A method and system for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture.

Abstract

This invention discloses a method and system for detecting defects on the inner wall of a centrifugal pump volute based on wear depth and texture. The method includes: acquiring three-dimensional point cloud data of the volute's inner wall; determining the current shooting position and acquiring a valid two-dimensional image of the volute's inner wall; preprocessing the three-dimensional point cloud data and the valid two-dimensional image of the volute's inner wall; registering a reference model and the three-dimensional point cloud; establishing a correspondence between the valid two-dimensional image and the reference three-dimensional model; constructing a wear field, a texture anomaly field, and a geometric gradient field; calculating a fusion defect index based on the wear field, texture anomaly field, and geometric gradient field; and calculating the overall risk value and regional risk value based on the fusion defect index to obtain the defect detection result of the volute's inner wall. This invention enables accurate identification, quantitative measurement, and risk assessment of defects on the inner wall of the volute, thereby improving the detection accuracy and operational reliability of key components of centrifugal pumps.

Description

Technical Field

[0001] This invention relates to the field of centrifugal pump volute defect detection technology, and in particular to a method and system for detecting defects in the inner wall of centrifugal pump volute based on wear depth and texture. Background Technology

[0002] In the field of fluid machinery, centrifugal pumps are widely used in industries such as petrochemicals, power, and water conservancy. The volute, as a key flow-through component of the centrifugal pump, endures complex conditions such as solid-liquid two-phase flow impact, slip wear, corrosion, and cavitation during long-term operation. This leads to failure of the inner wall material, thinning of the wall thickness, and the generation of defects, significantly affecting the pump's head, efficiency, and operational safety. Existing fluid mechanics and wear research indicates that the volute, especially the tongue and corner areas, is often the most severely worn region. The wear intensity is significantly affected by factors such as flow velocity, particle concentration, and particle size. Therefore, accurate identification and quantitative assessment of the volute wear state are of significant engineering importance for extending equipment life and improving reliability.

[0003] In current surface wear detection technology, conventional non-destructive testing methods such as 3D scanning and visual image inspection are the mainstays. 3D scanning involves acquiring internal point clouds using lasers or structured light and then identifying defects. This method extracts features such as normal changes and distances to neighboring points at a specific location in three dimensions to discover geometric defects like pits and protrusions, enabling surface defect identification based on real geometric shapes. Visual image inspection uses tools such as industrial cameras to acquire images of part surfaces and combines them with traditional machine learning and deep learning methods to classify or detect targets within a series of images. This method is more efficient than other methods in identifying surface texture defects. Furthermore, researchers conduct risk assessments of surface defects based on the identification and detection.

[0004] However, wear is a change in the height of a part's surface, essentially a three-dimensional aggregate defect. Image acquisition and detection can only record two-dimensional imaging information from the current viewpoint, lacking depth information. Therefore, image detection alone is likely to detect a large number of false wear patterns. Three-dimensional point clouds are essentially discrete sets of three-dimensional spatial coordinate points, which can only represent the geometric shape of an object. Texture is formed by uneven reflection of light when it shines on surfaces with different properties, thus it cannot record the surface texture state.

[0005] Therefore, existing research still has the following shortcomings: 3D point cloud detection technology typically uses only a single data type and cannot present the texture features of an object, making it difficult to comprehensively analyze the complex textures and geometric changes of a surface at the same time. Image detection is limited to a two-dimensional plane and lacks three-dimensional information such as depth and volume, so it cannot fully reflect the three-dimensional state of defects; Current risk assessment is developing towards a comprehensive risk modeling approach that integrates multi-source information, but it lacks a unified risk evaluation index system for complex curved surface internal structures. Summary of the Invention

[0006] To address the aforementioned shortcomings in existing technologies, this invention provides a method for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture. This method systematically solves the problems of existing technologies, which can only characterize the geometric shape of objects but cannot reflect surface texture information; two-dimensional image detection lacks three-dimensional geometric information such as depth and volume, resulting in incomplete defect representation; and existing multi-source detection methods are unable to establish a unified, interpretable, and quantifiable risk assessment index system in complex curved surfaces.

[0007] To achieve the aforementioned objectives, the technical solution adopted by this invention is as follows: a method for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture, comprising: Acquire 3D point cloud data of the inner wall of the volute; Determine the current shooting position and obtain an effective two-dimensional image of the inner wall of the volute; Preprocessing is performed on the 3D point cloud data of the inner wall of the volute and the effective 2D image of the inner wall of the volute. The preprocessed 3D point cloud data of the inner wall of the volute is registered with the reference 3D model; the effective 2D image of the inner wall of the preprocessed volute is used to solve or correct the camera pose, and the correspondence between the effective 2D image and the reference 3D model is established. The wear field is constructed based on the geometric deviation between the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model. A texture anomaly field is constructed based on the preprocessed effective two-dimensional image of the inner wall of the volute. Calculate the geometric gradient field based on the wear field; The fusion defect index is calculated based on the wear field, texture anomaly field, and geometric gradient field. The overall risk value and regional risk value are calculated based on the fusion defect index to obtain the defect detection results of the inner wall of the volute.

[0008] Secondly, the present invention also provides a system for implementing a method for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture, comprising: The point cloud acquisition module is used to acquire three-dimensional point cloud data of the inner wall of the volute. The image acquisition module is used to determine the current shooting position and acquire a valid two-dimensional image of the inner wall of the volute. The pose recording module is used to synchronously record the acquisition position, rotation angle, shooting time, and focal length information. The data preprocessing module is used to preprocess the three-dimensional point cloud data of the inner wall of the volute and the effective two-dimensional image of the inner wall of the volute. The registration module is used to register the preprocessed 3D point cloud data of the inner wall of the volute with the reference 3D model; it is also used to solve or correct the camera pose of the effective 2D image of the inner wall of the volute after preprocessing, and to establish the correspondence between the effective 2D image and the reference 3D model. The wear field construction module is used to construct the wear field based on the geometric deviation between the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model. The texture anomaly field construction module is used to construct a texture anomaly field based on the preprocessed effective two-dimensional image of the inner wall of the volute. The geometric analysis module is used to calculate the geometric gradient field based on the wear field; The fusion evaluation module is used to calculate the fusion defect index based on the wear field, texture anomaly field, and geometric gradient field; The risk assessment module is used to calculate the overall risk value and regional risk value based on the fusion defect index, and obtain the defect detection results of the inner wall of the volute.

[0009] The beneficial effects of this invention are as follows: (1) The present invention reduces the interference of large-area wear and local pits on the rigid body pose solution through the process of “initial determination of candidate stable region - evaluation of current sample normal deviation - update of actual stable region - check of stable region distribution - fine registration of stable region”, and reduces the possibility that the actual wear amount is partially offset by the registration process.

[0010] (2) The present invention adopts a closed-loop image acquisition method, which completes the real-time judgment of the expected focal length, actual focal length and target area image quality before the camera enters the next shooting position; unqualified images are adjusted and reshot in the original position, thereby reducing the number of blurry, overexposed, underexposed, poorly viewed or insufficiently covered images entering the PnP pose correction and texture anomaly field.

[0011] (3) The present invention uses the reference three-dimensional model of the inner wall of the volute as a unified benchmark, and maps the measured point cloud that has been registered in the stable region and the multi-view image that has been verified by focal length and quality to the same reference surface coordinate, so that the wear field, texture anomaly field and geometric gradient field can be fused and analyzed at the same position, thereby improving the integrity and spatial correspondence reliability of the detection of defects in complex inner surfaces.

[0012] (4) The present invention introduces a comprehensive confidence factor into the fusion defect index, which incorporates point cloud sampling quality, effective image mapping quality and registration reliability into the evaluation. At the same time, the fusion weight is dynamically adjusted according to the response of each feature field at different locations, which can suppress unreliable results at locations with poor data quality.

[0013] (5) This invention combines key areas such as the volute tongue, inlet section, and impeller outlet adjacent area to construct spatial weights, and conducts risk assessment on defect cluster areas and the overall volute surface; for areas where there are still no effective images or insufficient point cloud coverage after multiple retakes, it outputs insufficient data or low confidence prompts to avoid misinterpreting the lack of reliable data as low risk, and provides a more complete basis for judging the key points of re-inspection and maintenance priority. Attached Figure Description

[0014] Figure 1 Flowchart of the centrifugal pump volute inner wall defect detection system based on wear depth and texture provided for the embodiment; Figure 2 The movement route map for the point cloud acquisition module during 3D scanning. Detailed Implementation

[0015] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0016] The centrifugal pump volute inner wall defect detection system based on wear depth and texture provided by this invention includes a point cloud acquisition module, an image acquisition module, a pose recording module, a data preprocessing module, a registration module, a wear field construction module, a texture anomaly field construction module, a geometric analysis module, a fusion evaluation module, and a risk assessment module. The point cloud acquisition module acquires three-dimensional topographic data of the volute inner wall; the image acquisition module determines the current shooting position based on a reference three-dimensional model and encoder information, and acquires effective two-dimensional images according to the method of "image capture—quality evaluation—retake or move to the next position for shooting"; the pose recording module synchronously records the acquisition position, rotation angle, shooting time, and camera focal length information; the registration module, based on the original coarse alignment, point-to-surface fine registration, and PnP pose correction, adds candidate stable region evaluation, actual stable region determination, and effective images. The system operates in the following order: "data acquisition and quality control—data preprocessing—spatial registration—defect feature construction—fusion evaluation—risk assessment," as follows... Figure 1 As shown. This system is mainly designed for the inspection of complex internal curved surfaces on the inner wall of pump volutes, and can be extended to other shell-like components with similar complex internal curved surfaces based on similar principles.

[0017] In feasible embodiments of the present invention, the point cloud acquisition module can employ a line laser profile scanner, structured light scanner, or other device capable of acquiring three-dimensional topographic information of the inner wall of the volute; secondly, the image acquisition module preferably employs an industrial camera or lens assembly capable of outputting focal length information for each shot, and is configured with a lens, light source, and polarization or supplementary lighting components adapted to the actual working distance; finally, the pose recording module can employ a rotary encoder, displacement encoder, robotic arm pose feedback unit, or other information acquisition device capable of recording changes in the relative position of the equipment, while simultaneously associating and saving the shooting sequence number, shooting time, axial position, rotation angle, and focal length information with each image. The specific instrument type can be determined based on the volute size, detection accuracy, and on-site space conditions.

[0018] In one embodiment of the present invention, a method for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture includes the following steps: S1. Use the point cloud acquisition module to acquire three-dimensional point cloud data of the inner wall of the volute.

[0019] Before system operation, the casing to be tested is fixed on the testing fixture using a special tool, and the reference 3D model, internal and external parameters of the equipment, detection route, and threshold parameters are initialized. Then, the data acquisition phase begins, with the point cloud acquisition module performing multi-pose 3D scanning of the casing's inner wall.

[0020] If a line laser contour scanning method is used, the scanner moves gradually along the centerline of the volute flow channel, and the movement path is as follows: Figure 2 As shown, the laser beam rotates around the center line, allowing it to sequentially scan the entire inner wall. If structured light is used, point cloud regions with overlapping areas need to be acquired and stitched together. The pose recording module synchronously records the position and angle information at each sampling moment.

[0021] S2. Use the image acquisition module to determine the current shooting position and obtain an effective two-dimensional image of the inner wall of the volute.

[0022] The image acquisition module acquires images along a route matching the 3D point cloud acquisition path, forming a closed-loop acquisition at each shooting position. First, based on the reference 3D model, the combined positional relationship between the camera and the shooting surface, and the position and rotation angle recorded by the encoder, the current target surface region and the camera's initial pose are determined. The camera first acquires an image at the current position, then immediately reads the focal length information and evaluates the image quality of the target surface region. If the requirements are not met, the camera remains at the current shooting position, adjusts the focal length, working distance, shooting direction, exposure parameters, or light source, and re-shoots until the image quality meets the standard or the maximum number of retakes is reached. The system then saves the last candidate image as the original detection record, but marks it as invalid and proceeds to the next shooting position. Additionally, necessary overlapping areas are preserved between adjacent valid images, with a focus on acquiring images of the volute tongue, inlet section, and locally occluded areas.

[0023] Specifically: To avoid image blurring caused by the camera suddenly capturing a distant inner wall after entering the volute tongue or other structural change areas, the target surface position is determined before each shot based on the initial mechanical pose of the current centrifugal pump volute and the reference 3D model, and the expected working distance is calculated. The expression for this is:

[0024] In the formula, For the first m The estimated working distance for each shooting location; The coordinates of the camera's optical center position. The coordinates of the target surface position on the reference 3D model are determined by the current shooting direction; The system establishes a correspondence between working distance and applicable focal length through pre-calibration, and provides the expected focal length for the current position accordingly. This correspondence can be achieved through calibration lookup, segmented relationships, or fitting relationships.

[0025] Afterwards, the camera captures a candidate image at the current location. The system then reads the actual focal length of the captured image and compares it with the predicted focal length. The expression is:

[0026] In the formula, For the first m Focal length deviation index of candidate images To actually read the focal length, To estimate the focal length, To prevent positive constants with a denominator of zero, if the actual focal length deviates too much from the expected value, it indicates that the current lens state does not match the distance to the current target surface, and the candidate image is not considered a valid image.

[0027] The expression for calculating the sharpness of candidate images is as follows:

[0028] In the formula, For the first m Target region sharpness of candidate images The target surface image region is determined by the projection of the reference model. and These represent the image gradients in the horizontal and vertical directions, respectively. It is preferable to use image gradient energy to represent sharpness, avoiding misjudgments caused by averaging the entire image when nearby irrelevant areas are sharper than the inner wall of the target.

[0029] Further statistical analysis of the overexposure ratio, underexposure ratio, viewing angle evaluation value, and effective coverage within the target curved surface image region, combined with the candidate image sharpness, yields a comprehensive quality evaluation value for the candidate image:

[0030] In the formula, For the first m The overall quality evaluation value of the candidate images; The overexposure ratio of the target area; The darker ratio of the target area; This is a normalized evaluation value based on the observation perspective. The effective coverage of the target area; , , , and All are weighting coefficients, and the sum of the weights is 1; Candidate images are only deemed valid if they meet the requirements for focal length deviation, overall quality evaluation value, and all quality assessment criteria. If an image is deemed invalid, it is adjusted and retaken; if the maximum number of retakes is still insufficient, the corresponding reference surface area is marked as an area with insufficient image data.

[0031] Criteria for judgment:

[0032] in, For the first m A valid criterion for candidate images, This is the focal length deviation threshold. The overall image quality threshold; This indicates that the image is valid. If the image is invalid, a retake will be initiated. If the number of retakes has not reached the upper limit, the image will be retaken after adjusting the current position. If the upper limit for retakes is reached, the corresponding area will be marked as an area with insufficient image data.

[0033] The threshold can be determined based on lens and camera calibration results, repeated shooting data of healthy samples, permissible image registration error, and the minimum scale of texture defects that need to be identified. If a valid image is not obtained after reaching the maximum number of retakes, the system must not directly determine that the area is defect-free.

[0034] S3. Use the pose recording module to synchronously record the acquisition position, rotation angle, shooting time, and focal length information.

[0035] S4. Use the data preprocessing module to preprocess the three-dimensional point cloud data of the inner wall of the volute and the effective two-dimensional image of the inner wall of the volute.

[0036] After point cloud scanning and closed-loop acquisition of 2D images, the system enters the data preprocessing module, which requires denoising and standardization of the point cloud. Due to issues such as occlusion inside the volute, surface reflection, and local measurement gaps, outliers may appear in the acquired raw point cloud. Therefore, the system preferentially employs a statistical outlier removal method. For any point... Calculate the average distance from it to its r nearest neighbors. If the following formula is satisfied, the point is identified as an outlier and removed:

[0037] In the formula, Point The average distance to its neighboring points This represents the overall mean of the neighborhood distances of all points. Indicates the corresponding standard deviation. This represents the anomaly detection threshold, which depends on the overall sparsity of the point cloud map. (The parameter...) The larger the value, the more point cloud data is retained, but the ability to remove outliers weakens. Through the above processing, scattered points generated by reflective areas and measurement blind spots can be initially removed.

[0038] After outlier removal, the data preprocessing module performs voxel downsampling on the point cloud to control the overall point cloud density and reduce subsequent computation. This part divides the space containing the point cloud into sections with sides of length [missing information]. The voxel grid is used, and the centroid or representative point of each voxel is used as the downsampling result. It can be seen that... This indicates the downsampling scale. A value that is too large may weaken the information about small-sized pits or crack boundaries, while a value that is too small will increase the computational load. This value needs to be determined comprehensively based on the expected minimum detectable defect scale and the 3D acquisition effect.

[0039] After point cloud downsampling, the data preprocessing module continues with local smoothing and normal estimation to provide a geometric basis for subsequent point cloud-reference model registration and normal bias calculation. In this embodiment, a principal component analysis method based on the neighborhood covariance matrix is ​​used to estimate the local normal vector for each point. neighborhood point set Construct the covariance matrix:

[0040] in, Let the centroid of the neighborhood point set be . This represents the number of neighboring points. (Regarding the covariance matrix...) After eigenvalue decomposition, the eigenvector corresponding to the smallest eigenvalue can be used as the normal vector of that point.

[0041] While point cloud data processing is underway, the image preprocessing module performs brightness normalization and texture feature preprocessing only on the 2D images already deemed valid during the acquisition phase. This ensures that valid images from different acquisition times and viewpoints are on a relatively consistent basis. The focal length deviation index, comprehensive image quality evaluation value, and image validity flags calculated during the acquisition phase are saved along with the images for subsequent scenarios such as selecting multiple images to be mapped to the same surface position. Images with focal length deviations or substandard image quality that have not been successfully re-captured are not included in this stage.

[0042] S5. The registration module is used to register the preprocessed 3D point cloud data of the inner wall of the volute with the reference 3D model; it is also used to solve or correct the camera pose of the effective 2D image of the inner wall of the volute after preprocessing, and to establish the correspondence between the effective 2D image and the reference 3D model.

[0043] After the point cloud and the effective image are preprocessed respectively, the registration module first registers the measured point cloud with the reference 3D model. Considering that the actual wear area of ​​the inner wall of the volute interferes with the pose solution during point-to-surface registration, the system adds a stable region determination process between the original coarse alignment and fine alignment.

[0044] Specifically: Based on the theoretical design structure of the volute, theoretical wear distribution, or historical test results of the same model of volute, multiple candidate stable regions are first set on the reference 3D model. The candidate stable regions should meet the conditions of low wear and easy and accurate acquisition of point clouds. In addition, the candidate regions should be dispersed along the front, middle, and rear of the flow channel or at different cross-sectional circumferential positions, and the average normal direction of each region should be significantly different to avoid them all concentrating on a local approximate plane.

[0045] After point cloud preprocessing, the system still uses the initial position of the tooling, calibration parts, mounting reference plane or other known geometric relationships to complete the initial alignment, and then re-evaluates the candidate stable region based on the actual point cloud of the current sample.

[0046] The initial normal deviation of the candidate stable region is calculated using the following expression:

[0047] In the formula, For the first i Normal deviation of each measured point after initial alignment; for The unit normal vector of the reference surface at that location; The rotation matrix obtained from the initial alignment; These are the actual measurement points; For the corresponding point on the reference model; This is the translation vector obtained from the initial alignment.

[0048] Correcting the initial normal deviation:

[0049] In the formula, This is the corrected initial normal deviation; For the healthy volute sample at the reference position The baseline deviation at the reference position; when the system has healthy samples of the same model or historical normal test data, the health baseline deviation at the reference position can be deducted first; when there is no corresponding data, the health baseline deviation is zero.

[0050] The stability of individual candidate regions is then summarized into a regional stability index. This index comprehensively considers various indicators to determine whether a candidate region can be considered a stable region. This index considers both the overall deviation of the region and the variation in deviation within the region. It avoids incorrectly retaining uniformly worn regions due to their smooth surfaces, while also excluding regions with localized pits or abnormal undulations. The specific stability index is as follows:

[0051] In the formula, For the first j Stability indices for candidate stable regions; For the first j One candidate region; Preliminary normal deviation after correction to health baseline; For the revised first l Preliminary normal deviation at each measured point; median This indicates taking the median; This is the influence coefficient of the variation in deviation within the region. If... The smaller the value, the closer the region is to the healthy reference surface and the more stable its interior.

[0052] Candidate stable regions whose stability index does not exceed the stability threshold and whose number of valid points within the candidate stable region reaches a preset value are retained as actual stable regions. After completing the single-region screening, it is also necessary to check whether the actual stable region can provide sufficient geometric constraints for solving the 3D translation and rotation parameters between the point cloud and the reference model.

[0053] Calculate the distribution index of the actual stable region:

[0054] In the formula, This is a distributional indicator for the actual stable region; J This represents the actual number of stable subregions. For the first j A stable sub-region center; The average position of the centers of all stable sub-regions; L The feature length of the volute reference model; For the first j Transpose of the average unit normal vector of a stable subregion; η This is an adjustment coefficient for the degree of location dispersion and the degree of normal difference; if... The larger the value, the more favorable the location and surface orientation of the stable region are for registration.

[0055] The actual stable region whose distribution index meets the preset requirements is determined as the final stable region; when When the preset requirements are not met, the system prioritizes supplementing regions with more distant locations or larger differences in average normal direction from the remaining candidate regions; when Once the requirements are met, the final actual stable region is determined. The number of regions should not be limited to a fixed value in the main scheme, but it is preferable to include multiple mutually separated curved surface regions.

[0056] In the final fine alignment, only valid points in the actual stable regions are used, or corresponding registration weights are assigned based on the stability index of each stable region. The point clouds from all the final stable regions are then constructed into a point cloud set. .

[0057] Constructing a point cloud set and a reference 3D model for point-to-surface registration:

[0058] In the formula, The objective function is the point-to-surface registration based on the final stable region; A collection of point clouds constructed for all point clouds in the eventually stable regions; For registration weights; These are the actual measurement points; For the corresponding point of the reference model; The unit normal vector at the reference point; R and t Let be the rotation matrix and translation vector to be determined, respectively; Alignment of the preprocessed 3D point cloud data of the volute inner wall with the reference 3D model is achieved by minimizing the point-to-surface registration objective function based on the final stable region. After the final rotation matrix and translation vector are determined and kept fixed, the normal deviation and wear field are calculated across the entire area to be inspected.

[0059] S6. The wear field is constructed using the wear field construction module based on the geometric deviation between the preprocessed three-dimensional point cloud data of the inner wall of the volute and the reference three-dimensional model.

[0060] After the point cloud and reference model are finally registered, the system calculates or corrects the camera pose of the valid 2D images determined in the acquisition phase and establishes the correspondence between the images and the reference 3D model. Since the 2D images themselves only have pixel coordinates, the accurate camera extrinsic parameters for each valid image must be obtained based on the camera intrinsic parameters, the installation relationship between the camera and the fixture, the shooting position saved by the pose recording module, and the geometric relationship of the reference model.

[0061] satisfy Only valid images are used for 2D and 3D point extraction and PnP pose solving. Candidate images with focal length deviations or substandard image quality are discarded before registration, therefore their corresponding camera intrinsic parameters are not used for pose solving, nor are they involved in reference surface texture mapping and texture anomaly field construction. For the final valid images, the camera intrinsic parameter matrix calibrated at their effective focal length is used, or their focal length is guaranteed to be within the same calibration allowable range.

[0062] During the image pose solving or correction process, if the system has already acquired several 3D reference points and their corresponding pixels in the image Then the perspective-n-point method can be used to solve for the camera extrinsic parameters, satisfying:

[0063] in, K For the camera intrinsic parameter matrix, and They represent the first m The rotation matrix and translation vector corresponding to each image Represents two-dimensional image points in an image. This represents the corresponding three-dimensional point. The three-dimensional reference point... These can be derived from calibration targets placed near the detection fixture or the opening of the volute, or from stable feature positions obtained after registering 3D point clouds with a reference model; the corresponding 2D image points. This method can be obtained through image recognition, target identification, or manual verification. Based on the correspondence between 3D points and 2D image points, it establishes or corrects the camera pose parameters during image acquisition in the post-processing stage, ensuring accurate mapping of the 2D image onto the reference 3D model. The valid image markers generated during the acquisition stage should be associated with and saved along with the image's pose parameters and mapping results, enabling subsequent modules to identify whether the image has undergone focal length and quality verification.

[0064] Then the system's wear field construction module starts working. The wear field is mainly constructed by the geometric deviation between the measured 3D point cloud and the reference model. The 2D image does not directly participate in the solution of the wear depth value, but the 2D image will form texture anomaly information on the same reference surface and will be used together with the wear field in the subsequent fusion of defect index.

[0065] The registration module first transforms the measured point cloud to the reference model coordinate system. The wear field construction module then finds the corresponding position for each measured point on the reference model, calculates the normal deviation at that position, and then maps or interpolates the discrete deviation value onto the reference surface to form a continuous wear distribution.

[0066] Specifically: The normal deviation between any point in the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model is calculated using the following expression:

[0067] In the formula, For the first i The deviation of the measured point from the normal of the reference 3D model. If the material loss direction is defined as positive, then when the measured surface is concave or thinned relative to the reference surface along the material loss direction. If the normal of the reference surface is defined opposite to the material loss direction, the sign and direction can be unified during implementation.

[0068] To form a continuous distribution from discrete deviation samples, the system further uses surface interpolation based on... The wear field is constructed, and its expression is:

[0069] In the formula, Represents the reference surface parameter coordinates Wear field value at the location, The position of the sample point in the parameter domain of the reference surface. For bandwidth The kernel function; Used to determine the first i How close are each sampling point to the current location?

[0070] Considering that there may be manufacturing deviations between the actual good sample and the theoretical model, the system can introduce a deviation correction amount. The wear field is then corrected to obtain the final wear field:

[0071] In the formula, This represents the final wear field obtained after correction. Through this correction, the system obtains a wear distribution result that more closely approximates the actual material loss. Among these, This indicates the coordinates of the reference surface parameters for a good sample of the same model. The average offset relative to the nominal model is used to avoid misinterpreting existing structural differences as wear. If good samples or historically normal inspection samples are available, this can be obtained through statistical methods. When no such samples are available, the value is zero. The nominal model is the ideal geometric model, used as an absolute benchmark.

[0072] S7. Construct a texture anomaly field based on the preprocessed effective two-dimensional image of the inner wall of the volute using the texture anomaly field construction module.

[0073] Texture feature vectors are extracted at each location on the reference surface and compared with the texture statistical model at the same location on the good sample. The texture anomaly field is represented as:

[0074] In the formula, Indicates the degree of texture anomaly. Indicates the texture features at the current location. This represents the mean of texture features under normal conditions. This is the covariance matrix, representing the fluctuation range of normal texture features and the correlation between various features. Normal texture statistics can be obtained from working samples of the same model, normally inspected samples, or manually confirmed defect-free areas. A larger calculated result indicates a more significant deviation of the texture from its normal state at the current location.

[0075] S8. Calculate the geometric gradient field based on the wear field using the geometric analysis module.

[0076] To further characterize the changes in defect boundaries, the geometric analysis module can calculate the geometric gradient field based on the corrected wear field. The geometric gradient field is:

[0077] In the formula, This represents the local rate of change of the wear field on the curved surface.

[0078] S9. Calculate the fusion defect index based on the wear field, texture anomaly field and geometric gradient field using the fusion evaluation module.

[0079] After the wear field, texture anomaly field, and geometric gradient field have been established, the fusion evaluation module begins its work. Because the three types of features are different, the system needs to normalize each feature field before using it as parameters for constructing the fusion defect index. The normalization process specifically involves:

[0080] In the formula, Represents a certain characteristic field The result after normalization This represents the median. Indicates the absolute deviation of the median. To prevent constants with small values ​​whose denominators are zero; Based on the results after normalization, the fusion defect index is calculated:

[0081] In the formula, Indicates the fusion defect index. This represents the overall confidence factor. , , These represent the adaptive weights of the wear field, texture anomaly field, and geometric gradient field, respectively. If there are requirements for the wear characteristics of the curved surface, [the following can be specified]: High; if there are requirements for early surface anomaly conditions, the efficiency can be increased. If there are requirements for the pit edge or crack boundary, the efficiency can be increased. .

[0082] The weights are calculated as follows:

[0083]

[0084] in, Indicates the first The local contribution score of the class feature at the current position; This represents the data reliability coefficient for this type of feature; Represents a nonnegative response function; This represents the normalized corresponding feature field; Indicates the first Adaptive weights of class features at the current position; This represents the basic weight of this type of feature, used to prevent a feature from completely losing its contribution when its response is weak. To prevent small constants with zero denominators. As can be seen from the formula, the weights of each feature satisfy the non-negativity constraint and are comparable after normalization.

[0085] Comprehensive confidence factor The calculation formula is as follows:

[0086]

[0087]

[0088]

[0089] in, Indicates position The number of effective point clouds in the neighborhood. This represents the number of reference points required to satisfy a reliable estimate; Indicates position In the Whether it is visible in the image, and takes a value between 1 and 0; It can be determined by the angle between the camera's line of sight and the surface normal; the closer it is to the orthogonal angle, the higher its value. Indicates the first The overall quality evaluation value of the images; To prevent small constants with a denominator of zero; Indicates position Local registration residuals between the point cloud in the neighborhood and the reference 3D model. Indicates the permissible residual scale.

[0090] Whether the m-th image can provide effective information about the location of the reference surface depends not only on whether the location is visible, but also on whether the image was judged by focal length and image quality during the acquisition phase.

[0091]

[0092] For reference surface position (u,v) In the m Valid visible markers in the image; For the first m A valid identifier for a candidate image.

[0093] S10, the risk assessment module calculates the overall risk value and regional risk value based on the fusion defect index, and obtains the defect detection results of the inner wall of the volute.

[0094] Considering that the areas near the volute tongue, inlet section, and impeller outlet are prone to wear and are crucial for the safe operation of the machine, higher weights are assigned to these critical areas. These weights should be determined based on defect locations, surface geometry, and historically high-wear areas. The spatial weight at a specific location on the inner wall of the volute is defined as follows:

[0095] In the formula, It is the basic weight, which represents the basic spatial weight. It is used to ensure that defects in non-critical areas of the inner wall of the volute still participate in the risk assessment, and to avoid the risk being weakened to zero due to the low values ​​of each spatial influence factor. The position factor is related to key locations such as the volute tongue, inlet section, or impeller outlet adjacent area, and is determined based on the surface distance from the location to the key area. It represents the curvature or structural change factor, used to reflect structurally sensitive locations such as abrupt curvature changes; This refers to the sensitive area factor determined by historical monitoring results, statistics of high-incidence erosion areas, or fluid simulation. The weight coefficients representing each factor can be determined through empirical setting or expert evaluation. Spatial weights are determined in the above manner. It has a clear source and can demonstrate the engineering facts that the same defect may cause different risk impacts in different locations of the volute.

[0096] Key position factor The calculation formula can be defined as:

[0097]

[0098] in, Indicates the surface distance along the reference surface. Represents the set of key regions The position of the curved surface in the middle, This parameter represents the influence range of the critical area, and its value can be set according to the feature length of the reference surface.

[0099] Surface structure variation factor The formula used to reflect the strength of surface curvature or structural change at the current location is defined as follows:

[0100] in, This represents the structural response value at the current location; and These represent the minimum and maximum values ​​of the structural response within the reference surface range, respectively. This represents a small constant to prevent the denominator from being zero.

[0101] Sensitive region factor Used to incorporate historical inspection results, statistics on high-incidence erosion areas, maintenance records, or fluid simulation results, it can be represented as:

[0102] in, and These represent the minimum and maximum sensitivity values ​​on the reference surface, respectively. If no historical detection data or fluid simulation data is currently available, then... Correspondingly At this point, the spatial weight can still be calculated from the basic weight, the key position factor, and the surface structure change factor.

[0103] After obtaining the spatial weights, the risk assessment module calculates the local risk assessment density. The risk energy density described in this invention can be understood as an equivalent risk assessment quantity, used to characterize the combined impact of local defect severity and structural location importance.

[0104] After obtaining the spatial weights, the local risk assessment density is calculated based on the spatial weights and the fusion defect index. This density is used to characterize the combined influence of local defect severity and structural location importance; its expression is as follows:

[0105] In the formula, This represents the equivalent risk energy density of the local surface element. (Using...) The format is designed to make high-severity defects contribute more to risk assessment.

[0106] Subsequently, the system clusters adjacent high-risk units into defect cluster regions. Calculate the regional risk value : ; Integrating over the entire volute surface yields the overall risk value:

[0107] In the formula, Indicates the first k Regional risk value for each defect cluster; This represents the entire volute surface. Through risk calculations at both the local and global levels, the system can simultaneously provide risk assessment results at the single-point, regional, and system-wide levels.

[0108] This invention first determines candidate stable regions based on theoretical design or historical data, then updates the actual stable regions using the initial normal deviation of the current sample, and completes final fine registration after checking for positional dispersion and normal differences. Simultaneously, based on the reference model and encoder information, the expected working distance and expected focal length are determined for each shooting position. For each image captured, focal length deviation and target area image quality are judged; invalid images are re-captured in situ and not included in subsequent registration. After the qualified point cloud and valid images are uniformly mapped to the reference surface, a wear field, texture anomaly field, and geometric gradient field are constructed, and the defect cluster risk, overall volute risk, detection coverage, and low-confidence state are output.

[0109] This invention reduces the interference of large-area wear and local pits on the rigid body pose solution through the process of "initial determination of candidate stable regions based on theoretical or historical data - evaluation of current sample normal deviation - update of actual stable regions - check of stable region distribution - fine registration of stable regions", and reduces the possibility that the actual wear amount is partially offset by the registration process.

[0110] By adopting a frame-by-frame closed-loop image acquisition method, the expected focal length, actual focal length, and image quality of the target area are judged in real time before the camera moves to the next shooting position. Unqualified images are adjusted and reshot in the original position, thereby reducing the number of blurry, overexposed, underexposed, poorly viewed, or insufficiently covered images entering the PnP pose correction and texture anomaly field.

[0111] Using the 3D reference model of the inner wall of the volute as a unified benchmark, the measured point cloud that has been registered in the stable region and the multi-view images that have been verified by focal length and quality are mapped to the same reference surface coordinates, so that the wear field, texture anomaly field and geometric gradient field can be fused and analyzed at the same location, thereby improving the integrity and spatial correspondence reliability of the detection of defects in complex inner surfaces.

[0112] By introducing a comprehensive confidence factor into the fusion defect index, the quality of point cloud sampling, the quality of effective image mapping, and the reliability of registration are incorporated into the evaluation. At the same time, the fusion weights are dynamically adjusted according to the response of each feature field at different locations, which can suppress unreliable results at locations with poor data quality.

[0113] Spatial weights are constructed by combining key areas such as the volute tongue, inlet section, and impeller outlet adjacent area, and risk assessments are performed on defect cluster areas and the overall volute surface. For areas where no effective images are obtained after multiple retakes or where point cloud coverage is insufficient, insufficient data or low reliability prompts are output to avoid misinterpreting unreliable data as low risk, and to provide a more complete basis for judging the focus of re-inspection and maintenance priority.

Claims

1. A method for detecting defects in the inner wall of a centrifugal pump volute based on wear depth and texture, characterized in that, include: Acquire 3D point cloud data of the inner wall of the volute; Determine the current shooting position and obtain an effective two-dimensional image of the inner wall of the volute; Preprocessing is performed on the 3D point cloud data of the inner wall of the volute and the effective 2D image of the inner wall of the volute. The preprocessed 3D point cloud data of the inner wall of the volute is registered with the reference 3D model; the effective 2D image of the inner wall of the preprocessed volute is used to solve or correct the camera pose, and the correspondence between the effective 2D image and the reference 3D model is established. The wear field is constructed based on the geometric deviation between the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model. A texture anomaly field is constructed based on the preprocessed effective two-dimensional image of the inner wall of the volute. Calculate the geometric gradient field based on the wear field; The fusion defect index is calculated based on the wear field, texture anomaly field, and geometric gradient field. The overall risk value and regional risk value are calculated based on the fusion defect index to obtain the defect detection results of the inner wall of the volute.

2. The method according to claim 1, characterized in that, Determine the current shooting position and acquire a valid two-dimensional image of the inner wall of the volute, including: The target surface position is determined based on the initial mechanical pose of the centrifugal pump casing and the reference 3D model, and the expected working distance is calculated. The expression for this is: In the formula, For the first m The estimated working distance for each shooting location; The coordinates of the camera's optical center position. The coordinates of the target surface position on the reference 3D model are determined by the current shooting direction; The expected focal length is determined based on the expected working distance; Capture a candidate image at the current location, read the actual focal length, compare the actual focal length with the expected focal length, and obtain the focal length deviation index, which is expressed as: In the formula, For the first m Focal length deviation index of candidate images To actually read the focal length, To estimate the focal length, To prevent positive constants with a denominator of zero; The expression for calculating the sharpness of candidate images is as follows: In the formula, For the first m Target region sharpness of candidate images The target surface image region is determined by the projection of the reference 3D model. and These represent the image gradients in the horizontal and vertical directions, respectively. The overall quality evaluation value of the candidate image is obtained by statistically analyzing the overexposure ratio, underexposure ratio, viewing angle evaluation value, and effective coverage within the target curved surface image region, combined with the candidate image sharpness: In the formula, For the first m The overall quality evaluation value of the candidate images; The overexposure ratio of the target area; The darker ratio of the target area; This is a normalized evaluation value based on the observation perspective. The effective coverage of the target area; , , , and All are weighting coefficients, and the sum of the weights is 1; When the focal length deviation index and the comprehensive quality evaluation value of the candidate image meet the preset conditions, the current candidate image is used as the effective two-dimensional image of the inner wall of the volute.

3. The method according to claim 1, characterized in that, Preprocessing is performed on the 3D point cloud data and effective 2D images of the inner wall of the volute, including: Calculate the average distance from any point in the 3D point cloud data of the inner wall of the volute to r neighboring points. When the average distance from any point to its neighboring points meets the abnormality condition, the point is identified as an abnormal point and removed. The specific abnormal conditions are as follows: In the formula, Point The average distance to its neighboring points This represents the overall mean of the neighborhood distances of all points. Indicates the corresponding standard deviation. Indicates the anomaly detection threshold; Voxel downsampling was performed on the 3D point cloud data of the inner wall of the volute after outlier removal; Local smoothing and normal estimation are performed on the three-dimensional point cloud data of the inner wall of the volute after voxel downsampling to obtain the preprocessed three-dimensional point cloud data of the inner wall of the volute. Brightness normalization and texture feature preprocessing are performed on the effective two-dimensional images of the inner wall of the volute to make the effective two-dimensional images consistent at different acquisition times and from different viewpoints, thus obtaining the preprocessed effective two-dimensional images of the inner wall of the volute.

4. The method according to claim 3, characterized in that, The preprocessed 3D point cloud data of the inner wall of the volute is registered with the reference 3D model, including: Multiple candidate stable regions are defined on the reference 3D model; The pre-processed 3D point cloud data of the inner wall of the volute is initially aligned with the reference 3D model. The initial normal deviation of the candidate stable region is calculated using the following expression: In the formula, For the first i 1 measured point Normal deviation after initial alignment; for The unit normal vector of the reference surface at that location; express The transpose of the unit normal vector of the reference surface at that location; The rotation matrix obtained from the initial alignment; These are the actual measurement points; For the corresponding point on the reference model; The translation vector obtained from the initial alignment; Correcting the initial normal deviation: In the formula, For the revised first i Preliminary normal deviation at each measured point; For the healthy volute sample at the reference position Reference deviation at; Stability indices for candidate stable regions are calculated based on the corrected initial normal deviation. In the formula, For the first j Stability indices for candidate stable regions; median This indicates taking the median; For the first j One candidate region; For the revised first l Preliminary normal deviation at each measured point; The influence coefficient of deviation variation within the region; Candidate stable regions whose stability index does not exceed the stability threshold and whose number of valid points in the candidate stable region reaches the preset value are retained as actual stable regions. Calculate the distribution index of the actual stable region: In the formula, This is a distributional indicator for the actual stable region; η This is an adjustment coefficient for the degree of location dispersion and the degree of normal difference; J This represents the actual number of stable subregions. L The feature length of the volute reference model; For the first j A stable sub-region center; The average position of the centers of all stable sub-regions; For the first j Transpose of the average unit normal vector of a stable subregion; For the first k The average unit normal vector of a stable subregion; The actual stable region whose distribution index meets the preset requirements is determined as the final stable region; Construct a point cloud set from all the point clouds in the final stable regions; Constructing a point cloud set and a reference 3D model for point-to-surface registration: In the formula, The objective function is the point-to-surface registration based on the final stable region; A collection of point clouds constructed for all point clouds in the eventually stable regions; The unit normal vector at the reference point; R and t Let be the rotation matrix and translation vector to be determined, respectively; Alignment of the preprocessed 3D point cloud data of the inner wall of the volute with the reference 3D model is achieved by minimizing the point-to-surface registration objective function based on the final stable region.

5. The method according to claim 4, characterized in that, The wear field is constructed based on the geometric deviation between the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model, including: The normal deviation between any point in the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model is calculated using the following expression: In the formula, For the first i The normal deviation of each measured point relative to the reference 3D model; based on The wear field is constructed, and its expression is: In the formula, Represents the reference surface parameter coordinates Wear field value at the location; This refers to the position of the sample point in the reference 3D model within the parameter domain of the reference surface; For bandwidth The kernel function; By introducing a deviation correction factor, the wear field is corrected to obtain the final wear field: In the formula, This represents the final wear field obtained after correction; This indicates the coordinates of the reference surface parameters for a good sample of the same model. The average offset relative to the nominal model.

6. The method according to claim 5, characterized in that, Texture anomaly fields are represented as: In the formula, Indicates the degree of texture anomaly; Indicates the texture features at the current location; This represents the mean of texture features under normal conditions; Let be the covariance matrix.

7. The method according to claim 6, characterized in that, The geometric gradient field is: In the formula, This represents the local rate of change of the wear field on the curved surface, i.e., the geometric gradient field; Represents the surface gradient; This represents the Euclidean norm.

8. The method according to claim 7, characterized in that, The fusion defect index is calculated based on the wear field, texture anomaly field, and geometric gradient field, including: The wear field, texture anomaly field, and geometric gradient field are normalized; the normalization process specifically involves: In the formula, Represents a certain characteristic field The result after normalization; This represents the median. Indicates the absolute deviation of the median. To prevent positive constants with a denominator of zero; Based on the results after normalization, the fusion defect index is calculated: In the formula, Indicates the fusion defect index; Indicates the overall confidence factor; , , These represent the adaptive weights of the wear field, texture anomaly field, and geometric gradient field, respectively. This represents the normalized wear field value; This represents the normalized texture anomaly field value; This represents the normalized geometric gradient field value.

9. The method according to claim 8, characterized in that, Based on the fusion defect index, the overall risk value and regional risk value are calculated to obtain the defect detection results of the inner wall of the volute, including: Define the spatial weight at a certain location on the inner wall of the volute as: In the formula, for Spatial weights at each location; As the basic weight; The location factor represents the location associated with key locations, including the volute tongue, the inlet section, and the impeller outlet adjacency area. Indicates the curvature or structural variation factor; This refers to the sensitive area factor determined by historical detection results, statistics of high-incidence erosion areas, or fluid simulation. Indicates the weighting coefficients of each factor; Calculate the local risk assessment density based on spatial weights and fusion defect index: In the formula, This represents the equivalent risk energy density of the local surface element; Calculate the regional risk value: In the formula, Indicates the first k Regional risk value for a specific area; Represents the entire volute surface; express k A specific area; Integrating over the entire volute surface yields the overall risk value: In the formula, This represents the overall risk value.

10. A system for implementing the centrifugal pump volute casing inner wall defect detection method based on wear depth and texture as described in any one of claims 1 to 9, characterized in that, include: The point cloud acquisition module is used to acquire three-dimensional point cloud data of the inner wall of the volute. The image acquisition module is used to determine the current shooting position and acquire a valid two-dimensional image of the inner wall of the volute. The pose recording module is used to synchronously record the acquisition position, rotation angle, shooting time, and focal length information. The data preprocessing module is used to preprocess the three-dimensional point cloud data of the inner wall of the volute and the effective two-dimensional image of the inner wall of the volute. The registration module is used to register the preprocessed 3D point cloud data of the inner wall of the volute with the reference 3D model; to solve or correct the camera pose of the effective 2D image of the inner wall of the volute after preprocessing; and to establish the correspondence between the effective 2D image and the reference 3D model. The wear field construction module is used to construct the wear field based on the geometric deviation between the preprocessed 3D point cloud data of the inner wall of the volute and the reference 3D model. The texture anomaly field construction module is used to construct a texture anomaly field based on the preprocessed effective two-dimensional image of the inner wall of the volute. The geometric analysis module is used to calculate the geometric gradient field based on the wear field; The fusion evaluation module is used to calculate the fusion defect index based on the wear field, texture anomaly field, and geometric gradient field; The risk assessment module is used to calculate the overall risk value based on the fusion defect index and obtain the defect detection results of the inner wall of the volute.

Images