A method and system for predicting and controlling the quality of cheese dyeing
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- DONGHUA UNIV
- Filing Date
- 2026-05-21
- Publication Date
- 2026-06-19
Smart Images
Figure CN122243305A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of textile technology, specifically to a method and system for predicting and controlling the dyeing quality of yarn packages. Background Technology
[0002] In the production of yarn package dyeing, quality control is a core factor affecting product quality and production costs. Currently, the industry generally uses offline sampling and colorimetric testing after dyeing to evaluate dyeing quality. This post-dyeing testing method has significant time lag. Once quality problems such as uneven coloring or layer differences occur, the entire batch often has to be scrapped or re-dyed, severely impacting production efficiency and economic benefits. Furthermore, dyeing quality is strongly influenced by the nonlinear coupling of multiple factors such as process parameters, yarn properties, and dye and auxiliary agent parameters. Traditional root cause diagnosis methods based on human experience are insufficient to quickly and quantitatively determine whether the main cause of defects is insufficient dye liquor penetration or excessively rapid dye adsorption under complex operating conditions.
[0003] In recent years, artificial intelligence and purely data-driven methods have been applied in the field of quality prediction. However, traditional artificial neural networks or fuzzy systems, lacking constraints from first principles of physics, exhibit extremely poor interpretability. Especially when faced with out-of-distribution samples such as new processes, dyes, and yarns beyond the training sample range, purely data models are prone to outputting inaccurate predictions that violate the physicochemical laws of dyeing. Their black-box nature also makes it impossible to map model errors into clearly defined on-site process adjustment instructions. Therefore, existing technologies suffer from a technical deficiency in simultaneously achieving real-time prediction accuracy, physical interpretability, and process closed-loop controllability. Summary of the Invention
[0004] To address the shortcomings of existing methods and the needs of practical applications, this invention provides a method for predicting and controlling the dyeing quality of yarn packages, comprising the following steps: Acquire dyeing data of packaged yarn; construct a physical adversarial feature screening model for porous topology of packaged yarn to obtain intermediate state features; based on the intermediate state features, obtain color difference prediction values through a multi-scale physical-spatiotemporal manifold coupling dyeing quality prediction model; perform backpropagation operation on the mechanism manifold loss term according to the color difference prediction values to obtain the physical sensitivity gradient, and determine the physical root cause type leading to quality defects through the magnitude relationship of the physical sensitivity gradient; combine the quality evaluation index and the physical root cause type to trigger or generate process control instructions.
[0005] Optionally, the step of constructing a physical adversarial feature selection model for porous topology of bobbin yarn to obtain intermediate state features includes the following steps: In physical information neural networks, physical priors are incorporated into attention as attention biases to form a multi-head self-attention mechanism for physical perception. By introducing sparsity constraints to make the gating coefficients sparse, a physical adversarial feature selection model for porous topology of bobbin yarn is constructed. By combining the yarn dyeing generation data and the physical adversarial feature screening model, intermediate state features are obtained.
[0006] Optionally, obtaining the color difference prediction value based on the intermediate state characteristics through a multi-scale physical-spatiotemporal manifold coupled staining quality prediction model includes the following steps: Based on Fick's diffusion law, Langmuir's adsorption kinetics equation, and the all-time-space matter conservation equation, a physical prior-guided mechanism manifold loss term is constructed. By combining the data-driven loss term, the sparse loss function, and the aforementioned mechanism manifold loss term, a dynamic coupling total loss function is obtained; Based on the total loss function, combined with the intermediate state features and the multi-scale physical-spatiotemporal manifold coupled staining quality prediction model, the color difference prediction value is obtained.
[0007] Optionally, based on Fick's diffusion law, Langmuir's adsorption kinetics equation, and the all-time-space matter conservation equation, a physical prior-guided mechanistic manifold loss term is constructed, satisfying:
[0008] in, For the mechanistic manifold loss term, The radial nonstationary mass transfer constraint loss term is based on Fick's second law, and the constraint formula is as follows: , For a moment Radial position The concentration of the dye solution at that location. The effective diffusion coefficient; The constraint loss term for interfacial adsorption based on Langmuir dynamics is given by the constraint formula as follows: , For a moment The amount of dye adsorption, For maximum adsorption capacity, The adsorption rate constant is . Let be the desorption rate constant. This represents the main concentration of the dye solution; The loss term is the hard constraint term for matter conservation across all time and space domains, and the constraint formula is as follows: , The height of the yarn package. This represents the initial concentration of the dye. This represents the total volume of the dye solution. The radius of the yarn package; , , These are the preset weighting coefficients.
[0009] Optionally, the combination of the data-driven loss term, the sparse loss function, and the mechanistic manifold loss term yields a dynamically coupled total loss function that satisfies:
[0010] in, Let be the total loss function for training the neural network. It is a dynamic function based on information entropy gain and process sensitivity. For the mechanistic manifold loss term, For data-driven loss terms, This is the sparsity loss function.
[0011] Optionally, the intermediate state features and the multi-scale physical-spatiotemporal manifold coupled staining quality prediction model include: The physical manifold baseline layer is used to output preliminary predictions that conform to physical laws and implicit physical intermediate state variables. A time-domain unmodeled dynamic compensation layer is used to capture unmodeled dynamic residuals caused by pump frequency commutation, transient temperature drift, and physical information that cannot be expressed by ideal partial differential equations of neural networks; The spatial field topology consistency refinement layer is used to reconstruct the physical connectivity between yarn bobbin layers using graph structures.
[0012] Optionally, the physical sensitivity gradient is used to characterize the contribution of different physical dynamic factors to the total error; The physical sensitivity gradient includes: Permeation kinetics sensitivity gradient, used to quantify the sensitivity of error to the effective diffusion coefficient; Interfacial reaction sensitivity gradient, used to quantify the sensitivity of error to the adsorption rate constant; The process control commands include adjusting the frequency or commutation cycle of the main pump of the dyeing machine, or triggering the Bayesian posterior recalibration of the physical constants within the model.
[0013] Optionally, the quality evaluation index includes a joint risk operator for physical-data dual manifold boundary crossing. ,satisfy:
[0014] in, For the first Color difference value of the layer They represent the inner layer, middle layer, and outer layer, respectively. For absolute boundary thresholds, For the partial differential equation residuals of the corresponding layer, As a measure of the random uncertainty of data, and These are adaptive weights for data and physical risks, respectively. The quality evaluation indicators also include long-term state evolution evaluation metrics based on latent space information divergence. ,satisfy:
[0015] in, Let be the probability distribution of the hidden states in the current batch. Let be the hidden state probability distribution of the baseline batch. The second-order Wasserstein distance. For joint risk operators, , and These are preset parameters.
[0016] Optionally, the method for predicting and controlling the dyeing quality of packaged yarn further includes performing time alignment, data cleaning, and normalization on the data generated from the dyeing of packaged yarn to form a complete monitoring dataset, which serves as the input to the physical information neural network.
[0017] Secondly, to efficiently execute the yarn dyeing quality prediction and control method provided by this invention, this invention also provides a yarn dyeing quality prediction and control system, including: an input device, an output device, a processor, and a memory. The input device, output device, processor, and memory are interconnected. The memory stores program instructions used in the yarn dyeing quality prediction and control method. This yarn dyeing quality prediction and control system of the present invention has a compact structure and stable performance, and can stably execute the yarn dyeing quality prediction and control method provided by this invention, further improving the overall applicability and practical application capability of this invention.
[0018] This invention incorporates prior physical knowledge into the feature selection stage through a physical perception attention mechanism and an adaptive sparse gating mechanism. This ensures that the model maintains focus on key physical features even when data quality is poor or out-of-distribution samples are present, avoiding the risk of blindly pruning important features in purely data-driven models. Furthermore, by constructing a mechanism-temporal-spatial three-scale coupled prediction architecture, the physical manifold baseline layer internalizes first principles such as Fick's diffusion law, Langmuir adsorption kinetics, and the law of conservation of mass, which must be followed in the dyeing process, into network constraints based on a physical information neural network. This ensures that the prediction results always conform to the physicochemical laws of dyeing, fundamentally avoiding inaccurate predictions that may be output by pure data models that violate physical laws. The temporal unmodeled dynamics compensation layer uses a temporal convolutional network to specifically capture unmodeled dynamic residuals that cannot be expressed by ideal partial differential equations such as pump frequency commutation and transient temperature drift, achieving refined compensation for the core mechanism. The spatial field topological consistency refining layer uses a graph neural network to construct a dynamic adjacency matrix based on the actual winding density gradient of the yarn package, correcting the prediction inconsistencies between spatial layers. These three layers are coupled together, enabling the prediction architecture to achieve high-precision optimization through data-driven methods in regular mass production batches, and automatically degenerate to physical mechanism dominance when dealing with out-of-distribution samples, thus ensuring extrapolation stability.
[0019] Furthermore, this invention constructs a multidimensional quality manifold comprehensive evaluation index system and performs root cause diagnosis and process closed-loop feedback based on physical gradient spatial decomposition. This allows the prediction deviation to be directly mapped to the permeation kinetics sensitivity gradient and the interface reaction sensitivity gradient, thereby quantitatively diagnosing whether the main cause of quality defects is insufficient permeation or adsorption imbalance. It also generates quantitative process adjustment commands for the main pump frequency or temperature control valve, realizing a complete closed-loop control from result prediction to root cause tracing to autonomous process intervention, significantly improving the intelligent management and control level of the dyeing and finishing process and the consistency of product quality. Attached Figure Description
[0020] Figure 1 A flowchart of a method for predicting and controlling the dyeing quality of packaged yarn provided in an embodiment of the present invention; Figure 2 This is a framework diagram of a yarn dyeing quality prediction and control system provided in an embodiment of the present invention. Detailed Implementation
[0021] Specific embodiments of the present invention will now be described in detail. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the invention. In the following description, numerous specific details are set forth in order to provide a thorough understanding of the invention. However, it will be apparent to those skilled in the art that these specific details are not necessary to practice the invention. In other instances, well-known circuits, software, or methods have not been specifically described to avoid obscuring the invention.
[0022] Throughout this specification, references to an embodiment, example, or illustration mean that a particular feature, structure, or characteristic described in connection with that embodiment or example is included in at least one embodiment of the invention. Therefore, phrases appearing in various places throughout the specification, such as "in one embodiment," "in an embodiment," "an example," or "an illustration," do not necessarily refer to the same embodiment or example. Furthermore, specific features, structures, or characteristics can be combined in any suitable combination and / or sub-combination in one or more embodiments or examples. Moreover, those skilled in the art will understand that the illustrations provided herein are for illustrative purposes and are not necessarily drawn to scale.
[0023] Please see Figure 1 This invention provides a method for predicting and controlling the dyeing quality of yarn packages, comprising the following steps: S1. Obtain the data generated by dyeing the yarn package.
[0024] In one embodiment, real-time process parameters are acquired from the dyeing machine control system and online monitoring equipment. These real-time process parameters include at least one of the following: dyeing temperature profile, heating rate, holding time, dye liquor flow rate, main pump frequency, pH value, and liquor ratio. Yarn physical property parameters are acquired from the yarn specification sheet or incoming material inspection report. These yarn physical property parameters include at least one of the following: bobbin winding density, yarn count, and raw material type. Dye and auxiliary parameters are acquired from the dyeing process formula sheet or formula management system. These dye and auxiliary parameters include at least one of the following: initial dye concentration, auxiliary type, and auxiliary dosage. The acquired real-time process parameters, yarn physical property parameters, and dye and auxiliary parameters are aligned based on timestamps and normalized to form a monitoring dataset. Real-time process parameters can be directly read from the dyeing machine PLC and sensors via industrial communication protocols such as OPCUA or Modbus. The sampling frequency is set according to the parameter change rate, typically 1-10Hz. Yarn physical property parameters are static attributes and are acquired before the start of the batch by scanning barcodes or manually entering them into the system.
[0025] Furthermore, the isolated forest algorithm is used to detect abnormal sampling points in the monitoring dataset, and these abnormal sampling points are removed or repaired by linear interpolation based on adjacent time windows. For missing sensor data in the monitoring dataset, a temporal generative adversarial network is used to impute missing values. The temporal generative adversarial network generates substitute values based on the distribution characteristics of historical normal production data. A data drift detection mechanism is introduced to monitor the KL divergence between the input data distribution and the training set distribution in real time. When the KL divergence exceeds a preset drift threshold, the model is automatically updated or the physical constraints are enhanced.
[0026] In its implementation, the Isolation Forest algorithm randomly selects subsamples from the monitoring dataset to construct multiple binary search trees, calculates the average path length of each sample across all trees, converts it into anomaly scores, and identifies points with anomaly scores higher than a set quantile as outliers and removes them. For missing data at the location of anomalies, linear interpolation is performed using the valid data points before and after them to repair the missing data.
[0027] In temporal generative adversarial networks, the generator receives random noise and temporal features as input to generate filler values. The discriminator distinguishes between the generated values and the real values. Through adversarial training, the generator learns the dynamic distribution pattern of normal data, thereby generating substitute values that conform to physical laws for missing sensor channels.
[0028] Data drift detection uses a sliding window to estimate the probability density of the current input data and calculates its KL divergence with the training data density. If the KL divergence is greater than a preset drift threshold (e.g., 0.5), the model update process is triggered or the weight of the mechanistic manifold loss term in the total loss function is temporarily increased to ensure the robustness of the model during operating condition drift.
[0029] S2. Construct a physical adversarial feature selection model for porous topology of yarn bobbins, and then obtain intermediate state features.
[0030] In this embodiment, physical priors are incorporated as attention biases into the attention calculation process.
[0031] For the input feature matrix Where N is the number of samples and d is the original feature dimension, the query matrix Q, key matrix K, and value matrix V are first obtained through linear transformation:
[0032] Three learnable weight matrices , , Each of them is the original feature dimension d. Key vector dimension A real matrix is used to project the input feature X into a matrix.
[0033] Then introduce the physical perception bias matrix This matrix encodes the prior physical association between features, and for features i and j, the physical bias. Defined as:
[0034] in, Let i be the distance between feature i and feature j in the physical knowledge graph. For bandwidth parameters, This is a feature association strength function based on domain knowledge.
[0035] For example, temperature and heating rate have a strong physical correlation, while temperature and bath ratio have a weaker correlation. The formula for calculating physical perception attention is:
[0036] The introduction of the physical bias matrix allows the model to incorporate physical prior knowledge when calculating feature importance, so that the model can maintain its focus on key physical features even if some physical correlations in the data are not obvious.
[0037] Multi-head physical perception attention is computed in parallel using h attention heads, each with an independent parameter matrix. The results from each head are concatenated and then linearly transformed for output.
[0038] The multi-head mechanism computes h independent attention heads in parallel, each capable of focusing on learning different physical association priors. Concat is a concatenation operation that combines the output vectors of the h heads into a single attention head. Large-dimensional vectors, preserving the unique physical feature information learned by each head and avoiding information loss, may result in excessively large dimensional vectors after concatenation. This can be addressed by using a weight matrix. Perform dimensionality reduction or feature fusion to output the final specified features.
[0039] Furthermore, in order to control computational complexity while maintaining feature selection capability, this invention also designs an adaptive sparse gating mechanism, which dynamically filters important features through learnable gating coefficients and introduces sparse constraints to make the gating coefficients sparse.
[0040] The core formula for the gating coefficient calculation module is:
[0041] This is a gating coefficient calculation module. Its core function is to simultaneously integrate the original feature information and attention weight information to generate a learnable feature selection gating, allowing the model to adaptively retain important features and suppress redundant features. This is the gating coefficient vector. It is the sigmoid activation function. This is the original feature branch, used to extract the importance of the original features themselves. This is the attention weight branch.
[0042] To encourage sparsity of the gating coefficients, L0 norm regularization is introduced, and a gating-based L0 regularization approximation is adopted:
[0043] This is a sparsity loss function used to penalize the gating coefficients, encouraging the model to let more gating coefficients approach 0, thus achieving feature sparsity. Let be the gating coefficient for the i-th feature. It is the inverse function of Sigmoid. Here, is the temperature parameter, and sigmoid is the activation function.
[0044] L0 norm regularization and physical awareness attention form an adversarial closed loop. When sparse gating attempts to over-prune features, the physical bias matrix acts as a strong baseline, forcibly retaining parameters that are indispensable to the physical causal chain, even if the variance of the current batch of data is small (e.g., small fluctuations in the bath ratio are easily ignored in conventional data fitting, but their physical mass transfer weight is extremely high). This game between physical baseline protection and data pruning solves the fatal flaw of traditional data compression in losing key physical features when encountering unseen conditions.
[0045] S3. Based on the intermediate state characteristics, obtain the color difference prediction value through a multi-scale physical-spatiotemporal manifold coupling dyeing quality prediction model.
[0046] Yarn dyeing is not a steady-state fluid dynamics process, but a highly time-varying nonlinear process that fluctuates dramatically with the process curve. This invention proposes a decoupled and reconstructed architecture that integrates mechanism benchmark, time-domain dynamic compensation, and spatial consistency refinement. By anchoring the mechanism backbone through physical laws, heterogeneous operators are used to capture the physical residuals in the spatiotemporal dimensions, achieving deep structured coupling between mechanism and data.
[0047] First, a mechanistic manifold loss term guided by physical priors is constructed. This term aims to transform the first principles that the coloring process must follow into intrinsic search constraints of the neural network. Specifically, Radial nonstationary mass transfer constraint Based on an improved Fick's second law, the spatiotemporal evolution of dye concentration in the radial field is constrained to ensure that the prediction results conform to the physical monotonicity of mass diffusion.
[0048] Where C(r,t) is the concentration of the dye solution at time t and radial position r. The effective diffusion coefficient is the equation that describes the change of dye concentration with time and radial position, which must satisfy the diffusion law, that is, there is a definite physical relationship between the concentration gradient and the diffusion flux.
[0049] Interfacial adsorption kinetic constraints By using the Langmuir kinetic equation, the chemical equilibrium state of the fiber surface is forcibly embedded into the feature representation layer, thus constraining the saturation evolution of the adsorption amount q(t).
[0050] Where q(t) is the amount of dye adsorbed at time t. For maximum adsorption capacity, The adsorption rate constant is . Here, denoted by η, represents the desorption rate constant, and C represents the bulk concentration of the dye solution. This equation describes the dynamic equilibrium process of dye adsorption from the dye solution to the fiber surface, following the kinetics of monolayer adsorption.
[0051] Hard constraint of matter conservation in all time and space domains Throughout the dyeing cycle, the initial total mass of dye in the dye bath is equal to the sum of the mass of dye adsorbed on the fiber and the mass of dye remaining in the dye bath.
[0052] The global mass conservation equation considering radial distribution is:
[0053] Where C(r,t) represents the dye concentration at time t and radial position r, q(r,t) represents the amount of dye adsorbed at time t and radial position r, and H is the height of the yarn package. This indicates the initial concentration of the dye. Let represent the total volume of the dye solution. This constraint requires that the concentration field C(r,t) and adsorption field q(r,t) predicted by the neural network must satisfy the following at any time t: the total mass of dye within the entire yarn package volume is equal to the total mass of dye introduced at the initial time. This ensures that the model prediction does not violate the fundamental physical law of conservation of mass.
[0054] Secondly, by combining the data-driven loss term, the sparse loss function, and the aforementioned mechanism manifold loss term, the total loss function of dynamic coupling is obtained.
[0055] The complete mathematical expression for the mechanistic manifold loss term is:
[0056] The coloring process must simultaneously satisfy three physical laws, each of which corresponds to a loss term. During training, the four terms are weighted and summed as a penalty signal, forcing the neural network to learn to make predictions that conform to the physical laws. The larger the weight, the more the law is enforced.
[0057] The total loss function is not a simple summation of weights, but rather achieves a dynamic hedging between physical constraints and data fitting by introducing a physically-aware balance operator:
[0058] in, Let be the total loss function for training the neural network. It was designed as a dynamic function based on information entropy gain and process sensitivity. For the mechanistic manifold loss term, For data-driven loss terms, The loss function is sparsity. During periods of drastic physical changes, such as the heating phase, and when sensor noise surges, the model automatically switches to physical constraints (adjusting μ upwards) to suppress data noise using mechanistic stability. During steady-state phases, such as the heat preservation equilibrium phase, the model automatically switches to data correction (adjusting μ downwards) to compensate for simplification biases in the physical equations using measured data. This dynamic governance mechanism transforms the relationship between physical mechanisms and data-driven processes from a simple parallel one to a mutually boundarying, adaptively adjusting virtual-real mapping relationship.
[0059] Furthermore, based on the precise integration after decoupling the physical laws of dyeing at spatiotemporal scales, a multi-scale physical-spatiotemporal manifold coupled dyeing quality prediction model is constructed to characterize the mechanism-spatiotemporal three-scale coupled evolution.
[0060] The multi-scale physical-spatiotemporal manifold coupled staining quality prediction model includes: The physical manifold baseline layer, driven by a physical information neural network, has the core function of outputting preliminary predictions that conform to physical laws. and its implicit physical intermediate state variables These state variables carry deep physical characteristics and are passed to subsequent layers as a physical index, ensuring that subsequent calculations are always performed under the constraints of the physical manifold.
[0061] The temporal unmodeled dynamics compensation layer uses a temporal convolutional network as a residual physical quantity calibrator. It is specifically responsible for capturing unmodeled dynamic residuals caused by pump frequency commutation, transient temperature drift, and factors that cannot be expressed by an ideal PINN PDE. : .
[0062] in Selected features output by the attention-enhancing feature selection module. These are the intermediate physical state variables output from the first layer.
[0063] Non-superposition characteristic: The input of TCN is forced to include the implicit physical state of PINN. This embedded compensation of physical characteristics means that residual correction is not a blind data approximation, but a precise calibration based on the current physical state.
[0064] The spatial field topology consistency refinement layer uses a graph neural network as a spatial field topology regularizer to reconstruct the physical connectivity between yarn bobbins using graph structures.
[0065] Discretize the radial layer into nodes, whose adjacency matrix It is not static, but rather determined by the actual winding density gradient of the yarn package. Constructed permeability resistance weight core:
[0066] in, For the first The new feature matrix after the layer integrates neighborhood information. For activation function, The normalized adjacency matrix, Let be the degree matrix of the adjacency matrix. No. The learnable weight matrix of the layer, after passing through multiple graph convolutions, allows the node features to incorporate information from adjacent layers, thus correcting local anomaly predictions.
[0067] Finally, through logical fusion and Bayesian loop closure, a final staining quality prediction model is formed and color difference prediction values are obtained, satisfying:
[0068] in, This is the final color difference prediction value. The weighting coefficients satisfy the following conditions: Determined through Bayesian optimization on the validation set. These are preliminary forecast values. The compensation value is the output of the unmodeled dynamics compensation layer in the time domain. This is the refined value output by the spatial field topology consistency refinement layer.
[0069] When faced with out-of-distribution samples of new products and dyes, it can automatically degenerate to physical mechanism-driven, using the extrapolation stability of PINN as a baseline; in regular mass production batches, it can activate higher-order corrections through TCN and GNN to pursue high-precision ultimate optimization.
[0070] S4. Perform backpropagation on the mechanism manifold loss term based on the predicted color difference value to obtain the physical sensitivity gradient, and determine the type of physical root cause of the quality defect through the magnitude relationship of the physical sensitivity gradient.
[0071] In this embodiment, the inner, middle, and outer layer chromatic aberration tensors output by the prediction model are monitored in real time: The diagnostic logic is not a simple linear comparison, but rather a violation determination based on the quality and safety manifold: absolute boundary constraint detection: real-time verification of whether the color difference of each layer exceeds the process red line. .
[0072] Radial field gradient stability detection: Calculation of color difference gradient between inner and outer layers: .
[0073] If the gradient exceeds the dynamic relative threshold Even if the single-layer value meets the standard, the system will determine that the potential color difference exceeds the standard due to uneven field distribution; if the violation is determined, the system will automatically and seamlessly switch from the regular monitoring mode to the physical root cause tracing mode.
[0074] The core of root cause analysis lies in using the backpropagation algorithm to analyze the mechanistic manifold loss term of the total loss function. A deep analysis is performed. Instead of calculating the conventional weight gradient, this approach extracts the partial derivatives with respect to the physical latent variables: Permeation kinetics sensitivity gradient This gradient quantifies the total error with respect to the effective diffusion coefficient D. eff The degree of sensitivity essentially reflects the contribution of insufficient fluid dynamics transport capacity to the final quality defects.
[0075] Interface Response Sensitivity Gradient This gradient quantifies the total error with respect to the adsorption rate constant. The sensitivity of the color difference is used to characterize the impact of chemical equilibrium imbalance on color difference fluctuations.
[0076] The two gradient values mentioned above are not only used for diagnosis, but also serve as feedback operators for model self-evolution: Physical prior parameter calibration: When the gradient continues to point to a specific physical constant and the error does not decrease significantly after the actual process is adjusted, the system determines that the equipment mechanism is drifting. At this time, Bayesian posterior update is triggered to recalibrate the physical constants embedded in the model.
[0077] Dynamic weight evolution: The system dynamically adjusts the physical constraint weights in the total loss function based on the historical root cause distribution. With data-driven weights The proportion of this adaptive mechanism ensures that the system can adapt and evolve with equipment aging and environmental changes, significantly improving the robustness of intelligent management and control.
[0078] S5. Combine the quality evaluation indicators and the physical root cause type to trigger or generate process control instructions.
[0079] First, a topology-aware radial flow field consistency feature is defined. This feature introduces the local winding density field of the yarn package and its derived fluid permeation resistance weighting kernel. The second-order spatial partial derivative of the color difference tensor in the radial coordinate system is nonlinearly weighted and penalized to deeply quantify radial distribution differences.
[0080] in, This indicates the number of layers in the radial spatial discretization. For the normalized partition function, This represents the color difference gradient along the radial coordinate of the color difference tensor. For a local infinitesimal space, The core is the permeation resistance weighting core. The physical significance of this weighting core is that if the density gradient between two layers is large, the system applies a specific weighted penalty to the color difference gradient generated in that region, thereby mathematically decoupling the physical layer difference caused by the process preparation from the chemical layer difference caused by the failure of dyeing control.
[0081] Closed-loop feedback mechanism: When the system detects... When the flow deviates from the physical steady-state manifold (below a preset threshold), the system determines that the penetration is uneven. In this case, the system does not simply issue an alarm, but instead uses physical gradient reverse analysis to calculate the dynamic gain function between the main pump's frequency converter and the fluid penetration force in real time, automatically triggering the PLC to adjust the main pump's frequency converter or commutation cycle. At the same time, the evaluation results are simultaneously fed back to the previous process, suggesting that the winding density distribution gradient of the yarn in the next batch be corrected to improve mass transfer conditions from a physical perspective.
[0082] Furthermore, a physical information-driven process-quality sensitivity coefficient and dynamic gain feedback are constructed to satisfy:
[0083] in For process parameters, This represents the inner layer color difference value. This coefficient characterizes the absolute value of the partial derivative of the change in the inner layer color difference value when a certain process parameter changes slightly.
[0084] Closed-loop feedback mechanism: The system calculates the sensitivity of each parameter in real time. If the S of a certain parameter (such as pH value or flow rate) is... param If the value increases significantly, the feedback loop will automatically increase the sampling frequency and weighting coefficient of this parameter in the control system. Simultaneously, the system will issue a hardware self-test command to check the accuracy of the corresponding sensors, ensuring that the control focus remains locked within the critical process window where quality disturbances are greatest.
[0085] Traditional defect risk indices only integrate absolute threshold exceedance and relative threshold deviation. This invention upgrades them to a joint physical-data dual-manifold risk operator. This operator not only evaluates whether color difference exceeds process limits at the data fitting level, but also quantifies the residual deviation of the prediction results from mass conservation and dynamic partial differential equations in the latent feature space.
[0086] in, For the first Color difference value of the layer They represent the inner layer, middle layer, and outer layer, respectively. For absolute boundary thresholds, For the partial differential equation residuals of the corresponding layer, As a measure of the random uncertainty of data, and These are adaptive weights for data and physical risks, respectively. Closed-loop feedback mechanism: It becomes the guiding force in the dynamic game of virtual and real weights in the prediction model. When this operator abnormally increases and A significant increase indicates that the prediction violates physical laws. In this case, the feedback mechanism will immediately increase the weight of the physical constraint in the total loss function. The forced model regression mechanism logic is implemented; if the offset belongs to random noise, the feedback increases the fitting depth of the data-driven term, realizing the adaptive switching of the prediction mode.
[0087] For long-cycle continuous production scenarios, this invention reconstructs the batch quality stability score: a long-term state evolution evaluation metric based on latent space information divergence. This metric uses information theory tools to nonlinearly quantify and track the overall health degradation trajectory of the process system by calculating the Wasserstein distance between the current batch's physically perceived state and the baseline normal state.
[0088] in, Let be the probability distribution of the hidden states in the current batch. Let be the hidden state probability distribution of the baseline batch. The second-order Wasserstein distance. For joint risk operators, , and These are preset parameters.
[0089] Closed-loop feedback mechanism: Responsible for driving the system's long-term self-evolution, when the score shows a monotonically decreasing trend (indicating equipment deterioration or raw material drift), the system automatically triggers the incremental learning module to retrain the model and recalibrates the physical constants at the underlying level. Simultaneously, the system sends early warning signals to the equipment maintenance terminal, proactively tracking process consistency and issuing warnings about equipment pump efficiency deterioration or raw material batch drift, achieving a comprehensive closed loop from underlying algorithm self-calibration to macro-level production management.
[0090] It should be noted that the specific implementation methods described above, such as image processing, numerical simulation, and the construction and training of machine learning models, can all be accomplished by the processor by calling the corresponding computer program instructions stored in memory. Those skilled in the art can implement the above functions using algorithms and tools known in the prior art, according to actual needs.
[0091] Please see Figure 2In an embodiment, to efficiently execute the yarn dyeing quality prediction and control method provided by the present invention, the present invention also provides a yarn dyeing quality prediction and control system, including: an input device 1, an output device 2, a processor 3, and a memory 4. The input device 1, output device 2, processor 3, and memory 4 are interconnected. The memory 4 stores program instructions used to execute the steps of the yarn dyeing quality prediction and control method. The yarn dyeing quality prediction and control system of the present invention has a compact structure and stable performance, and can stably execute the yarn dyeing quality prediction and control method of the present invention, further improving the overall applicability and practical application capability of the present invention.
[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the present invention.
Claims
1. A method for predicting and controlling the dyeing quality of packaged yarn, characterized in that, Includes the following steps: Obtain the data generated from yarn dyeing; A physical adversarial feature selection model for porous topology of bobbin yarn is constructed to obtain intermediate state features; Based on the intermediate state characteristics, the color difference prediction value is obtained through a multi-scale physical-spatiotemporal manifold coupled dyeing quality prediction model; The backpropagation operation of the mechanism manifold loss term is performed based on the predicted color difference value to obtain the physical sensitivity gradient. The type of physical root cause leading to the quality defect is determined by the magnitude relationship of the physical sensitivity gradient. By combining the quality evaluation indicators and the physical root cause type, process control instructions are triggered or generated.
2. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 1, characterized in that, The construction of a physical adversarial feature selection model for porous topology of bobbin yarn, and the subsequent acquisition of intermediate state features, includes the following steps: In physical information neural networks, physical priors are incorporated into attention as attention biases to form a multi-head self-attention mechanism for physical perception. By introducing sparsity constraints to make the gating coefficients sparse, a physical adversarial feature selection model for porous topology of bobbin yarn is constructed. By combining the yarn dyeing generation data and the physical adversarial feature screening model, intermediate state features are obtained.
3. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 1, characterized in that, The step of obtaining the color difference prediction value based on the intermediate state characteristics through a multi-scale physical-spatiotemporal manifold coupled staining quality prediction model includes the following steps: Based on Fick's diffusion law, Langmuir's adsorption kinetics equation, and the all-time-space matter conservation equation, a physical prior-guided mechanism manifold loss term is constructed. By combining the data-driven loss term, the sparse loss function, and the aforementioned mechanism manifold loss term, a dynamic coupling total loss function is obtained; Based on the total loss function, combined with the intermediate state features and the multi-scale physical-spatiotemporal manifold coupled staining quality prediction model, the color difference prediction value is obtained.
4. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 3, characterized in that, Based on Fick's diffusion law, Langmuir's adsorption kinetics equation, and the all-time-space matter conservation equation, a physical prior-guided mechanistic manifold loss term is constructed, satisfying: in, For the mechanistic manifold loss term, The radial nonstationary mass transfer constraint loss term is based on Fick's second law, and the constraint formula is as follows: , For a moment Radial position The concentration of the dye solution at that location. The effective diffusion coefficient; The constraint loss term for interfacial adsorption based on Langmuir dynamics is given by the constraint formula as follows: , For a moment The amount of dye adsorption, For maximum adsorption capacity, The adsorption rate constant is . Let be the desorption rate constant. This represents the main concentration of the dye solution; The loss term is the hard constraint term for matter conservation across all time and space domains, and the constraint formula is as follows: , The height of the yarn package. This represents the initial concentration of the dye. This represents the total volume of the dye solution. The radius of the yarn package; , , These are the preset weighting coefficients.
5. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 3, characterized in that, The combined data-driven loss term, sparse loss function, and mechanistic manifold loss term yield a dynamically coupled total loss function that satisfies: in, Let be the total loss function for training the neural network. It is a dynamic function based on information entropy gain and process sensitivity. For the mechanistic manifold loss term, For data-driven loss terms, This is the sparsity loss function.
6. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 3, characterized in that, The intermediate state features and the multi-scale physical-spatiotemporal manifold coupled staining quality prediction model include: The physical manifold baseline layer is used to output preliminary predictions that conform to physical laws and implicit physical intermediate state variables. A time-domain unmodeled dynamic compensation layer is used to capture unmodeled dynamic residuals caused by pump frequency commutation, transient temperature drift, and physical information that cannot be expressed by ideal partial differential equations of neural networks; The spatial field topology consistency refinement layer is used to reconstruct the physical connectivity between yarn bobbin layers using graph structures.
7. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 1, characterized in that, The physical sensitivity gradient is used to characterize the contribution of different physical dynamic factors to the total error; The physical sensitivity gradient includes: Permeation kinetics sensitivity gradient, used to quantify the sensitivity of error to the effective diffusion coefficient; Interfacial reaction sensitivity gradient, used to quantify the sensitivity of error to the adsorption rate constant; The process control commands include adjusting the frequency or commutation cycle of the main pump of the dyeing machine, or triggering the Bayesian posterior recalibration of the physical constants within the model.
8. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 1, characterized in that, The quality evaluation indicators include the physical-data bimanifold boundary crossing joint risk operator. ,satisfy: in, For the first Color difference value of the layer They represent the inner layer, middle layer, and outer layer, respectively. For absolute boundary thresholds, For the partial differential equation residuals of the corresponding layer, As a measure of the random uncertainty of data, and These are adaptive weights for data and physical risks, respectively. The quality evaluation indicators also include long-term state evolution evaluation metrics based on latent space information divergence. ,satisfy: in, Let be the probability distribution of the hidden states in the current batch. Let be the hidden state probability distribution of the baseline batch. The second-order Wasserstein distance. For joint risk operators, , and These are preset parameters.
9. The method for predicting and controlling the dyeing quality of packaged yarn according to claim 1, characterized in that, It also includes time alignment, data cleaning, and normalization of the dyeing data generated from yarn packages to form a complete monitoring dataset, which serves as the input to the physical information neural network.
10. A yarn dyeing quality prediction and control system, characterized in that, The yarn dyeing quality prediction and control system includes: an input device, an output device, a processor, and a memory, wherein the input device, the output device, the processor, and the memory are interconnected, and the memory stores program instructions, which are used to execute the yarn dyeing quality prediction and control method according to any one of claims 1-9.