Intelligent control system for production tension of polyester fiber cloth

Abstract

The invention relates to the technical field of machine learning, in particular to an intelligent control system for production tension of polyester fiber cloth. Comprising a flexible strain acquisition module, an image acquisition module, a laser displacement acquisition module, a temperature monitoring module, a multi-modal time sequence fusion data set generation module, a multi-modal data fusion and evaluation module, a workshop section coupling analysis module and an MPC control decision module. The system acquires transverse tension distribution data, longitudinal tension distribution data, fiber cloth deformation data, fiber cloth surface images and real-time temperature data of each workshop section, and performs time synchronization and preprocessing to obtain a multi-modal time sequence fusion data set; generating a comprehensive tension evaluation value and a comprehensive deformation evaluation value by using a Stacking fusion model; and in combination with a workshop section coupling coefficient matrix, a control instruction is finally output by the MPC control decision module, so that intelligent, real-time and precise control on the tension in the polyester fiber cloth production process is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of machine learning, and in particular to an intelligent control system for the production tension of polyester fiber cloth. Background Art

[0002] In a series of production processes such as spinning, post-processing, heat setting and winding, polyester fiber cloth needs to be stretched and conveyed. In this process, tension is one of the important parameters that determine product quality: if the tension is not properly controlled, the fiber is prone to breakage, linting or deformation, or quality defects such as uneven surface, wrinkling and shrinkage may occur. However, due to factors such as high production speed, multiple process links and complex environment, production tension will be affected by many factors, including the physical properties of the fiber material itself, production line speed fluctuations, mechanical friction and temperature and humidity changes, making it difficult to measure and control tension in real time.

[0003] In traditional processes, tension control usually relies on mechanical or electrical sensors in conjunction with algorithms such as PID to maintain target tension, but such methods still have many shortcomings in high-speed or high-precision scenarios: mechanical sensors are easily disturbed by factors such as vibration and temperature drift, and the feedback speed is also slow, making it difficult to meet real-time and accurate control requirements. In addition, if only a single modal data is relied upon (such as simply monitoring tension), it can often only reflect local or surface state changes, and cannot fully cover multiple factors such as temperature, deformation, and surface defects. Once ambient temperature fluctuations, mechanical wear, or fabric deformation are not detected in time, local anomalies may be magnified, which in turn affects the overall production quality and efficiency. At the same time, the sensor equipment itself may also cause measurement deviations or noise due to vibration, temperature drift, etc., further weakening the degree of mastery of the actual working conditions, making it difficult to ensure the reliability of the system in a complex and changing production environment.

[0004] Therefore, an intelligent control system for the production tension of polyester fiber cloth is proposed. Summary of the invention

[0005] The purpose of the present invention is to provide an intelligent control system for the production tension of polyester fiber cloth, which combines a flexible strain sensor array, an industrial camera, a laser displacement sensor and a temperature sensor to collect multi-source data in real time, and comprehensively evaluates the transverse tension, longitudinal tension, cloth deformation data and surface image features through time synchronization and Stacking fusion model, and generates control instructions using section coupling analysis and MPC algorithm to achieve real-time dynamic control of multiple sections. Through multimodal cross-validation and full-process data fusion, this system improves the measurement accuracy and control response speed in complex production environments, and significantly enhances the reliability and stability of the polyester fiber cloth production process.

[0006] To achieve the above object, the present invention provides the following technical solutions:

[0007] An intelligent control system for the production tension of polyester fiber cloth, comprising:

[0008] Flexible strain acquisition module: The flexible strain sensor array is used to collect the tension data of the polyester fiber cloth at different positions in real time, and the lateral tension distribution data and longitudinal tension distribution data of each section are obtained;

[0009] Image acquisition module: collects image data of the surface of the polyester fiber cloth in real time through an industrial camera to obtain the surface image of the fiber cloth in each section;

[0010] Laser displacement drawing module: monitors the deformation of polyester fiber cloth in real time through a laser displacement sensor array, and obtains the deformation data of the fiber cloth in each section after CAD drawing;

[0011] Temperature monitoring module: monitors temperature changes in real time through temperature sensors to obtain real-time temperature data of each section;

[0012] A multimodal time series fusion data set generation module: performing time synchronization and preprocessing on the transverse tension distribution data, the longitudinal tension distribution data, the fiber cloth deformation data, the fiber cloth surface image and the real-time temperature data of each of the work sections to obtain a multimodal time series fusion data set;

[0013] Multimodal data fusion and evaluation module: input the multimodal time series fusion data set into the Stacking fusion model to generate a comprehensive tension evaluation value and a comprehensive deformation evaluation value for each of the work sections;

[0014] Work section coupling analysis module: performs work section coupling analysis on the work section to obtain a work section coupling coefficient matrix;

[0015] MPC control decision module: based on the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each section, combined with the section coupling coefficient matrix, the MPC algorithm is used to generate control instructions.

[0016] Furthermore, the working section includes: an unwinding section, a drafting section, a shaping section and a winding section.

[0017] Furthermore, the flexible strain sensor array is arranged on the guide roller surface of each of the work sections and distributed along the transverse and longitudinal directions of the guide roller.

[0018] Furthermore, the industrial camera is installed at a first designated position of each of the work sections; the first designated position is the entrance or exit of each of the work sections, and the industrial camera is installed above the direction perpendicular to the transmission direction of the polyester fiber cloth.

[0019] Furthermore, the laser displacement sensor array is installed at a second designated position of each of the work sections; the second designated position is on both sides of the fiber cloth transmission path of each of the work sections, and the laser displacement sensor array continuously detects the deformation of the polyester fiber cloth along the transmission direction, and the height of the laser displacement sensor array is parallel to the surface of the polyester fiber cloth.

[0020] Furthermore, the temperature sensor is installed at the third designated position of each of the work sections; the third designated position is the key control section of each of the work sections, and the key control sections include: the polyester fiber cloth entrance of the unwinding section, the stretching roller area of ​​the stretching section, the heating roller surface of the shaping section, and the winding machine area of ​​the winding section.

[0021] Furthermore, the Stacking fusion model includes:

[0022] The basic model layer includes four basic learning models, including: a transverse tension prediction model, which is used to receive the transverse tension distribution data and the real-time temperature data that have been time-synchronized and preprocessed through a random forest model, and predict the transverse tension prediction value at a specified time;

[0023] A longitudinal tension prediction model, used for receiving the longitudinal tension distribution data and the real-time temperature data after time synchronization and preprocessing through a random forest model, and predicting a predicted value of the longitudinal tension at the specified time;

[0024] A deformation prediction model, used for receiving the fiber cloth deformation data and the real-time temperature data after time synchronization and preprocessing through a decision tree regression model, and predicting a predicted value of the fiber cloth deformation data at the specified time;

[0025] An image feature prediction model is used to receive the fiber cloth surface image and the real-time temperature data that have been time-synchronized and preprocessed through a convolutional neural network, and predict the surface image prediction features at the specified time; the surface image prediction features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​deformation areas, texture contrast and texture entropy;

[0026] The meta-learner layer is used to use the gradient boosting regression model as a meta-learner, take the transverse tension prediction value, the longitudinal tension prediction value, the fiber cloth deformation data prediction value and the surface image prediction feature as input features, perform nonlinear fusion on the input features, and generate the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each of the work sections.

[0027] Furthermore, the system further comprises: a data verification and correction module, which is used to perform data verification and data correction on the multimodal time series fusion data set, comprising the following steps:

[0028] Inputting the multimodal time series fusion data set into a trained historical multimodal data anomaly detection model to obtain a first classification result, wherein the first classification result includes normal and abnormal;

[0029] If the first classification result is abnormal, perform the following steps:

[0030] Data correction step: identifying abnormal modal data and normal modal data in a multimodal time series fusion data set; inputting the normal modal data into a trained data correction model to obtain abnormal corrected data; the data correction model is specifically a random forest regression model; integrating the abnormal corrected data with the normal modal data to obtain a corrected multimodal time series fusion data set; inputting the corrected multimodal time series fusion data set into the trained historical multimodal data anomaly detection model to obtain a second classification result, the second classification result including normal and abnormal; if the second classification result is normal, using the corrected multimodal time series fusion data set as the multimodal time series fusion data set; if the second classification result is abnormal, repeating the data correction step.

[0031] Furthermore, the historical multimodal data anomaly detection model is obtained by training a deep neural network and includes the following structure:

[0032] An image feature extraction layer is used to receive the surface image of the fiber cloth after time synchronization and preprocessing, and use a convolutional neural network to extract the image features at each moment to obtain surface image features; the surface image features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​the deformation area, the texture contrast and the texture entropy;

[0033] An image feature processing layer, used for standardizing the surface image features to obtain standard surface image features;

[0034] The numerical feature processing layer is used to standardize and fuse the transverse tension distribution data, longitudinal tension distribution data, fiber cloth deformation data and real-time temperature data that have been time-synchronized and pre-processed to obtain standard numerical features;

[0035] A feature fusion layer, used for concatenating the standard surface image features and the standard numerical features to form a joint feature vector;

[0036] A fully connected layer, used for processing the joint feature vector to obtain a high-dimensional joint feature;

[0037] The output layer is used to receive the high-dimensional joint features and output the classification results through the Sigmoid activation function; the classification results are binary classification labels.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] 1. The intelligent control system for tension in the production of polyester fiber cloth proposed in the present invention can realize accurate monitoring and evaluation of tension and deformation in each work section by integrating multimodal data acquisition and data fusion technology. The system collects multidimensional data such as lateral tension, longitudinal tension, deformation, surface image and temperature in real time through flexible strain sensors, industrial cameras, laser displacement sensors and temperature monitoring modules. The Stacking fusion model is used to deeply fuse these multimodal time series data to generate comprehensive tension evaluation values ​​and deformation evaluation values, which significantly improves the accuracy and reliability of the evaluation. Through the work section coupling analysis, the system further reveals the relationship between the various work sections and enhances the accuracy and stability of the control decision. Finally, combined with the MPC (model predictive control) algorithm, the system generates optimal control instructions based on real-time data, realizing the precise regulation of tension and deformation in the production process of polyester fiber cloth. The system makes up for the limitations of single-mode sensors and improves the reliability and efficiency of the overall production process.

[0040] 2. The present invention proposes a Stacking fusion model for data prediction in each work section, which can effectively integrate multimodal data from different sensors and models to provide more accurate tension and deformation assessment. The basic model layer uses multiple basic prediction models to predict transverse tension, longitudinal tension, deformation data and surface image features, covering key multidimensional data in the production process. Through the meta-learner layer, these prediction results are nonlinearly fused, and the system can generate comprehensive tension evaluation values ​​and comprehensive deformation evaluation values, thereby providing more accurate information for subsequent control decisions. This method makes up for the shortcomings of a single model or a single data source, improves the accuracy and robustness of the system, makes the tension and deformation control in the production process of polyester fiber cloth more accurate and stable, and thus optimizes production efficiency and product quality.

[0041] 3. The production tension intelligent control system of the polyester fiber cloth effectively solves the impact of abnormal data in multimodal data on system performance by introducing a data verification and correction module. This module adopts a historical multimodal data anomaly detection model. First, it performs anomaly detection on multimodal time series data and promptly identifies abnormal data and normal data. If an anomaly is detected, the system automatically enters the data correction process and repairs the abnormal data through the random forest regression model to ensure the quality and accuracy of the data. The corrected data will be verified again by the anomaly detection model. If anomalies still exist, corrections will continue until data that meets the standards is obtained. This process not only improves the system's fault tolerance for abnormal data, but also can identify sensor failures, and promptly apply data from other modes for supplementation and repair, thereby avoiding the impact of missing or errors in single modal data on system performance.

[0042] 4. The present invention proposes a historical multimodal data anomaly detection model, which uses a convolutional neural network to extract the surface image features of the fiber cloth, and fuses the transverse tension, longitudinal tension, fiber cloth deformation data and temperature data. The model can process complex multimodal data, extract the effective features of each mode, and perform deep fusion through the fully connected layer and the feature fusion layer. Through this fusion method, the system can make full use of the data of each mode, make up for the shortcomings of a single sensor data, and improve the precise control of tension and deformation in complex production environments. This efficient data verification and correction mechanism not only ensures the high quality of data input, but also greatly improves the stability and reliability of the production process, and ensures the robustness of the intelligent control system under various faults and data anomalies. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Figure 1 An intelligent control system structure for the production tension of polyester fiber cloth provided by an embodiment of the present invention;

[0044] Figure 2 A schematic diagram of the Stacking fusion model structure provided by an embodiment of the present invention;

[0045] Figure 3 A schematic diagram of the structure of a historical multimodal data anomaly detection model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0046] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0047] With the continuous progress of the textile industry, polyester fiber cloth, as a widely used basic material, is becoming increasingly important in clothing, industrial fabrics and functional materials. In order to ensure the excellent performance of the product, the production process and quality control of polyester fiber cloth have become key concerns. From spinning to cloth, polyester fiber needs to go through multiple processes such as cooling, stretching, curling, heat setting and winding. Among these process links, the stability of tension is a key factor affecting the dimensional accuracy, mechanical properties and surface state of the fiber cloth. However, due to the physical properties of polyester fiber itself (such as elastic modulus, fineness) and complex interferences such as speed fluctuations, equipment friction and ambient temperature changes during production line operation, traditional tension monitoring and control methods face great technical challenges and are difficult to meet the requirements of modern production for efficiency and stability.

[0048] See also Figures 1 to 3The present invention provides an intelligent control system for the production tension of polyester fiber cloth, and the technical solution is as follows:

[0049] Embodiment 1

[0050] Textile Company A is a company specializing in the production of high-quality polyester fiber cloth, and its products are mainly used in high-end clothing and industrial fabrics. In recent years, as the market's requirements for product performance and appearance quality continue to increase, the company faces the following challenges: inaccurate tension control during the production process leads to breakage, deformation and surface defects in the fiber cloth during stretching, heat setting and winding; in addition, the traditional tension control system has a slow response speed and cannot adapt to the company's new high-speed production line's needs for real-time monitoring and dynamic regulation.

[0051] In order to solve these problems, textile company A introduced an intelligent control system for the production tension of polyester fiber cloth, such as Figure 1 As shown, including:

[0052] Flexible strain acquisition module: The tension data of polyester fiber cloth at different positions are collected in real time through the flexible strain sensor array to obtain the lateral tension distribution data and longitudinal tension distribution data of each section; further, the section includes: unwinding section, drafting section, shaping section and winding section. The four sections cover the complete process of polyester fiber cloth from the raw materials entering the production line to the finished product winding, ensuring that the tension and deformation data can be obtained in real time in each key process. The present invention divides the production process into unwinding section, drafting section, shaping section and winding section, and monitors and controls the tension and deformation state of each section respectively, which can accurately adapt to the process requirements of polyester fiber cloth production. The unwinding section ensures the stability of the tension of the raw material supply; the drafting section realizes the fiber orientation and stretching uniformity; the shaping section ensures the dimensional stability and physical properties of the fiber cloth under high temperature conditions by controlling deformation and tension; the winding section ensures the winding quality and flatness of the final product. Through this segmented design, not only can the characteristics of each section be finely controlled to reduce tension fluctuations caused by independent or coupled interference in each section, but the overall production process can also be optimized through multi-section linkage, thereby improving the quality consistency and production efficiency of the fiber cloth, reducing the defective rate, and meeting high-precision production requirements.

[0053] Furthermore, the flexible strain sensor array is arranged on the guide roller surface of each section, and is distributed along the lateral and longitudinal directions of the guide roller, so as to realize multi-point, real-time and accurate monitoring of the tension of the fiber cloth. This data acquisition method can fully capture the tension distribution state of the fiber cloth during the transmission process, especially the unevenness of the lateral tension and longitudinal tension. By arranging in the lateral direction, the tension difference of the fiber cloth at different width positions can be detected to avoid problems such as wrinkling and deformation caused by uneven distribution of lateral tension; by arranging in the longitudinal direction, the tension change of the fiber cloth in the transmission direction can be monitored in real time to prevent breakage or surface defects caused by abnormal local tension. This distribution method enhances the comprehensiveness and reliability of tension data, provides high-precision data support for accurate evaluation and dynamic control of tension, and significantly improves the production quality and process stability of the fiber cloth.

[0054] Image acquisition module: The image data of the surface of the polyester fiber cloth is collected in real time by an industrial camera to obtain the surface image of the fiber cloth in each section; further, the industrial camera is installed at the first designated position of each section; the first designated position is the entrance or exit of each section, and the industrial camera is installed above the direction perpendicular to the transmission direction of the polyester fiber cloth. This installation method can effectively capture the surface image of the fiber cloth when it enters or leaves each section, including defects, deformation areas, texture changes and other features, ensure the quality control of key positions, and realize full-width and real-time monitoring of the surface state of the fiber cloth. Installation at the entrance helps to monitor the state of the cloth before entering the section and provide initial conditions for the control of subsequent sections; installation at the exit can evaluate the surface state after processing and discover potential problems in time. The vertical installation method avoids the influence of angle deviation on the accuracy of image acquisition, ensures the accuracy and consistency of image features, thereby improving the system's defect detection and deformation recognition capabilities, and provides reliable image data support for multimodal fusion analysis and dynamic optimization control.

[0055] Laser displacement drawing module: The deformation of polyester fiber cloth is monitored in real time through the laser displacement sensor array, and the deformation data of the fiber cloth of each section is obtained after CAD drawing; further, the laser displacement sensor array is installed at the second designated position of each section; the second designated position is on both sides of the fiber cloth transmission path of each section, and the laser displacement sensor array continuously detects the deformation of polyester fiber cloth along the transmission direction, and the height of the laser displacement sensor array is parallel to the surface of polyester fiber cloth. This design can achieve high-precision, non-contact real-time monitoring of fiber cloth deformation. The sensors are arranged along the transmission direction to capture the dynamic changes of deformation during the transmission of fiber cloth, providing support for detecting instantaneous deformation peaks or local anomalies; the arrangement on both sides can obtain the deformation difference in the width direction of the fiber cloth, and identify problems such as width reduction or local bulging caused by uneven tension. The installation method of the sensor parallel to the surface minimizes the influence of angle deviation on measurement accuracy and ensures the accuracy and consistency of deformation data. This design provides all-round data support for deformation assessment.

[0056] Temperature monitoring module: The temperature change is monitored in real time by the temperature sensor to obtain the real-time temperature data of each section; further, the temperature sensor is installed at the third designated position of each section; the third designated position is the key control section of each section, and the key control section includes: the entrance of the polyester fiber cloth in the unwinding section, the drafting roller area in the drafting section, the heating roller surface in the setting section, and the winding machine area in the winding section. It can accurately monitor the temperature changes of the core parts of each section, so as to timely discover and deal with process problems caused by temperature fluctuations. This design fully considers the process characteristics of different sections: monitoring the temperature at the entrance of the unwinding section can ensure that the polyester fiber cloth is not affected by the external environment temperature when entering the production line; collecting temperature data in the drafting roller area of ​​the drafting section can effectively regulate the uniformity and stability of the fiber stretching process; real-time monitoring of the temperature on the surface of the heating roller in the setting section can directly ensure the accurate execution of the heat setting process; controlling the temperature in the winding machine area of ​​the winding section helps to avoid the problem of uneven winding or uneven tension of the fiber cloth due to temperature changes.

[0057] Multimodal time series fusion data set generation module: time synchronization and preprocessing of the transverse tension distribution data, the longitudinal tension distribution data, the fiber cloth deformation data, the fiber cloth surface image and the real-time temperature data of each work section to obtain a multimodal time series fusion data set; specifically, each work section independently generates a corresponding multimodal time series fusion data set, including a multimodal time series fusion data set of an unwinding section, a multimodal time series fusion data set of a drafting section, a multimodal time series fusion data set of a shaping section, and a multimodal time series fusion data set of a winding section.

[0058] Furthermore, the system further comprises: a data verification and correction module, which is used to perform data verification and data correction on the multimodal time series fusion data set, comprising the following steps:

[0059] Inputting the multimodal time series fusion data set into a trained historical multimodal data anomaly detection model to obtain a first classification result, wherein the first classification result includes normal and abnormal;

[0060] If the first classification result is abnormal, perform the following steps:

[0061] Data correction step: identify abnormal modal data and normal modal data in the multimodal time series fusion data set; input the normal modal data into the trained data correction model to obtain abnormal correction data; the data correction model is specifically a random forest regression model; integrate the abnormal correction data with the normal modal data to obtain a corrected multimodal time series fusion data set; input the corrected multimodal time series fusion data set into the trained historical multimodal data anomaly detection model to obtain a second classification result, the second classification result includes normal and abnormal; if the second classification result is normal, the corrected multimodal time series fusion data set is used as the multimodal time series fusion data set; if the second classification result is abnormal, repeat the data correction step. The introduction of the data verification and correction module can significantly improve the reliability and accuracy of the multimodal time series fusion data set. By classifying the data through the historical multimodal data anomaly detection model, abnormal data can be quickly identified, and the random forest regression model is used to correct the abnormal modal data, effectively restoring the credibility of the abnormal data. This process also verifies the corrected data repeatedly to ensure that the final data set is applied only after it reaches the "normal" standard, thereby significantly reducing the model deviation caused by sensor errors, environmental interference or data loss. This design ensures the robustness of the system in complex production environments, provides a high-quality data foundation for downstream evaluation and control modules, reduces the risk of misjudgment and miscontrol caused by data anomalies, and improves the system's automation level and the stability of the production process.

[0062] Furthermore, the historical multimodal data anomaly detection model is obtained by training a deep neural network, see Figure 3 , including the following structure:

[0063] An image feature extraction layer is used to receive the surface image of the fiber cloth after time synchronization and preprocessing, and use a convolutional neural network to extract the image features at each moment to obtain surface image features; the surface image features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​the deformation area, the texture contrast and the texture entropy;

[0064] An image feature processing layer, used for standardizing the surface image features to obtain standard surface image features;

[0065] The numerical feature processing layer is used to standardize and fuse the transverse tension distribution data, longitudinal tension distribution data, fiber cloth deformation data and real-time temperature data that have been time-synchronized and pre-processed to obtain standard numerical features;

[0066] A feature fusion layer, used for concatenating the standard surface image features and the standard numerical features to form a joint feature vector;

[0067] A fully connected layer, used for processing the joint feature vector to obtain a high-dimensional joint feature;

[0068] The output layer is used to receive the high-dimensional joint features and output the classification results through the Sigmoid activation function; the classification results are binary classification labels. The complementarity of image and numerical features is fully utilized to greatly improve the ability to detect abnormal data in complex production environments, ensuring the system's superior performance in real-time and accuracy, thereby providing a reliable foundation for data correction and subsequent evaluation modules, and further ensuring the stability of the production process and the effectiveness of intelligent control.

[0069] Multimodal data fusion and evaluation module: input the multimodal time series fusion data set into the Stacking fusion model to generate a comprehensive tension evaluation value and a comprehensive deformation evaluation value for each of the work sections;

[0070] Specifically, the unwinding section multimodal time series fusion data set, the drafting section multimodal time series fusion data set, the shaping section multimodal time series fusion data set, and the winding section multimodal time series fusion data set are respectively input into the corresponding Stacking fusion model to obtain the comprehensive tension evaluation value and comprehensive deformation evaluation value of the corresponding section. The correlation between the section data, model and output is shown in Table 1.

[0071] Table 1 Work section data - model - output association table

[0072]

[0073] Furthermore, if Figure 2 As shown, the Stacking fusion model includes:

[0074] The basic model layer includes four basic learning models, including: a transverse tension prediction model, which is used to receive the transverse tension distribution data and the real-time temperature data that have been time-synchronized and preprocessed through a random forest model, and predict the transverse tension prediction value at a specified time;

[0075] A longitudinal tension prediction model, used for receiving the longitudinal tension distribution data and the real-time temperature data that have been time-synchronized and preprocessed through a random forest model, and predicting a predicted value of the longitudinal tension at the specified time;

[0076] A deformation prediction model, used for receiving the fiber cloth deformation data and the real-time temperature data after time synchronization and preprocessing through a decision tree regression model, and predicting a predicted value of the fiber cloth deformation data at the specified time;

[0077] An image feature prediction model is used to receive the fiber cloth surface image and the real-time temperature data that have been time-synchronized and preprocessed through a convolutional neural network, and predict the surface image prediction features at the specified time; the surface image prediction features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​deformation areas, texture contrast and texture entropy;

[0078] The meta-learner layer is used to use the gradient boosting regression model as a meta-learner, take the transverse tension prediction value, the longitudinal tension prediction value, the fiber cloth deformation data prediction value and the surface image prediction feature as input features, perform nonlinear fusion on the input features, and generate the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each section. By adopting the Stacking fusion model, combined with the division of labor and cooperation between the basic model layer and the meta-learner layer, efficient fusion and accurate prediction of multimodal data are achieved. In the basic model layer, different modal data (such as transverse tension, longitudinal tension, fiber cloth deformation and surface image) are processed by random forest, decision tree regression and convolutional neural network respectively, fully mining the specific features of each modal data to ensure the prediction accuracy of a single modality. The meta-learner layer nonlinearly fuses the prediction results of each base learning model through the gradient boosting regression model, comprehensively considers the correlation and complementarity between different modal features, and thus generates more comprehensive and accurate comprehensive tension evaluation value and comprehensive deformation evaluation value. This design can not only perform refined processing on different modal data, but also improve the overall robustness of the system through the fusion of multimodal data, reducing the sensitivity to abnormalities of a single data source, and is particularly suitable for dynamic monitoring and control needs in complex production scenarios. Ultimately, the system significantly improved the assessment accuracy of tension and deformation states in the production process of polyester fiber cloth, providing strong technical support for high-quality and efficient production.

[0079] Work section coupling analysis module: performs work section coupling analysis on the work sections to obtain a work section coupling coefficient matrix between the work sections;

[0080] MPC control decision module: based on the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each section, combined with the section coupling coefficient matrix, the MPC algorithm is used to generate control instructions.

[0081] The polyester fiber cloth production tension intelligent control system of the present invention realizes real-time, accurate monitoring and intelligent regulation of tension and deformation in the production process through multimodal data collection and fusion analysis. Through the coordinated work of flexible strain sensors, industrial cameras, laser displacement sensors and temperature sensors, the multimodal time series fusion data set provides a comprehensive representation of the production status; the Stacking fusion model and the section coupling analysis module can accurately evaluate the comprehensive state of tension and deformation, and quantify the dynamic coupling relationship between multiple sections; combined with the MPC control decision module, the system can generate optimized dynamic control instructions, thereby effectively reducing tension fluctuations, deformation out of control and surface defects, etc., improving the stability and automation level of the production process, significantly improving the product quality and production efficiency of polyester fiber cloth, and reducing the defective rate and resource waste.

[0082] Embodiment 2

[0083] Textile Company B is an enterprise mainly engaged in the production of functional polyester fiber cloth. Its products are widely used in outdoor sportswear, medical fabrics, aviation materials and other fields, and have extremely high requirements for the performance stability and processing accuracy of the cloth. The main problems currently faced by the company are: the dynamic coupling effect between multiple sections in the production process is complex, and local tension fluctuations will gradually amplify, resulting in width shrinkage and uneven deformation of the fiber cloth in subsequent processing. At the same time, since the processing conditions of functional polyester fiber cloth have extremely low tolerance for temperature, tension and deformation, traditional single-mode monitoring and segmented control strategies cannot meet its high-precision requirements, which seriously affects production efficiency and order delivery capabilities.

[0084] In order to solve these problems, textile company B adopted the polyester fiber cloth production tension intelligent control system of the present invention, see Figure 1 The system used by Textile Company B has the same functions as the system mentioned in Example 1, except that:

[0085] As a specific and feasible method: First, the convolutional neural network (CNN) receives the preprocessed fiber cloth surface image, and extracts local features including edge, texture and image distribution features through multi-layer convolution and pooling operations; these features are further extracted through the fully connected layer to extract high-dimensional feature representation. The number of surface defects is realized through the target detection branch, by setting the bounding box and counting the number of targets; the number of dense and sparse areas is based on the pixel density clustering algorithm, and the area of ​​the deformed area is divided and counted on the feature map; the area of ​​the deformed area is realized through shape segmentation on the feature map, and is calculated by the connectivity and area of ​​the pixel area; the texture contrast is calculated by the grayscale distribution in the feature map, and the contrast formula is used for statistics; the texture entropy is analyzed by the grayscale co-occurrence matrix of the image features, and the distribution entropy value of the features is calculated. These results are integrated by the output layer of the convolutional neural network to form the final multi-index prediction result. The model is optimized based on the labeled data and the loss function (cross entropy) during the training phase to ensure the prediction accuracy of all output indicators.

[0086] Implementing Stacking in a multimodal scenario includes the following steps: First, align and preprocess the data (tension, image, deformation, and temperature) collected by different sensors according to the timestamp to ensure that each record contains the features corresponding to each modality; then, select an appropriate base model for each modality, train them separately, and obtain the prediction results for the same batch of samples; finally, input the predicted values ​​of the output of each base model for the same sample in the verification / testing phase as new features to the meta-learner (meta-model) for retraining or prediction. Through the comprehensive learning of the output of the multimodal model, the complementarity of different modal features is fully utilized to obtain more accurate and robust fusion prediction results. In this application, although data such as lateral tension, longitudinal tension, deformation, and image features come from different sensors or different modalities, they actually refer to samples of "the same production batch, the same time, or the same section". This information is first synchronized and preprocessed in time through the multimodal time series fusion data set generation module to form a batch of multimodal samples. Then each base model takes the required features for training and prediction, and finally inputs the prediction values ​​of each base model for the same batch of samples into the meta-learner to complete Stacking, which meets the requirements of Stacking. Among them, the Stacking fusion model is a machine learning model trained in advance based on the historical data of each work section.

[0087] A set of control experiments was also conducted. The experimental group adopted the intelligent control system proposed in the present invention, including a multimodal sensor, a Stacking fusion model, a data verification and correction module, and an MPC control decision module.

[0088] Control group 1: Only transverse and longitudinal tension data were collected, deformation, image and temperature data were removed, and the rest of the operations were the same as the experimental group

[0089] Control group 2: Collect multimodal data, but only use the random forest model to predict tension, without multi-model fusion, and the rest of the operations are the same as the experimental group

[0090] Control group 3: Abnormal data were directly removed without any correction operation, and the rest of the operations were the same as those of the experimental group;

[0091] Control group 4: After obtaining the comprehensive tension evaluation value and comprehensive deformation evaluation value of each section, traditional PID control is adopted without considering the coupling between sections.

[0092] Table 2 Comparative experimental results analysis table

[0093]

[0094] Referring to Table 2, the comparative experimental results analysis table clearly demonstrates the technical advantages of the polyester fiber cloth production tension intelligent control system of the present invention in key modules. Through multimodal data acquisition, the experimental group significantly reduced tension fluctuations (standard deviation from 0.36N to 0.12N), improved defect recognition rate (from 82.10% to 95.40%), and controlled deformation deviation within 0.02mm. The Stacking fusion model further optimizes the accuracy of comprehensive tension and deformation evaluation, reducing the evaluation errors from 0.08 and 0.10 to 0.03 and 0.04 respectively; the data verification and correction module shows an extremely high abnormal data correction success rate (98.60%), which effectively improves the prediction accuracy after correction (MAE from 0.12 to 0.03). In addition, through the MPC control decision module, the system greatly reduces tension fluctuations (standard deviation from 0.45N to 0.12N), and the control response time is shortened to 0.8 seconds (compared to 2.3 seconds). Experimental results show that the system of the present invention is significantly superior to traditional methods in multimodal data fusion, abnormal data processing and dynamic control, and significantly improves the stability and product quality of polyester fiber cloth production.

[0095] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An intelligent control system for the production tension of polyester fiber cloth, characterized in that: include: Flexible strain acquisition module: The flexible strain sensor array is used to collect the tension data of the polyester fiber cloth at different positions in real time, and the lateral tension distribution data and longitudinal tension distribution data of each section are obtained; Image acquisition module: collects image data of the surface of the polyester fiber cloth in real time through an industrial camera to obtain the surface image of the fiber cloth in each section; Laser displacement drawing module: monitors the deformation of polyester fiber cloth in real time through a laser displacement sensor array, and obtains the deformation data of the fiber cloth in each section after CAD drawing; Temperature monitoring module: monitors temperature changes in real time through temperature sensors to obtain real-time temperature data of each section; A multimodal time series fusion data set generation module: performing time synchronization and preprocessing on the transverse tension distribution data, the longitudinal tension distribution data, the fiber cloth deformation data, the fiber cloth surface image and the real-time temperature data of each of the work sections to obtain a multimodal time series fusion data set; Multimodal data fusion and evaluation module: input the multimodal time series fusion data set into the Stacking fusion model to generate a comprehensive tension evaluation value and a comprehensive deformation evaluation value for each of the work sections; Work section coupling analysis module: performs work section coupling analysis on the work section to obtain a work section coupling coefficient matrix; MPC control decision module: based on the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each section, combined with the section coupling coefficient matrix, the MPC algorithm is used to generate control instructions.

2. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The working sections include: an unwinding section, a drafting section, a shaping section and a winding section.

3. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The flexible strain sensor array is arranged on the guide roller surface of each section and distributed along the transverse and longitudinal directions of the guide roller.

4. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The industrial camera is installed at the first designated position of each of the work sections; the first designated position is the entrance or exit of each of the work sections, and the industrial camera is installed above the direction perpendicular to the transmission direction of the polyester fiber cloth.

5. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The laser displacement sensor array is installed at the second designated position of each of the work sections; the second designated position is on both sides of the fiber cloth transmission path of each of the work sections, and the laser displacement sensor array continuously detects the deformation of the polyester fiber cloth along the transmission direction, and the height of the laser displacement sensor array is parallel to the surface of the polyester fiber cloth.

6. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The temperature sensor is installed at the third designated position of each of the work sections; the third designated position is the key control section of each of the work sections, and the key control sections include: the polyester fiber cloth entrance of the unwinding section, the stretching roller area of ​​the stretching section, the heating roller surface of the shaping section, and the winding machine area of ​​the winding section.

7. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The Stacking fusion model includes: The basic model layer includes four basic learning models, including: a transverse tension prediction model, which is used to receive the transverse tension distribution data and the real-time temperature data that have been time-synchronized and preprocessed through a random forest model, and predict the transverse tension prediction value at a specified time; A longitudinal tension prediction model, used for receiving the longitudinal tension distribution data and the real-time temperature data after time synchronization and preprocessing through a random forest model, and predicting a predicted value of the longitudinal tension at the specified time; A deformation prediction model, used for receiving the fiber cloth deformation data and the real-time temperature data after time synchronization and preprocessing through a decision tree regression model, and predicting a predicted value of the fiber cloth deformation data at the specified time; An image feature prediction model is used to receive the fiber cloth surface image and the real-time temperature data that have been time-synchronized and preprocessed through a convolutional neural network, and predict the surface image prediction features at the specified time; the surface image prediction features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​deformation areas, texture contrast and texture entropy; The meta-learner layer is used to use the gradient boosting regression model as a meta-learner, take the transverse tension prediction value, the longitudinal tension prediction value, the fiber cloth deformation data prediction value and the surface image prediction feature as input features, perform nonlinear fusion on the input features, and generate the comprehensive tension evaluation value and the comprehensive deformation evaluation value of each of the work sections.

8. The intelligent control system for production tension of polyester fiber cloth according to claim 1, characterized in that: The system further includes: a data verification and correction module, which is used to perform data verification and data correction on the multimodal time series fusion data set, including the following steps: Inputting the multimodal time series fusion data set into a trained historical multimodal data anomaly detection model to obtain a first classification result, wherein the first classification result includes normal and abnormal; If the first classification result is abnormal, perform the following steps: Data correction step: identifying abnormal modal data and normal modal data in a multimodal time series fusion data set; inputting the normal modal data into a trained data correction model to obtain abnormal corrected data; the data correction model is specifically a random forest regression model; integrating the abnormal corrected data with the normal modal data to obtain a corrected multimodal time series fusion data set; inputting the corrected multimodal time series fusion data set into the trained historical multimodal data anomaly detection model to obtain a second classification result, the second classification result including normal and abnormal; if the second classification result is normal, using the corrected multimodal time series fusion data set as the multimodal time series fusion data set; if the second classification result is abnormal, repeating the data correction step.

9. The intelligent control system for production tension of polyester fiber cloth according to claim 8, characterized in that: The historical multimodal data anomaly detection model is obtained by training a deep neural network and includes the following structure: An image feature extraction layer, used for receiving the surface image of the fiber cloth after time synchronization and preprocessing, and extracting the image features at each moment using a convolutional neural network to obtain surface image features; The surface image features include: the number of surface defects, the number of dense areas, the number of sparse areas, the area of ​​deformed areas, texture contrast and texture entropy; An image feature processing layer, used for standardizing the surface image features to obtain standard surface image features; The numerical feature processing layer is used to standardize and fuse the transverse tension distribution data, longitudinal tension distribution data, fiber cloth deformation data and real-time temperature data that have been time-synchronized and pre-processed to obtain standard numerical features; A feature fusion layer, used for concatenating the standard surface image features and the standard numerical features to form a joint feature vector; A fully connected layer, used for processing the joint feature vector to obtain a high-dimensional joint feature; The output layer is used to receive the high-dimensional joint features and output the classification results through the Sigmoid activation function; the classification results are binary classification labels.

Images