Method and system for detecting tooth surface defects of precision gear

Abstract

The invention provides a tooth surface defect detection method and system for a precision gear. The method comprises the following steps: acquiring an original image sequence covering the whole circumferential direction of a tooth surface of the precision gear; preprocessing the original image sequence to generate a defect candidate region set; inputting a gear image with a fixed size in each region in the defect candidate region set into a deep neural network, and determining a preliminary detection result of the tooth surface of the precision gear; performing joint constraint regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box, and generating a regression correction bounding box set corresponding to the preliminary detection result; and judging the defect type of the tooth surface of the precision gear according to the spatial position information of the regression correction bounding box set and the corresponding category label. By establishing the index relationship between the image sequence and the gear rotation angle, image acquisition and accurate tooth position matching are realized, and the defect detection precision is effectively improved through cooperation of the deep neural network and the regression correction mechanism.

Description

Technical Field

[0001] The present application relates to the technical field of gear detection, and in particular to a method and system for detecting tooth surface defects of precision gears. Background Art

[0002] In modern high-end manufacturing, precision gears, as core transmission components, are widely used in key areas such as aerospace, intelligent equipment, CNC machine tools, and new energy vehicles. Their machining accuracy and surface quality directly impact the stability and service life of the entire machine. Taking high-speed gearbox manufacturing as an example, if tooth surface defects such as cracks, spalling, indentations, and missing teeth are not detected promptly, they can not only cause gear failure but also abnormal vibration of the entire machine and even damage the transmission system, posing serious safety and economic risks to industrial production. Therefore, building an efficient, stable, and automated tooth surface defect detection system during the gear production process has become a critical issue that the industry urgently needs to address. Currently, industrial sites widely use manual visual inspection or traditional machine vision methods based on simple image segmentation and contour analysis to perform visual inspection of tooth surfaces. These methods are prone to false detection and missed detection when faced with complex defect morphology, large scale variations, and strong local reflective interference backgrounds, and are unable to accurately classify and locate defects. Summary of the Invention

[0003] The present application provides a method and system for detecting tooth surface defects of precision gears, which are used to solve the problem in related technologies that traditional machine vision methods cannot accurately classify defects.

[0004] A first aspect of the present application provides a method for detecting tooth surface defects of a precision gear, the method comprising: Obtaining a sequence of original images covering the entire circumference of the precision gear tooth surface; Generating a defect candidate region set by preprocessing the original image sequence; Determining a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each region in the defect candidate region set into a deep neural network; Performing joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box to generate a regression corrected bounding box set corresponding to the preliminary detection result; The defect type of the precision gear tooth surface is determined according to the spatial position information of the regression corrected bounding box set and the corresponding category label.

[0005] Optionally, in a first implementation of the first aspect of the present application, the step of acquiring a sequence of original images covering the entire circumference of the precision gear tooth surface includes: Perform step angle control on precision gears according to the preset data acquisition path to generate a rotation angle sequence covering the tooth surface; Performing image acquisition at a specified angle position of the precision gear according to the rotation angle sequence to obtain an original image frame corresponding to the specified angle position; Constructing a mapping table between the original image frame and the gear angle position information according to the shooting angle identifier and the timestamp information of the original image frame; Binding the mapping table with the original image frame to generate an image index structure set; By sequentially arranging the image frames in the image index structure set, an original image sequence covering the entire circumference of the precision gear tooth surface is generated.

[0006] Optionally, in a second implementation of the first aspect of the present application, the step of generating a set of defect candidate regions by preprocessing the original image sequence includes: Obtaining a set of basic grayscale images corresponding to each image frame by performing grayscale processing and median filtering noise reduction operations on each image frame in the original image sequence; Generate an enhanced image set by performing multi-scale image enhancement processing and local contrast improvement processing on the basic grayscale image set; Determining a reference image for differential analysis based on pixel grayscale distribution characteristics of the enhanced image set, and generating a difference image set by calculating a difference matrix between the reference image and the enhanced image; By performing threshold segmentation on the difference map set, image regions with closed edge contours are extracted to generate a defect candidate region set.

[0007] Optionally, in a third implementation of the first aspect of the present application, the step of determining a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each area in the set of defect candidate areas into a deep neural network includes: Generate a fixed-size image subset that meets preset input specifications by performing a size normalization operation on all image regions of the defect candidate region set; Performing multi-scale feature extraction and adaptive convolution on the fixed-size image subset in a deep neural network to generate a feature expression graph set; Performing regression processing on target position parameters according to the spatial feature distribution extracted from the feature expression graph set to obtain a set of defect bounding boxes corresponding to all gear images; A preliminary detection result of the precision gear tooth surface is generated by classifying the feature vector corresponding to each bounding box in the defect bounding box set.

[0008] Optionally, in a fourth implementation of the first aspect of the present application, the step of performing a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the corresponding true value box coordinates to generate a regression-corrected bounding box set corresponding to the preliminary detection result includes: Constructing a multidimensional residual vector set based on the predicted coordinates of each defect boundary box in the preliminary detection results and the corresponding true value frame coordinates; Constructing a joint loss function by performing function fitting on the multidimensional residual vector set; wherein the joint loss function is composed of a bounding box coincidence function, a boundary offset function, and a size ratio function; Performing back-propagation training on the regression network according to the joint loss function to obtain a bounding box regression correction parameter set corresponding to the preliminary detection result; The predicted coordinates are numerically corrected according to the bounding box regression correction parameter set to generate a set of regression corrected bounding boxes that correspond one-to-one to the preliminary detection results.

[0009] Optionally, in a fifth implementation of the first aspect of the present application, the step of determining the defect type of the precision gear tooth surface based on the spatial position information of the regression corrected bounding box set and the corresponding category label includes: Constructing a topological relationship matrix of spatial distribution features of the bounding boxes based on the geometric center coordinates and size parameters of each bounding box in the regression-corrected bounding box set; Performing density cluster analysis on the topological relationship matrix by presetting a cluster radius threshold to generate a set of defect space clusters covering different areas of the tooth surface; Determining a candidate defect type subset based on density distribution characteristics of each defect space cluster in the defect space cluster set and statistical frequencies of corresponding category labels; The defect type determination result of the precision gear tooth surface is generated by determining the spatial similarity between the spatial distribution pattern of each defect type in the candidate defect type subset and the historical defect pattern in the preset defect type template library.

[0010] Optionally, in a sixth implementation of the first aspect of the present application, the method further includes: generating a three-dimensional topological map of tooth surface defects based on the coordinate information of the regression corrected bounding box set and the three-dimensional tooth surface model of the precision gear; By means of the correlation between the three-dimensional topological map and the gear meshing motion parameters, dynamic stress simulation is performed on the defective area to obtain stress distribution data of the defective area; Based on the stress distribution data and a preset defect evolution model, the defect expansion path during the gear operation is predicted to generate a defect expansion trajectory prediction diagram; By performing correlation analysis on the defect expansion trajectory prediction diagram and the fatigue life parameters of the gear material, a safety level assessment result of the tooth surface defect is generated.

[0011] A second aspect of the present application provides a device for detecting tooth surface defects of a precision gear, the device comprising: An acquisition module is used to acquire a sequence of original images covering the entire circumference of the precision gear tooth surface; A processing module, configured to generate a set of defect candidate regions by preprocessing the original image sequence; a determination module, configured to determine a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each region in the defect candidate region set into a deep neural network; a generation module, configured to perform a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box, to generate a regression-corrected bounding box set corresponding to the preliminary detection result; A judgment module is used to judge the defect type of the precision gear tooth surface according to the spatial position information of the regression corrected bounding box set and the corresponding category label.

[0012] A third aspect of an embodiment of the present application provides an electronic device, comprising a memory and a processor, wherein the processor is used to execute a computer program stored on the memory, and when the processor executes the computer program, it implements each step of the method for detecting tooth surface defects of precision gears provided in the first aspect of the embodiment of the present application.

[0013] The fourth aspect of the embodiment of the present application provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed by a processor, the steps of the method for detecting tooth surface defects of precision gears provided in the first aspect of the embodiment of the present application are implemented.

[0014] In summary, according to the present invention, a method and system for detecting tooth surface defects of precision gears is provided. An original image sequence covering the entire circumference of the tooth surface of the precision gear is obtained. A set of defect candidate regions is generated by preprocessing the original image sequence. A preliminary detection result of the tooth surface of the precision gear is determined by inputting a gear image of a fixed size in each region of the defect candidate region set into a deep neural network. A joint constrained regression is performed between the predicted coordinates of the defect bounding box in the preliminary detection result and the corresponding true value frame coordinates to generate a set of regression-corrected bounding boxes corresponding to the preliminary detection result. The defect type of the tooth surface of the precision gear is determined based on the spatial position information of the regression-corrected bounding box set and the corresponding category label. By establishing an index relationship between the image sequence and the gear rotation angle, accurate matching of image acquisition and tooth position is achieved, and the accuracy of defect detection is effectively improved by coordinating the deep neural network with the regression correction mechanism. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] Figure 1 A schematic flow chart of a method for detecting tooth surface defects of precision gears provided in an embodiment of the present application; Figure 2 A schematic diagram of a program module of a precision gear tooth surface defect detection device provided in an embodiment of the present application; Figure 3 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0016] In order to make the purpose, features, and advantages of the invention of this application more obvious and easy to understand, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are only part of the embodiments of this application, not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts shall fall within the scope of protection of this application.

[0017] In order to solve the problem that traditional machine vision methods in related technologies cannot accurately classify defects, the embodiment of the present application provides a method for detecting tooth surface defects of precision gears, such as Figure 1 The present invention provides a flowchart of a method for detecting tooth surface defects of precision gears. The method for detecting tooth surface defects of precision gears includes the following steps: Step 110: Acquire an original image sequence covering the entire circumference of the precision gear tooth surface.

[0018] Specifically, by constructing a collaborative control platform that includes a precision gear rotation control module and an image acquisition system, a stepper motor drives the gear to rotate around its axis at equal intervals during the data acquisition process, and an image acquisition is triggered every time it rotates a set angle. This process uses a high-resolution industrial camera to capture the tooth surface from a top or side view, and uses a uniformly distributed annular LED light source to fill in the reflective areas of the tooth surface to avoid local overexposure or shadow interference. Each frame of the image collected is automatically associated with the current rotation angle, timestamp, and image frame number through the program to generate an image-angle mapping index table. This mapping relationship ensures that each defect detected subsequently can be accurately mapped to the specific tooth position number of the gear, providing a spatial reference for defect recording, tracking, and marking.

[0019] In an optional implementation of this embodiment, the step of obtaining an original image sequence covering the entire circumference of the precision gear tooth surface includes: performing step angle control on the precision gear according to a preset data acquisition path to generate a rotation angle sequence covering the tooth surface; performing image acquisition on a specified angle position of the precision gear according to the rotation angle sequence to obtain an original image frame corresponding to the specified angle position; constructing a mapping table between the original image frame and the gear angle position information according to the shooting angle identifier and timestamp information of the original image frame; binding the mapping table to the original image frame to generate an image index structure set; and generating an original image sequence covering the entire circumference of the precision gear tooth surface by sequentially arranging the image frames in the image index structure set.

[0020] Specifically, a parameterized data acquisition path control system implements step-by-step angle control for precision gears. A stepper motor-driven turntable precisely positions the gear around its rotation axis at equal or unequal intervals according to a preset angle list. Each angle value in the list corresponds to the gear pose required for one image capture, covering the full 360-degree circumference of the tooth surface. For example, during gear quality inspection, to ensure that every tooth surface is captured, 360° can be divided into 120 sections, and the turntable can perform positioning in 3° increments. A high-precision optical encoder provides feedback on the current position, triggering the next stage of operation each time a specified angle is reached. Subsequently, using the rotation angle sequence as a reference, the industrial camera control module executes image acquisition commands at the specified angle positions. The industrial camera is equipped with a global shutter sensor, which allows for quick exposure and clear image output after stable gear rotation. If each gear tooth tip position is the capture target, the control system sends a "capture" signal to the camera during each 3° turntable pause. The gear state within the current field of view is captured as a raw image frame and stored in the edge computing unit. This process ensures that each raw image frame corresponds to a specific physical angle, laying the foundation for accurate mapping of subsequent images to gear positions. Next, a shooting angle identifier and timestamp are automatically added to each raw image frame, creating a mapping table between the image frame and gear angle position information. The shooting angle identifier is generated by combining the camera control signal and the stepper motor's current position feedback, while the timestamp records the system clock value at the moment of image acquisition. By storing corresponding fields such as angle and time in image metadata or a relational database, the physical position corresponding to each image frame can be queried based on this mapping table during subsequent processing. For example, when re-inspecting an abnormal image after a system restart, the corresponding rotation angle and time information can be quickly located, achieving precise traceability. The mapping table is then bound to the raw image frame to generate an image index structure set. An index structure set is a structured set that organizes the image file name, angle value, timestamp, and other auxiliary information, such as illumination parameters and exposure time, into a single data unit. Image file name, angle, and time are associated using a binary tuple or hash table, and all these associated records are aggregated and stored in a relational database or distributed file system. This allows efficient localization of contextual information for any frame within thousands of images without manual retrieval, providing data support for parallelized image processing. Finally, all image records in the aforementioned index structure are arranged in ascending order or in a customized sequence based on their angle values, yielding a raw image sequence covering the entire circumference of the precision gear tooth surface. The arrangement rules for this sequence can be adjusted based on the production line's tactile cycle or inspection strategy. The resulting image files can be retained as local cache or pushed sequentially to the back-end image processing pipeline via a message queue to ensure that the inspection algorithm reads the images strictly in the order of tooth surface rotation, thereby achieving precise spatiotemporal synchronous inspection and positioning.

[0021] Step 120: Generate a set of defect candidate regions by preprocessing the original image sequence.

[0022] Specifically, the collected raw image sequences are preprocessed. First, the color images are converted to grayscale to simplify feature extraction. Median filtering is then used to eliminate image noise, while image enhancement methods are employed to highlight tooth surface details. To further identify potential abnormal areas, a difference map is constructed to perform pixel-level comparisons between the reference image and the current image. Morphological operations such as dilation and erosion are then applied to extract regions with closed contour features. Ultimately, a collection of image regions with potential defects is obtained, which are then fed into the next stage of the detection model as targets for analysis.

[0023] In an optional implementation of the present embodiment, the step of generating a set of defect candidate areas by preprocessing the original image sequence includes: obtaining a set of basic grayscale images corresponding to each image frame by performing grayscale processing and median filtering noise reduction operations on each image frame in the original image sequence; generating a set of enhanced images by performing multi-scale image enhancement processing and local contrast enhancement processing on the basic grayscale image set; determining a reference image for differential analysis based on the pixel grayscale distribution characteristics of the enhanced image set, and generating a set of difference images by calculating the difference matrix between the reference image and the enhanced image; and generating a set of defect candidate areas by performing threshold segmentation on the difference image set to extract image areas with closed edge contours.

[0024] Specifically, in this embodiment, a weighted averaging method is used to fuse the RGB channels into a grayscale image based on the human eye's visual sensitivity ratio, thereby mapping the image's color information into intensity values. After grayscaling, a median filter algorithm is applied to remove noise. Median filtering is a nonlinear signal smoothing technique that removes salt-and-pepper noise and isolated impulse noise by taking a fixed-size sliding window around each pixel, sorting the grayscale values of all pixels within the window, and replacing the center pixel with the median value. Next, multiscale image enhancement and local contrast enhancement are performed on the resulting basic grayscale image set to enhance the visibility of tooth surface defects. Multiscale enhancement utilizes a Gaussian pyramid or Laplacian pyramid approach to generate image layers of varying resolutions, with high-frequency details enhanced at each layer. For example, using the Laplacian pyramid, a low-pass Gaussian filter is first used to generate a smoothed image, which is then subtracted from the original image to obtain a high-frequency detail layer. This high-frequency layer is then amplified by a factor of, for example, 1.5 or 2, before being reconstructed with the low-frequency components. This enhances crack edges and missing tooth outlines without exaggerating noise. At the same time, an adaptive histogram equalization algorithm is used to adjust the contrast of the local image grid. By mapping the local pixel distribution to a wider grayscale range, the originally weak defect signal is made more prominent in the enhanced image collection. Subsequently, by analyzing the overall pixel grayscale distribution of the enhanced image collection, the most representative reference image is determined, providing a standard template for subsequent differential analysis. The reference image can be selected from a defect-free initial calibration sample or a frame in the enhanced image collection. The reference image is automatically selected based on its grayscale histogram concentration in the middle grayscale region and the lowest standard deviation. For example, by calculating the average grayscale and standard deviation of each frame in ten consecutive image frames, the image with the lowest standard deviation and a mean close to middle gray can be selected as the reference. The selected reference image is then subtracted pixel by pixel from the remaining enhanced images in the collection to generate a difference matrix. The difference matrix is a two-dimensional array in which each element represents the grayscale difference between the enhanced image and the reference image. Taking the absolute value of the difference matrix generates a difference map, which highlights the local abnormal areas between the images. The difference map shows the distribution of pixels with the most significant contrast between light and dark along the edges of the contours. For tooth surface defects, grayscale changes at cracks or pits are often the most prominent, resulting in continuous blocks of high or low grayscale in the difference map. Finally, the difference map is binarized using an adaptive thresholding algorithm (such as the Otsu method) or a fixed thresholding method. Pixels above the threshold are set to white (value 1) and pixels below the threshold are set to black (value 0), thereby highlighting potential defect areas. Morphological dilation is then used to fill small gaps between cracks or noise points, and erosion is used to eliminate isolated noise, ultimately resulting in closed, connected regions.In the binary image, connected domain analysis identifies all closed regions. Based on geometric constraints such as area, perimeter, or aspect ratio, the system removes pseudo-contours that are too small or do not conform to the morphological characteristics of tooth surface defects. For example, noise regions with an area less than 20 pixels and an aspect ratio less than 2:1 are excluded. This allows the extraction of true cracks and missing teeth as candidate regions. The resulting candidate region set consists of a set of bounding boxes and their corresponding binary region masks. Each candidate region points to a possible defect location in the image, providing precise input for subsequent depth detection and positioning.

[0025] Step 130: Determine the preliminary detection result of the precision gear tooth surface by inputting the gear image of each region with a fixed size in the defect candidate region set into the deep neural network.

[0026] Specifically, in this embodiment, the defect candidate regions extracted in the previous stage are uniformly resized to meet the input requirements of the deep neural network model. The processed images are then fed into a neural network architecture capable of multi-scale feature extraction. A convolutional module within the network extracts spatial hierarchical features from the image, and a regression module outputs the location bounding boxes and category predictions for the candidate defects. This process achieves preliminary localization and classification of defect regions, and the output detection results include the location information, category label, and confidence score for each suspected defect.

[0027] In an optional implementation of this embodiment, the step of determining the preliminary detection result of the precision gear tooth surface by inputting the gear image of each area in the defect candidate area set into the deep neural network includes: performing a size normalization operation on all image areas of the defect candidate area set to generate a fixed-size image subset that meets the preset input specifications; performing multi-scale feature extraction and adaptive convolution on the fixed-size image subset in the deep neural network to generate a feature expression graph set; performing regression processing on the target position parameters according to the spatial feature distribution extracted from the feature expression graph set to obtain a defect bounding box set corresponding to all gear images; and generating the preliminary detection result of the precision gear tooth surface by classifying the feature vectors corresponding to each bounding box in the defect bounding box set.

[0028] Specifically, in this embodiment, in order to meet the requirements of the deep neural network for the input image size, the image segments in all defect candidate areas need to be size normalized. This process is usually completed using a bilinear interpolation algorithm by mapping the original pixel coordinates to a uniform target size. The interpolation process involves weighted averaging of adjacent pixels in the original image to obtain the grayscale or RGB value of the target location. This can not only preserve the texture details within the area but also avoid the introduction of obvious pixelation problems during scaling. Size normalization ensures that the network has the same receptive field size, so that the subsequent recognition of defect size and shape in different candidate areas is comparable. Next, multi-scale feature extraction and adaptive convolution operations are performed on the fixed-size image subset through a pre-designed convolutional neural network architecture. Multi-scale feature extraction refers to the use of convolution kernels of different sizes or pyramid structures (such as Feature Pyramid Network, FPN) within the network to process the overall outline and microscopic details of the image in parallel. For example, large 7×7 convolution kernels are used in the first few layers to obtain coarse-grained shape information, 3×3 convolution kernels are used in the middle layers to capture mid-scale texture cues, and 1×1 convolution is combined at the end to generate point-wise feature maps. Adaptive convolution refers to the dynamic adjustment of the convolution kernel weight distribution based on the local features of the input image during the prediction process. This can be achieved through a learnable routing function or spatial attention mechanism. For example, different weights are assigned to highlight areas and dark shadows on the tooth surface to mitigate the impact of reflective interference on detection accuracy. Through these modules, the network ultimately forms a set of multi-scale, multi-channel feature representation maps. Each feature map represents information about the input image at different levels of abstraction, including edge enhancement, high-frequency texture, and low-frequency contours. After obtaining the feature representation map, regression processing is performed on the target position parameters to locate the candidate defect areas. This regression process can adopt an end-to-end regression head structure, such as the regression head in a single-stage detector, which is input from feature maps at each layer of the feature pyramid. The regression head first uniformly maps the feature maps of each layer to a fixed number of channels, extracts spatial information through a series of convolutional layers, and then outputs a four-dimensional vector, which corresponds to the horizontal offset, vertical offset, and width and height scaling of the center point of the predicted bounding box. To guide the regression process, the loss function uses a joint loss with distance metric and overlap evaluation to ensure that the spatial position and shape size between the predicted box and the true box are highly consistent. Finally, the set of defect bounding boxes obtained after regression needs to be paired with the feature vector output by the classification branch to generate a complete detection result. The classification branch is common in single-stage detectors. It is responsible for estimating the category probability for each candidate box or each network grid point and outputting a probability vector with a length equal to the number of categories.This classification vector is typically compressed into a feature vector using global average pooling, then mapped using a fully connected layer or a separable convolutional layer. Finally, it is converted into class probabilities using the Softmax function. By applying the Argmax operation to the feature vector corresponding to each defect bounding box, its class label is determined. For example, if the output feature vector is [0.05, 0.12, 0.01, 0.78, 0.04], where the positions in the vector correspond to no defect, crack, scratch, missing tooth, and contamination, respectively, then argmax([0.05, 0.12, 0.01, 0.78, 0.04]) = 3, indicating that the third position has the highest probability, and the network predicts the category of this area as "missing tooth." Excessive overlapping boxes are then removed, ensuring that the final set of bounding boxes retains only the highest-scoring, non-overlapping detections.

[0029] Step 140 : Perform joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box to generate a regression corrected bounding box set corresponding to the preliminary detection result.

[0030] Specifically, in this embodiment, the bounding box coordinates output by the network are compared with the pre-annotated ground-truth box coordinates. A residual vector is constructed to describe the position and size differences between the predicted and true boxes. A joint loss function is then used to quantify the multi-dimensional error. The loss function includes multiple evaluation factors such as boundary overlap, boundary offset, and size ratio. During training, a backpropagation mechanism guides the model to optimize the bounding box prediction accuracy, ultimately generating a set of corrected bounding box coordinates that are closer to the actual defect location than the initial predictions, with higher consistency and stability.

[0031] In an optional implementation of the present embodiment, a step of performing a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the corresponding true value box coordinates to generate a regression corrected bounding box set corresponding to the preliminary detection result includes: constructing a multidimensional residual vector set based on the predicted coordinates of each defect bounding box in the preliminary detection result and the corresponding true value box coordinates; constructing a joint loss function by performing function fitting on the multidimensional residual vector set; wherein the joint loss function is composed of a bounding box coincidence function, a boundary offset function and a size ratio function; performing backpropagation training on the regression network according to the joint loss function to obtain a bounding box regression correction parameter set corresponding to the preliminary detection result; and performing numerical correction on the predicted coordinates according to the bounding box regression correction parameter set to generate a regression corrected bounding box set corresponding one-to-one to the preliminary detection result.

[0032] In this embodiment, by matching the predicted bounding box coordinates of each pair of preliminary inspection results with the manually annotated ground-truth bounding box coordinates, a multidimensional residual vector consisting of position and size differences is generated for each defect instance. This vector contains the horizontal and vertical offset differences of the center point, as well as the ratio difference between width and height, thus uniformly representing spatial position error and scale inconsistency as four-dimensional or multi-dimensional data. This process not only ensures that the network quantifies the geometric deviation of each predicted box but also provides an accurate error signal for subsequent regression optimization. For example, if the ground-truth bounding box center of a crack is at coordinates (100, 150) and the network predicts it as (110, 142), and the width is predicted to be 50 pixels while the true width is 45 pixels, the center offset (+10, –8) and the scale ratio of approximately 1.11 are recorded in the residual vector. Subsequently, a joint loss function is constructed through fitting operations for the resulting set of residual vectors, which comprehensively considers the three error evaluation aspects of coincidence, offset, and size ratio. The overlap function is based on the intersection over union (IoU) ratio, evaluating the ratio of the intersection area to the union area between the predicted and true boxes to reflect the degree of overlap. A low overlap indicates that increased gradient updates are needed to bring the predicted and true boxes closer together. The offset function focuses on the deviation in center point coordinates, penalizing offset errors by mapping the absolute or squared horizontal and vertical offsets to a loss domain. The size ratio function calculates the deviation in the ratio of the width and height of the predicted box to the true box and converts this difference into a gradient signal to correct for inaccuracies in the predicted box's length and width. Combining these three components with weights or adaptive weights to form a joint loss function not only prevents a single metric from overly dominating the regression process but also dynamically balances the errors based on the deviations of specific samples, ensuring that the network optimizes both spatial position and size information. Subsequently, backpropagation training of the regression branch using this joint loss function allows gradient descent updates of the network parameters. The regression branch network employs a multi-layer convolutional structure, whose weights generate the coordinate offset and size scaling factor of the candidate box during forward reasoning. Backpropagation, based on the output gradient of the joint loss function, transmits the error signal back to each convolution kernel and bias term, thereby adjusting the filter's response weights and translation bias. During each iteration, parameter updates are combined with momentum or adaptive learning rate strategies (such as the Adam optimizer) to accelerate convergence and reduce oscillations. Through multiple consecutive iterations, a regression network capable of effectively predicting deviation correction values is formed. Finally, after the regression network is trained, the initially predicted coordinates can be corrected based on the trained regression correction parameter set.This correction involves adding the offset output by the network directly to the original center coordinates, and multiplying the width and height by the scaling factor output by the network, thereby generating a regression-corrected bounding box that corresponds one-to-one with the preliminary prediction box. The corrected bounding box position more accurately covers the area where the actual defect is located, and the scaled size is more consistent with the target entity. This correction not only improves positioning accuracy, but also significantly reduces the geometric mismatch between the predicted box and the true value box. This results in a complete set of regression-corrected bounding boxes, providing high-precision spatial information support for subsequent defect type determination or fine measurement.

[0033] It should be noted that when the regression network optimizes the prediction box, a loss metric is needed that can evaluate the overlap between the prediction box and the true value box and comprehensively consider the center offset and size deviation. With the true value box Intersection-over-Union Converted into basic penalty item—— , to produce a larger gradient when the overlap is low. At the same time, the normalized Euclidean distance d and the geometric mean scaling factor s of the aspect ratio are defined for the center point offset and aspect ratio error, respectively, and they are integrated into an exponential modulation factor in a nonlinear coupling manner. Finally, the joint spatial consistency loss function (JSCL) is defined as: , in, and is a hyperparameter used to adjust the relative impact of center offset and size deviation on the overall loss. Specifically, first, the intersection-union ratio Measure the overlap between the prediction box and the true value box, and take its natural logarithm As a basic penalty term, it gives a larger gradient when the overlap is low. At the same time, the exponential factor The two geometric errors of center offset and scale scaling are coupled. In the back propagation process of the regression network, each predicted box first calculates the IoU to obtain the logarithmic loss, and then calculates the center offset d and the scale deviation (s-1) and calculates them according to the hyperparameters , After weighting, the network is subjected to an exponential function, so that even slight displacement or scaling will exponentially amplify or suppress the IoU loss. This multiplication differs from a simple weighted summation. Instead, it uses an "amplification-suppression" mechanism to automatically reduce the penalty for noise deviation when the overlap is high, while rapidly amplifying the gradient when the error is large, pushing the network to focus on correcting difficult-to-align areas. This allows the network to more finely correct the spatial position and size of each predicted box during the iterative process, generating a set of regressed corrected bounding boxes that correspond one-to-one with the initial detection results and have more precise geometric alignment.

[0034] Step 150 determines the defect type of the precision gear tooth surface based on the spatial position information of the regression-corrected bounding box set and the corresponding category label.

[0035] Specifically, this embodiment performs a comprehensive analysis based on the spatial location information within the corrected bounding box set and the corresponding defect category labels. A topological data structure is constructed by combining spatial features such as the bounding box's coordinate center, area, and density distribution. Predefined spatial similarity templates are then introduced to match the results. By comparing the spatial distribution patterns of historical defect types, the spatial behavior characteristics of each defect type on the gear surface are determined, resulting in a final determination of the specific defect type within the precision gear. This completes the entire process from image detection to type attribution, ensuring refined, traceable, and structured detection results.

[0036] In an optional implementation of this embodiment, the step of judging the defect type of the precision gear tooth surface based on the spatial position information of the regression corrected bounding box set and the corresponding category label includes: constructing a topological relationship matrix of the spatial distribution characteristics of the bounding box based on the geometric center coordinates and size parameters of each bounding box in the regression corrected bounding box set; performing density clustering analysis on the topological relationship matrix through a preset clustering radius threshold to generate a defect space cluster set covering different areas of the tooth surface; determining a candidate defect type subset based on the density distribution characteristics of each defect space cluster in the defect space cluster set and the statistical frequency of the corresponding category label; and generating a defect type judgment result for the precision gear tooth surface by determining the spatial similarity between the spatial distribution pattern of each defect type in the candidate defect type subset and the historical defect pattern in the preset defect type template library.

[0037] Specifically, the geometric center coordinates and size parameters of each bounding box obtained after regression correction are extracted to construct spatial distribution features. Each bounding box's center point is used as a node, and the corresponding width and height are used as node attributes. Next, a topological relationship matrix is generated by calculating the Euclidean distance between any two nodes and incorporating the size information. Each element of this matrix represents the spatial proximity between pairs of nodes. Pairs with smaller distances or similar sizes have larger values, while pairs with larger distances or significantly different sizes have smaller values. This matrix not only reflects the relative distribution of defect boxes on the tooth surface but also accommodates the detection requirements of different defect scales. Subsequently, a density clustering analysis is performed based on the topological relationship matrix to group bounding boxes that are spatially concentrated into the same cluster. Density clustering algorithms (such as DBSCAN) use a cluster radius threshold and a minimum number of neighbors to identify a set of nodes with high spatial topological weight as a spatial cluster. The algorithm first examines the number of neighboring nodes in the topological matrix for each node. If the number of neighbors exceeds a preset threshold, the node is labeled as a core node and all its neighbors are included in the same cluster. The algorithm then iteratively explores the neighboring nodes. Through continuous density expansion, several spatial clusters of defects covering different regions of the entire tooth surface are eventually formed. For example, if multiple scratches on a gear meshing surface are closely spaced or distributed in a striped pattern, they are clustered into the same spatial cluster and thus treated as a holistic entity in subsequent analysis. Next, statistics are collected on the internal density distribution characteristics of each spatial cluster and the frequency of class labels corresponding to each bounding box within the cluster to determine a subset of candidate defect types. Density distribution characteristics can be estimated by the average proximity of nodes within the cluster or the cluster radius, while class label frequencies are determined by the labeling output by the classification branch. After sorting the label frequencies, the most frequently occurring classes can be selected as candidate subsets to account for the diverse range of possible defect types. For example, if 12 of the 20 bounding boxes within a cluster are labeled "crack" and 6 are labeled "peel," the candidate subset may include both "crack" and "peel," allowing for further differentiation through subsequent spatial pattern analysis. Finally, the spatial distribution pattern of the candidate defect type subset needs to be compared with the pre-established defect type template library for spatial similarity to finally determine the tooth surface defect type. Each defect type in the template library corresponds to a historical pattern sample set, which can be composed of the cluster layout, directional preference and cluster shape statistical characteristics of known defects on the gear. During the comparison process, the similarity score between the candidate cluster and the template sample is calculated based on multi-dimensional indicators such as cluster center position, directional distribution, cluster shape diffusion range, etc., such as the cross-overlap degree or cluster center angle deviation measurement method based on position mapping. The template with the highest similarity score represents the final defect type determination result corresponding to the spatial cluster. For example, if the directional preference of a cluster is highly consistent with the crack pattern in the template library in terms of angle, length distribution and cluster width, all bounding boxes in the cluster are uniformly judged as "cracks".

[0038] In an optional implementation of this embodiment, a three-dimensional topological map of the tooth surface defect is generated based on the coordinate information of the regression corrected bounding box set and the three-dimensional tooth surface model of the precision gear; through the correlation relationship between the three-dimensional topological map and the gear meshing motion parameters, dynamic stress simulation is performed on the defect area to obtain stress distribution data of the defect area; based on the stress distribution data and a preset defect evolution model, the expansion path of the defect during the gear operation is predicted to generate a defect expansion trajectory prediction map; through the correlation analysis of the defect expansion trajectory prediction map and the fatigue life parameters of the gear material, a safety level assessment result of the tooth surface defect is generated.

[0039] Specifically, the regressed and corrected 2D bounding box coordinates are mapped back to the gear's 3D geometric model to construct a 3D topological map of the tooth surface defects. This mapping is accomplished by back-projecting the center point coordinates and dimensions of each bounding box in the 3D CAD model to specific vertices or facets on the 3D surface mesh. At this point, the four corners or center lines of the bounding box are matched one-to-one with the mesh nodes, generating closed polygonal regions distributed along the curvature of the tooth surface. By fusing the projection results of all candidate defect bounding boxes, a set of closed curves or facets can be drawn on the 3D tooth surface, forming a complete 3D topological structure. This 3D topological map not only preserves the defect's position in the planar image but also reflects its spatial distribution characteristics within the gear's physical form, achieving a seamless transition from image to physical form. The 3D topological map is then combined with the meshing motion parameters of the gear under actual operating conditions to perform dynamic stress simulation of the defect area. First, parameters such as the radial load, torque, and speed during gear meshing must be set. Finite element analysis (FEA) software is then used to create a mesh-refined finite element model of the entire gear or tooth surface area. A higher-density mesh (MeshRefinement) is used in the defect area to improve calculation accuracy. During the simulation, the contact line of the gear meshing is used as the load application area, and a combined working condition of pulsating contact force and torsional torque is applied, while considering parameters such as the elastic modulus and Poisson's ratio of the material. Through time-series iterative calculations, dynamic stress distribution data can be obtained, including principal stress, shear stress, and contact stress peak. Subsequently, based on the dynamic stress distribution data and the preset defect evolution model, the expansion path of the tooth surface defect during operation is predicted. The evolution model employs a fatigue crack growth formula based on the Paris law. Its core concept is the power function relationship between the crack growth rate and the stress intensity factor range ΔK\Delta KΔK. By calculating the stress intensity factor at the crack tip at each moment and iteratively updating the crack length, the crack growth trajectory under cyclic loading can be determined. The evolution model also incorporates a path deflection factor to account for the influence of tooth surface curvature on crack propagation direction. For example, the principal stress direction at each crack tip is vector-aligned with the local tangent direction of the mesh to determine the deflection angle of the propagation branch. After the simulation is complete, a three-dimensional crack trajectory is generated, simulating the crack's progressive extension along a spiral path along the tooth surface. The crack length and propagation angle at each node are sampled, ultimately forming a defect propagation trajectory prediction map, which visualizes the crack evolution path and propagation rate. Finally, the defect propagation trajectory prediction map is correlated with the fatigue life parameters of the gear material to generate a safety level assessment for the tooth surface defect.Fatigue life parameters include the material's fatigue limit, fatigue strength exponent, and the cycle threshold required for the Miner's Rule (Fatigue Damage Accumulation Method). The crack length increment in the trajectory prediction is compared with the critical value of the dangerous crack length. The cumulative damage value is calculated based on the actual number of operating cycles, and the risk level is assessed according to the C, B, and A safety grading standards. For example, when the cumulative damage value exceeds 0.5, it can be defined as a Class B warning state, and maintenance is recommended during the next maintenance cycle. When the crack length prediction reaches more than 80% of the material's critical crack length and the cumulative damage value exceeds 0.7, it can be assessed as a Class C high-risk state, requiring emergency shutdown and maintenance.

[0040] According to the method for detecting tooth surface defects of precision gears provided by the present application, an original image sequence covering the entire circumference of the precision gear tooth surface is obtained; a set of defect candidate regions is generated by preprocessing the original image sequence; a preliminary detection result of the precision gear tooth surface is determined by inputting a gear image of a fixed size in each region of the defect candidate region set into a deep neural network; a joint constrained regression is performed between the predicted coordinates of the defect bounding box in the preliminary detection result and the corresponding true value box coordinates to generate a regression-corrected bounding box set corresponding to the preliminary detection result; the defect type of the precision gear tooth surface is determined based on the spatial position information of the regression-corrected bounding box set and the corresponding category label. By establishing an index relationship between the image sequence and the gear rotation angle, accurate matching of image acquisition and tooth position is achieved, and the accuracy of defect detection is effectively improved by coordinating the deep neural network with the regression correction mechanism.

[0041] Figure 2 The embodiment of the present application provides a device for detecting tooth surface defects of precision gears, which can be used to implement the method for detecting tooth surface defects of precision gears in the aforementioned embodiment. Figure 2 As shown, the precision gear tooth surface defect detection device mainly includes: An acquisition module 10 is used to acquire a sequence of original images covering the entire circumference of the precision gear tooth surface; The processing module 20 is used to generate a defect candidate region set by preprocessing the original image sequence; a determination module 30 for determining a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each region of the defect candidate region set into a deep neural network; A generation module 40 is configured to perform a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box to generate a set of regression-corrected bounding boxes corresponding to the preliminary detection result; The judgment module 50 is used to judge the defect type of the precision gear tooth surface according to the spatial position information of the regression corrected bounding box set and the corresponding category label.

[0042] In an optional implementation of this embodiment, the acquisition module is specifically used to: perform step angle control on the precision gear according to a preset data acquisition path to generate a rotation angle sequence covering the tooth surface; perform image acquisition on the specified angle position of the precision gear according to the rotation angle sequence to obtain the original image frame corresponding to the specified angle position; construct a mapping table between the original image frame and the gear angle position information according to the shooting angle identifier and timestamp information of the original image frame; bind the mapping table to the original image frame to generate an image index structure set; and generate an original image sequence covering the entire circumference of the precision gear tooth surface by sequentially arranging the image frames in the image index structure set.

[0043] In an optional implementation of the present embodiment, the generation module is specifically used to: obtain a basic grayscale image set corresponding to each image frame by performing grayscale processing and median filtering noise reduction operations on each image frame in the original image sequence; generate an enhanced image set by performing multi-scale image enhancement processing and local contrast enhancement processing on the basic grayscale image set; determine a reference image for differential analysis based on the pixel grayscale distribution characteristics of the enhanced image set, and generate a difference image set by calculating the difference matrix between the reference image and the enhanced image; extract image areas with closed edge contours by performing threshold segmentation on the difference image set, and generate a defect candidate area set.

[0044] In an optional implementation of this embodiment, the determination module is specifically used to: generate a fixed-size image subset that meets preset input specifications by performing a size normalization operation on all image areas of the defect candidate area set; perform multi-scale feature extraction and adaptive convolution on the fixed-size image subset in a deep neural network to generate a feature expression graph set; regress the target position parameters according to the spatial feature distribution extracted from the feature expression graph set to obtain a defect bounding box set corresponding to all gear images; and generate preliminary detection results of the precision gear tooth surface by classifying the feature vectors corresponding to each bounding box in the defect bounding box set.

[0045] In an optional implementation of this embodiment, the processing module is specifically used to: construct a multidimensional residual vector set based on the predicted coordinates of each defect bounding box in the preliminary detection results and its corresponding true value box coordinates; construct a joint loss function by performing function fitting on the multidimensional residual vector set; wherein the joint loss function is composed of a bounding box coincidence function, a boundary offset function and a size ratio function; perform backpropagation training on the regression network according to the joint loss function to obtain a bounding box regression correction parameter set corresponding to the preliminary detection results; perform numerical correction on the predicted coordinates according to the bounding box regression correction parameter set to generate a regression corrected bounding box set that corresponds one-to-one to the preliminary detection results.

[0046] In an optional implementation of this embodiment, the judgment module is specifically used to: construct a topological relationship matrix of the spatial distribution characteristics of the bounding box based on the geometric center coordinates and size parameters of each bounding box in the regression corrected bounding box set; perform density clustering analysis on the topological relationship matrix through a preset clustering radius threshold to generate a defect space cluster set covering different areas of the tooth surface; determine a subset of candidate defect types based on the density distribution characteristics of each defect space cluster in the defect space cluster set and the statistical frequency of the corresponding category label; generate a defect type judgment result for the precision gear tooth surface by determining the spatial similarity between the spatial distribution pattern of each defect type in the candidate defect type subset and the historical defect pattern in the preset defect type template library.

[0047] In an optional implementation of this embodiment, the generating module is further configured to: generate a three-dimensional topological map of the tooth surface defect based on the coordinate information of the regression corrected bounding box set and the three-dimensional tooth surface model of the precision gear; Through the correlation between the three-dimensional topology map and the gear meshing motion parameters, dynamic stress simulation is performed on the defect area to obtain stress distribution data of the defect area; Based on stress distribution data and the preset defect evolution model, the defect expansion path during gear operation is predicted and a defect expansion trajectory prediction diagram is generated; The safety level assessment results of tooth surface defects are generated by correlating the defect expansion trajectory prediction diagram with the fatigue life parameters of the gear material.

[0048] According to the present invention, a device for detecting tooth surface defects of precision gears is provided. The device obtains an original image sequence covering the entire circumference of the tooth surface of the precision gear; generates a set of defect candidate regions by preprocessing the original image sequence; inputs a fixed-size gear image of each region in the defect candidate region set into a deep neural network to determine the preliminary detection results of the precision gear tooth surface; performs a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection results and the corresponding true value frame coordinates to generate a set of regression-corrected bounding boxes corresponding to the preliminary detection results; and determines the defect type of the precision gear tooth surface based on the spatial position information of the regression-corrected bounding box set and the corresponding category label. By establishing an index relationship between the image sequence and the gear rotation angle, accurate matching of image acquisition and tooth position is achieved, and the deep neural network and regression correction mechanism are coordinated to effectively improve the accuracy of defect detection.

[0049] According to the application plan Figure 3 An electronic device provided in an embodiment of the present application can be used to implement the method for detecting tooth surface defects of precision gears in the aforementioned embodiment, mainly comprising: Memory 301, processor 302, and computer program 303 stored in memory 301 and executable on processor 302. Memory 301 and processor 302 are connected via communication. When processor 302 executes computer program 303, the method for detecting tooth surface defects of precision gears described in the aforementioned embodiment is implemented. The number of processors may be one or more.

[0050] The memory 301 can be a high-speed random access memory (RAM) memory or a non-volatile memory such as a disk drive. The memory 301 is used to store executable program code. The processor 302 is coupled to the memory 301 .

[0051] Furthermore, the embodiment of the present application also provides a computer-readable storage medium, which can be provided in the electronic device in the above embodiments. The computer-readable storage medium can be the above Figure 3 Memory in the illustrated embodiment.

[0052] The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method for detecting tooth surface defects of precision gears in the aforementioned embodiment. Furthermore, the computer-readable storage medium may be any medium capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), RAM, a magnetic disk, or an optical disk.

[0053] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0054] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.

[0055] As described above, the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit them. Although the present application has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting tooth surface defects of precision gears, characterized in that: include: Obtaining a sequence of original images covering the entire circumference of the precision gear tooth surface; Generating a defect candidate region set by preprocessing the original image sequence; Determining a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each region in the defect candidate region set into a deep neural network; Performing joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box to generate a regression corrected bounding box set corresponding to the preliminary detection result; The defect type of the precision gear tooth surface is determined according to the spatial position information of the regression corrected bounding box set and the corresponding category label.

2. The method for detecting tooth surface defects of precision gears according to claim 1, characterized in that: The step of obtaining a sequence of original images covering the entire circumference of the precision gear tooth surface includes: Perform step angle control on precision gears according to the preset data acquisition path to generate a rotation angle sequence covering the tooth surface; Performing image acquisition at a specified angle position of the precision gear according to the rotation angle sequence to obtain an original image frame corresponding to the specified angle position; Constructing a mapping table between the original image frame and the gear angle position information according to the shooting angle identifier and the timestamp information of the original image frame; Binding the mapping table with the original image frame to generate an image index structure set; By sequentially arranging the image frames in the image index structure set, an original image sequence covering the entire circumference of the precision gear tooth surface is generated.

3. The method for detecting tooth surface defects of precision gears according to claim 1, characterized in that: The step of generating a set of defect candidate regions by preprocessing the original image sequence includes: Obtaining a set of basic grayscale images corresponding to each image frame by performing grayscale processing and median filtering noise reduction operations on each image frame in the original image sequence; Generate an enhanced image set by performing multi-scale image enhancement processing and local contrast improvement processing on the basic grayscale image set; Determining a reference image for differential analysis based on pixel grayscale distribution characteristics of the enhanced image set, and generating a difference image set by calculating a difference matrix between the reference image and the enhanced image; By performing threshold segmentation on the difference map set, image regions with closed edge contours are extracted to generate a defect candidate region set.

4. The method for detecting tooth surface defects of precision gears according to claim 3, characterized in that: The step of determining the preliminary detection result of the precision gear tooth surface by inputting the gear image of each region of the defect candidate region set with a fixed size into a deep neural network comprises: Generate a fixed-size image subset that meets preset input specifications by performing a size normalization operation on all image regions of the defect candidate region set; Performing multi-scale feature extraction and adaptive convolution on the fixed-size image subset in a deep neural network to generate a feature expression graph set; Performing regression processing on target position parameters according to the spatial feature distribution extracted from the feature expression graph set to obtain a set of defect bounding boxes corresponding to all gear images; A preliminary detection result of the precision gear tooth surface is generated by classifying the feature vector corresponding to each bounding box in the defect bounding box set.

5. The method for detecting tooth surface defects of precision gears according to claim 1, characterized in that: The step of performing joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the corresponding true value box coordinates to generate a regression corrected bounding box set corresponding to the preliminary detection result includes: Constructing a multidimensional residual vector set based on the predicted coordinates of each defect boundary box in the preliminary detection results and the corresponding true value frame coordinates; Constructing a joint loss function by performing function fitting on the multidimensional residual vector set; wherein the joint loss function is composed of a bounding box coincidence function, a boundary offset function, and a size ratio function; Performing back-propagation training on the regression network according to the joint loss function to obtain a bounding box regression correction parameter set corresponding to the preliminary detection result; The predicted coordinates are numerically corrected according to the bounding box regression correction parameter set to generate a set of regression corrected bounding boxes that correspond one-to-one to the preliminary detection results.

6. The method for detecting tooth surface defects of precision gears according to claim 1, characterized in that: The step of determining the defect type of the precision gear tooth surface according to the spatial position information of the regression corrected bounding box set and the corresponding category label includes: Constructing a topological relationship matrix of spatial distribution features of the bounding boxes based on the geometric center coordinates and size parameters of each bounding box in the regression-corrected bounding box set; Performing density cluster analysis on the topological relationship matrix by presetting a cluster radius threshold to generate a set of defect space clusters covering different areas of the tooth surface; Determining a candidate defect type subset based on density distribution characteristics of each defect space cluster in the defect space cluster set and statistical frequencies of corresponding category labels; The defect type determination result of the precision gear tooth surface is generated by determining the spatial similarity between the spatial distribution pattern of each defect type in the candidate defect type subset and the historical defect pattern in the preset defect type template library.

7. The method for detecting tooth surface defects of precision gears according to claim 1, characterized in that: The method further comprises: generating a three-dimensional topological map of tooth surface defects based on the coordinate information of the regression corrected bounding box set and the three-dimensional tooth surface model of the precision gear; By means of the correlation between the three-dimensional topological map and the gear meshing motion parameters, dynamic stress simulation is performed on the defective area to obtain stress distribution data of the defective area; Based on the stress distribution data and a preset defect evolution model, the defect expansion path during the gear operation is predicted to generate a defect expansion trajectory prediction diagram; By performing correlation analysis on the defect expansion trajectory prediction diagram and the fatigue life parameters of the gear material, a safety level assessment result of the tooth surface defect is generated.

8. A device for detecting tooth surface defects of precision gears, characterized in that: The precision gear tooth surface defect detection device comprises: An acquisition module is used to acquire a sequence of original images covering the entire circumference of the precision gear tooth surface; A processing module, configured to generate a set of defect candidate regions by preprocessing the original image sequence; a determination module, configured to determine a preliminary detection result of the precision gear tooth surface by inputting a gear image of a fixed size in each region in the defect candidate region set into a deep neural network; a generation module, configured to perform a joint constrained regression between the predicted coordinates of the defect bounding box in the preliminary detection result and the coordinates of the corresponding true value box, to generate a regression-corrected bounding box set corresponding to the preliminary detection result; A judgment module is used to judge the defect type of the precision gear tooth surface according to the spatial position information of the regression corrected bounding box set and the corresponding category label.

9. An electronic device, characterized in that: Comprising a memory and a processor, wherein: The processor is configured to execute a computer program stored in the memory; When the processor executes the computer program, the steps of the method for detecting tooth surface defects of precision gears according to any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method for detecting tooth surface defects of precision gears according to any one of claims 1 to 7 are implemented.

Images