An automated abrasion resistance detection system for a fabric

Abstract

The present application relates to the field of automatic detection, and more particularly to a kind of wear resistance automatic detection system for fabric fabric, the system includes: the image of fiber fabric after loom is off machine is through visual identification model, to generate processing quality grade mark grid chart;Fiber fabric processing parameters are through fully connected network, to generate fiber fabric processing parameter characteristics;Processing quality grade mark grid chart and fiber fabric processing parameter characteristics are through convolutional neural network model, to generate predicted fiber fabric wear resistance;Visual identification model, fully connected network and convolutional neural network model are through DeepSHAP algorithm, to determine the leading factor of the predicted fiber fabric wear resistance.The present application combines the textile processing effect of fabric with fabric process parameters to predict the wear resistance of fabric, and analyzes the leading factor of wear resistance judgment, realizes high-accuracy automatic detection of fabric wear resistance and its cause.

Description

Technical Field

[0001] This invention relates to the field of automated testing, and more particularly to an automated testing system for the abrasion resistance of fabrics. Background Technology

[0002] The abrasion resistance of textile fabrics is a core indicator for evaluating the intrinsic quality and service life of fabrics. It directly determines the end-user experience of clothing, home textiles, and industrial textiles made from the fabric, and is a core item for mandatory testing and enterprise quality control.

[0003] Currently, the industry's commonly used fabric abrasion resistance testing is based on the experimental method of "Textiles - Martindale Method for Determination of Fabric Abrasion Resistance". It relies on specialized abrasion resistance testing equipment to complete the friction cycle test, and then judges the abrasion resistance performance through manual visual rating and strength loss test.

[0004] However, with the advancement of intelligent transformation in the textile industry, fabric manufacturers have placed higher demands on the efficiency and real-time nature of abrasion resistance testing. Traditional experimental testing methods are offline destructive testing methods, which require hundreds to tens of thousands of friction cycle tests and have a testing cycle of several hours. These methods cannot meet the online real-time testing needs of weaving production lines and cannot provide accurate data support for optimizing fabric weaving and processing technology.

[0005] To address the real-time limitations of traditional experimental testing methods, automated fabric abrasion resistance testing technology based on machine vision has become a hot research topic in the industry. Existing technologies mostly employ convolutional neural networks (CNNs) as the core feature extraction model to achieve automated identification and rating of fabric wear appearance.

[0006] However, existing technologies have not achieved spatial alignment between global processing parameters and local visual features of fabrics. They cannot model the differential effects of fiber processing characteristics and fabric material characteristics on fabric abrasion resistance in different textile structural regions, such as warp and weft density, yarn evenness, surface hairiness index, and texture features. This leads to misjudgments in the detection of processing effects of the same fabric composition but different fabric processing techniques. Furthermore, they cannot fully explore the synergistic influence of fiber processing characteristics and fabric process parameters such as fabric composition on abrasion resistance, and the prediction accuracy is insufficient to meet the needs of industrial-grade testing.

[0007] Furthermore, existing technologies suffer from the inherent limitations of black-box models. These models can only output predicted results of abrasion resistance, but cannot quantify the contribution of various processing characteristics and fabric process parameters to abrasion resistance, nor can they accurately pinpoint the dominant factors affecting the abrasion resistance of fabrics.

[0008] In summary, how to combine the textile processing effect and fabric process parameters to predict the abrasion resistance of fabrics, and analyze the dominant factors in abrasion resistance judgment, so as to achieve highly accurate automated detection of fabric abrasion resistance and its causes, is a technical problem that needs to be solved. Summary of the Invention

[0009] Therefore, the present invention provides an automated abrasion resistance testing system for fabrics, which overcomes the problems in the prior art that cannot combine the textile processing effect and fabric process parameters to predict the abrasion resistance of the fabric, and cannot analyze the dominant factors for abrasion resistance judgment, thus failing to achieve high-accuracy automated detection of fabric abrasion resistance and its causes.

[0010] To achieve the above objectives, this invention proposes an automated abrasion resistance testing system for fabrics, comprising: The fabric image processing feature recognition module is used to generate a raster map with processing quality grade annotations from the image of the fiber fabric after it comes off the loom through a visual recognition model. The fabric parameter feature extraction module is used to extract the fiber fabric processing parameters through a fully connected network to generate fiber fabric processing parameter features. The textile fabric feature fusion detection module is used to combine the processing quality grade labeled raster image and fiber fabric processing parameter features through a convolutional neural network model to generate a predicted abrasion resistance of the fiber fabric. The abrasion resistance cause analysis module is used to use the DeepSHAP algorithm to determine the dominant factors in predicting the abrasion resistance of the fiber fabric by combining the visual recognition model, the fully connected network, and the convolutional neural network model.

[0011] Furthermore, the fabric image processing feature recognition module includes: The feature recognition submodule is used to extract features from the fiber fabric image using a pre-trained DeepLabv3 sub-model to generate a fabric processing feature map. The classification output submodule is used to generate the processing quality grade labeling raster map by passing the fabric processing feature map through a raster output network. The visual recognition model includes a pre-trained DeepLabv3 sub-model and a raster output network.

[0012] Furthermore, the classification output submodule includes: A global average pooling unit is used to map the fabric processing feature map to a raster image size through a global average pooling layer to generate a raster image size fabric processing feature map. The classification output unit is used to classify the fabric processing feature map of the grid size to the processing feature level through the convolution output layer to generate the processing quality level labeled grid map; The grid output network includes a global average pooling layer and a convolutional output layer.

[0013] Furthermore, the fabric image processing feature recognition module also includes: The frozen training submodule is used to process the input data of the customized dataset through the pre-trained DeepLabv3 sub-model and the raster output network to generate a raster map with processing quality level annotation. The cross-entropy loss function is calculated between the raster map with annotation of the customized dataset and the raster map with processing quality level annotation. After freezing the pre-trained DeepLabv3 sub-model, the raster output network is fine-tuned through the cross-entropy loss function. The Unfreeze Training submodule is used to unfreeze the pre-trained DeepLabv3 sub-model and then fine-tune the pre-trained DeepLabv3 sub-model and the raster output network using the cross-entropy loss function.

[0014] Furthermore, the fabric parameter feature extraction module includes: The high-dimensional feature extraction submodule is used to standardize the fiber fabric processing parameters and then generate high-dimensional fiber fabric processing parameter features through the first fully connected layer. The abstract feature extraction submodule is used to pass the high-dimensional fiber fabric processing parameter features through the second fully connected layer to generate abstract fiber fabric processing parameter features. The smooth feature output submodule is used to generate the fiber fabric processing parameter features by passing the abstract fiber fabric processing parameter features through a linear transformation fully connected layer. The fully connected network includes a first fully connected layer, a second fully connected layer, and a linear transformation fully connected layer.

[0015] Furthermore, the textile fabric feature fusion detection module includes: The attention fusion enhancement submodule is used to enhance the processing quality level annotation raster and fiber fabric processing parameter features through a multi-branch attention mechanism to generate processing feature level enhancement features and fiber process parameter enhancement features. The weighted gating fusion submodule is used to calculate the gating weight based on the processing feature level enhancement feature and the fiber process parameter enhancement feature, and to perform weighted fusion on the processing feature level enhancement feature and the fiber process parameter enhancement feature based on the gating weight to generate an initial fused feature; Abrasion resistance output submodule is used to pass the initial fusion features through the abrasion resistance output head to generate a predicted abrasion resistance of the fiber fabric. The convolutional neural network model includes a multi-branch attention mechanism, gating weights, and a wear-resistant output head.

[0016] Furthermore, the attention fusion submodule includes: The self-attention fusion unit is used to pass the processing quality level labeling raster and fiber fabric processing parameter features through a self-attention mechanism branch to generate processing feature self-enhancing features and fiber process parameter self-enhancing features. A cross-modal attention fusion unit is used to pass the processing quality level annotation raster map and fiber fabric processing parameter features through a cross-modal attention mechanism branch to generate cross-modal enhancement features of processing features and cross-modal enhancement features of fiber process parameters; The feature splicing unit is used to splice the self-reinforcing feature of the processing feature and the cross-modal enhancement feature of the processing feature to generate the level enhancement feature of the processing feature, and to splice the self-reinforcing feature of the fiber process parameter and the cross-modal enhancement feature of the fiber process parameter to generate the enhancement feature of the fiber process parameter. The multi-branch attention mechanism includes a self-attention mechanism branch and a cross-modal attention mechanism branch.

[0017] Furthermore, the wear resistance cause analysis module includes: The initial SHAP value calculation submodule is used to generate initial SHAP values ​​by passing the visual recognition model, fully connected network and convolutional neural network model through the DeepSHAP algorithm; The weight abrasion resistance relationship constraint submodule is used to obtain multiple initial SHAP values ​​corresponding to multiple weights, and to determine whether the weight is proportional to the initial SHAP value. If not, the initial SHAP value is fitted by ordinal regression to generate a corrected SHAP value. If so, the initial SHAP value is used as the corrected SHAP value. Based on the comparison of multiple corrected SHAP values ​​of processing feature level and fiber fabric processing parameters, the dominant factor is determined. The processing parameters of the fiber fabric include weight.

[0018] Furthermore, the wear resistance cause analysis module also includes: The SHAP cotton ratio interaction item calculation submodule is used to determine whether the cotton ratio exceeds the cotton ratio threshold. If so, it calculates the difference between the cotton ratio and the cotton ratio threshold, and calculates the corresponding corrected SHAP value based on the difference.

[0019] Furthermore, the automated wear resistance testing system also includes: The collaborative training module is used to perform collaborative training and optimization of the fully connected network and the convolutional neural network model through a combined loss function, wherein the combined loss function includes a gated fusion sparsity term and a mean squared error loss term.

[0020] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention uses a pre-trained DeepLabv3 sub-model as the backbone for feature extraction. Leveraging its multi-scale modeling capability of Spatial Pyramid Pooling (ASPP), it can accurately capture global structural features that determine the abrasion resistance of fabrics, such as warp and weft density and fabric structure, while also fully preserving microscopic details that determine the wear initiation point, such as yarn evenness, surface hairiness index, and weaving defects. Through the global average pooling layer and convolutional output layer of the raster output network, it directly generates a raster map of processing quality level annotations that is perfectly aligned with the input fabric image space, ensuring the real-time performance of visual detection. Through a two-step fine-tuning training strategy of freezing and thawing, it retains the general texture feature extraction capability of the pre-trained model and achieves accurate adaptation of textile-specific features through a customized fabric dataset. This solves the problems of limited labeled data and easy overfitting of models in the textile industry, and achieves accurate identification of the textile processing effect of fabrics.

[0021] In particular, this invention addresses the issue of mixed fabric process parameters by employing a three-level progressive fully connected coding structure—high-dimensional feature extraction, abstract feature mining, and smooth feature output—to automatically learn the inherent synergistic relationships between different process parameters such as fiber composition, yarn twist, and basis weight. This generates fiber fabric processing parameter features that are fully matched and semantically compatible with the dimensions of visual processing features, avoiding the inaccurate fusion of global process parameters and local visual features in traditional solutions. This enables the collaborative prediction of fabric abrasion resistance by combining the textile processing effect and fabric process parameters, and analyzes the dominant factors in abrasion resistance judgment, thereby achieving highly accurate automated detection of fabric abrasion resistance and its causes. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the structure of an automated abrasion resistance testing system for fabrics according to an embodiment of the present invention; Figure 2 This is a schematic flowchart of an automated abrasion resistance testing system for fabrics according to an embodiment of the present invention. Figure 3 This is a flowchart illustrating the visual recognition model of an automated abrasion resistance testing system for fabrics according to an embodiment of the present invention. Figure 4 This is a schematic diagram of the fully connected network of an automated abrasion resistance testing system for fabrics according to an embodiment of the present invention. Detailed Implementation

[0023] To make the objectives and advantages of the present invention clearer, the present invention will be further described below with reference to embodiments; it should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0024] Preferred embodiments of the present invention will now be described with reference to the accompanying drawings. Those skilled in the art should understand that these embodiments are merely illustrative of the technical principles of the present invention and are not intended to limit the scope of protection of the present invention.

[0025] It should be noted that in the description of this invention, the terms "upper", "lower", "left", "right", "inner", "outer", etc., which indicate directions or positional relationships, are based on the directions or positional relationships shown in the accompanying drawings. This is only for the convenience of description and is not intended to indicate or imply that the device or element must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation of this invention.

[0026] Furthermore, it should be noted that, in the description of this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0027] like Figures 1 to 4 As shown, the present invention provides an automated testing system for the abrasion resistance of fabrics, which overcomes the problems in the prior art that cannot combine the textile processing effect and fabric process parameters to predict the abrasion resistance of fabrics, and cannot analyze the dominant factors for abrasion resistance judgment, thus failing to achieve high-accuracy automated testing of fabric abrasion resistance and its causes.

[0028] like Figure 1 and 2 As shown, this embodiment proposes an automated abrasion resistance testing system for fabrics, comprising: The fabric image processing feature recognition module is used to generate a raster map with processing quality grade annotations from the image of the fiber fabric after it comes off the loom through a visual recognition model. The fabric parameter feature extraction module is used to extract the fiber fabric processing parameters through a fully connected network to generate fiber fabric processing parameter features. The textile fabric feature fusion detection module is used to combine the processing quality grade labeled raster image and fiber fabric processing parameter features through a convolutional neural network model to generate a predicted abrasion resistance of the fiber fabric. The abrasion resistance cause analysis module is used to use the DeepSHAP algorithm to determine the dominant factors in predicting the abrasion resistance of the fiber fabric by combining the visual recognition model, the fully connected network, and the convolutional neural network model.

[0029] Specifically, in this embodiment, the fibers include cotton, wool, silk, linen, synthetic fibers, and blended fibers. The fiber fabric image is acquired from the time the greige fabric comes off the loom, before entering the dyeing, finishing, or printing stages. Each grid of the processing quality grade labeling grid is labeled with the warp and weft density grade, yarn evenness grade, hairiness index grade, and texture type. The fabric at this stage is a natural-colored greige fabric, without dyeing, printing, or coating. The warp and weft interlacing structure and yarn morphology of the fabric are completely exposed, allowing for the accurate extraction of processing features including warp and weft density, yarn evenness, surface hairiness index, and texture characteristics. At the same time, the final abrasion resistance of the fabric can be predicted in advance, and unqualified fabrics can be rejected in advance to avoid subsequent ineffective investment.

[0030] Specifically, the fiber fabric processing parameters include the fiber composition (cotton, polyester, and nylon ratios), yarn twist, and the fabric weave (plain, twill, or satin) and weight. These processing parameters are all fully fixed before the fabric enters the dyeing and finishing process, and are core factors determining the fabric's abrasion resistance. They can be accurately obtained at the point where the fabric comes off the loom and are unaffected by subsequent dyeing and finishing processes.

[0031] It is understandable that warp and weft density, yarn evenness, hairiness index, and texture type are fabric processing quality characteristics. These are indicators that directly determine the stability of the fabric structure during the weaving process, and are also actual production fluctuations that cannot be covered by process parameters. Insufficient warp and weft density reduces fabric tightness, making yarns prone to slippage and disintegration during friction; poor yarn evenness creates weak points in abrasion resistance, leading to preferential wear and breakage; excessive hairiness accelerates fiber shedding during friction, shortening yarn lifespan; and different texture types directly disrupt the stable structure of the warp and weft interweaving. Therefore, combining fabric processing quality characteristics with process parameters to predict fabric abrasion resistance ensures that the predicted results closely match the actual abrasion resistance performance of the greige fabric, rather than relying solely on non-purely theoretical abrasion resistance values ​​based on process parameters.

[0032] like Figure 3 As shown, the fabric image processing feature recognition module further includes: The feature recognition submodule is used to extract features from the fiber fabric image using a pre-trained DeepLabv3 sub-model to generate a fabric processing feature map. The classification output submodule is used to generate the processing quality grade labeling raster map by passing the fabric processing feature map through a raster output network. The visual recognition model includes a pre-trained DeepLabv3 sub-model and a raster output network.

[0033] Specifically, the weights of the pre-trained DeepLabv3 sub-model are pre-trained on the large general dataset ImageNet. Existing research, "Semantic Segmentation Using DeepLabv3+ Model for Fabric Defect Detection," has confirmed that the DeepLabv3 model with MobileNetv2 as its backbone network achieves near-excellent performance in fabric defect detection. Therefore, this embodiment uses the DeepLabv3 sub-model and replaces its output head structure to meet the requirements for recognizing fabric image processing effect features such as warp and weft density levels, yarn evenness levels, hairiness index levels, and texture types.

[0034] Specifically, the pre-trained DeepLabv3 sub-model has a native encoder of the DeepLabv3 architecture. The encoder uses MobileNetv2 as the backbone network and uses the ASPP module to capture multi-scale contextual information. The fabric processing feature map output by the ASPP module is directly input into the raster output network.

[0035] like Figure 3 As shown, the classification output submodule further includes: A global average pooling unit is used to map the fabric processing feature map to a raster image size through a global average pooling layer to generate a raster image size fabric processing feature map. The classification output unit is used to classify the fabric processing feature map of the grid size to the processing feature level through the convolution output layer to generate the processing quality level labeled grid map; The grid output network includes a global average pooling layer and a convolutional output layer.

[0036] Specifically, the global average pooling layer is used to aggregate the fabric processing feature map output by ASPP into a grid of fixed size. The adaptive average pooling strategy ensures that a fixed number of spatial locations can be output regardless of the size of the input image, which facilitates subsequent grid-based analysis.

[0037] Specifically, the convolutional output layer uses a 1×1 convolution with a stride of 1 and zero padding, and incorporates a Softmax activation function. This maps the feature vector of each grid position to the respective level and category probabilities of four fabric image processing features. The category corresponding to the maximum probability is taken as the final labeling result for that grid. The four fabric image processing features include warp and weft density level, yarn evenness level, hairiness index level, and texture type. This enables multi-task output and avoids the prediction of each pixel by the native DeepLabv3, thereby achieving accurate feature recognition at the grid scale and ensuring real-time recognition of fabric image processing effect features.

[0038] Furthermore, the fabric image processing feature recognition module also includes: The frozen training submodule is used to process the input data of the customized dataset through the pre-trained DeepLabv3 sub-model and the raster output network to generate a raster map with processing quality level annotation. The cross-entropy loss function is calculated between the raster map with annotation of the customized dataset and the raster map with processing quality level annotation. After freezing the pre-trained DeepLabv3 sub-model, the raster output network is fine-tuned through the cross-entropy loss function. The Unfreeze Training submodule is used to unfreeze the pre-trained DeepLabv3 sub-model and then fine-tune the pre-trained DeepLabv3 sub-model and the raster output network using the cross-entropy loss function.

[0039] Specifically, in the first stage, the encoder of the pre-trained DeepLabv3 sub-model is frozen, and the raster output network is trained only through the cross-entropy loss function. A large learning rate of 1e-2 and a large batch size of 16 are used to quickly update the head learning task-related features.

[0040] Specifically, in the second stage, all network layers are unfrozen, and the entire visual recognition model is fine-tuned using a small learning rate of 1e-5 and a large batch size of 4. This allows for refined adjustment of all features to achieve optimal performance. Both the first and second stages use the Adam optimizer and employ a cosine annealing strategy for dynamic adjustment.

[0041] Understandably, the pre-trained DeepLabv3 sub-model, trained on the ImageNet large-scale general vision dataset or related industrial vision datasets, has learned general low- and mid-level visual features through massive amounts of data. These features have broad transfer applicability and serve as a general visual foundation for various vision tasks. In the first stage, before the task-specific head has learned a stable mapping, the gradient signal during backpropagation is quite noisy. If the DeepLabv3 sub-model is not frozen at this time and its encoder is not updated, the learned general features will be destroyed, leading to the loss of transfer knowledge. Freezing the encoder, however, limits the training scope to the raster output network of the head, utilizing the encoder's fixed feature extractor capability to allow the raster output network to quickly focus on learning task-related feature mappings. At this point, a large learning rate of 1e-2 and a large batch size of 16 are used. The large learning rate provides sufficient parameter update power for the randomly initialized head network, accelerating its mapping learning from general features to task outputs; the large batch size makes gradient estimation more stable, avoiding gradient oscillations caused by small batch sizes, and ensuring that the head network quickly converges to a usable state.

[0042] While the general features learned by the pre-trained DeepLabv3 model have transferability, the raw fabric images from the loom used as input for this task have strong domain specificity, preventing the features from accurately capturing these task-specific key visual cues. Therefore, after the first-stage head network has learned a stable task mapping, the pre-trained DeepLabv3 model needs to be unfrozen for the second stage. This allows it to co-optimize with the raster output network, adapting the general features to the specific domain of fabric processing feature recognition and extracting more targeted high-level features, such as strongly correlated features like the hairiness index level and yarn evenness level. The second stage uses a small learning rate of 1e-5 to fine-tune the pre-trained DeepLabv3 model, allowing its features to gradually transition from general to an optimal state that blends general and domain-specific features. The second stage uses a smaller batch size of 4. This is because the number of parameters increases dramatically during overall fine-tuning, requiring a smaller batch size due to memory limitations. Furthermore, a smaller batch size introduces noise regularization, preventing overfitting on the customized dataset.

[0043] Specifically, high-quality fiber fabric images are obtained by macro photography of the fabric using an industrial line-scan camera and a stable light source. The density values ​​at 10cm intervals of the fiber fabric images are labeled as low, medium, and high warp and weft density levels. Yarn evenness levels are labeled as excellent, good, and poor based on yarn diameter CV values. Hairiness index levels (1-5) are assigned based on the number of fibers larger than 0.5mm. Texture types such as plain weave, twill weave, satin weave, and others are also labeled to form the customized dataset. It is understood that the above levels and classification standards are also the annotation information of the processing quality level annotation raster map generated by the visual recognition model. After one-hot encoding of the annotation information, it is input into the convolutional neural network model. It is understood that the above levels and classification standards conform to the mechanisms of textile materials science and industrial quality inspection specifications, thereby achieving accuracy and interpretability in abrasion resistance prediction and avoiding the difficulty of directly corresponding pure numerical annotations with quality inspection levels.

[0044] like Figure 4 As shown, the fabric parameter feature extraction module further includes: The high-dimensional feature extraction submodule is used to standardize the fiber fabric processing parameters and then generate high-dimensional fiber fabric processing parameter features through the first fully connected layer. The abstract feature extraction submodule is used to pass the high-dimensional fiber fabric processing parameter features through the second fully connected layer to generate abstract fiber fabric processing parameter features. The smooth feature output submodule is used to generate the fiber fabric processing parameter features by passing the abstract fiber fabric processing parameter features through a linear transformation fully connected layer. The fully connected network includes a first fully connected layer, a second fully connected layer, and a linear transformation fully connected layer.

[0045] Specifically, the numerical values ​​of the fiber fabric processing parameters are standardized using Z-scores, and the fabric weave pattern is encoded using one-hot encoding. These are then integrated into a vector form and input into the first fully connected layer. The output dimension of the first fully connected layer is 256, undergoing high-dimensional expansion. A ReLU activation function is used to introduce a non-linear transformation, filtering negative features and retaining positively correlated high-dimensional interaction information. A Dropout layer with a dropout rate of 0.2 is also included to prevent overfitting on small-batch customized datasets. Therefore, the first fully connected layer releases the feature representation capability of low-dimensional parameters. After high-dimensional mapping, the combination relationships between parameters can be fully expanded in the 256-dimensional space, providing sufficient interaction information for subsequent deep feature extraction.

[0046] Specifically, the output dimension of the second fully connected layer is 128, which is used for dimensionality reduction and abstraction. The ReLU activation function is used for dimensionality reduction and abstraction, and deep nonlinearity is introduced to explore higher-order correlations between parameters. A Dropout layer with a dropout rate of 0.2 is also used to prevent the model from overfitting on small-batch customized datasets. Therefore, the second fully connected layer explores the deep nonlinear correlation between process parameters and abrasion resistance. The features output by the first fully connected layer are the preliminary combinations between parameters, while the second layer, through dimensionality reduction and deep nonlinearity, filters out redundant interactions that are weakly correlated with abrasion resistance and focuses on the core correlations, such as the impact of complex combinations of multiple parameters like 30% cotton, 70% polyester, plain weave, and 200g / m² on abrasion resistance, making the features closer to the core process factors that determine abrasion resistance.

[0047] Specifically, the output dimension of the linear transformation fully connected layer is 64, aligned with the number of channels in the processing quality grade labeling raster image. This linear transformation fully connected layer is a pure convolutional linear transformation without activation functions, preserving the linear smoothness of the features and eliminating dropout layers. Therefore, the linear transformation fully connected layer achieves feature smoothing and spatial alignment. The ReLU activation functions in the first two layers lead to sparsity of features with a large number of zero elements; the pure linear transformation can transform sparse abstract features into smooth, dense features, avoiding interference from feature fluctuations in subsequent fusion. Simultaneously, unifying the dimension to 64 fully adapts to the input requirements of convolutional neural networks, allowing process parameter features and the processing quality grade labeling raster image to co-learn in the same embedding space.

[0048] Furthermore, the textile fabric feature fusion detection module includes: The attention fusion enhancement submodule is used to enhance the processing quality level annotation raster and fiber fabric processing parameter features through a multi-branch attention mechanism to generate processing feature level enhancement features and fiber process parameter enhancement features. The weighted gating fusion submodule is used to calculate the gating weight based on the processing feature level enhancement feature and the fiber process parameter enhancement feature, and to perform weighted fusion on the processing feature level enhancement feature and the fiber process parameter enhancement feature based on the gating weight to generate an initial fused feature; Abrasion resistance output submodule is used to pass the initial fusion features through the abrasion resistance output head to generate a predicted abrasion resistance of the fiber fabric. The convolutional neural network model includes a multi-branch attention mechanism, gating weights, and a wear-resistant output head.

[0049] Specifically, the gating weights can dynamically adjust the contribution of the two types of features based on differences in fabric type and local weaving quality, solving the problem that fixed fusion cannot adapt to the differences in feature importance under different scenarios. The generation process of the gating weights is as follows: the processing feature level enhancement feature and the fiber process parameter enhancement feature are spliced ​​together to generate spliced ​​features. The spliced ​​features are transformed through a lightweight dual-branch 1×1 convolution to obtain two original gating weight values. For each grid, the two original gating weight values ​​are normalized using Softmax in the channel dimension to obtain two gating weights. At this point, the sum of the two gating weights for each grid is 1. The two gating weights are used as the weight coefficients of the processing feature level enhancement feature and the fiber process parameter enhancement feature, respectively, and weighted fusion is performed to generate the initial fused features. Therefore, the above grid-by-grid gating weights can be dynamically adjusted for the weaving quality of different regions. Defective areas automatically increase the weight of visual features to accurately capture abrasion-resistant weak points, while uniformly woven areas automatically increase the weight of process features to match the intrinsic abrasion resistance of the fabric.

[0050] Specifically, the abrasion resistance output head converts the rasterized spatial fusion features into abrasion resistance prediction results adapted to industrial quality inspection needs. The abrasion resistance output head sequentially sets up global average pooling, global max pooling, two fully connected layers, and a regression output layer. Average pooling captures the overall average abrasion resistance level of the fabric, while global max pooling captures the abrasion resistance limit of the weakest area of ​​the fabric, conforming to the "barrel effect" of fabric abrasion resistance performance. The barrel effect states that local weaving defects directly determine the abrasion resistance limit of that area, thus affecting the overall performance of the fabric. Both fully connected layers are equipped with ReLU activation functions to uncover the deep nonlinear correlation between fusion features and abrasion resistance performance, filtering redundant information and focusing on core influencing factors. The regression output layer is a fully connected layer without activation, outputting continuous predicted values ​​of the abrasion resistance of the fiber fabric, i.e., Martindale abrasion resistance counts, a core industry-standard indicator adapted to industrial quality inspection.

[0051] Furthermore, the attention fusion submodule includes: The self-attention fusion unit is used to pass the processing quality level labeling raster and fiber fabric processing parameter features through a self-attention mechanism branch to generate processing feature self-enhancing features and fiber process parameter self-enhancing features. A cross-modal attention fusion unit is used to pass the processing quality level annotation raster map and fiber fabric processing parameter features through a cross-modal attention mechanism branch to generate cross-modal enhancement features of processing features and cross-modal enhancement features of fiber process parameters; The feature splicing unit is used to splice the self-reinforcing feature of the processing feature and the cross-modal enhancement feature of the processing feature to generate the level enhancement feature of the processing feature, and to splice the self-reinforcing feature of the fiber process parameter and the cross-modal enhancement feature of the fiber process parameter to generate the enhancement feature of the fiber process parameter. The multi-branch attention mechanism includes a self-attention mechanism branch and a cross-modal attention mechanism branch.

[0052] Specifically, the self-attention mechanism branch mines the internal correlations between the two types of features without introducing cross-modal information. For the processing quality grade labeled raster map, it learns the spatial correlations between different quality inspection rasters. For the fiber fabric processing parameter features, it learns the interactive correlations between different process dimensions. Ultimately, it strengthens the core features that significantly affect abrasion resistance and suppresses redundant noise and irrelevant information. Both self-attention mechanism branches use 1×1 convolution to generate Q (Query), K (Key), and V (Value), as well as a Softmax activation function. The feature dimensions of Q and K are taken as one-eighth of the number of channels in the processing quality grade labeled raster map and the fiber fabric processing parameter features. Compared to the native full self-attention mechanism, the computational cost is reduced, making it suitable for industrial real-time requirements.

[0053] Specifically, the cross-modal attention mechanism establishes a correlation mapping between visual processing features (local weaving quality) and global process parameter features (intrinsic fabric attributes), achieving bidirectional guided enhancement. Global process features guide visual features, strengthening the core areas most affecting abrasion resistance under the corresponding process. Actual visual features guide process features, correcting the deviation between theoretical process values ​​and actual production conditions, thus solving the information blind spot problem of single-modal features. The cross-modal attention mechanism includes a linear layer and Softmax activation. The fiber fabric processing parameter features are passed through a linear layer to generate a first query vector. The processing quality grade labeled raster is passed through a linear layer to generate a first key vector and a first value vector. The first query vector, first key vector, and first value vector are used to calculate the first cross-modal weight. The first cross-modal weight is multiplied element-wise with the processing quality grade labeled raster to obtain the cross-modal enhancement feature of the processing features. The processing quality grade labeled raster map is used to generate a second query vector through a linear layer. The fiber fabric processing parameter features are used to generate a second key vector and a second value vector through a linear layer. The second query vector, the second key vector, and the second value vector are used to calculate the second cross-modal weight. The second cross-modal weight is multiplied element-wise with the fiber process parameters to obtain the cross-modal enhancement features of the fiber process parameters.

[0054] Furthermore, the wear resistance cause analysis module includes: The initial SHAP value calculation submodule is used to generate initial SHAP values ​​by passing the visual recognition model, fully connected network and convolutional neural network model through the DeepSHAP algorithm; The weight abrasion resistance relationship constraint submodule is used to obtain multiple initial SHAP values ​​corresponding to multiple weights, and to determine whether the weight is proportional to the initial SHAP value. If not, the initial SHAP value is fitted by ordinal regression to generate a corrected SHAP value. If so, the initial SHAP value is used as the corrected SHAP value. Based on the comparison of multiple corrected SHAP values ​​of processing feature level and fiber fabric processing parameters, the dominant factor is determined. The processing parameters of the fiber fabric include weight.

[0055] Specifically, as the fiber weight of a fabric increases, its abrasion resistance will necessarily increase; however, as the twist increases, its abrasion resistance will initially increase and then decrease. Therefore, the modified SHAP value can force the SHAP value to change proportionally with the weight.

[0056] Furthermore, the wear resistance cause analysis module also includes: The SHAP cotton ratio interaction item calculation submodule is used to determine whether the cotton ratio exceeds the cotton ratio threshold. If so, it calculates the difference between the cotton ratio and the cotton ratio threshold, and calculates the corresponding corrected SHAP value based on the difference.

[0057] Specifically, a cotton ratio exceeding 0.8 will drastically reduce abrasion resistance. Therefore, the difference between the cotton ratio exceeding the cotton ratio threshold of 0.8 is calculated, and the difference is re-determined using the DeepSHAP algorithm to correct the SHAP value, which conforms to the threshold rule of cotton ratio.

[0058] Specifically, the process of determining the dominant factor is as follows: the customized dataset also labels the fiber fabric processing parameters and the predicted abrasion resistance of the fiber fabric with Martindale abrasion resistance, and the customized dataset serves as the background dataset for DeepSHAP. The target sample to be analyzed is input into the initialized DeepSHAP interpreter. Backpropagation is used to calculate the predicted abrasion resistance value for each dimension of the attribution feature vector to obtain the original SHAP value. This attribution feature vector is obtained by concatenating the expanded vector of the processing quality grade grid map with the fiber fabric processing parameters. The absolute values ​​of the original SHAP values ​​for the four processing features at all grid positions are averaged to obtain four initial SHAP values. The nine original SHAP values ​​of the nine-dimensional process parameters are merged according to their original physical meaning. The three original SHAP values ​​for the three dimensions of cotton, polyester, and nylon proportions are merged into the total initial SHAP value for fiber composition. The original SHAP values ​​for the four-dimensional unique thermal codes of plain weave, twill weave, satin weave, and others are merged into the total initial SHAP value, finally obtaining the initial SHAP values ​​for five global process parameters. The initial SHAP values ​​are then processed through the above steps to obtain modified SHAP values. After normalizing multiple modified SHAP values, they are compared to determine the processing feature grade and fiber fabric processing parameters corresponding to the maximum value, which are then used as the dominant factors.

[0059] Furthermore, the automated wear resistance testing system also includes: The collaborative training module is used to perform collaborative training and optimization of the fully connected network and the convolutional neural network model through a combined loss function, wherein the combined loss function includes a gated fusion sparsity term and a mean squared error loss term.

[0060] Specifically, the dataset for collaborative training optimization includes a raster map labeled with processing quality levels, fiber fabric processing parameters, and true labels for fabric abrasion resistance. The true label for fabric abrasion resistance is the standard measured Martindale abrasion resistance count. The gated fusion sparsity term of the combined loss function is the sum of the products of the two gate weights for each raster across all raster and all samples, divided by the total number of raster cells in the full sample. Therefore, the gated fusion sparsity term avoids weight egalitarianism, ensuring that the abrasion resistance influencing factors for each raster are clearly attributed to weaving processing characteristics or process parameter characteristics. This improves the accuracy of subsequent DeepSHAP causal analysis and suppresses overfitting of the model to noise, aligning with the actual logic of textile production. That is, fabric abrasion resistance shortcomings are either inherent process design problems or acquired weaving defects, and it is extremely rare for both to contribute equally. The mean squared error loss term is used to calculate the deviation between the predicted fabric abrasion resistance and the true label for fabric abrasion resistance, ensuring prediction accuracy. The gating fusion sparse term is weighted by 0.1 and summed with the mean squared error loss term to calculate the combined loss function, avoiding excessive suppression of the core mean squared error loss term. The 0.1 weighting coefficient can be dynamically adjusted according to the validation set results.

[0061] Specifically, to verify the accuracy and real-time performance of the visual recognition model in extracting fabric processing features, a raw DeepLabv3 network without a raster output network and a two-step fine-tuning strategy was used as a control group. The pixel-level annotation accuracy (mIoU%) of the processing feature raster images of the control group and the visual recognition model were 90.1 and 94.2, respectively, and the single-image inference time (ms) were 26 and 28, respectively. It can be seen that the visual recognition model achieves accurate identification of the textile processing effect of the fabric while ensuring real-time performance.

[0062] Specifically, to verify the feature encoding effect of the three-level fully connected network on heterogeneous fabric process parameters and its adaptability to multimodal feature fusion, control group 2 used a two-layer fully connected layer for encoding. The feature cosine similarity between the process parameter features and visual features of the fully connected network and control group 2 was 0.758 and 0.894, respectively. It can be seen that the fully connected network can effectively uncover the intrinsic correlation between heterogeneous process parameters and generate high-dimensional features that are highly semantically matched with visual features.

[0063] Specifically, to verify the prediction accuracy and generalization ability of the convolutional neural network model with multi-branch attention and gated weighted fusion for fabric abrasion resistance, control group 3 used a simple channel splicing CNN model of visual features and process parameters (without multi-branch attention or gated weighting), control group 4 used a convolutional neural network model without cross-modal attention branches, retaining only self-attention branches, and control group 5 used a convolutional neural network model with no gated weighted fusion and only simple splicing. The MSE of control groups 3, 4, and 5 and the convolutional neural network model are as follows: , , , The accuracy rates (%) of the wear resistance grades were 88.5%, 92.1%, 93.3%, and 97.2%, respectively. This demonstrates that the architecture of the convolutional neural network model can accurately learn the synergistic influence of visual features and process parameters on wear resistance performance, achieving an order-of-magnitude improvement in prediction accuracy compared to existing technologies.

[0064] In this embodiment, a pre-trained DeepLabv3 sub-model is used as the backbone for feature extraction. Leveraging its multi-scale modeling capability of Spatial Pyramid Pooling (ASPP), it can accurately capture global structural features that determine the abrasion resistance of the fabric, such as warp and weft density and fabric structure, while also fully preserving microscopic details that determine the wear initiation point, such as yarn evenness, surface hairiness index, and weaving defects. Through the global average pooling layer and convolutional output layer of the raster output network, a processing quality grade annotation raster map that is perfectly aligned with the input fabric image space is directly generated, ensuring the real-time performance of visual detection. Through a two-step fine-tuning training strategy of freezing and thawing, the general texture feature extraction capability of the pre-trained model is preserved, and the fabric-specific features are accurately adapted through a customized fabric dataset. This solves the problems of limited labeled data and easy overfitting of the model in the textile industry, and achieves accurate identification of the textile processing effect of the fabric. To address the issue of mixed fabric process parameters, a three-level progressive fully connected coding structure—comprising high-dimensional feature extraction, abstract feature mining, and smooth feature output—is employed. This structure automatically learns the inherent synergistic relationships between different process parameters such as fiber composition, yarn twist, and basis weight. It generates fiber fabric processing parameter features that are fully matched and semantically compatible with the dimensions of visual processing features. This avoids the inaccurate fusion of global process parameters and local visual features found in traditional solutions. The approach combines the textile processing effect of the fabric with its process parameters to collaboratively predict the fabric's abrasion resistance and analyzes the dominant factors in abrasion resistance assessment. This enables highly accurate automated detection of fabric abrasion resistance and its causes.

[0065] Those skilled in the art will recognize that the modules and algorithm steps of the examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0066] The technical solution of the present invention has been described above with reference to the preferred embodiments shown in the accompanying drawings. However, it will be readily understood by those skilled in the art that the scope of protection of the present invention is obviously not limited to these specific embodiments. Without departing from the principles of the present invention, those skilled in the art can make equivalent changes or substitutions to the relevant technical features, and the technical solutions after these changes or substitutions will all fall within the scope of protection of the present invention.

[0067] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. An automated abrasion resistance testing system for fabrics, characterized in that, include: The fabric image processing feature recognition module is used to generate a raster map with processing quality grade annotations from the image of the fiber fabric after it comes off the loom through a visual recognition model. The fabric parameter feature extraction module is used to extract the features of fiber fabric processing parameters through a fully connected network to generate fiber fabric processing parameter features. The textile fabric feature fusion detection module is used to generate a predicted abrasion resistance of the fiber fabric by passing the processing quality grade labeled raster map and fiber fabric processing parameter features through a convolutional neural network model. The abrasion resistance cause analysis module is used to use the DeepSHAP algorithm to determine the dominant factors in predicting the abrasion resistance of the fiber fabric by combining the visual recognition model, fully connected network, and convolutional neural network model.

2. The automated abrasion resistance testing system for fabrics according to claim 1, characterized in that, The fabric image processing feature recognition module includes: The feature recognition submodule is used to extract features from the fiber fabric image using a pre-trained DeepLabv3 sub-model to generate a fabric processing feature map. The classification output submodule is used to generate the processing quality grade labeling raster map by passing the fabric processing feature map through a raster output network. The visual recognition model includes a pre-trained DeepLabv3 sub-model and a raster output network.

3. The automated abrasion resistance testing system for fabrics according to claim 2, characterized in that, The classification output submodule includes: A global average pooling unit is used to map the fabric processing feature map to a raster image size through a global average pooling layer to generate a raster image size fabric processing feature map. The classification output unit is used to classify the fabric processing feature map of the grid size to the processing feature level through the convolution output layer to generate the processing quality level labeled grid map; The grid output network includes a global average pooling layer and a convolutional output layer.

4. The automated abrasion resistance testing system for fabrics according to claim 2, characterized in that, The fabric image processing feature recognition module also includes: The frozen training submodule is used to process the input data of the customized dataset through the pre-trained DeepLabv3 sub-model and the raster output network to generate a raster map with processing quality level annotation. The cross-entropy loss function is calculated based on the raster map with annotation of the customized dataset and the raster map with processing quality level annotation. After freezing the pre-trained DeepLabv3 sub-model, the raster output network is fine-tuned through the cross-entropy loss function. The Unfreeze Training submodule is used to unfreeze the pre-trained DeepLabv3 sub-model and then fine-tune the pre-trained DeepLabv3 sub-model and the raster output network using the cross-entropy loss function.

5. The automated abrasion resistance testing system for fabrics according to claim 1, characterized in that, The fabric parameter feature extraction module includes: The high-dimensional feature extraction submodule is used to standardize the fiber fabric processing parameters and then generate high-dimensional fiber fabric processing parameter features through the first fully connected layer. The abstract feature extraction submodule is used to pass the high-dimensional fiber fabric processing parameter features through the second fully connected layer to generate abstract fiber fabric processing parameter features. The smooth feature output submodule is used to generate the fiber fabric processing parameter features by passing the abstract fiber fabric processing parameter features through a linear transformation fully connected layer. The fully connected network includes a first fully connected layer, a second fully connected layer, and a linear transformation fully connected layer.

6. The automated abrasion resistance testing system for fabrics according to claim 1, characterized in that, The textile fabric feature fusion detection module includes: The attention fusion enhancement submodule is used to enhance the processing quality level annotation raster and fiber fabric processing parameter features through a multi-branch attention mechanism to generate processing feature level enhancement features and fiber process parameter enhancement features. The weighted gating fusion submodule is used to calculate the gating weight based on the processing feature level enhancement feature and the fiber process parameter enhancement feature, and to perform weighted fusion on the processing feature level enhancement feature and the fiber process parameter enhancement feature based on the gating weight to generate an initial fused feature; Abrasion resistance output submodule is used to pass the initial fusion features through the abrasion resistance output head to generate a predicted abrasion resistance of the fiber fabric. The convolutional neural network model includes a multi-branch attention mechanism, gating weights, and a wear-resistant output head.

7. The automated abrasion resistance testing system for fabrics according to claim 6, characterized in that, The attention fusion enhancement submodule includes: The self-attention fusion unit is used to pass the processing quality level labeling raster and fiber fabric processing parameter features through a self-attention mechanism branch to generate processing feature self-enhancing features and fiber process parameter self-enhancing features. A cross-modal attention fusion unit is used to pass the processing quality level annotation raster map and fiber fabric processing parameter features through a cross-modal attention mechanism branch to generate cross-modal enhancement features of processing features and cross-modal enhancement features of fiber process parameters; The feature splicing unit is used to splice the self-reinforcing feature of the processing feature and the cross-modal enhancement feature of the processing feature to generate the level enhancement feature of the processing feature, and to splice the self-reinforcing feature of the fiber process parameter and the cross-modal enhancement feature of the fiber process parameter to generate the enhancement feature of the fiber process parameter. The multi-branch attention mechanism includes a self-attention mechanism branch and a cross-modal attention mechanism branch.

8. The automated abrasion resistance testing system for fabrics according to claim 1, characterized in that, The wear resistance cause analysis module includes: The initial SHAP value calculation submodule is used to generate initial SHAP values ​​by passing the visual recognition model, fully connected network and convolutional neural network model through the DeepSHAP algorithm; The weight abrasion resistance relationship constraint submodule is used to obtain multiple initial SHAP values ​​corresponding to multiple weights, and to determine whether the weight is proportional to the initial SHAP value. If not, the initial SHAP value is fitted by ordinal regression to generate a corrected SHAP value. If so, the initial SHAP value is used as the corrected SHAP value. Based on the comparison of multiple corrected SHAP values ​​of processing feature level and fiber fabric processing parameters, the dominant factor is determined. The processing parameters of the fiber fabric include weight.

9. The automated abrasion resistance testing system for fabrics according to claim 8, characterized in that, The wear resistance cause analysis module also includes: The SHAP cotton ratio interaction item calculation submodule is used to determine whether the cotton ratio exceeds the cotton ratio threshold. If so, it calculates the difference between the cotton ratio and the cotton ratio threshold, and calculates the corresponding corrected SHAP value based on the difference.

10. The automated abrasion resistance testing system for fabrics according to any one of claims 1 to 9, characterized in that, Also includes: The collaborative training module is used to perform collaborative training and optimization of the fully connected network and the convolutional neural network model through a combined loss function, wherein the combined loss function includes a gated fusion sparsity term and a mean squared error loss term.

Images