Production quality management method of high-elastic wrinkle-resistant and pilling-resistant fabric

Abstract

This invention discloses a production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabrics, belonging to the field of quality management technology. Specifically, it includes: extracting equipment fluctuation patterns and fabric response characteristics from historical logs of the finishing process to establish a fluctuation-response correlation trajectory library; constructing a fabric equivalent parameter inversion model and generating a raw material intrinsic parameter space; collecting real-time features of the leading roll to calculate real-time raw material intrinsic parameter vectors; generating a deviation parameter set, inputting it into a process compensation decision model to output process setting parameters for each process; performing high-frequency monitoring of each process to obtain dynamic deviation time sequences; constructing a joint attribution mapping relationship between raw material intrinsic parameters and the dynamic deviation of each process relative to local quality risks, identifying quality differentiation within the roll, and outputting risk judgment and sampling inspection lists. This invention achieves closed-loop quality management from raw material perception to full-process process compensation and precise sampling inspection, elevating quality control from batch-level passive inspection to in-roll segment-level proactive prediction.

Description

Technical Field

[0001] This invention relates to the field of quality management technology, specifically to a production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabrics. Background Technology

[0002] High-elasticity, wrinkle-resistant, and pilling-resistant fabrics have become an important development direction in the textile industry. By introducing functional components in fiber modification, fabric structure design, or finishing processes, fabrics can be endowed with single or combined functions such as radiant cooling, moisture wicking, active warmth retention, and antibacterial and deodorizing properties. For example, in the field of radiative cooling, existing research has shown that by growing SiO2 micro-nano particles in situ on the surface of cotton fabrics and using microsphere particle size control to construct micro-nano composite structures with high solar reflectivity and high infrared emissivity, passive radiative cooling of the fabric under direct sunlight is achieved. In the field of moisture wicking, the capillary moisture-wicking channels of the fabric are optimized by designing the fiber cross-section structure or modifying the hydrophilicity, combined with a gradient pore size unidirectional moisture-wicking layer structure, to achieve rapid unidirectional conduction of sweat from the inside to the outside. In the field of active warmth and antibacterial properties, antibacterial carbon quantum dots with photothermal activity are prepared using biomass materials such as rosin and lignin as carbon sources. These are then added to PA6 matrix to prepare photothermal active functional masterbatches, which are then melt-spun and knitted to obtain high-warmth fabrics with both photothermal conversion and antibacterial functions.

[0003] However, the final performance of the fabric imparted by the aforementioned functional modification methods is highly dependent on the precise control of the heat setting process. Since the fabric contains elastic components such as spandex, it requires a final heat setting process to eliminate internal stress accumulated in previous processing, stabilize width and weight, and impart the target elasticity and wrinkle resistance. The control level of process parameters such as temperature, tension, and machine speed in the heat setting process directly determines whether the finished fabric's width, weight, elastic recovery rate, and pilling resistance meet the standards.

[0004] In the actual operation of the heat setting process, the equipment process parameters are not constant but always exhibit a certain degree of dynamic fluctuation. Even if these fluctuations do not exceed the preset acceptable range of the process parameters, the different fluctuation amplitudes, frequencies, and waveform characteristics have significantly different effects on the internal microstructure of the fabric. In other words, the differences in the patterns of equipment fluctuation parameters will directly lead to differentiation in the thermo-mechanical coupling process experienced by the fabric, resulting in inconsistent quality risk levels in different sections within the same batch or even the same roll of fabric. Currently, quality control for the heat setting process mainly relies on two methods: process parameter threshold alarms and final quality sampling inspection. Process parameter threshold alarms, by setting upper and lower limits for parameters such as temperature and tension, trigger alarms only when the parameters exceed the acceptable range. Essentially, they treat all fluctuations within the acceptable range of process parameters as equivalent, ignoring the quality impact of the fluctuation pattern itself as an independent variable. Final quality sampling inspection, on the other hand, involves taking samples from the batch after heat setting for offline testing. Its detection effectiveness is highly dependent on the selection of the sampling location. When the quality differentiation within the roll exhibits localized dispersion characteristics, random sampling is prone to missing detections. Summary of the Invention

[0005] The purpose of this invention is to provide a production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabrics, solving the following technical problems:

[0006] Existing heat setting quality control methods for high-elasticity, wrinkle-resistant, and pilling-resistant fabrics lack the ability to model and analyze the relationship between equipment fluctuation patterns and fabric quality, and lack the ability to identify the risk of quality differentiation within the roll in real time. This makes it difficult to meet the requirements of high-quality stability for mass production while ensuring the effectiveness of testing.

[0007] The objective of this invention can be achieved through the following technical solutions: A production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabric includes the following steps: S1: Extract historical fluctuation pattern features and historical response features from the historical processing logs of the fabric finishing process, and establish a fluctuation-response correlation trajectory library; S2: Perform inverse modeling on the wave-response correlation trajectory library, construct a fabric equivalent parameter inversion model with wave pattern as excitation and response feature as observation, and solve the equivalent raw material intrinsic parameters corresponding to each wave-response correlation record based on the fabric equivalent parameter inversion model, and aggregate to form a raw material intrinsic parameter space; S3: When the lead roll of the current processing batch passes through the shaping process, collect the real-time fluctuation mode characteristics and real-time response characteristics and input them into the fabric equivalent parameter inversion model to calculate the real-time raw material intrinsic parameter vector of the fabric corresponding to the current processing batch. S4: Perform deviation analysis on the real-time raw material intrinsic parameter vector in the raw material intrinsic parameter space, generate a raw material deviation parameter set and input it into the pre-constructed process compensation decision model, and output the process setting parameter set for each process of the current batch of subsequent unprocessed fabric. S5: When the subsequent unprocessed fabrics of the current batch pass through each process in sequence according to the process setting parameter set, high-frequency equipment fluctuation monitoring is performed on each process to obtain the dynamic deviation time sequence of the actual fluctuation of each process relative to the corresponding process setting parameter. S6: Construct a joint attribution mapping relationship between the intrinsic parameter vector of raw materials and the dynamic deviation time sequence of each process relative to the local quality risk of the fabric. Through the joint attribution mapping relationship, identify the quality differentiation within the roll during each process, output the quality risk judgment result of each process, and determine the fabric to be sampled.

[0008] As a further aspect of the present invention: the specific process of establishing the wave-response correlation trajectory library in S1 is as follows: Obtain historical processing logs for the fabric finishing process. These logs include time series of equipment operating parameters collected during the finishing process for each historical batch and a sequence of fabric physical state parameters detected at the finishing process exit for each roll of fabric. The equipment operating parameters include temperature, tension, and machine speed, while the fabric physical state parameters include width and weight. The time series of the equipment operating parameters is segmented by a sliding window, and the fluctuation pattern features of the equipment operating parameters are extracted in each sliding window. The fluctuation pattern features include the fluctuation amplitude, fluctuation frequency and waveform feature parameters of each equipment operating parameter in the sliding window. The physical state parameter sequence of the fabric is grouped by roll identifier. The width and weight values ​​of each roll of fabric at the exit of the setting process are calculated relative to the initial width and weight values ​​at the entry of the setting process, and these are used as the response characteristics of the roll of fabric. The fluctuation pattern characteristics and response characteristics of each roll of fabric in the same historical processing batch are aligned and bound along the time axis to generate fluctuation-response association records; the fluctuation-response association records of all historical processing batches are aggregated, and the fluctuation-response association trajectory library is established with the roll identifier as the index.

[0009] As a further aspect of the present invention: in S2, the specific generation process of the fabric equivalent parameter inversion model is as follows: Laboratory test results of raw material samples from each roll of fabric in historical production batches are obtained. The measured values ​​of the intrinsic parameters of the raw materials used in each roll of fabric are extracted. These intrinsic parameters include the measured values ​​of spandex heat shrinkage rate, elastic modulus, and friction coefficient. Based on the roll identifier carried by each fluctuation-response correlation record in the fluctuation-response correlation trajectory library, the measured values ​​of the intrinsic parameters of the raw materials, the fluctuation pattern characteristics, and the response characteristics of the same roll of fabric are associated and bound to form a raw material intrinsic parameter-fluctuation-response correlation record. The raw material intrinsic parameter-fluctuation-response correlation records of all historical processing batches are aggregated. Using the fluctuation pattern characteristics and response characteristics in each correlation record as input data and the corresponding measured values ​​of the intrinsic parameters of the raw materials as output data, the fabric equivalent parameter inversion model is trained.

[0010] As a further aspect of the present invention: in step S2, the specific process for generating the intrinsic parameter space of the raw materials is as follows: Using each associated record in the fluctuation-response correlation trajectory library as input, the fabric equivalent parameter inversion model is invoked. After forward calculation by the fabric equivalent parameter inversion model, the equivalent raw material intrinsic parameters corresponding to the associated record are output. All associated records in the fluctuation-response correlation trajectory library are traversed to obtain the equivalent raw material intrinsic parameters corresponding to each roll of fabric in each historical processing batch, forming an equivalent raw material intrinsic parameter set. Statistical analysis is performed on the equivalent raw material intrinsic parameter set to calculate the mean vector and covariance matrix of the equivalent raw material intrinsic parameters in each dimension. Using the mean vector and the covariance matrix as the distribution parameters of the raw material intrinsic parameter space, a raw material intrinsic parameter space containing the equivalent raw material intrinsic parameter set and its distribution parameters is constructed.

[0011] As a further aspect of the present invention: the specific process for generating the raw material deviation parameter set in step S4 is as follows: Obtain the real-time raw material intrinsic parameter vector, as well as the mean vector and covariance matrix contained in the raw material intrinsic parameter space; calculate the component differences between the real-time raw material intrinsic parameter vector and the mean vector in each intrinsic parameter dimension to obtain the absolute deviation of each dimension; use the covariance matrix to calculate the Mahalanobis distance of the absolute deviation of each dimension to obtain the relative deviation of each dimension; use the relative deviation of each dimension as elements to obtain the raw material deviation parameter set.

[0012] As a further aspect of the present invention: the specific construction process of the process compensation decision model in S4 is as follows: Obtain the raw material deviation parameter set for each roll of fabric in historical production batches, as well as the actual process parameter records for each roll of fabric in each process and the corresponding finished product quality inspection results; the finished product quality inspection results include elastic recovery rate and anti-pilling grade; using the raw material deviation parameter set of the same roll of fabric as input features, the actual process parameter records for the roll of fabric in each process as output labels, and the finished product quality inspection results meeting the preset qualified standards as screening conditions, construct a process compensation training sample set; An initial mapping model is constructed, taking the raw material deviation parameter set of each sample in the process compensation training sample set as input, and constructing a loss function based on the deviation between the predicted process parameters of each process output by the initial mapping model and the actual process parameter records of the corresponding samples. The model parameters of the initial mapping model are iteratively updated based on the loss function until the preset convergence condition is met, thereby obtaining the process compensation decision model.

[0013] As a further aspect of the present invention: in step S5, the specific process for obtaining the dynamic deviation timing is as follows: Obtain the process setting parameter set for each process of the current batch of unprocessed fabric. The process setting parameter set includes the process control parameter setting value corresponding to each process. When the unprocessed fabric passes through each process in sequence according to the process setting parameter set, collect the time series of equipment operation parameters corresponding to each process at a preset sampling frequency. Calculate the deviation between the time series of equipment operation parameters for each process and the process control parameter setting value of the corresponding process at each sampling point to obtain the equipment operation parameter deviation sequence for each process, which serves as the dynamic deviation time series for that process.

[0014] As a further aspect of the present invention: in step S6, the specific process for generating the joint attribution mapping relationship is as follows: Obtain the intrinsic raw material parameter vector, dynamic deviation time sequence of each process, and quality inspection results of each section within the roll of fabric in each historical processing batch of the fluctuation-response correlation trajectory library; use the intrinsic raw material parameter vector and dynamic deviation time sequence of each process of the same section of the same roll of fabric as input features, and use the existence of local quality defects in the quality inspection results of the section as output labels to construct a joint attribution training sample set; An initial attribution model is constructed, taking the intrinsic parameter vectors of raw materials and the dynamic deviation time series of each process of each sample in the joint attribution training sample set as inputs. A loss function is constructed based on the deviation between the quality risk prediction result output by the initial attribution model and the output label of the corresponding sample. The model parameters of the initial attribution model are iteratively updated based on the loss function until the preset convergence condition is met, thereby obtaining the joint attribution mapping relationship.

[0015] As a further aspect of the present invention: In step S6, the process of outputting the quality risk assessment results of each process and determining the specific process of sampling the fabric is as follows: Obtain the real-time intrinsic parameter vector of the raw material for the current batch of fabric, and the dynamic deviation time sequence of the subsequent unprocessed fabric in the current batch at each process; align the real-time intrinsic parameter vector of the raw material and the dynamic deviation time sequence of each process according to fabric segments to obtain the joint input features of each segment in each process; The joint input features of each segment in each process are input into the joint attribution mapping relationship. After forward calculation by the joint attribution mapping relationship, the local quality risk prediction value of each segment in each process is output. The local quality risk prediction value of each segment in each process is compared with a preset risk threshold. Segments whose local quality risk prediction value exceeds the preset risk threshold are marked as risk segments, and the quality risk judgment result of each process is generated. The roll identifier and segment position information of the risk segments in each process are extracted to generate a sample inspection fabric list.

[0016] The beneficial effects of this invention are: 1) This invention extracts the fluctuation amplitude, frequency, and waveform characteristics of equipment operating parameters from historical finishing processing logs, and aligns and binds them with the changes in fabric width and weight at the fabric exit along the time axis to establish a fluctuation-response correlation trajectory library. Then, using fluctuation mode characteristics and response characteristics as inputs and measured values ​​of intrinsic parameters of the fabric raw material, such as the spandex heat shrinkage rate and elastic modulus, as outputs, a fabric equivalent parameter inversion model is trained. It is understandable that fluctuations in equipment parameters are not meaningless noise; different fluctuation amplitudes, frequencies, and waveforms correspond to drastically different thermo-mechanical coupling processes experienced by the fabric within the finishing machine. These differences in process will manifest as final changes in fabric width and weight. The degree of width and weight change is essentially determined by the inherent properties of the raw material itself, such as heat shrinkage characteristics and elastic modulus. The same temperature fluctuation, acting on spandex fabrics with high and low heat shrinkage rates, will result in drastically different final dimensional responses. This invention establishes a layer-by-layer mapping from fluctuation patterns to physical responses and then to the intrinsic parameters of raw materials. This enables the model to infer the inherent characteristics of raw materials that cannot be directly measured online from observable equipment fluctuations and macroscopic changes in fabrics. It establishes a quantitative correlation between process fluctuations and raw material states, expanding the assessment of fluctuations from a single criterion of whether they exceed limits to the analysis of the differences in raw material characteristics associated with fluctuations. This provides a physical basis for differentially judging the degree of influence of fluctuations.

[0017] 2) This invention performs inverse modeling and parameter decoupling on the fluctuation-response correlation trajectory library, extracting feature information related to fabric quality differentiation from equipment fluctuation patterns, and incorporating the fluctuation amplitude, frequency, and waveform characteristics of parameters such as temperature, tension, and speed into a unified analysis framework. It is understood that fluctuations in equipment process parameters within the acceptable range are not all equally valid; different combinations of fluctuation amplitude, frequency, and waveform have significantly different effects on the thermo-mechanical coupling of the fabric's internal microstructure. For example, the impact of rapid recovery after instantaneous overheating versus prolonged low-temperature deviation on residual stress in the fabric is drastically different. High-frequency equipment fluctuation monitoring is performed during each processing step to obtain the dynamic deviation time series of the actual equipment operating state relative to the process setpoint, fully recording the actual thermo-mechanical coupling process experienced by the fabric during processing. Furthermore, a joint attribution mapping relationship is constructed between the intrinsic parameter vector of the raw materials and the dynamic deviation time series of each process relative to the local quality risk of the fabric, quantitatively modeling the correlation between raw material characteristics, equipment fluctuation patterns, and local quality risk of the fabric. Therefore, the hidden quality damage caused by different fluctuation patterns within the qualified range to different sections of the fabric in the roll can be effectively identified, solving the blind spot in the existing technology where "the process parameters are not exceeded throughout the process but the quality within the roll is inconsistent".

[0018] 3) This invention aligns the real-time intrinsic parameter vector of raw materials with the dynamic deviation time sequence of each process according to fabric segments. It inputs a joint attribution mapping relationship to perform segment-by-segment risk prediction for each segment, marking segments with predicted values ​​exceeding a preset risk threshold as risk segments. A list of sampled fabrics is generated based on the roll identifier and segment location information of the risk segments. It is understandable that during the shaping and processing of high-elasticity, wrinkle-resistant, and pilling-resistant fabrics, even if the equipment operating parameters do not exceed the process qualification range throughout the entire process, different segments within the same roll of fabric experience differences in the amplitude, frequency, and waveform characteristics of fluctuations in parameters such as temperature, tension, and machine speed. This results in different thermal-mechanical coupling processes experienced by each segment, leading to a differentiation in the quality levels of different segments within the roll. Some segments have uniform residual stress distribution and meet elastic recovery rate standards, while other segments, although meeting width and weight requirements, have hidden quality defects. Traditional random sampling methods draw samples based on a fixed proportion, and the sampling location is independent of the actual distribution of risk segments, making it easy to miss high-risk segments while a large number of normal segments are invalidally inspected. This invention upgrades traditional post-event random sampling to risk-oriented precision sampling based on process awareness. This allows quality inspection resources to be concentrated on high-risk fabric sections identified through joint attribution mapping relationships. Under the same sampling workload, it significantly increases the detection probability of local quality defects within the roll, while avoiding excessive sampling of normal sections. This effectively alleviates the contradiction between the allocation of sampling resources and the distribution of quality risks in the prior art. Attached Figure Description

[0019] The invention will now be further described with reference to the accompanying drawings.

[0020] Figure 1 This is a schematic diagram of the production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] Please see Figure 1 As shown, this invention is a production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabrics, comprising the following steps: S1: Extract historical fluctuation pattern features and historical response features from the historical processing logs of the fabric finishing process, and establish a fluctuation-response correlation trajectory library; Detailed historical processing logs of the fabric setting process are obtained from the Manufacturing Execution System (MES). These logs contain two core data components: first, a time series of equipment operating parameters collected and stored by the equipment's programmable logic controller (PLC) at a second-level frequency during the setting process for each historical batch, specifically including the temperature of each zone in the setting oven, the tension of the needle plate chain or overfeed rollers, and the speed of the main drive; second, a sequence of physical state parameters of the fabric continuously detected and recorded by an online visual inspection system at the exit of the setting process for each roll, specifically including width and weight. The online visual inspection system employs a high-resolution line scan camera array, with the cameras horizontally covering the entire fabric width and the vertical scanning frequency synchronized with the machine speed, ensuring at least 100 line scan images are obtained per meter of fabric. The width value is calculated in real-time using an image edge detection algorithm, and the weight value is estimated in real-time using a grayscale-weight calibration model.

[0023] Feature engineering was performed on the time series of equipment operating parameters. A sliding window with a width of 30 seconds and a step size of 10 seconds was used to segment the signal sequences of temperature, tension, and vehicle speed. Within each sliding window, three sets of fluctuation mode features characterizing the instantaneous changes in equipment state were calculated: the first set was the fluctuation amplitude, obtained by calculating the standard deviation of the signal sequence; when a significant trend term was present, first-order differencing was performed to eliminate trend interference. The second set was the fluctuation frequency, extracted by performing a Fast Fourier Transform on the signal sequence within the window, extracting the frequency component with the largest power spectral density amplitude as the dominant frequency. The third set was waveform feature parameters, constructed by calculating the kurtosis and skewness values ​​of the signal sequence; the kurtosis value reflects the steepness of the fluctuation peak shape, and the skewness value reflects the degree of asymmetry of the fluctuation relative to the mean. The fluctuation amplitude, fluctuation frequency, and waveform feature parameters together constitute the fluctuation mode feature vector of the sliding window.

[0024] The system extracts response features from the sequence of physical state parameters of the fabric. Data is grouped by roll identifier. For any roll of fabric, the system retrieves the initial width and initial weight values ​​recorded by the previous process detection system at the entry point of the finishing process. The system then calculates the arithmetic mean of all width values ​​detected at the exit of the finishing process to obtain the exit average width value, and the arithmetic mean of all weight values ​​to obtain the exit average weight value. Subtracting the initial width value from the exit average width value yields the width change of the fabric roll; subtracting the initial weight value from the exit average weight value yields the weight change of the fabric roll. These width and weight changes together constitute the response feature vector of the fabric roll.

[0025] The fluctuation pattern features and response features of each roll of fabric in the same historical processing batch are aligned and bound along the time axis. Specifically, for each roll of fabric, the start and end times of its finishing process are extracted. The arithmetic mean of the fluctuation pattern feature vectors of all sliding windows within this time period is taken to obtain the average fluctuation pattern feature vector of that roll of fabric. This average fluctuation pattern feature vector is then paired with the response feature vector of that roll of fabric to generate a fluctuation-response association record. All association records from all historical processing batches are aggregated, indexed using a unique roll identifier as the primary key. The fluctuation pattern feature vector and response feature vector in each record are structured and stored in a fixed format to establish a systematic fluctuation-response association trajectory library.

[0026] S2: Perform inverse modeling on the wave-response correlation trajectory library, construct a fabric equivalent parameter inversion model with wave pattern as excitation and response feature as observation, and solve the equivalent raw material intrinsic parameters corresponding to each wave-response correlation record based on the fabric equivalent parameter inversion model, and aggregate to form a raw material intrinsic parameter space; Offline laboratory testing results were obtained for raw material samples retained from each roll of fabric in the aforementioned historical production batches. The offline laboratory testing was conducted according to relevant textile industry testing standards, specifically including: spandex heat shrinkage rate tested at 180℃ using the dry heat shrinkage method according to FZ / T50004 standard; elastic modulus tested using the strip method on a universal testing machine according to GB / T 3923.1 standard, measuring the tensile stress-strain curve and extracting the linear slope of the initial modulus range; and coefficient of friction tested using a friction coefficient meter according to FZ / T 01054 standard, measuring the dynamic friction coefficient between the fabric surface and a standard friction body. These three test results provide each roll of fabric with measured values ​​for spandex heat shrinkage rate, elastic modulus, and coefficient of friction, constituting a vector of measured intrinsic parameters of the raw material for that roll of fabric.

[0027] Using the roll identifier as the association key, the measured vectors of the raw material intrinsic parameters, the fluctuation mode feature vector, and the response feature vector are linked and bound to form a raw material intrinsic parameter-fluctuation-response association record. This association record from all historical processing batches is aggregated to form a supervised learning training set. Using the concatenated vector of the fluctuation mode feature vector and the response feature vector of each association record in this training set as input data, and the corresponding measured vector of the raw material intrinsic parameters as the output label, a multilayer perceptron deep neural network is trained as a fabric equivalent parameter inversion model. Specifically, the network structure is as follows: the number of nodes in the input layer equals the sum of the dimensions of the fluctuation mode feature vector and the response feature vector; the output layer contains three nodes corresponding to the predicted values ​​of spandex heat shrinkage rate, elastic modulus, and friction coefficient, respectively; and five fully connected hidden layers are contained in the middle, each containing 128 neural nodes. The ReLU nonlinear activation function is used between the hidden layers. The training loss function uses mean squared error loss, and the optimizer is the Adam optimizer. The initial learning rate is set to 0.001, and the training iterations are 200 epochs. In each epoch, the loss value is calculated on the validation set. When the validation set loss value no longer decreases for 20 consecutive epochs, an early stopping mechanism is triggered, and the model weight parameters with the lowest validation set loss value are saved as the final model. This final model is the fabric equivalent parameter inversion model, which essentially learns a complex nonlinear inverse mapping relationship between observable external stimulus-response patterns and intrinsic parameters of raw materials that cannot be directly and continuously measured.

[0028] The fluctuation pattern feature vector and response feature vector of each fluctuation-response correlation record stored in the fluctuation-response correlation trajectory library are extracted, concatenated in the same way as during training, and then sequentially input into the fabric equivalent parameter inversion model that has been trained. Through forward calculation of the model, a three-dimensional equivalent raw material intrinsic parameter vector is output for each correlation record. All correlation records in the fluctuation-response correlation trajectory library are traversed to obtain the equivalent raw material intrinsic parameter vectors corresponding to each roll of fabric in each historical processing batch, forming an equivalent raw material intrinsic parameter set. Multivariate statistical analysis is performed on this equivalent raw material intrinsic parameter set to calculate the mean vector and covariance matrix of the set in three-dimensional space, namely the spandex heat shrinkage rate dimension, the elastic modulus dimension, and the friction coefficient dimension. The three components of the mean vector are, respectively, the arithmetic mean of all equivalent spandex heat shrinkage rate values, the arithmetic mean of all equivalent elastic modulus values, and the arithmetic mean of all equivalent friction coefficient values ​​within the set; the covariance matrix is ​​a three-dimensional square matrix, where the element in the i-th row and j-th column is the covariance between the i-th dimension parameter and the j-th dimension parameter within the set. Using the calculated mean vector and covariance matrix as distribution parameters in the raw material intrinsic parameter space, a three-layer data structure is constructed, comprising an equivalent set of raw material intrinsic parameters, a mean vector, and a covariance matrix. This space not only stores the discrete distribution fingerprints of historical raw materials but also comprehensively describes the multivariate normal distribution statistical law followed by the historical raw materials as a whole through the distribution parameters, providing a precise quantitative statistical benchmark for deviation analysis of subsequent new batches of raw materials.

[0029] S3: When the lead roll of the current processing batch passes through the shaping process, collect the real-time fluctuation mode characteristics and real-time response characteristics and input them into the fabric equivalent parameter inversion model to calculate the real-time raw material intrinsic parameter vector of the fabric corresponding to the current processing batch. After a new production batch is started, the first 100 meters of the fabric roll in that batch is used as a lead roll, guiding it through the stenter. The system collects real-time data on the stenter's operating parameters and the physical state parameters of the fabric at the exit during the lead roll's passage. The operating parameter time series is segmented into sliding windows and features extracted in the same way as in S1, with a window width of 30 seconds and a step size of 10 seconds. The arithmetic mean of the fluctuation pattern feature vectors of all sliding windows during the lead roll's passage is taken to generate the real-time fluctuation pattern feature vector for that batch. The physical state parameters of the fabric at the exit are calculated in the same way as in S1, by calculating the average width and average weight of the lead roll at the stenter exit. These values ​​are then differed from the initial width and initial weight recorded by the system at the lead roll entrance to obtain the width change and weight change, generating the real-time response feature vector for that batch.

[0030] The real-time fluctuation pattern feature vector and the real-time response feature vector are concatenated and input into the fabric equivalent parameter inversion model trained in S2. The model performs one forward inference calculation, which involves multiplying the input vector sequentially by the weight matrices of each hidden layer of the network and transforming it through an activation function. Finally, a three-dimensional vector is output at the output layer. The three components of this three-dimensional vector are, in order, the predicted real-time spandex heat shrinkage rate, the predicted real-time elastic modulus, and the predicted real-time friction coefficient of the fabric for this batch. This three-dimensional vector is the real-time intrinsic parameter vector of the raw material for the fabric corresponding to the current processing batch.

[0031] S4: Perform deviation analysis on the real-time raw material intrinsic parameter vector in the raw material intrinsic parameter space, generate a raw material deviation parameter set and input it into the pre-constructed process compensation decision model, and output the process setting parameter set for each process of the current batch of subsequent unprocessed fabric. Obtain the real-time intrinsic parameter vector of the raw material output from S3, and the mean vector and covariance matrix contained in the intrinsic parameter space of the raw material in S2. Perform deviation analysis on the real-time intrinsic parameter vector of the raw material. First, calculate the component differences between the real-time intrinsic parameter vector and the mean vector in each intrinsic parameter dimension. The component difference is the real-time value minus the mean, and the deviations of spandex heat shrinkage rate, elastic modulus, and friction coefficient are obtained in sequence. These three deviation values ​​constitute the absolute deviation vector. To eliminate the interference of dimensional differences and correlations between parameter dimensions on the deviation measurement, Mahalanobis distance is used to calculate the relative deviation of each dimension. The specific calculation method is as follows: multiply the absolute deviation vector by the inverse of the covariance matrix, and then multiply by the transpose of the absolute deviation vector to obtain a scalar value, which is the overall Mahalanobis distance. Distribute the overall Mahalanobis distance according to the proportion of the absolute value of each dimension's absolute deviation component to the sum of the absolute values ​​of all components of the absolute deviation vector. The proportion allocated to each dimension is the relative deviation of each dimension. The raw material deviation parameter set is constructed using the relative deviation of each dimension as elements. This deviation parameter set quantifies the degree of specificity of the new batch of raw materials relative to the historical total of raw materials in a statistical sense.

[0032] The raw material deviation parameter set is input into a pre-built process compensation decision model. The pre-building process of this model is as follows: The calculation method for the raw material deviation parameter set of each roll of fabric in historical production batches is the same as the real-time deviation analysis method described above. That is, the equivalent raw material intrinsic parameter vector of the roll is used to replace the real-time raw material intrinsic parameter vector, and the calculation is performed based on the mean vector and covariance matrix in S2. The actual process parameter records for each roll of fabric at each process and the corresponding finished product quality inspection results are obtained. Each process includes pre-forming, napping, shearing, re-forming, and functional finishing processes. The actual process parameter records include the temperature setpoint, tension setpoint, machine speed setpoint, overfeed ratio setpoint, napping roller speed and napping pass number in the napping process, and auxiliary agent application rate in the functional finishing process. The finished product quality inspection results include the elastic recovery rate tested according to FZ / T 70006 standard and the anti-pilling grade tested according to GB / T 4802.2 standard using the Martindale method. The finished product quality inspection results, with elastic recovery rate and anti-pilling level meeting preset standards, were used as the qualification screening criteria to select all qualified samples. A process compensation training sample set was constructed using the raw material deviation parameter set of the same roll of fabric as input features and the actual process parameter records of the fabric in each process as output labels. A multi-task learning deep neural network was constructed as the initial mapping model. The bottom layer of this network contains three shared fully connected hidden layers, each containing 256 neural nodes. Above the shared layers, five task heads were set up, corresponding to the process parameter prediction of the pre-forming process, napping process, shearing process, re-forming process, and functional finishing process. Each task head contains two independent fully connected hidden layers, each containing 64 nodes. The number of nodes in the output layer equals the number of process parameters for the corresponding process. Using the raw material deviation parameter set of each sample in the process compensation training sample set as input, a mean squared error loss function is constructed based on the deviation between the predicted process parameters of each process output by the initial mapping model and the actual process parameter records of the corresponding samples. The total loss is obtained by weighting and summing the losses of each process according to preset weight coefficients, which are set according to the influence of each process on the final elastic recovery rate and anti-pilling grade. The Adam optimizer is used to iteratively update all model parameters of the initial mapping model. During the training process, an L2 regularization term is introduced to prevent overfitting, with a regularization coefficient of 0.001. Training stops when the total loss decreases less than a preset convergence threshold within 20 consecutive rounds, thus obtaining the process compensation decision model.

[0033] During the current batch production, the raw material deviation parameter set is input into the trained process compensation decision model. The model performs forward calculation, and its shared layer extracts deep correlation features from the deviation parameter set. Each task head independently predicts the optimized process parameters for each process based on the shared features. Finally, the model outputs a complete set of process setting parameters for the five processes of pre-setting, napping, shearing, re-setting, and functional finishing of the current batch of unprocessed fabric.

[0034] S5: When the subsequent unprocessed fabrics of the current batch pass through each process in sequence according to the process setting parameter set, high-frequency equipment fluctuation monitoring is performed on each process to obtain the dynamic deviation time sequence of the actual fluctuation of each process relative to the corresponding process setting parameter. Obtain the set of process settings parameters for each process step of the current batch of unprocessed fabric, output by the S4 process compensation decision model. This set of process settings parameters is stored in a structured format and contains all process control parameter settings corresponding to each process step. Specifically, these include the temperature, tension, speed, and overfeed ratio settings for the pre-forming process; the raising roller speed and number of raising passes settings for the raising process; the shearing blade distance and speed settings for the shearing process; the temperature, tension, speed, and overfeed ratio settings for the re-setting process; and the auxiliary agent application rate and drying oven temperature settings for the functional finishing process.

[0035] As the unprocessed fabric sequentially passes through each process step according to the set of process parameters, the corresponding equipment operating parameter time series are collected in real time from the equipment controllers of each process step via the OPC UA industrial communication protocol at a preset sampling frequency of 10Hz. For each process step, the deviation between its equipment operating parameter time series and the corresponding process control parameter set value is calculated point by point. For temperature parameters, the Celsius difference between the actual temperature value and the temperature set value is calculated; for tension parameters, the Newton difference between the actual tension value and the tension set value is calculated; for speed parameters, the meter-per-minute difference between the actual speed value and the speed set value is calculated; for overfeed ratio parameters, the percentage difference between the actual overfeed ratio value and the overfeed ratio set value is calculated. The deviation values ​​calculated point by point are arranged in the order of sampling time to form the equipment operating parameter deviation sequence for that process step. After synchronizing and aligning the deviation sequences of all controlled parameters for each process step along the time axis, they together constitute the dynamic deviation time series for that process step.

[0036] S6: Construct a joint attribution mapping relationship between the intrinsic parameter vector of raw materials and the dynamic deviation time sequence of each process relative to the local quality risk of the fabric. Through the joint attribution mapping relationship, identify the quality differentiation within the roll during each process, output the quality risk judgment result of each process, and determine the fabric to be sampled.

[0037] First, a spatiotemporal alignment system for fabric segments is constructed. A one-dimensional spatial coordinate axis is established along the fabric movement direction, with the stenter exit as the origin of the coordinate system. A virtual marker generator is set at the stenter exit. Whenever the cumulative fabric travel length reaches an integer multiple of 1 meter, a virtual marker with a precise timestamp and cumulative length value is generated. This virtual marker flows between processes along with the fabric production logistics information. The arrival and departure times of the virtual marker are recorded at the entrance and exit of each process. To eliminate length variations caused by fabric stretching or shrinkage between processes, a linear scaling factor is calculated based on the historical fabric length-to-input ratio for each process. The cumulative length value of each virtual marker is then linearly corrected according to this scaling factor. After correction, the virtual marker system can achieve precise segment-level registration and alignment within the data space, including the raw material intrinsic parameter vector in the raw material intrinsic parameter space, the dynamic deviation sequence of the corresponding segment during equipment fluctuations in each process, and the local quality inspection results of the corresponding segment during finished product inspection.

[0038] Secondly, a joint attribution mapping relationship is constructed. This involves obtaining the equivalent raw material intrinsic parameter vectors of each roll of fabric from each historical processing batch in the fluctuation-response correlation trajectory library, the dynamic deviation time-series segments of each process extracted after registration according to the aforementioned segment spatiotemporal alignment system, and the corresponding finished product quality inspection results for each segment within the roll. The finished product quality inspection results are obtained by continuous meter-by-meter inspection of the finished fabric using an offline inspection station, including meter-by-meter records of local values ​​for elastic recovery rate and anti-pilling grade. Segments with local values ​​for elastic recovery rate below the acceptable threshold or local values ​​for anti-pilling grade below the acceptable grade are marked as having local quality defects and used as positive sample labels; the remaining segments are marked as not having local quality defects and used as negative sample labels. A joint attribution training sample set is constructed by splicing the equivalent raw material intrinsic parameter vectors of the same segment of the same roll of fabric and the segment segments of dynamic deviation time-series of each process as the input feature vector, and using the presence or absence of local quality defects in that segment as the output label. An initial attribution model based on a Transformer encoder architecture is constructed. Its input layer maps the input feature vector to a fixed-dimensional embedding space through a linear projection layer, followed by four stacked Transformer encoder layers. Each layer includes a multi-head self-attention mechanism and a feedforward fully connected network. The multi-head self-attention mechanism has 8 heads and can automatically learn which process and at which moment, given specific raw material intrinsic parameters, has a stronger causal association weight with the final local quality defect. The encoder output is fed into a binary classification output layer after global average pooling. The output layer uses a sigmoid activation function to output the local quality risk prediction probability value. Using the binary cross-entropy loss between the model's output quality risk prediction probability value and the sample's true label as the loss function, the AdamW optimizer is used to iteratively update the model parameters. During training, a cosine annealing learning rate scheduling strategy is used, iterating until a preset convergence condition is met to obtain the joint attribution mapping relationship.

[0039] In practical applications, the real-time intrinsic parameter vector of the current batch of fabric (output by S3) and the dynamic deviation time sequence of subsequent unprocessed fabric in the current batch at each process (output by S5) are obtained. Based on the virtual labeling system and linear scaling correction, the real-time intrinsic parameter vector and the dynamic deviation time sequence of each process are aligned segment by segment according to the fabric's physical segments. For each 1-meter-long fabric segment, its corresponding intrinsic parameter vector is extracted as the intrinsic feature input of that segment. The dynamic deviation time sequence statistical features within the corresponding time segment during each process, including the mean deviation and maximum deviation, are extracted as the process deviation feature input of that segment. These two features are concatenated to obtain the joint input feature vector of that segment. The joint input feature vector is input into the joint attribution mapping relationship segment by segment. After forward calculation by the model, the local quality risk prediction probability value corresponding to each segment is output. A preset risk threshold of 0.8 is set. Segments with a local quality risk prediction probability value exceeding 0.8 are marked as risk segments, and segments with a local quality risk prediction probability value not exceeding 0.8 are marked as normal segments. The risk assessment labels for all segments within each process are compiled to generate quality risk assessment results for each process. The roll identifier, starting cumulative length position, and ending cumulative length position information corresponding to all risky segments in each process are extracted to generate a sample fabric list containing precise positioning information. Based on this sample fabric list, quality inspectors can directly locate the corresponding specific segment position on the finished fabric roll for targeted, destructive, or sensory sampling inspections, achieving precise capture of quality-differentiated areas within the roll.

[0040] In a preferred embodiment of the present invention, the specific process of establishing the fluctuation-response correlation trajectory library in step S1 is as follows: Obtain historical processing logs for the fabric setting process, including time series of equipment operating parameters collected during the setting process for each historical batch and a sequence of fabric physical state parameters detected at the exit of the setting process for each roll of fabric. Equipment operating parameters include temperature, tension, and machine speed. Fabric physical state parameters include width and weight. Temperature is collected at fixed intervals by thermocouples installed in each heating zone of the drying oven and uploaded to the manufacturing execution system database. Tension is measured by a load cell installed in the overfeed roller bearing housing to measure the tension of the fabric on the roller or by a torque sensor of the needle plate chain drive motor to measure the load torque. The speed is calculated by converting the angular displacement of the shaft into pulse counts via the encoder of the main drive motor, and then converting it into linear velocity based on the diameter of the drive roller. The width is calculated by a line-scan camera installed at the exit of the stenter scanning line by line at a frequency synchronized with the machine speed as the fabric passes through continuously. The left and right boundaries of the fabric are identified by an edge detection algorithm, and then converted into physical width based on the pixel equivalent coefficient. The weight is calculated by emitting beta rays through the fabric using an online areal density measuring instrument and estimating the mass per unit area based on the attenuation of the rays, or by regressing the average grayscale value of the fabric transmitted light grayscale image acquired by the line-scan camera with the offline weighing weight. The relationship is estimated; the time series of equipment operating parameters is divided into sliding windows with fixed window width and fixed step size. Adjacent windows overlap partially because the step size is smaller than the window width to ensure the continuity of fluctuation characteristics over time. For temperature, tension, and vehicle speed signals within each window, the fluctuation amplitude is obtained by calculating the root mean square (RMS) value, i.e., the standard deviation, of the signal segment within the window. The standard deviation reflects the overall dispersion of the signal around its center level. When the signal exhibits a clear trend, a first-order difference is first performed on the signal, i.e., subtracting the previous sample value from the subsequent sample value to obtain a difference sequence fluctuating around zero. Then, the standard deviation is calculated on the difference sequence. The standard deviation is used to eliminate trend interference. The fluctuation frequency is obtained by performing a fast Fourier transform on the signal segment within the window to decompose the time-domain signal into a superposition of sine waves of different frequencies. After obtaining the spectrum, the frequency component with the largest amplitude in the spectrum is taken as the main frequency. The level of the main frequency reflects the speed of the periodic vibration of the equipment parameters. The waveform characteristic parameters are obtained by calculating the kurtosis and skewness values ​​of the probability density distribution of the signal segment within the window. The kurtosis value describes the concentration of the signal values ​​and the thickness of the tail, which can distinguish between smooth fluctuations and impact fluctuations containing sudden spikes. The skewness value describes the asymmetry of the signal distribution, which can determine the amount of positive jumps and negative drops in the fluctuation.The physical state parameter sequence of the fabric is grouped by roll identifier. The initial width and initial weight values ​​of the fabric roll recorded by the previous process before entering the setting process are retrieved from the database. The width and weight values ​​of all sampling points at the setting exit of the roll are arithmetically averaged to eliminate measurement noise and short-range fluctuations, yielding the average width and average weight values ​​at the exit. The width change is obtained by subtracting the initial width value from the average width value at the exit, reflecting the combined effect of the transverse tensile force and longitudinal tension of the needle plate chain on the plastic deformation and shape fixation of the thermoplastic fibers during the setting process. The weight change is obtained by subtracting the initial weight value from the average weight value at the exit, reflecting the weight change per unit area caused by the combined effect of longitudinal tension and overfeed on the warp and weft densities during the setting process. The changes in total fiber accumulation, width, and weight are collectively used as the response characteristics of the fabric roll. The average fluctuation pattern feature vector is obtained by taking the arithmetic mean of the fluctuation pattern features of all sliding windows within the same time period of the same fabric roll across all dimensions. This average fluctuation pattern feature vector represents the overall equipment excitation level experienced by the fabric roll during the setting process. This average fluctuation pattern feature vector is paired with the response feature vector of the fabric roll to generate fluctuation-response correlation records. The underlying logic is that the same fabric roll experienced equipment fluctuations during the setting process, resulting in corresponding physical state changes, and there is a causal relationship between the two. All fluctuation-response correlation records from all historical processing batches are aggregated, and a fluctuation-response correlation trajectory library is established using the roll identifier as an index.

[0041] In another preferred embodiment of the present invention, the specific generation process of the fabric equivalent parameter inversion model in step S2 is as follows: Laboratory test results were obtained from raw material samples retained from each roll of fabric in historical production batches. The measured intrinsic parameters of the raw materials used in each roll of fabric were extracted. These measured intrinsic parameters included the measured values ​​of spandex heat shrinkage rate, elastic modulus, and coefficient of friction. The spandex heat shrinkage rate was obtained by measuring the ratio of the length difference before and after shrinkage to the initial length after heat treatment of the spandex yarns disassembled from the fabric in a dry heat shrinkage tester at a specified temperature and time. The elastic modulus was recorded by tensile testing of standard-sized strips of fabric on a universal testing machine. The strain curve is obtained by taking the slope of the initial linear segment. The friction coefficient is obtained by using a friction coefficient meter to slide a standard friction body on the fabric surface at a constant speed and pressure, and measuring the ratio of frictional force to normal pressure. These three test results, from the three dimensions of heat shrinkage characteristics, mechanical rigidity, and surface slip characteristics, together constitute the measured values ​​of the intrinsic parameters of the raw material reflecting its inherent quality. Based on the roll identifier carried by each fluctuation-response correlation record in the fluctuation-response correlation trajectory library, the measured values ​​of the intrinsic parameters of the raw material, fluctuation mode characteristics, and response characteristics of the same roll of fabric are associated and bound using the roll identifier as the association key. The implementation logic is as follows. Starting from the roll identifier of the laboratory test record, the system retrieves associated records with the same roll identifier from the trajectory database and pairs them to form a complete record. The first half of this record describes the equipment fluctuation pattern experienced by the roll of fabric during the finishing process and the resulting changes in fabric width and weight. The second half describes the physicochemical property test results of the raw material of the roll of fabric in the offline laboratory, thus forming a raw material intrinsic parameter-fluctuation-response association record. This association record of all historical processing batches is aggregated to form a supervised learning training set. The concatenation of fluctuation pattern features and response features in each association record is used as input data, and the corresponding measured values ​​of raw material intrinsic parameters are used as output labels for training. The basic principle of training is to allow the model to learn the statistical mapping law between observable external stimulus response patterns and raw material internal properties that cannot be directly and continuously measured in a large number of samples. That is, when the equipment fluctuation pattern and the fabric response pattern exhibit a certain combination of features, the corresponding raw material intrinsic parameters will fall into a certain numerical range. The model continuously narrows the gap between the raw material intrinsic parameters predicted by it based on the input and the laboratory measured values ​​by repeatedly adjusting the internal parameters, and finally completes the training of the fabric equivalent parameter inversion model.

[0042] The heat shrinkage rate of spandex in the raw material determines the fabric's sensitivity to temperature and time during heat setting; the elastic modulus determines the fabric's resistance to deformation under tension; and the coefficient of friction affects the fabric's processing behavior in subsequent surface treatment processes such as napping and shearing. These three parameters are the fundamental variables affecting the final quality of high-elasticity, wrinkle-resistant, and pilling-resistant fabrics. However, traditional testing methods can only sample a few items after raw materials are received or finished products are produced, resulting in significant lag and sampling limitations, failing to guide the actual processing of each roll of fabric. By constructing a fabric equivalent parameter inversion model, the heat setting process—a necessary step for all fabrics—is designed as a large-scale online raw material characteristic analysis instrument. Utilizing the equipment fluctuation excitation and fabric physical state response naturally generated during heat setting of each roll of fabric, its intrinsic raw material parameters are inversely calculated. This allows the core characteristics of each roll of fabric to be digitally characterized at the very beginning of processing and enter the subsequent decision-making process in the form of a raw material intrinsic parameter vector, providing input for subsequent deviation analysis and full-process process compensation.

[0043] In another preferred embodiment of the present invention, the specific process for generating the intrinsic parameter space of the raw materials in step S2 is as follows: Using each correlation record in the wave-response correlation trajectory library as input, the pre-trained fabric equivalent parameter inversion model is invoked. The wave pattern features and response features in the correlation record are concatenated and fed into the model for a forward calculation. Internally, the model maps the input to the output space through multiple nonlinear transformations to obtain a three-dimensional vector. The three components of this vector correspond to the equivalent spandex heat shrinkage rate, equivalent elastic modulus, and equivalent friction coefficient of the fabric roll, respectively. These are called equivalent raw material intrinsic parameters because they are not directly measured by physical instruments in the laboratory, but are the most efficient solution calculated based on the external behavior of the fabric roll during the setting process under the statistical laws of the inversion model. The implicit attribute values ​​of its input-output relationship are interpreted. Their physical meaning is consistent with the laboratory measured values, but they are called equivalent values ​​because they include model errors and measurement noise in the calculation process. All associated records in the fluctuation-response correlation trajectory library are traversed, and the above operation is performed on each record to obtain the equivalent raw material intrinsic parameter vectors corresponding to each roll of fabric in each historical processing batch. All vectors are aggregated to form the equivalent raw material intrinsic parameter set. Each point in this set represents the mapping position of the raw material characteristics of a historically processed roll of fabric in three-dimensional space. The distribution of points is determined by various raw material batches actually used in historical production. Spandex and nylon or polyester yarns from different origins and batches differ due to their... Differences in processing technology and spinning conditions will occupy different regions in the intrinsic parameter space, naturally forming several clusters of points with varying density. Statistical analysis of this equivalent raw material intrinsic parameter set is performed. First, the arithmetic mean of all points within the set is calculated in each dimension. The mean values ​​of spandex heat shrinkage rate, elastic modulus, and friction coefficient are combined into a mean vector. This mean vector represents the overall average characteristics of the historical raw material, i.e., the central position of the raw material's intrinsic parameter space. Then, the covariance between each pair of the three dimensions is calculated. The covariance between spandex heat shrinkage rate and elastic modulus reflects whether they change in the same direction, i.e., whether a raw material with a high heat shrinkage rate also has a high elastic modulus. The covariance between elastic modulus and friction coefficient... The covariance between the two reflects the statistical correlation between mechanical stiffness and surface slip, while the covariance between thermal shrinkage rate and friction coefficient reflects the intrinsic connection between thermal properties and surface properties. All covariances form a three-row, three-column covariance matrix, which fully describes the correlation direction and strength between various properties of historical raw materials, as well as the degree of variation of each property itself. Finally, the mean vector is used as the center position parameter of the raw material intrinsic parameter space, and the covariance matrix is ​​used as the distribution shape and scattering direction parameter of the raw material intrinsic parameter space. The equivalent set of raw material intrinsic parameters is used as the historical sample point group in the space. The three together constitute a complete raw material intrinsic parameter space containing discrete point groups and continuous statistical distribution information.

[0044] In another preferred embodiment of the present invention, the specific process for generating the raw material deviation parameter set in step S4 is as follows: The system obtains the real-time intrinsic parameter vector of the raw material, as well as the mean vector and covariance matrix contained in the intrinsic parameter space. The mean vector represents the average characteristics of all historical batches of raw material in each intrinsic parameter dimension, i.e., the distribution center of the historical normal state. The covariance matrix records the correlation structure between various characteristics of historical raw materials. The absolute deviation of the heat shrinkage rate is obtained by subtracting the spandex heat shrinkage rate component of the real-time vector from the spandex heat shrinkage rate component of the mean vector, the absolute deviation of the elastic modulus is obtained by subtracting the mean elastic modulus from the elastic modulus component, and the absolute deviation of the friction coefficient is obtained by subtracting the mean friction coefficient from the friction coefficient component. These three differences constitute the absolute deviation vector, which describes the new batch of raw material relative to the historical average in each dimension. The direction and magnitude of the deviation are determined by calculating the Mahalanobis distance of the absolute deviations of each dimension using the covariance matrix. The principle is to perform a linear transformation on the original space using the inverse of the covariance matrix, which is equivalent to rotating the coordinate system to eliminate the correlation between dimensions and scaling according to the historical variation of each dimension. This automatically reduces the deviation in dimensions with large historical fluctuations and automatically amplifies the deviation in dimensions with small historical fluctuations, so that the degree of deviation reflects the statistical degree of abnormality rather than the absolute numerical difference. The overall Mahalanobis distance is decomposed according to the contribution ratio of the absolute deviation of each dimension to obtain the relative deviation of each dimension, which is used as the raw material deviation parameter set.

[0045] In another preferred embodiment of the present invention, the specific construction process of the process compensation decision model in step S4 is as follows: This method acquires the raw material deviation parameter set for each roll of fabric from historical production batches, along with the actual process parameter records for each roll at each stage and the corresponding finished product quality inspection results. The finished product quality inspection results include elastic recovery rate and anti-pilling grade. Elastic recovery rate, measured by applying a fixed elongation stretch to the finished fabric using standard testing methods, reflects the fabric's ability to resist deformation and return to its original shape during wear. Anti-pilling grade, assessed by subjecting the finished fabric to a specified number of rubbing and tumbling cycles using the Martindale method or pilling box method, is compared with standard rating samples to reflect the degree of fiber entanglement on the fabric surface. Using the raw material deviation parameter set of the same roll of fabric as input features, the actual process parameter records for each stage of the fabric roll as output labels, and the finished product quality inspection results meeting preset qualification standards as screening conditions, a process compensation training sample set is constructed. The basic logic of the screening is to retain only samples with qualified final product quality for training, because the process parameters corresponding to these samples are effective parameters verified in practice. Process parameters corresponding to unqualified samples, even if related to raw material deviation, have no learning value. To ensure the model learns a positive decision-making strategy for obtaining qualified finished products under given raw material deviation conditions, a multi-layer mapping network is constructed as the initial mapping model. The raw material deviation parameter set of each sample in the training sample set is used as input, allowing the model to calculate and output the predicted process parameters for each process layer by layer. The predicted process parameters are compared with the actual process parameters of the sample, and the deviation values ​​of the two at each process are calculated. The deviations of all processes are summarized into a loss function, which measures the gap between the model-recommended process scheme and historical successful experience. The training objective is to minimize this gap, that is, to make the model-recommended process parameters as close as possible to the actual process parameters that have produced qualified finished products. Based on the loss function, all parameters of the initial mapping model are iteratively updated. In each iteration, the gradient direction of the loss function with respect to each parameter is calculated first, i.e., the trend of the loss value increasing or decreasing with a small increase or decrease in the parameter. Then, the parameters are adjusted along the gradient descent direction to gradually reduce the loss. When the deviation between the model-recommended process parameters and the actual successful experience no longer decreases significantly, i.e., the preset convergence condition is met, training stops, and the process compensation decision model is obtained.

[0046] In another preferred embodiment of the present invention, the specific process for obtaining the dynamic deviation timing in step S5 is as follows: Obtain the process setting parameter set for each process of the current batch of unprocessed fabric. The process setting parameter set includes the process control parameter setting value corresponding to each process. When the unprocessed fabric passes through each process in sequence according to the process setting parameter set, collect the time series of equipment operation parameters corresponding to each process at a preset sampling frequency. Calculate the deviation between the time series of equipment operation parameters for each process and the process control parameter setting value of the corresponding process at each sampling point to obtain the equipment operation parameter deviation sequence for each process, which serves as the dynamic deviation time series for that process.

[0047] In another preferred embodiment of the present invention, the specific process of generating the joint attribution mapping relationship in step S6 is as follows: The system acquires the equivalent intrinsic parameter vectors of raw materials for each roll of fabric in each historical processing batch from the fluctuation-response correlation trajectory library, as well as the dynamic deviation sequence of each process and the corresponding quality inspection results for each segment within the roll. The quality inspection results for each segment within the roll are obtained through continuous meter-by-meter testing of the finished fabric or sampling testing at fixed length intervals. The testing items are consistent with the finished product quality standards, including elastic recovery rate and anti-pilling grade. Each segment is assigned a quality judgment label, indicating either the presence or absence of local quality defects. The judgment is based on whether the elastic recovery rate of that segment is lower than a preset acceptable threshold or whether the anti-pilling grade is lower than a preset acceptable grade. If any one of the criteria is not met, it is marked as defective. This fine-grained quality labeling at the paragraph level is to enable the model to learn the spatial distribution pattern of quality defects within the fabric roll, rather than simply judging whether the entire roll is acceptable. A joint attribution training sample set is constructed using the equivalent raw material intrinsic parameter vector of the same paragraph of the same fabric roll and the splicing of the dynamic deviation time sequence of each process as input features, and the quality judgment label of that paragraph as the output label. The key to this construction lies in the spatiotemporal alignment at the paragraph level, which requires ensuring that the dynamic deviation time sequence segment in the input features precisely corresponds to the equipment fluctuation record experienced by that section of fabric during each process. The formula uses the cumulative fabric travel length recorded at the inlet and outlet of each process, combined with the fabric stretch coefficient of each process, to map the spatial position of the roll segment to the corresponding time period on the time axis of each process, thereby extracting the deviation time sequence segment of that segment in each process; an initial attribution model with an attention mechanism at its core is constructed. When processing input features, this model can automatically learn the correlation strength between different input parts. Taking the equivalent raw material intrinsic parameter vector of each sample in the joint attribution training sample set and the dynamic deviation time sequence of each process as input, the model automatically evaluates which process which deviation time sequence is correct under the given specific raw material intrinsic parameters through internal attention calculation. The device deviation at a given moment contributes most to the prediction of the final quality defect. The model outputs the predicted probability value of the local quality defect in that paragraph. The predicted probability value is compared with the true defect label of the sample to calculate the deviation between the two and construct a loss function. The loss function measures the accuracy of the model in predicting the quality risk of each paragraph in the volume. Based on the loss function, the model parameters of the initial attribution model are iteratively updated. In each iteration, the weights of the connections between the layers inside the model are adjusted so that the next prediction is closer to the true label. When the model's accuracy in identifying quality defect paragraphs no longer improves significantly, i.e., the preset convergence condition is met, the joint attribution mapping relationship is obtained.

[0048] In actual production, different sections of a roll of fabric may exhibit different final qualities due to microscopic differences in raw material properties and fluctuations in equipment operation at different times. The traditional approach is to have technicians reverse-engineer the cause after problems are found during finished product inspection. However, due to the numerous and intertwined influencing factors, it is often difficult to accurately pinpoint whether the cause is a deviation in the intrinsic properties of the raw material or a momentary deviation in the equipment at a certain process. By constructing a joint attribution mapping relationship, the intrinsic parameter vector of the raw material and the dynamic deviation sequence of each process are simultaneously introduced into the model as explanatory variables. The attention mechanism is used to automatically discover the higher-order interaction relationship between the input variables and quality defects. For example, the model may learn that when the heat shrinkage rate of spandex is high and the roller speed of the napping process has a positive deviation, the risk of pilling increases significantly. However, a high heat shrinkage rate of spandex alone or a deviation in the napping roller speed alone is not enough to cause defects. This kind of conditional combination of defect causes is difficult to discover using traditional single-factor analysis methods. From the perspective of the ultimate goal of the whole solution, the joint attribution mapping relationship is the diagnosis and feedback link of the entire closed-loop quality control system. It can not only predict in real time which sections in the roll have quality risks in the new batch production, guide accurate sampling inspection and reduce full inspection costs, but more importantly, when quality risks are found, it can trace whether the problem is caused by the deviation of the raw material characteristics themselves, the equipment execution deviation of a certain process, or a specific combination of the two, thus providing a clear direction for continuous process improvement and equipment maintenance.

[0049] In another preferred embodiment of the present invention, step S6, which involves outputting the quality risk assessment results of each process and determining the specific process for sampling the fabric, is as follows: The system acquires the real-time intrinsic parameter vector of the raw material for the current batch of fabric and the dynamic deviation time sequence of subsequent unprocessed fabric in each process. It then aligns the real-time intrinsic parameter vector with the dynamic deviation time sequence of each process by fabric segment. The basic method for segment alignment is to first divide the entire roll of fabric into several continuous virtual segments of fixed length along the fabric's travel direction and assign a unique segment identifier to each segment. Then, based on the cumulative fabric travel length recorded at the inlet and outlet of each process and the inlet / outlet fabric length ratio (expansion coefficient) of that process, the spatial position of each fabric segment is mapped to the corresponding time interval on the dynamic deviation time sequence time axis of that process. The deviation sequence within that time interval is extracted from the dynamic deviation time sequence of that process as the deviation feature of that segment in that process. Finally, the real-time intrinsic parameter vector of the same segment is concatenated with the deviation features of each process to form the joint input feature of that segment. In this way, each fabric segment obtains a structured description containing its inherent raw material characteristics and the equipment execution deviations throughout the processing. The joint input features of each segment in each process are input into the trained joint attribution mapping relationship. The model then internally... Multi-layer nonlinear transformation and attention-weighted calculation automatically assess the degree of interaction between equipment deviations in each process and quality risk under given raw material characteristics, and output the local quality risk prediction value of that segment. The prediction value is a probability value between zero and one, and the larger the value, the higher the possibility of local quality defects in that segment. The local quality risk prediction value of each segment is compared with a preset risk threshold. The preset risk threshold is determined by combining the distribution of predicted values ​​of quality defect segments in historical data with the company's tolerance for missed and false inspections. Segments with predicted values ​​exceeding the threshold are marked as risk segments, and segments with predicted values ​​not exceeding the threshold are marked as normal segments. The risk judgment labels of all segments in each process are summarized to generate the quality risk judgment result of each process. This result indicates which segments of the fabric have high local quality risks within the processing range of which processes. The roll identifier and the start and end length positions corresponding to the risk segments in each process are extracted to generate a sample inspection fabric list. Based on this list, quality inspectors do not need to conduct full inspection of the entire roll of fabric, but only need to find the corresponding specific position on the roll for fixed-point sampling inspection to efficiently capture potential quality problems.

[0050] Traditional sampling methods typically involve taking samples from the beginning and end of fabric rolls or at fixed intervals. However, localized quality defects within a roll often occur in random segments unrelated to these predetermined locations, leading to low sampling efficiency and a high risk of missed defects. By using a joint attribution mapping relationship to quantify the comprehensive risk of each segment and filtering high-risk segments according to thresholds, sampling resources can be concentrated on the most likely problem locations, significantly improving the detection rate of quality issues without increasing the sampling workload. Simultaneously, the generated sampling list includes precise roll identification and segment location information, providing clear and executable instructions for quality inspection operations. This transforms quality control from passively waiting for problems to surface to proactively predicting and conducting targeted verification. From the perspective of the entire solution's ultimate goal, this step translates front-end raw material perception and end-to-end equipment monitoring into actionable quality control actions, achieving a complete closed loop of quality management from data collection to analysis and decision-making to physical execution.

[0051] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A production quality management method for high-elasticity, wrinkle-resistant, and pilling-resistant fabric, characterized in that, Includes the following steps: S1: Extract historical fluctuation pattern features and historical response features from the historical processing logs of the fabric finishing process, and establish a fluctuation-response correlation trajectory library; S2: Perform inverse modeling on the wave-response correlation trajectory library, construct a fabric equivalent parameter inversion model with wave pattern as excitation and response feature as observation, and solve the equivalent raw material intrinsic parameters corresponding to each wave-response correlation record based on the fabric equivalent parameter inversion model, and aggregate to form a raw material intrinsic parameter space; S3: When the lead roll of the current processing batch passes through the shaping process, collect the real-time fluctuation mode characteristics and real-time response characteristics and input them into the fabric equivalent parameter inversion model to calculate the real-time raw material intrinsic parameter vector of the fabric corresponding to the current processing batch. S4: Perform deviation analysis on the real-time raw material intrinsic parameter vector in the raw material intrinsic parameter space, generate a raw material deviation parameter set and input it into the pre-constructed process compensation decision model, and output the process setting parameter set for each process of the current batch of subsequent unprocessed fabric. S5: When the subsequent unprocessed fabrics of the current batch pass through each process in sequence according to the process setting parameter set, high-frequency equipment fluctuation monitoring is performed on each process to obtain the dynamic deviation time sequence of the actual fluctuation of each process relative to the corresponding process setting parameter. S6: Construct a joint attribution mapping relationship between the intrinsic parameter vector of raw materials and the dynamic deviation time sequence of each process relative to the local quality risk of the fabric. Through the joint attribution mapping relationship, identify the quality differentiation within the roll during each process, output the quality risk judgment result of each process, and determine the fabric to be sampled.

2. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In S1, the specific process of establishing the fluctuation-response correlation trajectory library is as follows: Obtain historical processing logs for the fabric finishing process. These logs include time series of equipment operating parameters collected during the finishing process for each historical batch and a sequence of fabric physical state parameters detected at the finishing process exit for each roll of fabric. The equipment operating parameters include temperature, tension, and machine speed, while the fabric physical state parameters include width and weight. The time series of the equipment operating parameters is segmented by a sliding window, and the fluctuation pattern features of the equipment operating parameters are extracted in each sliding window. The fluctuation pattern features include the fluctuation amplitude, fluctuation frequency and waveform feature parameters of each equipment operating parameter in the sliding window. The physical state parameter sequence of the fabric is grouped by roll identifier. The width and weight values ​​of each roll of fabric at the exit of the setting process are calculated relative to the initial width and weight values ​​at the entry of the setting process, and these are used as the response characteristics of the roll of fabric. The fluctuation pattern characteristics and response characteristics of each roll of fabric in the same historical processing batch are aligned and bound along the time axis to generate fluctuation-response association records; the fluctuation-response association records of all historical processing batches are aggregated, and the fluctuation-response association trajectory library is established with the roll identifier as the index.

3. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In S2, the specific generation process of the fabric equivalent parameter inversion model is as follows: Laboratory test results of raw material samples from each roll of fabric in historical production batches are obtained. The measured values ​​of the intrinsic parameters of the raw materials used in each roll of fabric are extracted. These intrinsic parameters include the measured values ​​of spandex heat shrinkage rate, elastic modulus, and friction coefficient. Based on the roll identifier carried by each fluctuation-response correlation record in the fluctuation-response correlation trajectory library, the measured values ​​of the intrinsic parameters of the raw materials, the fluctuation pattern characteristics, and the response characteristics of the same roll of fabric are associated and bound to form a raw material intrinsic parameter-fluctuation-response correlation record. The raw material intrinsic parameter-fluctuation-response correlation records of all historical processing batches are aggregated. Using the fluctuation pattern characteristics and response characteristics in each correlation record as input data and the corresponding measured values ​​of the intrinsic parameters of the raw materials as output data, the fabric equivalent parameter inversion model is trained.

4. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In S2, the specific process for generating the intrinsic parameter space of the raw materials is as follows: Using each associated record in the fluctuation-response correlation trajectory library as input, the fabric equivalent parameter inversion model is invoked. After forward calculation by the fabric equivalent parameter inversion model, the equivalent raw material intrinsic parameters corresponding to the associated record are output. All associated records in the fluctuation-response correlation trajectory library are traversed to obtain the equivalent raw material intrinsic parameters corresponding to each roll of fabric in each historical processing batch, forming an equivalent raw material intrinsic parameter set. Statistical analysis is performed on the equivalent raw material intrinsic parameter set to calculate the mean vector and covariance matrix of the equivalent raw material intrinsic parameters in each dimension. Using the mean vector and the covariance matrix as the distribution parameters of the raw material intrinsic parameter space, a raw material intrinsic parameter space containing the equivalent raw material intrinsic parameter set and its distribution parameters is constructed.

5. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 4, characterized in that, In step S4, the specific process for generating the raw material deviation parameter set is as follows: Obtain the real-time raw material intrinsic parameter vector, as well as the mean vector and covariance matrix contained in the raw material intrinsic parameter space; calculate the component differences between the real-time raw material intrinsic parameter vector and the mean vector in each intrinsic parameter dimension to obtain the absolute deviation of each dimension; use the covariance matrix to calculate the Mahalanobis distance of the absolute deviation of each dimension to obtain the relative deviation of each dimension; use the relative deviation of each dimension as elements to obtain the raw material deviation parameter set.

6. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In S4, the specific construction process of the process compensation decision model is as follows: Obtain the raw material deviation parameter set for each roll of fabric in historical production batches, as well as the actual process parameter records for each roll of fabric in each process and the corresponding finished product quality inspection results; the finished product quality inspection results include elastic recovery rate and anti-pilling grade; Using the raw material deviation parameter set of the same roll of fabric as input features, the actual process parameter records of the roll of fabric in each process as output labels, and the finished product quality inspection results meeting the preset qualified standards as screening conditions, a process compensation training sample set is constructed. An initial mapping model is constructed, taking the raw material deviation parameter set of each sample in the process compensation training sample set as input, and constructing a loss function based on the deviation between the predicted process parameters of each process output by the initial mapping model and the actual process parameter records of the corresponding samples. The model parameters of the initial mapping model are iteratively updated based on the loss function until the preset convergence condition is met, thereby obtaining the process compensation decision model.

7. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In step S5, the specific process for obtaining the dynamic deviation timing is as follows: Obtain the process setting parameter set for each process of the current batch of unprocessed fabric. The process setting parameter set includes the process control parameter setting value corresponding to each process. When the unprocessed fabric passes through each process in sequence according to the process setting parameter set, collect the time series of equipment operation parameters corresponding to each process at a preset sampling frequency. Calculate the deviation between the time series of equipment operation parameters for each process and the process control parameter setting value of the corresponding process at each sampling point to obtain the equipment operation parameter deviation sequence for each process, which serves as the dynamic deviation time series for that process.

8. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In step S6, the specific process for generating the joint attribution mapping relationship is as follows: Obtain the raw material intrinsic parameter vector, dynamic deviation time sequence of each process, and quality inspection results of each section within the roll of fabric in each historical processing batch in the fluctuation-response correlation trajectory library; Using the intrinsic parameter vector of the raw material and the dynamic deviation time sequence of each process in the same roll of fabric as input features, and the presence of local quality defects in the quality inspection results of the same section as output labels, a joint attribution training sample set is constructed. An initial attribution model is constructed, taking the intrinsic parameter vectors of raw materials and the dynamic deviation time series of each process of each sample in the joint attribution training sample set as inputs. A loss function is constructed based on the deviation between the quality risk prediction result output by the initial attribution model and the output label of the corresponding sample. The model parameters of the initial attribution model are iteratively updated based on the loss function until the preset convergence condition is met, thereby obtaining the joint attribution mapping relationship.

9. The production quality management method for a high-elasticity, wrinkle-resistant, and pilling-resistant fabric according to claim 1, characterized in that, In step S6, the process of outputting the quality risk assessment results of each process and determining the specific fabric to be sampled is as follows: Obtain the real-time intrinsic parameter vector of the raw material for the current batch of fabric, and the dynamic deviation time sequence of the subsequent unprocessed fabric in the current batch at each process; align the real-time intrinsic parameter vector of the raw material and the dynamic deviation time sequence of each process according to fabric segments to obtain the joint input features of each segment in each process; The joint input features of each segment in each process are input into the joint attribution mapping relationship. After forward calculation by the joint attribution mapping relationship, the local quality risk prediction value of each segment in each process is output. The local quality risk prediction value of each segment in each process is compared with a preset risk threshold. Segments whose local quality risk prediction value exceeds the preset risk threshold are marked as risk segments, and the quality risk judgment result of each process is generated. Extract the roll identifier and section position information of the risk sections in each process to generate a list of fabrics to be sampled for inspection.

Images