Universal joint seal ring wear state monitoring method and system based on image recognition

Abstract

The present application relates to the technical field of wear state image recognition, in particular to a gimbal seal ring wear state monitoring method and system based on image recognition. The specific implementation process includes: obtaining a seal ring frozen frame image by using a stroboscopic synchronous exposure technology; performing macro and micro dimensional analysis on the seal ring frozen frame image in parallel to generate a geometric attenuation contour image and a topological degradation fingerprint image respectively; constructing an image wear monitoring model to perform wear state feature fusion and generate a seal ring wear morphology image; positioning the wear morphology image in a wear morphology feature space, outputting a wear grade label and monitoring the wear state by combining time series analysis. The present application combines information of macro geometric wear and micro topological damage in two dimensions, and uses image recognition to achieve more comprehensive and accurate evaluation of the wear state, improve the early warning capability of early faults, and ensure the reliability and safety of equipment operation.

Description

Technical Field

[0001] This invention relates to the field of wear condition image recognition technology, and in particular to a method and system for monitoring the wear condition of universal joint seals based on image recognition. Background Technology

[0002] In the field of universal joint seal wear monitoring, existing technologies generally employ tools such as endoscopes, microscopes, and 3D scanners, or rely on manual visual inspection to directly observe and measure surface images. Machine vision technology, used for quality inspection of products on automated production lines, provides a basis for image recognition of static, standardized products under controllable environmental conditions such as lighting and cleanliness.

[0003] However, while existing image recognition monitoring technologies provide intuitive results, they compromise real-time performance. This not only incurs high time and manpower costs but also fails to identify the dynamic evolution of wear and tear. Environmental factors such as oil and dust can randomly obscure image content and introduce noise; uneven and dynamically changing lighting can result in low image contrast and inconsistent grayscale distribution; and mechanical vibrations can cause motion blur. Traditional image processing algorithms based on threshold segmentation, edge detection, or template matching exhibit a sharp decline in stability when processing low-quality, contaminated image data. These factors severely impact the quality of image acquisition and the stability of algorithm analysis, significantly limiting their application scope and hindering the ability to perform non-destructive, in-situ, real-time, and intuitive image recognition of internal equipment images.

[0004] To address this, a method and system for monitoring the wear condition of universal joint seals based on image recognition are proposed. Summary of the Invention

[0005] To address the problems in existing technologies such as random occlusion and noise caused by environmental oil and dust in universal joint sealing ring images, low image contrast and uneven grayscale distribution caused by dynamic lighting changes, and motion blur caused by mechanical vibration.

[0006] The technical solution provided by this invention is as follows: A method for monitoring the wear condition of universal joint seals based on image recognition includes: Receive the image strobe pulses triggered synchronously when the universal joint rotates, perform image strobe exposure on the contact surface of the sealing ring lip, and generate a frozen frame image of the sealing ring without dynamic blur under operating conditions; Edge detection and morphological calculations are performed on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; texture primitive decomposition and spectral analysis are applied to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. An image wear monitoring model is constructed, which receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to fuse wear state features to generate a seal ring wear morphology image; a wear morphology feature space is set, the seal ring wear morphology image is located, and the wear level label is calculated and output. The wear level label categorizes the wear status of the sealing ring into safe images, warning images, and fault images; the wear level label is monitored in real time, and the wear status of the sealing ring is updated and reported.

[0007] Preferably, the specific implementation process of generating the frozen frame image of the sealing ring includes: A pulse signal corresponding to the real-time rotational speed of the universal joint is acquired; the pulse signal is used to drive the exposure start point of the image sensor to provide instantaneous high-brightness illumination to the contact surface of the sealing ring lip, and a momentary image of the sealing ring morphology synchronized with the strobe of the light source is captured; the acquired multi-frame instantaneous images are filtered and fused to output a frozen frame image of the sealing ring lip with no motion blur effect and complete display of static details.

[0008] Preferably, the specific implementation process of generating a geometric attenuation profile image containing macroscopic wear of the sealing ring includes: A Gaussian filter is applied to smooth the frozen frame image of the sealing ring, and the gradient intensity and direction of each pixel in the image are calculated to identify potential edge locations. A non-maximum suppression algorithm is applied to scan pixels along the gradient direction to refine the set of pixels constituting the edge. The suppressed gradient data is input into a dual-threshold detector, and strong edge pixels are divided and connected using two thresholds (high and low) to generate a binary edge map representing the current shape of the sealing ring. Morphological operations are performed on the binary edge map to fill in any small discontinuities and holes. The morphologically optimized contour is spatially registered and contour difference is calculated with a pre-stored standard digital contour template of a wear-free sealing ring to quantify the pixel deviation at the corresponding position, generating a geometric attenuation contour image representing the macroscopic changes in the width, height, and shape of the sealing ring lip.

[0009] Preferably, the specific implementation process for generating the topologically degraded fingerprint image containing microscopic damage to the surface of the sealing ring includes: The frozen frame image of the sealing ring is input into a multi-channel Gabor filter, and the frozen frame image of the sealing ring is subjected to parallel filtering and convolution at multiple scales and directions to separate the texture information in the image with different directional and spatial frequency characteristics. The separated texture feature maps of each channel are subjected to fast Fourier transform to convert the image signal in the spatial domain into spectral data in the frequency domain that includes surface roughness and periodic texture changes. Energy calculation and analysis are performed on the spectral data of each channel to extract spectral feature vectors that characterize micro-damage modes. The extracted multi-channel spectral feature vectors are fused to generate a topological degradation fingerprint image.

[0010] Preferably, the specific implementation process of constructing an image wear monitoring model and generating a topological degradation fingerprint image of the seal ring wear morphology image includes: An image wear monitoring model incorporating a dual-path convolutional neural network is constructed. The geometric attenuation contour image is processed through convolutional and pooling layers to extract and abstract deep geometric feature maps concerning the overall contour deformation, width, and height attenuation of the sealing ring layer by layer. Convolutional operations are performed on the topological degradation fingerprint image containing microscopic damage information to capture semantic features of local texture and topological changes caused by surface scratches, pits, and material peeling. The output feature maps are then stitched together along the channel dimension to generate a sealing ring wear morphology image containing both macroscopic geometric and microscopic topological wear information.

[0011] Preferably, the specific implementation process of calculating wear image level labels and monitoring the wear status of the sealing ring in real time includes: In a sample image library of sealing rings labeled with wear levels, feature vectors are extracted from each image to construct a wear morphology feature space. The class center vector of each wear level is selected as a typical representative of the level. The Euclidean distance between the feature vector of the sealing ring wear morphology image and all pre-calculated class center vectors in the feature space is calculated, and the wear level corresponding to the class center vector with the smallest distance is taken as the classification result. The wear image level label representing the current morphology of the sealing ring is output. The wear level label is subjected to time series analysis. When a transition of the wear image level label is detected and it remains in an unsafe image, the wear process is dynamically identified and tracked, and the wear status of the sealing ring is updated and reported.

[0012] The image recognition-based universal joint seal wear condition monitoring system includes: The sealing ring image generation module receives the image strobe pulse synchronously triggered when the universal joint rotates, performs image strobe exposure on the contact surface of the sealing ring lip, and generates a frozen frame image of the sealing ring without dynamic blur under operating conditions. The dual-scale wear image analysis module performs edge detection and morphological calculations on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; and applies texture primitive decomposition and spectral analysis to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. The wear condition positioning and monitoring module constructs an image wear monitoring model, receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to perform wear condition feature fusion to generate a seal ring wear morphology image; it sets a wear morphology feature space, locates the seal ring wear morphology image, calculates and outputs wear level labels; the wear level labels divide the wear condition of the seal ring into safe images, early warning images, and fault images; it monitors the wear level labels in real time, updates and reports the seal ring wear condition.

[0013] The beneficial effects of the technical solution provided by this invention include: 1. This invention combines a rotary encoder with synchronously triggered high-speed stroboscopic exposure technology, enabling the acquisition of frozen frame images of the sealing ring without dynamic blurring, without interfering with the normal operation of the equipment. By combining multi-frame fusion and smoothing filtering techniques, sensor noise and environmental interference are effectively filtered out, clearly restoring the static details of the sealing ring lip, thus improving the accuracy and reliability of in-situ, real-time monitoring of dynamic components.

[0014] 2. This invention employs a dual-scale wear image analysis strategy, constructing a dual-path convolutional neural network model. It combines macroscopic geometric wear (geometric attenuation contours) with microscopic surface topological damage (topological degradation fingerprints, such as micro-scratches and pits) for feature fusion analysis. This effectively avoids potential omissions and misjudgments that may arise from single-source information analysis, improving the comprehensiveness and accuracy of wear condition assessment and enhancing the early warning capability for early-stage seal failures.

[0015] 3. This invention constructs a wear morphology feature space, objectively quantifying complex image information into clear level labels such as "safety," "early warning," and "fault." Simultaneously, it introduces a time-series analysis mechanism, effectively filtering out misjudgments caused by transient interference. An alarm is only triggered when a continuous and trending deterioration in the wear condition is detected. This improves the stability of condition monitoring and the reliability of early warnings, effectively avoiding unexpected equipment downtime and economic losses due to seal failure. Attached Figure Description

[0016] Figure 1 A flowchart of a universal joint seal wear condition monitoring method based on image recognition provided in an embodiment of the present invention; Figure 2This is a structural diagram of a universal joint seal wear condition monitoring system based on image recognition provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the universal joint seal wear monitoring process provided in an embodiment of the present invention. Detailed Implementation

[0017] Reference Figures 1 to 3 This invention proposes a method for monitoring the wear condition of universal joint seals based on image recognition. The technical solution is as follows:

[0018] Example 1:

[0019] Reference Figure 1 This embodiment proposes a method for monitoring the wear condition of universal joint seals based on image recognition, including: Receive the image strobe pulses triggered synchronously when the universal joint rotates, perform image strobe exposure on the contact surface of the sealing ring lip, and generate a frozen frame image of the sealing ring without dynamic blur under operating conditions; Edge detection and morphological calculations are performed on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; texture primitive decomposition and spectral analysis are applied to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. An image wear monitoring model is constructed, which receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to fuse wear state features to generate a seal ring wear morphology image; a wear morphology feature space is set, the seal ring wear morphology image is located, and the wear level label is calculated and output. The wear level label categorizes the wear status of the sealing ring into safe images, warning images, and fault images; the wear level label is monitored in real time, and the wear status of the sealing ring is updated and reported.

[0020] Furthermore, the specific implementation process for generating the frozen frame image of the sealing ring includes: A pulse signal corresponding to the real-time rotational speed of the universal joint is acquired; the pulse signal is used to drive the exposure start point of the image sensor to provide instantaneous high-brightness illumination to the contact surface of the sealing ring lip, and a momentary image of the sealing ring morphology synchronized with the strobe of the light source is captured; the acquired multi-frame instantaneous images are filtered and fused to output a frozen frame image of the sealing ring lip with no motion blur effect and complete display of static details.

[0021] Specifically, the universal joint drive shaft operates at a stable speed of 1500 revolutions per minute, and a rotary encoder is coaxially connected to the drive shaft. This rotary encoder is specifically an incremental encoder that outputs 2048 pulses per revolution, capable of precisely corresponding to the universal joint's rotational speed in real time and generating a series of pulse signals. The rising edge of the pulse signal drives an image sensor for exposure; the image sensor is specifically an industrial camera with a global shutter function, and the exposure time is set to 50 microseconds. At the instant each pulse signal triggers exposure, a high-brightness LED strobe light source synchronously illuminates the sealing ring lip contact surface for a brief, 40-microsecond burst of intense light.

[0022] Ten consecutive instantaneous images were captured within a tiny angular change of less than 10 degrees from the gimbal rotation. These ten images were processed using an image registration algorithm. Using the first frame as a reference, the subsequent nine frames underwent minor translation and rotation corrections to eliminate subtle positional deviations caused by mechanical vibrations and other factors. A pixel-based grayscale fusion algorithm was then used to fuse the aligned multi-frame images into a single frozen frame image. After this series of processing steps, the final frozen frame image not only eliminated motion blur caused by high-speed rotation but also effectively suppressed random noise through multi-frame fusion. This allowed for the complete and clear display of static details such as early microscopic scratches and material peeling on the sealing ring lip contact surface, resulting in an image signal-to-noise ratio improvement of approximately 15% compared to a single instantaneous image.

[0023] This embodiment combines a rotary encoder with high-speed imaging technology, effectively improving the motion blur problem that exists when monitoring the surface condition of dynamic components. By acquiring frozen frame images, the accuracy and reliability of monitoring the wear condition of sealing rings are improved.

[0024] Furthermore, the specific implementation process for generating the geometric attenuation profile image containing the macroscopic wear of the sealing ring includes: A Gaussian filter is applied to smooth the frozen frame image of the sealing ring, and the gradient intensity and direction of each pixel in the image are calculated to identify potential edge locations. A non-maximum suppression algorithm is applied to scan pixels along the gradient direction to refine the set of pixels constituting the edge. The suppressed gradient data is input into a dual-threshold detector, and strong edge pixels are divided and connected using two thresholds (high and low) to generate a binary edge map representing the current shape of the sealing ring. Morphological operations are performed on the binary edge map to fill in any small discontinuities and holes. The morphologically optimized contour is spatially registered and contour difference is calculated with a pre-stored standard digital contour template of a wear-free sealing ring to quantify the pixel deviation at the corresponding position, generating a geometric attenuation contour image representing the macroscopic changes in the width, height, and shape of the sealing ring lip.

[0025] Specifically, the acquired frozen frame image of the sealing ring is a grayscale image with a resolution of 2048*2048 pixels. To suppress sensor Gaussian noise introduced during image acquisition, a 5*5 pixel kernel Gaussian filter was applied to the frozen frame image for smoothing. The standard deviation sigma value of this filter was set to 1.5. This effectively filters out noise while preserving the detailed information of the sealing ring lip edge to the maximum extent. After processing, the gradient intensity and direction of each pixel in the image were obtained by calculating the grayscale change rate in the horizontal and vertical directions, thereby initially identifying potential edge locations with drastic grayscale value changes.

[0026] To refine blurry edges with a certain width into precise contours of a single pixel width, all potential edge pixels are then scanned along the gradient direction. The non-maximum suppression algorithm suppresses pixels whose gradient intensity is not the maximum value between its two adjacent pixels along the gradient direction, resulting in a set of candidate edge pixels. The suppressed gradient data is input into a dual-threshold detector. The selection of the high and low thresholds is determined by a combination of adaptive calculation and empirical values. Specifically, the high threshold is automatically calculated using Otsu's method based on the grayscale histogram of the current image to ensure optimal threshold separation; the low threshold is set based on the high threshold, typically set to 0.5 * high threshold. Under typical conditions, the high threshold ranges from 100 to 200, and the low threshold ranges from 50 to 100, allowing for fine-tuning by technicians within this range. Pixels with gradient intensities higher than the high threshold are identified as strong edge pixels, while those between the high and low thresholds are identified as weak edge pixels. Starting from a strong edge pixel, all adjacent weak edge pixels are connected until no further expansion is possible, generating a binary edge map representing the current complete shape of the sealing ring. Due to the presence of minute oil stains obscuring the edges in actual working conditions, causing discontinuities, a morphological operation with a 3x3 pixel circular structuring element as the kernel is performed on the binarized edge image to fill in the minute discontinuities and holes. An iterative nearest-point algorithm is used to spatially register the morphologically optimized contour with the standard digital contour template of the sealing ring, eliminating translation and rotation errors during imaging. After registration, the normal distance deviation of each point on the optimized contour relative to the corresponding point on the standard template is calculated; for example, a maximum deviation of 8 pixels is detected at the top of the main lip of the sealing ring. Based on a pre-calibrated scale (e.g., each pixel represents 0.02 mm), this deviation corresponds to 0.16 mm of actual material wear. These quantified pixel deviation values ​​are mapped to grayscale values ​​to generate a geometrical attenuation contour image. The position of each point on the contour in the image corresponds to the actual position of the sealing ring, and its grayscale value intuitively represents the macroscopic wear degree at that point.

[0027] This embodiment utilizes image processing to extract the macroscopic contour of the sealing ring from raw, noisy dynamic images and digitally compares it with an ideal state. By transforming the abstract concept of wear into a quantifiable geometric attenuation contour image, the reliability of sealing ring monitoring is significantly improved.

[0028] Furthermore, the specific implementation process for generating the topologically degraded fingerprint image containing microscopic damage to the surface of the sealing ring includes: The frozen frame image of the sealing ring is input into a multi-channel Gabor filter, and the frozen frame image of the sealing ring is subjected to parallel filtering and convolution at multiple scales and directions to separate the texture information in the image with different directional and spatial frequency characteristics. The separated texture feature maps of each channel are subjected to fast Fourier transform to convert the image signal in the spatial domain into spectral data in the frequency domain that includes surface roughness and periodic texture changes. Energy calculation and analysis are performed on the spectral data of each channel to extract spectral feature vectors that characterize micro-damage modes. The extracted multi-channel spectral feature vectors are fused to generate a topological degradation fingerprint image.

[0029] Specifically, the frozen frame image of the sealing ring is input into a multi-channel Gabor filter bank. This Gabor filter bank consists of 40 independent Gabor filters, with parameters designed to cover five different scales (i.e., spatial frequencies) and eight different directions. The input frozen frame image of the sealing ring undergoes parallel filtering convolution operations with these 40 filters, separating and enhancing the fine scratches along specific directions and periodically distributed pits in the original image into 40 corresponding texture feature maps. A Fast Fourier Transform is then performed on each of these 40 images representing texture features in the spatial domain, converting their signals from the spatial domain to the frequency domain.

[0030] The 40 sets of spectral data obtained after conversion can intuitively reflect the changes in the original surface roughness (represented by the energy of high-frequency components) and the strength of texture periodicity (represented by peak values ​​at specific frequencies) through their energy distribution. Energy calculations are performed on each set of spectral data, and the integral value of the spectral energy within the frequency band is statistically analyzed. Quantized spectral feature values ​​are extracted for each channel, collectively forming a 40-dimensional spectral feature vector. The frequency band is a high-frequency band because microscopic damage such as scratches and pits on the sealing ring surface mainly manifests as high-frequency information in the image signal. The range of this frequency band is set to 50% to 100% of the highest frequency of the Fourier transform spectrum. For example, if the highest frequency determined by image sampling is f_max, then the frequency range of the integral is (0.5*f_max, f_max). Selecting this frequency band can most effectively separate the energy features related to microscopic damage from background noise, thereby accurately quantifying the degree of damage.

[0031] The spectral energy extracted by the filter channel corresponding to the rotational tangent direction of the gimbal (e.g., 90 degrees) increased by 25% compared to the baseline value in the initial undamaged state, indicating significant circumferential wear scratches on the sealing ring surface. Data fusion was performed on the spectral feature vectors to generate a topological degradation fingerprint image. Specifically, the data fusion process involved generating a blank single-channel image of size 256*256 pixels. Using a method of inter-channel maximum value projection, each pixel coordinate (x, y) of the blank image was traversed. At each coordinate point, the pixel response value at (x, y) was obtained from all 40 texture feature maps, and the maximum value was selected from these 40 response values. This maximum value was used as the pixel grayscale value at coordinate (x, y) of the topological degradation fingerprint image. After calculating all pixels, the entire image was linearly normalized to adjust its grayscale value range to the standard 0-255 range.

[0032] In this fingerprint image, the grayscale values ​​of pixels no longer represent physical contours, but directly map the severity of microscopic damage, making early damage such as scratches, pits, and material peeling clearly visible as highlighted areas.

[0033] This embodiment introduces Gabor filtering technology and frequency domain analysis methods to separate the topological degradation fingerprint image, which is difficult to identify directly, from the frozen frame image of the sealing ring. This enables a comprehensive assessment of the wear state from both macroscopic deformation and microscopic topology dimensions, greatly improving the early warning capability of the sealing ring wear state.

[0034] Furthermore, the specific implementation process of constructing an image wear monitoring model and generating a topological degradation fingerprint image of the seal wear morphology image includes: An image wear monitoring model incorporating a dual-path convolutional neural network is constructed. The geometric attenuation contour image is processed through convolutional and pooling layers to extract and abstract deep geometric feature maps concerning the overall contour deformation, width, and height attenuation of the sealing ring layer by layer. Convolutional operations are performed on the topological degradation fingerprint image containing microscopic damage information to capture semantic features of local texture and topological changes caused by surface scratches, pits, and material peeling. The output feature maps are then stitched together along the channel dimension to generate a sealing ring wear morphology image containing both macroscopic geometric and microscopic topological wear information.

[0035] Specifically, the constructed image wear monitoring model comprises two parallel convolutional neural network paths specifically designed to handle different types of wear features. The first path is configured to process geometric attenuation contour images with a resolution of 256*256 pixels. This path consists of four concatenated convolutional modules, each containing a convolutional layer with a 3*3 pixel kernel, a rectified linear unit (ReLU) activation function, and a 2*2 pixel pooling layer. The number of channels in the convolutional layers increases progressively, to 32, 64, 128, and 256 respectively. Through this deep structure, the network processes the input geometric contour image layer by layer, automatically and progressively extracting and abstracting a series of deep geometric features, such as the asymmetric deformation of the overall contour of the sealing ring and the attenuation of the lip width and radial height. The final output is a deep geometric feature map with dimensions of 16*16*256.

[0036] Meanwhile, the second path receives a topologically degraded fingerprint image with the same 256*256 pixel resolution, containing microscopic damage information. This path employs a convolutional network structure that focuses more on capturing local details. Through a series of convolutional operations, it accurately captures and encodes the local topological changes defined by the density and direction of surface micro-scratches, the size distribution of pits, and the edge texture of material peeling areas, transforming them into features with clear semantic information. For example, when processing a sample image containing numerous circumferential scratches, the filter response of this path is significantly enhanced in the corresponding direction. This path ultimately outputs a semantic feature map with the same dimensions as the first path, namely 16*16*256. Subsequently, the feature maps output by these two paths are concatenated along the channel dimension. For example, the first 16*16*256 feature map is concatenated with the second 16*16*256 feature map to generate a single, more information-rich fused feature map with dimensions of 16*16*512. This fused feature map is the sealing ring wear morphology image defined in this invention. It simultaneously encodes macroscopic geometric attenuation information and microscopic topological degradation information at each spatial location. In a validation dataset containing 10,000 wear samples, the accuracy of wear level classification using this fused morphology image is improved by 18% and 23% respectively compared to the scheme using only a single geometric or topological image.

[0037] This embodiment constructs a dual-path neural network structure to achieve parallel processing and efficient fusion of wear feature information from two different modalities and scales. This method overcomes the limitations of single-source analysis, namely, analyzing geometric contours alone may overlook fatal early microscopic damage, while analyzing surface texture alone cannot assess the overall structural wear degree. By generating a comprehensive wear morphology image that simultaneously contains macroscopic geometric and microscopic topological wear information, the comprehensiveness, accuracy, and reliability of wear condition monitoring are significantly improved.

[0038] Furthermore, the specific implementation process of calculating wear image level labels and monitoring the wear status of the sealing ring in real time includes: In a sample image library of sealing rings labeled with wear levels, feature vectors are extracted from each image to construct a wear morphology feature space. The class center vector of each wear level is selected as a typical representative of the level. The Euclidean distance between the feature vector of the sealing ring wear morphology image and all pre-calculated class center vectors in the feature space is calculated, and the wear level corresponding to the class center vector with the smallest distance is taken as the classification result. The wear image level label representing the current morphology of the sealing ring is output. The wear level label is subjected to time series analysis. When a transition of the wear image level label is detected and it remains in an unsafe image, the wear process is dynamically identified and tracked, and the wear status of the sealing ring is updated and reported.

[0039] Specifically, the wear morphology feature space is based on an offline modeling of a sample image library of sealing rings containing precisely labeled wear levels (divided into "safe images," "warning images," and "fault images"). For each image in the library, a dual-path convolutional neural network is used to extract the wear morphology feature vector of the sealing ring sample image. A high-dimensional feature space is constructed based on the wear morphology feature vectors of all samples. In this space, by averaging all feature vectors within the same level, a class center vector that can represent the most typical feature of the three wear levels, "safe," "warning," and "fault," is calculated and stored.

[0040] In the real-time monitoring process, after acquiring a new image of the wear morphology of the sealing ring and extracting its corresponding 512-dimensional feature vector, the Euclidean distance between this vector and all pre-calculated class center vectors in the feature space is immediately calculated. For example, if the Euclidean distances between the feature vector calculated at a certain moment and the three class center vectors of "safety," "warning," and "fault" are 15.8, 8.2, and 21.4 respectively, based on the principle of minimum distance, the "warning level" corresponding to the class center vector with the smallest current distance (8.2) is determined to be the current wear state of the sealing ring, and this wear image level label is output. The wear level label output every 10 minutes is subjected to time-series analysis. When a transition from "safety" to "warning" is detected, and this "warning" state is stably confirmed for three consecutive monitoring cycles (i.e., for a continuous 30 minutes), it is determined that the wear process has undergone substantial deterioration. At this time, the tracking of the wear process is dynamically identified and initiated, the health status of the sealing ring in the database is updated, and a warning report containing timestamps and wear level information is automatically sent to the equipment maintenance management center.

[0041] This embodiment achieves rapid, objective, and automated classification of wear states by constructing a wear morphology feature space and class centers. The introduction of a time-series analysis mechanism effectively filters out misjudgments caused by transient interference, triggering alarms only when the wear state shows a continuous and trending deterioration. This significantly improves the stability of wear state monitoring and the reliability of early warnings.

[0042] This embodiment improves upon the technical challenge of achieving clear imaging of universal joints under high-speed rotation by employing synchronously triggered stroboscopic exposure technology. It enables the acquisition of dynamic-blur-free images of the sealing ring without interfering with normal equipment operation. A dual-scale wear image analysis strategy combines macroscopic geometric wear with microscopic surface topological damage for comprehensive analysis, enhancing the comprehensiveness of wear condition assessment and avoiding potential misjudgments and omissions caused by single-scale analysis. By constructing an image wear monitoring model, complex image information is objectively quantified into wear level labels of "safety," "early warning," and "fault," enabling reliable early warnings at the early stages of fault occurrence and achieving dynamic monitoring and trend judgment of the wear process.

[0043] Example 2: This embodiment provides an image recognition-based universal joint seal wear condition monitoring system for monitoring the wear condition of seals, referring to... Figure 2 The system includes a sealing ring image generation module, a dual-scale wear image analysis module, and a wear condition positioning and monitoring module.

[0044] The sealing ring image generation module receives the image strobe pulse synchronously triggered when the universal joint rotates, performs image strobe exposure on the contact surface of the sealing ring lip, and generates a frozen frame image of the sealing ring without dynamic blur under operating conditions. The dual-scale wear image analysis module performs edge detection and morphological calculations on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; and applies texture primitive decomposition and spectral analysis to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. The wear condition positioning and monitoring module constructs an image wear monitoring model, receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to perform wear condition feature fusion to generate a seal ring wear morphology image; it sets a wear morphology feature space, locates the seal ring wear morphology image, calculates and outputs wear level labels; the wear level labels divide the wear condition of the seal ring into safe images, early warning images, and fault images; it monitors the wear level labels in real time, updates and reports the seal ring wear condition.

[0045] Furthermore, in the sealing ring image generation module, the monitored universal joint drive shaft rotates stably at a high speed of 1500 revolutions per minute. The system acquires a sequence of pulse signals synchronized with the universal joint rotation angle in real time through an incremental rotary encoder that is rigidly connected to the drive shaft and can output 2048 pulses per revolution. The rising edge of each pulse signal is used as a hardware trigger signal to drive the exposure start point of the image sensor. In this embodiment, the image sensor used is an industrial camera with a global shutter function, and its exposure time is precisely set to 50 microseconds. At the moment of triggering exposure, a high-brightness LED strobe light source synchronously illuminates the contact surface of the sealing ring lip for a duration of 40 microseconds. Its extremely short duration and extremely high brightness ensure that even under high-speed rotation, the single-frame instantaneous image captured by the camera has no ghosting. To further improve the signal-to-noise ratio of the image and ensure the integrity of details, the module continuously acquires 10 such instantaneous images within a very small time window (corresponding to an angle range of less than 10 degrees of universal joint rotation). These 10 images were screened and registered to eliminate inter-frame displacement deviations caused by minor equipment vibrations, and then fused into a single, high-resolution frozen frame image. The final output frozen frame image not only eliminates motion blur, but also improves the signal-to-noise ratio by approximately 15% compared to any single instantaneous image, and can completely and clearly display the morphological features of the sealing ring lip in static details.

[0046] Furthermore, the dual-scale wear image analysis module receives a 2048*2048 pixel frozen frame image of the sealing ring provided by the sealing ring image generation module and performs processing on two parallel paths. In the first path, to analyze macroscopic wear, the module applies a Gaussian filter with a 5*5 pixel kernel to smooth and denoise the frozen frame image. Then, by calculating the gradient, applying non-maximum suppression, and using double thresholding, it extracts a single-pixel width binarized edge map representing the current shape of the sealing ring. It fills the small discontinuities in the edge with morphological operations using a 3*3 pixel circular structural element as the kernel. It spatially registers the optimized contour with a wear-free standard digital contour template and calculates the normal pixel deviation between the two. This deviation is quantized (e.g., an 8-pixel deviation corresponds to 0.16 mm of actual wear) and mapped to grayscale values ​​to generate a geometric attenuation contour image that intuitively represents the macroscopic changes in the width, height, and shape of the lip. In the second path, to analyze microscopic surface damage, the module inputs the frozen frame image of the sealing ring into a multi-channel Gabor filter bank consisting of 40 filters covering 5 scales and 8 directions for parallel filtering and convolution. This separates the microscopic texture information (e.g., scratches, pits) with different directional and spatial frequency characteristics into 40 independent texture feature maps. Each of these 40 feature maps undergoes a Fast Fourier Transform to be converted to the frequency domain. By calculating the energy distribution of each spectrum, a 40-dimensional spectral feature vector that can quantitatively characterize the microscopic damage pattern is extracted. For example, a channel with a 25% increase in energy along the rotational tangent direction compared to the baseline accurately characterizes the appearance of circumferential wear scratches. After fusing this multi-channel feature vector, a topological degradation fingerprint image is generated. In the topological degradation fingerprint image, the pixel grayscale value directly corresponds to the severity of the microscopic damage.

[0047] Furthermore, the wear state positioning and monitoring module constructs an image wear monitoring model. This model receives geometric attenuation contour images and topological degradation fingerprint images output in parallel by the dual-scale wear image analysis module. The image wear monitoring model uses a dual-path convolutional neural network structure to perform parallel deep feature extraction on two input images, each with a resolution of 256*256 pixels. The 16*16*256 feature maps output from the two paths, representing macroscopic geometric features and microscopic semantic features respectively, are stitched together along the channel dimension to generate a single 16*16*512 dimension image of the sealing ring wear morphology. To achieve the positioning and classification of wear states, the module first constructs a wear morphology feature space offline based on a sample image library containing 5000 images labeled as "safe," "warning," and "fault" levels. In this space, the mean of all sample sealing ring wear morphology images within each level is calculated, and class center vectors representing the typical features of these three levels are stored. After the module generates a new wear morphology image, it stretches it into a high-dimensional feature vector and calculates the Euclidean distance between this vector and all class center vectors in the feature space. For example, at a certain monitoring moment, the calculated distances between the current feature vector and the three class center vectors of "Safety," "Warning," and "Fault" are 15.8, 8.2, and 21.4, respectively. Based on the principle of minimum distance, the module determines the current wear state of the sealing ring as "Warning" and outputs a wear image level label characterizing the current wear degree of the sealing ring, i.e., the "Warning Image." The module performs real-time monitoring and time-series analysis on this series of level labels output in chronological order. When a transition from "Safety" to "Warning" is detected, and this "Warning" state is stably confirmed for three consecutive monitoring cycles (i.e., for 30 minutes), the system automatically updates the wear state of the sealing ring and sends a warning report containing a timestamp and wear level information to the equipment maintenance management center.

[0048] Example 3: This embodiment deploys the aforementioned image recognition-based universal joint seal wear condition monitoring method and system entirely on a drive shaft durability testing production line in Factory X, referring to... Figure 3 This enables real-time monitoring of the wear condition of the universal joint seal ring under high-speed operation.

[0049] On the drive shaft test bench, the monitored universal joint operates at a constant speed of 1500 revolutions per minute. An industrial camera with a global shutter function is mounted 200 mm from the contact surface of the sealing ring lip, with its lens optical axis strictly perpendicular to the universal joint's rotation axis. The camera uses a 35 mm fixed-focus industrial lens to ensure sufficient depth of field while balancing light intake and image sharpness. The light source is a 20-watt high-brightness white ring LED strobe light source coaxially mounted with the camera lens. This arrangement ensures uniform and shadow-free illumination of the sealing ring surface and provides a center illuminance of over 15,000 lux at a working distance of 200 mm, guaranteeing a high signal-to-noise ratio and detailed image within an extremely short exposure time of 50 microseconds.

[0050] An incremental rotary encoder capable of outputting 2048 pulses per revolution is rigidly connected coaxially to the drive shaft. In terms of control logic, the encoder's pulse signal output is connected to the camera's external hardware trigger input port, while the camera's flash synchronization signal output port is connected to the controller of the LED strobe light source. This connection method utilizes the rising edge of the encoder pulse to precisely trigger the camera's exposure, and the camera outputs a signal to illuminate the LED light source at the instant of exposure, thus achieving microsecond-level synchronization between camera exposure and light source flashing.

[0051] When the gimbal rotates stably, the camera receives a sequence of pulse signals from the rotary encoder. For each trigger pulse, the camera performs an exposure lasting 50 microseconds, while the LED light source provides illumination for 40 microseconds. Within a very small time window corresponding to a gimbal rotation angle of less than 10 degrees, the system continuously acquires 10 instantaneous images with a resolution of 2048*2048 pixels. To eliminate inter-frame displacement caused by minute mechanical vibrations, the system employs a phase-correlation-based image registration algorithm, using the first frame as a reference, and performing sub-pixel-level translation and rotation corrections on the remaining 9 frames. After calibration, the grayscale values ​​of corresponding pixels in the 10 frames are averaged to create a high-definition frozen frame image of the sealing ring that effectively suppresses random noise and is free of motion blur.

[0052] The system initiates parallel dual-scale wear image analysis. In the first scale, a geometrical attenuation contour image containing macroscopic wear is generated. This process requires a baseline. This baseline, the "wear-free sealing ring standard digital contour template," is pre-prepared as follows: a brand-new, quality-certified sealing ring of the same model is acquired and fused into a high signal-to-noise ratio standard image under identical hardware and geometric configuration as the monitoring system. Its contour is then extracted using an edge detection algorithm and stored as the standard template. The system smooths the acquired frozen frame image of the sealing ring using a 5x5 pixel kernel Gaussian filter (standard deviation sigma of 1.5). By calculating the image gradient, applying a non-maximum suppression algorithm, and using a double-threshold connection method, a single-pixel-width binarized edge map representing the current shape of the sealing ring is extracted. To ensure the integrity of the contour, the system performs morphological operations with a 3x3 pixel circular structuring element as the kernel to fill in any small discontinuities in the edges. The optimized contour is spatially registered with the pre-stored standard template, and the normal distance deviation between the two is calculated point-by-point. These deviation values ​​(e.g., a maximum deviation of 8 pixels detected at the top of the main lip, which corresponds to 0.16 mm of actual material wear according to a pre-calibrated scale) are mapped to grayscale values ​​from 0 to 255, thereby generating a 256*256 pixel geometric attenuation profile image, in which the position of each point in the image corresponds to the actual sealing ring position, and its brightness intuitively reflects the macroscopic wear degree of that point.

[0053] In the parallel second scale, the system generates a topological degradation fingerprint image containing surface micro-damage. The frozen frame image of the sealing ring is input to a multi-channel filter bank consisting of 40 independent Gabor filters, designed to cover five different spatial frequencies (scales) and eight different orientations (from 0 degrees to 157.5 degrees, spaced 22.5 degrees apart). After parallel filtering and convolution operations, 40 texture feature maps highlighting specific directional and scale-specific texture features (e.g., micro-scratches or periodic pits) are obtained. The system performs a Fast Fourier Transform on each of these 40 feature maps, transforming them from the spatial domain to the frequency domain. By calculating the energy integral of each spectrogram in the high-frequency band, a 40-dimensional spectral feature vector is extracted, which can quantitatively describe the micro-roughness and damage patterns of the sealing ring surface. The channel energy along the rotational tangent direction (90 degrees) increases by 25% compared to the initial baseline, indicating the presence of circumferential wear scratches on the sealing ring. The system fuses this multi-channel spectral feature vector to generate a 256*256 pixel topological degradation fingerprint image. In this image, the grayscale values ​​of pixels no longer represent physical contours, but directly map the severity of microscopic damage, making early damage clearly visible as highlighted areas.

[0054] The system utilizes an image wear monitoring model incorporating a dual-path convolutional neural network for wear state localization and monitoring. The specific architecture of the image wear monitoring model is as follows: Path 1 receives the geometric attenuation contour image and consists of four concatenated convolutional modules. Each module contains a 3x3 convolutional layer, a ReLU activation function, and a 2x2 max-pooling layer, with channel numbers of 32, 64, 128, and 256 respectively. Path 2 receives the topological degradation fingerprint image and employs a network containing six convolutional layers. The first four convolutional layers use 3x3 kernels with a stride of 1, and channel numbers of 32, 64, 128, and 128 respectively. A 2x2 max-pooling layer is connected after every two convolutional layers. To better capture details, the last two convolutional layers use 1x1 kernels, increasing the channel number to 256, and no longer use pooling layers to maintain high feature resolution. A ReLU activation function is used after all convolutional layers. The outputs of the two paths (feature maps of 16*16*256 dimensions each) are concatenated along the channel dimension to form a 16*16*512 fused feature map. This map is then flattened and connected to a 512-node fully connected layer (including Dropout to prevent overfitting) and a Softmax classification layer with a final output of 3 nodes, corresponding to the three wear levels: "Safe," "Warning," and "Fault." This model is trained offline on a database of 5000 precisely labeled sealing ring sample images. The Adam optimizer is used during training with an initial learning rate of 0.001 and a loss function of classification cross-entropy, for a total of 100 iterations. During the offline phase, the system also calculates and stores the average feature vector of all samples at each level in the deep feature space of the network, serving as the "class center vector" for that level.

[0055] In the real-time monitoring process, after the newly generated geometric attenuation contour image and topological degradation fingerprint image are input into the model and stitched into a wear morphology image, the system calculates its feature vector in the deep feature space and calculates the Euclidean distance between this vector and three pre-stored "class center vectors". For example, at a certain monitoring moment, the calculated distances are: 15.8 with the "safe" class center, 8.2 with the "warning" class center, and 21.4 with the "fault" class center. Based on the minimum distance principle, the system determines the current wear state of the seal as "warning" and outputs this level label. The system performs time-series analysis on the continuously output level labels. When a transition from "safe" to "warning" is detected, and this "warning" state is stably confirmed within the subsequent three consecutive monitoring cycles (i.e., for 30 minutes), the system determines that the wear process has substantially deteriorated. At this time, the system will automatically update the health status of the seal in the database and send an alarm containing a timestamp and "warning" level information to the equipment maintenance management center.

[0056] This embodiment effectively improves the dynamic blurring problem of high-speed rotating components by using stroboscopic synchronous exposure technology, ensuring the clarity of image acquisition. Employing a dual-path image analysis strategy, combining information from both macroscopic wear and microscopic surface damage dimensions, and fusing it through a deep learning model, significantly enhances the comprehensiveness and accuracy of wear condition assessment, avoiding potential misjudgments from single-scale analysis. Establishing a wear morphology feature space and combining it with temporal analysis effectively prevents unexpected equipment downtime and economic losses due to seal failure.

[0057] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for monitoring the wear condition of universal joint seals based on image recognition, characterized in that, include: Receive the image strobe pulses triggered synchronously when the universal joint rotates, perform image strobe exposure on the contact surface of the sealing ring lip, and generate a frozen frame image of the sealing ring without dynamic blur under operating conditions; Edge detection and morphological calculations are performed on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; texture primitive decomposition and spectral analysis are applied to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. An image wear monitoring model is constructed, which receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to fuse wear state features to generate a seal ring wear morphology image; a wear morphology feature space is set, the seal ring wear morphology image is located, and the wear level label is calculated and output. The wear level label categorizes the wear status of the sealing ring into safe images, warning images, and fault images; the wear level label is monitored in real time, and the wear status of the sealing ring is updated and reported.

2. The method for monitoring the wear condition of universal joint seals based on image recognition according to claim 1, characterized in that: The specific implementation process of generating the frozen frame image of the sealing ring includes: acquiring a pulse signal corresponding to the real-time rotation speed of the universal joint; using the pulse signal to drive the exposure start point of the image sensor to instantaneously illuminate the contact surface of the sealing ring lip with high brightness, capturing an instantaneous image of the sealing ring morphology synchronized with the strobe of the light source; filtering and fusing the acquired multiple frames of instantaneous images to output a frozen frame image without motion blur effect, fully displaying the static details of the sealing ring lip.

3. The method for monitoring the wear condition of universal joint seals based on image recognition according to claim 1, characterized in that: The specific implementation process for generating a geometric attenuation contour image containing macroscopic wear of the sealing ring includes: applying a Gaussian filter to smooth the frozen frame image of the sealing ring, and calculating the gradient intensity and direction of each pixel in the image to identify potential edge positions; applying a non-maximum suppression algorithm to scan pixels along the gradient direction to refine the set of pixels constituting the edge; inputting the suppressed gradient data into a dual threshold detector, using high and low thresholds to divide and connect strong edge pixels and weak edge pixels to generate a binary edge map representing the current shape of the sealing ring; performing morphological operations on the binary edge map to fill in the small discontinuities and holes present in the edge map; spatially registering and calculating the contour difference between the morphologically optimized contour and a pre-stored standard digital contour template of the wear-free sealing ring, quantifying the pixel deviation at the corresponding position, and generating a geometric attenuation contour image representing the macroscopic changes in the width, height, and shape of the sealing ring lip.

4. The method for monitoring the wear condition of universal joint seals based on image recognition according to claim 1, characterized in that: The specific implementation process for generating a topological degradation fingerprint image containing microscopic damage to the surface of the sealing ring includes: inputting a frozen frame image of the sealing ring into a multi-channel Gabor filter; performing parallel filtering and convolution on the frozen frame image of the sealing ring at multiple scales and directions to separate texture information with different directional and spatial frequency characteristics in the image; performing a fast Fourier transform on the texture feature maps of each channel obtained after separation to convert the image signal in the spatial domain into spectral data in the frequency domain containing the periodic changes in surface roughness and texture; performing energy calculation and analysis on the spectral data of each channel to extract spectral feature vectors characterizing the microscopic damage mode; and fusing the extracted multi-channel spectral feature vectors to generate a topological degradation fingerprint image.

5. The method for monitoring the wear condition of universal joint seals based on image recognition according to claim 1, characterized in that: The specific implementation process of constructing an image wear monitoring model and generating a seal ring wear morphology image includes: constructing an image wear monitoring model containing a dual-path convolutional neural network; processing the geometric attenuation contour image through convolutional and pooling layers to extract and abstract deep geometric feature maps about the overall contour deformation, width, and height attenuation of the seal ring layer by layer; performing convolution operations on the topological degradation fingerprint image containing micro-damage information to capture the semantic features of local texture and topological structure changes caused by surface scratches, pits, and material peeling; and stitching the output feature maps along the channel dimension to generate a seal ring wear morphology image containing macro-geometric and micro-topological wear information.

6. The method for monitoring the wear condition of universal joint seals based on image recognition according to claim 1, characterized in that: The specific implementation process of calculating wear image level labels and monitoring the wear status of sealing rings in real time includes: extracting feature vectors from a sample image library of sealing rings labeled with wear levels and constructing a wear morphology feature space; selecting the class center vector of each wear level as a typical representative of the level; calculating the Euclidean distance between the feature vector of the sealing ring wear morphology image and all pre-calculated class center vectors in the feature space, and taking the wear level corresponding to the class center vector with the smallest distance as the classification result, and outputting a wear image level label representing the current morphology of the sealing ring; performing time-series analysis on the wear level label, and when a transition of the wear image level label is detected and it remains in an unsafe image, dynamically identifying and tracking the wear process, updating and reporting the wear status of the sealing ring.

7. A universal joint seal wear condition monitoring system based on image recognition, characterized in that, include: The sealing ring image generation module receives the image strobe pulse synchronously triggered when the universal joint rotates, performs image strobe exposure on the contact surface of the sealing ring lip, and generates a frozen frame image of the sealing ring without dynamic blur under operating conditions. The dual-scale wear image analysis module performs edge detection and morphological calculations on the frozen frame image of the sealing ring to generate a geometric attenuation contour image containing macroscopic wear of the sealing ring; and applies texture primitive decomposition and spectral analysis to the frozen frame image of the sealing ring to generate a topological degradation fingerprint image containing microscopic damage on the surface of the sealing ring. The wear condition positioning and monitoring module constructs an image wear monitoring model, receives the geometric attenuation contour image and the topological degradation fingerprint image, and uses dual-path image state analysis to perform wear condition feature fusion to generate a seal ring wear morphology image; it sets a wear morphology feature space, locates the seal ring wear morphology image, calculates and outputs wear level labels; the wear level labels divide the wear condition of the seal ring into safe images, early warning images, and fault images; it monitors the wear level labels in real time, updates and reports the seal ring wear condition.

Images