A yarn quality prediction and process parameter intelligent recommendation system based on big data

Abstract

This invention relates to the field of yarn quality management technology, and more particularly to a big data-based yarn quality prediction and intelligent process parameter recommendation system. The system includes a multi-process production data assetization and causal graph construction module, a time-series state-space prediction module, a multi-process game theory recommendation module, and a production decision-making closed-loop and knowledge accumulation module. The multi-process production data assetization and causal graph construction module constructs a multi-process quality genetic causal graph; the time-series state-space prediction module uses a state-space model incorporating fiber rheology priors to output a probability distribution of yarn quality; the multi-process game theory recommendation module outputs globally optimal process parameters based on counterfactual reasoning and Pareto front exploration; and the production decision-making closed-loop and knowledge accumulation module uses real production data to back-update the causal graph weights and utility functions. This invention achieves probabilistic quality prediction, physical prior fusion, cross-process global optimization, and adaptive closed-loop, significantly improving the accuracy of yarn quality prediction and the effectiveness of process parameter recommendation.

Description

Technical Field

[0001] This invention relates to the field of yarn quality management technology, specifically to a yarn quality prediction and intelligent process parameter recommendation system based on big data. Background Technology

[0002] Yarn quality is a core factor affecting the final quality of textiles. Yarn production involves multiple continuous processes such as opening, carding, drawing, roving, spinning, and winding. The process parameters of each process are interdependent, and quality fluctuations in upstream processes can be inherited and amplified downstream along the production flow, forming a complex quality inheritance effect. Therefore, how to accurately predict yarn quality and intelligently recommend the optimal combination of process parameters for each process based on the prediction results is a key technical challenge that urgently needs to be solved in the field of intelligent textile manufacturing.

[0003] Currently, yarn quality prediction mainly employs the following two methods:

[0004] The first category is the traditional method based on statistical process control. This type of method sets upper and lower control limits by statistically analyzing the distribution characteristics of historical quality data and alarms for abnormal points that exceed the control limits. However, the SPC method is essentially a post-event detection method and cannot make forward predictions before quality anomalies occur. At the same time, the SPC method assumes that each process is independent of each other and ignores the transmission and coupling effects of quality fluctuations between multiple processes. When faced with the complex quality inheritance relationships between processes in modern textile production, the prediction accuracy and early warning lead time are seriously insufficient.

[0005] The second category is data-driven machine learning prediction methods. These methods utilize historical production data to train a black-box mapping model of process parameters and quality indicators, which improves prediction accuracy to some extent. However, existing machine learning methods still have the following shortcomings:

[0006] The model can only output deterministic quality predictions and cannot quantify the uncertainty of the prediction results, making it difficult to provide risk boundaries for production decisions. The model training relies entirely on data-driven methods and lacks the integration of physical mechanisms such as fiber rheology, resulting in poor generalization ability and weak physical interpretability of the model under production conditions outside the coverage of the training data. The model usually only predicts a single quality indicator and fails to comprehensively represent the overall condition of future yarn quality from the perspective of probability distribution.

[0007] In summary, existing technologies have significant shortcomings in areas such as causal modeling of quality inheritance relationships across multiple processes, probabilistic quality prediction incorporating physical priors, recommendation of globally optimal process parameters across processes, and adaptive optimization driven by production feedback. Therefore, a new technological solution that can systematically address these issues is urgently needed. Summary of the Invention

[0008] The purpose of this invention is to provide a yarn quality prediction and process parameter intelligent recommendation system based on big data, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] A big data-based yarn quality prediction and process parameter intelligent recommendation system includes:

[0011] The multi-process production data assetization and causal graph construction module is used to acquire historical process parameters, environmental parameters and corresponding yarn quality inspection data of each process in the entire textile process. Based on time series alignment, a causal discovery algorithm is used to construct a multi-process quality genetic causal graph that represents the quality inheritance and parameter coupling relationship between processes.

[0012] The temporal state-space prediction module includes a long sequence feature encoding unit and a rheological state transition unit. The long sequence feature encoding unit is used to map the node features in the multi-process quality genetic causal graph into a high-dimensional hidden state sequence. The rheological state transition unit uses a state-space model to predict the evolution of the high-dimensional hidden state sequence and introduces the fiber rheological prior equation as a regularization constraint in the calculation of the state transition matrix, outputting the probability distribution of yarn quality within the future preset production cycle.

[0013] The multi-process game recommendation module treats the probability distribution of yarn quality as the environmental state and the configuration of process parameters for each process as the decision action, constructing a joint utility function that includes quality compliance rate, energy consumption cost, and production cycle balance. The multi-process game recommendation module evaluates the potential intervention effect of different combinations of process parameters on the quality fluctuation of downstream processes through a counterfactual reasoning network, and outputs the globally optimal process parameter resource allocation scheme based on a multi-objective Pareto front exploration algorithm.

[0014] The production decision-making closed loop and knowledge accumulation module is used to collect real yarn quality data after actual production execution, calculate the deviation measure between the data and the yarn quality probability distribution, and use the deviation measure to update the edge weights of the multi-process quality genetic causal graph and the utility weights of the joint utility function in reverse, forming a continuously evolving production decision-making closed loop.

[0015] As can be seen from the technical solution provided by the present invention above, the yarn quality prediction and process parameter intelligent recommendation system based on big data provided by the present invention has the following beneficial effects:

[0016] The system systematically collects, detects anomalies, fills in missing information, and aligns time series of multi-source heterogeneous production data across the entire textile process through a multi-process production data assetization and causal graph construction module, enabling asset-based management of production data. It automatically mines the intrinsic causal relationships between process parameters, environmental parameters, and yarn quality from the data using a constraint-based PC algorithm and a greedy equivalence search algorithm. This constructs a multi-process quality genetic causal graph representing the quality inheritance and parameter coupling relationships between processes, solving the problem of traditional methods neglecting the transmission and coupling effects of quality fluctuations between processes. This provides reliable causal support for subsequent quality prediction and process parameter recommendations, making the quality prediction results causally interpretable rather than relying solely on statistical correlation.

[0017] By using the long-sequence feature encoding unit and rheological state transition unit in the temporal state space prediction module, the structured causal topological information in the multi-process quality genetic causal graph is deeply fused with multi-scale temporal features to generate a high-dimensional hidden state sequence containing causal enhancement representation. In the calculation of the state transition matrix, the fiber rheological prior equation is innovatively introduced as a physical regularization constraint, so that the state transition process conforms to the mechanical behavior law of fiber materials in the processing process, overcoming the defects of weak physical interpretability and poor generalization ability of pure data-driven models. By outputting the probability distribution of future yarn quality through Monte Carlo sampling and kernel density estimation, the problem that traditional single-value prediction methods cannot quantify prediction uncertainty is solved, providing a clear risk boundary for production decisions.

[0018] By transforming the process parameter configurations of each process into quantifiable decision-making actions through a multi-process game recommendation module, a multi-dimensional joint utility function integrating quality compliance rate, energy consumption cost, and production cycle balance is constructed, solving the problem of single optimization objectives in traditional methods. Counterfactual reasoning networks are used to accurately assess the potential causal intervention effects of different process parameter combinations on downstream process quality fluctuations, overcoming the bottleneck of traditional recommendation methods that only focus on local optima in a single process and cannot quantify cross-process impacts. Based on a multi-objective Pareto front exploration algorithm, non-dominated optimal solution sets are found among effective intervention parameter combinations, and the diversity of solutions is maintained by using crowding distance, outputting a globally optimal rather than locally optimal process parameter resource allocation scheme, achieving synergistic optimization of yarn quality, production cost, and production efficiency.

[0019] Through the production decision-making closed loop and knowledge accumulation module, the actual yarn quality data after actual production execution is fed back to the system. A multi-dimensional deviation measurement system is constructed using KL divergence and interval coverage deviation to comprehensively evaluate the performance of the prediction model and recommendation algorithm. The causal attribution network is used to locate the key causal relationship edges that cause prediction deviations, so as to achieve accurate reverse update of the edge weights of the multi-process quality genetic causal graph. A multi-objective reinforcement learning agent network is used to dynamically adjust the utility weights of each objective in the joint utility function, so that the system can adapt to the dynamic changes in the production environment such as raw material characteristics, equipment status and market demand, forming a continuous evolution closed loop from data collection to quality prediction, process recommendation, production execution and feedback update, ensuring the stability and effectiveness of the system in the long-term operation. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the structure of a yarn quality prediction and process parameter intelligent recommendation system based on big data according to the present invention. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0022] To better understand the above technical solutions, the following will provide a detailed description of the technical solutions in conjunction with the accompanying drawings and specific embodiments.

[0023] like Figure 1 As shown, this embodiment of the invention provides a yarn quality prediction and process parameter intelligent recommendation system based on big data, including:

[0024] The multi-process production data assetization and causal graph construction module is used to acquire historical process parameters, environmental parameters and corresponding yarn quality inspection data of each process in the entire textile process. Based on time series alignment, a causal discovery algorithm is used to construct a multi-process quality genetic causal graph that represents the quality inheritance and parameter coupling relationship between processes.

[0025] The temporal state-space prediction module includes a long sequence feature encoding unit and a rheological state transition unit. The long sequence feature encoding unit is used to map the node features in the multi-process quality genetic causal graph into a high-dimensional hidden state sequence. The rheological state transition unit uses a state-space model to predict the evolution of the high-dimensional hidden state sequence and introduces the fiber rheological prior equation as a regularization constraint in the calculation of the state transition matrix, outputting the probability distribution of yarn quality within the future preset production cycle.

[0026] The multi-process game recommendation module treats the probability distribution of yarn quality as the environmental state and the configuration of process parameters for each process as the decision action, constructing a joint utility function that includes quality compliance rate, energy consumption cost, and production cycle balance. The multi-process game recommendation module evaluates the potential intervention effect of different combinations of process parameters on the quality fluctuation of downstream processes through a counterfactual reasoning network, and outputs the globally optimal process parameter resource allocation scheme based on a multi-objective Pareto front exploration algorithm.

[0027] The production decision-making closed loop and knowledge accumulation module is used to collect real yarn quality data after actual production execution, calculate the deviation measure between the data and the yarn quality probability distribution, and use the deviation measure to update the edge weights of the multi-process quality genetic causal graph and the utility weights of the joint utility function in reverse, forming a continuously evolving production decision-making closed loop.

[0028] In this embodiment, the multi-process production data assetization and causal graph construction module realizes the assetization management of production data through the systematic collection and standardized processing of multi-source heterogeneous production data throughout the textile process; it utilizes causal discovery algorithms to mine the intrinsic causal relationships between process parameters, environmental parameters, and yarn quality in each process, and constructs a multi-process quality genetic causal graph that can accurately characterize the quality inheritance laws and parameter coupling effects between processes, providing reliable causal relationship support for subsequent quality prediction and process parameter recommendation, specifically including:

[0029] Multi-source heterogeneous data acquisition and preprocessing unit: comprehensively acquires historical process parameters, environmental parameters and corresponding yarn quality test data of each process in the entire textile process; after standardizing all the acquired data according to a unified coding standard, it is stored in a distributed data warehouse to build a multi-source heterogeneous original production dataset covering the entire spinning process;

[0030] The Isolation Forest algorithm is used to score anomalies in the multi-source heterogeneous original production dataset for each process. The formula for calculating the anomaly score is as follows: ,in, For the sample Abnormal scores, The number of samples in the dataset. For the sample The average path length of all trees in an isolated forest. For a given The average path length for each sample; samples with anomaly scores exceeding a preset anomaly threshold are marked as outliers and removed to obtain the dataset after outlier removal;

[0031] For missing values ​​remaining in the dataset after outlier removal, a spatial correlation between processes is used. Nearest neighbor interpolation method is used for filling; The formula for calculating nearest neighbor interpolation is: ,in, These are estimates of the missing values. The number of nearest neighbor samples, For the first The observed values ​​of the nearest neighbor samples are used to obtain the cleaned production dataset.

[0032] Time series alignment and feature engineering unit: Automatically identify the inherent sampling frequency of each process data in the post-washing production dataset, and determine the inherent time delay parameters between each process based on the physical characteristics of the textile process; the inherent time delay parameters reflect the time delay required for parameter changes in the upstream process to be transmitted to the downstream process;

[0033] Using the highest sampling frequency in the dataset as a benchmark, a dynamic time warping algorithm combined with inherent time delay parameters is employed to map process data with different sampling frequencies onto a unified time axis; the distance calculation formula for dynamic time warping is: ,in, For sequence with sequence The dynamic time-warped distance between them For sequence Length, For sequence Length, For sequence The Middle Elements and sequences The Middle The Euclidean distance between each element; the time-normalized dataset is resampled using cubic spline interpolation to give all process data the same time resolution, resulting in a time-aligned production dataset;

[0034] A sliding window method is used to segment the time-aligned production dataset, extracting steady-state statistical features and dynamic time-series features within each window. The steady-state statistical features include mean, variance, and skewness, while the dynamic time-series features include autocorrelation coefficient and trend slope. Domain features closely related to yarn quality are extracted by combining fiber rheological properties, including fiber draw ratio fluctuation rate and twist transmission attenuation coefficient. The steady-state statistical features, dynamic time-series features, and yarn quality-related domain features are then spliced ​​and fused to construct a multi-process, multi-dimensional feature matrix.

[0035] Causal framework construction and direction determination unit: A causal direction constraint matrix is ​​constructed based on mature textile technology knowledge; the causal direction constraint matrix clearly defines the possible causal transmission directions between each process, prohibits the occurrence of reverse causal relationships that violate physical laws, thereby narrowing the search space for causal discovery and improving the efficiency and accuracy of causal discovery.

[0036] A constraint-based PC algorithm combined with a causal direction constraint matrix is ​​used to perform conditional independence tests on the multi-process multi-dimensional feature matrix. By testing the independence between variables under the condition of other variables, causal relationship variable pairs that meet the preset significance level are selected to generate an initial causal skeleton. The initial causal skeleton only contains undirected connection relationships between variables and does not contain causal direction information.

[0037] A score-based greedy equivalence search algorithm is used to determine the direction of the initial causal skeleton. The algorithm searches all possible combinations of causal directions, calculates the score function value corresponding to each combination, and selects the causal direction combination with the highest score as the optimal solution to obtain the initial set of causal relationship edges.

[0038] Cause-effect graph pruning and stability verification unit: The initial set of causal edge relationships is scored using the Bayesian information criterion; the formula for calculating the Bayesian information criterion is as follows: ,in, Score the Bayesian information criterion. The likelihood function value of the model. The number of parameters in the model. The number of samples in the dataset is denoted as ; based on the scoring results, redundant causal relationship edges are pruned, removing edges with lower scores, to obtain the pruned set of causal relationship edges;

[0039] Bootstrap Resampling Stability Verification: The stability of the pruned causal edge set is verified using the Bootstrap resampling method. Multiple subsets are randomly selected from the original dataset with replacement, and the causal discovery process is repeated on each subset. The frequency of each causal edge in all subsets is counted. Causal edges with a frequency exceeding a preset stability threshold are retained to ensure the reliability and robustness of the causal graph.

[0040] The retained causal edges are assigned weights representing the strength of causality; these weights are obtained by calculating the conditional mutual information between variables, using the following formula: ,in, For a given variable Time variable With variables Mutual information between conditions For variables , , The joint probability distribution, For a given hour The conditional probability distribution, For a given hour The conditional probability distribution; ultimately forming a complete multi-process quality genetic causal diagram;

[0041] Among them, multi-source heterogeneous data preprocessing technology is based on data cleaning theory and improves data quality through two core steps: outlier detection and missing value imputation; the Isolation Forest algorithm isolates outlier samples by randomly partitioning the data space. Due to their rarity and specialness, outlier samples are isolated more quickly and therefore have higher outlier scores; this algorithm does not require assumptions about the distribution of the data and can effectively handle the problem of outlier detection in high-dimensional data; the K-nearest neighbor interpolation method utilizes the spatial correlation between adjacent processes in textile production to identify outlier samples by finding the closest neighbor to the missing value sample. Using the observations of the nearest neighbor samples to estimate missing values ​​can effectively fill in missing values ​​while preserving the original distribution characteristics of the data.

[0042] Time series alignment technology, based on the dynamic time warping algorithm, solves the problems of inconsistent sampling frequencies and time delays between processes in textile production. Traditional Euclidean distance calculation requires two sequences to have the same length and a one-to-one correspondence between time points, which cannot handle sequences with time scaling and offset. The dynamic time warping algorithm finds the optimal time curvature path between two sequences, minimizing the cumulative distance between the two sequences on this path, thereby achieving alignment of sequences of different lengths and sampling frequencies. The cubic spline interpolation method constructs a piecewise cubic polynomial function to resample the data, ensuring the continuity and smoothness of the interpolation curve and avoiding the introduction of additional noise and distortion.

[0043] Causal discovery algorithms, based on graph model theory, automatically mine causal relationships between variables from observational data. Constraint-based PC algorithms progressively remove undirected edges between variables through conditional independence tests to construct an initial causal framework, offering advantages such as high computational efficiency and applicability to high-dimensional data. Scoring-based greedy equivalence search algorithms determine causal directions by maximizing the score function, effectively handling cases with latent variables and selection bias. Combining these two algorithms leverages the efficiency of PC algorithms for rapid initial framework construction while utilizing the accuracy of greedy equivalence search to determine causal directions. The Bayesian information criterion, by introducing a penalty term for model complexity, balances model fit and complexity, avoiding overfitting. The Bootstrap resampling method verifies the stability of causal relationships through repeated sampling, effectively reducing the impact of random errors on causal discovery results and improving the reliability of the causal graph.

[0044] In this embodiment, the temporal state space prediction module receives the multi-process quality genetic causal graph output by the multi-process production data assetization and causal graph construction module. By integrating causal topological information and fiber rheological physical priors, it achieves accurate probabilistic prediction of yarn quality within a future preset production cycle. The temporal state space prediction module effectively solves the problems of traditional prediction methods, such as difficulty in handling long sequence dependencies, lack of physical interpretability, and inability to quantify prediction uncertainty, providing a reliable basis for quality prediction for subsequent intelligent recommendation of process parameters.

[0045] The temporal state-space prediction module is primarily responsible for transforming the structured features in the multi-process quality genetic causal graph into a high-dimensional hidden state representation that can be used for temporal prediction, and modeling the evolution of the hidden states based on a physically constrained state-space model. It extracts deep features containing causal topological information and multi-scale temporal information through a long-sequence feature encoding unit, and then introduces fiber rheology prior knowledge to constrain the state transition process through a rheological state transition unit. Finally, it outputs a yarn quality probability distribution containing quality fluctuation ranges and probability density functions, comprehensively characterizing the uncertainty features of future yarn quality, specifically including:

[0046] Long sequence feature encoding unit: Extracts attribute features and causal edge information of each process node from the multi-process quality genetic causal graph; classifies the attribute features into process parameter features, environmental parameter features and quality genetic features according to the preset process feature classification standard; normalizes the three types of features and concatenates them; then vectorizes them through a fully connected layer to obtain the initial node feature vector; organizes the causal edge information according to the process topology and causal propagation direction to obtain the weighted causal adjacency matrix;

[0047] Based on the causal strength weights in the weighted causal adjacency matrix, a directed attention weight matrix is ​​constructed between nodes in each process step; the formula for calculating the directed attention coefficient is: ,in, For nodes For nodes Attention coefficient For the activation function of the linear unit with leakage correction, For attention weight vectors, It is a linear transformation matrix. For nodes The initial node feature vector, For nodes The initial node feature vector, For feature splicing operations;

[0048] The attention coefficient is normalized and calculated using the following formula: ,in, The attention weights are normalized. For nodes The set of upstream causal neighbor nodes;

[0049] A directed graph attention network is adopted, with a weighted causal adjacency matrix as the topological constraint. The initial node feature vector of each process node and its initial node feature vector of its upstream causal neighbor nodes are aggregated asymmetrically by the directed attention weight matrix to obtain the causal topology enhancement feature vector of each process node.

[0050] The causal topology enhancement feature vectors are sorted according to the sequence of each process in the textile process to construct a temporal causal node sequence. A multi-scale one-dimensional convolutional network is used to extract features within a local window for each causal topology enhancement feature vector in the temporal causal node sequence. The convolutional kernel sizes are set to short-range convolutional kernels covering a single process, medium-range convolutional kernels covering three adjacent processes, and long-range convolutional kernels covering the entire process. The features extracted by convolutional kernels of different scales are spliced ​​and fused to obtain a multi-scale temporal feature vector.

[0051] Calculate the relative position encoding vector and the process type embedding vector for the multi-scale temporal feature vector respectively; the relative position encoding vector is used to represent the relative positional relationship of different process nodes on the time axis, and the process type embedding vector is used to represent the inherent process characteristics of different processes; the relative position encoding vector, the process type embedding vector and the multi-scale temporal feature vector are added element by element to construct the spatiotemporal embedding feature sequence.

[0052] An encoder stacking structure containing multiple self-attention layers and feedforward neural network layers is adopted to perform global feature interaction calculation and residual connection operation on the spatiotemporal embedded feature sequence. The self-attention layer can capture the long-range dependencies between different process nodes, and the residual connection operation can effectively alleviate the gradient vanishing problem in the training process of deep network. The feature sequence after global feature interaction calculation and residual connection operation is mapped to the preset hidden state dimension through the linear projection layer, and a high-dimensional hidden state sequence is output.

[0053] Rheological state transition unit: Obtain the high-dimensional hidden state sequence output by the long sequence feature encoding unit, slide the high-dimensional hidden state sequence into blocks according to the preset time step to obtain the hidden state block at the current time and the historical hidden state block; extract the fiber rheological constitutive relation data of textile materials, linearize the fiber rheological constitutive relation data, and construct the fiber rheological prior equation;

[0054] A gated recurrent network is used to extract the temporal dependency features of the current hidden state block and the historical hidden state blocks to obtain the hidden state temporal feature vector; tensor expansion is performed on the fiber rheology prior equation to obtain the rheology physical constraint matrix; the hidden state temporal feature vector is projected onto the physical subspace represented by the rheology physical constraint matrix to obtain the physical prior features; the hidden state temporal feature vector and the physical prior features are concatenated and fused to obtain the state transition fusion feature.

[0055] The state transition fusion features are input into a multilayer perceptron for nonlinear mapping to obtain the basic state transition matrix. Stiffness and viscosity parameters are extracted from the fiber rheological prior equations, and a diagonal regularization matrix characterizing the degree of parameter attenuation is constructed. The final formula for calculating the state transition matrix is: ,in, This is the final state transition matrix. Based on the state transition matrix, This is an element-wise multiplication operation. It is a diagonal regularized matrix;

[0056] Multi-step iterative prediction with Gaussian noise injection: The hidden state block at the current time step is multiplied by the final state transition matrix in multiple iterations to obtain the predicted hidden state vectors for each time step within the preset production cycle. The calculation formula for multi-step iterative prediction is as follows: ,in, For the future The predicted hidden state vector at each time step To predict the time step, This is the hidden state block at the current moment;

[0057] A Gaussian noise matrix is ​​generated based on the variance of the historical hidden state blocks. The Gaussian noise matrix is ​​then added element by element to the predicted hidden state vector to obtain the noisy predicted hidden state vector. The injection of Gaussian noise can simulate random fluctuations in the production process, making the prediction results more consistent with the actual production situation.

[0058] The Monte Carlo sampling method is used to randomly sample the noisy predicted hidden state vector multiple times to obtain a set of future hidden state samples. This set of future hidden state samples is then input into a fully connected decoding layer for dimensionality reduction mapping, yielding a set of yarn quality prediction values. Kernel density estimation is performed on this set of yarn quality prediction values ​​to obtain the yarn quality probability density function. The formula for kernel density estimation is as follows: ,in, Let be the probability density function of yarn quality. The sample size for the predicted yarn quality values. The bandwidth of the kernel function. For Gaussian kernel function, For the first Predicted values ​​for yarn quality;

[0059] The quality fluctuation range under a preset confidence level is calculated based on the yarn quality probability density function. The quality fluctuation range is combined with the yarn quality probability density function to output the yarn quality probability distribution within the preset production cycle in the future.

[0060] Among them, the long-sequence causal feature encoding technology, based on graph neural network and temporal convolutional network theory, achieves deep fusion of structured information and temporal information in the multi-process quality genetic causal graph; the directed graph attention network can perform asymmetric weighted aggregation of the features of upstream nodes according to the causal transmission direction, effectively capturing the causal dependencies between processes and avoiding interference from irrelevant node features; the multi-scale one-dimensional convolutional network extracts short-range features within a single process, medium-range features between adjacent processes, and long-range features of the entire process simultaneously through convolutional kernels of different sizes, comprehensively representing the quality change law at different time scales during the spinning process; the spatiotemporal embedding mechanism integrates positional information and process type information into the feature representation, enabling the model to distinguish the characteristics of different processes and their relative positions on the time axis, further improving the effectiveness of feature representation; the multi-layer self-attention encoder can capture long-range dependencies across processes, solving the gradient vanishing and information forgetting problems existing in traditional recurrent neural networks when processing long sequences;

[0061] Physically constrained state-space prediction technology, based on state-space models and physical information neural network theory, introduces prior knowledge of fiber rheology into data-driven prediction models, achieving a deep integration of data and physics. Traditional data-driven prediction models often lack physical interpretability and are prone to producing prediction results that violate physical laws. By transforming fiber rheological constitutive relations into regularization constraints, the model can be guided to learn state transition processes that conform to physical laws, improving the reliability and generalization ability of prediction results. State-space models have good mathematical properties and can efficiently perform multi-step iterative predictions, avoiding the error accumulation problem of autoregressive models in long-sequence predictions. Gated recurrent networks can effectively extract temporal dependency features from historical hidden state sequences, providing rich historical information for the calculation of state transition matrices. Diagonal regularization matrices, by introducing the decay characteristics of fiber rheological parameters, can simulate the changes in the mechanical behavior of fibers during processing, making the state transition process more consistent with the physical essence of the spinning process.

[0062] The probabilistic quality distribution generation technology, based on Monte Carlo sampling and kernel density estimation theory, achieves a quantitative characterization of yarn quality uncertainty. Traditional single-point prediction methods can only provide a definite quality prediction value, failing to reflect the uncertainty of the prediction result and making it difficult to meet the risk assessment needs of production decisions. By injecting Gaussian noise consistent with actual production during the prediction process and using the Monte Carlo sampling method to generate a large number of prediction samples, it can comprehensively cover all possible future yarn quality situations. The kernel density estimation method can estimate a continuous probability density function from discrete prediction samples, intuitively showing the probability of different quality levels occurring. The quality fluctuation range calculated based on the probability density function can provide a clear risk boundary for production decisions, helping production managers to formulate more scientific and reasonable production plans and process adjustment schemes.

[0063] In this embodiment, the multi-process game recommendation module takes over the yarn quality probability distribution output by the temporal state space prediction module, transforms the process parameter configuration of each process into quantifiable decision actions, and comprehensively balances three conflicting core objectives: quality compliance rate, energy consumption cost, and production cycle balance, constructing a multi-dimensional joint utility function. Through counterfactual reasoning technology, it accurately evaluates the causal intervention effect of process parameter adjustment on downstream process quality fluctuations, breaking through the bottleneck of traditional recommendation methods that only focus on the local optimum of a single process and cannot quantify the cross-process impact. Finally, based on the multi-objective Pareto front exploration algorithm, it outputs the globally optimal process parameter resource allocation scheme, realizing the synergistic optimization of yarn quality, production cost, and production efficiency.

[0064] The multi-process game-theoretic recommendation module is primarily responsible for transforming yarn quality prediction results into executable process parameter recommendation schemes. It first transforms the yarn quality probability distribution into standardized environmental state characteristics, constructing a global decision space covering all physically feasible process parameter combinations. Based on this, it establishes a multi-objective joint utility function integrating quality, cost, and efficiency, quantifying the comprehensive benefits of different process parameter combinations. Through counterfactual reasoning techniques, it simulates the causal impact of process parameter adjustments on the overall process quality, screening out effective parameter combinations with significant intervention effects. Finally, it employs a multi-objective evolutionary optimization algorithm to find the Pareto optimal solution among the effective parameter combinations and selects the globally optimal process parameter configuration scheme through a diversity maintenance strategy, providing a scientific and flexible basis for production decisions. Specifically, this includes:

[0065] Joint utility function construction unit: Extract the probability density function and quality fluctuation range from the yarn quality probability distribution, and then concatenate the probability density function and quality fluctuation range after numerical normalization to obtain the environmental state feature vector; the environmental state feature vector comprehensively characterizes the uncertainty of future yarn quality under the current production conditions, providing a basic input for subsequent decision-making;

[0066] Feature decoupling analysis is performed on the environmental state feature vector to obtain the set of key control parameters corresponding to each process; boundary constraints are set on the set of key control parameters based on the textile process knowledge base to clarify the physically feasible value range of each parameter; the parameter value range of each process is divided into grids according to the preset discretization step size to obtain the discretized decision space of each process; the discretized decision spaces of each process are combined by Cartesian product according to the sequence of textile process flow to obtain a global joint decision action space covering all possible process parameter combinations;

[0067] The process parameter configurations of each process are extracted as decision actions from the global joint decision action space. These decision actions are then concatenated with the environmental state feature vector to obtain state-action feature pairs. These state-action feature pairs are input into a pre-trained yarn quality prediction model to predict the predicted yarn quality value corresponding to each decision action. The formula for calculating the quality compliance rate is as follows: ,in, To achieve the quality compliance rate, This is the lower limit of the yarn quality standard. This represents the upper limit of yarn quality standards. For a given decision action and environmental conditions yarn quality The conditional probability density function;

[0068] Features corresponding to decision actions with a quality compliance rate lower than a preset compliance threshold are removed to obtain candidate state action feature pairs, effectively narrowing the search scope for subsequent optimization.

[0069] From the candidate state action feature pairs, extract the equipment operating parameters and production cycle balance parameters corresponding to the decision actions of each process; input the equipment operating parameters into the pre-built energy consumption prediction model for calculation to obtain the energy consumption value of each process; sum the energy consumption values ​​of each process to obtain the global energy consumption value; the formula for calculating energy consumption cost is: ,in, For energy consumption costs, This is the global energy consumption value. This is the preset energy consumption baseline value;

[0070] The production cycle balance parameter is used to calculate the cycle deviation between adjacent processes, and the cycle deviation is input into the preset cycle balance model to obtain the production cycle balance. The production cycle balance value ranges from 0 to 1. The larger the value, the better the production cycle balance between each process is matched, and the higher the overall efficiency of the production line.

[0071] Dynamic weighting is applied to quality compliance rate, energy cost, and production cycle balance to obtain the utility weight of each indicator. The utility weight is dynamically adjusted based on the priority of the current production task: when quality is the core of the production task, the weight of quality compliance rate is increased; when cost reduction and efficiency improvement are the core of the production task, the weights of energy cost and production cycle balance are increased. The formula for calculating the joint utility function is as follows: ,in, For joint utility value, The utility weight for the quality compliance rate, As a utility weight for energy consumption costs, The utility weight for production rhythm balance, and satisfying , For production cycle balance;

[0072] The joint utility function comprehensively quantifies the overall benefits of different combinations of process parameters, providing a unified evaluation standard for subsequent multi-objective optimization;

[0073] Counterfactual Reasoning and Global Optimization Unit: Quality compliance rate, energy consumption cost, and production cycle balance are extracted from the joint utility function as multi-objective optimization indicators to construct a multi-objective optimization objective vector. Based on the multi-objective optimization objective vector, the network structure and loss function of the counterfactual reasoning network are determined, and the counterfactual reasoning network is constructed. The counterfactual reasoning network consists of three parts: a fact encoder, an intervention generator, and a counterfactual verifier. The fact encoder encodes the current process parameters into a factual state vector; the intervention generator performs intervention operations on the process parameters to generate counterfactual intervention vectors; and the counterfactual verifier calculates the counterfactual yarn quality prediction value based on the factual state vector and the counterfactual intervention vector.

[0074] Each decision action combination in the global joint decision action space is used as the current process parameter combination and input into the fact encoder of the counterfactual reasoning network for encoding to obtain the fact state vector corresponding to each decision action combination. The fact state vector is input into the yarn quality prediction model to obtain the fact yarn quality prediction value corresponding to each decision action combination. Single-parameter and multi-parameter intervention operations are performed on each decision action combination in the global joint decision action space to obtain the process parameter combination after intervention. The process parameter combination after intervention is input into the intervention generator of the counterfactual reasoning network for encoding to obtain the counterfactual intervention vector. The fact state vector and the counterfactual intervention vector are concatenated and input into the counterfactual verifier of the counterfactual reasoning network to obtain the counterfactual yarn quality prediction value corresponding to each process parameter combination after intervention.

[0075] The difference between the predicted actual yarn quality and the predicted counterfactual yarn quality is calculated to obtain the potential intervention effect value for each combination of process parameters; the formula for calculating the potential intervention effect value is as follows: ,in, This represents the potential intervention effect value. This is a counterfactual yarn quality prediction. This is a predicted value for the actual yarn quality;

[0076] By comparing the potential intervention effect value with the preset intervention effect threshold, the process parameter combinations with potential intervention effect values ​​greater than the intervention effect threshold are selected to obtain the effective intervention parameter combination set. The effective intervention parameter combination set includes all process parameter adjustment schemes that can significantly affect yarn quality, effectively narrowing the search space of multi-objective optimization.

[0077] The set of effective intervention parameter combinations and the corresponding multi-objective optimization objective vector are input into a multi-objective Pareto front exploration algorithm. A non-dominated sorting method is used to perform a non-dominated sorting of each process parameter combination in the set of effective intervention parameter combinations. For any two process parameter combinations… and If the combination Not inferior to the combination in all objectives And it is superior to the combination in at least one objective. Then it is called a combination. Dominant Combination Based on the dominance relationship, all combinations of process parameters are divided into different Pareto levels. The combinations of process parameters with the first Pareto level constitute the Pareto optimal solution set. No solution in the Pareto optimal solution set can improve the performance of a certain objective without reducing the performance of other objectives.

[0078] Crowding Distance Calculation and Global Optimal Solution Selection: The crowding distance calculation method is applied to the Pareto optimal solution set to calculate the crowding distance for each combination of process parameters. The crowding distance measures the density of other solutions surrounding the Pareto optimal solution; a larger value indicates a sparser and more diverse range of solutions around that solution. The formula for calculating the crowding distance is: ,in, For the first The crowding distance of each Pareto optimal solution To optimize the number of indicators for multiple objectives, For the first The solution is at the th solution. The value that can be taken on each target For the first The solution is at the th solution. The value that can be taken on each target For all Pareto optimal solutions in the th... The maximum value on each target For all Pareto optimal solutions in the th... Minimum value on each objective;

[0079] Based on the crowding distance, the Pareto optimal solution set is maintained for diversity, and the process parameter combination with the largest crowding distance is selected as the globally optimal process parameter resource allocation scheme. This scheme has the best diversity and robustness while ensuring optimal multi-objective performance.

[0080] Among them, the multi-objective joint utility modeling technology, based on utility theory, transforms multiple conflicting optimization objectives into a single comprehensive utility index. In the yarn production process, there is an inherent conflict between the three objectives of quality compliance rate, energy consumption cost, and production cycle balance. Improving quality often requires increasing energy consumption or reducing production efficiency, while reducing energy consumption may lead to a decline in quality. By constructing a joint utility function, the weights of each objective can be dynamically allocated according to the priority of different production tasks, realizing the trade-off and synergy among multiple objectives. The joint utility function uses quality compliance rate as a positive indicator, energy consumption cost as a negative indicator, and production cycle balance as a positive indicator, comprehensively quantifying the comprehensive benefits of different combinations of process parameters, and providing a unified evaluation standard for multi-objective optimization.

[0081] Counterfactual reasoning intervention effect assessment technology, based on causal inference theory, can quantify the potential causal impact of process parameter adjustments on downstream process quality fluctuations. Traditional process parameter recommendation methods are often based on correlation analysis, which cannot distinguish between causal and correlational relationships, easily leading to unexpected negative effects from recommended parameter adjustment schemes. Counterfactual reasoning technology, by constructing counterfactual scenarios of what would happen if different process parameters were adopted, can accurately assess the causal effects of different process parameter combinations on the overall process quality. By screening out effective parameter combinations with significant intervention effects, it can not only narrow the search space of multi-objective optimization and improve optimization efficiency, but also ensure that the recommended process parameter schemes have clear causal basis, improving the reliability and interpretability of the recommendation results.

[0082] Multi-objective Pareto front exploration technology, based on evolutionary computation theory, can find the optimal trade-off solution among multiple conflicting objectives. Traditional single-objective optimization methods can only obtain one optimal solution, which cannot meet the comprehensive needs of multiple objectives in the production process. The Pareto optimal solution set includes all non-dominated solutions, each representing a trade-off scheme between objectives. The non-dominated sorting method can quickly divide all solutions into different Pareto levels and filter out the Pareto optimal solution set. The crowding distance calculation method can select the solution with the best diversity from the Pareto optimal solution set, avoiding excessive concentration of solutions in the solution set and ensuring that the recommended solution has better robustness and adaptability. The multi-objective Pareto front exploration algorithm can optimize multiple objectives simultaneously, providing more comprehensive and flexible choices for production decisions.

[0083] In this embodiment, the production decision-making closed loop and knowledge accumulation module receives the process parameter resource allocation scheme output by the multi-process game recommendation module. By collecting real yarn quality data after actual production execution, it quantifies the deviation between the prediction results and the actual production results. The deviation information is used to drive the iterative update of the system's core model, forming a complete closed loop of data collection, prediction, decision-making, execution, and feedback. At the same time, this module transforms the effective experience accumulated in the production process into reusable knowledge assets, continuously improving the system's prediction accuracy and recommendation effect, enabling the system to adapt to the dynamic changes in the production environment.

[0084] The production decision-making closed-loop and knowledge accumulation module is primarily responsible for the system's feedback updates and knowledge accumulation. It completes the entire closed-loop operation from real production data collection and deviation quantification calculation to reverse updating of model parameters. By constructing a multi-dimensional deviation measurement system, it comprehensively evaluates the performance of the prediction model and recommendation algorithm. It utilizes causal attribution technology to locate the key causal relationships leading to prediction deviations and accurately updates the edge weights of the multi-process quality genetic causal graph. It employs multi-objective reinforcement learning technology to dynamically adjust the utility weights of the joint utility function, enabling the system to adapt to the target requirements of different production stages. Ultimately, it achieves continuous self-optimization and knowledge accumulation, ensuring the system's stability and effectiveness during long-term operation. Specifically, this includes:

[0085] The comprehensive deviation measurement calculation unit acquires the actual yarn quality data corresponding to the globally optimal process parameter resource configuration scheme after actual production execution; it aligns the actual yarn quality data with timestamps to ensure that the actual data and the predicted data correspond one-to-one in the time dimension; at the same time, it performs working condition matching, filters out the actual quality samples that are consistent with the production working conditions at the time of prediction, removes invalid data under abnormal working conditions, and obtains the actual quality sample sequence corresponding to the future preset production cycle.

[0086] Extract the probability distribution of yarn quality within a future preset production cycle, and extract the probability density function of yarn quality from the probability distribution of yarn quality; construct an empirical distribution function using a sequence of real quality samples, and calculate the KL divergence distance between the empirical distribution function and the probability density function of yarn quality to obtain a measure of probability distribution deviation; the formula for calculating KL divergence is: ,in, For empirical distribution With predicted distribution KL divergence between them For real quality samples The probability in the empirical distribution For real quality samples The probability in the predicted probability density function The number of true quality samples; the smaller the KL divergence value, the smaller the difference between the predicted distribution and the true distribution, and the higher the prediction accuracy.

[0087] Extract the quality fluctuation range under a preset confidence level from the yarn quality probability distribution; count the number of samples in the actual quality sample sequence that fall within the quality fluctuation range and calculate the sample proportion; calculate the difference between the sample proportion and the preset confidence level to obtain the interval coverage deviation measure; the formula for calculating the interval coverage deviation measure is as follows: ,in, As a measure of interval coverage deviation, This represents the proportion of real samples falling within the quality fluctuation range. To preset the credit level;

[0088] The interval coverage deviation measure reflects the degree to which the prediction interval covers the actual quality fluctuations. A positive value indicates that the actual coverage is higher than the preset confidence level, and the prediction interval is too conservative; a negative value indicates that the actual coverage is lower than the preset confidence level, and the prediction interval is too optimistic.

[0089] The probability distribution deviation measure and the interval coverage deviation measure are weighted and fused to construct a comprehensive deviation measure tensor. The weights are dynamically adjusted according to the system's requirements for prediction accuracy and interval reliability. The comprehensive deviation measure tensor fully represents the overall difference between the prediction results and the actual production results, providing a quantitative basis for subsequent model back-update.

[0090] Reverse Update and Knowledge Accumulation Unit: Obtain the comprehensive deviation metric tensor and input it into a pre-constructed causal attribution network; the causal attribution network calculates the gradient of the comprehensive deviation metric tensor with respect to the weights of each causal relationship edge in the multi-process quality genetic causal graph using the backpropagation algorithm, and extracts the gradient contribution value of each causal relationship edge; the formula for calculating the gradient contribution value is: ,in, For causal relationship gradient contribution value, The loss function corresponding to the comprehensive deviation metric tensor. For causal relationship The current weight value;

[0091] Calculate the weight decay factor and weight enhancement factor for each causal relationship edge based on the sign and magnitude of the gradient contribution value. When the gradient contribution value is positive, it indicates that the weight of the causal relationship edge is too large, causing prediction bias, and the weight needs to be reduced. When the gradient contribution value is negative, it indicates that the weight of the causal relationship edge is too small, causing prediction bias, and the weight needs to be increased. Combine the weight decay factor and weight enhancement factor with the current weight value of the corresponding causal relationship edge to obtain the updated causal relationship edge weight value, thus completing the reverse update of the edge weights of the multi-process quality genetic causal graph.

[0092] From the comprehensive deviation metric tensor, probability distribution deviation metric and interval coverage deviation metric are separated to construct a deviation penalty evaluation matrix. This matrix is ​​then input into a multi-objective reinforcement learning agent network. Based on the current deviation situation and production targets, the agent network calculates the utility weight adjustment step size for quality compliance rate, energy consumption cost, and production cycle balance in the next production cycle. The formula for calculating the utility weight adjustment step size is as follows: ,in, For the first The step size for adjusting the utility weights of each objective. For learning rate, For the first The deviation reward signal corresponding to each target. For joint utility function For the first Individual target utility weights The gradient;

[0093] The initial utility weights corresponding to each objective in the joint utility function are dynamically adjusted according to the step size of the utility weight adjustment to obtain the updated utility weights of the joint utility function; when the deviation of a certain objective is large, the utility weight of that objective is increased so that the system pays more attention to the performance of that objective in the subsequent recommendation process.

[0094] The system manages the updated multi-process quality genetic causal graph and joint utility function in a versioned manner, recording the update time, update reason, and performance improvement. Simultaneously, it extracts combinations of process parameters from actual production data that can significantly improve yarn quality, reduce energy consumption, or increase production efficiency, forming an optimal process knowledge base. This optimal process knowledge base is categorized and stored according to yarn type, raw material type, and production equipment, providing direct reference for subsequent process parameter recommendations. Through continuous knowledge accumulation, the system can continuously build up production experience and gradually form a process knowledge system covering all scenarios.

[0095] Among them, the multi-dimensional probability deviation measurement technology is based on probability and statistics theory, and comprehensively evaluates the performance of probability prediction from two dimensions: distribution shape and interval coverage. Traditional single-point prediction deviation measures, such as mean square error, can only evaluate the average error of the predicted value and cannot reflect whether the uncertainty of the prediction result is accurate. KL divergence can quantify the overall difference between two probability distributions and evaluate the degree of fit between the predicted distribution and the true distribution. Interval coverage deviation measure can evaluate the reliability of the prediction interval and ensure that the prediction result can provide a reasonable risk boundary for production decision-making. By fusing the two deviation measures, the performance of the probability prediction model can be comprehensively and accurately evaluated, providing a reliable quantitative basis for the reverse update of the model.

[0096] The causal attribution reverse update technique, based on gradient attribution theory, can locate key causal relationships leading to prediction bias and achieve accurate updates to the causal graph. The multi-process quality genetic causal graph contains numerous causal edges, each with varying degrees of influence on the prediction results. By calculating the gradient of the comprehensive bias metric tensor with respect to the weights of each causal edge, the contribution of each edge to the prediction bias can be quantified. Targeted adjustments to the weights of causal edges based on the sign and magnitude of the gradient contribution effectively correct inaccurate causal strength representations in the causal graph, improving its accuracy and reliability. This update method avoids retraining the entire causal graph, significantly improving model update efficiency.

[0097] Multi-objective reinforcement learning utility update technology, based on reinforcement learning theory, can dynamically adjust the weights of the joint utility function according to the actual production situation, achieving adaptive balance among multiple objectives. During the production process, factors such as raw material characteristics, equipment status, and market demand are constantly changing, leading to changes in the priority of objectives at different production stages. Through continuous interaction with the production environment, the multi-objective reinforcement learning agent network automatically adjusts the utility weights of each objective based on prediction bias and changes in production objectives. When the performance of an objective fails to meet the standard, the weight of that objective is increased, making the system pay more attention to the optimization of that objective in subsequent decision-making processes. This dynamic adjustment mechanism enables the system to adapt to changes in the production environment and always maintain optimal overall performance.

[0098] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A yarn quality prediction and process parameter intelligent recommendation system based on big data, characterized by: include: The multi-process production data assetization and causal graph construction module is used to acquire historical process parameters, environmental parameters and corresponding yarn quality inspection data of each process in the entire textile process. Based on time series alignment, a causal discovery algorithm is used to construct a multi-process quality genetic causal graph that represents the quality inheritance and parameter coupling relationship between processes. The temporal state-space prediction module includes a long sequence feature encoding unit and a rheological state transition unit. The long sequence feature encoding unit is used to map the node features in the multi-process quality genetic causal graph into a high-dimensional hidden state sequence. The rheological state transition unit uses a state-space model to predict the evolution of the high-dimensional hidden state sequence and introduces the fiber rheological prior equation as a regularization constraint in the calculation of the state transition matrix, outputting the probability distribution of yarn quality within the future preset production cycle. The multi-process game recommendation module treats the probability distribution of yarn quality as the environmental state and the configuration of process parameters for each process as the decision action, constructing a joint utility function that includes quality compliance rate, energy consumption cost, and production cycle balance. The multi-process game recommendation module evaluates the potential intervention effect of different combinations of process parameters on the quality fluctuation of downstream processes through a counterfactual reasoning network, and outputs the globally optimal process parameter resource allocation scheme based on a multi-objective Pareto front exploration algorithm. The production decision-making closed loop and knowledge accumulation module is used to collect real yarn quality data after actual production execution, calculate the deviation measure between the data and the yarn quality probability distribution, and use the deviation measure to update the edge weights of the multi-process quality genetic causal graph and the utility weights of the joint utility function in reverse, forming a continuously evolving production decision-making closed loop.

2. The big data based yarn quality prediction and process parameter intelligent recommendation system according to claim 1, characterized in that: The multi-process production data assetization and cause-effect graph construction module includes: The data preprocessing unit is used to detect and remove outliers from historical process parameters, environmental parameters, and yarn quality inspection data using the isolated forest algorithm, and to fill in missing values ​​using the K-nearest neighbor interpolation method to obtain the cleaned production dataset. The time series alignment unit is used to perform time series alignment and resampling on the cleaned production dataset by using a dynamic time warping algorithm combined with the inherent time delay parameters between processes, so as to obtain a time-aligned production dataset. The feature extraction unit is used to extract steady-state statistical features, dynamic temporal features, and domain features based on fiber rheological properties from time-aligned data, and construct a multi-process, multi-dimensional feature matrix. The causal graph construction unit is used to construct an initial set of causal relationship edges from the multi-process multi-dimensional feature matrix based on a preset causal direction constraint matrix, using a constraint-based PC algorithm and a greedy equivalence search algorithm. After pruning and stability verification, the retained causal relationship edges are assigned causal strength weights to obtain a multi-process quality genetic causal graph.

3. The big data based yarn quality prediction and process parameter intelligent recommendation system as claimed in claim 1 wherein: The long sequence feature encoding unit includes: The causal topology enhancement subunit is used to perform causal topology enhancement on the features of each process node based on the causal strength weights in the multi-process quality genetic causal graph and a directed graph attention network. A multi-scale temporal feature extraction subunit is used to extract temporal features from the causal topology-enhanced feature sequence through multi-scale convolution. The spatiotemporal embedding and coding subunit is used to spatiotemporally embed the extracted temporal features and output a high-dimensional hidden state sequence through an encoder stack structure.

4. The big data based yarn quality prediction and process parameter intelligent recommendation system according to claim 1, characterized in that: The rheological state transition unit includes: The physical prior fusion subunit is used to extract the temporal dependency features of high-dimensional hidden state sequences using a gated recurrent network, and combine them with the physical constraints of the fiber rheology prior equation to obtain physical prior features. The state transition prediction subunit is used to calculate the basic state transition matrix based on physical prior features, and multiply it with the diagonal regularization matrix constructed based on rheological parameters to obtain the final state transition matrix. It performs multi-step iterative prediction on the current hidden state to obtain the predicted hidden state at each future time step. The probability distribution generation sub-unit is used to inject Gaussian noise into the predicted hidden state and then use Monte Carlo sampling to obtain a sample set, and output the yarn quality probability distribution by estimating the kernel density.

5. The big data based yarn quality prediction and process parameter intelligent recommendation system according to claim 1, characterized in that: The process of constructing the joint utility function in the multi-process game recommendation module includes: Environmental state feature vectors are extracted from the probability distribution of yarn quality, and the process parameters of each process are configured as decision actions to construct a global joint decision action space. For each decision action, the quality compliance rate is calculated based on the yarn quality probability distribution, the energy consumption cost is calculated based on the energy consumption prediction model, and the production cycle balance is calculated based on the cycle balance model. The quality compliance rate, energy consumption cost, and production cycle balance are dynamically weighted and summed to obtain the joint utility function.

6. The big data based yarn quality prediction and process parameter intelligent recommendation system according to claim 5, characterized in that: The multi-process game recommendation module outputs the globally optimal process parameter resource allocation scheme in the following way: Construct a counterfactual reasoning network to evaluate the potential intervention effect of different combinations of process parameters on the quality fluctuations of downstream processes, and screen out effective intervention parameter combinations; A multi-objective Pareto front exploration algorithm is used to perform non-dominated sorting of effective intervention parameter combinations to obtain the Pareto optimal solution set; The globally optimal process parameter resource allocation scheme is selected from the Pareto optimal solution set based on the congestion distance.

7. The big data based yarn quality prediction and process parameter intelligent recommendation system according to claim 1, characterized in that: The process of calculating the deviation metric in the production decision-making closed loop and knowledge accumulation module includes: The collected real yarn quality data is time-aligned and operating condition-matched with the yarn quality probability distribution; The probability distribution deviation metric is obtained by calculating the distribution divergence between the empirical distribution of actual yarn quality data and the probability distribution of yarn quality. The proportion of actual yarn quality data falling within the quality fluctuation range of the yarn quality probability distribution is statistically analyzed and compared with a preset confidence level to obtain a measure of interval coverage deviation. By integrating probability distribution deviation measure and interval coverage deviation measure, a comprehensive deviation measure is constructed.

8. The yarn quality prediction and process parameter intelligent recommendation system based on big data according to claim 7, characterized in that: The process of reverse updating using deviation measurement in the production decision-making closed loop and knowledge accumulation module includes: The comprehensive deviation metric is input into the causal attribution network to extract the gradient contribution value of each causal relationship edge in the multi-process quality genetic causal graph, and the weight of each causal relationship edge is updated according to the gradient contribution value. The comprehensive bias metric is input into the multi-objective reinforcement learning agent network to calculate the adjustment step size of each utility weight in the joint utility function, and the utility weights are updated accordingly.

Images