Textile intelligent production management system and method

Abstract

The invention discloses an intelligent textile production management system and method, and belongs to the technical field of textile process control. The method comprises the following steps: constructing a microscopic abnormal response map based on different cloth types, process formulas and equipment configuration; slicing the formal production instruction into a plurality of process fragment units, and matching potential abnormal probability vectors; constructing a process variation transfer graph used for simulating cascade propagation of anomalies in a process chain; in combination with the operation state parameters of the production equipment, process instability prediction is carried out by using the graph neural network model, and an early warning factor is output; adjusting production line resource configuration based on the early warning factor, and generating an optimal process routing path; operation feedback data in the path execution process is collected, error backspacing training is carried out on the prediction model, and self-learning updating is achieved; according to the invention, dynamic perception and closed-loop control of process abnormity are realized, and the method has the advantages of high fault prediction precision, timely scheduling response, adaptive optimization of the model and the like, and is suitable for flexible textile production scenes.

Description

Technical Field

[0001] This invention relates to the field of textile process control technology, specifically to a textile intelligent production management system and method. Background Technology

[0002] In the modern textile industry, with the rapid development of flexible production and personalized customization, the traditional assembly line production model can no longer meet the demand for orders with multiple varieties, small batches, and rapid delivery. Especially in the production of high-end textiles, the characteristics of different raw materials, differences in dyeing and finishing processes, the aging condition of equipment, and differences in personnel operation can all lead to dynamic fluctuations in the production process, thereby affecting the consistency of the final product quality.

[0003] Although some companies have already attempted to introduce MES (Manufacturing Execution System) for information management, the following problems still exist in the actual implementation: Existing systems can only perform static scheduling based on planned tasks, failing to identify the micro-fluctuations in looms and dyeing / finishing equipment across different shifts. Furthermore, some quality issues only occur in specific batches or process paths, and existing systems lack mechanisms to identify such "occasional" and "structural" problems. Textile field data is highly heterogeneous, with data from looms, dyeing machines, and quality inspection equipment using different formats, resulting in a lack of unified data integration and a basis for decision-making. Summary of the Invention

[0004] The purpose of this invention is to provide a smart textile production management system and method to address the shortcomings of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a textile intelligent production management method, comprising: A microscopic anomaly response spectrum is constructed for each process in the textile production process. The microscopic anomaly response spectrum represents the anomaly patterns formed under trial production conditions for different fabric types, process formulas and equipment configurations. The production instructions in the formal production task are sliced ​​and divided into multiple process segment units P1, P2, ..., Pn, where n is the total number of process segment units. The potential anomaly probability vector V of each process segment unit is matched based on the micro-anomaly response map. Based on the potential anomaly probability vector V of each process segment, a textile process anomaly transfer diagram G is constructed to simulate the probability propagation path of the cascading effect of local anomalies in the process chain. Based on the current operating status parameters F of the production equipment and the textile process anomaly transfer diagram G, a graph neural network model is trained to predict process instability and output an early warning factor R. The warning factor R is mapped to the current production line resource configuration map, the resource allocation priority is adjusted, and the optimal matching process routing path C is generated; Data on equipment micro-vibration characteristics, temperature and humidity gradients, and tension fluctuations are collected during the execution path C of the production line. The collected results are then input into a graph neural network model for error back-off training, enabling the model to learn and update itself dynamically.

[0006] Preferably, the construction of the microscopic anomaly response map includes: Collect trial production data under different fabric types, process formulas and equipment configurations. The trial production data includes fabric images, tension fluctuation curves, temperature and humidity records and equipment vibration signals. Feature extraction was performed on the trial production data to identify abnormal features such as weft skew, fabric shrinkage, and yarn breakage, and continuous process abnormal features were extracted based on the changes in tension and temperature and humidity curves. Structural anomaly features are fused with continuous process anomaly features to construct a multidimensional anomaly feature vector; Based on multidimensional anomaly feature vectors, clustering analysis algorithms are used to classify anomaly patterns of process nodes, forming a microscopic anomaly response map between various process parameter combinations and their corresponding anomaly patterns in textile production.

[0007] Preferably, the step of matching the potential anomaly probability vector V of each process segment unit based on the microscopic anomaly response map includes: The complete production instructions in the formal production task are analyzed. Based on the equipment operation stage, the material characteristic conversion point and the key process parameter change point, the production instructions are divided into multiple continuous process segment units, and each process segment unit corresponds to an independent process execution interval. For each process segment unit, the corresponding fabric type, process formula number and equipment configuration parameters are extracted and used as query conditions to search for matching records in the pre-constructed microscopic anomaly response map. Based on the historical abnormal feature vectors retrieved from the map, feature similarity is calculated for the current process segment unit, and the cosine similarity index is used to compare the degree of similarity. Matching results with similarity higher than a set threshold are fused together to generate a potential anomaly probability vector for each process segment unit.

[0008] Preferably, the step of constructing a textile process anomaly transfer map G based on the potential anomaly probability vector V of each process segment includes: Obtain potential anomaly probability vectors for multiple process segment units, and use each anomaly probability in each potential anomaly probability vector as a process node attribute to form a process node sequence with anomaly probability attributes. Based on the sequential order of process segment units, a directed connection relationship is established between process nodes, and anomaly propagation weights are constructed based on the difference in the probability of the same anomaly type in adjacent process nodes. The anomaly propagation weights of adjacent process nodes are positively normalized. Using process nodes and their corresponding directed connection paths as graph structure elements, construct a textile process variation transfer graph G.

[0009] Preferably, the step of predicting process instability by training a graph neural network model and outputting a warning factor R includes: Collect the operating status parameters of the production equipment corresponding to each process segment. The operating status parameters include real-time tension fluctuation value, spindle vibration frequency, drive power change rate and temperature and humidity disturbance amplitude, and bind the status parameters to the process segment nodes one by one. The process node sequence bound with equipment operating status parameters and the textile process variation transition graph are jointly input into the graph neural network model, which encodes the graph structure based on node attributes and edge weights. The graph neural network model is trained using supervised learning. The training data includes the instability labels of historical process segments and their corresponding operating state features. The training objective is to minimize the binary cross-entropy loss between the predicted value and the actual instability state. After the model converges, the graph neural network is used to output the process instability probability of each process segment, and an early warning factor is generated based on the set threshold.

[0010] Preferably, the step of mapping the early warning factor R to the current production line resource configuration map, adjusting resource allocation priority, and generating the optimal matching process routing path C includes: Map the early warning factors of each process segment to the process resource nodes in the current production line resource configuration diagram one by one; Based on the load status, remaining available window, and warning factor value of the corresponding process segment of each resource node, a resource availability score is calculated, and the resource nodes are prioritized according to the score results. Based on the priority ranking of resource nodes and combined with the temporal dependencies of each process segment, a dynamic programming algorithm is used to search for continuous feasible paths in the resource allocation graph to form a candidate set of process routes that meet the condition of minimizing instability risk. In the process route candidate set, the path with the highest cumulative resource score and the lowest total path risk value is selected as the optimal matching process route path C.

[0011] Preferably, the process of collecting equipment micro-vibration characteristics, temperature and humidity gradients, and tension fluctuation data during production line execution path C, and inputting the collected results into a graph neural network model for error backoff training, includes: During the execution of process routing path C, the micro-vibration characteristics of the corresponding equipment, the ambient temperature and humidity gradient, and the tension fluctuation curve are collected in real time to construct the operation feedback feature vector corresponding to each process segment node. The running feedback feature vector is compared with the predicted instability probability of each process segment output by the graph neural network model before path generation. The real labels are marked according to whether the actual process instability event has occurred, forming a supervision sample set. Using the supervised sample set as input, the binary cross-entropy loss between the predicted probability and the actual label is calculated. The loss function is used to characterize the error deviation between the current model prediction and the actual production result. The error bias is backpropagated to the graph neural network model via the backpropagation algorithm, and the model parameters are incrementally trained based on the gradient update strategy.

[0012] The present invention also provides a textile intelligent production management system, comprising: Anomaly Modeling Module: Constructs micro-anomaly response maps for each process in textile production. These micro-anomaly response maps represent anomaly patterns formed under trial production conditions for different fabric types, process formulas, and equipment configurations. Feature matching module: Slices the production instructions in the formal production task into multiple process segment units P1, P2, ..., Pn, where n is the total number of process segment units, and matches the potential anomaly probability vector V of each process segment unit based on the micro-anomaly response map; Anomaly propagation mapping module: Based on the potential anomaly probability vector V of each process segment, construct a textile process anomaly transfer map G to simulate the probability propagation path of the cascading effect of local anomalies in the process chain. Early warning generation module: Based on the current operating status parameter F of the production equipment and the textile process anomaly transfer diagram G, the module predicts process instability by training a graph neural network model and outputs an early warning factor R. Path optimization module: Maps the early warning factor R to the current production line resource configuration map, adjusts the resource allocation priority, and generates the optimal matching process routing path C; Feedback optimization module: Collects equipment micro-vibration characteristics, temperature and humidity gradients, and tension fluctuation data when the production line executes path C, and inputs the collected results into the graph neural network model for error back-off training to achieve self-learning and dynamic updating of the model.

[0013] The technical effects and advantages provided by the present invention in the above technical solution are as follows: 1. The intelligent textile production management method provided by this invention achieves structured modeling and propagation path simulation of potential anomalies throughout the entire production process by constructing microscopic anomaly response maps and process variation transfer maps, solving the problem of traditional textile management methods being unable to quantify the correlation of anomalies across multiple processes. By combining graph neural network models to predict the instability probability of process segments, it enables the forward-looking identification of process faults under complex and multivariate conditions, improving the sensitivity and response speed of the production line to anomaly risks, and significantly reducing fabric defect rates and downtime risks.

[0014] 2. This invention employs a resource scheduling optimization strategy driven by early warning factors. Combining real-time operating status parameters and production feedback data, it dynamically generates the optimal process routing path and constructs a model self-learning mechanism, enabling continuous optimization of the prediction model within the production closed loop. The overall solution balances local process fluctuations with global path optimization, exhibiting high scalability and adaptability. It is suitable for flexible, multi-variety textile production scenarios and can be widely applied in high-end textile management platforms within the context of intelligent manufacturing and the Industrial Internet. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the method of the present invention.

[0017] Figure 2 This is a flowchart of the system modules of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Example 1, please refer to Figure 1 As shown in this embodiment, a textile intelligent production management method includes: A microscopic anomaly response spectrum is constructed for each process in the textile production process. The microscopic anomaly response spectrum represents the anomaly patterns formed under trial production conditions for different fabric types, process formulas and equipment configurations.

[0020] Under different combinations of fabric types (such as pure cotton, polyester, viscose, etc.), process formulas (including dyeing and finishing ratios, sizing concentrations, sizing times, etc.), and equipment configuration conditions (such as loom models, speed settings, pressure rollers, etc.), targeted trial production operations are carried out.

[0021] During the trial production process, real-time fabric image data was acquired through a fabric image acquisition device; tension fluctuation curves of warp and weft yarns were recorded using a tension sensor; temperature and humidity change curves in the working environment were recorded synchronously using a temperature and humidity recorder; and vibration signals of the equipment within the corresponding time period were recorded using a vibration accelerometer installed at key parts of the loom.

[0022] The collected fabric image data were processed using an image edge detection algorithm and a grayscale histogram comparison analysis method. By analyzing the differences in pixel distribution, linear offset, and repeated texture destruction of the fabric images, three types of structural abnormalities were identified: weft skew, fabric shrinkage, and yarn breakage.

[0023] Specifically, weft skew is determined by the angular deviation between the fitted straight line at the edge of the fabric and the standard vertical direction. When this angle exceeds a set threshold (e.g., 3 degrees), it is determined that weft skew exists. Fabric shrinkage is detected by the change in texture direction gradient. When the variance of the direction gradient per unit area exceeds a set upper limit, it is identified as a shrinkage area. Yarn breakage is achieved by high-frequency signal noise detection. A sudden change in gray value in a local area of ​​the image and a contrast with the background exceeding 30% can be regarded as a breakage point.

[0024] Meanwhile, based on the tension fluctuation curve and temperature and humidity change data, the moving average filtering and discrete trend analysis methods are used to extract continuous process anomaly characteristics.

[0025] Taking tension fluctuation as an example, if the mean of the first difference of the tension curve exceeds the range of ±σ (the standard deviation of the tension baseline) within N consecutive time units (e.g., 100), then it is determined that there is an abnormal tension disturbance during that time period. The temperature and humidity curve is judged by the trend of the change of the second derivative. If the rate of change per unit time exceeds twice the historical average, it is determined to be an abnormal disturbance of the process environment.

[0026] The structural anomaly features and continuous process anomaly features obtained above are normalized based on a unified feature space to construct a multidimensional anomaly feature vector.

[0027] The vector dimension includes dimensions such as image anomaly score, tension fluctuation amplitude, temperature and humidity gradient change value, and abnormal amplitude of equipment vibration frequency. Each dimension is represented by a value in the range of [0,1], which facilitates subsequent clustering processing.

[0028] Feature fusion employs weighted principal component analysis (PCA), where the weights of each feature dimension are determined based on its significance in historical abnormal datasets related to product defects.

[0029] After the feature vectors are constructed, the K-means clustering algorithm is used to classify abnormal patterns for the feature vector set corresponding to each process.

[0030] The number of clusters K is automatically determined by the silhouette coefficient method. In this embodiment, the value of K is set to be between 2 and 6, and finally the cluster centers with the largest average silhouette coefficient are selected.

[0031] The combination of process parameters corresponding to each cluster center is regarded as a typical abnormal pattern.

[0032] The abnormal pattern is combined with its corresponding fabric type, process formula, and equipment configuration to form a mapping relationship, and finally a micro-abnormal response map between each process combination and abnormal type in the textile production process is established.

[0033] The production instructions in the formal production task are sliced ​​into multiple process segment units P1, P2, ..., Pn, where n is the total number of process segment units. The potential anomaly probability vector V of each process segment unit is matched based on the microscopic anomaly response map.

[0034] First, the production instructions in the formal production task are analyzed in a structured manner. The production instructions include information such as fabric number, dyeing and finishing process formula number, equipment operation sequence and process parameter settings.

[0035] After analysis, the continuous production task is divided into multiple process segment units based on the following three types of slicing criteria: Key points in equipment operation: such as a loom transitioning from low-speed commissioning to high-speed production or from shutdown to startup; Fabric characteristic changeover points: such as changing the type or batch of fabric within the same production task; Key process parameter changes: such as sudden changes in sizing concentration, dye concentration, or tension setting.

[0036] Each segment of the process is considered an independent process execution interval, numbered P1, P2...Pn, where n is the total number of process segments.

[0037] For each process segment unit, extract its corresponding fabric type number, process formula number, and equipment configuration parameter set.

[0038] The above three types of information constitute a composite query vector, which serves as an index condition for matching and retrieval in a pre-constructed microscopic anomaly response map.

[0039] Each record in the graph contains a process anomaly feature vector and historical occurrence probability under a specific combination of fabric, formula, and equipment. All graph records that satisfy the three information requirements of complete or near-complete matching will be extracted as a candidate matching set.

[0040] For each historical feature vector in the candidate matching set, a similarity calculation is performed between it and the production feature vector of the current process segment.

[0041] The feature vector includes multiple dimensions such as tension fluctuation amplitude, fabric anomaly score, temperature and humidity change value, and equipment vibration spectrum, all of which are normalized to the [0,1] interval.

[0042] The similarity calculation uses the cosine similarity algorithm, and the calculation formula is as follows: In the formula, Sim(A, B) represents the cosine similarity between vector A and vector B, with a value ranging from 0 to 1. The closer the value is to 1, the more similar the two vectors are. This represents the normalized outlier feature value of the current process segment unit in the i-th feature dimension; d represents the normalized abnormal feature value of the historical process segment unit matched by the micro-anomaly response map on the i-th feature dimension; d represents the total number of feature dimensions.

[0043] When the similarity value is greater than the set threshold T1 (e.g., 0.85), the historical record is considered to have a high similarity to the current process segment.

[0044] The probability of occurrence of the corresponding anomaly type in all historical feature vectors that meet the similarity threshold is weighted and fused, with the weights determined based on the similarity value.

[0045] For example, if the similarity of three matching records is 0.91, 0.88 and 0.86 respectively, and their corresponding skew anomaly probabilities are 0.3, 0.2 and 0.4, then the weighted average of these records will generate the skew anomaly probability of the process segment as (0.91×0.3 + 0.88×0.2 + 0.86×0.4) / (0.91 + 0.88 + 0.86).

[0046] Based on the potential anomaly probability vector V of each process segment, a textile process anomaly transfer graph G is constructed to simulate the probability propagation path of the cascading effect of local anomalies in the process chain.

[0047] First, obtain the potential anomaly probability vectors for all process segment units in the current production task.

[0048] Each potential anomaly probability vector consists of multiple normalized anomaly type probabilities, denoted as V=[P1,P2,...,Pm], where P1 to Pm represent the predicted probabilities of m anomaly types such as weft skew, fabric shrinkage, and yarn breakage, with values ​​ranging from 0 to 1.

[0049] Each process segment unit is represented as an independent process node, and the anomaly probability of each dimension in its corresponding potential anomaly probability vector is assigned as a node attribute value, resulting in a set of attributed nodes arranged in process order, denoted as N1, N2, ..., Nn, where n is the total number of process segments.

[0050] Based on the execution order of the process segments in the production instructions, directed connection paths are established sequentially between nodes N1→N2→…→Nn. A directed edge is formed between each pair of adjacent nodes, representing the temporal evolution relationship of the process state.

[0051] To quantify the potential propagation strength of anomalies between adjacent process nodes, anomaly propagation weights are calculated on each directed edge, using the following method: For a certain anomaly type k (k∈[1,m]), let the probability of this anomaly at node Ni be Pik, and the probability at node N(i+1) be P(i+1)k. Then, what is the propagation increment of this anomaly type between Ni and N(i+1)? If Δk is positive, it indicates that the anomaly type has an amplification trend, and Δk is used as the original value of the propagation weight; if Δk is negative or 0, the anomaly type does not have an effective propagation trend on this path, and its propagation weight is recorded as 0. This processing strategy can shield the anomaly self-decay path and enhance the sensitivity of the graph structure to positive anomaly amplification.

[0052] To ensure the comparability of anomaly propagation weights across different paths, positive normalization is applied to the non-zero anomaly propagation weights on all directed edges.

[0053] Let the maximum value of all non-zero propagation weights in the entire graph be Wmax, and the minimum value of non-zero propagation weights be Wmin. Then, for any propagation weight W on a path, the normalized value W′ is calculated as follows: The normalized propagation weight W′ is limited to the range of 0 to 1, and is used to uniformly assess the degree of abnormal growth under different paths.

[0054] Using the aforementioned process nodes with abnormal attributes as node elements in the graph structure, and the connection paths between nodes with positive propagation weights as directed edge elements, a textile process mutation transfer graph G is constructed.

[0055] Graph G is a directed weighted graph where nodes represent process segments in the production process, edges represent the forward propagation paths of anomalies between adjacent segments, and edge weights represent the intensity of anomaly increments. Once the graph structure is established, the cumulative weight of each path in the graph can serve as a measure of the cumulative effect of anomalies, providing structured input for subsequent process instability prediction, scheduling adjustments, or anomaly intervention decisions.

[0056] Based on the current operating status parameters F of the production equipment and the textile process anomaly transfer graph G, a graph neural network model is trained to predict process instability and output an early warning factor R.

[0057] For each process segment unit in the current production task, the operating status parameters of the corresponding equipment are collected in real time. The operating status parameters include: Real-time tension fluctuation value: refers to the average first derivative of the warp or weft tension curve per unit time, reflecting tension stability; Spindle vibration frequency: The main vibration frequency of the spindle in the high-frequency range (e.g., 50 to 200 Hz) is measured by a vibration accelerometer; Drive power change rate: The change in drive motor power per unit time is calculated by the power monitoring module; Temperature and humidity disturbance amplitude: The maximum change in the ambient temperature and humidity curve within a specified time window.

[0058] After normalization, the above parameters are used as multidimensional state vectors and bound to the corresponding process segment nodes, so that each node has dual features of process anomaly probability attribute and operating state characteristic attribute.

[0059] The sequence of process nodes bound with running status parameters is input into the graph neural network model along with the completed textile process variation transition graph.

[0060] Graph neural networks employ a graph convolutional network structure, using node feature vectors as input feature matrices and the propagation weights of edges in the directed graph as adjacency matrices for encoding.

[0061] The node feature vector is composed of the following two parts concatenated: Anomaly probability attribute vector: predicted probability of anomaly types such as weft skew, fabric shrinkage, and yarn breakage; Operating status feature vector: real-time parameters such as tension fluctuation value, vibration frequency, power change rate, temperature and humidity disturbance.

[0062] The weight of each edge in the adjacency matrix represents the anomaly propagation intensity. During the forward propagation process, the model aggregates the information of each node using the connection relationships in the graph structure, thereby achieving joint modeling of process timing and structural relationships.

[0063] To enable effective model learning, a training sample set is constructed based on historical production data.

[0064] Each training sample corresponds to a process segment node, including its bound operating status parameters and anomaly probability attributes as input features, and the historical instability states of the process segment as supervision labels. The instability states are records of abnormal events such as fabric defects, process interruptions, or major parameter deviations.

[0065] The model training objective is to minimize the binary cross-entropy loss function between the predicted instability probability and the actual instability state, calculated as follows: For each sample, let the true label be y (which takes the value 0 or 1) and the predicted probability be p, then the single-sample loss value is: L represents the loss value for a single sample, used to measure the deviation between the model's predicted output and the true label; y represents the true label of the sample, with a value of 0 or 1, where 1 indicates that an instability event has occurred in the process segment and 0 indicates that it has not occurred; p represents the probability that the graph neural network model predicts that the sample will become unstable, with a value ranging from 0 to 1; the total loss value is the weighted average of the loss values ​​of all training samples, and the model parameters are iteratively optimized through backpropagation and gradient descent algorithms until the loss function converges or the accuracy of the validation set reaches the preset target (such as an accuracy of not less than 90%).

[0066] After the model training is completed and converged, in the actual production process, all process segment nodes under the current task are input into the graph neural network model, and the predicted probability of process instability for each process segment is output, ranging from 0 to 1.

[0067] Set an instability warning threshold T (e.g., 0.65). If the predicted instability probability of a certain process segment is greater than T, it is determined that it has potential instability risk.

[0068] In the model output, segments exceeding the threshold will be assigned a warning factor R, which is a Boolean value. A value of 1 indicates that early intervention is required, while a value of 0 indicates that the operation is normal.

[0069] The early warning factor R can be used for production scheduling optimization, equipment parameter self-adjustment, or manual inspection priority ranking to achieve proactive identification and response to potential process risks in textile production.

[0070] The warning factor R is mapped to the current production line resource configuration map, the resource allocation priority is adjusted, and the optimal matching process routing path C is generated.

[0071] Obtain the early warning factors for all process segments in the current production task. The early warning factors are Boolean variables, with a value of 1 indicating that the process segment has an instability risk and a value of 0 indicating that the risk is acceptable.

[0072] Establish a resource configuration diagram for the current production line. The resource configuration diagram consists of multiple process resource nodes. Each node represents a type of or a specific piece of equipment resource. The resource node contains the following attribute information: The current equipment load status indicates the percentage of scheduled tasks for this equipment. The remaining available time window represents the idle time period of the device in the current schedule; Historical failure frequency indicates the statistical number of historical instability events of the equipment per unit time.

[0073] Each process segment is mapped to an available resource node that meets its process requirements (such as weaving, dyeing, and finishing), forming a "process segment – ​​resource node" mapping pair.

[0074] For each of the above resource nodes, a resource availability score is calculated based on its operating status and the early warning factor value of the mapped process segment. The score calculation method is as follows: Let the load ratio of resource node i be Li, the remaining available time be Ti, the historical failure rate be Fi, and the warning factor of the mapped process segment be Ri; The resource availability score Si can be calculated using the following expression: Here, α, β, γ, and δ are adjustable weighting coefficients used to control the influence of each factor on the score. For example, in high-risk scenarios, the weight of δ can be increased to enhance the suppression effect of instability warning. The higher the score Si, the more suitable the resource node is for executing the corresponding process segment task. Based on the score results of each resource node, they are sorted in descending order to obtain a resource priority list, providing a sorting basis for subsequent path search.

[0075] Based on the resource priority list and the temporal dependencies between process segments, a resource scheduling path search problem is constructed.

[0076] Using dynamic programming, a state transition table is constructed step by step from front to back in the resource allocation graph. Each process segment is taken as a stage, and its matching resource nodes are taken as states. The transition paths between states are scored and accumulated.

[0077] During state transition, for process segments with a warning factor of 1, only resource nodes with a score higher than the set threshold T are allowed to be matched, thereby controlling the possibility of high-risk nodes entering the path.

[0078] All feasible paths are retained in the process route candidate set, and each path contains a sequence of resource nodes and their corresponding scores and cumulative instability risk values.

[0079] For each path in the entire process route candidate set, calculate two key metrics: The overall availability score of a path represents the weighted average of the scores of all resource nodes in the path. The total risk value of the path represents the sum of the warning factors of all process segments in the path, reflecting the overall potential degree of instability.

[0080] Finally, the path with the highest resource score and the lowest total risk value is selected from the candidate paths as the optimal matching process routing path C.

[0081] Path C will be used to guide resource scheduling and equipment deployment for subsequent process segments, and can be dynamically replanned based on real-time status changes.

[0082] Data on equipment micro-vibration characteristics, temperature and humidity gradients, and tension fluctuations are collected during the execution path C of the production line. The collected results are then input into a graph neural network model for error back-off training, enabling the model to learn and update itself dynamically.

[0083] During the execution of the optimally matched process routing path C, for each process segment node in the path, key operating status data of its associated equipment are collected in real time, including: Equipment micro-vibration characteristics: High-frequency vibration signals of key parts of the equipment are measured using a triaxial accelerometer to extract the main frequency distribution and fluctuation amplitude; Ambient temperature and humidity gradient: The temperature and humidity changes per unit time are recorded by a digital thermometer and hygrometer, and the temperature and humidity derivatives within the time window are calculated. Tension fluctuation curve: Obtain the trend of tension change of warp or weft yarn per unit time, and calculate its first difference and fluctuation standard deviation.

[0084] After normalizing the above data, we construct a structured set of operational feedback feature vectors by mapping time to process segment nodes, representing the operational performance of process path C during actual execution.

[0085] The constructed operational feedback feature vector is paired and compared with the instability prediction probability output by the graph neural network model for each process segment node before path C is generated.

[0086] Based on the output quality and on-site event records after the actual execution of path C, determine whether an instability event occurred in each process segment. If it occurred, such as fabric weft skew, tension yarn breakage, or emergency equipment shutdown, mark it with a true label of 1, indicating an unstable state; if it did not occur, mark it with a label of 0. The true label and the predicted probability constitute the supervised sample set. Each sample includes: Input: Run feedback feature vector; Predicted value: The probability of instability output by the model; Real label: The binary state of whether instability has actually occurred.

[0087] This supervised sample set is used for model performance evaluation and error correction training.

[0088] The predicted value of each sample in the supervised sample set is compared with the true label, and the error is measured using the binary cross-entropy loss function. The overall bias level of the model's current prediction error is obtained by averaging the loss values ​​of all samples.

[0089] The calculated loss function value is used as the optimization objective. The parameter gradient of the graph neural network model is calculated through the backpropagation algorithm, and the model weight parameters are incrementally updated using an adaptive gradient optimizer (such as the Adam optimization algorithm).

[0090] Each training round only updates the local subgraph parameters of the nodes and adjacency structures involved in the current path feedback, in order to control the training cost and ensure that the model can be iterated online in production.

[0091] After the update, the predictive ability of the graph neural network model will gradually adapt to the actual performance of the new process path C, thereby improving the self-learning ability and dynamic adaptability of the predictive model.

[0092] Example 2, please refer to Figure 2 As shown in this embodiment, a textile intelligent production management system includes: Anomaly Modeling Module: Constructs micro-anomaly response maps for each process in textile production. These micro-anomaly response maps represent anomaly patterns formed under trial production conditions for different fabric types, process formulas, and equipment configurations. Feature matching module: Slices the production instructions in the formal production task into multiple process segment units P1, P2, ..., Pn, where n is the total number of process segment units, and matches the potential anomaly probability vector V of each process segment unit based on the micro-anomaly response map; Anomaly propagation mapping module: Based on the potential anomaly probability vector V of each process segment, construct a textile process anomaly transfer map G to simulate the probability propagation path of the cascading effect of local anomalies in the process chain. Early warning generation module: Based on the current operating status parameter F of the production equipment and the textile process anomaly transfer diagram G, the module predicts process instability by training a graph neural network model and outputs an early warning factor R. Path optimization module: Maps the early warning factor R to the current production line resource configuration map, adjusts the resource allocation priority, and generates the optimal matching process routing path C; Feedback optimization module: Collects equipment micro-vibration characteristics, temperature and humidity gradients, and tension fluctuation data when the production line executes path C, and inputs the collected results into the graph neural network model for error back-off training to achieve self-learning and dynamic updating of the model.

[0093] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for textile intelligent production management, characterized in that: The method comprises the following steps: constructing a micro-abnormal response graph of each process in the textile production process, wherein the micro-abnormal response graph represents abnormal patterns formed under test production conditions of different cloth types, process recipes and equipment configurations; slicing the production instructions in the formal production task into multiple process segment units P1, P2,..., Pn, n being the total number of process segment units, and matching a potential abnormal probability vector V of each process segment unit based on the micro-abnormal response graph; constructing a textile process abnormality transfer graph G according to the potential abnormal probability vector V of each process segment, which is used to simulate the probability propagation path of the cascade effect of local abnormalities in the process chain; based on the running state parameters F of the current production equipment and the textile process abnormality transfer graph G, performing process instability prediction by training a graph neural network model, and outputting a warning factor R; mapping the warning factor R to the current production line resource configuration graph, adjusting the resource allocation priority, and generating an optimally matched process routing path C; collecting device micro-vibration characteristics, temperature and humidity gradient and tension fluctuation data when the production line executes the path C, and inputting the collection results into the graph neural network model for error rollback training to realize self-learning and dynamic updating of the model.

2. The method for intelligent production management of textiles according to claim 1, characterized in that: Wherein, the construction of the micro-abnormal response graph comprises: collecting test production data under different cloth types, process recipes and equipment configuration conditions, wherein the test production data includes cloth image, tension fluctuation curve, temperature and humidity record and equipment vibration signal; extracting features from the test production data, identifying weft skew, cloth surface shrinkage and yarn breakage abnormal features, and extracting continuity process abnormal features based on the tension and temperature and humidity curve changes; performing feature fusion on the structural abnormal features and the continuity process abnormal features to construct a multi-dimensional abnormal feature vector; based on the multi-dimensional abnormal feature vector, using a clustering analysis algorithm to classify abnormal patterns of process nodes to form a micro-abnormal response graph between each process parameter combination in the textile production process and its corresponding abnormal pattern.

3. The method of claim 1, wherein: Wherein, the matching of the potential abnormal probability vector V of each process segment unit based on the micro-abnormal response graph comprises: analyzing the complete production instructions in the formal production task, dividing the production instructions into multiple continuous process segment units according to the equipment running stage, cloth characteristic conversion point and key process parameter change point, each process segment unit corresponding to an independent process execution interval; for each process segment unit, extract its corresponding cloth type, process recipe number and equipment configuration parameters as query conditions to search for matching records in the pre-constructed micro-abnormal response graph; based on the historical abnormal feature vectors retrieved in the graph, perform feature similarity calculation on the current process segment unit, and use cosine similarity index to compare the similarity degree; fuse the matching results with similarity higher than the set threshold to generate the potential abnormal probability vector corresponding to each process segment unit.

4. The method of claim 1, wherein: Wherein, the construction of the textile process abnormality transfer graph G according to the potential abnormal probability vector V of each process segment comprises: obtain a plurality of potential abnormal probability vectors of process segment units, take each abnormal probability in each potential abnormal probability vector as a process node attribute, and form a process node sequence with abnormal probability attributes; establish a directed connection relationship between process nodes according to the sequence of the process segment units, and construct an abnormal propagation weight according to the difference of the same abnormal type probability in adjacent process nodes; perform forward normalization processing on the abnormal propagation weight of adjacent process nodes; take the process nodes and the corresponding directed connection paths as graph structure elements, and construct a textile process anomaly transfer graph G.

5. The method for intelligent production management of textiles according to claim 1, characterized in that: In the formula, the process of predicting process instability by training the graph neural network model and outputting a warning factor R includes: collecting production equipment running state parameters corresponding to each process segment, the running state parameters including real-time tension fluctuation values, spindle vibration frequencies, driving power change rates, and temperature and humidity disturbance amplitudes, and binding the state parameters with the process segment nodes one by one; inputting the process node sequence bound with the equipment running state parameters and the textile process anomaly transfer graph into the graph neural network model, the graph neural network model performing graph structure coding based on node attributes and edge weights; training the graph neural network model in a supervised learning manner, the training data including instability labels and corresponding running state features of historical process segments, and the training target being to minimize the binary cross-entropy loss between the predicted value and the real instability state; after the model converges, outputting the process instability probability of each process segment by the graph neural network, and generating a warning factor based on a set threshold.

6. The method of claim 1, wherein: In the formula, the process of mapping the warning factor R to a current production line resource configuration graph, adjusting the resource allocation priority, and generating an optimally matched process routing path C includes: mapping the warning factors of each process segment with the process resource nodes in the current production line resource configuration graph one by one; calculating a resource availability score according to the load state, the remaining available window of each resource node, and the warning factor value of the corresponding process segment, and performing priority sorting on the resource nodes according to the score result; based on the priority sorting result of the resource nodes, combining the time sequence dependency relationship of each process segment, and searching for a continuous feasible path in the resource configuration graph by using a dynamic programming algorithm to form a process routing candidate set that meets the condition of minimizing instability risk; in the process routing candidate set, selecting a path with the highest cumulative resource score and the lowest total risk value as the optimally matched process routing path C.

7. The method of claim 1, wherein: In the formula, the process of collecting equipment micro-vibration features, temperature and humidity gradient, and tension fluctuation data when the production line executes the path C, and inputting the collection results into the graph neural network model for error backtracking training includes: in the execution process of the process routing path C, real-time collection of micro-vibration features, environmental temperature and humidity gradient, and tension fluctuation curve of the corresponding equipment to construct a running feedback feature vector corresponding to the process segment node one by one; comparing the running feedback feature vector with the predicted instability probability of each process segment output by the graph neural network model before path generation, labeling a real label according to whether an actual process instability event occurs, and forming a supervised sample set; Taking a supervised sample set as input, a binary cross-entropy loss between a predicted probability and an actual label is calculated, and the loss function is used to represent the error deviation between the current model prediction and the actual production result; The error deviation is returned to the graph neural network model through a back propagation algorithm, and the model parameters are incrementally trained based on a gradient update strategy.

8. A textile intelligent production management system for implementing the textile intelligent production management method of any one of claims 1-7, characterized in that: Comprise: An abnormal modeling module: construct a micro abnormal response atlas of each process in the textile production process, which represents the abnormal mode formed under the test production conditions of different fabric types, process recipes and equipment configurations; A feature matching module: the production instructions in the formal production task are sliced and divided into multiple process segment units P1, P2,..., Pn, n is the total number of process segment units, and the potential abnormal probability vector V of each process segment unit is matched based on the micro abnormal response atlas; An abnormal conduction mapping module: according to the potential abnormal probability vector V of each process segment, a textile process anomaly transfer graph G is constructed, which is used to simulate the probability propagation path of the cascading effect of local anomalies in the process chain; An early warning generation module: based on the running state parameters F of the current production equipment and the textile process anomaly transfer graph G, the process instability is predicted by training the graph neural network model, and the early warning factor R is output; A path optimization module: the early warning factor R is mapped to the current production line resource configuration graph, the resource allocation priority is adjusted, and the optimal matching process routing path C is generated; A feedback optimization module: collect the device micro-vibration characteristics, temperature and humidity gradient and tension fluctuation data when the production line executes the path C, and input the collected results into the graph neural network model for error rollback training, so as to realize the self-learning and dynamic updating of the model.

Images