Carbon fiber cloth surface flaw detection device

Abstract

The invention relates to a carbon fiber cloth surface defect detection device which solves the problems that the accuracy and speed of carbon fiber cloth surface defect detection are improved, defects of rolled carbon fiber cloth finished products are not easy to find in the using process of downstream customers, and normal production and product quality are affected when defective carbon fiber cloth is used. The device comprises an encoder, a CIS line-scan digital camera, a marking machine and a controller, the encoder is configured to be connected to a rotating roller of carbon fiber cloth production equipment, the CIS line-scan digital camera is arranged above to-be-detected carbon fiber cloth, the marking machine is arranged at the position close to the carbon fiber cloth, and the controller is connected with the CIS line-scan digital camera. The marking machine is located behind the CIS line-scan digital camera in the moving direction of the carbon fiber cloth to be detected, a certain distance exists between the marking machine and the CIS line-scan digital camera, the encoder is electrically connected with the controller through a signal line, and the CIS line-scan digital camera is electrically connected with the controller. The method is widely applied to a carbon fiber cloth production process.

Description

Technical Field

[0001] The present invention relates to the technical field of defect detection in a carbon fiber cloth production process, and in particular to a device for detecting surface defects of a carbon fiber cloth. Background Art

[0002] Carbon fiber cloth is also known as carbon fiber cloth, carbon fiber cloth, carbon fiber woven cloth, carbon fiber prepreg cloth, carbon fiber reinforced cloth, carbon fiber fabric, carbon fiber tape, etc. With the continuous development of the carbon fiber cloth industry, the application fields of carbon fiber cloth are becoming more and more extensive.

[0003] Carbon fiber cloth is an advanced composite material made by impregnating carbon fibers with a matrix resin. The production process involves weaving and impregnating multiple carbon fiber strands through appropriate equipment. For details on the relevant production equipment and processes, please refer to the invention patent application with publication number CN117005209A, the utility model patent with grant publication number CN217972035U, and the utility model patent with grant publication number CN216613292U.

[0004] Due to various reasons, finished carbon fiber cloth is prone to surface defects such as white spots and cracks. Therefore, factories need to detect these surface defects during the production process and take measures when defects are found. Conventional detection methods (such as visual inspection, ultrasonic testing, and X-ray testing) have low accuracy and slow detection speed. Therefore, the detection methods need to be improved to meet actual production needs.

[0005] In addition, the finished carbon fiber cloth is rolled into a roll by a winding device and then used by downstream customers. It is not easy for downstream customers to find defects during use. The use of defective carbon fiber cloth will affect normal production and product quality. Summary of the Invention

[0006] The present invention aims to solve the technical problems of how to improve the accuracy and speed of carbon fiber cloth surface defect detection, and how to prevent defects of rolled carbon fiber cloth products from being difficult to be discovered during use by downstream customers, and the use of defective carbon fiber cloth affecting normal production and product quality. The present invention provides a carbon fiber cloth surface defect detection device with high detection accuracy, fast speed, and the ability to mark defects.

[0007] The present invention provides a carbon fiber cloth surface defect detection device, comprising an encoder, a CIS linear array camera, a marking machine, an image acquisition card and a controller;

[0008] The encoder is configured to be connected to a rotating roller of a carbon fiber cloth production device. The CIS line array camera is set above the carbon fiber cloth to be inspected. The marking machine is set near the carbon fiber cloth. The marking machine is located behind the CIS line array camera along the moving direction of the carbon fiber cloth to be inspected, and there is a certain distance between the marking machine and the CIS line array camera.

[0009] The encoder is electrically connected to the controller through a signal line, the image acquisition card is electrically connected to the controller, and the CIS line array camera is electrically connected to the image acquisition card.

[0010] Preferably, the working process of the carbon fiber cloth surface defect detection device is:

[0011] The carbon fiber cloth production equipment and winding device are running, the roller of the carbon fiber cloth production equipment rotates, the front end of the carbon fiber cloth is connected to the winding device, and the carbon fiber cloth continuously moves toward the winding device on the horizontal plane; the encoder generates a pulse signal and sends it to the controller, which receives the pulse signal and performs calculations;

[0012] The CIS linear array camera acquires an image of the carbon fiber cloth and sends the image data to the image acquisition card 6. The image acquisition card then transmits the image data to the controller, which then uses the YOLOv8 object detection model to identify defects in the image.

[0013] When the YOLOv8 target detection model identifies a defect, the controller starts counting the number n of pulse signals output by the encoder. When the number n reaches N, the controller controls the marking machine to mark the edge of the carbon fiber cloth.

[0014] Preferably, the YOLOv8 object detection model is built in the following way:

[0015] The first step is to build a database;

[0016] Use CIS line array cameras to collect a large number of images of carbon fiber cloth on the production line;

[0017] The second step is to label the large number of images mentioned above.

[0018] Use annotation tools to annotate images in detail and create a dataset of 820 annotated images.

[0019] The third step is to train the target detection model;

[0020] Based on the conventional YOLOv8 model, the model's convolutional layer uses depthwise separable convolution to replace the traditional convolution operation. Depthwise separable convolution includes depthwise convolution and pointwise convolution. The depthwise convolution process performs convolution operations on each input channel independently, while the pointwise convolution process uses a 1×1 convolution kernel to integrate the features of each channel. The standard convolution layer is decomposed into two stages: depthwise convolution and pointwise convolution.

[0021] The model is trained using the dataset generated in the second step. During the training process, CIoULoss and BCELoss are used as loss functions for the bounding box regression and category prediction branches, respectively, and the AdamW optimizer is adopted.

[0022] After the marking machine has been working for a period of time, the controller processes the identified defects, forms a diagram showing the defect location, and displays the diagram on the screen. The operator observes the diagram on the screen and conducts targeted inspections of the corresponding processes based on the distribution of the defect locations in the diagram in order to find the cause of the defect.

[0023] The beneficial effects of the present invention are: using a CIS linear array camera to capture images ensures the rapidity and real-time nature of the detection process; and the YOLOv8 target detection model enables high detection accuracy, high precision, and high speed.

[0024] The improved YOLOv8 target detection model introduces Depthwise Separable Convolution, which decomposes the standard convolution layer in the network structure of the conventional YOLOv8 target detection model into two steps: depthwise convolution and point-by-point convolution. This reduces the number of model parameters and the computational complexity while maintaining detection accuracy.

[0025] The invention meets the high standard requirements of carbon fiber cloth industrial production and is beneficial to improving production efficiency and reducing production costs.

[0026] Further features and aspects of the present invention will be clearly described in the following description of specific embodiments with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 This is a schematic diagram of the structure of the CIS line array camera detecting carbon fiber cloth;

[0028] Figure 2 It is the principle diagram of the detection and control process;

[0029] Figure 3 This is a sample diagram of crack defects;

[0030] Figure 4 yes Figure 1 In the structure shown, the positional relationship between the CIS line array camera, marking machine, and carbon fiber cloth;

[0031] Figure 5 It is a defect location map shown on the display.

[0032] Explanation of symbols in the figure:

[0033] 1. Roller, 2. CIS line array camera, 3. Marking machine, 4. Carbon fiber cloth, 5. Encoder, 6. Image acquisition card, 7. Controller, 8. Winding device, 9. Defect. DETAILED DESCRIPTION

[0034] The present invention will be further described in detail below with reference to the accompanying drawings using specific embodiments.

[0035] like Figure 1 As shown, the carbon fiber cloth production equipment includes a roller 1, and the roller 1 outputs the finished carbon fiber cloth. The technical solution of the present invention connects an encoder 5 to the roller 1. The production workshop is also provided with a winding device 8, and the carbon fiber cloth 4 output from the roller 1 is wound on the winding device 8, and the winding device 8 rolls the finished carbon fiber cloth 4 into a roll. The carbon fiber cloth between the winding device 8 and the roller 1 is flat and located in a horizontal plane. The technical solution of the present invention is provided with a CIS line array camera 2 and a marking machine 3. The CIS line array camera 2 is located above the carbon fiber cloth, and the detection field of the CIS line array camera 2 is downward facing the carbon fiber cloth. The marking machine 3 is close to the carbon fiber cloth, and the carbon fiber cloth is arranged according to Figure 1 Moving in the direction indicated by the middle arrow, the marking machine 3 is located in front of the winding device 8 and behind the CIS line array camera 2 , with a certain distance between the marking machine 3 and the CIS line array camera 2 .

[0036] like Figure 2 As shown, the encoder 5 is electrically connected to the controller 7 via a signal line, the image acquisition card 6 is electrically connected to the controller 7, and the CIS line array camera 2 is electrically connected to the image acquisition card 6. The CIS line array camera 2 sends the captured image to the image acquisition card 6, and the image acquisition card 6 then sends the image data to the controller 7. The controller 7 controls the marking machine 3 to mark the carbon fiber cloth.

[0037] The marking machine 3 may be a laser marking machine, a coding marking machine, a labeling marking machine or other common marking machines with specific structures.

[0038] During the inspection process, the CIS line array camera 2 and the marking machine 3 remain fixed. The carbon fiber cloth production equipment operates, with roller 1 rotating. The front end of the carbon fiber cloth is connected to the winding device 8. As the equipment continues to operate, the carbon fiber cloth continuously moves toward the winding device 8, performing continuous production. As roller 1 rotates, encoder 5 generates a pulse signal and sends it to controller 7, which receives the pulse signal and performs calculations.

[0039] The CIS line array camera 2 acquires an image of the carbon fiber cloth and sends the image data to the image acquisition card 6, which then transmits the image data to the controller 7. The controller 7 then processes the image data.

[0040] The CIS line scan camera 2 has its own light source and can be used in environments with different brightness conditions.

[0041] The controller 7 processes the image data, finds defects on the carbon fiber cloth based on the image, and identifies the defect type, such as cracks, white spots, holes, hair balls, wrinkles, etc. Defects can be identified through the target detection model. The specific process is as follows:

[0042] YOLOv8 object detection model based on fused deep separable convolution.

[0043] The first step is to build the database.

[0044] The CIS line scan camera 2 captures images of the carbon fiber cloth on the production line. The image acquisition process can include image acquisition under three different lighting conditions. A large amount of image data forms a database for training the model.

[0045] The second step is to label the large number of images mentioned above.

[0046] Use annotation tools to annotate the image in detail, marking the locations of defects such as cracks, holes, hair balls, wrinkles, etc. Figure 3 For example, a dataset of 820 labeled images is produced.

[0047] The third step is to train the target detection model.

[0048] Based on the infrastructure of the conventional YOLOv8 model, depthwise separable convolution is used to replace traditional convolution operations in the convolutional layer of the model. Depthwise separable convolution includes depthwise convolution and pointwise convolution. The depthwise convolution process performs convolution operations on each input channel independently, while the pointwise convolution process uses a 1×1 convolution kernel to integrate the features of each channel to achieve cross-channel information fusion.

[0049] The standard convolution layer is decomposed into two stages: depthwise convolution and pointwise convolution. In the depthwise convolution stage, each convolution kernel performs convolution operation only on a single channel, which significantly reduces the number of convolution kernel parameters. For depthwise convolution, assuming that the input feature map I has C in channels, depthwise convolution decomposes I into C in single-channel feature maps, each single-channel feature map I c Apply a convolution kernel K c , where c is the channel index. c =I c *K cOperation, where Oc is the cth channel of the output feature map, * indicates the convolution operation. Subsequently, the point convolution stage uses a 1x1 convolution kernel to further reduce the number of model parameters and computational complexity, while integrating the information of all channels on a single feature point to complete feature extraction; point-by-point convolution is actually a 1×1 convolution, which takes the output O of the depthwise convolution as input, where O has C in Channels. Point-by-point convolution uses a set of 1×1 convolution kernels K′ to combine the channels in O to generate the final output feature map P, with C out channels.

[0050]

[0051] P i,j is the pixel at position (i, j) in the output feature map P, O i,j,c is the pixel of the cth channel at position (i, j) in the depth convolution output feature map O, K′ c,i,j Is the weight of the cth channel at position (i, j) in the point-by-point convolution kernel K′. Combining depthwise convolution and point-by-point convolution, the output P of depthwise separable convolution can be expressed as:

[0052] P=Pointwise(Depthwise(I)·F)

[0053] Here I is the input feature map, F is the set of all convolution kernels in the depthwise convolution, Depthwise represents the depthwise convolution operation, and Pointwise represents the point-by-point convolution operation.

[0054] Depthwise separable convolution differs from the 3D convolution kernel in standard convolution. The convolution kernel in depthwise separable convolution has a channel count of 1. Therefore, during depthwise convolution, each convolution kernel only considers the channel it is responsible for, rather than the semantic information of all channels as in standard convolution. Depthwise convolution performs convolution operations on a single channel, reducing the number of convolution kernel parameters compared to standard convolution.

[0055] Point convolution uses a 1×1 convolution form.

[0056] Use the dataset formed in the second step above to train the model.

[0057] (1) Applying SPPF (Spatial Pyramid Pooling) to integrate multi-scale contextual information through the SPPF layer helps improve the model's ability to detect objects of different sizes.

[0058] (2) Feature map concatenation (Concat), combining feature maps from different layers of the network, so that the model can utilize deep and shallow features to enhance the accuracy of detection.

[0059] (3) Configure training parameters: set parameters such as learning rate, batch size, and training cycle.

[0060] (4) Loss function selection: Use a loss function suitable for target detection, such as cross entropy loss combined with bounding box regression loss.

[0061] (5) Optimizer selection: Select an optimizer such as Adam or SGD to optimize the model parameters.

[0062] (6) Test the performance of the model on an independent validation dataset, focusing on key indicators such as accuracy and recall, and obtain verification results.

[0063] (7) According to the verification results, the learning rate, regularization parameters, etc. are adjusted to optimize the model performance, and the weight file is exported to form an improved YOLOv8 target detection model based on fused deep separable convolution.

[0064] During training, CIoULoss and BCELoss can be used as loss functions for bounding box regression and category prediction, respectively, and the AdamW optimizer can be used to accelerate model convergence. After sufficient training, such as achieving an mAP of 0.719 during validation, the model can be deployed.

[0065] Step 4: Reference Figure 1 The carbon fiber cloth 4 moves toward the winding device 8 , and the CIS line array camera 2 collects an image of the carbon fiber cloth 4 , and the collected image is sent to the controller 7 .

[0066] In the fifth step, the controller 7 pre-processes the image data.

[0067] The specific process of preprocessing can be, step (1), dehazing processing: first calculate the dark channel DC in the image, which is usually the area with the least fogging effect. The calculation method of dark channel DC is as follows: DC(x) = min(R(x), G(x), B(x)), where R(x), G(x) and B(x) are the pixel values ​​of the red, green and blue channels of the image at position x, respectively. Secondly, use the calculated dark channel DC to generate a dark channel map, which shows the area with the minimum brightness in the image. Then, the atmospheric light is estimated through the image. The atmospheric light A is the maximum value in the dark channel of all pixels in the image. The atmospheric light is usually considered to be constant and colorful. A = max(R, G, B), where A is a vector, each element of which is the maximum value of the red, green and blue channels of the corresponding pixel; finally, calculate the transmission map The transmission map T(x) represents the degree of defogging of each pixel. It is the ratio of the dark channel to the atmospheric light. The transmission map reflects the degree of fogging of each pixel in the image. Using the estimated atmospheric light and the transmission map, the original color of the image can be restored. and Where O(x) is the color channel of the observed haze image.

[0068] Step (2), histogram equalization. Adaptive histogram equalization is used to solve the problem of how to enhance the contrast of data images under different lighting conditions. The specific process is to first calculate the cumulative distribution function CDF, and then calculate the corresponding cumulative distribution function from the histogram to balance the brightness of the image. The cumulative distribution function CDF(i) represents the cumulative probability of pixels with grayscale less than or equal to i. The calculation formula is: Among them, MN is the total number of pixels in the image; secondly, calculate the original histogram: calculate the grayscale histogram of the image, that is, the number of pixels at each grayscale level. The histogram H(i) is a function that counts the frequency of occurrence of each grayscale level i in the image. For grayscale level i, the value of its histogram is calculated as Where I(x,y) is the grayscale value of the image at position (x,y), M and N are the width and height of the image respectively; then equalization is performed to adjust the grayscale of the image according to the cumulative distribution function so that the contrast of the entire image is enhanced and adapted to different lighting conditions. The equalization mapping T(i) is defined as: T(i) = min(CDF(i) (L-1), L-1), where L is the maximum grayscale value (for 8-bit images, L = 255).

[0069] Histogram homogenization stretches the histogram distribution of the image to the entire representable brightness range as much as possible. This method can help enhance the contrast between fibers, making the visual effect of the entire image more uniform, which is helpful for subsequent image analysis and feature extraction.

[0070] Step (3), edge enhancement. The Sobel gradient algorithm is used to enhance the edge of the image. Carbon fiber prepreg has a unique texture structure. The Sobel algorithm can make these textures more obvious by enhancing the edges of the image, which is helpful for subsequent texture analysis and feature recognition. The formula for Sobel horizontal edge detection is: Among them, I x is the edge strength in the horizontal direction, I p-1 ,I p0 ,I p+1 are the intensities of the left, center, and right pixels of the current pixel p, respectively. n-1， I n0 ,I n+1are the intensities of the adjacent pixels directly below the current pixel p. Then vertical edge detection is performed, and the formula for Sobel vertical edge detection is:

[0071]

[0072] In this formula, I y is the edge strength in the vertical direction, I top-left ,I top-center ,I top-right I are the pixel intensity values ​​on the left, center, and right side of the current pixel point p. bottom-left , I bottom-center ,I bottom-right are the pixel intensity values ​​to the left, center, and right of the current pixel point p.

[0073] Step (4), smoothing. Using the mean filter method for smoothing can effectively reduce image noise and eliminate unnecessary details, which helps to reduce the impact of random noise or stray light. The calculation formula of the mean filter is expressed as: Where P′(x,y) is the pixel value of the filtered image at point (x,y), P(i,j) is the pixel value of point (i,j) in the original image, and N is the total number of pixels in the neighborhood.

[0074] Edge enhancement through Sobel gradient algorithm and Frangi filter can help highlight the edges in the carbon fiber prepreg image, making the texture of the carbon fiber and possible structural defects in the carbon fiber prepreg image more obvious. This is especially important for the subsequent defect detection stage, because clear edges can help improve the accuracy of defect identification; using methods such as mean filtering and Gaussian filtering to achieve image smoothing can effectively reduce image noise and eliminate unnecessary details. In carbon fiber image processing, smoothing helps reduce the influence of random noise or stray light in the image.

[0075] The sixth step is to segment the image after the sixth step into 640x640 pixels, and prepare to detect each segmented image separately. Of course, you can also detect the entire image without segmentation.

[0076] In the seventh step, the segmented image is input into the YOLOv8 target detection model based on the fusion of deep separable convolution. The YOLOv8 target detection model can identify defects such as holes, cracks, hair balls, and wrinkles.

[0077] The identified defects can be visually displayed on the display interface.

[0078] The target detection model can effectively identify defects such as white spots, holes, cracks, hair balls, and wrinkles with high detection accuracy and fast speed. It can also identify tiny white spots, holes, cracks, hair balls, and wrinkles with high detection accuracy.

[0079] In the eighth step, the corresponding position of the carbon fiber cloth is marked by the marking machine 3.

[0080] Because roller 1 rotates continuously, encoder 5 continuously outputs pulse signals, and the forward speed V of carbon fiber cloth 4 remains constant. Since the distance L between CIS line scan camera 2 and marker 3 is constant (for example, L = 2 meters), a portion of carbon fiber cloth 4 within the field of view of CIS line scan camera 2 moves to marker 3 after a period of time s (assuming s is 10 seconds). During this period of time s, the number of pulses N generated by encoder 5 remains constant.

[0081] Therefore, when the target detection model identifies a defect, the controller 7 starts to count the number n of pulse signals output by the encoder 5. When the number n reaches N (that is, when n=N), it indicates that the defective part of the carbon fiber cloth has moved to the position of the marking machine 3. At this time, the controller 7 controls the marking machine 3 to work, and the marking machine 3 marks the edge of the carbon fiber cloth (such as Figure 4 As shown, the edge of the carbon fiber cloth corresponding to the defect 9), that is, the edge of the carbon fiber cloth corresponding to the defective portion of the carbon fiber cloth is marked.

[0082] In this way, the carbon fiber cloth 4 is finally rolled up under the action of the winding device 8, and the rolled carbon fiber cloth is shipped to the customer. When the customer puts the rolled carbon fiber cloth into production and uses it, he can easily find out which parts have defects through the markings, and then take measures to deal with the defects, such as cutting off the defective parts, which provides great convenience for the customer's production.

[0083] In addition, during the operation of the CIS line array camera 2 and the marking machine 3 for a period of time (for example, 10 minutes), the controller processes the identified defects, forms a map indicating the defect location, and displays the map on the display screen, such as the reference image. Figure 5 , the dotted box represents the identified defects. It can be seen that the five defects are basically located in a straight line along the Y-axis direction. The operator observes the display screen and finds that the five defects are basically located in a straight line along the Y-axis direction. Then, the operator can conduct targeted inspections on the corresponding processes and find the causes of the defects. For example, the operator can inspect the input station of 60 carbon fiber raw materials, find several carbon fiber raw materials in the 60 carbon fiber raw materials that are in the same Y-axis direction as the five defects, and find the causes for these carbon fiber raw materials. The operator can also find the causes in the straight line position where the five defects are located in the impregnation station.

[0084] The above description is only about the preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations.

Claims

1. A carbon fiber cloth surface defect detection device, characterized in that: Including encoder, CIS line array camera, marking machine, image acquisition card and controller; The encoder is configured to be connected to a rotating roller of a carbon fiber cloth production device, the CIS line array camera is arranged above the carbon fiber cloth to be inspected, and the marking machine is arranged near the carbon fiber cloth. The marking machine is located behind the CIS line array camera along the moving direction of the carbon fiber cloth to be inspected, and there is a certain distance between the marking machine and the CIS line array camera; The encoder is electrically connected to the controller via a signal line, the image acquisition card is electrically connected to the controller, and the CIS line array camera is electrically connected to the image acquisition card.

2. The carbon fiber cloth surface defect detection device according to claim 1, characterized in that: The working process of the carbon fiber cloth surface defect detection device is: The carbon fiber cloth production equipment and winding device are running, the roller of the carbon fiber cloth production equipment rotates, the front end of the carbon fiber cloth is connected to the winding device, and the carbon fiber cloth continuously moves toward the winding device on the horizontal plane; the encoder generates a pulse signal and sends it to the controller, which receives the pulse signal and performs calculations; The CIS linear array camera acquires an image of the carbon fiber cloth and sends the image data to the image acquisition card 6. The image acquisition card then transmits the image data to the controller, which then uses the YOLOv8 object detection model to identify defects in the image. When the YOLOv8 target detection model identifies a defect, the controller starts counting the number n of pulse signals output by the encoder. When the number n reaches N, the controller controls the marking machine to mark the edge of the carbon fiber cloth.

3. The carbon fiber cloth surface defect detection device according to claim 2, characterized in that: The YOLOv8 target detection model is built in the following way: The first step is to build a database; Use CIS line array cameras to collect a large number of images of carbon fiber cloth on the production line; The second step is to label the large number of images mentioned above. Use annotation tools to annotate images in detail and create a dataset of 820 annotated images. The third step is to train the target detection model; Based on the conventional YOLOv8 model, the model's convolutional layer uses depthwise separable convolution to replace the traditional convolution operation. Depthwise separable convolution includes depthwise convolution and pointwise convolution. The depthwise convolution process performs convolution operations on each input channel independently, while the pointwise convolution process uses a 1×1 convolution kernel to integrate the features of each channel. The standard convolution layer is decomposed into two stages: depthwise convolution and pointwise convolution. The model is trained using the dataset generated in the second step. During the training process, CIoULoss and BCELoss are used as loss functions for the bounding box regression and category prediction branches, respectively, and the AdamW optimizer is adopted. After the marking machine has been working for a period of time, the controller processes the identified defects, forms a diagram showing the defect location, and displays the diagram on the screen. The operator observes the diagram on the screen and conducts targeted inspections of the corresponding processes based on the distribution of the defect locations in the diagram in order to find the cause of the defect.

Images