A machine vision-based inner floating disc seal monitoring method and system

Abstract

This invention discloses a machine vision-based method and system for monitoring the seal of an internal floating roof, belonging to the field of machine vision technology. It includes: acquiring laser stripe images collected by a monitoring unit; performing binarization and connected component analysis on the laser stripe images to delineate the region of interest, and extracting the center line of the laser stripes with sub-pixel precision using a grayscale centroid algorithm; establishing a triangulation model based on a pre-calibrated camera intrinsic parameter matrix and laser plane equation, converting the center line into a three-dimensional contour point cloud of the seal, and performing time-series filtering; retrieving the reference contour point cloud corresponding to the circumferential encoder angle, calculating the normal deviation value and tangential gap value; determining the fault type based on multi-level comparison results with preset thresholds, and generating a fault spatial index by fusing circumferential azimuth angle and liquid level data. This invention obtains high-contrast stripes under extremely dark lighting and metal reflection environments in storage tanks, enabling quantitative measurement of the three-dimensional deformation of the seal and preventative maintenance.

Description

Technical Field

[0001] This application belongs to the field of machine vision technology, specifically relating to a machine vision-based method and system for monitoring the sealing of an internal floating disk. Background Technology

[0002] With the increasing demands for safety and environmental protection in the petrochemical industry, monitoring the integrity of the internal floating roof edge sealing structure has become a crucial aspect of ensuring the sealing performance of the floating roof. Utilizing machine vision technology to replace traditional manual inspections and achieve non-contact automatic assessment of the sealing status has become an important development direction in this field.

[0003] However, existing machine vision-based sealing monitoring solutions have the following shortcomings in practical applications: On the one hand, the interior of the storage tank is in a dark environment for a long time, and the tank wall and floating roof are made of metal, which easily generates complex specular reflection interference, making it difficult for conventional imaging systems to obtain clear, high-contrast target images; on the other hand, the seals are mostly made of black rubber with scarce surface texture, making it difficult for traditional two-dimensional image analysis methods to accurately extract defect features, and even more difficult to achieve three-dimensional quantitative measurement of the normal deformation and tangential gap of the seals, thus limiting the monitoring accuracy and the feasibility of preventive maintenance.

[0004] Therefore, there is an urgent need for a monitoring solution that can adapt to the harsh optical environment of storage tanks and can quantitatively assess the three-dimensional deformation of seals. Summary of the Invention

[0005] The purpose of this invention is to provide a machine vision-based method and system for monitoring the sealing of an internal floating disk, which can effectively solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0007] A machine vision-based method for monitoring the seal of an internal floating disk includes the following specific steps:

[0008] Acquire laser stripe images captured by an optical detection module integrated into the edge of an inner floating disk that rises and falls with the liquid level. The optical detection module includes a line laser, an industrial camera with a narrow-band filter, and a circumferential encoder for recording the rotational position.

[0009] The laser stripe image is binarized and connected component analysis is performed to delineate a rectangular region of interest containing the complete laser stripe. The gray-scale centroid algorithm is then used to extract the center line of the laser stripe with sub-pixel precision from the region of interest.

[0010] Based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, a triangulation model is established to convert the center line of the laser stripe with sub-pixel precision into a three-dimensional contour point cloud of the seal in the camera coordinate system, and time series filtering is performed on the three-dimensional contour point cloud of multiple consecutive frames.

[0011] Retrieve the reference contour point cloud corresponding to the current circumferential encoder angle value from the memory. Use the nearest point iteration algorithm to finely register the real-time three-dimensional contour point cloud after time series filtering with the reference contour point cloud to eliminate global coordinate offset. Calculate the normal deviation value of the real-time contour relative to the reference contour and the tangential gap value between the sealing edge and the tank wall after registration.

[0012] Based on the multi-level comparison results of the normal deviation value and the preset deviation threshold, and the tangential clearance value and the preset clearance threshold, the fault type of the seal is determined, and the circumferential azimuth angle data provided by the circumferential encoder and the current liquid level height data provided by the liquid level gauge are integrated to generate a fault space index.

[0013] Furthermore, in the step of acquiring the laser stripe image collected by the optical detection module, the line laser is selected as a visible red laser with a wavelength of 650nm, and the narrowband filter is specifically configured with a center wavelength of 650nm and a bandwidth of 10nm, so as to obtain high-contrast laser stripes in extremely dark lighting and metal reflection environments.

[0014] The circumferential encoder is used to output angular position signals in real time when the floating disk undergoes circumferential motion or rotation, providing a basis for azimuth positioning for subsequent fault space indexing.

[0015] Furthermore, the extraction of the laser stripe center line with sub-pixel precision from the region of interest using the gray-scale centroid algorithm specifically involves: traversing line by line along the normal direction of the laser stripes within the region of interest, extracting continuous pixel segments in each line with gray-scale values ​​higher than a preset threshold as the effective cross-section of the stripes, and then applying the formula... Calculate the sub-pixel center coordinates on this cross section, where The x-coordinate of the sub-pixel center. The grayscale value of the pixel after filtering. The original discrete x-coordinate of the pixel. The total number of pixels involved in the calculation is represented by the sub-pixel center coordinates of each row, which are arranged in row index order to form the complete laser stripe center line.

[0016] Furthermore, the establishment of a triangulation model based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, converting the sub-pixel precision laser stripe centerline into a three-dimensional contour point cloud of the sealing component in the camera coordinate system, specifically includes: pre-obtaining the intrinsic parameter matrix of the industrial camera according to the Zhang Zhengyou calibration method, including the equivalent focal length. and principal point coordinates ;

[0017] According to the rotation matrix Translation vector By calibrating the spatial position of the laser plane in the camera coordinate system, the standard form of the laser plane equation is obtained. For any point on the center line of the laser stripe with sub-pixel precision Construct the parameter equations of the measurement ray starting from the optical center of the camera. Solve for the depth parameter by simultaneously solving the laser plane equation and the measured ray parameter equation. Then, substituting back into the equation for the measurement ray parameters, the corresponding three-dimensional coordinates are obtained. All three-dimensional coordinate points constitute the three-dimensional contour point cloud of the seal.

[0018] Furthermore, the step of using the nearest neighbor iteration algorithm to perform fine registration between the real-time 3D contour point cloud after time series filtering and the reference contour point cloud is as follows: taking the reference contour point cloud as the target point set and the real-time contour point cloud as the source point set, iteratively performing nearest neighbor search to construct a set of matching point pairs, estimating the rigid body transformation matrix of the current iteration step by the least squares method and applying it to the source point set to update its position, until the average distance change of the matching point pairs between two iterations is less than a preset convergence threshold, thus obtaining the optimal rigid body transformation matrix used to correct the global pose of the real-time contour.

[0019] The calculation of the normal deviation value is specifically as follows: for any sampling point on the real-time contour after registration, the three nearest neighbor points on the reference contour surface are determined to fit a local plane, and the directed distance from the sampling point to the plane along the normal vector direction of the local plane is calculated and defined as the normal deviation value.

[0020] The calculation of the tangential gap value is specifically as follows: In the registered three-dimensional contour point cloud, the curvature change position and grayscale jump position are detected along the laser projection direction. The position that simultaneously satisfies the curvature change and grayscale jump is identified as the edge break point of the seal. The Euclidean distance between the break point and the point on the side surface of the tank wall on the horizontal projection plane is calculated and defined as the tangential gap value.

[0021] Furthermore, the specific criteria for determining the fault type of the seal include: when the normal deviation value exceeds the preset first deviation threshold of 3mm in a local area with a continuous circumferential span of not less than 50mm, and the sign of the normal deviation value in this area changes alternately with a wave-like undulation feature, it is determined to be a local wrinkle.

[0022] When the absolute value of the difference between the normal deviation values ​​of adjacent sampling points exceeds the preset jump threshold of 2mm, and physical breakage features with a spacing of more than 1mm appear at the corresponding position in the three-dimensional contour point cloud, it is judged as tearing damage.

[0023] When the tangential gap value exceeds the preset gap threshold of 5mm, it is determined to be a partial detachment.

[0024] Furthermore, after generating the fault space index, the step of preventive maintenance analysis is also included: the arithmetic mean of the absolute values ​​of the normal deviations of all sampling points collected daily is stored in the health status database as the average profile deviation.

[0025] The least squares method was used to perform linear regression on the sequence of average profile deviations stored over the past 30 to 90 days, and a linear equation was obtained with the cumulative number of days as the independent variable and the average deviation as the dependent variable.

[0026] Based on the linear equation, the deviation value is predicted for the next 15 days. When the predicted deviation value reaches or exceeds the preset safety red line value of 3 mm, a preventive replacement recommendation is issued.

[0027] Furthermore, the time-series filtering of the three-dimensional contour point cloud for multiple consecutive frames is specifically as follows: a circular buffer with a preset capacity of 15 consecutive frames is opened in memory. When a new three-dimensional contour point cloud arrives, if the buffer is not full, it is written directly; if it is full, the frame data with the earliest timestamp is overwritten.

[0028] For all valid frames in the buffer at the current time, calculate the arithmetic mean of their three-dimensional coordinates point by point to filter out random errors caused by liquid surface fluctuations and mechanical vibrations.

[0029] Furthermore, the optical detection module is fixed to the annular beam at the edge of the inner floating plate by a flexible stainless steel bracket. The flexible stainless steel bracket integrates a damping and shock absorption module composed of a metal spring and a high-damping rubber pad, which is used to absorb the mechanical vibration caused by sliding friction and liquid surface fluctuations during the lifting and lowering of the floating plate, and maintain the stability of the optical axis of the line laser and the industrial camera.

[0030] A machine vision-based internal floating disk seal monitoring system includes:

[0031] The optical detection module, integrated into the edge of the inner floating plate that rises and falls with the liquid level, includes a line laser, an industrial camera with a narrow-band filter, a circumferential encoder, and explosion-proof protection components. It is used to acquire high-contrast laser stripe images in extremely dark lighting and metal reflection environments inside the storage tank and output circumferential azimuth signals in real time.

[0032] The centerline extraction module is used to perform binarization processing and connected component analysis on the laser stripe image to delineate the region of interest, and to extract the centerline of the laser stripe with sub-pixel precision using a gray-scale centroid algorithm.

[0033] The 3D reconstruction module is used to establish a triangulation model based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, convert the center line of the laser stripe with sub-pixel precision into a 3D contour point cloud of the seal, and perform time-series filtering on the 3D contour point cloud of multiple consecutive frames through a ring buffer.

[0034] The registration and calculation module is used to retrieve the corresponding reference contour point cloud based on the circumferential encoder signal, and to perform fine registration between the real-time 3D contour point cloud and the reference contour point cloud using the nearest point iteration algorithm, and to calculate the normal deviation value and tangential gap value after registration.

[0035] The fault diagnosis and location module is used to compare the normal deviation value and the tangential gap value with preset multi-level fault thresholds to determine the fault type of wrinkling, tearing or detachment, and to fuse the determination result with the circumferential encoder data and the liquid level height data received from the external liquid level gauge to generate a fault space index that includes fault type, severity, circumferential azimuth angle and current liquid level height.

[0036] In summary, this application includes at least one of the following beneficial technical effects:

[0037] 1. This invention constructs a closed optical detection channel by configuring a narrowband filter of a specific wavelength at the industrial camera end and combining it with high-intensity line laser projection. This design completely solves the imaging problem under extremely low lighting conditions inside large storage tanks, while effectively suppressing complex specular reflection interference caused by the metal walls and internal floating roof structure of the tank. Compared with traditional visual solutions that rely on natural light or ordinary supplemental lighting, this invention can still obtain high-contrast feature fringes in harsh optical environments, ensuring the stable operation of the monitoring system in all weather conditions.

[0038] 2. This invention elevates the pixel analysis of traditional two-dimensional images to the measurement of physical quantities in three-dimensional space. By establishing a rigorous triangulation model and a stripe extraction algorithm with predetermined precision, the system can directly output the specific numerical value of the sealing gap, rather than a vague initial visual judgment. This quantitative capability allows security personnel to accurately grasp the deformation state of the seal, exhibiting extremely high sensitivity in capturing defects that are difficult to detect in two-dimensional images, such as micro-tears and normal fatigue indentations, significantly reducing the missed detection rate of manual inspections.

[0039] 3. This invention not only identifies existing, visible faults but also captures the physical fatigue trends of the sealing rubber material through long-term trend analysis of three-dimensional profile data. By comparing minute offsets in the baseline profile, the system can identify early signs of deteriorating sealing performance before the seals completely fail or substantial oil and gas leaks occur. This data-driven early warning mechanism transforms the maintenance mode of storage tanks from post-fault repair to preventative maintenance, significantly improving the safety level of petrochemical storage operations.

[0040] 4. The monitoring unit of this invention is installed on the edge of the inner floating roof and adopts a non-contact measurement method, which will not cause any physical obstruction to the normal movement of the sealing components or the sealing performance. The system has a compact structure and can be adapted to the retrofitting of storage tanks of different specifications and sealing structures. Through linkage with the circumferential encoder and level gauge, it achieves full spatial coverage monitoring of the sealing status of the storage tank. This automated monitoring method replaces the high-risk manual inspection of the tank top, reducing the operational risks of personnel entering confined spaces. Attached Figure Description

[0041] Figure 1 This is a schematic diagram of the overall technical solution for the machine vision-based internal floating disk sealing monitoring method.

[0042] Figure 2 This is a schematic diagram illustrating the core principle of locating and three-dimensionally representing the surface features of a sealing component based on a triangulation model.

[0043] Figure 3 A flowchart illustrating the logic of quantifying seal damage through precise registration of real-time 3D contours with reference contours;

[0044] Figure 4 A schematic diagram of a multi-level interactive data flow for early warning processing of multi-dimensional fault criteria and health status database. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 The present invention will be further described in detail below with reference to specific embodiments.

[0046] Firstly, the machine vision-based internal floating roof sealing monitoring method disclosed in this application, in practical engineering applications, achieves real-time, quantitative monitoring of the sealing status of the internal floating roof edge in storage tanks by deploying a high-precision optical sensing system and combining it with an edge computing and cloud analytics architecture. The specific implementation steps of the method are detailed below:

[0047] The first step, S1, involves constructing a non-contact optical detection front-end capable of adapting to the harsh environment inside the storage tank. This is achieved by installing a sealed, explosion-proof unit integrating a specific wavelength laser and a narrowband filtered industrial camera on the edge of the inner floating roof, which rises and falls with the liquid level. The spatial geometric relationship between the two is precisely fixed, enabling the stable acquisition of high-contrast laser stripe images covering the entire sealed area even under extremely low light conditions and strong metal reflection interference. Specifically, step S1 involves the deployment and debugging of the detection front-end through the following sub-steps.

[0048] Step S101: The monitoring unit is installed and integrated with the device. The entire monitoring unit is fixed to the annular beam at the edge of the inner floating disc, allowing it to rise and fall synchronously with the floating disc and the liquid level. This ensures that each data acquisition occurs within the same predetermined monitoring range between the seal and the tank wall. The monitoring unit integrates a line laser, an industrial camera with a narrow-band filter, a circumferential encoder for recording rotational position, and explosion-proof protection components. These three components are encapsulated as a non-contact monitoring unit, facing the contact area between the seal and the tank wall. The circumferential encoder is mechanically coupled to the monitoring unit. When the floating disc moves circumferentially along the tank wall or undergoes a slight rotation, the circumferential encoder outputs an angular position signal in real time for subsequent fault space indexing and orientation location.

[0049] In step S102, the parameters of the laser and industrial camera are determined. A visible red laser with a wavelength of 650nm is selected as the line laser, and the output power is set between 50mW and 100mW. The line width is precisely controlled between 0.5mm and 0.8mm at a working distance of 300mm, so as to provide a sufficiently bright laser stripe while meeting the safety restrictions of being filled with oil and gas, and at the same time keep the energy highly concentrated.

[0050] The industrial camera uses a high dynamic range CMOS sensor chip with a resolution of no less than 1280×1024 pixels and a frame rate set to 30fps to 60fps. This acquisition rate can sensitively capture the dynamic deformation of the seals during the lifting and lowering of the floating table, avoiding motion blur that causes measurement distortion.

[0051] Step S103: Establishing the geometric relationship and mechanical stabilization of optical triangulation. After selecting the components, adjust the output plane of the line laser so that it obliquely strikes the contact surface between the seal and the tank wall at an incident angle of 35 to 55 degrees. This angle is selected based on the triangulation sensitivity optimization results. When the seal undergoes a radial deformation of 1 mm, the pixel shift of the laser stripe on the image sensor can reach 5 to 10 pixels, thereby amplifying the sub-millimeter-scale physical deformation into a clearly discernible image displacement.

[0052] Simultaneously, the industrial camera is adjusted so that its optical axis points towards the laser projection area and maintains a 30-60 degree acquisition angle with the laser plane normal, ensuring that the field of view completely covers the laser beam and avoids the specular reflection blind zone of the metal can wall. After adjustment, it is verified that the laser stripes can completely cover the test area of ​​the seal, which includes a primary compensation seal (first seal) and a secondary oil-scraping seal (second seal), forming continuous laser stripes. Here, the first seal is specifically a primary compensation seal, and the second seal is specifically a secondary oil-scraping seal.

[0053] The monitoring unit is fixed by a flexible stainless steel bracket, which integrates a damping and shock absorption module composed of a metal spring and a high-damping rubber pad.

[0054] During the raising and lowering of the floating platform, this shock absorption module is responsible for absorbing mechanical vibrations caused by sliding friction and liquid surface fluctuations, reducing disturbances to the stability of the optical system axis.

[0055] Step S104: Narrowband Filtering and Explosion-proof Encapsulation. A narrowband filter with a center wavelength of 650nm and a bandwidth of 10nm is installed in front of the industrial camera lens. Its peak transmittance is greater than 92%, which allows efficient passage of effective laser signals at a wavelength of 650nm. At the same time, it can effectively block stray light in the visible light band other than laser light, including natural light entering through the light-collecting hole on the top of the tank, artificial maintenance lighting, and secondary reflection light from the metal tank wall. This ensures that the camera only performs monochromatic acquisition at the target wavelength and obtains a striped image with a high signal-to-noise ratio.

[0056] The entire monitoring unit was then encapsulated in an IP67-rated explosion-proof enclosure, and the circuit interface was sealed with an explosion-proof stuffing box to meet the explosion-proof compliance requirements of hazardous areas in the petrochemical industry, enabling the optical sensing device to operate safely for a long time inside storage tanks containing flammable oil and gas mixtures.

[0057] Thus, through step S1, a compact, tightly protected, and complex optical interference-suppressing non-contact active optical detection front-end has been successfully constructed inside the storage tank. The clear, high-contrast laser stripe image generated by this front-end will serve as a high-quality pixel-level data source. In subsequent step S2, the center line of the laser stripes with sub-pixel precision will be extracted from this image using a grayscale centroid algorithm, laying the data foundation for cross-dimensional reconstruction from two-dimensional image coordinates to the three-dimensional sealed profile. Furthermore, the system establishes a real-time communication connection with the tank's built-in level gauge via a data interface. The current liquid level height data output by the level gauge is periodically transmitted to the edge computing module in the form of a 4-20mA analog signal or Modbus digital protocol, providing liquid level dimensional information for the fault space index in subsequent step S5.

[0058] The next step, S2, follows the high-contrast laser stripe image output from S1. Subpixel center lines of the laser stripes are precisely extracted from this image, providing a reliable data foundation for subsequent 3D reconstruction. Once the monitoring system is activated, the industrial camera synchronously acquires images of the laser stripes modulated by the geometric topography of the sealed component surface.

[0059] Since the laser plane is projected onto the surface of the seal at a certain angle, the undulations, wrinkles or radial displacement on the seal will cause the originally continuous laser stripes to produce corresponding distortions on the camera's photosensitive surface. These distortions are the original signal carriers for subsequent extraction of deformation information.

[0060] Step S201: Divide the acquired raw image into regions of interest. Each acquired frame is first sent to the edge computing module, which performs binarization processing on the raw image. The binarization threshold is set to 60% to 80% of the maximum gray value of the image to initially separate the high-brightness laser stripe region.

[0061] Subsequently, a connected component analysis algorithm was used to identify the upper and lower boundary positions of the laser stripes in the image. Based on these upper and lower boundaries, a rectangular region of interest (ROI) containing the complete laser stripes was automatically delineated by extending outwards by 50 pixels. By performing subsequent calculations only within the ROI, the amount of computational data for a single frame image can be reduced by more than 80%, significantly alleviating the processing burden on the edge computing module.

[0062] Step S202: Spatial domain smoothing filtering is performed on the ROI region. The system calls a Gaussian filter operator with a preset size of 5×5 to perform convolution smoothing processing on the ROI region. The 5×5 Gaussian convolution kernel is generated as follows: within a 5×5 neighborhood centered on the current pixel, the weight coefficient of each position is calculated according to a two-dimensional Gaussian function. The standard deviation σ of the two-dimensional Gaussian function is set between 0.8 and 1.2, and the sum of all weight coefficients in the convolution kernel is normalized to 1.

[0063] This filtering operation can specifically eliminate thermal noise and laser speckle noise generated by the CMOS sensor in the high-temperature environment inside the tank, making the edges of the filtered laser stripes smooth and significantly improving the signal-to-noise ratio, providing a cleaner grayscale distribution for subsequent center extraction.

[0064] Step S203: The sub-pixel center coordinates of the laser stripes are extracted using a grayscale centroid algorithm. This algorithm operates on the stripe region within the filtered ROI, performing centroid calculation row by row along the normal direction of the laser stripes within the ROI region. Specifically, starting with the top row 1 of the ROI, the algorithm iterates downwards to the bottom row m.

[0065] For the c-th row of the current iteration, extract all consecutive pixel segments in that row whose grayscale value is higher than a preset threshold T as the effective cross-section of the stripes on that row. The threshold T is set to 50% to 70% of the maximum grayscale value of that row. Within the defined effective cross-section range, calculate the sub-pixel level coordinates of the stripe center on that cross-section according to the grayscale centroid formula. After calculating the c-th row, increment the row index c by 1 and continue processing the next row until all rows within the ROI have been processed. Arrange the center point coordinates calculated for all effective rows in row index order to form the complete sub-pixel precision laser stripe center line.

[0066] The standard mathematical formula for calculating grayscale centroid is expressed as follows:

[0067]

[0068] In this formula: This represents the subpixel center coordinates of the laser stripe calculated on the c-th row section. Specifically, it is the horizontal coordinate value in the image coordinate system, in pixels. The subscript c is the row index number, which is an integer from 1 to m, where m is the total number of rows in the ROI region. This represents the grayscale value of the i-th pixel in the cross section after Gaussian filtering in step S202, and is a dimensionless value. This represents the original discrete x-coordinate of the i-th pixel in the image coordinate system, in pixels; This represents the total number of pixels involved in the calculation within the cross-section of that row, i.e., the number of pixels contained in a continuous pixel segment in that row whose grayscale value is higher than a preset threshold T; summation symbol. This indicates that the grayscale weights of the first to nth pixels in the cross-section are accumulated sequentially.

[0069] The grayscale centroid algorithm can achieve a center positioning accuracy of 0.1 pixels, which can convert the stripe offset that is difficult to distinguish with the naked eye in the original image into a quantifiable coordinate change, providing a high-precision two-dimensional input data source for establishing a triangulation model and realizing three-dimensional deformation measurement at the millimeter or even micrometer level in the subsequent step S3.

[0070] In summary, step S2 completes the entire extraction process from the original laser stripe image to the sub-pixel precision centerline. The obtained centerline point sequence accurately carries the modulation information of the geometric topography of the sealing surface. These center coordinates will be combined with the pre-calibrated camera intrinsic parameter matrix and laser plane equation in step S3, and the true three-dimensional coordinates of the corresponding points in physical space will be solved through a triangulation model, completing the cross-dimensional reconstruction from a two-dimensional image to a three-dimensional contour.

[0071] Step S3 then transforms the sub-pixel precision laser stripe centerline output in step S2 from two-dimensional image coordinates into three-dimensional physical space coordinates, thereby achieving a true spatial description of the geometric contour of the sealing surface. This cross-dimensional transformation relies on a pre-established camera imaging model and laser plane equations. The spatial intersection coordinates are solved point by point using the principle of triangulation, and dynamic environmental disturbances are eliminated through time series filtering, ultimately outputting a stable and quantifiable three-dimensional contour point cloud of the sealing component.

[0072] Step S301: Calibrate system parameters and establish a mapping benchmark from image to space. Before the system is officially put into operation, the intrinsic parameter matrix of the industrial camera is obtained using the Zhang Zhengyou calibration method. During the calibration process, a checkerboard calibration board is used to take multiple images in different orientations within the camera's field of view. The focal length is then calculated using corner point extraction and optimization iteration. Principal point coordinates The intrinsic parameter matrix of the camera consists of at least third-order radial distortion coefficients and second-order tangential distortion coefficients.

[0073] Represents the equivalent focal length along the x-axis in the image coordinate system, in pixels; Represents the equivalent focal length along the y-axis of the image coordinate system, in pixels; The x-coordinate of the intersection point of the camera's optical axis and the image plane, in pixels; The vertical coordinate of the intersection point of the camera's optical axis and the image plane is expressed in pixels.

[0074] Simultaneously, the spatial pose relationship between the camera coordinate system and the laser plane is determined using an external reference object. This external reference object can be a high-precision calibration block with a flat surface, placed within the overlapping area of ​​the laser plane and the camera's field of view. Images of the laser stripes on the calibration block's surface are captured using the camera. The center of the stripes is extracted, and combined with the known geometric dimensions of the calibration block, the rotation matrix of the laser plane in the camera coordinate system is calculated. Translation vector Rotation matrix Let be a 3x3 orthogonal matrix describing the rotational transformation of the laser plane coordinate system relative to the camera coordinate system; translation vector. This is a 3x1 column vector describing the translation position of the origin of the laser plane coordinate system in the camera coordinate system, with its element unit being millimeters.

[0075] Step S302: Establish the laser plane equation and the measurement ray equation, and construct a triangulation model. Based on the rotation matrix obtained in step S301... Translation vector The spatial position of the laser plane in the camera coordinate system can be expressed as a standard mathematical formula using plane equations:

[0076]

[0077] In this formula: Represents the three coordinate components of a point in space in the camera coordinate system, in millimeters; The normal vector components of the laser plane are given by the rotation matrix. The corresponding column is derived, and the three components satisfy... The unit vector constraint has no physical units; This is a constant term in the plane equation, and its absolute value is the perpendicular distance from the origin of the camera coordinate system to the laser plane, in millimeters. The specific values ​​are determined by the translation vector. It is obtained by taking the negative of the dot product with the normal vector.

[0078] After establishing the laser plane equation, the center coordinates of each sub-pixel of the laser stripe output in step S2 are... Based on the principle of perspective projection, a measurement ray is constructed originating from the optical center of the camera and passing through the image point. Let the coordinates of the image point be... ,in , This represents the ordinate of the image corresponding to the current row index, in pixels. The parametric equation of this measurement ray is expressed in standard mathematical form in the camera coordinate system as follows:

[0079]

[0080] In this system of parametric equations: This represents the three coordinate components of a spatial point on the measurement ray in the camera coordinate system, in millimeters; The x-coordinate of the image representing the center point of the laser stripe is taken from the output of step S2. The unit is pixels; The vertical coordinate of the center point of the laser stripe is the row number of the row containing the center point in the image, in pixels. These are the coordinates of the camera's principal point, derived from the calibration results in step S301, in pixels. The equivalent focal length of the camera is derived from the calibration result in step S301, and the unit is pixels. The depth coordinates of spatial points on the ray are used as parameters, with units in millimeters.

[0081] By combining the above laser plane equations and the measurement ray parameter equations—that is, by substituting the ray parameter equations into the plane equations—the depth parameters can be solved. Substituting these equations back into the ray parameter equations, the three-dimensional coordinates of the sampling point on the seal surface corresponding to the center point of the laser fringe in the camera coordinate system can be calculated. .

[0082] The above simultaneous solution operation is performed on all the center points of the laser stripes output in step S2 one by one. All the obtained three-dimensional coordinate points are arranged in the order of the original image rows to form a three-dimensional contour point cloud of the sealing surface at the laser projection section.

[0083] Step S303: Perform time-series filtering on the 3D contour data to eliminate instantaneous disturbances under dynamic conditions. The system allocates a circular buffer in memory with a preset capacity of 15 consecutive frames of 3D contour data. The data update mechanism of the circular buffer is as follows: when a new frame of 3D contour data arrives, if the buffer is not full, it is written directly; if the buffer is full, the earliest timestamp of the written frame is overwritten.

[0084] The specific algorithm for time series filtering is as follows: For all valid frames in the buffer at the current time step, calculate the arithmetic mean of their three-dimensional coordinates point by point. Let the number of valid frames currently stored in the buffer be... , The value range is 10 to 15, for the first [value] in the contour point cloud. The coordinates of each sampling point, after filtering, are expressed as a standard mathematical formula:

[0085]

[0086] In this formula: Indicates the first The three-dimensional coordinate vector of each sampling point after time series filtering contains These three components are in millimeters; Indicates the number of valid frames participating in the averaging calculation, with a value between 10 and 15; Indicates the first The first frame data The original three-dimensional coordinate vector of each sampling point contains These three components are in millimeters; summation symbol. Indicates the first frame to the second frame. The coordinate vectors of the corresponding sampling points of the frame are accumulated sequentially.

[0087] This filtering process can effectively filter out random error interference caused by slight fluctuations in the liquid surface and instantaneous micro-vibrations of the floating plate due to liquid level inertia in single-frame measurement, so that the repeatability error of the measurement value is controlled within 0.05mm, ensuring that the true geometric characteristics of the seal can still be stably expressed under the complex dynamic conditions inside the storage tank.

[0088] At this point, step S3 completes the full transformation and stabilization process from the two-dimensional stripe center coordinates to the three-dimensional physical contour. The output three-dimensional contour point cloud accurately describes the true spatial morphology of the seal surface at the current moment, containing both the overall curvature information of the seal and retaining the detailed features of local deformation. This three-dimensional contour data will be directly fed into the subsequent step S4 for fine registration and deviation calculation with the pre-stored reference contour, thereby quantifying the normal deformation and tangential clearance of the seal and providing direct geometric basis for the final fault determination.

[0089] The next step is S4, which involves a detailed comparison between the real-time 3D contour point cloud output in step S3 and the pre-stored reference contour in the system. After eliminating global pose deviation, the deformation degree and gap size of the seal are quantified from the normal and tangential dimensions, respectively, to provide direct geometric basis for subsequent fault determination. This includes the following steps.

[0090] Step S401: Acquire baseline profile data. The system memory pre-stores baseline profile point clouds under standard sealing conditions. These baseline profiles are collected and saved under ideal sealing conditions during the initial operation of the storage tank. They can also be re-collected and updated through the system's self-learning function after replacing the sealing components.

[0091] During the acquisition of the baseline profile, the system control monitoring unit moves along the inner floating disk circumferentially at preset angular steps and triggers the acquisition and reconstruction processes in steps S1 to S3 point by point. The three-dimensional profile point clouds acquired at each circumferential angular position are stitched together according to the circumferential angular index order to form a complete circumferential baseline profile dataset. This baseline profile accurately records the curvature distribution, radial position, and initial contact relationship between the seal and the tank wall under the design compression ratio, serving as a reference benchmark for subsequent deviation calculations.

[0092] Step S402: Register the real-time contour with the reference contour to eliminate global coordinate offset. During the real-time monitoring phase, the system compares and fits the current frame 3D contour point cloud output in step S3 with the reference contour point cloud at the corresponding circumferential angle position in step S401.

[0093] The part of the reference contour point cloud that corresponds to the real-time contour is retrieved by indexing the circumferential angle value provided by the circumferential encoder integrated in the monitoring unit in step S1. The circumferential encoder has an angular resolution of 0.1 degrees to 1 degree.

[0094] Because the floating roof undergoes minor rotation and translational drift during operation, a rigid body transformation-type global coordinate offset exists between the real-time profile and the reference profile. To eliminate this offset, the system employs the Intermediate Closest Point (ICP) algorithm for fine registration.

[0095] The specific process of ICP registration is as follows: In the first iteration, the reference contour point cloud is used as the target point set and the real-time contour point cloud is used as the source point set. For each sampling point in the source point set, the nearest corresponding point in the target point set is found through nearest neighbor search to form an initial set of matching point pairs. The rigid body transformation matrix of the current iteration step is estimated using the least squares method using the set of matching point pairs, including the rotation transformation increment and translation transformation increment. The estimated rigid body transformation matrix is ​​applied to the source point set to update the position of the source point set. The average distance change of the matching point pairs between the previous and next iterations is used as the convergence criterion. When the change is less than the preset convergence threshold of 0.01mm, the iteration stops. Otherwise, the next iteration continues to perform nearest neighbor search and transformation matrix estimation.

[0096] After ICP registration is completed, the optimal rigid body transformation matrix is ​​obtained, which maximizes the overlap between the real-time contour and the reference contour in the non-fault region. This matrix is ​​then applied to the entire real-time contour point cloud to complete the correction of global rotation and drift offset.

[0097] Step S403: Calculate the normal deviation value. After registration, the system determines the corresponding normal vector direction on the reference contour surface for each sampling point on the real-time contour. Specifically, for the p-th sampling point on the real-time contour, the three points with the closest spatial distance to the sampling point are extracted from the registered reference contour point cloud. A local plane is fitted using these three reference points, and the normal vector of this local plane is the normal vector direction corresponding to the sampling point.

[0098] Then, the directed distance from the sampling point along this normal vector direction to the aforementioned local plane is calculated. This distance is defined as the normal deviation value of the sampling point. If the deviation value is greater than zero, it indicates that the seal at the sampling point bulges towards the tank wall relative to the reference state; if the deviation value is less than zero, it indicates that the seal is recessed inward relative to the reference state.

[0099] Step S404: Calculate the tangential clearance value. The calculation of the tangential clearance value focuses on the physical gap between the edge of the seal and the inner surface of the tank wall. The calculation process is as follows: In the registered 3D contour point cloud, scan point by point along the laser projection direction from the seal side to the tank wall side, and detect the abrupt changes in curvature between adjacent sampling points in the point cloud and the corresponding jumps in grayscale values ​​in the original image.

[0100] Curvature abrupt change is determined by calculating the local curve turning angle formed by three adjacent sampling points. When the local turning angle exceeds the preset angle threshold of 45 degrees, it is considered a curvature abrupt change. Gray level jump is determined by the gray level gradient of the corresponding pixel position in the ROI image before filtering in step S2. When the gray level difference between adjacent pixels exceeds the preset gray level threshold of 50, it is considered a gray level jump.

[0101] When a location simultaneously satisfies both the abrupt change in curvature and the abrupt change in grayscale, that location is identified as the break point at the edge of the seal. Points on the tank wall surface are identified in the same manner. After determining the coordinates of the break point on the seal side and the surface point on the tank wall side on the horizontal projection plane, the Euclidean distance between them is calculated. This distance corresponds to the physical gap between the seal and the tank wall, in millimeters.

[0102] Thus, step S4 completes the entire calculation process from the 3D contour point cloud to the quantitative sealing deviation and gap. The output normal deviation value sequence and tangential gap value directly quantify the degree to which the seal deviates from the ideal state at different locations, including both the overall deformation trend and the details of local defects. These quantitative deviation and gap data will be sent to the multi-level fault judgment logic in step S5, and compared with the preset deviation thresholds and gap thresholds at each level to achieve automatic identification and spatial positioning of sealing fault types such as wrinkles, tears, and detachments.

[0103] Finally, in step S5, the normal deviation value sequence and tangential gap value output in step S4 are received. Through a preset multi-level judgment logic, the quantitative geometric deviation is transformed into a clear sealing fault type and spatial location information. At the same time, a fault spatial index with circumferential orientation and liquid level height is generated to provide directly usable diagnostic conclusions for tank maintenance decisions. Specifically, the steps are as follows.

[0104] Step S501: Local wrinkling determination is performed on the normal deviation value. The system compares the normal deviation value sequence calculated in step S4 with a preset first deviation threshold. This first deviation threshold is set to 3mm. When the normal deviation value exceeds 3mm in a continuously distributed local area along the circumference, and the span of this area is not less than 50mm, and the deviation values ​​of adjacent sampling points in this area alternate in sign, exhibiting a wave-like undulation characteristic with alternating peaks and troughs, it is determined that local wrinkling has occurred in the seal within this area. After the determination is completed, the system records the circumferential start and end azimuth angles and the radial deviation peak value of the wrinkled area.

[0105] Step S502: Determine tear damage based on the normal deviation value. When a sudden, discontinuous numerical jump occurs in the normal deviation value sequence between adjacent sampling points, the system further examines the deviation value characteristics on both sides of the jump position. The specific determination condition is: the absolute value of the difference between the normal deviation values ​​of two adjacent sampling points exceeds a preset jump threshold of 2mm, and in the three-dimensional contour point cloud output in step S4 corresponding to this jump position, a breakpoint with a spacing exceeding 1mm appears in the continuous point sequence along the laser projection direction, indicating that the contour line exhibits physical fracture characteristics at this location. When both conditions are met simultaneously, it is determined that the seal has tear damage at this location. The system records the circumferential azimuth angle and tear width corresponding to the tear position.

[0106] Step S503: A partial detachment determination is performed on the tangential clearance value. The system compares the tangential clearance value calculated in step S4 with a preset clearance threshold. This clearance threshold is set to 5mm. When the tangential clearance value at a sampling location is greater than 5mm, it is determined that the seal at that location has partially detached, meaning that the edge of the seal has separated from the inner surface of the tank wall beyond the allowable range, potentially posing a risk of oil and gas leakage.

[0107] Step S504: Generate a fault space index. The system integrates the various fault results determined in steps S501 to S503 with the data from the circumferential encoder on the monitoring unit and the liquid level gauge on the storage tank to automatically generate a fault space index. The circumferential encoder provides the circumferential azimuth angle of the fault point in the circumferential direction of the inner floating roof, with an angle range of 0 degrees to 360 degrees, using the fixed reference mark of the storage tank as the zero point; the liquid level gauge provides the current height value of the medium liquid level in the storage tank at the time of the fault, in meters.

[0108] The generated fault information includes the following fields: fault type, which takes at least one of wrinkle, tear, or detachment, corresponding to the judgment results of steps S501, S502, and S503, respectively; severity level, which is divided into three levels: mild, moderate, and severe, based on the extent to which the deviation or gap value exceeds the limit; circumferential azimuth angle, a floating-point number in degrees; and current liquid level height, a floating-point number in meters. This fault information is stored in the system database in timestamp order, supporting retrieval and statistical analysis by fault type and severity.

[0109] At this point, step S5 completes the entire judgment process from geometric deviation to a clear fault diagnosis conclusion. The generated fault space index accurately locates the abnormal state of the seal to its specific position and liquid level within the tank, providing structured fault event input for subsequent preventive maintenance analysis in the health status database. Combined with the long-term accumulated contour deviation trend data in the health status database, the system can further distinguish between sudden failures and progressive material fatigue degradation, issuing preventive replacement recommendations before the seal completely fails, thus upgrading the tank maintenance mode from passive post-failure repair to proactive preventive maintenance.

[0110] The system also integrates a health status database for performing preventative maintenance analysis. The daily average profile deviation is stored in the database; specifically, this average deviation is the arithmetic mean of the absolute values ​​of the normal deviations of all sampling points on that day, measured in millimeters. The system uses the least squares method to perform linear regression on the daily average profile deviation sequence stored over the past 30 to 90 days, fitting a straight line with the cumulative number of days as the independent variable and the average deviation as the dependent variable. The slope and intercept of this line are determined by the least squares error minimization principle. Based on this linear equation, the system substitutes the cumulative number of days corresponding to the 15th day after the current date to calculate the predicted deviation value 15 days later. The system has a preset safety threshold of 3 millimeters. When the predicted deviation value reaches or exceeds 3 millimeters, the control room receives an advance preventative replacement recommendation along with a 3D reconstructed image of the fault area, providing intuitive data support for the development of maintenance plans.

[0111] On the other hand, the machine vision-based internal floating disk sealing monitoring system disclosed in this application includes:

[0112] The optical detection module, integrated into the edge of the inner floating plate that rises and falls with the liquid level, includes a line laser, an industrial camera with a narrow-band filter, a circumferential encoder, and explosion-proof protection components. It is used to acquire high-contrast laser stripe images in extremely dark lighting and metal reflection environments inside the storage tank and output circumferential azimuth signals in real time.

[0113] The centerline extraction module is used to perform binarization processing and connected component analysis on the laser stripe image to delineate the region of interest, and to extract the centerline of the laser stripe with sub-pixel precision using a gray-scale centroid algorithm.

[0114] The 3D reconstruction module is used to establish a triangulation model based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, convert the center line of the laser stripe with sub-pixel precision into the 3D contour point cloud of the seal, and perform time series filtering on the 3D contour point cloud of multiple consecutive frames through a ring buffer.

[0115] The registration and calculation module is used to retrieve the corresponding reference contour point cloud based on the circumferential encoder signal, and to perform fine registration between the real-time 3D contour point cloud and the reference contour point cloud using the nearest point iteration algorithm, and to calculate the normal deviation value and tangential gap value after registration.

[0116] The fault diagnosis and location module compares the normal deviation value and tangential clearance value with preset multi-level fault thresholds to determine the fault type of wrinkling, tearing or detachment. The determination result is then fused with the circumferential encoder data and the liquid level height data received from the external liquid level gauge to generate a fault space index that includes the fault type, severity, circumferential azimuth angle and current liquid level height.

[0117] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0118] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A machine vision-based internal floating roof seal monitoring method, characterized by, Includes the following steps: Acquire laser stripe images captured by an optical detection module integrated into the edge of an inner floating disk that rises and falls with the liquid level. The optical detection module includes a line laser, an industrial camera with a narrow-band filter, and a circumferential encoder for recording the rotational position. The laser stripe image is binarized and connected component analysis is performed to delineate a rectangular region of interest containing the complete laser stripe. The gray-scale centroid algorithm is then used to extract the center line of the laser stripe with sub-pixel precision from the region of interest. Based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, a triangulation model is established to convert the center line of the laser stripe with sub-pixel precision into a three-dimensional contour point cloud of the seal in the camera coordinate system, and time series filtering is performed on the three-dimensional contour point cloud of multiple consecutive frames. Retrieve the reference contour point cloud corresponding to the current circumferential encoder angle value from the memory. Use the nearest point iteration algorithm to finely register the real-time three-dimensional contour point cloud after time series filtering with the reference contour point cloud to eliminate global coordinate offset. Calculate the normal deviation value of the real-time contour relative to the reference contour and the tangential gap value between the sealing edge and the tank wall after registration. Based on the multi-level comparison results of the normal deviation value and the preset deviation threshold, and the tangential clearance value and the preset clearance threshold, the fault type of the seal is determined, and the circumferential azimuth angle data provided by the circumferential encoder and the current liquid level height data provided by the liquid level gauge are integrated to generate a fault space index.

2. The machine vision-based inner floating roof seal monitoring method of claim 1, wherein, In the step of acquiring the laser stripe image collected by the optical detection module, the line laser is a visible red laser with a wavelength of 650nm, and the narrowband filter is specifically configured with a center wavelength of 650nm and a bandwidth of 10nm to obtain high-contrast laser stripes in extremely dark lighting and metal reflection environments. The circumferential encoder is used to output angular position signals in real time when the floating disk undergoes circumferential motion or rotation, providing a basis for azimuth positioning for subsequent fault space indexing.

3. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, A grayscale centroid algorithm is used to extract the center line of the laser stripe with sub-pixel precision from the region of interest. Specifically, the algorithm involves traversing the region of interest row by row along the normal direction of the laser stripe, extracting continuous pixel segments with grayscale values ​​higher than a preset threshold in each row as the effective cross-section of the stripe, and then applying the formula... Calculate the sub-pixel center coordinates on this cross section, where The x-coordinate of the sub-pixel center. The grayscale value of the pixel after filtering. The original discrete x-coordinate of the pixel. The total number of pixels involved in the calculation is represented by the sub-pixel center coordinates of each row, which are arranged in row index order to form the complete laser stripe center line.

4. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, Based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, a triangulation model is established to convert the sub-pixel precision laser stripe centerline into a three-dimensional contour point cloud of the sealing component in the camera coordinate system. Specifically, this includes: pre-obtaining the intrinsic parameter matrix of the industrial camera, including the equivalent focal length, according to Zhang Zhengyou's calibration method. and principal point coordinates ; According to the rotation matrix Translation vector By calibrating the spatial position of the laser plane in the camera coordinate system, the standard form of the laser plane equation is obtained. For any point on the center line of the laser stripe with sub-pixel precision Construct the parameter equations of the measurement ray starting from the optical center of the camera. Solve for the depth parameter by simultaneously solving the laser plane equation and the measured ray parameter equation. Then, substituting back into the equation for the measurement ray parameters, the corresponding three-dimensional coordinates are obtained. All three-dimensional coordinate points constitute the three-dimensional contour point cloud of the seal.

5. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, The nearest neighbor iterative algorithm is used to finely register the real-time 3D contour point cloud after time series filtering with the reference contour point cloud. Specifically, the reference contour point cloud is used as the target point set and the real-time contour point cloud is used as the source point set. The nearest neighbor search is performed iteratively to construct a set of matching point pairs. The rigid body transformation matrix of the current iteration step is estimated by the least squares method and applied to the source point set to update its position until the average distance change of the matching point pairs between two iterations is less than the preset convergence threshold, thus obtaining the optimal rigid body transformation matrix for correcting the global pose of the real-time contour. The calculation of the normal deviation value is specifically as follows: for any sampling point on the real-time contour after registration, the three nearest neighbor points on the reference contour surface are determined to fit a local plane, and the directed distance from the sampling point to the plane along the normal vector direction of the local plane is calculated and defined as the normal deviation value. The calculation of the tangential gap value is specifically as follows: In the registered three-dimensional contour point cloud, the curvature change position and grayscale jump position are detected along the laser projection direction. The position that simultaneously satisfies the curvature change and grayscale jump is identified as the edge break point of the seal. The Euclidean distance between the break point and the point on the side surface of the tank wall on the horizontal projection plane is calculated and defined as the tangential gap value.

6. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, The specific types of seal failures include: when the normal deviation value exceeds the preset first deviation threshold of 3mm in a local area with a continuous circumferential span of not less than 50mm, and the normal deviation value in this area changes alternately with a wave-like undulation feature, it is determined to be a local wrinkle. When the absolute value of the difference between the normal deviation values ​​of adjacent sampling points exceeds the preset jump threshold of 2mm, and physical breakage features with a spacing of more than 1mm appear at the corresponding position in the three-dimensional contour point cloud, it is judged as tearing damage. When the tangential gap value exceeds the preset gap threshold of 5mm, it is determined to be a partial detachment.

7. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, After generating the fault space index, the process also includes a preventive maintenance analysis step: the arithmetic mean of the absolute values ​​of the normal deviations of all sampling points collected daily is stored in the health status database as the average profile deviation. The least squares method was used to perform linear regression on the sequence of average profile deviations stored over the past 30 to 90 days, and a linear equation was obtained with the cumulative number of days as the independent variable and the average deviation as the dependent variable. Based on the linear equation, the deviation value is predicted for the next 15 days. When the predicted deviation value reaches or exceeds the preset safety red line value of 3 mm, a preventive replacement recommendation is issued.

8. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, The time series filtering of the three-dimensional contour point cloud for multiple consecutive frames is specifically as follows: a circular buffer with a preset capacity of 15 consecutive frames is opened in memory. When a new three-dimensional contour point cloud arrives, if the buffer is not full, it is written directly; if it is full, the frame data with the earliest timestamp is overwritten. For all valid frames in the buffer at the current time, calculate the arithmetic mean of their three-dimensional coordinates point by point to filter out random errors caused by liquid surface fluctuations and mechanical vibrations.

9. The machine vision-based internal floating disk sealing monitoring method according to claim 1, characterized in that, The optical detection module is fixed to the annular beam at the edge of the inner floating plate by a flexible stainless steel bracket. The flexible stainless steel bracket integrates a damping and shock absorption module composed of a metal spring and a high-damping rubber pad, which is used to absorb the mechanical vibration caused by sliding friction and liquid surface fluctuation during the raising and lowering of the floating plate, and maintain the stability of the optical axis of the line laser and the industrial camera.

10. A machine vision-based internal floating disk sealing monitoring system, characterized in that, For performing the method according to any one of claims 1-9, comprising: The optical detection module, integrated into the edge of the inner floating plate that rises and falls with the liquid level, includes a line laser, an industrial camera with a narrow-band filter, a circumferential encoder, and explosion-proof protection components. It is used to acquire high-contrast laser stripe images in extremely dark lighting and metal reflection environments inside the storage tank and output circumferential azimuth signals in real time. The centerline extraction module is used to perform binarization processing and connected component analysis on the laser stripe image to delineate the region of interest, and to extract the centerline of the laser stripe with sub-pixel precision using a gray-scale centroid algorithm. The 3D reconstruction module is used to establish a triangulation model based on the pre-calibrated camera intrinsic parameter matrix and laser plane equation, convert the center line of the laser stripe with sub-pixel precision into a 3D contour point cloud of the seal, and perform time-series filtering on the 3D contour point cloud of multiple consecutive frames through a ring buffer. The registration and calculation module is used to retrieve the corresponding reference contour point cloud based on the circumferential encoder signal, and to perform fine registration between the real-time 3D contour point cloud and the reference contour point cloud using the nearest point iteration algorithm, and to calculate the normal deviation value and tangential gap value after registration. The fault diagnosis and location module is used to compare the normal deviation value and the tangential gap value with preset multi-level fault thresholds to determine the fault type of wrinkling, tearing or detachment, and to fuse the determination result with the circumferential encoder data and the liquid level height data received from the external liquid level gauge to generate a fault space index that includes fault type, severity, circumferential azimuth angle and current liquid level height.

Images