SYSTEM FOR TRAINING DEEP LEARNING ALGORITHMS AND RELATED METHODS
The use of deep learning algorithms for defect detection in ceramic rolling elements addresses inefficiencies and human error by automating the classification process, improving production speed and accuracy.
Patent Information
- Application Number
- JP2024525899
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2021-11-08
- Filing Date
- 2022-10-26
- Publication Date
- 2024-11-19
AI Technical Summary
Existing methods for detecting defects in ceramic rolling elements, such as foreign inclusions and porosity, are inefficient and prone to human error, leading to production line slowdowns and potential misinterpretation, which can cause critical failures in mechanical systems.
A method and system using deep learning algorithms, specifically a convolutional neural network, to classify ceramic rolling element images into defect classes by first filtering and generating datasets, then training the algorithm with statistical learning to enhance accuracy and automate defect detection.
The deep learning algorithm improves production speed and reduces misinterpretation risks, ensuring rapid and accurate detection of defects in ceramic rolling elements, enhancing the reliability of mechanical systems.
Smart Images

Figure 2024543023000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to ceramic rolling elements, and more particularly to the detection of manufacturing defects in such rolling elements. [Background technology]
[0002] To guide the rotation of mechanical assemblies, it is generally proposed to use rolling bearings equipped with ceramic or steel rolling elements, which can be, for example, balls, but also cylindrical, tapered or spherical rollers.
[0003] These rolling elements of the rolling bearing make it possible to limit the frictional forces and ensure a circular movement of the shaft relative to the fixed element.
[0004] The rolling elements can be used in reversible electric motors or combustion engines in both the automotive and aviation sectors.
[0005] However, because ceramic rolling elements are produced by sintering, they may contain foreign inclusions, non-uniform agglomerations of material, or porosity.
[0006] These defects are prone to cause failure of the rolling elements and therefore damage to products comprising rolling bearings equipped with such rolling elements.
[0007] For example, when a rolling bearing with failed rolling elements is positioned on a rotating shaft, the rolling elements may peel off and / or overheat, making it increasingly difficult for the shaft to rotate.
[0008] Furthermore, the rolling elements are prone to completely detaching from the other components of the bearing, thus disassociating the rotating shaft from its surrounding mechanical system, which can prove critically important in aeronautics.
[0009] In order to prevent failure of the rolling elements, it is advantageous to detect these defects immediately after their manufacture.
[0010] It is known to use X-rays to obtain radiographic images of the rolling elements.
[0011] These images are processed and then analyzed by a human operator with a view to recognizing defects in the rolling elements.
[0012] However, the procedure for the recognition of defects in rolling elements by an operator is not suitable for mass-produced rolling elements.
[0013] Since operators have limited ability to process images, the defect recognition procedure can slow down the entire rolling element production line.
[0014] Additionally, operators may be misled in their interpretation of the images.
[0015] It is known to use deep learning algorithms, for example neural networks, to recognize certain features in images.
[0016] However, for deep learning algorithms to perform well, they must be trained. Summary of the Invention [Problem to be solved by the invention]
[0017] In view of the above, it is an object of the present invention to overcome the aforementioned limitations. [Means for solving the problem]
[0018] Thus, one subject of the present invention is a method for training a deep learning algorithm for the detection of defects in ceramic rolling elements, comprising: - capturing a set of digital radiographic images of rolling elements containing defects and rolling elements not containing any defects; - first filtering each image of the set of images to improve the contrast of each image; - generating a data set based on the set of filtered radiological images; - classifying each image of the set of radiological images into a class of suspicious rolling objects or into a class of non-suspicious rolling objects using a statistical learning algorithm based on the dataset; - training a deep learning algorithm such that, based on the set of radiological images, the deep learning algorithm classifies each image of the set of radiological images as a class of suspicious rolling objects or a class of non-suspicious rolling objects; - comparing the classification of the images of the set of images performed by the statistical learning algorithm with the classification of the images of the set of images performed by the deep learning algorithm for the purpose of determining the accuracy of the classification of the deep learning algorithm relative to the classification performed by the statistical learning algorithm; The method includes:
[0019] By analysing the data of the classification performed by the learning algorithm, it is possible to interpret the variability of the weights of the deep learning algorithm when the deep learning algorithm comprises a neural network.
[0020] When the accuracy rate of the deep learning algorithm is sufficient, the procedure of recognition of defects in rolling elements can be automated in order to increase the production speed of lines for producing rolling elements while minimizing the risk of misinterpretation of the image set.
[0021] Preferably, the generation of the training data set comprises: - partitioning each image into various regions; - re-filtering each region of each image of the set of images to obtain a data vector, the data vector forming a training data set; Includes.
[0022] Advantageously, the classification of each image of the set of radiological images by the statistical learning algorithm comprises: - calculating a value of an image analysis indicator based on an indicative data set for each image of the set of filtered images; - comparing the analysis indicator to a threshold value; - Classifying each image according to the results of the comparison Includes.
[0023] Preferably, the image analysis indices include calculating the mean and / or standard deviation and / or median and / or greyscale gradient.
[0024] Advantageously, the training of the deep learning algorithm comprises: - partitioning each filtered image into different regions; - processing each region of each image using a deep learning algorithm for the purpose of detecting at least one defect based on a selection criterion; - detecting defects according to selection criteria specific to each area; - Classify each image and Includes.
[0025] Preferably, the selection criteria include the geometry of the rolling elements within said region and / or the texture within said region and / or the contrast within said region.
[0026] Advantageously, the deep learning algorithm comprises a convolutional neural network.
[0027] Preferably, the comparison of the classification of the images of the set of images performed by the statistical learning algorithm with the classification of the images of the set of images performed by the deep learning algorithm comprises: - comparing, for each identical image of the set of radiological images, the classification performed by the statistical learning algorithm with the classification performed by the deep learning algorithm; - determining the number of images classified into the same class by the statistical learning algorithm and the deep learning algorithm; - determining the accuracy of the classification of the deep learning algorithms by calculating a ratio between the number of images classified into the same class by the two algorithms and the total number of images in the set of radiological images; Includes.
[0028] Another subject of the invention is a system for training a deep learning algorithm for the detection of defects in ceramic rolling elements, comprising: - means for capturing a set of digital radiographic images of rolling elements containing defects and rolling elements not containing any defects; - filtering means configured to improve the contrast of each image of the set of images; - means for generating a data set based on the set of filtered radiological images; - execution means configured to execute a statistical learning algorithm configured to classify each image of the image set into a class of suspicious rolling objects or into a class of non-suspicious rolling objects based on the dataset; - training means configured to train a deep learning algorithm based on the set of radiological images such that the deep learning algorithm classifies each image of the set of radiological images into a class of suspicious rolling objects or into a class of non-suspicious rolling objects; - a comparison means configured to compare a classification of the images of the set of images performed by the statistical learning algorithm with a classification of the images of the set of images performed by the deep learning algorithm for the purpose of determining an accuracy of the classification of the deep learning algorithm relative to the classification performed by the statistical learning algorithm; The system comprises:
[0029] Other objects, features and advantages of the present invention will appear on reading the following description, given purely by way of non-limiting example and with reference to the accompanying drawings, in which: [Brief description of the drawings]
[0030] [Figure 1] FIG. 1 illustrates a system for training a deep learning algorithm for detection of defects in ceramic rolling elements, according to an embodiment of the present invention. [Diagram 2] FIG. 2 illustrates one mode of implementation of the training system according to the present invention. [Diagram 3] FIG. 2 shows an example of a filtered image according to the present invention. [Figure 4] FIG. 13 is a diagram showing an example of segmentation of a filtered image. [Diagram 5] FIG. 2 shows an example of a data set according to the present invention. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0031] FIG. 1 shows a system for training a deep learning algorithm ALG1 for the detection of defects in ceramic rolling elements 2.
[0032] The system 1 comprises means 3 for capturing a set DATA1 of two-dimensional digital radiographic images of rolling elements 2 containing defects and of rolling elements 2 without any defects.
[0033] The capture means 3 comprise a photographic sensor 4 and a generator 5 of high-frequency electromagnetic waves R1 arranged to emit said waves R1 towards the rolling elements 2.
[0034] These electromagnetic waves R 1 are then absorbed by the photographic sensor 4 .
[0035] Thus, a two-dimensional digital radiographic image of the rolling element 2 is formed.
[0036] It should be noted that the generator 5 typically takes the form of a microfocus X-ray tube.
[0037] According to another embodiment, the capture means 3 generate a three-dimensional digital radiographic image.
[0038] The system 1 further comprises filtering means 6 connected to the capturing means 3 and adapted to process images of the set of images DATA1 captured by the capturing means 3 in order to improve the contrast of each image of the set of images. The filtering means 6 delivers a set of filtered radiological images DATA2.
[0039] The system 1 also comprises means 7 for generating a data set DATA3 from the set DATA2 of radiological images filtered by the filtering means 6, and execution means 8 for executing a statistical learning algorithm ALGO1 configured to classify each image of the set of filtered images DATA2 into a class CL1 of suspicious rolling objects 2 or into a class CL2 of non-suspicious rolling objects 2 based on the data set DATA3. The generation means 7 are connected to the filtering means 6 and to the execution means 8.
[0040] The classification performed by the statistical learning algorithm ALGO1 is stored, for example, in the memory 9.
[0041] The statistical learning algorithm ALGO1 includes, for example, a decision tree algorithm, a random forest algorithm, a support vector machine algorithm, a K-nearest neighbor algorithm, or a logistic regression algorithm.
[0042] It should be noted that a suspect rolling element is a rolling element 2 whose probability of having at least one defect is higher than its probability of not having any defects.
[0043] Defects include, for example, foreign inclusions, non-uniform agglomerations of material, or porosity.
[0044] The system 1 further comprises training means 10 for training the deep learning algorithm ALGO2 on the set of filtered radiographic images DATA2 such that the deep learning algorithm classifies each image of the set of filtered radiographic images DATA2 into a class CL1 of suspicious rolling objects or into a class CL2 of non-suspicious rolling objects. The training means 10 are connected to the filtering means 6.
[0045] The deep learning algorithm ALGO2 comprises, for example, a convolutional neural network.
[0046] The classification performed by the deep learning algorithm ALGO2 is stored, for example, in memory 11.
[0047] The system 1 further comprises a comparison means 12 for comparing the classification performed by the statistical learning algorithm ALGO1 and the classification performed by the deep learning algorithm ALGO2 in order to determine the accuracy of the deep learning algorithm based on the classification of the statistical learning algorithm. The comparison means 12 is connected to the execution means 8 and the training means 10 and for example reads the contents of the memory 9 of the execution means storing the classification performed by the statistical learning algorithm ALGO1 and the contents of the memory 11 of the training means 10 storing the classification performed by the deep learning algorithm ALGO2.
[0048] The system 1 further comprises a processing unit UT using capture means 3 for capturing a set of images DATA1, filtering means, means 7 for generating a data set DATA3 based on a set of filtered radiological images DATA2, execution means 8, training means 10 and comparison means 12.
[0049] The comparison means 12 supplies a data set DATA4 indicative of the accuracy of the classification of the deep learning algorithm relative to the classification performed by the statistical learning algorithm.
[0050] By analyzing the dataset DATA4, it is possible to verify the operation of the deep learning algorithm ALGO2.
[0051] If the level of accuracy is insufficient, a new dataset DATA1 is generated for the purpose of continuing training the deep learning algorithm ALGO2.
[0052] As a variant, when a set of images DATA1 of sufficient size is available, the system 1 is implemented on the basis of said set with the aim of continuing the training of the deep learning algorithm ALGO2 and improving the accuracy of said algorithm.
[0053] FIG. 2 shows an example of an implementation of the system 1.
[0054] In step 20, the capture means 3 capture a set DATA1 of two-dimensional digital radiographic images of rolling elements 2 containing defects and of rolling elements 2 not containing any defects.
[0055] The set of images DATA1 comprises, for example, 100 images, making it possible to obtain a sufficient classification accuracy from the algorithm ALGO1.
[0056] When the set of images DATA1 is complete, in step 22, the filtering means 6 filters the images of the set of images DATA1 to improve the contrast of each image and supplies the set of filtered images DATA2 to the generating means 7 and the training means 10.
[0057] Upon receiving the set of filtered images DATA2, the generating means 7 generates a data set DATA3 based on the set of filtered images DATA2 (step 24), and then the executing means 8 executes the statistical learning algorithm ALGO1 such that the statistical learning algorithm ALGO1 classifies each image of the set of filtered images DATA2 into a class CL1 of a suspicious rolling object or into a class CL2 of a non-suspicious rolling object (step 26).
[0058] Furthermore, upon receiving the set of filtered images DATA2, the training means 10 trains the deep learning algorithm ALGO2 such that the deep learning algorithm classifies each image of the set of filtered images DATA2 into a class CL1 of a suspicious rolling object or into a class CL2 of a non-suspicious rolling object (step 28).
[0059] The two steps 24 and 26, as well as step 28, are, for example, performed in parallel.
[0060] As a variant, steps 24 and 26, and step 28 are performed sequentially.
[0061] In a step 24 of generating a data set DATA3, the generating means 7 partition each image of the set of filtered images DATA2 into various regions.
[0062] Then, for each image region of the set of filtered images DATA2, the means 7 generate a data vector containing a coding of each image region of the set of images DATA2.
[0063] The data set DATA3 contains this data vector.
[0064] FIG. 3 shows an image of the rolling element 2 filtered by the filtering means 6 .
[0065] FIG. 4 shows how the filtered image shown in FIG. 3 is divided into three different regions Z1, Z2, Z3, each region being characterized by a different grayscale level.
[0066] Each image is filtered by grayscale level.
[0067] Of course, each image may be divided into more than at least three different regions.
[0068] FIG. 5 shows an example of a data set DATA3 modeled by the coding line Lc.
[0069] Each line Lc represents data coding an image of the filtered image set DATA2.
[0070] When the execution means 8 receives the data set DATA3, in step 26 an algorithm ALGO1 executed by the execution means 8 calculates, on the basis of the data set DATA3, a value of an image analysis index for each image of the set of images DATA2.
[0071] Image analysis metrics include, for example, calculating the mean and / or standard deviation and / or median and / or greyscale gradient.
[0072] The value of the image index is then compared to a threshold value.
[0073] If the index value of an image of the set of images DATA2 is greater than a threshold value, the image is classified, for example, into a class CL1 of suspicious rolling objects, and if the index value of an image is less than the threshold value, the image is classified, for example, into a class CL2 of non-suspicious rolling objects.
[0074] In step 28, the training means 10 partitions each image of the set of filtered images DATA2 into a number of regions, each region corresponding to one selection criterion comprising, for example, the geometry of the rolling element 2 within said region and / or the texture within said region and / or the contrast within said region.
[0075] An algorithm ALGO2 executed by the training means 10 then processes each region of each image in order to detect at least one defect based on the selection criteria, detects any defects in the rolling element 2 and classifies said rolling element into one of the classes CL1 and CL2 depending on whether the rolling element 2 is considered suspicious or not.
[0076] For example, if the geometry of a rolling element 2 is not circular within a given tolerance compared to the geometry of the rolling element 2 in that region, the algorithm ALGO2 classifies said rolling element 2 into class CL2.
[0077] Once all images of the set of filtered images DATA2 have been classified by the algorithms ALGO1 and ALGO2 into one of the classes CL1 and CL2, in step 30, the comparison means 12 compare, for each given image of the set of images DATA1, the classification performed by the statistical learning algorithm ALGO1 and stored in the memory 9 of the means 8 with the classification performed by the deep learning algorithm ALGO2 and stored in the memory 11.
[0078] The comparison means 12 determines the number of images classified into the same classes CL1, CL2 by the statistical learning algorithm ALGO1 and the deep learning algorithm ALGO2, and determines the classification accuracy of the deep learning algorithm ALGO2 by calculating the ratio of the number of images classified into the same class by both algorithms ALGO1, ALGO2 to the total number of images in the set of radiological images DATA1.
[0079] The comparison means 12 then supplies the data set DATA4.
[0080] By analysing the data DATA4 of the classification performed by the statistical learning algorithm ALGO1 and stored in the memory 9, it becomes possible to interpret the variability of the weights of the deep learning algorithm ALGO2 when the deep learning algorithm ALGO2 comprises a neural network.
[0081] When the accuracy rate of the deep learning algorithm ALGO2 is sufficient, the procedure of recognition of defects in rolling elements can be automated in order to increase the production speed of the line for producing rolling elements while minimizing the risk of misinterpretation of the image set DATA1. [Explanation of symbols]
[0082] 1 System 2. Ceramic rolling elements 3. Import Method 4 Photo Sensor 5. Generator 6 Filtering Methods 7 Generation means 8. Means of Implementation 9. Memory 10. Training Methods 11. Memory 12 Means of comparison ALG1 Deep Learning Algorithm ALGO1 Statistical Learning Algorithm ALGO2 Deep Learning Algorithm CL1 Class of suspect rolling element CL2 Class of rolling elements not suspect DATA1 Set of 2D digital radiographic images DATA2 A set of filtered radiological images DATA3 Training dataset DATA4 Dataset Lc coding line R1 High frequency electromagnetic waves UT Processing Unit Z1 area Z2 area Z3 region
Claims
1. A method for training a deep learning algorithm (ALGO2) for the detection of defects in ceramic rolling elements (2), comprising: - capturing a set of digital radiographic images (DATA1) of a rolling element (2) containing a defect and of a rolling element (2) not containing any defect; first filtering each image of said set of radiological images (DATA1) to improve the contrast of each image; generating a data set (DATA3) based on the set of filtered radiographic images (DATA2); classifying each image of the set of radiological images (DATA1) based on the data set (DATA3) into a class of suspicious rolling objects (CL1) or a class of non-suspicious rolling objects (CL2) using a statistical learning algorithm (ALGO1); training said deep learning algorithm (ALGO2) based on said set of radiological images (DATA1) such that said deep learning algorithm classifies each image of said set of radiological images into said class of suspicious rolling objects (CL1) or into said class of non-suspicious rolling objects (CL2); comparing the classification of the images of the set of radiographic images performed by the statistical learning algorithm (ALGO1) with the classification of the images of the set of radiographic images performed by the deep learning algorithm (ALGO2) with a view to determining the accuracy of the classification of the deep learning algorithm with respect to the classification performed by the statistical learning algorithm; A method comprising:
2. The step of generating a training data set (DATA3) comprises: Partitioning each image into various regions (Z1, Z2, Z3); re-filtering each region of each image of the set of radiological images to obtain a data vector, said data vector forming said training data set (DATA3); 2. The method of claim 1, comprising:
3. said step of classifying each image of said set of radiological images using a statistical learning algorithm (ALGO1), calculating a value of an image analysis index for each image of said set of filtered radiological images (DATA2) based on said index data set (DATA3); comparing the analysis index to a threshold; classifying each image according to the results of the comparison; 3. The method of claim 1 or 2, comprising:
4. The method of claim 3, wherein the image analysis metrics include calculating a mean and / or a standard deviation and / or a median and / or a greyscale gradient.
5. training a deep learning algorithm (ALGO2), partitioning each filtered image into a number of regions (Z1, Z2, Z3); processing each region of each image using said deep learning algorithm (ALGO2) for the purpose of detecting at least one defect based on a selection criterion; detecting defects according to selection criteria specific to each region; classifying each image; 5. The method according to claim 1 , comprising:
6. The method of claim 5 , wherein the selection criteria comprises the geometry of the rolling elements in the region and / or the texture in the region and / or the contrast in the region.
7. 5. The method of claim 1, wherein the deep learning algorithm (ALGO2) comprises a convolutional neural network.
8. said step of comparing said classification of images of said set of radiological images performed by said statistical learning algorithm (ALGO1) with said classification of images of said set of radiological images performed by said deep learning algorithm (ALGO2), - comparing, for each identical image of the set of radiological images, the classification performed by the statistical learning algorithm (ALGO1) with the classification performed by the deep learning algorithm (ALGO2); determining the number of images classified into the same class (CL1, CL2) by the statistical learning algorithm and the deep learning algorithm; determining the accuracy of the classification of the deep learning algorithms by calculating the ratio of the number of images classified into the same class by two algorithms to the total number of images in the set of radiological images; 8. The method of claim 1 , comprising:
9. A system (1) for training a deep learning algorithm (ALGO2) for the detection of defects in ceramic rolling elements (2), comprising: means (3) for capturing a set of digital radiographic images (DATA1) of a rolling element (2) containing a defect and of a rolling element (2) not containing any defect; filtering means (6) configured to improve the contrast of each image of said set of radiation images (DATA1); means (7) for generating a data set (DATA3) based on the set of filtered radiographic images (DATA2); execution means (8) configured to execute a statistical learning algorithm (ALGO1) configured to classify each image of the set of radiological images (DATA1) into a class of suspicious rolling objects (CL1) or into a class of non-suspicious rolling objects (CL2) based on said data set (DATA3); training means (10) configured to train the deep learning algorithm (ALGO2) based on the set of radiological images (DATA1) such that the deep learning algorithm classifies each image of the set of radiological images into the class of suspicious rolling objects (CL1) or into the class of non-suspicious rolling objects (CL2); a comparison means (12) adapted to compare the classification of the images of the set of radiographic images performed by the statistical learning algorithm (ALGO1) with the classification of the images of the set of radiographic images performed by the deep learning algorithm (ALGO2) with a view to determining the accuracy of the classification of the deep learning algorithm relative to the classification performed by the statistical learning algorithm; A system (1) comprising:
Citation Information
Patent Citations
Inspection method of rolling element, manufacturing method of rolling element and rolling element
JP2013044717A
Rolling element inspection method, rolling element manufacturing method, and rolling element
US20140205063A1
