A bearing defect detection method based on neural network fusion strategy

Abstract

The present invention relates to the technical field of bearing defect detection, and in particular to a bearing defect detection method based on a neural network fusion strategy, comprising the following steps: collecting images for annotation and classification, and performing preprocessing; training a bearing defect detector with annotated files and preprocessed images; designing two network parallel structures to jointly decide on the location and category of defects, and to divide defects into confirmed defects 1 and suspected defects; training a local defect classification model C, re-judging suspected defects, and then combining three models to make decisions on suspected defects to obtain confirmed defects 2; and finally outputting confirmed defects 1 and confirmed defects 2. The present invention uses the fusion concept of multiple networks, including the parallel connection of multiple detection networks, plus the series connection of classification networks to jointly decide on the location and category of detection, solving the problem that a single network cannot cover all defects, and the fusion scheme can also take into account over-killing and missed detection.

Description

Technical Field

[0001] The present invention relates to the technical field of bearing defect detection, and in particular to a bearing defect detection method based on a neural network fusion strategy. Background Art

[0002] The earliest use of bearings can be traced back to the ancient Egyptian era, and now the application of bearings can be seen everywhere, ranging from door locks and parts to automobiles, equipment, aerospace, etc. At the same time, some application areas have high requirements for the quality of bearings. However, in the bearing manufacturing process, whether it is a problem of raw materials or process problems, various surface defects and structural defects will occur in the bearings, so the assessment of the severity of defects becomes very valuable.

[0003] At present, the main bearing defect detection scheme on the market is manual inspection, large parts are identified by naked eyes, and small parts are identified by auxiliary tools such as magnifying glasses. It is difficult to have a more detailed quantitative standard, and most of them are determined by human experience and subjectivity. In addition, there are some solutions that use traditional image processing for detection, but due to the changeable bearing styles and heavy surface textures, there are also large differences in the brightness of the detection surface caused by uneven lighting when encountering large bearings. The above situation makes the traditional image processing algorithm have poor compatibility, uneven defect detection effects, and inability to classify and other disadvantages.

[0004] With the development of deep learning, there have been a small number of examples of using deep learning for detection recently, such as using Faster RCNN or YOLO for target detection, but these solutions are still in the initial stage, and they treat the bearing defect detection task and the natural scene object detection task equally, without designing for the characteristics of bearings, and require a large amount of error sample data. In addition, a single model cannot actually achieve a good effect for all types of defects at the same time. Therefore, there is currently no better mature solution for bearing defect detection. Therefore, the present invention proposes a detection scheme based on a deep learning fusion strategy, and according to the characteristics of bearings, the data is processed and expanded in a targeted manner to solve the problem of insufficient data. The present invention is suitable for the detection and classification of bearing defects of various types and sizes, has strong compatibility, and can cover conventional defect types such as scratches, cracks, missing corners, stains, etc. Summary of the invention

[0005] Based on the technical problems existing in the background technology, the present invention proposes a bearing defect detection method based on a neural network fusion strategy, which uses the fusion idea of ​​multiple networks, including the parallel connection of multiple detection networks and the series connection of classification networks to jointly decide the detection location and category, solving the problem in the prior art that a single network cannot cover all bearing defects.

[0006] The present invention provides the following technical solution: a bearing defect detection method based on a neural network fusion strategy, comprising the following steps:

[0007] S1. Collect different types of bearing images with defects, label and classify the collected images, and pre-process the labeled images;

[0008] S2. Use the annotated files and preprocessed images to train the bearing defect detector to locate the defect position and category. The bearing defect detector uses a fusion of two different types of target detection network models A and B. Design two parallel network structures to jointly decide the location and category of defects and classify defects into confirmed defects 1 and suspected defects.

[0009] S3, train the local defect classification model C, re-judge the suspected defects, and classify the suspected defects by the trained local defect classification model C. Then, the three models are combined to make a decision on the suspected defects, and the confirmed defect 2 is obtained;

[0010] S4. Finally, confirm defect 1 and confirm defect 2 are output.

[0011] Preferably, the preprocessing in step S1 includes: performing a polar coordinate transformation on the bearing image, and transforming the bearing into a rectangle after the polar coordinate transformation.

[0012] Preferably, the preprocessing in step S1 further includes: randomly intercepting images of defective areas on different bearings and attaching them to non-defective backgrounds of different bearings.

[0013] Preferably, the preprocessing in step S1 also includes: using Fourier transform to convert the image to the frequency domain, filtering the image using a filter to eliminate texture information on the bearing image, then using Fourier transform to convert the image back to the spatial domain, removing the texture of the bearing, leaving surface defects and structural information, and finally performing a contrast stretch to highlight the characteristics of the defects again.

[0014] Preferably, the filter is a differential filter, which is used to filter out the frequency portion corresponding to the repeated texture of the bearing, leaving only the defective portion.

[0015] Preferably, the final output in step S2 is the confidence level of the locations and categories of defect 1 and suspected defects.

[0016] The present invention provides a bearing defect detection method based on a neural network fusion strategy, which uses the fusion concept of multiple networks, including the parallel connection of multiple detection networks and the series connection of classification networks to jointly decide the detection location and category, solving the problem that a single network cannot cover all defects, and the fusion solution can also take into account over-killing and missed detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a flow chart of the overall scheme of the present invention;

[0018] Figure 2 This is a flow chart of bearing image preprocessing of the present invention;

[0019] Figure 3 This is a plane optical bearing detection effect diagram of the present invention;

[0020] Figure 4 This is a local detection effect diagram of the high-angle optical bearing of the present invention;

[0021] Figure 5 This is a diagram showing the local detection effect of the combined optical bearing of the present invention. DETAILED DESCRIPTION

[0022] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0023] like Figure 1 As shown, the present invention provides a technical solution: a bearing defect detection method based on a neural network fusion strategy, comprising the following steps:

[0024] S1. Collect hundreds of images of different types of bearings with defects (the number can be adjusted according to demand), perform conventional annotation and classification on the collected images, and annotate them with rectangular frames. Annotate and classify the collected images, and preprocess the annotated images;

[0025] like Figure 2 As shown, the preprocessing includes:

[0026] A polar coordinate transformation is performed on the bearing image. Since the bearing is basically circular, the experimental results show that the defect detection effect on the circular pattern is not good. After the polar coordinate transformation, the bearing is transformed into a rectangle.

[0027] Randomly intercept defective area images on different bearings and attach them to non-defective backgrounds of different bearings. This method can effectively increase the amount of negative sample data and solve the small sample problem to a certain extent. Because the number of existing defective samples is small, and the neural network is built on big data, by artificially attaching defects to the background, on the one hand, the number of samples is increased, and on the other hand, the network can see more defects under different backgrounds, which is conducive to network learning.

[0028] The image is converted to the frequency domain using Fourier transform, and a special filter is designed to filter the image to eliminate the texture information on the bearing image. The special filter here is determined based on the characteristics of the current image. For example, in this experiment, in order to remove repeated textures, a differential filter is designed to filter out the frequency part corresponding to the repeated texture, leaving only the defective part.

[0029] Then, the image is converted back to the spatial domain using Fourier transform to remove the bearing texture, leaving the surface defects and structural information. Finally, a contrast stretch is performed to highlight the characteristics of the defects again. There are many methods for contrast stretching, such as histogram equalization, Laplace sharpening, Log transformation, etc. Histogram equalization is used in this experiment.

[0030] S2. Use the annotated files and preprocessed images to train the bearing defect detector to locate the defect position and category. The bearing defect detector uses a fusion of two different types of target detection network models A and B. Design two parallel network structures to jointly decide the location and category of defects and classify defects into confirmed defects 1 and suspected defects.

[0031] The present invention uses a fusion strategy. The previous stage is the fusion of the target detection network. There are many network choices. Yolov4 and an improved version of Faster RCNN are used in this experiment. The fusion effect of networks with completely different structures will be better.

[0032] The two networks are trained using the labeled files and the preprocessed images mentioned above (there is still conventional data enhancement during training, such as flipping, rotation, etc.). The backbone network can be selected at will according to the effect, and finally the detection models of the two networks are obtained.

[0033] According to the experimental results, two parallel networks are designed to jointly decide the location and category of defects, and finally classify the defects into confirmed defects and suspected defects. Suspected defects are left for the next classification network to re-judge. The final output of this part is the confidence of the location and category of confirmed defects and suspected defects.

[0034] S3, train the local defect classification model C, re-judge the suspected defects, and classify the suspected defects by the trained local defect classification model C. Then, the three models are combined to make a decision on the suspected defects, and the confirmed defect 2 is obtained;

[0035] Classification networks include but are not limited to the ResNet series and the Inception series. ResNet18 is used in this experiment. The trained classification model classifies suspected defects, and then makes decisions on suspected defects by combining the three networks.

[0036] S4. Finally, confirm defect 1 and confirm defect 2 are output.

[0037] like Figure 3 The plane optical bearing detection effect diagram shown; Figure 4 The high-angle optical bearing local detection effect diagram shown; and Figure 5 The combined optical bearing local inspection effect diagram is shown. The rectangular frame in the figure is blue or other conspicuous colors, and the rectangular frame is also accompanied by defect classification labels.

[0038] In the present invention, according to the characteristics of bearing images (circular, heavy texture of the whole image), a new image preprocessing scheme is designed, including polar coordinate transformation and Fourier transform de-texturing, which greatly improves the detection effect.

[0039] The problem of insufficient data from small samples was solved by using the method of local defect cutout and random application.

[0040] The idea of ​​multi-network fusion is used, including the parallel connection of multiple detection networks and the series connection of classification networks to jointly decide the detection location and category, which solves the problem that a single network cannot cover all defects. In addition, the fusion solution can also take into account both over-killing and missed detection.

[0041] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A bearing defect detection method based on a neural network fusion strategy, characterized in that: The steps include: S1. Collect different types of bearing images with defects, label and classify the collected images, and pre-process the labeled images; The pre-processing in step S1 includes: Perform a polar coordinate transformation on the bearing image, and transform the bearing into a rectangle after the polar coordinate transformation; Randomly capture defective area images on different bearings and attach them to non-defective backgrounds of different bearings; The image is converted to the frequency domain using Fourier transform, and the image is filtered using a filter to eliminate the texture information on the bearing image. The image is then converted back to the spatial domain using Fourier transform to remove the texture of the bearing, leaving the surface defects and structural information. Finally, a contrast stretch is performed to highlight the characteristics of the defects again. The filter is a differential filter, which is used to filter out the frequency part corresponding to the repeated texture of the bearing, leaving only the defective part. S2. Use the annotated files and preprocessed images to train the bearing defect detector to locate the defect position and category. The bearing defect detector uses a fusion of two different types of target detection network models A and B. Design two parallel network structures to jointly decide the location and category of defects and classify defects into confirmed defects 1 and suspected defects. S3, train the local defect classification model C, re-judge the suspected defects, and classify the suspected defects by the trained local defect classification model C. Then, the three models are combined to make a decision on the suspected defects, and the confirmed defect 2 is obtained; S4. Finally, confirm defect 1 and confirm defect 2 are output.

2. A bearing defect detection method based on a neural network fusion strategy according to claim 1, characterized in that: The final output in step S2 is the confidence level of the locations and categories of defect 1 and suspected defects.

Images