A yarn quality prediction system and method

Abstract

The application discloses a yarn quality prediction system, relates to the fields of computer vision and textile technology, and comprises a data acquisition module, a data storage module, a data preprocessing module, a defect detection module, a mathematical model module and a comprehensive quality prediction module. The application also discloses a yarn quality prediction method, which comprises the following steps: S100, preparation, S200, yarn conveying, S300, yarn original data acquisition, S400, yarn original data storage, S500, yarn original data preprocessing, S600, yarn detection result acquisition, S700, yarn quality index calculation, and S800, yarn quality comprehensive prediction. The application guarantees the effectiveness of the yarn quality prediction result and improves the accuracy of the yarn comprehensive quality prediction result.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of computer vision and textile technology, and in particular to a yarn quality prediction system and method. BACKGROUND

[0002] In recent years, with the substantial increase in demand for the production capacity of the wool textile industry and the increasing demand for high-quality products, the process of industrial intelligence has gradually accelerated, and the use of sensor equipment for yarn quality detection has also begun to be popular. The most traditional yarn quality detection uses a blackboard detector to assist manual detection, which is processed by observing and referring to the quality evaluation standard with the human eye. This method greatly consumes human resources, and the detection results are easily affected by subjective factors. At present, most yarn quality detection uses capacitive sensors or photoelectric sensors to detect yarn diameter using electrical signals and light energy signals, which can ensure large-scale production of yarn, but has certain limitations in quality detection accuracy and cannot meet the demand for high-quality wool yarn products.

[0003] With the development of artificial intelligence, deep learning methods have also been applied to the wool textile field. The yarn quality prediction method based on deep learning can handle more complex scenarios and has no special requirements for the hairiness in the yarn, making it suitable for most yarn quality prediction work.

[0004] The patent "Device and method for yarn quality monitoring" (application number: CN201880088432.9) discloses a device and method for detecting the quality of fabric yarn, which directly irradiates an image sensor with a primary light source through a focusing element to obtain a fabric yarn image, and then judges the yarn quality. The defects of this patent are: lack of advanced computer vision algorithms, using open-loop image processing method to judge yarn defects, ignoring some features affecting yarn quality, low yarn diameter measurement accuracy, and may lead to misjudgment due to complex yarn conditions.

[0005] The patent "Fabric weaving online defect detection method based on linear laser" (application number: CN201910034418.7) discloses a fabric weaving online defect detection method based on linear laser, which uses a two-dimensional spectrum formed by the bright spots of laser and yarn interweaving for neural network-based defect detection. The defects of this patent are: the bright spots obtained by using laser are approximately binary images, ignoring some features of fabric yarn, and only detecting defects caused by fabric strip diameter and deformation, making it difficult to obtain comprehensive indicators of yarn quality.

[0006] Therefore, the skilled in the art is committed to developing a yarn quality prediction system and method. SUMMARY

[0007] In view of the above defects of the prior art, the technical problem to be solved by the present application is how to overcome the problems of difficulty in predicting yarn quality, defect misjudgment and uneven color in the prior art.

[0008] The inventors propose a target detection method for strengthening feature information, introduce deformable convolution to improve the feature extraction capability of the convolutional neural network, use a feature fusion network composed of a feature pyramid network (FPN) and a balanced feature pyramid (BFP) to fuse the feature information of the collected real-time yarn images, and the real-time yarn images are single-target or multi-target images. The inventors use a linear array camera and various sensor devices to collect yarn original data, obtain yarn quality indicators through mathematical models, and predict the yarn quality by comprehensively considering various factors.

[0009] In one embodiment of the present application, a yarn quality prediction system is provided, comprising:

[0010] a data acquisition module for acquiring yarn original data;

[0011] a data storage module for storing the yarn original data;

[0012] a data preprocessing module for preprocessing the yarn original data;

[0013] a defect detection module for constructing a deep learning model, using the deep learning model for defect processing, and generating a yarn detection result;

[0014] a mathematical model module for calculating yarn quality indicators using the yarn original data;

[0015] a comprehensive quality prediction module for predicting the yarn quality by comprehensively considering the yarn detection result and the yarn quality indicators;

[0016] The data acquisition module, the data storage module and the data preprocessing module are sequentially communicatively connected, the data preprocessing module is communicatively connected with the defect detection module and the mathematical model module, and the comprehensive quality prediction module is communicatively connected with the defect detection module and the mathematical model module.

[0017] In response to the yarn delivery of the automatic doffing machine, the data acquisition module acquires the yarn original data, sends the yarn original data to the data storage module for storage, and then sends the yarn original data to the data preprocessing module for preprocessing, the preprocessed yarn original data is subjected to the deep learning model constructed by the defect detection module to obtain the yarn detection result, the mathematical model module calculates the yarn quality indicators, and the comprehensive quality prediction module predicts the yarn quality according to the yarn detection result and the yarn quality indicators.

[0018] Optionally, in the yarn quality prediction system in the above embodiment, the data acquisition module comprises a linear array camera and a sensor.

[0019] Further, in the yarn quality prediction system in the above embodiment, the yarn raw data comprises real-time yarn images and sensor data.

[0020] Further, in the yarn quality prediction system in the above embodiment, the sensor data comprises yarn linear density, yarn twist and yarn tensile data.

[0021] Further, in the yarn quality prediction system in the above embodiment, the sensor comprises a linear density meter, a twist meter and a strength tester.

[0022] Further, in the yarn quality prediction system in the above embodiment, the linear array camera shoots real-time yarn images, the linear density meter collects yarn linear density, the twist meter collects yarn twist and the strength tester collects yarn tensile data.

[0023] Further, in the yarn quality prediction system in the above embodiment, the linear array image sensor of the linear array camera is based on CCD (Charge Coupled Device).

[0024] Optionally, in the yarn quality prediction system in any of the above embodiments, the preprocessing comprises real-time yarn image preprocessing and sensor data preprocessing.

[0025] Further, in the yarn quality prediction system in the above embodiment, the real-time yarn image preprocessing comprises image denoising, image enhancement, determining defect range, semi-automatic labeling of true value box and defect category, generating training set and verification set according to a set proportion, and normalizing yarn detection results.

[0026] Further, in the yarn quality prediction system in the above embodiment, the determination of defect range adopts morphological method and region growing algorithm.

[0027] Further, in the yarn quality prediction system in the above embodiment, the set proportion is preferably 8:2.

[0028] Further, in the yarn quality prediction system in the above embodiment, the yarn detection results comprise yarn linear density and color unevenness indexes derived from defect detection results.

[0029] Further, in the yarn quality prediction system in the above embodiment, the sensor data preprocessing comprises data extraction and removal of error data with excessively large deviation.

[0030] Optionally, in the yarn quality prediction system in any of the above embodiments, constructing the deep learning model comprises: inputting the training set and the validation set into a convolutional neural network for feature extraction, into a region generation network, and continuing the training of the yarn target detection network through a classification branch; through a regression branch, the candidate region is moved closer to the real region; finally, through back propagation of a loss function and weight update, the optimal target detection result is obtained.

[0031] Optionally, in the yarn quality prediction system in any of the above embodiments, the yarn quality indicators include yarn strip unevenness and yarn tensile property.

[0032] Further, in the yarn quality prediction system in the above embodiments, the yarn strip unevenness calculation uses the Martindale formula as follows:

[0033]

[0034]

[0035] wherein CV is the yarn strip unevenness, C D is the diameter dispersion coefficient of the wool fiber, c is the yarn linear density, d is the diameter of the wool fiber, and n is an intermediate variable.

[0036] Further, in the yarn quality prediction system in the above embodiments, the yarn tensile property calculation uses the Bearale formula as follows:

[0037]

[0038] wherein T represents the yarn tensile property, t w is the yarn twist.

[0039] Optionally, in the yarn quality prediction system in any of the above embodiments, predicting the yarn quality comprises predicting the yarn defect type and quantity, the yarn strip unevenness, and the yarn tensile property.

[0040] Based on any of the above embodiments, another embodiment of the present application provides a yarn quality prediction method, comprising the following steps:

[0041] S100, preparation, a data acquisition module acquires yarn original data, sends the yarn original data to a data storage module for storage, and then sends the yarn original data to a data preprocessing module to generate a training set and a validation set, and trains a deep learning model constructed by a defect detection module;

[0042] S200, yarn conveying, an automatic doffing machine conveys the yarn;

[0043] S300, yarn original data acquisition, a data acquisition module acquires yarn original data, and sends the yarn original data to a data storage module;

[0044] S400, yarn original data storage, the data storage module receives and stores the yarn original data, and sends to the data preprocessing module;

[0045] S500, yarn original data preprocessing, the data preprocessing module pre-processes the yarn original data, and sends to the defect detection module and the mathematical model module respectively;

[0046] S600, yarn detection result obtaining, in response to the pre-processed yarn original data, the defect detection module obtains the yarn detection result through the deep learning model, and sends to the comprehensive quality prediction module;

[0047] S700, yarn quality index calculation, in response to the pre-processed yarn original data, the mathematical model module calculates the yarn quality index, and sends to the comprehensive quality prediction module;

[0048] S800, yarn quality comprehensive prediction, the comprehensive quality prediction module calculates the comprehensive yarn quality prediction result by giving weights to the yarn detection result and the yarn quality index.

[0049] Optionally, in the yarn quality prediction method in the above embodiment, the data acquisition module comprises a line array camera and a sensor.

[0050] Optionally, in the yarn quality prediction method in any of the above embodiments, the yarn original data comprises real-time yarn images and sensor data.

[0051] Further, in the yarn quality prediction method in the above embodiment, the sensor data comprises yarn linear density, yarn twist and yarn tensile data.

[0052] Optionally, in the yarn quality prediction method in the above embodiment, step S100 comprises:

[0053] S110, training set and validation set generation, the data acquisition module acquires yarn original data, sends to the data storage module for storage, and then sends to the data preprocessing module for preprocessing to generate a training set and a validation set;

[0054] S120, deep learning model training, the defect detection module constructs a deep learning model, inputs the training set and the validation set for training, and when the loss function of the target detection result is the minimum, the deep learning model training is completed.

[0055] Optionally, in the yarn quality prediction method in any of the above embodiments, the ratio of the training set and the validation set in step S110 is 8:2.

[0056] Optionally, in the yarn quality prediction method in any of the above embodiments, the constructing the deep learning model in step S120 comprises: inputting the training set and the validation set into a convolutional neural network for feature extraction, entering a region generation network, and continuing the training of the yarn target detection network through a classification branch; the candidate regions are brought closer to the real regions through a regression branch; and finally, the optimal target detection result is obtained through the back propagation of the loss function and the weight update.

[0057] Optionally, in the yarn quality prediction method in any of the above embodiments, step S120 comprises:

[0058] S121, constructing a deep learning model network, setting initial values of network training hyperparameters, taking Cascade RCNN as the yarn target detection network, taking ResNet50 as the feature extraction backbone network, the yarn target detection network further comprising a region generation network, a classification regression network and a feature fusion network, obtaining a feature image from the feature extraction backbone network, performing feature fusion and enhancement through the feature fusion network, and performing defect positioning and identification through the region generation network and the classification regression network;

[0059] S122, improving ResNet50, replacing the original convolution kernel in ResNet50 with a deformable convolution, dynamically adjusting according to different targets to be detected, so that the extracted features are concentrated on the target, and the feature extraction capability of the feature extraction backbone network for defects with complex shapes is enhanced;

[0060] S123, extracting yarn feature information of real-time yarn images, adding FPN and BFP for the Cascade RCNN network, balancing and fusing features of different levels, and extracting yarn feature information of real-time yarn images using the feature extraction backbone network;

[0061] S124, training the deep learning model, calculating the error loss of the yarn target detection network, and updating the weight parameters to minimize the error of the deep learning model;

[0062] S125, completing the training of the deep learning model, repeating steps S121 to S124 for not less than 200 times, and completing the training of the deep learning model when the loss function of the target detection result is minimized.

[0063] Further, in the yarn quality prediction method in the above embodiments, the hyperparameters include the number of iterations, the learning rate, Batch, and Epoch.

[0064] Optionally, in the yarn quality prediction method in any of the above embodiments, the preprocessing comprises real-time yarn image preprocessing and sensor data preprocessing.

[0065] Further, in the yarn quality prediction method in the above embodiment, the real-time yarn image preprocessing includes image denoising, image enhancement, defect range determination, semi-automatic labeling of true value box and defect category, training set and validation set generation according to a set proportion, and yarn detection result normalization processing.

[0066] Further, in the yarn quality prediction method in the above embodiment, the defect range determination adopts a morphological method and a region growing algorithm.

[0067] Further, in the yarn quality prediction method in the above embodiment, the set proportion is preferably 8:2.

[0068] Further, in the yarn quality prediction method in the above embodiment, the sensor data preprocessing includes data extraction and removal of error data with excessively large deviation.

[0069] Optionally, in the yarn quality prediction method in any of the above embodiments, the step S300 includes:

[0070] S310, a line array camera shoots a real-time yarn image,

[0071] S320, a yarn evenness instrument collects yarn evenness, a twist meter collects yarn twist, and a strength tester collects yarn tensile data.

[0072] S330, a data acquisition module sends the real-time yarn image and the yarn evenness, yarn twist and yarn tensile data to a data storage module.

[0073] Optionally, in the yarn quality prediction method in any of the above embodiments, the weight proportion of the yarn detection result and the yarn quality index in the step S800 is 1:1.

[0074] The present application introduces a deep learning model into yarn quality prediction, realizes yarn comprehensive quality prediction, improves target detection result precision, guarantees the effectiveness of the yarn quality prediction result, and improves yarn comprehensive quality prediction result accuracy.

[0075] The concept, specific structure and generated technical effects of the present application will be further described below with reference to the accompanying drawings, so as to fully understand the purpose, features and effects of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0076] Figure 1 is a structural schematic diagram of a yarn quality prediction system according to an exemplary embodiment;

[0077] Figure 2 is a flowchart of a yarn quality prediction method according to an exemplary embodiment;

[0078] Figure 3is a diagram illustrating a deep learning model training flowchart according to an exemplary embodiment. DETAILED DESCRIPTION

[0079] The technical content of the present application will be more clearly understood and facilitated to be understood by referring to the following description of the drawings, which introduces a plurality of preferred embodiments of the present application. The present application can be embodied in many different forms of embodiments, and the protection scope of the present application is not limited to the embodiments mentioned herein.

[0080] In the drawings, components of the same structure are denoted by the same reference numerals, and components having similar structures or functions are denoted by similar reference numerals. The size and thickness of each component shown in the drawings are arbitrarily shown, and the present application is not limited to the size and thickness of each component. In order to make the drawing clearer, the thickness of the components is appropriately exaggerated in some places in the drawing.

[0081] The inventors designed a yarn quality prediction system, as shown in Figure 1 comprises:

[0082] A data acquisition module acquires yarn original data; the data acquisition module comprises a line array camera and a sensor, the line array image sensor of the line array camera is based on a CCD (Charge Coupled Device), and the sensor comprises a yarn evenness tester, a twist tester, and a strength tester; the yarn original data comprises real-time yarn images and sensor data, and the sensor data comprises yarn evenness, yarn twist, and yarn tensile data; the line array camera shoots real-time yarn images, the yarn evenness tester acquires yarn evenness, the twist tester acquires yarn twist, and the strength tester acquires yarn tensile data;

[0083] A data storage module stores the yarn original data;

[0084] A data preprocessing module pre-processes the yarn original data, and the pre-processing comprises real-time yarn image pre-processing and sensor data pre-processing; the real-time yarn image pre-processing comprises image denoising, image enhancement, determination of defect range, semi-automatic labeling of true value box and defect category, generation of training set and verification set according to a set proportion, normalization processing of yarn detection results, determination of defect range by using a morphological method and a region growing algorithm, and the set proportion is preferably 8:2; the sensor data pre-processing comprises data extraction and removal of error data with excessively large deviation;

[0085] The defect detection module constructs a deep learning model, uses the deep learning model for defect processing, and generates a yarn detection result, which includes yarn evenness and color unevenness indexes derived from the defect detection result; constructing the deep learning model includes: inputting the training set and the validation set into a convolutional neural network for feature extraction, entering a region generation network, and continuing the training of the yarn target detection network through a classification branch; the candidate region is brought closer to the real region through a regression branch; finally, the optimal target detection result is obtained through the back propagation of the loss function and the weight update;

[0086] The mathematical model module calculates yarn quality indexes using the yarn original data, which include yarn evenness unevenness and yarn tensile properties; the yarn evenness unevenness is calculated using the Martindale formula as follows:

[0087]

[0088]

[0089] wherein, CV is the yarn evenness unevenness, C D is the diameter dispersion coefficient of the wool fiber, c is the yarn linear density, d is the diameter of the wool fiber, and n is an intermediate variable;

[0090] The yarn tensile properties are calculated using the Bearale formula as follows:

[0091]

[0092] wherein, T represents the yarn tensile properties, t w is the yarn twist.

[0093] The comprehensive quality prediction module integrates the yarn detection result and the yarn quality indexes to predict the yarn quality, including predicting the yarn defect type and quantity, the yarn evenness unevenness, and the yarn tensile properties.

[0094] The data acquisition module, the data storage module, and the data preprocessing module are sequentially connected in communication, the data preprocessing module is connected in communication with the defect detection module and the mathematical model module, and the comprehensive quality prediction module is connected in communication with the defect detection module and the mathematical model module;

[0095] In response to the yarn delivery of the automatic doffer, the data acquisition module acquires the yarn original data, sends it to the data storage module for storage, and then to the data preprocessing module for preprocessing, the preprocessed yarn original data is obtained through the deep learning model constructed by the defect detection module to obtain the yarn detection result, the yarn quality indexes are calculated through the mathematical model module, and the comprehensive quality prediction module predicts the yarn quality according to the yarn detection result and the yarn quality indexes.

[0096] Based on the above embodiments, the inventors provide a yarn quality prediction method, as shown in Figure 2 comprising the following steps:

[0097] S100, preparation, the data acquisition module acquires yarn original data, sends to the data storage module for storage, and then sends to the data preprocessing module to generate a training set and a validation set, and the defect detection module constructs a deep learning model; the data acquisition module includes a line array camera and a sensor, the yarn original data includes real-time yarn images and sensor data, and the sensor data includes yarn linear density, yarn twist and yarn tensile data; specifically including:

[0098] S110, training set and validation set generation, the data acquisition module acquires yarn original data, sends to the data storage module for storage, and then sends to the data preprocessing module for preprocessing, generates a training set and a validation set, the ratio of the training set and the validation set is 8:2, the preprocessing includes real-time yarn image preprocessing and sensor data preprocessing, the real-time yarn image preprocessing includes image denoising, image enhancement, determining defect range, semi-automatic labeling of true value box and defect category, generating training set and validation set according to the set ratio, normalizing yarn detection results, determining defect range by morphological method and region growing algorithm, and the set ratio is preferably 8:2;

[0099] Sensor data preprocessing includes data extraction, removing error data with large deviation;

[0100] S120, deep learning model training, the defect detection module constructs a deep learning model, including: inputting the training set and the validation set into the convolutional neural network for feature extraction, entering the region generation network, and continuing the training of the yarn target detection network through the classification branch; the regression branch is used to move the candidate region to the real region; finally, the optimal target detection result is obtained by using the back propagation of the loss function and the weight update; input the training set and the validation set for training, when the loss function of the target detection result is the minimum, the deep learning model training is completed, specifically including:

[0101] S121, constructing a deep learning model network, setting the initial value of network training hyperparameters, the hyperparameters include iteration times, learning rate, Batch, Epoch, taking Cascade RCNN as the yarn target detection network, and taking ResNet50 as the feature extraction backbone network, the yarn target detection network also includes the region generation network, the classification regression network and the feature fusion network, the feature image is obtained from the feature extraction backbone network, the feature fusion enhancement is performed through the feature fusion network, and the defect positioning and identification are performed through the region generation network and the classification regression network;

[0102] S122, improve ResNet50, replace the original convolution kernel in ResNet50 with a deformable convolution, dynamically adjust according to different detection targets, concentrate the extracted features on the target, and enhance the feature extraction ability of the feature extraction backbone network for defects with complex morphology;

[0103] S123, extract yarn feature information of real-time yarn image, add FPN and BFP for Cascade RCNN network, balance and fuse features at different levels, and use the feature extraction backbone network to extract yarn feature information of real-time yarn image;

[0104] S124, train the deep learning model, calculate the error loss of the yarn target detection network, and update the weight parameters to minimize the error of the deep learning model;

[0105] S125, complete the training of the deep learning model, repeat steps S121 to S124 for not less than 200 times, and when the loss function of the target detection result is minimized, the training of the deep learning model is completed;

[0106] S200, yarn conveying, the automatic doffer conveys the yarn;

[0107] S300, yarn original data acquisition, the data acquisition module acquires yarn original data and sends it to the data storage module; including:

[0108] S310, line array camera shoots real-time yarn image,

[0109] S320, slub tester acquires yarn slub fineness, twist tester acquires yarn twist, and strength tester acquires yarn tensile data;

[0110] S330, the data acquisition module sends the real-time yarn image and the yarn slub fineness, yarn twist and yarn tensile data to the data storage module.

[0111] S400, yarn original data storage, the data storage module receives and stores the yarn original data, and sends it to the data preprocessing module;

[0112] S500, yarn original data preprocessing, the data preprocessing module pre-processes the yarn original data and sends it to the defect detection module and the mathematical model module respectively, the preprocessing includes real-time yarn image preprocessing and sensor data preprocessing, the real-time yarn image preprocessing includes image denoising, image enhancement, determining defect range, semi-automatic labeling of true value box and defect category, generating training set and verification set according to the set proportion, normalizing the yarn detection result, determining the defect range by morphological method and region growing algorithm, and the set proportion is preferably 8:2; sensor data preprocessing includes data extraction and removal of error data with large deviation;

[0113] S600, yarn detection result is obtained, in response to the pretreated yarn original data, the defect detection module obtains the yarn detection result through the deep learning model, and sends to the comprehensive quality prediction module;

[0114] S700, yarn quality index calculation, in response to the pretreated yarn original data, the mathematical model module calculates the yarn quality index, and sends to the comprehensive quality prediction module;

[0115] S800, yarn quality comprehensive prediction, the comprehensive quality prediction module calculates the yarn detection result and the yarn quality index with weights, the weight ratio is 1:1, and the comprehensive yarn quality prediction result is obtained.

[0116] The preferred embodiments of the application are described in detail above. It should be understood that those skilled in the art can make many modifications and changes without creative labor based on the concept of the present application. Therefore, any technical solution obtained by logical analysis, reasoning or limited experiment based on the existing technology within the scope of the present application should be within the protection scope determined by the claims.

Claims

1. A yarn quality prediction system characterized by, The method comprises the following steps: a data acquisition module acquires yarn original data; a data storage module stores the yarn original data; a data preprocessing module preprocesses the yarn original data; a defect detection module constructs a deep learning model, uses the deep learning model for defect processing, and generates a yarn detection result; The defect detection module constructs a deep learning model, inputs a training set and a validation set for training, and completes the deep learning model training when the loss function of the target detection result is the minimum. Specifically, the method comprises the following steps: S121, a deep learning model network is constructed, and the initial value of the network training hyperparameter is set. Cascade RCNN is used as the yarn target detection network, and ResNet50 is used as the feature extraction backbone network. The yarn target detection network further comprises a region generation network, a classification regression network, and a feature fusion network. The feature image is obtained from the feature extraction backbone network, the feature fusion enhancement is performed through the feature fusion network, and the defect positioning and identification are performed through the region generation network and the classification regression network; S122, the ResNet50 is improved, and the original convolution kernel in the ResNet50 is replaced with a deformable convolution. The deformable convolution is dynamically adjusted according to different target to be detected, so that the extracted features are concentrated on the target, and the feature extraction capability of the feature extraction backbone network for defects with complex and variable shapes is enhanced; S123, yarn feature information of a real-time yarn image is extracted. For the Cascade RCNN network, FPN and BFP are added to balance the fusion of features at different levels. The yarn feature information of the real-time yarn image is extracted by using the feature extraction backbone network; S124, the deep learning model is trained, the error loss of the yarn target detection network is calculated, and the weight parameters are updated to minimize the error of the deep learning model; S125, the deep learning model training is completed. The training is repeated for no less than 200 times. When the loss function of the target detection result is the minimum, the deep learning model training is completed; a mathematical model module calculates yarn quality indexes by using the yarn original data. The yarn quality indexes include yarn evenness and yarn tensile property. The yarn evenness is calculated by using the Martindale formula as follows: (1) (2) wherein, is the yarn evenness, is the wool fibre diameter dispersion coefficient, is the yarn linear density, is the wool fibre diameter, is the intermediate variable; The yarn tensile property is calculated by using the Bearale formula as follows: (3) wherein, represents the yarn tensile properties, is the yarn twist; A comprehensive quality prediction module predicts the yarn quality by comprehensively considering the yarn detection result and the yarn quality indexes. The yarn quality prediction includes predicting the yarn defect type and quantity, the yarn evenness, and the yarn tensile property; The data acquisition module, the data storage module, and the data preprocessing module are sequentially connected in communication. The data preprocessing module is connected in communication with the defect detection module and the mathematical model module. The comprehensive quality prediction module is connected in communication with the defect detection module and the mathematical model module. In response to yarn conveying of the automatic doffer, the data acquisition module acquires the yarn original data, sends the yarn original data to the data storage module for storage, and then sends the yarn original data to the data preprocessing module for preprocessing, the yarn original data after preprocessing is sent to the deep learning model constructed by the defect detection module to obtain the yarn detection result, the yarn quality index is calculated by the mathematical model module, and the comprehensive quality prediction module predicts the yarn quality according to the yarn detection result and the yarn quality index.

2. The yarn quality prediction system of claim 1, wherein, The data acquisition module comprises a linear array camera and a sensor.

3. The yarn quality prediction system of claim 2, wherein, The yarn original data comprises real-time yarn images and sensor data.

4. The yarn quality prediction system of claim 3, wherein, The preprocessing comprises real-time yarn image preprocessing and sensor data preprocessing.

5. The yarn quality prediction system of claim 1, wherein, The deep learning model is constructed by inputting a training set and a validation set into a convolutional neural network for feature extraction, entering a region generation network, continuing the training of a yarn target detection network through a classification branch, and through a regression branch, the candidate region is moved closer to the real region, and finally, the optimal target detection result is obtained by using the back propagation of the loss function and the weight update.

6. A yarn quality prediction method using the yarn quality prediction system according to any one of claims 1 to 5, characterized by, The method comprises the following steps: S100, preparation, the data acquisition module acquires yarn original data, sends the yarn original data to the data storage module for storage, and then sends the yarn original data to the data preprocessing module to generate a training set and a validation set, and trains the deep learning model constructed by the defect detection module; S200, yarn conveying, the automatic doffer conveys the yarn; S300, yarn original data acquisition, the data acquisition module acquires yarn original data and sends the yarn original data to the data storage module; specifically comprising: S310, the linear array camera shoots real-time yarn images, S320, the yarn evenness is acquired by a evenness tester, the yarn twist is acquired by a twist tester, and the yarn tensile data is acquired by a strength tester; S330, the data acquisition module sends the real-time yarn images and the yarn evenness, the yarn twist and the yarn tensile data to the data storage module; S400, yarn original data storage, the data storage module receives and stores the yarn original data, and sends the yarn original data to the data preprocessing module; S500, yarn original data preprocessing, the data preprocessing module pre-processes the yarn original data and sends the yarn original data to the defect detection module and the mathematical model module respectively; S600, yarn detection result acquisition, in response to the pre-processed yarn original data, the defect detection module obtains the yarn detection result through the deep learning model and sends the yarn detection result to the comprehensive quality prediction module; S700, yarn quality index calculation, in response to the pre-processed yarn original data, the mathematical model module calculates the yarn quality index and sends the yarn quality index to the comprehensive quality prediction module; S800, comprehensive yarn quality prediction, the comprehensive quality prediction module calculates the yarn detection result and the yarn quality index with weights to obtain a comprehensive yarn quality prediction result.

7. The yarn quality prediction method of claim 6, wherein, The step S100 comprises: S110, training set and validation set generation, the data acquisition module acquires yarn original data, sends to the data storage module for storage, and then sends to the data preprocessing module for preprocessing, to generate training set and validation set; S120, deep learning model training, the defect detection module constructs a deep learning model, inputs the training set and the validation set for training, and when the loss function of the target detection result is minimum, the deep learning model training is completed.

8. The yarn quality prediction method of claim 7, wherein, The step S120 comprises: S121, constructing a deep learning model network, setting the initial value of the hyperparameter of network training, taking Cascade RCNN as the yarn target detection network, and taking ResNet50 as the feature extraction backbone network, the yarn target detection network further comprising a region generation network, a classification regression network and a feature fusion network, obtaining a feature image from the feature extraction backbone network, performing feature fusion enhancement through the feature fusion network, and performing defect positioning and identification through the region generation network and the classification regression network; S122, improving ResNet50, using deformable convolution to replace the original convolution kernel in the ResNet50, dynamically adjusting with different target to be detected, making the extracted features focus on the target, and enhancing the feature extraction ability of the feature extraction backbone network for defects with complex and variable shapes; S123, extracting yarn feature information of real-time yarn image, adding FPN and BFP for the Cascade RCNN network, balancing and fusing features of different levels, and using the feature extraction backbone network to extract yarn feature information of real-time yarn image; S124, training the deep learning model, calculating the error loss of the yarn target detection network, updating the weight parameters, and making the error of the deep learning model minimum; S125, completing the deep learning model training, repeating steps S121 to S124 for training not less than 200 times, and when the loss function of the target detection result is minimum, the deep learning model training is completed.

9. A yarn quality prediction method according to any of claims 6 to 8, characterised in that, The weight ratio of the yarn detection result and the yarn quality index in the step S800 is 1:1.

Images