A cross-scale abrasive particle image segmentation method for online ferrography
Patent Information
- Application Number
- CN202410910724.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-09
- Publication Date
- 2026-10-16
- Estimated Expiration
- 2044-07-09
AI Technical Summary
[0004]为了解决极端尺度磨粒碎片难以分割的问题，本发明提供一种针对在线铁谱的跨尺度磨粒图像分割方法，包括：
[0044]This invention captures images of multi-scale wear fragments under reflected and transmitted light to obtain images containing different dimensions of information. The reflected light image mainly reflects the texture and color information of the object's surface, while the transmitted light image may reveal the internal structure or defects of the object. The adaptive Threshold function can effectively handle images with uneven illumination or large contrast variations. The segmentation image created by binarizing and adding the reflected and transmitted images combines the effective information from both reflected and transmitted light images, enhancing the features in the image and improving the accuracy of segmentation.
Smart Images
Figure CN118747754B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of mechanical equipment monitoring and maintenance, and in particular relates to a cross-scale abrasive grain image segmentation method for online ferrography. Background Technology
[0002] By monitoring the information of lubricating oil in mechanical equipment, the operating status and service life of the equipment can be predicted. The information of abrasive fragments carried in the lubricating oil is the key to analyzing the condition of mechanical equipment. Monitoring information such as the size, concentration and color of abrasive fragments can effectively analyze the operating status of mechanical equipment.
[0003] Currently, the scale of abrasive fragments in captured images varies greatly, with most abrasive fragments exhibiting extreme scales, meaning their scales are polarized and highly variable. Existing deep learning methods are mostly designed for image segmentation of objects with relatively small scale variations. When existing deep learning networks are applied to segment images of abrasive fragments at extreme scales, some details or features in the images inevitably cannot be correctly captured by the deep learning algorithm, thus affecting the accuracy of segmentation and consequently impacting subsequent analysis of the operating status of mechanical equipment. This results in the lack of an ideal model that can effectively segment abrasive fragment images at extreme scales. Summary of the Invention
[0004] To address the challenge of segmenting extreme-scale abrasive fragments, this invention provides a cross-scale abrasive image segmentation method for online ferrography, comprising:
[0005] S1: Acquire images of multi-scale wear debris under reflected and transmitted light, using reflection and transmission images of equal pixel size;
[0006] S2: Apply the adaptive Threshold function to the reflection image and the transmission image respectively for binary discrimination, then add them together to obtain the fused target segmentation map;
[0007] S3: Combine the fused target segmentation image and reflection image into an image pair, and use the image pair as input to train the wear debris segmentation model;
[0008] S4: Use the trained wear debris segmentation model to segment the multi-scale wear debris image to obtain the segmented image.
[0009] Preferably, acquiring images of multi-scale wear debris under reflected and transmitted light includes:
[0010] The lubricating oil inside the pipe was photographed using online visual ferrography (OLVF) under reflected and transmitted light to obtain a 1:1 pixel-sized image of the reflected and transmitted light.
[0011] Preferably, the step of performing binary discrimination on the reflected image and the transmitted image using the adaptive Threshold function and then adding them includes:
[0012] S21: Read the reflection and transmission images;
[0013] S22: Apply the adaptive threshold function to convert the reflected image and the transmitted image into binary images respectively;
[0014] S23: The two binarized images, after conversion of the reflection image and the transmission image, are added together at the corresponding pixels to obtain the fused target segmentation map.
[0015] Preferably, training the wear debris segmentation model using the fused target segmentation map and reflection image includes:
[0016] The reflection image is input into the wear debris segmentation model for forward propagation. Based on the segmentation results of the reflection image and the target segmentation map of the wear debris segmentation model, the BCE loss function is used for iterative processing. Based on the results of the BCE loss function, the SGD optimizer algorithm is used to update the parameters of the wear debris segmentation model until the loss function converges, thus completing the training of the wear debris segmentation model.
[0017] Preferably, the wear fragmentation model adopts a U-net structure, which includes: an encoding part, N NTN feature fusion layers, and a decoding part; the encoding part includes: N cascaded encoding layers; the decoding part includes N+1 decoding layers;
[0018] The input to the first coding layer is the reflection image, and the output features of the Nth coding layer are the input features of the first decoding layer.
[0019] The input features of the first NTN feature fusion layer include the output features of the Nth encoding layer and the output features of the first decoding layer; the output features of the first NTN feature fusion layer are the input features of the second decoding layer.
[0020] The input features of the nth NTN feature fusion layer include the output features of the (N+1-n)th coding layer and the output features of the nth decoding layer.
[0021] The output features of the (N+1)th decoding layer are processed by convolution and used as the segmented image of the wear fragmentation model, where n is an integer between 1 and N.
[0022] Preferably, each coding layer includes: a first convolutional layer and M first residual blocks connected in series; wherein the input features and output features of each first residual block are residually concatenated to serve as the input features of the next first residual block; wherein the input features of the first first residual block, the input features of the Nth first residual block, and the output features of the Nth first residual block are residually concatenated to serve as the output features of the coding layer; the input features of the first convolutional layer are the input features of the coding layer, and the output features of the first convolutional layer are the input features of the first first residual block.
[0023] Preferably, each decoding layer includes: a first second residual block, a self-attention layer, and a second second residual block, which are sequentially connected in series; the input features of the first second residual block are the input features of the decoding layer; the input features and output features of the first second residual block are residually concatenated to serve as the input features of the self-attention layer; the output features of the self-attention layer are the input features of the second second residual block; and the input features of the second second residual block, the output features of the second second residual block, and the input features of the first second residual block are residually concatenated to serve as the output features of the decoding layer.
[0024] Preferably, both the first residual block and the second residual block include: K cascaded RT modules; the input and output features of each RT module are residually concatenated to serve as the input features of the next RT module; the input features of the first RT module are the input features of either the first residual block or the second residual block; the output features of the Kth RT module, the input features of the Kth RT module, and the input features of the first RT module are residually concatenated to serve as the output features of the residual block; the RT module includes: B cascaded feature processing modules; each feature processing module includes a sequentially cascaded activation function and a second convolutional layer; the input and output features of each feature processing module are residually concatenated to serve as the input features of the next feature processing module; the input features of the first feature processing module are the input features of the RT module; the input features of the Bth feature processing module, the output features of the Bth feature processing module, and the input features of the first feature processing module are residually concatenated to serve as the output features of the RT module.
[0025] Preferably, the features obtained by the NTN feature fusion layer from the encoding layer are used as low-level feature images, and the features obtained by the NTN feature fusion layer from the decoding layer are used as high-level feature images, and the low-level feature images and high-level images are fused together;
[0026] The NTN feature fusion layer includes two feature processing branches, an upper and an lower one. The upper feature processing branch is used to process the low-level feature image, including:
[0027] The intermediate feature f1 is obtained by convolution processing the low-level feature image;
[0028] The intermediate feature f1 is subjected to max pooling and average pooling respectively to obtain intermediate feature f11 and intermediate feature f12;
[0029] The intermediate features f11 and f12 are input into the multilayer perceptron for processing to obtain the intermediate features f21 and f22.
[0030] The intermediate feature f2 is obtained by adding the intermediate feature f21 and the intermediate feature f22.
[0031] Binarize the intermediate feature f2 to obtain the intermediate feature f3;
[0032] The intermediate feature f4 is obtained by matrix multiplication of intermediate feature f3 and intermediate feature f1;
[0033] The lower feature processing branch is used to process the high-level feature image, including:
[0034] The intermediate feature g1 is obtained by upsampling or convolution of the high-level feature image;
[0035] The intermediate feature g1 is subjected to max pooling and average pooling respectively to obtain intermediate feature g11 and intermediate feature g12;
[0036] The intermediate feature g11 and intermediate feature g12 are concatenated to obtain the intermediate feature g2;
[0037] The intermediate feature g2 is then subjected to max pooling and average pooling to obtain the intermediate feature g3.
[0038] The intermediate feature g3 is convolved to obtain the intermediate feature g4;
[0039] Binarize the intermediate feature g4 to obtain the intermediate feature g5;
[0040] The intermediate feature g6 is obtained by matrix multiplication of intermediate feature g1 and intermediate feature g5.
[0041] The intermediate features f4 and g6 are concatenated to obtain the output features of the NTN feature fusion layer.
[0042] Another aspect of the present invention provides a computer storage medium storing a program that, when executed by a processor, implements the aforementioned method for cross-scale abrasive grain image segmentation for online ferrography.
[0043] The present invention has at least the following beneficial effects
[0044] This invention captures images of multi-scale wear fragments under reflected and transmitted light to obtain images containing different dimensions of information. The reflected light image mainly reflects the texture and color information of the object's surface, while the transmitted light image may reveal the internal structure or defects of the object. The adaptive Threshold function can effectively handle images with uneven illumination or large contrast variations. The segmentation image created by binarizing and adding the reflected and transmitted images combines the effective information from both reflected and transmitted light images, enhancing the features in the image and improving the accuracy of segmentation.
[0045] This invention improves the traditional U-net model structure, enabling the recovery of spatial information lost during the encoding process. It can more clearly and accurately segment images with obvious wear debris at extreme scales, solving the problem of unclear or lost targets after segmentation of images with wear debris at extreme scales. Attached Figure Description
[0046] Figure 1 This is a schematic diagram of the method flow of the present invention;
[0047] Figure 2 This is a schematic diagram of the network structure of the wear debris segmentation model of the present invention;
[0048] Figure 3 This is a schematic diagram of the specific structure of the residual block of the present invention;
[0049] Figure 4 This is a schematic diagram of the specific structure of the NTN feature fusion layer of the present invention;
[0050] Figure 5 This is a schematic diagram comparing the segmentation effects of the wear debris segmentation model of the present invention;
[0051] Figure 6 This is a schematic diagram comparing the segmentation effects of the wear debris segmentation model of the present invention; Detailed Implementation
[0052] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.
[0053] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures. They should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.
[0054] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.
[0055] Please see Figure 1 This invention provides a method for cross-scale abrasive grain image segmentation based on online ferrography, comprising:
[0056] S1: Acquire images of multi-scale wear debris under reflected and transmitted light, using reflection and transmission images of equal pixel size;
[0057] S2: Apply the adaptive Threshold function to the reflection image and the transmission image respectively for binary discrimination, then add them together to obtain the fused target segmentation map;
[0058] S3: Combine the fused target segmentation image and reflection image into an image pair, and use the image pair as input to train the wear debris segmentation model;
[0059] S4: Use the trained wear debris segmentation model to segment the multi-scale wear debris image to obtain the segmented image.
[0060] Preferably, acquiring images of multi-scale wear debris under reflected and transmitted light includes:
[0061] The lubricating oil inside the pipe was photographed using online visual ferrography (OLVF) under reflected and transmitted light to obtain a 1:1 pixel-sized image of the reflected and transmitted light.
[0062] In this embodiment, images are taken of the lubricating oil in the pipeline using online visual ferrography (OLVF); the operating status of the equipment can be obtained in real time, facilitating timely maintenance of the mechanical equipment.
[0063] Preferably, the step of performing binary discrimination on the reflected image and the transmitted image using the adaptive Threshold function and then adding them includes:
[0064] S21: Read the reflection and transmission images;
[0065] S22: Apply the adaptive threshold function to convert the reflected image and the transmitted image into binary images respectively;
[0066] S23: The two binarized images, after conversion of the reflection image and the transmission image, are added together at the corresponding pixels to obtain the fused target segmentation map.
[0067] In this example, the effective information from both reflected and transmitted light images is combined by fusing the reflected and transmitted light images, thereby enhancing the features in the image and improving the accuracy of segmentation.
[0068] Preferably, training the wear debris segmentation model using the fused target segmentation map and reflection image includes:
[0069] The reflection image is input into the wear debris segmentation model for forward propagation. Based on the segmentation results of the reflection image and the target segmentation map of the wear debris segmentation model, the BCE loss function is used for iterative processing. Based on the results of the BCE loss function, the SGD optimizer algorithm is used to update the parameters of the wear debris segmentation model until the loss function converges, thus completing the training of the wear debris segmentation model.
[0070] Please see Figure 2 Preferably, the wear fragmentation model adopts a U-net structure, which includes: an encoding part, N NTN feature fusion layers, and a decoding part; the encoding part includes: N cascaded encoding layers; the decoding part includes N+1 decoding layers;
[0071] The input to the first coding layer is the reflection image, and the output features of the Nth coding layer are the input features of the first decoding layer.
[0072] The input features of the first NTN feature fusion layer include the output features of the Nth encoding layer and the output features of the first decoding layer; the output features of the first NTN feature fusion layer are the input features of the second decoding layer.
[0073] The input features of the nth NTN feature fusion layer include the output features of the (N+1-n)th coding layer and the output features of the nth decoding layer.
[0074] The output features of the (N+1)th decoding layer are processed by convolution and used as the segmented image of the wear fragmentation model, where n is an integer between 1 and N.
[0075] In this embodiment, the value of N can be, for example, 2, 3, 4, or 5; those skilled in the art can determine a suitable value through a limited number of experiments according to actual needs; in this embodiment, the input of one encoding layer is a reflection image (i.e., an image of wear debris taken using only reflected light); multi-scale features are extracted through the encoding part, and then the feature maps of the encoder and decoder are fused through the NTN feature fusion layer to make full use of feature information from different scales, thereby improving the accuracy and robustness of segmentation. By fusing the feature maps of the encoder and decoder through the NTN feature fusion layer, the spatial information lost during the encoding process can be recovered, which can effectively fuse multi-scale features and achieve accurate segmentation of wear debris.
[0076] Please see Figure 3 Preferably, each coding layer includes: a first convolutional layer and M first residual blocks connected in series; wherein the input features and output features of each first residual block are residually concatenated to serve as the input features of the next first residual block; wherein the input features of the first first residual block, the input features of the Nth first residual block, and the output features of the Nth first residual block are residually concatenated to serve as the output features of the coding layer; the input features of the first convolutional layer are the input features of the coding layer, and the output features of the first convolutional layer are the input features of the first first residual block.
[0077] In this embodiment, the value of M can be, for example, 2, 3, 4 or 5; those skilled in the art can determine a suitable value through a limited number of experiments according to actual needs; in this embodiment, the coding layer combines the convolutional layer and multiple residual blocks, allowing the network to directly pass input features to deeper layers, which helps to alleviate the gradient vanishing problem and improve training efficiency and model performance.
[0078] Preferably, each decoding layer includes: a first second residual block, a self-attention layer, and a second second residual block, which are sequentially connected in series; the input features of the first second residual block are the input features of the decoding layer; the input features and output features of the first second residual block are residually concatenated to serve as the input features of the self-attention layer; the output features of the self-attention layer are the input features of the second second residual block; and the input features of the second second residual block, the output features of the second second residual block, and the input features of the first second residual block are residually concatenated to serve as the output features of the decoding layer.
[0079] In this embodiment, each decoding layer incorporates residual blocks and a self-attention layer to enhance the model's feature extraction capabilities and its ability to capture global information. This enables the U-net model to more accurately understand image content and generate high-quality segmentation results in segmentation tasks.
[0080] Preferably, both the first residual block and the second residual block include: K cascaded RT modules; the input and output features of each RT module are residually concatenated to serve as the input features of the next RT module; the input features of the first RT module are the input features of either the first residual block or the second residual block; the output features of the Kth RT module, the input features of the Kth RT module, and the input features of the first RT module are residually concatenated to serve as the output features of the residual block; the RT module includes: B cascaded feature processing modules; each feature processing module includes a sequentially cascaded activation function and a second convolutional layer; the input and output features of each feature processing module are residually concatenated to serve as the input features of the next feature processing module; the input features of the first feature processing module are the input features of the RT module; the input features of the Bth feature processing module, the output features of the Bth feature processing module, and the input features of the first feature processing module are residually concatenated to serve as the output features of the RT module.
[0081] In this embodiment, the value of K can be, for example, 2, 3, 4 or 5; those skilled in the art can choose a suitable value through a limited number of experiments according to actual needs; in this embodiment, the activation function is, for example, the ReLU function; residual connection is represented by adding corresponding elements of two matrices, which is a basic operator in matrix operations and a commonly used technique in the field of deep learning, so it will not be elaborated on in this article.
[0082] In this embodiment, the residual block adopts a residual structure to mitigate gradient descent or gradient explosion and accelerate model convergence.
[0083] Please see Figure 4 Preferably, the features obtained by the NTN feature fusion layer from the encoding layer are used as low-level feature images, and the features obtained by the NTN feature fusion layer from the decoding layer are used as high-level feature images, and the low-level feature images and high-level images are fused together;
[0084] The NTN feature fusion layer includes two feature processing branches, an upper and an lower one. The upper feature processing branch is used to process the low-level feature image, including:
[0085] The intermediate feature f1 is obtained by convolution processing the low-level feature image;
[0086] The intermediate feature f1 is subjected to max pooling and average pooling respectively to obtain intermediate feature f11 and intermediate feature f12;
[0087] The intermediate features f11 and f12 are input into the multilayer perceptron for processing to obtain the intermediate features f21 and f22.
[0088] The intermediate feature f2 is obtained by adding the intermediate feature f21 and the intermediate feature f22.
[0089] Binarize the intermediate feature f2 to obtain the intermediate feature f3;
[0090] The intermediate feature f4 is obtained by matrix multiplication of intermediate feature f3 and intermediate feature f1;
[0091] The lower feature processing branch is used to process the high-level feature image, including:
[0092] The intermediate feature g1 is obtained by upsampling or convolution of the high-level feature image;
[0093] The intermediate feature g1 is subjected to max pooling and average pooling respectively to obtain intermediate feature g11 and intermediate feature g12;
[0094] The intermediate feature g11 and intermediate feature g12 are concatenated to obtain the intermediate feature g2;
[0095] The intermediate feature g2 is then subjected to max pooling and average pooling to obtain the intermediate feature g3.
[0096] The intermediate feature g3 is convolved to obtain the intermediate feature g4;
[0097] Binarize the intermediate feature g4 to obtain the intermediate feature g5;
[0098] The intermediate feature g6 is obtained by matrix multiplication of intermediate feature g1 and intermediate feature g5.
[0099] The intermediate features f4 and g6 are concatenated to obtain the output features of the NTN feature fusion layer.
[0100] In this embodiment, the NTN feature fusion layer repeatedly extracts and fuses the features of the segmented image, effectively fusing the low-level feature image provided by the encoding layer and the high-level feature image provided by the decoding layer. At the same time, the NTN feature fusion layer fuses the feature maps of the encoder and decoder to recover the spatial information lost during the encoding process, which can effectively fuse multi-scale features and achieve accurate segmentation of wear debris.
[0101] Please see Figure 5 and Figure 6 , Figure 5 The left side shows an image with large-scale wear debris; the middle side shows a segmentation image of extremely large-scale wear debris processed by the wear debris segmentation model of this invention; and the right side shows a segmentation image of extremely large-scale wear debris processed by a traditional segmentation model. Figure 5 As shown, the present invention can more clearly and accurately segment images with obvious large-scale wear fragments, solving the problem of unclear or lost targets after segmentation of large-scale wear fragment images; Figure 6 The image on the left shows wear debris with extremely small scale. Figure 6 The image in the middle shows the wear debris segmentation model of this invention, which processes extremely small-scale wear debris. Figure 6 The right side shows the segmentation diagram of wear debris at extremely small scales processed by traditional segmentation models, such as... Figure 6 As shown, this invention acquires images containing different dimensions of information by capturing images of multi-scale wear debris under reflected and transmitted light. The reflected light image primarily reflects the texture and color information of the object's surface, while the transmitted light image may reveal the object's internal structure or defects. The adaptive Threshold function effectively handles images with uneven illumination or large contrast variations. The segmentation image, created by binarizing and adding the reflected and transmitted images, combines the effective information from both images, enhancing image features and improving segmentation accuracy. This invention improves the traditional U-net model structure, recovering spatial information lost during encoding and more clearly and accurately segmenting images with obvious extreme-scale wear debris, thus solving the problem of unclear or lost targets after segmentation of extreme-scale wear debris images.
[0102] Another aspect of the present invention provides a computer storage medium storing a program that, when executed by a processor, implements the aforementioned method for cross-scale abrasive grain image segmentation for online ferrography.
[0103] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.
[0104] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
[0105] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for cross-scale abrasive grain image segmentation for online ferrography, characterized in that, include: S1: Acquire images of multi-scale wear debris under reflected and transmitted light, using reflection and transmission images of equal pixel size; S2: Apply the adaptive Threshold function to the reflection image and the transmission image respectively for binary discrimination, then add them together to obtain the fused target segmentation map; S3: Combine the fused target segmentation image and reflection image into an image pair, and use the image pair as input to train the wear debris segmentation model; The wear fragmentation model adopts a U-net structure, which includes: an encoding part, N NTN feature fusion layers, and a decoding part; the encoding part includes: N cascaded encoding layers; the decoding part includes N+1 decoding layers; The input to the first coding layer is the reflection image, and the output features of the Nth coding layer are the input features of the first decoding layer. The input features of the first NTN feature fusion layer include the output features of the Nth encoding layer and the output features of the first decoding layer; the output features of the first NTN feature fusion layer are the input features of the second decoding layer. The input features of the nth NTN feature fusion layer include: the output features of the (N+1-n)th coding layer and the output features of the nth decoding layer; The output features of the (N+1)th decoding layer are processed by convolution and used as the image segmented by the wear debris segmentation model, where n is an integer between 1 and N; The NTN feature fusion layer obtains features from the encoding layer as a low-level feature image, and obtains features from the decoding layer as a high-level feature image, and then fuses the low-level feature image and the high-level feature image. S4: Use the trained wear debris segmentation model to segment the multi-scale wear debris image to obtain the segmented image.
2. The method for cross-scale abrasive grain image segmentation based on online ferrography according to claim 1, characterized in that, The acquisition of images of multi-scale wear debris under reflected and transmitted light includes: The lubricating oil inside the pipe was photographed using online visual ferrography (OLVF) under reflected and transmitted light to obtain a 1:1 pixel-sized image of the reflected and transmitted light.
3. The method for cross-scale abrasive grain image segmentation for online ferrography according to claim 1, characterized in that, The step of performing binary discrimination on the reflected and transmitted images using the adaptive Threshold function and then adding them together includes: S21: Read the reflection and transmission images; S22: Apply the adaptive threshold function to convert the reflected image and the transmitted image into binary images respectively; S23: The two binarized images, after conversion of the reflection image and the transmission image, are added together at the corresponding pixels to obtain the fused target segmentation map.
4. The method for cross-scale abrasive grain image segmentation based on online ferrography according to claim 1, characterized in that, The method of training the wear debris segmentation model using the fused target segmentation map and reflection image includes: The reflection image is input into the wear debris segmentation model for forward propagation. Based on the segmentation results of the reflection image and the target segmentation map of the wear debris segmentation model, the BCE loss function is used for iterative processing. Based on the results of the BCE loss function, the SGD optimizer algorithm is used to update the parameters of the wear debris segmentation model until the loss function converges, thus completing the training of the wear debris segmentation model.
5. The method for cross-scale abrasive grain image segmentation based on online ferrography according to claim 1, characterized in that, Each coding layer includes: a first convolutional layer and M first residual blocks connected in series; wherein, the input features and output features of each first residual block are residually concatenated to serve as the input features of the next first residual block; wherein, the input features of the first first residual block, the input features of the Nth first residual block, and the output features of the Nth first residual block are residually concatenated to serve as the output features of the coding layer; the input features of the first convolutional layer are the input features of the coding layer, and the output features of the first convolutional layer are the input features of the first first residual block.
6. The method for cross-scale abrasive grain image segmentation for online ferrography according to claim 5, characterized in that, Each decoding layer includes: a first second residual block, a self-attention layer, and a second second residual block, which are sequentially connected in series; the input features of the first second residual block are the input features of the decoding layer; the input features and output features of the first second residual block are residually concatenated to serve as the input features of the self-attention layer; the output features of the self-attention layer are the input features of the second second residual block; the input features of the second second residual block, the output features of the second second residual block, and the input features of the first second residual block are residually concatenated to serve as the output features of the decoding layer.
7. The method for cross-scale abrasive grain image segmentation for online ferrography according to claim 6, characterized in that, Both the first and second residual blocks include: K cascaded RT modules; the input and output features of each RT module are residually concatenated to serve as the input features of the next RT module; the input features of the first RT module are the input features of either the first or second residual block; the output features of the Kth RT module, the input features of the Kth RT module, and the input features of the first RT module are residually concatenated to serve as the output features of the residual block; each RT module includes: B cascaded feature processing modules; each feature processing module includes a sequentially cascaded activation function and a second convolutional layer; the input and output features of each feature processing module are residually concatenated to serve as the input features of the next feature processing module; the input features of the first feature processing module are the input features of the RT module; the input features of the Bth feature processing module, the output features of the Bth feature processing module, and the input features of the first feature processing module are residually concatenated to serve as the output features of the RT module.
8. The method for cross-scale abrasive grain image segmentation based on online ferrography according to claim 1, characterized in that, The NTN feature fusion layer includes two feature processing branches, an upper and an lower one. The upper feature processing branch is used to process the low-level feature image, including: The intermediate feature f1 is obtained by convolution processing the low-level feature image; The intermediate feature f1 is subjected to max pooling and average pooling respectively to obtain intermediate feature f11 and intermediate feature f12; The intermediate features f11 and f12 are input into the multilayer perceptron for processing to obtain the intermediate features f21 and f22. The intermediate feature f2 is obtained by adding the intermediate feature f21 and the intermediate feature f22. Binarize the intermediate feature f2 to obtain the intermediate feature f3; The intermediate feature f4 is obtained by matrix multiplication of intermediate feature f3 and intermediate feature f1; The lower feature processing branch is used to process the high-level feature image, including: The intermediate feature g1 is obtained by upsampling or convolution of the high-level feature image; The intermediate feature g1 is subjected to max pooling and average pooling respectively to obtain intermediate feature g11 and intermediate feature g12; The intermediate feature g11 and intermediate feature g12 are concatenated to obtain the intermediate feature g2; The intermediate feature g2 is then subjected to max pooling and average pooling to obtain the intermediate feature g3. The intermediate feature g3 is convolved to obtain the intermediate feature g4; Binarize the intermediate feature g4 to obtain the intermediate feature g5; The intermediate feature g6 is obtained by matrix multiplication of intermediate feature g1 and intermediate feature g5. The intermediate features f4 and g6 are concatenated to obtain the output features of the NTN feature fusion layer.
9. A computer storage medium storing a program, characterized in that, When the program is executed by the processor, it implements the cross-scale abrasive grain image segmentation method for online ferrography as described in any one of claims 1-8.