A cold-rolled tension roller surface cleaning system and method based on visual detection

Abstract

This invention proposes a visual inspection-based cold rolling tension roll surface cleaning system and method, belonging to the field of visual roll surface cleaning technology. It includes a frame, fixedly mounted on one side of the tension roll body, with components spaced parallel to the axial direction of the tension roll; an axial feed device straddling the frame, equipped with a first movable unit for driving a radial feed device on the first movable unit to move along the axial direction of the roll body; a second movable unit on the radial feed device, which drives a grinding head and a camera module to move radially along the roll body; the grinding head for performing a circumferential cleaning of the tension roll surface; the camera module for capturing images of the roll body surface; and a controller, which, through a built-in algorithm, acquires images of the roll body surface to evaluate the degree of contamination, thereby driving the grinding head to the contaminated location for cleaning. After the grinding head has cleaned the contaminated location, the controller again evaluates the cleaning effect of the grinding head based on the images of the roll body surface.

Description

Technical Field

[0001] This invention relates to the field of visual roller surface cleaning technology, and in particular to a cold rolling tension roll surface cleaning system and method based on visual inspection. Background Technology

[0002] Cold-rolled galvanizing / continuous annealing lines utilize numerous roller conveyors as guide rollers, tension rollers, etc. After prolonged operation, as cold-rolled thin sheets pass through these conveyors, residual iron powder, zinc powder, and dust adhere to the roller surfaces. With continuous production, this contaminant accumulates on the roller surfaces, easily polluting the steel sheet surface and causing problems such as strip scratches, foreign object indentation, and roller marks. Because tension rollers experience significant radial forces during operation, surface contamination leading to surface defects in the strip is particularly pronounced. Therefore, it is necessary to clean impurities from the roller surfaces online or periodically to prevent secondary contamination of the thin strip surface, thereby reducing the rejection rate of finished products.

[0003] Tension rollers come with built-in scraper cleaning devices, which offer advantages such as simple structure, low cost, and real-time roller surface cleaning. However, traditional scraper roller surface cleaning devices are only effective at removing large particles of impurities and are ineffective at cleaning small particles such as zinc powder, easily forming ring-shaped particle bands and scratching the roller surface. Furthermore, scraper replacement and maintenance are complex, requiring downtime, which increases labor intensity and production costs.

[0004] Therefore, it is essential to provide a vision-based cold rolling tension roll surface cleaning system and method. This system cleans the entire roll surface by feeding the composite cleaning unit in both the axial and radial dimensions relative to the roll body. Simultaneously, an improved Transformer image detection algorithm is used to detect the cleaning results, and the cleaning strategy is dynamically adjusted based on the cleaning results. This achieves efficient and intelligent roll surface cleaning. Summary of the Invention

[0005] In view of this, the present invention proposes a vision-based cold rolling tension roll surface cleaning system and method that achieves cleaning of the entire roll surface by feeding the composite cleaning unit relative to the roll body in both axial and radial dimensions, and simultaneously uses an improved Transformer image detection algorithm to detect the cleaning results, and dynamically adjusts the cleaning strategy according to the cleaning results.

[0006] On one hand, the present invention provides a visual inspection-based cold rolling tension roll surface cleaning system, comprising: The frame is fixedly installed on one side of the tension roller body and is arranged at intervals along the axial direction of the tension roller. An axial feed device is mounted across the frame. The axial feed device is equipped with a first movable unit, which moves linearly in a direction parallel to the axial direction of the tension roller. A radial feed device is mounted on the first movable unit, and a second movable unit is mounted on the radial feed device. The second movable unit moves linearly in a direction parallel to the radial direction of the tension roller. The grinding head, located on the second movable unit, is used to approach the surface of the tension roller and perform a circular cleaning of the tension roller surface. A camera module is mounted on the second active unit, with its viewfinder facing the tension roller body, for taking pictures of the roller body surface; The controller communicates with the camera module, axial feed device, radial feed device and grinding head. Through built-in algorithms, it obtains images of the roller surface to evaluate the degree of contamination, thereby driving the grinding head to the contaminated location for cleaning. On the other hand, after the grinding head grinds the contaminated location, the controller evaluates the cleaning effect of the grinding head again based on the images of the roller surface.

[0007] Based on the above technical solutions, preferably, the axial feed device includes a mounting base, a coupling, a servo motor, a slide table, and a synchronous belt. The mounting base is used to be fixedly connected to at least one pair of frames. Synchronous pulleys are rotatably mounted on both ends of the mounting base. The servo motor is connected to the synchronous pulleys via the coupling. The synchronous belt is sleeved on the synchronous pulleys and is rolledly connected to them. A slide table is fixedly mounted on the synchronous belt. The slide table serves as the first movable unit, driving the radial feed device, grinding head, and camera module to move linearly in a direction parallel to the axial direction of the tension roller.

[0008] Preferably, the radial feed device includes a cylinder and a piston rod. The cylinder is fixedly connected to the slide table and has a medium inlet and a medium outlet. One end of the piston rod extends into the cylinder and is slidably connected to the cylinder. The other end of the piston rod extends toward the surface of the tension roller. The end of the piston rod extending out of the cylinder serves as a second movable unit, used to drive the grinding head and camera module to move linearly in a direction parallel to the radial direction of the tension roller under the drive of the pressure medium.

[0009] Preferably, the grinding head includes a grinding head body and an eccentric grinding disc. The grinding head body is fixedly connected to the second movable unit, and a detachable grinding head material is provided on the side of the eccentric grinding disc away from the grinding head body.

[0010] On the other hand, the present invention provides a method for cleaning the surface of cold rolling tension rolls based on visual inspection, comprising the following steps: Configure the above-mentioned vision-based cold rolling tension roll surface cleaning system; the controller includes a PLC and an image workstation; an incremental encoder is configured on the rotating shaft of the tension roll; After the equipment is started, the camera module continuously acquires images of the roller surface and sends the images to the image workstation. The image workstation performs real-time stitching and processing of the images based on the improved feature recognition image stitching algorithm, and then identifies the contamination rate of the roller surface through the image defect detection algorithm architecture. The image defect detection algorithm architecture includes a multi-layer feature extractor, an improved feature encoder, an improved feature decoder, a feature classifier, and an anomaly score smoother. The original image is fed into the multi-layer feature extractor, which maps the original pixel space to a high-dimensional feature space, extracting semantic and detail information at different levels to obtain a multi-scale feature map. Subsequently, the multi-scale feature map is input into the improved feature encoder for unified encoding, modeling the local structure and global context relationship in the multi-scale feature map to obtain high-dimensional encoded features. After encoding, the improved feature encoder outputs two branches: one branch connects to the feature classifier, which is used to discriminate and learn the high-dimensional encoded features to distinguish the probability distribution of normal and abnormal samples at the overall semantic level; the other branch connects to the improved feature decoder, which reconstructs and models the high-dimensional encoded features to obtain a reconstructed feature map. The anomaly score smoother measures the difference between the reconstructed feature map and the multi-scale feature map, and outputs an anomaly score heatmap that measures the contamination rate of the roller surface. If the image workstation detects that the contamination rate on the roller surface does not exceed the preset threshold, the axial feed device will drive the radial feed device, grinding head, and camera module to move to the next detection position in a direction parallel to the axial direction of the tension roller by a preset step size. If the image workstation detects that the contamination rate on the roller surface exceeds the preset threshold, the PLC will pause the axial feed device, allowing the grinding head to press against the roller surface for cleaning. If the contamination rate still exceeds the preset threshold after cleaning, the PLC will activate the radial feed device, extending the second movable unit and increasing the grinding pressure of the grinding head. After grinding, the contamination rate on the roller surface will be reassessed. If the radial feed device reaches the maximum radial feed stroke or pressure limit, and the contamination rate on the roller surface still exceeds the preset threshold after grinding, the PLC will pause the cleaning operation and issue an alarm through the human-machine interface, prompting the replacement of the grinding head or manual intervention.

[0011] Preferably, the image workstation performs real-time image stitching and processing based on an improved feature recognition image stitching algorithm. Specifically, this includes rotating the shaft of the tension roller and causing the roller body to rotate at least one revolution, with the camera module acquiring video of the roller surface at a fixed position to obtain a video slice of the entire roller body. The image workstation performs frame preprocessing on the video slice based on the displacement of the incremental encoder, finds feature points in the frame images, marks the feature points in the frame images using descriptors, sends the feature points to a matcher, matches the feature points of two adjacent frame images, generates a list of matching feature points, removes duplicate areas in adjacent frame images, and so on, until all frame images of the video slice are stitched together to obtain the original image of the roller surface after unfolding.

[0012] More preferably, the multi-layer feature extractor includes four consecutive convolutional layers, upsampling, and a multi-scale fusion stage, with the output specifications of the four consecutive convolutional layers being respectively... , , , , X and Y These represent the width and height of the original image, respectively. Adjacent consecutive convolutional layers are connected via residual blocks, and the outputs of the third and fourth consecutive convolutional layers are upsampled to the normal size. The output of the second consecutive convolutional layer is then fed into a multi-scale fusion stage to obtain a multi-scale feature map. T .

[0013] Preferably, the improved feature encoder includes a first convolutional layer, a first batch normalization layer, a first ReLU activation layer, a second convolutional layer, a second batch normalization layer, a first adder, a second ReLU activation layer, a first feature block, and a second feature block, forming a multi-scale feature map. T After processing by the first convolutional layer and the first batch of normalization layers, nonlinearity is introduced through the first ReLU activation layer, followed by processing by the second convolutional layer and the second batch of normalization layers to obtain the residual mapping. F ( T Then residual mapping F ( T ) and multi-scale feature maps TAfter element-wise addition by the first adder, a second ReLU activation layer is applied to obtain an intermediate feature map, which is then fed into the first feature block. The first feature block includes a first-layer normalization unit, a convolutional attention layer, and a second adder. The second feature block includes a second-layer normalization unit, a convolutional multilayer perceptron, and a third adder. The first feature block sequentially performs layer normalization and convolutional self-attention processing on the input intermediate feature map. The processing result is then added element-wise with the input of the first feature block by the second adder and fed into the second feature block. The second feature block performs layer normalization and convolutional multilayer perceptron processing on the input and feeds the processing result into the third adder for element-wise addition. Finally, a high-dimensional encoded feature that enhances the expression of the contamination pattern on the roller surface is output.

[0014] In a further preferred embodiment, the improved feature decoder includes deconvolution operations, a third batch normalization layer, and a third ReLU activation layer. Through deconvolution operations and the third batch normalization layer, the third ReLU activation layer introduces nonlinearity, reconstructing the high-dimensional encoded features output by the improved feature encoder into a multi-scale feature map. T Reconstructed feature maps of the same size.

[0015] Preferably, the feature classifier includes a first multilayer perceptron (MLP), a softmax mapping, and a linear classifier. The first MLP includes two fully connected layers and a nonlinear activation function, used to model the importance of the high-dimensional encoded features output by the improved feature encoder according to sequence features and re-aggregate them. The high-dimensional encoded features are rewritten into feature vectors in sequence form. The first MLP with shared parameters scores the feature vectors in the sequence to characterize the importance of different feature vectors in global anomaly detection. s The feature vector scoring results are normalized by Softmax mapping along the sequence dimension to obtain the weight distribution of the feature vectors. The weights of each feature vector are then weighted and converged with the original sequence-form feature vectors to obtain the global feature representation. The global feature representation is then fed into a linear classifier, and the abnormal and normal probability distributions of the entire roller surface are output through the Softmax function to provide discrimination constraints.

[0016] Preferably, the anomaly score smoother first estimates the variance of the reconstructed feature map using a second multilayer perceptron based on the reconstruction error between the reconstructed feature map and the multi-scale feature map. The output of the second multilayer perceptron... δ This is used to correct reconstruction errors and convert the reconstruction errors into anomaly score heatmaps. S Anomaly scoring heatmap is defined as , It is a reconstruction error. It is the L2 norm. This represents a division operation on the number of pixels. The resulting anomaly score heatmaps are processed through convolutional layers with kernel sizes of 1, 4, and 16 to obtain anomaly score maps with different smoothness levels. The final anomaly score heatmap is obtained by adding the anomaly score maps with different smoothness levels pixel by pixel and taking the average. S 0. By obtaining the pixel points whose defect intensity in the final abnormal score heatmap is greater than the design hue threshold, the contamination rate of the roller surface is identified.

[0017] The present invention provides a visual inspection-based cold rolling tension roll surface cleaning system and method, which has the following advantages compared with the prior art: 1. This invention provides a tension roller surface cleaning system, in which the movement of each component is controlled by a PLC. The grinding head grinds the dirty areas of the roller body to eliminate the potential contamination of the steel plate surface by the dirt. Combined with visual means, the dirt is located and the effect before and after cleaning is compared to achieve closed-loop control of the cleaning process. When the cleaning effect is not satisfactory after multiple cleanings, the system can remind the management personnel to replace the grinding head or adjust the position of the grinding head to better improve the grinding effect.

[0018] 2. In the identification of defects on the roller surface, this invention feeds the original image into a multi-layer feature extractor, mapping the image from pixel space to a high-dimensional feature space. This process extracts semantic and detailed information at different levels to obtain multi-scale features, thereby providing a more discriminative feature representation for subsequent anomaly detection and avoiding the instability caused by directly modeling anomalies in pixel space. Subsequently, the extracted multi-scale features are input into an improved feature encoder for unified encoding. The encoding process aims to model the local structure and global context relationship in the feature map, enabling the feature representation to maintain spatial correlation while incorporating long-range dependency information, providing more sufficient contextual basis for distinguishing between abnormal and normal features. After encoding, the network structure is divided into two parallel functional branches to address different levels of anomaly detection requirements. One branch introduces a feature classifier to perform discriminative learning on the encoded global features, guiding the model to distinguish between normal and abnormal samples at the overall semantic level, thereby improving the stability and discriminative ability of global anomaly detection. Another branch reconstructs and models the encoded features through the feature decoding process to restore the original feature space representation. In the local anomaly detection process, the model constructs an initial anomaly score map by measuring the difference between the reconstructed features and the original features, introduces an anomaly score smoothing mechanism, imposes spatial consistency constraints on the anomaly distribution, and finally generates a continuous and more interpretable final anomaly score heatmap for accurately locating abnormal regions in the image.

[0019] 3. By converting the final anomaly score heatmap into the HSV color space, calculating the hue, and obtaining the defect intensity at the image location, the degree of contamination on the roller surface and the grinding effect of the roller before and after cleaning are evaluated by statistically analyzing all cases where the defect intensity is greater than the set threshold. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a perspective view of a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention. Figure 2 This is a perspective view of the tension roll body of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 3 This is a perspective view of the axial feeding device of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 4 This is a perspective view of the radial feeding device of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 5 This is a perspective view of the grinding head of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 6 This is a perspective view of the frame of a cold rolling tension roll surface cleaning system and method based on vision inspection according to the present invention. Figure 7 This is a perspective view of the combined state of the cleaning system and the tension roll body in the cold rolling tension roll surface cleaning system and method based on vision detection of the present invention. Figure 8 This is a framework diagram of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 9 This is a flowchart of the servo control process for cleaning the roll body of a cold rolling tension roll surface cleaning system and method based on vision detection according to the present invention. Figure 10 This is a video stitching flowchart of a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention; Figure 11 This is a diagram illustrating the image defect detection algorithm architecture of a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention. Figure 12 This is a flowchart of the algorithm of an improved feature encoder for a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention. Figure 13 This is a flowchart of the algorithm for an improved feature decoder of a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention. Figure 14 This is a closed-loop control flowchart of the cleaning efficiency and defect rate of a visual inspection-based cold rolling tension roll surface cleaning system and method according to the present invention.

[0022] Reference numerals: 1. Frame; 2. Radial feed device; 3. Grinding head; 4. Axial feed device; 5. Camera module; 100. Tension roller frame; 200. Roller body; 300. Bearing housing; 400. Flange; 41. Mounting base; 42. Coupling; 43. Servo motor; 44. Slide table; 45. Synchronous belt; 21. Piston rod; 22. Cylinder body; 23. Media inlet; 24. Media outlet; 31. Grinding head body; 32. Eccentric grinding disc; 11. Pad plate; 12. Aluminum profile; 13. Angle bracket. Detailed Implementation

[0023] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0024] Traditional scraper roller cleaning devices are only effective at removing large particles of impurities, and are ineffective at removing small particles such as zinc powder, easily forming ring-shaped particle bands and scratching the roller surface. Furthermore, scraper replacement and maintenance are complex, requiring downtime, increasing labor intensity and production costs. Even after using conventional methods to clean foreign objects from the roller surface, a steel plant still experienced a 1.5% scratch rate due to zinc powder on the roller surface in 2024. Traditional cleaning methods urgently need improvement.

[0025] In view of this, such as Figures 1-7 As shown, in one aspect, the present invention provides a visual inspection-based cold rolling tension roll surface cleaning system, comprising: The frame 1 is fixedly installed on one side of the tension roller body and is arranged at intervals along the axial direction of the tension roller. An axial feed device 4 is mounted across the frame 1. A first movable unit is provided on the axial feed device 4. The first movable unit moves linearly in a direction parallel to the axial direction of the tension roller. Radial feed device 2 is mounted on the first movable unit, and a second movable unit is mounted on the radial feed device 2. The second movable unit moves linearly in a direction parallel to the radial direction of the tension roller. Grinding head 3 is mounted on the second movable unit and is used to approach the surface of the tension roller and perform circumferential cleaning on the surface of the tension roller. Camera module 5 is mounted on the second active unit, with its viewfinder facing the tension roller body, and is used to capture images of the roller body surface. The controller is connected to the camera module 5, the axial feed device 4, the radial feed device 2 and the grinding head 3. Through the built-in algorithm, it obtains images of the roller surface to evaluate the degree of contamination, thereby driving the grinding head 3 to reach the contaminated position for cleaning. On the other hand, after the grinding head 3 grinds the contaminated position, the controller evaluates the cleaning effect of the grinding head 3 again based on the images of the roller surface.

[0026] like Figure 1 As shown, in the above structure, the axial feed device 4 drives the radial feed device 2, the grinding head 3, and the camera module 5 to move linearly along a direction parallel to the axial direction of the tension roller. The stroke of the axial feed device 4 is greater than the axial length of the roller to achieve cleaning of the entire roller surface. The radial feed device 2 can drive the grinding head 3 and the camera module 5 to move linearly along a direction parallel to the radial direction of the roller. In one embodiment, this radial direction is the direction of gravitational acceleration, allowing the grinding head 3 to touch the roller surface and apply a certain pressure. When the grinding head 3 contacts the roller surface, it will begin cleaning. Through the low-speed rotation of the roller, the roller surface can be cleaned in a circular motion. The camera module 5 will take pictures and slice the roller surface every few revolutions to obtain images of the roller surface, which are then sent to the controller to evaluate the cleaning effect.

[0027] like Figure 2 As shown in the figure, the roller structure of a steel plant is as follows. The tension roller is mounted on the tension roller frame 100. The roller body 200 is made of tungsten carbide, which has extremely high hardness and wear resistance. The roller body is 2300mm long and 1000mm in diameter. Bearing seats 300 are set at both ends of the roller body 200 and are fixed to the frame by four hexagonal bolts. The bearing seats 300 are encapsulated by flanges 400 to prevent the bearings from being affected by small particles in the workshop environment.

[0028] like Figure 1 , Figure 3 and Figure 7As shown, the axial feed device 4 includes a mounting base 41, a coupling 42, a servo motor 43, a slide 44, and a synchronous belt 45. The mounting base 41 is used to fix and connect to at least one pair of frames 1. Synchronous pulleys are rotatably mounted on both ends of the mounting base 41. The servo motor is connected to the synchronous pulleys via the coupling 42. The synchronous belt 45 is sleeved on the synchronous pulleys and is rolled and connected to them. The slide 44 is fixedly mounted on the synchronous belt 45. The slide 44 serves as the first movable unit, driving the radial feed device 2, the grinding head 3, and the camera module 5 to move linearly in a direction parallel to the axial direction of the tension roller. In this embodiment, the core component of the axial feed device 4 can be a synchronous belt slide module mounted on the frame 1. The synchronous pulleys are connected to the output shaft of the servo motor 43 via the coupling 42. The power source for the axial feed device 4 is an external 200W servo motor 43. The mounting base 41 is mounted on the frame 1 by trapezoidal bolts.

[0029] like Figure 4 As shown, the radial feed device 2 includes a cylinder 22 and a piston rod 21. The cylinder 22 is fixedly connected to the slide table 44, and a medium inlet 23 and a medium outlet 24 are provided on the cylinder 22. One end of the piston rod 21 extends into the cylinder 22 and is slidably connected to the cylinder. The other end of the piston rod 21 extends toward the surface of the tension roller. The end of the piston rod 21 extending out of the cylinder 22 serves as a second movable unit, used to drive the grinding head 3 and the camera module 5 to move linearly in a direction parallel to the radial direction of the tension roller under the drive of the pressure medium. The piston rod adopts a dual-axis structure, which has a larger contact area with the second movable unit and makes the movement more stable. The end of the second movable unit can adopt a plate-like structure as shown in the figure, which facilitates the installation of the grinding head 3 and the camera module 5 by bolts. The radial feed device 2 is driven by a pneumatic source, and the extension and retraction stroke of the piston rod is adjusted by inputting or discharging compressed gas.

[0030] Selecting a synchronous belt module for the axial feed device 4 can minimize the contamination of the internal precision mechanical structure by material dust in the cold rolling workshop, thereby reducing manual maintenance costs. Simultaneously, using a synchronous belt as the axial feed device offers advantages such as low friction, high efficiency, and high precision, ensuring the stability and accuracy of the cylinder feed system during long-term operation. Compared to traditional lead screws, synchronous belt modules effectively reduce energy loss, provide a smoother feed force, and also offer lower noise and a longer service life. Combined with the cylinder's position adjustment function, precise control can be achieved to meet the needs of high-precision polishing operations, making it suitable for roll surface repair and precision machining. By adjusting the air pressure and feed speed, flexible feed is achieved, reducing the impact caused by rigid contact during grinding and improving the surface quality. Compared to traditional mechanical feed, the cylinder structure is simple, control is flexible, and feed characteristics can be optimized by combining flow valves or buffer mechanisms to ensure a uniform and stable grinding process.

[0031] like Figure 5 As shown, the grinding head 3 includes a grinding head body 31 and an eccentric grinding disc 32. The grinding head body 31 is fixedly connected to the second movable unit, and a detachable grinding head material is provided on the side of the eccentric grinding disc 32 away from the grinding head body 31. In this embodiment, the grinding head 3 can use an automated pneumatic grinding module. Through uniform grinding by the eccentric grinding disc and fine polishing by using the grinding head body of a pneumatic polishing machine, it can also reduce overheating during the grinding process, provide continuous and stable power, and is suitable for long-term, high-efficiency operation. Its low impact force and high-precision control help avoid surface damage and extend the service life of the roll. Combined with appropriate polishing materials, the pneumatic polishing machine provides an ideal solution for surface repair and detail finishing of cold-rolled rolls.

[0032] like Figure 6 As shown, the custom-structured frame 1 includes two sets of pads 11, aluminum profiles 12, and angle brackets 13. The two sets of pads 11 are used to connect to the top of the tension roller frame 100. The aluminum profile 12 uses European standard 4080Z-2.5 aluminum profile with a wall thickness of 2.5mm, a weight of 2.85kg / m, and a total length of 2620mm, which can cover the entire synchronous belt module. The synchronous belt module can be fastened to the aluminum profile 12 with bolts. The aluminum profile 12 is fixed to the two sets of pads 11 by angle brackets 13. Figure 7 As shown, when the visual inspection-based cold rolling tension roll surface cleaning system is correctly installed, the axial feed device 4, in conjunction with the radial feed device 2, can move within a range that allows the grinding head to reach any position on the roll surface. In one embodiment, the frame 1 ensures that the feed axis of the radial feed device 2 is exactly orthogonal to the axis of the roll.

[0033] On the other hand, the present invention provides a method for cleaning the surface of cold rolling tension rolls based on visual inspection, comprising the following steps: Configure the above-mentioned vision-based cold rolling tension roll surface cleaning system; the controller includes a PLC and an image workstation; and an incremental encoder is configured on the rotation shaft of the tension roll.

[0034] The control system framework diagram is as follows: Figure 8As shown, the entire equipment is powered by 220V. The servo motor uses single-phase 220V, while the PLC, HMI, camera module 5, and proportional valves are powered by DC power converted from a switching power supply. Compressed air drives the pneumatic grinding head to rotate for cleaning, and the grinding head speed is controlled by an electronic proportional control valve. The pressure of the grinding head pressing against the roller is also dynamically adjusted by the electronic proportional valve. The equipment is also equipped with an emergency stop button for emergency shutdown in case of emergency, ensuring personal and equipment safety. The control system can operate in two modes: one is that the vision system monitors the degree of contamination on the roller surface in real time, using a fast cleaning method for areas with relatively light contamination to shorten cleaning time; the other is that heavily contaminated areas are cleaned slowly to prioritize cleaning effectiveness. Users can select the working mode with a single click through the HMI according to the actual working conditions. The HMI transmits control commands to the PLC via serial port, which then drives the PLC, saving cleaning time. An incremental encoder is used to obtain the angular displacement of the rotating shaft.

[0035] like Figure 9 As shown in the diagram, the servo control flowchart is illustrated. After the equipment starts, the grinding head starts and rotates, and then the axial feed device drives the grinding head to accurately determine the initial working coordinates. The camera module 5 continuously acquires images of the roller surface and sends the images to the image workstation. The image workstation performs real-time stitching and processing of the images based on an improved feature recognition image stitching algorithm.

[0036] like Figure 10 The diagram shows a flowchart of an image stitching algorithm based on improved feature recognition for an image workstation. Image stitching techniques include direct stitching and feature-based stitching. Direct stitching achieves this by minimizing differences between pixels, while feature-based stitching extracts and matches features. Because direct methods are slower and less robust than feature-based methods, they are increasingly used in current image stitching tasks. Therefore, this embodiment adopts an image stitching algorithm based on improved feature recognition. This algorithm works in conjunction with an incremental encoder, which calculates the relative displacement between captured frames, generating realistic unfolded images even for cylindrical rollers. The inter-frame distance depends on the roller's rotational speed and the frame rate of the camera module 5. In this embodiment, a Raspberry Pi v2 camera module is used to capture images of the roller surface at a resolution of 1920×1080 pixels and a frame rate of 30fps. During cleaning operations, the roller rotates at a low speed of 3 rpm.

[0037] Since the camera module 5 can capture a complete circle of the roller cleaning area with just one rotation of the roller, the video recording needs to be truncated according to the rotational speed of the roller to be cleaned during recording to obtain a video slice of the entire roller. The image workstation performs frame preprocessing on the video slice based on the displacement of the incremental encoder to find significant feature points. Subsequently, descriptors are used to mark the feature points in the frame images, and the same process is repeated for the previous frame image. For this purpose, a region corresponding to the width of the current frame needs to be selected from the previously stitched panoramic image, which means that the first frame will be selected twice in the first stitching iteration. After the feature points are selected, these feature points are sent to the matcher. By matching the feature points in the two images, a list of matching feature points is generated, and duplicate regions in adjacent frame images are removed. This process is repeated until all frame images of the video slice are stitched together, and the original image of the unfolded roller surface is obtained.

[0038] like Figure 9 As shown, the contamination rate of the roller surface is identified through an image defect detection algorithm architecture. If the image workstation detects that the contamination rate of the roller surface does not exceed a preset threshold, the axial feed device 4 will drive the radial feed device 2, the grinding head 3, and the camera module 5 to move at a preset step size to the next detection position in a direction parallel to the axial direction of the tension roller, until all detection positions are processed. If the image workstation detects that the contamination rate of the roller surface exceeds the preset threshold, the PLC will pause the axial feed device 4, allowing the grinding head 3 to be held against the roller surface for cleaning. If the contamination rate still exceeds the preset threshold after cleaning, the PLC will start the radial feed device 2, causing the second movable unit to extend and increase the grinding pressure of the grinding head 3. After grinding, the contamination rate of the roller surface will be reassessed. If the radial feed device 2 reaches the maximum radial feed stroke or pressure limit, and the contamination rate of the roller surface still exceeds the preset threshold after grinding, the PLC will pause the cleaning operation and issue an alarm through the human-machine interface, prompting the replacement of the grinding head 3 or manual intervention.

[0039] like Figure 11As shown, the image defect detection algorithm architecture mentioned above includes a multi-layer feature extractor, an improved feature encoder, an improved feature decoder, a feature classifier, and an anomaly score smoother. The original image is fed into the multi-layer feature extractor, which maps the original pixel space to a high-dimensional feature space, extracting semantic and detail information at different levels to obtain a multi-scale feature map. Subsequently, the multi-scale feature map is input into the improved feature encoder for unified encoding, modeling the local structure and global context relationships in the multi-scale feature map to obtain high-dimensional encoded features. After encoding, the improved feature encoder outputs two branches: one branch connects to the feature classifier, used for discriminative learning of the high-dimensional encoded features to distinguish the probability distribution of normal and abnormal samples at the overall semantic level; the other branch connects to the improved feature decoder, reconstructing and modeling the high-dimensional encoded features to obtain a reconstructed feature map. The anomaly score smoother measures the difference between the reconstructed feature map and the multi-scale feature map, outputting an anomaly score heatmap that measures the contamination rate of the roller surface.

[0040] In the local anomaly detection process, the image defect detection algorithm architecture constructs an initial anomaly score map by measuring the difference between the reconstructed feature map and the original feature map. Considering the potential discreteness and noise interference in the anomaly response in the feature space, this embodiment further introduces an anomaly score smoothing mechanism to constrain the spatial consistency of the anomaly distribution, ultimately generating a continuous and more interpretable final anomaly score heatmap. S 0 is used to accurately locate abnormal areas in an image.

[0041] Assuming the original image is , Z The original image has [number] channels. The multi-scale feature map obtained after processing by the multi-layer feature extractor is [number]. , representing multi-scale feature maps T Classified as The block is embedded, and the width of the block is... X 1. The height of the block is Y 1. Number of channels is Z 1. Then convert the multi-scale feature map T The features are sequentially fed into the improved feature encoder and the improved feature decoder for reconstruction, and the reconstructed feature map is as follows: . R Represents the real number field.

[0042] The process of multi-scale feature map reconstruction is represented as follows: , Formula 1, This represents the functional form of the reconstruction. The mean squared loss (MSE) is used to measure the loss function. Formula 2, where As reconstruction error Using the L2 norm, since the reconstruction error can vary significantly on normal features, it is necessary to minimize the reconstruction error on normal features and amplify the error on abnormal features. This embodiment uses a multilayer perceptron (MLP) network to estimate the variance of the reconstructed feature map, and the output of the improved feature decoder... This is used to correct reconstruction errors and convert them into an anomaly scoring heatmap, which is defined as follows: , Formula 3, This represents a division operation on the number of pixels.

[0043] To further optimize the training process, the reconstruction error of the multilayer perceptron network (MLP) and the output of the improved feature decoder are incorporated as part of the MSE loss. The MLP loss is expressed as follows: Formula 4. To further process global defect detection, this embodiment incorporates Label Smoothing Cross Entropy Loss (LSCE) to optimize the training process. The LSCE loss is... Formula 5, where N The number of categories representing the samples. m i as well as l i Representing the classification number respectively i The predicted value and label value of the current sample. p i Is for the first i Predicted probabilities for each category, label values l i Expressed as follows: Formula 6, where α The parameter for the predicted value can be set to 0.05; t This represents the current sample number. To ensure that the result is absolute when a certain loss is significant, the three loss functions mentioned above need to be normalized. First, the network parameters... θ Calculate the gradient: Formula 7, where L MSE , L MLP and L LSCE These are MSE loss, MLP loss, and LSCE loss, respectively. Network parameters θ Find the gradient. g MSE , g MLP and g LSCEThese are the MSE loss gradient, MLP loss gradient, and LSCE loss gradient. The target gradient is calculated based on these three loss gradients. Formula 8, where Let be the gradients of the MSE loss and the MLP loss. Calculate the normalized loss as follows: , formula 9; Formula 10, where For MSE loss and MLP loss, This parameter, with a value of 1e-8, is used to prevent the denominator from being zero or the gradient from being too small, which could lead to an explosion. Finally, all losses are combined to form the total normalized loss function: , Formula 11, To adjust the weight coefficients of the loss function, which control the weights of regression loss and classification loss respectively, they are set to 0.3 in this embodiment.

[0044] The components of the image defect detection algorithm architecture will now be explained in detail.

[0045] 1. For example Figure 11 As shown, the first step in improving the Transformer image detection algorithm is to extract multi-scale feature maps from the input roll surface unfolded image. The first three layers of a ResNet with a pyramid structure are employed, combined with upsampling and skip connections to obtain semantic features at different receptive field scales. This structure seamlessly integrates with the subsequent structure, the improved feature encoder, and serves as the foundational module providing local and high-level contextual information throughout the entire reconstruction framework.

[0046] The multi-layer feature extractor includes four consecutive convolutional layers, upsampling, and multi-scale fusion steps to extract features from the original image. P Mapping to multi-scale feature maps T The output specifications of the four consecutive convolutional layers are as follows: , , , , X and Y These represent the width and height of the original image, respectively. Adjacent consecutive convolutional layers are connected via residual blocks, and the outputs of the third and fourth consecutive convolutional layers are upsampled to the normal size. The output of the second consecutive convolutional layer is then fed into a multi-scale fusion stage to obtain a multi-scale feature map. T Table 1 below shows the parameters and operations of four consecutive convolutional layers.

[0047] Table 1. Parameters and operation of the multilayer feature extractor

[0048] 2. For example Figure 12The diagram illustrates the structure of the improved feature encoder. The improved feature encoder includes a first convolutional layer, a first batch normalization layer, a first ReLU activation layer, a second convolutional layer, a second batch normalization layer, a first adder, a second ReLU activation layer, a first feature block, and a second feature block, along with multi-scale feature maps. T After processing by the first convolutional layer and the first batch of normalization layers, nonlinearity is introduced through the first ReLU activation layer, followed by processing by the second convolutional layer and the second batch of normalization layers to obtain the residual mapping. F ( T Then residual mapping F ( T ) and multi-scale feature maps T After element-wise addition by the first adder, a second ReLU activation layer is applied to obtain an intermediate feature map.

[0049] Formula 12, where the parameters Conv1, BN1, ReLU, Conv2, and BN2 correspond to the sequential processing of the first convolutional layer, the first batch of normalization layers, the first ReLU activation layer, the second convolutional layer, and the second batch of normalization layers, respectively. Formula 13, where the plus sign and ReLU correspond to the sequential processing of the first adder and the second ReLU activation layer.

[0050] See Figure 12 The lower half, then the intermediate feature map T The input is fed into the first feature block, which includes a first-layer normalization unit, a convolutional attention layer, and a second adder. The second feature block includes a second-layer normalization unit, a convolutional multilayer perceptron, and a third adder. The first feature block sequentially performs layer normalization and convolutional self-attention processing on the input intermediate feature map. The second adder then adds the processing result element-wise to the input of the first feature block, and the result is fed into the second feature block. The second feature block performs layer normalization and convolutional multilayer perceptron processing on the input, and the result is added element-wise to the input of the second feature block by the third adder. The final output is a high-dimensional encoded feature that enhances the expression of the contamination pattern on the roller surface. The processing process can be described by the following formula: , Formula 14, Formula 15, where T 1 and T 2 represents the output of the first feature block and the second feature block, respectively. LN ( ) represents the independent variable T 0 or T 1. Perform layer normalization. A conv Represents the independent variable T 0. Use the convolutional attention module. Mconv Independent variable of the representative team T 1. Processing is performed using a convolutional multilayer perceptron.

[0051] 3. For example Figure 13 As shown, the improved feature decoder includes deconvolution operations, a third batch normalization layer, and a third ReLU activation layer. Through deconvolution operations and the third batch normalization layer, the third ReLU activation layer introduces nonlinearity, reconstructing the high-dimensional encoded features output by the improved feature encoder into a multi-scale feature map. T Reconstructed feature maps of the same size are used. The parameter settings for the improved feature decoder are shown in Table 2. To better preserve the inductive bias of the model, and since both feature map extraction and encoding processes involve convolution, we choose to use deconvolution as the decoder for our proposed method. The inductive feature decoder includes deconvolution, batch normalization (BN), and the ReLU activation function.

[0052] Table 2. Parameter settings for improved feature encoder and decoder

[0053] 4. Figure 11 The feature classifier branch in the algorithm is used to handle global anomaly detection tasks. This branch improves the ability to distinguish between abnormal and normal features through auxiliary supervision. In application scenarios that require determining whether there are anomalies on the entire roller surface, the feature classifier maps the sequence features output by the encoder to the global category space, thereby outputting the probability distribution of anomalies and normal features.

[0054] The feature classifier includes a first multilayer perceptron (MLP), a softmax mapping, and a linear classifier. The MLP consists of two fully connected layers and a non-linear activation function, used to model the importance of the high-dimensional encoded features output by the improved feature encoder as sequence features and re-aggregate them, thus rewriting the high-dimensional encoded features into a sequence-form feature vector. The sequence features all include batch size. b Sequence length q and embedding dimension d The first multilayer perceptron (MLP) with shared parameters is used to analyze the feature vectors in the sequence. X Scoring is performed to characterize the importance of different feature vectors in global anomaly detection. s , Formula 16: The feature vector scoring results are normalized along the sequence dimension using Softmax mapping to obtain the feature vector weight distribution. Formula 17 calculates the weights of each feature vector and aggregates them with the original sequence-form feature vectors to obtain the global feature representation. Formula 18; then the global feature representation is fed into a linear classifier, and the abnormal and normal probability distributions of the entire roller surface are output through the Softmax function to provide discrimination constraints.

[0055] 5. The anomaly score smoother first estimates the variance of the reconstructed feature map based on the reconstruction error between the reconstructed feature map and the multi-scale feature map using a second multilayer perceptron. The output of the second multilayer perceptron... δ The system is used to correct reconstruction errors and convert these errors into anomaly score heatmaps. These heatmaps are then processed through convolutional layers with kernel sizes of 1, 4, and 16 to obtain anomaly score maps with different smoothness levels. The final anomaly score heatmap is obtained by summing the anomaly score maps pixel-by-pixel and averaging the results. S 0, by obtaining the final anomaly score heatmap S The contamination rate of the roller surface is identified by identifying pixels with a defect intensity greater than the design hue threshold.

[0056] The following methods are used to identify the contamination rate of the roller surface and to confirm the cleaning effect of the roller surface: Obtain the final anomaly score heatmap containing defect information. S 0. By comparing images after the previous cleaning or comparing images of the same roller ring before and after cleaning, it can be determined whether the roller is dirty or the cleaning effect is achieved. Since the heatmap uses three-channel RGB representation, the coordinates ( x , y The heatmap color at point ) is represented by the following formula: Formula 19, where R ( x , y () represents the red parameter in RGB. G ( x , y ) represents the green parameter in RGB. B ( x , y ) represents the blue parameter in RGB. p ( x , y The color represents the current coordinate system. Because the RGB color code table has semantic mixing issues and the generated results are not robust enough, it is necessary to convert RGB pixels to the HSV color code table to calculate the defect score. Calculate the normalized RGB values. Formula 20 retrieves the maximum and minimum values ​​of the RGB values. , Formula 21, , Formula 22, Formula 23, calculate brightness. Formula 24, calculate saturation Formula 25, Calculate hue Formula 26; where The remainder after dividing by 360 represents the area in the image defect heatmap. Red indicates the center of the defect, green indicates the edge, and blue indicates a non-defect area. Based on these characteristics, the degree of defect at that location is represented by the following formula: , Formula 27, Representing coordinates ( x , y Pixel hue at ) To calculate the defect intensity, it's crucial to understand that in real-world scenarios, the non-defective area on the roller surface is significantly larger than the defective area. This translates to a much larger blue-toned area than a red-toned area on the heatmap. Calculating the entire heatmap would drastically increase computation time and fail to effectively reflect these differences. Therefore, a hue threshold is designed for each pixel on the heatmap. If the hue of a pixel is below this threshold, it is excluded from subsequent calculations. In this embodiment... Set it to 0.2.

[0057] Assuming the current cleanup location is determined, for the first... j The defect intensities of all pixels with a defect intensity greater than 0.2 on the unfolded image of the roller after the second cleaning are summarized and represented as follows: Formula 28: The contamination rate of the roller surface can be calculated based on the ratio of the number of pixels with a defect intensity greater than 0.2 to the total number of pixels in the fully expanded image of the roller at the current position. Based on the current cleaning position... j Circle and the first j The cleaning efficiency of the grinding head on the roller body was calculated by comparing the defect strength of -1 ring: Formula 29, S j-1 It is the first j -1. The defect intensity of all pixels with a defect intensity greater than 0.2 on the rolled image after cleaning. In one embodiment, the pass standard for the current cleaning area can be set as the number of pixels with a defect intensity greater than 0.2 being less than 1% of the total number of pixels. Based on the cleaning efficiency per revolution and the change in the proportion of defect points, it can be determined whether the cleaning effect of the grinding head meets the requirements, and feedback can be provided to the operator through the human-machine interface (HMI).

[0058] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A cold rolling tension roll face cleaning system based on visual inspection, characterized by, include: The frame is fixedly installed on one side of the tension roller body and is arranged at intervals along the axial direction of the tension roller. An axial feed device is mounted across the frame. The axial feed device is equipped with a first movable unit, which moves linearly in a direction parallel to the axial direction of the tension roller. A radial feed device is mounted on the first movable unit, and a second movable unit is mounted on the radial feed device. The second movable unit moves linearly in a direction parallel to the radial direction of the tension roller. The grinding head, located on the second movable unit, is used to approach the surface of the tension roller and perform a circular cleaning of the tension roller surface. A camera module is mounted on the second active unit, with its viewfinder facing the tension roller body, for taking pictures of the roller body surface; The controller communicates with the camera module, axial feed device, radial feed device and grinding head. Through built-in algorithms, it obtains images of the roller surface to evaluate the degree of contamination, thereby driving the grinding head to the contaminated location for cleaning. On the other hand, after the grinding head grinds the contaminated location, the controller evaluates the cleaning effect of the grinding head again based on the images of the roller surface.

2. A cold rolling tension roll face cleaning system based on visual detection according to claim 1, characterized in that, The axial feed device includes a mounting base, a coupling, a servo motor, a slide table, and a synchronous belt. The mounting base is used to be fixedly connected to at least one pair of frames. Synchronous pulleys are rotatably mounted on both ends of the mounting base. The servo motor is connected to the synchronous pulleys via the coupling. The synchronous belt is sleeved on the synchronous pulleys and is rolledly connected to them. A slide table is fixedly mounted on the synchronous belt. The slide table serves as the first movable unit, driving the radial feed device, grinding head, and camera module to move linearly in a direction parallel to the axial direction of the tension roller.

3. A cold rolling tension roll face cleaning system based on visual detection according to claim 2, characterized in that, The radial feed device includes a cylinder and a piston rod. The cylinder is fixedly connected to the slide table and has a medium inlet and a medium outlet. One end of the piston rod extends into the cylinder and is slidably connected to the cylinder. The other end of the piston rod extends toward the surface of the tension roller. The end of the piston rod extending out of the cylinder serves as a second movable unit, which drives the grinding head and camera module to move linearly in a direction parallel to the radial direction of the tension roller under the drive of the pressure medium.

4. A cold rolling tension roll face cleaning system based on visual detection according to claim 2, characterized in that, The grinding head includes a grinding head body and an eccentric grinding disc. The grinding head body is fixedly connected to the second movable unit, and a detachable grinding head material is provided on the side of the eccentric grinding disc away from the grinding head body.

5. A method for cleaning the surface of a cold rolling tension roll based on visual inspection, characterized in that Includes the following steps: The system is configured as described in any one of claims 1-4, which is a vision-based cold rolling tension roll surface cleaning system; the controller includes a PLC and an image workstation; and an incremental encoder is configured on the rotating shaft of the tension roll. After the equipment is started, the camera module continuously acquires images of the roller surface and sends the images to the image workstation. The image workstation performs real-time stitching and processing of the images based on the improved feature recognition image stitching algorithm, and then identifies the contamination rate of the roller surface through the image defect detection algorithm architecture. The image defect detection algorithm architecture includes a multi-layer feature extractor, an improved feature encoder, an improved feature decoder, a feature classifier, and an anomaly score smoother. The original image is fed into the multi-layer feature extractor, which maps the original pixel space to a high-dimensional feature space, extracting semantic and detail information at different levels to obtain a multi-scale feature map. Subsequently, the multi-scale feature map is input into the improved feature encoder for unified encoding, modeling the local structure and global context relationship in the multi-scale feature map to obtain high-dimensional encoded features. After encoding, the improved feature encoder outputs two branches: one branch connects to the feature classifier, which is used to discriminate and learn the high-dimensional encoded features to distinguish the probability distribution of normal and abnormal samples at the overall semantic level; the other branch connects to the improved feature decoder, which reconstructs and models the high-dimensional encoded features to obtain a reconstructed feature map. The anomaly score smoother measures the difference between the reconstructed feature map and the multi-scale feature map, and outputs an anomaly score heatmap that measures the contamination rate of the roller surface. If the image workstation detects that the contamination rate on the roller surface does not exceed the preset threshold, the axial feed device will drive the radial feed device, grinding head, and camera module to move to the next detection position in a direction parallel to the axial direction of the tension roller by a preset step size. If the image workstation detects that the contamination rate on the roller surface exceeds the preset threshold, the PLC will pause the axial feed device, allowing the grinding head to press against the roller surface for cleaning. If the contamination rate still exceeds the preset threshold after cleaning, the PLC will activate the radial feed device, extending the second movable unit and increasing the grinding pressure of the grinding head. After grinding, the contamination rate on the roller surface will be reassessed. If the radial feed device reaches the maximum radial feed stroke or pressure limit, and the contamination rate on the roller surface still exceeds the preset threshold after grinding, the PLC will pause the cleaning operation and issue an alarm through the human-machine interface, prompting the replacement of the grinding head or manual intervention.

6. A method of cleaning the surface of a cold rolling tension roll based on visual detection according to claim 5, characterized in that, The image workstation performs real-time image stitching and processing based on an improved feature recognition image stitching algorithm. Specifically, it rotates the shaft of the tension roller and drives the roller body to rotate at least one revolution. The camera module acquires video of the roller body surface from a fixed position, obtaining a video slice of the entire roller body. The image workstation performs frame preprocessing on the video slice based on the displacement of the incremental encoder, finds feature points in the frame images, marks the feature points in the frame images with descriptors, sends the feature points to the matcher, matches the feature points of two adjacent frame images, generates a list of matching feature points, removes duplicate areas in adjacent frame images, and so on, until all frame images of the video slice are stitched together to obtain the original image of the roller body surface after unfolding.

7. A method for cleaning the surface of a cold rolling tension roll based on visual inspection according to claim 6, characterized in that, The multi-layer feature extractor includes four consecutive convolutional layers, upsampling, and a multi-scale fusion stage. The output specifications of the four consecutive convolutional layers are as follows: , , , , X and Y These represent the width and height of the original image, respectively. Adjacent consecutive convolutional layers are connected via residual blocks, and the outputs of the third and fourth consecutive convolutional layers are upsampled to the normal size. The output of the second consecutive convolutional layer is then fed into a multi-scale fusion stage to obtain a multi-scale feature map. T .

8. A method for cleaning the surface of a cold rolling tension roll based on visual inspection according to claim 7, characterized in that, The improved feature encoder includes a first convolutional layer, a first batch normalization layer, a first ReLU activation layer, a second convolutional layer, a second batch normalization layer, a first adder, a second ReLU activation layer, a first feature block, a second feature block, and a multi-scale feature map. T After processing by the first convolutional layer and the first batch of normalization layers, nonlinearity is introduced through the first ReLU activation layer, followed by processing by the second convolutional layer and the second batch of normalization layers to obtain the residual mapping. F ( T Then residual mapping F ( T ) and multi-scale feature maps T After element-wise addition by the first adder, a second ReLU activation layer is applied to obtain an intermediate feature map, which is then fed into the first feature block. The first feature block includes a first-layer normalization unit, a convolutional attention layer, and a second adder. The second feature block includes a second-layer normalization unit, a convolutional multilayer perceptron, and a third adder. The first feature block sequentially performs layer normalization and convolutional self-attention processing on the input intermediate feature map. The processing result is then added element-wise with the input of the first feature block by the second adder and fed into the second feature block. The second feature block performs layer normalization and convolutional multilayer perceptron processing on the input and feeds the processing result into the third adder for element-wise addition. Finally, a high-dimensional encoded feature that enhances the expression of the contamination pattern on the roller surface is output.

9. A method for cleaning the surface of a cold rolling tension roll based on visual inspection according to claim 8, characterized in that, The feature classifier includes a first multilayer perceptron (MLP), a softmax mapping, and a linear classifier. The first MLP consists of two fully connected layers and a non-linear activation function, used to model the importance of the high-dimensional encoded features output by the improved feature encoder according to sequence features and re-aggregate them. The high-dimensional encoded features are rewritten into sequence-form feature vectors. The first MLP with shared parameters scores the feature vectors in the sequence to characterize the importance of different feature vectors in global anomaly detection. s The feature vector scoring results are normalized by Softmax mapping along the sequence dimension to obtain the weight distribution of the feature vectors. The weights of each feature vector are then weighted and converged with the original sequence-form feature vectors to obtain the global feature representation. The global feature representation is then fed into a linear classifier, and the abnormal and normal probability distributions of the entire roller surface are output through the Softmax function to provide discrimination constraints.

10. A method for cleaning the surface of a cold rolling tension roll based on visual inspection according to claim 9, characterized in that, The anomaly fraction smoother first estimates the variance of the reconstructed feature map using a second multilayer perceptron based on the reconstruction error between the reconstructed feature map and the multi-scale feature map. The output of the second multilayer perceptron... δ This is used to correct reconstruction errors and convert the reconstruction errors into anomaly score heatmaps, which are defined as follows: , It is a reconstruction error. It is the L2 norm. This represents a division operation on the number of pixels; the resulting anomaly score heatmap... S Anomaly score maps with different smoothness levels are obtained by processing the data through convolutional layers with kernel sizes of 1, 4, and 16. The final anomaly score heatmap is obtained by summing the anomaly score maps with different smoothness levels pixel by pixel and taking the average. S 0. By obtaining the pixel points whose defect intensity in the final abnormal score heatmap is greater than the design hue threshold, the contamination rate of the roller surface is identified.

Images