Machine learning-based denim prescoring process adaptive control method and system

Abstract

This invention discloses an adaptive control method and system for denim pre-shrinking process based on machine learning, belonging to the field of fabric pre-shrinking technology. The method includes: acquiring denim production process data and real-time images, and inputting them into a pre-established lightweight physical simulation network to obtain a virtual state of the fabric surface; inputting the virtual state of the fabric surface into an adaptive control network for optimization processing to obtain pre-shrinking process parameters; based on the pre-shrinking process parameters, executing a double-loop learning process to obtain sorting control commands for controlling the pre-shrinking machine; and driving the pre-shrinking machine to perform actions according to the sorting control commands, completing the adaptive control of the denim pre-shrinking process. This invention uses a lightweight physical simulation network as a differentiable forward simulator, which can accurately deduce the virtual moisture content, shrinkage rate, and smoothness distribution of the fabric surface, providing accurate virtual quality feedback for parameter optimization. Combined with the double-loop learning mechanism, it ensures the uniformity and high standards of pre-shrinking quality from both the source and the process.

Description

Technical Field

[0001] This invention relates to the field of fabric pre-shrinking technology, and in particular to an adaptive control method and system for denim pre-shrinking process based on machine learning. Background Technology

[0002] Denim pre-shrinking technology is a core finishing process in denim production. It applies heat, moisture, and force to the fabric during production to eliminate internal stress generated during weaving and dyeing, thus pre-shrinking the fabric dimensions. This directly determines key quality indicators such as dimensional stability, surface smoothness, and moisture content uniformity of the finished denim product. It is a crucial step in ensuring the precision of subsequent cutting and sewing processes and improving the quality of finished denim garments.

[0003] Most existing pre-shrinking machines use pre-set fixed parameter sets for batch production, or simply adjust single parameters such as temperature and pressure using PID control for localized feedback correction. This lack of comprehensive perception of the entire production process and multi-parameter collaborative optimization capabilities easily leads to unstable pre-shrinking effects, resulting in quality problems such as excessive fabric shrinkage, uneven moisture content distribution, and localized wrinkling. Furthermore, when the core properties of the incoming fabric fluctuate within or between batches, or when the pre-shrinking machine's status changes with production time, the fixed parameters may fail to match the changing operating conditions in real time. Summary of the Invention

[0004] The purpose of this invention is to provide an adaptive control method and system for denim pre-shrinking processes based on machine learning. A lightweight physical simulation network is used as a differentiable forward simulator, which can accurately deduce the virtual moisture content, shrinkage rate, and smoothness distribution of the fabric surface, providing precise virtual quality feedback for parameter optimization. Combined with a dual-loop learning mechanism, the uniformity and high standards of pre-shrinking quality are guaranteed from both the source and the process. This effectively solves the problems in existing technologies, which lack comprehensive perception of the entire production process and the ability to collaboratively optimize multiple parameters, easily leading to unstable pre-shrinking effects and quality issues such as excessive fabric shrinkage, uneven moisture content distribution, and localized wrinkling of the fabric surface.

[0005] To achieve the above objectives, in a first aspect, the present invention provides an adaptive control method for denim pre-shrinking process based on machine learning, the method comprising: Acquire denim production process data and real-time images; The production process data and real-time images are input into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The virtual state of the fabric is input into a pre-established adaptive control network for optimization processing to obtain the pre-shrinking process parameters; Based on the pre-shrinking process parameters, a dual-loop learning process is executed to obtain the sorting control command for controlling the pre-shrinking machine; According to the sorting control command, the pre-shrinking machine is driven to perform actions to complete the adaptive control of the denim pre-shrinking process.

[0006] Furthermore, the production process data includes measurable process parameters, fabric properties, and characteristic fluctuation parameters of the raw fabric. The measurable process parameters include the temperature and pressure data of the pre-shrinker's rubber blanket compression zone, the steam pressure of the pre-shrinker, the rubber blanket rotation speed, and the pre-shrinking time; The fabric properties include the fiber composition, thickness, initial moisture content, warp yarn density, weft yarn density, and fabric weave type of the denim. The characteristic fluctuation parameters of the raw fabric include yarn count fluctuation, density fluctuation, and weft skew fluctuation.

[0007] Furthermore, the real-time images are used to acquire information related to the surface condition of the fabric; The information related to the surface condition of the fabric includes the smoothness of the denim surface, surface defects, color and luster, texture characteristics, weft skew and arc degree, and surface gloss distribution.

[0008] Furthermore, the lightweight physics simulation network is a forward simulation model; The forward simulation model uses the production process data and real-time images to extrapolate the virtual state of the fabric.

[0009] Furthermore, the virtual fabric state includes virtual moisture content distribution, virtual shrinkage distribution, and virtual fabric smoothness distribution.

[0010] Furthermore, the dual-loop learning process includes inner-loop simulation optimization and outer-loop online adaptation; The inner ring simulation optimization includes performing virtual quality deduction based on the pre-shrinking process parameters, and determining whether the virtual state of the fabric obtained by the deduction meets the preset quality standard. If the target is not met, the pre-shrinking process parameters are iteratively optimized until the virtual quality simulation result meets the target. Then, the pre-shrinking process parameters optimized by the inner loop are used as sorting control commands. The outer ring online adaptation includes acquiring sparse quality feedback data during the actual production process of the pre-shrinking machine, and calculating the deviation between the sparse quality feedback data and the quality data predicted by the virtual state of the fabric. The lightweight physics simulation network and the adaptive control network are updated using the aforementioned deviations; The sparse quality feedback data refers to the actual fabric quality data obtained at a frequency lower than the production cycle and confirmed through offline testing.

[0011] Furthermore, the actions performed by the drive pre-shrinking machine include: The rubber blanket compression zone of the pre-shrinking machine is discretized into multiple intelligent particle units. These intelligent particle units make collaborative decisions based on local sensor data and through distributed algorithms, and dynamically adjust the pressure distribution field.

[0012] Furthermore, the adaptive control network is trained using a reinforcement learning paradigm under security constraints; Among them, the pre-shrinking process environment, parameter adjustment actions, and fabric quality reward function are defined as the three elements of reinforcement learning in the reinforcement learning paradigm.

[0013] Furthermore, the three elements of reinforcement learning include the following: The pre-shrinking process environment includes the overall working conditions during the denim pre-shrinking production process, including the pre-shrinking machine operating status, production process data, real-time image acquisition of the fabric surface status, and the virtual status of the fabric surface, forming a comprehensive scene. The parameter adjustment actions include the adjustment behavior of the adaptive control network on the initial pre-shrinking process parameters. The adjustment range is adapted to the pre-shrinking process environment and safety constraints. Specifically, it includes the adjustment of the temperature, pressure, steam pressure, rubber blanket rotation speed, and pre-shrinking time in the pre-shrinking machine's rubber blanket compression zone. The fabric quality reward function includes a function for quantifying the quality of parameter adjustment actions. Its input is the deviation between the actual fabric quality data and the preset quality standard. The smaller the deviation, the higher the reward value. When the deviation exceeds the safety constraint range, a negative reward is given, which guides the adaptive control network to learn the optimal parameter adjustment strategy.

[0014] Secondly, this invention provides an adaptive control system for denim pre-shrinking processes based on machine learning, the system comprising: The data acquisition module is used to acquire denim production process data and real-time images; The simulation module is used to input the production process data and real-time images into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The control module is used to input the virtual state of the fabric into a pre-established adaptive control network for optimization processing to obtain the pre-shrinking process parameters; The learning module, based on the pre-shrinking process parameters, performs a double-loop learning process to obtain the sorting control command for controlling the pre-shrinking machine; The execution module is used to drive the pre-shrinking machine to perform actions according to the sorting control instructions, so as to complete the adaptive control of the denim pre-shrinking process.

[0015] Furthermore, the simulation module includes: The extraction unit is used to extract key parameters from production process data and fabric surface condition information from real-time images, and to preprocess and normalize the key parameters and fabric surface condition information. The calculation unit is used to input the preprocessed and normalized key parameters and fabric surface state information into the lightweight physical simulation network, perform forward inference calculation, and obtain the forward inference calculation results. The generation unit is used to generate the virtual state of the fabric surface based on the forward extrapolation calculation results and transmit it synchronously to the control module.

[0016] Furthermore, the execution module includes: The mapping unit is used to receive the sorting control instructions transmitted by the learning module and map the sorting control instructions into pressure distribution field adjustment instructions for the pre-shrinker rubber blanket compression zone and action instructions for each actuator. The decision unit is used to control multiple intelligent particle units in the rubber blanket compression zone of the pre-shrinker, receive local sensing data from multiple intelligent particle units, and drive the distributed algorithm to run. The coordination unit is used to coordinate multiple intelligent particle units to make collaborative decisions, dynamically adjust the pressure distribution field, drive the linkage of various execution components of the pre-shrinking machine, and complete the pre-shrinking action.

[0017] Compared with the prior art, the embodiments of this application have at least the following beneficial effects: 1. This invention employs a lightweight physical simulation network as a differentiable forward simulator, which can accurately deduce the virtual moisture content, shrinkage rate, and flatness distribution of the fabric surface, providing precise virtual quality feedback for parameter optimization. Combined with a dual-loop learning mechanism, the inner loop simulation optimization ensures virtual quality meets standards through iterative calibration, while the outer loop online adaptation continuously narrows the gap between simulation and reality using sparse measured data. This ensures the uniformity and high standards of pre-shrinking quality from both the source and the process. Simultaneously, the compression zone of the pre-shrinking machine's rubber blanket is discretized into intelligent particle units and distributed collaboratively controlled, realizing dynamic adjustment of the pressure distribution field. This effectively solves the problems of inconsistent pre-shrinking and weft skew in the fabric surface, improving the quality of the final product.

[0018] 2. This invention, based on production process data and real-time images, combined with a lightweight physical simulation network designed specifically for edge computing, ensures the real-time performance of simulation and control. Based on an adaptive control network and employing a reinforcement learning paradigm under safety constraints, it can automatically and accurately adjust core process parameters such as temperature, pressure, and rotation speed within the safety constraints according to the deviation between the virtual state of the fabric and the target. This achieves dynamic matching of fluctuations in fabric properties and equipment status, overcoming the problems of fixed and poorly adaptable traditional process parameters.

[0019] 3. This invention deeply integrates quality, safety, and stability objectives into the reward function and iterative strategy of reinforcement learning, and adopts penalty terms, pruning functions, and small step size optimization strategies to strictly limit process parameters within a safe range at the algorithm level and ensure smooth parameter adjustment, thereby avoiding equipment damage and production accidents. At the same time, the system constructs a full-process fault tolerance and anomaly response mechanism covering sensors and execution units, which can quickly isolate and handle anomalies, minimize production interruptions and quality losses, and improve the continuity and stability of production. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments or examples of the present invention, the drawings used in the embodiments or examples will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained according to these drawings without creative effort.

[0021] Figure 1 This is a flowchart illustrating the adaptive control method for denim pre-shrinking process based on machine learning, as described in this invention. Figure 2 This is an architecture diagram of the adaptive control system for denim pre-shrinking process based on machine learning, as described in this invention. Figure 3 This is an architecture diagram of the execution module of the present invention; Figure 4 This is an architecture diagram of the simulation module of the present invention. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0024] Most existing pre-shrinking machines use a pre-set fixed parameter set for mass production, or simply adjust the temperature, pressure and other single parameters through PID control to achieve local feedback correction. They lack comprehensive perception of the entire production process and the ability to optimize multiple parameters in a coordinated manner, which can easily lead to unstable pre-shrinking effect and quality problems such as excessive fabric shrinkage, uneven moisture content distribution and localized wrinkling of the fabric surface.

[0025] For example, if the core properties of the same batch of greige fabric, such as yarn count and density, fluctuate naturally, the fabric's tightness will change accordingly. However, the pre-shrinking machine will still use the original fixed parameters for production. The matching PID control can only make small local corrections to a single temperature or pressure, and cannot coordinate the adjustment of key parameters such as the rubber blanket speed and pre-shrinking time. Ultimately, this will cause the fabric shrinkage rate to exceed the standard and fail to meet the qualification requirements.

[0026] To address the shortcomings of existing technologies, such as a lack of comprehensive perception and multi-parameter collaborative optimization capabilities across the entire production process, which can lead to unstable pre-shrinking effects and quality issues like excessive fabric shrinkage, uneven moisture content distribution, and localized wrinkling, this application provides a machine learning-based adaptive control method and system for denim pre-shrinking processes. The core technical features are: employing a lightweight physical simulation network as a differentiable forward simulator, capable of accurately deriving the virtual moisture content, shrinkage, and smoothness distribution of the fabric surface, providing precise virtual quality feedback for parameter optimization; combining a dual-loop learning mechanism, the inner loop simulation optimization ensures virtual quality compliance through iterative calibration, while the outer loop online adaptation continuously narrows the gap between simulation and reality using sparse measured data, thus guaranteeing the uniformity and high standards of pre-shrinking quality from both the source and process perspectives; simultaneously, discretizing the pre-shrinking machine's rubber blanket compression zone into intelligent particle units and performing distributed collaborative control enables dynamic adjustment of the pressure distribution field, effectively solving problems of inconsistent fabric pre-shrinking and weft skew, and improving the quality of the final product.

[0027] Example 1 See appendix Figure 1 This invention discloses an adaptive control method for denim pre-shrinking process based on machine learning, the method comprising: Acquire denim production process data and real-time images; Production process data and real-time images are input into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The virtual state of the fabric is input into a pre-established adaptive control network for optimization to obtain the pre-shrinking process parameters; Based on the pre-shrinking process parameters, a dual-loop learning process is executed to obtain the sorting control command for controlling the pre-shrinking machine; According to the sorting control instructions, the pre-shrinking machine is driven to perform actions to complete the adaptive control of the denim pre-shrinking process.

[0028] In this embodiment, production process data, including temperature, pressure, yarn count, density, weft skew, moisture content, and rotation speed, are acquired at a frequency of 10Hz; real-time image acquisition frequency is 5fps, synchronized with the production process data acquisition frequency, triggered by a PLC synchronization signal. The acquisition range and accuracy include, but are not limited to: temperature: 0-200℃; pressure: 0-1MPa; yarn count: 10-40 English count; density: 100-300 yarns / 10cm; weft skew: 0-5°; moisture content: 5-20%; rubber blanket rotation speed: 10-30r / min; pre-shrinking time: 0.5-5s. The acquired data is transmitted in real-time to an edge server, stored locally, and simultaneously backed up to a USB flash drive. When a sensor fails to collect data (no data received for 3 consecutive times), the PLC immediately issues an alarm signal (alarm light: red, flashing frequency 1Hz; alarm sound: buzzer, frequency 500Hz), and simultaneously activates the backup sensor (each core sensor is equipped with one backup, the backup sensor is connected in parallel with the main sensor, and automatically switches); if the industrial camera fails to collect images (no images for 5 consecutive frames), the edge server issues an alarm and simultaneously suspends the pre-shrinker feeding (the PLC controls the feeding roller to stop rotating). After the engineer checks the camera connection or lens obstruction, the acquisition is manually restarted.

[0029] Real-time image acquisition and processing involves installing two industrial cameras at the feed end of the pre-shrinking machine, using a stitched method to cover the full width of the fabric and avoid blind spots. The camera lenses are perpendicular to the fabric surface to ensure distortion-free images. The acquisition frequency matches the production cycle, acquiring one frame at regular intervals to ensure real-time performance while avoiding data redundancy. During acquisition, factors such as light reflection and dust obstruction must be avoided to ensure clear and discernible images. After acquisition, the images undergo simple processing to remove invalid information, facilitating subsequent surface condition extraction. First, the color image is converted to grayscale to simplify image information and reduce processing load. Then, filtering removes noise from the image, preventing noise caused by dust, light fluctuations, etc., from affecting judgment. Next, invalid areas at the image edges are cropped, retaining the valid fabric area. Finally, tilted images are corrected to ensure the fabric texture is parallel to the image coordinate axes, improving the accuracy of surface condition extraction. From the preprocessed images, key surface state information of the fabric is extracted, specifically including fabric smoothness, surface defects, fabric color, fabric texture features, weft skew and curvature, and surface gloss distribution. The extraction process utilizes visual recognition technology, focusing on anomalies in various states. Fabric defects are identified by recognizing abnormal color blocks and damaged areas in the image to determine the type and severity of defects; fabric smoothness is assessed by identifying undulating edges to determine whether the fabric is smooth. Specifically, the surface state information extracted from the collected production process data and real-time images undergoes preprocessing and fusion to remove invalid data and standardize data. For production process data, outliers and unstable data are addressed. Conventional outlier detection methods are used to filter out data exceeding reasonable ranges and replace them with the recent average of normal data to prevent outliers from affecting model input. Simultaneously, all data is normalized to adjust data of different magnitudes and units to the same range for easier identification and processing. For the surface state information extracted from the images, invalid surface state information feature data is filtered out. For example, state information that cannot be identified due to image blur is supplemented using valid data from adjacent frames. The preprocessed production process data and surface state information are fused together to form a complete feature set, which serves as the input to the subsequent lightweight physical simulation network. The fusion process assigns weights based on the reliability of various data types; the more stable and reliable the data, the greater the weight, ensuring that the fused data can comprehensively and accurately reflect the current state of the pre-shrinking process.

[0030] In some embodiments, production process data includes measurable process parameters, fabric properties, and characteristic fluctuation parameters of the raw fabric. Measurable process parameters include temperature and pressure data in the pre-shrinker's rubber blanket compression zone, steam pressure in the pre-shrinker, rubber blanket rotation speed, and pre-shrinking time; Fabric properties include the fiber composition, thickness, initial moisture content, warp yarn tightness, weft yarn tightness, and fabric weave type of denim; The characteristic fluctuation parameters of the raw fabric include yarn count fluctuation, density fluctuation, and weft skew fluctuation.

[0031] In some embodiments, real-time images are used to acquire information related to the surface condition of the fabric. Information related to the surface condition of fabrics includes the smoothness of the denim surface, surface defects, color and luster, texture characteristics, weft skew and curvature, and surface gloss distribution.

[0032] In some embodiments, the lightweight physics simulation network is a forward simulation model; The forward simulation model uses production process data and real-time images to extrapolate the virtual state of the fabric.

[0033] In some embodiments, the virtual state of the fabric surface includes virtual moisture content distribution, virtual shrinkage distribution, and virtual fabric surface smoothness distribution.

[0034] In some embodiments, the dual-loop learning process includes inner loop simulation optimization and outer loop online adaptation; The inner ring simulation optimization includes virtual quality deduction based on pre-shrinking process parameters, and judging whether the virtual state of the fabric obtained by the deduction meets the preset quality standard. If the target is not met, the pre-shrinking process parameters are iteratively optimized until the virtual quality simulation results meet the target. Then, the pre-shrinking process parameters optimized by the inner loop are used as sorting control commands. The outer loop online adaptation includes acquiring sparse quality feedback data during the actual production process of the pre-shrinking machine, and calculating the deviation between the sparse quality feedback data and the quality data predicted by the virtual state of the fabric. The lightweight physics simulation network and adaptive control network are updated using bias; Among them, sparse quality feedback data refers to actual fabric quality data obtained at a frequency lower than the production cycle and confirmed through offline testing.

[0035] In this embodiment, production process data and real-time images are input into a lightweight physical simulation network to obtain the virtual state of the fabric. The lightweight physical simulation network is a forward simulation model. Based on the input production process data and real-time image information, the forward simulation model simulates the physical process of the pre-shrinking process and extrapolates forward to obtain the virtual state of the fabric. The focus is on simulating the physical changes during the pre-shrinking process, providing a reference for subsequent parameter optimization. Specific details are as follows: The forward simulation model is lightweight and adaptable to the computing power of edge servers, ensuring that the simulation speed is synchronized with the production cycle and will not affect the continuity of pre-shrinking production. The core principle of the forward simulation model is based on the physical mechanism of the pre-shrinking process, simulating the influence of process parameters such as temperature, pressure, and steam on the fabric surface state, as well as the effect of fluctuations in the fabric's own properties and characteristics on the pre-shrinking effect, so as to achieve accurate simulation of the fabric surface state.

[0036] Furthermore, the input consists of fused feature data (various parameters during the production process and fabric surface condition information); the output is a virtual fabric state, specifically including three categories: virtual moisture content distribution, virtual shrinkage distribution, and virtual fabric smoothness distribution, each corresponding one-to-one with the actual pre-shrinked fabric state, accurately reflecting the possible state of the fabric under the current process parameters. After inputting the fused feature data (various parameters during the production process and fabric surface condition information) into a trained lightweight physical simulation network, the forward simulation model simulates the physical environment of the pre-shrinking machine's rubber blanket compression zone, reproducing the effects of temperature and pressure parameters on the fabric, as well as the physical phenomena of moisture content changes, dimensional shrinkage, and smoothness changes during the pre-shrinking process. The simulation process requires no manual intervention; the forward simulation model completes it automatically, with simulation time controlled within a reasonable range to ensure real-time performance. After simulation, the specific data of the virtual fabric state are output, clearly defining the virtual moisture content, virtual shrinkage, and virtual smoothness of each region, providing clear reference targets for subsequent parameter optimization. After the simulation is completed, the system automatically verifies the rationality of the virtual state data. If the data exceeds the preset reasonable range, the simulation is judged to have failed. The backup forward simulation model is immediately activated for re-simulation, and an alarm signal is issued to remind engineers to check the input data or the simulation model itself. If the simulation fails multiple times in a row, the pre-shrinking machine is suspended to avoid ineffective production. Among them, the three categories of parameters of the fabric virtual state have clear meanings: virtual moisture content distribution reflects the moisture content of different areas of the fabric after pre-shrinking, which is key to ensuring that the moisture content does not cause the fabric to yellow or deform; virtual shrinkage distribution reflects the dimensional shrinkage of different areas of the fabric after pre-shrinking, which ensures that the shrinkage rate is within a reasonable range and is the core of ensuring the dimensional stability of the fabric; virtual fabric flatness distribution reflects the flatness of the fabric after pre-shrinking, which directly determines the appearance quality of the fabric.

[0037] The lightweight physical simulation network of this invention is a differentiable forward physical simulator. By embedding the thermal-humidity-mechanical coupling physical mechanism of the pre-shrinking process, it achieves accurate deduction of the fabric state while maintaining lightweight computing characteristics to adapt to the real-time requirements of edge servers. The calculation process includes the following: Feature extraction formula: hyv = ReLU(W1z + b1); In the formula, hyv represents the hidden state output by the feature extraction layer, which is a high-dimensional feature representation of the input features after linear transformation and nonlinear activation, and serves as the input to the physical simulation layer; ReLU is the rectified linear unit activation function, used to introduce nonlinearity, enhance the network's expressive power, and avoid gradient vanishing; W1 is the weight matrix of the feature extraction layer, used to perform linear transformation on the input features; z is the input fused feature vector, including the preprocessed production process data x. proc(Temperature, pressure, yarn count, density, etc.), fabric surface state features extracted from images x img (Smoothness, defects, weft skew, etc.) and pre-shrinking process parameters a; b1 is the bias vector of the feature extraction layer, used to adjust the offset of the linear transformation; This feature extraction formula is a neural network feature extraction module that maps z to h suitable for physical simulation through linear transformation and nonlinear activation, providing a high-dimensional feature foundation for subsequent embedding of physical evolution formulas; Physical evolution formula: z=[x proc ,x img ,a]=W0(h+Δt·g(h,a;W p ))+b0; In the formula, z is the input fused feature vector; x proc This refers to preprocessed production process data; x img The image extracts the fabric surface condition features; 'a' represents the pre-shrinking process parameters; g(h,a;W) represents the pre-shrinking process parameters. p This is a lightweight physical layer that embeds a learnable module based on the thermo-humidity-mechanical coupling mechanism. Based on the current hidden state h and process parameter a, it outputs the rate of change of fabric surface state (moisture content, shrinkage rate, smoothness), approximating the physical evolution of the pre-shrinking process. h + Δt·g(h,a;W p ) is the state update term, where h+Δt·g(h,a;W p The formula uses Euler's integral method to discretize and simulate the pre-shrinking process, where: h represents the current state of the fabric; Δt·g(·) represents the change in the fabric state within the virtual time step Δt; the sum of the two represents "current state + change in state within the time step", which yields the new state h after the pre-shrinking process evolves by one virtual time step. next To realize the time evolution of physical processes. W0(h+Δt·g(h,a;W p ))+b0 is a state mapping, which maps the physical evolution to the new state h. next After a linear transformation, it is mapped to the updated feature vector z, or directly mapped to the final output virtual state of the cloth surface; Where, h+Δt·g(h,a;W p The meaning of ) is element-wise addition of vectors; the two are added element-wise to obtain the new state vector h after the pre-shrinking process has evolved for one virtual time step. next This is the discretization update of the Euler integral method, corresponding to the differential equation. The numerical solution, simulating the evolution of the fabric state under the thermal-humid-mechanical coupling during the pre-shrinking process, is the core operation of physical simulation; in, denoted as h, the first derivative of the fabric state with respect to time t, representing the instantaneous rate of change of the fabric state during pre-shrinking (e.g., rate of change of moisture content, rate of change of shrinkage, rate of change of smoothness); h is the hidden feature representation of the current fabric state, extracted from the preceding network layers, including key state information such as the current fabric moisture content, shrinkage, and smoothness; a is the pre-shrinking process parameter vector (e.g., temperature, pressure, rubber blanket rotation speed, pre-shrinking time, etc.), which represents the external conditions driving the pre-shrinking physical process; W p The lightweight physical layer is a learnable parameter set used to fit the approximate physical laws of the thermal-humidity-mechanical coupling during the pre-shrinking process; g(·) is the lightweight physical layer function, which takes the current state h and process parameter a as input and outputs the rate of change of state. The physical mechanisms of the approximate pre-shrinking process (such as heat transfer, moisture migration, and the mechanical response of fiber shrinkage). W0(h+Δt·g(h,a;W p The meaning of ))+b0 is the element-wise addition of vectors. The two are added element by element to complete the offset adjustment of the linear layer. This is the standard operation of the linear layer of the neural network. By shifting the state after physical evolution through the bias b0, the linear symmetry is broken, the network's ability to map the state of the fabric is improved, and finally the updated feature vector z (including production process data, image features, and process parameters) is output. Finally, the output is defined as yv = [moisture content distribution, shrinkage distribution, smoothness distribution]. T ; The virtual fabric state output by the lightweight physical simulation network includes the distribution of moisture content, shrinkage rate and flatness in different areas of the pre-shrinked fabric, which corresponds one-to-one with the actual fabric quality after pre-shrinking, providing accurate virtual quality feedback for the subsequent adaptive control network.

[0038] The overall technical process is as follows: input the fused production process data, image features and initial process parameters; generate hidden states through the feature extraction layer to prepare for physical simulation; embed a lightweight physical layer and simulate the thermal-humid-mechanical coupling evolution of the pre-shrinking process through Euler integral; map the evolved state to the fabric virtual state and output the distribution of moisture content, shrinkage rate and flatness to provide a basis for optimizing the pre-shrinking process parameters.

[0039] In this embodiment, the virtual state of the fabric is input into a pre-established adaptive control network for optimization processing to obtain pre-shrinking process parameters, specifically including: The adaptive control network is trained using a reinforcement learning paradigm under safety constraints. Based on the obtained virtual state of the fabric, it optimizes the pre-shrinking process parameters to ensure that the pre-shrinked fabric quality meets the standards, while avoiding parameters exceeding the process safety range. This is based on the synergistic effect of the three elements of reinforcement learning and the reasonable adjustment of parameters. Specific details are as follows: Network training clearly defines the three key elements of reinforcement learning: the pre-shrinking process environment, parameter adjustment actions, and fabric quality reward function. These three elements work together to guide the network to learn the optimal parameter adjustment strategy, ensuring that the optimized process parameters not only meet safety constraints but also achieve the best pre-shrinking effect.

[0040] The pre-shrinking process environment is a comprehensive scenario encompassing the overall operating conditions during the pre-shrinking production process. It includes four main categories of information: the pre-shrinking machine's operating status, production process data, real-time image acquisition of the fabric surface condition, and the fabric's virtual state. The environmental status is updated in real-time and synchronized with the production process. Based on the current environmental status, the network determines the rationality of the current process parameters and generates corresponding parameter adjustment actions. When the pre-shrinking machine malfunctions, data fluctuations exceed the range, or the virtual state fails to meet standards, the environmental status is promptly fed back, guiding the network to adjust its action strategy.

[0041] The parameter adjustment action refers to the network's adjustment of the initial pre-shrinking process parameters. The adjustment range strictly adapts to the pre-shrinking process environment and safety constraints to avoid the adjusted parameters exceeding the allowable range, which could lead to equipment damage or substandard fabric quality. The specific parameters to be adjusted include five categories: temperature and pressure in the pre-shrinking machine's rubber blanket compression zone, steam pressure, rubber blanket rotation speed, and pre-shrinking time, each corresponding to a pre-shrinking process parameter. During adjustment, the corresponding parameters are adjusted specifically based on the deviation of the fabric's virtual state. For example, if the virtual shrinkage rate exceeds the reasonable range, the pressure and pre-shrinking time are adjusted as a key focus to ensure precise and effective adjustments.

[0042] The fabric quality reward function is used to quantify the merits of parameter adjustment actions and guide the network to learn the optimal parameter adjustment strategy. The core evaluation criterion is the deviation between the actual fabric quality data and the preset quality standard. The smaller the deviation, the better the adjustment action, and the higher the reward value. If the deviation exceeds the safety constraints, it indicates a violation of the adjustment action, which may lead to serious fabric quality defects or equipment damage. In this case, a negative reward is given to guide the network to forget such violations. Simultaneously, when serious defects appear on the fabric after adjustment or the equipment malfunctions, the minimum negative reward is directly given, and a shutdown command is triggered to prevent further losses.

[0043] In some embodiments, driving the pre-shrinking machine to perform actions includes: The rubber blanket compression zone of the pre-shrinking machine is discretized into multiple intelligent particle units. These intelligent particle units make collaborative decisions based on local sensor data and through distributed algorithms, and dynamically adjust the pressure distribution field.

[0044] In some embodiments, the adaptive control network is trained using a reinforcement learning paradigm under security constraints; Among them, the pre-shrinking process environment, parameter adjustment actions, and fabric quality reward function are defined as the three elements of reinforcement learning in the reinforcement learning paradigm.

[0045] In some embodiments, the three elements of reinforcement learning include the following: The pre-shrinking process environment includes the overall working conditions during the denim pre-shrinking production process, including the pre-shrinking machine's operating status, production process data, real-time image acquisition of the fabric surface condition, and the virtual state of the fabric surface, forming a comprehensive scene. The parameter adjustment actions include the adjustment behavior of the adaptive control network on the initial pre-shrinking process parameters. The adjustment range is adapted to the pre-shrinking process environment and safety constraints. Specifically, it includes the adjustment of the temperature, pressure, steam pressure, rubber blanket speed, and pre-shrinking time in the pre-shrinking machine's rubber blanket compression zone. The fabric quality reward function includes a function to quantify the quality of parameter adjustment actions. Its input is the deviation between the actual fabric quality data and the preset quality standard. The smaller the deviation, the higher the reward value. When the deviation exceeds the safety constraint range, a negative reward is given, which guides the adaptive control network to learn the optimal parameter adjustment strategy.

[0046] Furthermore, based on the deviation between the virtual fabric state and the preset quality standard, the process parameters are continuously adjusted through a reinforcement learning network until the optimal parameter combination is obtained. The specific process is as follows: The collected fabric attributes are used as a benchmark, combined with the historical best parameters of the production line, to set initial pre-shrinking process parameters, which are then input into the adaptive control network. The adaptive control network receives the virtual fabric state, compares it with the preset quality standard, determines whether various virtual state parameters (virtual moisture content, virtual shrinkage rate, virtual smoothness) meet the standards, calculates the deviation between the virtual state and the quality standard, and identifies the deficiencies of the current process parameters. Based on the deviation, the network generates corresponding parameter adjustment actions, specifically adjusting the five major process parameters, with the adjustment range strictly controlled within the safety constraints. If the parameter adjustment exceeds the safety range, the system automatically corrects it to the constraint boundary value to ensure parameter safety. The adjusted process parameters are fed back to the lightweight physical simulation network, which re-deduces the new virtual fabric state and compares it again with the preset quality standard to determine whether it meets the standard. If it does not meet the standard, the above deviation judgment and parameter adjustment steps are repeated until the virtual state meets the standard, resulting in the optimized pre-shrinking process parameters. After the optimized pre-shrinking process parameters are verified by the system and confirmed to meet safety constraints and virtual state standards, they are transmitted to the subsequent dual-loop learning process to generate sorting control instructions.

[0047] The adaptive regulation network employs security-constrained reinforcement learning, which includes the policy output formula: π(s t ;θ A ):s t →Δ at ; π(s t ;θ A) represents the policy network; π(·) represents the policy function of the adaptive control network, which is the core mapping of reinforcement learning and represents the decision rule from state to action; s t Let θ be the state vector at time t, including comprehensive working condition information such as the pre-shrinking machine's operating status, production process data, fabric surface condition, and virtual fabric state; A The set of learnable parameters for the policy network is continuously updated during training to optimize decision-making; Δ at This represents the adjustment amount of the process parameters at time t, corresponding to the adjustment values ​​of parameters such as temperature, pressure, steam pressure, rubber blanket rotation speed, and pre-shrinking time. The meaning of the policy output formula is that the policy network, based on the current production state s, t Output the corresponding process parameter adjustment amount Δ at This enables precise mapping from operating conditions to control actions; Furthermore, based on the training objective formula, the training objective is to maximize the long-term reward under security constraints, and its training objective formula is: maxθ A E[∑ t γ t R(s t ,a t )],sta min ≤a t ≤a max ; maxθ A To maximize the objective, the policy network parameters θ are optimized. A To achieve this; E[·] is the expectation operator, which takes the expectation of the long-term reward for all possible production trajectories, reflecting the average performance; ∑ t γ t R(s t ,a t R(s) represents the cumulative reward with discounts, where γ is the discount factor (0 < γ < 1), used to balance the weights of current and future rewards; t ,a t ) is the immediate reward at time t; sta min ≤a t ≤a max For safety constraints, a t These are the pre-shrinking process parameters; a t =a t -1+Δ at The adjusted process parameters must be within the safety upper and lower limits a. min and a max Between these parameters, avoid exceeding the limits, which could lead to equipment damage or substandard quality; a min Here, represents the lower limit of the adjustable parameters in the pre-shrinking process, and represents the minimum allowable value of the process parameter; a maxThis represents the safe upper limit of each adjustable parameter in the pre-shrinking process, and the maximum allowable value of the process parameter. The meaning of the training objective formula is that, under the safety constraints of process parameters, the training objective is to maximize the long-term cumulative reward and guide the policy network to learn the optimal parameter adjustment strategy that is both safe and efficient. Furthermore, based on the reward function formula, the objectives of quality, safety, and stability are integrated into a unified reward signal through weighted penalty terms. This guides the policy network to maximize pre-shrinking quality while ensuring safety and stability. The reward function formula is as follows: R=-λ1‖e q || 2 -λ2∑[ReLU 2 (a t -a max )+ReLU 2 (a min -a t )]-λ3‖a t -a t-1 || 2 ; In the formula, R is the instantaneous reward function, quantifying the state s at time t. t and action a t The quality of the parameters is the core signal for strategy optimization; λ1, λ2, and λ3 are non-negative weighting coefficients used to balance the importance of the three optimization objectives: quality, safety, and stability; e q The fabric quality deviation vector represents the deviation between the virtual fabric state (moisture content, shrinkage, smoothness) and the preset quality standard target value; ReLU(·) is the rectified linear unit; a t-1 For the process parameters at time; ||a t -a t-1 || 2 The squared L2 norm of the parameter variation represents the stability of parameter adjustment, avoiding frequent and large adjustments; a min Here, represents the lower limit of the adjustable parameters in the pre-shrinking process, and represents the minimum allowable value of the process parameter; a max This represents the safe upper limit of each adjustable parameter in the pre-shrinking process, and the maximum allowable value of the process parameter. Where -λ1‖e q || 2 To penalize deviations in fabric quality from the target, the target is to improve pre-shrinkage quality; -λ2∑[ReLU 2 (a t -a max )+ReLU 2 (a min -a t To penalize process parameters that exceed safe ranges and ensure production safety; -λ3‖a t -at-1 || 2 To penalize frequent and significant adjustments to parameters and ensure stable production processes; the sum of these three items transforms multi-objective optimization into a single-objective reward function, enabling the policy network to simultaneously consider quality, safety, and stability, thus achieving adaptive control of the pre-shrinking process; Specifically, the adaptive control network employs the following steps in safety-constrained reinforcement learning: Policy Network π(s) t ;θ A Based on the current operating condition (st), the process parameter adjustment amount Δ is output. at ; The training objective is to ensure safety constraint a min ≤a t ≤a max Under the following conditions, maximize the long-term cumulative reward E[∑ t γ t R(s t ,a t )]; The reward function R integrates three objectives: quality, safety, and stability, guiding the network to learn the optimal parameter adjustment strategy to ensure that the pre-shrinking process is both compliant and safe and stable.

[0048] Furthermore, based on the pre-shrinking process parameters, a dual-loop learning process is executed to obtain the sorting control commands for controlling the pre-shrinking machine. The dual-loop learning process is the core link to ensure the accuracy and stability of the pre-shrinking process parameters. It consists of two parts: inner-loop simulation optimization and outer-loop online adaptation. The two work together to ensure that the parameters meet the standards in the virtual scenario and can adapt to the fluctuations in actual production, ensuring the feasibility and accuracy of the commands. Specifically, it includes the following: The inner-loop simulation optimization further calibrates the obtained pre-shrinking process parameters to ensure that the parameters can achieve the optimal pre-shrinking effect in the virtual scenario, avoiding unqualified actual production due to parameter deviations. The specific process is as follows: The optimized pre-shrinking process parameters are input again into the lightweight physical simulation network to perform forward inference and obtain the virtual state of the fabric after inference, which serves as virtual quality data. This simulates the subtle changes in actual production and ensures that the virtual quality data can accurately reflect the actual pre-shrinking effect.

[0049] Furthermore, preset quality standard thresholds are used to determine whether the virtual quality data meets the standards. The core indicators for this determination include virtual moisture content distribution, virtual shrinkage distribution, and virtual fabric smoothness distribution. The specific determination rules are as follows: The moisture content in each region of the virtual moisture content distribution is within a reasonable range, and the difference in moisture content between regions is small; The shrinkage rates in each region of the virtual shrinkage distribution are within acceptable limits, and the fluctuations in shrinkage rates are slight. The flatness of each area in the virtual fabric surface distribution meets the standard, with no obvious undulations.

[0050] Only when all three indicators mentioned above are met can the virtual quality simulation result be considered to meet the standard. If any indicator is not met, it is considered a failure and enters the iterative optimization phase. If the virtual quality simulation result fails to meet the standard during the iterative optimization phase, the substandard virtual quality data is fed back to the adaptive control network. The network, based on the deviation, generates parameter adjustment actions again, with the adjustment amount being half of the previous adjustment, to avoid frequent parameter fluctuations and iterative oscillations. The new pre-shrinking process parameters are obtained after adjustment and are input into the lightweight physical simulation network for simulation again. This process is repeated until the virtual quality simulation result meets the standard. If the standard is not met even after reaching the maximum number of iterations, an alarm signal is triggered, the pre-shrinking machine operation is suspended, and engineers investigate or input data issues. If the virtual quality meets the standard during the iteration process, the iteration is immediately stopped, and the current optimized pre-shrinking process parameters are encapsulated into sorting control instructions. These sorting control instructions include target values ​​for various process parameters, control types (normal control or abnormal shutdown), etc., in a standardized format for easy identification and execution by subsequent execution modules.

[0051] Specifically, the inner loop simulation optimization uses an iterative update formula to optimize the pre-shrinking process parameter a. t (e.g., temperature, pressure), enabling the lightweight physics simulation network to predict the virtual state y of the fabric surface. v,t Approaching the target y target : Its iterative update formula is: a t+1 =Clip(a t +η·π(s t ;θ A ),a min ,a max ); In the formula, a t η is the pre-shrinking process parameter vector at the t-th iteration (including adjustable parameters such as temperature, pressure, steam pressure, rubber blanket rotation speed, and pre-shrinking time); η is the learning rate (step size coefficient), used to control the magnitude of each parameter update to avoid excessively large step sizes leading to iteration oscillations or non-convergence; π(s t ;θ A ) is an adaptive control network (policy network) in state s t The output process parameter adjustment amount, θ A For the learnable parameters of the policy network; a min Here, represents the lower limit of the adjustable parameters in the pre-shrinking process, and represents the minimum allowable value of the process parameter; a max , where is the safe upper limit for each adjustable parameter in the pre-shrinking process, and is the maximum allowed value for the process parameter; Clip is the clipping function that limits the updated parameters to the safe range [amin ,a max [Inside, to avoid equipment damage or quality abnormalities; a] t+1 The pre-shrinkage process parameter vector for the (t+1)th iteration is the parameter after strategy adjustment and safe trimming; Wherein, η·π(s) t ;θ A (a) is an element-wise multiplication of a scalar and a vector, with the learning rate used to scale the values, avoiding excessively large step sizes that could cause iterative oscillations or non-convergence, thus making parameter updates more stable; t +η·π(s t ;θ A This is a vector element-wise addition method. Based on the current process parameters, it updates the parameters according to the direction and step size given by the strategy network, so that the parameters are adjusted towards the optimization goal (approaching the target fabric state). The iterative update formula updates the current process parameters based on the adjustment direction given by the policy network, and trims them to a safe range to obtain the process parameters for the next round of iteration. Furthermore, the formula for the iterative process parameters is: s t =[x t ,y v,t ,t,a t ],y v,t =F(x t ,a t ;θ F ); In the formula, s t Let x be the system state vector at the t-th iteration, which serves as the input to the policy network π(·), including the core information of the current operating condition; t y is the fused feature vector at the t-th iteration, including production process data (temperature, pressure, yarn count, etc.) and fabric surface conditions extracted from the image (smoothness, defects, etc.); v,t F(x) represents the virtual fabric state (moisture content distribution, shrinkage distribution, and smoothness distribution) predicted by the lightweight physical simulation network at iteration t; t is the current iteration number used to track the iteration progress; t ,a t ;θ F ) is the forward simulation function of the lightweight physics simulation network, with the current feature x as input. t and process parameter a t Output the virtual state y of the cloth surface v,t θ F These are the learnable parameters of the physical simulation network; Among them, s t =[x t ,y v,t ,t,a t[] is vector concatenation, which combines key information from different dimensions into a complete state vector s. t As the input to the adaptive control network, it fully reflects the current iteration's operating conditions and virtual quality state; y v,t =F(x t ,a t ;θ F () is the separator for function parameters. Function F uses x as the separator. t a t As input, in parameter θ F The forward inference is performed to output the virtual state y of the cloth surface. v,t This refers to the virtual quality prediction after pre-shrinking under the current pre-shrinking process parameters; based on the lightweight physical simulation network F, and according to the current production characteristics x t and process parameter a t Predict the virtual state y of the fabric v,t； The production characteristics, virtual state, iteration count, and process parameters are concatenated into a state vector s. t This provides complete operational feedback for subsequent parameter adjustments of the strategy network; The iterative process parameter formula constructs the virtual state of the fabric surface, which is predicted by the physical simulation network based on the current parameters and operating conditions, providing feedback for strategy adjustment. Furthermore, the iteration termination condition is: when ||y|| v,t -y target ||<τ in When, output a t To optimize parameter a ; In the formula, y v,t For the t-th iteration, the virtual state of the fabric surface (moisture content distribution, shrinkage distribution, and smoothness distribution) predicted by the lightweight physical simulation network; y target The target value for the fabric surface condition (preset quality standards, such as target moisture content, shrinkage, and smoothness); τ in The inner loop has an allowable deviation threshold. When the deviation is less than this threshold, the virtual state is considered to have met the standard, and the iteration terminates. The optimal process parameters after optimization, i.e., 'a' at the end of the iteration. t ; Among them, y v,t -y target The vector is subtracted element by element. The difference vector is obtained by subtracting the two elements one by one, which represents the difference between the virtual state and the target state in each quality index. The iteration termination condition means that when the deviation between the fabric virtual state predicted by the lightweight physical simulation network and the target state is small enough, the iteration stops and the current process parameters are output as the optimal solution.

[0052] The core iterative steps for the above inner loop simulation optimization are as follows: Based on the current pre-shrinking process parameter a t and operating condition characteristics x t As input, the virtual state y of the cloth surface is predicted using a lightweight physics simulation network F(·). v,t ; Build status s t The parameter adjustment amount is obtained by inputting the adaptive control network π(·); Update the process parameters according to the step size and trim to a safe range to obtain a t+1 ; Repeat the iteration until the virtual state y is reached. v,t With the target state y target Deviation less than threshold τ in Output the optimal process parameter a .

[0053] The purpose of the outer-loop online adaptation is to adapt to fluctuations in actual production. By acquiring sparse mass feedback data from actual production, the lightweight physical simulation network and adaptive control network are calibrated to ensure continuous adaptation to actual production conditions and avoid the virtual scenario from becoming disconnected from actual production. The specific process is as follows: The sparse quality feedback data is actual fabric quality data, acquired through offline testing equipment. The testing frequency is lower than the production cycle to avoid impacting actual production progress. During testing, multiple fabric samples are randomly selected at the pre-shrinking machine's output. The actual shrinkage rate, actual moisture content, and actual flatness of the samples are measured using offline testing equipment. The average of multiple samples is taken as the final sparse quality feedback data to ensure data representativeness. After testing, the validity of the data must be verified. If the sample data fluctuates excessively, the test is considered invalid and retested. If multiple consecutive tests are invalid, an alarm is triggered on the testing equipment. The deviation between the sparse quality feedback data and the virtual quality data is calculated to clarify the difference between the virtual scenario and actual production. Based on the magnitude of the deviation, it is categorized into three types: small deviation, medium deviation, and large deviation. Different deviations require different handling methods: a small deviation indicates minimal difference between the virtual scenario and actual production, requiring no adjustment; a medium deviation indicates some difference, requiring slight parameter updates; a large deviation indicates significant difference, requiring a complete parameter update and triggering an early warning signal to remind engineers to monitor the production status. The updates target both lightweight physical simulation networks and adaptive control networks. The update process is tailored to the magnitude of deviations to ensure adaptation to actual production conditions. For the adaptive control network, a rapid fine-tuning approach is used, updating only the parameters of the core layers without requiring full retraining, thus improving update efficiency. For the lightweight physical simulation network, the focus is on updating the network's transfer coefficients and adjusting the accuracy of the virtual simulation to ensure the simulation results more closely reflect actual production. After each update, the system automatically backs up the model parameters, retaining recent updates for easy rollback in case of problems. The update frequency is synchronized with the acquisition frequency of sparse quality feedback data; an update is performed each time sparse quality feedback data is acquired. If a large deviation occurs, an update is performed immediately without waiting for the next feedback, ensuring timely adaptation to actual production fluctuations.

[0054] Specifically, based on the outer ring online adaptation, sparse measured quality data y real Calibrate network parameters θ F ,θ A ; The formula for the deviation judgment condition is as follows: ‖y real -F(x;θ F )‖>τ out ; In the formula, y real The data represents sparse measured quality data, i.e., the actual fabric quality (moisture content, shrinkage, and smoothness distribution) obtained through offline testing, and is a feedback signal from actual production; F(x;θ) F This is a lightweight physics simulation network. The inputs are the current production feature x (production process data + image features) and the network parameters θ. F Output the predicted virtual state of the fabric surface: yv; τ outThe outer loop deviation threshold is used to determine that the virtual model differs significantly from actual production when the deviation exceeds this threshold, and the network parameters need to be calibrated. Among them, y real -F(x;θ F ) is a vector subtraction operation. Subtracting the two elements one by one yields the deviation vector, which represents the difference between the measured quality and the virtual prediction in each quality dimension. The deviation judgment formula determines whether network parameter calibration needs to be triggered. If the deviation exceeds the threshold, it indicates that the prediction accuracy of the physical simulation network is insufficient, and the parameters need to be updated to match actual production. Furthermore, the formula for updating the physical simulation network parameters is: ; In the formula, θ F A learnable parameter set for lightweight physics simulation networks; α F The learning rate of the physical simulation network controls the step size of parameter updates, avoiding excessively large step sizes that could lead to model instability. For the physical simulation network loss function L F The gradient of L F It is usually defined as L F =‖y real -F(x;θ F )‖ 2 The gradient represents the direction of parameter adjustment, which minimizes the loss function; The "step size × direction" for parameter updates: gradient Since the gradient term points in the direction that the loss function increases, subtracting it adjusts the parameters in the direction that the loss function decreases, thereby minimizing the deviation between the measured and predicted values. This update method allows the network to continuously "learn" from the measured data, gradually narrowing the gap between virtual predictions and actual production. in, Using the learning rate α F Scaling gradient The size of the parameter update step size is controlled to avoid excessive update amplitude, which may cause model oscillation or non-convergence. This is the core update rule of the gradient descent method. Gradient For the loss function L F The direction of increase, therefore from the current parameter θ F By subtracting the gradient term, the parameters are adjusted in the direction of decreasing the loss function, thereby minimizing the deviation between the measured quality and the virtual prediction and improving the accuracy of the physical simulation network. The physical simulation network parameter update formula uses the gradient descent method to update the physical simulation network parameters, making the network's prediction results closer to the measured quality data and improving the accuracy of the virtual state; Based on the sorting control commands, the pre-shrinking machine is driven to perform actions, completing the adaptive control of the denim pre-shrinking process. This includes translating the sorting control commands into specific actions of each actuator of the pre-shrinking machine, dynamically adjusting the pressure distribution field through the coordinated action of multiple intelligent particle units, driving the linkage of each component, and completing the pre-shrinking action to ensure the accurate implementation of pre-shrinking process parameters. The specific details are as follows, corresponding to the three units of the system's execution module: The mapping unit receives the sorting control commands transmitted by the learning module and maps them into pressure distribution field adjustment commands for the pre-shrinker's rubber blanket compression zone and action commands for each actuator, ensuring that the commands can be directly recognized and executed by the actuators of the pre-shrinker. The specific operation is as follows: The mapping unit receives sorting control commands through a dedicated interface. The receiving frequency is synchronized with the data acquisition frequency to ensure the real-time performance of the commands. If no command is received for several consecutive times during the receiving process, an alarm signal is triggered, and the last valid command is activated to avoid production interruption. After receiving the command, the format of the command is verified to ensure that the command is error-free and complete.

[0055] The pressure parameters in the sorting control command are mapped to the target pressure value for each zone based on the physical partitioning of the pre-shrinking machine's rubber blanket compression zone. The rubber blanket compression zone is discretized into multiple intelligent particle units, each corresponding to a physical partition. During mapping, the target pressure value is reasonably allocated according to the location of each partition to ensure uniform pressure distribution and avoid excessive or insufficient local pressure, which could lead to unqualified pre-shrinking of the fabric.

[0056] The various process parameters in the sorting control instructions are mapped to the corresponding action instructions of the actuators. For example, the temperature parameter is mapped to the target temperature instruction of the temperature control module, specifying the allowable temperature error; the steam pressure parameter is mapped to the steam valve opening instruction, specifying the valve opening size; the rubber blanket rotation speed parameter is mapped to the drive motor rotation speed instruction, specifying the rotation speed error range; and the pre-shrinking time parameter is achieved by adjusting the speed difference between the feed roller and the discharge roller, ensuring accurate pre-shrinking time.

[0057] The mapped instructions are transmitted to the decision-making unit, the coordination unit, and each execution component through a dedicated interface. The output delay is controlled within a reasonable range to ensure synchronized actions. After the instruction is output, the system automatically records relevant information, including instruction type, target value, output time, and corresponding execution component, which facilitates subsequent troubleshooting and production traceability.

[0058] Furthermore, the decision-making unit controls multiple intelligent particle units in the pre-shrinker's rubber blanket compression zone, receives local sensor data from each unit, drives the distributed algorithm to run, provides decision-making basis for the collaborative unit, and ensures accurate adjustment of the pressure distribution field. The specific operation is as follows: Each intelligent particle unit has built-in pressure and temperature sensors, corresponding to a physical partition of the rubber blanket compression zone. They are arranged in a distributed architecture, allowing data exchange between units. A failure in one unit does not affect the operation of other units. The decision-making unit controls the operating status of each intelligent particle unit in real time, ensuring that each unit can collect data and respond to commands normally. If a unit fails, the failure status is immediately marked, a backup unit is activated, and an alarm signal is issued. The sensor data from the failed unit is replaced by the average of the adjacent units, ensuring data continuity.

[0059] The decision-making unit receives local pressure and temperature data collected by each intelligent particle unit in real time, with the collection frequency synchronized with the overall data collection frequency. After receiving the data, it preprocesses and fuses the data, combining the local data with the collected denim production process data and real-time global image data, and uses a weighted average method to fuse the data to improve data accuracy. The fused data is compared with the target value output by the mapping unit. If the deviation is too large, it is judged as data anomaly and immediately fed back to the mapping unit to trigger the instruction for fine adjustment.

[0060] The decision-making unit drives the distributed algorithm to achieve coordination and consistency among multiple intelligent particle units, ensuring a uniform pressure distribution field. During algorithm execution, the target pressure and temperature values ​​output by the mapping unit are used as initial values. Each intelligent particle unit combines its own local data and data from neighboring units to perform collaborative calculations, gradually adjusting its own state to make the pressure and temperature data of each unit tend to be consistent. During the algorithm's operation, it judges in real time whether the data converges. If convergence is achieved, the final decision value of each unit is output; if convergence is not achieved, the number of calculations is increased, and if convergence is still not achieved, an alarm is triggered.

[0061] The decision values ​​of each intelligent particle unit after the distributed algorithm converges are transmitted to the collaborative unit as the core basis for the dynamic adjustment of the pressure distribution field. The format of the decision result is consistent with the format of the instruction output by the mapping unit, ensuring that the collaborative unit can directly recognize and use it.

[0062] Furthermore, the collaborative unit coordinates multiple intelligent particle units to make collaborative decisions, dynamically adjusts the pressure distribution field, and drives the various actuators of the pre-shrinking machine to work together to complete the pre-shrinking action, ensuring the precise execution of the pre-shrinking process parameters. The specific operation is as follows: The collaborative unit receives decision values ​​from each intelligent particle unit output by the decision-making unit, and simultaneously receives action commands from each execution component output by the mapping unit. Combining this with the collected fabric surface state characteristic values, it performs collaborative decision-making. The core of the decision-making process is to determine the rationality of pressure distribution field adjustments and the actions of each execution component based on the fabric surface state and the decision values ​​of each unit. For example, if the fabric surface is slightly uneven, the pressure value and rubber blanket rotation speed in the corresponding area are fine-tuned to improve the fabric surface state; if the fabric surface is severely uneven, a pause action is triggered to re-optimize the process parameters. The decision results are updated in real time, synchronized with the data acquisition frequency, ensuring dynamic adaptation to changes in fabric surface state and process parameters.

[0063] Based on the collaborative decision-making results, the collaborative unit controls the pressure regulating valves corresponding to each intelligent particle unit, dynamically adjusting the valve opening to achieve uniform adjustment of the pressure distribution field. During adjustment, the adjustment speed and accuracy of the valve opening are controlled to ensure timely pressure adjustment response, and the pressure deviation of each unit is controlled within a reasonable range after adjustment. After adjustment, the uniformity of the pressure distribution field is verified by the local pressure data collected by the intelligent particle unit. If it is not uniform, it is fine-tuned again until the target is met. If the pressure in a certain area cannot be adjusted to the target value, the backup valve is immediately activated, and an alarm is triggered simultaneously.

[0064] The coordinated unit drives the various actuators of the pre-shrinking machine to work together to complete the pre-shrinking action according to a preset process. The coordinated process conforms to the pre-shrinking process requirements, specifically including the following: The steam system is started, and the steam pressure is adjusted to the target value and stabilized for a period of time. Then, the rubber blanket drive motor is started, and the speed is adjusted to the target value. Next, the feed roller is started, and the feed speed is matched with the rubber blanket speed. At the same time, the industrial camera and sensors are started to collect data in real time. After the fabric enters the rubber blanket compression zone, the pressure distribution field and temperature are dynamically adjusted to ensure that the pre-shrinking time meets the standard. After the pre-shrinking is completed, the discharge roller is started to transport the fabric to the offline testing equipment to obtain sparse quality feedback data. When the production line stops feeding, each actuator is shut down in sequence to complete one pre-shrinking linkage process.

[0065] If a component malfunctions during the linkage process, its operation is immediately stopped, a backup component is activated, and feedback is sent to the system along with an alarm, recording the fault information. If a linkage anomaly occurs, the pre-shrinking action is immediately paused, all components are adjusted to the baseline state, the process parameters are re-optimized, and the linkage is restarted. After each pre-shrinking action is completed, the execution result is fed back to the learning module and combined with sparse quality feedback data to form a control closed loop, verifying the pre-shrinking effect. If the target is not met, the model is updated and the parameters are re-optimized until the pre-shrinking effect meets the target, thus completing the adaptive control of the denim pre-shrinking process.

[0066] This embodiment features fully automated closed-loop control from data acquisition, virtual simulation, parameter optimization to instruction execution. Dual-loop learning and model updates are executed automatically at a speed that matches the production cycle. Compared to traditional manual trial and error, this significantly shortens the parameter optimization cycle. The distributed intelligent particle unit architecture of the execution module supports local failures without affecting the overall system. All execution components are linked and automated, enabling intelligent continuous operation of pre-production, thereby reducing manual intervention, lowering labor costs, and improving production efficiency.

[0067] In another embodiment, the machine learning-based adaptive control method for denim pre-shrinking process further includes employing a federated learning framework to improve the pre-shrinking process parameters and the generality and adaptability of the optimized pre-shrinking process parameters, specifically: Each denim pre-shrinking production line trains a local model on a local PLC or edge server, and periodically uploads the updated encrypted gradients or parameter changes of the local model to the central cloud server for aggregation to generate a more general global model. The cloud-based central server distributes the global model to each production line, updates the local models of each production line, realizes the aggregation of production experience and knowledge across the entire domain, and at the same time protects the data privacy and security of each factory.

[0068] The dual-loop learning mechanism forms a continuous iterative closed loop of "virtual simulation - actual production - model calibration," enabling the model to adaptively evolve alongside the production process and maintain long-term adaptability. The introduction of a federated learning framework allows for the aggregation of experience from multiple production lines to generate a more universal global model while protecting the data privacy of each production entity. This effectively solves the problem of overfitting from a single model, making this solution adaptable to different fabric specifications, different pre-shrinking machine models, and diverse production conditions, thus enhancing the technology's industry universality and feasibility for implementation.

[0069] In another embodiment, the machine learning-based adaptive control method for denim pre-shrinking processes also includes a process root cause tracing procedure to optimize parameter adjustment direction when pre-shrinking quality fails to meet standards. Construct a cause-and-effect diagram of the pre-shrinking process, and use knowledge of denim pre-shrinking and production process data to clarify the causal relationship between various process variables such as raw material moisture content, steam pressure, rubber blanket elasticity, and upstream process parameters; When the final shrinkage rate of the denim fabric is found to be substandard, the root cause of the substandard shrinkage rate is deduced by tracing the causal chain along the process cause-effect diagram. Based on the root cause, corresponding process compensation suggestions are given, or early warning information is sent to the upstream process to assist in optimizing the pre-shrinking process parameters. The preset pre-shrinking quality standards are: shrinkage rate deviation ≤ ±0.1%, fabric surface flatness reaching the preset flatness threshold, and virtual moisture content distribution deviation ≤ ±5%. The sparse quality feedback data includes actual shrinkage rate detection values ​​and actual fabric surface flatness detection values, collected 1-3 times per batch, providing accurate basis for deviation calculation and model updates in the dual-loop learning process.

[0070] Example 2 See appendix Figure 2-4 Based on the same inventive concept as the machine learning-based adaptive control method for denim pre-shrinking process in Embodiment 1 above, this invention also provides a machine learning-based adaptive control system for denim pre-shrinking process, the system comprising: The data acquisition module is used to acquire denim production process data and real-time images; The simulation module is used to input production process data and real-time images into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The control module is used to input the virtual state of the fabric into a pre-established adaptive control network for optimization processing to obtain the pre-shrinking process parameters; The learning module, based on the pre-shrinking process parameters, executes a double-loop learning process to obtain the sorting control commands for controlling the pre-shrinking machine; The execution module is used to drive the pre-shrinking machine to perform actions according to the sorting control instructions, and to complete the adaptive control of the denim pre-shrinking process.

[0071] The acquisition module, simulation module, control module, learning module, and execution module illustrated in the embodiments of this application (see reference) Figure 2 The system utilizes a communication link built via the Industrial Ethernet protocol to achieve data interaction and command transmission. Each module works collaboratively to ensure the stable and efficient operation of the intelligent production process for denim pre-shrinking. The system adopts a modular, collaborative architecture, with unified and highly independent interfaces for each functional module (acquisition, simulation, control, learning, and execution), facilitating targeted upgrades or replacements (such as sensors and algorithms) without requiring a complete system refactoring. Simultaneously, through the recording of commands and action feedback from the execution modules, as well as the complete storage of production data, the system achieves traceability throughout the entire production process. This provides a solid data foundation for troubleshooting, process review, and continuous optimization, enhancing the system's practicality and lifecycle management capabilities.

[0072] In some embodiments, the simulation module includes: The extraction unit is used to extract key parameters from the production process data and information related to the fabric surface condition from real-time images, and to preprocess and normalize the key parameters and information related to the fabric surface condition. The calculation unit is used to input the preprocessed and normalized key parameters and fabric surface state information into the lightweight physical simulation network, perform forward inference calculation, and obtain the forward inference calculation results. In some embodiments, the execution module includes: The mapping unit is used to receive the sorting control commands transmitted by the learning module and map the sorting control commands into pressure distribution field adjustment commands for the pre-shrinker rubber blanket compression zone and action commands for each actuator. The decision unit is used to control multiple intelligent particle units in the rubber blanket compression zone of the pre-shrinker, receive local sensing data from multiple intelligent particle units, and drive the distributed algorithm to run. The coordination unit is used to coordinate multiple intelligent particle units to make collaborative decisions, dynamically adjust the pressure distribution field, drive the linkage of various execution components of the pre-shrinking machine, and complete the pre-shrinking action.

[0073] It is understandable that, such as Figure 1 The content of the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiments is applicable to the embodiments of this system. The specific functions implemented in the embodiments of this system are the same as those shown in the examples. Figure 1 The embodiment of the adaptive control method for denim pre-shrinking process based on machine learning shown is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiment are the same.

[0074] Another embodiment of the present invention provides a computer device, including: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the machine learning-based adaptive control method for denim pre-shrinking process as described in any of the above methods.

[0075] It is understandable that, such as Figure 1 The content of the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiments is applicable to this embodiment, and the specific functions implemented in this embodiment are the same as those shown in the examples. Figure 1 The embodiment of the adaptive control method for denim pre-shrinking process based on machine learning shown is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiment are the same.

[0076] Another embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the machine learning-based adaptive control method for denim pre-shrinking process as described in any of the above methods.

[0077] It is understandable that, such as Figure 1 The content of the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiments is applicable to this embodiment, and the specific functions implemented in this embodiment are the same as those shown in the examples. Figure 1 The embodiment of the adaptive control method for denim pre-shrinking process based on machine learning shown is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the machine learning-based adaptive control method for denim pre-shrinking process shown in the embodiment are the same.

[0078] The adaptive control system for denim pre-shrinking process based on machine learning provided in this application relates to the field of fabric pre-shrinking technology. This machine learning-based adaptive control system for denim pre-shrinking process can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, smart speaker, smartwatch, or vehicle terminal, but is not limited to these. The server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network. The software can be an application implementing the machine learning-based adaptive control system for denim pre-shrinking process, but is not limited to the above forms.

[0079] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0080] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0081] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0082] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0083] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0084] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0085] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0086] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0087] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0088] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0089] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0090] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. An adaptive control method for denim pre-shrinking process based on machine learning, characterized in that, The method includes: Acquire data and real-time images of the denim production process; The production process data and real-time images are input into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The virtual state of the fabric is input into a pre-established adaptive control network for optimization processing to obtain the pre-shrinking process parameters; Based on the pre-shrinking process parameters, a dual-loop learning process is executed to obtain the sorting control command for controlling the pre-shrinking machine; According to the sorting control command, the pre-shrinking machine is driven to perform actions to complete the adaptive control of the denim pre-shrinking process.

2. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 1, characterized in that, The production process data includes measurable process parameters, fabric properties, and characteristic fluctuation parameters of the raw fabric. The measurable process parameters include the temperature and pressure data of the pre-shrinker's rubber blanket compression zone, the steam pressure of the pre-shrinker, the rubber blanket rotation speed, and the pre-shrinking time; The fabric properties include the fiber composition, thickness, initial moisture content, warp yarn density, weft yarn density, and fabric weave type of the denim. The characteristic fluctuation parameters of the raw fabric include yarn count fluctuation, density fluctuation, and weft skew fluctuation.

3. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 2, characterized in that, The real-time images are used to acquire information related to the surface condition of the fabric. The information related to the surface condition of the fabric includes the smoothness of the denim surface, surface defects, color and luster, texture characteristics, weft skew and arc degree, and surface gloss distribution.

4. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 3, characterized in that, The lightweight physics simulation network is a forward simulation model; The forward simulation model uses the production process data and real-time images to extrapolate the virtual state of the fabric.

5. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 1, characterized in that, The dual-loop learning process includes inner loop simulation optimization and outer loop online adaptation. The inner ring simulation optimization includes performing virtual quality deduction based on the pre-shrinking process parameters, and determining whether the virtual state of the fabric obtained by the deduction meets the preset quality standard. If the target is not met, the pre-shrinking process parameters are iteratively optimized until the virtual quality simulation result meets the target. Then, the pre-shrinking process parameters optimized by the inner loop are used as sorting control commands. The outer ring online adaptation includes acquiring sparse quality feedback data during the actual production process of the pre-shrinking machine, and calculating the deviation between the sparse quality feedback data and the quality data predicted by the virtual state of the fabric. The lightweight physics simulation network and the adaptive control network are updated using the aforementioned deviations; The sparse quality feedback data refers to the actual fabric quality data obtained at a frequency less than the production cycle time and confirmed through offline testing.

6. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 1, characterized in that, The actions performed by the drive pre-shrinking machine include: The rubber blanket compression zone of the pre-shrinking machine is discretized into multiple intelligent particle units. These intelligent particle units make collaborative decisions based on local sensor data and through distributed algorithms, and dynamically adjust the pressure distribution field.

7. The adaptive control method for denim pre-shrinking process based on machine learning according to claim 1, characterized in that, The adaptive control network is trained using reinforcement learning under security constraints. Among them, the pre-shrinking process environment, parameter adjustment actions, and fabric quality reward function are defined as the three elements of reinforcement learning in the reinforcement learning paradigm.

8. An adaptive control system for denim pre-shrinking process based on machine learning, characterized in that, The system includes: The data acquisition module is used to acquire denim production process data and real-time images; The simulation module is used to input the production process data and real-time images into a pre-established lightweight physical simulation network to obtain the virtual state of the fabric surface; The control module is used to input the virtual state of the fabric into a pre-established adaptive control network for optimization processing to obtain the pre-shrinking process parameters; The learning module, based on the pre-shrinking process parameters, performs a double-loop learning process to obtain the sorting control command for controlling the pre-shrinking machine; The execution module is used to drive the pre-shrinking machine to perform actions according to the sorting control instructions, so as to complete the adaptive control of the denim pre-shrinking process.

9. The adaptive control system for denim pre-shrinking process based on machine learning according to claim 8, characterized in that, The simulation module includes: The extraction unit is used to extract key parameters from production process data and fabric surface condition information from real-time images, and to preprocess and normalize the key parameters and fabric surface condition information. The calculation unit is used to input the preprocessed and normalized key parameters and fabric surface state information into the lightweight physical simulation network, perform forward inference calculation, and obtain the forward inference calculation results. The generation unit is used to generate the virtual state of the fabric surface based on the forward extrapolation calculation results and transmit it synchronously to the control module.

10. The adaptive control system for denim pre-shrinking process based on machine learning according to claim 9, characterized in that, The execution module includes: The mapping unit is used to receive the sorting control instructions transmitted by the learning module and map the sorting control instructions into pressure distribution field adjustment instructions for the pre-shrinker rubber blanket compression zone and action instructions for each actuator. The decision unit is used to control multiple intelligent particle units in the rubber blanket compression zone of the pre-shrinker, receive local sensing data from multiple intelligent particle units, and drive the distributed algorithm to run. The coordination unit is used to coordinate multiple intelligent particle units to make collaborative decisions, dynamically adjust the pressure distribution field, drive the linkage of various execution components of the pre-shrinking machine, and complete the pre-shrinking action.

Images