Textile jacquard detection system and method for distinguishing defects

Abstract

The invention relates to the technical field of textile detection, in particular to a textile jacquard detection system and method for distinguishing defects, and the method comprises the steps: collecting and recording yarn tension data and yarn vibration frequency, and collecting image data and multispectral imaging data; extracting pattern texture features in the image data, and detecting color consistency; performing texture feature extraction on the multispectral imaging data, and performing texture uniformity analysis; carrying out binarization processing on the image data, and detecting an edge gradient; extracting geometrical characteristics of the reed mark and the square eye, and carrying out vibration spectrum analysis; utilizing a YOLO algorithm to mark the position coordinates of the temple faults, and establishing a relation model of the temple pressure and the occurrence probability of the temple faults; and establishing a multi-modal data fusion model based on a neural network by combining the steps, and judging whether the jacquard fabric finished product has defects or not. According to the method, the internal relation of different types of data can be fully mined through the multi-modal data fusion model, the detection accuracy and robustness are improved, and the detection result is more reliable.

Description

Technical Field

[0001] The present invention relates to the technical field of textile detection, and particularly to a textile jacquard detection system and method for distinguishing defects. Background Art

[0002] The quality inspection of textile jacquard products is crucial for the textile industry. Traditional inspection methods mainly rely on manual labor, with low efficiency and the accuracy being easily affected by subjective factors. Although automated inspection has developed to some extent, the existing methods based on single image recognition or simple physical parameter monitoring are difficult to cope with the complex and diverse defect types in the textile jacquard production process. The textile jacquard production involves multiple links such as yarn, pattern weaving, and equipment technology. Any problem in any link may lead to product defects. For example, the breakage of yarn, warp breakage and weft breakage, the disorder of patterns, color deviation, as well as reed marks and temple defects caused by equipment technology. These defects not only affect the appearance of the product but also may reduce the service performance of the product.

[0003] The existing inspection methods have serious deficiencies in the comprehensiveness and correlation analysis of inspection parameters. Relying solely on a single inspection parameter, such as only detecting pattern defects through image recognition, cannot comprehensively consider the influence of yarn state, equipment operation parameters, etc. on product quality. At the same time, the detection accuracy and robustness for complex defects are poor. In the actual production environment, affected by interference factors such as light changes and equipment vibrations, the accuracy of the detection results is greatly reduced. There is a lack of effective multi-modal data fusion and dynamic adjustment mechanisms. When facing different production conditions and defect types, it cannot adaptively adjust the detection parameters and models, resulting in the detection efficiency and accuracy being unable to meet the requirements of high-quality and high-efficiency production in the modern textile industry. Summary of the Invention

[0004] The present invention significantly improves the detection accuracy and robustness by establishing a multi-modal data fusion model, so as to reliably detect textile jacquard defects even under complex working conditions, providing a more reliable quality inspection guarantee for production.

[0005] The technical solution proposed by the present invention is: a textile jacquard detection method for distinguishing defects, the method comprising: Real-time collecting and recording the yarn tension data and yarn vibration frequency during the textile process of the product through sensors, and collecting the image data and multi-spectral imaging data of the jacquard fabric finished product; Extracting the pattern texture features in the image data by using the LBP algorithm, and detecting the color consistency through color space conversion and color difference calculation; Extracting the texture features from the multi-spectral imaging data, and performing texture uniformity analysis through a Gabor filter; Performing binarization processing on the image data of the jacquard fabric finished product, and detecting the edge gradient through a Canny operator; The geometric features of reed marks and square eyes are extracted by using edge detection algorithms and Hough transform, and vibration signals are collected through an acceleration sensor and analyzed by Fourier transform for vibration spectrum analysis; The position coordinates of temple defects are marked by using the YOLO algorithm, and a relationship model between temple pressure and the occurrence probability of temple defects is established in combination with historical data; Based on the above steps, a multi-modal data fusion model is established based on a neural network, and the data collected is input to determine whether there are defects in the jacquard fabric finished product.

[0006] Preferably, the process of extracting the pattern texture features is as follows: The image data is grayscale processed to convert the color image into a grayscale image; Gaussian filtering is performed to remove the noise in the image; taking each pixel point as the center, a neighborhood is selected; the gray value of the neighborhood pixels is compared with the gray value of the center pixel. If the gray value of the neighborhood pixels is greater than the gray value of the center pixel, the corresponding position is recorded as 1, otherwise it is recorded as 0 to obtain an eight-bit binary number; the binary number is converted into a decimal number as the LBP value of the center pixel point; all pixel values are traversed to obtain the LBP feature map; the occurrence probability of different LBP values is calculated to generate the LBP feature histogram.

[0007] Preferably, the specific content of the texture uniformity analysis is as follows: The multi-spectral imaging data set is converted into a grayscale image; the grayscale image is convolved by a designed Gabor filter; the response values of each pixel point in Gabor filters with different parameters are calculated to obtain a set of texture feature vectors; the distribution of the texture feature vectors in the grayscale image is analyzed to judge the uniformity of the fabric surface texture. If the texture feature vectors in a certain area are significantly different from those in other areas, it is determined that there is an abnormal blending ratio in this area.

[0008] Preferably, the specific process of detecting the edge gradient is as follows: The image data is binarized to distinguish the warp and weft yarns of the fabric from the background; morphological operations are used to enhance the connectivity of the yarns; the binary image is traversed to calculate the number of intersections of warp and weft yarns per unit area; the image data is preprocessed by grayscale and Gaussian filtering; the Canny operator calculates the gradient amplitude and direction of the image, and high and low thresholds are set for edge extraction.

[0009] Preferably, the process of extracting the geometric features of reed marks and square eyes is as follows; The image data is processed by using a gradient-based edge detection algorithm to extract the edge information of the fabric surface; morphological thinning is performed on the edge image to make the edge lines clearer; the Hough transform is used to detect straight line segments, and the angles and spacings of the sawteeth are calculated through the angular and positional relationships of the straight line segments. If these geometric data parameters exceed the normal range, it is determined that there are reed mark and square eye defects.

[0010] Preferably, the process of establishing the relationship model between temple pressure and the occurrence probability of temple defects is as follows: Collect the temple pressure value data over a period of time, denoted as , and the occurrence of temple defects during the corresponding time; Represent whether the temple defects occur with 0 and 1, denoted as ; Taking the temple pressure value as the independent variable and the occurrence of temple defects as the dependent variable, use the linear regression analysis method to construct a mathematical model: ; Wherein: is the regression coefficient, is the intercept, is the random error term; Fit the collected data by the least squares method to calculate the values of and .

[0011] Preferably, the specific content of establishing the multi-modal data fusion model is as follows: Adopt the wavelet denoising method to denoise the yarn tension data and the yarn vibration frequency; Perform image preprocessing on the image data and the hyperspectral imaging data; Construct a fully connected neural network, the number of nodes in the input layer is determined according to the dimension of the sensor data, set 2-3 hidden layers, the number of nodes in each layer is adjusted between 32-128, use the ReLU activation function, the number of nodes in the output layer is 16, and output a 16-dimensional feature vector, representing the feature representation of the sensor data; Based on the VGG16 network, remove the last fully connected classification layer, the input image is processed through a series of convolutional layers and pooling layers to extract the high-level semantic features of the image, and obtain a 512-dimensional image feature vector; The fusion layer outputs a 128-dimensional final fusion feature vector Preferably, the multi-modal data fusion model has an adaptive dynamic model adjustment, and the adaptive dynamic model adjustment includes the following contents: Define the state space, action space and reward function. The state space consists of the feature statistical information of the current detection data, the accuracy rate and recall rate of the detection results, reflecting the current detection state; The action space includes adjusting the detection parameter priority and changing the model structure, using discrete integer coding; The reward function is designed based on the degree of improvement of the detection performance. Positive rewards are given for the improvement of the detection performance, and negative rewards are given otherwise.

[0012] The present invention also provides a textile jacquard detection system for distinguishing defects, and the system is used to execute the textile jacquard detection method for distinguishing defects described above.

[0013] The present invention also provides a computer-readable storage medium storing a computer program, which is executed by a processor to implement the textile jacquard detection method for distinguishing defects as described above.

[0014] Advantages of the present invention: Multi-source data are comprehensively collected. The collection of yarn-related data provides strong support for ensuring yarn quality and production stability by accurately monitoring vibration frequency, tension, testing breaking strength and elongation rate, and efficiently detecting broken warp / weft. For the collection of pattern and weaving data, with the help of advanced technologies such as image texture feature extraction, color consistency detection, multi-spectral imaging and texture uniformity analysis, various problems such as pattern defects, different yarns / different warps / different wefts, and cobwebs and thick / thin places can be accurately identified, effectively improving the appearance and internal quality of products. In terms of the collection of equipment and process data, by accurately extracting the geometric features of reed marks and square eyes, in-depth vibration spectrum analysis, accurate marking of temple defects and abnormal pressure judgment, equipment process problems can be discovered in time, product defects caused by equipment failures can be reduced, the smooth progress of the entire textile jacquard production process can be ensured, and production efficiency and product quality can be improved.

[0015] In the data preprocessing stage, operations such as denoising and normalization of sensor data, and graying, filtering, binarization and size normalization of image data effectively improve the data quality and usability, laying a good foundation for subsequent model processing. In the construction of the multi-modal data fusion model, the sensor data sub-model and the image data sub-model effectively extract the data features of their own, and the fusion layer cleverly splices and processes the feature vectors, fully exploring the internal connections between different data. After training with a large number of labeled samples, using the cross-entropy loss function and the Adam optimizer, the model can accurately learn the defect features, significantly improving the detection accuracy and robustness, so that textile jacquard defects can be reliably detected under complex working conditions, providing a more reliable quality detection guarantee for production.

[0016] Based on the reinforcement learning algorithm, the state space is scientifically defined, comprehensively reflecting the detection state by integrating the current detection data feature statistics, accuracy rate and recall rate, etc., providing rich basis for model decision-making. The action space is carefully defined, covering flexible operations such as adjusting the priority of detection parameters and changing the model structure, enabling the model to quickly make strategy adjustments according to the actual detection situation. The reward function is reasonably designed, and incentives are given according to the degree of improvement of detection performance, guiding the model to continuously learn and optimize, and developing in the direction of improving detection efficiency and accuracy. Through this series of adaptive dynamic adjustments, the detection system can better adapt to complex and changeable production working conditions and diverse defect types, continuously improve the detection effect, and provide a more efficient and intelligent quality detection service for textile jacquard production. Description of the Drawings

[0017] Figure 1 This is the flowchart of a textile jacquard detection system and method for differentiating defects according to the present invention; Figure 2 This is the flowchart of model adaptive adjustment of a textile jacquard detection system and method for differentiating defects according to the present invention. Specific embodiments

[0018] The following description is used to disclose the present invention so that those skilled in the art can implement the present invention. The preferred embodiments in the following description are only examples, and other obvious variations can be conceived by those skilled in the art. The basic principles defined in the following description can be applied to other implementation schemes, variant schemes, improvement schemes, equivalent schemes, and other technical schemes that do not deviate from the spirit and scope of the present invention.

[0019] It can be understood that the term "a" should be understood as "at least one" or "one or more". That is, in one embodiment, the number of an element can be one, while in other embodiments, the number of the element can be multiple. The term "a" cannot be understood as a limitation on the quantity.

[0020] As Figure 1 shown, a high-precision sensor is used to monitor the vibration frequency of the yarn in real time. Abnormal yarn breakage will cause high-frequency vibration. By capturing the change in vibration frequency, potential yarn breakage hazards can be detected in a timely manner. At the same time, a tension sensor is used to collect yarn tension data in real time. The tension will drop sharply at the moment of yarn breakage, providing a key basis for detecting yarn breakage. In addition, an electronic tensile testing machine is used to regularly test the breaking strength and elongation rate of the yarn during the stretching process to evaluate the yarn quality and provide a reference for subsequent production. For warp / weft breakage defects, in addition to monitoring the tension fluctuation (the tension signal mutates when the warp or weft yarn breaks), image recognition technology is also combined to accurately locate the breakage position. For example, the tension data is collected once per second, and when the tension drops suddenly by ≥10%, it is marked as a potential breakage event.

[0021] The sensor selected is a high-precision piezoelectric vibration sensor, which has high sensitivity and wide-frequency response characteristics and can accurately capture the subtle vibration changes of the yarn during operation. The sensor is tightly fixed beside the yarn running track through a special fixture. For example, it is set about 1 - 2 centimeters away from the yarn to ensure that the sensor can effectively receive the mechanical waves generated by the yarn vibration. The sensor converts the received vibration signal into an electrical signal and transmits it to the data acquisition card through a shielded cable. The data acquisition card collects the signal at a sampling frequency of 10kHz - 50kHz. Such a high sampling frequency can capture the high-frequency vibration characteristics generated at the moment of yarn breakage. The collected data is temporarily stored in the local cache waiting for subsequent processing.

[0022] Install a strain - type tension sensor on the yarn tension adjusting device (such as a tensiometer). By connecting the sensor to the key force - bearing components of the tension adjusting device, it can sense the change in yarn tension in real - time. During installation, it is necessary to ensure that the force - receiving direction of the sensor is consistent with the yarn tension direction to ensure the accuracy of measurement. When the yarn tension changes, the resistance value of the strain gauge of the tension sensor changes, and the resistance change is converted into a voltage signal through a Wheatstone bridge. This voltage signal is amplified by a signal amplifier and then collected by a data acquisition card at a frequency of 1 kHz - 5 kHz. The collected data is first stored in the local cache and synchronized with the vibration frequency data for subsequent processing. For example, the vibration data is sampled every 10 ms, and high - frequency acquisition (100 Hz) is started when the vibration amplitude exceeds the threshold, and the threshold can be set to 50 μm.

[0023] Regularly intercept a certain length (such as 20 - 50 cm) of yarn samples from the production line and install them on the fixture of an electronic tensile testing machine. Set the tensile speed of the tensile testing machine, generally 50 - 200 mm / min. This speed can not only ensure the accuracy of the test results but also simulate the tensile situation that the yarn may withstand in actual production. Start the tensile testing machine. During the process of stretching the yarn until it breaks, the sensors built into the testing machine collect data on the tensile force and the elongation of the yarn in real - time. The control system of the tensile testing machine records these data and transmits them to the computer through a data cable. The data analysis software calculates the breaking strength of the yarn based on the collected data. The calculation formula is: Breaking strength = Reading of the tensile force sensor at the time of breakage Elongation rate=(Length of the yarn at break - Initial length of the yarn) / Initial length of the yarn×100% The calculated breaking strength and elongation rate data are used to evaluate the quality of the yarn. If the breaking strength is lower than the standard value or the elongation rate exceeds the reasonable range, it may indicate that the yarn is prone to breakage and other problems in subsequent production, and it is necessary to adjust the yarn source or the production process.

[0024] After using a high - resolution camera to take an image of the jacquard fabric, first perform grayscale processing on the image to convert the color image into a grayscale image, simplifying subsequent calculations. Then perform Gaussian filtering to remove the noise in the image. For example, the size of the convolution kernel for Gaussian filtering is set to 3×3 or 5×5, and the standard deviation is adjusted between 0.5 - 2 according to the image noise situation.

[0025] The strain - type tension sensor is shared for tension fluctuation monitoring and yarn tension data acquisition. When warp breakage or weft breakage occurs, the tension of the corresponding yarn drops instantaneously, and the tension sensor captures this mutation signal. By setting a tension change threshold (for example, a 30% - 50% drop in the normal tension value), when the detected tension value is lower than the threshold, a warp / weft breakage warning is triggered.

[0026] The specific implementation method of image recognition assisted positioning is to install a high-resolution industrial camera at the key positions of the loom (such as the interweaving area of warp and weft yarns), and capture fabric images at a frame rate of 5-10 frames per second. The edge detection algorithm (such as the Canny algorithm) is used to process the captured images to identify the edges of the yarns. When the tension sensor triggers the warning of broken warp / weft, the system quickly analyzes the continuity of the yarn edges in the recent few frames of images. If it is found that the edge of a certain yarn is interrupted, combined with the image coordinate system, the position of the broken warp / weft can be accurately located, providing a basis for quick repair.

[0027] The specific implementation method of jacquard pattern defect monitoring is as follows: First, image texture feature extraction is carried out, which is implemented by using LBP in this scheme. After using a high-resolution camera to capture the jacquard fabric image, first perform grayscale processing on the image to convert the color image into a grayscale image, simplifying subsequent calculations. Then perform Gaussian filtering to remove the noise in the image. The convolution kernel size of Gaussian filtering is generally set to 3×3 or 5×5, and the standard deviation is adjusted between 0.5-2 according to the image noise situation. Then perform LBP feature calculation. For the preprocessed image, take each pixel point as the center and select a neighborhood (such as an 8-neighborhood). For the central pixel point, compare the gray value of its neighborhood pixels with the gray value of the central pixel. If the gray value of the neighborhood pixel is greater than the gray value of the central pixel, the corresponding position is recorded as 1, otherwise it is recorded as 0. In this way, an 8-bit binary number is obtained, and it is converted into a decimal number, which is the LBP value of this central pixel point. Traverse all pixel points in the image to obtain the LBP feature map of the entire image. By counting the occurrence frequencies of different LBP values in the LBP feature map, an LBP feature histogram is generated, which is used as the texture feature description of the image.

[0028] Next, color consistency detection is carried out. Color consistency detection includes three steps. The first step is image color space conversion. Convert the collected fabric image from the RGB color space to the LAB color space. The LAB color space is more in line with the human eye's perception of color and has more advantages in color difference calculation. The conversion formula is as follows: RGB to XYZ conversion: ; ; ; XYZ to LAB conversion: ; ; ; ; Where: , , are the tristimulus values under the standard illuminant.

[0029] The second step is color difference calculation. The design pattern image is also converted to the LAB color space, and then the color difference between the fabric image and the design pattern image in the LAB color space is calculated pixel by pixel. The calculation formula is as follows: ; Where: , , are the differences between the corresponding pixels of the fabric image and the design pattern image in the L, a, and b channels respectively.

[0030] The third step is color deviation judgment. Set the color difference threshold. When the calculated color difference exceeds the threshold, it is judged that there is a color deviation, that is, there may be problems such as different yarns / different warps. For example, set the color difference threshold to 5. When is calculated, it can be determined that there are problems such as different yarns / different warps in the textile.

[0031] Then perform different yarns / different warps / different wefts detection. Select a multispectral imaging device suitable for textile material detection to collect multispectral imaging data, which generally covers the visible light to near-infrared band (such as 400nm - 1000nm). Install the multispectral imaging device above the fabric, 30 - 50 cm away from the fabric, to ensure that the entire fabric area can be photographed. Before collecting the image, calibrate the device using a standard whiteboard and a blackboard for reflectivity and absorptivity calibration. During collection, the device obtains fabric images at different wavelengths at a certain wavelength interval (such as 10nm - 20nm) to form a multispectral image dataset.

[0032] After completing the multispectral data collection, perform texture uniformity analysis. This step is achieved through a Gabor filter. The Gabor filter is a band-pass filter, and its mathematical expression is: ; Where: , , is the wavelength, is the direction, is the standard deviation of the Gaussian envelope, is the aspect ratio, is the phase shift. Set a series of Gabor filters with different parameters according to the fabric texture characteristics. For example, varies between 4 - 16 pixels, takes 0°, 45°, 90°, 135°, is adjusted between 2 - 8 pixels, Take 0.5 - 1.5.

[0033] Perform texture feature extraction. After converting the multi - spectral image dataset into a grayscale image, perform a convolution operation on the image using the designed Gabor filter. For each pixel, calculate its response values under Gabor filters with different parameters to obtain a set of texture feature vectors. By analyzing the distribution of these texture feature vectors in the image, judge the uniformity of the fabric surface texture. If the texture feature vectors in some areas are significantly different from those in other areas, there may be an abnormal blending ratio caused by different yarns / different warp threads / different weft threads.

[0034] Then perform cobweb and thick / thin path detection. The texture density calculation process is as follows: Perform binarization on the collected fabric image to separate the warp and weft yarns of the fabric from the background. Then use morphological operations (such as dilation and erosion) to further enhance the connectivity of the yarns. By traversing the binarized image, calculate the number of intersections of warp and weft yarns per unit area (such as 1 square centimeter). For a normal fabric, the texture density is within a certain range. If the calculated texture density is significantly higher or lower than this range, it corresponds to cobweb or thin / thick density defects. After pre - processing the fabric image by grayscale conversion and Gaussian filtering, use the Canny operator for edge detection. The Canny operator extracts edges by calculating the gradient magnitude and direction of the image and setting high and low thresholds. When detecting the edge of the fabric surface, if an abnormal mutation area is found in the edge (such as discontinuous edge, too large or too small edge gradient), it may be a cobweb and thick / thin path defect caused by unclear shedding or malfunction of the weft detector. By analyzing the position and shape of the edge mutation area, the type and severity of the defect can be further determined.

[0035] In addition to the above content detection, reed mark and square eye detection and temple defect detection are also required. Reed mark and square eye detection are achieved through a gradient-based edge detection algorithm. The gradient-based edge detection algorithm (such as the Sobel operator) is used to process the fabric image to extract the edge information of the fabric surface. For reed mark and square eye defects, they appear as serrated marks in the image. By analyzing the geometric parameters of the serrated edges in the edge image, such as the angle and spacing of the serrations. Specifically, when implementing, first perform morphological thinning on the edge image to make the edge lines clearer, and then use the Hough transform to detect straight line segments. Calculate the angle and spacing of the serrations through the angle and position relationship of the straight line segments. If these geometric parameters exceed the normal range, it is determined that there are reed mark and square eye defects. Accelerometers are installed on key components of the loom (such as the reed and the loom main shaft) to collect the vibration signals during the operation of the loom at a sampling frequency of 1 kHz - 10 kHz. The collected vibration signals are transformed into the frequency domain through Fourier transform to obtain the vibration spectrum. A normally operating loom has specific vibration spectrum characteristics. If abnormal frequency components appear in the spectrum or the amplitudes of certain frequency components increase abnormally, it may correspond to problems such as reed deformation or tension imbalance, which may lead to reed mark and square eye defects. By establishing a normal loom vibration spectrum model and a defective loom vibration spectrum model, use pattern recognition algorithms (such as support vector machines) to classify the collected vibration spectra to determine whether the loom is abnormal.

[0036] The Sobel operator is used for edge detection. For each pixel in the image , the Sobel operator is applied for convolution calculation in the horizontal and vertical directions respectively. The Sobel operator in the horizontal direction is , and the Sobel operator in the vertical direction is . By calculating and with the convolution of the 3×3 neighborhood pixels centered on , the gradient approximations in the horizontal and vertical directions and are obtained. Then calculate the gradient magnitude of this pixel point: ; and the gradient direction: ; By performing such calculations on all pixels in the image, a preliminary edge image is obtained. Morphological thinning operations employ morphological operations such as erosion and dilation. Taking erosion as an example, an edge image is processed using a structuring element (such as a 3×3 square structuring element). For each pixel in the edge image, if the pixel and its neighborhood are all edge pixels (with a value of 1) under the coverage of the structuring element, then the pixel remains an edge pixel in the eroded image; otherwise, it becomes a background pixel (with a value of 0). The dilation operation is the opposite. By processing the image with the structuring element, the background pixels adjacent to the edge pixels are turned into edge pixels, thereby filling the holes and breaks in the edge image and making the edge lines clearer.

[0037] The edge image after morphological thinning is used for detecting line segments by the Hough transform. In the Hough space, a line can be represented by ($ $, $ $), where $ $ is the perpendicular distance from the origin to the line, and $ $ is the angle between the line and the $ $-axis. For each edge point ($ $, $ $) in the edge image, a sine curve is drawn in the Hough space: $ $; $ $ By performing such operations on all edge points, votes are accumulated in the Hough space, and the ($ $, $ $) pairs with the number of votes exceeding a certain threshold are found, and these pairs correspond to the line segments in the image. For the serrated edge formed by reed marks and square eye defects, the angle and spacing of the serrations are calculated based on the detected line segments. Suppose the angles of two adjacent detected line segments are $ $ and $ $ respectively, then the serration angle $ $. For the serration spacing, it is determined by calculating the positional relationship (such as the distance in the horizontal or vertical direction) of two adjacent line segments in the image. For example, if the projected distance of two adjacent line segments in the horizontal direction is $ $, then the serration spacing is approximately $ $. The calculated serration angle and spacing are compared with the corresponding standard ranges of normal fabrics. If they exceed the range, it is determined that there are reed marks and square eye defects.

[0038] After installing an acceleration sensor on the key components of the loom to collect vibration signals $ $, they are transformed into the frequency domain through the discrete Fourier transform (DFT). The formula for the discrete Fourier transform is $ $, where $ $ is the number of sampling points, $ $, $ . Through DFT calculation, the spectral values of the vibration signal at different frequency points are obtained to obtain the vibration spectrum .

[0039] Collect a large amount of vibration spectrum data during the normal operation of the loom, and calculate its statistical characteristics, such as the mean and the standard deviation to establish a vibration spectrum model for the normal loom. For the vibration spectrum data of the loom to be detected collected , calculate the difference between it and the normal spectrum model. For example, calculate the Euclidean distance: ; Use the support vector machine (SVM) for classification. Take the vibration spectrum data of the normal loom as one type of sample, and the vibration spectrum data of the loom with abnormal conditions such as reed marks and square eye defects as another type of sample to train the SVM. After training, input the vibration spectrum data to be detected into the trained SVM model, and the model outputs a judgment result, that is, whether there is an abnormality in the loom. If it is judged as abnormal, and combined with the abnormal geometric parameters of the sawtooth edge detected in the geometric feature extraction, it is further determined that there are reed marks and square eye defects

[0040] The temple defect detection is realized by using the object detection algorithm. With the help of the AI quality inspection system, use the object detection algorithm (such as the YOLO series algorithm based on deep learning) to monitor the cloth edge in real time. When training the YOLO model, collect a large number of cloth edge images with temple defect annotations and mark the position coordinates of the needle marks. The trained model can quickly identify the temple defects in the real-time collected cloth edge images and mark their grid coordinates in the images. Combine historical data to establish a relationship model between the temple pressure and the occurrence probability of temple defects. For example, by analyzing the temple pressure values and the corresponding temple defect occurrences in the past period of time, use methods such as regression analysis to establish a mathematical model. When the AI quality inspection system detects a temple defect, query the current temple pressure value, and judge whether the temple pressure is abnormal according to the established model. If the temple pressure exceeds the normal range and the frequency of temple defects is relatively high, it is prompted to adjust the temple pressure to reduce the generation of temple defects. The process of constructing the relationship model between temple defects and edge pressure is as follows: Collect the temple pressure value data within a period of time, denoted as , and the situation of temple defects occurring in the corresponding time. The occurrence of temple defects can be represented by 0 (not occurred) and 1 (occurred), denoted as . Taking the temple pressure value as the independent variable and the situation of temple defects occurring as the dependent variable, use the linear regression analysis method to construct a mathematical model. Assume the model is , where is the regression coefficient, is the intercept is the random error term. By using methods such as the least squares method to fit the collected data, calculate and values, so as to determine the relationship model between temple pressure and the occurrence probability of temple defects.

[0041] When the AI quality inspection system detects a temple defect ( 1), query the temple pressure value at this time . Substitute . into the established relationship model to calculate the theoretically predicted probability of temple defects . If is much greater than the occurrence probability of temple defects under normal conditions (the normal probability range can be determined according to historical data statistics), and the current temple pressure value exceeds the historical normal temple pressure range (assuming the normal range is , if ( < ) or ( > ), then it is judged that the temple pressure is abnormal. For example, through historical data statistics, the occurrence probability of temple defects under normal conditions is (5%), and the calculated value is ( = 30%), and at the same time ( > ), then it can be determined that the temple pressure is abnormal, and it is prompted to adjust the temple pressure to reduce the generation of temple defects After obtaining the above detection data, multi-modal data fusion is required. First, data preprocessing is carried out, including sensor data preprocessing and image data preprocessing. The specific content of sensor data preprocessing is as follows: For sensor data such as vibration frequency and tension, first perform denoising processing using the wavelet denoising method. Decompose the sensor data by wavelet, and according to the distribution characteristics of noise in the wavelet coefficients, set an appropriate threshold to process the wavelet coefficients, remove the noise components, and then perform wavelet reconstruction to obtain the denoised data. Then perform normalization processing to map the data to the [0,1] interval. The formula is: ; where: is the original data, and are the minimum and maximum values of this group of data respectively.

[0042] The specific content of image data preprocessing is as follows: For the collected image data, in addition to the operations such as grayscale conversion, filtering, and binarization mentioned above, size normalization is also performed. Images with different resolutions are uniformly adjusted to the same size (such as 256×256 pixels), and the bilinear interpolation algorithm is used for image scaling to meet the input requirements of the subsequent model.

[0043] Construct a multi-modal data fusion model based on a neural network fusion model. Build a sensor sub-model: Construct a fully connected neural network, and the number of input layer nodes is determined according to the dimension of the sensor data (such as the dimension after splicing data such as vibration frequency and tension). Set 2 - 3 hidden layers, and the number of nodes in each layer is adjusted between 32 and 128. The ReLU activation function is used, and the specific formula is: ; The number of output layer nodes is 16, and a 16-dimensional feature vector is output, representing the feature representation of the sensor data. The data input to the input layer includes yarn data, image data, and equipment data. The yarn data includes vibration frequency, tension, and temperature. The image data includes LBP data. The equipment data includes vibration frequency, temple pressure, and reed mark angle.

[0044] For image data, a convolutional neural network (CNN) is used. Based on the VGG16 network, the last fully connected classification layer is removed. The input image is processed through a series of convolutional layers and pooling layers to extract the high-level semantic features of the image. The convolutional layer uses a 3×3 convolutional kernel, the stride is 1, and the padding is 1. The pooling layer uses 2×2 max pooling. Finally, a 512-dimensional image feature vector is obtained.

[0045] The design of the fusion layer is to splice the 16-dimensional feature vector output by the sensor data sub-model and the 512-dimensional feature vector output by the image data sub-model to obtain a 528-dimensional fusion feature vector. Then, a fully connected layer is used to further process the fusion feature vector. The number of nodes in the fully connected layer is 128, and the ReLU activation function is used to output a 128-dimensional final fusion feature vector. The correlation matrix between modalities is calculated as follows: ; Among them: is the cosine similarity, 、 are different modality data respectively.

[0046] The fusion model is trained using a large number of multi-modal data samples with defective annotations. The loss function uses the cross-entropy loss function, the optimizer selects the Adam optimizer, and the learning rate is set to 0.001 - 0.0001. During the training process, the model parameters are continuously adjusted to enable the model to accurately learn defective features from the fused data and improve the robustness of detection under complex working conditions. The calculation formula of the loss function for training is as follows: ; where: is the cross-entropy loss, is the Dice coefficient, balances classification and localization.

[0047] As Figure 2 shown, the model also has adaptive dynamic model adjustment, which is achieved through a reinforcement learning algorithm. The state space consists of the feature statistics of the current detection data, the accuracy and recall rate of the detection results, etc. For example, the occurrence frequencies of different types of defects in the current batch of detection data, the mean and standard deviation of various sensor data, the statistics of image features, etc. are used as part of the state space. Represented by a vector as where are different feature statistics. The action space includes operations such as adjusting the priority of detection parameters and changing the model structure. For example, for adjusting the priority of detection parameters, the action can be to increase the weight of the yarn vibration frequency parameter by 10%, or to adjust a certain parameter (such as the LBP neighborhood size) of the image texture feature extraction algorithm. For changing the model structure, the action can be to increase or decrease the number of nodes in a certain layer of the neural network. Different actions are represented by discrete integer encodings, such as 0 representing increasing the weight of the yarn vibration frequency parameter, and 1 representing changing the LBP frequency domain size, etc. The reward function is defined according to the improvement degree of detection performance. The specific calculation formula of the reward function is as follows: ; where: , , motivating the model to focus on high-priority defects. After each batch of detections, the accuracy and recall rate are statistically calculated, the reward function is calculated. If the reward improvement in five consecutive batches is greater than 5%, the model weight update is triggered. If the misdetection rate of a certain defect type is greater than 15%, the priority of its detection parameters is increased by one level.

[0048] Embodiments disclosed in the present invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. Embodiments disclosed in the present invention include a computer program product, which includes a computer program carried on a computer-readable medium, and the computer program contains program codes for executing the methods shown in the flowcharts. In such an embodiment, the computer program can be downloaded and installed from the network through a communication part, and / or installed from a removable medium. When the computer program is executed by a central processing unit (CPU), the above-mentioned functions defined in the methods of the present application are executed. It should be noted that the computer-readable medium in the present application can be a computer-readable signal medium, a computer-readable storage medium, or any combination of the two. A computer-readable storage medium can be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination of the above. More specific examples of a computer-readable storage medium can include, but are not limited to: an electrical connection with one or more wire segments, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above. In the present application, a computer-readable storage medium can be any tangible medium that contains or stores a program, and the program can be used by or combined with an instruction execution system, apparatus, or device. In the present application, a computer-readable signal medium can include a data signal propagated in a baseband or as part of a carrier wave, which carries computer-readable program codes. Such a propagated data signal can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination of the above. A computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, and the computer-readable medium can send, propagate, or transmit a program for use by or combined with an instruction execution system, apparatus, or device. The program codes contained on the computer-readable medium can be transmitted by any appropriate medium, including but not limited to: wireless segments, wire segments, optical cables, RF, etc., or any suitable combination of the above.

[0049] The flowcharts and block diagrams in the accompanying drawings illustrate the possible architectures, functions, and operations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in the flowchart or block diagram may represent a module, a segment of a program, or a portion of code that contains one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions noted in the blocks may occur in a different order than that noted in the accompanying drawings. For example, two consecutive blocks shown may actually be executed substantially in parallel, or they may sometimes be executed in the reverse order, depending on the functions involved. It should also be noted that each block in the block diagram and / or flowchart, and combinations of blocks in the block diagram and / or flowchart, may be implemented by a dedicated hardware-based system that performs the specified functions or operations, or may be implemented by a combination of dedicated hardware and computer instructions.

[0050] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are only examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been demonstrated and illustrated in the embodiments. Without departing from the said principles, the embodiments of the present invention may have any variations or modifications.

Claims

1. A textile jacquard detection method for distinguishing defects, characterized in that, The method includes: Collecting and recording in real time the yarn tension data and yarn vibration frequency during the textile process of the product by sensors, and collecting the image data and multi-spectral imaging data of the jacquard fabric finished product; Extracting the pattern texture features in the image data by using the LBP algorithm, and detecting the color consistency through color space conversion and color difference calculation; Extracting texture features from the multi-spectral imaging data, and performing texture uniformity analysis through Gabor filters; Performing binary processing on the image data of the jacquard fabric finished product, and detecting the edge gradient through the Canny operator; Adopting an edge detection algorithm and Hough transform to extract the geometric features of reed marks and square eyes, and collecting vibration signals through an acceleration sensor for vibration spectrum analysis through Fourier transform; Using the YOLO algorithm to label the position coordinates of temple defects, and establishing a relationship model between temple pressure and the occurrence probability of temple defects in combination with historical data; Based on the above steps, establishing a multi-modal data fusion model based on a neural network, and inputting the collected data to determine whether there are defects in the jacquard fabric finished product.

2. The textile jacquard detection method for distinguishing defects according to claim 1, characterized in that, The process of extracting the pattern texture features is as follows: Performing grayscale processing on the image data to convert the color image into a grayscale image; performing Gaussian filtering to remove the noise in the image; taking each pixel point as the center and selecting a neighborhood; comparing the gray values of the neighborhood pixels with the gray value of the central pixel; If the gray value of the neighborhood pixel is greater than the gray value of the central pixel, the corresponding position is recorded as 1, otherwise it is recorded as 0, obtaining an eight-bit binary number; Converting the binary number into a decimal number as the LBP value of the central pixel point; Traversing all pixel values to obtain an LBP feature map; calculating the occurrence probability of different LBP values to generate an LBP feature histogram.

3. The textile jacquard detection method for distinguishing defects according to claim 2, wherein, The specific content of the texture uniformity analysis is as follows: Converting the multi-spectral imaging data set into a grayscale image; performing convolution operation on the grayscale image through a designed Gabor filter; calculating the response values of each pixel point in different parameter Gabor filters to obtain a set of texture feature vectors; analyzing the distribution of the texture feature vectors in the grayscale image to judge the uniformity of the fabric surface texture. If the texture feature vectors in a certain area are significantly different from those in other areas, it is determined that there is an abnormal blending ratio in that area.

4. A textile jacquard detection method for distinguishing defects according to claim 3, characterized in that The specific process of detecting the edge gradient is as follows: Performing binary processing on the image data to separate the warp and weft yarns of the fabric from the background; using morphological operations to enhance the connectivity of the yarns; traversing the binary image and calculating the number of intersections of warp and weft yarns per unit area; Performing preprocessing of grayscale and Gaussian filtering on the image data; The Canny operator calculates the gradient amplitude and direction of the image, and sets high and low thresholds for edge extraction.

5. The textile jacquard detection method for distinguishing defects according to claim 4, characterized in that, The process of extracting the geometric features of reed marks and square eyes is as follows; Processing the image data by using an edge detection algorithm based on gradient to extract the edge information of the fabric surface; performing morphological thinning on the edge image to make the edge lines clearer; using the Hough transform to detect straight line segments, and calculating the angle and spacing of the sawteeth through the angle and position relationship of the straight line segments. If these geometric data parameters exceed the normal range, it is judged that there are reed mark and square eye defects.

6. A textile jacquard detection method for distinguishing defects according to claim 5, characterized in that The establishment process of the relationship model between temple pressure and the occurrence probability of temple defects is as follows: Collect the data of temple pressure values over a period of time, denoted as , and the occurrence of temple defects during the corresponding time; Whether or not the temple defect occurs is represented by 0 and 1, denoted as ; With the temple pressure value as the independent variable and the occurrence situation of the temple defect as the dependent variable, a mathematical model is constructed using the linear regression analysis method: ； Wherein: is the regression coefficient, is the intercept, is the random error term; The collected data is fitted by the least squares method to calculate the and values.

7. A textile jacquard detection method for distinguishing defects according to claim 6, characterized in that The specific content of the establishment of the multi-modal data fusion model is as follows: The wavelet denoising method is used to denoise the yarn tension data and the yarn vibration frequency; image preprocessing is performed on the image data and the hyperspectral imaging data; a fully connected neural network is constructed. The number of nodes in the input layer is determined according to the dimension of the sensor data. 2-3 hidden layers are set, and the number of nodes in each layer is adjusted between 32 and 128. The ReLU activation function is used, and the number of nodes in the output layer is 16, outputting a 16-dimensional feature vector, representing the feature representation of the sensor data; Based on the VGG16 network, the last fully connected classification layer is removed. The input image is processed through a series of convolutional layers and pooling layers to extract the high-level semantic features of the image, obtaining a 512-dimensional image feature vector; the fusion layer outputs a 128-dimensional final fusion feature vector.

8. A textile jacquard detection method for differentiating defects according to claim 7, characterized in that The multi-modal data fusion model has adaptive dynamic model adjustment, and the adaptive dynamic model adjustment includes the following content: The state space, action space, and reward function are defined. The state space consists of the feature statistical information of the current detection data, the accuracy rate and recall rate of the detection results, reflecting the current detection state; the action space includes adjusting the priority of the detection parameters and changing the model structure, using discrete integer coding; The reward function is designed based on the degree of improvement in detection performance. Positive rewards are given for improved detection performance, and negative rewards are given otherwise.

9. A textile jacquard detection system for distinguishing defects, characterized in that, The system is used to execute a textile jacquard detection method for distinguishing defects according to any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, and the computer program is executed by a processor to implement a textile jacquard detection method for distinguishing defects according to any one of claims 1-8 above.

Images