Quality evaluation method and device based on non-woven fabric defect detection

Abstract

The invention relates to a non-woven fabric defect detection-based quality evaluation method and device, and the method comprises the following steps: selecting a setting mode of a camera and a light source according to the material information of a non-woven fabric needing defect detection; acquiring environment data, non-woven fabric image data and corresponding label data on a non-woven fabric production line, and performing data analysis processing to obtain a non-woven fabric data set for training; constructing a quality evaluation model for defect detection, and training and optimizing the quality evaluation model in combination with the non-woven fabric data set to obtain a non-woven fabric quality evaluation model; and obtaining a target non-woven fabric image to be identified on the non-woven fabric production line and environment real-time data, and performing quality evaluation and analysis in combination with the non-woven fabric quality evaluation model. According to the quality evaluation method based on non-woven fabric defect detection, automation of the whole process from image acquisition, data analysis and processing to quality evaluation is achieved, manual intervention is reduced, and the detection accuracy and stability are improved.

Description

Technical Field

[0001] The present application relates to the technical field of non-woven fabric defect detection, and particularly relates to a quality evaluation method and device based on non-woven fabric defect detection. Background Art

[0002] As an important material, non-woven fabrics are widely used in many fields such as medical and health, packaging, clothing, and home furnishings. With the continuous growth of the demand for non-woven fabrics in various industries, higher requirements are put forward for their quality. The existence of defects will seriously affect the performance and quality of non-woven fabrics. For example, defects such as holes, broken filaments, stains, and wrinkles will reduce the strength, air permeability, and appearance beauty of non-woven fabrics. Therefore, it is crucial to accurately detect non-woven fabric defects and conduct quality evaluation. Traditional non-woven fabric defect detection methods mainly rely on manual visual inspection. Manual visual inspection has many drawbacks. The inspectors are prone to fatigue after long-term work, resulting in a decrease in the detection accuracy. Moreover, the subjective judgments of different inspectors vary greatly, making it difficult to ensure the consistency of the detection results. In addition, manual detection has low efficiency and cannot meet the rapid detection needs of large-scale production.

[0003] Therefore, there is an urgent practical need for an efficient, accurate, and comprehensive quality evaluation method based on non-woven fabric defect detection that can consider multiple factors. Summary of the Invention

[0004] In order to solve the above problems existing in the prior art, the purpose of the present application is to provide a quality evaluation method and device based on non-woven fabric defect detection. The quality evaluation method based on non-woven fabric defect detection realizes the whole process automation from image acquisition, data analysis and processing to quality evaluation, reduces manual intervention, reduces the errors caused by human factors, and improves the accuracy and stability of detection.

[0005] A quality evaluation method based on non-woven fabric defect detection described in the present application includes the following steps:

[0006] S1. Select the setting method of the camera and the light source according to the material information of the non-woven fabric to be detected for defects;

[0007] S2. Obtain the environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and conduct data analysis and processing to obtain a non-woven fabric data set for training;

[0008] S3. Construct a quality evaluation model for defect detection, and train and optimize the quality evaluation model in combination with the non-woven fabric data set to obtain a non-woven fabric quality evaluation model;

[0009] S4. Obtain the target non-woven fabric image to be recognized on the non-woven fabric production line and the real-time environmental data, and conduct quality evaluation and analysis in combination with the non-woven fabric quality evaluation model.

[0010] Preferably, the step S1 specifically includes:

[0011] The setting modes of the camera and the light source include forward setting and backward setting;

[0012] The forward setting means that the camera and the light source are on the same side of the non-woven fabric surface, and the optical axis of the camera forms an included angle α with the light rays emitted by the light source;

[0013] The backward setting means that the camera and the light source are respectively on both sides of the non-woven fabric, and the camera and the light source are oppositely arranged so that the optical axis of the camera is collinear with the light rays emitted by the light source;

[0014] Obtain the proportion C of chemical fibers, cf the proportion C of natural fibers, nf the thickness D and the surface roughness R a of the non-woven fabric, satisfying C cf + C nf = 1, and calculate the comprehensive index S for forward setting and the comprehensive index S for backward setting according to the following formulas f and the comprehensive index S for backward setting b :

[0015] S f = μ1 * C cf + μ2 * (1 - D) + μ3 * (1 - R a )

[0016] S b = ω1 * C nf + ω2 * D + ω3 * R a

[0017] where μ1, μ2, μ3, ω1, ω2 and ω3 are all weight coefficients, satisfying: μ1 + μ2 + μ3 = 1 and ω1 + ω2 + ω3 = 1;

[0018] If S f > S b , then select the forward setting;

[0019] If S f < S b , then select the backward setting;

[0020] If S f = S b , then select the forward setting or the backward setting.

[0021] Preferably, the step S2 specifically includes:

[0022] Obtain the environmental data on the non-woven fabric production line, and the environmental data includes temperature data, humidity data and tension data;

[0023] Obtain non-woven fabric image data and corresponding label data on a non-woven fabric production line. The non-woven fabric image data includes image samples with defect-free features, hole features, broken wire features, stain features, and wrinkle features. The labels corresponding to the defect-free features in the label data are 0, the labels for the hole features are 1, the labels for the broken wire features are 2, the labels for the stain features are 3, and the labels for the wrinkle features are 4;

[0024] Perform outlier processing and smoothing processing on the environmental data to obtain the processed environmental data;

[0025] Perform grayscale processing on the non-woven fabric image data using the weighted average method, remove the noise in the image using the median filtering algorithm, enhance the image contrast through histogram equalization technology, and extract the image edges using the Canny edge detection algorithm to obtain the processed non-woven fabric image data.

[0026] Preferably, step S2 further includes:

[0027] Construct a convolutional neural network model based on the LeNet structure, input the processed non-woven fabric image data into the convolutional neural network model for convolution operation, extract the local features of the non-woven fabric image, and obtain a feature map;

[0028] Perform downsampling on the feature map and integrate the extracted local features to obtain the spatial feature vector of the non-woven fabric image;

[0029] Construct a single-layer long short-term memory network model, and organize the continuously acquired non-woven fabric image sequences in the non-woven fabric image data into a time series data format to obtain time series data;

[0030] Input the time series data into the single-layer long short-term memory network model and process the time series data using the gating mechanism to obtain the time feature vector of the non-woven fabric image;

[0031] Calculate the statistical features of the processed environmental data. The statistical features include mean, variance, and change rate, to obtain the environmental data statistical features;

[0032] Connect and fuse the spatial feature vector, the time feature vector, and the environmental data statistical features to obtain the spatio-temporal fusion feature data;

[0033] Divide the spatio-temporal fusion feature data into a training set and a test set according to the ratio of A1:A2, and pair and integrate the image data and the corresponding label data of the training set and the test set to construct a non-woven fabric dataset for training.

[0034] Preferably, step S3 specifically includes:

[0035] Construct a quality evaluation model of a three-layer fully connected neural network, where the quality evaluation model includes an input layer, a hidden layer, and an output layer;

[0036] Determine the dimension of the spatio-temporal fusion feature data according to the principal component analysis method, and set the number of input neurons in the input layer to be the same as the dimension of the spatio-temporal fusion feature data;

[0037] Set the number of hidden neurons in the hidden layer to 1.5 times the number of input neurons, and use ReLU as the activation function of the hidden layer;

[0038] Set the number of output neurons in the output layer to 5, and use the Softmax function as the activation function of the output layer;

[0039] Set the number of training epochs to Q and the training batch size to e;

[0040] Initialize the parameters of the quality evaluation model, input the training set into the quality evaluation model according to the training batch size. The training set passes through the input layer, the hidden layer, and the output layer in sequence to obtain the predicted probability distribution of each sample corresponding to each label;

[0041] Use the cross-entropy loss function to analyze the difference between the probability distribution predicted by the quality evaluation model and the true label, and calculate the average loss value of each batch;

[0042] According to the calculated average loss value, use the backpropagation algorithm to calculate the gradients of the loss function with respect to each connection weight and bias in the model;

[0043] According to the calculated gradients of each connection weight and bias, use the stochastic gradient descent optimization algorithm to update the connection weights and biases of the model;

[0044] Repeat the training until the training epochs are completed.

[0045] Preferably, step S3 further includes:

[0046] Test the quality evaluation model that has completed the test rounds through the test set to obtain the recognition result, and calculate the accuracy rate E, recall rate R, F1 value, and test loss value L;

[0047] The calculation formula for the accuracy rate E is as follows:

[0048]

[0049] The calculation formula for the recall rate R is as follows:

[0050]

[0051] The calculation formula of the F1 value is as follows:

[0052]

[0053] The cross-entropy loss function is used to calculate the test loss value L;

[0054] where D represents the number of correctly recognized image samples; A represents the total number of image samples in the test set; X i represents the number of image samples that correctly recognize the i-th type of feature; M i represents the total number of image samples of the i-th type of feature; i = 1, 2, 3, 4, 5; i = 1 represents the non-defect feature; i = 2 represents the hole feature; i = 3 represents the broken wire feature; i = 4 represents the stain feature; i = 5 represents the wrinkle feature;

[0055] The model test comprehensive coefficient K is calculated by the following formula:

[0056] K = C1 * E + C2 * R + C3 * F1 + C4 * L

[0057] where C1, C2, C3, and C4 are all weight coefficients, satisfying C1 + C2 + C3 + C4 = 1;

[0058] If K ≥ 0.8, it is determined that the quality assessment model meets the requirements, and the quality assessment model is used as the non-woven fabric quality assessment model;

[0059] If K < 0.8, it is determined that the quality assessment model does not meet the requirements, and retraining is performed until K ≥ 0.8 is satisfied.

[0060] Preferably, the step S4 specifically includes:

[0061] Obtain the target non-woven fabric image to be recognized on the non-woven fabric production line, recognize the target non-woven fabric image through the non-woven fabric quality assessment model, and statistically obtain the first area ratio S K of the hole feature, the first number Z of the broken wire feature, the second area ratio S W of the stain feature, and the first length Y cm of the wrinkle feature;

[0062] If S K < 5%, Z < 5, S W < 3%, and Y < 10 are all satisfied, it is determined that the quality of the non-woven fabric is qualified;

[0063] If S K < 5%, Z < 5, S WIf any one of < 3% and Y < 10> is not satisfied, it is determined that the quality of the non-woven fabric is unqualified, and defect correlation analysis is carried out in combination with the real-time environmental data.

[0064] Preferably, the defect correlation analysis includes:

[0065] Obtain the real-time environmental data on the non-woven fabric production line, and the real-time environmental data includes the real-time temperature value W S , the real-time humidity value H S and the real-time tension value F S , and obtain the temperature change rate dW, the humidity change rate dH and the tension change rate dF in the ΔT time period;

[0066] If W S > W max and dW > 2, it is determined that the temperature is abnormal. When Z ≥ 5, it is determined that the temperature has an impact on the generation of the broken wire feature. When S W ≥ 3%, it is determined that the temperature has an impact on the generation of the stain feature;

[0067] If H S ≠ [H min , H max and dH > 5%, it is determined that the humidity is abnormal. When S K > 5%, it is determined that the humidity has an impact on the generation of the hole feature. When Y > 10, it is determined that the humidity has an impact on the generation of the wrinkle feature;

[0068] If F S ≠ [F min , F max and dF > 1.5, it is determined that the tension is abnormal. When Z ≥ 5, it is determined that the tension has an impact on the generation of the broken wire feature. When S K > 5%, it is determined that the tension has an impact on the generation of the hole feature;

[0069] wherein, W max represents the theoretical maximum temperature value; H min and H max respectively represent the theoretical minimum humidity value and the theoretical maximum humidity value; F min and F max respectively represent the standard minimum tension value and the standard maximum tension value; the unit of temperature is °C; the unit of tension is N.

[0070] Preferably, the defect correlation analysis further includes:

[0071] When there is the temperature abnormality, the temperature on the non-woven fabric production line is processed until the temperature returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is re-identified to obtain the second number Z' of the broken wire features and the third area ratio S W';

[0072] If Z' < 5, it is determined that the generation of the broken wire feature is related to temperature; otherwise, it is determined that the generation of the broken wire feature is not related to temperature, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0073] If S W ' < 3%, it is determined that the generation of the stain feature is related to temperature; otherwise, it is determined that the generation of the stain feature is not related to temperature, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0074] When there is the above humidity abnormality, the humidity on the non-woven fabric production line is processed until the humidity returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is recognized again to obtain the fifth area ratio S K ' of the hole feature and the second length Y' cm of the wrinkle feature;

[0075] If S K ' < 5%, it is determined that the generation of the hole feature is related to humidity; otherwise, it is determined that the generation of the hole feature is not related to humidity, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0076] If Y' < 10, it is determined that the generation of the wrinkle feature is related to humidity; otherwise, it is determined that the generation of the wrinkle feature is not related to humidity, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0077] When there is the above tension abnormality, the tension on the non-woven fabric production line is processed until the tension returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is recognized again to obtain the third quantity Z″ of the broken wire feature and the sixth area ratio S K ″;

[0078] If Z″ < 5, it is determined that the generation of the broken wire feature is related to tension; otherwise, it is determined that the generation of the broken wire feature is not related to tension, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0079] If S K ″ < 5%, it is determined that the generation of the hole feature is related to tension; otherwise, it is determined that the generation of the hole feature is not related to tension, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics.

[0080] The present application also provides a quality evaluation device based on non-woven fabric defect detection, including:

[0081] A component setting module, configured to select the setting method of the camera and the light source according to the material information of the non-woven fabric to be detected for defects;

[0082] A data acquisition and analysis module, which is used to obtain environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and perform data analysis and processing to obtain a non-woven fabric data set for training;

[0083] A model construction module, which is used to construct a quality evaluation model, and train and optimize the quality evaluation model in combination with the non-woven fabric data set to obtain a non-woven fabric quality evaluation model;

[0084] A quality evaluation and analysis module, which is used to obtain the target non-woven fabric image to be recognized on the non-woven fabric production line and real-time environmental data, and perform quality evaluation and analysis in combination with the non-woven fabric quality evaluation model.

[0085] The quality evaluation method and device based on non-woven fabric defect detection described in this application have the following advantages:

[0086] 1. The quality evaluation method based on non-woven fabric defect detection in this application can ensure the high quality and accuracy of image acquisition by selecting the setting methods of the camera and light source according to the material information of the non-woven fabric, which helps to reduce problems such as uneven illumination and reflection, thereby improving the accuracy of defect detection; by obtaining the environmental data, non-woven fabric image data and corresponding label data on the production line and performing data analysis and processing, a high-quality training data set can be generated, which helps the quality evaluation model to learn more accurate defect features, thereby improving the detection accuracy; by using the constructed quality evaluation model to train and optimize the training data set, an efficient and accurate non-woven fabric quality evaluation model can be obtained, which can quickly identify and analyze the defects on the non-woven fabric, significantly improving the detection efficiency. The quality evaluation method based on non-woven fabric defect detection realizes the full automation from image acquisition, data analysis and processing to quality evaluation, reduces manual intervention, reduces errors caused by human factors, and improves the accuracy and stability of detection.

[0087] 2. The quality evaluation device based on non-woven fabric defect detection in this application realizes the high automation and intelligence of non-woven fabric defect detection and quality evaluation through the collaborative work of the component setting module, data acquisition and analysis module, model construction module and quality evaluation and analysis module; it can select the setting methods of the camera and light source according to the material information of the non-woven fabric to be detected for defects, so that the device can adapt to the detection requirements of different types and specifications of non-woven fabrics; the quality evaluation and analysis module can perform high-precision quality evaluation and analysis on the target non-woven fabric image in combination with the non-woven fabric quality evaluation model; it can obtain the data on the non-woven fabric production line in real time and perform real-time quality evaluation and analysis, so that the problems in the production process can be discovered and solved in time, avoiding the production of unqualified products; and it can realize the optimal allocation of production resources, improving production efficiency and product quality. Description of the Drawings

[0088] Figure 1 It is a schematic flow chart of a quality assessment method based on non-woven fabric defect detection described in this application;

[0089] Figure 2 It is a schematic front view of the camera and the light source in a quality assessment method based on non-woven fabric defect detection described in this application;

[0090] Figure 3 It is a schematic rear view of the camera and the light source in a quality assessment method based on non-woven fabric defect detection described in this application.

[0091] Explanation of reference numerals:

[0092] 10 - Non-woven fabric;

[0093] 20 - Camera;

[0094] 30 - Light source. Detailed implementation manners

[0095] As Figure 1 shown, a quality assessment method based on non-woven fabric defect detection described in this application includes the following steps:

[0096] S1. Select the setting mode of the camera 20 and the light source 30 according to the material information of the non-woven fabric 10 to be detected for defects;

[0097] S2. Obtain the environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and perform data analysis and processing to obtain a non-woven fabric data set for training;

[0098] S3. Build a quality assessment model for defect detection, and train and optimize the quality assessment model in combination with the non-woven fabric data set to obtain a non-woven fabric quality assessment model;

[0099] S4. Obtain the target non-woven fabric image to be recognized on the non-woven fabric production line and the real-time environmental data, and perform quality assessment and analysis in combination with the non-woven fabric quality assessment model.

[0100] Furthermore, in this embodiment, step S1 specifically includes:

[0101] The setting modes of the camera 20 and the light source 30 include front setting and rear setting;

[0102] The front setting means that the camera 20 and the light source 30 are on the same side of the surface of the non-woven fabric 10, and the optical axis of the camera 20 forms an included angle α with the light emitted by the light source 30; Figure 2As shown, it is a forward setting, that is, both the camera 20 and the light source 30 are located above the surface of the non-woven fabric 10. The optical axis of the camera 20 points to the surface of the non-woven fabric 10 where the image needs to be collected, and the light rays emitted by the light source 30 also point to the surface of the non-woven fabric 10 where the image needs to be collected. Moreover, the optical axis of the camera 20 and the light rays emitted by the light source 30 form an included angle α. The forward setting can collect a relatively clear image of the non-woven fabric surface by increasing the reflected light on the surface of the non-woven fabric 10;

[0103] The back setting is that the camera 20 and the light source 30 are respectively located on both sides of the non-woven fabric 10, and the camera 20 and the light source 30 are relatively set so that the optical axis of the camera 20 and the light rays emitted by the light source 30 are collinear; Figure 3 As shown, it is a back setting, that is, the camera 20 is located above the surface of the non-woven fabric 10, and the light source 30 is located below the non-woven fabric 10, forming a relative setting of the camera 20 and the light source 30. The optical axis of the camera 20 points to the surface of the non-woven fabric 10 where the image needs to be collected, and the light rays emitted by the light source 30 point to the back of the non-woven fabric 10 and penetrate the non-woven fabric 10 and point to the camera 20. The back setting can improve the clarity and contrast of the collected non-woven fabric surface image by increasing the transmitted light;

[0104] Obtain the chemical fiber proportion C of the non-woven fabric 10 cf 、the natural fiber proportion C nf 、the thickness D and the surface roughness R a , satisfying C cf +C nf =1, and calculate the comprehensive index S of the forward setting according to the following formula f and the comprehensive index S of the back setting b :

[0105] S f =μ1*C cf +μ2*(1 - D)+μ3*(1 - R a )

[0106] S b =ω1*C nf +ω2*D+ω3*R a

[0107] Among them, μ1, μ2, μ3, ω1, ω2 and ω3 are all weight coefficients, satisfying: μ1 + μ2 + μ3 = 1 and ω1 + ω2 + ω3 = 1;

[0108] If S f >S b , then select the forward setting;

[0109] If S f <S b , then select the back setting;

[0110] If Sf = S b , select the forward setting or the backward setting, that is, arbitrarily select the forward setting or the backward setting according to the actual situation;

[0111] The angle formed by the optical axis of the camera 20 and the light emitted by the light source 30 is selected as follows:

[0112] Measure the gloss G on the surface of the non-woven fabric 10 using a gloss meter, measure the light intensity in the non-woven fabric production environment using an illuminometer, obtain the running speed V of the non-woven fabric 10 on the non-woven fabric production line, and according to the angle selection coefficient B α , the formula is as follows:

[0113] B α = λ1 * G + λ2 * U + λ3 * V

[0114] Among them, λ1 represents the gloss influence coefficient; λ2 represents the light intensity influence coefficient; λ3 represents the running speed influence coefficient;

[0115] If B α < 6, then the angle α is selected as 0 < α ≤ 15°; the optimal value is 15°;

[0116] If 6 ≤ B α < 8, then the angle α is selected as 15° < α ≤ 30°; the optimal value is 30°;

[0117] If 8 ≤ B α < 10, then the angle α is selected as 30° < α ≤ 45°; the optimal value is 45°;

[0118] If 10 ≤ B α < 12, then the angle α is selected as 45° < α ≤ 60°; the optimal value is 60°;

[0119] If 12 ≤ B α < 14, then the angle α is selected as 60° < α ≤ 75°; the optimal value is 70°;

[0120] If B α ≥ 14, then the angle α is selected as 75° < α ≤ 90°; the optimal value is 80°;

[0121] Examples are as follows:

[0122] Measure the proportion C of chemical fibers in the non-woven fabric 10 using an infrared spectrometer cf = 0.6 and the proportion C of natural fibers nf = 0.4, measure the thickness D of the non-woven fabric 10 as 0.3 mm using a thickness gauge, and measure the surface roughness R of the non-woven fabric 10 a = 0.2;

[0123] μ1 = 0.3, μ2 = 0.4, μ3 = 0.3, ω1 = 0.2, ω2 = 0.5, ω3 = 0.3;

[0124] S f = 0.3 * 0.6 + 0.4 * (1 - 0.3) + 0.3 * (1 - 0.2) = 0.7;

[0125] S b = 0.2 * 0.4 + 0.5 * 0.3 + 0.2 * 0.2 = 0.27;

[0126] So S f > S b , then select the forward setting;

[0127] Measure the glossiness G = 30 on the surface of the non-woven fabric 10 with a glossiness meter, measure the light intensity U = 100 Lux in the production environment of the non-woven fabric with an illuminometer, and obtain the running speed V = 10 m / min of the non-woven fabric 10 on the non-woven fabric production line;

[0128] λ1 = 0.1, λ2 = 0.05, λ3 = 0.1;

[0129] B α = 0.1 * 30 + 0.05 * 100 + 0.1 * 10 = 9, so the angle α is selected as 30° < α ≤ 45°, and the optimal choice is 45°.

[0130] Further, in this embodiment, step S2 specifically includes:

[0131] Obtain the environmental data on the non-woven fabric production line, and the environmental data includes temperature data, humidity data and tension data;

[0132] Obtain the non-woven fabric image data and the corresponding label data on the non-woven fabric production line. The non-woven fabric image data includes image samples with defect-free features, hole features, broken wire features, stain features and wrinkle features. The labels corresponding to the defect-free features in the label data are 0, the labels for hole features are 1, the labels for broken wire features are 2, the labels for stain features are 3, and the labels for wrinkle features are 4;

[0133] Perform outlier processing and smoothing processing on the environmental data to obtain the processed environmental data;

[0134] Perform gray-scale processing on the non-woven fabric image data using the weighted average method, remove the noise in the image using the median filtering algorithm, enhance the image contrast through the histogram equalization technique, and extract the image edges using the Canny edge detection algorithm to obtain the processed non-woven fabric image data.

[0135] Further, in this embodiment, step S2 further includes:

[0136] Construct a convolutional neural network model based on the LeNet structure, input the processed non-woven fabric image data into the convolutional neural network model for convolution operation, extract the local features of the non-woven fabric image, and obtain a feature map;

[0137] Perform downsampling on the feature map and integrate the extracted local features to obtain the spatial feature vector of the non-woven fabric image;

[0138] Construct a single-layer long short-term memory network model, and organize the continuously collected non-woven fabric image sequences in the non-woven fabric image data into a time series data format to obtain time series data;

[0139] By inputting the time series data into the single-layer long short-term memory network model and using a gating mechanism to process the time series data, obtain the time feature vector of the non-woven fabric image;

[0140] Calculate the statistical features of the processed environmental data. The statistical features include mean, variance, and change rate, and obtain the environmental data statistical features;

[0141] Connect and fuse the spatial feature vector, time feature vector, and environmental data statistical features to obtain spatio-temporal fusion feature data;

[0142] Divide the spatio-temporal fusion feature data into a training set and a test set according to the ratio of A1:A2, and pair and integrate the image data and the corresponding label data of the training set and the test set to construct a non-woven fabric dataset for training; where A1:A2 = 8:2, that is, A1 = 8 and A2 = 2.

[0143] Further, in this embodiment, step S3 specifically includes:

[0144] Construct a quality evaluation model of a three-layer fully connected neural network. The quality evaluation model includes an input layer, a hidden layer, and an output layer;

[0145] Determine the dimension of the spatio-temporal fusion feature data according to the principal component analysis method, and set the number of input neurons in the input layer to be the same as the dimension of the spatio-temporal fusion feature data;

[0146] Set the number of hidden neurons in the hidden layer to be 1.5 times the number of input neurons, and use ReLU as the activation function of the hidden layer;

[0147] Set the number of output neurons in the output layer to be 5, and use the Softmax function as the activation function of the output layer;

[0148] Set the number of training epochs to Q and the number of training batches to e; select Q = 50 times and e = 32 batches;

[0149] Initialize the parameters of the quality assessment model, input the training set into the quality assessment model in training batches. The training set passes through the input layer, hidden layer, and output layer in sequence to obtain the predicted probability distribution of each sample corresponding to each label.

[0150] Use the cross-entropy loss function to analyze the difference between the probability distribution predicted by the quality assessment model and the true labels, and calculate the average loss value for each batch.

[0151] According to the calculated average loss value, use the backpropagation algorithm to calculate the gradients of the loss function with respect to each connection weight and bias in the model.

[0152] According to the calculated gradients of each connection weight and bias, use the stochastic gradient descent optimization algorithm to update the connection weights and biases of the model.

[0153] Repeat the training until the training rounds are completed.

[0154] Furthermore, in this embodiment, step S3 further includes:

[0155] Test the quality assessment model that has completed the test rounds through the test set to obtain the recognition results, and calculate the accuracy rate E, recall rate R, F1 value, and test loss value L.

[0156] The calculation formula for the accuracy rate E is as follows:

[0157]

[0158] The calculation formula for the recall rate R is as follows:

[0159]

[0160] The calculation formula for the F1 value is as follows:

[0161]

[0162] Use the cross-entropy loss function to calculate the test loss value L.

[0163] Among them, D represents the number of correctly recognized image samples; A represents the total number of image samples in the test set; X i represents the number of image samples that correctly recognize the i-th type of feature; M i represents the total number of image samples of the i-th type of feature; i = 1, 2, 3, 4, 5; i = 1 represents the non-defect feature; i = 2 represents the hole feature; i = 3 represents the broken wire feature; i = 4 represents the stain feature; i = 5 represents the wrinkle feature.

[0164] Calculate the model test comprehensive coefficient K through the following formula:

[0165] K = C1 * E + C2 * R + C3 * F1 + C4 * L

[0166] Among them, C1, C2, C3, and C4 are all weight coefficients, satisfying C1 + C2 + C3 + C4 = 1;

[0167] If K ≥ 0.8, it is determined that the quality evaluation model meets the requirements, and the quality evaluation model is used as the non-woven fabric quality evaluation model;

[0168] If K < 0.8, it is determined that the quality evaluation model does not meet the requirements, and retraining is carried out until K ≥ 0.8 is satisfied;

[0169] The example is as follows:

[0170] The total number of image samples A in the test set = 200, and the number of correctly recognized image samples D = 180, then E = 90%;

[0171] M1 = 40, X1 = 37, M2 = 30, X2 = 26, M3 = 30, X3 = 28, M4 = 50, X4 = 45, M5 = 50, X5 = 44, then R = 90.1%;

[0172] F1 = 90%, L = 0.2, C1 = 0.4, C2 = 0.4, C3 = 0.1, C4 = 0.1;

[0173] Then K = 0.83 > 0.8, it is determined that the quality evaluation model meets the requirements, and the quality evaluation model is used as the non-woven fabric quality evaluation model.

[0174] Furthermore, in this embodiment, step S4 specifically includes:

[0175] Obtain the target non-woven fabric image to be recognized on the non-woven fabric production line, identify the target non-woven fabric image through the non-woven fabric quality evaluation model, and statistically obtain the first area ratio S of the hole feature K , the first number Z of broken wire features, the second area ratio S of the stain feature W and the first length Y cm of the wrinkle feature;

[0176] If S K < 5%, Z < 5, S W < 3% and Y < 10 are all satisfied, it is determined that the non-woven fabric is of qualified quality;

[0177] If S K < 5%, Z < 5, S W < 3% and Y < 10 are not satisfied for any one item, it is determined that the non-woven fabric is of unqualified quality, and defect correlation analysis is carried out in combination with the real-time environmental data;

[0178] The example is as follows:

[0179] The first area ratio S of the hole feature is statistically obtained K = 3%, the first quantity Z of the broken wire feature = 3 pieces, the second area ratio S of the stain feature W = 2% and the first length Y of the wrinkle feature = 8 cm;

[0180] Then S K < 5%, Z < 5, S W < 3% and Y < 10 are all satisfied, and it is determined that the quality of the non-woven fabric is qualified.

[0181] Furthermore, in this embodiment, the defect correlation analysis includes:

[0182] Obtain the real-time environmental data on the non-woven fabric production line. The real-time environmental data includes the real-time temperature value W S , the real-time humidity value H S and the real-time tension value F S , and obtain the temperature change rate dW, humidity change rate dH and tension change rate dF in the ΔT time period;

[0183] If W S > W max and dW > 2, it is determined that the temperature is abnormal. When Z ≥ 5, it is determined that the temperature has an impact on the generation of the broken wire feature. When S W ≥ 3%, it is determined that the temperature has an impact on the generation of the stain feature; among them, the environment being at a high temperature or the temperature being unstable is a factor leading to broken wires; the environment being at a high temperature will cause some substances to adhere to the surface of the non-woven fabric more easily, thus forming stains;

[0184] If H S ≠ [H min , H max and dH > 5%, it is determined that the humidity is abnormal. When S K > 5%, it is determined that the humidity has an impact on the generation of the hole feature. When Y > 10, it is determined that the humidity has an impact on the generation of the wrinkle feature; among them, inappropriate humidity will affect the forming process of the non-woven fabric, resulting in the generation of hole features; humidity changes will cause the non-woven fabric material to shrink or expand unevenly, thus generating wrinkle features;

[0185] If F S ≠ [F min , F max and dF > 1.5, it is determined that the tension is abnormal. When Z ≥ 5, it is determined that the tension has an impact on the generation of the broken wire feature. When S K > 5%, it is determined that the tension has an impact on the generation of the hole feature; among them, unstable tension on the non-woven fabric on the equipment will cause fiber breakage or damage to the non-woven fabric structure, and it is easy to generate broken wire features and hole features;

[0186] Among them, W max represents the theoretical maximum temperature value; Hmin and H max represent the theoretical minimum humidity value and the theoretical maximum humidity value respectively; F min and F max represent the standard minimum tension value and the standard maximum tension value respectively; the unit of temperature is °C; the unit of tension is N;

[0187] Examples are as follows:

[0188] The real-time temperature value W S = 40 °C, the real-time humidity value H S = 55%, the real-time tension value F S = 35 N;

[0189] ΔT = 10 min, W max = 35 °C, H min = 20%, H max = 50%, F min = 15 N, F max = 30;

[0190] Within the time of ΔT, the temperature rises from 20 °C to 45 °C, then the temperature change rate dW = 2.5 °C / min, so W S > W max and dW > 2, it is determined that the temperature is abnormal;

[0191] The first quantity Z of the broken wire feature = 6 pieces > 5 pieces, then it is determined that the temperature has an impact on the generation of the broken wire feature;

[0192] The second area ratio S of the stain feature W = 4% > 3%, then it is determined that the temperature has an impact on the generation of the stain feature

[0193] Within the time of ΔT, the humidity changes from 35% to 55%, then the humidity change rate dH = 2% / min, so H S ≠ [H min , H max and dH > 5%, then it is determined that the humidity is abnormal;

[0194] The first area ratio S of the hole feature K = 6% > 5%, then it is determined that the humidity has an impact on the generation of the hole feature;

[0195] The first length Y of the wrinkle feature = 12 cm > 10 cm, it is determined that the humidity has an impact on the generation of the wrinkle feature;

[0196] Within the time of ΔT, the tension changes from 15 N to 35 N, then the tension change rate dF = 2 N / min, so dF ≠ [F min , F max and dF > 1.5, then it is determined that the tension is abnormal;

[0197] If the first quantity Z of the broken wire feature = 6 pieces > 5 pieces, it is determined that the tension has an impact on the generation of the broken wire feature;

[0198] The first area proportion S of the hole feature K = 6% > 5%, it is determined that the tension has an impact on the generation of the hole feature.

[0199] Furthermore, in this embodiment, the defect correlation analysis further includes:

[0200] When there is a temperature anomaly, the temperature on the non-woven fabric production line is processed until the temperature returns to normal, and the new target non-woven fabric image of the same batch of non-woven fabrics is recognized again to obtain the second quantity Z' of the broken wire feature and the third area proportion S of the stain feature W ';

[0201] If Z' < 5, it is determined that the generation of the broken wire feature is related to the temperature, otherwise it is determined that the generation of the broken wire feature is not related to the temperature, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0202] If S W ' < 3%, it is determined that the generation of the stain feature is related to the temperature, otherwise it is determined that the generation of the stain feature is not related to the temperature, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0203] When there is a humidity anomaly, the humidity on the non-woven fabric production line is processed until the humidity returns to normal, and the new target non-woven fabric image of the same batch of non-woven fabrics is recognized again to obtain the fifth area proportion S of the hole feature K ' and the second length Y' cm of the wrinkle feature;

[0204] If S K ' < 5%, it is determined that the generation of the hole feature is related to the humidity, otherwise it is determined that the generation of the hole feature is not related to the humidity, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0205] If Y' < 10, it is determined that the generation of the wrinkle feature is related to the humidity, otherwise it is determined that the generation of the wrinkle feature is not related to the humidity, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0206] When there is a tension anomaly, the tension on the non-woven fabric production line is processed until the tension returns to normal, and the new target non-woven fabric image of the same batch of non-woven fabrics is recognized again to obtain the third quantity Z'' of the broken wire feature and the sixth area proportion S of the hole feature K '';

[0207] If Z'' < 5, it is determined that the generation of the broken wire feature is related to the tension, otherwise it is determined that the generation of the broken wire feature is not related to the tension, and it is determined that there is a quality unqualified problem with this batch of non-woven fabrics;

[0208] If S K ″ < 5%, it is determined that the generation of the hole feature is related to the tension; otherwise, it is determined that the generation of the hole feature is not related to the tension, and it is determined that there is a quality non - compliance problem with this batch of non - woven fabrics;

[0209] Examples are as follows:

[0210] The real - time temperature value W of the non - woven fabric production line is detected S = 40 °C, the theoretical maximum temperature value W max = 35 °C, and within the time period of ΔT = 10 min, the temperature change rate dW = 2.5 °C, satisfying the temperature anomaly determination condition W S > W max and dW > 2;

[0211] At this time, the temperature of the non - woven fabric production line is processed to restore it to the normal range, for example, 25 °C, and then the new target non - woven image in the same batch of non - woven fabrics is recognized again, and the second quantity Z′ of the broken - wire feature is obtained as 3 pieces, and the third area proportion S W ′ = 2%;

[0212] Because Z′ = 3 < 5, it is determined that the generation of the broken - wire feature is related to the temperature;

[0213] Because S W ′ = 2% < 3%, it is determined that the generation of the stain feature is related to the temperature;

[0214] The real - time humidity value H of the non - woven fabric production line is detected S = 55%, the theoretical minimum humidity value H min = 20%, the theoretical maximum humidity value H max = 50%, and within the time period of ΔT = 10 min, the humidity change rate is 2% / min, satisfying the humidity anomaly determination condition H S ≠ [H min , H max and dH > 5%;

[0215] At this time, the humidity of the non - woven fabric production line is adjusted to restore it to normal, for example, adjusted to 30%, and then the new target non - woven image in the same batch of non - woven fabrics is recognized again, and the fifth area proportion S K ′ = 4%, and the second length Y′c of the wrinkle feature is 8 cm;

[0216] Because S K ′ = 4% < 5%, it is determined that the generation of the hole feature is related to the humidity;

[0217] Because Y′ = 8 < 10, it is determined that the generation of the wrinkle feature is related to the humidity;

[0218] The real-time value F of the tension in the non-woven fabric production line is detected S = 35 N, the minimum standard tension F min = 15 N, the maximum standard tension F max = 30, and within the time period ΔT = 10 min, the tension change rate dF = 2 N / min, satisfying the abnormal tension determination condition dF ≠ [F min , F max , and dF > 1.5;

[0219] At this time, the tension of the non-woven fabric production line is adjusted to restore it to the normal range, such as 22 N, and then the new target non-woven fabric image in the same batch of non-woven fabrics is recognized again, and the third quantity Z″ of the broken wire feature = 4 pieces, and the sixth area ratio S of the hole feature K ″ = 4%.

[0220] Because Z″ = 4 < 5, it is determined that the generation of the broken wire feature is related to the tension;

[0221] Because S K ″ = 4% < 5%, it is determined that the generation of the hole feature is related to the tension.

[0222] This application also proposes a quality evaluation device based on non-woven fabric defect detection, including:

[0223] A component setting module for selecting the setting method of the camera and the light source according to the material information of the non-woven fabric to be detected for defects;

[0224] A data acquisition and analysis module for obtaining the environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and performing data analysis and processing to obtain a non-woven fabric data set for training;

[0225] A model construction module for constructing a quality evaluation model, training and optimizing the quality evaluation model in combination with the non-woven fabric data set to obtain a non-woven fabric quality evaluation model;

[0226] A quality evaluation and analysis module for obtaining the target non-woven fabric image to be recognized on the non-woven fabric production line and the real-time environmental data, and performing quality evaluation and analysis in combination with the non-woven fabric quality evaluation model.

[0227] In the description of the present application, it should be understood that the orientation or positional relationship indicated by orientation words such as "front, back, up, down, left, right", "horizontal, vertical, level" and "top, bottom", etc. is usually based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present application and simplifying the description. Without contrary explanation, these orientation words do not indicate or imply that the device or element referred to must have a specific orientation or be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation on the protection scope of the present application.

[0228] For those skilled in the art, various corresponding changes and deformations can be made according to the technical solutions and concepts described above, and all such changes and deformations should fall within the protection scope of the claims of the present application.

Claims

1. A quality assessment method based on non-woven fabric defect detection, characterized in that: The following steps are involved: S1. Selecting a setting method of a camera (20) and a light source (30) according to material information of a non-woven fabric (10) to be inspected for defects; S2, obtaining environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and performing data analysis and processing to obtain a non-woven fabric data set for training; S3, constructing a quality assessment model for defect detection, training and optimizing the quality assessment model in combination with the non-woven fabric data set, and obtaining a non-woven fabric quality assessment model; S4. Acquire the target non-woven fabric image to be identified on the non-woven fabric production line and the real-time environmental data, and perform quality assessment and analysis in combination with the non-woven fabric quality assessment model.

2. The quality assessment method based on nonwoven fabric defect detection according to claim 1, characterized in that: The step S1 specifically includes: The camera (20) and the light source (30) are arranged in a forward direction and a backward direction; The forward direction is set such that the camera (20) and the light source (30) are located on the same side of the surface of the non-woven fabric (10), and the optical axis of the camera (20) and the light emitted by the light source (30) form an angle α of a certain angle; The back-facing arrangement is such that the camera (20) and the light source (30) are respectively located on two sides of the non-woven fabric (10), and the camera (20) and the light source (30) are arranged relative to each other so that the optical axis of the camera (20) is collinear with the light emitted by the light source (30); Get the chemical fiber proportion C of the nonwoven fabric (10) cf 、Natural fiber proportion C nf , thickness D and surface roughness R a , satisfying C cf +C nf =1, and calculate the forward setting comprehensive index S according to the following formula f And the reverse setting comprehensive indicator S b : S f =μ1*C cf +μ2*(1-D)+μ3*(1-R a ) S b =ω1*C nf +ω2*D+ω3*R a Among them, μ1, μ2, μ3, ω1, ω2 and ω3 are all weight coefficients, satisfying: μ1+μ2+μ3=1 and ω1+ω2+ω3=1; If S f >S b , then select the forward setting; If S f <S b , then select the back-facing setting; If S f =S b , then select the forward setting or the reverse setting.

3. The quality assessment method based on nonwoven fabric defect detection according to claim 1, characterized in that: The step S2 specifically includes: Acquiring environmental data on a nonwoven fabric production line, wherein the environmental data includes temperature data, humidity data, and tension data; Acquire non-woven fabric image data and corresponding label data on a non-woven fabric production line, wherein the non-woven fabric image data includes image samples with a no-defect feature, an image sample with a hole feature, an image sample with a broken thread feature, an image sample with a stain feature, and an image sample with a wrinkle feature, wherein the label corresponding to the no-defect feature in the label data is 0, the label corresponding to the hole feature is 1, the label corresponding to the broken thread feature is 2, the label corresponding to the stain feature is 3, and the label corresponding to the wrinkle feature is 4; Performing outlier processing and smoothing processing on the environmental data to obtain processed environmental data; The non-woven fabric image data is gray-scaled by using a weighted average method, the noise in the image is removed by using a median filter algorithm, the image contrast is enhanced by using a histogram equalization technique, and the image edge is extracted by using a Canny edge detection algorithm to obtain processed non-woven fabric image data.

4. The quality assessment method based on nonwoven fabric defect detection according to claim 3 is characterized in that: The step S2 further comprises: Constructing a convolutional neural network model based on the LeNet structure, inputting the processed non-woven fabric image data into the convolutional neural network model for convolution operation, extracting local features of the non-woven fabric image, and obtaining a feature map; Downsampling the feature map, and integrating the extracted local features to obtain a spatial feature vector of the non-woven fabric image; A single-layer long short-term memory network model is constructed, and the continuously collected non-woven fabric image sequence in the non-woven fabric image data is sorted into a time series data format to obtain time series data; The time series data is input into the single-layer long short-term memory network model, and the time series data is processed by a gating mechanism to obtain a time feature vector of the non-woven fabric image; Calculating statistical characteristics of the processed environmental data, wherein the statistical characteristics include mean, variance and change rate, to obtain statistical characteristics of the environmental data; Connecting and fusing the spatial feature vector, the temporal feature vector and the environmental data statistical feature to obtain spatiotemporal fusion feature data; The spatiotemporal fusion feature data is divided into a training set and a test set according to the ratio of A1:A2, and the image data of the training set and the test set are paired and integrated with the corresponding label data to construct a non-woven fabric dataset for training.

5. The quality assessment method based on nonwoven fabric defect detection according to claim 4 is characterized in that: The step S3 specifically includes: Constructing a quality assessment model of a three-layer fully connected neural network, wherein the quality assessment model includes an input layer, a hidden layer, and an output layer; Determine the dimension of the spatiotemporal fusion feature data according to a principal component analysis method, and set the number of input neurons in the input layer to be the same as the dimension of the spatiotemporal fusion feature data; The number of hidden neurons in the hidden layer is set to 1.5 times the number of input neurons, and ReLU is used as the hidden layer activation function; The number of output neurons in the output layer is set to 5, and the Softmax function is used as the activation function of the output layer; Set the training round to Q and the training batch to e; Initializing the parameters of the quality assessment model, inputting the training set into the quality assessment model according to the training batches, and sequentially passing the training set through the input layer, the hidden layer, and the output layer to obtain the predicted probability distribution of each sample corresponding to each label; Using the cross entropy loss function to analyze the difference between the probability distribution predicted by the quality assessment model and the true label, and calculating the average loss value of each batch; Based on the calculated average loss value, the back propagation algorithm is used to calculate the gradient of the loss function relative to the weights and biases of each connection in the model; According to the calculated gradients of each connection weight and bias, the stochastic gradient descent optimization algorithm is used to update the connection weights and biases of the model; The training is repeated until the training round is completed.

6. The quality assessment method based on nonwoven fabric defect detection according to claim 5, characterized in that: The step S3 further comprises: The quality assessment model that has completed the test round is tested using the test set to obtain a recognition result, and the accuracy E, recall R, F1 value, and test loss value L are calculated; The calculation formula of the accuracy rate E is as follows: The recall rate R is calculated as follows: The calculation formula of the F1 value is as follows: The test loss value L is calculated using a cross entropy loss function; Where D represents the number of correctly identified image samples; A represents the total number of image samples in the test set; X i Indicates the number of image samples that correctly identify the i-th type of features; M i The total number of image samples representing the i-th type of feature; i=1, 2, 3, 4, 5; i=1 represents the defect-free feature; i=2 represents the hole feature; i=3 represents the broken wire feature; i=4 represents the stain feature; i=5 represents the wrinkle feature; The comprehensive coefficient K of the model test is calculated by the following formula: K=C1*E+C2*R+C3*F1+C4*L Among them, C1, C2, C3 and C4 are all weight coefficients, satisfying C1+C2+C3+C4=1; If K≥0.8, the quality assessment model is determined to meet the requirements, and the quality assessment model is used as the non-woven fabric quality assessment model; If K is less than 0.8, it is determined that the quality assessment model does not meet the requirements and is retrained until K≥0.8 is satisfied.

7. The quality assessment method based on nonwoven fabric defect detection according to claim 6, characterized in that: The step S4 specifically includes: Obtain a target nonwoven fabric image to be identified on a nonwoven fabric production line, identify the target nonwoven fabric image using the nonwoven fabric quality assessment model, and obtain the first area ratio S of the hole feature by statistics. K , the first number Z of the broken wire feature, the second area proportion S of the stain feature W and a first length Ycm of the wrinkle feature; If S K <5%、Z<5、S W If both <3% and Y <10 are satisfied, the nonwoven fabric is judged to be of qualified quality; If S K <5%、Z<5、S W If any one of <3% and Y<10 is not satisfied, the nonwoven fabric is judged to be unqualified in quality, and defect correlation analysis is performed in combination with the environmental real-time data.

8. The quality assessment method based on nonwoven fabric defect detection according to claim 7, characterized in that: The defect correlation analysis includes: Acquire the real-time environmental data on the non-woven fabric production line, the real-time environmental data includes the real-time temperature value W S , humidity real-time value H S And tension real-time value F S , and obtain the temperature change rate dW, humidity change rate dH and tension change rate dF during the ΔT period; If W S >W max If dW>2, the temperature is considered abnormal. If Z≥5, the temperature is considered to have an impact on the occurrence of the broken wire feature. W When ≥3%, it is determined that the temperature has an effect on the generation of the stain characteristics; If H S ≠[H min , H max ] and dH>5%, the humidity is judged to be abnormal. K When Y>5%, it is determined that the humidity has an effect on the generation of the hole feature, and when Y>10, it is determined that the humidity has an effect on the generation of the wrinkle feature; If F S ≠[F min , F max ] and dF>1.5, the tension is judged to be abnormal. When Z≥5, the tension is judged to have an impact on the generation of the broken wire feature. When S K >5% determines that the tension has an effect on the generation of the hole characteristics; Among them, W max Indicates the theoretical maximum temperature value; H min and H max Respectively represent the theoretical minimum humidity value and the theoretical maximum humidity value; F min and F max They represent the minimum and maximum standard tension respectively; the unit of temperature is ℃; the unit of tension is N.

9. The quality assessment method based on nonwoven fabric defect detection according to claim 8, characterized in that: The defect correlation analysis also includes: When the temperature anomaly exists, the temperature on the non-woven fabric production line is processed until the temperature returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is re-identified to obtain the second number Z' of the broken thread feature and the third area ratio S of the stain feature. W ′; If Z′<5, it is determined that the generation of the broken wire feature is related to the temperature, otherwise it is determined that the generation of the broken wire feature is not related to the temperature, and it is determined that the batch of non-woven fabrics has a quality problem; If S W '<3%, it is determined that the generation of the stain characteristics is related to temperature, otherwise it is determined that the generation of the stain characteristics is not related to temperature, and it is determined that the batch of non-woven fabrics has a quality problem; When the humidity is abnormal, the humidity on the non-woven fabric production line is processed until the humidity returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is re-identified to obtain the fifth area ratio S of the hole feature. K ' and a second length Y'cm of the wrinkle feature; If S K '<5%, it is determined that the generation of the hole characteristics is related to humidity, otherwise it is determined that the generation of the hole characteristics is not related to humidity, and it is determined that the batch of non-woven fabrics has a quality problem; If Y′<10, it is determined that the generation of the wrinkle feature is related to humidity, otherwise it is determined that the generation of the wrinkle feature is not related to humidity, and it is determined that the batch of non-woven fabrics has a quality problem; When the tension is abnormal, the tension on the non-woven fabric production line is processed until the tension returns to normal, and the new target non-woven fabric image in the same batch of non-woven fabrics is re-identified to obtain the third number Z″ of the broken wire feature and the sixth area ratio S of the hole feature. K ″; If Z″<5, it is determined that the generation of the broken wire feature is related to the tension, otherwise it is determined that the generation of the broken wire feature is not related to the tension, and it is determined that the batch of non-woven fabrics has a quality problem; If S K If the value is less than 5%, it is determined that the generation of the hole characteristics is related to the tension; otherwise, it is determined that the generation of the hole characteristics is not related to the tension, and it is determined that the batch of non-woven fabrics has a quality problem.

10. A quality assessment device based on non-woven fabric defect detection, characterized in that: include: A component setting module, used for selecting a setting mode of a camera (20) and a light source (30) according to material information of a non-woven fabric (10) to be inspected for defects; The data acquisition and analysis module is used to obtain environmental data, non-woven fabric image data and corresponding label data on the non-woven fabric production line, and perform data analysis and processing to obtain a non-woven fabric data set for training; A model building module, used to build a quality assessment model, train and optimize the quality assessment model in combination with the non-woven fabric data set, and obtain a non-woven fabric quality assessment model; The quality assessment and analysis module is used to obtain the target non-woven fabric image to be identified on the non-woven fabric production line and the real-time environmental data, and perform quality assessment and analysis in combination with the non-woven fabric quality assessment model.

Images