Cotton yarn single yarn strength prediction method and system based on machine learning

CN122927592APending Publication Date: 2026-10-16XINJIANG PUMEI MEDICAL TECHNOLOGY CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202610810650.5
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-06-05
Publication Date
2026-10-16

AI Technical Summary

Technical Problem

[0004]针对现有技术存在的无法将原料与环境的交互关系下沉到物理交互层面进行表征与推理,导致模型在未知工况下产生物理悖论及无法实时进化的问题,本申请通过基于机器学习的棉纱单纱强力预测方法及系统,构建内化物理交互机理的物理状态计算图并配备局部在线演化机制,实现预测过程的物理一致性与模型知识的实时安全进化

Benefits of technology

[0015]This application provides a machine learning-based method and system for predicting the strength of single cotton yarn. It transforms strength prediction into a computational deduction process that allows environmental media to penetrate the matrix microstructure under the constraint of a physical potential energy field, generating a mechanical response. This ensures that even under extreme or unseen conditions, the reasoning logic of the physical state calculation graph inevitably triggers a significant inhibitory correction increment in the second calculation unit. This correction is subject to the mandatory constraint of the physical potential energy function, resulting in a stable strength prediction that reflects a decay trend. This fundamentally eliminates the possibility of erroneous predictions of strength increasing against the trend of humidity under cross-conditions and elevates the physical reliability of the extrapolation to an extremely high level. Unlike existing solutions that only vaguely suggest curve shapes in the loss function, this solution internalizes physical laws as hard constraints within the calculation graph, eliminating any abnormal outputs that violate thermodynamic trends. Meanwhile, the physical state computation graph, leveraging its internalized cross-scale physical causal reasoning chain and practical experience accumulated in local evolution, can perform first-principles reasoning roams in the state space of any combination of raw materials and environment, spontaneously generating an infinite amount of high-fidelity training data containing complete physical deduction trajectories. This data naturally avoids the physical taboos that statistical models are prone to commit. When used for consolidation training of any downstream model, it can effectively improve its generalization boundary, thus eliminating the need for expensive temperature and humidity chamber experiments and the construction of unstable generative adversarial networks during the training process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122927592A_ABST
    Figure CN122927592A_ABST
Patent Text Reader

Abstract

The application relates to the field of quality control and machine learning of the textile industry, and provides a cotton yarn single yarn strength prediction method and system based on machine learning, which comprises the following steps: acquiring characteristic data and environmental medium parameters of a prediction object; constructing a physical state calculation graph, and performing iterative reasoning on the data to obtain a prediction value, wherein the calculation graph comprises state nodes representing inherent properties, environmental medium penetration and target responses and calculation units simulating physical interaction effects; in the iterative reasoning, a physical potential energy constraint is applied to the target response node, so that when the environmental medium factor is forced to be raised, the upper limit of the target response is reduced; the time sequence traces of the calculation units are recorded, and based on the measured error and the time sequence traces, the strong time sequence traces are calculated, and the local directional update of the internal weights of the calculation units is performed. The application internalizes the physical law as a hard constraint of the calculation graph, eliminates the physical paradox, and realizes the local directional safe real-time evolution of the model knowledge.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of quality control and machinery in the textile industry, and in particular to a method and system for predicting the strength of cotton yarn single yarn based on machine learning. Background Technology

[0002] Currently, in the field of quality control in the textile industry, using machine learning models to predict the strength of cotton yarn is a core technological path to replace traditional destructive testing and achieve full inspection. Existing technologies typically collect online characteristics of cotton yarn during the production process (such as yarn diameter variation coefficient, unevenness, hairiness index, and workshop temperature and humidity) and offline raw material characteristics (such as micronaire value, short fiber rate, and breaking strength from HVI high-volume fiber test data). These multi-source data are concatenated into a high-dimensional feature vector, which is then input into black-box regression models such as Gradient Boosting Tree (GBDT) and Deep Neural Network (DNN) to train them to fit the statistical mapping relationship between the feature vector and the single yarn strength. To address the coupled influence of raw materials and environmental humidity on strength, existing technologies have developed methods such as explicitly constructing interactive features, multi-model integration, and physical information neural networks.

[0003] However, in long-term deployment on actual production lines, when changing raw material batches, the existing models initially show a significantly higher strength prediction for the new batch, exhibiting a dangerous tendency of "statistical averaging," incorrectly transferring historical experience with high-maturity raw materials to current low-maturity raw materials. Furthermore, under the combined conditions of "low micronaire value × high moisture regain," even with the introduction of physical loss function constraints, the existing models sometimes still output predictions that violate physical laws, i.e., the predicted strength increases against the trend as moisture regain increases. When using training data based on simple polynomials or generative adversarial networks to synthesize training data under extreme conditions, the model's prediction accuracy actually decreases because the synthesized data only statistically mimics macroscopic labels and does not internalize the causal mechanism of the interaction between moisture and fiber microstructure. Summary of the Invention

[0004] To address the problem that existing technologies cannot characterize and reason about the interaction between raw materials and the environment at the physical interaction level, leading to physical paradoxes and the inability to evolve in real time under unknown working conditions, this application proposes a machine learning-based method and system for predicting the strength of single cotton yarn. This method constructs a physical state calculation graph that internalizes the physical interaction mechanism and is equipped with a local online evolution mechanism, thereby achieving physical consistency in the prediction process and real-time secure evolution of model knowledge.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: Firstly, a machine learning-based method for predicting the strength of single cotton yarn includes: acquiring feature data and environmental medium parameters of the target object; constructing a physical state calculation graph and iteratively reasoning on the feature data and environmental medium parameters to obtain predicted values; the physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and calculation units, which are used to simulate the physical interaction effects between state nodes; in the iterative reasoning, applying physical potential energy constraints to the third state node to force a decrease in the upper limit of the target response when the environmental medium factor increases; recording the temporal traces of the calculation units, and based on the measured error and temporal traces corresponding to the predicted values, performing localized targeted updates of the internal weights of calculation units with strong temporal traces.

[0006] In conjunction with the first aspect described above, in one possible implementation, the computational unit includes a first computational unit and a second computational unit. The first state node is generated from feature data through an encoded mapping. The second state node is generated from environmental medium parameters through a dynamic mapping that includes learnable attenuation parameters. The first computational unit takes the first and second state nodes as inputs and, through gated interactive computation, outputs a correction increment for the second state node. The second computational unit takes the updated second and third state nodes as inputs, calculates the attenuation coefficient, and outputs a suppressive correction increment for the third state node.

[0007] In conjunction with the first aspect mentioned above, in one possible implementation, the encoding mapping is achieved through a self-supervised pre-trained encoder. The encoder's pre-training process specifically includes: inputting the first state node output by the encoder along with given environmental medium parameters into the decoder; predicting the baseline response value of the object under the baseline environmental medium parameters; and training the encoder by minimizing the error between the predicted baseline response value and the true baseline response value. The kinetic mapping is achieved through a hygroscopic kinetic mapping function, the mathematical form of which is: .in, Let be the water infiltration state vector of the second state node. It is the hyperbolic tangent function. The real-time moisture regain scalar value in the environmental medium parameters. The decay rate matrix is ​​a learnable matrix. The average of historical moisture retention data. The standard deviation of historical moisture content data. For bias terms, For element-wise multiplication, It is an activation function that approximates ReLU with smoothness.

[0008] In conjunction with the first aspect mentioned above, in one possible implementation, the physical potential energy constraint is a physical potential energy function, which measures the degree of deviation of the components of the third state node from a threshold that is negatively correlated with the components of the second state node. After updating the third state node, gradient descent is used to pull the third state node in the direction that the physical potential energy function decreases.

[0009] In conjunction with the first aspect mentioned above, in one possible implementation, the mathematical expression for the physical potential energy function is: .in, This is the brute force state vector of the third state node. Let be the water infiltration state vector of the second state node. The physical potential energy function, for and vector dimension, and Let i represent the force state component and the water permeability state component in dimension i, respectively. This is a configurable positive management coefficient used to adjust the upper limit of the suppression of the strong state by the water permeability state. The update formula for pulling the third state node in the direction of decreasing physical potential energy function is: .in, This is the updated brute force state vector of the third state node. The preset potential energy decay step size, The potential energy function for the third state node The gradient.

[0010] In conjunction with the first aspect mentioned above, in one possible implementation, the process of acquiring the temporal trace specifically includes: in the current round of iterative inference, obtaining the sum of the absolute values ​​of the correction vectors output by the computational unit as the instantaneous activation value of that computational unit; generating a temporal decay factor that follows an exponential decay law based on the difference between the current round and the preset total number of iterations; and denoting the instantaneous activation value and the temporal decay factor as the temporal trace of the computational unit in the current round of iterative inference.

[0011] In conjunction with the first aspect mentioned above, in one possible implementation, the process of locally directionally updating the internal weights of computational units with strong temporal traces specifically includes: acquiring the prediction error and generating a global modulation signal proportional to the prediction error through an evolutionary controller; accumulating all temporal traces from the current prediction process for each computational unit to obtain a total temporal trace; scaling the gradient of the predicted value with respect to the internal weights of the corresponding computational unit using the product of the global modulation signal and the total temporal trace as a scaling factor to generate a local weight update amount; and accumulating the local weight update amount to the internal weights of the computational unit to complete the local directional update.

[0012] In conjunction with the first aspect described above, one possible implementation further includes: during offline periods, sampling the intrinsic property states and virtual environment medium parameters based on the physical state computation graph. The sampled data is then input into the physical state computation graph for complete graph diffusion inference, generating corresponding predicted response values. An augmented dataset, composed of the intrinsic property states, virtual environment medium parameters, and predicted response values, is used for offline consolidation training of downstream models.

[0013] In conjunction with the first aspect mentioned above, in one possible implementation, the object to be predicted is cotton yarn. The inherent property state is the microstructure state of the fiber matrix, the environmental medium is moisture, the target response state is the strength state, the characteristic data is HVI test data, and the environmental medium parameter is moisture regain.

[0014] Secondly, a machine learning-based cotton yarn strength prediction system is provided, comprising: a data acquisition unit for acquiring cotton yarn feature data and environmental medium parameters; a graph construction and inference unit for constructing a physical state calculation graph and iteratively inferring the feature data and environmental medium parameters to obtain predicted values; the physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and a calculation unit for simulating the physical interaction effects between state nodes; a physical potential energy constraint unit for applying physical potential energy constraints to the third state node during iterative inference to force a decrease in the upper limit of the target response when the environmental medium factor increases; and a time-series trace recording and updating unit for recording the time-series traces of the calculation unit and, based on the measured error and time-series traces corresponding to the predicted values, performing localized targeted updates of the internal weights of calculation units with strong time-series traces.

[0015] This application provides a machine learning-based method and system for predicting the strength of single cotton yarn. It transforms strength prediction into a computational deduction process that allows environmental media to penetrate the matrix microstructure under the constraint of a physical potential energy field, generating a mechanical response. This ensures that even under extreme or unseen conditions, the reasoning logic of the physical state calculation graph inevitably triggers a significant inhibitory correction increment in the second calculation unit. This correction is subject to the mandatory constraint of the physical potential energy function, resulting in a stable strength prediction that reflects a decay trend. This fundamentally eliminates the possibility of erroneous predictions of strength increasing against the trend of humidity under cross-conditions and elevates the physical reliability of the extrapolation to an extremely high level. Unlike existing solutions that only vaguely suggest curve shapes in the loss function, this solution internalizes physical laws as hard constraints within the calculation graph, eliminating any abnormal outputs that violate thermodynamic trends. Meanwhile, the physical state computation graph, leveraging its internalized cross-scale physical causal reasoning chain and practical experience accumulated in local evolution, can perform first-principles reasoning roams in the state space of any combination of raw materials and environment, spontaneously generating an infinite amount of high-fidelity training data containing complete physical deduction trajectories. This data naturally avoids the physical taboos that statistical models are prone to commit. When used for consolidation training of any downstream model, it can effectively improve its generalization boundary, thus eliminating the need for expensive temperature and humidity chamber experiments and the construction of unstable generative adversarial networks during the training process. Attached Figure Description

[0016] Figure 1 A flowchart illustrating a machine learning-based method for predicting the strength of single cotton yarns, provided in an embodiment of this application; Figure 2 A detailed flowchart of physical state calculation graph construction and iterative reasoning in a machine learning-based method for predicting the strength of single cotton yarn provided in this application embodiment; Figure 3 The physical state calculation diagram contains the internal structure and interaction relationship diagram of a machine learning-based method for predicting the strength of a single cotton yarn, as provided in this application embodiment. Figure 4 A flowchart of the self-supervised encoder pre-training process in a machine learning-based method for predicting the strength of a single cotton yarn provided in this application embodiment; Figure 5 In a machine learning-based method for predicting the strength of a single cotton yarn provided in this application embodiment, the moisture absorption dynamics are mapped to a physical and mathematical duality diagram. Figure 6 The principle and execution flowchart of the physical potential energy hard constraint mechanism in a machine learning-based method for predicting the strength of a single cotton yarn provided in this application embodiment; Figure 7A comparative analysis of the effects of physical hard constraints and traditional soft constraints in a machine learning-based method for predicting the strength of single cotton yarn provided in this application embodiment; Figure 8 A flowchart of a local directional update mechanism based on time-series traces in a machine learning-based method for predicting the strength of a single cotton yarn provided in this application embodiment; Figure 9 In a machine learning-based method for predicting the strength of a single cotton yarn provided in this application embodiment, a time-series diagram of time-series trace recording and locally directional update data flow is included. Figure 10 The flowchart of offline enhanced data generation in a machine learning-based method for predicting the strength of single cotton yarn provided in this application embodiment; Figure 11 A schematic diagram of the structure of a machine learning-based cotton yarn strength prediction system provided in an embodiment of this application; Figure 12 This diagram illustrates the adaptive prediction and safety evolution process of a machine learning-based cotton yarn strength prediction system under extreme conditions, as provided in an embodiment of this application. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terms “comprising” or “including” and similar words as used in the specification and claims of this invention mean that the element preceding the word encompasses the element following the word, without excluding other elements not listed. The terms “first,” “second,” and similar words used herein are used only to distinguish between similar things and do not imply any particular order, importance, or arrangement.

[0019] Example 1: like Figure 1 , 2As shown in Figures 6 and 8, this embodiment provides a method for predicting the strength of single cotton yarn based on machine learning. This method abandons the static paradigm of the traditional black-box model that directly maps macroscopic features to strength values, and instead constructs a dynamic computing architecture that internalizes physical interaction mechanisms. It is equipped with a local online evolution mechanism based on time-series tracing, thereby establishing a defense depth of physical consistency and continuous adaptation at the macroscopic process level.

[0020] It should be understood that the process architecture described in this embodiment is explanatory rather than restrictive, and is intended to provide sufficient public support for the widest range of protection. Specific algorithm formulas and parameter decomposition will be further elaborated in subsequent embodiments.

[0021] Step 100: Obtain the feature data and environmental medium parameters of the object to be predicted.

[0022] The predicted object refers to the physical entity whose target response state needs to be estimated. In the preferred application scenario of the present invention, it is cotton yarn, but in other embodiments it can also be other fiber aggregates or porous materials affected by the permeation of environmental media.

[0023] Feature data refers to the original data set that can characterize the inherent physical properties of the object being predicted, such as HVI test data for cotton yarn, which includes macroscopic indicators such as micronaire value, short fiber rate, and breaking strength.

[0024] Environmental media parameters refer to physical quantities that characterize the real-time state of the external permeation environment, such as the moisture regain or temperature of the workshop.

[0025] In some implementations, the above data can be acquired by sensors deployed on the production line (such as online moisture regain detectors) and offline testing equipment (such as HVI high-capacity fiber testers), and then transmitted to the processor's storage unit for subsequent use.

[0026] Step 200: Construct a physical state calculation graph and perform iterative reasoning on the feature data and environmental medium parameters to obtain predicted values. The physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and a calculation unit. The calculation unit is used to simulate the physical interaction effects between the state nodes.

[0027] The physical state computation graph is not a static knowledge graph that records entity relationships in the traditional sense, but a differentiable dynamic computation network with physical interaction mechanisms as its framework. It transforms macroscopic input parameters into a continuous evolution process of microscopic physical states.

[0028] The first state node is a high-dimensional vector representation of the inherent microstructural properties of the predicted object that do not change instantaneously with the environment, such as the crystallinity and orientation of the fiber matrix.

[0029] The second state node is a dynamic vector representation of the state of penetration and binding of the environmental medium within the object, such as the differentiated adsorption state of water in the amorphous region and crystalline region of the fiber.

[0030] The third state node is a vector representation of the response state of the predicted object to external forces under the combined action of the current microstructure and media penetration, such as the strong force state.

[0031] The computational unit is a miniature neural network module embedded in the computational graph, used to simulate the physical causal interactions between nodes. For example, it is an interactive unit that simulates the modification of microstructure state by water adsorption, and an interactive unit that simulates the suppression of strong state by water plasticization.

[0032] In some implementations, three types of state nodes can be initialized first based on feature data and environmental medium parameters. The interaction effects between states can be simulated through multiple iterations by the computing unit, and the second and third state nodes can be continuously updated. Finally, the third state node that has converged in the iteration is mapped to a scalar prediction value.

[0033] This computational graph iterative reasoning differs from traditional black-box mapping. Traditional black-box mapping only establishes statistical correlations from input to output and lacks physical interpretability of intermediate states. In contrast, the computational graph in this scheme decomposes the reasoning process into a micro-state evolution chain that conforms to physical causality, making the changes in the predicted value at each step traceable and fundamentally blocking abnormal deduction paths that violate physical laws.

[0034] Step 300: In the iterative reasoning, a physical potential energy constraint is applied to the third state node to force the upper limit of the target response to decrease when the environmental medium factor increases.

[0035] Among them, the physical potential energy constraint is a mathematical potential energy field constructed based on thermodynamic trends, which measures the degree of anomaly in the deviation of the third state node from the physical negative correlation law.

[0036] In some implementations, after each iteration updates the third state node, the deviation of the node component from the environmental medium suppression threshold is calculated in real time, and based on the descent principle, the third state node is directly pulled back to the direction of decreasing potential energy.

[0037] It is important to note that this physical potential energy constraint is a "hard constraint" in the iterative inference process of the computational graph, rather than the "soft suggestion" in traditional physical information neural networks that merely adds a penalty term to the loss function. Traditional soft suggestions only apply macroscopic fuzzy guidance to the model parameters during the training phase; during the inference phase, the model may still output a strong physical paradox of rising humidity due to data distribution shifts. In contrast, the hard constraint in this scheme intervenes directly at the state space level of the inference computational graph. Once the state vector attempts to exceed the physical negative correlation threshold, the potential energy field immediately generates a gradient to forcibly pull it back, thus eliminating any possibility of physical paradox output at the underlying mechanism level and raising the physical reliability of the extrapolation to an extremely high level.

[0038] Step 104: Record the time series traces of the computing unit, and based on the measured error and time series traces corresponding to the predicted value, perform local directional updates of the internal weights for computing units with strong time series traces.

[0039] Among them, the temporal trace is a dynamic scalar that simulates the causal tracing mechanism of neural synapses, quantifying the real-time contribution and causal weight of the computing unit to the final prediction result during this iterative inference process.

[0040] Localized targeted updates refer to making minor parameter adjustments only to computational units that leave strong temporal traces in the current inference, rather than making indiscriminate modifications to the entire model.

[0041] In some implementations, during each iteration of forward inference, the magnitude of the correction vector output by each computational unit and its temporal position are automatically recorded and comprehensively calculated as a temporal trace. When the production line laboratory returns the measured strength value and calculates the prediction error, the system generates a global modulation signal proportional to the error, and uses the product of this signal and the temporal trace as a scaling factor to update the gradient only on the internal weights of the strong trace units.

[0042] This localized, targeted update mechanism effectively addresses the pain point of catastrophic forgetting in traditional online learning. It can accurately locate the causal responsibility unit for the current inference based on temporal traces, and only fine-tune these units, achieving millisecond-level adaptation to new scenarios while perfectly preserving the inference capability for historical scenarios, ensuring the model's continuous and safe evolution throughout its entire lifecycle.

[0043] Based on the above technical solution, a macroscopic defense depth system was established, encompassing the microscopic evolution of physical states, the hard constraint guarantee of potential energy, and the local evolution of temporal traces. This solution not only elevates strong prediction from statistical fitting to physical causal inference, eliminating physical paradoxes, but also, within milliseconds of receiving real feedback from each laboratory, it only requires one forward propagation to calculate the temporal traces and one backpropagation to update a few computational units to internalize the interaction law of new raw materials and environment. This locally targeted update mechanism reduces the model's adaptation latency to new cotton blends and new temperature and humidity from hours to milliseconds, while updates are strictly limited to computational units that causally contribute to the current inference, completely avoiding the catastrophic forgetting risk common in traditional online learning, and achieving stable and continuous evolution of the model throughout the entire production line lifecycle. Combined with the zero-cost, unlimited, physically consistent, high-fidelity data generation of the physical state computation graph, the model is endowed with the ability to continuously grow endogenously under complex production line conditions, achieving a leapfrog improvement in prediction accuracy and low-risk adaptive effects.

[0044] Example 2: like Figure 2-5 As shown, this embodiment expands upon step 200 of embodiment 1 by further developing the process of constructing the graph ontology and initializing nodes of the physical state calculation graph.

[0045] It should be understood that the upper-level architecture described in Embodiment 1 is explanatory rather than restrictive. This embodiment will deeply bind the architecture with specific physical objects and algorithm logic to establish a mapping defense line from macroscopic test data to microscopic physical state vectors.

[0046] First, the preferred application scenario for the physical semantic mapping of this invention is clarified. The prediction object is cotton yarn; The inherent property state is the microstructure state of the fiber matrix. The environmental medium is moisture. The target response state is the strength state. This fiber matrix microstructure state does not refer to a single macroscopic scalar index (such as micronaire value), but rather to the set of inherent microstructural properties within the cotton fiber aggregate that affect its interaction with moisture and mechanical behavior. These properties include cellulose crystallinity, microfibril orientation, accessibility of amorphous regions, and the initial cohesion morphology between fibers, among other comprehensive physical states. These microscopic states are difficult to measure directly with conventional instruments, but they are the fundamental intrinsic factors determining the differences in strength decay of cotton yarn at different moisture regain rates. The characteristic data is HVI test data, which includes standard test indicators such as micronaire value, maturity ratio, short fiber percentage, breaking strength, and breaking elongation, serving as macroscopically measurable input. The environmental medium parameter is moisture regain, which refers to the macroscopic statistical proportion of environmental moisture in the fiber aggregate, and serves as a real-time parameter of the environmental medium.

[0047] This invention can endow abstract computation graph nodes with specific physical semantics of cotton yarn, so that subsequent algorithm iterations are no longer purely mathematical operations, but digital twin deductions of the real physical process of water penetrating into specific microstructures and causing strong attenuation.

[0048] The specific generation process of the three types of state nodes and the internal structure of the two types of computing units include: The first state node is generated by encoding and mapping feature data. This first computational unit physically simulates the water adsorption and storage capacity of the fiber microstructure, and its internal network structure is preferably a fully connected network containing a single hidden layer.

[0049] In some implementations, the network first calculates the gating interactions between the first state node and the second state node in each dimension, and generates a correction vector with the same dimension as the second state node as the correction increment.

[0050] For example, the dimension corresponding to "accessibility of amorphous regions" allows the water permeation state vector to rapidly approach saturation in this dimension, while the dimension corresponding to "inaccessibility of crystalline regions" suppresses the increment of the water permeation state vector in this dimension to near zero, thereby accurately simulating the differentiated constraints of different microstructural features on water adsorption. The second state node is generated from environmental medium parameters through a dynamic mapping that includes learnable decay parameters; The first computing unit takes the first state node and the second state node as input, and calculates the correction increment for the second state node through gating interaction. The second calculation unit takes the updated second and third state nodes as inputs, calculates the attenuation coefficient, and outputs the inhibitory correction increment for the third state node.

[0051] The second computational unit physically simulates the plasticizing, swelling, and weakening effects of moisture ingress on fiber mechanical strength. Internally, it is preferably a fully connected network containing a single hidden layer. This network calculates the attenuation coefficients of the updated moisture penetration state vector on each dimension of the strength state vector, generating negative inhibitory correction increments, thereby achieving a digital expression of the core physical causal law that "high humidity leads to strength attenuation."

[0052] This design, which explicitly divides the computational units into two sub-networks—gated interaction and attenuation suppression—effectively eliminates the physical meaningless general attention mechanism found in traditional black-box models. Instead, it adopts structured computational modules that are strictly dual to the real physical link of "adsorption-plasticization-mechanical response," ensuring that every state transition within the computational graph has clear microscopic mechanism support, fundamentally blocking abnormal deduction paths that violate physical laws.

[0053] Furthermore, the algorithmic details of encoding mapping and dynamic mapping are explored in detail.

[0054] The encoding mapping is achieved through a self-supervised pre-trained encoder. The pre-training process of the encoder specifically includes: inputting the first state node output by the encoder along with the given environmental medium parameters into the decoder, predicting the benchmark response value of the object under the benchmark environmental medium parameters, and training the encoder by minimizing the error between the predicted benchmark response value and the true benchmark response value.

[0055] The self-supervised pre-trained encoder is a small neural network composed of fully connected layers. Its input is the complete HVI test vector of a batch of raw materials, and its output is the first state node vector with a fixed dimension (e.g., 16-dimensional or 32-dimensional). The preferred standard moisture regain parameter is 7.5%.

[0056] During pre-training, a corresponding decoder network is constructed. The first state node output by the encoder, along with a given moisture regain of 7.5%, is used as the input to the decoder to predict the strength value of the raw material at the standard moisture regain. The encoder is trained by backpropagation by minimizing the error between the predicted strength and the actual strength, forcing the first state node vector to necessarily contain potential microstructural information directly related to the strength response that cannot be expressed by a single macroscopic indicator.

[0057] This allows us to utilize the physical fact that different microstructures exhibit different strengths under the same baseline humidity as a monitoring signal. Without the need for expensive microstructure slicing tests, we can inversely derive the implicit microstructure state vector from macroscopic HVI data, achieving a leapfrog mapping from macroscopic appearance to microscopic essence.

[0058] The kinetic mapping is achieved through a hygroscopic kinetic mapping function, the mathematical form of which is: ; in, Let be the water infiltration state vector of the second state node. It is the hyperbolic tangent function. The real-time moisture regain scalar value in the environmental medium parameters. The decay rate matrix is ​​a learnable matrix. The average of historical moisture retention data. The standard deviation of historical moisture content data. For bias terms, For element-wise multiplication, It is an activation function that approximates ReLU with smoothness.

[0059] In some implementations, the measured real-time moisture regain scalar is first... Standardization is performed by subtracting the historical mean μ and dividing by the standard deviation σ, transforming it into a statistical distribution with zero mean and unit variance. This improves the stability and convergence speed of subsequent neural network calculations. Then, the hyperbolic tangent function tanh is used to standardize the real-time moisture regain scalar. Nonlinear compression is applied to map it to the interval (-1, 1). In this case, the physical meaning of the tanh function is to simulate the saturation trend during the water adsorption process. When the moisture regain is extremely low, the adsorption rate increases approximately linearly. When the moisture regain is extremely high, the fiber's moisture absorption capacity tends to saturate, and the increase in adsorption gradually decreases and approaches the limit. This perfectly matches the S-shaped physical curve characteristic of the cotton fiber moisture absorption isotherm.

[0060] Subsequently, a learnable decay rate matrix can be introduced. Perform an element-wise product with the output of tanh. At this point, the... The matrix is ​​preferably a square matrix with dimensions of 16×16 or 32×32, which matches the dimension of the state vector. Its physical meaning is to simulate the differentiated adsorption rates of water by microstructural features of different dimensions. For example, In the matrix, the element values ​​corresponding to the "accessibility of amorphous regions" dimension are relatively large, indicating that this dimension is sensitive to water adsorption and has a fast rate; while the element values ​​corresponding to the "inaccessibility of crystalline regions" dimension are extremely small, indicating that this dimension is insensitive to water adsorption and has a slow rate.

[0061] This element-wise multiplication operation allows a single scalar moisture regain input to be decoupled into a multi-dimensional, heterogeneous moisture permeation state, truly reflecting the differences in the depth of moisture permeation in different microstructural regions within the fiber.

[0062] Finally, a bias term b is added, and the activation function Softplus, which approximates ReLU with smoothness, is input, i.e., log(1+exp(x)). The physical meaning of the Softplus function is to ensure that the output water permeability state vector... It is strictly non-negative in all dimensions because the amount of water penetration and binding cannot be negative physically, and its smoothness avoids the gradient dead zone of the ReLU function at zero, ensuring that the computation graph is differentiable across the entire range, thus supporting subsequent gradient iterative inference and local directional updates.

[0063] Through the above-mentioned sophisticated mathematical and physical dual design, the originally coarse macroscopic moisture regain scalar can be transformed into a high-dimensional state vector that internalizes the moisture absorption dynamics and the differential response of microstructures. This provides a microscopic input with great physical tension for the gating interaction and attenuation suppression of subsequent computational units, greatly enhancing the model's extrapolation and generalization ability and physical consistency for unseen operating conditions.

[0064] Based on the above technical solution, the advanced branches of map ontology construction and node initialization successfully transform macroscopic HVI data and hygroscopic parameters into state vectors containing microscopic mechanisms, and are equipped with computational units of strictly dual physical links. This design not only provides an operable state-space foundation for hard constraints on physical potential energy, but also frees the entire prediction model from dependence on statistical correlation, instead building it on a solid foundation of cross-scale causal inference, fundamentally solving the extrapolation distortion and physical paradox problems caused by shallow fusion in existing technologies.

[0065] Example 3: like Figure 6 , 7 As shown, this embodiment further expands the mathematical implementation mechanism of physical potential energy constraint based on step 300 of embodiment 1.

[0066] It should be understood that the concept of hard constraint of potential energy described in Example 1 is explanatory rather than restrictive. This example will sink this higher concept into a rigorous mathematical formula and parameter defense logic to prove why traditional soft constraints fail while the hard constraint of this scheme can successfully eliminate physical paradoxes.

[0067] The physical potential energy constraint is the physical potential energy function, which measures the degree of deviation of the components of the third state node from the threshold of negative correlation with the components of the second state node. After updating the third state node, the third state node is pulled in the direction of decreasing physical potential function based on gradient descent.

[0068] Among them, the negative correlation threshold refers to the physical boundary established based on thermodynamic trends. That is, the higher the permeation state component of the environmental medium (such as water), the lower the upper limit inhibition threshold of its response state component to the target (such as force).

[0069] The degree of deviation refers to the magnitude of the deviation that violates the laws of physics when a certain dimensional component of the third state node attempts to break through the inhibition threshold and shows an abnormal trend of strong counter-increase in the environmental medium. The physical potential energy function will quantify this deviation and assign it a positive potential energy value.

[0070] In some implementations, after calculating a new third state node in each iteration, the result is not directly received. Instead, it is substituted into the physical potential energy function to calculate the potential energy value. If the potential energy value is greater than zero, it indicates a physical deviation. Then, based on the gradient descent principle, the third state node is directly pulled in the direction of decreasing potential energy function in the state space until it falls within the physically permissible negative correlation threshold range. This allows for a mechanism to directly intervene in the state vector during iterative reasoning in the computational graph.

[0071] Furthermore, we delve into the specific mathematical expression and physical meaning of the parameters of the physical potential energy function. The mathematical expression of the physical potential energy function is: ; in, This is the brute force state vector of the third state node. Let be the water infiltration state vector of the second state node. The physical potential energy function, for and vector dimension, and Let i represent the force state component and the water permeability state component in dimension i, respectively. It is a configurable positive management coefficient used to regulate the upper limit of the inhibition strength of the water permeability state on the strong state.

[0072] This formula constructs the potential energy field by comparing the water infiltration component with the strong state component dimension by dimension. The max(0,...) operation in the formula ensures that only when... Greater than Potential energy penalties are only applied when a strong force attempts to exceed the suppression limit set by the water state. If the strong force naturally falls below this limit, the potential energy is zero and no intervention is applied. This allows the model to avoid blindly suppressing all strong force outputs during use, and instead precisely penalize only "out-of-bounds" behaviors that violate the thermodynamic negative correlation trend, thus preserving the model's ability to freely fit within the legal physical space.

[0073] It is important to note the physical meaning and defensive expansion interpretation of the positive management coefficient α. The so-called positive management coefficient α is a hyperparameter used to regulate the upper limit of the inhibition of the force state by the water infiltration state. Its physical essence is to define the sensitivity multiplier of water infiltration to the attenuation of force.

[0074] To demonstrate the rationality and ingenuity of the parameter settings in this scheme, the following is provided: Examples of endpoint values ​​and outbound comparisons.

[0075] when When the value is 0.1, it represents a weak suppression strategy. At this time, the moisture penetration state needs to reach an extremely high value in order to slightly suppress the upper limit of strength. This strategy is suitable for special synthetic fiber blending scenarios that are not sensitive to moisture absorption. when When the value is 1.0, it represents a moderate suppression strategy. At this time, the moisture penetration component and the upper limit of strength have a standard linear negative correlation, which is suitable for most conventional cotton fibers in the scenario of moisture absorption strength reduction. when =5.0 represents a strong suppression strategy. At this point, even a small increase in moisture penetration will significantly lower the upper limit of strength. This is suitable for scenarios where the strength of low-maturity, high-short-fiber inferior cotton fibers decreases sharply under extreme high humidity conditions.

[0076] As an out-of-bounds comparison, if set If the value is 0, then the physical potential energy function is always equal to zero, meaning the system completely abandons the upper limit suppression of strength by moisture. Under the condition of "low Micronaire value × high moisture regain," the interactive inference of the computational unit loses the constraint of the physical boundary, easily outputting the physical paradox that strength increases against the trend of humidity. This proves... The logic requires a positive value.

[0077] It should be understood that although the specific values ​​of 0.1, 1.0, and 5.0 are listed above, in actual deployment... It can be dynamically configured based on the characteristics of raw material batches, and any batch that meets the requirements... Values ​​greater than 0 all fall within the protection scope of this invention.

[0078] Next, the update formula for pulling the third state node in the direction of decreasing physical potential energy function is: ; This update formula essentially performs a forced pull-back operation based on a thermodynamic gradient in the state space of the computation graph, where, This is the powerful state vector of the updated third state node. The potential energy function for the third state node The gradient indicates the direction in which the potential energy function changes most dramatically with the force state vector. and The preset potential energy decay step size controls the amplitude of each pull-back operation. It can be used to simulate the relaxation rate of a fiber system returning from a non-equilibrium state to a thermodynamic equilibrium state. However, if the value of η is too large, it may cause the strong force state vector to oscillate at the potential energy trough. If the value of η is too small, multiple iterations are required to eliminate the physical deviation. This application preferably takes a value of η between 0.01 and 0.1 to balance the convergence speed and stability.

[0079] Therefore, the system can be directly modified after each iteration using this formula. The value is then forcibly projected into the valid solution space allowed by the physical potential energy field.

[0080] To more clearly illustrate the ingenuity of the hard constraint mechanism of this invention in eliminating physical paradoxes, the following uses the extreme working condition of "low Markovnikov value × high moisture regain" as an example to plan a rigorous comparative example to compare the essential differences in output between the soft constraints of the traditional Physical Information Neural Network (PINN) and the hard constraints of this scheme.

[0081] In this comparative scenario, the input raw material is set as a batch of inferior cotton with a low micronaire value of 3.8 and a high short fiber content of 15%, while the environment experiences a high moisture regain of 9.5%. Under these conditions, the physical reality is that a large amount of moisture enters the amorphous region of the low-maturity fibers, causing macromolecular chain slippage, and the strength will inevitably decrease significantly.

[0082] The traditional PINN soft constraint scheme's processing and output: The traditional scheme only adds a physical penalty term to the loss function during model training, suggesting the model output conform to a curve shape of "high moisture, low strength." However, when the trained model is subjected to inference under extreme conditions, the soft constraint of the loss function can no longer influence the inference calculation process. Due to the scarcity of historical training data for low-quality cotton, the black-box mapping network inside the model is still influenced by historical data on high-maturity cotton, and its hidden layer activation values ​​tend to output an excessively high strength prediction value. At this point, the soft constraint cannot correct this abnormal activation during inference, ultimately leading to the physical paradox of the model outputting strength increasing against the trend with increasing moisture regain, with the predicted strength value even higher than that under medium moisture regain.

[0083] The processing and output of the hard constraint scheme in this proposal: When the same extreme working condition data is input into the physical state calculation diagram of this proposal, the second state node... The third state node generates an extremely high moisture permeation vector through hygroscopic kinetic mapping. In the early stages of iteration, the influence of historical data may lead to an overly high strong state vector.

[0084] But in each iteration update Then, the system immediately calculates the physical potential energy function. .because Extremely high weight The weight is too high, and it appears in many dimensions. The violation caused a sharp increase in potential energy. The system then performed a gradient descent update. The excessively high values ​​are directly adjusted within the calculation graph. The component is forcibly lowered until it falls below... The set suppression threshold.

[0085] After multiple iterations and potential energy recovery, the final output powerful prediction value is strictly suppressed within the physically permissible decay range, making it impossible for the paradox of increasing against the trend to occur.

[0086] The above comparison clearly demonstrates that the soft constraints of traditional PINN are merely macroscopic suggestions during the training phase and cannot withstand data distribution shifts during the inference phase. In contrast, the hard constraints of this scheme are directly embedded in the iterative inference state space of the computation graph, using potential energy gradient pull-back to enforce physical laws. This fundamentally eliminates any anomalous outputs that violate thermodynamic trends, elevating the physical reliability of the extrapolation to an extremely high level. Consequently, it provides unshakeable underlying logical support for preventing physical paradoxes.

[0087] Example 4: like Figure 8 , 9 As shown, this embodiment elaborates on the advanced branch of the local online evolution mechanism in detail, based on step 400 of embodiment 1.

[0088] It should be understood that the temporal traces and local directional updates described in Example 1 are explanatory rather than restrictive. This example will bring this higher-level concept down to a rigorous computational logic and parameter defense, explaining how to simulate the causal tracing of neural synapses through mathematical means and establish an update logic defense line to prevent evasion.

[0089] The process of acquiring time-series traces specifically includes: In the current iteration of inference, the sum of the absolute values ​​of the correction vectors output by the computational unit is obtained as the immediate activation value of that computational unit. Based on the difference between the current iteration and the preset total number of iterations, a temporal decay factor following an exponential decay law is generated. Finally, the immediate activation value and the temporal decay factor are recorded as the temporal trace of the computational unit in the current iteration of inference.

[0090] The immediate activation value is a direct measure of the output activity of a computational unit in the current round of inference. It is obtained by summing the absolute values ​​of the correction vectors (e.g., the correction increment of the first computational unit's output to the second state node, or the inhibitory correction increment of the second computational unit's output to the third state node) across all dimensions. The larger this sum of absolute values, the stronger the perturbation generated by the computational unit in this round of state evolution, and the greater its immediate causal contribution to the final prediction result.

[0091] In some implementations, the weighting coefficient of the time-decay factor, which increases exponentially with the number of iterations, is preferably mathematically expressed as follows: Where T is the preset total number of iterations, and t is the current iteration. The attenuation coefficient is greater than zero.

[0092] It should be noted that the logic of this factor is that the closer to the end of inference, that is, the closer t is to T, the smaller the difference (Tt), and the closer the value of the time decay factor is to 1 (i.e., the stronger the trace). Conversely, activations generated in the early stages of inference have a longer direct causal link to the final output due to the superposition and erosion of subsequent rounds of states, and their trace weight is decayed exponentially.

[0093] This design mimics the causal tracing mechanism of biological neural synapses. In the neural impulse conduction chain, the closer a synapse is to the final action potential emission, the higher its triggering necessity, and its weight should be prioritized for memorization. By multiplying the instantaneous activation amplitude by the temporal proximity, the system generates a precise temporal trace scalar that quantifies the causal contribution of each computational unit in each iteration, providing irrefutable traceability evidence for subsequent safe local updates.

[0094] The process of locally oriented updating the internal weights of computational units with strong temporal traces specifically includes: First, the prediction error is acquired, and a global modulation signal proportional to the prediction error is generated through an evolutionary controller. Then, for each computational unit, all time-series traces from the current prediction process are accumulated to obtain a total time-series trace. The gradient of the predicted value with respect to the weights within the corresponding computational unit is then scaled using the product of the global modulation signal and the total time-series trace as a scaling factor, generating a local weight update. Finally, the local weight update is accumulated into the internal weights of the computational unit, completing the local directional update.

[0095] The evolution controller is a function module that dynamically generates modulation intensity based on errors. The generated global modulation signal G is preferably generated using the formula... Calculation, where This represents the prediction error between the measured strength value and the predicted strength value. The global learning rate scalar is preset. The preferred value for the global learning rate γ is a minimum value such as 1e-4. Its physical meaning is to strictly control the upper limit of the step size for each adaptive update, preventing severe oscillations in the model's internal state space due to excessively large single errors.

[0096] The total time series trace is the sum of the time series traces generated by a specific computational unit in all T iterations of the current prediction process, thus obtaining the overall causal weight of that unit to the final prediction result.

[0097] During the update process, the system does not modify all parameters of the entire model indiscriminately. Instead, it uses the product of the global modulation signal G and the total time series trace as a scaling factor to scale only the gradient of the predicted value with respect to the weights within the computational unit, generating a local weight update amount, which is then added to the internal weights of the unit.

[0098] It should be noted that traditional online learning or fine-tuning solutions typically employ global gradient backpropagation. When the production line returns an error signal for a new batch, the weights of all network layers in the model are modified. This global update suffers from a fatal forgetting mechanism: because the data distribution of the new batch often differs from that of historical batches, the global update forcibly distorts the internal representation that the model originally built to adapt to historical stable operating conditions. As a result, while the model learns the new operating conditions, it quickly loses its ability to accurately predict the old operating conditions.

[0099] This solution's localized targeted updates precisely locate the causal responsibility units for the current inference based on temporal traces. Only these specific interactive subnetworks that leave strong traces are fine-tuned, while units that did not participate in the current inference or whose traces are extremely weak remain frozen. This enables millisecond-level adaptation to new scenarios while perfectly preserving the inference capability for historical scenarios, ensuring the model's continuous and safe evolution throughout its entire lifecycle and fundamentally preventing the risk of catastrophic amnesia.

[0100] It should be understood that, although the above lists the global learning rate The specific value of γ is 1e-4, but in actual deployment, γ can be dynamically configured according to the trade-off between adaptive speed and stability in the production line, provided that 0 < γ. Small values ​​less than 0.01 all fall within the protection scope of this invention, so as to maintain the minimally invasive characteristics of local updates.

[0101] Based on the above technical solution, the advanced branches of temporal trace calculation and local directional update successfully transform the abstract model adaptation requirements into an evolutionary algorithm with rigorous mathematical logic and biologically inspired mechanisms. This design not only endows the physical state computation graph with the ability to continuously grow endogenously under complex production line conditions, but also ensures that each evolution is safe, controllable, and physically consistent through an anti-avoidance update logic defense line, completely solving the technical problems of existing models being static, unable to adapt in real time, and prone to forgetting.

[0102] Example 5: like Figure 10 As shown, this embodiment elaborates on the advanced branch of offline enhanced data generation based on embodiments 1 to 4.

[0103] It should be understood that the aforementioned embodiments mainly focus on online reasoning and local adaptive evolution processes, while this embodiment is independent of online reasoning and demonstrates the downstream application scheme of the physical state computation graph in offline time periods. It aims to reflect the reuse value of structured knowledge graphs and establish a rigorous defense logic for data fidelity.

[0104] During offline periods, the inherent attribute states and virtual environment medium parameters are sampled based on the physical state computation graph. These offline periods refer to times when production line equipment is not performing real-time prediction tasks, or when the processor is in a maintenance cycle with idle computing resources. During this time, the system can utilize the constructed physical state computation graph, which has been continuously improved through local online evolution, to perform large-scale internal state space deduction, thereby spontaneously generating high-quality training data.

[0105] Inherent property state sampling refers to the pooling of fiber matrix microstructure state vectors corresponding to various historically accumulated raw material characteristic data, i.e. Random sampling is performed. This vector pool stores the microstructural state characteristics of cotton yarns of different batches and qualities, covering a wide range from high-maturity premium cotton to low-micronaire-value inferior cotton.

[0106] Virtual environment medium parameter sampling refers to the system uniformly sampling virtual moisture regain values ​​within a reasonable range that conforms to physical reality. For example, uniform grid sampling or random sampling can be performed within the moisture regain range of 2% to 15% to ensure that the generated data can cover the entire range of operating conditions from extremely dry to extremely high humidity, especially those extreme cross-operating conditions that rarely occur in real production lines but are of great testing value.

[0107] The sampled data is input into the physical state computation graph for complete graph diffusion inference, generating corresponding predicted response values. An augmented dataset, consisting of inherent property states, virtual environment medium parameters, and predicted response values, is used for offline consolidation training of downstream models.

[0108] Among them, complete graph diffusion reasoning refers to the sampling of... and As input, the physical state computation graph is invoked to execute the complete forward computation process described in the foregoing embodiments, which includes multiple rounds of iterative interaction and hard constraints on physical potential energy, and finally outputs the corresponding predicted strong response value. Therefore, each sampling combination will generate a triplet data point that strictly follows the internal physical causality. , , The augmented dataset is formed by the aggregation of a large number of triples.

[0109] This augmented dataset can not only be used for offline consolidation training of the model itself to enhance its robustness to rare operating conditions, but also serve as high-fidelity prior knowledge to be output to other downstream supervised models on the production line (such as traditional deep neural networks or gradient boosting tree models) for offline pre-training, thereby comprehensively improving the generalization boundary of the overall prediction system of the production line.

[0110] It should be noted that existing GAN schemes generate macroscopic label distributions through game-theoretic training. Internally, it is a black box mapping that lacks physical mechanisms. The generated data only imitates real samples in terms of statistical features. Once it encounters extreme working conditions that exceed the training distribution, GAN is very likely to generate distorted data with pattern collapse, or even output data that strongly increases against the trend and violates the laws of thermodynamics.

[0111] Existing simple polynomial enhancement methods merely perform algebraic combinations and interpolations on the original features, essentially representing a Cartesian product-like mechanical extension that completely fails to internalize the causal mechanism of the interaction between moisture and fiber microstructure. In contrast, the enhanced data generation method of this scheme is a dynamic reenactment of the "adsorption-plasticization-mechanical response" causal reasoning chain that has been internalized in the physical state computation graph.

[0112] In each graph diffusion inference, the generation of the second state node is constrained by the hygroscopic dynamic mapping function, the interaction between the first and second computational units is constrained by the microstructure gating and attenuation suppression mechanism, and the iteration of the third state node is subject to the hard constraint of the physical potential energy function.

[0113] This means that each generated triplet is not a blind imitation at the statistical level, but a necessary result rigorously derived from the evolution of microscopic physical states. It naturally obeys the constraints of the physical potential field and has irrefutable physical consistency. This data fidelity based on first principles deduction is unattainable by any black-box generation model or shallow algebraic extension.

[0114] As another alternative implementation, in addition to uniform sampling, a key sampling mechanism guided by the physical potential field gradient can also be used for sampling strategies of virtual environment medium parameters.

[0115] Specifically, the system can calculate the sensitivity of the physical potential energy function under different moisture regain settings, and prioritize dense sampling in the range where the potential energy gradient changes drastically (such as the paradoxical high-risk zone where low micronaire values ​​and high moisture regain intersect), thereby generating the most valuable challenging data for consolidating downstream models with minimal computational cost.

[0116] It should be understood that this gradient-guided sampling is interpretive rather than restrictive, and any sampling strategy that can cover the operating conditions falls within the scope of protection of this invention.

[0117] Based on the above technical solution, by generating advanced branches of offline augmented data, the system successfully expands the physical state computation graph from a single online prediction tool into a zero-cost, unlimited, and physically consistent high-fidelity data generator. This design not only demonstrates the reusability value of structured knowledge graphs, but also, through rigorous causal recapitulation logic, completely solves the technical problems of data augmentation distortion and model degradation in existing technologies, providing a solid data foundation for the continuous improvement of prediction accuracy throughout the entire production line lifecycle.

[0118] Example 6: like Figure 11 As shown, this embodiment provides a cotton yarn single yarn strength prediction system based on machine learning, based on the method flow of Embodiment 1.

[0119] It should be understood that the aforementioned embodiments 1 to 5 mainly describe the prediction method of the present invention from the perspective of algorithm logic and mathematical mechanism, while this embodiment maps the method to a specific software and hardware functional architecture, aiming to provide corresponding system support for the method's proprietary rights and to realize the binary binding of functional units and hardware / software logic.

[0120] The system referred to in this embodiment is for illustrative purposes only and not for limitation. Any combination of software and hardware capable of performing the following data flow and logical judgment falls within the protection scope of this invention.

[0121] The cotton yarn strength prediction system specifically comprises four functional units: a data acquisition unit, a graph construction and inference unit, a physical potential energy constraint unit, and a time-series trace recording and updating unit. These four units are not isolated code modules, but rather an organic collaborative system comprised of a processor, memory, and external sensors. The data acquisition unit serves as the portal for the system's interaction with the physical world. Its hardware and software implementation includes an online moisture regain detector deployed on the spinning workshop production line, an HVI high-capacity fiber tester for offline testing, and a front-end microcontroller responsible for receiving and performing preliminary filtering and formatting of these sensor signals. The graph construction and inference unit is the core computing engine of the system. It is implemented using a central processing unit or graphics processing unit with high parallel computing capabilities, along with the physical state calculation graph construction program and iterative inference algorithm stored in the memory loaded on top of it. Furthermore, the physical potential energy constraint unit is a hard constraint executor embedded in the inference engine. Through the floating-point arithmetic logic unit within the same processor, it calculates the potential energy function in real time after each iteration and performs gradient pullback operations. Finally, its time-series trace recording and update unit is the adaptive evolution hub of the system. Its carrying form includes a random access memory for storing time-series trace scalars and local weight update quantities, and a microprocessor responsible for triggering the evolution controller to calculate and perform directional gradient updates after receiving measured errors.

[0122] The data acquisition unit is used to acquire characteristic data of cotton yarn and environmental medium parameters. When performing this function, the data acquisition unit is connected to external sensors via a wired or wireless communication bus. For example, it reads the current moisture regain scalar output by the online moisture regain meter in real time as an environmental medium parameter through a serial communication interface, and simultaneously receives HVI test vectors containing indicators such as micronaire value, short fiber content, and breaking strength uploaded by an offline HVI tester via an industrial Ethernet network as characteristic data.

[0123] It should be understood that although this embodiment lists a moisture regain meter and an HVI meter as specific sensor forms, in other embodiments, the data acquisition unit can also be connected to other types of online or offline testing equipment such as temperature and humidity sensors and yarn diameter meters, as long as they can acquire physical quantities characterizing the inherent properties of cotton yarn and the state of the environmental medium. After completing data acquisition, the data acquisition unit packages the feature data and environmental medium parameters into a standard data structure and transmits it to the input port of the graph construction and inference unit via the internal bus.

[0124] The graph construction and reasoning unit is used to construct the physical state calculation graph and perform iterative reasoning on the feature data and environmental medium parameters to obtain the predicted value. The physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and a calculation unit. The calculation unit is used to simulate the physical interaction effects between the state nodes.

[0125] Upon receiving the data packet from the data acquisition unit, the processor of the graph construction and inference unit first calls the pre-stored encoder and dynamic mapping functions in memory to map the HVI test vector to the first state node, the moisture regain scalar to the second state node, and initializes the third state node. Subsequently, the processor enters a multi-round iterative inference loop. In each round, it sequentially calls the weight matrices and computational logic of the first and second computational units to simulate the adsorption interaction and attenuation suppression effect between state nodes, outputting the correction increments for the second and third state nodes. During the intervals between iterative inferences, the graph construction and inference unit temporarily stores the updated third state node data in a cache and issues an interrupt signal for review by the physical potential energy constraint unit.

[0126] The physical potential energy constraint unit is used to apply physical potential energy constraints to the third-state node during iterative inference, forcing a decrease in the upper limit of the target response when environmental medium factors increase. Upon receiving an interrupt signal and temporarily stored third-state node data, the physical potential energy constraint unit immediately reads the positive management coefficient and potential energy decay step size from memory and calculates the deviation of the current third-state node component from the negative correlation threshold. If the deviation is greater than zero, the unit's operational logic directly generates a pullback correction amount based on the gradient descent formula and overwrites this correction amount into the cache of the graph construction and inference unit, forcibly pulling the third-state node in the direction of decreasing potential energy function.

[0127] This hardware-level real-time overwrite operation ensures that the hard constraints of physical laws are executed without delay in every micro-step of inference computation, eliminating any physical paradoxes that may be caused by logical bypassing or asynchronous delays at the software level.

[0128] The temporal trace recording and updating unit is used to record the temporal traces of the computing unit, and based on the measured error and temporal traces corresponding to the predicted value, performs local directional updates of the internal weights of computing units with strong temporal traces.

[0129] This unit performs a dual role during system operation: During the forward inference phase, it listens to the output bus of the graph construction and inference unit to obtain the sum of the absolute values ​​of the correction vectors in real time as the instantaneous activation value, and calculates the timing decay factor by combining the difference between the current round and the total number of iterations. The product of the two is recorded as a timing trace in a dedicated trace memory.

[0130] When the graph construction and inference unit outputs the final strong prediction value, and the production lab returns the measured strong value after a period of time, the processor of the time-series trace recording and update unit will calculate the prediction error, call the evolution controller to generate a global modulation signal, and then extract the total time-series traces of each computation unit from the trace memory. Using the product of the global modulation signal and the total time-series traces as a scaling factor, this unit only calculates the local gradient update amount for the internal weight matrix of the strong trace unit, and accumulates this update amount to the model parameter storage area of ​​the graph construction and inference unit through a write operation, completing a minimally invasive local directional update.

[0131] It should be understood that this local update mechanism based on trace tracking and error modulation isolates the modification path of global parameters in the hardware architecture and only opens the weight write permission of specific computing units, thus blocking the risk of catastrophic forgetting at the physical level.

[0132] Through the coordinated operation of the four functional units and the binary binding of software and hardware, the system in this embodiment not only fully executes the method flow described in Embodiment 1, but also translates the abstract physical state evolution, potential energy hard constraints, and local adaptive mechanisms into industrial-grade predictive equipment with clear data flow logic and hardware support. During operation, the data acquisition unit is responsible for perceiving the physical world, the graph construction and reasoning unit is responsible for internalizing causal deduction, the physical potential energy constraint unit is responsible for safeguarding the physical bottom line, and the time-series trace recording and updating unit is responsible for driving safety evolution. These four units form a closed-loop defense system from perception, reasoning, constraint to evolution, enabling cotton yarn single-yarn strength prediction to not only possess extremely high physical consistency but also achieve millisecond-level safety adaptation under complex production line conditions. This completely solves the system-level pain points of existing technology models being static, prone to paradoxes, and unable to evolve in real time.

[0133] Example 7: like Figure 12As shown, in order to more clearly illustrate the commercial application value and creative effect of the cotton yarn single yarn strength prediction method and system based on machine learning in the real industry pain points, the following uses the extreme working condition adaptive application scenario of textile production line as an example to substitute the abstract schemes of the above embodiments 1 to 6 into the specific spinning workshop operating environment for detailed explanation.

[0134] It should be understood that the specific parameters and scenarios described in this embodiment are illustrative and not restrictive, and are intended to demonstrate how the system can eliminate physical paradoxes and achieve millisecond-level safety evolution in actual production. Any similar extreme raw material and environmental cross-operating conditions fall within the protection scope of this invention.

[0135] In this specific scenario, the spinning workshop has just switched to a new batch of raw materials. This batch of raw materials is of poor quality, characterized by a low micronaire value of 3.8 and a high short fiber content of 15%. At the same time, the workshop is experiencing a high-humidity environment during the rainy season, with the online moisture regain meter measuring an environmental medium parameter, i.e., a moisture regain rate as high as 9.5%. This combination of "low micronaire value × high moisture regain rate" is precisely the high-risk area where existing traditional black box models are prone to physical paradoxes.

[0136] Step 701: The data acquisition unit acquires the characteristic data and environmental medium parameters of the cotton yarn. The system uses a front-end microcontroller and sensor bus to read in real-time the HVI test vector uploaded by the HVI high-capacity fiber tester, which includes indicators such as a micronaire value of 3.8 and a short fiber content of 15%, as characteristic data. Simultaneously, it reads the 9.5% scalar output from the online moisture regain detector as the environmental medium parameter. The system packages these macroscopic physical quantities into a standard data structure and transmits it to the input port of the graph construction and inference unit for later use.

[0137] Step 702: The graph construction and inference unit constructs a physical state computation graph and iteratively infers the feature data and environmental medium parameters to obtain predicted values. The processor first calls the self-supervised pre-trained encoder to map the low-quality HVI test vector into a fiber matrix microstructure state vector containing poor microstructure information such as low maturity and high amorphous region accessibility, as the first state node. Next, it calls the hygroscopic kinetic mapping function to map the 9.5% high moisture regain scalar into a water permeation state vector with extremely high permeability components in all dimensions, as the second state node. Finally, it initializes the strength state vector as the third state node.

[0138] In the multi-round iterative inference loop, the first computing unit simulates the rapid adsorption and containment of moisture by the inferior microstructure, and outputs the correction increment for the second state node, so that the moisture penetration state is further saturated; the second computing unit simulates the severe plasticization and weakening of fiber mechanical strength by high moisture penetration, and outputs the inhibitory correction increment for the third state node, attempting to suppress the strength state.

[0139] In step 703, the physical potential energy constraint unit applies a physical potential energy constraint to the third state node during iterative reasoning, so as to force the upper limit of the target response to decrease when the environmental medium factor increases.

[0140] In the early stages of iterative inference, due to the residual influence of historical high-quality cotton data on the model's internal weights, the third state node may generate an excessively high strong state component, attempting to break through the suppression limit set by the extremely high moisture penetration state. At this point, the physical potential energy constraint unit intervenes in real time to calculate the physical potential energy function. .

[0141] Due to a moisture regain of 9.5%, The quantity is extremely high, and The weight is too high, and it appears in many dimensions. The potential energy value increases sharply due to the violation.

[0142] At this point, the physical potential energy constraint unit immediately performs gradient descent update. The excessively high values ​​are directly adjusted within the calculation graph. The component is forcibly lowered until it falls below... The set suppression threshold.

[0143] After multiple iterations and repeated pullbacks due to potential energy hard constraints, the final output strength prediction value is strictly suppressed within the physically permissible attenuation range, stably outputting a reasonable low value that reflects the sharp decline in strength of inferior cotton caused by high humidity, making it impossible for the physical paradox of strength increasing against the trend of humidity to occur.

[0144] Step 704: The timing trace recording and updating unit records the timing trace of the calculation unit, and based on the measured error and timing trace corresponding to the predicted value, performs local directional updates of the internal weights of the calculation unit with strong timing trace.

[0145] During the aforementioned forward inference process, the timing trace recording and update unit records the instantaneous activation values ​​and timing decay factors of the first and second computing units in each iteration in real time by listening to the output bus of the computing unit, and calculates them into timing traces.

[0146] Since this inference faces an extreme high humidity condition, the inhibitory correction increment output by the second computational unit is extremely large and is most frequently activated in rounds near the end of the inference. Therefore, the total time-series trace accumulated by the second computational unit is very strong, indicating that its causal contribution to the low-strength prediction result is dominant.

[0147] After a period of time, the actual strength value of the batch of cotton yarn measured in the production line laboratory is returned to the system. The processor of the time-series trace recording and update unit calculates the prediction error between the measured value and the predicted value, and calls the evolution controller to generate a global modulation signal G proportional to the error. Subsequently, using the product of the global modulation signal G and the total time-series trace as a scaling factor, only the gradient of the predicted value with respect to the internal weights of the second computational unit is scaled to generate a local weight update, which is then accumulated into the internal weight matrix of that unit, completing a minimally invasive local directional update. For the first computational unit, whose traces were extremely weak in this inference, and other network layers that did not participate, their weight parameters are completely frozen and not modified in any way.

[0148] This localized, targeted update mechanism was validated in a closed-loop manner during subsequent production line operation: when the workshop continued to use this batch of substandard cotton, and the moisture regain rate hovered around 9.5% again, the system executed the prediction process again. Since the previous update only fine-tuned the second calculation unit with the largest causal contribution, this unit had internalized the new interaction law of "the strength of substandard cotton rapidly decreasing under high humidity," significantly reducing the error between the predicted strength value and the measured value, achieving a millisecond-level instantaneous improvement in prediction accuracy. Simultaneously, because the previous update strictly froze irrelevant units, the system could still output high-precision prediction values ​​when faced with the workshop subsequently switching back to the historical high-quality cotton operating conditions, completely avoiding the catastrophic forgetting phenomenon common in traditional online learning.

[0149] Therefore, through the adaptive application scenario of extreme working conditions in textile production lines demonstrated in this embodiment, the system successfully translates the abstract physical state evolution, potential energy hard constraints, and local directional update mechanism into a closed-loop defense equipment that solves real industry pain points. When facing unknown extreme working conditions that are prone to inducing physical paradoxes, the system eradicates the abnormal output that strongly increases against the trend by relying on potential energy hard constraints. At the same time, when facing production line condition drift, it achieves millisecond-level safe adaptation by relying on time-series trace tracing and local directional updates, and perfectly retains the memory of historical stable working conditions. It completely solves the three major technical problems of existing technical models in long-term deployment on production lines: averaged prediction distortion, physical paradox output, and inability to evolve in real time, demonstrating extremely high commercial application value and industrial reliability.

[0150] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto.

[0151] It should be understood that the algorithm formulas, parameter values, physical semantic mappings, and software and hardware implementation forms listed in the foregoing embodiments are all defensive interpretations of the abstract technical features in the claims, intended to provide sufficient disclosure support for the widest scope of protection of the present invention, rather than limiting the scope of protection of the present invention.

[0152] Any variations or substitutions readily conceived by those skilled in the art within the scope of the technology disclosed in this invention, such as migrating the internalized physical interaction mechanism and locally oriented evolution architecture of this solution to other fields of material performance prediction affected by environmental medium penetration, or making equivalent substitutions to the internal network structure and specific form of the potential energy function of the computing unit, should be covered within the protection scope of this invention, as long as it essentially still constructs a physical state calculation graph containing three types of state nodes: inherent properties, environmental medium penetration, and target response, and applies a hard constraint on physical potential energy that forces a decrease in the upper limit of the response when the medium increases during iterative inference, and performs local oriented updates based on time traces and measured errors. Therefore, the protection scope of this invention should be determined by the scope of the claims.

Claims

1. A method for predicting the strength of single cotton yarn based on machine learning, characterized in that, include: Obtain the feature data and environmental parameters of the object to be predicted; A physical state calculation graph is constructed, and the feature data and environmental medium parameters are iteratively inferred to obtain predicted values. The physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and a calculation unit. The calculation unit is used to simulate the physical interaction effects between the state nodes. In the iterative reasoning, a physical potential energy constraint is applied to the third state node to force the upper limit of the target response to decrease when the environmental medium factor increases. Record the time series traces of the computing unit, and based on the measured error corresponding to the predicted value and the time series traces, perform local directional updates of the internal weights of the computing unit with strong time series traces.

2. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, The computing unit includes a first computing unit and a second computing unit; The first state node is generated by encoding and mapping the feature data; The second state node is generated from the environmental medium parameters through a dynamic mapping that includes learnable decay parameters; The first calculation unit takes the first state node and the second state node as input, and calculates the correction increment for the second state node through gating interaction; The second calculation unit takes the updated second state node and the third state node as input, calculates the attenuation coefficient, and outputs the inhibitory correction increment for the third state node.

3. The method for predicting the strength of single cotton yarn based on machine learning according to claim 2, characterized in that, The encoding mapping is implemented through a self-supervised pre-trained encoder, and the pre-training process of the encoder specifically includes: The first state node output by the encoder, together with the given environmental medium parameters, is input into the decoder to predict the baseline response value of the object under the baseline environmental medium parameters, and the encoder is trained by minimizing the error between the predicted baseline response value and the true baseline response value. The kinetic mapping is achieved through a hygroscopic kinetic mapping function, the mathematical form of which is: ; in, Let be the water infiltration state vector of the second state node. It is the hyperbolic tangent function. This refers to the real-time moisture regain scalar value among the environmental medium parameters. The decay rate matrix is ​​a learnable matrix. The average of historical moisture content data. The standard deviation of historical moisture content data. For bias terms, For element-wise product, It is a smooth approximation of ReLU activation function.

4. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, The physical potential energy constraint is a physical potential energy function, which measures the degree of deviation of the components of the third state node from the threshold of negative correlation with the components of the second state node. After updating the third state node, the third state node is pulled in the direction of decreasing physical potential function based on gradient descent.

5. The method for predicting the strength of single cotton yarn based on machine learning according to claim 4, characterized in that, The mathematical expression for the physical potential energy function is: ; in, This is the brute force state vector of the third state node. Let be the water infiltration state vector of the second state node. The physical potential energy function, for and vector dimension, and Let i represent the force state component and the water permeability state component in dimension i, respectively. It is a configurable positive management coefficient used to adjust the upper limit of the inhibition strength of the water permeability state on the strong state; The update formula for pulling the third state node in the direction of decreasing physical potential energy function is: ; in, This is the updated brute force state vector of the third state node. The preset potential energy decay step size, The potential energy function for the third state node The gradient.

6. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, The process of acquiring the time-series traces specifically includes: In the current round of iterative inference, the sum of the absolute values ​​of the correction vectors output by the computing unit is obtained as the instantaneous activation value of the computing unit. Based on the difference between the current round and the preset total number of iterations, a time decay factor that follows an exponential decay law is generated. Based on the instantaneous activation value and the temporal decay factor, it is recorded as the temporal trace of the computing unit in the current round of iterative inference.

7. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, The process of locally updating the internal weights of computational units with strong temporal traces specifically includes: The prediction error is obtained, and a global modulation signal proportional to the prediction error is generated through an evolution controller; For the aforementioned calculation unit, all the time-series traces from this prediction process are summed to obtain the total time-series traces; Using the product of the global modulation signal and the total time trace as a scaling factor, the predicted value is scaled with respect to the gradient of the weights within the corresponding computing unit to generate local weight update amounts; The local weight update amount is accumulated into the internal weight of the calculation unit to complete the local targeted update.

8. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, Also includes: During offline periods, the inherent attribute states and virtual environment medium parameters are sampled based on the physical state calculation graph. The sampled data is input into the physical state computation graph to perform complete graph diffusion inference and generate the corresponding predicted response value. The augmented dataset, consisting of the inherent attribute states, virtual environment medium parameters, and predicted response values, is used for offline consolidation training of downstream models.

9. The method for predicting the strength of single cotton yarn based on machine learning according to claim 1, characterized in that, The object to be predicted is cotton yarn; the inherent property state is the microstructure state of the fiber matrix; the environmental medium is moisture; the target response state is the strength state; the feature data is HVI test data; and the environmental medium parameter is moisture regain.

10. A machine learning-based system for predicting the strength of single cotton yarn, characterized in that, The method for predicting the strength of a single cotton yarn based on machine learning, as described in any one of claims 1-9, specifically includes: The data acquisition unit is used to acquire characteristic data of cotton yarn and environmental medium parameters; The graph construction and reasoning unit is used to construct a physical state calculation graph and perform iterative reasoning on the feature data and environmental medium parameters to obtain predicted values. The physical state calculation graph includes a first state node representing the inherent property state, a second state node representing the environmental medium in the permeation state, a third state node representing the target response state, and a calculation unit. The calculation unit is used to simulate the physical interaction effects between the state nodes. A physical potential energy constraint unit is used to apply a physical potential energy constraint to the third state node in the iterative reasoning, so as to force the upper limit of the target response to decrease when the environmental medium factor increases. The time-series trace recording and updating unit is used to record the time-series traces of the computing unit, and based on the measured error corresponding to the predicted value and the time-series traces, to perform local directional updates of the internal weights of computing units with strong time-series traces.