Bearing roller surface damage detection method based on damage region fusion transformation

Abstract

This invention discloses a method for detecting bearing roller surface damage based on damage region fusion transformation. The method comprises the following steps: constructing a roller surface damage feature sample detection device, collecting side images of the roller working surface and using them as the research object; applying digital image processing methods such as gradient extraction and contour detection, and cropping the core detection area of ​​the sample based on the distribution characteristics of the roller outer fillet in the image; screening key damage areas of appropriate size and aspect ratio from negative samples in the dataset based on clustering pre-selection boxes and damage assessment functions, performing different affine transformations on the selected areas, and fusing them with randomly sampled positive samples to generate new enhanced sample data; inputting the data-enhanced images into a ResNeXt101 network for detection and classification, and verifying the effectiveness of the proposed method from the perspectives of accuracy and precision. This invention effectively improves detection accuracy and precision, providing technical support for the actual production process of roller surface damage detection.

Description

Technical Field

[0001] The present invention belongs to the technical field of bearing roller surface damage detection, and in particular relates to a bearing roller surface damage detection method based on damage area fusion transformation. Background Art

[0002] Rolling bearings are widely used in industry and are crucial transmission components. Their dynamic performance significantly impacts mechanical systems. As the most critical component in a bearing, the roller's surface quality determines its service life and kinematic performance. All rollers undergo rigorous surface defect inspection before shipment to improve the roller's pass rate. Currently, the main technologies used to detect roller surface defects include manual inspection, eddy current testing, ultrasonic testing, and machine vision.

[0003] Manual inspection methods rely on a large amount of manpower, resulting in a waste of human resources and increasing the overall inspection cost. In addition, manual inspection is prone to false detection and missed detection, resulting in low overall inspection efficiency.

[0004] Eddy current testing is a nondestructive testing method that evaluates the performance of a workpiece by measuring changes in eddy currents induced in it. However, this method is not ideal for detecting some types of defects. Furthermore, because it operates in a magnetic field, it is susceptible to external interference. Consequently, eddy current testing is not widely used.

[0005] The principle of ultrasonic testing is that ultrasonic waves form different waveforms when propagating in different media. When ultrasonic waves strike the surface of a damaged bearing roller, the presence and type of damage can be determined by detecting the specific waveform returned. A limitation of ultrasonic testing is that it requires high-quality testing equipment.

[0006] In machine vision inspection, neural network models are typically trained and then applied to classify and identify surface damage during inspection. This approach often offers high accuracy and robustness. However, a limitation of machine vision inspection methods is that model training often requires a large number of samples, and an imbalance in sample types can lead to reduced detection effectiveness. Summary of the Invention

[0007] The purpose of this invention is to propose a bearing roller surface damage detection method based on damage area fusion transformation. As a machine vision detection method, it can enrich different types of samples and damage characteristics by enhancing the data of existing training samples, thereby improving the problems caused by sample imbalance and making the model more adaptable to new damage samples.

[0008] To achieve the above object, the technical solution of the present invention is as follows:

[0009] A bearing roller surface damage detection method based on damage area fusion transformation includes the following steps:

[0010] Step 1: Build a device to shoot the roller's side, end, and chamfer images, including a chassis, side camera, end camera, and chamfer camera. Use the device to shoot various parts of the roller to obtain side, end, and chamfer datasets. The process is as follows:

[0011] Step 1.1: Build the roller surface damage detection device, including a chassis, a turning mechanism, a conveyor belt, a side camera, an end camera, and a chamfer camera. The side camera uses a 360° external viewing optical lens, capable of capturing 360° images of the roller end face and a circular lateral surface. The end and chamfer cameras use standard area array cameras.

[0012] Step 1.2: Use the constructed device to photograph various parts of the roller part. After loading, the roller is driven by the chassis to the side, end, and chamfer photography stations in sequence. After a one-second pause, the corresponding camera takes a picture, and the chassis then drives it to the next station. After completing three shots, the roller is flipped by the flipping device and then driven by the chassis to complete the side, end, and chamfer photography at the other end before unloading at the unloading station.

[0013] Step 2: Based on step 1, since the side surface of the roller is the main working surface when it is working, the side surface of the roller is taken as the research object. The obtained side surface dataset is annotated and divided into positive samples and negative samples. The negative samples are classified according to the damage type and the damaged area of ​​the negative sample image is annotated. The process is as follows:

[0014] Step 2.1: Label the side photos of the roller. Label the damaged ones as negative samples and the undamaged ones as positive samples.

[0015] Step 2.2: According to the damage types in the negative samples, the negative samples are divided into several categories, such as chip damage, cut damage, scratch damage, corrosion, burn damage, dent damage, turning damage, convex spots, and wear near the end surface.

[0016] Step 2.3: For all types of negative sample images, use fine polygons to mark the damaged area.

[0017] Step 3: Preprocess the obtained side dataset by using digital image processing methods to remove irrelevant pixels around the image and crop the key central area image. The process is as follows:

[0018] Step 3.1: Use the Sobel operator to calculate the image gradient.

[0019] Step 3.2: Use Gaussian filtering for noise reduction. For a Gaussian kernel with a side length of 2k+1, the value of the element in the i-th row and j-th column of the Gaussian kernel is as shown in formula (1):

[0020]

[0021] Where σ represents the standard deviation of the Gaussian distribution. The Gaussian kernel is selected to be 9×9 and σ is 1.7.

[0022] Step 3.3: Select a threshold of 90 and perform binarization on the image.

[0023] Step 3.4: Use morphological methods to remove spots and peripheral residual noise formed in the damaged parts of the image. The morphological methods include corrosion and expansion processing.

[0024] Step 3.5: Use the Suzuki contour tracking algorithm to extract the inner and outer contours.

[0025] Step 3.6: Based on the extracted lower outer corner contour, use the area corresponding to its small circumscribed square as the retained area to crop the image, and finally retain a square area of ​​900×900 pixels.

[0026] Step 4: For each type of damage in the negative sample, use the k-means++ clustering algorithm to set three aspect ratios of pre-selected boxes to complete the unified extraction of the damage area in the negative sample, and design a damage assessment function based on the intersection of the pre-selected box and the damage area and the aspect ratio. The extracted damage areas are screened and evaluated, sorted from small to large according to the evaluation value, and the top N areas are selected as key damage areas for subsequent damage transformation operations. The process is as follows:

[0027] Step 4.1: Select the k-means++ clustering algorithm classification number k, which is 3 based on the three relationships between the length of the annotated damage area being greater than the width, equal to the width, and less than the width. Because the annotated area may have an excessively large aspect ratio, to prevent the noise generated by this situation from affecting the initial center selection of the clustering algorithm, the damage area with an aspect ratio greater than 5 is centered and segmented.

[0028] Step 4.2: Initialize the cluster center. The process is as follows: First, select a damaged area with an aspect ratio close to 1 as the first cluster center, whose aspect ratio is τ1, and then calculate the distance function D(x) = τ between the aspect ratio of each sample and τ1. x / τ1 to obtain the probability P(x) of it being selected as the next center point, which is calculated by formula (2):

[0029]

[0030] where τ xis the aspect ratio of the current sample, χ is the set of damaged areas. The area with the largest P(x) is selected as the next center point. Finally, the three aspect ratio cluster centers are obtained, namely τ1, τ2, and τ3.

[0031] Step 4.3: Set the length of the preselected box to its length in the critical case of the inner side of the outer circle and the outer side of the inner circle. Let the length of the preselected box be τh, the height be h, the distance from the center of the image be r, the radius of the upper outer corner be r1, the radius of the lower outer corner be r2, and the average of the two circle radii be r3. When r∈(r1,r3), select the inner critical box. The height h of the preselected box can be obtained by formula (3):

[0032]

[0033] θ is the angle between the line connecting the center of the upper outer corner to the critical contact point of the preselection box and the x-axis. When r∈[r3,r2), the outer critical box is selected, and the preselection box height h can be obtained by formula (4):

[0034]

[0035] γ is the angle between the line from the center of the preselected box to the center of the upper outer corner and the x-axis.

[0036] Step 4.4: Design a damage assessment function based on the difference between the pre-selected box and the damaged area. Select the key damaged area similar to the pre-selected box and define the damage assessment function as shown in formula (5):

[0037] P eval =-log(kv)+1 (5)

[0038] k is the balance factor, and v is the factor that controls the aspect ratio, which is defined as follows:

[0039]

[0040] w A 、h A is the width and height of the damaged area, i is the cluster center number, i∈{1, 2, 3}, and k is defined as follows:

[0041]

[0042] Where IoU is the intersection over union ratio of two rectangular regions.

[0043] For the damaged areas selected by using the unified aspect ratio specification, an evaluation function is used to evaluate and screen the damaged areas to make the subsequent transformation effect more reasonable. For every 3 negative samples, a total of N damaged areas are used as the damage source of a new negative sample. All damaged areas are calculated and sorted using the evaluation function, and the top N / 3 damaged areas with the largest results are selected for subsequent transformation and collage processing.

[0044] Step 5: Based on step 4, perform collage and transformation operations on the selected key damage to generate some new samples to achieve data enhancement.

[0045] Step 5.1: For damage with an aspect ratio less than 2 and a long side not exceeding 85 pixels, no transformation is performed and it is directly retained, preserving some of the original features.

[0046] Step 5.2: For damage that does not meet the conditions in Step 5.1, random shearing is performed in different directions based on its aspect ratio. For damage areas that are taller than wide, shearing is performed in the x-direction; otherwise, shearing is performed in the y-direction. The shearing angle α is a random angle between 5° and 25°.

[0047] Step 5.3: For damage that does not meet the conditions in step 5.1, randomly rotate it. The rotation angle β of each damage area is between 0-360°.

[0048] Step 5.4: For damage that does not meet the conditions in step 5.1, randomly scale it. The scaling factor S in the x and y directions is x ,S y A random value between 0.7 and 0.9.

[0049] Step 5.5: Determine the distance r between the center of the damaged area to be tiled and the center of the new image. Add a uniformly distributed random perturbation Δr~U[-5,5] to the distance r0 from the damaged area to the center of the circularity, i.e., r = r0 + Δr.

[0050] Step 5.6: Randomly select the rotation angle of the damaged area in the new image. Randomly select an angle γ between 0 and 360 degrees as the rotation angle. In the tiling step, γ is the angle between the line connecting the damage center and the image center and the positive x-axis.

[0051] Step 5.7: Determine whether the damage exceeds the upper outer corner boundary. Damage that exceeds the upper outer corner boundary is discarded. If any damage exceeds the image boundary, it is trimmed to within the boundary. If the collage damage overlaps, the latter will cover the former.

[0052] Step 6: Based on step 4, the data-augmented roller image is input into the ResNeXt101 network for training.

[0053] The process is as follows:

[0054] The roller image dataset that has undergone image processing and data augmentation is fed into the ResNeXt101_32×8d network for training. The loss function is modified to meet production needs. The original cross entropy loss function is adjusted, and a balance factor α is added to control the loss of positive and negative samples and the weight factor (1-p t ) to control the loss of high confidence samples, and the final loss function is as follows:

[0055]

[0056] Where p is the predicted probability, and y indicates whether the final result is a positive sample or a negative sample.

[0057] Step 7: Input the image to be detected containing the target points on the component surface into the trained damage recognition model to obtain the recognition result.

[0058] The beneficial effects of the present invention are as follows:

[0059] 1) The present invention uses a digital image method to preprocess the training samples, and then adopts a pre-selection box based on a clustering algorithm and a damage assessment function to determine the key damage areas in the negative samples for feature fusion data enhancement, trains the detection model, and finally obtains the roller surface damage detection classification results through testing.

[0060] 2) This data augmentation method effectively enriches the damage features of the roller surface, alleviating the problem of poor training results caused by the imbalance of positive and negative samples. In addition, this data augmentation method significantly improves the accuracy of roller detection, with the final accuracy reaching 95.2%.

[0061] 3) This paper experimentally determined the coefficients of the loss function when the precision rate was high, thereby adjusting the loss function. Furthermore, by raising the threshold for determining a sample as a positive, the precision rate was significantly improved, reaching 99.8%, while maintaining an acceptable detection accuracy. This result can provide a reference for actual production processes. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] Figure 1 Schematic diagram of a roller surface damage detection device according to the present invention;

[0063] Figure 2 This is a schematic diagram of the actual comparison between the roller side surface samples collected in the present invention and the actual roller;

[0064] Figure 3 This is a schematic diagram of the roller image after gradient extraction and Gaussian filtering of the present invention;

[0065] Figure 4 This is a schematic diagram of the roller image after binarization of the present invention;

[0066] Figure 5 This is a schematic diagram of the roller image after morphological analysis of the present invention;

[0067] Figure 6 This is a schematic diagram of the roller image after target cropping according to the present invention;

[0068] Figure 7 This is a schematic diagram of the roller overall detection process of the present invention;

[0069] Figure 8 A schematic diagram of a roller defect image damage collage of the present invention;

[0070] Figure 9 Schematic diagram of a negative sample of the roller after damage transformation according to the present invention;

[0071] In the figure: 1. Rotating chassis; 2. Side camera; 3. End camera; 4. Chamfer camera; 5. Flipping device; 6. Bearing roller to be inspected. DETAILED DESCRIPTION

[0072] In order to make the above-mentioned objects and advantages of the present invention more obvious and easy to understand, the specific embodiments of the present invention are described in detail below with reference to the accompanying drawings.

[0073] Example 1

[0074] This embodiment provides a bearing roller surface damage detection method based on damage area fusion transformation, including:

[0075] S1: See attached Figure 1 The bearing roller damage detection device of the present invention mainly includes the following components: a rotating chassis 1, a side camera 2, an end camera 3, a chamfer camera 4, and a flipping device 5. The bearing roller 6 to be inspected is transported via the rotating chassis 1 to a fixed imaging station, where the three cameras obtain image information of the corresponding surface.

[0076] Since the side of the roller is its main working surface, the experiment uses side images as training samples. The roller side camera is a 360° external viewing optical lens that can capture the roller end face and the 360° side image in a circular shape. The actual comparison diagram of the collected roller side samples and the actual roller is shown in the attached figure. Figure 2 shown.

[0077] S2: Taking the side of the main working surface of the roller as the research object, the obtained side dataset is annotated and divided into positive samples and negative samples. The negative samples are classified according to the damage type, and the damaged area of ​​the negative sample image is annotated. Specifically, it includes: annotating the side photos of the roller taken. The damaged ones are marked as negative samples, and the undamaged ones are marked as positive samples. According to the damage type in the negative samples, the negative samples are divided into several categories such as flake damage, cut damage, scratches, corrosion, burns, dents, turning marks, bumps, and wear near the end face. For all types of negative sample images, fine polygons are used to select and annotate the damaged area.

[0078] S3: Preprocess the obtained side dataset by using digital image processing methods to remove irrelevant pixels around the image and crop the key central area image. The process is as follows:

[0079] This invention uses the Sobel operator to extract image gradients. After gradient extraction, the edges and contours of the roller image have larger pixel values, which are more easily classified as white pixels during binarization, facilitating contour extraction. Due to the effects of scratches on the background board around the outer contour and the halo caused by light refraction, the original image contains a certain amount of noise after gradient extraction. Therefore, Gaussian filtering is used to reduce the noise of the image. For a Gaussian kernel with a side length of 2k+1, the values ​​of each element in the Gaussian kernel satisfy the following relationship:

[0080]

[0081] Where σ represents the standard deviation of the Gaussian distribution. The present invention selects a Gaussian kernel of 9×9 and σ of 1.7. The image after gradient extraction and Gaussian filtering is as follows Figure 3 shown.

[0082] To extract the outer corner contour of the roller bottom side, the image needs to be binarized. After binarization, the image contour is clearer and the outer ring-shaped halo is removed. The binarization result is as follows: Figure 4 As shown. Morphological methods are then used to remove spots and peripheral residual noise formed in the damaged area of ​​the image, specifically including corrosion and expansion processing. After the image has been corroded several times, the peripheral noise is completely removed, and only a small amount of damage remains after processing, which is conducive to the detection of the lower outer corner contour. The expansion processing after corrosion eliminates the slight discontinuity of the contour caused by corrosion, making the lower outer corner contour more complete and clear. The contour features finally extracted are shown as follows Figure 5 shown.

[0083] The contour extraction algorithm uses the Suzuki contour tracking algorithm. Only the required lower outer corner contour is retained among all contours, and finally the image is cropped according to the minimum circumscribed square of the extracted lower outer corner contour to realize the target cropping process. The area after target cropping eliminates the interference of irrelevant pixels, and its length and width are only about 900×900 pixels, which is greatly reduced compared to the original image's 1920×1200 pixel scale. This is not only beneficial to the network's requirements for input scale, but also ensures that the damaged pixels are not too small when the image is scaled. Finally, a black border is added before inputting the network. The target cropping effect is as follows Figure 6 .

[0084] S4: Select the number of clusters k for the k-means++ clustering algorithm, which is 3 based on the three relationships between the length of the labeled damage area being greater than the width, equal to the width, and less than the width. Because the labeled area may have an excessively large aspect ratio, to prevent the noise generated by this situation from affecting the initial center selection of the clustering algorithm, the damaged area with an aspect ratio greater than 5 is centered and segmented.

[0085] Initialize the cluster center. The process is as follows: first select a damaged area with an aspect ratio close to 1 as the first cluster center, whose aspect ratio is τ1, and then calculate the distance function D(x) = τ between the aspect ratio of each sample and τ1 x / τ1 to obtain the probability P(x) of it being selected as the next center point, P(x) is as follows:

[0086]

[0087] where τ x is the aspect ratio of the current sample, χ is the set of damaged areas. The area with the largest P(x) is selected as the next center point. Finally, the three aspect ratio cluster centers are obtained, namely τ1, τ2, and τ3.

[0088] The side length of the preselection box is set to its length at the critical point between the inside and outside of the circle. Let the side length of the preselection box be τh, the height be h, the distance from the center of the image be r, the radius of the upper outer corner be r1, the radius of the lower outer corner be r2, and the average of the two circle radii be r3.

[0089] When r∈(r1,r3), the inner critical box is selected and the preselected box height h can be obtained as follows:

[0090]

[0091] θ is the angle between the line connecting the center of the upper outer corner to the critical contact point of the preselected box and the x-axis. When r∈[r3,r2), the outer critical box is selected, and the preselected box height h can be obtained by the following formula:

[0092]

[0093] γ is the angle between the line from the center of the preselected box to the center of the upper outer corner and the x-axis.

[0094] According to the difference between the pre-selected box and the damaged area, a damage assessment function is designed to select the key damaged area similar to the pre-selected box. The damage assessment function is defined as follows:

[0095] P eval =-log(kv)+1

[0096] k is the balance factor, and v is the factor that controls the aspect ratio, which are defined as follows:

[0097]

[0098] w A 、h A is the width and height of the damaged area, i is the cluster center number, i∈{1, 2, 3}, and k is defined as follows:

[0099]

[0100] Where IoU is the intersection over union ratio of two rectangular regions.

[0101] The larger the damage assessment function value of a damaged area, the more likely it is to be selected as a critical damage. For every three negative samples, a total of N damaged areas are used as the damage source of a new negative sample. All damaged areas are calculated and sorted using the assessment function, and the top N / 3 damaged areas with the largest results are selected for subsequent transformation and tiling.

[0102] S5: Transform and tile the selected key damages. Damage transformation includes random shearing, random rotation, and random scaling. Damage tile includes the following steps:

[0103] For damage with an aspect ratio less than 2 and a long side not exceeding 85 pixels, no transformation is performed and it is directly retained, preserving some of the original features.

[0104] If the above conditions are not met, random shearing in different directions is performed according to the aspect ratio. For damaged areas with a height greater than the width, shearing in the x direction is used, and vice versa, shearing in the y direction is used. The shearing angle α is a random angle between 5° and 25°. In addition, it is randomly rotated. The rotation angle β of each damaged area is between 0-360°. In addition, it is randomly scaled. The scaling factor S in the x and y directions is x ,S y A random value between 0.7 and 0.9.

[0105] When making a collage, Figure 8As shown in Figure 1, the distance r between the center of the damaged area to be tiled and the center of the new image is determined. r is the distance r0 from the damaged area to the center of the circularity, and a random perturbation Δr~U[-5,5] that satisfies a uniform distribution is added, that is, r=r0+Δr.

[0106] Randomly select the rotation angle of the damaged area in the new image. Randomly select an angle γ between 0-360° as the rotation angle. In the tiling step, γ is the angle between the line connecting the damage center and the image center and the positive x-axis.

[0107] Determine whether the damage exceeds the upper outer corner boundary. Damage that exceeds the upper outer corner boundary is discarded. If any damage exceeds the image boundary, it is trimmed to within the boundary. If the collage damage is exactly overlapping, the latter will cover the former.

[0108] The sample after damage transformation is as follows Figure 9 shown.

[0109] S6: Adjust the classification network and the loss function used. The classification network uses the ResNeXt101_32×8d network. The present invention sets the input image length and width to 512 pixels and changes the number of fully connected outputs at the end of the network to 2 to perform classification operations on high-level semantic features.

[0110] The original intention of the model design is to have a high precision rate, while sacrificing a certain degree of recall rate and accuracy. The cross entropy loss function used in this invention is as follows:

[0111]

[0112] Adjusting α adjusts the loss ratio of positive and negative samples. When α is between 0 and 1, the loss of positive samples is suppressed and the loss of negative samples is amplified. Therefore, the model tends to judge positive samples as negative samples and strictly judge negative samples as negative samples. Finally, α is set to 0.5.

[0113] The roller image enhanced by damage transformation data is input into the network for training, and the classification network obtained by training can be used for roller damage detection.

[0114] Example 2

[0115] This embodiment provides a bearing roller surface damage detection method based on damage area fusion transformation with high precision and high detection speed.

[0116] This paper experiments on the model's online computational time. The time required for online detection of a single image includes the image processing time for object cropping and adding a black border, as well as the image inference time within the model. The time consumption for each component is shown in Table 1. As can be seen, the average processing time for a single image is 217.1ms, which meets the real-time requirements of the online detection process.

[0117] Table 1 Single image time consumption

[0118]

[0119] The present invention performs a precision test on the model. In the output of the network, the precision index can be improved by increasing the threshold for classifying as qualified products. The network output obtained by the softmax function is as follows:

[0120]

[0121] Because the quality of rollers is a binary classification, the default threshold t for classifying positive samples is 0.5. Increasing the threshold can improve the precision. Table 2 shows the corresponding accuracy and precision rates for varying the threshold t. As can be seen, as the threshold increases, the accuracy decreases, but the precision rate continues to improve. When the threshold reaches 0.9, the precision rate approaches 1, demonstrating that increasing the classification standard can effectively improve the precision rate, approaching production standards.

[0122] Table 2 The impact of threshold on accuracy and precision

[0123]

[0124] The above embodiments are only preferred embodiments of the present invention and are not limitations on the technical solutions of the present invention. Any technical solution that can be implemented on the basis of the above embodiments without creative work should be deemed to fall within the scope of protection of the patent of the present invention.

Claims

1. A bearing roller surface damage detection method based on damage area fusion transformation, characterized in that: The following steps are involved: Step 1: Build a roller surface damage detection device and use it to photograph various parts of the roller parts to obtain side, end, and chamfer data sets; Step 2: Based on step 1, take the roller side as the research object, annotate the obtained side dataset, divide it into positive samples and negative samples, classify the negative samples by damage type, and annotate the damaged areas of the negative sample images; Step 3: Preprocess the obtained side dataset by using digital image processing methods to remove irrelevant pixels around the image and crop the key central area image; Step 4: For each type of damage in the negative sample, use the k-means++ clustering algorithm to set pre-selected boxes with three aspect ratios to complete the unified extraction of the damage area in the negative sample. Then, design a damage assessment function based on the intersection-over-union ratio and aspect ratio of the pre-selected boxes and the damage area. Screen and evaluate the extracted damage areas, sort them from small to large according to the evaluation value, and select the top N areas as key damage areas for subsequent damage transformation operations. The damage assessment function is defined as formula (5): P eval =-log(kv)+1 (5) k is the balance factor, and v is the factor that controls the aspect ratio, which is defined as follows: w A 、h A is the width and height of the damaged area, i is the cluster center number, i∈{1, 2, 3}, and k is defined as follows: Where IoU is the intersection-over-union ratio of two rectangular regions; Step 5: Based on step 4, perform tiling and transformation operations on the selected key damage areas to generate some new samples and achieve data enhancement; Step 6: Based on step 4, the data-enhanced roller image is input into the ResNeXt101 network for training; Step 7: Input the image to be detected containing the target points on the component surface into the trained damage recognition model to obtain the recognition result.

2. The method for detecting bearing roller surface damage based on damage area fusion transformation according to claim 1, characterized in that: The specific steps of step 1 are as follows: Step 1.1: Build a roller surface damage detection device, including a chassis, a turning device, a conveyor belt, a side camera, an end camera, and a chamfer camera. The side camera uses a 360° external viewing optical lens to capture the roller end face and a 360° annular side image. The end and chamfer cameras use ordinary area array cameras. Step 1.2: Use the constructed device to shoot various parts of the roller parts. After loading, the roller is driven by the chassis to the side, end, and chamfer shooting stations in turn and pauses. After being photographed by the corresponding camera, it is driven by the chassis to the next station. After completing three shots, the roller is flipped through the flipping device, and then driven by the chassis to complete the side, end, and chamfer shooting of the other end, and unloaded at the unloading point.

3. The method for detecting bearing roller surface damage based on damage area fusion transformation according to claim 1, characterized in that: The specific steps of step 2 are as follows: Step 2.1: Label the side photos of the rollers, marking damaged ones as negative samples and undamaged ones as positive samples; Step 2.2: Based on the damage types in the negative samples, the negative samples are classified into chip damage, cut damage, scratch damage, corrosion damage, burn damage, dent damage, turning damage, convex damage, and wear damage near the end surface; Step 2.3: For all types of negative sample images, use fine polygons to mark the damaged area.

4. The method for detecting bearing roller surface damage based on damage area fusion transformation according to claim 1, characterized in that: The specific steps of step 3 are as follows: Step 3.1: Use the Sobel operator to calculate the image gradient; Step 3.2: Use Gaussian filtering to reduce noise. For a Gaussian kernel with a side length of 2k+1, the value of the element in the i-th row and j-th column of the Gaussian kernel is as shown in formula (1): Where σ represents the standard deviation of the Gaussian distribution; Step 3.3: Select a threshold and perform binarization on the image; Step 3.4: Using morphological methods to remove spots and peripheral residual noise formed in the damaged parts of the image, the morphological methods include erosion and dilation processing; Step 3.5: Use Suzuki contour tracking algorithm to extract inner and outer contours; Step 3.6: Based on the extracted lower outer corner contour, use the area corresponding to its small circumscribed square as the retained area, crop the image, and finally retain the square area.

5. The method for detecting bearing roller surface damage based on damage area fusion transformation according to claim 1, characterized in that: The specific steps of step 4 are as follows: Step 4.1: Select the number of clusters k for the k-means++ clustering algorithm. Based on the three relationships of the annotated damage area (length greater than width, equal to width, and less than width), the annotated area may have an excessively large aspect ratio. To prevent the noise generated by this situation from affecting the initial center selection of the clustering algorithm, the damage area with an aspect ratio greater than 5 is split at the center. Step 4.2: Initialize the cluster center. The process is as follows: First, select a damaged area with an aspect ratio closest to 1 as the first cluster center. Its aspect ratio is τ1. Then calculate the distance function D(x) = τ between the aspect ratio of each sample and τ1. x / τ1 to obtain the probability P(x) of it being selected as the next center point, which is calculated by formula (2): where τ x is the aspect ratio of the current sample, χ is the set of damaged areas. The area with the largest P(x) is selected as the next center point, and the final clustering results in three aspect ratio cluster centers, τ1, τ2, and τ3; Step 4.3: The length of the preselected box is set to its length in the critical case of the inner side of the outer circle and the outer side of the inner circle. Let the length of the preselected box be τh, the height be h, the distance from the center of the image be r, the radius of the upper outer corner be r1, the radius of the lower outer corner be r2, and the average of the two circle radii be r3. When r∈(r1,r3), the inner critical box is selected. The height h of the preselected box can be obtained by formula (3): θ is the angle between the line connecting the center of the upper outer corner to the critical contact point of the preselected box and the x-axis. When r∈[r3,r2), the outer critical box is selected, and the preselected box height h can be obtained by formula (4): γ is the angle between the line from the center of the preselected box to the center of the upper outer corner and the x-axis; Step 4.4: Design a damage assessment function based on the difference between the pre-selected box and the damaged area to select the key damaged area similar to the pre-selected box; for the damaged area selected using a unified aspect ratio specification, use an evaluation function to evaluate and screen the damaged area to make the subsequent transformation effect more reasonable. For every 3 negative samples, a total of N damaged areas are used as the damage source of a new negative sample. All damaged areas are calculated and sorted using the evaluation function, and the top N / 3 damaged areas with the largest results are selected for subsequent transformation and collage processing.

6. The method for detecting bearing roller surface damage based on damage area fusion transformation according to claim 1, characterized in that: The specific steps of step 5 are as follows: Step 5.1: For damage with an aspect ratio less than 2 and a long side not exceeding 85 pixels, no transformation is performed and it is retained directly, preserving some of the original features; Step 5.2: For damage that does not meet the conditions in step 5.1, perform shearing in different directions according to its aspect ratio. For damage areas with height greater than width, shearing in the x-direction is used, otherwise, shearing in the y-direction is used. The shearing angle α is a random angle between 5° and 25°. Step 5.3: For damage that does not meet the conditions in step 5.1, randomly rotate it, and the rotation angle β of each damage area is between 0-360°; Step 5.4: For damage that does not meet the conditions in step 5.1, randomly scale it; the scaling factor S in the x and y directions x ,S y A random value between 0.7 and 0.9; Step 5.5: Determine the distance r between the center of the damaged area to be tiled and the center of the new image. Add a random perturbation Δr~U[-5,5] that satisfies a uniform distribution to the distance r0 from the damaged area to the center of the circularity, i.e., r=r0+Δr; Step 5.6: Randomly select the rotation angle of the damaged area in the new image, and randomly select an angle γ between 0-360° as the rotation angle. In the collage step, γ is the angle between the line connecting the damage center and the image center and the positive x-axis; Step 5.7: Determine whether the damage exceeds the upper outer corner boundary. Discard the damage that exceeds the upper outer corner boundary. If the damage exceeds the image boundary, trim it to within the boundary. If the collage damage happens to overlap, the latter will cover the former.

7. The method for detecting bearing roller surface damage based on damage region fusion transformation according to claim 1, characterized in that: The specific steps of step 6 are as follows: The roller image dataset that has undergone image processing and data augmentation is input into the ResNeXt101_32×8d network for training. The loss function is modified to meet production needs: the original cross entropy loss function is adjusted, and a balance factor α is added to control the loss of positive and negative samples and a weight factor is added to control the loss of high-confidence samples. The final loss function is shown in Equation (8): Where p is the predicted probability, and y indicates whether the final result is a positive sample or a negative sample.

Images