Deep learning-based tail angle recognition method for hot-rolled coiled strip

Abstract

The application discloses a tail angle identification method for hot-rolled coiled strip steel based on deep learning, which comprises the following steps: 1, establishing a strip steel tail image data set, which is divided into a training data set and a verification data set; 2, constructing a deep learning model, sending the training data set and the verification data set to the deep learning model for training, and obtaining an identification model for the strip steel tail position; 3, setting a field environment, collecting field images, identifying a tail category, calculating an initial strip steel tail position angle, and starting a tail tracking task; 4, collecting field images at intervals, using the identification model to track the strip steel tail position in real time, identifying the tail category and calculating the strip steel tail position angle; 5, calculating an actual angle value according to the tail category identification result in the strip steel coiling process; and 6, whether the coiling is completed, if yes, ending the image collection, and if not, returning to step 4. The application can realize accurate identification and tracking of the strip steel tail position based on the deep learning identification model.

Description

Technical Field

[0001] This invention relates to a strip coiling control method, and more particularly to a deep learning-based method for identifying the tail angle of hot-rolled strip coiling. Background Technology

[0002] During the coiling of hot-rolled strip steel at the tail end, a hot metal detector or laser detector installed in front of the pinch rolls tracks the position of the strip tail. By calculating the remaining length of the strip tail and the radius of the coil, the angle of the strip tail is obtained in real time. An encoder installed on the drum motor controls the rotation angle, ensuring the strip tail reaches the target angle position. This target angle position is set manually or automatically to allow for smooth uncoiling and bundling of the coil. However, in actual positioning, various factors may arise, such as tracking errors in the strip tail position, calculation errors in the coil radius, and inability of the coiler torque to lock, sometimes causing the strip tail to deviate too much from the target position. Currently, the confirmation and adjustment of the final position of the strip tail on the coiler largely relies on manual operation, which is time-consuming and labor-intensive.

[0003] Chinese utility model patent CN200920103390.X discloses a strip position detection device for a coiler. Two laser detectors and one hot metal detector are installed near the coiler's pinch rolls. The two laser detectors are connected in parallel and then in series with the hot metal detector. This detection device is mainly used to improve the speed and reliability of tail-end tracking, but it cannot solve the problem of inaccurate tail-end tracking during the coiling process.

[0004] Japanese patent application JP2011184085 discloses a winding device and method for hot-rolled strip steel. This winding device is used for winding thick (≥10mm) hot-rolled strip steel. Multiple auxiliary winding rollers are arranged at predetermined intervals along the circumference of the coiler to prevent the tail end of the thick hot-rolled strip steel from drooping to the outside of the coil. Then, based on the tail position detected by a laser sensor, a two-stage position control method is used during the winding stage to achieve tail positioning of the strip steel. This winding device is mainly used to improve stability during the winding process. While the two-stage control method allows for easy manual intervention, it cannot improve the automation rate of tail-end positioning control. Summary of the Invention

[0005] The purpose of this invention is to provide a deep learning-based method for identifying the tail angle of hot-rolled coiled strip, which can accurately identify and track the tail position of the strip based on a deep learning-based identification model.

[0006] This invention is implemented as follows:

[0007] A deep learning-based method for identifying the tail angle of hot-rolled coiled strip steel, characterized by the following steps:

[0008] Step 1: Establish a strip tail image dataset and divide it into a training dataset and a validation dataset;

[0009] Step 2: Build a deep learning model. Feed the training dataset and validation dataset into the deep learning model for training to obtain a recognition model suitable for the tail position of hot-rolled strip coiling.

[0010] Step 3: Set up the site environment, collect on-site images of the strip during coiling, identify the tail category of the strip in the on-site images using the recognition model, calculate the initial tail position angle of the strip, and start the tail tracking task;

[0011] Step 4: The image acquisition device periodically acquires on-site images, and the recognition model tracks the position of the strip tail in real time based on the on-site images, identifies the tail category, and calculates the position angle of the strip tail.

[0012] Step 5: During the strip coiling process, determine whether each on-site image has a tail category identification result, and calculate the actual angle value based on the tail category and the tail position angle of the strip.

[0013] Step 6: Determine whether the strip winding has ended. If yes, end image acquisition and exit the control program. If not, return to step 4.

[0014] Step 1 includes:

[0015] Step 1.1: When the coiler is coiling the strip steel, an image of the final section of the coiled strip steel is acquired using an image acquisition device;

[0016] Step 1.2: Based on the position of the strip tail, the images are divided into the following two categories:

[0017] (I) Taking the center of the coiler drum as the reference center, if the tail end of the strip appears in the unobstructed area to the left of the reference center, the image is recorded as the left tail image.

[0018] (II) Taking the center of the coiler drum as the reference center, if the tail end of the strip appears in the unobstructed area to the right of the reference center, the image is recorded as the right tail image.

[0019] Step 1.3: Delineate the outline of the strip tail in the classified image as a closed area, and add the label "left tail image" or "right tail image" to the closed area according to the position of the strip tail.

[0020] Step 1.4: Using the reference circle center as the center and a square window as the clipping border, clip the labeled image and save it as a dataset image;

[0021] Step 1.5: Repeat steps 1.1-1.4 to save all dataset images, forming a strip tail image dataset;

[0022] Step 1.6: Divide all the images in the strip tail image dataset into training dataset and validation dataset according to the number of images.

[0023] In step 1.4, the cropped pixel size is a multiple of 64; in step 1.6, the ratio of the number of images in the training dataset to the number of images in the validation dataset is 8:2 to 9:1.

[0024] Step 2 includes:

[0025] Step 2.1: Select RestNet50 of mask-rcnn as the deep learning network architecture;

[0026] Step 2.2: Set the resizing mode of the dataset images to "square", that is, set the short side length value and the long side length value of the dataset images to be equal;

[0027] Step 2.3: Feed the dataset images into the deep learning model, and use the deep learning model to identify the tail of the strip in the dataset images, and apply the tail category loss L. cls Bounding box loss L box Tail shape mask loss L mask As three targets for identification;

[0028] Step 2.4: Establish the model for the loss function L: L = L cls +L box +L mask ;

[0029] Step 2.5: Select the ResNet50 weights on the original open-source dataset COCO of Mask-R-CNN as the initial training weights of the deep learning model. Set the training batch according to the recognition accuracy, and set the number of iterations in each batch according to the number of training samples and the device memory size. Perform iterative training to obtain the loss function trend plot of the training dataset and the loss function trend plot of the validation dataset.

[0030] Step 2.6: Select the optimal weights on the loss function trend graph to obtain a recognition model suitable for the tail position of hot-rolled strip coiling. The input of the recognition model is the cropped on-site image, and the output is the tail category and the coordinates of the four corner points of the bounding box.

[0031] In step 2.6, the selection criterion for the optimal weight is that the loss function values ​​of the training dataset and the validation dataset are both at their minimum values ​​in the trend graphs of the loss function of the training dataset and the validation dataset.

[0032] In step 3, the requirements for adjusting the on-site environment include:

[0033] a. Without affecting the material properties, during the strip tail winding process, shut off the cooling water for the upstream guide plate and the winding machine auxiliary roll;

[0034] b. Install a fan next to the winding machine to disperse water mist and smoke;

[0035] c. Adaptively change the exposure time and increase the light source.

[0036] Step 3 includes:

[0037] Step 3.1: The image acquisition device acquires on-site images, with the center of the winding machine drum as the center of the square window, and the on-site image is cropped;

[0038] Step 3.2: Send the cropped scene image to the recognition model, and the recognition model outputs the tail category and the coordinates of the four corner points of the bounding box;

[0039] Step 3.3: The control system determines whether there is a tail category recognition result. If yes, proceed to step 3.4; otherwise, return to step 3.1.

[0040] Step 3.4: Calculate the initial tail position angle of the strip and start the tail tracking task.

[0041] In step 3.4, the initial calculation method for the tail position angle of the strip is as follows:

[0042] With the center of the winding machine drum as the reference center (x center y center Using 0 o'clock as the 0° angle, and proceeding clockwise, the region [0°, 90°) in the cropped field image is defined as the first quadrant, the region [90°, 180°) as the second quadrant, the region [180°, 270°) as the third quadrant, and the region [270°, 360°) as the fourth quadrant. The position angle θ of the strip tail is calculated based on the tail category and the coordinates of the four corner points of the bounding box. dec ;

[0043] (I) When the tail category is the right tail image:

[0044] Use the top right corner of the bounding box as the tail positioning point (x dec y dec ), calculate the position angle θ of the strip tail at the tail positioning point. dec ;

[0045] If the tail positioning point is located in the first quadrant, the angle θ of the strip tail position of that tail positioning point is... dec The calculation formula is:

[0046]

[0047] If the tail positioning point is located in the second quadrant, the angle θ of the strip tail position of that tail positioning point is... dec The calculation formula is:

[0048]

[0049] (II) When the tail category is the left tail image:

[0050] Use the left midpoint of the bounding box as the tail positioning point (x dec y dec ), calculate the position angle θ of the strip tail at the tail positioning point. dec ;

[0051] If the tail positioning point is located in the third quadrant, the angle θ of the strip tail position at that tail positioning point is... dec The calculation formula is:

[0052]

[0053] If the tail positioning point is located in the fourth quadrant, the angle θ of the strip tail position at that tail positioning point is... dec The calculation formula is:

[0054]

[0055] In step 5,

[0056] During the strip coiling process, the actual angle value θ output by the control system ang The default angle is 0°. The actual angle value θ output by the control system is determined when the tail section is positioned to the left or right of the reference center, after a tail category identification result is obtained. ang =θ dec And calculate the estimated angle value θ for the next on-site image. pre The calculation formula is:

[0057] θ pre =θ ang +θ cal (5)

[0058] Where, θ cal The angle corresponding to the current angular velocity w of the winding machine is calculated using the formula: θ cal =w·t s ·360° / 60, unit is ° / second;

[0059] As the strip continues to be wound, the control system calculates the actual angle value θ based on the tail category recognition result of the i-th on-site image. ang And calculate the estimated angle value θ for the next on-site image.pre :

[0060] (I) When the i-th on-site image has a tail category recognition result, the actual angle value θ output by the control system is... ang The calculation formula is:

[0061] θ ang =θ dec ·(1-α)+θ prei ·α (6)

[0062] Where α is the weight value, and the value of α ranges from 0 to 1; θ prei Let be the estimated angle value of the i-th on-site image;

[0063] Calculate the estimated angle value θ for the next on-site image. pre The calculation formula is: θ pre =θ ang +θ cal ;

[0064] (II) When the i-th on-site image has no tail category recognition result, the actual angle value θ output by the control system is... ang The calculation formula is:

[0065] θ ang =θ prei (7);

[0066] Calculate the estimated angle value θ for the next on-site image. pre The calculation formula is: θ pre =θ dec +θ cal .

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] 1. This invention, based on deep learning to construct a recognition model and visual recognition, can accurately identify the tail position of the strip based on the acquired images of the strip being wound by the coiler. Combined with the angular velocity of the coiler, it can predict the tail position of the strip in the next image, achieving real-time tracking of the strip tail position. The coiler can use the predicted angle value of the strip tail position to automatically correct the tail position during strip winding, providing reliable data support for real-time adjustment and correction of the angular velocity during strip winding. This fundamentally solves various problems such as inaccurate tracking of the strip tail after leaving the pinch roll, inaccurate calculation of the coil radius, and inability of the coiler torque to lock.

[0069] 2. This invention employs a mask-R-CNN neural network to construct a deep learning model for training and selects training results based on the trend of the loss function. It can determine the strip tail category and bounding box coordinates based on strip coiling images acquired at appropriate locations, thereby calculating the strip tail angle based on different strip tail positions. Simultaneously, by adjusting the on-site environment such as water, fog, and light sources, it can effectively improve the recognition accuracy of the strip tail position, with fast calculation speed. It eliminates the need for manual visual positioning of the strip tail, thereby improving the automation rate and operational efficiency of strip coiling control.

[0070] This invention can accurately identify and track the tail position of strip steel based on a deep learning recognition model combined with the angular velocity of the coiler. This provides reliable data support for automatic tail correction of strip steel during coiling, reduces manual intervention, and improves the automation rate of strip steel coiling control. It is suitable for identifying the tail position of hot continuous rolling strip steel. Attached Figure Description

[0071] Figure 1 This is a flowchart of the tail angle recognition method for hot-rolled coiled strip steel based on deep learning, according to the present invention.

[0072] Figure 2 This is a flowchart of steps 1 and 2 in the deep learning-based tail angle recognition method for hot-rolled coiled strip steel of the present invention;

[0073] Figure 3 This is a dataset image from the deep learning-based tail angle recognition method for hot-rolled coiled strip steel in this invention;

[0074] Figure 4 This is a trend graph of the loss function of the training dataset in Embodiment 1 of the deep learning-based tail angle recognition method for hot-rolled coiled strip steel of the present invention.

[0075] Figure 5 This is a trend chart of the loss function of the verification dataset in Example 1 of the deep learning-based tail angle recognition method for hot-rolled coiled strip steel of the present invention.

[0076] In the diagram, 11 is the winding machine drum, 12 is the winding machine support arm, 100 is the reference center, 2 is the strip tail end, 3 is the enclosed area, and 4 is the cutting border. Detailed Implementation

[0077] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0078] Please see the appendix Figure 1 A deep learning-based method for identifying the tail angle of hot-rolled coiled strip steel includes the following steps:

[0079] Step 1: Establish a strip tail image dataset and divide it into a training dataset and a validation dataset.

[0080] Please see the appendix Figure 2 Step 1 includes:

[0081] Step 1.1: When the coiler is coiling the strip steel, the image of the final section of the coiled strip steel is acquired by the image acquisition device.

[0082] Preferably, the image acquisition device can be a color CCD camera to acquire clear images of the coiler drum and strip. However, due to limitations in the performance of the color CCD camera and environmental interference, real-time tail position detection solely through visual algorithms may result in significant errors. For example, when the winding speed is too fast, the outline of the strip tail in the image may be unclear, potentially leading to multiple images of what appears to be the strip tail, or even the strip tail itself becoming unidentifiable. To accurately and stably identify the strip tail position, when creating the strip tail image dataset, images are manually selected by visual inspection. Clear and stable images of the strip tail are retained; that is, images that clearly identify the strip tail and its outline are kept, while unclear or unstable images of the strip tail are deleted.

[0083] Step 1.2: Based on the position of the strip tail, the images are divided into the following two categories:

[0084] Please see the appendix Figure 3 (I) Taking the center of the coiler drum 11 as the reference center 100, if the tail end 2 of the strip appears in the unobstructed area to the left of the reference center 100, the image is recorded as the left tail image. The unobstructed area refers to the area where the tail end 2 of the strip is not obstructed by the coiler support arm 12, ensuring the accuracy of the identification of the position of the tail end 2 of the strip.

[0085] (II) Taking the center of the winding machine drum 11 as the reference center 100, if the tail end 2 of the strip appears in the unobstructed area to the right of the reference center 100, the image is recorded as the right tail image.

[0086] Step 1.3: Delineate the outline of the strip tail in the classified image as a closed region 3, and add a label to the closed region 3 according to the location of the strip tail 2: "Left tail image" or "Right tail image".

[0087] Step 1.4: Cropping the labeled image with the reference circle center 100 as the center and a square window as the cropping border 4, and saving it as a dataset image. Preferably, the cropping pixel size is compatible with the pre-trained model. When Mask-RCNN is selected as the deep learning model, the cropping pixel size can be set to a multiple of 64.

[0088] Step 1.5: Repeat steps 1.1-1.4 to save all dataset images, forming a strip tail image dataset.

[0089] Step 1.6: Divide all images in the strip tail image dataset into training and validation datasets according to a certain image ratio. Preferably, the image ratio of the training and validation datasets can be 8:2 to 9:1, which can be set according to actual needs.

[0090] Please see the appendix Figure 2 Step 2: Construct a deep learning model. Send the training dataset and validation dataset to the deep learning model for training to obtain a recognition model suitable for the tail position of hot-rolled strip coiling.

[0091] Step 2.1: Preferably, the deep learning model can be the RestNet50 of mask-rcnn as the deep learning network architecture. In order to improve the computing speed, the number of nodes in the fully connected layer is set to 100, and the number of classifications is set to 4: including background, left tail of the center, right tail of the center, and other cases.

[0092] Step 2.2: Set the image resizing mode (IMAGE_RESIZE_MODE) of the dataset to "square", that is, set the shorter side length value (IMAGE_MIN_DIM) and the longer side length value (IMAGE_MAX_DIM) of the dataset image to be equal in size, so as to avoid information loss and computation caused by image compression or stretching.

[0093] Step 2.3: Feed the dataset images into the deep learning model to identify the tail of the strip in the dataset images. The identification accuracy needs to take into account the tail category loss L. cls Bounding box loss L box Tail shape mask loss L mask Three objectives. After inputting the dataset images into the recognition model, the tail category loss L is automatically calculated using the algorithm built into Mask-R-CNN. cls Bounding box loss L box Tail shape mask loss L mask There are three objectives; the algorithm process will not be elaborated upon here.

[0094] Step 2.4: Establish the model for the loss function L: L = L cls +L box +L mask .

[0095] Step 2.5: Select the ResNet50 weights on the original open-source dataset COCO of Mask-R-CNN as the initial training weights for the deep learning model. Set the training batch according to the recognition accuracy requirements, and set the number of iterations in each batch according to the number of training samples and the device memory size. Perform iterative training to obtain the loss function trend plot of the training dataset and the loss function trend plot of the validation dataset.

[0096] Preferably, the training batches can be adjusted according to the changing trend of the loss function L during the actual training process. If the loss function of the training dataset continues to decrease while the loss function of the validation dataset does not change significantly, the number of training batches can be increased. The number of iterations can be set to 300-500.

[0097] Step 2.6: Select the optimal weights on the loss function trend graph to obtain a recognition model suitable for the tail position of hot-rolled strip coiling. The input of the recognition model is the cropped on-site image, and the output is the tail category (i.e., the left tail image or the right tail image) and the coordinates of the four corner points of the bounding box.

[0098] The bounding box is defined by the deep learning model. For example, when using Mask-R-CNN as the deep learning model, if the tail category is "left tail image", then a rectangular bounding box is drawn for the tail of the strip, and the tail end 2 is placed at the midpoint of the left side of the bounding box; if the tail category is "right tail image", then a rectangular bounding box is drawn for the tail of the strip, and the tail end 2 is placed at the upper right corner of the bounding box.

[0099] The optimal weight selection condition is that the loss function values ​​of the training dataset and the validation dataset are both at their minimum values ​​in the loss function trend graphs of the training dataset and the validation dataset.

[0100] Step 3: Set up the site environment, collect on-site images of the strip during coiling, identify the tail category of the strip in the on-site images using the recognition model, and start the tail tracking task.

[0101] Because the winding process of the winding machine takes place in a complex industrial environment, it will inevitably cause some interference to the identification of the tail position, thus affecting the implementation of the subsequent control system. Therefore, it is necessary to adjust the relevant on-site environment. The on-site environment adjustment requirements in step 3 include:

[0102] 1. Reduce the interference of water on the strip surface on the identification of strip position in the field images. Generally, water on the strip surface has the greatest impact on the identification of the tail position, mainly from the cooling water of the upstream guide plate and the cooling water of the coiler's auxiliary winding rolls. Without affecting material properties, the cooling water of the upstream guide plate and the coiler's auxiliary winding rolls can be turned off during the tail winding process of the strip.

[0103] 2. Reduce interference from water mist and smoke on strip position identification in on-site images. Generally, the area around the coiler is in a relatively high-temperature environment, which easily generates water mist and smoke. A fan can be installed next to the coiler to disperse the water mist and smoke.

[0104] 3. Reduce interference from insufficient light on strip position identification in on-site images. Since the coiler operates continuously for extended periods, light intensity varies at different times. To create a suitable lighting environment for image acquisition, the adaptive adjustment of the CCD camera's exposure time can be used to maintain relatively stable light intensity in the image acquisition environment. If necessary, additional light sources can be added to supplement the lighting in the shooting area.

[0105] Depending on the actual situation on site, one or more of the above requirements can be selected to adjust the site environment.

[0106] Step 3 includes:

[0107] Step 3.1: The image acquisition device acquires on-site images, with the center of the winding machine drum at the center of the square window, and the on-site image is cropped. Preferably, the cropped pixel size is a multiple of 64.

[0108] Step 3.2: Send the cropped scene image to the recognition model. The recognition model outputs the tail category and the coordinates of the four corner points of the bounding box.

[0109] Step 3.3: The control system determines whether there is a tail category recognition result. If yes, proceed to step 3.4; otherwise, return to step 3.1.

[0110] Step 3.4: Calculate the initial position of the strip tail and start the tail tracking task.

[0111] The method for calculating the initial strip tail position angle is as follows: taking the center of the coiler drum as the reference center (x... center y center Using 0 o'clock as the 0° angle, and proceeding clockwise, the region [0°, 90°) in the cropped field image is defined as the first quadrant, the region [90°, 180°) as the second quadrant, the region [180°, 270°) as the third quadrant, and the region [270°, 360°) as the fourth quadrant. The position angle θ of the strip tail is calculated based on the tail category and the coordinates of the four corner points of the bounding box. dec .

[0112] (I) When the tail category is the right tail image:

[0113] Use the top right corner of the bounding box as the tail positioning point (x dec y dec), calculate the position angle θ of the strip tail at the tail positioning point. dec .

[0114] If the tail positioning point is located in the first quadrant, the angle θ of the strip tail position of that tail positioning point is... dec The calculation formula is:

[0115]

[0116] If the tail positioning point is located in the second quadrant, the angle θ of the strip tail position of that tail positioning point is... dec The calculation formula is:

[0117]

[0118] (II) When the tail category is the left tail image:

[0119] Use the left midpoint of the bounding box as the tail positioning point (x dec y dec ), calculate the position angle θ of the strip tail at the tail positioning point. dec .

[0120] If the tail positioning point is located in the third quadrant, the angle θ of the strip tail position at that tail positioning point is... dec The calculation formula is:

[0121]

[0122] If the tail positioning point is located in the fourth quadrant, the angle θ of the strip tail position at that tail positioning point is... dec The calculation formula is:

[0123]

[0124] Step 4: The image acquisition device operates at intervals t. s Collect on-site images, use a recognition model to track the position of the strip tail in real time based on the on-site images, identify the tail category, and calculate the position angle of the strip tail.

[0125] The method for determining the tail position angle of the strip is the same as that for determining the initial tail position angle of the strip in step 3, and will not be repeated here.

[0126] Preferred, t s The value range is 10-200ms.

[0127] Step 5: During the strip coiling process, the control system determines whether there is a tail category identification result for each on-site image. The identification model determines the tail category identification result and the tail position angle θ of the strip. dec Calculate the actual angle value θ ang .

[0128] The recognition model is run in a control system that can use existing computer equipment, such as a GPU server configured with a 1080Ti.

[0129] During the strip coiling process, the actual angle value θ output by the control system ang The default value is 0°, until a tail category is identified, i.e., θ. dec When the angle is ≠ 0, the tail of the strip is located to the left or right of the reference center, and the actual angle value θ output by the control system is... ang =θ dec And calculate the estimated angle value θ for the next on-site image. pre The calculation formula is:

[0130] θ pre =θ ang +θ cal (5)

[0131] Where, θ cal The angle corresponding to the current angular velocity w of the winding machine is calculated using the formula: θ cal =w·t s • 360° / 60, in degrees per second.

[0132] As the strip continues to be wound, the control system calculates the actual angle value θ based on the tail category recognition result of the i-th on-site image. ang Where i is a natural number, the value of which is determined based on the number of on-site images, and the estimated angle value θ for the next on-site image is calculated. pre :

[0133] (I) When the i-th on-site image has a tail category recognition result, the output recognition value is the tail position angle θ of the strip. dec The actual angle value θ output by the control system ang That is, the weighted sum of the predicted values, and the actual angle value θ. ang The calculation formula is:

[0134] θ ang =θ dec ·(1-α)+θ prei α (6)

[0135] Where α is the weight value, and the value of α ranges from 0 to 1.

[0136] θ prei Let θ be the estimated angle value of the i-th on-site image. prei The actual angle value θ can be obtained from the (i-1)th image, i.e., the previous image. ang The estimate was obtained.

[0137] Calculate the estimated angle value θ for the next (i+1)th image. pre The calculation formula is: θ pre =θ ang +θ cal .

[0138] (II) When the i-th on-site image has no tail category recognition result, the actual angle value θ output by the control system is... ang The calculation formula is:

[0139] θ ang =θ prei (7);

[0140] Calculate the estimated angle value θ for the next (i+1)th image. pre The calculation formula is: θ pre =θ dec +θ cal The control system outputs the estimated angle value θ. pre That is, the predicted value.

[0141] It can be based on the actual angle value θ ang and the estimated angle value θ pre The angular velocity w of the coiler is adaptively adjusted to improve the automation rate of strip coiling. The adjustment of the angular velocity w of the coiler is executed by the control model built into the coiler, which will not be elaborated here.

[0142] Step 6: The control system determines whether the strip winding has ended. If yes, it ends image acquisition and exits the control program; otherwise, it returns to step 4.

[0143] Example 1: Please refer to the appendix. Figure 1 The tail angle identification method of the present invention was applied to a hot-rolled strip steel production unit.

[0144] Step 1: Establish a strip tail image dataset and divide it into a training dataset and a validation dataset.

[0145] Please see the appendix Figure 2 Step 1 includes:

[0146] Step 1.1: While the two coilers are coiling the strip, an image of the final section of the strip is captured using a color CCD camera (image acquisition device). Images are then manually selected by visual inspection, retaining only clear and stable images of the strip's tail section.

[0147] Please see the appendix Figure 3 Step 1.2: Based on the position of the strip tail, the images are divided into the following two categories:

[0148] (I) Taking the center of the winding machine drum 11 as the reference center 100, the coordinates of the reference center 100 are (973, 495). The tail end 2 of the strip appears in the unobstructed area to the left of the reference center 100. This image is recorded as the left tail image.

[0149] (II) With the center of the winding machine drum 11 as the reference center 100, if the tail end 2 of the strip appears in the unobstructed area to the right of the reference center 100, then this image is recorded as the right tail image.

[0150] Step 1.3: Delineate the outline of the strip tail in the classified image as a closed region 3, and add the label "left tail image" or "right tail image" to the closed region 3 according to the position of the strip tail.

[0151] Step 1.4: Using the reference circle center 100 as the center, use a 960*960 square window as the clipping border 4 to clip the labeled image so that the clipped image includes the steel tail end 2, and save it as a dataset image.

[0152] Step 1.5: Repeat steps 1.1-1.4 to save 1158 dataset images, forming the strip tail image dataset.

[0153] Step 1.6: Divide all the images in the strip tail image dataset into a training dataset (926 images) and a validation dataset (232 images) according to a ratio of 8:2.

[0154] Step 2: Build a deep learning model. Feed the training dataset and validation dataset into the deep learning model for training to obtain a recognition model suitable for the tail position of hot-rolled strip coiling.

[0155] Step 2.1: The deep learning model selected is the RestNet50 of Mask-R-CNN as the deep learning network architecture. In order to improve the computing speed, the number of nodes in the fully connected layer is set to 100, and the number of classifications is set to 4: including background, left tail of the center, right tail of the center, and other cases.

[0156] Step 2.2: Set the resizing mode of the dataset image to "square", that is, set the length of the longer side of the dataset image to be equal to the length of the shorter side, both to 960, to avoid information loss and computation caused by image compression or stretching.

[0157] Step 2.3: Feed the dataset images into the deep learning model to identify the tail of the strip in the dataset images. The identification accuracy needs to take into account the tail category loss L. cls Bounding box loss L box Tail shape mask loss L maskThree objectives. Using the dataset images as input, the tail category loss L is automatically calculated using the algorithm built into Mask-RCNN. cls Bounding box loss L box Tail shape mask loss L mask There are three objectives; the algorithm process will not be elaborated upon here.

[0158] Step 2.4: Establish the model for the loss function L: L = L cls +L box +L mask .

[0159] Step 2.5: Select the ResNet50 weights from the original open-source dataset COCO of Mask-R-CNN as the initial training weights for the deep learning model. Set the training batch size to 500 based on the required recognition accuracy. Based on the number of training samples and the GPU memory size, set the number of iterations within each batch to 500. Perform iterative training to obtain the loss function trend plots for the training dataset and the validation dataset. Please refer to the appendix. Figure 4 and attached Figure 5 Among them, attached Figure 4 A trend chart of the loss function for the training dataset is attached. Figure 5 To verify the trend of the loss function of the dataset.

[0160] Step 2.6: Select the optimal weights on the loss function trend graph, from the attached... Figure 4 and attached Figure 5 It can be seen that, taking the 256th training iteration, the loss function values ​​of both the training dataset and the validation dataset are at their minimum values. The training dataset and the validation dataset both perform well, resulting in a recognition model suitable for the tail position of hot-rolled strip coiling. The input of the recognition model is the cropped on-site image, and the output is the tail category and the coordinates of the four corner points of the bounding box.

[0161] Step 3: Set up the site environment, collect on-site images of the strip during coiling, identify the tail category of the strip in the on-site images using the recognition model, and start the tail tracking task.

[0162] 1. Without affecting the material properties, during the strip tail winding process, shut off the cooling water for the upstream guide plate and the winding machine auxiliary roll.

[0163] 2. Install a fan next to the winding machine to disperse water mist and smoke.

[0164] 3. By using the adaptive change of the CCD camera's exposure time to maintain a relatively stable light intensity in the image acquisition environment, the acquired image as a whole has a brightness visible to the naked eye. A light source is added above the crossbeam in front of each winding machine to increase the brightness of the strip tail.

[0165] Step 3 includes:

[0166] The recognition model is run on a GPU server configured with a 1080Ti. The mask-rcnn model takes about 160ms to complete one recognition task, so two GPU servers are required, with one server processing two images at a time.

[0167] Step 3.1: The image acquisition device acquires on-site images, with the center of the winding machine drum at the center of the 960*960 square window, and the on-site image is cropped.

[0168] Step 3.2: Send the cropped scene image to the recognition model. The recognition model outputs the tail category and the coordinates of the four corner points of the bounding box.

[0169] Step 3.3: If there is a tail category recognition result, proceed to step 3.4.

[0170] Step 3.4: Calculate the initial tail position angle of the strip and start the tail tracking task.

[0171] Step 4: The image acquisition device operates at intervals t. s = Acquire on-site images every 40ms, use the recognition model to track the position of the strip tail in real time based on the on-site images, and calculate the position angle of the strip tail.

[0172] With the center of the winding machine drum as the reference center (x center y center Using 0 o'clock as the 0° angle, and proceeding clockwise, [0°, 90°) is defined as the first quadrant, [90°, 180°) as the second quadrant, [180°, 270°) as the third quadrant, and [270°, 360°) as the fourth quadrant. The position angle θ of the strip tail is calculated based on the tail category and the coordinates of the four corner points of the bounding box. dec .

[0173] Step 5: Determine if there is a tail category identification result, based on the tail category and the tail position angle θ of the strip. dec Calculate the actual angle value θ ang .

[0174] During the strip coiling process, the actual angle value θ output by the control system ang The default angle is 0°. The actual angle value θ output by the control system is determined when the tail section is positioned to the left or right of the reference center, after a tail category identification result is obtained. ang =θ dec =65°, and calculate the estimated angle value θ for the next on-site image. pre =θ ang +θ cal= 65° + 1.5° (the current angular velocity w of the winding machine is 6.5 revolutions per minute).

[0175] As the strip continues to be wound, the control system calculates the actual angle value θ based on the tail category recognition result of the fifth on-site image. ang And calculate the estimated angle value θ of the 6th on-site image. pre :

[0176] When the fifth on-site image shows a tail category identification result, the current angular velocity w of the coiler is 6 revolutions per minute, and the tail position angle θ of the strip is... dec The angle is 75°, the weight value α is 0.5, and the actual angle value θ output by the control system is... ang =θ dec ·(1-α)+θ prei • α = 73°, and calculate the estimated angle value θ of the 6th on-site image. pre =θ ang +θ cal =73° + 1.4° = 74.4°.

[0177] When the sixth on-site image does not have a tail category recognition result, the actual angle value θ output by the control system is... ang The angle is 74.4°; the current angular velocity w of the winding machine is 6 revolutions per minute. Calculate the estimated angle value θ of the 6th on-site image. pre =θ dec +θ cal =74.4° + 1.4° = 75.8°.

[0178] It can be based on the actual angle value θ ang and the estimated angle value θ pre The angular velocity w of the coiler is adaptively adjusted to improve the automation rate of strip coiling.

[0179] Step 6: Return to step 4 and track the tail position of the strip in real time until the strip coiling is completed.

[0180] The above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A deep learning-based tail angle identification method for hot-rolled coiled steel strips, characterized by: The method comprises the following steps: Step 1: establishing a strip tail image data set and dividing it into a training data set and a verification data set; Step 2: constructing a deep learning model, feeding the training data set and the verification data set into the deep learning model for training, and obtaining a recognition model suitable for recognizing the tail position of a hot-rolled strip coil; Step 3: setting up a field environment, collecting field images during the coiling of the strip, and identifying the tail type of the strip in the field images by using the recognition model, calculating an initial tail position angle of the strip, and starting a tail tracking task; Step 4: collecting field images at intervals by using an image collection device, tracking the tail position of the strip in real time by using the recognition model according to the field images, identifying the tail type, and calculating a tail position angle of the strip; Step 5: during the coiling of the strip, determining whether there is an identification result of the tail type for each field image, and calculating an actual angle value according to the identification result of the tail type and the tail position angle of the strip; Step 6: determining whether the coiling of the strip is completed, and if so, ending the image collection and exiting the control program, and if not, returning to step 4. 2.The tail angle recognition method of hot-rolled coiled strip steel based on deep learning according to claim 1, characterized in that: The step 1 comprises: Step 1.1: collecting images of the tail end of the strip during the coiling of the strip by using an image collection device; Step 1.2: dividing the images into two types according to the position of the tail of the strip: (I) taking the center of the coiler drum (11) as a reference center (100), if the tail end (2) of the strip appears in the unobstructed area on the left side of the reference center (100), the image is recorded as a left tail image; (II) taking the center of the coiler drum (11) as a reference center (100), if the tail end (2) of the strip appears in the unobstructed area on the right side of the reference center (100), the image is recorded as a right tail image; Step 1.3: delineating the tail contour of the strip on the classified image into a closed area (3), and adding a label "left tail image" or "right tail image" to the closed area (3) according to the tail position of the strip; Step 1.4: taking the reference center (100) as the center and a square window as a clipping frame (4) to clip the labeled image, and saving it as a data set image; Step 1.5: repeating steps 1.1-1.4 to complete the saving of all data set images, and forming a strip tail image data set; Step 1.6: dividing all data set images in the strip tail image data set into a training data set and a verification data set according to the image quantity ratio. 3.The tail angle recognition method of hot-rolled coiled strip steel based on deep learning according to claim 2, characterized in that: In the step 1.4, the clipping pixel size is a multiple of 64; in the step 1.6, the image quantity ratio in the training data set and the verification data set is 8:2-9:

1. 4.The tail angle recognition method of hot-rolled coiled steel strip based on deep learning according to claim 1, characterized in that: The step 2 comprises: Step 2.1: selecting RestNet50 of mask-rcnn as a deep learning network architecture; Step 2.2: setting the size adjustment mode of the data set image to "square", that is, setting the short side length value and the long side length value of the data set image to be equal in size; Step 2.3: send the dataset image to the deep learning model, identify the strip tail in the dataset image through the deep learning model, and take the tail class loss L cls , the bounding box loss L box , and the tail shape mask loss L mask as the three targets of identification; Step 2.4: Model the loss function L: L = L cls + L box + L mask ; Step 2.5: Select the resnet50 weight of the original open source dataset COCO of mask-rcnn as the initial training weight of the deep learning model, set the training batch according to the recognition accuracy, set the number of iterations in each batch according to the number of training samples and the size of the device memory, and perform iterative training to obtain the loss function trend graph of the training dataset and the loss function trend graph of the validation dataset; Step 2.6: Select the optimal weight on the loss function trend graph to obtain an identification model suitable for the tail position of the hot-rolled strip coiling, and the input of the identification model is the cropped on-site image, and the output is the tail class and the coordinates of the four corner points of the bounding box.

5. The tail angle recognition method of a hot-rolled coil based on deep learning according to claim 4, characterized in that: In step 2.6, the selection condition of the optimal weight is that the loss function value of the training dataset and the loss function value of the validation dataset are minimum values in the loss function trend graph of the training dataset and the loss function trend graph of the validation dataset. 6.The tail angle recognition method of hot-rolled coiled strip steel based on deep learning according to claim 1, characterized in that: In step 3, the setting requirements of the on-site environment include: a. Under the condition of not affecting the material performance, turn off the cooling water of the upstream side guide plate and the coiling machine during the coiling process of the strip tail; b. Install a fan beside the coiler to blow away the water mist and smoke; c. Use adaptive exposure time and increase the light source. 7.The tail angle recognition method of hot-rolled coiled strip steel based on deep learning according to claim 1, characterized in that: Step 3 includes: Step 3.1: The image acquisition device acquires the on-site image, and the center of the coiler drum is the center of the square window, and the on-site image is cropped; Step 3.2: The cropped on-site image is sent to the identification model, and the identification model outputs the tail class and the coordinates of the four corner points of the bounding box; Step 3.3: The control system judges whether there is an identification result of the tail class, if yes, step 3.4 is executed, if not, step 3.1 is returned; Step 3.4: Calculate the initial strip tail position angle and start the tail tracking task. 8.The tail angle identification method of a hot-rolled coil based on deep learning according to claim 7, characterized in that: In step 3.4, the calculation method of the initial strip tail position angle is: Taking the center of the winding drum of the coiler as a reference circle center (x center , y center ), taking the 0 point of the clock as the 0° angle, in the clockwise direction, the region of [0°, 90°) in the cut field image is defined as the first quadrant, the region of [90°, 180°) is defined as the second quadrant, the region of [180°, 270°) is defined as the third quadrant, and the region of [270°, 360°) is defined as the fourth quadrant, and the tail part position angle θ dec of the strip steel is calculated according to the tail part category and the coordinates of the four corner points of the bounding box. (I) When the tail class is a right tail image: The right upper corner point of the boundary box is taken as the tail positioning point (x dec , y dec ), and the strip tail position angle θ dec of the tail positioning point is calculated. If the tail positioning point is located in the first quadrant, the strip tail position angle θ of the tail positioning point dec is calculated by the following formula: If the tail positioning point is located in the second quadrant, the strip tail position angle θ of the tail positioning point dec is calculated as follows: (II) When the tail class is a left tail image: The left edge midpoint of the boundary frame is taken as the tail positioning point (x dec , y dec ), and the strip tail position angle θ dec of the tail positioning point is calculated. If the tail positioning point is located in the third quadrant, the strip tail position angle θ of the tail positioning point is dec The calculation formula is: If the tail positioning point is located in the fourth quadrant, the strip tail position angle θ of the tail positioning point dec is calculated by the following formula: 9.The tail angle recognition method of hot-rolled coiled steel strip based on deep learning according to claim 1, characterized in that: In step 5, The actual angle value θ output by the control system in the strip coiling process ang The actual angle value θ output by the control system is 0° by default until the tail category has a recognition result, and the strip tail is located on the left or right side of the reference circle center ang = θ dec The estimated angle value θ of the next field image is calculated pre The calculation formula is: θ pre = θ ang + θ cal (5) Wherein, θ cal is the angle corresponding to the angular velocity w of the current coiler, and the conversion formula is: θ cal = w·t s ·360° / 60, unit: ° / s; With the continuous coiling of the strip, the control system calculates the actual angle value θ according to the tail category recognition result of the i-th field image ang , and calculates the estimated angle value θ of the next field image pre : (I) When the recognition result of the tail category of the i-th field image is obtained, the control system outputs the actual angle value θ ang The calculation formula is: θ ang = θ dec • (1 - a) + θ prei • a (6) Wherein, a is a weight value, a value range is 0-1; θ prei is the estimated angle value of the i-th live image; Calculate the estimated angle value θ for the next on-site image. pre The calculation formula is: θ pre =θ ang +θ cal ; (II) When the recognition result of the tailless category of the i-th live image is correct, the control system outputs the actual angle value θ ang The calculation formula is: θ ang = θ prei (7); calculating an estimated angle value θ of the next live image pre , the calculation formula is: pre = θ dec + θ cal .

Images