Machine learning-based multi-parameter intelligent detection and evaluation method and system for special-shaped pipes

Abstract

The application discloses a special-shaped pipe multi-parameter intelligent detection and evaluation method and system based on machine learning, relates to the technical field of computer vision and machine learning, and comprises the following steps: a structured light projector dynamically adjusts the density and coding strategy of a projected stripe according to the curvature distribution of a pipe surface; a three-dimensional surface model is constructed by preprocessing collected multi-view images, and the construction process comprises a curvature distortion compensation step; contour size parameters, geometric shape parameters, surface defect parameters and three-dimensional topography parameters are synchronously extracted; the extracted multi-parameter features are input into a multi-task joint learning network model, and a comprehensive quality evaluation result is output and a detection and evaluation report is generated; and the system comprises an image acquisition module, an image preprocessing and three-dimensional construction module, a multi-parameter feature extraction module, an intelligent evaluation module, a report generation module and a control and data processing unit. Through curvature partition adaptive measurement and distortion compensation, the three-dimensional reconstruction precision of the special-shaped pipe and the efficiency of multi-parameter comprehensive evaluation are improved.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and machine learning technology, and in particular to a machine learning-based intelligent multi-parameter detection and evaluation method and system for irregularly shaped pipes. Background Technology

[0002] With the rapid development of high-end manufacturing industries such as aerospace, automobile manufacturing, and nuclear power equipment, the application of irregularly shaped pipes as key structural components is becoming increasingly widespread. Their geometries are becoming more complex, and the requirements for dimensional accuracy and surface quality are becoming increasingly stringent. Currently, the quality inspection of irregularly shaped pipes mainly relies on coordinate measuring machine sampling, manual visual inspection, and traditional machine vision methods. These methods suffer from problems such as slow inspection speed, high subjectivity, single parameters, and poor adaptability to complex curved surfaces, making it difficult to meet the needs of modern mass production for online full inspection and multi-parameter comprehensive evaluation.

[0003] In existing technologies, structured light projectors typically employ fixed density and a single encoding strategy for full-surface projection. This fails to adjust measurement parameters based on variations in pipe surface curvature, leading to data redundancy due to overly dense stripes in curvature abrupt changes and insufficient accuracy due to sparse stripes in straight areas. This makes it difficult to achieve consistent 3D reconstruction accuracy across the entire surface of irregularly shaped pipes. Furthermore, existing technologies lack pre-scanning and curvature zoning mechanisms. Before formal scanning, the system cannot determine the curvature distribution of different areas on the pipe surface, preventing the structured light projector from making targeted adjustments based on curvature characteristics. This results in arbitrary setting of stripe parameters during formal scanning. Additionally, existing technologies do not consider the distortion caused by pipe surface curvature to the projected stripes, nor do they perform reverse deformation compensation based on the design model. This leads to deviations between the projected pattern of structured light stripes in curvature abrupt changes and the ideal pattern, ultimately resulting in systematic measurement errors in the 3D reconstruction at curved sections. Regarding parameter detection, existing technologies typically only detect single parameters, such as dimensions or defects. They lack the ability to simultaneously extract and comprehensively characterize multiple parameters, including contour dimensions, geometry, surface defects, and 3D morphology, making it difficult to achieve a comprehensive assessment of the quality of irregularly shaped pipes. In multi-parameter evaluation, existing technologies typically employ multiple independent models to process each parameter separately. The lack of feature sharing between these models leads to wasted computational resources and low evaluation efficiency. Furthermore, there is a lack of unified quantitative standards and comprehensive judgment mechanisms for the evaluation results of each parameter. Regarding the overall workflow, existing technologies often involve independent processes such as image acquisition, 3D reconstruction, feature extraction, and quality assessment. Data formats are inconsistent, and workflow connections rely on manual intervention, making it difficult to form an automated closed loop from raw data to evaluation conclusions. This results in low detection efficiency, large human errors, and an inability to meet the needs of industrial-scale batch testing.

[0004] Therefore, this invention proposes a multi-parameter intelligent detection and evaluation method and system for irregularly shaped pipes based on machine learning. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-parameter intelligent detection and evaluation method and system for irregularly shaped pipes based on machine learning, in order to solve the problems in the prior art where the structured light projector cannot adaptively adjust the measurement parameters according to the curvature change of the pipe surface, resulting in insufficient accuracy of three-dimensional reconstruction, lack of multi-parameter synchronous extraction and comprehensive evaluation capabilities, and the independent nature of each link in the detection process, making it difficult to form an automated closed loop.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] A machine learning-based intelligent multi-parameter detection and evaluation method for irregularly shaped pipes includes:

[0008] Step 1: Multi-source image data acquisition: Surface image data of the irregularly shaped pipe under test is acquired through a multi-view imaging device to obtain multi-view images. The multi-view imaging device includes at least four industrial cameras arranged at different angles along the circumference of the pipe and at least one structured light projector. The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes according to the curvature distribution of the irregularly shaped pipe surface: a first density stripe and phase encoding method are used in the curvature abrupt change region, a second density stripe and Gray code encoding method are used in the curvature gradually changing region, and a third density stripe and triangulation method are used in the curvature flat region. The first density is greater than the second density, and the second density is greater than the third density.

[0009] Step 2: Constructing a 3D surface model after image preprocessing: Preprocess the multi-view images acquired in Step 1 to obtain preprocessed multi-view 2D images, and construct a 3D surface model of the irregular pipe based on the multi-view 2D images; construct a model including curvature distortion compensation steps, extract theoretical curvature information based on the design model of the irregular pipe, calculate the theoretical distortion mode of the projected fringes under each curvature, and perform reverse deformation compensation on the projected fringes.

[0010] Step 3: Multi-parameter feature extraction: Based on the three-dimensional surface model obtained in Step 2 and the preprocessed multi-view two-dimensional images, extract multiple detection parameter features of the irregularly shaped pipe. The multiple detection parameters include contour size parameters, geometric shape parameters, surface defect parameters and three-dimensional morphology parameters.

[0011] Step 4: Multi-parameter intelligent evaluation based on machine learning: Input the multi-parameter features extracted in step 3 into the trained multi-task joint learning network model, and output the comprehensive quality evaluation results of the irregular-shaped pipe.

[0012] Step 5: Generation of Inspection and Evaluation Report: Based on the comprehensive quality evaluation results of Step 4, an inspection and evaluation report for the irregular-shaped pipe is generated.

[0013] Preferably, the imaging method of the multi-view imaging device in step 1 is as follows: at least four industrial cameras are arranged at non-uniform angle intervals along the circumference of the irregular pipe, a preset baseline distance is maintained between the structured light projector and each industrial camera, and each industrial camera is triggered to acquire surface images of the pipe from different circumferential angles.

[0014] Preferably, the method for dividing the curvature abrupt region, the curvature gradually changing region, and the curvature flat region in step 1 is as follows: a low-density stripe is projected by a structured light projector to pre-scan the irregularly shaped pipe, obtain the coarse contour data of the pipe surface, and calculate the absolute value of curvature of each region; using a first threshold and a second threshold as the dividing boundary, the region with the absolute value of curvature greater than the first threshold is the curvature abrupt region, the region with the absolute value of curvature between the first threshold and the second threshold is the curvature gradually changing region, and the region with the absolute value of curvature less than the second threshold is the curvature flat region, wherein the first threshold is greater than the second threshold.

[0015] Preferably, the preprocessing in step 2 includes image denoising, contrast enhancement, distortion correction, and epipolar correction.

[0016] Preferably, in step 3, the contour dimension parameters include the outer diameter, inner diameter, wall thickness and their uniformity of distribution of the pipe; the geometric shape parameters include the roundness, ellipticity, curvature of the pipe cross-section and the characteristic dimensions of irregular cross-sections; the surface defect parameters include the location, size and morphological characteristics of cracks, pits, scratches, bubbles and wrinkles; and the three-dimensional morphology parameters include the waviness, roughness and local geometric deviations of the pipe surface.

[0017] Preferably, the training method for the multi-task joint learning network model in step 4 includes: constructing a multi-parameter labeled dataset for irregularly shaped pipes, wherein the labeled dataset contains multi-view images and corresponding contour size labels, geometric shape labels, surface defect labels and three-dimensional topography labels; using the labeled dataset to perform end-to-end training of the multi-task joint learning network model; and using a transfer learning strategy, pre-training with a general pipe detection dataset and then fine-tuning with the multi-parameter labeled dataset for irregularly shaped pipes.

[0018] Preferably, the comprehensive quality assessment results in step 4 include: quantitative scores of each detection parameter, weight coefficients of each detection parameter, overall quality level, and comprehensive health score.

[0019] Preferably, the inspection and evaluation report in step 5 includes: basic information about the pipe, a comparison table of measured values ​​and standard values ​​of each inspection parameter, a visual map of defect distribution, a radar chart of scores for each parameter, and a comprehensive quality level assessment conclusion.

[0020] This invention also proposes a machine learning-based intelligent multi-parameter detection and evaluation system for irregularly shaped pipes, including:

[0021] The image acquisition module is used to perform multi-view image acquisition and transmit the acquired multi-view images to the image preprocessing and 3D construction module.

[0022] The image preprocessing and 3D construction module is used to perform image preprocessing and 3D surface model construction, and transmits the 3D surface model and the preprocessed multi-view 2D image to the multi-parameter feature extraction module.

[0023] The multi-parameter feature extraction module is used to perform multi-parameter feature extraction and transmit the extracted multi-parameter features to the intelligent evaluation module.

[0024] The intelligent evaluation module is used to perform multi-parameter comprehensive evaluation based on a multi-task joint learning network model and transmit the comprehensive quality evaluation results to the report generation module.

[0025] The report generation module is used to generate test and evaluation reports and transmit the generated reports to the control and data processing unit.

[0026] The control and data processing unit communicates with the above modules and is used to control the collaborative work of the modules and to manage the data in a unified manner.

[0027] The display and interaction module communicates with the control and data processing unit to visualize the detection and evaluation results and receive user operation commands.

[0028] Compared with the prior art, the beneficial technical effects of the present invention are as follows:

[0029] This invention uses a structured light projector to dynamically adjust the density and coding strategy of the projected stripes according to the curvature distribution of the irregular pipe surface. Different stripe densities and corresponding coding methods are used in curvature abrupt change regions, curvature gradual change regions, and curvature straight regions, so that the measurement accuracy of each region matches the stripe density, thereby improving the accuracy of three-dimensional reconstruction of the irregular pipe surface.

[0030] This invention obtains the rough outline of the pipe surface through pre-scanning and divides the curvature region by threshold, so that the structured light projector can know the curvature distribution of each region of the pipe before the formal scan, providing a decision basis for dynamically adjusting the stripe density and encoding strategy, and realizing the zoned adaptive measurement of different curvature characteristics of the irregular pipe surface.

[0031] This invention uses a curvature distortion compensation step to extract theoretical curvature information based on the design model of the irregularly shaped pipe and calculate the theoretical distortion mode of the projected stripes. It then performs reverse deformation compensation on the projected stripes, eliminating the structured light stripe distortion error introduced by the curvature of the pipe surface and improving the geometric accuracy of the three-dimensional surface model.

[0032] This invention achieves comprehensive quantitative characterization of multiple quality dimensions of irregularly shaped pipes by simultaneously extracting contour size parameters, geometric shape parameters, surface defect parameters, and three-dimensional morphology parameters from a three-dimensional surface model and preprocessed multi-view two-dimensional images, providing a multi-dimensional data foundation for comprehensive quality assessment.

[0033] This invention uses a multi-task joint learning network model to comprehensively evaluate the extracted multi-parameter features. It shares the underlying feature extraction network and outputs evaluation results for different categories through multiple task branches. This reduces computational overhead while maintaining evaluation accuracy and improves the efficiency of multi-parameter comprehensive evaluation.

[0034] This invention achieves full automation from raw image input to detection and evaluation report output through a complete process of image acquisition, preprocessing and 3D construction, multi-parameter feature extraction, intelligent evaluation and report generation, as well as a system architecture in which each module operates in coordination. This improves the efficiency and intelligence level of quality inspection of irregular-shaped pipes. Attached Figure Description

[0035] Figure 1 This is a schematic diagram of the process of the intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning provided in an embodiment of the present invention;

[0036] Figure 2 This is a schematic diagram of the process of the intelligent multi-parameter detection and evaluation system for irregularly shaped pipes based on machine learning provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0038] Please refer to Figure 1Multi-source image data acquisition involves using a multi-view imaging device to acquire surface image data of the irregularly shaped pipe under test. The multi-view imaging device includes four industrial cameras and a structured light projector. The four industrial cameras are arranged at non-uniform angular intervals along the circumference of the pipe, with angles of 0°, 75°, 165°, and 255°. The minimum angular interval between adjacent cameras is 60°, and the maximum angular interval is 120°. This non-uniform interval arrangement effectively covers areas with different curvature characteristics along the circumference of the pipe, avoiding overlapping blind spots caused by uniform arrangement. The structured light projector is positioned directly in front of the pipe, with a baseline distance of 200 mm between it and each camera. The industrial cameras are synchronously triggered to acquire images at a frame rate of 30 frames per second and an exposure time of 5000 microseconds, obtaining surface images of the pipe from different circumferential angles, resulting in four two-dimensional images from different perspectives.

[0039] The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes based on the curvature distribution of the irregularly shaped pipe surface. In regions with abrupt curvature changes, a first-density stripe and phase encoding method are used, with a first density of 5 stripes per millimeter. In regions with gradual curvature changes, a second-density stripe and Gray code encoding method are used, with a second density of 3 stripes per millimeter. In regions with flat curvature, a third-density stripe and triangulation method are used, with a third density of 1 stripe per millimeter. The first density is greater than the second density, and the second density is greater than the third density. Regions with abrupt curvature changes are those with an absolute curvature value greater than 0.05 mm⁻¹, regions with gradual curvature changes are those with an absolute curvature value between 0.02 mm⁻¹ and 0.05 mm⁻¹, and regions with flat curvature are those with an absolute curvature value less than 0.02 mm⁻¹.

[0040] The method for dividing curvature abrupt change regions, curvature gradually changing regions, and curvature flat regions is as follows: A structured light projector first projects low-density stripes to pre-scan the irregularly shaped pipe. The low-density stripes are 0.5 stripes per millimeter, acquiring coarse contour data of the pipe surface. Based on this coarse contour, the absolute value of curvature for each region is calculated. A first threshold and a second threshold are used as the dividing boundaries. The first threshold is 0.05 mm⁻¹, and the second threshold is 0.02 mm⁻¹. Regions with an absolute curvature value greater than the first threshold are defined as curvature abrupt change regions; regions with an absolute curvature value between the first and second thresholds are defined as curvature gradually changing regions; and regions with an absolute curvature value less than the second threshold are defined as curvature flat regions.

[0041] The phase encoding method employs a four-step phase shift approach, projecting four sinusoidal fringe patterns with a 90° phase difference. The phases of each pattern are 0°, 90°, 180°, and 270°, respectively. The wrapping phase of each pixel is then calculated using a four-step phase shift algorithm. The Gray code encoding method uses a seven-bit Gray code pattern, which is decoded to determine the fringe order of each pixel. The triangulation method, based on the principle of laser triangulation, calculates the spatial height value of the pixel based on the fringe offset.

[0042] The four-step phase-shifting method projects four sinusoidal fringe patterns with a phase difference of π / 2. The phases of the four patterns are 0, π / 2, π, and 3π / 2, respectively. The grayscale distribution formula for each pixel is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] in, These represent the grayscale values ​​of each pixel in the four stripe images. The non-uniform reflectivity of the pipe surface. Background light intensity, For the amplitude of the stripe modulation, The phase principal value to be solved is directly related to the height information of the pipe surface. Solving the above four equations simultaneously can eliminate... , , The formulas for calculating the principal phase values ​​of the three external factors are as follows:

[0048]

[0049] in, It is the arctangent function in the fourth quadrant, with a range of . The wrap-around phase of each pixel is calculated using the above formula, and then a phase unwrapping algorithm is used to obtain continuous absolute phase, thereby reconstructing the three-dimensional height information of the pipe surface. In this embodiment, the phase encoding method is only applied to regions with abrupt curvature changes, in conjunction with the first density stripes, to provide the high spatial resolution required for these regions.

[0050] Gray code encoding uses a seven-bit Gray code pattern. Gray code is a binary code where adjacent numbers differ by only one bit. Projecting seven Gray code stripe patterns allows for the encoding of... Each pixel column position is encoded. The black and white stripes in each Gray code pattern correspond to binary bits 1 and 0, respectively. After the camera acquires the Gray code image modulated on the pipe surface, each pixel in the image is binarized. The grayscale value of each pixel is converted into a binary sequence, then restored to a decimal number to determine the projector column coordinates corresponding to that pixel, i.e., the stripe level. In this embodiment, Gray code encoding is applied to regions with gradually changing curvature, in conjunction with second density stripes.

[0051] The triangulation method is based on the principle of laser triangulation. A structured light projector projects a light stripe onto the pipe surface. The light stripe deforms along the contours of the pipe surface. A camera captures images of the deformed light stripe, and the three-dimensional coordinates of the measured point are determined through geometric relationships. Based on the principle of similar triangles, the height of a point on the pipe surface is calculated. Offset of the light stripe on the imaging plane The calculation formula between them is as follows:

[0052]

[0053] in, The baseline distance from the optical center of the structured light projector to the optical center of the industrial camera is set to 200 mm. The distance from the optical center of the structured light projector to the reference plane is set to 500 mm. The angle between the optical axis of the structured light projector and the optical axis of the industrial camera is set to 30°. This represents the offset of the projected stripes at each pixel, expressed in pixels. By measuring the stripe offset at each pixel, the spatial height value of that point can be calculated. In this embodiment, the triangulation method is applied to regions with flat curvature, and combined with third-density stripes, processing efficiency is improved while ensuring measurement accuracy.

[0054] Image preprocessing and 3D surface model construction: The acquired multi-view images are preprocessed to obtain preprocessed multi-view 2D images. Preprocessing includes image denoising, contrast enhancement, distortion correction, and epipolar correction.

[0055] Image denoising employs a median filtering algorithm with a neighborhood window size of 5×5 pixels, effectively removing noise while preserving image edge information. Contrast enhancement utilizes an adaptive histogram equalization method, dividing the image into 8×8 local regions and performing histogram equalization on each region, increasing the contrast gain to 1.5 to 2 times that of the original image, thus improving local contrast. Distortion correction corrects lens distortion based on camera calibration parameters, restoring the image's true geometric relationships. Epipolar correction employs a fundamental matrix-based method, correcting multi-view images to a parallel epipolar state, compressing the stereo matching search range from a two-dimensional plane to a one-dimensional horizontal scan line.

[0056] The 3D surface model construction includes a curvature distortion compensation step. Theoretical curvature information is extracted from the design model of the irregularly shaped pipe. The design model is a CAD 3D model of the pipe, containing theoretical curvature data for each region. The theoretical distortion mode of the projected fringes under each curvature is calculated; that is, based on the geometric relationship of the projection system, the deformation shape of the fringe pattern should be simulated under a given curvature. Reverse deformation compensation is performed on the projected fringes; that is, the fringe pattern is pre-deformed at the projection end according to the inverse mode of theoretical distortion, so that the fringe pattern projected onto the pipe surface is consistent with the theoretical fringe pattern of the design model, thereby eliminating measurement errors introduced by surface curvature.

[0057] Based on compensated multi-view 2D images, a multi-view stereo matching algorithm is used to reconstruct the 3D surface model of the pipe. Using the 0° view image as a reference image, the corresponding pixels in the other three view images are searched along the epipolar direction. A matching cost is constructed based on the gray-level similarity of the pixel neighborhood. After cost aggregation and disparity optimization, a dense disparity map is obtained. According to camera parameters and disparity values, the 3D spatial coordinates of each pixel are calculated using triangulation principles, generating 3D point cloud data with a density of approximately 200 points per square millimeter. Outliers in the 3D point cloud data are removed by statistical filtering with a filtering threshold equal to one standard deviation of the average distance. A Poisson surface reconstruction algorithm is then used to generate a triangular mesh model, which is the final 3D surface model.

[0058] Multi-parameter feature extraction. Based on the obtained 3D surface model and preprocessed multi-view 2D images, multiple detection parameter features of the irregularly shaped pipe are extracted. These multiple detection parameters include contour size parameters, geometric shape parameters, surface defect parameters, and 3D morphology parameters.

[0059] The profile dimensions include the pipe's outer diameter, inner diameter, wall thickness, and their uniformity of distribution. A cross-sectional profile is extracted every 5 mm along the pipe's axis, for a total of 600 profiles. Ellipse fitting is performed on each profile to obtain the outer and inner diameter values; the wall thickness is half the difference between the outer and inner diameters. The wall thickness data for all profiles are statistically analyzed. The maximum wall thickness is 8.12 mm, the minimum is 7.88 mm, the average is 8.02 mm, and the standard deviation is 0.05 mm.

[0060] Geometric parameters include the roundness, ellipticity, curvature, and characteristic dimensions of the pipe cross-section. Ellipticity is calculated by fitting an ellipse to each cross-sectional profile; ellipticity is the ratio of the major axis to the minor axis. Curvature is the maximum distance the pipe axis deviates from the ideal straight line. This is calculated by extracting the coordinates of the center points of each cross-section and fitting a straight line to calculate the maximum deviation of each center point from the fitted axis.

[0061] Surface defect parameters include the location, size, and morphological features of cracks, pits, scratches, bubbles, and wrinkles. An improved Canny edge detection algorithm is used to extract defect edges from preprocessed multi-view 2D images. The Canny algorithm has a low threshold of 50 and a high threshold of 150. Morphological closing operations are used to connect fracture edges and fill closed regions. The structuring element of the closing operation is a 5×5 pixel circular structuring element to obtain the connected regions of each defect. Combined with the depth information of the 3D surface model, pseudo-defect regions caused by illumination variations are eliminated to determine the true location, size, and depth of each defect. Defects larger than 0.5 mm are marked and classified.

[0062] The three-dimensional topography parameters include the waviness, roughness, and local geometric deviations of the pipe surface. The three-dimensional surface model is registered with the pipe design model using an iterative nearest-point algorithm to calculate the deviation of each point relative to the design model and generate a deviation distribution map. Surface waviness is separated from the roughness component using Gaussian filtering with a cutoff wavelength of 0.8 mm, extracting the amplitude and wavelength characteristics of the waviness. Roughness is calculated by applying a high-pass filter to the three-dimensional surface model and then calculating the arithmetic mean deviation of the surface profile with a sampling length of 0.25 mm.

[0063] Multi-parameter intelligent evaluation based on machine learning. The extracted multi-parameter features are input into a trained multi-task joint learning network model, which outputs a comprehensive quality evaluation result for irregularly shaped pipes.

[0064] The multi-task joint learning network model employs an encoder-decoder architecture. The encoder is a shared low-level feature extraction network using a ResNet-50 structure, containing 5 convolutional blocks with a total of approximately 23.5 million parameters. It is used to extract shared high-level semantic features from the input multi-parameter features, outputting 2048-dimensional features. The decoder contains four independent task branches, corresponding to contour size evaluation, geometric shape evaluation, surface defect evaluation, and 3D morphology evaluation. Each task branch consists of three fully connected layers, with 512, 256, and 1 neurons per layer, respectively. After inputting the shared features, each layer outputs the evaluation result for its corresponding category.

[0065] The training method for the multi-task joint learning network model is as follows: First, a multi-parameter labeled dataset for irregularly shaped pipes is constructed. This dataset contains 5000 sets of multi-view images and corresponding labels for contour dimensions, geometric shapes, surface defects, and 3D topography. The labeled data is obtained through actual measurements of pipe samples using a coordinate measuring machine and a high-precision profilometer; the measured values ​​serve as the true values ​​for each parameter label. Then, the multi-task joint learning network model is trained end-to-end using the labeled dataset. The multi-task joint loss function is a weighted sum of the losses of each task branch, with each task branch using mean squared error loss and a weight of 0.25. During training, a transfer learning strategy is employed for optimization. The model is first pre-trained on the ImageNet dataset to learn general image feature extraction capabilities, and then fine-tuned using the multi-parameter labeled dataset for irregularly shaped pipes to focus on the specific detection features of irregularly shaped pipes. The pre-training process consisted of 200 training epochs, the fine-tuning process consisted of 50 training epochs, the batch size was 32, the initial learning rate was 0.001, and the learning rate was reduced to 0.9 times the original value every 20 epochs.

[0066] The comprehensive quality assessment results include quantitative scores for each testing parameter, weighting coefficients for each parameter, overall quality level, and comprehensive health score. The quantitative score for each testing parameter is obtained by normalizing the raw output values ​​of each task branch and mapping them to a score range of 0 to 100. The weighting coefficients for each testing parameter are pre-set based on the application scenario and design requirements of the pipe material, with a sum of 1. The overall quality level is determined by the weighted sum of the scores for each parameter, and is divided into four levels: A, B, C, and D. Level A indicates a comprehensive score of no less than 90 points; Level B indicates a comprehensive score of no less than 80 points and less than 90 points; Level C indicates a comprehensive score of no less than 70 points and less than 80 points; and Level D indicates a comprehensive score less than 70 points. The comprehensive health score is a comprehensive score obtained by weighted summing the scores for each parameter and mapping them to a score range of 0 to 100.

[0067] Inspection and evaluation report generation. Based on the comprehensive quality assessment results, an inspection and evaluation report for the irregular-shaped pipe is generated. The report includes basic pipe information, a comparison table of measured values ​​and standard values ​​for each inspection parameter, a visualization map of defect distribution, a radar chart of each parameter's score, and a comprehensive quality level assessment conclusion. Basic pipe information includes pipe number, specifications, material, production batch, and inspection date. The comparison table of measured values ​​and standard values ​​for each inspection parameter lists the measured value, standard value, deviation value, and deviation percentage for each parameter in tabular form. The visualization map of defect distribution uses a 3D surface model as a base map, marking the location of various defects with different colors on the model. The radar chart of each parameter's score visually presents the scoring status of each inspection parameter in radar chart form. The comprehensive quality level assessment conclusion includes the quality level determination result and subsequent processing recommendations.

[0068] Example 2

[0069] This embodiment provides a machine learning-based multi-parameter intelligent detection and evaluation method for irregularly shaped pipes, used for quality inspection of square stainless steel pipes with a length of 2000 mm, a cross-sectional size of 100 mm × 100 mm, and a wall thickness of 6 mm.

[0070] The multi-view imaging device includes four industrial cameras and a structured light projector. The four industrial cameras are arranged at non-uniform angular intervals along the circumference of the tube, with angles of 0°, 90°, 180°, and 270°. The four cameras are located directly opposite the four sides of the square tube, with adjacent cameras spaced at a uniform 90° interval. The baseline distance between the structured light projector and each camera is 250 mm. Each camera is synchronously triggered to acquire surface images of the tube from different circumferential angles, thus obtaining multi-view images.

[0071] The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes based on the surface curvature distribution of the irregularly shaped tube. The edges of the square tube exhibit significant curvature abrupt changes, belonging to the curvature abrupt change region. A first-density stripe method and phase encoding are used, with a first density of 6 stripes per millimeter. The transition region between the edges and the plane is a curvature gradually changing region, employing a second-density stripe method and Gray code encoding, with a second density of 3 stripes per millimeter. The planar region has near-zero curvature, belonging to the curvature flat region, employing a third-density stripe method and triangulation, with a third density of 1 stripe per millimeter.

[0072] The method for dividing the curvature abrupt change region, the curvature gradual change region, and the curvature flat region is as follows: The structured light projector first projects low-density stripes onto the square tube for pre-scanning. The low-density stripes are 0.5 stripes per millimeter, acquiring coarse contour data of the tube surface, and calculating the absolute value of curvature for each region. A first threshold and a second threshold are used as the dividing boundaries. The first threshold is 0.08 mm⁻¹, and the second threshold is 0.03 mm⁻¹. Regions with an absolute curvature value greater than 0.08 mm⁻¹ are defined as curvature abrupt change regions, i.e., edges and their adjacent areas; regions with an absolute curvature value between 0.03 mm⁻¹ and 0.08 mm⁻¹ are defined as curvature gradual change regions, i.e., the transition region between edges and planes; regions with an absolute curvature value less than 0.03 mm⁻¹ are defined as curvature flat regions, i.e., the planar regions of each side surface.

[0073] The acquired multi-view images undergo preprocessing, including image denoising, contrast enhancement, distortion correction, and epipolar correction. Image denoising employs a median filtering algorithm with a neighborhood window size of 3×3 pixels. Contrast enhancement utilizes an adaptive histogram equalization method. Distortion correction corrects lens distortion based on camera calibration parameters. Epipolar correction restores the multi-view images to a state of epipolar parallelism.

[0074] The 3D surface model construction includes a curvature distortion compensation step. Theoretical curvature information is extracted from the design model of the square tube. The curvature at the edges of the square tube is theoretically known and can be directly used as a reference parameter. The theoretical distortion mode of the projected fringes under each curvature is calculated. Reverse deformation compensation is performed on the projected fringes to ensure that the fringe pattern projected onto the surface of the square tube matches the theoretical fringe pattern of the design model. Based on the compensated multi-view 2D images, a multi-view stereo matching algorithm is used to reconstruct the 3D surface model of the square tube, with a point cloud density of approximately 180 points per square millimeter.

[0075] Based on a 3D surface model and preprocessed multi-view 2D images, multiple detection parameter features of square pipes are extracted. Contour dimension parameters include the pipe's side length, wall thickness, and its uniformity. A cross-sectional profile is extracted every 5 mm along the pipe's axis, for a total of 400 profiles. The side length and wall thickness of each profile are calculated, and the uniformity of wall thickness distribution is statistically analyzed. Geometric parameters include the pipe's cross-sectional right-angle deviation, flatness, and curvature. Right-angle deviation is the difference between the measured angle between adjacent planes at each edge and 90°; flatness is the smoothness of each side surface, assessed by calculating the maximum distance from each point on the plane to the least-squares fitted plane. Curvature is the maximum distance the pipe's axis deviates from the ideal straight line. Surface defect parameters include the location, size, and morphological characteristics of cracks, pits, scratches, bubbles, and wrinkles. 3D morphology parameters include the pipe's surface waviness, roughness, and local geometric deviations.

[0076] The extracted multi-parameter features are input into a trained multi-task joint learning network model, which outputs a comprehensive quality assessment result for the square pipe. The comprehensive quality assessment result includes the quantitative score of each detection parameter, the weight coefficient of each detection parameter, the overall quality level, and the comprehensive health score.

[0077] Based on the comprehensive quality assessment results, a testing and evaluation report for the square pipes is generated. The report includes basic information about the pipes, a comparison table of measured values ​​and standard values ​​for each testing parameter, a visualization map of defect distribution, a radar chart of scores for each parameter, and a comprehensive quality level assessment conclusion.

[0078] Example 3

[0079] This embodiment provides a machine learning-based intelligent multi-parameter detection and evaluation method for irregularly shaped pipes, used for quality inspection of corrugated pipes with a length of 1500 mm, a corrugation height of 25 mm, a corrugation pitch of 50 mm, a wall thickness of 4 mm, and 30 corrugations.

[0080] The multi-view imaging device comprises six industrial cameras and two structured light projectors. The six industrial cameras are arranged at non-uniform angular intervals along the circumference of the pipe, with angles of 0°, 50°, 110°, 170°, 230°, and 290°. The minimum angular interval between adjacent cameras is 50°, and the maximum angular interval is 70°. The two structured light projectors are positioned directly in front of and behind the pipe, respectively, with a baseline distance of 180 mm from each camera. The two structured light projectors project stripes alternately at a frequency of 15 Hz to eliminate occlusion shadows caused by projection from a single direction. Each camera synchronously triggers acquisition to obtain surface images of the pipe from different circumferential angles, thus obtaining multi-view images.

[0081] The corrugated pipe surface has a periodic corrugated structure. The curvature changes drastically at the crests and troughs, forming abrupt curvature changes, with absolute curvature values ​​ranging from approximately 0.15 mm⁻¹ to 0.20 mm⁻¹. The curvature changes more gently in the wave-side regions, forming a gradually changing curvature region, with absolute curvature values ​​ranging from approximately 0.04 mm⁻¹ to 0.08 mm⁻¹. The curvature of the connecting areas between corrugations is close to zero, forming a flat curvature region. The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes based on the above curvature distribution. The crest and trough regions use a first-density stripe and phase encoding method, with a first density of 8 stripes per millimeter. The wave-side regions use a second-density stripe and Gray code encoding method, with a second density of 4 stripes per millimeter. The flat regions use a third-density stripe and triangulation method, with a third density of 1 stripe per millimeter.

[0082] The method for dividing the curvature abrupt change region, the curvature gradually changing region, and the curvature flat region is as follows: A structured light projector first projects low-density stripes to pre-scan the corrugated pipe. The low-density stripes are 0.5 stripes per millimeter, acquiring coarse contour data of the pipe surface, and calculating the absolute value of curvature for each region. A first threshold and a second threshold are used as the dividing boundaries. The first threshold is 0.10 mm⁻¹, and the second threshold is 0.03 mm⁻¹. Regions with an absolute curvature value greater than 0.10 mm⁻¹ are defined as curvature abrupt change regions, i.e., crest and trough regions. Regions with an absolute curvature value between 0.03 mm⁻¹ and 0.10 mm⁻¹ are defined as curvature gradually changing regions, i.e., waveside regions. Regions with an absolute curvature value less than 0.03 mm⁻¹ are defined as curvature flat regions, i.e., the connection regions between corrugations.

[0083] The acquired multi-view images underwent preprocessing, including image denoising, contrast enhancement, distortion correction, and epipolar correction. Image denoising employed a non-local mean filtering algorithm with a search window size of 21×21 pixels, smoothing noise while better preserving the subtle textures of the corrugated pipe surface. Contrast enhancement utilized an adaptive histogram equalization method. Distortion correction addressed lens distortion based on camera calibration parameters. Epipolar correction brought the multi-view images to a state of epipolar parallelism.

[0084] The 3D surface model construction includes a curvature distortion compensation step. Theoretical curvature information is extracted from the bellows design model. The bellows' crests, troughs, and sides exhibit periodically repeating curvature characteristics. The theoretical distortion patterns of the projected fringes under each curvature are calculated. Reverse deformation compensation is performed on the projected fringes to ensure that the fringe pattern projected onto the bellows surface matches the theoretical fringe pattern of the design model. Based on the compensated multi-view 2D images, a multi-view stereo matching algorithm is used to reconstruct the 3D surface model of the bellows, achieving a point cloud density of approximately 220 points per square millimeter.

[0085] Based on a 3D surface model and preprocessed multi-view 2D images, multiple detection parameter features of the corrugated pipe are extracted. Profile dimensional parameters include the pipe's corrugation height, pitch, wall thickness, and uniformity of distribution. The positions of the crests and troughs of each corrugation are extracted along the pipe's axis, and the corrugation height and pitch of each corrugation are calculated. The maximum deviation of the corrugation height and the coefficient of variation of the pitch are statistically analyzed. Geometric parameters include the corrugation symmetry, crest fillet radius, and trough fillet radius. Surface defect parameters include the location, size, and morphological characteristics of cracks, pits, scratches, bubbles, and wrinkles. 3D morphology parameters include the corrugation degree, roughness, and local geometric deviations of the pipe surface.

[0086] The extracted multi-parameter features are input into a trained multi-task joint learning network model, which outputs a comprehensive quality assessment result for the bellows. The comprehensive quality assessment result includes the quantitative score of each detection parameter, the weight coefficient of each detection parameter, the overall quality level, and the comprehensive health score.

[0087] Based on the comprehensive quality assessment results, a test and evaluation report for the corrugated pipe is generated. The report includes basic information about the pipe, a comparison table of measured values ​​and standard values ​​for each test parameter, a visualization map of defect distribution, a radar chart of scores for each parameter, and a comprehensive quality level assessment conclusion.

[0088] Example 4

[0089] Please refer to Figure 2 This embodiment provides a machine learning-based multi-parameter intelligent detection and evaluation system for irregularly shaped pipes. The system includes an image acquisition module, an image preprocessing and 3D construction module, a multi-parameter feature extraction module, an intelligent evaluation module, a report generation module, a control and data processing unit, and a display and interaction module.

[0090] The image acquisition module is used to perform multi-view image acquisition. It includes four industrial cameras arranged at non-uniform angular intervals along the circumference of the pipe, a structured light projector, and a rotating platform. Each of the four industrial cameras has a resolution of 1920×1080 pixels and a frame rate of 30 frames per second. The structured light projector has a projection power of 5 watts and a projection wavelength of 450 nanometers. The rotating platform has a rotation angle range of 0° to 360° and an adjustable rotation speed range of 1° / second to 30° / second. The industrial cameras are used to capture multi-view images of the pipe surface, the structured light projector projects structured light stripes onto the pipe surface, and the rotating platform supports and rotates the pipe to achieve image acquisition from different angles. The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes according to the curvature distribution of the pipe surface. The image acquisition module is connected to the image preprocessing and 3D construction module via Gigabit Ethernet, transmitting the acquired multi-view images to the image preprocessing and 3D construction module.

[0091] The image preprocessing and 3D construction module performs image preprocessing and 3D surface model construction. It includes a preprocessing unit, a curvature distortion compensation unit, and a 3D construction unit. The preprocessing unit performs denoising, contrast enhancement, distortion correction, and epipolar correction on the acquired multi-view images to obtain preprocessed multi-view 2D images. The curvature distortion compensation unit performs reverse deformation compensation on the projected fringes based on the theoretical curvature information of the pipe design model. The 3D construction unit constructs a 3D surface model of the pipe based on the compensated multi-view 2D images. The 3D construction unit runs a multi-view stereo matching algorithm, reconstructing a point cloud density of approximately 200 points per square millimeter. After model construction, it is automatically saved as an STL format file. The image preprocessing and 3D construction module communicates with the multi-parameter feature extraction module, transmitting the 3D surface model and the preprocessed multi-view 2D images to the multi-parameter feature extraction module.

[0092] The multi-parameter feature extraction module is used to perform multi-parameter feature extraction. This module includes a contour size extraction unit, a geometry extraction unit, a surface defect extraction unit, and a 3D morphology extraction unit. The contour size extraction unit extracts the outer diameter, inner diameter, wall thickness, and their uniformity of distribution from the 3D surface model, with a sampling interval of 5 mm. The geometry extraction unit extracts the roundness, ellipticity, curvature, and characteristic dimensions of irregular cross-sections from the 3D surface model. The surface defect extraction unit extracts the location, size, and morphological features of cracks, pits, scratches, bubbles, and wrinkles from the 3D surface model and the preprocessed 2D image, with a minimum detectable defect size of 0.3 mm. The 3D morphology extraction unit extracts the waviness, roughness, and local geometric deviations of the pipe surface from the 3D surface model. The multi-parameter feature extraction module communicates with the intelligent evaluation module, transmitting the extracted multi-parameter features to the intelligent evaluation module.

[0093] The intelligent evaluation module performs multi-parameter comprehensive evaluation based on a multi-task joint learning network model. The module includes a feature fusion unit and an inference unit. The feature fusion unit concatenates and fuses the features of each parameter to form a joint feature vector, which has a dimension of 1024. The inference unit stores the trained multi-task joint learning network model and, using the joint feature vector as input, outputs the comprehensive quality evaluation result of the irregularly shaped pipe. The comprehensive quality evaluation result includes the quantitative score of each detection parameter, the weight coefficient of each detection parameter, the overall quality level, and the comprehensive health score. The intelligent evaluation module communicates with the report generation module, transmitting the comprehensive quality evaluation result to the report generation module.

[0094] The report generation module is used to generate testing and evaluation reports. It includes a report template storage unit and a report assembly unit. The report template storage unit stores preset report templates, which define the report's structure and content framework. The report assembly unit retrieves the comprehensive quality evaluation results from the intelligent evaluation module, fills in the corresponding positions according to the report template structure, and generates a complete testing and evaluation report. The report generation module communicates with the control and data processing unit, transmitting the generated report to it. The generated report supports exporting to PDF and Excel formats for easy archiving and printing.

[0095] The control and data processing unit is connected to the image acquisition module, image preprocessing and 3D construction module, multi-parameter feature extraction module, intelligent evaluation module, and report generation module via Ethernet communication. This connection controls the collaborative operation of each module and provides unified data management. The control and data processing unit utilizes an industrial-grade computer equipped with an Intel Xeon processor with a clock speed of 3.2GHz, 32GB of memory, and a 512GB solid-state drive.

[0096] In summary, this invention dynamically adjusts the density and encoding strategy of projected fringes based on the curvature distribution of the pipe surface using a structured light projector. Different fringes density and corresponding encoding methods are employed in curvature abrupt change regions, gradually changing curvature regions, and flat curvature regions, improving the accuracy of 3D reconstruction of irregularly shaped pipes. By pre-scanning to obtain the rough contour of the pipe surface and dividing the curvature regions using thresholds, zonal measurement of different curvature characteristics of the pipe surface is achieved. Through a curvature distortion compensation step, theoretical curvature information is extracted based on the design model, and the theoretical distortion mode of the projected fringes is calculated. Reverse deformation compensation is performed on the projected fringes, eliminating the structured light fringes distortion error introduced by the pipe surface curvature. By simultaneously extracting contour size parameters, geometric shape parameters, surface defect parameters, and 3D morphology parameters from the 3D surface model and preprocessed multi-view 2D images, simultaneous extraction of multiple parameters of irregularly shaped pipes is achieved. A multi-task joint learning network model is used to comprehensively evaluate the extracted multi-parameter features. The underlying feature extraction network is shared, and different categories of evaluation results are output through multiple task branches, improving the efficiency of multi-parameter comprehensive evaluation while maintaining evaluation accuracy. This invention achieves full automation from raw image input to inspection and evaluation report output through the coordinated operation of the aforementioned processes and system modules. This invention is applicable to the quality inspection of irregularly shaped pipes in aerospace, automotive manufacturing, petrochemical, and nuclear power equipment industries, and has promising application prospects.

[0097] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A machine learning-based intelligent multi-parameter detection and evaluation method for irregularly shaped pipes, characterized in that, Includes the following steps: Step 1: Multi-source image data acquisition: Surface image data of the irregularly shaped pipe under test is acquired through a multi-view imaging device to obtain multi-view images. The multi-view imaging device includes at least four industrial cameras arranged at different angles along the circumference of the pipe and at least one structured light projector. The structured light projector dynamically adjusts the density and encoding strategy of the projected stripes according to the curvature distribution of the irregularly shaped pipe surface: a first density stripe and phase encoding method are used in the curvature abrupt change region, a second density stripe and Gray code encoding method are used in the curvature gradually changing region, and a third density stripe and triangulation method are used in the curvature flat region. The first density is greater than the second density, and the second density is greater than the third density. Step 2: Constructing a 3D surface model after image preprocessing: Preprocess the multi-view images acquired in Step 1 to obtain preprocessed multi-view 2D images, and construct a 3D surface model of the irregular pipe based on the multi-view 2D images; construct a model including curvature distortion compensation steps, extract theoretical curvature information based on the design model of the irregular pipe, calculate the theoretical distortion mode of the projected fringes under each curvature, and perform reverse deformation compensation on the projected fringes. Step 3: Multi-parameter feature extraction: Based on the three-dimensional surface model obtained in Step 2 and the preprocessed multi-view two-dimensional images, extract multiple detection parameter features of the irregularly shaped pipe. The multiple detection parameters include contour size parameters, geometric shape parameters, surface defect parameters and three-dimensional morphology parameters. Step 4: Multi-parameter intelligent evaluation based on machine learning: Input the multi-parameter features extracted in step 3 into the trained multi-task joint learning network model, and output the comprehensive quality evaluation results of the irregular-shaped pipe. Step 5: Generation of Inspection and Evaluation Report: Based on the comprehensive quality evaluation results of Step 4, an inspection and evaluation report for the irregular-shaped pipe is generated.

2. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The imaging method of the multi-view imaging device in step 1 is as follows: at least four industrial cameras are arranged at non-uniform angle intervals along the circumference of the irregular pipe, and a preset baseline distance is maintained between the structured light projector and each industrial camera. Each industrial camera is triggered to acquire surface images of the pipe from different circumferential angles.

3. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The method for dividing the curvature abrupt change region, curvature gradual change region, and curvature flat region in step 1 is as follows: a low-density stripe is projected by a structured light projector to pre-scan the irregularly shaped pipe, obtain the coarse contour data of the pipe surface, and calculate the absolute value of curvature of each region; using a first threshold and a second threshold as the dividing boundary, the region with the absolute value of curvature greater than the first threshold is the curvature abrupt change region, the region with the absolute value of curvature between the first threshold and the second threshold is the curvature gradual change region, and the region with the absolute value of curvature less than the second threshold is the curvature flat region, and the first threshold is greater than the second threshold.

4. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The preprocessing in step 2 includes image denoising, contrast enhancement, distortion correction, and epipolar correction.

5. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The contour dimension parameters in step 3 include the outer diameter, inner diameter, wall thickness and uniformity of distribution of the pipe. Geometric parameters include the roundness, ellipticity, curvature of the pipe cross-section, and the characteristic dimensions of irregular cross-sections; Surface defect parameters include the location, size, and morphological characteristics of cracks, pits, scratches, bubbles, and wrinkles; The three-dimensional morphology parameters include the waviness, roughness, and local geometric deviations of the pipe surface.

6. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The training method for the multi-task joint learning network model in step 4 includes: A multi-parameter labeled dataset for irregularly shaped pipes was constructed. The labeled dataset includes multi-view images and corresponding contour size labels, geometric shape labels, surface defect labels and three-dimensional topography labels. The multi-task joint learning network model was trained end-to-end using a labeled dataset. A transfer learning strategy was adopted, which involved pre-training on a general pipe inspection dataset and then fine-tuning on a multi-parameter labeled dataset of irregular pipes.

7. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The comprehensive quality assessment results in step 4 include: the quantitative score of each detection parameter, the weight coefficient of each detection parameter, the overall quality level, and the comprehensive health score.

8. The intelligent multi-parameter detection and evaluation method for irregularly shaped pipes based on machine learning according to claim 1, characterized in that: The inspection and evaluation report in step 5 includes: basic information about the pipe, a comparison table of measured values ​​and standard values ​​of each inspection parameter, a visual map of defect distribution, a radar chart of scores for each parameter, and a comprehensive quality level assessment conclusion.

9. A machine learning-based intelligent multi-parameter detection and evaluation system for irregularly shaped pipes, used to implement the method as described in any one of claims 1 to 8, characterized in that: include: The image acquisition module is used to perform multi-view image acquisition and transmit the acquired multi-view images to the image preprocessing and 3D construction module. The image preprocessing and 3D construction module is used to perform image preprocessing and 3D surface model construction, and transmits the 3D surface model and the preprocessed multi-view 2D image to the multi-parameter feature extraction module. The multi-parameter feature extraction module is used to perform multi-parameter feature extraction and transmit the extracted multi-parameter features to the intelligent evaluation module. The intelligent evaluation module is used to perform multi-parameter comprehensive evaluation based on a multi-task joint learning network model and transmit the comprehensive quality evaluation results to the report generation module. The report generation module is used to generate test and evaluation reports and transmit the generated reports to the control and data processing unit. The control and data processing unit communicates with the above modules and is used to control the coordinated work of the modules and to manage the data in a unified manner. The display and interaction module communicates with the control and data processing unit to visualize the detection and evaluation results and receive user operation commands.

Images