a laminar load balancing solution module, which further optimizes the solution by minimizing the sum of feature vector distances between adjacent discrete load pulse units.

4. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 2, characterized in that, When constructing a robust assignment model, the constraint boundary expansion submodule targets the first... For each discrete load pulse unit, the time slot inequality constraint is constructed as follows: ,in, The planned resource allocation time slot for the i-th discrete load pulse unit is defined by the system. Let μ be the physical reference time for the i-th discrete load pulse unit, and μ be the preset global resource conversion efficiency coefficient. The capacity relaxation variable is generated by the entropy-capacity margin mapping submodule based on the semantic cognitive entropy of the i-th discrete load pulse unit.

5. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 1, characterized in that, The system also includes: a real-time load response feedback and deviation calibration module, used to collect execution phase deviation data during the actual operation of the resource network; and an unstructured load waveform semantic parsing module, which is used to: correct the output confidence parameters of the large language model based on the execution phase deviation data, and dynamically update the calculation benchmark of semantic cognitive entropy in subsequent parsing, thereby forming a closed-loop self-evolution mechanism for load prediction accuracy.

6. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 3, characterized in that, The load feature vector distance calculation module includes: a physical medium attribute extraction submodule, used to extract key load physical attributes, including medium impedance characteristics, connection topology type, and auxiliary resource specifications, from the unstructured load demand data stream; a switching impedance matrix generation submodule, used to quantify and generate a state switching impedance matrix based on the feature vector distance, representing the resource reset time data required for switching between different load access requests; and a laminar load balancing solution module that directly calls the state switching impedance matrix as the impedance penalty term parameter for the full sequence of resource load allocation.

7. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 1, characterized in that, The system also includes: a nonlinear coupling correlation analysis module, used to identify the physical coupling relationship between adjacent units in a discrete load pulse sequence; and a resource capacity spatiotemporal constraint dynamic reconstruction module, which, when generating capacity relaxation variables, also introduces a transfer superposition factor based on the physical coupling relationship, used to cascade and amplify the capacity relaxation variables of continuous high-entropy load units on critical resource paths.

8. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 1, characterized in that, The unstructured load waveform semantic parsing module is specifically used to: accumulate or weight the logarithmic probability of each token output by the large language model to obtain the sequence-level semantic cognitive entropy; and perform sensitivity marking on the fuzzy qualifiers contained in the unstructured load demand data stream. When a fuzzy qualifier is detected, the numerical level of the semantic cognitive entropy of the corresponding discrete load pulse unit is forcibly increased.

9. The garment intelligent manufacturing process dynamic optimization management system based on a large model according to claim 1, characterized in that, The system also includes: a multimodal data verification interface module, used to receive image stream data or video stream data from the trial production verification stage; and an unstructured load waveform semantic parsing module that cross-compares the image stream data or video stream data with the unstructured load demand data stream, and adjusts the final output value of semantic cognitive entropy based on the consistency of the comparison results.

10. The dynamic optimization management system for intelligent garment manufacturing process based on a large model according to claim 1, characterized in that, The laminar load balancing solution module includes: a mixed integer programming solution submodule, which is used to find the solution vector with the shortest total job cycle or the highest resource utilization rate under the premise of satisfying the occupancy time slot inequality constraint; and a dynamic reallocation triggering submodule, which is used to trigger the mixed integer programming solution submodule to perform local resource reallocation calculation based on the remaining job load when the actual occupied time slots are detected to exceed the range covered by the capacity slack variable during the operation of the resource network.

Images