Method and system for detecting metal particle defects in lubricating oil

Abstract

The application provides a metal abrasive particle defect detection method and system in lubricating oil, and the method comprises the following steps: scanning the lubricating oil sample collected by a machine by using a multi-spectral imaging system to obtain a double-channel image in a visible light band and a near-infrared band; the double-channel image comprises a visible light image and a near-infrared image; the visible light image and the near-infrared image are enhanced respectively, and the enhanced images are fused and analyzed to obtain a double-channel fused image; a sliding window is used for the double-channel fused image to construct a direction constraint factor; the double-channel fused image is threshold segmented by using the direction constraint factor to obtain a metal abrasive particle region; and an actual metal abrasive particle defect image is collected to analyze the metal abrasive particle region to obtain a recognition result. The visible light image and the near-infrared image are enhanced and fused for recognition, the contour and the texture are double-recognized, the accuracy is improved, and the misjudgment rate is reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of lubricating oil quality detection, and particularly relates to a method and system for detecting metal abrasive particle defects in lubricating oil. BACKGROUND

[0002] Lubricating oil plays a vital role in various mechanical equipment and engines, and its main functions include reducing friction, reducing wear, dissipating heat, and preventing direct contact between metal parts. However, after long-term use, the sampled lubricating oil is found to contain metal abrasive particle residues; in order to ensure the reliability of the equipment and prolong the service life, monitoring the number and morphology of metal abrasive particles in the lubricating oil has become an important indicator for equipment condition assessment. The presence of metal abrasive particles usually indicates that there is excessive wear, hidden trouble or abnormal operating condition inside the device, so high-precision detection of metal abrasive particles in lubricating oil has important technical significance.

[0003] Traditional metal abrasive particle detection methods rely on manual observation or chemical analysis. These methods often have the disadvantages of long detection time, poor accuracy, low efficiency, and are difficult to achieve real-time and accurate detection of metal abrasive particles in lubricating oil. In order to improve the detection efficiency and accuracy, researchers have tried different image processing and optical imaging technologies, including infrared imaging, optical microscopic imaging, etc., but these technologies have the problems of limited waveband selection, incomplete image processing, etc., which makes it difficult to extract surface features and deep information when detecting metal abrasive particles in lubricating oil, and cannot effectively identify abrasive particles. SUMMARY

[0004] The purpose of the present application is to provide a method and system for detecting metal abrasive particle defects in lubricating oil, which solves the above technical problems pointed out in the prior art.

[0005] The present application provides a method for detecting metal abrasive particle defects in lubricating oil, comprising the following operation steps:

[0006] The lubricating oil sample collected by the machine is scanned by a multispectral imaging system to obtain dual-channel images in the visible light band and the near-infrared band; the dual-channel images include visible light images and near-infrared images;

[0007] The visible light images and the near-infrared images are respectively enhanced, and the enhanced images are fused and analyzed to obtain a dual-channel fusion image; a sliding window is applied to the dual-channel fusion image to construct a direction constraint factor; the dual-channel fusion image is threshold segmented by the direction constraint factor to obtain a abrasive particle region; the actual metal abrasive particle defect image is collected to analyze the abrasive particle region to obtain a recognition result.

[0008] Compared with the prior art, the present application has at least the following technical advantages:

[0009] It can be known from the above-mentioned metal abrasive particle defect detection method and system of the lubricating oil provided by the application that, in specific application, the visible light image is divided into multiple small areas through the adaptive histogram equalization processing method, and local enhancement is performed block by block, which effectively avoids the problems of over-enhancement or under-enhancement that may occur in global histogram equalization, especially for the noise amplification and detail loss problems in the abrasive particle image, more accurate image enhancement is provided, the dynamic contrast stretching amplitude technology prevents over-enhancement after the gray histogram is cropped, so as to ensure the balance of the enhanced image, especially when the edge details of the abrasive particle are enhanced, the edge definition and detail performance of the abrasive particle can be effectively improved; the near-infrared image is subjected to multi-scale Gaussian filtering through the multi-scale Retinex enhancement algorithm, the illumination component and the reflection component of different scales are extracted, and the local details and definition of the image are enhanced; after the illumination component is removed, the influence of light can be eliminated, and the true form of the abrasive particle can be accurately presented; through multi-scale fusion, the details of the near-infrared image are effectively retained, and the interference of light changes on the form and texture of the abrasive particle is reduced; through the fusion of the gradient amplitude graph of the visible light image and the reflection component of the near-infrared image, the edge and texture details in the image can be accurately highlighted, and the form of the abrasive particle can be accurately identified; the fusion of the two channels combines the details of the visible light image and the true surface information of the near-infrared image, further improves the image contrast and detail resolution, and reduces the noise interference; the gradient direction of the image is calculated through the sliding window technology, and the direction constraint factor is constructed, and the segmentation threshold surface is generated by combining the convolution operation, so that the abrasive particle region can be accurately segmented, and strong support is provided for the local feature analysis of the abrasive particle region, which helps to improve the accuracy of abrasive particle identification, especially when complex forms and similar abrasive particles are processed, the defects can be more effectively distinguished and identified;

[0010] Further, the gray mean and variance in each window are calculated by using the fixed window sliding method, a dynamic threshold benchmark graph is constructed, the interference of uneven light and noise can be effectively eliminated, the local contrast of the image is improved, and the subsequent abrasive particle edge extraction is more fine; through the use of the dynamic threshold, the characteristics of different regions can be more accurately processed, and the shortcomings of the traditional fixed threshold are avoided; secondly, the image gradient direction is calculated by using the Sobel operator, and the direction consistency between adjacent pixels is judged, the direction constraint factor is constructed, the edge region of the abrasive particle can be accurately extracted, the identification ability of important structures in the image is enhanced, the direction consistency helps to eliminate noise and misidentification, and the identification accuracy of the abrasive particle edge is improved;

[0011] In the aspect of segmentation, the dynamic threshold benchmark is combined with the direction constraint factor through convolution to obtain a segmentation threshold surface, thereby improving the accuracy of image segmentation, being able to flexibly adapt to different regional characteristics, and combining the gray scale and direction information to ensure the coherence of the abrasive grain edge and the accurate extraction of local details; after edge extraction, the adjacent pixels in the edge binary image are connected by using the connected domain algorithm to accurately extract the abrasive grain region, ensuring the completeness of the abrasive grain boundary and avoiding the recognition difficulty caused by scattered edge information; finally, the local feature image of the abrasive grain is obtained through the multi-angle polarized light imaging technology, and the contour and texture analysis of the real metal abrasive grain are collected to construct a cascade classifier, which makes the classification of abrasive grain defects more intelligent, can automatically identify different types of defects (such as wear), and improves the accuracy and efficiency of detection; combined with three-dimensional topography analysis and surface texture constraint, the final output recognition result is more in-depth, and different abrasive grain defects can be accurately classified and analyzed. BRIEF DESCRIPTION OF DRAWINGS

[0012] Figure 1 The main flow of a metal abrasive grain defect detection method in lubricating oil of embodiment one is shown in the figure.

[0013] Figure 2 The flow chart of the double-channel image enhancement of a metal abrasive grain defect detection method in lubricating oil of embodiment one is shown in the figure.

[0014] Figure 3 The double-channel image enhancement schematic diagram of a metal abrasive grain defect detection method in lubricating oil of embodiment one is shown in the figure.

[0015] Figure 4 The flow chart of the extraction of the abrasive grain region of a metal abrasive grain defect detection method in lubricating oil of embodiment one is shown in the figure.

[0016] Figure 5 The flow chart of the collection of the actual metal abrasive grain image for the abrasive grain region identification of a metal abrasive grain defect detection method in lubricating oil of embodiment one is shown in the figure.

[0017] Figure 6 The flow chart of a metal abrasive grain defect detection system in lubricating oil of embodiment two is shown in the figure.

[0018] Label: collection module 10; identification module 20. DETAILED DESCRIPTION

[0019] The technical solutions of the present application will be described clearly and completely in combination with the drawings. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of protection of the present application.

[0020] The application will be described in further detail below with specific embodiments and in conjunction with the drawings.

[0021] Embodiment one

[0022] As shown in the figure, the embodiment of the application provides a method for detecting metal abrasive particles in lubricating oil, comprising the following operation steps: Figure 1

[0023] S1: Scanning the lubricating oil sample collected by the machine using a multi-spectral imaging system to obtain dual-channel images in the visible light band and the near-infrared band; the dual-channel images include visible light images and near-infrared images;

[0024] It should be noted that the multi-spectral imaging system is used to scan the lubricating oil sample to collect images in the visible light band and the near-infrared band; the images in these two bands contain different types of information, which helps to capture the characteristics of metal abrasive particles in the lubricating oil sample;

[0025] The visible light image provides high-resolution surface details, while the near-infrared image can penetrate deeper areas to obtain different information about abrasive particles and lubricating oil; it is usually used to capture the details and color differences on the surface of the lubricating oil, and has good performance for the shape and surface characteristics of abrasive particles;

[0026] The near-infrared band can obtain hidden information in the lubricating oil, and provide more information about the distribution and hidden abrasive particles in the metal abrasive particles in the lubricating oil, especially the spectral differences of different materials such as metal and oil;

[0027] By using images in two different bands, the recognition ability of the system for abrasive particles can be enhanced, especially since metal abrasive particles in the lubricating oil may have different physical characteristics, using information in these two bands helps to more accurately identify and classify;

[0028] S2: respectively enhancing the visible light image and the near-infrared image, and performing fusion analysis on the enhanced images to obtain a dual-channel fusion image; performing sliding window on the dual-channel fusion image to construct a direction constraint factor; performing threshold segmentation on the dual-channel fusion image through the direction constraint factor to obtain an abrasive particle region; collecting actual metal abrasive particle defect images to analyze the abrasive particle region to obtain a recognition result;

[0029] It should be noted that the collected visible light image and near-infrared image are respectively enhanced; image enhancement processing can be performed by improving the contrast, sharpening, and denoising of the image, increasing the details in the image, and making the characteristics of the abrasive particle region more obvious;

[0030] ​Subsequently, the enhanced visible light image and the near-infrared image are fused and analyzed to generate a dual-channel fused image, so that the advantages of the two wavebands can be fused to further improve the information quantity and clarity of the image; image enhancement is to improve the quality of the image and make the details clearer, which is particularly important for the detection of abrasive particle defects; fusion analysis is to complement the information in the visible light and near-infrared images and combine the advantages of the two to provide more comprehensive features;

[0031] The enhanced image can clearly show the contour, shape and texture features of the abrasive particles, so that the subsequent image analysis and abrasive particle detection are more accurate; and image fusion combines the advantages of the two wavebands to help identify more types of abrasive particle defects;

[0032] For the dual-channel fused image, the gradient direction of each pixel point in each window is calculated; the gradient direction information is used to determine the direction of the key features in the image, which helps to emphasize the features in a specific direction to further refine the image processing, thereby constructing a direction constraint factor; the introduction of the direction constraint factor makes the identification of abrasive particle features more directional, especially the shape and arrangement of the abrasive particles;

[0033] The dual-channel fused image is threshold segmented by the direction constraint factor, and the purpose is to accurately extract the abrasive particle region from the dual-channel fused image through threshold segmentation. The direction constraint factor plays a guiding role (i.e., reflecting the shape according to the direction) to ensure that only the regions that meet the abrasive particle features are segmented out;

[0034] Specifically, as shown in Figure 2 As shown in Figure 3 In step S2, the visible light image and the near-infrared image are enhanced respectively, and the enhanced images are fused and analyzed to obtain a dual-channel fused image; a sliding window is performed on the dual-channel fused image to construct a direction constraint factor; the dual-channel fused image is threshold segmented by the direction constraint factor to obtain an abrasive particle region; an actual metal abrasive particle defect image is collected to analyze the abrasive particle region to obtain an identification result, and the specific operation steps are as follows:

[0035] In steps S21-S23, it is briefly illustrated that the adaptive histogram equalization algorithm is used to calculate the pixel gray value of the visible light image to construct a gray histogram; the pixel point distribution state is judged by the pixel gray value, and the pixel points with high threshold gray values in the gray histogram are clipped and divided according to the set threshold to obtain the execution process of the enhanced visible light channel image;

[0036] Meanwhile, the multi-scale Retinex enhancement algorithm is used to perform convolution operation on the near-infrared image, so as to identify the illumination component and the reflection component in the near-infrared image; the near-infrared image is decomposed and reconstructed by using the illumination component and the reflection component, so as to obtain an enhanced near-infrared channel image;

[0037] The visible light channel image and the near-infrared channel image are normalized respectively, and then are fused by assigning weights, so as to obtain a dual-channel fusion image, which is specifically as follows:

[0038] S21: The adaptive histogram equalization algorithm is used to divide the visible light image into uniform region blocks, so as to obtain local region blocks;

[0039] The internal pixel gray values of each local region block are calculated, and a gray histogram of each local region block is constructed;

[0040] The amplitude threshold is set by using dynamic restriction of contrast stretching amplitude;

[0041] It is judged whether the pixel gray value of each gray histogram is greater than the amplitude threshold;

[0042] If yes, the gray histogram is clipped, and the pixel gray values of the clipped part of the histogram are evenly added to other pixel points in the clipped gray histogram (i.e. the remaining pixel points in the gray histogram after clipping);

[0043] The gray histograms are spliced after the amplitude threshold is set for each gray histogram, so as to obtain an enhanced visible light channel image;

[0044] It should be noted that the whole visible light image is divided into multiple small regions for independent processing, so as to solve the problem of local over-enhancement or under-enhancement caused by global histogram equalization; for edge blocks (such as the four boundaries of the visible light image), the part beyond the range of the visible light image is filled by mirror copying adjacent pixels, so as to avoid artificial artifacts at the boundary and maintain the continuity of the abrasive grain edge (i.e. the visible light image is divided into multiple small regions, and local enhancement is performed block by block, which helps to avoid the problem of local over-enhancement or under-enhancement caused by global histogram equalization; especially in the abrasive grain image, over-enhancement may cause noise amplification, and under-enhancement may cause detail loss; by setting the amplitude threshold, the local region is prevented from being over-enhanced, and for the details of the abrasive grain, over-enhancement may introduce noise, so that the abrasive grain features are not clear, and it is ensured that the pixel points in the gray histogram do not exceed the threshold, so as to ensure the balance of enhancement);

[0045] Dynamic limit contrast stretching amplitude is a key technology in image enhancement, and its core purpose is to prevent local area over-enhancement from leading to noise amplification or detail loss. For example, the gray level histogram of each local area block has 20 pixel points for a gray value of 50 and 50 pixel points for a gray value of 100. A dynamic limit contrast stretching amplitude (i.e., a clipping threshold) is set to indicate that the number of pixel points of any gray value cannot exceed 30. The pixel points exceeding 30 in the histogram are truncated (e.g., the original 50 pixel points for a gray value of 100 are retained as 30), and then redistributed: the 20 pixel points that are clipped are evenly added to all gray values (each gray level + 20 / 256 ≈ 0.08 pixel points, which are the “other pixel points in the clipped gray level histogram” in the step), to enhance the gray difference between the abrasive grain edge and the oil film background, and to enhance the edge details to make the abrasive grain boundary clearer.

[0046] S22: The multi-scale Retinex enhancement algorithm is used to perform a multi-scale Gaussian filter kernel smoothing on the near-infrared image, and a multi-scale corresponding distribution convolution kernel operation is performed on the smoothed near-infrared image to generate an illumination component and a near-infrared smoothed image (i.e., the smoothed near-infrared image);

[0047] The near-infrared image and the near-infrared smoothed image are subjected to logarithmic operation by using the Retinex formula of the multi-scale Retinex enhancement algorithm, and the illumination component is removed at the same time, to calculate a reflection component;

[0048] The reflection components of multiple scales are assigned weights and merged to obtain a near-infrared image of a fusion component;

[0049] The illumination component and the reflection component of the near-infrared image of the fusion component are decomposed and reconstructed to obtain an enhanced near-infrared channel image through combination;

[0050] It should be noted that the multi-scale Retinex enhancement algorithm applies different scales of Gaussian filters to the near-infrared image to smooth the near-infrared image, to obtain multiple scales of illumination components and reflection components. The illumination component contains the illumination information of the near-infrared image, and the reflection component represents the real reflection characteristics of the object. The multi-scale processing of the near-infrared image can extract different levels of details from different scales, and thus enhance the local features of the image;

[0051] The reflection component and the illumination component are separated from the near-infrared image that may contain abrasive grains, so that the real morphology of the abrasive grains can be better recognized instead of being affected by illumination or shadow. Through multi-scale fusion, the details and clarity of the near-infrared image are enhanced, and removing the illumination component helps to eliminate the changes in illumination, so that the shape and texture of the abrasive grains can be accurately recognized;

[0052] S23: calculating gradient values in horizontal and vertical directions of each pixel point in the visible light channel image; generating a gradient map of the visible light channel image according to the gradient values; calculating gradient amplitudes of each pixel point according to the gradient map, and constructing a visible light gradient amplitude map;

[0053] normalizing the visible light gradient amplitude map and the reflection component of the near-infrared channel image, and fusing them according to weights to obtain a dual-channel fusion image;

[0054] It should be noted that by calculating the gradient values in the horizontal and vertical directions of each pixel point, a gradient map of the image is obtained, and then a gradient amplitude map is further calculated to emphasize the edges and texture details of the image. After normalizing the gradient amplitude map of the visible light image and the reflection component of the near-infrared image, fusing them according to weights, the useful information of the two channels can be combined to obtain a fusion image.

[0055] The gradient amplitude map emphasizes the edges and texture details of the abrasive particles, while the reflection component highlights the real surface information of the abrasive particles. The fusion of the two channels, especially in the case of enhanced edges and textures, helps to more accurately distinguish the morphology of the abrasive particles and reduce the interference of light and noise. By reasonable weighting, the characteristics of the visible light image and the near-infrared image can be better balanced to enhance the details and contrast of the abrasive particle region.

[0056] S24: performing sliding window on the dual-channel fusion image, calculating the gradient direction of each window pixel, and constructing a direction constraint factor; performing convolution operation on the direction constraint factor to obtain a segmentation threshold surface, analyzing the dual-channel fusion image using the segmentation threshold surface to obtain an abrasive particle region; performing specific analysis of local features on the abrasive particle region to form a local feature region; collecting actual metal abrasive defect images to identify the local feature region;

[0057] It should be noted that by calculating the gradient direction of each window using sliding window, a direction constraint factor is constructed for analyzing the direction information in the image. The direction constraint factor is convolved to generate a segmentation threshold surface, thereby segmenting the abrasive particle region. Through analysis of the abrasive particle region, the defect information of the abrasive particle (i.e., the local feature region) is obtained to assist in defect recognition.

[0058] Local feature analysis refers to in-depth analysis of the details inside the abrasive particle region to identify the specific shape, texture or other features related to the abrasive particle. Local feature analysis of the abrasive particle region can accurately capture the unique morphology and texture of the abrasive particle, which helps to improve the classification and recognition accuracy of the abrasive particle, especially for similar morphology abrasive particles.

[0059] By collecting actual metal abrasive particle defect images, combined with analysis of local feature regions, higher precision abrasive particle recognition can be realized; by comparing actual images with local features, it can be determined whether the recognition result of the abrasive particle region exists abrasive particles;

[0060] Specifically, as shown in Figure 4 the step S24, the sliding window of the double-channel fusion image is performed, the gradient direction of each window pixel point is calculated, and the direction constraint factor is constructed; the direction constraint factor is subjected to convolution operation to obtain a segmentation threshold surface, and the double-channel fusion image is analyzed by using the segmentation threshold surface to obtain an abrasive particle region; the abrasive particle region is specifically analyzed and local features are formed to form a local feature region; the actual metal abrasive particle defect image is collected to identify the local feature region, and the specific operation steps are as follows:

[0061] S241: The fixed window sliding window of the double-channel fusion image is performed, and the gray mean value and the gray variance of the pixel gray value of each fixed window pixel point are calculated;

[0062] The gray mean value and the gray variance calculated by using each fixed window are mapped to construct a dynamic threshold reference map;

[0063] It should be noted that the double-channel fusion image is divided into a plurality of small windows, the pixels in each window are statistically calculated, the gray mean value and the gray variance of each window are obtained, which can help to capture the local area light and contrast change; the gray mean value represents the average brightness in the window, and the gray variance represents the brightness fluctuation degree of the image (i.e. if the high variance may represent the edge or texture region in the image);

[0064] By calculating the gray mean value and the variance of each window, a dynamic threshold reference map can be constructed for threshold setting in subsequent image segmentation or edge detection; by statistically calculating the gray value of each fixed window, the influence of uneven light or noise can be effectively eliminated, the dynamic threshold reference map can flexibly adapt to the characteristics of different regions, avoid the over-enhancement or under-enhancement problems that may be encountered when using fixed threshold, and help to improve the contrast between abrasive particles and background through more accurate threshold setting, so that the subsequent abrasive particle edge extraction is more fine;

[0065] S242: The gradient direction of each pixel point in the double-channel fusion image is calculated by using the Sobel operator;

[0066] The direction difference value of the gradient direction of each pixel point and the gradient direction of the adjacent pixel point is calculated;

[0067] A preset direction consistency threshold u is set; it is judged whether the direction difference value is greater than the direction consistency threshold u;

[0068] If yes, it is determined that the directions between adjacent pixel points are consistent;

[0069] A direction constraint factor is constructed according to the consistent directions between adjacent pixel points;

[0070] It should be noted that the gradient direction of each pixel in the dual-channel fusion image is calculated, that is, the local change trend of the pixel; the gradient direction reflects the direction information of the edge in the dual-channel fusion image, which is helpful for positioning the edge region;

[0071] By calculating the gradient direction difference between adjacent pixel points, it can be determined whether they belong to the same direction; when the consistency of the directions between adjacent pixels is strong, it indicates that they may belong to the same object or region;

[0072] A threshold is set, and when the gradient direction difference of adjacent pixels is less than the threshold, it is considered that the directions are consistent, and a direction constraint factor is constructed to constrain the edge region; the gradient direction calculation is helpful for accurately extracting the edge of the abrasive grain, and the direction consistency judgment is helpful for eliminating noise and misidentification, avoiding false edges due to local changes, and better identifying the structure and texture of the abrasive grain by strengthening the direction consistent region;

[0073] S243: convolve the dynamic threshold reference map and the direction constraint factor to obtain a segmentation threshold surface;

[0074] It should be noted that the dynamic threshold reference map and the direction constraint factor are convolved to generate a segmentation threshold surface, which combines the gray level characteristics and edge direction characteristics of the dual-channel fusion image, thereby improving the accuracy of image segmentation;

[0075] The surface sets different segmentation thresholds for each region of the dual-channel fusion image, which can flexibly handle the characteristics of different regions; combined with the gray level and direction information, the coherence and local details of the abrasive grain edge can be considered in the segmentation process. This provides strong support for the accurate segmentation of the abrasive grain region in the subsequent process, ensuring that the abrasive grain can still be accurately identified in a complex background;

[0076] S244: filter the pixel points with consistent directions between adjacent pixel points in the gradient direction of the pixel points of the dual-channel fusion image calculated by the Sobel operator as the suspected edge pixel points of the dual-channel fusion image;

[0077] An edge gradient map is established for the suspected edge pixel points;

[0078] The edge gradient map is binarized using the segmentation threshold surface to obtain an edge binary map;

[0079] It should be noted that the gradient direction calculated by the Sobel operator is used to determine which pixel points are suspected edge pixels, and usually these pixel points have larger gradient values and obvious changes. The gradient values of the suspected edge pixels are constructed into an image to highlight the possible edge region;

[0080] By screening suspected edge pixels and establishing an edge gradient map, the edge features of the abrasive particles can be effectively extracted, the key contour information in the abrasive particle image can be found, and the basis for positioning and analyzing the abrasive particles is provided, the interference of the background and noise is reduced, and the edge of the abrasive particle is more clear and visible;

[0081] S245: edge pixel extraction and connection of the edge binary image are performed by using a connected domain algorithm to obtain an abrasive particle region;

[0082] It should be noted that the algorithm is used to extract and connect adjacent edge pixels from the edge binary image to form a complete region. This algorithm can effectively identify continuous and similar regions in the image;

[0083] By extracting these connected regions, the abrasive particle region in the image can be accurately positioned; by connected domain analysis, the dispersed edge information is connected into a complete region to ensure that the boundary of the abrasive particle can be completely identified; the connected domain algorithm effectively extracts the coherent region of the abrasive particle, which helps to accurately distinguish the abrasive particle from the background and provides reliable region data for subsequent defect analysis;

[0084] S246: multi-angle polarized light imaging is performed on the abrasive particle region to obtain a local feature region of the abrasive particle region; an actual metal abrasive particle defect image is collected for analysis of the contour, and a contour topological condition constraint is constructed; the contour of the actual metal abrasive particle defect image is further analyzed and reconstructed to generate a three-dimensional abrasive particle morphology; height field analysis is performed on the three-dimensional abrasive particle morphology to construct a surface texture condition constraint; a cascade classifier is constructed based on the contour topological condition constraint and the surface texture condition constraint, and an output recognition result is performed on the local feature region;

[0085] It should be noted that multi-angle polarized light imaging uses the change of the reflection of polarized light on the surface of an object in different directions to obtain more abundant surface information, which can clearly reveal the surface texture, contour and defect region, especially on the surface of a metal abrasive particle. It helps to enhance the contrast of surface details and identify small defects such as wear and cracks; the image obtained by multi-angle polarized light imaging has strong contrast, and more detailed information about the local abrasive particle can be extracted, such as surface microstructure, shape and texture;

[0086] The actual metal abrasive defect image is collected for profile and texture analysis, and a cascade classifier is constructed; this step analyzes the profile and texture of the actual collected metal abrasive image; by extracting the geometric shape, edge smoothness, surface texture and other features of the abrasive, the morphological characteristics of the abrasive can be described more clearly; profile and texture analysis mainly focuses on the geometric shape (such as circular, elliptical, etc.) and surface structure (such as smooth, rough, etc.) of the abrasive; these features are crucial for distinguishing different types of abrasive defects, such as cutting abrasive, sliding abrasive, fatigue spalling abrasive, etc.; analyzing the profile and texture features helps to capture the microscopic morphology of the abrasive, and thus helps to more accurately determine the defect type; texture analysis can reveal the microscopic changes on the surface, while profile analysis provides specific features of the macroscopic morphology;

[0087] By constructing a cascade classifier, different types of abrasive defects can be efficiently classified according to the shape and texture features of the abrasive, which can improve the intelligent level of the entire defect recognition system, enabling it to automatically classify abrasives and ultimately output the recognition result of the abrasive defect type;

[0088] It is found through research that by using multi-angle polarized light imaging and generating a normal vector distribution map, the detailed information of the surface abrasive area can be captured more accurately, thereby improving the recognition accuracy of abrasive defects. The original scheme may only rely on a single image processing technology or traditional visual detection method, while the expanded scheme combines polarized light imaging, normal vector calculation and three-dimensional topography reconstruction, so that the local features of the defect surface are analyzed more comprehensively and accurately. The specific analysis is as follows:

[0089] Specifically, as shown in Figure 5 step S246, multi-angle polarized light imaging is performed on the abrasive area to obtain a local feature area of the abrasive area; the actual metal abrasive defect image is collected for profile analysis, and a profile topology condition constraint is constructed; and the profile of the actual metal abrasive defect image is further analyzed and reconstructed to generate a three-dimensional abrasive topography; height field analysis is performed on the three-dimensional abrasive topography to construct a surface texture condition constraint; based on the profile topology condition constraint and the surface texture condition constraint, a cascade classifier is constructed, and the local feature area is output to obtain a recognition result. The specific operation steps are as follows:

[0090] S2461: Use multi-angle polarized light imaging to collect multi-angle polarized light of the abrasive area to obtain multi-angle polarized images under different polarization states;

[0091] The multi-angle polarized image contains the normal vector of the reflected component;

[0092] The normal vector direction of each pixel point inside the multi-angle polarized image is calculated;

[0093] A normal vector distribution map is formed using the normal vector direction of each pixel.

[0094] It should be noted that by changing the polarization angle of the light source, the target area (abrasive area) is illuminated with light of different polarization states (i.e., linear polarization and circular polarization), and reflected light images are obtained at different angles. The intensity and properties of the reflected light will change under different polarization angles, providing more information for obtaining surface details.

[0095] In multi-angle polarized light imaging, not only is the reflected image acquired, but the normal vector can also be calculated through the reflection characteristics of polarized light. The normal vector describes the normal direction of each point on the surface (i.e., a vector perpendicular to the surface, which fully describes the local direction and tilt of that point). Under different reflection states, the change of the normal vector can reflect the changes in the surface shape and angle. Therefore, multi-angle polarized images can contain reflection components. In multi-angle polarized images, the reflection components can be used to derive the normal vector of each point on the surface, thereby describing the microstructure and morphology of the surface.

[0096] By processing the reflection information of multi-angle polarized images, each pixel in the image is analyzed to calculate the direction of its normal vector. The direction of the normal vector describes the degree of tilt (i.e., the degree of tilt is an important indicator of the local unevenness of the surface, which can reflect the shape and size of the microstructure (e.g., abrasive grains)) and orientation of each point on the surface, and can reveal the local features of each tiny point on the surface. The constructed normal vector distribution map summarizes and plots the normal vectors of each pixel, which can visualize the surface features of different regions, especially the differences between different abrasive grains, worn areas, etc.

[0097] S2462: Normalize the normal vector distribution map, and calculate the gradient information (i.e., gradient magnitude and gradient direction) in the x and y directions (i.e., the direction of the pixel) of the normal vector distribution map using the direction of the normal vector, to obtain the components of the normal vector. The calculation formula is as follows:

[0098] ;

[0099] In the formula, This is represented as the average value of the normal vectors of the pixels in the normal vector distribution map (i.e., when...). A larger value indicates a higher average rate of change of the current surface normal vector (larger surface tilt); a smaller value indicates a relatively flat surface, used to measure the overall characteristics of the current surface morphology.

[0100] The average value of the gradient amplitude of the normal vector in the normal vector distribution map (i.e., the intensity of the change in the direction of the normal vector gradient, reflecting the details and complexity of the surface features, when the gradient amplitude is large, it indicates that the surface inclination or concave-convex changes sharply; when the gradient amplitude is small, it indicates that the surface is relatively smooth);

[0101] The gradient change rate in the direction of the pixel point ; The gradient change rate in the direction of the pixel point ;

[0102] The gradient change rate in the direction of the pixel point ;

[0103] The total number of pixel points as a normalization factor;

[0104] The reference weight (i.e., assigning weights to different factors in the formula, adjusting the importance of the calculation result, by adjusting the value of , the influence of local and global features can be balanced, and different analysis requirements can be adapted);

[0105] The dispersion degree of the normal vector distribution map (i.e., used to measure the overall change range of the current normal vector distribution, a larger indicates that the change range of the current surface features is larger, and the surface may be more complex; a smaller indicates that the surface features are relatively smooth or regular);

[0106] The correlation factor of the normal vector gradient in the and directions (i.e., used to measure the correlation of the normal vector gradient in two directions, reflecting the overall change trend of the surface features; when the following covariance is large and the variance is small, it indicates that the change trends of the two direction gradients are consistent; when the covariance is small or negative, it indicates that the change trends of the two direction gradients are relatively independent or opposite);

[0107] The covariance of the normal vector gradient in the and directions;

[0108] The variance of the normal vector gradient in the direction;

[0109] The variance of the normal vector gradient in the direction of ;

[0110] Norm [ ] represents the normalization function (i.e., used to normalize the calculation results to a certain range for comparison and analysis; ensure the consistency of the range of formula output values, avoid the distortion of the results caused by the too large or too small range of values);

[0111] The components of the normal vector are used to establish a Poisson equation, and the Poisson equation is solved (i.e., the solution of the Poisson equation can be solved by using discrete Fourier transform or fast Poisson solver), to obtain the height field z (x, y) (i.e., the obtained height field z (x, y) is the same size as the abrasive particle area, and the height field reflects the three-dimensional topographic information of the possible abrasive particle surface);

[0112] It should be noted that different lighting conditions, reflection angles, etc. may affect the size of the normal vector, and normalization makes all normal vectors have the same scale, ensuring the accuracy of the calculation;

[0113] By calculating the gradients of the normal vector distribution map in the x and y directions, the change trend of the surface in these two directions can be obtained; specifically, the gradient describes the spatial variation of the normal vector, that is, the inclination and change direction of the surface at different positions; the components of the normal vector are the projections of the gradient in different directions, which are usually divided into x and y components, and at the same time, the x component reflects the inclination or "slope" information in the x direction of the surface, and the y component reflects the inclination information in the y direction of the surface; the inclination information of the x component and the y component can reflect the inclination of the vertical direction n (i.e., the normal vector is a unit vector perpendicular to the local tangent plane at the point (i.e., the pixel point), therefore, according to the inclination information of the x direction and the y direction, the inclination information of the vertical direction n can be reflected); calculating these components can help us understand the height changes of the surface in the horizontal direction (x, y) and the vertical direction n;

[0114] The Poisson equation is a common partial differential equation, which is widely used in physics, especially in the calculation related to fields (such as electric field, gravitational field, etc.); here, the Poisson equation is used to describe the relationship between the normal vector component and the surface height (i.e., the z direction height), so a Poisson equation related to the height field (i.e., the surface topography) is constructed, and the three-dimensional morphology of the surface is derived through the normal vector direction (i.e., the relationship between the gradients of the x and y directions and the vertical direction n);

[0115] S2463: Numerical differentiation of the height field z (x, y) to calculate the partial derivative (i.e., the first-order partial derivative and the second-order partial derivative );

[0116] calculating a Gaussian curvature of each pixel point in the abrasive particle region by using the partial derivative;

[0117] calculating an average curvature of the pixel point by using the Gaussian curvature;

[0118] screening the pixel points of the local maximum Gaussian curvature and the pixel points of the local minimum Gaussian curvature in the abrasive particle region by using the average curvature of the pixel point;

[0119] respectively clustering the pixel points of the local maximum Gaussian curvature and the pixel points of the local minimum Gaussian curvature by using the surrounding Gaussian curvature of the pixel point, to obtain a local feature region (i.e. a certain microscopic structure detail of the abrasive particle (i.e. a cutting abrasive particle, a sliding abrasive particle, a fatigue spalling abrasive particle) that may exist);

[0120] It should be noted that by calculating the partial derivative of the height field, the first and second partial derivatives are obtained to describe the rate of change of the surface shape; the degree of change of each pixel point on the surface can be quantified, thereby more geometric information is extracted to help analyze the morphological characteristics of the abrasive particle; the Gaussian curvature helps to identify the concave-convex, texture and other features of the surface, and identify the local details of the abrasive particle; by calculating the Gaussian curvature and the average curvature, the maximum and minimum values in the abrasive particle region are screened to find the pixel points with significant features; based on the clustering of the local maximum and minimum values, the regions that may represent different abrasive particle morphologies are obtained;

[0121] S2464: collecting an actual metal abrasive particle defect image;

[0122] preprocessing the metal abrasive particle defect image to obtain a plurality of defect abrasive particle contours (i.e. the metal abrasive particle defect image is known to have a defect abrasive particle, and the defect abrasive particle contour can be directly extracted by using an edge detection algorithm);

[0123] labeling the morphological defect feature statistics of the plurality of defect abrasive particle contours to obtain an abrasive particle defect type (i.e. the plurality of defect abrasive particle contours are cutting abrasive particles, sliding abrasive particles, and fatigue spalling abrasive particles);

[0124] extracting preliminary geometric features (i.e. such as the closure of the contour, the edge smoothness, the local convex-concave condition, etc.) of each abrasive particle defect type;

[0125] constructing a contour topological condition constraint for each abrasive particle defect type according to the extracted preliminary geometric features;

[0126] It should be noted that by analyzing the characteristics of the metal abrasive particle in terms of shape, contour, etc., a classification system including cutting abrasive particles, sliding abrasive particles, fatigue spalling abrasive particles and other defect types is established, and corresponding contour topological constraint conditions are defined for each type;

[0127] The profile of the metal abrasive particle is obtained by the collected actual metal abrasive particle defect image. The profile of the metal abrasive particle can more intuitively know the defect morphology of the actual metal abrasive particle. According to the defect morphology of the actual metal abrasive particle, the cutting abrasive particle, the sliding abrasive particle and the fatigue spalling abrasive particle possibly existing in the lubricating oil are marked. The preliminary geometric features of various abrasive particles, that is, the closure of the profile, the edge smoothness and the local convex-concave condition are extracted. The constraint conditions of the defects of various types of cutting abrasive particles, sliding abrasive particles and fatigue spalling abrasive particles are constructed according to the closure of the profile, the edge smoothness and the local convex-concave condition, that is, the detection standard is defined.

[0128] That is, for the cutting abrasive particle, the edge is usually sharp and regular, and the profile may require small curvature fluctuation and high symmetry. For the sliding abrasive particle, the profile may present a certain arc or irregular elongated shape, and the edge requires a certain continuity but a large local curvature change amplitude. For the fatigue spalling abrasive particle, the profile is often incomplete with obvious defect area, and the topological feature may be profile fracture and more missing areas.

[0129] In steps S2465-S2466, the minimum circumscribed ellipse fitting of the plurality of defect abrasive particle profiles is mainly performed according to the preliminary geometric features, the elliptic center coordinates and the major axis and the minor axis of the defect abrasive particle profile are calculated, the angle between the major axis and the minor axis of the defect abrasive particle profile is calculated, and the elliptic main direction angle of the defect abrasive particle profile is analyzed. The defect abrasive particle profile is reconstructed by the elliptic center coordinates, the major axis and the minor axis and the elliptic main direction angle of the defect abrasive particle profile, and the three-dimensional abrasive particle morphology is obtained.

[0130] The Gaussian curvature of the three-dimensional abrasive particle morphology is calculated to obtain the energy distribution. The surface texture condition constraint is constructed according to the energy distribution. The specific steps are as follows:

[0131] S2465: The minimum circumscribed ellipse fitting of the plurality of defect abrasive particle profiles is performed according to the preliminary geometric features, and the elliptic center coordinates of the defect abrasive particle profile are obtained.

[0132] The longest distance and the shortest distance of the elliptic center coordinates of the edge pixel points of the defect abrasive particle profile are calculated as the major axis and the minor axis of the defect abrasive particle profile.

[0133] The angle of the included angle of the major axis and the minor axis of the defect abrasive particle profile is calculated according to the elliptic center coordinates, and the elliptic main direction angle (that is, the profile direction) of the defect abrasive particle profile is determined according to the angle of the included angle of the major axis and the minor axis.

[0134] a length ratio of a long axis and a short axis of the defect abrasive particle profile is calculated to obtain an abrasive particle shape elongation index (i.e. the elongation degree of the shape of the abrasive particle profile is determined by the length ratio, that is, whether it is close to a circular shape, the greater the value of the abrasive particle shape elongation index, the higher the elongation degree of the shape of the abrasive particle; the lower the value of the abrasive particle shape elongation index, the closer the shape of the abrasive particle is to a circular shape);

[0135] The three-dimensional abrasive particle morphology is reconstructed by using the abrasive particle shape elongation index and the elliptical major direction angle of the defect abrasive particle profile, to obtain a three-dimensional abrasive particle morphology (i.e. the three-dimensional abrasive particle morphology reflects the reconstruction of the defect abrasive particle profile, that is, the profile area formed after reconstruction, reflecting the shape and structural characteristics of the abrasive particle defect; the core of the three-dimensional morphology reconstruction is to restore the height field of the abrasive particle surface in the subsequent step, from which the local geometric information (such as Gaussian curvature, average curvature, etc.) is extracted to describe the convexity, concavity, and texture characteristics, reflecting the micro-morphology, details, and overall shape structure (i.e. step 2466));

[0136] It should be noted that the elliptical center, long axis, short axis and their included angle of the abrasive particle profile are calculated by the elliptical fitting technology, and then the directionality of the abrasive particle is obtained; fitting an ellipse helps to capture the morphological characteristics of the abrasive particle, especially to determine the main direction and shape elongation index of the abrasive particle, which helps to qualitatively analyze the abrasive particle; the three-dimensional abrasive particle morphology is reconstructed by using the shape elongation index and the elliptical major direction angle; through the three-dimensional reconstruction, the morphological characteristics of the abrasive particle defect can be more truly restored, and the three-dimensional structure of the abrasive particle can be more intuitively analyzed;

[0137] S2466: calculating the height field of the three-dimensional abrasive particle morphology;

[0138] The Gaussian curvature of the height field of the three-dimensional abrasive particle morphology is calculated.

[0139] The peak points and valley points (i.e. the maximum and minimum extreme points) in the height field are counted to determine the spatial distribution of the three-dimensional abrasive particle morphology of the peak points and valley points.

[0140] The three-dimensional abrasive particle morphology is directionally analyzed by using wavelet transform and the spatial distribution obtained.

[0141] The pixel points in the three-dimensional abrasive particle morphology are subjected to multi-scale and multi-directional Gaussian curvature transform frequency domain signals, and the amplitude of each pixel point is calculated.

[0142] The energy distribution is obtained by calculating the energy of each pixel point (i.e., the energy is usually represented by the square of its amplitude;), which is the texture direction distribution of the three-dimensional abrasive grain morphology (i.e., the directionality analysis is to process the curvature of the three-dimensional abrasive grain morphology, and the energy distribution in the frequency domain is calculated by the multi-angle and multi-direction calculation of the wavelet transform point, which can reflect the texture characteristics in different directions, thereby serving as the direction distribution);

[0143] The average energy value of the energy distribution is calculated, and the maximum energy value is found by the energy distribution; the energy concentration ratio is calculated by the average energy value and the maximum energy value (i.e., the energy concentration ratio reflects whether the energy is concentrated in a main direction, that is, the description of the direction symmetry and dispersion of the surface texture);

[0144] The energy standard deviation of the energy distribution is calculated; the energy concentration ratio and the energy standard deviation are encoded to form the surface texture condition constraint;

[0145] It should be noted that the Gaussian curvature of the height field is calculated on the three-dimensional abrasive grain morphology; the Gaussian curvature can help identify the direction of surface change and extract the texture features of the abrasive grain surface; the energy distribution is obtained by analyzing the frequency domain signal of the Gaussian curvature through wavelet transform, thereby analyzing the texture direction of the abrasive grain surface; the directionality analysis helps reveal the texture rules of the abrasive grain surface and provides texture features for classification; the energy concentration ratio and the energy standard deviation are calculated to reflect the directionality features of the texture; through the energy distribution analysis, the symmetry and dispersion of the abrasive grain surface texture can be judged, further improving the accuracy of the classification;

[0146] S2467: The cascade classifier is constructed by using the contour topological condition constraint and the surface texture condition constraint;

[0147] It should be noted that the decision tree is used to select the geometric features (such as the major axis ratio of the minimum circumscribed ellipse and the principal direction angle) of the contour topological condition constraint to construct the first-level decision tree; through the training data set, the judgment threshold of each feature is set, and the defects are preliminarily divided according to the basic type (such as cutting, sliding or fatigue peeling); the contour topological condition constraint reflects the shape and appearance of the abrasive grain defect type;

[0148] For the appearance of each basic defect type screened out in the first stage, the surface texture encoding (including the curvature, the topological relationship of the saddle point and the peak and valley point, and the anisotropy coefficient of the surface ripple direction (i.e., the energy concentration ratio reflects the main direction, that is, the description of the direction symmetry and dispersion of the surface texture, so according to the symmetry and dispersion, different directions are known, that is, the anisotropy coefficient)) is extracted to construct the second-level decision tree; the training samples are used to further subdivide the defect subtypes, and the defects with texture detail differences in the same basic type are classified in detail;

[0149] The two-stage classifier is cascaded to form a whole decision tree cascade model, that is, a cascade classifier, which can detect each type of detected abrasive particle defect; through step-by-step subdivision, the accuracy of abrasive particle classification can be improved, and complex abrasive particle defect types can be processed;

[0150] S2468: inputting the local feature area of the lubricating oil sample into the cascade classifier, and outputting the identification result of the lubricating oil sample (that is, outputting whether the lubricating oil sample has an abrasive particle defect, and if so, outputting the type of the abrasive particle);

[0151] It should be noted that the cascade classifier is used to identify and classify the abrasive particles in the lubricating oil sample; finally, through the cascade classifier, it can be determined whether the lubricating oil sample has an abrasive particle defect, and the type of the abrasive particle is determined, which helps to realize automatic abrasive particle defect detection;

[0152] Embodiment Two

[0153] As shown in Figure 6 The present application also provides a metal abrasive particle defect detection system in lubricating oil, comprising: a collection module 10; an identification module 20;

[0154] The collection module 10 is used for scanning the lubricating oil sample collected by the machine by using a multi-spectral imaging system, to obtain double-channel images in the visible light band and the near-infrared band;

[0155] The double-channel image includes a visible light image and a near-infrared image;

[0156] The identification module 20 respectively enhances the visible light image and the near-infrared image, and performs fusion analysis on the enhanced images to obtain a double-channel fusion image; a sliding window is used for the double-channel fusion image to construct a direction constraint factor; the double-channel fusion image is threshold segmented by using the direction constraint factor to obtain a abrasive particle region; an actual metal abrasive particle defect image is collected to analyze the abrasive particle region to obtain an identification result.

[0157] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present application, but not to limit them; a person skilled in the art can modify the technical solutions described in the above embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for detecting metal particle defects in lubricating oil, characterized by, The method comprises the following operation steps: The lubricating oil sample collected by the machine is scanned by using a multispectral imaging system to obtain double-channel images in visible light and near-infrared bands; the double-channel images comprise visible light images and near-infrared images; The visible light images and the near-infrared images are respectively enhanced, and the enhanced images are fused and analyzed to obtain a double-channel fused image; A sliding window is used for the double-channel fused image to construct a direction constraint factor; the double-channel fused image is threshold segmented by using the direction constraint factor to obtain a wear particle region; actual metal wear particle defect images are collected to analyze the wear particle region to obtain a recognition result; The visible light images are calculated by using an adaptive histogram equalization algorithm to obtain pixel gray values and construct a gray histogram; the pixel distribution state is judged by using the pixel gray values; the pixel gray values of the high threshold in the gray histogram are clipped and divided according to a set threshold to obtain an enhanced visible light channel image; A multiscale Retinex enhancement algorithm is used for convolution operation on the near-infrared image to identify the illumination component and the reflection component in the near-infrared image; the near-infrared image is decomposed and reconstructed by using the illumination component and the reflection component to obtain an enhanced near-infrared channel image; The visible light channel image and the near-infrared channel image are respectively normalized and weighted fused to obtain a double-channel fused image; The double-channel fused image is subjected to sliding window operation to construct a direction constraint factor; the double-channel fused image is threshold segmented by using the direction constraint factor to obtain a wear particle region; actual metal wear particle defect images are collected to analyze the wear particle region to obtain a recognition result; the specific operation steps are as follows: The double-channel fused image is subjected to sliding window operation to calculate the gradient direction of the pixel points in each window and construct a direction constraint factor; the direction constraint factor is subjected to convolution operation to obtain a segmentation threshold surface; the double-channel fused image is analyzed by using the segmentation threshold surface to obtain a wear particle region; the wear particle region is subjected to local feature analysis to form a local feature region; actual metal wear particle defect images are collected to analyze the local feature region to obtain a recognition result; The double-channel fused image is subjected to sliding window operation to calculate the gradient direction of the pixel points in each window and construct a direction constraint factor; the direction constraint factor is subjected to convolution operation to obtain a segmentation threshold surface; the specific operation steps are as follows: The double-channel fused image is subjected to fixed window sliding window operation; the pixel gray value gray mean value and the pixel gray value gray variance of the pixel points in each fixed window are calculated; The gray mean value and the gray variance calculated by using each fixed window are mapped to construct a dynamic threshold reference map; the gradient direction of each pixel point in the double-channel fused image is calculated by using a Sobel operator; ​ Calculate the direction difference of the gradient direction of each pixel point and the gradient direction of adjacent pixel points; a preset direction consistency threshold u; determine whether the direction difference is greater than the direction consistency threshold u; if yes, determine that the directions of adjacent pixel points are consistent; Construct a direction constraint factor according to the direction consistency between the adjacent pixel points; Convolve the dynamic threshold reference map and the direction constraint factor to obtain a segmentation threshold surface.

2. The method of claim 1, wherein the metal particle defect in the lubricating oil is detected by the method comprising: Analyze the dual-channel fusion image using the segmentation threshold surface to obtain a wear particle region; perform local feature specific analysis on the wear particle region to form a local feature region; collect actual metal wear particle defect images to identify the local feature region, and the specific operation steps are as follows: ​ Screen the pixel points with consistent directions between adjacent pixel points from the pixel gradient directions of the dual-channel fusion image calculated using the Sobel operator, as the suspected edge pixel points of the dual-channel fusion image; Establish an edge gradient map for the suspected edge pixel points; perform a binary operation on the edge gradient map using the segmentation threshold surface to obtain an edge binary map; Extract and connect the edge pixel points of the edge binary map using a connected domain algorithm to obtain a wear particle region; Perform multi-angle polarized light imaging on the wear particle region to obtain a local feature region of the wear particle region; collect actual metal wear particle defect images to perform contour analysis, construct a contour topological condition constraint; and further analyze and reconstruct the contour of the actual metal wear particle defect image to generate a three-dimensional wear particle morphology; perform height field analysis on the three-dimensional wear particle morphology to construct a surface texture condition constraint; construct a cascade classifier based on the contour topological condition constraint and the surface texture condition constraint, and output the identification result of the local feature region.

3. The method of claim 2, wherein the step of detecting the metal particles in the lubricating oil is characterized by, Perform multi-angle polarized light imaging on the wear particle region to obtain a local feature region of the wear particle region, and the specific operation steps are as follows: Collect multi-angle polarized light images of different polarization states using multi-angle polarized light imaging on the wear particle region; The multi-angle polarized light images contain the normal vector of the reflection component; Calculate the normal vector direction of each pixel point in the multi-angle polarized light image; form a normal vector distribution map using the normal vector direction of each pixel point; Normalize the normal vector distribution map, calculate the gradient information of the normal vector distribution map in the x and y directions using the direction of the normal vector, and obtain the components of the normal vector; establish a Poisson equation using the components of the normal vector, solve the Poisson equation, and obtain a height field z(x, y); Perform numerical differential partial derivative calculation on the height field z(x, y); Calculate the Gaussian curvature of each pixel point in the wear particle region using the partial derivative; calculate the average curvature of the pixel points through the Gaussian curvature; Screen the pixel points with local maximum Gaussian curvature and the pixel points with local minimum Gaussian curvature in the wear particle region through the average curvature of the pixel points; Cluster the pixel points with local maximum Gaussian curvature and the pixel points with local minimum Gaussian curvature respectively to obtain a local feature region.

4. The method of claim 3, wherein the step of detecting the metal particle defect in the lubricating oil is characterized by, The specific steps for acquiring actual metal abrasive defect images and performing contour analysis to construct contour topological constraints are as follows: Acquire actual images of metal abrasive defects; The metal abrasive defect image is preprocessed to obtain multiple defect abrasive profiles; the multiple defect abrasive profiles are annotated with morphological defect feature statistics to obtain the abrasive defect type; Preliminary extraction of geometric features for each type of abrasive defect; Contour topological constraints are constructed for each abrasive defect type based on the extracted preliminary geometric features.

5. The method of claim 4, wherein the step of detecting the metal particles in the lubricating oil is characterized by, The contours of the actual metal abrasive defect images are further analyzed and reconstructed to generate a three-dimensional abrasive morphology. Height field analysis is then performed on the three-dimensional abrasive morphology to construct surface texture constraints. The specific steps are as follows: Based on the preliminary geometric features, the minimum bounding ellipse fitting is performed on the multiple defect abrasive grain profiles to calculate the ellipse center coordinates, major axis, and minor axis of the defect abrasive grain profiles. Calculate the angle between the major axis and minor axis of the defective abrasive grain profile, and analyze the elliptical principal direction angle of the defective abrasive grain profile; reconstruct the defective abrasive grain profile using the elliptical center coordinates, the major axis and minor axis, and the elliptical principal direction angle to obtain the three-dimensional abrasive grain morphology. The Gaussian curvature of the three-dimensional abrasive morphology is calculated to obtain the energy distribution; surface texture condition constraints are constructed based on the energy distribution.

6. The method of claim 5, wherein the step of detecting the metal particles in the lubricating oil is characterized by, A cascaded classifier is constructed based on contour topology constraints and surface texture constraints to identify the local feature regions and output the recognition results. The specific operation steps are as follows: A cascaded classifier is constructed using the aforementioned contour topology constraints and surface texture constraints; The local feature regions of the lubricating oil sample are input into a cascade classifier, which outputs the identification result of the lubricating oil sample.

Images