Process model variable rank fine tuning method and device for textile weaving process

Abstract

The invention discloses a process model variable rank fine tuning method and device for a textile weaving process, relates to the technical field of weaving process optimization, and solves the problem that an existing'static rank 'mechanism is difficult to meet working condition change requirements of strong nonlinearity, strong coupling and time-varying characteristics in the weaving process. The method comprises the following steps: constructing a multi-source heterogeneous weaving comprehensive data system, calculating a process complexity increment and a production stability increment based on the multi-source heterogeneous weaving comprehensive data system, calculating a total rank value of a process model under a current working condition based on a basic initial rank of an LoRA module, the process complexity increment and the production stability increment, and calculating the total rank value of the process model under the current working condition. The current total rank value is used for driving fine adjustment updating of the LoRA-d of the process model, so that self-adaptive adjustment of structure parameters is achieved under the double constraints of complexity driving and disturbance leading, self-adaptive adjustment of rank increasing according to needs and steady-state rank reducing is achieved under the complex working condition of the weaving process, and energy efficiency optimization and calculation power saving are achieved while analysis precision is guaranteed.

Description

Technical Field

[0001] This application relates to the field of weaving process optimization technology, and in particular to a variable rank fine-tuning method and apparatus for a process model of textile weaving process. Background Technology

[0002] Weaving process optimization and intelligent decision-making have been gaining momentum globally in recent years, gradually becoming a key research direction for improving the quality and efficiency of textile manufacturing. Early studies largely relied on mechanistic modeling methods, analyzing the impact of process parameters on weaving quality and energy efficiency from the perspectives of loom drive systems, warp and weft yarn kinematics, and transmission chain dynamics. While these models can effectively characterize the mapping relationship between process and equipment under single, steady-state conditions, their adaptability is significantly insufficient in variable weaving environments such as frequent product changes, multiple styles in small quantities, and complex weave structures. With the development of artificial intelligence technology, data-driven weaving process optimization has gradually become mainstream. Researchers are constructing time-series optimization models around multi-dimensional indicators such as process parameters, weaving quality, and production cycle time to characterize nonlinear and time-varying features. Algorithms such as LSTM, GRU, and Transformer have significantly improved the accuracy and robustness of process configuration in loom parameter setting, process window optimization, and quality analysis. Meanwhile, interpretable artificial intelligence (XAI) has been introduced into the weaving process modeling system to reveal the causal chain between key process variables (such as machine speed, tension, and shedding sequence) and weaving defects, breakage rate, and stability indicators. However, these models still suffer from "black box" problems and high computational overhead to varying degrees, and their generalization and transfer capabilities in complex multi-condition weaving scenarios are limited.

[0003] In recent years, Large Language Models (LLMs), leveraging their cross-task generalization and knowledge extraction capabilities, have been widely explored for intelligent manufacturing decision-making and process optimization. Zhong et al. pointed out that mainstream large models (such as LLaMA and GPT) perform excellently on text, code, and general time-series tasks, but their large parameter scale, high training costs, and significant performance degradation in highly vertical industrial process scenarios are significant. Wang et al. further demonstrated that even with the introduction of parameter efficient fine-tuning (PEFT) methods, such as LoRA or Adapter, which can reduce parameter tuning costs to some extent, their "static rank" mechanism struggles to meet the demands of highly nonlinear, strongly coupled, and time-varying operating conditions in weaving processes. Existing research largely focuses on intelligent control and process modeling for macro-level scenarios such as buildings, transportation, and energy networks, while research on dedicated vertical process large models that systematically characterize the "process-stability-quality / energy efficiency" coupling mechanism in weaving production sites remains largely unexplored. Furthermore, general-purpose large models generally lack a deep understanding of the specialized semantics of the weaving field, making it difficult to effectively analyze parameter expressions with process characteristics in the weaving context (such as "weft stop," "warp elasticity tension," and "tension recovery process inducements"). This leads to distortion or failure in knowledge transfer, feature interpretation, and process suggestion generation when implemented. Therefore, constructing a vertical model system for weaving process optimization that combines high domain adaptability, low computational cost, and an interpretable process logic chain has become a key breakthrough in promoting the shift of weaving processes from experience-driven to data-knowledge fusion-driven in the current context of intelligent manufacturing. Summary of the Invention

[0004] The purpose of this application is to overcome the problem that the existing "static rank" mechanism is difficult to meet the requirements of the working conditions of strong nonlinearity, strong coupling and time-varying characteristics in the weaving process, and to provide a variable rank fine-tuning method and device for the process model of textile weaving.

[0005] Firstly, a variable-rank fine-tuning method for a process model in textile weaving is provided, including:

[0006] Construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset;

[0007] The incremental complexity of the process and the incremental stability of production are calculated based on a multi-source heterogeneous weaving integrated data system.

[0008] The total rank of the process model under the current operating condition is calculated based on the initial rank of the LoRA module, the process complexity increment, and the production stability increment.

[0009] The LoRA-d of the process model is fine-tuned and updated using the current total rank value.

[0010] In some possible implementations, the process analysis dataset uses production shifts as sample units, including the number of looms, shift duration, total running time, total downtime, total output, workshop average efficiency, average operating speed, total weft downtime, total warp downtime, total number of weft downtimes, and total number of warp downtimes. The process information dataset uses fabric category as sample units, including standard process parameters and actual processing status parameters.

[0011] In some possible implementations, the standard process parameters include warp density, weft density, nominal fabric width, warp and weft shrinkage, warp and weft tightness, standard moisture regain of warp and weft yarns, loom speed, weft insertion length, and warp tension. The actual processing state process parameters include running, stopping time, output, efficiency, rotation speed, weft stop, warp stop time, and number of warp stops.

[0012] In some possible implementations, the calculation of the process complexity increment includes:

[0013] Calculate the nonlinear interaction energy term between process parameters:

[0014]

[0015] in, This refers to the nonlinear interaction energy term between process parameters. For cross-coupling coefficients, and For each process parameter, It is a saturation factor;

[0016] The chaos index for calculating the process state:

[0017]

[0018] Where cp is the chaos index of the process state. Let m be the probability of fluctuation of the k-th parameter in ascending order, and m be the size of the time window for the process parameters.

[0019] Calculate the dynamic process complexity index:

[0020]

[0021] in, Entropy-driven weighting coefficients;

[0022] Based on a dynamic process complexity index, a sigmoid-type response function is constructed to output a variable-rank adjustment:

[0023]

[0024] in, This is the complexity increment based on the input process parameters. To activate historical thresholds, It is a moving average. For maximum rank increase, k is the nonlinear gain coefficient.

[0025] In some possible implementations, the calculation of the production stability increment includes:

[0026] Calculate the volatility dispersion index based on downtime per unit output:

[0027]

[0028] Where Dis is a volatility dispersion index based on the downtime per unit output. The downtime per unit output of the equipment during the i-th shift; Standard deviation; This is the average of the unit output list;

[0029] Calculate the anomalous chaotic term:

[0030]

[0031] in, This is an abnormally chaotic term. The average downtime per unit output of equipment within a shift; The longest single stop time for the equipment in the i-th shift; For use as an indicator of whether an abnormal time threshold is triggered The indicator function, whose value is 0 or 1. The number of looms;

[0032] Construct a stability score function:

[0033]

[0034] in, This is a stability score function used to characterize the overall operational stability of the weaving system. and These are the weighting coefficients for each item;

[0035] Calculate the stability increment based on equipment production status:

[0036]

[0037] in, This is a stability increment based on equipment production status. To activate historical thresholds, It is a moving average. For maximum rank increase, k is the nonlinear gain coefficient.

[0038] In some possible implementations, the formula for calculating the total rank value of the process model under the current operating condition is:

[0039]

[0040] in, This represents the total rank value of the process model under the current operating conditions. This is the basic initial rank of the LoRA module.

[0041] In some possible implementations, the LoRA-d fine-tuning update of the process model is driven by the current total rank value, including: using the total rank value as the rank configuration of the LoRA-d of the process model, performing fine-tuning training updates of the process model, and dynamically guiding rank value adjustment to achieve synchronous adaptive optimization of the process model structure and process features.

[0042] Secondly, a variable-rank fine-tuning device for a process model in textile weaving is provided, comprising:

[0043] A construction module is used to construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset;

[0044] The first calculation module is used to calculate the increment of process complexity and the increment of production stability based on the multi-source heterogeneous weaving integrated data system.

[0045] The second calculation module is used to calculate the total rank value of the process model under the current operating condition based on the basic initial rank of the LoRA module, the process complexity increment, and the production stability increment.

[0046] The fine-tuning update module is used to drive the fine-tuning update of the LoRA-d of the process model with the current total rank value.

[0047] Thirdly, a computer-readable storage medium is provided that stores program code for execution by a device, the program code including steps for performing a method as described in any of the implementations of the first aspect above.

[0048] Fourthly, an electronic device is provided, the electronic device including a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any of the implementations of the first aspect above.

[0049] This application offers the following advantages: The proposed variable-rank LoRA can achieve adaptive adjustment of "increasing rank on demand and decreasing rank in steady state" under complex weaving conditions. While ensuring analytical accuracy, it achieves optimal energy efficiency and computational savings, realizing a balance between "high precision and low redundancy." It ensures continuous expression of process characteristics and dynamic adaptation to production efficiency, enabling the variable-rank process model to possess analytical accuracy, computational efficiency, and stability. It achieves high-fidelity modeling of operational behavior and process response in complex weaving scenarios, providing key technical support for intelligent control and adaptive energy efficiency optimization of looms. Attached Figure Description

[0050] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments of this application and their descriptions are used to explain this application and do not constitute an undue limitation of this application.

[0051] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0052] Figure 1 This is a flowchart of the variable rank fine-tuning method for the process model of textile weaving process according to Embodiment 1 of this application;

[0053] Figure 2 This is a flowchart of the variable rank calculation in the variable rank fine-tuning method of the process model for textile weaving process in Embodiment 1 of this application;

[0054] Figure 3(a) shows the effectiveness and computational complexity analysis of the variable rank change in Embodiment 1 of this application. Figure 1 ;

[0055] Figure 3(b) is an analysis of the effectiveness and computational complexity of variable rank transformation in Embodiment 1 of this application. Figure 2 ;

[0056] Figure 3(c) is a diagram analyzing the effectiveness and computational complexity of variable rank changes in Embodiment 1 of this application.

[0057] Figure 3(d) is an analysis of the effectiveness and computational complexity of the variable rank change in Embodiment 1 of this application. Figure 4 ;

[0058] Figure 3(e) is an analysis of the effectiveness and computational complexity of variable rank transformation in Embodiment 1 of this application. Figure 5 ;

[0059] Figure 4This is a graph showing the analysis results of production efficiency and number of downtimes of the variable-rank process model in Embodiment 1 of this application under ten typical production shifts;

[0060] Figure 5 This is a structural block diagram of the variable rank fine-tuning device for the process model of textile weaving process according to Embodiment 2 of this application;

[0061] Figure 6 This is a schematic diagram of the internal structure of the electronic device according to Embodiment 4 of this application.

[0062] Figure label:

[0063] 100. Construction Module; 200. First Calculation Module; 300. Second Calculation Module; 400. Fine-tuning and Update Module. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] Example 1

[0066] like Figure 1 As shown in Embodiment 1 of this application, a variable-rank fine-tuning method for a process model in textile weaving includes:

[0067] S100. Construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset.

[0068] In this embodiment, the process analysis dataset (855 records) uses "one 8-hour production shift (covering the collaborative production process of multiple looms in the workshop)" as the sample unit. The fields include: number of looms, shift duration; total running time, total downtime, total output, workshop average efficiency, average operating speed, total weft / warp stop time, and total number of weft / warp stops. The target output is the total workshop process (kWh).

[0069] In this embodiment, the process information dataset (610 process combinations) uses "1 fabric variety (standard process parameters + historical actual production records)" as the sample unit, including standard process parameters (such as warp density / weft density / nominal fabric width / warp and weft shrinkage / warp and weft tightness / standard moisture regain of warp and weft yarns / weaving machine speed / weft insertion length / warp tension, etc.) and actual processing status parameters (running / stopping time, output, efficiency, speed, weft stop / warp stop time and number); the evaluation output includes processing complexity (high / moderate / low) and production stability (good / moderate / poor).

[0070] S200 calculates the increment of process complexity and production stability based on a multi-source heterogeneous weaving integrated data system.

[0071] In loom process analysis, the diversity of weaving process parameters and their significant nonlinear coupling relationships make it difficult for the static structure of traditional models to adapt to real-time changes in dynamic operating conditions. To improve the model's ability to represent process responses under complex conditions, this embodiment proposes a process complexity increment calculation system that integrates multi-parameter interactive coupling identification and entropy-driven control mechanisms. This system dynamically guides rank adjustment during the fine-tuning of the process LLM (i.e., process model), achieving synchronous adaptive optimization of the model structure and process characteristics. The calculation of process complexity increment includes:

[0072] S211. Calculate the interaction energy term between process parameters: There are synergistic and conflicting effects between process parameters. The interaction energy term is characterized by the "cross-coupling coefficient", which is calculated by the Hilbert-Schmidt independence test. Tanh saturation is introduced to avoid gradient explosion caused by extreme parameter combinations. At the same time, the nonlinear driving force of key coupling structures (such as speed-tension, air pressure-yarn density combination) on complexity is explicitly characterized.

[0073] Specifically, there are significant synergistic and conflicting effects among the process parameters. Their nonlinear interactions determine the complexity of the energy transfer path in the weaving system, and the definition of their nonlinear interactive energy terms is as follows. for:

[0074]

[0075] in, The cross-coupling coefficient is calculated by the Hilbert-Schmidt independence test. and These are various process parameters (such as warp density, weft density, weaving machine speed, and yarn tension). The saturation factor is used. This model avoids the gradient explosion problem under extreme parameter combinations by using the tanh function and explicitly characterizes the nonlinear driving effect of key coupling structures (such as velocity-tension and air pressure-yarn density combinations) on process complexity. Therefore, the model can quantify the contribution of different parameters to system complexity.

[0076] S212. Calculate the chaos compensation term driven by permutation entropy: For implicit disturbances that cannot be directly observed, such as yarn jump, tension oscillation, and air pressure fluctuation, permutation entropy is used as an index of process state chaos. The permutation entropy is calculated based on the historical permutation probability distribution of the parameter time window and is used as a statistical compensation for unmodelable disturbances, so that the model can still capture the process change trend under high noise conditions.

[0077] Specifically, in actual production, disturbances such as yarn bounce, tension oscillations, and air pressure fluctuations are often not directly observable, but their frequency and amplitude significantly affect process fluctuations and analytical stability. To describe the chaotic characteristics of these hidden disturbances, permutation entropy is introduced as a chaos index cp for the process state:

[0078]

[0079] in, Let be the probability of fluctuations in the history of the k-th parameter arranged in ascending order (i.e., from smallest to largest), and m be the size of the time series window for the process parameters. This mechanism provides statistical compensation for "unmodelable perturbations," enabling the model to accurately capture process change trends even under high-noise conditions. Furthermore, a dynamic process complexity index is defined. for:

[0080]

[0081] in, Entropy-driven weighting coefficients.

[0082] S213. Calculation of Nonlinear Sigmoid-Type Variable Rank Adjustment: To achieve variable rank adjustment based on complexity exponent, a Sigmoid-type response function is constructed, whose output is a "complexity-driven variable rank adjustment." This function integrates key parameters such as activation history thresholds (e.g., historical 85th percentile), moving average, nonlinear gain coefficient k, and maximum rank increase. Through the nonlinear characteristics of the Sigmoid function, this mapping can maintain rank stability in the low-to-medium complexity range, while achieving rapid rank expansion in the high-complexity range. This design effectively avoids rank oscillations that may occur at complexity boundaries.

[0083] Specifically, based on the dynamic process complexity index Construct a sigmoid-type response function to output a variable-rank adjustment. :

[0084]

[0085] in, To activate historical thresholds (such as the historical 85th percentile). It is a moving average. The maximum rank that can be increased (e.g., 128) is given by k, a nonlinear gain coefficient used to control the growth rate. This function achieves a balance between sensitivity and stability through a sigmoid response structure: maintaining a stable rank value in the low-to-medium complexity range and achieving rapid expansion in the high-complexity range, thus avoiding oscillations in the LoRA rank in the boundary region.

[0086] The incremental calculation of production stability is systematically carried out from three aspects: equipment discrete fluctuation assessment, abnormal chaos identification, and stability scoring fusion modeling. It aims to achieve a comprehensive assessment of the loom's operational consistency, abnormal states, and global steady-state level through the construction and quantification of multi-dimensional disturbance features. The incremental calculation of production stability includes:

[0087] S221. Modeling the Fluctuation and Dispersion of Equipment Group Consistency: To evaluate the operational consistency and coordination of a group of looms across different shifts, a fluctuation and dispersion index based on downtime per unit output is introduced. This index measures the degree of operational dispersion of the group of equipment: a larger value indicates weaker coordination and decreased overall scheduling stability; a smaller value indicates better group coordination and energy efficiency.

[0088] Specifically, to evaluate the operational consistency and structural synergy of the loom group across different production shifts, this embodiment introduces a fluctuation dispersion index Dis based on the downtime per unit output, the calculation formula of which is as follows:

[0089]

[0090] in, The downtime per unit output of the equipment during the i-th shift; Standard deviation; This represents the average output per unit. The index Dis measures the dispersion of operating states and the level of structural entropy among a group of machines. A large Dis value indicates significant differences in machine operation, weak group coordination, and decreased overall system scheduling stability; conversely, a small Dis value indicates coordinated operation of the loom group, concentrated process distribution, and better energy efficiency.

[0091] S222. Modeling of Chaotic Terms for Extreme Equipment Anomalies: For sudden disturbances in local equipment (long-term downtime or frequent shutdowns), construct abnormal chaotic terms: use the average downtime per unit output of equipment within a shift and the longest single downtime of equipment within a shift, and use an indicator function to determine whether an abnormal time threshold is triggered. By combining the number of looms, the proportion and severity of "overtime shutdown" equipment can be identified, providing early warning signals for variable rank regulation.

[0092] Specifically, in the weaving process, sudden disturbances in local equipment (such as prolonged downtime or frequent shutdowns) often become the main source of system instability. To identify and quantify such extreme behaviors, this embodiment constructs an anomalous chaotic term. :

[0093]

[0094] in, The average downtime per unit output of equipment within a shift; The longest single stop time for the equipment in the i-th shift; For use as an indicator of whether an abnormal time threshold is triggered Indicator function (i.e., whether an abnormal time threshold is triggered) (e.g., 30 minutes), the indicator function has values ​​of 0 and 1. If the longest single stop time of the equipment in the i-th shift exceeds the triggering abnormal time threshold... ,but The value is 1 if it is 1, otherwise it is 0. This refers to the number of weaving machines. This indicator... Focusing on the detection of small-scale, high-intensity disturbances, this study constructs a micro-level chaotic compensation mechanism by identifying the proportion and severity of "overtime shutdown" equipment, providing early warning signals for fine-tuning the variable-rank adjustment of the process model.

[0095] S223, Stability Score Weighted Fusion Calculation: A stability score function is constructed based on the weighted fusion of multiple features. The higher the score, the more unstable and the stronger the disturbance. Furthermore, a nonlinear Sigmoid mapping is designed to make the model's disturbance response exhibit dynamic characteristics of high agility under unstable conditions and structural convergence under steady-state conditions.

[0096] Specifically, to integrate stability information from different dimensions, this embodiment constructs a stability scoring function based on the idea of ​​multi-feature weighted fusion. :

[0097]

[0098] in, This is a stability score function used to characterize the overall operational stability of the weaving system. and These are the weighting coefficients for each item.

[0099] Furthermore, to achieve a dynamic response mapping from stability score to fine-tuned rank value, this embodiment designs a nonlinear Sigmoid function:

[0100]

[0101] in, This is a stability increment based on equipment production status. To activate historical thresholds, It is a moving average. For maximum rank increase, k is the nonlinear gain coefficient. This mapping function ensures that the model exhibits dynamic behavior characteristics of "hysteresis start-up-fast response-smooth saturation" during the disturbance response process, enabling the model to have high agility under unstable conditions and maintain structural convergence under steady-state conditions.

[0102] S300. Calculate the total rank of the process model under the current operating condition based on the initial rank of the LoRA module, the process complexity increment, and the production stability increment.

[0103] In actual weaving production, the performance of the loom depends not only on the complexity of the process settings but also on the significant impact of production stability disturbances. To achieve adaptive enhancement of the process analysis model under multiple operating conditions and disturbance scenarios, this embodiment proposes a variable-rank LoRA fine-tuning mechanism for the process model oriented towards weaving process characteristics, so as to achieve adaptive adjustment of structural parameters under the dual constraints of complexity-driven and disturbance-dominated approaches.

[0104] Specifically, this embodiment constructs a variable-rank LoRA regulation framework focusing on the two factors of "process complexity – production stability", such as Figure 2 As shown, the process complexity increment describes the degree of nonlinear coupling of the loom under different fabric types, weave structures, tension control, and speed settings, reflecting the complexity of system energy transfer and control response. The production stability increment characterizes the dynamic disturbance characteristics in the production process, such as stopping frequency, downtime, and disturbance distribution, reflecting the energy loss and power fluctuation level of the loom in the non-steady-state stage. From a physical perspective, process complexity determines the steady-state process baseline of the system, while production stability dominates the transient deviation and dynamic loss of the process. The synergistic effect between the two constitutes the dynamic evolution mechanism of the weaving process and provides a theoretical basis for the model's balance between structural flexibility and learning ability. Accordingly, the variable rank calculation formula focusing on the coupling effect of weaving process complexity and production disturbance can be defined as follows:

[0105]

[0106] in, This represents the total rank value of the process model under the current operating conditions. The basic initial rank of the LoRA module, This is the complexity increment based on the input process parameters. This is a stability increment based on equipment production status.

[0107] S400, drives the fine-tuning update of LoRA-d of the process model with the current total rank value.

[0108] Specifically, a LoRA fine-tuning module is introduced into the process LLM (process model), and the basic initial rank of the LoRA module is set as the baseline term for variable rank calculation.

[0109] Calculate the variable rank total rank value for each training batch / training step: For the process parameters and equipment production status corresponding to the current training batch, execute the following steps in sequence: obtain the total rank value (basic initial rank + complexity increment + stability increment) under the current working condition by following steps S200-S300.

[0110] Drive LoRA-d fine-tuning update with the current total rank value: In this training step, the total rank value obtained in step S9 is used as the rank configuration of LoRA-d to perform fine-tuning training update of process LLM, dynamically guiding rank adjustment and realizing synchronous adaptive optimization of model structure and process features.

[0111] Based on the process LLM trained in the previous steps, the input fields described in Table 1 are used as inference inputs for the target working condition to be analyzed (such as the production record corresponding to a certain production shift or a certain fabric variety): including standard process parameters and actual processing status parameters, and / or shift-level workshop multi-loom association data fields, and finally the model's evaluation output results on the current weaving process situation are obtained.

[0112] To verify the effectiveness of the proposed loom process analysis and interpretable modeling method, this embodiment constructs a comprehensive weaving data system containing multi-source heterogeneous data:

[0113] Process information JSON dataset: As shown in Table 1, the process information dataset constructed in this embodiment focuses on the standard process configuration and actual processing status of a single fabric variety. It includes 13 core process parameters and corresponding operating characteristics. This subset contains 610 typical process combinations, providing solid data support for the multi-source fusion and interpretable analysis of subsequent models.

[0114] Table 1: Composition of Loom Technology and Technology Information Dataset

[0115]

[0116] Figures 3(a), 3(b), 3(c), 3(d), and 3(e) demonstrate the adaptability and robustness of the variable-rank mechanism from a process perspective. The upper part shows the rank changes and computational load response of the model under different training steps. When the tension is high, the rotation speed is fast, or the weave structure is complex, the rank value increases to enhance the ability to capture nonlinear features; when the working conditions stabilize (low stopping rate, reduced tension fluctuations), the rank value decreases to achieve low-power and efficient operation. Batch 1 shows the model's high sensitivity to complex disturbances, Batch 2 achieves a preliminary balance between complexity identification and process allocation, Batch 3 achieves optimal matching in three-dimensional indicators, and Batch 4 exhibits typical low-power steady-state characteristics. Overall results show that the variable-rank LoRA can achieve adaptive adjustment of "increasing rank on demand and decreasing rank in steady state" under complex working conditions in the weaving process, achieving optimal energy efficiency and computational savings while ensuring analytical accuracy.

[0117] This embodiment further conducts performance tests on the variable-rank fine-tuned process LLM (ProcessLLM) to evaluate its generalization ability and industrial adaptability in weaving process parameter analysis and operational behavior modeling. Figure 4 The analysis results of the variable-rank process model on production efficiency (left figure) and number of shutdowns (right figure) under ten typical production shifts are presented. Overall, the variable-rank model (LoRA-d) shows better consistency and robustness in both types of indicators. In terms of production efficiency analysis, the results of the LoRA-d model are in high agreement with the actual values, with the average deviation controlled within ±1%, significantly better than the fixed low-rank (LoRA-32) and medium-rank (LoRA-64) models. In contrast, although LoRA-128 has a stronger fitting ability, it is slightly overlearned under low-fluctuation conditions, resulting in slightly higher local errors. The variable-rank mechanism achieves a balance of "high precision and low redundancy" by adjusting the rank value in real time, actively increasing the rank under complex process conditions (such as high tension or high weaving speed samples) and automatically decreasing the rank under steady-state conditions, ensuring the continuous expression of process characteristics and dynamic adaptation of production efficiency. In the analysis of the number of shutdowns, the variable-rank model also shows a significant advantage. Fixed low-rank models generally overestimate the frequency of stoppages (with an average deviation of approximately 10%–20%), while high-rank models, although offering improved accuracy, suffer from higher computational costs and slight fluctuations in some shifts. In contrast, LoRA-d's analysis results are almost synchronized with actual trends, accurately capturing stoppage peaks in high-disturbance shifts (such as Cycle 4 and Cycle 9) and effectively suppressing overfitting errors in steady-state cycles (such as Cycle 5 and Cycle 8), exhibiting the lowest error volatility. Overall, variable-rank process LLM combines analytical accuracy, computational efficiency, and stability, achieving high-fidelity modeling of operational behavior and process response in complex weaving scenarios, providing key technical support for intelligent control and adaptive energy efficiency optimization of looms.

[0118] Example 2

[0119] like Figure 5 As shown, Embodiment 2 of this application relates to a variable-rank fine-tuning device for a process model in textile weaving, comprising:

[0120] The construction module 100 is used to construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset;

[0121] The first calculation module 300 is used to calculate the increment of process complexity and the increment of production stability based on the multi-source heterogeneous weaving integrated data system.

[0122] The second calculation module 300 is used to calculate the total rank value of the process model under the current operating condition based on the basic initial rank of the LoRA module, the process complexity increment, and the production stability increment.

[0123] Fine-tuning update module 400 is used to drive fine-tuning updates of LoRA-d of the process model with the current total rank value.

[0124] It should be noted that other specific implementations of the variable rank fine-tuning device for the process model of textile weaving process in this embodiment can be found in the specific implementations of the variable rank fine-tuning method for the process model of textile weaving process described above. To avoid redundancy, they will not be repeated here.

[0125] Example 3

[0126] This application relates to a computer-readable storage medium in embodiment 3, which stores program code for execution by a device, the program code including steps for performing the method as described in any implementation of embodiment 1 of this application;

[0127] The computer-readable storage medium may be a read-only memory (ROM), a static storage device, a dynamic storage device, or a random access memory (RAM); the computer-readable storage medium may store program code, and when the program stored in the computer-readable storage medium is executed by a processor, the processor is used to perform the steps of the method in any of the implementations of Embodiment 1 of this application.

[0128] Example 4

[0129] like Figure 6As shown, an electronic device according to Embodiment 4 of this application includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor. When the program or instructions are executed by the processor, they implement the method in any of the implementations in Embodiment 1 of this application.

[0130] The processor can be a general-purpose central processing unit (CPU), microprocessor, application-specific integrated circuit (ASIC), graphics processing unit (GPU), or one or more integrated circuits, used to execute related programs to implement the method in any of the implementations of Embodiment 1 of this application.

[0131] The processor can also be an integrated circuit electronic device with signal processing capabilities. In implementation, each step of the method in any of the implementations of Embodiment 1 of this application can be completed by the integrated logic circuitry in the processor's hardware or by software instructions.

[0132] The aforementioned processor can also be a general-purpose processor, digital signal processor, application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component. It can implement or execute the methods, steps, and logic block diagrams disclosed in the embodiments of this application. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the methods disclosed in the embodiments of this application can be directly embodied in the execution of a hardware decoding processor, or executed by a combination of hardware and software modules in the decoding processor. The software modules can be located in random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, or other mature storage media in the art. The storage medium is located in memory; the processor reads information from the memory and, in conjunction with its hardware, completes the functions required by the units included in the data processing apparatus of the embodiments of this application, or executes the methods in any implementation of Embodiment 1 of this application.

[0133] The above are merely preferred embodiments of this application; however, the scope of protection of this application is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in this application, based on the technical solution and its improved concept, should be covered within the scope of protection of this application.

Claims

1. A variable-rank fine-tuning method for a process model used in textile weaving, characterized in that, include: Construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset; The incremental complexity of the process and the incremental stability of production are calculated based on a multi-source heterogeneous weaving integrated data system. The total rank of the process model under the current operating condition is calculated based on the initial rank of the LoRA module, the process complexity increment, and the production stability increment. The LoRA-d of the process model is fine-tuned and updated using the current total rank value.

2. The variable-rank fine-tuning method for the process model of textile weaving as described in claim 1, characterized in that, The process analysis dataset uses production shifts as the sample unit and includes the number of looms, shift duration, total running time, total downtime, total output, workshop average efficiency, average operating speed, total weft downtime, total warp downtime, total weft downtime, and total warp downtime. The process information dataset uses fabric type as the sample unit and includes standard process parameters and actual processing status parameters.

3. The variable-rank fine-tuning method for the process model of textile weaving as described in claim 2, characterized in that, The standard process parameters include warp density, weft density, nominal fabric width, warp and weft shrinkage, warp and weft tightness, standard moisture regain of warp and weft yarns, loom speed, weft insertion length, and warp tension. The actual processing state process parameters include running, stopping time, output, efficiency, rotation speed, weft stop, warp stop time, and number of warp stops.

4. The variable-rank fine-tuning method for a process model used in textile weaving processes according to any one of claims 1-3, characterized in that, The calculation of the process complexity increment includes: Calculate the nonlinear interaction energy term between process parameters: in, This refers to the nonlinear interaction energy term between process parameters. For cross-coupling coefficients, and For each process parameter, It is a saturation factor; The chaos index for calculating the process state: Where cp is the chaos index of the process state. Let m be the probability of fluctuation of the k-th parameter in ascending order, and m be the size of the time window for the process parameters. Calculate the dynamic process complexity index : in, Entropy-driven weighting coefficients; Based on a dynamic process complexity index, a sigmoid-type response function is constructed to output a variable-rank adjustment: in, This is the complexity increment based on the input process parameters. To activate historical thresholds, It is a moving average. For maximum rank increase, k is the nonlinear gain coefficient.

5. The variable-rank fine-tuning method for a process model used in textile weaving as described in claim 4, characterized in that, The calculation of the production stability increment includes: Calculate the volatility dispersion index based on downtime per unit output: Where Dis is a volatility dispersion index based on the downtime per unit output. The downtime per unit output of the equipment during the i-th shift; Standard deviation; This is the average of the unit output list; Calculate the anomalous chaotic term: in, This is an abnormally chaotic term. The average downtime per unit output of equipment within a shift; The longest single stop time for the equipment in the i-th shift; For use as an indicator of whether an abnormal time threshold is triggered The indicator function, whose value is 0 or 1. The number of looms; Construct a stability score function: in, This is a stability score function used to characterize the overall operational stability of the weaving system. and These are the weighting coefficients for each item; Calculate the stability increment based on equipment production status: in, This is a stability increment based on equipment production status. To activate historical thresholds, It is a moving average. For maximum rank increase, k is the nonlinear gain coefficient.

6. The variable-rank fine-tuning method for the process model used in textile weaving as described in claim 5, wherein the formula for calculating the total rank value of the process model under the current operating condition is: in, This represents the total rank value of the process model under the current operating conditions. This is the basic initial rank of the LoRA module.

7. The variable-rank fine-tuning method for a process model in textile weaving as described in claim 1, driving the fine-tuning update of the LoRA-d of the process model with the current total rank value, includes: Using the total rank value as the rank configuration of the LoRA-d process model, fine-tuning training and updating of the process model are performed, dynamically guiding the rank value adjustment, so as to achieve synchronous adaptive optimization of the process model structure and process characteristics.

8. A variable-rank fine-tuning device for a process model in textile weaving, characterized in that, include: A construction module is used to construct a multi-source heterogeneous weaving integrated data system, wherein the multi-source heterogeneous weaving integrated data system includes a process analysis dataset and a process information dataset; The first calculation module is used to calculate the increment of process complexity and the increment of production stability based on the multi-source heterogeneous weaving integrated data system. The second calculation module is used to calculate the total rank value of the process model under the current operating condition based on the basic initial rank of the LoRA module, the process complexity increment, and the production stability increment. The fine-tuning update module is used to drive the fine-tuning update of the LoRA-d of the process model with the current total rank value.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program code for execution by the device, the program code including steps for performing the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes a processor, a memory, and a program or instructions stored in the memory and executable on the processor, wherein the program or instructions, when executed by the processor, implement the method as described in any one of claims 1-7.

Images