A production quality online prediction method for a carding process

Abstract

The application belongs to the technical field of production quality management, and provides a production quality online prediction method for a cotton carding process. The method comprises the following steps: constructing and running a fiber motion mechanism model based on physical laws and a data-driven model based on machine learning; inputting a first quality prediction result and a second quality prediction result into a dynamic weight fusion mechanism, which performs the following operations to generate a high-precision digital twin model: deploying the digital twin model integrated with an optimal weight configuration on an industrial operation and maintenance platform, driving the digital twin model using real-time collected cotton carding process production data, and realizing online dynamic prediction of production quality. The application not only retains the physical interpretability of the fiber motion mechanism model, but also fully utilizes the high-precision advantage of the data-driven model, and significantly improves the prediction accuracy and robustness of the digital twin model under different working conditions through adaptive weight adjustment.

Description

Technical Field

[0001] This invention relates to the field of production quality management technology, and more specifically, to an online prediction method for production quality in the carding process. Background Technology

[0002] Spinning is a fundamental step in the textile industry, and its production quality directly determines the quality and value of subsequent textiles. Carding, as the core of the spinning process, plays a crucial role in breaking down fiber bundles, removing impurities, reducing neps, and forming a uniform cotton web. The quality of the carding process directly affects the evenness, strength, and defect rate of the finished yarn.

[0003] Currently, modeling and prediction methods for carding production quality are mainly divided into three categories: mechanistic modeling, experimental fitting, and data modeling. However, all of these methods have significant limitations. Regarding mechanistic modeling, existing research is primarily based on classical mechanics and aerodynamics, attempting to describe the motion and forces acting on fibers within the carding machine through mathematical equations. However, the sheer number of fibers inside the carding machine and the highly uncertain and nonlinear nature of their motion make it difficult for existing fundamental theories to accurately quantify the microscopic behavior of fiber assemblies. Furthermore, mechanistic modeling typically requires the introduction of numerous idealized assumptions, leading to significant deviations between the established mathematical model and actual production conditions, thus failing to effectively support precise equipment design and online quality control.

[0004] Regarding experimental fitting: This method fits empirical formulas by designing specific experiments and measuring key parameters. Although the experimental method can intuitively reflect some physical laws, its experimental schemes are complex, costly, and the results are subject to chance, making it difficult to cover all complex operating conditions. More importantly, this method is essentially an offline, lagging verification approach, which cannot meet the needs of real-time prediction and control in the production process.

[0005] In terms of data modeling: Data-driven methods such as neural networks have been attempted for yarn quality prediction, which can alleviate the reliance on basic theories to some extent. However, quality data in the carding process mainly relies on offline detection, resulting in a very small number of effective samples available for modeling. Simple data models are prone to overfitting and poor generalization under "small sample" conditions, and the models lack physical interpretability.

[0006] More importantly, existing technologies lack a modeling framework that can effectively integrate the advantages of mechanistic models and data models. Existing simple combination methods often use fixed weights or empirical formulas for fusion, which cannot be dynamically adjusted according to the complex and ever-changing production conditions of the carding process. This results in a significant decrease in the predictive accuracy and robustness of the model when data distribution changes, equipment status fluctuates, or new products are produced.

[0007] Therefore, developing a fusion modeling method that can intelligently balance the reliability of mechanisms and the accuracy of data, and achieve dynamic optimization of weights, has become an urgent need for accurate prediction and control of the quality of the carding process. Summary of the Invention

[0008] Therefore, this invention aims to overcome the shortcomings of the prior art and proposes an online prediction method for the production quality of the carding process. By deeply integrating the mechanism model and the data model, a high-precision digital twin model is constructed to achieve real-time, dynamic, and accurate prediction and optimization of the carding production quality.

[0009] This invention provides an online production quality prediction method for the carding process, comprising the following steps: constructing and running a fiber motion mechanism model based on physical laws and a data-driven model based on machine learning; wherein, the fiber motion mechanism model receives carding process parameters and outputs a first quality prediction result, and the data-driven model receives the same carding process parameters and historical data and outputs a second quality prediction result; inputting the first and second quality prediction results into a dynamic weight fusion mechanism, which performs the following operations to generate a high-precision digital twin model: calculating the uncertainties of the first and second quality prediction results based on the information entropy weight method, and assigning initial weights to the fiber motion mechanism model and the data-driven model accordingly; dynamically adjusting the initial weights of the fiber motion mechanism model and the data-driven model using gradient descent with real-time production data as feedback; globally optimizing the dynamically adjusted initial weight combination using a genetic algorithm to search for the optimal weight configuration; deploying the digital twin model integrated with the optimal weight configuration on an industrial operation and maintenance platform, and using real-time collected carding process production data to drive the digital twin model to achieve online dynamic prediction of production quality.

[0010] The beneficial technical effects of this invention are at least as follows: By constructing a prediction architecture that allows the fiber motion mechanism model and the data-driven model to run in parallel, and employing a three-level dynamic weight fusion mechanism based on information entropy weighting, gradient descent, and genetic algorithms, this invention effectively overcomes the limitations of traditional single modeling methods. This scheme retains the physical interpretability of the fiber motion mechanism model while fully leveraging the high precision advantages of the data-driven model. Through adaptive weight adjustment, it significantly improves the prediction accuracy and robustness of the digital twin model under different operating conditions. The resulting online dynamic prediction system provides a reliable quality monitoring method for the carding process, helping to improve the stability and control efficiency of production quality. Attached Figure Description

[0011] Figure 1 This is a flowchart of an online production quality prediction method for the carding process disclosed in an embodiment of the present invention.

[0012] Figure 2 This is another flowchart of an online production quality prediction method for the carding process disclosed in an embodiment of the present invention.

[0013] Figure 3 This is a flowchart illustrating how the gradient of the second loss function with respect to the weights of the fiber motion mechanism model and the data-driven model is calculated using the gradient descent method, as disclosed in an embodiment of the present invention. Detailed Implementation

[0014] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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.

[0016] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. These embodiments do not constitute a limitation on the embodiments of the present invention.

[0017] like Figure 1 , Figure 2 As shown, this embodiment of the invention discloses an online production quality prediction method 100 for the carding process, including the following steps: Step 10, constructing and running a fiber motion mechanism model based on physical laws and a data-driven model based on machine learning; wherein, the fiber motion mechanism model receives carding process parameters and outputs a first quality prediction result, and the data-driven model receives the same carding process parameters and historical data and outputs a second quality prediction result.

[0018] In this step, the fiber motion mechanism model is a white-box model based on physical laws. This model uses fundamental theories such as elasticity and fluid mechanics as its core, establishing a series of mathematical equations to describe the macroscopic physical processes such as the fiber's trajectory, force conditions, and morphological changes within the carding machine. The model receives carding process parameters such as cylinder speed and flats spacing as input, and outputs a first-order quality prediction result through theoretical calculations. Understandably, this first-order quality prediction result provides a prediction benchmark that conforms to physical laws and has strong interpretability, ensuring the basic reliability of the output, especially under conditions of data scarcity or new operating conditions.

[0019] Data-driven models are gray-box or black-box models based on machine learning algorithms. These models use historical production data as their learning object to extract complex, latent patterns and nonlinear relationships that are difficult for mechanistic models to accurately describe. The model receives the same carding process parameters as the mechanistic model and combines them with historical data to infer and output a second quality prediction result. Under sufficient data, this second quality prediction result can provide high-precision prediction compensation, effectively compensating for deviations between the pure mechanistic model and actual production due to idealized assumptions.

[0020] Step 20: Input the first quality prediction result and the second quality prediction result into the dynamic weight fusion mechanism. The dynamic weight fusion mechanism performs the following operations to generate a high-precision digital twin model: calculate the uncertainty of the first quality prediction result and the second quality prediction result based on the information entropy weight method, and assign initial weights to the fiber motion mechanism model and the data-driven model accordingly; use the gradient descent method with real-time production data as feedback to dynamically adjust the initial weights of the fiber motion mechanism model and the data-driven model; and use a genetic algorithm to globally optimize the dynamically adjusted initial weight combination to search for the optimal weight configuration.

[0021] In this step, a multi-level optimization mechanism is used to achieve intelligent fusion of the fiber motion mechanism model and the data-driven model, constructing a high-precision digital twin model. Specifically, based on the mixed-input physical phenomena, the complex engineering problem is decoupled from three dimensions: the precise interpretability of the mathematical equations, the fuzzy interpretability of the finite element model, and the non-interpretability of the data model. Specifically: First, the information entropy weighting method is used to calculate the information entropy of the first and second quality prediction results based on the initial experimental data, measuring the degree of uncertainty of the output results of the fiber motion mechanism model and the data-driven model. An initial weight allocation function is established based on the information entropy value to assign initial weights to the fiber motion mechanism model and the data-driven model, ensuring that model fusion is initiated under theoretical guidance.

[0022] Secondly, a dynamic weight adjustment mechanism is established using the gradient descent method. Real-time collected production quality data serves as feedback signals, and the error between the comprehensive prediction results of the digital twin model and the actual measured values ​​is used as the loss function. The weight coefficients of the fiber motion mechanism model and the data-driven model are dynamically adjusted through the backpropagation algorithm. This mechanism enables the digital twin model to adapt to different operating conditions such as raw material changes and equipment status fluctuations, achieving data-driven online optimization.

[0023] Finally, a global optimization mechanism is established using a genetic algorithm. The prediction accuracy and robustness of the digital twin model on the validation set are used as the fitness function. Through operations such as selection, crossover, and mutation, a global search is performed in the weight parameter space to find the optimal weight configuration. This process ensures an optimal balance between the interpretability of the fiber motion mechanism model and the generalization ability of the data-driven model, significantly improving the overall predictive performance of the digital twin model.

[0024] Step 30: Deploy the digital twin model with the optimal weight configuration on the industrial operation and maintenance platform, and use the real-time collected production data of the carding process to drive the digital twin model to achieve online dynamic prediction of production quality.

[0025] This step enables the engineering application of digital twin models and real-time monitoring of production quality. In practice, based on an industrial IoT architecture, it integrates sensor networks, PLC controllers, and edge computing devices, employing industrial communication protocols such as OPC UA and MQTT to establish a real-time data acquisition channel for production equipment operating parameters.

[0026] The weighted and optimized digital twin model is deployed on the industrial operation and maintenance platform. Through standardized data interfaces, it is deeply integrated with manufacturing execution systems and enterprise resource management systems, breaking down information silos and enabling collaborative management of production data, quality data, and equipment status data. A visual monitoring interface is built at the platform level, integrating functional modules such as real-time dynamic simulation, historical data traceability, and quality early warning.

[0027] By continuously driving the digital twin model with real-time collected production data from the carding process, a precise mapping between the physical factory and the virtual space is constructed, forming a complete closed loop from data collection and intelligent analysis to feedback control. Ultimately, this enables online dynamic prediction and optimized control of production quality in the carding process, significantly improving the quality stability and production efficiency of the carding process.

[0028] This invention constructs a prediction architecture that allows a fiber motion mechanism model and a data-driven model to operate in parallel. It employs a three-level dynamic weight fusion mechanism based on information entropy weighting, gradient descent, and genetic algorithms, effectively overcoming the limitations of traditional single-modeling methods. This approach retains the physical interpretability of the fiber motion mechanism model while fully leveraging the high precision of the data-driven model. Adaptive weight adjustment significantly improves the prediction accuracy and robustness of the digital twin model under different operating conditions. The resulting online dynamic prediction system provides a reliable quality monitoring tool for the carding process, contributing to improved production quality stability and control efficiency.

[0029] As an example, a fiber motion mechanism model based on physical laws is constructed and run, including: Step 11, defining the initial spatial configuration and connection constraints of the fiber assembly based on topological theory, constructing constitutive equations describing fiber tension, bending and shear deformation based on elasticity theory, and constructing dynamic equations describing the force state of the fiber in the airflow field based on fluid mechanics theory.

[0030] In this step, topological theory is used to mathematically abstract the complex spatial structure of the fiber assembly, defining its initial spatial configurations such as entanglement and intersection, as well as the mutual connection constraints, thus establishing an accurate geometric starting point for subsequent mechanical analysis.

[0031] Based on the theory of elasticity, constitutive equations were constructed to quantitatively describe the basic deformations of single fibers and fiber bundles under the mechanical action of a carding machine, including tension, bending, and shearing. These equations establish the relationship between stress and strain in fibers and form the mathematical core for analyzing their morphological changes.

[0032] Meanwhile, considering fluid mechanics theory, a dynamic equation is constructed to describe the force state of fibers in the high-speed airflow field inside the carding machine, such as traction force and resistance, thus incorporating the influence of airflow into a unified mechanical analysis framework.

[0033] Step 12: Establish a carding cloth-fiber contact mechanics model based on Hertz contact theory to quantify the interaction force between the fiber and the carding teeth; establish an internal aerodynamic model of the carding machine based on the Navier-Stokes equations to calculate the airflow field distribution; and establish a fluid-structure interaction model based on the arbitrary Lagrange-Euler method to solve the motion trajectory and velocity distribution of the fiber in the coupled field.

[0034] In this step, a carding cloth-fiber contact mechanics model based on Hertzian contact theory is established to address the solid-solid interaction between the fiber and the equipment. This model utilizes classical contact mechanics to quantify the collision, compression, and friction forces occurring between the fiber and the carding cloth teeth of the carding machine. These forces directly determine the carding and transfer effects of the fiber and the formation of neps.

[0035] To address the airflow environment inside the equipment, an aerodynamic model of the carding machine based on the Navier-Stokes equations is established. By solving this core fluid dynamics equation, the velocity, pressure, and other distributions of the airflow around components such as the cylinder and licker-in can be calculated.

[0036] A fluid-structure interaction model based on the arbitrary Lagrange-Euler method is established. This method can handle the interaction between large deformation flow and solid motion, and is used to solve the accurate motion trajectory and velocity distribution of fibers under the coupling effect of airflow (flow field) and needle teeth (solid), realizing the dynamic simulation of the entire process of fiber motion.

[0037] Step 13: Establish a fiber breakage mechanism model based on fracture mechanics theory, and construct a quantitative relationship function between fiber breakage probability and carding quality indicators.

[0038] In this step, based on fracture mechanics theory, the mechanism by which internal defects in fibers expand and lead to fracture after being subjected to the complex loads described in the previous steps is analyzed, and a fiber breakage mechanism model is established accordingly. This model allows for the calculation of the fiber breakage probability under specific stress conditions. Finally, through theoretical derivation and data fitting, a clear quantitative relationship function is constructed, directly linking microscopic physical quantities such as the fiber breakage probability with macroscopic carding quality indicators (such as the short fiber content of the sliver), thereby enabling the mechanism model to output quantitative quality prediction values.

[0039] The fiber motion mechanism model of this invention defines the spatial configuration of the fiber through topology, constructs the fiber deformation and force equations based on elasticity and fluid mechanics, quantifies the fiber-needle interaction and airflow field distribution using Hertz contact theory and Navier-Stokes equations, achieves fluid-structure interaction solution using arbitrary Lagrange-Euler method, and finally establishes the quantitative relationship between fiber fracture and quality indicators through fracture mechanics. This enables an accurate description of the fiber motion trajectory, stress state, and quality formation mechanism during the carding process, providing a reliable theoretical basis for production quality prediction.

[0040] As an example, building and running a machine learning-based data-driven model includes: Step 14, building a deep prediction network architecture containing physical hidden layers; and building a first loss function based on physical knowledge to constrain the output of the data-driven model to conform to the basic laws of fiber motion.

[0041] In this step, based on the relevant mathematical equations of the mechanistic model, regularized activation functions and physical attention layers are introduced as key physical hidden layers in deep learning architectures such as LSTM (Long Short-Term Memory) networks. These layer structures are designed to preferentially learn and retain physical features related to fiber mechanical behavior, ensuring that the network's internal computations conform to fundamental physical laws.

[0042] Simultaneously, a first loss function based on physical knowledge is constructed. This first loss function not only includes traditional prediction error terms but also supplements the mechanistic equations through the constructed mathematical theoretical equations, constructs an annealing iteration mechanism based on the annealing algorithm, and ultimately forms a physical constraint term. This constraint term penalizes prediction outputs that violate known physical laws (such as mass conservation and energy conservation), thereby ensuring that the output of the data-driven model conforms to the basic laws of fiber motion from the optimization objective level, effectively reducing the black box effect.

[0043] Step 15: Train the deep prediction network architecture based on simulation data and experimental data, and use optimization algorithms to optimize the model hyperparameters to obtain a fully trained data-driven model.

[0044] In this step, a large number of simulation data samples obtained through finite element analysis and discrete element method are used to train the model on mechanistic data, establishing the model's basic understanding of physical laws. Then, the model is fine-tuned by combining some actual experimental data to ensure its adaptability to actual production processes. This training strategy can effectively expand the scale of training samples and improve the model's generalization ability. A correction factor is constructed to guide the training process, addressing the error between the mechanistic model's calculation results and the actual experimental data.

[0045] Furthermore, PSO (Particle Swarm Optimization) is used to further optimize the model's hyperparameters, while multi-granularity feature extraction methods are employed to extract key data features and optimize the prediction model structure. This optimization process significantly improves the prediction accuracy, stability, and robustness of the data-driven model under small sample conditions, ultimately resulting in a fully trained data-driven model. Optimization algorithms are used to find the optimal model hyperparameters.

[0046] As an example, the fiber motion mechanism model receives carding process parameters and outputs a first quality prediction result, while the data-driven model receives the same carding process parameters and historical data and outputs a second quality prediction result. This includes: inputting the carding process parameters into the fiber motion mechanism model to obtain the first quality prediction result predicted by the fiber motion mechanism model, including the theoretical predicted values ​​of sliver neps and sliver short fiber rate.

[0047] Input the carding process parameters and historical data into the data-driven model to obtain the second quality prediction result of the data-driven model prediction output, including the data-driven prediction values ​​of sliver neps and sliver short fiber rate.

[0048] The carding process parameters include one or more of the following: cylinder speed, flats spacing, doffer speed, licker-in speed, and feed rate; the historical data includes historical values ​​of the carding process parameters, as well as corresponding historical values ​​of sliver neps and sliver short fiber content.

[0049] In this implementation, key carding process parameters from the actual production process, including one or more parameters such as cylinder speed, flats spacing, doffer speed, licker-in speed, and feed rate, are input into a pre-constructed fiber motion mechanism model. Based on its internal physical laws and mathematical equations, the fiber motion mechanism model performs real-time calculations and simulations on these input parameters, and directly outputs the first quality prediction result through theoretical expressions such as established quality characterization functions.

[0050] The first quality prediction result specifically includes the theoretical predicted values ​​of sliver neps and sliver short fiber content. These prediction values ​​are derived purely from physical mechanisms and serve as a prediction benchmark for production quality analysis that conforms to physical laws and has a clear theoretical basis.

[0051] Next, the same real-time carding process parameters, along with the stored historical data, are input into the fully trained data-driven model. The historical data specifically includes historical values ​​of the carding process parameters and their corresponding historical detection values ​​of sliver neps and sliver short fiber percentage obtained through offline detection. The data-driven model utilizes the complex nonlinear mapping relationships learned from massive amounts of historical data, combined with the real-time parameters, to perform comprehensive reasoning and judgment, outputting a second quality prediction result.

[0052] The second quality prediction result specifically includes data-driven predictions of sliver neps and sliver short fiber percentage. These predictions are derived from historical patterns mined by the data-driven model, and can provide high-precision prediction compensation under conditions of sufficient data, effectively capturing potential patterns that mechanistic models may overlook.

[0053] As an example, the uncertainty of the first quality prediction result and the second quality prediction result are calculated based on the information entropy weighting method, and initial weights are assigned to the fiber motion mechanism model and the data-driven model accordingly. This includes: Step 21, calculating the first deviation sequence between the first quality prediction result and the actual quality measurement value, and the second deviation sequence between the second quality prediction result and the actual quality measurement value, based on the initial experimental data.

[0054] This step is based on initial experimental data, which includes the actual production results obtained by running the carding machine under specific process parameters. In practice, actual quality measurements are obtained using specialized testing instruments, including the true measurement results of key quality indicators such as sliver neps and sliver short fiber content. The same experimental operating parameters are then input into both the fiber motion mechanism model and the data-driven model to obtain the corresponding first and second quality prediction results.

[0055] By calculating the deviation between the predicted value and the actual measured value for each experimental sample point, a first deviation sequence characterizing the prediction error of the mechanistic model and a second deviation sequence characterizing the prediction error of the data-driven model are formed. It can be understood that these two deviation sequences objectively record the prediction accuracy of each model in the initial stage.

[0056] Step 22: Based on the distribution of the first deviation sequence and the second deviation sequence, calculate their information entropy respectively to measure the uncertainty of the prediction results of the fiber motion mechanism model and the data-driven model.

[0057] This step analyzes the probability distribution characteristics of the first and second deviation sequences. In practice, the numerical range of each deviation sequence is divided into several intervals, and the frequency of deviation values ​​occurring in each interval is counted to obtain the empirical probability distribution of the deviation sequence.

[0058] Based on this empirical probability distribution, the information entropy of the two deviation sequences is calculated respectively. In information theory, information entropy is an important indicator for measuring the uncertainty of random variables. The more dispersed the distribution of the deviation sequence, i.e., the greater the fluctuation of the prediction results, the higher its information entropy value, indicating that the uncertainty of the model's prediction results is higher; conversely, the more concentrated the distribution of the deviation sequence, i.e., the more stable the prediction results, the lower the information entropy value, indicating that the uncertainty of the model is lower.

[0059] Step 23: Assign high initial weights to models with lower uncertainty based on their information entropy values, and assign low initial weights to models with higher uncertainty.

[0060] In this step, the information entropy values ​​of the corresponding deviation sequences of the fiber motion mechanism model and the data-driven model are compared. Based on the principle that lower uncertainty equates to higher reliability, the model that performs more stably and has lower predictive uncertainty in the initial experiment is assigned a higher initial weight, giving it greater influence in the initial stage of the fusion model. Correspondingly, models with higher predictive uncertainty are assigned lower initial weights.

[0061] The information entropy-based weight allocation method used in this embodiment ensures that the dynamic weight fusion mechanism is established on the basis of an objective evaluation of the model's relative reliability at startup, providing a rational and robust starting point for subsequent online dynamic optimization.

[0062] As an example, the gradient descent method is used to dynamically adjust the initial weights of the fiber motion mechanism model and the data-driven model with real-time production data as feedback. This includes step 24, using the real-time collected production quality data as the true value, and calculating the prediction error between the comprehensive prediction result of the digital twin model and the true value.

[0063] In this step, an online detection system deployed on the carding machine collects real-time production quality data representing the current production status, such as real-time measurements of indicators like sliver neps and sliver short fiber rate. These data serve as the benchmark for evaluating the accuracy of predictions. Simultaneously, the system acquires the combined prediction results from a digital twin model calculated based on the current optimal weight configuration, derived from the fiber motion mechanism model and the data-driven model based on real-time process parameters.

[0064] The real-time prediction error is obtained by calculating the difference between the comprehensive prediction result and the actual value. This error quantifies the prediction accuracy of the current digital twin model and serves as the driving signal for subsequent dynamic adjustments to the weights.

[0065] Step 25: Construct a second loss function based on the prediction error, and calculate the gradient of the second loss function with respect to the weights of the fiber motion mechanism model and the weights of the data-driven model using the gradient descent method.

[0066] In this step, a second loss function (such as mean squared error) is constructed based on the prediction error calculated in step 24. The value of this function directly reflects the comprehensive prediction performance of the digital twin model under the current weight configuration. Subsequently, the gradient descent method is used to calculate the gradient of this second loss function with respect to the weights of the fiber motion mechanism model and the weights of the data-driven model, respectively. Here, the gradient is a vector whose direction indicates the direction in which each weight should be adjusted to make the loss function value decrease the fastest; its magnitude indicates the urgency of the adjustment.

[0067] Step 26: Adjust the weights of the fiber motion mechanism model and the data-driven model in reverse according to the gradient direction and learning rate to minimize the second loss function.

[0068] In this step, based on the gradient direction calculated in step 25 and combined with the preset learning rate, the weights of the fiber motion mechanism model and the data-driven model are adjusted in the opposite direction of the gradient, i.e., along the opposite direction of the gradient, because this direction is the direction that makes the loss function decrease. Specifically, (gradient × learning rate) is subtracted from the current weight value to obtain the updated weight value. In this way, the adjustment of the model weights can be larger and more aggressive when the prediction error is large.

[0069] By iterating the above process, the mechanism can drive the continuous optimization of the weight configuration of the two models. The ultimate goal is to minimize the second loss function, that is, to make the comprehensive prediction result of the digital twin model as close as possible to the real production quality data, thereby enabling the model to adapt to complex and ever-changing production conditions.

[0070] As an example, such as Figure 3 As shown, the gradient of the second loss function with respect to the weights of the fiber motion mechanism model and the data-driven model is calculated using the gradient descent method, including: step 251, calculating the initial gradient of the second loss function with respect to the comprehensive prediction result.

[0071] In this step, the partial derivative of the second loss function (such as mean squared error) with respect to the overall prediction result of the digital twin model is first calculated to obtain the initial gradient. It can be understood that this initial gradient reflects the extent to which a small change in the overall prediction result will cause a change in the value of the second loss function, and its direction indicates the direction in which the overall prediction result should be adjusted to reduce the loss function.

[0072] Step 252: The initial gradient is optimized in multiple steps using a diffusion strategy. Specifically, random noise is added to the current gradient value in each step, and then an optimized gradient that is closer to the true gradient distribution is predicted by a denoising network.

[0073] In this step, in complex production systems, the initial gradient may be uncertain or point to a local optimum due to data noise or model nonlinearity. To address this issue, this invention further introduces a diffusion policy to perform a multi-step iterative optimization of the initial gradient inspired by a diffusion model. This process simulates a denoising process, aiming to recover a more reliable gradient direction that points to the global optimum from noisy or unstable initial gradients.

[0074] In each iteration, random noise of a specific intensity is injected into the current gradient estimate. This operation enables the algorithm to actively explore the gradient neighborhood, helping it escape potential gradient plateaus or local optima.

[0075] A pre-trained denoising network (whose training goal is to learn to recover a smoother and more directional gradient from a noisy gradient) processes the noisy gradient and predicts an optimized gradient. This optimized gradient is closer to the true gradient distribution of the loss function in a global sense, reducing noise and misleading information that may exist in the initial gradient.

[0076] Step 253: The final gradient after multi-step optimization is backpropagated along the path of the comprehensive prediction result, and the first gradient of its weights on the fiber motion mechanism model and the second gradient of its weights on the data-driven model are calculated based on the chain rule.

[0077] In this step, after multiple optimizations using the diffusion strategy, a higher-quality final gradient is obtained. This gradient carries the optimal adjustment direction that the digital twin model should make to minimize the second loss function and the overall prediction results.

[0078] The final gradient is then backpropagated along the computational graph that generates the integrated prediction result. Specifically, the integrated prediction result is a weighted sum of the output of the fiber motion mechanism model (multiplied by its weights) and the output of the data-driven model (multiplied by its weights).

[0079] Based on the chain rule, the total gradient is decomposed through backpropagation. The first gradient of the final gradient with respect to the weights of the fiber motion mechanism model and the second gradient with respect to the weights of the data-driven model are calculated. These two gradients are used for subsequent weight updates to ensure that adjustments to both the fiber motion mechanism model and the data-driven model are made in a way that optimizes the overall performance of the digital twin model.

[0080] This embodiment introduces a diffusion strategy to optimize the gradient in multiple steps, which can significantly improve the accuracy and stability of the weight adjustment direction and effectively overcome the shortcomings of traditional gradient descent methods in complex production environments, such as susceptibility to noise interference and getting trapped in local optima.

[0081] As an example, the diffusion strategy includes a forward diffusion process and a reverse denoising process; random noise is added to the current gradient value in each step, and then an optimized gradient that is closer to the true gradient distribution is predicted by a denoising network, including: step 2521, according to a preset linear noise scheduling table, the intensity of Gaussian noise is gradually increased in the forward diffusion process, and the initial gradient is gradually destroyed into approximately pure noise within T time steps.

[0082] In this step, a linear noise scheduling table with T time steps is defined, where each time step t corresponds to a noise intensity parameter β. t Among them, β t The preset value is incremented with each time step to ensure that the noise intensity increases gradually and smoothly.

[0083] During the forward diffusion process, starting from the initial gradient g0, at each time step, the current noise intensity β is used as the basis for the diffusion. t Gaussian noise is added to the current gradient value. After all T time steps, the initial gradient signal is completely destroyed, transforming into a state g that is approximately pure noise. T .

[0084] Step 2522: The noisy gradient of the current time step, the time step embedding vector, and the real-time operating condition vector are concatenated along the channel dimension to form the combined input of the denoising network.

[0085] This step is the information fusion phase of the reverse denoising process, designed to provide the denoising network with comprehensive contextual information. The combined input to the denoising network consists of three parts: the noisy gradient at the current time step, i.e., the object to be denoised.

[0086] Time step embedding vector: A vector representing which inverse denoising time step the network is currently in, informing the network of the current progress of the denoising process.

[0087] Real-time operating condition vector: A vector that encodes the current carding production status (such as cylinder speed, cotton feed rate, etc.), providing specific physical scene constraints for gradient denoising.

[0088] The three vectors are concatenated along the channel dimension to form a combined input containing noise data, process information, and scenario conditions, enabling the denoising network to perform conditional gradient optimization adapted to specific production conditions.

[0089] Step 2523: The combined input is processed by a denoising network with a U-Net structure, where the encoder extracts multi-scale gradient features by downsampling, and the decoder reconstructs the optimized gradient by upsampling and skip connections.

[0090] In this step, the U-Net network is used to complete the mapping from noisy gradients to optimized gradients, as follows: The encoder part gradually compresses the feature map size through convolution and downsampling operations, while increasing the number of channels, thereby extracting multi-scale gradient features from noisy gradients, from local details to global semantics.

[0091] The decoder gradually recovers the spatial dimensions of the feature maps through transposed convolutions and upsampling operations. Key skip connections directly pass high-resolution feature maps from the corresponding encoder layers to the decoder, helping the network retain more detail when reconstructing and optimizing gradients.

[0092] The entire U-Net takes the aforementioned combined inputs as conditions and finally outputs the denoised result of the current noisy gradient, that is, the optimized gradient estimate.

[0093] Repeat steps 2522 to 2523, updating the current gradient estimate at each time step using the gradient predicted by the denoising network, and obtaining the final optimized gradient after all T time steps.

[0094] This step recovers a high-quality gradient through multiple iterations. Specifically, the reverse denoising process recovers the gradient from pure noise g. T Begin. At each time step t, execute steps 2522 and 2523: input the current noisy gradient, time step information, and operating conditions into the U-Net; the network predicts the optimized gradient for that time step. Using this prediction result, update the current gradient estimate according to the derivation of the diffusion model, obtaining the gradient estimate g for time step t-1. t-1 This update process can be viewed as gradually peeling away a layer of noise from the noise itself, making the gradient estimation increasingly clear.

[0095] The above process is repeated T times until a complete reconstruction from pure noise to a clear gradient is achieved. The final result is the optimized gradient after multiple steps of optimization using the diffusion strategy. It is smoother, more stable, and closer to the globally optimal descent direction than the initial gradient.

[0096] This implementation method constructs a complete forward diffusion and backward denoising process, combined with U-Net gradient optimization guided by operating conditions, to achieve deep purification and reconstruction of the initial gradient, significantly improving the accuracy of weight update direction and model convergence stability.

[0097] As an example, a genetic algorithm is used to globally optimize the dynamically adjusted initial weight combination and search for the optimal weight configuration, including: Step 27, encoding the weight parameters of the fiber motion mechanism model and the data-driven model into chromosomes, and constructing an initial population containing multiple chromosomes.

[0098] In this step, the two parameters to be optimized, Wm (weights of the fiber motion mechanism model) and Wd (weights of the data-driven model), are combined into a complete chromosome using binary encoding or real-number encoding. Each chromosome represents a possible weight configuration scheme.

[0099] Multiple such chromosomes are randomly generated to form an initial population. Each individual (chromosome) in the population represents a weighted combination of the fiber movement mechanism model and the data-driven model, providing a rich starting point for subsequent evolutionary calculations.

[0100] Step 28: Using the prediction accuracy and robustness of the digital twin model on the validation dataset as the fitness function, evaluate the fitness of each chromosome in the population; iteratively evolve the population through selection, crossover and mutation operations, retain chromosomes with high fitness, and explore new weight combinations.

[0101] In this step, a fitness function is defined, whose input is a chromosome (i.e., a set of weight configurations), and whose output is the overall performance of the digital twin model under that weight configuration on an independent validation dataset. This overall performance, for example, needs to take into account both prediction accuracy (such as the average error with respect to the true value) and robustness (such as the variance of the error across all validation samples).

[0102] Selection: Based on fitness scores, individuals with high fitness are selected from the current population using methods such as roulette or tournaments, so that they have a higher probability of passing on their genes (weighted configuration features) to the next generation.

[0103] Crossover: Simulating gene recombination in biological inheritance, two parent chromosomes are randomly selected, and their partial codes (weight parameters) are exchanged to generate new offspring chromosomes. This operation aims to combine the advantages of different optimal weight configurations.

[0104] Mutation: Randomly altering the value of one or more gene loci (weight parameters) in a chromosome with a small probability. This operation aims to introduce new genes into the population, helping the algorithm escape local optima and explore a larger weight space.

[0105] Step 29: When the preset termination condition is reached, the chromosome with the highest fitness is decoded to obtain the optimal weight configuration of the fiber movement mechanism model and the data-driven model.

[0106] In this step, the genetic algorithm iterates continuously until a preset termination condition is met. Common termination conditions include: reaching the maximum number of iterations, the highest fitness in the population no longer significantly improving over multiple generations, or the fitness reaching a satisfactory threshold.

[0107] When the iteration terminates, the chromosome with the highest fitness is selected from the final population. Decoding this chromosome (i.e., the reverse conversion encoding process) yields its specific weight values. This set of weights represents the optimal weight configuration obtained through global optimization search using a genetic algorithm. Assigning this optimal weight configuration to the fiber motion mechanism model and the data-driven model generates a high-precision digital twin model with (approximately) globally optimal overall predictive performance.

[0108] By using chromosome encoding, the weight optimization problem is transformed into an evolutionarily solvable search problem. A fitness function is used to comprehensively evaluate the prediction accuracy and robustness of the digital twin model, effectively avoiding getting trapped in local optima. Selection, crossover, and mutation operations are used to achieve iterative evolution of the population, exploring the optimal weight combination globally and significantly improving the overall prediction performance of the digital twin model. This implementation method can adapt to different production conditions, ensuring that the mechanistic model and the data-driven model achieve optimal fusion, providing high-precision and high-reliability quality prediction capabilities for the carding process.

[0109] As an example, the prediction accuracy is assessed by the mean absolute error between the predicted values ​​of the digital twin model and the actual measured values, and the robustness is assessed by the coefficient of variation of the prediction results of the digital twin model in different production batches.

[0110] In this embodiment, prediction accuracy reflects the accuracy of a single prediction made by the digital twin model, quantifying the average deviation between the model's predicted value and the actual value. Specifically, it is evaluated using the mean absolute error, calculated as follows: ,in, This represents the actual measured value of the i-th sample (such as the true value of the sliver neps or short fibers obtained through offline detection). The MAE value represents the model's predicted value for the i-th sample, where n is the total number of samples used for evaluation. A smaller MAE value indicates that the digital twin model's predictions are closer to the true values, meaning higher prediction accuracy. This metric is insensitive to outliers and robustly reflects the average predictive power of the digital twin model.

[0111] Robustness reflects the ability of a digital twin model to maintain stable predictive performance across different production batches and operating conditions, i.e., its generalization ability and reliability. Specifically, it is evaluated using the coefficient of variation, calculated as follows: ,in, It is the standard deviation of the model's predicted values ​​across multiple different production batch datasets, reflecting the degree of fluctuation in the prediction results; This is the mean of these predicted values. The CV value is a dimensionless relative indicator that eliminates the influence of the mean size on the degree of dispersion. The smaller the CV value, the less the prediction results of the digital twin model fluctuate across different batches, indicating stronger robustness.

[0112] By simultaneously optimizing MAE (pursuing high accuracy) and CV (pursuing high robustness), the fitness function of the genetic algorithm can guide the search direction, ultimately obtaining a high-performance digital twin model that not only performs well on known data but can also adapt to unknown future production conditions.

[0113] Although the invention has been specifically shown and described with reference to preferred embodiments, those skilled in the art will understand that various modifications in form and detail may be made without departing from the spirit and scope of the invention. Accordingly, the disclosed invention should be considered merely illustrative and limited only by the scope specified in the appended claims.

Claims

1. A method for online prediction of production quality in the carding process, characterized in that, Includes the following steps: A fiber motion mechanism model based on physical laws and a data-driven model based on machine learning are constructed and run. The fiber motion mechanism model receives carding process parameters and outputs a first quality prediction result, while the data-driven model receives the same carding process parameters and historical data and outputs a second quality prediction result. The first quality prediction result and the second quality prediction result are input into a dynamic weight fusion mechanism, which performs the following operations to generate a high-precision digital twin model: The uncertainty of the first quality prediction result and the second quality prediction result are calculated based on the information entropy weight method, and initial weights are assigned to the fiber motion mechanism model and the data-driven model accordingly. The initial weights of the fiber motion mechanism model and the data-driven model are dynamically adjusted using the gradient descent method with real-time production data as feedback. The optimal weight configuration is obtained by globally optimizing the dynamically adjusted initial weight combination using a genetic algorithm. The digital twin model, which integrates the optimal weight configuration, is deployed on the industrial operation and maintenance platform. The digital twin model is driven by the real-time collected production data of the carding process, so as to realize online dynamic prediction of production quality. Construct and run a fiber motion mechanism model based on physical laws, including: Based on the theories of elasticity, fluid mechanics and topology, a motion and deformation model of fiber assemblies is established to analyze the tensile, bending and shearing effects on fibers during the carding process. Based on the fluid-solid and solid-solid coupling mechanisms, key models of contact mechanics, aerodynamics, and fluid-solid coupling are established to study the motion trajectory, velocity distribution, and flow law of fibers in the carding machine. Based on fracture mechanics theory, we analyze the breakage mechanism model of fiber bundles under tensile, shear or combined loads, and establish a theoretical expression corresponding to the quality indicators of the carding process. The first quality prediction result includes theoretical predictions of sliver neps and sliver short fiber content; the second quality prediction result includes data-driven predictions of sliver neps and sliver short fiber content; the carding process parameters include one or more of cylinder speed, flats spacing, doffer speed, licker-in speed, and feed rate; the historical data includes historical values ​​of the carding process parameters and corresponding historical detection values ​​of sliver neps and sliver short fiber content.

2. The online production quality prediction method for the carding process according to claim 1, characterized in that: Building and running machine learning-based data-driven models includes: Construct a deep prediction network architecture that includes physical hidden layers; construct a first loss function based on physical knowledge to constrain the output of the data-driven model to conform to the basic laws of fiber motion; The deep prediction network architecture is trained based on simulation and experimental data, and the hyperparameters of the model are optimized using an optimization algorithm to obtain a fully trained data-driven model.

3. The online production quality prediction method for the carding process according to claim 1, characterized in that: The fiber motion mechanism model receives carding process parameters and outputs a first quality prediction result. The data-driven model receives the same carding process parameters and historical data and outputs a second quality prediction result, including: Input the carding process parameters into the fiber motion mechanism model to obtain the first quality prediction result of the fiber motion mechanism model prediction output; The carding process parameters and historical data are input into the data-driven model to obtain the second quality prediction result of the data-driven model prediction output.

4. The online production quality prediction method for the carding process according to claim 1, characterized in that: The uncertainties of the first and second quality prediction results are calculated based on the information entropy weighting method, and initial weights are assigned to the fiber motion mechanism model and the data-driven model accordingly, including: Based on the initial test data, calculate the first deviation sequence between the first quality prediction result and the actual quality measurement value, and the second deviation sequence between the second quality prediction result and the actual quality measurement value. Based on the distribution of the first deviation sequence and the second deviation sequence, their information entropy is calculated respectively to measure the uncertainty of the prediction results of the fiber motion mechanism model and the data-driven model. Based on the information entropy value, models with lower uncertainty are assigned higher initial weights, while models with higher uncertainty are assigned lower initial weights.

5. The online production quality prediction method for the carding process according to claim 4, characterized in that: Using gradient descent with real-time production data as feedback, the initial weights of the fiber motion mechanism model and the data-driven model are dynamically adjusted, including: Using real-time collected production quality data as the true value, the prediction error between the comprehensive prediction result of the digital twin model and the true value is calculated. A second loss function is constructed based on the prediction error, and the gradient of the second loss function with respect to the weights of the fiber motion mechanism model and the weights of the data-driven model is calculated using the gradient descent method. Based on the gradient direction and learning rate, the weights of the fiber motion mechanism model and the data-driven model are adjusted inversely to minimize the second loss function.

6. The online production quality prediction method for the carding process according to claim 5, characterized in that: The gradient of the second loss function with respect to the weights of the fiber motion mechanism model and the data-driven model is calculated using the gradient descent method, including: Calculate the initial gradient of the second loss function with respect to the comprehensive prediction result; The initial gradient is optimized in multiple steps using a diffusion strategy. Specifically, random noise is added to the current gradient value in each step, and then an optimized gradient that is closer to the true gradient distribution is predicted by a denoising network. The final gradient after multiple optimization steps is backpropagated along the path that constitutes the comprehensive prediction result. Based on the chain rule, the first gradient of its weights on the fiber motion mechanism model and the second gradient of its weights on the data-driven model are calculated respectively.

7. The online production quality prediction method for the carding process according to claim 6, characterized in that: The diffusion strategy includes a forward diffusion process and a backward denoising process; random noise is added to the current gradient value at each step, and then an optimized gradient that is closer to the true gradient distribution is predicted through a denoising network, including: According to the preset linear noise scheduling table, the intensity of Gaussian noise is gradually increased during the forward diffusion process, and the initial gradient is gradually destroyed into approximately pure noise within T time steps. The noisy gradient at the current time step, the time step embedding vector, and the real-time operating condition vector are concatenated along the channel dimension to form the combined input of the denoising network. The combined input is processed by a denoising network with a U-Net structure, where the encoder extracts multi-scale gradient features through downsampling, and the decoder reconstructs the optimized gradient through upsampling and skip connections. Repeat the above steps, updating the current gradient estimate at each time step using the gradient predicted by the denoising network, and obtain the final optimized gradient after all T time steps.

8. A method for online prediction of production quality in the carding process according to any one of claims 1-7, characterized in that: A genetic algorithm is used to globally optimize the dynamically adjusted initial weight combination to find the optimal weight configuration, including: The weight parameters of the fiber motion mechanism model and the data-driven model are encoded into chromosomes to construct an initial population containing multiple chromosomes. The prediction accuracy and robustness of the digital twin model on the validation dataset are used as the fitness function to evaluate the fitness of each chromosome in the population. The population is iteratively evolved through selection, crossover and mutation operations to retain chromosomes with high fitness and explore new weight combinations. When the preset termination condition is reached, the chromosome with the highest fitness is decoded to obtain the optimal weight configuration of the fiber movement mechanism model and the data-driven model.

9. The online production quality prediction method for the carding process according to claim 8, characterized in that: The prediction accuracy is assessed by the mean absolute error between the predicted values ​​of the digital twin model and the actual measured values, and the robustness is assessed by the coefficient of variation of the prediction results of the digital twin model in different production batches.

Images