A textile yarn intelligent production control method, system, device and medium

Abstract

This application relates to a method, system, equipment, and medium for intelligent production control of textile yarns. The method includes: collecting yarn diffraction fringes, triaxial vibration spectra, environmental temperature and humidity, and surface image data to form a multi-dimensional sensor dataset, which is then input into a coupled model. Combined with fiber material parameters, the dataset outputs a real-time quality score and a dynamic weight vector. The method also includes receiving global model parameters from federated learning, inputting the score and global parameters into an anomaly decision tree to generate emergency braking or parameter optimization commands. In response to the optimization commands, the method inputs the current process parameter set and the dynamic weight vector into a multi-objective optimization algorithm to obtain an optimized process parameter set. This optimized set is then converted into speed and displacement control signals for the spinning equipment, driving the equipment to perform adjustments. Finally, the sensor data, optimized parameters, quality score, and weight vector are concatenated into a raw string, and a hash value is calculated as a quality DNA identifier, which is then written into a blockchain to generate a traceability block. This application improves the quality stability and production efficiency of yarn products.

Description

Technical Field

[0001] This application relates to the technical field of intelligent manufacturing in the textile industry, and particularly to an intelligent production control method, system, device and medium for textile yarns. Background Art

[0002] As an important basic manufacturing industry, the quality of textile yarns directly determines the quality and value of subsequent fabrics and end products. With the development of industrial Internet, big data and artificial intelligence technologies, textile production is transforming from a traditional mode relying on workers' experience to an intelligent mode centered on data-driven. In the basic link of yarn production in the textile industry, modern spinning equipment has generally equipped with programmable logic controllers, servo drive systems and various on-line detection devices, which can achieve a certain degree of automated production and quality monitoring. For example, the yarn diameter is monitored by capacitive or photoelectric sensors, the operating status of the equipment is collected by vibration sensors, or the appearance image of the yarn is obtained by means of cameras. The application of these technologies aims to improve production efficiency, stabilize product quality and reduce labor costs, representing the mainstream direction of the current development of textile production control technologies.

[0003] However, in the process of moving towards higher-level intelligence, the existing technical system still faces a series of technical problems that need to be solved urgently. First of all, in the face of complex production conditions and diverse raw material characteristics, traditional quality control models often rely on fixed rules or static mathematical models, and it is difficult to adaptively adjust the attention priorities for different quality indicators (such as diameter uniformity, hairiness, strength). Especially when dealing with sudden disturbances or raw material fluctuations, there is a lack of dynamic and flexible multi-objective trade-off mechanisms, which is prone to "attend to one thing and lose sight of another" and cannot achieve the global optimal process adjustment. Moreover, the production experience and knowledge precipitation of a single device are local, and it is impossible to carry out safe and effective collaborative learning and knowledge sharing with the same type of devices under the premise of protecting data privacy. As a result, the improvement of the intelligence level of each device is limited by its own limited data accumulation, and there are bottlenecks in the recognition and disposal capabilities for rare faults or complex conditions, and the overall robustness and learning and evolution capabilities of the system are insufficient. Finally, there may be a risk of being tampered with in the storage and management methods of the massive process data and quality judgment results generated during the production process, and it is difficult to build a real, complete and non-repudiable quality data chain, thus unable to meet the growing demand of the high-end textile market for trustworthy traceability of the full life cycle quality. Summary of the Invention

[0004] In order to address the above technical problems, this application provides an intelligent production control method, system, device and medium for textile yarns.

[0005] In the first aspect, this application provides an intelligent production control method for textile yarns, adopting the following technical solutions:

[0006] A method for intelligent production control of textile yarn, the method comprising:

[0007] By triggering the sensor group deployed on the spinning equipment through a synchronous clock source, yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data, and yarn surface image data are collected to generate a timestamp-aligned multidimensional sensor dataset.

[0008] The multidimensional sensing dataset is input into a pre-built coupling model, and combined with the fiber material characteristic parameters pre-stored in the database, the real-time yarn diameter, vibration energy integral value and hair gradient value are calculated and generated, and the real-time quality score and dynamic weight vector are output.

[0009] Receive global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into a pre-constructed anomaly decision tree, and generate emergency braking instructions or parameter optimization instructions;

[0010] In response to the parameter optimization instruction, the current process parameter set and the dynamic weight vector are input into a multi-objective optimization algorithm to generate an optimized process parameter set;

[0011] The optimized process parameter set is converted into spinning equipment control signals to drive the textile equipment to perform adjustments; wherein, the spinning equipment control signals include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide;

[0012] The multidimensional sensor dataset, optimized process parameter set, real-time quality score, and dynamic weight vector are concatenated to generate an original string. The hash value is calculated as a quality DNA identifier and written into the blockchain network to generate a quality traceability block.

[0013] By adopting the above technical solutions, synchronous sensing, federated learning, dynamic multi-objective optimization, and blockchain traceability are deeply integrated into the spinning process. This not only enables millisecond-level real-time monitoring and adaptive process optimization of yarn quality and equipment status, significantly improving the consistency of yarn quality and production efficiency, but also, through the introduction of federated learning, allows individual devices to benefit from collective intelligence, enhancing the system's robustness in dealing with complex operating conditions and unknown disturbances. Finally, by using blockchain technology to solidify the data fingerprints of the entire production chain, a solid technical foundation is provided for reliable quality assurance and transparent supply chain management of high-end textiles, representing an advanced direction for the intelligent and digital integration of the textile industry.

[0014] Secondly, this application provides an intelligent production control system for textile yarns, which adopts the following technical solution:

[0015] A smart production control system for textile yarn, comprising:

[0016] The synchronous sensing module is used to trigger the sensor group deployed on the spinning equipment through a synchronous clock source to collect yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data and yarn surface image data, and generate a timestamp-aligned multidimensional sensing dataset.

[0017] The coupling evaluation module is used to input the multidimensional sensing dataset into a pre-built coupling model, combine it with the fiber material characteristic parameters pre-stored in the database, calculate and generate the real-time yarn diameter, vibration energy integral value and hair gradient value, and output the real-time quality score and dynamic weight vector.

[0018] The intelligent decision-making module is used to receive the global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into the pre-built anomaly decision tree, and generate emergency braking instructions or parameter optimization instructions.

[0019] The multi-objective optimization module is used to respond to the parameter optimization command by inputting the current process parameter set and the dynamic weight vector into the multi-objective optimization algorithm to generate an optimized process parameter set;

[0020] A servo control module is used to convert the optimized process parameter set into spinning equipment control signals to drive the textile equipment to perform adjustments; wherein, the spinning equipment control signals include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide;

[0021] The full-chain evidence storage module is used to concatenate the multi-dimensional sensor dataset, optimized process parameter set, real-time quality score and dynamic weight vector to generate an original string, calculate the hash value as a quality DNA identifier, and write it into the blockchain network to generate a quality traceability block.

[0022] Thirdly, this application provides a computer device, which adopts the following technical solution:

[0023] A computer device includes a memory, a processor, and a computer program stored in the memory, the processor executing the computer program to perform the steps of the method as described in the first aspect.

[0024] Fourthly, this application provides a computer-readable storage medium, which adopts the following technical solution:

[0025] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as in any of the methods in the first aspect.

[0026] In summary, this application includes at least one of the following beneficial technical effects: by using multi-dimensional sensor data fusion and AI algorithms, it achieves fully automated quality monitoring and process optimization in yarn production, enabling real-time monitoring of key indicators such as yarn diameter, vibration state, and surface quality, automatically identifying production anomalies and intelligently adjusting parameters, significantly improving the stability of yarn product quality and production efficiency. At the same time, it establishes a complete quality traceability system through blockchain technology, ensuring the verifiability of product quality and transparent management of the production process. Attached Figure Description

[0027] Figure 1 This is a schematic diagram of the first process of a smart production control method for textile yarn according to one embodiment of this application.

[0028] Figure 2 This is a schematic diagram of the second process of a smart production control method for textile yarn according to one embodiment of this application.

[0029] Figure 3 This is a schematic diagram of the third process of a smart production control method for textile yarn according to one embodiment of this application.

[0030] Figure 4 This is a schematic diagram of the fourth process of a smart production control method for textile yarn according to one embodiment of this application.

[0031] Figure 5 This is a schematic diagram of the fifth process of a smart production control method for textile yarn according to one embodiment of this application.

[0032] Figure 6 This is a schematic diagram of the sixth process of a smart production control method for textile yarn according to one embodiment of this application.

[0033] Figure 7 This is a schematic diagram of the seventh process of a smart production control method for textile yarn according to one embodiment of this application. Detailed Implementation

[0034] To make the purpose, technical solution, and advantages of this application clearer, the following description is provided in conjunction with the appendix. Figures 1-7 The present application will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the application.

[0035] This application discloses an intelligent production control method for textile yarns.

[0036] Reference Figure 1 A method for intelligent production control of textile yarn, the method comprising:

[0037] Step S101: The sensor group deployed on the spinning equipment is triggered by the synchronous clock source to collect yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data and yarn surface image data, and generate a timestamp-aligned multidimensional sensor dataset.

[0038] The synchronization clock source serves as the time reference for the entire data acquisition system. Its core function is to ensure that sensor groups distributed across different spinning machines are triggered at an absolutely uniform time. This synchronization mechanism eliminates signal correlation errors caused by acquisition time delays or misalignments, providing time consistency guarantees for subsequent data fusion and model analysis.

[0039] In this embodiment, the sensor group is a collection of sensing devices specifically deployed at key nodes in the yarn production process. It includes a laser interferometer for measuring changes in yarn structure, which analyzes the diffraction fringe deformation matrix generated when the yarn is irradiated by a laser to reflect subtle changes in the internal structure of the yarn, such as diameter and twist. It also includes a microelectromechanical system (MEMS) vibration sensor, which can collect acceleration spectrum data of spinning equipment (such as spindles and rollers) on the three orthogonal axes of X, Y, and Z to analyze the mechanical smoothness and abnormal vibration modes of the equipment. An environmental temperature and humidity sensor monitors the temperature and humidity parameters of the production environment, which directly affect the physical properties of the fiber (such as moisture regain) and the yarn forming process. Furthermore, a machine vision system equipped with an industrial camera captures real-time images of the yarn surface and extracts morphological features such as hair density and contour index through image processing algorithms.

[0040] Driven by a synchronized clock, all sensors collect multi-dimensional and heterogeneous raw data, including geometric, mechanical, visual, and environmental data. The data is then timestamped and preliminarily integrated to form a multi-dimensional sensor dataset with a unified time index. This dataset forms the foundational data layer for all subsequent intelligent analysis and control decisions.

[0041] Step S102: Input the multidimensional sensing dataset into the pre-built coupling model, combine it with the fiber material characteristic parameters pre-stored in the database, calculate and generate the real-time yarn diameter, vibration energy integral value and hair gradient value, and output the real-time quality score and dynamic weight vector.

[0042] The model is a comprehensive analysis system that integrates physical mechanisms and data-driven approaches. It first combines pre-stored fiber material characteristic parameters (such as fiber type, length, fineness, initial modulus, and moisture regain) in a database, using the constitutive relationship of the fiber as prior knowledge to interpret the physical meaning and extract features from the original sensor signals. Specifically, the model analyzes the diffraction fringe data of the laser interferometer, combines it with the material's optical properties to calculate the real-time diameter and rate of change of the yarn; it integrates the triaxial vibration spectrum to obtain the vibration energy integral value representing the energy state of the equipment's operation; simultaneously, it performs gradient analysis and edge detection on the yarn surface image to calculate the feathering gradient value, characterizing the amount and uniformity of the yarn surface feathering.

[0043] Based on these extracted core quality indicators, the model further implements a comprehensive evaluation. The real-time quality score is a comprehensive quantitative value calculated by normalizing and weighting multiple indicators such as diameter uniformity, vibration energy, and yarn hairiness, used to quickly determine the overall level of current yarn quality. The generation of the dynamic weight vector is even more intelligent; it is not fixed but dynamically adjusts the focus of each quality indicator according to the real-time production situation.

[0044] In this embodiment, the environmental coupling coefficient is first calculated based on real-time temperature and humidity and fiber moisture regain to reflect the degree of environmental influence on yarn performance. Then, the diameter weighting factor is determined based on the drastic change in the real-time yarn diameter. Finally, according to the predetermined vector formula W=[α·γ,β,γ], parameters such as the diameter weighting factor α, vibration energy coefficient γ, and environmental coupling coefficient γ are fused to generate a three-dimensional dynamic weight vector. This vector reflects the different priorities of the system in responding to the three dimensions of "diameter control," "vibration suppression," and "environmental compensation" under the current specific operating conditions, providing precise weight guidance for subsequent optimization and decision-making.

[0045] Step S103: Receive the global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into the pre-built anomaly decision tree, and generate an emergency braking command or parameter optimization command.

[0046] This step introduces a collaborative intelligence mechanism based on federated learning networks. A federated learning network is a distributed machine learning framework that allows each spinning terminal (local device) to periodically (e.g., every 5 minutes) upload encrypted local model parameters (such as decision tree splitting thresholds, node weights, etc.) without sharing raw data. A central server or node securely aggregates (e.g., uses a weighted average) the encrypted parameters from multiple terminals, generating optimized global model parameters, which are then distributed to each terminal.

[0047] Next, the local system inputs the output real-time quality score along with the received global model parameters into the anomaly decision tree. The anomaly decision tree is a classification model whose node splitting conditions (i.e., decision thresholds) can be dynamically adjusted by the global model parameters, thereby incorporating the production experience and failure modes of other similar devices across the network.

[0048] Specifically, the decision tree's judgment logic is as follows: First, it determines whether the rate of decrease or fluctuation in the real-time quality score exceeds a preset threshold calibrated by global experience. Simultaneously, it considers whether the components reflecting vibration issues (such as the β component) in the dynamic weight vector exceed a critical value. If both conditions are met, an emergency situation such as yarn breakage or severe equipment instability is identified, and an emergency braking command is generated. If the emergency conditions are not met, the decision tree generates parameter optimization commands based on the optimization direction provided by the global model. Essentially, this combines the real-time status of a single device with collective intelligence to make more robust and forward-looking decisions.

[0049] Step S104: In response to the parameter optimization instruction, the current process parameter set and dynamic weight vector are input into the multi-objective optimization algorithm to generate an optimized process parameter set;

[0050] When the system receives a parameter optimization command, it initiates the process parameter optimization process. This step inputs the current set of process parameters (such as spinning speed, draft ratio, twist setting, etc.) along with the generated dynamic weight vector into a multi-objective optimization algorithm (such as genetic algorithm, particle swarm optimization, etc.). In the algorithm, process parameters are encoded as optimization variables (such as chromosome genes). The core of the algorithm is the fitness function, the design of which directly reflects the guiding role of the dynamic weight vector.

[0051] For example, the function might be constructed as a weighted difference between yarn strength (a positive indicator) and yarn hairiness and energy consumption (negative indicators), with the weights (W1, W2, W3) directly derived from the dynamic weight vector or its derivatives. In this way, the optimization process no longer seeks a fixed "optimal solution," but rather finds the optimal balance among multiple conflicting objectives such as strength, smoothness, and efficiency, under the priority defined by the current dynamic weights and targeting specific weak points (e.g., high vibration requiring reduced energy consumption, and excessive hairiness requiring strong control). Through iterative calculation, the algorithm ultimately outputs an optimized set of process parameters that performs best overall under the current operating conditions, achieving real-time, adaptive adjustment of process parameters.

[0052] Step S105: Convert the optimized process parameter set into spinning equipment control signals to drive the textile equipment to perform adjustments; wherein, the spinning equipment control signals include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide;

[0053] Understandably, this step is responsible for accurately mapping optimization decisions from the digital world to control actions of physical equipment. By receiving a set of optimized process parameters, it generates control signals that can directly drive the actuators based on a series of preset conversion rules that reflect the electromechanical characteristics of the equipment.

[0054] Specifically, based on the twist setting value in the optimized parameters, a pre-calibrated twist-displacement mapping table is consulted to determine the precise position where the pneumatic yarn guide needs to move. This generates a corresponding pulse control signal to drive its action, thereby altering the yarn twisting stroke and effect. Simultaneously, based on the machine speed setting value in the optimized parameters, a proportional-integral-derivative (PID) control algorithm is used to generate a speed adjustment signal for the variable frequency motor, ensuring the motor speed smoothly and quickly tracks the set value. This step achieves a closed loop from optimized process parameters to the underlying drive signals of the equipment, a crucial link in generating practical production benefits from intelligent decision-making.

[0055] Step S106: The multidimensional sensor dataset, optimized process parameter set, real-time quality score and dynamic weight vector are concatenated to generate the original string, the hash value is calculated as the quality DNA identifier, and written into the blockchain network to generate a quality traceability block.

[0056] To construct an immutable and fully reliable quality archive, this step involves fusing and storing all key process data. Heterogeneous data, including the original multidimensional sensor dataset, the optimized process parameter set based on decision results, and the real-time quality score and dynamic weight vector of the comprehensive evaluation results, are sequentially concatenated to form a complete original string.

[0057] Next, a cryptographic hash function (such as SHA-256) is applied to the string to generate a unique and irreversible hash value. This hash value acts as a digital fingerprint of all production status and decision-making information for that batch of yarn at that moment, and is named the "quality DNA identifier." This identifier and related data index are written into the blockchain network to generate a quality traceability block. The blockchain's distributed ledger, chain structure, and consensus mechanism ensure that once this information is on the chain, it cannot be unilaterally tampered with or deleted, providing a highly reliable digital traceability certificate for the entire production process of each piece of yarn, from raw materials, real-time working conditions, adjustment decisions to final quality rating.

[0058] In the above implementation, synchronous sensing, federated learning, dynamic multi-objective optimization, and blockchain traceability are deeply integrated into the spinning process. This not only enables millisecond-level real-time monitoring and adaptive process optimization of yarn quality and equipment status, significantly improving the consistency of yarn quality and production efficiency, but also, through the introduction of federated learning, allows individual devices to benefit from collective intelligence, enhancing the system's robustness in dealing with complex operating conditions and unknown disturbances. Finally, by using blockchain technology to solidify the data fingerprints of the entire production chain, a solid technical foundation is provided for reliable quality assurance and transparent supply chain management of high-end textiles, representing an advanced direction for the intelligent and digital integration of the textile industry.

[0059] Reference Figure 2 As one implementation of step S102, the steps of inputting a multidimensional sensing dataset into a pre-constructed coupled model, combining it with fiber material characteristic parameters pre-stored in a database, calculating and generating the real-time yarn diameter, vibration energy integral value, and hairiness gradient value, and outputting a real-time quality score and dynamic weight vector include:

[0060] Step S201: Analyze the multidimensional sensor dataset and separate the yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data, and yarn surface image data.

[0061] Step S202: Gaussian filtering is applied to the yarn diffraction stripe deformation data for noise reduction. Based on the noise-reduced data, the real-time yarn diameter is calculated using a morphological inversion algorithm.

[0062] Specifically, this step focuses on extracting the core geometric dimensions of the yarn from the optical signal. Besides the effective signal caused by changes in yarn diameter, the diffraction fringe deformation data is also mixed with interference from equipment vibration, air disturbances, and electronic noise. Therefore, Gaussian filtering is first used for noise reduction. This is a linear smoothing filter that effectively suppresses high-frequency noise through convolution operations while preserving the edges of the fringe image (i.e., phase transition information), providing clean input for subsequent accurate analysis.

[0063] Subsequently, based on the denoised fringe data, a morphological inversion algorithm is applied to calculate the real-time yarn diameter. The algorithm's logic is to solve an "inverse problem": given the far-field diffraction pattern (i.e., fringe morphology) formed after laser light passes through the yarn, the cross-sectional shape and size of the yarn that caused this pattern are deduced in reverse. This process typically involves establishing a physical model of optical diffraction, iteratively comparing the observed fringe matrix with the predicted fringes generated by the model (e.g., using optimization algorithms such as iterative nearest point), and continuously adjusting the yarn diameter parameters in the model until the solution with the highest matching degree to the measured data is found. This solution is the inverted real-time diameter.

[0064] In some embodiments, the specific steps of the morphology inversion algorithm include: constructing a two-dimensional coordinate mapping relationship of the yarn diffraction stripe deformation matrix; using an iterative nearest-point algorithm to match a preset yarn cross-sectional morphology template; and calculating the real-time diameter change based on the optimal matching result.

[0065] Step S203: Perform a fast Fourier transform on the triaxial vibration spectrum data, perform spectrum energy integration calculation within a preset frequency band, and generate vibration energy integral values;

[0066] Specifically, the spectrum energy integration calculation includes: dividing the 0-1kHz frequency band into multiple sub-bands; calculating the energy density integral of each sub-band; and vector superimposing the triaxial integration results to generate the total vibration energy value.

[0067] This step aims to transform the time-domain vibration waveform into a quantitative energy index characterizing the overall operating status of the equipment. First, a Fast Fourier Transform (FFT) is performed on the separated triaxial vibration spectrum data (actually time-domain acceleration signals). FFT is an efficient algorithm that converts time-series signals from the time domain to the frequency domain, decomposing the change of vibration over time into a superposition of sinusoidal components of different frequencies, and revealing the amplitude (or energy) distribution of each frequency component, thus forming a spectrum.

[0068] Subsequently, spectral energy integration is performed within a preset frequency band (e.g., 0-1kHz, which typically covers the main operating frequencies and common fault characteristic frequencies of spinning equipment). The logic is as follows: the power spectral density curve within a specific frequency band in the spectrum is integrated. Power spectral density reflects the distribution of signal power across frequencies; integration yields the total energy contained in the vibration within that frequency band. Since the data comes from the X / Y / Z axes, the integration results for each axis need to be vector-synthesized or weighted summed to ultimately generate a scalarized vibration energy integral value. This value comprehensively reflects the overall vibration intensity of the equipment within the key frequency band and is the core quantitative basis for judging the equipment's operational stability and whether its mechanical condition is abnormal.

[0069] Step S204: Perform edge detection and contour analysis on the yarn surface image data, extract the hair distribution feature vector, and calculate the hair gradient value;

[0070] Specifically, the steps for calculating the hairline gradient value include: extracting the yarn edge gradient using the Sobel operator; sampling a set number of feature points on the contour curve; and calculating the standard deviation of the feature point offset as the hairline gradient value.

[0071] This step utilizes digital image processing technology to evaluate the surface quality of the yarn. For the separated yarn surface image data, edge detection is first performed (using operators such as Sobel or Canny). This is based on the characteristic that image grayscale changes drastically at the edges of the yarn fibers. By calculating the gradient magnitude and direction of pixels, the contour boundaries of the yarn body and the protruding fibers are delineated. Next, contour analysis is performed, which involves tracking and mathematically describing the detected edge contours.

[0072] Based on this, a feature vector of yarn hair distribution is extracted, which may include a set of feature values ​​such as hair length, quantity, spatial distribution density, and azimuth. Calculating the hair gradient value is a core comprehensive evaluation indicator, its logic being to quantify the irregularity of the yarn profile: a series of feature points are sampled at equal intervals on the extracted profile curve, the offset of each point from the ideal smooth profile baseline is calculated, and then the standard deviation of these offsets is calculated as the hair gradient value. A larger value indicates more severe profile fluctuations, more abundant and longer hairs, or a more uneven distribution; conversely, a smaller value indicates a smoother yarn surface.

[0073] Step S205: Input the real-time yarn diameter, vibration energy integral value, hair gradient value and ambient temperature and humidity data into the pre-built coupling model, and load the fiber material characteristic parameters pre-stored in the database.

[0074] This process involves aggregating and using real-time yarn diameter (representing geometric dimensions), vibration energy integral (representing mechanical state), hairiness gradient (representing apparent quality), and environmental temperature and humidity data (representing environmental conditions) as input. Simultaneously, pre-stored fiber material characteristic parameters (such as fiber type, standard moisture regain, specific heat capacity, and coefficient of friction) are loaded into the database. These material parameters are prior, relatively static knowledge, providing crucial context for understanding and interpreting the aforementioned dynamic characteristics. For example, the same diameter fluctuation may have different meanings for cotton and synthetic fibers; the same environmental humidity will affect fibers with different hygroscopic properties to varying degrees. Therefore, this step essentially aligns and correlates the real-time sensed dynamic data stream with the existing material knowledge base, providing a comprehensive and context-rich decision input matrix for subsequent weighted fusion.

[0075] Step S206: Calculate the fiber moisture regain based on the ambient temperature and humidity data, determine the humidity influence factor through the material moisture absorption characteristic curve in the fiber material characteristic parameters, and use the humidity influence factor as the dynamic environmental calibration coefficient.

[0076] The dynamic environment calibration coefficient is an adaptive adjustment parameter of the weighted fusion layer of the coupled model. It is dynamically generated based on the real-time ambient temperature and humidity and the hygroscopic properties of the fiber material, and is used to adjust the weights of different features in the fusion process in real time. For example, in a high-humidity environment, the weights of features related to fiber plasticity and viscosity (which may affect diameter uniformity) will be adaptively increased.

[0077] Specifically, firstly, based on real-time monitored ambient temperature and humidity data, the instantaneous moisture regain of the fiber under the current working conditions is calculated using empirical formulas or physical models. This transforms macroscopic environmental state quantities (temperature and humidity) into key intrinsic variables that directly affect the fiber's mechanical properties (such as softness, strength, and coefficient of friction) and process adaptability, thus establishing the first-level quantitative correlation between the environment and material properties. Subsequently, the system calls upon a pre-stored material moisture absorption characteristic curve characterizing the inherent properties of the specific fiber. This curve represents the characteristic mapping relationship between the standard moisture regain of the fiber material and ambient humidity. By substituting the calculated instantaneous moisture regain into this characteristic curve for comparison and interpolation, a quantified humidity influence factor can be accurately determined. This factor is ultimately assigned as a dynamic environmental calibration coefficient, the purpose of which is to summarize the complex influence of ambient humidity on yarn forming quality into an adjustment parameter that can be embedded in the subsequent weighted fusion model, enabling the system's quality assessment and decision-making to adaptively correct based on environmental changes.

[0078] Step S207: Through the weighted fusion layer of the coupled model, combined with the dynamic environment calibration coefficient, output the real-time quality score and dynamic weight vector.

[0079] In this model, the multidimensional features and material parameters are integrated and calculated in the weighted fusion layer. The core logic of this layer is not a simple weighted average, but a nonlinear synthesis that considers the interaction between factors. Finally, the weighted fusion layer integrates all inputs and outputs two key results: first, a real-time quality score, which is a single comprehensive score that integrates multiple indicators such as diameter uniformity, vibration energy, and yarn hairiness under the guidance of dynamic weights, intuitively reflecting the overall quality level of the yarn; second, a dynamic weight vector, in which each component (such as diameter weight factor, vibration concern coefficient, and environmental compensation factor) clarifies the priority and control focus of the system on various quality indicators under the current specific working conditions (defined by the environment, material, and real-time status).

[0080] In the above embodiments, Gaussian filtering and morphological inversion are used to achieve precise online optical measurement of yarn diameter, frequency domain integration is used to achieve objective quantification of equipment vibration energy, and machine vision and contour analysis are used to achieve digital characterization of yarn hairiness. Finally, with the help of the weighted fusion layer of the coupled model, the above features are deeply integrated with fiber material properties and real-time environmental parameters. The evaluation center of gravity is adaptively adjusted through dynamic environmental calibration coefficients, which not only outputs a real-time score reflecting the overall quality status, but also generates a dynamic weight vector to guide subsequent differentiated control. This technical solution elevates traditional isolated detection to systematic and correlated intelligent perception and evaluation, improving the accuracy, comprehensiveness, and contextual awareness of quality judgment.

[0081] Reference Figure 3 As one implementation of step S103, the steps of receiving global model parameters periodically issued by the federated learning network, inputting the real-time quality score and global model parameters into a pre-constructed anomaly decision tree, and generating an emergency braking command or parameter optimization command include:

[0082] Step S301: parse the encrypted data packets transmitted by the federated learning network, decrypt them using a preset key to obtain the local model parameters of each spindle terminal, perform a weighted average operation on the decrypted local model parameters, and generate global model parameters.

[0083] Federated learning networks, as a distributed machine learning system, achieve collaborative model optimization without aggregating the raw, sensitive data from each spindle terminal. The periodically distributed encrypted data packets encapsulate local model parameters (such as split point thresholds and node weights in decision trees) trained by each terminal based on its own production data. After parsing and decryption using a pre-set key, the system obtains a set of parameters processed using homomorphic encryption or secure multi-party computation techniques, rather than the original data. This fundamentally ensures the privacy and security of production data.

[0084] Subsequently, a weighted average calculation is performed on the decrypted local model parameters of each terminal (for example, assigning different weights based on the data volume, model quality, or production stability of each terminal), thereby integrating the operational experience and knowledge of multiple production units across the network to generate a set of optimized global model parameters. This set of parameters essentially carries the ability to identify fault modes and quality evaluation standards under a wider range of operating conditions. When distributed to local terminals, it is equivalent to injecting "collective intelligence" into individual devices, enabling their decision-making benchmarks to no longer be limited to their own limited historical data, significantly improving the model's generalization ability and predictability for rare anomalies.

[0085] Step S302: Input the real-time quality score and global model parameters into the pre-built anomaly decision tree, and generate a judgment result based on the dynamic threshold rule;

[0086] Among them, the anomaly decision tree is a predefined, tree-structured classification model. Its inputs are the real-time quality score output by the coupled model and the global model parameters issued by federated learning. The dynamic threshold rule is the key to the decision tree's ability to adapt to different production stages and environments.

[0087] Specifically, the construction of dynamic threshold rules includes: loading predefined baseline quality thresholds; adjusting the baseline thresholds based on the distribution characteristics in the global model parameters; and generating dynamic judgment thresholds by combining the short-term rate of change of real-time quality scores.

[0088] Understandably, the critical thresholds used in the anomaly decision tree to determine whether a state is abnormal (e.g., the lower limit of the quality score, the limit of the rate of decline, etc.) are not fixed, but are dynamically adjusted by the loaded baseline threshold based on the distribution characteristics in the global model parameters (such as the mean, variance, skewness, and other statistics of the overall quality scores). For example, when the global model parameters reflect that the current batch of raw materials generally leads to a slight decrease in scores, the system will automatically and appropriately relax the anomaly threshold of the local score to avoid false alarms caused by common raw material characteristics; conversely, it will tighten the threshold to improve detection sensitivity. The decision tree uses these dynamically calibrated thresholds layer by layer to make logical judgments on real-time quality scores and their short-term change rates, and finally outputs a judgment result on whether the current production state is abnormal.

[0089] In some embodiments, the execution logic of the anomaly decision tree includes: determining whether the real-time quality score is lower than the dynamic judgment threshold at the first-level node; detecting whether the rate of change of the quality score exceeds the critical slope at the second-level node; verifying whether the vibration energy integral value is abnormally exceeded at the third-level node; and triggering the anomaly condition and marking it as an anomaly state when all three levels of nodes meet the conditions.

[0090] Step S303: Determine whether the judgment result triggers an abnormal condition; if yes, proceed to step S304; if no, proceed to step S305.

[0091] Step S304: Generate an emergency braking command;

[0092] Specifically, when the anomaly decision tree's judgment result meets a preset combination of anomaly conditions representing a high-risk state (e.g., the real-time quality score is not only below the dynamic judgment threshold, but its rate of decline also exceeds the critical slope, and the vibration energy integral value also confirms an abnormal over-limit), the system will immediately generate an emergency braking command. This command is designed to respond to emergencies such as impending yarn breakage or serious mechanical failure of critical equipment components, immediately interrupting the process to prevent secondary damage or safety accidents. Its goal is to transmit control signals to the underlying drivers of the spinning equipment (such as frequency converters and servo systems) within milliseconds to execute emergency stops or specific protective actions, thereby minimizing production risks and material losses.

[0093] Step S305: Output parameter optimization instructions to the multi-target optimization module.

[0094] Specifically, when the anomaly decision tree determines that the current state does not meet the emergency braking standard but may be within an optimizable range (e.g., the quality score is within the acceptable range but close to the lower threshold, or the dynamic weight vector indicates a weakness in a certain aspect), the system will automatically trigger the parameter optimization process. The parameter optimization instruction generated at this time is not a simple adjustment suggestion, but a structured, executable command package. This instruction typically encapsulates the version number of the current process parameter set (for traceability and consistency verification), a precise timestamp (for synchronization with the control cycle), and a dynamic weight vector calculated by the coupled model (indicating the quality dimension that needs to be prioritized for optimization). This instruction is transmitted to the downstream multi-objective optimization module as one of the direct trigger signals and core inputs for initiating a new round of process parameter optimization calculations, guiding the optimization algorithm to adaptively optimize towards improving efficiency, reducing energy consumption, or improving specific quality indicators while ensuring safety and quality.

[0095] The above implementation not only enhances the early identification and warning capabilities of individual equipment for complex faults and performance degradation, thus strengthening the overall robustness of the production system, but also achieves full coverage from emergency safety braking in extreme situations to continuous fine-tuning of process parameters during normal operation through a tiered response mechanism. This solution enables yarn production control to shift from passive monitoring and experience-based adjustments to an intelligent mode of proactive early warning, collaborative optimization, and adaptive decision-making. While steadily improving the consistency of yarn quality, it provides crucial technical support for maximizing production efficiency and equipment safety.

[0096] Reference Figure 4 As one implementation of step S104, in response to the parameter optimization instruction, the step of inputting the current process parameter set and dynamic weight vector into a multi-objective optimization algorithm to generate an optimized process parameter set includes:

[0097] Step S401: In response to the parameter optimization command, read the currently running set of process parameters and the dynamic weight vector output by the coupled model;

[0098] The parameter optimization command indicates that the system, after analysis, has determined that the current production status is in a stage of "operable but with room for optimization," rather than an abnormal state requiring emergency braking. The current set of process parameters is the collection of core process settings being executed by the spinning equipment, such as spinning speed, draft ratio between drafting rollers, and twist determined by the movement of the spindle and ring ring. These parameters are direct variables controlling yarn formation and determining the final product performance, forming the starting point and benchmark for optimization operations. The dynamic weight vector is a weight allocation scheme dynamically calculated by the coupled model based on real-time collected data such as yarn diameter, vibration energy, hairiness gradient, and ambient temperature and humidity, combined with fiber material characteristics. It quantifies the priority and importance of various quality indicators (such as strength, smoothness, and energy consumption) in the optimization objectives under the current specific production situation.

[0099] Step S402: Encode the process parameter set into the initial solution space of the multi-objective optimization algorithm, and at the same time construct the multi-objective fitness function based on the dynamic weight vector;

[0100] Specifically, the set of process parameters is encoded into the initial solution space of the optimization algorithm. Since intelligent optimization algorithms such as genetic algorithms and particle swarm optimization algorithms usually deal with encoded "individual" or "position" vectors, it is necessary to map process parameters with different physical meanings and dimensions, such as speed, draw ratio, and twist, into a unified, algorithm-operable encoding sequence through normalization and other methods. This encoding set defines the starting range for the algorithm to search and iterate and the expression form of possible solutions.

[0101] Secondly, the multi-objective fitness function is used to evaluate the merits of each candidate solution (i.e., a set of process parameter combinations). In yarn production, optimization objectives are often multiple conflicting indicators. For example, increasing yarn strength may require reducing machine speed or increasing twist, which may contradict the objectives of reducing energy consumption and increasing production efficiency.

[0102] In this embodiment, a dynamic weight vector is used to construct the fitness function in real time. The specific construction formula includes: F = w1·S + w2·(1 / H) + w3·(1 / E), where w1, w2, and w3 are dynamic weight components, S is the intensity value, H is the feathering index, and E is the energy consumption value. The fitness function is constructed by weighting and summing the intensity, the reciprocal of the feathering index (representing feathering suppression), and the reciprocal of the energy consumption value (representing efficiency), with the weight coefficients being the components of the dynamic weight vector. This means that when a significant feathering problem is detected in real time, the corresponding weight w2 in the vector will automatically increase, thus naturally guiding the search direction of the constructed fitness function in this optimization iteration towards the process parameter region that can more effectively suppress feathering.

[0103] Step S403: Under preset process constraints, perform iterative calculations of a multi-objective optimization algorithm to generate a Pareto optimal solution set;

[0104] These process constraints are physical or safety limitations that cannot be violated in actual production processes. For example, the speed must be within the feasible range of equipment capacity and product quality requirements (e.g., 50-150 m / min), the adjustment range of the draft ratio cannot exceed the allowable deviation threshold of the equipment's mechanical precision (e.g., ±0.15), and the fluctuation of twist must be controlled within a certain tolerance (e.g., ≤5%) to avoid causing instability in the yarn structure.

[0105] Specifically, during the iterative process, the algorithm continuously generates new candidate process parameter codes (i.e., new "solutions") and evaluates them using a constructed fitness function. Simultaneously, it rigorously checks whether these candidate solutions satisfy all process constraints, eliminating invalid solutions that violate the constraints. Through generations of "selection-crossover-mutation" (using a genetic algorithm as an example), the algorithm searches within the feasible region defined by the constraints, in the direction indicated by the fitness function. Its ultimate goal is not to find a single best solution (because such a solution usually does not exist under multi-objective conditions), but rather to find a series of Pareto optimal solutions, which constitute a Pareto optimal solution set. In this solution set, any further improvement in any objective inevitably leads to the deterioration of at least one other objective. This solution set represents the set of all possible optimal trade-offs under the existing process constraints, providing decision-makers with a comprehensive optimal choice surface.

[0106] Step S404: Select the target solution from the Pareto optimal solution set using a preset strategy, and decode to generate an optimized process parameter set.

[0107] Since all solutions in the Pareto solution set are "optimal", a target solution needs to be selected for actual production based on a preset strategy.

[0108] In some embodiments, the Euclidean distance between each solution in the solution set and an "ideal point" (i.e., a virtual point where each sub-objective individually reaches its optimum) can be calculated. The solution with the smallest distance is then selected as the target solution, which is considered to have achieved the best global balance among multiple competing objectives. After selecting the target solution, a decoding operation, the reverse of encoding, needs to be performed. This involves inverting the encoded sequence (such as a chromosome gene sequence) within the algorithm and restoring it to specific process parameters with defined physical units and values. For example, restoring the value at a certain gene locus to a specific vehicle speed of "145 m / min". The optimized process parameter set generated by decoding is the new set of process instructions obtained from this optimization calculation and ready to be issued for execution.

[0109] In the above implementation, a dynamic weight vector reflecting real-time production status is deeply embedded into the construction of the multi-objective optimization model. This transforms the optimization objective from a fixed dogma into an intelligent guide that can dynamically adjust to factors such as yarn quality fluctuations and environmental changes. Within a strict process constraint framework, this method utilizes intelligent optimization algorithms to search for the Pareto optimal frontier and selects the optimal equilibrium point based on a preset strategy. Finally, it outputs a decoded set of optimized process parameters, enhancing the production system's adaptive control capability and overall quality efficiency in the face of complexity, variability, and multi-objective conflicts.

[0110] Reference Figure 5 As one implementation of step S106, the step of concatenating the multidimensional sensor dataset, optimized process parameter set, real-time quality score, and dynamic weight vector to generate the original string includes:

[0111] Step S501: Encode the multidimensional sensor dataset into a binary stream;

[0112] The raw data formats of multidimensional sensor datasets may include unstructured or semi-structured data such as image matrices, spectrum arrays, and time-series waveforms. Recording them directly as text would result in the loss of a significant amount of precision information or structural relationships. Therefore, binary stream encoding is used to preserve these complex data according to their original, underlying byte sequences.

[0113] For example, an image of a yarn surface can be converted into a byte sequence following a specific format (such as RAW or a bitmap with a specific header file), and an array of vibration spectra can be converted into a byte representation of sequentially arranged double-precision floating-point numbers. This encoding method can completely preserve the original precision and structure of the data, avoiding information loss or distortion caused by format conversion, and allowing the original sensing information to be completely reconstructed from the binary stream, providing the most original evidence basis for traceability.

[0114] Step S502: Convert the optimized process parameter set into key-value pair strings;

[0115] The optimized process parameter set (such as spinning speed, draft ratio, twist, etc.) is structured data, with each parameter having a clear name (key) and a calculated value (value). By converting it into key-value pair strings, this format has strong readability and parsability, clearly revealing the specific content of each decision parameter.

[0116] Step S503: Append the dynamic weight vector in the form of a floating-point matrix;

[0117] The dynamic weight vector, generated by the coupled model, reflects the current priority given to each mass dimension (such as diameter, vibration, and feathering). It is appended as a floating-point matrix to maintain its precise numerical characteristics and multidimensional structure as a mathematical vector. Each floating-point number in the weight vector carries a specific physical meaning and relative importance; serializing it in matrix form ensures that its numerical precision and order are accurately preserved.

[0118] Step S504: Combine all data segments in timestamp order to obtain the original string.

[0119] Based on the high-precision timestamps carried or associated with each data segment and generated by a unified clock source, they are concatenated into a long, continuous byte sequence, i.e., a string, in strict chronological order. The timestamp order is fundamental to reconstructing the causal logic of events; for example, sensor data input must precede weights calculated by the correlation model, and only then can optimization parameters be derived. This sequential combination ensures that the generated original string is linear and logically coherent in the time dimension.

[0120] In the above implementation, the discrete and heterogeneous "data islands" (raw signals, decision parameters, evaluation weights) in the production site are systematically integrated into a single data entity with strict temporal logic and complete semantic information. This solution not only solves the standardization problem of multi-source data fusion, but also creates a data object that can unambiguously represent the full state of a specific production moment.

[0121] Reference Figure 6 As a further implementation of the intelligent production control method, after the step of generating the quality traceability block, the method further includes:

[0122] Step S601: Parse the historical sequence of the multidimensional sensor dataset from the quality traceability block to locate the time node of yarn quality mutation;

[0123] Since the data is stored on the blockchain, its authenticity and temporal integrity are cryptographically guaranteed, avoiding distortion of the analytical basis due to accidental data modification or contamination. The positioning logic adopts the idea of ​​time series change point detection, which assumes that in a normal and stable production process, the statistical characteristics of quality scores (such as mean and fluctuation range) should remain relatively stable.

[0124] Specifically, the system divides a continuous time series into multiple adjacent data segments using a sliding window and calculates the differences in statistical characteristics (such as mean and variance) between adjacent windows. The significance of these differences is quantified by calculating statistical measures such as standard deviation offset. When the offset at a window boundary exceeds a preset tolerance threshold (e.g., ±2.5σ based on a historical normal distribution), it indicates a significant shift in statistical characteristics in the production process before and after that point, i.e., a quality mutation. The system marks this time point as a quality mutation node.

[0125] Step S602: Extract the yarn diffraction stripe deformation data and hair gradient value corresponding to the positioning time node, input the pre-constructed defect root cause analysis model, and output the dominant defect type and influencing factor weight.

[0126] The system extracts two key original data that best reflect the internal and external structural state of the yarn at the time before and after the located mutation node: diffraction stripe deformation data (acquired by a laser interferometer, directly reflecting the fiber arrangement density, uniformity and defects inside the yarn) and hairy gradient value (calculated by a machine vision system, characterizing the degree of dispersion and disorder of fibers on the yarn surface). These two types of data describe the yarn state from two physical perspectives: internal structure and surface morphology.

[0127] In this embodiment, the pre-constructed defect root cause analysis model is a multimodal fusion deep learning model (such as a convolutional-attention network). The specific execution steps of the model include: First, using a convolutional neural network to extract local features from the diffraction fringe image, identifying microscopic distortion patterns such as fringe breakage, distortion, and abrupt spacing changes. These patterns correspond to physical defects such as internal yarn breakage, knots, and fiber entanglement. Simultaneously, an attention mechanism is applied to the hairline gradient value sequence, allowing the model to focus on abnormal gradient jumps near abrupt nodes, enhancing the perception of sudden surface problems. Finally, the model fuses and comprehensively judges the extracted internal structural features and enhanced surface morphology features, outputting the dominant defect type (such as "yarn breakage," "hairline aggregation," "coarse and fine details," etc.) and the influence factor weights of various defects at the fully connected layer.

[0128] Step S603: Based on the fiber material characteristic parameters pre-stored in the database, a set of recommended process compensation schemes is matched according to the dominant defect type;

[0129] In some embodiments, the operation of matching the recommended process compensation scheme set includes: loading the draft ratio strengthening scheme set when the dominant defect is yarn breakage; and loading the twist optimization scheme set when the dominant defect is hair aggregation.

[0130] The system retrieves pre-stored fiber material characteristic parameters (including fiber type, moisture regain, strength, elongation, coefficient of friction, etc.) based on the diagnosed dominant defect type. For example, when diagnosed as "hair accumulation," the system will not simply give a general suggestion to "increase twist," but will make a decision based on the moisture regain (moisture content) characteristics of the current yarn raw material. For cotton fibers with high moisture regain, increasing twist combined with reducing ambient humidity may be an effective solution; while for polyester fibers with poor moisture absorption, the focus may be on adjusting spinning tension and traveler configuration.

[0131] Understandably, this parameter-driven matching mechanism essentially establishes a knowledge base mapping between "defect type - material properties - process parameters". The system matches or dynamically derives a set of recommended process compensation schemes from this knowledge base. This set of schemes clarifies which process parameters (such as machine speed, draw ratio, twist, temperature and humidity settings) should be adjusted in what direction to address the currently diagnosed defects occurring in a specific material.

[0132] Step S604: Compare the recommended process compensation scheme set with the optimized process parameter set, generate a model update request data packet, and upload it to the federated learning network.

[0133] The system performs a structured comparison between the generated set of recommended process compensation schemes (representing the "theoretically optimal" or "empirical" adjustment directions derived from the analysis of historical mutation root causes and material properties) and the set of optimized process parameters actually in operation at the time of the mutation (representing the "actually executed" parameters calculated in real time by the intelligent control system at that time). This comparison generates a difference vector that quantifies the deviation between theoretical recommendations and actual execution on key process parameters. For example, the theoretical recommendation might significantly increase twist, while the actual system only makes minor adjustments.

[0134] Subsequently, this discrepancy vector, along with defect type, material properties, and environmental context, is encapsulated into a model update request data packet. This data packet is uploaded to the federated learning network, reporting to other nodes or the central server in the network: "Under certain materials and defect conditions, the local system's optimization decision at that time differs from the theoretically optimal recommendation based on post-analysis."

[0135] Understandably, this step provides federated learning with extremely valuable incremental learning samples with physical semantic annotations, enabling the global model to learn how to adjust its internal parameters (such as the threshold of the decision tree and the weight of the optimization objective) under specific operating conditions so that the real-time optimization output of the local system is closer to the theoretical direction that is proven to be better after the fact, thereby continuously approaching a better control strategy at the group level.

[0136] Step S605: In response to the updated global model parameters issued by the federated learning network, dynamically adjust the dynamic threshold rules of the anomaly decision tree.

[0137] The updated global model parameters incorporate the learning outcomes of all nodes in the network after experiencing similar defects and differences, reflecting the statistically optimal decision-making tendency of the group when facing a certain type of problem. The local system uses these new parameters to dynamically adjust the node threshold rules of the anomaly decision tree.

[0138] For example, in some embodiments, the global model may discover that for a certain type of fiber, when the vibration energy rises slightly in a certain frequency band, even if it does not reach the original threshold, it is very likely to cause yarn breakage. Therefore, the updated parameters will include information for fine-tuning this threshold. After receiving the updated parameters, the local system will immediately reconstruct its decision logic, for example, by slightly lowering the warning threshold for vibration energy or increasing the sensitivity of the rate of change in quality score.

[0139] In the above implementation, the moment of quality mutation is accurately located from the blockchain, a multimodal fusion model is used to diagnose the physical root cause of the defect, and a physically feasible process compensation plan is generated by combining it with a material property knowledge base. By comparing the difference between theoretical compensation and actual execution, knowledge increments are formed to drive the optimization of the federated model. Finally, the collective intelligence obtained from federated learning is fed back to the local real-time control system in the form of dynamically adjusted decision thresholds. This technical solution enables the entire intelligent production system to continuously learn and evolve from past experience, realizing a fundamental improvement in quality control from passive response to proactive prediction, and from isolated decision-making to group collaboration, laying a technical foundation for building an industrial intelligent agent with continuous self-optimization capabilities.

[0140] Reference Figure 7 As a further implementation of the intelligent production control method, after the step of generating the quality traceability block, the method further includes:

[0141] Step S701: Obtain multiple quality traceability blocks generated within a preset historical period from the blockchain network, and extract the quality DNA identifier, optimized process parameter set, and real-time quality score encapsulated therein.

[0142] Specifically, all quality traceability blocks generated within a historical period (such as the past week or month) are traced back and retrieved. Each block permanently stores the key data fingerprint (quality DNA identifier) ​​of the corresponding production moment and its associated core process and result data (optimized process parameter set, real-time quality score). These data are extracted to obtain a historical data sample set with a sufficiently long time span, covering multiple operating conditions, and possessing extremely high reliability.

[0143] Step S702: On-chain verification of the extracted quality DNA identifier, and based on the verified data, association and restoration of the corresponding multidimensional sensor dataset, constructing a local historical verification dataset containing sensor data, process parameters and quality scores.

[0144] Specifically, first, the hash value of the original data extracted from the block is recalculated and compared with the "quality DNA identifier" recorded on the chain. If they match, the verification is successful. This leverages the collision-resistant properties of cryptographic hash functions to ensure that the data used for training is completely consistent with the data initially uploaded to the chain.

[0145] Secondly, based on the mapping relationship between the "quality DNA identifier" and the original production snapshot (this mapping relationship may be maintained in a local database or off-chain storage), the abstract identifier is restored to specific, high-dimensional original sensor data (such as diffraction fringes, vibration spectra, etc.). Finally, the validated sensor data, their corresponding historical process parameters, and the quality scores obtained at that time are aligned and integrated to construct a local historical validation dataset.

[0146] Step S703: Based on the local historical validation dataset, perform the first round of training and optimization on the parameters in the coupled model to generate the first model optimization parameters;

[0147] The training process uses historical "multidimensional sensor datasets" and "fiber material property parameters" as inputs, and historical "real-time quality scores" and "dynamic weight vectors" as training targets (or labels). Backpropagation is performed to adjust the parameters within the coupled model (such as neural network weights, feature extractor parameters, weighted fusion coefficients, etc.). The goal of training is to make the model fit historical data more accurately, that is, to make the scores and weights calculated by the model based on historical sensor signals as close as possible to the historically generated and proven effective scores and weights. Through this round of optimization, the model can better extract features related to the final quality from the complex sensor signals, more accurately assess the state, and generate more reasonable dynamic weights. The first model optimization parameters represent the parts of the coupled model whose internal state has changed after this training, which can be understood as the model's new knowledge.

[0148] Step S704: Based on the data sequences corresponding to the abnormal events in the local historical verification dataset, perform a second round of training and optimization on the node judgment logic in the abnormal decision tree to generate the second model optimization parameters;

[0149] Specifically, data sequences marked as anomalous events or containing sharp quality fluctuations or triggered commands are selected from the local historical verification dataset. Using the time-series features of these event sequences, such as real-time quality scores and vibration energy integral values, along with the global model parameters issued at the time, a second round of training and optimization is performed on the structure of the anomaly decision tree (e.g., node splitting conditions, thresholds). For example, by analyzing historical false alarms (braking without anomalies) and false negatives (not braking despite anomalies), the judgment thresholds of each node in the decision tree are adjusted to make it more sensitive and accurate in identifying true anomalies. The second round of model optimization parameters represents the specific adjustments made to the decision tree's judgment rules after this learning process.

[0150] In some embodiments, the branch logic for generating optimization instructions in the decision tree can also be optimized by learning the pre-state features corresponding to parameter optimization instructions that have successfully avoided failures in the past. The second round of training optimization can adopt algorithms such as reinforcement learning or decision tree pruning and growth.

[0151] Step S705: Load the first model optimization parameters and the second model optimization parameters into the local verification environment, and drive the simulation model in the local verification environment to run, so as to obtain simulated production process data;

[0152] Specifically, applying new, unverified model parameters directly to operating spinning equipment carries high risks, potentially leading to production fluctuations or even equipment damage. Therefore, offline simulation in a local validation environment is necessary. This environment is a software system containing digital simulation models of the equipment, processes, and yarn forming process. Loading the optimized "first model optimization parameters" and "second model optimization parameters" into this environment means driving the simulation with the updated coupled model and anomaly decision tree. By inputting a series of simulated or historical sensor data to drive the simulation model, we can observe how the system makes process adjustment decisions (simulated optimized process parameter set) under the new model logic, as well as the predicted quality results (simulated quality score). The simulated production process data generated in this process provides safe, low-cost, and repeatable experimental data for evaluating the optimization effect.

[0153] Step S706: Calculate the overall performance gain of this model optimization based on the simulated production process data;

[0154] The comprehensive performance index is used to measure the degree of performance improvement before and after model optimization. For example, performance gain may be calculated by comparing the changes in real-time quality score prediction accuracy, early warning time for abnormal events, and the virtual quality improvement brought about by the recommended optimized process parameter set in simulation on the same test dataset between the old and new models. The calculation formula comprehensively considers multiple dimensions such as quality, efficiency, and stability. The calculated comprehensive performance gain is a quantified value score used to objectively judge whether the optimization is effective and to what extent.

[0155] Step S707: When the overall performance gain meets the preset conditions, the first model optimization parameters and the second model optimization parameters are encrypted and encapsulated to generate a local model parameter update package.

[0156] The optimization process involves setting a "preset condition" (such as the overall performance gain must be greater than zero or exceed a certain minimum meaningful threshold). Only when the calculated gain meets this condition is the optimization considered successful and beneficial. Subsequently, the validated "first model optimization parameters" and "second model optimization parameters" are merged and encrypted using encryption techniques (such as asymmetric encryption) to generate a local model parameter update package. This update package represents the model improvement results learned by the local device based on its historical experience and internally validated.

[0157] Step S708: Upload the local model parameter update package to the federated learning network to participate in the periodic aggregation and distribution of global model parameters.

[0158] Each local device uploads its verified optimization experience (i.e., an encrypted "local model parameter update package") to the federated learning network, instead of uploading raw data, fundamentally protecting the privacy of its production data. The network's central server or coordinating node periodically collects update packages from numerous devices and merges these local optimization parameters using secure aggregation algorithms (such as weighted average or federated average algorithms) to generate a new set of "global model parameters" that incorporates the latest and most effective experience from the entire network. This new set of parameters is then distributed to all participating devices, allowing the successful optimization experience of a single device to be securely adopted by all similar devices. This enables the entire production system to evolve collaboratively, jointly improving its ability to handle complex operating conditions and rare faults, and continuously optimizing production quality and efficiency.

[0159] In the above embodiments, intelligent control of textile yarn production is no longer a static, preset program, but an organic entity capable of continuously learning, summarizing, and optimizing from its own historical practices, and absorbing wisdom from the collective experience of peers. This technical solution not only dynamically improves the modeling accuracy of the coupled model for complex quality correlations and the forward-looking judgment capability of the anomaly decision tree for risks, but also establishes a sustainable global knowledge sharing and evolutionary ecosystem under the premise of ensuring data privacy, fundamentally enhancing the overall robustness and long-term adaptability of the production system in the face of unknown disturbances, raw material fluctuations, and process innovations.

[0160] This application also discloses an intelligent production control system for textile yarns.

[0161] A smart production control system for textile yarns, specifically comprising:

[0162] The synchronous sensing module is used to trigger the sensor group deployed on the spinning equipment through a synchronous clock source to collect yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data and yarn surface image data, and generate a timestamp-aligned multidimensional sensing dataset.

[0163] The coupling evaluation module is used to input multidimensional sensor datasets into a pre-built coupling model, combine fiber material characteristic parameters stored in the database, calculate and generate real-time yarn diameter, vibration energy integral value and hair gradient value, and output real-time quality score and dynamic weight vector.

[0164] The intelligent decision-making module is used to receive global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into the pre-built anomaly decision tree, and generate emergency braking instructions or parameter optimization instructions.

[0165] The multi-objective optimization module is used to respond to parameter optimization instructions by inputting the current process parameter set and dynamic weight vector into the multi-objective optimization algorithm to generate an optimized process parameter set.

[0166] The servo control module is used to convert the optimized process parameter set into control signals for the spinning equipment to drive the textile equipment to perform adjustments; the control signals for the spinning equipment include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide.

[0167] The full-chain evidence storage module is used to concatenate multi-dimensional sensor datasets, optimized process parameter sets, real-time quality scores, and dynamic weight vectors to generate an original string, calculate a hash value as a quality DNA identifier, and write it into the blockchain network to generate a quality traceability block.

[0168] The intelligent production control system for textile yarn according to the embodiments of this application can implement any of the above methods, and the specific working process of each module in the system can refer to the corresponding process in the above method embodiments.

[0169] In the several embodiments provided in this application, it should be understood that the provided methods and systems can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the division of a certain module is merely a logical functional division, and in actual implementation there may be other division methods, such as multiple modules can be combined or integrated into another system, or some features can be ignored or not executed.

[0170] This application also discloses a computer device.

[0171] A computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement a method for intelligent production control of textile yarn as described above.

[0172] This application also discloses a computer-readable storage medium.

[0173] A computer-readable storage medium storing a computer program that can be loaded by a processor and executed as described above in any of the intelligent production control methods for textile yarns.

[0174] The computer-readable storage medium can be any tangible medium that contains or stores a program that can be used by or in connection with an instruction execution system, apparatus, or device; the program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.

[0175] The above are all preferred embodiments of this application and are not intended to limit the scope of protection of this application. Any feature disclosed in this specification (including the abstract and drawings) may be replaced by other equivalent or similar features unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is only one example of a series of equivalent or similar features.

Claims

1. A method for intelligent production control of textile yarn, characterized in that, Intelligent production control methods include: By triggering the sensor group deployed on the spinning equipment through a synchronous clock source, yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data, and yarn surface image data are collected to generate a timestamp-aligned multidimensional sensor dataset. The multidimensional sensing dataset is input into a pre-built coupling model, and combined with the fiber material characteristic parameters pre-stored in the database, the real-time yarn diameter, vibration energy integral value and hair gradient value are calculated and generated, and the real-time quality score and dynamic weight vector are output. Receive global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into a pre-constructed anomaly decision tree, and generate emergency braking instructions or parameter optimization instructions; In response to the parameter optimization instruction, the current process parameter set and the dynamic weight vector are input into a multi-objective optimization algorithm to generate an optimized process parameter set; The optimized process parameter set is converted into spinning equipment control signals to drive the textile equipment to perform adjustments; wherein, the spinning equipment control signals include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide; The multidimensional sensor dataset, optimized process parameter set, real-time quality score, and dynamic weight vector are concatenated to generate an original string. The hash value is calculated as a quality DNA identifier and written into the blockchain network to generate a quality traceability block.

2. The intelligent production control method for textile yarn according to claim 1, characterized in that, The steps of inputting the multidimensional sensing dataset into a pre-constructed coupled model, combining it with fiber material characteristic parameters pre-stored in the database, calculating and generating real-time yarn diameter, vibration energy integral value, and hairiness gradient value, and outputting real-time quality score and dynamic weight vector include: The multidimensional sensor dataset was analyzed to separate yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data, and yarn surface image data. The yarn diffraction stripe deformation data is subjected to Gaussian filtering for noise reduction, and the real-time yarn diameter is calculated based on the noise-reduced data using a morphological inversion algorithm. Perform a fast Fourier transform on the triaxial vibration spectrum data, and perform spectral energy integration within a preset frequency band to generate vibration energy integral values; Edge detection and contour analysis are performed on the yarn surface image data to extract the hair distribution feature vector and calculate the hair gradient value; The real-time diameter of the yarn, the integral value of vibration energy, the hair gradient value, and the ambient temperature and humidity data are input into the pre-constructed coupled model, while the fiber material characteristic parameters pre-stored in the database are loaded. The fiber moisture regain rate is calculated based on the environmental temperature and humidity data. The humidity influence factor is determined by the material moisture absorption characteristic curve in the fiber material characteristic parameters. The humidity influence factor is used as the dynamic environmental calibration coefficient. The weighted fusion layer of the coupled model, combined with the dynamic environment calibration coefficients, outputs a real-time quality score and a dynamic weight vector.

3. The intelligent production control method for textile yarn according to claim 1, characterized in that, The steps of receiving global model parameters periodically distributed by the federated learning network, inputting the real-time quality score and global model parameters into a pre-constructed anomaly decision tree, and generating emergency braking instructions or parameter optimization instructions include: The encrypted data packets transmitted in the federated learning network are parsed, and the local model parameters of each spindle terminal are obtained by decryption using a preset key. The weighted average operation is then performed on the decrypted local model parameters to generate global model parameters. The real-time quality score and global model parameters are input into a pre-constructed anomaly decision tree, and a judgment result is generated based on dynamic threshold rules. When the judgment result triggers an abnormal condition, an emergency braking command is generated; When the judgment result does not trigger the abnormal condition, the parameter optimization instruction is output to the multi-objective optimization module.

4. The intelligent production control method for textile yarn according to claim 3, characterized in that, In response to the parameter optimization instruction, the step of inputting the current process parameter set and the dynamic weight vector into a multi-objective optimization algorithm to generate an optimized process parameter set includes: In response to the parameter optimization command, the currently running set of process parameters and the dynamic weight vector output by the coupled model are read. The process parameter set is encoded into the initial solution space of a multi-objective optimization algorithm, and a multi-objective fitness function is constructed based on the dynamic weight vector. Under preset process constraints, perform iterative calculations using a multi-objective optimization algorithm to generate a Pareto optimal solution set; The target solution is selected from the Pareto optimal solution set using a preset strategy, and then decoded to generate an optimized process parameter set.

5. The intelligent production control method for textile yarn according to claim 1, characterized in that, The steps of concatenating the multidimensional sensor dataset, optimized process parameter set, real-time quality score, and dynamic weight vector to generate the original string include: The multidimensional sensor dataset is encoded into a binary stream; Convert the optimized process parameter set into key-value pair strings; The dynamic weight vector is appended in the form of a floating-point matrix; Combine all data segments in timestamp order to obtain the original string.

6. The intelligent production control method for textile yarn according to claim 5, characterized in that, Following the step of generating the quality traceability block, the following is also included: The historical sequence of the multidimensional sensor dataset is parsed from the quality traceability block to locate the time node of yarn quality mutation; Extract the yarn diffraction stripe deformation data and hair gradient value corresponding to the time node of the location, input them into the pre-constructed defect root cause analysis model, and output the dominant defect type and the weight of the influencing factor. Based on the fiber material characteristic parameters pre-stored in the database, a set of recommended process compensation schemes is matched according to the dominant defect type; The recommended process compensation scheme set is compared with the optimized process parameter set, and a model update request data packet is generated and uploaded to the federated learning network. In response to the updated global model parameters issued by the federated learning network, the dynamic threshold rules of the anomaly decision tree are dynamically adjusted.

7. A method for intelligent production control of textile yarn according to any one of claims 1 to 6, characterized in that, Following the step of generating the quality traceability block, the following is also included: Multiple quality traceability blocks generated within a preset historical period are obtained from the blockchain network, and the quality DNA identifier, the optimized process parameter set, and the real-time quality score encapsulated therein are extracted. The extracted quality DNA identifiers are verified on-chain, and based on the verified data, the corresponding multidimensional sensor datasets are associated and restored to construct a local historical verification dataset containing sensor data, process parameters, and quality scores. Based on the local historical verification dataset, the parameters in the coupled model are trained and optimized in the first round to generate the first model optimization parameters. Based on the data sequences corresponding to the abnormal events in the local historical verification dataset, the node judgment logic in the abnormal decision tree is trained and optimized in the second round to generate the second model optimization parameters. The first model optimization parameters and the second model optimization parameters are loaded into the local verification environment, and the simulation model in the local verification environment is driven to run to obtain simulated production process data. Based on the simulated production process data, calculate the overall efficiency gain of this model optimization; When the overall performance gain meets the preset conditions, the first model optimization parameters and the second model optimization parameters are encrypted and encapsulated to generate a local model parameter update package; The local model parameter update package is uploaded to the federated learning network to participate in the periodic aggregation and distribution of the global model parameters.

8. A smart production control system for textile yarn, characterized in that, Intelligent production control systems include: The synchronous sensing module is used to trigger the sensor group deployed on the spinning equipment through a synchronous clock source to collect yarn diffraction stripe deformation data, triaxial vibration spectrum data, ambient temperature and humidity data and yarn surface image data, and generate a timestamp-aligned multidimensional sensing dataset. The coupling evaluation module is used to input the multidimensional sensing dataset into a pre-built coupling model, combine it with the fiber material characteristic parameters pre-stored in the database, calculate and generate the real-time yarn diameter, vibration energy integral value and hair gradient value, and output the real-time quality score and dynamic weight vector. The intelligent decision-making module is used to receive the global model parameters periodically issued by the federated learning network, input the real-time quality score and global model parameters into the pre-built anomaly decision tree, and generate emergency braking instructions or parameter optimization instructions. The multi-objective optimization module is used to respond to the parameter optimization command by inputting the current process parameter set and the dynamic weight vector into the multi-objective optimization algorithm to generate an optimized process parameter set; A servo control module is used to convert the optimized process parameter set into spinning equipment control signals to drive the textile equipment to perform adjustments; wherein, the spinning equipment control signals include the speed control signal of the variable frequency motor and the displacement control signal of the pneumatic yarn guide; The full-chain evidence storage module is used to concatenate the multi-dimensional sensor dataset, optimized process parameter set, real-time quality score and dynamic weight vector to generate an original string, calculate the hash value as a quality DNA identifier, and write it into the blockchain network to generate a quality traceability block.

9. A computer device, characterized in that: The method includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer program is stored that can be loaded by a processor and executed as described in any one of claims 1 to 7.

Images