A polyester thread tensile strength intelligent detection system for production field artificial intelligence

Abstract

The application relates to the field of artificial intelligence, and particularly discloses a polyester thread tensile strength intelligent detection system in the field of production, aiming to solve the problems of low detection efficiency, insufficient real-time performance, incapability of continuous monitoring, and difficulty in early warning and intervention of quality fluctuation of the existing polyester thread tensile strength detection; the system integrates a plurality of modules such as multi-modal sensor data acquisition, data preprocessing, feature engineering, deep learning intelligent detection model, real-time prediction, result feedback and process optimization, and knowledge base, realizes non-contact, high-precision, real-time online prediction of the polyester thread tensile strength and self-adaptive optimization of the production process, thereby greatly improving product quality stability and production efficiency, and realizing early warning of quality abnormalities.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to an intelligent detection system for the tensile strength of polyester yarn in the production field. Background Technology

[0002] In the manufacturing sector, product quality control is a core element in ensuring production efficiency and market competitiveness. This is especially true for polyester yarn production in the textile industry, where tensile strength, as a key physical property, directly determines the quality and lifespan of downstream products. However, existing methods for testing and managing the tensile strength of polyester yarn still face significant challenges: traditional testing methods rely heavily on manual sampling or destructive laboratory testing, which are time-consuming and inefficient, failing to meet the needs of real-time, continuous monitoring on production lines; although some automated testing equipment has been introduced, it is often limited by single testing principles and data processing capabilities, making it difficult to cope with complex and changing production environments and material properties, and lacking sufficient accuracy and real-time response capability in identifying subtle defects or strength anomalies.

[0003] In particular, the complexity of polyester yarn tensile strength testing and the multi-dimensional conflict between the testing objectives further exacerbate the shortcomings of existing technologies. For example, there may be a contradiction between pursuing extremely high testing accuracy and maintaining high-speed operation of the production line, while the differences in characteristics of different batches and specifications of polyester yarn increase the difficulty of universalizing the testing strategy. In addition, existing testing systems generally lack the ability to integrate and analyze multi-source data, making it difficult to effectively correlate tensile strength data with production process parameters, environmental conditions, etc., resulting in the inability to achieve the leap from "testing" to "quality prediction and optimization," which can easily lead to hidden quality risks or decreased production efficiency.

[0004] Existing technologies for testing the tensile strength of polyester yarn generally suffer from problems such as low efficiency, insufficient accuracy, poor real-time performance, and limited levels of intelligence. These shortcomings seriously affect the quality stability and production efficiency of polyester yarn products. Therefore, there is an urgent need for an intelligent polyester yarn tensile strength testing system that can achieve high precision, real-time performance, and intelligent artificial intelligence in the production field. Summary of the Invention

[0005] To address the technical problems of low efficiency, insufficient real-time performance, inability to continuously monitor the entire product line, and difficulty in predicting and proactively intervening in quality fluctuations during existing polyester yarn production processes, this invention proposes an intelligent polyester yarn tensile strength detection system based on artificial intelligence for the production field. This system integrates multimodal sensor data acquisition, advanced data preprocessing technology, a deep learning-based intelligent detection model, and a closed-loop feedback control mechanism. This enables non-contact, high-precision, real-time online prediction of polyester yarn tensile strength and adaptive optimization of the production process, thereby significantly improving product quality stability and production efficiency.

[0006] According to one aspect of the present invention, an intelligent detection system for the tensile strength of polyester yarn using artificial intelligence in the production field is provided, comprising: a data acquisition module, a data preprocessing module, a feature engineering and selection module, an intelligent detection model construction and training module, a real-time prediction and evaluation module, a result feedback and process optimization module, and a knowledge base and experience accumulation module.

[0007] In one embodiment of the present invention, a data acquisition module is used to acquire multi-dimensional physical characteristic data of polyester yarn on the production line in real time. The data acquisition module includes: an optical imaging submodule, a spectral analysis submodule, a laser profile measurement submodule, a tension detection submodule, and an environmental parameter detection submodule. The optical imaging submodule acquires surface images of the polyester yarn using a high-speed linear array camera or area array camera. These images are used to capture the diameter uniformity, surface defects, gloss, and texture features of the polyester yarn. The spectral analysis submodule acquires chemical composition and molecular structure information of the polyester yarn using a near-infrared spectrometer or Raman spectrometer. This information is directly related to the crystallinity, orientation, and molecular weight distribution of the polymer. The laser profile measurement submodule acquires precise cross-sectional profile data and diameter variation data of the polyester yarn using a laser displacement sensor or a triangulation laser sensor. This data reflects the geometric uniformity of the polyester yarn. The tension detection submodule acquires real-time tension data of the polyester yarn during the production process using a high-precision tension sensor. This data reflects the stability of the process conditions. The environmental parameter detection submodule acquires temperature and humidity data of the production site using temperature and humidity sensors. This data is used to correct sensor drift or as auxiliary input for the model. The data acquisition module continuously acquires the above data at a preset sampling frequency and sends the raw data to the data preprocessing module through the data transmission interface.

[0008] Furthermore, the data preprocessing module is used to clean, synchronize, calibrate, and perform preliminary processing on the raw data acquired by the data acquisition module. The data preprocessing module includes: a data synchronization submodule, a noise reduction submodule, a data calibration submodule, an outlier detection and processing submodule, and a data normalization submodule. The data synchronization submodule synchronizes heterogeneous data from different sensors to a unified time base using a timestamp alignment algorithm, ensuring temporal consistency between data. The noise reduction submodule denoises the raw data using a digital filtering algorithm, reducing the impact of random noise on data quality. The data calibration submodule performs linear or nonlinear calibration on the sensor data using a preset calibration model, compensating for errors caused by sensor drift and environmental factors. The outlier detection and processing submodule identifies and processes outliers in the data using statistical methods or machine learning algorithms, ensuring the validity and reliability of the data. The data normalization submodule scales the preprocessed data to a uniform numerical range to eliminate the influence between data of different dimensions, facilitating subsequent model training and processing. The data preprocessing module then sends the processed, normalized data to the feature engineering and selection module.

[0009] In one embodiment of the present invention, a feature engineering and selection module is used to extract effective features highly correlated with the tensile strength of polyester yarn from preprocessed data and optimize these features. The feature engineering and selection module includes a multi-dimensional feature extraction submodule, a feature dimensionality reduction submodule, and a feature optimization submodule. The multi-dimensional feature extraction submodule extracts visual features, including texture features, shape features, color features, and brightness distribution features, from optical images. Texture features are calculated using methods such as gray-level co-occurrence matrix, local binary mode, or wavelet transform. Shape features are calculated using edge detection and region analysis methods. The multi-dimensional feature extraction submodule extracts spectral features, including absorbance at a specific wavelength, transmittance, spectral peak intensity, peak position, and peak width, from spectral data. The multi-dimensional feature extraction submodule extracts geometric features, including mean diameter, diameter variance, roundness deviation, and surface roughness, from laser profile data. The multi-dimensional feature extraction submodule extracts statistical features from tension data and environmental parameters. The feature dimensionality reduction submodule uses algorithms such as principal component analysis, linear discriminant analysis, or manifold learning to project data from a high-dimensional feature space into a lower-dimensional space, thereby reducing data redundancy and preserving core information. The feature optimization submodule uses algorithms based on correlation, mutual information, or recursive feature elimination to select the subset of features that contributes most to tensile strength prediction, thus optimizing the model's input. The feature engineering and selection module sends the optimized feature vectors to the intelligent detection model construction and training module or the real-time prediction and evaluation module.

[0010] Furthermore, the intelligent detection model construction and training module is used to build and optimize the polyester yarn tensile strength prediction model based on historical data and physical test results. This module includes: a model selection submodule, a model training submodule, a model evaluation and optimization submodule, and a model deployment submodule. The model selection submodule selects a deep learning model architecture based on the feature data type and prediction task requirements. This architecture includes convolutional neural networks, recurrent neural networks, long short-term memory networks, or Transformer models, and employs a multimodal fusion network based on the characteristics of multimodal data. Convolutional neural networks are used to process image and contour features, extracting spatial hierarchical features. Recurrent neural networks or long short-term memory networks are used to process time-series features, capturing temporal dependencies. Transformer models model long-sequence dependencies through a self-attention mechanism. Multimodal fusion networks fuse features from different modalities, learning the complex relationships between them. The model training submodule uses historically collected feature data and corresponding laboratory physical tensile strength test results as label data, and iteratively trains the model parameters using backpropagation algorithms and optimizers such as Adam or SGD to minimize the error between predicted and true values. The model evaluation and optimization submodule evaluates the model's performance using cross-validation, including prediction accuracy, generalization ability, and stability. Based on the evaluation results, it adjusts and optimizes the model architecture, hyperparameters, and training strategies. The model deployment submodule encapsulates the trained and performance-compliant model and loads it into the real-time prediction and evaluation module for online inference.

[0011] In one embodiment of the present invention, a real-time prediction and evaluation module is used to perform real-time, continuous, and intelligent prediction of the tensile strength of polyester yarn on a production line. The real-time prediction and evaluation module includes: a data receiving interface, a prediction inference engine, a quality evaluation submodule, and an early warning generation submodule. The data receiving interface receives real-time optimized feature vectors from the feature engineering and selection module. The prediction inference engine loads the training model deployed by the intelligent detection model construction and training module and performs high-speed inference calculations on the received feature vectors to generate a predicted value of the tensile strength of the polyester yarn at the current moment and the corresponding prediction confidence level. The quality evaluation submodule compares the predicted tensile strength value with a preset quality standard or acceptable range to determine the current quality status of the polyester yarn. The early warning generation submodule immediately generates an early warning signal when the predicted tensile strength value is below the lower quality limit, above the upper quality limit, or shows significant fluctuations, and sends it to the human-computer interaction and visualization module and the result feedback and process optimization module through the early warning interface. The early warning signal includes the early warning level, occurrence time, predicted strength value, and possible cause analysis.

[0012] Furthermore, the result feedback and process optimization module is used to perform closed-loop control of polyester yarn production process parameters based on real-time prediction results, achieving adaptive optimization of the production process. The result feedback and process optimization module includes: a control strategy generation submodule, a parameter adjustment submodule, and an optimization effect evaluation submodule. The control strategy generation submodule generates corresponding process parameter adjustment strategies based on real-time predicted tensile strength values, quality assessment results, and early warning information, combined with preset process rules and optimization objectives. The strategy generation process employs advanced control algorithms, including proportional-integral-derivative control, model predictive control, or reinforcement learning control. The parameter adjustment submodule, based on the adjustment strategy output by the control strategy generation submodule, sends control commands to relevant actuators on the production line via an industrial control interface or digital-to-analog converter, precisely adjusting key process parameters such as spinning temperature, stretch ratio, winding tension, cooling rate, and heat setting temperature. The optimization effect evaluation submodule continuously monitors the production line status after parameter adjustment and subsequent tensile strength prediction results, evaluates the effectiveness of the optimization strategy, and adaptively adjusts the algorithm parameters of the control strategy generation submodule based on the evaluation results, achieving continuous learning and performance improvement of the system.

[0013] In one embodiment of the present invention, a knowledge base and experience accumulation module is used to store, manage, and utilize the system's historical operating data, model versions, and optimization experience. The knowledge base and experience accumulation module includes: a historical data storage submodule, a model version management submodule, an optimization experience base submodule, and a knowledge reasoning submodule. The historical data storage submodule persistently stores the raw data acquired by the data acquisition module, the data processed by the data preprocessing module, the features extracted by the feature engineering and selection module, the training data and labels from the intelligent detection model construction and training module, and the prediction results and quality assessment reports from the real-time prediction and evaluation module. The model version management submodule manages different versions of the model generated by the intelligent detection model construction and training module, including the model's training parameters, performance indicators, and deployment status, ensuring the model's traceability and backtrackability. The optimization experience base submodule stores various control strategies, parameter adjustment records, and corresponding optimization effects generated by the result feedback and process optimization module, forming reusable knowledge assets. The knowledge reasoning submodule uses artificial intelligence reasoning technology to perform in-depth analysis of historical data and experience, uncovering potential process patterns and quality influencing factors, providing decision support for the iterative upgrading of the intelligent detection model and the improvement of process optimization strategies.

[0014] This invention also includes a human-computer interaction and visualization module, which provides users with intuitive system status monitoring, data query, result display, early warning information prompts, and parameter setting functions. The human-computer interaction and visualization module displays the tensile strength curve, historical trends, quality distribution map, and various key process parameters of the polyester yarn in real time through a graphical user interface, and supports users in generating customized reports and exporting data.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0016] This invention achieves non-contact, high-precision, real-time online intelligent detection of the tensile strength of polyester yarn. By integrating multimodal sensor technology, it overcomes the shortcomings of traditional destructive testing methods, such as poor sample representativeness, long detection cycles, and inability to fully cover the production line. The system can continuously acquire multi-dimensional information on the polyester yarn during the production process, including optical, spectral, geometric, tension, and environmental data. Combined with advanced artificial intelligence algorithms, it accurately predicts the tensile strength, improving the comprehensiveness and timeliness of the detection.

[0017] This invention constructs an intelligent detection model based on deep learning. This model can learn and mine complex nonlinear relationships and potential features related to the tensile strength of polyester yarn from massive amounts of multimodal heterogeneous data. Through a multimodal fusion network, it effectively integrates the complementary advantages of data from different sensors, improving the robustness and accuracy of the prediction model, far exceeding the detection capabilities of traditional models based on single features or simple statistical models.

[0018] This invention's system possesses powerful data-driven and adaptive learning capabilities. Through the synergistic effect of the intelligent detection model construction and training module, as well as the knowledge base and experience accumulation module, the system can continuously accumulate production data and detection experience, constantly optimizing and iterating the prediction model and control strategies. Model version management and optimization experience base ensure the system's continuous evolution and performance improvement, enabling it to adapt to changes in different polyester yarn product types and process conditions.

[0019] This invention achieves closed-loop control and proactive optimization of the polyester yarn production process. Through the result feedback and process optimization module, the system can intelligently generate and execute process parameter adjustment strategies based on real-time tensile strength prediction results, such as precisely controlling spinning temperature and stretching ratio. This proactive, real-time process intervention capability significantly reduces the risk of product quality fluctuations, reduces the generation of defective products, and thus improves production efficiency and material utilization.

[0020] This invention enables early detection and rapid response to quality anomalies through an early warning generation submodule. When the predicted tensile strength value deviates from the normal range, the system can immediately issue an early warning and provide possible cause analysis, allowing operators to quickly locate the problem and take corrective measures, thus avoiding batch quality problems and ensuring product quality stability.

[0021] This invention deeply integrates artificial intelligence technology into the entire quality control chain of polyester yarn production, constructing a highly automated and intelligent production quality management system from data acquisition, feature extraction, intelligent prediction to process optimization. This system not only provides accurate quality inspection, but more importantly, it realizes a shift from passive inspection to proactive control, bringing significant economic and social benefits to manufacturing enterprises. Attached Figure Description

[0022] Figure 1 This is a schematic diagram of the overall technical solution architecture of the present invention;

[0023] Figure 2 This is a schematic diagram of the core principle framework of the deep learning-based multimodal fusion intelligent detection model in this invention. Detailed Implementation

[0024] Please refer to Figure 1 and Figure 2 The intelligent polyester yarn tensile strength testing system based on artificial intelligence in the production field, as described in this embodiment, has a core architecture including a data acquisition module, a data preprocessing module, a feature engineering and selection module, an intelligent detection model construction and training module, a real-time prediction and evaluation module, a result feedback and process optimization module, and a knowledge base and experience accumulation module. In addition, the system also includes a human-computer interaction and visualization module to support users in system monitoring and data management.

[0025] The data acquisition module is responsible for acquiring multi-dimensional physical characteristic data of polyester yarn on the production line in real time. This module is equipped with an optical imaging submodule, a spectral analysis submodule, a laser profile measurement submodule, a tension detection submodule, and an environmental parameter detection submodule to ensure the comprehensiveness and diversity of the data.

[0026] The optical imaging submodule integrates a high-speed linear or area scan camera, deployed above the polyester production line to continuously acquire surface images of the polyester yarn in a non-contact manner. This image acquisition system operates in the visible light band and can be configured with various light sources, such as diffused or parallel light sources, to adapt to polyester yarns with different surface characteristics. High-speed imaging ensures clear capture of rapidly moving polyester yarns on the production line, avoiding motion blur. The acquired image data contains information about the polyester yarn's diameter uniformity, surface defects such as fuzz, spots, and foreign objects, gloss, and fine texture features. This image information is transmitted via a high-resolution digital interface; the raw images are typically in grayscale or color bitmap format. Typical sampling frequencies can be set to hundreds to thousands of frames per second to ensure uninterrupted monitoring of the continuous production line. After passing through an internal frame buffer, the image data is sent to the data preprocessing module via a data transmission interface. The data transmission interface uses Gigabit Ethernet or a dedicated high-speed serial communication protocol to ensure stable real-time transmission of large amounts of data.

[0027] The spectral analysis submodule is equipped with a near-infrared spectrometer or a Raman spectrometer. This spectrometer operates non-contactly by illuminating the surface of the polyester thread with a beam of light of a specific wavelength and collecting the reflected or scattered light signals. Near-infrared spectroscopy is primarily used to analyze the chemical bond vibrations of the polymer, revealing key intrinsic properties of the polyester thread such as its chemical composition, crystallinity, orientation, and molecular weight distribution. Raman spectroscopy provides more detailed molecular structure information. Spectral data is output as wavelength versus intensity or absorbance curves, reflecting the characteristics of the interaction between light and the polyester thread at different wavelengths. Typical spectral ranges for the instrument are, for example, 800 nm to 2500 nm. The spectrometer's built-in fiber optic probe is aimed at the polyester thread and features automatic calibration to compensate for drift caused by environmental changes. The data acquisition frequency is matched to the production line speed to ensure representative spectral data is obtained for each section of the polyester thread. Spectral data is transmitted to the data preprocessing module via a universal serial bus or Ethernet interface.

[0028] The laser profile measurement submodule employs laser displacement sensors or triangulation laser sensors. These sensors precisely emit laser beams and receive reflected light, acquiring high-precision cross-sectional profile data and diameter variation data of the polyester thread by measuring the optical path travel time or using triangulation geometry principles. This geometric data directly reflects the physical dimensional uniformity, roundness, and potential microscopic geometric defects of the polyester thread. Sensors can be deployed in arrays along the vertical or horizontal direction of the polyester thread to obtain profile information from multiple angles or all directions. For example, real-time three-dimensional point cloud data of the polyester thread can be constructed using multiple sensor arrays. The sensors achieve measurement accuracy down to the micrometer level and have a high response frequency, sufficient to capture minute fluctuations in the diameter of the polyester thread. The measurement data includes a series of discrete point coordinates or diameter values. The laser profile measurement submodule transmits the raw geometric data to the data preprocessing module via a dedicated high-speed data acquisition card or industrial Ethernet protocol.

[0029] The tension detection submodule is equipped with a high-precision tension sensor. This sensor is typically integrated into the production line as a non-contact roller or fixed guide wheel to sense real-time tension changes in the polyester yarn during the production process. Tension is a key parameter in processes such as spinning, stretching, and winding; its stability directly affects the molecular orientation and crystallinity of the polyester yarn, thus being closely related to the final tensile strength. The sensor converts mechanical tension into an electrical signal, which is then digitized and output through an analog-to-digital converter. The sensor's sampling frequency is much higher than the rate of process change, for example, data is collected every millisecond to ensure that instantaneous tension fluctuations are captured. Real-time tension data is transmitted to the data preprocessing module via an industrial fieldbus such as Profinet or EtherCAT.

[0030] The environmental parameter detection submodule includes temperature and humidity sensors. These sensors are deployed at key locations on the production site to monitor the temperature and humidity of the production area in real time. Ambient temperature and humidity affect the physical properties of the polyester yarn, sensor performance, and the stability of data acquisition. For example, temperature fluctuations may cause slight changes in the material properties of the polyester yarn or drift in the optical sensors. The collected environmental data is used in the data preprocessing module for sensor drift correction or as auxiliary input features for the intelligent detection model to enhance its robustness. The environmental parameter data is acquired at a relatively low frequency, such as once per second or once per minute, and transmitted to the data preprocessing module via a wireless sensor network or industrial fieldbus.

[0031] The data acquisition module continuously collects all the above data at a preset sampling frequency and sends the raw data to the data preprocessing module through a unified data transmission interface, ensuring the coordination of data flow between the sub-modules.

[0032] The data preprocessing module is responsible for cleaning, synchronizing, calibrating, and initially processing the raw multimodal data acquired by the data acquisition module. Its purpose is to eliminate data noise, compensate for sensor errors, and standardize the data format, providing high-quality input for subsequent feature engineering. This module includes a data synchronization submodule, a noise reduction submodule, a data calibration submodule, an outlier detection and processing submodule, and a data normalization submodule.

[0033] The data synchronization submodule uses a high-precision timestamp alignment algorithm to synchronize heterogeneous data from different sensors, such as optical imaging, spectral analysis, laser profilometry, tension detection, and environmental parameter detection, to a unified time reference. This process is crucial because different sensors may have different sampling frequencies, data transmission delays, and internal clock drifts. By parsing the timestamp embedded in each data packet and combining it with interpolation algorithms such as linear interpolation or cubic spline interpolation, the data synchronization submodule maps all data points onto a preset common time axis, ensuring the consistency of different modal data in the time dimension, thereby accurately reflecting the comprehensive physical state of the polyester yarn at a specific moment.

[0034] The noise cancellation submodule employs various digital filtering algorithms to denoise the raw data. For optical image data, median filtering, Gaussian filtering, or wavelet denoising methods can be used to eliminate image sensor noise or ambient light spots. For spectral data, Savitzky-Golay smoothing filtering or Fourier transform denoising can be used to remove random noise and enhance spectral features. For laser profile and tension data, moving average filters, Kalman filters, or adaptive filters can be applied to smooth high-frequency noise and reduce the impact of measurement fluctuations on data quality. The purpose of noise cancellation is to improve the signal-to-noise ratio, ensuring that subsequent analysis is based on cleaner and more reliable data.

[0035] The data calibration submodule performs linear or nonlinear calibration on sensor data using a pre-defined calibration model. This model, built upon historical calibration data, compensates for inherent sensor drift, aging effects, and measurement errors caused by environmental factors such as temperature and humidity. For example, a calibration lookup table or regression model is constructed by performing multiple measurements on a known standard sample and recording the deviations. For optical images, brightness, contrast, and geometric distortion corrections are available. For spectral data, baseline and scattering corrections are performed. The calibration process ensures the accuracy and comparability of the sensor output data.

[0036] The outlier detection and handling submodule identifies and processes outliers in the data using statistical methods or machine learning algorithms. Statistical methods, such as those based on the three-sigma criterion, interquartile range, or box plot analysis, identify data points that significantly deviate from the normal data distribution. Machine learning algorithms, such as those based on isolated forests, local outliers, or Gaussian mixture models, detect data with anomalous patterns. For identified outliers, various processing strategies can be employed, such as direct deletion, replacement with the mean or median of neighboring values, imputation through interpolation, or using more complex models for predictive imputation, to ensure the validity and reliability of the data and avoid negative impacts on subsequent model training.

[0037] The data normalization submodule scales the preprocessed data to a uniform numerical range to eliminate the influence between data of different dimensions. For example, the pixel values ​​of optical images are typically between 0 and 255, while tension data may range from 0 to thousands of Newtons. Common normalization methods include min-max normalization, which maps the data to the interval between 0 and 1, or Z-score normalization, which transforms the data into a distribution with a mean of 0 and a variance of 1. Normalization ensures that features of different types and dimensions have the same weight or influence in subsequent model training, preventing certain features with large numerical ranges from dominating the model training process, thereby accelerating model convergence and improving model performance.

[0038] The data preprocessing module sends the normalized data, which has undergone cleaning, synchronization, calibration, anomaly handling, and normalization, to the feature engineering and selection module via the internal data bus.

[0039] The feature engineering and selection module is responsible for extracting effective features highly correlated with the tensile strength of polyester yarn from the preprocessed multimodal data and optimizing these features to construct high-quality input vectors for use by the intelligent detection model. This module includes a multidimensional feature extraction submodule, a feature dimensionality reduction submodule, and a feature optimization submodule.

[0040] The multidimensional feature extraction submodule extracts representative features from various preprocessed data. From optical images, it extracts texture features such as gray-level co-occurrence matrix, local binary mode, or energy, entropy, and contrast calculated by wavelet transform; shape features such as diameter, roundness, and aspect ratio calculated through edge detection and region analysis; color features such as color mean and variance in RGB or HSV space; and brightness distribution features such as histogram statistics. From spectral data, it extracts spectral features such as absorbance at specific wavelengths, transmittance, spectral peak intensity, peak position, and peak width, which are directly related to the molecular structure and crystallinity of polyester yarn. From laser profile data, it extracts geometric features, including mean diameter, diameter variance, roundness deviation, cross-sectional area, and surface roughness parameters such as arithmetic mean roughness Ra. From tension data and environmental parameters, it extracts statistical features such as mean, variance, peak value, and slope in the time domain, or extracts frequency domain features through Fourier transform.

[0041] The feature dimensionality reduction submodule projects data from a high-dimensional feature space into a lower-dimensional space using algorithms such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), or manifold learning. The goal of this process is to reduce data redundancy, computational complexity, and mitigate overfitting, while preserving as much core information as possible relevant to tensile strength prediction. PCA transforms the original features into a set of linearly uncorrelated features called principal components through orthogonal transformations. LDA, on the other hand, reduces intra-class variability while preserving significant inter-class differences, making it suitable for supervised learning scenarios. Manifold learning, such as isometric feature mapping or local linear embedding, is suitable for discovering nonlinear data structures. For example, when applying PCA, for data containing... Raw data of each feature For example One sample, each sample have The dimensional features can be obtained by calculating their covariance matrix and performing eigenvalue decomposition. Each eigenvalue and its corresponding eigenvector. Before selection... The projection matrix is ​​composed of the eigenvectors corresponding to the largest eigenvalues. For example k dimensions are sufficient to project the original high-dimensional features onto k-dimensional space. k-dimensional space.

[0042] The dimensionality reduction formula based on principal component analysis is expressed as follows:

[0043] ;

[0044] in, This represents the original high-dimensional feature vector. This represents the principal component transformation matrix, whose column vectors are the selected principal eigenvectors. This represents the eigenvectors after dimensionality reduction. Through this transformation, the main variance of the data is preserved in the lower-dimensional features.

[0045] The feature optimization submodule selects the subset of features that contribute most to the tensile strength prediction using algorithms based on correlation, mutual information, or recursive feature elimination. Correlation analysis assesses the linear relationship between a single feature and the target variable, tensile strength. Mutual information measures the degree of non-linear dependence between a feature and the target variable. Recursive feature elimination is a model-based feature selection method that finds the optimal feature subset by iteratively training the model and removing the features that contribute the least. The optimization process aims to remove redundant and irrelevant features, further optimizing the model's input and improving the model's prediction accuracy and efficiency.

[0046] The feature engineering and selection module sends the optimized feature vectors to the intelligent detection model building and training module or the real-time prediction and evaluation module through a standardized data interface.

[0047] The intelligent detection model construction and training module is responsible for building and optimizing a polyester yarn tensile strength prediction model based on historically collected data and corresponding laboratory physical tensile strength test results. This module includes a model selection submodule, a model training submodule, a model evaluation and optimization submodule, and a model deployment submodule.

[0048] The model selection submodule selects a deep learning model architecture based on the type of multimodal feature data, such as images, sequences, numerical values, and the requirements of the prediction task. This architecture can include convolutional neural networks, recurrent neural networks, long short-term memory networks, or Transformer models, and a multimodal fusion network is employed based on the characteristics of the multimodal data. Convolutional neural networks are particularly suitable for processing spatially correlated feature data such as optical images and laser profiles, extracting spatial hierarchical features and local patterns through multiple convolutional kernels. Recurrent neural networks or long short-term memory networks excel at processing time-series features such as tension data and environmental parameters, capturing temporal dependencies and long-term memory. Transformer models, through their unique self-attention mechanism, can effectively model long-sequence dependencies, and are particularly suitable for handling complex interactions and global correlations in multimodal data. The multimodal fusion network effectively fuses features from different modalities, such as image features, spectral features, and geometric features, for example, through early fusion, late fusion, or intermediate fusion strategies, learning the complex correlations between them, thereby constructing a unified model with comprehensive predictive capabilities for the tensile strength of polyester yarn.

[0049] The model training submodule utilizes historically collected feature data and corresponding laboratory physical tensile strength test results as label data. This label data consists of accurately measured true tensile strength values. The model parameters are iteratively trained using backpropagation algorithms and optimizers such as Adam or SGD to minimize the error between predicted and true values. During training, the model continuously adjusts its internal weights and biases, gradually learning the complex nonlinear mapping relationship between features and tensile strength. For example, in regression tasks, mean squared error or mean absolute error is often used as the loss function to measure prediction error.

[0050] The mean squared error loss function is expressed as follows:

[0051] ;

[0052] in, This represents the value of the loss function. This represents the total number of samples. This represents the true tensile strength value, while This represents the tensile strength value predicted by the model. The optimizer updates the model parameters based on the gradient of the loss function until the model reaches the preset convergence condition on the training set.

[0053] The model evaluation and optimization submodule evaluates the model's performance using cross-validation, assessing prediction accuracy (e.g., root mean square error, mean absolute error, coefficient of determination R-squared), generalization ability, and stability. Based on the evaluation results, the model architecture (e.g., number of layers, number of neurons), hyperparameters (e.g., learning rate, batch size, regularization coefficient), and training strategies (e.g., learning rate scheduling, early stopping mechanism) are adjusted and optimized. This iterative process aims to improve the model's robustness, enabling it to maintain high prediction accuracy even on unseen new data.

[0054] The model deployment submodule encapsulates the trained and performance-compliant optimized model, for example, by converting it to ONNX or TensorRT format, and loads it into the real-time prediction and evaluation module for online inference. The deployment process ensures efficient model operation and low-latency response.

[0055] The real-time prediction and evaluation module is used to perform real-time, continuous, and intelligent prediction of tensile strength on polyester production lines, as well as to evaluate quality and generate early warnings. This module includes a data receiving interface, a predictive inference engine, a quality evaluation submodule, and an early warning generation submodule.

[0056] The data receiving interface receives real-time optimized feature vectors from the feature engineering and selection module. These feature vectors are the latest data, acquired in real-time, preprocessed, and feature-engineered, representing the overall physical state of the polyester yarn on the current production line. The data receiving interface employs a high-speed, low-latency communication protocol to ensure timely data transmission and meet the needs of real-time forecasting.

[0057] The predictive inference engine loads the trained model deployed by the intelligent detection model building and training module and performs high-speed inference calculations on the received feature vectors. Utilizing optimized models, such as GPU acceleration, the inference engine generates a predicted tensile strength value for the polyester yarn at the current moment, along with the corresponding prediction confidence score, with extremely low latency. The predicted value is a continuous numerical value representing the expected tensile strength of the polyester yarn, while the confidence score reflects the reliability of the prediction result. The inference engine is designed with concurrent processing capabilities in mind, enabling it to process feature data from multiple polyester yarns or multiple time points simultaneously.

[0058] The quality assessment submodule compares the predicted tensile strength value with preset quality standards or acceptable ranges to determine the current quality status of the polyester yarn. Quality standards are typically set by process engineers, including upper and lower limits for tensile strength, as well as more refined quality grade classifications. The quality assessment submodule compares the predicted value with these thresholds to determine whether the current product is qualified, meets high-quality standards, or poses a quality risk. Assessment results, such as "qualified," "unqualified," or "low warning," are transmitted to the warning generation submodule and the human-computer interaction and visualization module.

[0059] The early warning generation submodule immediately generates an early warning signal when the predicted tensile strength value is below the lower limit of mass, above the upper limit of mass, or exhibits significant fluctuations, such as multiple consecutive predicted values ​​deviating from the average level. The early warning signal includes not only the warning level (e.g., slight, moderate, severe), the time of occurrence, and the predicted intensity value, but also a potential causal analysis. This causal analysis is based on the model's analysis of feature contributions, indicating whether the intensity anomaly is caused by optical texture abnormalities, spectral composition changes, or tension fluctuations. The early warning signal is sent via the early warning interface to the human-computer interaction and visualization module for operator viewing, and simultaneously to the results feedback and process optimization module to trigger automatic adjustments.

[0060] The results feedback and process optimization module is responsible for closed-loop control of the polyester yarn production process parameters based on real-time prediction results, achieving adaptive optimization of the production process. This module includes a control strategy generation submodule, a parameter adjustment submodule, and an optimization effect evaluation submodule.

[0061] The control strategy generation submodule generates corresponding process parameter adjustment strategies based on real-time predicted tensile strength values, quality assessment results, and early warning information, combined with preset process rules and optimization objectives such as maximizing the yield rate and minimizing energy consumption. The strategy generation process employs advanced control algorithms, including proportional-integral-derivative (PID) control, model predictive control, and reinforcement learning control. PID adjusts based on error, error integral, and error rate of change. Model predictive control optimizes control actions by building a process model and predicting future states. Reinforcement learning control learns the optimal control strategy through interaction with the production environment. For example, when the predicted tensile strength is low, the control strategy generation submodule may, according to preset rules, suggest increasing the tensile strength ratio or adjusting the heat setting temperature to improve product strength.

[0062] The parameter adjustment submodule generates an adjustment strategy based on the control strategy and sends control commands to relevant actuators on the production line via an industrial control interface or a digital-to-analog converter. These actuators include, for example, the spinneret heating device of the spinning machine, the frequency converter speed controller of the drawing roller of the stretching machine, the tension controller of the winding machine, the airflow regulating valve of the cooling zone, and the heating element of the heat setting furnace. The parameter adjustment submodule precisely adjusts key process parameters such as spinning temperature, stretch ratio, winding tension, cooling rate, and heat setting temperature, achieving refined control of the production process. Control commands are transmitted securely and reliably in digital or analog signal form via industrial fieldbuses such as Modbus or OPCUA.

[0063] The optimization effect evaluation submodule continuously monitors the production line status after parameter adjustment and subsequent tensile strength prediction results to assess the effectiveness of the optimization strategy. By comparing the predicted tensile strength, yield rate, and energy consumption before and after parameter adjustment, it evaluates whether the current optimization strategy has achieved the expected results. Based on the evaluation results, if the optimization effect is unsatisfactory or new problems are introduced, the optimization effect evaluation submodule will provide feedback to the control strategy generation submodule and adaptively adjust its algorithm parameters, such as the controller's gain coefficient and learning rate, to achieve continuous learning and performance improvement of the system, thereby ensuring that the system can operate stably and continuously optimize in the long term.

[0064] The knowledge base and experience accumulation module is used to store, manage, and utilize the system's historical operational data, model versions, and optimization experience to support the system's continuous evolution. This module includes a historical data storage submodule, a model version management submodule, an optimization experience base submodule, and a knowledge reasoning submodule.

[0065] The historical data storage submodule is responsible for persistently storing the raw data acquired by the data acquisition module, the data processed by the data preprocessing module, the features extracted by the feature engineering and selection module, the training data and labels from the intelligent detection model building and training module, and the prediction results and quality assessment reports from the real-time prediction and evaluation module. Data storage utilizes distributed databases such as HDFS and Cassandra, or relational databases such as PostgreSQL, ensuring high availability and scalability. All data is accompanied by detailed timestamps and metadata for traceability and analysis.

[0066] The model version management submodule manages the different versions of the model generated by the intelligent detection model building and training module. This includes recording the model's training parameters, performance metrics such as accuracy, recall, F1 score, generalization ability, and deployment status such as whether it is online and which version is currently running. Version management ensures the traceability and reversibility of the model, facilitating rollback to a stable version or A / B testing of a new model in case of problems.

[0067] The experience base submodule stores feedback results and various control strategies, parameter adjustment records, and corresponding optimization effects generated by the process optimization module. These experiences include successful optimization cases, failed attempts and their root cause analysis, and optimal adjustment schemes under specific product or process conditions, forming reusable knowledge assets. The experience base supports structured storage and semantic retrieval.

[0068] The knowledge reasoning submodule utilizes artificial intelligence reasoning techniques such as rule-based reasoning, expert systems, ontology reasoning, or case-based reasoning to conduct in-depth analysis of historical data and experience. Its aim is to uncover potential process patterns, quality influencing factors, and complex nonlinear relationships between parameters. The reasoning results provide decision support for the iterative upgrading of the intelligent detection model, such as providing new feature combination suggestions, updating the model architecture, and improving process optimization strategies, such as discovering new control rules and optimizing control parameters, thereby achieving continuous intelligent evolution of the system.

[0069] The human-computer interaction and visualization module provides users with intuitive system status monitoring, data querying, result display, early warning information prompts, and parameter setting functions. Through a graphical user interface, this module displays the tensile strength curve, historical trends, quality distribution map, and key process parameters of the polyester yarn in real time, such as spinning temperature and draw ratio. Users can generate customized reports and export data through the interface, facilitating quality analysis and production management. The user interface provides multi-level information display, from a macro-level overview of the production line to micro-level details of individual polyester yarn data, meeting the needs of different users. Early warning information is displayed prominently in pop-up or highlighted formats, providing detailed warning content and suggested actions to help operators react quickly.

[0070] In summary, the intelligent polyester yarn tensile strength detection system for the production field proposed in this embodiment uses multi-modal sensors to collaboratively collect multi-dimensional data on polyester yarn, including optical, spectral, geometric, tension, and environmental data. The data preprocessing module cleans, synchronizes, calibrates, and normalizes heterogeneous data to form high-quality input. The feature engineering and selection module extracts and optimizes feature vectors highly correlated with tensile strength. The intelligent detection model construction and training module, based on a deep learning architecture, uses historical data to build a high-precision prediction model. The real-time prediction and evaluation module continuously predicts tensile strength, evaluates quality, and generates early warnings on the production line. The result feedback and process optimization module uses advanced control algorithms to perform closed-loop adjustment of process parameters based on the prediction results, achieving adaptive optimization of the production process. The knowledge base and experience accumulation module accumulates and infers all data, models, and optimization experience, supporting continuous learning and performance improvement of the system. The human-computer interaction and visualization module provides an intuitive monitoring and management interface. This system fundamentally solves the limitations of traditional testing methods, enabling non-contact, high-precision, real-time online intelligent testing of polyester yarn tensile strength and proactive optimization of the production process. It significantly improves product quality stability and production efficiency, bringing significant economic and social benefits to polyester yarn manufacturers.

Claims

1. An intelligent testing system for the tensile strength of polyester yarn using artificial intelligence in the production field, characterized in that, include: The data acquisition module is used to acquire multi-dimensional physical characteristic data of polyester yarn on the production line in real time; The data preprocessing module is used to clean, synchronize, calibrate, and perform preliminary processing on the raw data acquired by the data acquisition module; The feature engineering and selection module is used to extract effective features that are highly correlated with the tensile strength of polyester yarn from the preprocessed raw data and to optimize the features. The intelligent detection model building and training module is used to build and optimize the polyester yarn tensile strength prediction model based on historical data and physical detection results; The real-time prediction and evaluation module is used to perform real-time and continuous intelligent prediction, quality evaluation and early warning of the tensile strength of polyester yarn on the production line, thereby forming real-time prediction results. The result feedback and process optimization module is used to perform closed-loop control of the polyester yarn production process parameters based on the real-time prediction results, so as to achieve adaptive optimization of the production process. The knowledge base and experience accumulation module is used to store, manage, and utilize the system's historical operating data, model versions, and optimization experience. The data acquisition module includes: The optical imaging submodule is used to acquire surface images of polyester yarns using a high-speed linear array camera or area array camera to form optical images. These images are used to capture the diameter uniformity, surface defects, gloss, and texture features of the polyester yarns. The spectral analysis submodule is used to acquire information on the chemical composition and molecular structure of polyester yarn through a near-infrared spectrometer or a Raman spectrometer, forming spectral data. This information is directly related to the crystallinity, orientation, and molecular weight distribution of the polymer. The laser profile measurement submodule is used to acquire precise cross-sectional profile data and diameter variation data of polyester thread through a laser displacement sensor or a triangulation laser sensor, forming laser profile data. The precise cross-sectional profile data and diameter variation data reflect the geometric uniformity of the polyester thread. The tension detection submodule is used to acquire real-time tension data of polyester yarn during the production process through a high-precision tension sensor. The real-time tension data reflects the stability of the process conditions. The environmental parameter detection submodule is used to acquire temperature and humidity data of the production site through temperature and humidity sensors to form environmental parameters. The temperature and humidity data are used to correct sensor drift or as auxiliary input for the model. The intelligent detection model construction and training module includes: The model selection submodule is used to select a deep learning model architecture based on the data type of the feature data and the requirements of the prediction task. The deep learning model architecture includes a convolutional neural network, a recurrent neural network, a long short-term memory network, or a Transformer model. A multimodal fusion network is adopted according to the characteristics of multimodal data. The convolutional neural network is used to process image and contour features, the recurrent neural network or the long short-term memory network is used to process time series features, and the Transformer model achieves modeling of long sequence dependencies through a self-attention mechanism. The model training submodule is used to use historically collected feature data and corresponding laboratory physical tensile strength test results as label data, and to iteratively train the model parameters through backpropagation algorithm and optimizer Adam or SGD to minimize the error between the predicted value and the true value. The model evaluation and optimization submodule is used to evaluate the model's performance using cross-validation, including prediction accuracy, generalization ability, and stability, and to adjust and optimize the model architecture, hyperparameters, and training strategies based on the evaluation results; and The model deployment submodule is used to encapsulate and load trained and performance-compliant models into the real-time prediction and evaluation module for online inference.

2. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 1, characterized in that, The data preprocessing module includes: The data synchronization submodule is used to synchronize heterogeneous data from different sensors to a unified time base using a high-precision timestamp alignment algorithm, ensuring time consistency between data. The noise cancellation submodule is used to denoise the original data using digital filtering algorithms, such as median filtering, Gaussian filtering, or Savitzky-Golay smoothing filtering. The data calibration submodule is used to perform linear or nonlinear calibration on sensor data using a preset calibration model to compensate for the sensor’s inherent drift, aging effect and measurement errors caused by environmental factors. The outlier detection and processing submodule is used to identify and process outliers in the data using statistical methods or machine learning algorithms, based on the three sigma criterion, isolated forest, or local outliers, to ensure the validity and reliability of the data. The data normalization submodule is used to scale the preprocessed data to a uniform numerical range to eliminate the influence between data of different units, which facilitates the training and processing of subsequent models.

3. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 1, characterized in that, The feature engineering and selection module includes: A multi-dimensional feature extraction submodule is used to extract effective features that are highly correlated with the tensile strength of polyester yarn from the preprocessed data. The feature dimensionality reduction submodule is used to project data from a high-dimensional feature space to a lower-dimensional space using principal component analysis, linear discriminant analysis, or manifold learning algorithms, thereby reducing data redundancy while preserving core information; and The feature optimization submodule is used to select the subset of features that contribute the most to the prediction of tensile strength by using correlation, mutual information or recursive feature elimination algorithms, thereby optimizing the input of the model.

4. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 3, characterized in that, The multidimensional feature extraction submodule is specifically used for: Visual features, including texture features gray-level co-occurrence matrix, features calculated by local binary mode or wavelet transform, features obtained by shape feature edge detection and region analysis, color features, and brightness distribution features, are extracted from the optical image. Spectral features, including absorbance at a specific wavelength, transmittance, spectral peak intensity, peak position, and peak width, are extracted from the spectral data. These spectral features are associated with the molecular structure and crystallinity of the polyester yarn. Geometric features, including mean diameter, diameter variance, roundness deviation, and surface roughness, are extracted from the laser profile data. Statistical features are extracted from the real-time tension data and environmental parameters, including mean, variance, peak value, or slope.

5. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 1, characterized in that, The real-time prediction and evaluation module includes: A data receiving interface is used to receive real-time optimized feature vectors from the feature engineering and selection module; The predictive inference engine is used to load the training model deployed by the intelligent detection model construction and training module, and to perform high-speed inference calculations on the received feature vectors to generate the predicted tensile strength of the polyester thread at the current moment and the corresponding prediction confidence. The quality assessment submodule is used to determine the current quality status of the polyester yarn by comparing the predicted tensile strength value with a preset quality standard or acceptable range; and The early warning generation submodule is used to immediately generate an early warning signal when the predicted tensile strength value is lower than the lower limit of mass, higher than the upper limit of mass, or shows significant fluctuations, and send it to the human-computer interaction and visualization module and the result feedback and process optimization module through the early warning interface.

6. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 5, characterized in that, The warning generation submodule is used to generate the warning signal, which includes the warning level, occurrence time, predicted intensity value, and possible cause analysis. The possible cause analysis is based on the model's analysis of feature contribution to indicate whether the intensity abnormality is caused by optical texture anomaly, spectral composition change, or tension fluctuation.

7. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 1, characterized in that, The result feedback and process optimization module includes: The control strategy generation submodule is used to generate corresponding process parameter adjustment strategies based on the real-time predicted tensile strength value, quality assessment results and early warning information, combined with preset process rules and optimization objectives. The strategy generation process adopts advanced control algorithms including proportional-integral-derivative control, model predictive control or reinforcement learning control. The parameter adjustment submodule is used to generate an adjustment strategy output by the submodule based on the control strategy. This strategy is then sent to relevant actuators on the production line via an industrial control interface or a digital-to-analog converter to precisely adjust key process parameters, including spinning temperature, draw ratio, winding tension, cooling rate, and heat setting temperature. The optimization effect evaluation submodule is used to continuously monitor the production line status after parameter adjustment and the subsequent tensile strength prediction results, evaluate the effectiveness of the optimization strategy, and adaptively adjust the algorithm parameters of the control strategy generation submodule according to the evaluation results, so as to realize the continuous learning and performance improvement of the system.

8. The intelligent detection system for tensile strength of polyester yarn using artificial intelligence in the production field according to claim 1, characterized in that, The knowledge base and experience accumulation module includes: The historical data storage submodule is used to persistently store the raw data acquired by the data acquisition module, the data processed by the data preprocessing module, the features extracted by the feature engineering and selection module, the training data and labels of the intelligent detection model construction and training module, and the prediction results and quality assessment reports of the real-time prediction and evaluation module. The model version management submodule is used to manage different versions of the model generated by the intelligent detection model construction and training module, including the model's training parameters, performance metrics accuracy, generalization ability, and deployment status, to ensure the model's traceability and backtrackability. The optimization experience base submodule stores various control strategies, parameter adjustment records, and corresponding optimization effects generated by the result feedback and process optimization module, forming reusable knowledge assets; and The knowledge reasoning submodule is used to conduct in-depth analysis of the historical data and experience using artificial intelligence reasoning technology, rule-based reasoning, expert systems, or case-based reasoning, to uncover potential process patterns and quality influencing factors, and to provide decision support for the iterative upgrading of the intelligent detection model and the improvement of process optimization strategies.

Images