WO2022222120A1

WO2022222120A1 - Bearing three-dimensional defect detection method and system

Info

Publication number: WO2022222120A1
Application number: PCT/CN2021/089130
Authority: WO
Inventors: 徐刚; 赵明; 肖江剑; 许根; 王菊
Original assignee: 中国科学院宁波材料技术与工程研究所
Priority date: 2021-04-23
Filing date: 2021-04-23
Publication date: 2022-10-27

Abstract

The present application discloses a bearing three-dimensional defect detection method and system. The method comprises: under the action of a combined light source and a rotating platform, acquiring N two-dimensional bearing images to be detected; preprocessing said N bearing image to obtain 1/N preprocessed images for defect detection; and inputting said 1/N preprocessed images into a pre-trained N-channel deep learning model to obtain a defect detection result. According to the present application, by means of the cooperation of a 2D camera and the rotating platform, and by replacing a current expensive 3D camera detection method with a detection algorithm for a YOLO network structure improved by a multi-view, multi-channel and multi-attention mechanism, detection accuracy and detection efficiency are improved, and system costs are reduced.

Description

A kind of bearing three-dimensional defect detection method and system

technical field

The application belongs to the technical field of machine vision defect detection, and in particular relates to a bearing three-dimensional defect detection method and system.

Background technique

Bearings are important basic components of various mechanical equipment, and their accuracy, performance, life and reliability play a decisive role in the accuracy, performance, life and reliability of the host. Among mechanical products, bearings are high-precision products, which not only require comprehensive support from mathematics, physics and many other disciplines, but also require material science, heat treatment technology, precision machining and measurement technology, numerical control technology, effective numerical methods and powerful computers. Technology and many other disciplines serve it, so the bearing is a product that represents the national scientific and technological strength.

Especially for small bearings, the appearance defect detection usually has defects that cannot be extracted from a single direction, especially scratches, pits, burrs and other three-dimensional features. It is difficult to distinguish from one direction, resulting in the existing 2D defects. Bearing inspection equipment or systems cannot fully detect the three-dimensional defects of bearings; and most of the existing 3D cameras are used for this type of defects, but because 3D cameras rely on imports and are expensive, they cannot meet the actual application needs of enterprises.

Therefore, how to provide a bearing detection scheme that can solve the problems existing in the three-dimensional defect detection technology faced by the existing bearing manufacturers under the premise of low cost is an urgent problem to be solved.

SUMMARY OF THE INVENTION

The main purpose of the present application is to provide a bearing three-dimensional defect detection method and system, so as to overcome the problems of the prior art bearing surface three-dimensional defect detection technology, the existing 2D camera solution has low detection accuracy and efficiency, and the 3D sensor solution is expensive.

In order to achieve the aforementioned purpose of the invention, the technical solution adopted in this application includes: a bearing three-dimensional defect detection method, the method includes:

S100, the bearing to be detected is fixed on the rotating platform, and under the corresponding combined light source, the rotating platform rotates according to the set rotation step length, and the bearing surface is photographed at N angles while rotating to obtain N pieces of the bearing surface to be detected. Two-dimensional bearing image, where N is a natural number greater than or equal to 3;

S200, preprocessing the N bearing images to be detected to obtain 1/N preprocessing images for defect detection;

S300: Input the 1/N preprocessed images into a pre-trained N-channel deep learning model to obtain a defect detection result.

In a preferred embodiment, in S100, for bearings of different specifications, the 2D image acquisition module selects different combined light sources, and uses a rotating platform to set different rotation step sizes, and the combined light source adjusts the bearings according to the rotation step sizes. The surface is photographed at N angles under the light source, and the rotation step = 360°/N.

In a preferred embodiment, in S200, the methods used in the preprocessing are one or more of filtering, character template positioning, channel separation and fusion, and splicing and matching methods.

In a preferred embodiment, the method further includes:

S400, store the defect detection result and 1/N preprocessed images, and optimize and modify the deep learning model according to the stored defect detection results and 1/N preprocessed images.

In a preferred embodiment, the method further includes:

S10, train a YOLO multi-channel neural network to obtain an N-channel detection model.

In a preferred embodiment, the 10 includes:

S11, label the 1/N preprocessed images obtained in step S200 as training images;

S12, inputting the training image into a multi-channel feature extraction network, and using the feature extraction network to extract high-level feature information of the image to obtain a high-level feature image of the bearing to be tested;

S13, a multi-attention mechanism is introduced to enhance the weight of the defective parts of the bearing, and the attention mechanism includes a spatial attention mechanism, a channel attention mechanism and a convolution kernel adaptive selection mechanism;

S14, performing differential processing between the high-level feature image of the bearing to be tested and the feature image of the same layer of the standard bearing to obtain a high-level feature differential image of the bearing to be tested;

S15, the high-level feature difference image is input into the detection network to detect the defect target, and finally the improved YOLO detection model is output to step S300.

In a preferred embodiment, the S400 includes:

S16, the stored defect detection results and 1/N preprocessed images are handed over to the re-inspection workers for re-inspection, and the detection result maps with detection errors are eliminated;

S17, input the image with the correct defect detection result and the corresponding annotation data into the improved YOLO detection network for retraining, optimize and modify the parameters of the deep learning model, improve the accuracy of the defect target, and greatly reduce the number of manually annotated samples.

The embodiment of the present application also provides a 2D image acquisition module, which includes:

Rotating platform, used to fix the bearing to be tested and used for 360-degree rotation of the bearing in any step length;

Combined light source, used to provide the light source required for image acquisition;

The 2D camera and lens are used to photograph the bearing surface at N angles while the rotating platform drives the bearing to rotate and the combined light source illuminates, to obtain N two-dimensional bearing images to be inspected.

The embodiment of the present application also provides a bearing three-dimensional defect detection system, which includes:

The 2D image acquisition module is used to fix the bearing to be detected on the rotating platform. Under the corresponding combined light source, the rotating platform rotates according to the set rotation step, and the bearing surface is photographed at N angles while rotating to obtain N two-dimensional bearing images to be detected, where N is a natural number greater than or equal to 3;

an image preprocessing module, used for preprocessing the N bearing images to be detected to obtain 1/N preprocessing images for defect detection;

The image defect detection module is used for inputting the 1/N preprocessed images into a pre-trained N-channel target detection model to obtain defect detection results.

In a preferred embodiment, the 2D image acquisition module includes:

In a preferred embodiment, the combined light source adopts a set of adjustable coaxial light sources and ring light sources.

In a preferred embodiment, the method used by the image preprocessing module is one or more of filtering, character template positioning, channel separation and fusion, and splicing and matching methods.

In a preferred embodiment, the system further includes: a deep learning model optimization module for storing the defect detection results and 1/N preprocessed images, and according to the stored defect detection results and 1/N images Preprocess the image, optimize and modify the deep learning model, and the optimization and modification of the deep learning model include:

The stored defect detection results and 1/N pre-processed images are handed over to the re-inspection workers for re-inspection, and the detection result pictures with detection errors are eliminated;

Input the images with correct defect detection results and the corresponding labeled data into the improved YOLO detection network for retraining, optimize and modify the parameters of the deep learning model, improve the accuracy of the defect target, and greatly reduce the number of manually labeled samples.

In a preferred embodiment, the system further includes: a deep learning model training module connected between the image preprocessing module and the image defect detection module, and the deep learning model training module includes:

The multi-channel image loading module is connected to the image preprocessing module, and is used for loading the 1/N preprocessed images obtained by the image preprocessing module into the N-channel feature extraction network, so as to extract high-level image feature information;

The image high-level feature information extraction module is connected with the image loading module, and is used to extract the high-level feature information of the preprocessed image to obtain the high-level feature image of the bearing to be tested;

The multi-attention mechanism module is connected to the image high-level feature information extraction module and is used to enhance the weight of the defective parts of the bearing in the process of model training. The attention mechanism includes a spatial attention mechanism, a channel attention mechanism and a convolution kernel. Adaptive selection mechanism;

The high-level feature image difference module is connected to the multi-attention mechanism module, and is used to differentiate the high-level feature image of the bearing to be tested and the feature image of the same layer of the standard bearing to obtain a high-level feature difference image;

The defect detection module is connected to the high-level feature image difference module, and is used for inputting the high-level feature difference image to the detection network to detect the defect target, and finally output the improved YOLO detection model to the image defect detection module.

Compared with the prior art, the beneficial effect of the present application is at least as follows: the present application uses the 2D camera and the rotating platform to cooperate, and uses the multi-view multi-channel multi-attention mechanism module to improve the detection algorithm of the YOLO network structure, replacing the current expensive detection algorithm. 3D camera detection method, improve detection accuracy and detection efficiency, and reduce system cost.

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments described in this application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

Fig. 1 is the overall flow schematic diagram of the detection method in one embodiment of the present application;

2 is a structural block diagram of a detection system in an embodiment of the present application;

3 is a schematic structural diagram of a 2D image acquisition module in an embodiment of the present application;

FIG. 4 is an image acquired by setting a 90° step size on a rotating platform in an embodiment of the present application;

FIG. 5 is an image of the detection result of three-dimensional defects on the bearing surface according to an embodiment of the present application.

Detailed ways

The application will be more fully understood from the following detailed description, which should be read in conjunction with the accompanying drawings. Detailed embodiments of the present application are disclosed herein; however, it is to be understood that the disclosed embodiments are merely exemplary of the present application, which may be embodied in various forms. Therefore, specific functional details disclosed herein are not to be interpreted as limiting, but merely as a basis for the claims and for teaching one skilled in the art to vary in virtually any suitable detailed embodiment. The approach adopts the representative basis of this application.

A bearing three-dimensional defect detection method and system disclosed in the embodiments of the present application relate to the technical field of industrial inspection, and in particular to a multi-view three-dimensional appearance defect detection method and system for a small bearing based on a 2D camera and a rotating platform. Including but not limited to pits, scratches, burrs, etc.

With reference to FIG. 1 and FIG. 3 , a bearing three-dimensional defect detection method disclosed in the embodiment of the present application includes the following steps:

S100, the bearing to be detected is fixed on the rotating platform, and under the corresponding combined light source, the rotating platform rotates according to the set rotation step length, and the bearing surface is photographed at N angles while rotating to obtain N pieces of the bearing surface to be detected. Two-dimensional bearing image, where N is a natural number greater than or equal to 3.

Specifically, the bearing to be detected is fixed on the center of the rotating platform, and the rotating platform can drive the bearing to rotate 360 degrees in any step length. In this embodiment, the rotation step length of the rotating platform=360°/N, where N is a natural number greater than or equal to 3.

And the combined light source is used to illuminate the bearing on the rotating platform. In this embodiment, the combined light source adopts a set of adjustable coaxial light sources and ring light sources, wherein the coaxial light source is located above the rotating platform and is connected to the rotating platform and the bearing. All are arranged coaxially, and the annular optical axis is located above the side of the rotating platform, and is arranged at a certain inclination angle with the top plane of the rotating platform. In this application, different combined light sources are selected for bearings of different specifications. During implementation, the combined light source is controlled by a light source controller (not shown in the figure), and the combined light source may be a combination of two or more light sources among parallel light, coaxial light, and ring light.

The rotating platform rotates according to the set step length. While rotating, the bearing surface on the rotating platform of the light source is combined to illuminate, and the bearing is photographed according to the rotation step set by the rotating platform to complete the N angles of the bearing surface under the light source. Obtain N two-dimensional bearing images to be detected, if N is 4, the rotation step of the rotating platform is 90°, and the rotating platform rotates one circle to achieve 4 angles of the bearing surface, as shown in Figure 4, as Turn the platform to set the acquired image in 90° steps. In this embodiment, all 2D cameras are used to photograph the bearings, and the cost is lower than that of 3D cameras.

S200: Preprocess N bearing images to be detected to obtain 1/N preprocessed images for defect detection.

Specifically, during implementation, the preprocessing method of the bearing image in the present application may be one or more of filtering, character template positioning, channel separation and fusion, and splicing and matching methods.

S300: Input 1/N preprocessed images into a pre-trained deep learning model of N channels to obtain a defect detection result.

Specifically, in this embodiment, the training process of the deep learning model specifically includes:

S13: Introduce a multi-attention mechanism to enhance the weight of the defective parts of the bearing. The attention mechanism includes a spatial attention mechanism, a channel attention mechanism, and a convolution kernel adaptive selection mechanism. Among them, the spatial attention mechanism and the channel attention mechanism The force mechanism sequentially infers the attention map along two independent dimensions (channel and space), and then multiplies the attention map with the input feature map for adaptive feature optimization. The convolution kernel adaptive selection mechanism can be adaptively adjusted according to the input Its receptive field size captures target objects with different scales;

As shown in FIG. 5 , it is an image of the detection result of three-dimensional defects on the bearing surface in an embodiment of the present application.

Specifically, in this embodiment, optimizing and modifying the deep learning model specifically includes:

S16, optimize and modify the deep learning model according to the stored defect detection results and 1/N preprocessed images, and according to the stored defect detection results and 1/N preprocessed images; Workers conduct re-inspection and eliminate the detection result map with detection error;

S17 , input the image with correct defect detection result and the corresponding labeling data into the improved YOLO detection network for retraining, optimize and modify the parameters of the deep learning model, improve the accuracy of the defect target, and greatly reduce the number of manually labeled samples.

Referring to FIG. 3 , an aspect of the embodiments of the present application further provides a 2D image acquisition module, the module includes:

In a preferred embodiment, the coaxial light source is located above the rotating platform, and is coaxially arranged with the rotating platform and the bearing.

In a preferred embodiment, the annular optical axis is located above the side of the rotating platform, and is arranged at a certain inclination angle to the top plane of the rotating platform.

In a preferred embodiment, the rotating platform is mounted on a fixed bracket, and the top platform thereof is horizontally arranged.

Please refer to FIG. 2 and FIG. 3 , corresponding to the above-mentioned method for detecting three-dimensional defects of bearings, another aspect of the embodiments of the present application further provides a three-dimensional defect detection system for bearings, which includes:

2D image acquisition module;

In some preferred embodiments, the 2D image acquisition module is used to fix the bearing to be detected on the rotating platform. Under the corresponding combined light source, the rotating platform rotates according to the set rotation step, and the bearing surface is rotated while rotating. Shooting at N angles is performed to obtain N two-dimensional bearing images to be detected, where N is a natural number greater than or equal to 3.

Specifically, in this embodiment, the 2D image acquisition module includes a rotating platform 10 mounted on a fixed bracket 20, a combined light source, a 2D camera and a lens 30, wherein the rotating platform 10 may specifically be a rotating platform driven by a stepping motor, The top platform of the rotating platform 10 is set horizontally, which can drive the bearing (not shown) to rotate 360 degrees in any step length. The bearing to be detected is fixed on the central bearing fixing base 11 on the rotating platform 10 . In this embodiment, the rotation step length of the rotating platform=360°/N, where N is a natural number greater than or equal to 3.

The combined light source is used to illuminate the bearing on the rotating platform 10 . In this embodiment, the combined light source adopts a set of adjustable coaxial light sources 40 and/or ring light sources 50 , wherein the coaxial light source 40 is located above the rotating platform 10 , and are arranged coaxially with the rotating platform 10 and the bearing. The annular optical axis 50 is located above the side of the rotating platform 10 and is arranged at a certain inclination angle with the top plane of the rotating platform 10 . In this application, different combined light sources are selected for bearings of different specifications. During implementation, the combined light source is controlled by a light source controller (not shown in the figure), and the combined light source may be a combination of two or more light sources among parallel light, coaxial light, and ring light.

In some preferred embodiments, the image preprocessing module is configured to preprocess the N bearing images to be inspected to obtain 1/N preprocessed images for defect inspection.

In this embodiment, the preprocessing method used by the image preprocessing module is one or more of filtering, character template positioning, channel separation and fusion, and splicing and matching methods.

In some preferred embodiments, the image defect detection module is configured to input the 1/N pre-processed images into a pre-trained N-channel deep learning model to obtain defect detection results.

Further, the bearing three-dimensional defect detection system further includes a deep learning model training module connected between the image preprocessing module and the image defect detection module, and the deep learning model training module includes:

The feature image difference module is connected with the multi-attention mechanism module, and is used to differentiate the high-level feature image of the bearing to be tested and the feature image of the same layer of the standard bearing to obtain a high-level feature difference image;

The defect detection module is connected with the feature image difference module, and is used for inputting the high-level feature difference image to the detection network to detect the defect target, and finally output the improved YOLO detection model to the image defect detection module.

The deep learning model optimization module is used to store the defect detection results and 1/N preprocessed images, and optimize and modify the deep learning model according to the stored defect detection results and 1/N preprocessed images.

The bearing three-dimensional defect detection method and system disclosed in the present application can solve the technical problems of high price, low detection accuracy and poor robustness of the system in the surface three-dimensional defect detection technology of existing bearing manufacturers; The technical solution of the present application can achieve higher detection accuracy and robustness under the premise of lower system costs acceptable to enterprises.

It should be understood that the above-mentioned embodiments are only intended to illustrate the technical concept and characteristics of the present application, and the purpose thereof is to enable those who are familiar with the technology to understand the content of the present application and implement accordingly, and cannot limit the protection scope of the present application. All equivalent changes or modifications made according to the spirit and spirit of this application shall be covered within the protection scope of this application.

Claims

A bearing three-dimensional defect detection method, characterized in that the method comprises:

S100, the bearing to be detected is fixed on the rotating platform, and under the corresponding combined light source, the rotating platform rotates according to the set rotation step length, and the bearing surface is photographed at N angles while rotating to obtain N pieces of the bearing surface to be detected. Two-dimensional bearing image, where N is a natural number greater than or equal to 3;

S200, preprocessing the N images of the bearing to be detected to obtain 1/N preprocessing images for defect detection;

S300: Input the 1/N preprocessed images into a pre-trained N-channel deep learning model to obtain a defect detection result.
A bearing three-dimensional defect detection method according to claim 1, characterized in that: in S100, for bearings of different specifications, the 2D image acquisition module selects different combined light sources, and uses a rotating platform to set different rotation step sizes, The combined light source photographs the bearing surface at N angles under the light source according to the rotation step size, where the rotation step size=360°/N.
A bearing three-dimensional defect detection method according to claim 1, characterized in that: in S200, the method used in the preprocessing is one of filtering, character template positioning, channel separation and fusion, and splicing and matching methods. or more.
A bearing three-dimensional defect detection method according to claim 1, characterized in that, the method further comprises: S400, storing the defect detection result and 1/N preprocessed images, and detecting the defect according to the stored defect Results and 1/N preprocessed images to optimize and modify deep learning models.
A bearing three-dimensional defect detection method according to claim 1, wherein the training process of the deep learning model comprises:

S11, label the 1/N preprocessed images obtained in step S200 as training images;

S12, inputting the training image into a multi-channel feature extraction network to extract high-level feature information of the image to obtain a high-level feature image of the bearing to be tested;

S13, in the process of model training, a multi-attention mechanism is introduced to enhance the weight of the defective parts of the bearing, and the multi-attention mechanism includes a spatial attention mechanism, a channel attention mechanism, and a convolution kernel adaptive selection mechanism;

S14, performing differential processing between the high-level feature image of the bearing to be tested and the feature image of the same layer of the standard bearing to obtain a differential feature image of the bearing to be tested;

S15, the differential feature image input detection network, carry out model training, and finally output the improved YOLO detection model to step S300.
A bearing three-dimensional defect detection system, characterized in that the system includes:

The 2D image acquisition module is used to fix the bearing to be detected on the rotating platform. Under the corresponding combined light source, the rotating platform rotates according to the set rotation step, and the bearing surface is photographed at N angles while rotating to obtain N two-dimensional bearing images to be detected, where N is a natural number greater than or equal to 3;

an image preprocessing module for preprocessing the N bearing images to be tested to obtain 1/N preprocessing images for defect detection;

The image defect detection module is used for inputting the 1/N preprocessed images into a pre-trained N-channel deep learning model to obtain defect detection results.
A bearing three-dimensional defect detection system according to claim 6, wherein the 2D image acquisition module comprises:

Rotating platform, used to fix the bearing to be tested and used for 360-degree rotation of the bearing in any step length;

Combined light source, used to provide the light source required for image acquisition;

The 2D camera and lens are used to photograph the bearing surface at N angles while the rotating platform drives the bearing to rotate and the combined light source illuminates, to obtain N two-dimensional bearing images to be inspected.
A bearing three-dimensional defect detection system according to claim 6, wherein the preprocessing method used by the image preprocessing module is one of filtering, character template positioning, channel separation and fusion, and splicing and matching methods. one or more.
A bearing three-dimensional defect detection system according to claim 6, characterized in that, the system further comprises: a deep learning model optimization module for storing the defect detection results and 1/N preprocessed images, and according to The stored defect detection results and 1/N preprocessed images optimize and modify the deep learning model, and the optimized and modified deep learning model includes:

The stored defect detection results and 1/N pre-processed images are handed over to the re-inspection workers for re-inspection, and the detection result pictures with detection errors are eliminated;

Input the images with correct defect detection results and the corresponding labeled data into the improved YOLO detection network for retraining, optimize and modify the parameters of the deep learning model, improve the accuracy of the defect target, and greatly reduce the number of manually labeled samples.
A bearing three-dimensional defect detection system according to claim 9, wherein the system further comprises: a deep learning model training module connected between the image preprocessing module and the image defect detection module, the deep learning model Training modules include:

The multi-channel image loading module is connected with the image preprocessing module, and is used for loading the 1/N preprocessed images obtained by the image preprocessing module into the N-channel feature extraction network, so as to perform image high-level feature extraction;

The image high-level feature information extraction module is connected with the image loading module, and is used to extract the high-level feature information of the preprocessed image to obtain the high-level feature image of the bearing to be tested;

The multi-attention mechanism module is connected to the image high-level feature information extraction module and is used to enhance the weight of the defective parts of the bearing in the process of model training. The attention mechanism includes a spatial attention mechanism, a channel attention mechanism and a convolution kernel. Adaptive selection mechanism;

The high-level feature image difference module is connected to the multi-attention mechanism module, and is used to differentiate the high-level feature image of the bearing to be tested and the feature image of the same layer of the standard bearing to obtain a high-level feature difference image;

The defect detection module is connected to the high-level feature image difference module, and is used for inputting the high-level feature difference image to the detection network to detect the defect target, and finally output the improved YOLO detection model to the image defect detection module.