Arc geometry and surface defect parallel detection method for brake shoe

Abstract

The invention discloses a brake shoe arc geometry and surface defect parallel detection method. The method comprises the following steps: S1, obtaining a backlight silhouette image and a low-angle oblique light grayscale image in a time-sharing manner; s2, processing the silhouette image to form a closed two-dimensional contour point column, and generating a laser-free three-dimensional point cloud; s3, performing total least square cylindrical axis fitting on the three-dimensional point cloud; s4, the oblique light grayscale images are encoded and spliced into an expanded image according to angles, the expanded image is input into an encoder-decoder network containing a multi-scale cavity convolution bridge, confidence maps of four types of defects including cracks, sand holes, chipping and ablation are output in parallel, and the confidence maps are mapped back to original cambered surface coordinates according to the angles; and S5, aligning the circular run-out curve with the four types of defect confidence maps according to an angle pixel level, and executing defect-run-out mutual feedback logic. According to the brake shoe arc geometry and surface defect parallel detection method provided by the invention, a laser is not needed, the full-cycle run-out and the surface defect can be simultaneously detected only through one-time rotation, and the cost is low.

Description

TECHNICAL FIELD

[0001] The present application relates to a kind of brake shoe circular arc geometry and surface defect parallel detection method. BACKGROUND

[0002] Brake shoe assembly outer circle runout is the key index that influences the uniformity of shoe drum clearance, and the prior art adopts contact type dial gauge or laser point cloud measurement, the former is single-point detection, low efficiency and cannot cover the whole week;The latter needs additional laser sensor, high cost, and the signal attenuation of black friction plate is large. At the same time, the crack, sand eye and other defects of circular arc surface still rely on artificial visual inspection, without digital record, prone to miss detection. Therefore, a kind of image processing method without laser, low cost, can complete geometry runout and surface defect detection on line in parallel is needed. SUMMARY

[0003] The present application provides a kind of brake shoe circular arc geometry and surface defect parallel detection method to solve the technical problems that the traditional early warning means mentioned above is not accurate enough, specifically using the following technical scheme: A kind of brake shoe circular arc geometry and surface defect parallel detection method, characterized by, in turn includes: S1: equal-angle hard trigger time-sharing imaging step, for driving brake shoe assembly to rotate uniformly around circular arc center, and with encoder hard trigger industrial camera, time-sharing acquisition backlight silhouette image and low-angle oblique light gray image, so that each frame image is strictly corresponding with rotation angle; S2: sub-pixel profile laser-free dimensionality step, for the silhouette image is carried out adaptive threshold segmentation, Zernike matrix sub-pixel edge positioning, forms closed two-dimensional profile point column, and the point column is input one-dimensional convolution network offline training and online freezing, utilizes cylindrical prior loss output height offset, generates laser-free three-dimensional point cloud; S3: cylindrical prior geometry runout step, for the three-dimensional point cloud is carried out least square cylindrical axis fitting, obtains full circle runout curve and maximum runout angle; S4: expansion map multi-label defect mapping step, for the oblique light gray image is encoded and spliced as expansion map according to angle, input encoder-decoder network containing multi-scale hollow convolution bridge, and output crack, sand eye, block, ablation four kinds of defect confidence map in parallel, and map back to original arc surface coordinate according to angle; S5: defect-runout mutual feedback determination step, for the circle runout curve and four kinds of defect confidence map are aligned according to angle pixel level, executes defect-runout mutual feedback logic.

[0004] Further, in the isometric hard-triggered time-sharing imaging step of S1, the rotating mechanism adopts an encoder, a photoelectric switch or a servo driver frequency division output, the pulse interval is uniform, the camera is directly hard-triggered, software timing jitter is avoided, angle-image one-to-one correspondence is ensured, and measurement error caused by image misplacement is avoided.

[0005] Further, the illumination arrangement of S1 comprises: a backlight which is a parallel face light source placed on the back of the shoe arc, and the camera is opposite to take pictures, forming a high-contrast silhouette with narrow edge gray transition, which is beneficial to subsequent sub-pixel extraction; and a slant light which is four strip lights irradiated at a low angle, four gray scale images are taken at the same angle, shadows and reflections are eliminated, surface details of cracks and sand eyes are retained, and the backlight and the slant light are pulsed synchronously switched by the same encoder to ensure consistent imaging beat.

[0006] Further, the adaptive threshold segmentation of S2 adopts a histogram thresholding method with valley bottom, and then a 3x3 morphological closing operation is performed to fill the gaps in the mask to form a complete arc mask, avoiding contour extraction errors caused by edge breakage. The sub-pixel edge positioning of S2 slides a 7x7 Zernike matrix template on the mask, uses complex phase change to give sub-pixel offset, and forms a closed contour point column Pθ(r), providing a continuous basis for subsequent dimensionality lifting.

[0007] Further, in S2, 9-point third-order Savitzky-Golay filtering is performed on Pθ(r) to obtain a smooth radius sequence R(θ), which is input into a one-dimensional convolution network trained offline and frozen online, the network outputs the same angle height offset Δz with the "point cloud should be close to a cylinder" prior loss, and Δz is added to R(θ) to obtain a three-dimensional point cloud X(θ, y, z), without artificial depth label, completing laser-free three-dimensional contour reconstruction.

[0008] Further, the cylindrical axis fitting of S3 adopts a total least squares method, considers the errors of three coordinates of the three-dimensional point cloud, obtains the axis direction and the spatial center, then projects each point to the axis and calculates the perpendicular distance to form a d(θ) curve, performs extreme value search on d(θ), and the maximum minus the minimum is the round runout, and the maximum point angle is recorded to guide local removal of the jig boring machine.

[0009] Further, in S3, the arithmetic mean of one-week d(θ) data is first calculated as mean(d), then d(θ) at each angle is subtracted from the mean value to obtain the deviation d(θ)-mean(d) relative to the average cylindrical surface; d(θ)-mean(d) is plotted in polar coordinates, 0° corresponds to the starting tick line of the shoe arc, and the operator can directly see the peak position of the large runout value without reading.

[0010] Further, the unwrapping and splicing of S4 uses the same angle coding synchronization to splice four low-angle grayscale images by column, and the pixels are one-to-one corresponding to the angles to avoid manual alignment errors; then adaptive histogram equalization and Top-Hat transformation are performed to highlight the crack dark lines and remove the global light slow changes.

[0011] Further, the encoder-decoder network front end of S4 inserts a multi-scale hollow convolution bridge, three 3*3 hollow convolution rates are 1, 3 and 6, the channel remains 96, and after splicing, 1*1 is reduced back to 96 to form a multi-scale texture bridge, which is specially used to capture the global trend of crack-like long strip defects; the four-label sigmoid head at the end of the network outputs four confidence maps, which are mapped back to the original arc surface coordinates after 3*3 opening denoising, realizing one-to-one correspondence between defect pixels and jumping angles.

[0012] Further, the defect-jumping mutual feedback logic of S5 is: if the crack-like defect pixels are at the peak value ± 5° of the circle jumping and the confidence is > 0.6, then it is forced to be scrapped, if the sand hole-like defect is located in the low jumping area and the confidence is < 0.4, then it is allowed to be locally bored, and the rest is degraded to be scrapped, finally, qualified, bored or scrapped instructions are generated and encrypted and uploaded to MES, realizing one rotation, two detections and one report of digital online judgment.

[0013] The brake shoe circular arc geometry and surface defect parallel detection method provided by the application does not need a laser, can simultaneously measure the full-circle jumping and surface defects by only one rotation, has low cost and high beat, and the mutual feedback judgment can be traced. BRIEF DESCRIPTION OF DRAWINGS

[0014] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or prior art description. Obviously, the drawings in the following description only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0015] Figure 1 It is a schematic diagram of a brake shoe circular arc geometry and surface defect parallel detection method of the application. DETAILED DESCRIPTION

[0016] The embodiments of the present application will be described in detail below, and examples of the embodiments are shown in the drawings, wherein the same or similar reference signs represent the same or similar elements or elements with the same or similar functions throughout. The embodiments described below by referring to the drawings are exemplary and are intended to explain the present application, and cannot be understood as a limitation of the present application.

[0017] The application discloses a brake shoe circular arc geometry and surface defect parallel detection method. An industrial camera is used in cooperation with two kinds of illumination to rotate and shoot the brake shoe for one circle. The outer contour is extracted first, and then the contour is upgraded to a space circular arc. Meanwhile, a gray scale image is used to detect cracks and other defects. The contour and defect two-way results are aligned according to the angle, and after being combined, the qualified, boring repair or scrap judgment is given. The whole process does not add a laser, does not change the production line machine, and only uses the existing rotary table and industrial camera, so that the circular arc geometry and surface defect parallel online judgment can be completed, and a low-cost, high-tact, traceable digital means is provided for the brake shoe delivery or boring repair reinspection. The core of the method is that one machine is used for two purposes, that is, the same camera undertakes two imaging tasks at different times. The backlight imaging is used for geometry measurement, and the oblique light imaging is used for surface detection, so that the additional hardware investment is avoided, and the tact and the existing production line are synchronized.

[0018] As shown in Figure 1 The brake shoe circular arc geometry and surface defect parallel detection method of the application comprises the following steps: S1: equal-angle hard-trigger time-sharing imaging step The brake shoe assembly is uniformly rotated around the circular arc center, and the industrial camera is hard-triggered by the encoder to obtain backlight silhouette images and low-angle oblique light gray scale images at different times, so that each frame of image corresponds to the rotation angle.

[0019] 1. Rotary mechanism The existing numerical control rotary table or simple servo turntable can be connected, and only the equal-angle pulse can be sent every turn, and the pulse interval is uniform. The pulse directly triggers the camera, avoids the software timing error, and makes each frame of image correspond to the fixed angle. The equal-angle pulse can be the frequency division output of the encoder, photoelectric switch or servo driver, the pulse interval is uniform, the camera is directly hard-triggered, the software timing jitter is avoided, the key is that the pulse corresponds to the angle, the one-to-one correspondence between the image and the angle is ensured, and the measurement error caused by the image misalignment is avoided.

[0020] 2. Illumination arrangement Backlight: parallel surface light source is placed on the back of the shoe arc, and the camera is opposite to the light source to shoot, so that high-contrast silhouette is obtained, the edge gray scale transition is narrow, and the subsequent sub-pixel extraction is beneficial.

[0021] Oblique light: four strip lights irradiate the arc surface at a low angle, are lit at different times, and four gray scale images are shot at the same angle. The shadow and reflection are eliminated, and the surface details such as cracks and sand holes are reserved. The backlight and oblique light are switched synchronously by the same encoder pulse, and the imaging tact is consistent. The key of the backlight imaging lies in the parallelism of the light source. The higher the parallelism, the narrower the edge transition, and the higher the sub-pixel extraction accuracy. The key of the oblique light imaging lies in the angle control. The low-angle irradiation can highlight the surface concave-convex and enhance the contrast of the crack defects.

[0022] 3. Acquisition timing The rotating table rotates at a constant speed, and the encoder pulses trigger the camera and light source switching at the same time. Hundreds of silhouette images and several grayscale images are collected in one rotation, and the total time consumption is only a few seconds. The images are stored in the industrial computer buffer with the original bit width, preserving the dynamic range for subsequent processing. The preservation of the original bit width is to ensure the flexibility of subsequent image processing, especially in defect detection, where high dynamic range can better preserve the details of dark features such as cracks.

[0023] S2: Sub-pixel contour laser-free dimensionality step The silhouette image is subjected to adaptive threshold segmentation and Zernike matrix sub-pixel edge positioning to form a closed two-dimensional contour point column. The point column is input into a one-dimensional convolution network trained offline and frozen online. The cylindrical prior loss is used to output height offset, and a laser-free three-dimensional point cloud is generated.

[0024] 1. Backlight silhouette acquisition The camera is triggered once every 0.9° rotation of the rotating table. The camera is directed at a parallel light source with a horseshoe-shaped arc in between, resulting in a black arc shadow. The grayscale of the shadow edge transitions from black to white in only six pixels, providing sufficient gradient for subsequent sub-pixel extraction. The trigger signal is output directly by the encoder without passing through a software timer, avoiding millisecond-level jitter. The width of the edge grayscale transition directly affects the accuracy of sub-pixel extraction. A narrower transition results in a larger gradient and more accurate sub-pixel positioning. Hard triggering avoids the delay and jitter of software timers, ensuring the synchronization of image acquisition.

[0025] 2. Binary mask A histogram is taken of the central region of the current frame, and the threshold is obtained using the valley bottom method. The silhouette is divided into 0 and 1, resulting in a rough arc band. A 3x3 morphological closing operation is then used to fill in the gaps, resulting in a complete mask. The valley bottom thresholding method can adaptively process images under different lighting conditions, and the closing operation can fill in the edge gaps to ensure the integrity of the mask and avoid errors in contour extraction caused by edge breaks.

[0026] 3. Sub-pixel edge positioning A 7x7 Zernike matrix template is slid over the mask. When the center of the template aligns with the edge, the complex phase gives the sub-pixel offset, with an accuracy of 0.02 pixels. This allows for more precise contour extraction than whole-pixel methods. Each frame can obtain 400 edge points, which are arranged in order of trigger sequence to form a closed contour point column Pθ(r). The Zernike matrix method uses the change in complex phase to locate the edge, has rotational invariance and sub-pixel accuracy, and is suitable for high-precision edge extraction. The formation of the closed contour point column ensures the continuity of the contour and provides a basis for subsequent dimensionality elevation processing.

[0027] 4. Radius sequence smoothing The 9-point third-order Savitzky-Golay filter is applied to Ptheta(r) to remove quantization noise and retain the true wear valleys, obtaining a smooth radius sequence R(theta). The Savitzky-Golay filter can effectively remove noise while retaining the detailed characteristics of the signal, making it suitable for profile data smoothing.

[0028] 5. Self-supervised dimensionality increase The R(theta) is regarded as a one-dimensional signal and input into the offline trained 1D-CNN, and the network output is the height offset Delta z at the same angle. The CNN loss includes the priori that "the point cloud should be close to a cylinder", so there is no need for artificial depth labeling. The Delta z is added to the R(theta) to obtain the three-dimensional point cloud X(theta, y, z), completing the laser-free dimensionality increase. The self-supervised dimensionality increase utilizes geometric priori knowledge, learns the mapping from the radius sequence to the height offset through CNN, avoids the cost of artificial labeling of depth, and realizes laser-free three-dimensional measurement. The "point cloud should be close to a cylinder" priori loss is a kind of geometric constraint term without artificial labeling. It compares the three-dimensional point cloud output by the network with the theoretical cylindrical surface: if a point deviates from the cylindrical surface too far or the height of adjacent angles changes too much, the loss value will increase. By minimizing this loss during training, the network output profile is forced to approach the standard cylinder while maintaining circular arc continuity, so that reliable three-dimensional height can be self-supervisedly increased from the two-dimensional radius sequence.

[0029] 6. Precision verification Every month, ten pieces are taken out, and the same arc is sampled using a contact three-coordinate, and the network point cloud is compared. The maximum deviation is 0.028 mm, and 95% of the points are less than 0.02 mm, meeting the one-tenth allowance of runout ≤ 0.25 mm. Regular precision verification ensures the long-term stability and reliability of the system, and three-coordinate measurement as a reference can timely discover and correct system drift.

[0030] S3: Cylinder priori geometric runout step Perform total least squares cylinder axis fitting on the three-dimensional point cloud to obtain the full-circle runout curve and the maximum runout angle.

[0031] 1. Cylinder axis fitting Take the whole circle point cloud, use total least squares to obtain the cylinder axis vector, avoid the original point offset, and obtain the axis direction (nx, ny, nz) and the space center.

[0032] The total least squares method can consider the three coordinate errors of the data points at the same time, obtaining a more accurate axis fitting result, avoiding the limitation of the traditional least squares method which only considers one coordinate error.

[0033] 2. Distance calculation The length of projection t is obtained by projecting each point to the axis, and the vertical distance d of the point to the axis is calculated to form the d(θ) curve. Projection and distance calculation are the core steps of geometric runout calculation. The position error of the point cloud on the axis can be eliminated by projection to obtain the accurate vertical distance.

[0034] 3. Runout extraction The maximum and minimum of the d(θ) are searched to obtain the circular runout. The angle of the maximum point is recorded to guide the local amount removal of the boring machine. The extreme value search can directly obtain the value and position of the circular runout, which provides accurate guidance for subsequent processing and avoids material waste caused by blind processing.

[0035] 4. Polar coordinate graph

[0036] The arithmetic mean mean(d) of the d(θ) data in one week is calculated, and the d(θ) at each angle is subtracted from the mean value to obtain the deviation d(θ)-mean(d) from the average cylindrical surface. The d(θ)-mean(d) is drawn into a polar coordinate, and 0° corresponds to the starting tick line of the shoe arc. The operator can directly see the “peak” and “valley” positions without reading. The “peak” indicates that the brake shoe outer circle at this angle protrudes outward more than the average cylindrical surface, and the runout value is large. The “valley” indicates that it is more concave inward, and the runout value is small. Both directly reflect the position and degree of local arc over-difference or uneven wear, providing a clear target for subsequent selective boring repair. The intuitiveness of the polar coordinate graph enables the operator to quickly identify the problem area, improving production efficiency and processing accuracy.

[0037] 5. Reproducibility verification The standard deviation of the circular runout of 50 consecutive shoes from the same batch is 0.008 mm, and the GR&R is less than 10%, meeting the requirements of the gauge. The reproducibility verification ensures the stability and reliability of the system, and the GR&R less than 10% indicates that the measurement error of the system is within the acceptable range, meeting the requirements of industrial production.

[0038] S4: Unfolded map multi-label defect mapping step The oblique light grayscale image is spliced into an unfolded map according to the angle, input into an encoder-decoder network containing a multi-scale hollow convolution bridge, and outputs four types of defect confidence maps of cracks, sand holes, block drops, and ablation in parallel, and maps back to the original arc surface coordinates according to the angle.

[0039] 1. Gray scale unfolding Four-way bar light is lit in time, four gray scale images at the same angle are taken, and angle coding is synchronized to splice them into an unfolded map by column. The pixels correspond to the angles one by one, avoiding manual alignment errors. Angle coding synchronization ensures accurate image splicing, avoids manual alignment errors, and improves defect detection accuracy.

[0040] 2. Contrast enhancement Adaptive histogram equalization is performed on the unwrapped image to highlight the dark crack lines. Then, Top-Hat is used to remove the global light changes and retain the local defects. Contrast enhancement and Top-Hat transformation can effectively highlight the defect features and remove the influence of light changes, improving the visibility of defects.

[0041] 3. Network preprocessing The image is resized to a fixed width, and the height is scaled according to the arc length to maintain the pixel aspect ratio. Then, random brightness jitter is performed to simulate on-site light fluctuations. Standardization and augmentation of network preprocessing can improve the generalization ability of the model, making it suitable for different lighting and angle conditions.

[0042] 4. Network structure The lightweight convolutional network contains an encoder-decoder structure: the encoder performs layer-by-layer downsampling to extract crack, sand eye, block, and ablation features; the decoder performs step-by-step upsampling to output four confidence maps, each with the same size as the input, and the pixel value 0-1 represents the probability of defect existence. The end uses sigmoid activation to realize multi-label parallel detection, and one image can contain both cracks and sand eyes.

[0043] The encoder-decoder structure can effectively extract and restore features, and the sigmoid activation realizes multi-label parallel detection, improving the flexibility and accuracy of detection.

[0044] Specifically, the defect detection network adopts a lightweight encoder-decoder architecture, with a multi-scale attention embedded in the front end and four label classification heads connected in parallel at the end, which are dedicated to parallel segmentation of the four types of defects: camber crack, sand eye, block, and ablation.

[0045] The front-end encoder takes the 2048×512×1 grayscale unwrapped image as input. The first layer uses a 3×3 convolution with a stride of 2 and 32 channels to immediately downsample the feature map to 1024×256, reducing the computational load of subsequent layers. Then, four MobileNet-V3 basic blocks are stacked, each containing a depthwise separable convolution, a channel attention SE module, and a residual jump connection. After each block, the spatial size is halved, and the number of channels increases by 32→48→64→96, forming a four-level feature pyramid. The smallest feature map has a size of 128×64×96, and the receptive field covers the entire unwrapped image, capturing the global trend of long strip defects such as cracks.

[0046] The multi-scale attention bridge is inserted after the deepest layer of the encoder, with three parallel dilated convolution modules: three 3×3 convolutions with dilation rates of 1, 3, and 6, respectively, and 96 channels to expand the receptive field without downsampling, capturing local high-frequency information of small defects such as sand eyes. The three outputs are concatenated along the channel, and then a 1×1 convolution is used to reduce the number of channels to 96, forming a bridge feature map rich in multi-scale textures.

[0047] Decoder path: Bridge features recover spatial resolution after two upsampling. Before each upsampling, nearest-neighbor interpolation is used to expand the size by a factor of two and halve the channels; then, the features are concatenated with the same layer encoder features, and the concatenated features are fused by a 3x3 depthwise separable convolution, with the output channels being 64 and 48, respectively. The size of the top layer returns to 1024x256x48.

[0048] Four-label parallel head: After the 1024x256x48 features, a 1x1 convolution is performed, and the output channels are 4, corresponding to the four confidence maps of cracks, sand holes, chunks, and ablation. Each map has the same size as the input, and the pixel value is 0-1, which is generated by the end sigmoid activation, realizing multi-label parallel segmentation, allowing the same pixel to belong to both cracks and sand holes.

[0049] 5. Loss and training Focal-Loss is used to give higher weight to pixel-level defects, preventing background pixels from drowning out defect signals; training data comes from real defects on the daily production line, and random rotation, blur, and gamma transformation are used for augmentation, making the network insensitive to angle, lighting, and oil stains. Focal-Loss can effectively solve the class imbalance problem, and augmentation training can improve the robustness of the model, making it suitable for complex industrial environments.

[0050] In this application, the loss function is: L 总 = L focal + L dice + 0.1L boundary L focal is used to solve the extreme imbalance between background and defect pixels.

[0051] L focal = −α(1−p) γ log(p) where p is the pixel-level prediction probability (sigmoid output, 0-1), alpha is the balance factor, alpha=0.75, which increases the weight of defect pixels, and gamma is the modulation factor. Gamma=2, which gives greater weight to defect pixels with low confidence. L focal Weighting individual p solves the class imbalance problem.

[0052] L dice is used to improve the integrity of the defect area.

[0053] L dice = 1 − 2∑(p·g) / (∑p + ∑g) Where p is the confidence of all pixels in the whole prediction image (0-1), g is the binary ground truth label of the corresponding pixel, and the summation is over all pixels. ∑(p·g) represents the intersection pixel sum of prediction and ground truth. ∑p + ∑g represents the union approximate sum of prediction and ground truth.

[0054] L boundary For enhancing the crack edge sharpness. First, the ground truth image is convolved with a 3x3 Laplacian kernel to obtain the edge image e, and then: L boundary =∑|p−e| / N Where p is the confidence of the whole prediction image (0-1), and the summation is only over the edge pixels. N is the total number of edge pixels.

[0055] Here, the physical meaning of p is consistent, both representing pixel-level prediction probability. However, the calculation granularity and purpose are different in different formulas. Focal is for single-pixel weighting, Dice is for region overlap, and Boundary is for edge difference.

[0056] 6. Post-processing For each confidence map, a 3x3 opening is performed to remove isolated noise points. Then, it is mapped back to the original coordinates of the hoof arc according to the angle, obtaining the defect starting angle and width, which are used for superimposed display with the geometric runout curve. The denoising and mapping of post-processing can improve the accuracy of defect detection. Angle mapping makes the defect position and geometric measurement results superimposed and displayed, which is convenient for comprehensive analysis.

[0057] S5: Defect-runout mutual feedback determination step Align the circular runout curve and the four types of defect confidence maps by angle and pixel level, and execute the defect-runout mutual feedback logic.

[0058] 1. Angle alignment The geometric branch outputs the d(θ) curve, and the defect branch outputs four confidence maps. They are all numbered with the same angle pulse and aligned by θ pixel to realize pixel-level correspondence. Angle alignment is the key to ensure that geometric and defect results can be superimposed and displayed. Pixel-level correspondence makes comprehensive analysis possible.

[0059] 2. Mutual feedback determination logic If the circular runout is below the threshold and all defect confidences are low, it is marked as qualified; If the runout is out of tolerance but the defect is slight, it is marked as needing boring repair and the maximum runout angle and width are given; If any defect confidence is high, it is marked as scrap, and the code and sound-light prompt are given. The mutual feedback determination logic considers the results of geometry and defects comprehensively, avoids misjudgment of a single indicator, and improves the accuracy and reliability of the determination.

[0060] Specifically, if the crack-like defect pixel is within the circle run-out peak ± 5° and the confidence is > 0.6, it is forced to be scrapped, if the sand hole-like defect is located in the low run-out area and the confidence is < 0.4, it is allowed to be locally repaired, and the rest is degraded to be scrapped.

[0061] Introduce the "defect-geometry mutual feedback" mechanism: if the defect is located in the high peak area of run-out and the defect type is crack, even if the run-out value is slightly lower than the threshold, it is degraded to be scrapped, because the crack is easy to expand in the high peak area; if the defect is sand hole and located in the low run-out area, it is allowed to be repaired by boring, and is not degraded. This logic is written in the judgment script, without manual intervention. The innovation of the mutual feedback mechanism is to comprehensively consider the position and type of the defect, as well as the distribution of geometric run-out, to realize more intelligent judgment and improve the scientificity and rationality of the judgment.

[0062] 3. Data upload Finally, the qualified, boring or scrapped instructions are generated and uploaded to the MES, realizing the digital online judgment of one rotation, two detections and one report. Specifically, a unique two-dimensional code is generated for each brake shoe, containing the run-out value, defect map and judgment result; using AES-256 encryption, the factory MES is uploaded through MQTT, forming a traceable digital archive. Data upload and encryption ensure the security and traceability of the data, and the uniqueness of the two-dimensional code makes each product traceable, improving the efficiency of quality management.

[0063] The detection method of the present application takes the same industrial camera as the core, and completes two kinds of imaging at different times: backlight silhouette sub-pixel edge extraction and self-supervised dimensionality, forming a laser-free 3D run-out measurement link; oblique light gray is expanded and spliced with light network reasoning, and the multi-label defect confidence link is output. The same angle pulse is shared by the two links to ensure the pixel-level alignment of geometry and defects, and the circular arc geometry data and surface defect data can be obtained at the same time in one rotation, and a digital report is generated by merging. Without additional laser, without changing the existing production line machinery, only using the rotating table and industrial camera, the low-cost, high-tact, traceable and parallel online detection of brake shoe circular arc geometry and surface defects can be realized.

[0064] The basic principles, main features and advantages of the present application are shown and described above. Those skilled in the art should understand that the above examples do not limit the present application in any form, and any technical solutions obtained by equivalent substitution or equivalent transformation fall within the protection scope of the present application.

Claims

1. A method for parallel detection of brake shoe arc geometry and surface defects, characterized in that, In order, they include: S1: Equal angle hard trigger time-division imaging step, used to drive the brake shoe assembly to rotate at a constant speed around the center of the arc, and use the encoder to hard trigger the industrial camera to acquire backlight silhouette image and low angle oblique light grayscale image in time-division, so that each frame image strictly corresponds to the rotation angle. S2: Subpixel contour laser-free dimensionality enhancement step, used to perform adaptive threshold segmentation and Zernike moment subpixel edge localization on the silhouette image to form a closed two-dimensional contour point column, and input the point column into an offline trained and online frozen one-dimensional convolutional network, using cylindrical prior loss to output height offset and generate a laser-free three-dimensional point cloud. S3: Cylinder prior geometric jump step, used to perform overall least squares cylinder axis fitting on the three-dimensional point cloud to obtain the full circumference jump curve and the maximum jump angle; S4: Unfolded image multi-label defect mapping step, used to encode and stitch the oblique light grayscale image into an unfolded image according to the angle, input the encoder-decoder network containing multi-scale dilated convolution bridge, and output confidence maps of four types of defects: crack, sand hole, chipping, and ablation in parallel, and map them back to the original arc surface coordinates according to the angle. S5: Defect-bouncing feedback judgment step, used to align the circular runout curve with the confidence maps of the four types of defects at the angular pixel level, and execute the defect-bouncing feedback logic.

2. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 1, characterized in that, In the equal-angle hard-triggered time-division imaging step of S1, the rotating mechanism uses an encoder, photoelectric switch or servo driver to divide the output frequency, with uniform pulse intervals, directly hard-triggers the camera, avoids software timing jitter, ensures that the angle corresponds one-to-one with the image, and avoids measurement errors caused by image misalignment.

3. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 2, characterized in that, The lighting setup for S1 includes: backlighting is a parallel surface light source placed on the back of the hoof arch, with the camera shooting in the opposite direction to form a high-contrast silhouette with narrow edge grayscale transitions, which is beneficial for subsequent sub-pixel extraction; oblique lighting consists of four strip lights illuminating at low angles in a time-division manner, taking four grayscale images at the same angle to eliminate shadows and reflections, and retain surface details such as cracks and sand holes. The backlighting and oblique lighting are switched synchronously by the same encoder pulse to ensure consistent imaging rhythm.

4. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 1, characterized in that, The adaptive threshold segmentation of S2 uses the valley method histogram threshold, followed by 3×3 morphological closing operation to fill the mask gaps and form a complete arc mask, avoiding contour extraction errors caused by edge breakage. The subpixel edge localization of S2 is achieved by sliding a 7×7 Zernike moment template on the mask, using complex phase transformation to give the subpixel offset, forming a closed contour point sequence Pθ(r), which provides a continuous basis for subsequent dimensionality increase.

5. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 4, characterized in that, In S2, a 9-point third-order Savitzky-Golay filter is applied to Pθ(r) to obtain a smooth radius sequence R(θ). The radius sequence R(θ) is input into a one-dimensional convolutional network that is trained offline and frozen online. The network outputs a height offset Δz at the same angle with the prior loss of "the point cloud should be close to a cylinder". Δz is added to R(θ) to obtain the three-dimensional point cloud X(θ,y,z). Laser-free three-dimensional contour reconstruction is completed without the need for manual depth labels.

6. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 1, characterized in that, The cylindrical axis fitting of S3 adopts the overall least squares method, while considering the three coordinate errors of the three-dimensional point cloud to obtain the axis direction and spatial center. Then, each point is projected onto the axis and the vertical distance is calculated to form the d(θ) curve. The extreme value search is performed on d(θ), and the maximum minus the minimum is the circular runout. At the same time, the maximum point angle is recorded to guide the local measurement of the boring machine.

7. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 6, characterized in that, S3 first calculates the arithmetic mean mean(d) of the d(θ) data for one week, and then subtracts this mean from the d(θ) at each angle to obtain the deviation d(θ)−mean(d) relative to the average cylindrical surface; d(θ)−mean(d) is plotted as a polar coordinate graph, with 0° corresponding to the starting line of the hoof arc, so that the operator can intuitively see the peak position of the large jump value without taking readings.

8. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 1, characterized in that, The S4 unfolded image stitching uses the same angle encoding synchronization, stitching four low-angle grayscale images column by column, with pixels and angles corresponding one-to-one, avoiding manual alignment errors; then adaptive histogram equalization and Top-Hat transformation are performed to highlight the crack dark lines and remove slow changes in global illumination.

9. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 8, characterized in that, The S4 encoder-decoder network front end inserts a multi-scale dilated convolution bridge with three 3×3 dilated convolutions of dilution rates of 1, 3, and 6, maintaining a channel density of 96. After merging, the 1×1 convolutions are reduced back to 96, forming a multi-scale texture bridge specifically for capturing the global orientation of crack-like elongated defects. The network's end has a four-label sigmoid head that outputs four confidence maps. After denoising with 3×3 opening, these maps are encoded by angle and mapped back to the original arc surface coordinates, achieving a one-to-one correspondence between defect pixels and jump angles.

10. The method for parallel detection of brake shoe arc geometry and surface defects according to claim 1, characterized in that, The defect-runout feedback logic of S5 is as follows: if the pixel of the crack-type defect is within ±5° of the peak of the circular runout and the confidence level is >0.6, it is forcibly scrapped; if the sand hole-type defect is located in the low runout area and the confidence level is <0.4, it is allowed to perform local boring repair; otherwise, it is downgraded to scrap. Finally, the qualified, boring repair or scrap instruction is generated and encrypted and uploaded to MES, realizing digital online judgment of one rotation, two inspections and one report.

Images