Textile production quality control method

Abstract

The invention relates to the technical field of image processing, and discloses a textile production quality control method, which comprises the following steps: acquiring multi-modal data of a textile, the multi-modal data comprising fabric image data and multi-channel process parameter time sequence data; constructing a space-time diagram attention network; extracting and fusing the spatial features of the fabric image data and the time features of the process parameter time sequence data to obtain a fusion representation vector; establishing a Gaussian process regression model, constructing a flaw propagation field, and calculating a flaw density gradient and a state transition probability; constructing a structural causal model of the production process, and determining an intervention priority of the key process parameters; and establishing an optimization model, solving the model to obtain an optimal process parameter adjustment combination, and generating an equipment control instruction. According to the method, the fabric image and the process parameter multi-modal data are obtained and deeply fused, the internal mapping relation between the apparent quality and the production process state is constructed, and the problem that in a traditional method, analysis of reasons caused by data islands is cumbersome is solved.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a method for controlling the quality of textile production. Background Art

[0002] In the modern textile industry, product quality is a core factor determining a company's competitiveness. To ensure and improve textile quality, various monitoring systems are typically deployed on production lines. Existing quality control technologies rely primarily on two main branches: machine vision-based fabric surface defect detection, such as automated optical inspection (AOI) systems. These systems capture fabric images with high-speed cameras and use image processing algorithms to identify local defects such as broken warps, broken wefts, and stains. Sensor-based production process monitoring, such as statistical process control (SPC), monitors key process parameters such as yarn tension, machine speed, and temperature and humidity to determine whether the production process is stable. However, in practice, these two technologies often operate independently, creating data silos. Image inspection systems can only identify the "identity" of defects, while process parameter monitoring only reflects the "process status." The lack of effective and in-depth dynamic correlation between the two makes it difficult to directly link apparent quality issues with specific process parameter fluctuations. In addition, traditional defect detection and analysis is often limited to static, isolated defect points, and lacks a macroscopic characterization of the spatial distribution, aggregation, and evolution trends of defects on the fabric plane, making the tracing and prediction of quality problems extremely difficult and heavily dependent on the personal experience of front-line engineers.

[0003] Existing technologies also have significant limitations in fault diagnosis and process optimization. When quality anomalies are detected, traditional methods often use correlation analysis to identify possible causes, analyzing the statistical relationship between defect rates and changes in various process parameters. However, correlation does not equate to causation. Numerous process parameters are coupled and influence each other during production. Simple correlation analysis is likely to reveal spurious correlations, making it difficult to pinpoint the root drivers of quality fluctuations. When making process adjustments, operators often rely on experience or fixed rulebooks, making tentative adjustments to single variables or within a small range. This partial optimization approach, which treats the symptoms but not the root cause, ignores the complex physical constraints (such as tension balance) and synergistic effects between process parameters. This not only results in inefficient adjustments, but can also disrupt the overall balance of the production system through inappropriate interventions, potentially leading to new and more complex quality issues. Therefore, the industry urgently needs a new intelligent quality control approach that can deeply integrate multi-source data, reveal the causal mechanisms of quality issues, and provide globally optimal control strategies to overcome current technical bottlenecks. Summary of the Invention

[0004] The present invention provides a textile production quality control method to solve the problem in the above-mentioned prior art that image detection and parameter monitoring lack deep dynamic correlation, cannot directly link apparent defects with fluctuations in specific process parameters, and makes it difficult to trace quality problems.

[0005] The textile production quality control method of the present invention comprises the following steps: Acquire multimodal data on predetermined nodes of a textile production line, wherein the multimodal data includes fabric image data continuously collected along the fabric running direction and multi-channel process parameter time series data collected synchronously with the fabric image data; construct a spatiotemporal graph attention network, wherein the nodes of the spatiotemporal graph attention network are process parameter sensors and image segmentation areas of the fabric image, and the edges of the spatiotemporal graph attention network are determined based on the mapping relationship between the physical position of the sensor and the image area; utilize the spatiotemporal graph attention network to extract and fuse the spatial features of the fabric image data and the temporal features of the process parameter time series data to obtain a fused representation vector of the fabric state; establish a Gaussian process regression model based on the The fused characterization vector is used as input to construct a defect propagation field that characterizes the spatial distribution probability of fabric defects, and the defect density gradient and the state transition probability of defect types between adjacent areas are calculated; a structural causal model of the production process is constructed based on historical data, and the causal effect of each process parameter on the defect density gradient is quantified through causal inference, and the intervention priority of key process parameters is determined; the key process parameters to be adjusted are determined according to the intervention priority, and an optimization model is established with minimizing the overall entropy of the defect propagation field as the optimization goal and with a physical model containing the yarn tension balance equation as the constraint. The model is solved to obtain the optimal process parameter adjustment combination, and equipment control instructions are generated.

[0006] Preferably, the construction of the spatiotemporal graph attention network includes: dividing the fabric image along the fabric running direction and width direction into multiple image segmentation areas as image nodes, and each process parameter sensor as a parameter node; establishing connecting edges between nodes based on the coverage relationship between the sensor physical monitoring range and the image segmentation area; assigning image features extracted by a pre-trained convolutional neural network to each image node, and assigning time series features extracted by a long short-term memory network to each parameter node as the initial representation of the node.

[0007] Preferably, the spatial-temporal graph attention network is used to extract and fuse the spatial features of the fabric image data and the temporal features of the process parameter time series data to obtain a fused representation vector of the fabric state, including: aggregating the feature information of the adjacent nodes of each node through the graph attention layer of the spatial-temporal graph attention network to update the node representation; all updated node representations are spliced ​​and input into the fully connected network to obtain the fused representation vector of the fabric state.

[0008] Preferably, the Gaussian process regression model is established, and the fused characterization vector is used as input to construct a defect propagation field that characterizes the spatial distribution probability of fabric defects, including: taking the fused characterization vector as input and the defect type and position in the historical data as output for training; using the trained model to predict the posterior probability distribution of the defect type for each current image segmentation area, thereby constructing the defect propagation field.

[0009] Preferably, the calculation of the defect density gradient and the state transition probability of the defect type between adjacent areas includes: calculating the change in the posterior probability of a specific defect type between adjacent image segmentation areas along the fabric running direction as the defect density gradient; and calculating the state transition probability of the defect type between adjacent image segmentation areas based on the posterior probability distribution.

[0010] Preferably, the determination of the intervention priority of the key process parameters includes: using a causal discovery algorithm to learn a structural causal model between each process parameter and the defect density gradient from historical data; based on the structural causal model, quantifying the causal effect of each process parameter on the defect density gradient through causal inference, and sorting the process parameters according to the absolute value of the causal effect to obtain the intervention priority.

[0011] Preferably, the objective function of the optimization model is to minimize the total entropy of the defect probability distribution of all image segmentation areas in the defect propagation field; the constraints of the optimization model include the value range constraints of each process parameter to be adjusted, and the physical model constraints based on the yarn tension balance equation.

[0012] Preferably, the objective function of the optimization model is to minimize the total entropy of the defect probability distribution of all image segmentation areas in the defect propagation field; the constraints of the optimization model include the value range constraints of each process parameter to be adjusted and the physical model constraints based on the yarn tension balance equation.

[0013] Preferably, solving the optimization model to obtain the optimal process parameter adjustment combination includes: using a sequential quadratic programming algorithm to solve the optimization model to obtain a set of optimal process parameters that minimize the objective function and meet all constraints.

[0014] Preferably, the pre-trained convolutional neural network is a residual network.

[0015] The beneficial effects of the present invention are as follows: the present invention constructs an intrinsic and structured mapping relationship between apparent quality and production process status by acquiring and deeply fusing multimodal data of fabric images and process parameters, thus overcoming the problem of superficial cause analysis caused by data islands in traditional methods. By converting discrete defect points into continuous defect propagation fields for modeling, the spatial distribution trend of defects is quantified and predicted from a macroscopic perspective, providing a forward-looking perspective for quality control. The present invention uses causal inference to replace simple correlation analysis, which can accurately locate the root driving factors that lead to quality problems and significantly improve the accuracy of fault tracing. On this basis, the constructed global optimization adjustment strategy comprehensively considers the integrity of the defect propagation field and the physical constraints between process parameters. The parameter adjustment combination given can not only efficiently eliminate defects, but also ensure the overall stability of the production system, avoiding new problems that may be caused by local and tentative adjustments, thereby realizing refined and systematic control of textile production quality. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 A schematic diagram of a process for controlling the quality of textile production provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0017] The embodiments of the present invention are described in detail below, and examples of the embodiments are shown in the accompanying drawings. The embodiments described below with reference to the accompanying drawings are exemplary and intended to be used to explain the present invention, but should not be understood as limiting the present invention.

[0018] like Figure 1 As shown, the textile production quality control method provided by the embodiment of the present invention specifically includes the following steps: S1. Acquire multimodal data at a predetermined node of a textile production line, wherein the multimodal data includes fabric image data continuously collected along a fabric running direction and multi-channel process parameter time series data collected synchronously with the fabric image data.

[0019] Specifically, a linear array camera is installed at the loom entrance to continuously capture fabric images, while tension, speed, and temperature sensors are deployed at the warp creel, let-off beam, and take-up roller. Using a unified clock signal from the main drive shaft encoder, each frame of image data and the corresponding sensor reading are time-stamped, forming a one-to-one corresponding image-parameter data pair.

[0020] S2. Construct a spatiotemporal graph attention network, where the nodes of the spatiotemporal graph attention network are the process parameter sensors and the image segmentation areas of the fabric image, and the edges of the spatiotemporal graph attention network are determined based on the mapping relationship between the physical position of the sensor and the image area; use the spatiotemporal graph attention network to extract and fuse the spatial features of the fabric image data and the temporal features of the process parameter time series data to obtain a fused representation vector of the fabric state.

[0021] Specifically, a single fabric image is divided into M by N grid-like image segments as image nodes, and each process parameter sensor is a parameter node. If a sensor physically affects a certain image segment, an edge is established between the corresponding nodes. For example, the left-side warp-feeding beam tension sensor node establishes edges with all image nodes, but the connection weight with the left-side image segment node is higher. The residual network (ResNet) is used to extract the image features of each image segment, and the long short-term memory (LSTM) network is used to extract the temporal features of each process parameter. These features are then used as the initial node embedding and input into the graph attention network (GAT). The multi-head attention mechanism aggregates the information of neighboring nodes, and finally outputs a fused representation vector for each image segment node.

[0022] S3. Establish a Gaussian process regression model, use the fused representation vector as input, construct a defect propagation field that represents the spatial distribution probability of fabric defects, and calculate the defect density gradient and the state transition probability of defect types between adjacent image segmentation areas.

[0023] Specifically, the fused representation vector output by the spatiotemporal graph attention network is used as the input to a Gaussian process regression model, with the two-dimensional coordinates of the fabric serving as indices. The Gaussian process regression model outputs the mean and variance of the probability of a defect occurring at each coordinate point. These probabilities together constitute a continuous defect propagation field. Taking the partial derivative of the probability mean function of this defect propagation field yields the defect density gradient along the warp and weft directions of the fabric. Based on the joint probability distribution characteristics of the Gaussian process, the posterior probability of the defect type for any two adjacent image segments is calculated, resulting in a transition probability matrix for the defect type from one image segment to another.

[0024] S4. Construct a structural causal model of the production process based on historical data, quantify the causal effect of each process parameter on the defect density gradient through causal inference, and determine the intervention priority of key process parameters.

[0025] Specifically, an initial causal directed acyclic graph (DAG) is constructed based on textile process knowledge, where nodes represent process parameters and defect density gradients. This graph structure is then verified and modified using historical production data, using methods such as the PC algorithm or the LiNGAM model, to produce the final structural causal model. Using a backdoor adjustment formula, intervention simulations are performed on the structural causal model to calculate the average causal effect on the defect density gradient for each process parameter change per unit. These effect values ​​are sorted from highest to lowest to obtain a priority list for interventions on each process parameter.

[0026] S5. Determine the key process parameters to be adjusted based on the intervention priority, establish an optimization model with minimizing the overall entropy of the defect propagation field as the optimization goal and a physical model containing the yarn tension balance equation as a constraint, solve the optimization model to obtain the optimal process parameter adjustment combination, and generate equipment control instructions.

[0027] Specifically, the first N key process parameters in the intervention priority list are selected as decision variables. The overall entropy of the defect propagation field is obtained by calculating the Shannon entropy of the defect probability distribution at each point in the field and integrating it. The tension relationship between the yarn transmitted between the various components of the loom, such as the balance relationship between the warp tension, the cloth-fell tension and the winding tension, is expressed as a set of algebraic or differential equations as physical constraints. Then, a nonlinear programming problem is constructed and solved using a sequential quadratic programming algorithm or an interior point method to obtain a set of optimal process parameter settings that minimize the objective function and satisfy all constraints. Finally, this set of settings is converted into control signals for the corresponding actuators, such as inverters or servo motors, and sent to the equipment controller.

[0028] In an optional embodiment, constructing a spatiotemporal graph attention network includes: The fabric image is segmented into multiple image segmentation areas along the fabric running direction and width direction as image nodes, and each process parameter sensor is used as a parameter node; connecting edges between nodes are established based on the coverage relationship between the sensor's physical monitoring range and the image segmentation area; image features extracted by a pre-trained convolutional neural network are assigned to each image node, and time series features extracted by a long short-term memory network are assigned to each parameter node as the initial representation of the node.

[0029] Specifically, this step aims to unify multi-source, heterogeneous production data into a single graph structure. For example, a fabric image with a width of 1920 pixels and a length of 1080 pixels along the running direction can be evenly segmented into 192 x 108 image segments of 10 x 10 pixels along the running and width directions, with each segment representing an image node. Furthermore, assuming that five sensors are deployed on the production line to monitor warp tension, weft tension, loom speed, let-off rate, and temperature, these five sensors serve as parameter nodes.

[0030] Specifically, the connection relationship between nodes reflects the scope of influence in the physical world. For example, a warp tension sensor monitors the area between the 500th and 1000th pixels in the width direction of the fabric. Then, the parameter node corresponding to the sensor will establish a connection edge with all image nodes within this width range. Subsequently, the nodes are assigned initial features. For each image node, a ResNet50 network pre-trained on the ImageNet dataset is used to extract a 2048-dimensional feature vector of its 10 by 10 pixel image block. For each parameter node, it is assumed that it collects data once per second in the past 60 seconds, forming a time series of length 60. After the sequence is processed by a long short-term memory network, a 128-dimensional feature vector is generated as its initial representation.

[0031] In an optional embodiment, the spatial features of the fabric image data and the temporal features of the process parameter time series data are extracted and fused using the spatiotemporal graph attention network to obtain a fused representation vector of the fabric state, including: aggregating the feature information of the adjacent nodes of each node through the graph attention layer of the spatiotemporal graph attention network to update the node representation; splicing all updated node representations and inputting them into a fully connected network to obtain the fused representation vector of the fabric state.

[0032] Specifically, the graph attention layer is the core of information aggregation. For any node in the graph, such as an image node representing a small area in the center of the fabric, this layer calculates the contribution of all its adjacent nodes, including image nodes above, below, left, and right, as well as process parameter nodes that may cover this area, to its feature update. For example, if an adjacent tension sensor node exhibits abnormal fluctuations, the graph attention mechanism assigns a higher weight to the parameter node's features, thereby incorporating more information about the abnormal tension when updating the representation of the current image node.

[0033] After iterative updates through one or more graph attention layers, each node acquires a new representation that incorporates information from its local spatiotemporal neighborhood. Assume that the graph contains 20,736 image nodes and 5 parameter nodes, and the representation of each node is updated to 256 dimensions. At this point, these 20,741 256-dimensional vectors are concatenated in a predetermined order to form a single, very high-dimensional long vector. This long vector is then input into a fully connected network consisting of three hidden layers. The network uses nonlinear transformations and layer-by-layer dimensionality reduction to ultimately output a fused representation vector, for example, of 1024 dimensions. This fused representation vector highly condenses the visual information of the entire fabric image at the current moment and the dynamic information of all key process parameters.

[0034] In an optional embodiment, the Gaussian process regression model is established, and the fused characterization vector is used as input to construct a defect propagation field that characterizes the spatial distribution probability of fabric defects, including: taking the fused characterization vector as input and the defect type and location in historical data as output for training; using the trained model to predict the posterior probability distribution of the defect type for each current image segmentation area, thereby constructing the defect propagation field.

[0035] Specifically, the training phase uses historical production data. 10,000 samples are selected from the historical database. Each sample corresponds to a 1024-dimensional fused representation vector for a point in time and the actual defect map corresponding to that point in time. This map records the defect type for each segmented area in the image, using a numerical code such as 0 for normal, 1 for warp breakage, and 2 for weft skew. The model is trained using the fused representation vector as input and the defect map as output, learning the complex mapping relationship between the two.

[0036] When new production data generates a current fusion representation vector, the trained Gaussian process regression model predicts it. The model is unique in that it does not directly output a specific defect type, but instead outputs a posterior probability distribution of the defect type for each image segmentation region. For example, for the image segmentation region in the i-th row and j-th column, the model might predict a 70% probability of warp breakage, a 20% probability of weft skewness, and a 10% probability of normality. Combining these probability distributions for all image segmentation regions creates a defect propagation field covering the entire fabric surface, visually displaying the spatial distribution trends and severity of various defects on the fabric in the form of a probability cloud map.

[0037] In an optional embodiment, the calculation of the defect density gradient and the state transition probability of the defect type between adjacent areas includes: calculating the change in the posterior probability of a specific defect type between adjacent image segmentation areas along the fabric running direction as the defect density gradient; and calculating the state transition probability of the defect type between adjacent areas based on the posterior probability distribution.

[0038] Specifically, the defect density gradient quantifies the rate of defect evolution along the production direction. Assume the fabric is running along the Y-axis, focusing on warp breakage. If the average posterior probability of a warp break in a segmented area with a Y coordinate of 100 is 40%, and in the adjacent row with a Y coordinate of 101, the average probability rises to 45%, then the density gradient of the warp breakage defect in this area is 5%. A high gradient value indicates that the defect is rapidly developing or spreading in that area, making it a region of particular concern.

[0039] State transition probability describes the propagation characteristics of defects from another perspective. Based on historical data and the current posterior probability distribution, the probability of transitioning from one region's state to that of its downstream neighbor can be calculated. For example, through analysis of extensive data, it was found that when the probability of warp breakage in one region exceeds 80%, the probability of warp breakage in its downstream neighbor reaches 90%, while the probability of weft skewness is only 1%. This state transition probability matrix reveals the mutual influence and propagation patterns of different defect types, providing a quantitative basis for predicting the future direction of defects.

[0040] In an optional embodiment, determining the intervention priority of key process parameters includes: using a causal discovery algorithm to learn a structural causal model between each process parameter and the defect density gradient from historical data; based on the structural causal model, quantifying the causal effect of each process parameter on the defect density gradient through causal inference, and sorting the process parameters according to the absolute value of the causal effect to obtain the intervention priority.

[0041] Specifically, this step aims to fundamentally identify the causes of defect deterioration. Historical data is collected, showing time-varying sequences of process parameters such as warp tension and loom speed, along with corresponding sequences of changes in the density gradient of warp break defects. Causal discovery techniques, such as the PC algorithm or the FCI algorithm, are used to analyze this time series data and construct a directed graph representing the causal relationships between them. For example, the model might discover that changes in warp tension are the direct cause of changes in the warp break density gradient, while loom speed indirectly affects defects by affecting warp tension.

[0042] After establishing the causal structure model, causal inference techniques are used to quantify the strength of this effect. For example, through the do calculus, it can be calculated that, under ideal intervention conditions, increasing the warp tension by 1 Newton, while keeping all other parameters constant, will result in an average decrease in the warp density gradient by 0.05 units. Similarly, it is calculated that decreasing the loom speed by 10 revolutions per minute has a causal effect on the warp density gradient of 0.02 units. Comparing the absolute values ​​of these two causal effects, 0.05 is greater than 0.02, so adjusting the warp tension has a higher priority than adjusting the loom speed. By ranking the causal effects of all process parameters, a clear list of intervention priorities is obtained.

[0043] In an optional embodiment, the objective function of the optimization model is to minimize the total entropy of the defect probability distribution of all image segmentation areas in the defect propagation field; the constraints of the optimization model include the value range constraints of each process parameter to be adjusted, and the physical model constraints based on the yarn tension balance equation.

[0044] Specifically, the optimization model aims to ensure that the fabric quality state most consistently approaches a defect-free state. Information entropy is a measure of uncertainty. The more dispersed the defect probability distribution in a region, the higher its entropy, indicating a more uncertain state. For example, if a region has a 50% probability of both warp breakage and normality, its entropy is very high. The optimization goal is to adjust the process parameters so that this distribution approaches a state with a 99% probability of normality and a 1% probability of warp breakage, at which point the entropy is extremely low. Minimizing the total entropy across all regions means driving the entire fabric surface to a deterministic state of defect-free quality with a high probability.

[0045] The optimization process must adhere to practical production constraints. The first type is parameter range constraints. For example, the warp tension adjustment range must be between 20 and 50 Newtons, which is determined by equipment capabilities and material tolerances. The second type is physical law constraints. For example, the yarn tension balance equation describes the interrelationship between let-off rate, loom speed, and warp tension. When searching for the optimal parameter combination, the values ​​of these three parameters must always satisfy this equation. This ensures that the optimization results are physically feasible and stable, avoiding violations of physical laws caused by adjusting a single parameter, which could lead to more serious problems.

[0046] The implementation principle of the textile production quality control method of the embodiment of the present invention is: by acquiring and deeply fusing the multimodal data of fabric images and process parameters, an intrinsic and structured mapping relationship between the apparent quality and the production process status is constructed, which overcomes the problem of superficial cause analysis caused by data islands in traditional quality control methods. By converting discrete defect points into continuous defect propagation fields for modeling, it is possible to quantify and predict the spatial distribution trend of defects from a macro perspective, providing a forward-looking perspective for quality control. In addition, the use of causal inference instead of simple correlation analysis can accurately locate the root driving factors that lead to quality problems, significantly improving the accuracy of fault tracing. Moreover, the constructed global optimization adjustment strategy comprehensively considers the integrity of the defect propagation field and the physical constraints between process parameters. The parameter adjustment combination given can not only effectively eliminate defects, but also ensure the overall stability of the production system, avoiding new problems that may be caused by tentative adjustments, thereby realizing refined and systematic control of textile production quality.

[0047] Although the embodiments of the present invention have been shown and described above, it will be understood that the above embodiments are illustrative and are not to be construed as limitations on the present invention. A person skilled in the art may change, modify, replace and modify the above embodiments within the scope of the present invention.

Claims

1. A textile production quality control method, characterized in that: The method comprises the following steps: acquiring multimodal data at predetermined nodes of a textile production line, the multimodal data comprising fabric image data continuously collected along a fabric running direction and multi-channel process parameter time series data collected synchronously with the fabric image data; constructing a spatiotemporal graph attention network, wherein the nodes of the spatiotemporal graph attention network are process parameter sensors and image segmentation regions of the fabric image, and the edges of the spatiotemporal graph attention network are determined based on a mapping relationship between the physical position of the sensor and the image region; and utilizing the spatiotemporal graph attention network to extract and fuse the spatial features of the fabric image data with the temporal features of the process parameter time series data to obtain a fused representation vector of the fabric state. A Gaussian process regression model is established, which uses the fused representation vector as input to construct a defect propagation field that represents the spatial distribution probability of fabric defects, and calculates the defect density gradient and the state transition probability of defect types between adjacent regions; A structural causal model of the production process is constructed based on historical data. The causal effect of each process parameter on the defect density gradient is quantified through causal inference, and the intervention priority of key process parameters is determined. The key process parameters to be adjusted are determined according to the intervention priority, and an optimization model is established with minimizing the overall entropy of the defect propagation field as the optimization goal and with a physical model containing the yarn tension balance equation as the constraint. The optimization model is solved to obtain the optimal process parameter adjustment combination, and equipment control instructions are generated.

2. The textile production quality control method according to claim 1, characterized in that: The construction of the spatiotemporal graph attention network includes: dividing the fabric image along the fabric running direction and width direction into multiple image segmentation areas as image nodes, and each process parameter sensor as a parameter node; establishing connecting edges between nodes based on the coverage relationship between the sensor's physical monitoring range and the image segmentation area; assigning image features extracted by a pre-trained convolutional neural network to each image node, and assigning time series features extracted by a long short-term memory network to each parameter node as the initial representation of the node.

3. The textile production quality control method according to claim 2, characterized in that: The method of extracting and fusing the spatial features of the fabric image data and the temporal features of the process parameter time series data using the spatiotemporal graph attention network to obtain a fused representation vector of the fabric state includes: Through the graph attention layer of the spatiotemporal graph attention network, the feature information of the adjacent nodes of each node is aggregated to update the node representation; all updated node representations are spliced ​​and input into the fully connected network to obtain the fused representation vector of the fabric state.

4. The textile production quality control method according to claim 1, characterized in that: The Gaussian process regression model is established, and the fused representation vector is used as input to construct a defect propagation field representing the spatial distribution probability of fabric defects, including: The fused representation vector is used as input and the defect type and location in the historical data is used as output for training; the trained model is used to predict the posterior probability distribution of the defect type for each current image segmentation area, thereby constructing the defect propagation field.

5. The textile production quality control method according to claim 4, characterized in that: The calculation of the defect density gradient and the state transition probability of the defect type between adjacent areas includes: Along the fabric running direction, the change in the posterior probability of a specific defect type between adjacent image segmentation areas is calculated as the defect density gradient; based on the posterior probability distribution, the state transition probability of the defect type between adjacent image segmentation areas is calculated.

6. The textile production quality control method according to claim 5, characterized in that: The determination of intervention priorities for key process parameters includes: A causal discovery algorithm is used to learn a structural causal model between each process parameter and the defect density gradient from historical data. Based on the structural causal model, the causal effect of each process parameter on the defect density gradient is quantified through causal inference, and the process parameters are sorted according to the absolute value of the causal effect to obtain the intervention priority.

7. The textile production quality control method according to claim 6, characterized in that: The causal discovery algorithm is a PC algorithm or a FCI algorithm.

8. The textile production quality control method according to claim 1, characterized in that: The objective function of the optimization model is to minimize the total entropy of the defect probability distribution of all image segmentation areas in the defect propagation field; the constraints of the optimization model include the value range constraints of each process parameter to be adjusted and the physical model constraints based on the yarn tension balance equation.

9. The textile production quality control method according to claim 8, characterized in that: Solving the optimization model to obtain the optimal process parameter adjustment combination includes: using a sequential quadratic programming algorithm to solve the optimization model to obtain a set of optimal process parameters that minimize the objective function and meet all constraints.

10. The textile production quality control method according to claim 2, characterized in that: The pre-trained convolutional neural network is a residual network.

Images