A method for detecting the production process of covered yarn based on lightweight objectives

Abstract

The present invention relates to the technical field of yarn detection, and specifically to a method for detecting a coated yarn production process based on a lightweight goal. The present invention analyzes the historical coated yarn image data with current problems through a coated yarn composite analysis model, divides the data into multiple levels and generates corresponding labels, generates a problem association matrix by associating corresponding historical processing parameters and converting them into image data and associating corresponding labels; identifies abnormal processing parameter features through the coated yarn composite analysis model, and optimizes the corresponding problem association matrix weights; generates new category labels for historical coated yarn image data based on the new association matrix weights, and inputs them into the coated yarn composite analysis model for training; through this method, while identifying the defects of the coated yarn, the corresponding processing parameters of the coated yarn are associated, so that the optimized model can also automatically feedback the corresponding processing parameter problems when identifying the defect features, so as to facilitate subsequent optimization and reduce the probability of problematic coated yarns.

Description

Technical Field

[0001] The invention relates to the technical field of yarn detection, and in particular to a method for detecting a coated yarn production process based on a lightweight target. Background Art

[0002] Covered yarn is a composite yarn composed of multiple fibers. The covered yarn mainly includes a yarn core and an outer fiber. Since the covered yarn can inherit the advantages of the multiple fibers it combines, such as elasticity and comfort, it is widely used in the field of clothing. With the advancement of technology, the emergence of various new fibers has made the covered yarn not only have the original characteristics, but also realize the lightweight process, and also have the characteristics of heat resistance and wear resistance. Therefore, the covered yarn can not only produce lightweight clothing, but also can be used in protective equipment, realizing the dual needs of protective performance and lightweight.

[0003] Hollow spindle technology is one of the important technologies for producing covered yarn to achieve the goal of lightweighting. This technology enables the outer fiber to be wrapped around the yarn core in an efficient and stable manner, while ensuring its tension and stability while having the advantage of lightweighting. However, in the production process, due to the complex processing parameters of hollow spindle technology, problems such as uneven coating or breakage are prone to occur during the production of lightweight covered yarn.

[0004] In the prior art, deep learning algorithms such as image recognition are used to automatically detect and mine the above-mentioned problems arising in the production process. However, in this process, only the coated yarn with defects can be mined, and it is impossible to provide feedback on the processing parameters that cause the problems in the coated yarn.

[0005] Therefore, a detection method for the production process of covered yarn based on the lightweight goal is proposed. Summary of the invention

[0006] The purpose of the present invention is to provide a method for detecting the production process of coated yarn based on the goal of lightweighting, which analyzes the historical coated yarn image data with current problems through a coated yarn composite analysis model, divides the image data into multiple levels according to the characteristics of the image data and generates corresponding labels, generates a problem association matrix by associating the corresponding historical processing parameters and converting them into image data and associating the corresponding labels; identifies abnormal processing parameter features through a coated yarn composite analysis model based on the image data converted by the processing parameters, and optimizes the corresponding problem association matrix weights; re-labels the second coated yarn processing parameter set based on the new association matrix weights, generates new category labels for the historical coated yarn image data, and inputs them into the coated yarn composite analysis model for training; through this method, while identifying the defects of the coated yarn, the corresponding processing parameters of the coated yarn are associated, so that the optimized model can also automatically feedback the corresponding processing parameter problems when identifying the defect features, so as to facilitate subsequent optimization and reduce the probability of problematic coated yarns.

[0007] To achieve the above object, the present invention provides the following technical solutions:

[0008] A method for detecting a production process of a covered yarn based on a lightweight target, comprising:

[0009] Obtaining historical covered yarn image data of problems that occurred during the production process, and extracting the historical covered yarn image data to obtain covered yarn problem characteristics based on a covered yarn composite analysis model or multiple image analysis models;

[0010] Perform feature classification and labeling according to the covered yarn problem features to generate a set of multiple labeled problem images;

[0011] Further, the corresponding tags include a first covered yarn problem tag and a second covered yarn problem tag;

[0012] Divide the covered yarn problem into multiple levels according to the number of the first covered yarn problem labels, and divide the covered yarn problem into multiple levels according to the number of the second covered yarn problem labels; divide the label problem image sets according to the covered yarn problem levels to generate multiple label problem image sets;

[0013] Associating historical processing parameters of production hardware in which a plurality of the labeled problem image sets appear, and generating a first set of covered yarn processing parameters based on the historical processing parameters;

[0014] Performing data feature conversion according to the first covered yarn processing parameter set, and generating a second covered yarn processing parameter set according to the first covered yarn processing parameter set after the data feature conversion;

[0015] Further, all data of the first covered yarn processing parameter set are subjected to fast Fourier transform to obtain multiple frequency data and multiple amplitude data of each covered yarn problem data, each frequency data corresponds to one amplitude data; the first covered yarn processing parameter set is a plurality of time series data; the total number of the first covered yarn processing parameter set is m;

[0016] Sort the amplitude data from large to small, and select the corresponding frequency data of the first n amplitude data as reference frequency data , obtain the cycle of each covered yarn problem data ,calculate Get n segmentation cycles of each of the covered yarn problem data ,in accordance with Divide each of the covered yarn problem data and splice the divided covered yarn problem data to obtain a quantity of The second coated yarn processing parameter set, each of the coated yarn image data size is ;

[0017] Establishing and initializing a problem association matrix of a plurality of the labeled problem image sets and the second covered yarn processing parameter set, wherein the weight of the problem association matrix is ​​the label of the plurality of the labeled problem image sets;

[0018] Further, the second covered yarn processing parameter set data is encoded, and the corresponding level of the problem feature annotation is added to the data encoding as the problem association matrix weight to generate multiple associated data, and the multiple associated data are sorted to generate the problem association matrix; the data length of each associated data is obtained, and the maximum data length is used as reference data to supplement each associated data until the data length of each associated data is consistent, and the problem association matrix weight is initialized to complete the initialization of the problem association matrix;

[0019] The abnormal processing parameter features of the second covered yarn processing parameter set are extracted by the covered yarn composite analysis model, and the weight of the problem association matrix is ​​optimized according to the abnormal processing parameter features:

[0020] ;

[0021] in, is the weight of the problem association matrix after the i-th optimization, n is the total number of the covered yarn image problem data formed after each of the covered yarn problem data is segmented, is the weight of the j-th corresponding problem association matrix, is the quantitative statistics of abnormal features of the processing parameters, abnormal characteristics of the processing parameters identified by the composite analysis model of the covered yarn;

[0022] Further, the coated yarn composite analysis model includes a coated yarn image processing unit, a coated yarn image feature extraction unit, a coated yarn image feature classification unit and a coated yarn classification level output unit;

[0023] The covered yarn image processing unit is used to pre-process the input image data to generate feature data to be extracted; the input image data includes historical covered yarn image data and a second covered yarn processing parameter set;

[0024] Further, the covered yarn image processing unit adopts different preprocessing methods according to the type of input data;

[0025] If the input data is the historical covered yarn image data, data denoising is performed by Gaussian filtering, data enhancement is performed on the denoised image according to histogram equalization, and the enhanced data is normalized to generate the feature data to be extracted of the historical covered yarn image data;

[0026] If the input data is a second covered yarn processing parameter set, the data dimensions are unified by data supplementation, and the unified data are normalized to generate the feature data to be extracted of the second covered yarn processing parameter set;

[0027] The coated yarn image feature extraction unit extracts data features of the feature data to be extracted to generate coated yarn feature data; the coated yarn feature data includes the covered yarn problem features and the processing parameter abnormal features; the coated yarn image feature extraction unit includes an Attention mechanism, an expansion convolution network, a residual convolution network and a feature fusion layer;

[0028] The number of Attention mechanisms is 1; the number of dilated convolution networks is 3, and the dilated convolution adopts The convolution kernel has expansion interval parameters of 1, 2 and 3; the number of residual convolution networks is 3, and the residual convolution network includes 3 The convolution kernel and 2 The convolution kernel of , the feature fusion layer is relu;

[0029] The covered yarn image feature classification unit performs classification according to the feature category of the covered yarn feature data to generate a covered yarn feature category and a covered yarn feature category probability; the covered yarn image feature classification unit is a Bayesian classifier;

[0030] The covered yarn classification level output unit generates a weight level corresponding to the covered yarn feature category according to the covered yarn feature category probability;

[0031] Re-labeling the second covered yarn processing parameter set according to the optimized problem association matrix weights to generate a new category label;

[0032] Further, obtaining each covered yarn problem level of the first covered yarn problem label and each covered yarn problem level of the second covered yarn problem label;

[0033] According to the same problem label and the same covered yarn problem level, all the problem association matrix weights corresponding to each of the historical processing parameters are obtained; according to the weights of the multiple problem association matrices of the same historical processing parameters, they are normalized, and if the normalized problem association matrix weight is greater than 0.5, the problem association matrix weight is re-labeled as 1; if the normalized problem association matrix weight is not greater than 0.5, the problem association matrix weight is re-labeled as 0;

[0034] Re-labeling the corresponding labels of the multiple label problem image sets according to the new category labels, and inputting the labels into the covered yarn composite analysis model for data training and feature learning, and optimizing the covered yarn composite analysis model for production inspection;

[0035] Further, secondary data encoding is performed according to the data encoding of the second covered yarn processing parameter set and the re-labeled labels, the category labels of different levels are generated, and the label problem image set is marked;

[0036] The re-labeled label problem image set is divided into 70% training set and 30% test set, and is input into the coated yarn composite analysis model for training and testing to generate new production detection model parameters.

[0037] Compared with the prior art, the present invention has the following beneficial effects:

[0038] 1. The present invention can extract the characteristics of coated yarn problems through the coated yarn composite analysis model, and can also perform analysis based on the time series data obtained from the processing parameters, and can also perform classification and identification based on the subsequent category data; this method can enable the coated yarn composite analysis model to realize a multi-dimensional data feature learning process, and realize the reuse of the model through the common characteristics in the method, which greatly reduces the investment cost of building the model. The model is also helpful to explore the characteristics of the occurrence of coated yarn problems and subsequent automatic recognition training, and provides a model basis and data feature basis for the feedback of coated yarn problems and subsequent data analysis.

[0039] 2. The present invention is based on processing parameter data and uses the fast Fourier algorithm to mine the cycles in the time series data formed by the processing parameter data, and forms sub-time series data of multiple dimensions according to the cycles and forms corresponding image feature data; this method can convert the time series data into image data, and upgrade the data to provide more data features, which is helpful to improve the recognition accuracy of problem data, so as to provide a high-quality processing data analysis basis and ensure the accuracy of feedback on the causes of coated yarn problems; at the same time, conversion into image data also facilitates the reuse of models and reduces the investment in building new models.

[0040] 3. The present invention establishes a problem association matrix, associates corresponding processing parameters based on multiple levels of coated yarn image data features, optimizes the weights of the problem association matrix through analysis of the processing parameter process, and optimizes the weights to generate information category label data to mark the corresponding coated yarn image data for training. Through this method, when a coated yarn problem of a problematic degree occurs, it can be classified into parameter data of the processing problem based on the category label, so as to analyze the image data defects and associate the processing parameters that cause the problem for feedback, so as to assist subsequent related technical personnel to make quick adjustments and reduce the occurrence of related coated yarn problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 It is a schematic diagram of the method flow of the present invention;

[0042] Figure 2 It is a corresponding diagram of a virtual device of the composite analysis model of the covered yarn of the present invention;

[0043] In the figure: 1. a coated yarn image processing unit; 2. a coated yarn image feature extraction unit; 3. a coated yarn image feature classification unit; 4. a coated yarn classification level output unit;

[0044] Figure 3 It is an algorithm flow chart of a covered yarn image feature extraction unit of the covered yarn composite analysis model of the present invention;

[0045] Figure 4 It is an algorithm flow chart of the residual convolution network of the covered yarn composite analysis model of the present invention;

[0046] Figure 5 It is a schematic diagram of some data of the processing parameter standardization of the present invention;

[0047] Figure 6 It is a data display diagram of two-dimensional data formed by splicing one-dimensional data of processing parameters of the present invention;

[0048] Figure 7 It is the training set and test accuracy result diagram of the present invention;

[0049] Figure 8 This is a graph of the loss function results of the training set and test set of the present invention. DETAILED DESCRIPTION

[0050] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0051] With the development of technology, the lightweight of coated yarn can not only reduce the weight of the original clothing while meeting the original clothing characteristics, but also realize the wearing of lighter clothing; hollow spindle technology is the key technology of lightweight coated yarn. Due to the complexity of processing parameters in the production process, the coated yarn may have uneven coating or breakage problems in the production process; but in the prior art, only the problematic coated yarn can be identified, and the processing parameters with problems cannot be fed back. Therefore, the present invention provides a detection method for the production process of coated yarn based on the lightweight target, referring to Figure 1 As shown, the technical solution is as follows:

[0052] Obtain historical covered yarn image data with problems in the production process, and extract the historical covered yarn image data to obtain covered yarn problem features based on a covered yarn composite analysis model or multiple image analysis models; perform feature classification and label corresponding tags based on the covered yarn problem features to generate multiple labeled problem image sets;

[0053] Associating historical processing parameters of production hardware in which a plurality of the labeled problem image sets appear, and generating a first set of covered yarn processing parameters based on the historical processing parameters;

[0054] Performing data feature conversion according to the first covered yarn processing parameter set, and generating a second covered yarn processing parameter set according to the first covered yarn processing parameter set after the data feature conversion; establishing and initializing a problem association matrix of a plurality of the labeled problem image sets and the second covered yarn processing parameter set, wherein the weights of the problem association matrix are the labels of the plurality of the labeled problem image sets;

[0055] Extracting abnormal processing parameter features of the second covered yarn processing parameter set by using the covered yarn composite analysis model, and optimizing the problem association matrix weights according to the abnormal processing parameter features;

[0056] Relabel the second coated yarn processing parameter set according to the optimized problem association matrix weights to generate new category labels, relabel the corresponding labels of the multiple label problem image sets according to the new category labels, input the coated yarn composite analysis model for data training and feature learning, and optimize the coated yarn composite analysis model for production inspection.

[0057] By collecting and identifying the image data of the covered yarn with problems, the data is classified according to the problem characteristics to divide the problem covered yarn image sets of different levels under different problems, so as to facilitate the corresponding processing data; the data set formed by the corresponding historical processing parameters is used for data feature conversion, the time series data is converted into image data, the data format is unified, the cost of building a time series model is reduced, and it is also convenient for feature identification; the corresponding problem association matrix is ​​established according to the processing parameter characteristics and the covered yarn image set, the problem management matrix is ​​optimized according to the recognition results of the abnormal processing parameter features extracted by the model, and the processing parameter problem labels of the image problem data sets of different levels are converted according to the weights of the matrix, and data training and feature learning are carried out; through this method, in the project of identifying the covered yarn image data, the corresponding processing parameters with problems can be fed back according to the level, so that relevant technical personnel can make adjustments and improve the analysis dimension of the data.

[0058] Embodiment 1

[0059] In order to specifically explain the matrix building process of the processing parameter data and the problem image data and the identification process of each data, the following examples are used for illustration:

[0060] Obtain historical coated yarn image data of problems in the production process, and extract the historical coated yarn image data to obtain the covered yarn problem features based on the coated yarn composite analysis model or multiple image analysis models; the covered yarn problem features include breakage features and uneven coating features; the multiple image analysis models include the coated yarn composite analysis model and other multiple image analysis models (for example, but not limited to YOLO, GANs, Mask R-CNN, etc.);

[0061] Perform feature classification and labeling according to the covered yarn problem features to generate a set of multiple labeled problem images;

[0062] Further, the corresponding tags include a first covered yarn problem tag and a second covered yarn problem tag;

[0063] The first covered yarn problem label is a broken label corresponding to a broken feature, and the second covered yarn problem label is an uneven coating label corresponding to an uneven coating feature;

[0064] Divide the covered yarn problem into multiple levels according to the number of the first covered yarn problem labels, and divide the covered yarn problem into multiple levels according to the number of the second covered yarn problem labels; divide the label problem image sets according to the covered yarn problem levels to generate multiple label problem image sets;

[0065] The calculation of the covered yarn problem level is as follows:

[0066] ;

[0067] in, is the problem level of the kth type of covered yarn, k=1 or k=2, 1 is the breakage problem, 2 is the uneven covering problem, is the kth image feature obtained by the composite analysis model of coated yarn, Parameters for level division; The number of historical coated yarn breakage or uneven coating in multiple batches is identified in the coated yarn composite analysis model. The number of coated yarn breakage is used as the horizontal axis and the number of uneven coated yarn is used as the vertical axis to generate corresponding coordinates and cluster them using the k-means algorithm. The number corresponding to the k cluster center results is used as The number of k is set based on expert experience, and can also be adjusted based on expert experience during the specific implementation process. make corrections;

[0068] Optionally, in the process of calculating the covered yarn grade, multiple models can be added for auxiliary calculation to improve the accuracy of recognition and improve work efficiency; the covered yarn problem grade calculation adding multiple models is as follows:

[0069] ;

[0070] Among them, M is the total number of models added, is the level weight of the i-th model; It is obtained based on the proportion of the recognition accuracy of the model in all models. The specific calculation is as follows:

[0071] ;

[0072] in, is the recognition accuracy of the sth model, is the recognition accuracy of the i-th model, is the weight correction coefficient, The setting is made based on the actual recognition accuracy calculation results combined with expert experience;

[0073] Table 1 shows the corresponding level data in some problem data sets. In this embodiment, the level data is 0-5, the minimum level is 0, and the maximum level is 5. During the calculation process, the values ​​exceeding 5 are marked as 5. The specific contents are as follows:

[0074] Table 1 Problem levels of each label problem image set

[0075]

[0076] Since the problems of the covered yarn are different, and they may appear in the production process, such as only one break, multiple breaks, uneven coating in one place, and uneven coating in multiple places; and the corresponding processing parameters of these problems are also different, for this reason, the data are graded and classified according to the characteristics of the original image, so as to conduct different data analysis for different problems, improve the accuracy of image problem data classification, facilitate the subsequent analysis process of corresponding time series data, and provide a stable and accurate data basis for the subsequent label data generation;

[0077] Associating historical processing parameters of production hardware in which a plurality of the labeled problem image sets appear, and generating a first set of covered yarn processing parameters based on the historical processing parameters;

[0078] The processing parameters are the input voltage of the servo motor, the speed of the speed controller and the acceleration of the speed change process during the use of the hollow spindle technology; data feature conversion is performed according to the first covered yarn processing parameter set, and the second covered yarn processing parameter set is generated according to the first covered yarn processing parameter set after the data feature conversion; the first covered yarn processing parameter set is a timing processing parameter set, and the second covered yarn processing parameter set is an image processing parameter set converted based on the timing processing parameter set;

[0079] Further, all data of the first covered yarn processing parameter set are subjected to fast Fourier transform to obtain multiple frequency data and multiple amplitude data of each covered yarn problem data, each frequency data corresponds to one amplitude data; the first covered yarn processing parameter set is a plurality of time series data; the total number of the first covered yarn processing parameter set is m;

[0080] Sort the amplitude data from large to small, and select the corresponding frequency data of the first n (for example, but not limited to 1, 2, 3) amplitude data as reference frequency data , obtain the cycle of each covered yarn problem data ,calculate Get n segmentation cycles of each of the covered yarn problem data ,in accordance with Divide each of the covered yarn problem data and splice the divided covered yarn problem data to obtain a quantity of The second coated yarn processing parameter set, each of the coated yarn image data size is ;

[0081] For example, data A selects the first three frequency data as reference frequency data; the first frequency data is divided into 36 division periods, the second frequency data is divided into 44 division periods, and the third frequency data is divided into 60 division periods;

[0082] Data splicing into image data reference Figure 6 As shown;

[0083] Furthermore, before image stitching, all data will be standardized, referring to Figure 5 As shown, Figure 5 The vertical axis is the standardized data size, compressed to between 0 and 1, and the horizontal axis is the unit time. In this embodiment, 5 seconds is used as the unit time for sorting. For example, the first 5 seconds is used as the unit time 1, the second 5 seconds is used as the unit time 2, and so on. Figure 5 Some of the data after standardization are shown;

[0084] The method converts the time series data formed by the processing parameters into image data. Through this method, the dimension of the data can be increased on the basis of the time series data to increase the number of data features. At the same time, based on the data of different periods after segmentation, it is also convenient to enhance the data features and analyze the abnormal processing parameters in the data, so as to facilitate the subsequent weight optimization calculation process, improve the accuracy of the optimization data and label production, and improve the accuracy of the subsequent image data training based on the label.

[0085] Establishing and initializing a problem association matrix of a plurality of the labeled problem image sets and the second covered yarn processing parameter set, wherein the weight of the problem association matrix is ​​the label of the plurality of the labeled problem image sets;

[0086] Further, the second covered yarn processing parameter set data is encoded, the corresponding level of the problem feature annotation is added to the data encoding as the problem association matrix weight to generate a plurality of associated data, and the plurality of associated data are sorted to generate the problem association matrix;

[0087] The associated data will be entered into the data number and the corresponding level. The format of data encoding is as follows:

[0088] Data number 1, data number 2, ..., data number t, fracture grade weight K1, coating uneven weight K2;

[0089] The data number specifically includes the problem classification number, the problem level number, the processing parameter position number, the processing parameter type number, the unified processing parameter number, the completion check number, and the abnormality check number. Each number can be set according to the specific implementation situation, such as decimal, binary, hexadecimal, etc. The function of each number is shown in Table 2:

[0090] Table 2 Data number, function and content introduction table

[0091]

[0092] Obtaining the data length of each piece of associated data, using the maximum data length as reference data, supplementing each piece of associated data until the data length of each piece of associated data is consistent, and initializing the weight of the problem association matrix to complete the initialization of the problem association matrix;

[0093] The initialization process is performed according to data and weight respectively. For example, if the B data has 2 frequency data, the first frequency data has 72 segmentation cycles, and the second frequency data has 74 segmentation cycles, the alignment result is:

[0094] Data number 1,..., data number 72,0,0;

[0095] The first frequency data will be padded with 2 zeros.

[0096] For weighted data, for example, C data includes fracture level weight and uneven coverage weight, and D data only has fracture level weight, so 0 will be added to the uneven coverage weight of D data. The initialization process is as follows:

[0097] C data number, C data break level weight, C data coverage uneven weight;

[0098] D data number, D data break level weight, 0;

[0099] Generate corresponding data according to different frequencies, and associate data weights through these data to generate weights corresponding to different image data. Since the processing parameter extraction frequency data is different, the processing parameter data is different. At the same time, the same processing parameter may correspond to multiple problems. Therefore, it is necessary to initialize the matrix through data supplementation and weights to facilitate data recording and subsequent calculation processes of the problem association matrix weight data, improve data integrity, and provide stable data for subsequent optimization calculations and label generation processes;

[0100] The abnormal processing parameter characteristics of the second covered yarn processing parameter set are extracted by the covered yarn composite analysis model; Table 3 shows the recognition accuracy of some processing parameters. Some abnormal processing parameters are marked by experts, and 4 groups of data are generated. Each group of data includes normal data and abnormal data. These data have not been used in the training process of the model. By inputting these data into the model to simulate the real environment recognition data, the recognition results are mainly the abnormal recognition accuracy and the abnormal recognition precision. The abnormal recognition accuracy is the proportion of the abnormal and normal correct data to all data, and the abnormal recognition precision is the proportion of the abnormal correct data to all data identified as abnormal, as shown in Table 3:

[0101] Table 3. Accuracy and precision of abnormal machining parameter recognition

[0102]

[0103] As can be seen from Table 3, the accuracy of identifying anomalies is around 89%, but the accuracy of anomaly identification is higher;

[0104] When the abnormal processing parameter characteristics are identified, the abnormal number will be marked on the corresponding data number;

[0105] Optimize the problem association matrix weights according to the abnormal characteristics of the processing parameters:

[0106] ;

[0107] in, is the weight of the problem association matrix after the i-th optimization, n is the total number of the covered yarn image problem data formed after each of the covered yarn problem data is segmented, is the weight of the j-th corresponding problem association matrix, is the quantitative statistics of abnormal features of the processing parameters, abnormal characteristics of the processing parameters identified by the composite analysis model of the covered yarn;

[0108] By identifying the abnormal features of processing parameters with corresponding grade weights and analyzing the number of abnormalities generated, the abnormal data that causes the corresponding problems can be obtained from a large amount of initial data. This method can mine the abnormal processing parameters that produce problematic image data in the existing data, improve the corresponding weights based on a large amount of data analysis, facilitate subsequent label generation, and provide an accurate label data basis for subsequent model training;

[0109] Furthermore, the coated yarn composite analysis model comprises a coated yarn image processing unit 1, a coated yarn image feature extraction unit 2, a coated yarn image feature classification unit 3 and a coated yarn classification level output unit 4, referring to Figure 2 As shown, the coated yarn image feature classification unit and the coated yarn classification level output unit will make a judgment. When the execution step is the label generation step of the image and processing parameters, the coated yarn image feature classification unit will proceed to the coated yarn classification level output unit after execution; otherwise, the data of the coated yarn image feature classification unit will be directly output; this step is for the subsequent training of image data containing new labels, so that the model can be directly output to realize the reuse of the model in the process;

[0110] The covered yarn image processing unit is used to pre-process the input image data to generate feature data to be extracted; the input image data includes historical covered yarn image data and a second covered yarn processing parameter set; the historical covered yarn image data is image data, and the second covered yarn processing parameter set is time series data of processing parameters;

[0111] Further, the covered yarn image processing unit adopts different preprocessing methods according to the type of input data;

[0112] If the input data is the historical covered yarn image data, data denoising is performed by Gaussian filtering, data enhancement is performed on the denoised image according to histogram equalization, and the enhanced data is normalized to generate the feature data to be extracted of the historical covered yarn image data;

[0113] If the input data is a second covered yarn processing parameter set, the data dimensions are unified by data supplementation, and the unified data are normalized to generate the feature data to be extracted of the second covered yarn processing parameter set;

[0114] Through different preprocessing methods, the model can adopt different preprocessing methods to enhance data features when acquiring different data. It is also convenient for model learning and analysis, improves the accuracy of model output data, and further enhances the reusability of data problems, reduces data processing errors caused by using the same processing method, reduces the accuracy of model recognition results, and provides a stable data foundation for subsequent training;

[0115] The coated yarn image feature extraction unit extracts the data features of the feature data to be extracted to generate coated yarn feature data; the coated yarn feature data includes the coated yarn problem features and the processing parameter abnormal features; the coated yarn image feature extraction unit includes an Attention mechanism, an expansion convolution network, a residual convolution network and a feature fusion layer; refer to Figure 3 As shown; Attention mechanism corresponds to Figure 3 Attention mechanism in

[0116] The number of Attention mechanisms is 1; the number of dilated convolution networks is 3, and the dilated convolution adopts The convolution kernel has expansion interval parameters of 1, 2 and 3; the number of residual convolution networks is 3, and the residual convolution network includes 3 The convolution kernel and 2 The convolution kernel of the residual convolution is referenced Figure 4 As shown in the figure; the feature fusion layer is relu; the attention mechanism can be used to mark the key points in the covered yarn processing parameters and the problem image data, the global image problem data features are extracted by dilated convolution, and the local image problem data features are extracted based on residual convolution; at the same time, the dilated convolution can also analyze the change trend information of the processing parameters based on global information, and the residual convolution can extract the local abnormal change information; so as to more comprehensively analyze the characteristics of the image problem data and the processing parameters, improve the recognition accuracy, and facilitate the subsequent label generation and image level division;

[0117] The covered yarn image feature classification unit performs classification according to the feature category of the covered yarn feature data to generate a covered yarn feature category and a covered yarn feature category probability; the covered yarn image feature classification unit is a Bayesian classifier;

[0118] The covered yarn classification level output unit generates a weight level corresponding to the covered yarn feature category according to the covered yarn feature category probability;

[0119] This model can quickly identify problematic covered yarn image data in image data, and can also be associated with time series data formed by processing parameter data; this method can reduce the process of establishing a time series model, and use a single model to generate different model parameters and weights for storage using different data to reduce the model establishment cost and training cost, while also reducing the time of the actual detection process, quickly extracting image problems and generating problem labels corresponding to processing parameters, thereby improving the recognition efficiency of the detection process.

[0120] Embodiment 2

[0121] In order to illustrate the training process of image data and corresponding processed data labels, the following example is used for illustration:

[0122] Re-labeling the second covered yarn processing parameter set according to the optimized problem association matrix weights to generate a new category label;

[0123] Further, obtaining each covered yarn problem level of the first covered yarn problem label and each covered yarn problem level of the second covered yarn problem label;

[0124] According to the same problem label and the same covered yarn problem level, obtaining all the problem association matrix weights corresponding to each of the historical processing parameters;

[0125] Normalizing the weights of the multiple problem association matrices of the same historical processing parameters, if the normalized problem association matrix weight is greater than 0.5, the problem association matrix weight is re-marked as 1; if the normalized problem association matrix weight is not greater than 0.5, the problem association matrix weight is re-marked as 0;

[0126] Considering the efficiency of image data recognition, the same level division is adopted through the recognized image data features to reduce the processing process of model recognition. At the same time, through the binary classification model, the image data of the coated yarn can be quickly classified on the basis of data analysis, and rapid recognition and feedback can be performed according to different problems.

[0127] Re-labeling the corresponding labels of the multiple label problem image sets according to the new category labels, and inputting the labels into the covered yarn composite analysis model for data training and feature learning, and optimizing the covered yarn composite analysis model for production inspection;

[0128] Further, secondary data encoding is performed according to the data encoding of the second covered yarn processing parameter set and the re-labeled labels, the category labels of different levels are generated, and the label problem image set is marked;

[0129] The re-labeled label problem image set is divided into 70% training set and 30% test set, and input into the coated yarn composite analysis model for training and testing to generate new production inspection model parameters;

[0130] We selected data based on expert experience and obtained data categories according to the steps, and input them into the model of this embodiment for training and testing. Table 4 shows the recognition accuracy of data training and testing. The key parameters are set to epoch 150, batch_size 16, and an early stopping mechanism is added. The loss function is L1, and some model parameters are modified for multiple training and testing. The results of 9 training and testing are shown in Table 4:

[0131] Table 4 The highest accuracy of training with various model parameters

[0132]

[0133] From Table 4, we can see that the accuracy of the overall training data is around 90%. The highest training accuracy of K07 is 91.25%. The test accuracy of K07 is also the highest, reaching 89.94%. Both the training set and the test set have reached an accuracy of about 90%, which is a good effect. The accuracy result curve of the training set and the test set of K07 is shown in Figure 4. Figure 7 As shown, the loss function result curves of the training set and test set of K07 refer to Figure 8 As shown in the figure, the division of different orders of magnitude has an impact on the feature learning of the model and the final accuracy data results. Therefore, we also divide the training set and test set of the data into 6:4, 7:3 and 8:2 ratios respectively. Table 5 shows the highest accuracy results of each data after multiple training and testing after dividing the training set and test set into different ratios:

[0134] Table 5 Accuracy results of different ratios of training set and test set division

[0135]

[0136] From the results in Table 5, it can be seen that the accuracy is highest when the training set and the test set are divided into 7:3. During the training process, different encoding types are used to further reduce the analysis steps of image data and processing parameters in the production and identification process of the covered yarn. Different encodings are used for output, which can output different results based on different levels based on binary classification data, and reduce the feature engineering of the data, thereby improving the feedback efficiency of the model in training and practical application.

[0137] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for detecting the production process of covered yarn based on the lightweight target, characterized in that: include: Obtaining historical covered yarn image data of problems that occurred during the production process, and extracting the historical covered yarn image data to obtain covered yarn problem characteristics based on a covered yarn composite analysis model or multiple image analysis models; Perform feature classification and labeling according to the covered yarn problem features to generate a set of multiple labeled problem images; Associating historical processing parameters of production hardware in which a plurality of the labeled problem image sets appear, and generating a first set of covered yarn processing parameters based on the historical processing parameters; Performing data feature conversion according to the first covered yarn processing parameter set, and generating a second covered yarn processing parameter set according to the first covered yarn processing parameter set after the data feature conversion; establishing and initializing a problem association matrix of a plurality of the labeled problem image sets and the second covered yarn processing parameter set, wherein the weights of the problem association matrix are the labels of the plurality of the labeled problem image sets; Extracting abnormal processing parameter features of the second covered yarn processing parameter set by using the covered yarn composite analysis model, and optimizing the problem association matrix weights according to the abnormal processing parameter features; Relabel the second coated yarn processing parameter set according to the optimized problem association matrix weights to generate new category labels; relabel the corresponding labels of the multiple label problem image sets according to the new category labels, and input them into the coated yarn composite analysis model for data training and feature learning, and optimize the coated yarn composite analysis model for production inspection.

2. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 1, characterized in that: Extracting the historical covered yarn image data to obtain covered yarn problem characteristics based on the covered yarn composite analysis model and / or other feature models; Extracting abnormal processing parameter features of the second covered yarn processing parameter set by the covered yarn composite analysis model includes: The covered yarn composite analysis model includes a covered yarn image processing unit, a covered yarn image feature extraction unit, a covered yarn image feature classification unit and a covered yarn classification level output unit; The covered yarn image processing unit is used to pre-process the input image data to generate feature data to be extracted; the input image data includes historical covered yarn image data and a second covered yarn processing parameter set; The coated yarn image feature extraction unit extracts the data features of the feature data to be extracted to generate coated yarn feature data; the coated yarn feature data includes the coated yarn problem features and the processing parameter abnormal features; The coated yarn image feature extraction unit includes an Attention mechanism, a dilated convolutional network, a residual convolutional network and a feature fusion layer; The covered yarn image feature classification unit performs classification according to the feature category of the covered yarn feature data to generate a covered yarn feature category and a covered yarn feature category probability; the covered yarn image feature classification unit is a Bayesian classifier; The covered yarn classification level output unit generates a weight level corresponding to the covered yarn feature category according to the covered yarn feature category probability.

3. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 2, characterized in that: The number of Attention mechanisms is 1; the number of dilated convolution networks is 3, and the dilated convolution adopts The convolution kernel has expansion interval parameters of 1, 2 and 3; the number of residual convolution networks is 3, and the residual convolution network includes 3 The convolution kernel and 2 The convolution kernel of the feature fusion layer is relu.

4. A method for detecting a covered yarn production process based on a lightweight target according to claim 2, characterized in that: The coated yarn image processing unit also includes: The covered yarn image processing unit adopts different preprocessing methods according to the type of input data; If the input data is the historical covered yarn image data, data denoising is performed by Gaussian filtering, data enhancement is performed on the denoised image according to histogram equalization, and the enhanced data is normalized to generate the feature data to be extracted of the historical covered yarn image data; If the input data is the second covered yarn processing parameter set, the data dimensions are unified through data supplementation, and the unified data are normalized to generate the feature data to be extracted of the second covered yarn processing parameter set.

5. A method for detecting a covered yarn production process based on a lightweight target according to claim 1, characterized in that: According to the characteristics of the covered yarn problem, feature classification is performed and corresponding labels are marked to generate a variety of label problem image sets including: The corresponding tags include a first covered yarn problem tag and a second covered yarn problem tag; Divide the coated yarn problem into multiple levels according to the number of first coated yarn problem labels, and divide the coated yarn problem into multiple levels according to the number of second coated yarn problem labels; divide the label problem image sets according to the coated yarn problem levels to generate multiple label problem image sets.

6. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 1, characterized in that: Performing data feature conversion according to the first covered yarn processing parameter set to generate a second covered yarn processing parameter set includes: All data of the first covered yarn processing parameter set are subjected to fast Fourier transform to obtain a plurality of frequency data and a plurality of amplitude data of each covered yarn problem data, each frequency data corresponds to one amplitude data; the first covered yarn processing parameter set is a plurality of time series data; the total number of the first covered yarn processing parameter set is m; Sort the amplitude data from large to small, and select the corresponding frequency data of the first n amplitude data as reference frequency data , obtain the cycle of each covered yarn problem data ,calculate Get n segmentation cycles of each of the covered yarn problem data ,in accordance with Divide each of the covered yarn problem data and splice the divided covered yarn problem data to obtain a quantity of The second covered yarn processing parameter set, each of the covered yarn image data size is .

7. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 1, characterized in that: Establish and initialize a problem association matrix for multiple labeled problem image sets and the second coated yarn processing parameter set, wherein the problem association matrix weights are the labels of the multiple labeled problem image sets, including: encoding the data of the second coated yarn processing parameter set, adding the corresponding levels of the problem feature annotations as the problem association matrix weights to the data encoding to generate multiple associated data, and sorting the multiple associated data to generate the problem association matrix; obtaining the data length of each associated data, using the maximum data length as reference data, supplementing each associated data until the data length of each associated data is consistent, and initializing the problem association matrix weights to complete the initialization of the problem association matrix.

8. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 1, characterized in that: Optimizing the problem association matrix weight according to the abnormal characteristics of the processing parameters includes: ； in, is the weight of the problem association matrix after the i-th optimization, n is the total number of the covered yarn image problem data formed after each of the covered yarn problem data is segmented, is the weight of the j-th corresponding problem association matrix, is the quantitative statistics of abnormal features of the processing parameters, The abnormal characteristics of the processing parameters identified by the composite analysis model of the covered yarn.

9. A method for detecting a covered yarn production process based on a lightweight target according to claim 1, characterized in that: Re-labeling the second covered yarn processing parameter set according to the optimized problem association matrix weights to generate new category labels including: Obtaining each covered yarn problem level of the first covered yarn problem label and each of the covered yarn problem levels of the second covered yarn problem label; According to the same problem label and the same covered yarn problem level, all the problem association matrix weights corresponding to each of the historical processing parameters are obtained; the multiple problem association matrix weights of the same historical processing parameters are normalized, and if the normalized problem association matrix weight is greater than 0.5, the problem association matrix weight is relabeled as 1; if the normalized problem association matrix weight is not greater than 0.5, the problem association matrix weight is relabeled as 0.

10. A method for detecting a production process of a covered yarn based on a lightweight target according to claim 1, characterized in that: Re-labeling the corresponding labels of the multiple label problem image sets according to the new category labels, and inputting the labels into the covered yarn composite analysis model for data training and feature learning, and optimizing the covered yarn composite analysis model for production detection includes: Performing secondary data encoding according to the data encoding of the second covered yarn processing parameter set and the re-labeled labels, generating the category labels of different levels, and marking the label problem image set; The re-labeled label problem image set is divided into 70% training set and 30% test set, and is input into the coated yarn composite analysis model for training and testing to generate new production detection model parameters.

Images