Fabric weave structure optimization method and system based on deep learning

Abstract

The invention discloses a fabric weave structure optimization method and system based on deep learning. The method comprises the steps of S00, data acquisition; s10, data preprocessing; s20, constructing and training a prediction model; s30, predicting the hand feeling of the fabric and optimizing the structure of the fabric; s40, production and quality control: measuring physical characteristics of the fabric in detail, collecting sensory cognition data, performing standardization processing on the data, training a prediction model by adopting a deep learning algorithm, performing verification and optimization through an independent data set, and in a design stage, performing production and quality control on the fabric. The model is used for simulating the influence of different fibers, yarns and weaving methods on the hand feeling, the fabric structure is adjusted through iterative optimization so as to achieve the optimal hand feeling index, the test result of an actual production sample can be fed back to the model so as to continuously improve the prediction precision, and the technical scheme can improve the design efficiency, reduce the cost, reduce the resource waste and improve the product quality. And market requirements and challenges of new materials are met, and sustainable textile design and production are supported.

Description

Technical Field

[0001] This invention relates to the field of textile manufacturing technology, and in particular to a method and system for optimizing fabric structure based on deep learning. Background Technology

[0002] In the field of textile production and design, the tactile quality of fabrics is a crucial attribute, directly impacting consumer satisfaction and perceived quality. Tactile quality, sometimes referred to as "feel," refers to the sensation produced when touching or holding a textile. The tactile quality of a fabric may include its softness, smoothness, warmth, weight, elasticity, and other sensory characteristics.

[0003] In the textile industry, the design of fabric structure to optimize the fabric feel is a core issue. In traditional solutions, the design of fabric structure relies on the designer's experience and intuition, as well as optimization through actual trial production and testing. That is, the designer needs to consider factors such as fiber type, yarn thickness, weaving method and finishing treatment, and predict how they will work together to affect the feel of the final product.

[0004] As can be seen from the above, traditional methods for designing and optimizing fabric feel mainly rely on the designer's experience and subjective personal preferences, making it difficult to predict the general public's experience. Moreover, the prediction process requires multiple trials to evaluate and adjust the fabric structure, which is time-consuming and costly. In recent years, with the development of technology, the advancement of machine learning technology has made it possible to develop a faster, more reliable, and more cost-effective method for designing and optimizing fabric feel. This method can extract patterns and relationships from large amounts of data to predict and optimize the impact of fabric structure on tactile feel. However, current methods are mostly aimed at the identification, detection, and quality assessment (rating) of fabric defects or flaws. Therefore, a new method is needed that can improve the accuracy of fabric feel prediction (making it more suitable for the majority of people), increase design efficiency, reduce resource consumption, and ensure product feel consistency and high quality. Summary of the Invention

[0005] In order to overcome the shortcomings of the prior art, the technical problem to be solved by the present invention is to propose a method and system for optimizing fabric structure based on deep learning, so as to improve the accuracy of fabric touch prediction, help improve design efficiency, reduce resource consumption, and ensure product hand feel consistency and high quality.

[0006] To achieve this objective, the present invention adopts the following technical solution: This invention provides a deep learning-based method for optimizing fabric structure, comprising the following steps: S00: Data acquisition, selecting fabric samples, measuring the physical characteristics of the fabric samples and collecting sensory perception data, establishing the physical characteristic dataset and sensory dataset corresponding to the fabric, and objectifying the sensory perception data. S10: Data preprocessing, preprocessing the physical property dataset and sensory dataset to form a structured dataset; S20: Prediction model construction and training, dividing the structured dataset into training set, validation set and test set, and further training a deep learning tactile prediction model to establish a mapping relationship from the physical characteristics and design parameters of the fabric to the tactile sensory perception score. S30: Fabric hand feel prediction and fabric structure optimization. The design parameters of the target fabric are input into the trained deep learning model to obtain the predicted tactile sensory perception score. The design parameters are iteratively optimized based on the predicted tactile sensory perception score to generate an optimized fabric structure design scheme. S40: Production and quality control: Produce samples according to the optimized fabric structure design, collect physical characteristic data and sensory perception evaluation data of new samples, and feed the new data back to the structured dataset for updating and training the deep learning model.

[0007] The following is a further explanation of each of the above steps: (1) During the data acquisition process, in step S00, the physical characteristic data includes one or more of thickness, weight, tensile properties, bending stiffness and surface friction characteristics, and the sensory perception data includes one or more of softness, smoothness and warmth.

[0008] In step S00, the objectification processing steps for the sensory cognitive data are as follows: S01: Select evaluators for sensory perception data, conduct sensory sensitivity tests on evaluators before data collection, and then provide uniform training to evaluators. Use "standard fabric samples" to calibrate evaluators' understanding of the concepts of softness, smoothness, and warmth, and ensure that all evaluators score under the same "evaluation system". S02: Organize trained evaluators to conduct blind tests on fabric samples under the same standard evaluation environment and evaluation process to obtain preliminary standardized sensory perception evaluation data. S03: Use group ratings for evaluators, take the mean or median, and conduct a consistency assessment.

[0009] In step S02, while evaluating the feel of the fabric sample through sensory testing by evaluators, detailed physical tests (softness, smoothness, and warmth) are also conducted using the Kawabata evaluation system. The test results and evaluator test results are compared with the prediction results of the machine learning model to analyze the accuracy of the predictions.

[0010] (2) In the data preprocessing process, in step S10, the preprocessing steps of the physical characteristic dataset and the sensory dataset include encoding, cleaning and normalization processes to make the structured dataset suitable for the input of the machine learning model.

[0011] (3) In the process of predictive model construction and training, in step S20, the deep learning model is a long short-term memory network (LSTM). The deep learning model contains a recurrent neural network with at least one LSTM layer to process the sequence data in the structured dataset. Dropout layer and / or L1 and L2 regularization terms are introduced into the deep learning model to prevent overfitting. At the same time, the Adam optimizer and mean squared error loss function are used to train the deep learning model in combination with the early stopping mechanism.

[0012] (4) In the process of fabric hand feel prediction and fabric structure optimization, in step S30, the steps of using the deep learning model to predict fabric hand feel and optimize fabric structure are as follows: S31: Use a trained deep learning model to simulate the impact of different combinations of design parameters on sensory perception scores. The design parameters include one or more of fiber type, yarn specification, warp and weft density, and fabric structure. S32: Using the expected tactile target as the optimization objective, the design parameters are iteratively adjusted multiple times until the prediction output of the deep learning model meets the design requirements.

[0013] (5) In the production and quality control process, in step S40, the quality of the fabric is monitored during the production process, and the physical feature dataset and sensory dataset of the product are collected and compared with the prediction results of the deep learning model. If the difference exceeds a predetermined threshold, the update training process of the deep learning model is triggered to achieve continuous optimization of the model.

[0014] A deep learning-based fabric structure optimization system, used to implement the deep learning-based fabric structure optimization method described above, includes the following modules: The data acquisition module is used to acquire physical feature datasets and sensory datasets of fabric samples. The data acquisition module includes standard textile testing instruments for automatically measuring the physical characteristics of the fabric, and a human-computer interaction interface for structured recording of sensory evaluation data. The data processing and model training module is used to preprocess the data and perform deep learning model training tasks. The tactile prediction and design optimization module is used to load the trained model and perform tactile prediction and parameter optimization for new fabric designs. The production monitoring and feedback module is used to monitor production quality and feed back new data obtained during production and testing to the data processing and model training module.

[0015] The beneficial effects of this invention are as follows: In summary, this project involves detailed measurement of the physical characteristics of fabrics and collection of sensory perception data. This data is then standardized, and a predictive model is trained using deep learning algorithms based on selected key features. The model is validated and optimized using independent datasets. During the design phase, the model simulates the impact of different fibers, yarns, and weaving methods on the feel of the fabric. The fabric structure is iteratively optimized to achieve the best feel indicators. Test results from actual production samples are fed back into the model to continuously improve prediction accuracy. This approach enhances design efficiency, reduces costs, minimizes resource waste, and addresses the challenges of market demands and new materials, supporting sustainable textile design and production. The following section further elaborates on the beneficial effects in conjunction with the technical solution: (1) A dataset that unifies physical measurement data with human sensory evaluation was constructed to address the mapping problem between "subjective perception" and "objective data". The complex and nonlinear mapping relationship between the two was learned through a deep learning model, and the quantitative prediction from measurable physical quantities to subjective feel was realized. (2) This case constructs a complete closed loop of “design-prediction-verification-optimization”, namely, forming a model to guide the design, making samples for verification, feeding back the verification results to the model for retraining, and continuously optimizing the model; in terms of design efficiency, it can reduce the number of trials and errors and shorten the development cycle; by using a deep learning model to achieve quantitative prediction of tactile attributes, the number of sample productions can be reduced, saving raw materials and production costs. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating a deep learning-based fabric structure optimization method provided in a specific embodiment of the present invention. Detailed Implementation The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0017] To provide a new method for improving the accuracy of fabric feel prediction (more suitable for most people), increasing design efficiency, reducing resource consumption, and ensuring consistent product feel and high quality to assist in actual production, this invention provides a deep learning-based fabric structure optimization method and system.

[0018] Example 1: A deep learning-based method for optimizing fabric structure, comprising the following steps: S00: Data Acquisition. This involves selecting a series of representative fabric samples covering different fiber compositions, weaving methods, and post-processing techniques. The physical characteristics of the fabric samples are measured, and sensory perception data is collected. A physical characteristic dataset and a sensory dataset are established for each fabric. Under standardized environmental conditions, professional textile testing equipment is used to perform detailed measurements of the physical characteristics of the samples. The physical characteristic data includes one or more of thickness, weight, tensile properties, flexural stiffness, and surface friction characteristics. The sensory perception data includes one or more of softness, smoothness, and warmth. Furthermore, regarding sensory perception data, it should be noted that the fabric's tactile feel... Essentially, this is a psychophysical phenomenon, namely the relationship between physical stimuli (fabric properties) and psychological perception (feel). This perception is inherently subjective, and the traditional industry dilemma lies in its inability to overcome this subjectivity, leading to instability in the design process. In this case, by quantifying and utilizing this collective subjectivity, its negative impact is minimized. That is, a system and method are constructed that can learn stable patterns from noisy subjective data to predict the "average preference" or "consensus feel" of the target consumer group. Therefore, in order to reduce the impact of subjectivity on the evaluation results and models, the sensory perception data needs to be objectively processed. Specifically, the steps for objectively processing the sensory perception data are as follows: S01: Select evaluators for sensory perception data. Before collecting data, conduct sensory sensitivity tests on the evaluators. This step is to screen the evaluators. The evaluators need to fit the target user profile (such as age and gender) and pass sensory sensitivity tests (such as tactile threshold tests) so that the evaluators have relatively consistent cognition and judgment when perceiving the same object. Then, the evaluators are uniformly trained, that is, using "standard fabric samples" to calibrate the evaluators' understanding of the concepts of softness, smoothness and warmth, ensuring that all evaluators score under the same "evaluation system". S02: Organize trained evaluators to conduct blind tests on fabric samples under the same standard evaluation environment and process to obtain preliminary standardized sensory perception evaluation data. For example, the tests can be conducted in a temperature and humidity controlled laboratory to avoid the influence of ambient temperature and humidity on the tactile sensation, and standardized lighting can be used to eliminate visual interference (i.e., blind testing). At the same time, the presentation order of the fabric samples should be randomized to avoid the order effect. In terms of the evaluation process, a structured scoring table should be designed, using clear scales (such as a 1-10 numerical scale or a semantic difference scale), and uniform operating gestures (such as the speed and force of touching and stroking with fingertips) should be specified. Preferably, in step S02, while evaluating the feel of the fabric sample through sensory testing by the evaluator, a detailed physical test (softness, smoothness, and warmth) is also conducted using the Kawabata evaluation system, and the test results and the evaluator's test results are compared with the prediction results of the machine learning model to analyze the accuracy of the prediction. The above steps all involve "standardization" of sensory data before the assessment, starting from the source of the evaluation. Next, it is necessary to mitigate individual noise from the perspective of data processing. The ratings of a single evaluator are subjective and unstable, but the average rating of a sufficiently large group of evaluators can approximate the "consensus perception" of the target population. This average can serve as a stable target value for model training, i.e., S03: Take group ratings for evaluators, take the average or median, and conduct a consistency assessment. When checking the consistency of the assessment, use statistical methods (such as the Cronbach's alpha coefficient) to calculate the consistency among evaluators. If the consistency is too low, it indicates that the assessment standards are not uniform or the rating table design is problematic. The data needs to be re-collected or cleaned, or the data of "abnormal evaluators" who are significantly inconsistent with the group rating pattern over a long period of time can be identified and removed.

[0019] By performing the above-mentioned objective processing on sensory cognitive data, subjectivity becomes a systemic feature that needs to be "managed." Through this processing, scattered and noisy subjective evaluations are aggregated and refined into stable target values ​​that can be used for machine learning and represent group preferences.

[0020] After data acquisition, step S10 can be performed: data preprocessing, which involves preprocessing the physical characteristic dataset and sensory dataset to form a structured dataset. The preprocessing steps for the physical characteristic dataset and sensory dataset include encoding, cleaning, and normalization to make the structured dataset suitable as input for machine learning models. Specifically: S11: Encoding. Since machine learning models typically require numerical input, non-numerical data needs to be encoded. This is mainly divided into encoding of categorical data and encoding of ordinal data. For unordered categorical data, such as fabric types, one-hot encoding is usually used, creating a binary column for each category, with existing categories marked as 1 and others as 0. For ordered ordinal data, such as feel ratings (very rough, rough, medium, smooth, very smooth), ordinal encoding can be used to convert this ordered relationship into an integer sequence. S12: Data cleaning is the process of ensuring data accuracy and integrity. It mainly includes handling missing values: identifying missing values ​​for each field and filling them with the average value; identifying and handling outliers: identifying outliers in the data through statistical analysis or visualization and deciding whether to correct, delete, or retain these values; and standardizing data formats: ensuring that all data are consistent in format and units. S13: Normalization processing. Normalization transforms numerical data to a common scale, typically between 0 and 1. Min-max normalization is used to process the numerical data: Min-max normalization: also known as scaling to the interval [0,1], is achieved by subtracting the minimum value from the original data (X) and then dividing by the maximum value (...). ) and minimum value ( This is achieved through the difference between ), that is:

[0021] After completing the above preprocessing, the data will be in a clean, uniformly formatted state that is suitable for use by machine learning models.

[0022] After the data is processed, step S20 can be performed: prediction model construction and training. The structured dataset is divided into a training set, a validation set, and a test set, and a deep learning tactile prediction model is further trained to establish a mapping relationship from the physical characteristics and design parameters of the fabric to the tactile sensory perception score. Specifically, the deep learning model is a Long Short-Term Memory (LSTM) network, which contains a recurrent neural network with at least one LSTM layer to process the sequential data in the structured dataset. Dropout layers and / or L1 and L2 regularization terms are introduced into the deep learning model to prevent... To prevent overfitting, the deep learning model is trained using the Adam optimizer and mean squared error loss function, combined with an early stopping mechanism. Specifically, each LSTM layer is configured with 128 units, and Dropout and recurrent Dropout are added to reduce the risk of overfitting. The Dropout ratio is set to 0.3 to further reduce overfitting. Next, a fully connected layer is added, using the ReLU activation function and applying L1 and L2 regularization. L1 / L2 regularization is an additional regularization term added to the loss function to limit the size of the network weights; this coefficient is set to 0.0005. In the output layer, a linear activation function is used. When compiling the model, the Adam optimizer is selected, and the learning rate is set to 0.001. The loss function is set to mean squared error, a common loss function for regression problems. The model is then trained with a batch size of 64 and 50 training epochs. Validation data is used to monitor model performance. Furthermore, an early stopping mechanism is employed to prevent overfitting during training; training stops if the performance on the validation set no longer improves or begins to deteriorate over several consecutive training epochs.

[0023] S30: Fabric hand feel prediction and fabric structure optimization. The design parameters of the target fabric are input into the trained deep learning model to obtain the predicted tactile sensory perception score. Based on the predicted tactile sensory perception score, the design parameters are iteratively optimized to generate an optimized fabric structure design scheme. Specifically, the steps of using the deep learning model for fabric hand feel prediction and fabric structure optimization are as follows: S31: Use a trained deep learning model to simulate the impact of different combinations of design parameters on sensory perception scores. The design parameters include one or more of fiber type, yarn specification, warp and weft density, and fabric structure. S32: Using the expected tactile target as the optimization objective, the design parameters are iteratively adjusted multiple times until the prediction output of the deep learning model meets the design requirements.

[0024] In summary, in step S30, the test structure of the Kawabata evaluation system and the human perception test results of the evaluators are further collected. These results are then compared and analyzed with the output of the deep learning model, using statistical indicators such as accuracy, recall, and F1 score to evaluate model performance. If the prediction results do not match the actual data, the reasons for the discrepancies need to be identified, and the model adjusted accordingly. This is usually an iterative process that needs to be repeated until the model predictions reach acceptable accuracy. Finally, all findings, actions taken, and final results are recorded to form a detailed report for future reference. Based on the differences between the sample test data and the prediction results, this new data is input into the model for retraining. This helps correct the model's prediction bias and improve its accuracy and reliability in future predictions. Over time and with the continuous accumulation of new data, the model will continue to optimize, and its learning ability will be continuously enhanced.

[0025] S40: Production and Quality Control. Based on the optimized fabric structure design, sample production is carried out. Physical characteristic data and sensory perception evaluation data of the new samples are collected, and the new data is fed back to the structured dataset for updating and training the deep learning model. Specifically, the quality of the fabric is monitored during production to ensure that the produced fabric matches the prediction results of the design model. The physical characteristic dataset and sensory dataset of the product are collected and compared with the prediction results of the deep learning model. If the difference exceeds a predetermined threshold, the update and training process of the deep learning model is triggered to achieve continuous model optimization. That is, the model is updated using new data obtained during production to cope with factors such as changes in raw materials and consumer demand, maintaining the timeliness and accuracy of the model.

[0026] Example 2: A deep learning-based fabric structure optimization system, used to implement the deep learning-based fabric structure optimization method described above, includes the following modules: The data acquisition module is used to acquire physical feature datasets and sensory datasets of fabric samples. Preferably, the data acquisition module includes a standard textile testing instrument for automatically measuring the physical features of the fabric, and a human-computer interaction interface for structured recording of sensory evaluation data. The data processing and model training module is used to preprocess the data and perform the training task of the deep learning model. The tactile prediction and design optimization module is used to load the trained model and perform tactile prediction and parameter optimization for new fabric designs. The production monitoring and feedback module is used to monitor production quality and feed back new data obtained in production and testing to the data processing and model training module.

[0027] This invention has been described through preferred embodiments. Those skilled in the art will understand that various changes or equivalent substitutions can be made to these features and embodiments without departing from the spirit and scope of the invention. This invention is not limited to the specific embodiments disclosed herein; other embodiments falling within the scope of the claims are also within the protection scope of this invention.

Claims

1. A deep learning-based fabric structure optimization method, characterized in that, The method comprises the following steps: S00: data collection, selecting fabric samples, measuring physical characteristic data of the fabric samples and collecting sensory perception data, establishing a fabric corresponding physical characteristic data set and sensory data set, and objectively processing the sensory perception data; S10: data preprocessing, preprocessing the physical characteristic data set and the sensory data set to form a structured data set; S20: prediction model construction and training, dividing the structured data set into a training set, a validation set, and a test set, and further training a deep learning tactile sensation prediction model to establish a mapping relationship from the fabric physical characteristics and design parameters to the tactile sensation sensory perception score; S30: fabric hand feeling prediction and fabric structure optimization, inputting the design parameters of the target fabric into the trained deep learning model to obtain a predicted tactile sensation sensory perception score, and iteratively optimizing the design parameters based on the predicted tactile sensation sensory perception score to generate an optimized fabric structure design scheme; S40: production and quality control, producing samples according to the optimized fabric structure design scheme, collecting physical characteristic data and sensory perception evaluation data of new samples, and feeding the new data back to the structured data set for updating and training the deep learning model.

2. The fabric structure optimization method based on deep learning according to claim 1, wherein, In step S00, the physical characteristic data includes one or more of thickness, weight, tensile properties, bending stiffness, and surface friction characteristics, and the sensory perception data includes one or more of softness, smoothness, and warmth.

3. The fabric structure optimization method based on deep learning according to claim 2, characterized in that, In step S00, the objective processing step of the sensory perception data is: S01: selecting evaluators for sensory perception data, conducting sensory sensitivity tests on the evaluators before collecting data, then uniformly training the evaluators, using "standard fabric samples" to calibrate the evaluators' understanding of softness, smoothness, and warmth concepts, and ensuring that all evaluators score under the same set of "evaluation systems"; S02: organizing trained evaluators to conduct blind tests on fabric samples under the same standard evaluation environment and evaluation process, and preliminarily obtaining standardized sensory perception evaluation data; S03: taking group scores of the evaluators, taking the average or median, and performing consistency evaluation.

4. The fabric structure optimization method based on deep learning according to claim 3, characterized in that, In step S02, while evaluating the hand feeling of the fabric samples through the evaluators' sensory tests, detailed physical tests (softness, smoothness, and warmth) are also conducted using the Kawabata evaluation system, and the test results and the evaluators' test results are compared with the prediction results of the machine learning model to analyze the accuracy of the prediction.

5. The fabric structure optimization method based on deep learning according to claim 3, characterized in that, In step S10, the preprocessing steps of the physical characteristic data set and the sensory data set include encoding, cleaning, and normalization processing to make the structured data set suitable for input of the machine learning model.

6. The fabric structure optimization method based on deep learning according to claim 5, wherein, In step S20, the deep learning model is a long short-term memory network (LSTM), which comprises a recurrent neural network with at least one LSTM layer to process sequence data in the structured data set, and incorporates Dropout layers and / or L1, L2 regularization terms in the deep learning model to prevent overfitting, while using an Adam optimizer and a mean square error loss function in combination with an early stopping mechanism to train the deep learning model.

7. The fabric structure optimization method based on deep learning according to claim 6, characterized in that, In step S30, the steps of fabric hand prediction and fabric structure optimization using the deep learning model are: S31: using the trained deep learning model to simulate the impact of different combinations of design parameters, including one or more of fiber type, yarn specification, warp and weft density, and fabric structure, on sensory perception scores; S32: taking the expected target tactile sensation as the optimization objective, iteratively adjusting the design parameters until the predicted output of the deep learning model meets the design requirements.

8. The fabric structure optimization method based on deep learning according to claim 1, wherein, In step S40, the quality of the fabric is monitored during production, and the physical characteristic data set and sensory data set of the product are collected and compared with the prediction results of the deep learning model. If the difference exceeds a predetermined threshold, the update training process of the deep learning model is triggered to achieve continuous optimization of the model.

9. A deep learning-based fabric weave structure optimization system for implementing the deep learning-based fabric weave structure optimization method according to any one of claims 1 to 8, characterized in that, The system comprises the following modules: a data acquisition module for obtaining the physical characteristic data set and sensory data set of the fabric sample; a data processing and model training module for pre-processing the data and performing the training task of the deep learning model; a tactile sensation prediction and design optimization module for loading the trained model and performing tactile sensation prediction and parameter optimization for new fabric designs; a production monitoring and feedback module for monitoring production quality and feeding new data obtained during production and testing back to the data processing and model training module.

10. The deep learning-based fabric structure optimization system of claim 9, wherein, The data acquisition module comprises standard textile testing instruments for automatically measuring fabric physical characteristics, and a human-computer interaction interface for structurally recording sensory evaluation data.

Images