Intelligent Detection Method Based on Nondestructive Testing Equipment

Abstract

The present invention discloses an intelligent detection method based on a nondestructive testing device, which relates to the technical field of nondestructive testing. The method includes: establishing a twin digital model. The internal image of the component to be tested is collected through an endoscope to obtain the original surface image. Multidimensional feature extraction is performed on the original surface image, and the original surface image is screened to obtain a standard surface image. The motion parameter sequence of the endoscope is obtained, and the standard surface image is fitted to the twin digital model based on the motion parameter sequence to generate a three-dimensional restored image. Based on the shape characteristics of the component model to be tested, the grid of the three-dimensional restored image is adaptively divided to obtain a meshed three-dimensional image. Combining with a data dimensionality reduction method, a multi-channel classification detector is constructed to perform intelligent flaw detection on the meshed three-dimensional image. The multi-channel classification detector includes multiple flaw detection channels for performing different types of flaw detection tasks. Furthermore, the technical effects of reducing the detection cost and decreasing the operation difficulty are achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of nondestructive testing, and particularly to an intelligent detection method based on a nondestructive testing device. Background Art

[0002] Nondestructive Testing (NDT) technology refers to a method for detecting and evaluating internal and surface defects or properties of an object without damaging or affecting its service performance and appearance. Existing nondestructive testing technologies include ultrasonic testing, radiographic testing, magnetic particle testing, eddy current testing, etc. Although these technical methods can effectively detect internal defects of materials, they usually require professional operators and the detection process is complex. There are technical problems such as high detection cost and great operation difficulty. Summary of the Invention

[0003] The present invention provides an intelligent detection method based on a nondestructive testing device to solve the technical problems of high detection cost and great operation difficulty in the prior art, and achieve the technical effects of reducing the detection cost and decreasing the operation difficulty.

[0004] In a first aspect, the present invention provides an intelligent detection method based on a nondestructive testing device, wherein the method includes:

[0005] Establish a twin digital model, including a model of the component to be tested and an endoscope model.

[0006] Collect internal images of the component to be tested through an endoscope to obtain original surface images.

[0007] Perform multi-dimensional feature extraction on the original surface images to obtain a multi-dimensional feature set, and perform image screening on the original surface images based on the multi-dimensional feature set to obtain standard surface images.

[0008] An interactive control component obtains a sequence of motion parameters of the endoscope optical component, and fits the standard surface images to the twin digital model based on the sequence of motion parameters to obtain three-dimensional restored images.

[0009] Based on the shape characteristics of the model of the component to be tested, perform adaptive mesh division on the three-dimensional restored images to obtain meshed three-dimensional images.

[0010] Combined with a data dimensionality reduction method, construct a multi-channel classification detector to perform intelligent flaw detection on the meshed three-dimensional images, wherein the multi-channel classification detector includes multiple flaw detection channels for performing different types of flaw detection tasks.

[0011] In a feasible implementation manner, collecting internal images of the component to be tested through an endoscope to obtain original surface images includes:

[0012] Based on the model of the component to be measured, multi-level acquisition parameters are defined; the image acquisition path of the model of the component to be measured is planned in combination with the multi-level acquisition parameters, and an internal image acquisition scheme is obtained, where the internal image acquisition scheme includes a plurality of serialized acquisition points; the endoscope is activated based on the internal image acquisition scheme for image acquisition, and an original surface image is generated.

[0013] In a feasible implementation, multi-dimensional feature extraction is performed on the original surface image to obtain a multi-dimensional feature set, and the original surface image is screened based on the multi-dimensional feature set to obtain a standard surface image, including

[0014] The original surface image is grayed out, and the gray-level co-occurrence matrix of the gray image is solved; based on the gray-level co-occurrence matrix, the gray information entropy and the first contrast are calculated; a scatter coordinate system is established, and the gray information entropy and the first contrast are used as the first coordinate and the second coordinate of the scatter coordinate system respectively to perform the scatter distribution of the original surface image; clustering analysis is performed on the scatter distribution result, and the original surface image is screened according to the clustering analysis result, and multiple original images included in the abnormal clustering cluster are removed to generate the standard surface image.

[0015] In a feasible implementation, in combination with a data dimensionality reduction method, a multi-channel classification detector is constructed, including:

[0016] The historical flaw detection records of the target environment are obtained, where the historical flaw detection records include historical surface images and associated historical damage information; the flaw detection task information is obtained by interacting with the target flaw detection scenario, and a flaw detection target set is extracted; the broad-spectrum feature data of the historical surface images are extracted, and based on the flaw detection target set, the broad-spectrum feature data are grouped and dimensionally reduced to obtain a multi-target key feature set; according to the multi-target key feature set, a plurality of classification detection channels are constructed, and supervised training is performed based on the historical flaw detection records, and the trained multiple classification detection channels are integrated to obtain a multi-channel classification detector.

[0017] In a feasible implementation, the broad-spectrum feature data of the historical surface images are extracted, and based on the flaw detection target set, the broad-spectrum feature data are grouped and dimensionally reduced to obtain a multi-target key feature set, and further includes:

[0018] Obtain a first flaw detection target based on the flaw detection target set, where the first flaw detection target corresponds to any one type of damage; according to the first flaw detection target, traverse the historical flaw detection records to extract the first historical flaw detection record, where the first historical flaw detection record includes a first historical surface image and first historical damage information; solve the gray level co-occurrence matrix, gray level difference, and color moment of the first historical surface image, and perform feature extraction to obtain first broad-spectrum feature data, where the first broad-spectrum feature data at least includes gray information entropy, first contrast, energy, differential entropy, second contrast, average value, first-order moment, and second-order moment; perform dimensionality reduction analysis on the first broad-spectrum feature data, extract the main features, and store them as the first target key feature set; traverse the flaw detection target set to obtain a multi-target key feature set.

[0019] In a feasible implementation manner, according to the multi-target key feature set, construct multiple classification detection channels, and perform supervised training based on the historical flaw detection records, and integrate the trained multiple classification detection channels to obtain a multi-channel classification detector. It further includes:

[0020] Based on the first target key feature set, construct a first classification detection model according to the classification algorithm; use the first historical flaw detection record as the training data set to perform supervised training on the first classification detection model to obtain a first classification detection channel; based on the multi-target key feature set and the historical flaw detection records, construct and train multiple classification detection channels; obtain the division constraints of the adaptive grid division, and based on the division constraints, configure an input allocator, connect multiple classification detection channels to the input allocator, and construct the multi-channel classification detector.

[0021] In a feasible implementation manner, for the intelligent flaw detection of the meshed three-dimensional image, it further includes:

[0022] Obtain the flaw detection result, parse to obtain multiple edge damages, and output them as an edge damage set; based on the damage type, damage location, damage direction, and damage size, perform continuity analysis on the edge damage set to obtain a continuity analysis result, where the continuity analysis result includes multiple edge damage groups; based on the continuity analysis result, perform damage merging processing on the edge damage set to obtain a merged edge damage set; update the merged edge damage set to the flaw detection result.

[0023] In a feasible implementation manner, after performing continuity analysis on the edge damage set based on the damage type, damage location, damage direction, and damage size, it further includes:

[0024] Analyze the results of the continuity analysis, discriminate the type differences of multiple edge damage groups, and extract multiple differential edge damage groups; combine multiple differential edge damage groups to obtain multiple differential damage partitions; activate the multi-channel classification detector to perform updated detection on multiple differential damage partitions, and obtain an updated damage type set; combine the updated damage type set to perform damage combination processing on the edge damage set.

[0025] The present invention discloses an intelligent detection method based on a non-destructive testing device, including: establishing a twin digital model, including a model of a component to be tested and an endoscope model; collecting internal images of the component to be tested through the endoscope to obtain original surface images; performing multi-dimensional feature extraction on the original surface images to obtain a multi-dimensional feature set, and screening the original surface images based on the multi-dimensional feature set to obtain standard surface images; the interactive control component obtains a sequence of motion parameters of the endoscope optical component, and fits the standard surface images to the twin digital model based on the sequence of motion parameters to obtain three-dimensional restored images; based on the shape characteristics of the model of the component to be tested, perform adaptive mesh division on the three-dimensional restored images to obtain gridified three-dimensional images; combine data dimensionality reduction methods to construct a multi-channel classification detector to perform intelligent flaw detection on the gridified three-dimensional images, where the multi-channel classification detector includes multiple flaw detection channels for performing different types of flaw detection tasks. The intelligent detection method based on a non-destructive testing device disclosed by the present invention solves the technical problems of high detection cost and great operation difficulty, and achieves the technical effects of reducing the detection cost and decreasing the operation difficulty. Brief Description of the Drawings

[0026] Figure 1 It is a schematic flow chart of the intelligent detection method based on a non-destructive testing device of the present invention.

[0027] Figure 2 It is a schematic flow chart of obtaining standard surface images in the intelligent detection method based on a non-destructive testing device of the present invention. Detailed Description of the Embodiment

[0028] In the technical solutions provided in the embodiments of the present invention, to solve the technical problems of high detection cost and great operation difficulty existing in the prior art, the overall idea adopted is as follows:

[0029] First, obtain the historical welding records of the carbon dioxide cylinders, where the historical welding records include welding parameter records and welding quality evaluation records. Next, according to the welding quality evaluation records, extract the welding quality coefficients for two-way division, and divide the historical welding records into a positive solution set and a negative solution set. Then, conduct an influencing factor analysis based on the positive solution set to obtain the positive key factor set, and construct a welding quality evaluation function based on the positive key factor set and the welding quality coefficients. Next, according to the welding task information, traverse the positive solution set for two-objective matching to obtain the two-objective matching result, and initialize the first population and the second population. Then, combine the welding quality evaluation function to construct a shared path, and perform collaborative optimization on the first population and the second population to iteratively obtain the optimal welding parameter combination, where the shared path is configured with shared constraints. Finally, perform welding control based on the optimal welding parameter combination.

[0030] The following will combine the accompanying drawings of the specification and specific implementation manners to elaborate on the above technical solutions in detail for a better understanding of the above technical solutions. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments of the present invention. It should be understood that the present invention is not limited to the example embodiments that are only used to explain the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts belong to the scope of protection of the present invention. In addition, it should be noted that for the sake of description, only the parts related to the present invention are shown in the drawings rather than all.

[0031] Embodiment

[0032] Figure 1 is a schematic flow chart of the intelligent detection method based on the non-destructive testing equipment of the present invention, where the method includes

[0033] Establish a twin digital model, including a model of the component to be tested and an endoscope model.

[0034] Optionally, the twin digital model includes a model of the component to be tested and an endoscope model registered in the same digital space. In other words, the coordinate systems of the model of the component to be tested and the endoscope model are aligned so that the two can be accurately matched in the same digital space. Among them, the component to be tested is an industrial production component with an inner surface and requires inner surface control.

[0035] Optionally, based on the design data of the component to be tested (such as CAD drawings, design stability, etc.), obtain the geometric shape and size of the component to be tested, and use 3D modeling software (such as SolidWorks, CATIA, etc.) to establish the geometric model of the component to be tested to obtain the model of the component to be tested.

[0036] Optionally, a geometric model of the endoscope is established based on the physical size and structure of the endoscope. The endoscope model includes parts such as a probe, an optical fiber bundle, and a lens. Then, based on the movement constraints of the movement components of the endoscope, the movement range and movement mode of the endoscope are defined, such as rotation, forward movement, backward movement, etc., to ensure that the endoscope can cover all areas of the component to be measured.

[0037] Optionally, the endoscope model also defines the optical center position, phase plane coordinates, and internal parameter matrix of the endoscope lens or image acquisition component. The optical center position determines the field of view center of the endoscope during detection. The phase plane coordinates refer to the coordinates of the endoscope lens or image acquisition component on the imaging plane, which are used to describe the position and size of the imaging picture. Through the optical center position, phase plane coordinates, and internal parameter matrix, the image mapping path of the endoscope lens can be determined in combination with the external parameter matrix, which helps to subsequently map the image to the component model to be measured. Among them, the external parameter matrix describes the position and orientation of the endoscope in space, including the rotation matrix and the translation vector.

[0038] Exemplarily, in the optical system of the endoscope, the optical center position and the internal parameter matrix are determined through optical calibration technology. The optical center refers to the center point of the endoscope lens, and the internal parameter matrix is used to describe the internal parameters of the endoscope imaging system, including the focal length, principal point coordinates (the position of the optical center on the imaging plane), and distortion coefficients, etc. Specifically, the optical calibration technology includes laser calibration, calibration plate, etc.

[0039] Exemplarily, the phase plane refers to the plane where the image sensor is located. In the endoscope, the phase plane is usually the plane of the image sensor chip. The position and orientation of the phase plane are determined through the optical design parameters of the endoscope.

[0040] Furthermore, the position and orientation of the phase plane are defined in the same three-dimensional coordinate system as the optical center to ensure the clear spatial relationship between the two.

[0041] Internal images of the component to be measured are collected by the endoscope to obtain the original surface image.

[0042] Optionally, the endoscope traverses the internal space of the component to be measured for image collection based on a preset collection route and collection parameters. The optical signal is converted into an electrical signal by the image sensor, and the electrical signal is processed by the embedded ISP unit, and the processing result is output as the original surface image.

[0043] Optionally, the original surface image is a series of local surface images with acquisition marks, and the acquisition marks include an acquisition sequence mark and an acquisition orientation mark. Among them, the acquisition sequence mark is used to mark the acquisition sequence of the images; the acquisition orientation mark is used to mark the orientation information during image acquisition, such as the angle and position of the endoscope. In other words, the acquisition orientation mark represents the external parameter matrix when the endoscope acquires images. The acquisition marks including the acquisition sequence mark and the acquisition orientation mark are helpful for the management of the original surface images and the three-dimensional restoration of subsequent images.

[0044] In some embodiments, acquiring the internal image of the component to be measured by an endoscope and obtaining the original surface image includes:

[0045] Defining multi-level acquisition parameters based on the component model to be measured.

[0046] Combining the multi-level acquisition parameters to plan the image acquisition path of the component model to be measured, and obtaining an internal image acquisition scheme, where the internal image acquisition scheme includes a plurality of serialized acquisition points.

[0047] Activating the endoscope for image acquisition based on the internal image acquisition scheme to generate the original surface image.

[0048] Optionally, for regions with different shape complexities of the component to be measured, multi-level acquisition parameters are adaptively configured. Specifically, the higher the shape complexity and the stricter the damage control requirements for a region or part, the higher the accuracy of the corresponding acquisition parameters. Exemplarily, the acquisition parameters include: a resolution parameter, that is, an appropriate resolution is set for different regions, and a higher resolution is required for high-precision regions; an acquisition angle parameter, that is, the optimal angle of the endoscope at each acquisition point is determined to ensure full coverage of the inner surface; an acquisition distance parameter, that is, the distance between the endoscope and the acquisition point is set according to the size and shape of the component to ensure image quality. A lighting parameter, including the intensity and direction of the endoscope light source, to ensure uniform lighting in different regions and improve image clarity.

[0049] Further, combining the multi-level acquisition parameters, plan the image acquisition path of the component model to be measured to obtain an internal image acquisition scheme. First, use a suitable path planning algorithm (such as the A* algorithm, Dijkstra algorithm, or genetic algorithm) to generate an image acquisition path to ensure that the endoscope can cover all inner surface regions. Then, based on the multi-level acquisition parameters, adaptively distribute acquisition points on the image acquisition path, and sort the generated acquisition points to form serialized acquisition points to ensure the optimal path. Finally, according to the path planning result, the serialized acquisition points, and the multi-level acquisition parameters, form a complete image acquisition scheme. This image acquisition scheme includes parameter settings for each acquisition point, including information such as position, angle, distance, and lighting.

[0050] Optionally, based on the internal image acquisition scheme, move the endoscope to the first acquisition point, adjust the acquisition angle and distance, then activate the endoscope camera and light source to acquire the image of the first acquisition point and obtain the original surface image. Next, move the endoscope to each acquisition point in sequence, repeating the positioning and image acquisition process until the acquisition of all points is completed, and store the acquired original surface images to obtain the original surface images.

[0051] Through the above method steps, perform reasonable path planning and multi-level parameter settings to ensure that the endoscope can fully cover the inner surface of the component to be tested, and adjust the acquisition parameters according to the detection requirements of different regions to ensure the imaging quality of high-precision regions and meet the detection requirements.

[0052] Perform multi-dimensional feature extraction on the original surface image to obtain a multi-dimensional feature set, and perform image screening on the original surface image based on the multi-dimensional feature set to obtain a standard surface image.

[0053] Optionally, during the image acquisition process, due to the complex operating environment of the endoscope, various ineffective pictures with poor acquisition effects may be generated. To ensure the accuracy and reliability of subsequent analysis, it is necessary to filter these ineffective pictures. Exemplarily, ineffective pictures with poor acquisition effects include abnormal exposure (overexposure or underexposure), glare spots, distortion (geometric distortion or geometric distortion), and blur (motion blur or defocus blur).

[0054] Optionally, the multi-dimensional feature set refers to a set of feature indicators related to the gray information of the original image. Exemplarily, it includes mean value, contrast, information entropy, energy, second-order moment, third-order moment, etc. The quality and details of the original surface image can be comprehensively described through the multi-dimensional feature set.

[0055] In some embodiments, as Figure 2 shown, performing the multi-dimensional feature extraction on the original surface image to obtain a multi-dimensional feature set, and performing image screening on the original surface image based on the multi-dimensional feature set to obtain a standard surface image includes:

[0056] Grayscale the original surface image and solve to obtain the gray-level co-occurrence matrix of the grayscale image.

[0057] Based on the gray-level co-occurrence matrix, calculate and obtain the gray information entropy and the first contrast.

[0058] Establish a scatter coordinate system, with the gray information entropy and the first contrast as the first coordinate and the second coordinate of the scatter coordinate system respectively, and perform the scatter distribution of the original surface image.

[0059] Perform clustering analysis on the scatter distribution results, and screen the original surface images according to the clustering analysis results, excluding multiple original images included in abnormal clustering clusters to generate the standard surface image.

[0060] Specifically, obtain the multi-dimensional feature set of the original surface image. First, perform grayscale processing on the original surface image to convert the original surface image that is a chromaticity image into a grayscale image. Among them, the grayscale processing methods include the average method, weighted average method, component method, summation method, etc. Exemplarily, grayscale based on the RGB model or HIS model. Among them, based on the RGB model, the grayscale value is obtained by weighted summation of the RGB three channels, which is applicable to most cases. Preferably, perform grayscale processing based on the IHS model, and obtain the grayscale value by extracting the intensity component, so as to retain the image brightness information.

[0061] Optionally, based on the grayscale image, solve the gray-level co-occurrence matrix (GLCM). The gray-level co-occurrence matrix is a matrix that describes the spatial relationship between pixel value pairs in a grayscale image and can reflect the comprehensive information of the image and gray level about the direction and the change range of the adjacent interval. Specifically, the elements in the gray-level co-occurrence matrix represent the probability that a certain gray value appears around a pixel with gray value i.

[0062] Optionally, based on statistical methods, calculate and obtain the gray information entropy and the first contrast based on the gray-level co-occurrence matrix. Among them, the gray information entropy is used to reflect the randomness and complexity of the image gray value, and the first contrast is used to measure the change degree of the image gray value. Specifically, the gray information entropy is calculated based on the following formula:

[0063] ;

[0064] Among them, represents the normalized value of the i-th row and j-th column in the gray-level co-occurrence matrix.

[0065] Specifically, the first contrast is calculated based on the following formula:

[0066] ;

[0067] Furthermore, establish a scatter coordinate system. Use the gray information entropy as the first coordinate and the contrast as the second coordinate, and represent each original surface image as a point in the scatter coordinate system. In other words, use the gray information entropy and contrast of each image as coordinates and represent them in the scatter plot. Then, use a clustering algorithm to analyze the scatter plot to identify different categories of images, and each image category corresponds to a clustering cluster. Exemplarily, through density-based DBSCAN clustering, divide the scatter points in the scatter plot into clusters of arbitrary shapes and detect noise points.

[0068] Optionally, the original images corresponding to the abnormal clustering clusters in the clustering results are removed, and the remaining images after removal are the standard surface images. Through the above method, high-quality standard surface images can be effectively screened out from the original surface images, and the invalid images that may be generated during the acquisition process are removed. This not only improves the quality of the image data, but also provides a more reliable data basis for subsequent detection and analysis.

[0069] The interaction control component obtains the motion parameter sequence of the endoscope optical component, and fits the standard surface image to the twin digital model based on the motion parameter sequence to obtain a three-dimensional restored image.

[0070] Optionally, based on the position of the determined phase plane and the optical center, the imaging image can be restored and mapped to the inner surface of the component to be measured. Specifically, first, according to the optical system parameters of the endoscope (i.e., the internal parameter matrix), the imaging image is geometrically corrected through a polynomial distortion model or other optical distortion correction algorithms to eliminate the image distortion caused by optical distortion. Then, through the interaction control component, the motion parameters during the image acquisition process of the endoscope are called to obtain the motion parameter sequence, and the initial position of the coordinate system of the endoscope optical component is initialized based on the motion parameter sequence. Among them, the motion parameter sequence includes serialized multi-group position parameters, angle parameters, etc. The position parameters define the specific coordinates of the optical center of the endoscope optical component in three-dimensional space. The angle parameters refer to the rotation angles of the coordinate system of the endoscope optical component relative to the model coordinate system (such as pitch angle, yaw angle, and roll angle).

[0071] Furthermore, based on the position of the optical center and the phase plane, the pixel coordinates in the imaging image are mapped to the inner surface of the component to be measured. This includes converting each pixel coordinate in the imaging image to the phase plane coordinate system to obtain the spatial coordinates of the pixel on the phase plane. Then, starting from the optical center, ray tracing is performed through the spatial coordinates of each pixel to determine the intersection point of the ray on the inner surface of the component to be measured. Furthermore, the spatial coordinates of the ray intersection point are converted to the inner surface coordinate system of the component to be measured to obtain the mapping position of the imaging image on the inner surface.

[0072] Optionally, the mapping results of multiple imaging images are fused to generate a complete inner surface image of the component to be measured. Exemplarily, they are superimposed and stitched according to the endoscope motion trajectory to form a complete three-dimensional inner surface image of the component to be measured. During the fusion process, image stitching, image registration and other technologies can be used to ensure seamless connection between the images.

[0073] Through the above method, the images collected are accurately mapped into the digital model using the motion parameters of the endoscope, realizing the three-dimensional precise restoration of the inner surface. The obtained three-dimensional restored image can more intuitively reflect the internal defect conditions of the component to be measured, helping to discover and handle potential problems.

[0074] Based on the shape features of the component to be measured model, perform adaptive mesh division on the three-dimensional restored image to obtain a meshed three-dimensional image.

[0075] Optionally, for the adaptive mesh division of the three-dimensional restored image, first, based on the three-dimensional model of the component to be measured, extract its shape feature information, including surface features (the curvature change of the model, used to identify the main surfaces and inflection points); boundary features (used to identify the boundary lines and boundary points of the model, especially the edges with sharp changes); angle features: detect the regions with significant angle changes in the model, such as sharp corners or edges.

[0076] Optionally, according to the overall size of the component to be measured model, perform preliminary mesh division. The initial mesh can be uniform or divided based on the rough features of the model. Then, based on the above shape features, perform adaptive adjustment on the initial mesh. The adjustment process includes: refining the mesh, in the regions with large curvature changes, significant boundary and angle features, refine the mesh division to increase the mesh density. Optimize the mesh shape, according to the feature points and feature lines, optimize the shape and size of the mesh to make the mesh better adapt to the shape features of the model.

[0077] Furthermore, map the mesh nodes to the coordinate system of the three-dimensional restored image. According to the mesh division, cut the three-dimensional restored image into multiple mesh units.

[0078] Combined with the data dimensionality reduction method, construct a multi-channel classification detector to perform intelligent flaw detection on the meshed three-dimensional image, where the multi-channel classification detector includes multiple flaw detection channels for performing different categories of flaw detection tasks.

[0079] In some embodiments, combined with the data dimensionality reduction method, constructing a multi-channel classification detector includes:

[0080] Obtain the historical flaw detection records of the target environment, where the historical flaw detection records include historical surface images and associated historical damage information.

[0081] Interact with the target flaw detection scenario to obtain flaw detection task information and extract the flaw detection target set.

[0082] Extract the broad-spectrum feature data of the historical surface images, and based on the flaw detection target set, perform grouped dimensionality reduction on the broad-spectrum feature data to obtain a multi-target key feature set.

[0083] According to the multi-target key feature set, construct multiple classification detection channels, and perform supervised training based on the historical flaw detection records, and integrate the trained multiple classification detection channels to obtain a multi-channel classification detector.

[0084] Specifically, first, collect and organize the historical flaw detection records of the target environment, including historical surface images and associated historical damage information, and clean and organize the collected historical flaw detection records to ensure data consistency and integrity. Then, interact with the target flaw detection scenario to obtain the current flaw detection task information. Exemplarily, the flaw detection task information includes the type of the component to be detected, flaw detection requirements, types of detected damage, etc. Next, set the types of detected damage in the flaw detection task information as flaw detection targets to form a flaw detection target set. Exemplarily, the types of damage include cracks, corrosion, dents, deformations, etc.

[0085] Furthermore, use image processing techniques to extract the broad-spectrum feature data of multiple historical surface images. This broad-spectrum feature data includes all the feature index items in the historical pictures of the component to be detected. Exemplarily, the broad-spectrum feature data includes texture features, color features, morphological features, gray-scale features, etc.

[0086] Optionally, according to the flaw detection target set, group the broad-spectrum feature data. Each group of feature indexes corresponds to a flaw detection target. In other words, each group of feature indexes is the discriminant feature indexes for a possible type of damage. Then, use dimensionality reduction methods such as principal component analysis (PCA) to perform dimensionality reduction on each group of feature indexes respectively, extract the key features, and output them as a multi-target key feature set. Through the above steps, the feature indexes that make an active contribution to damage recognition can be obtained, which helps to improve the efficiency of the subsequent detection and analysis process and reduce irrelevant interference features.

[0087] Optionally, construct a classification detection channel for each flaw detection target according to the multi-target key feature set. The input of the channel is the key features after dimensionality reduction. Then, use the labeled data in the historical flaw detection records to perform supervised training on each classification detection channel. During the training process, adjust the parameters of the classifier to improve the detection accuracy. Then, integrate the trained multiple classification detection channels to form a multi-channel classification detector, and further optimize and verify the integrated multi-channel classification detector to ensure its accuracy and robustness in various flaw detection tasks.

[0088] In some implementation manners, extracting the broad-spectrum feature data of the historical surface images and performing grouped dimensionality reduction on the broad-spectrum feature data based on the flaw detection target set to obtain a multi-target key feature set further includes:

[0089] Obtain a first flaw detection target based on the flaw detection target set, and the first flaw detection target corresponds to any one type of damage.

[0090] According to the first flaw detection target, traverse the historical flaw detection records to extract the first historical flaw detection record, and the first historical flaw detection record includes the first historical surface image and the first historical damage information.

[0091] Solve the gray-level co-occurrence matrix, gray-level difference, and color moments of the first historical surface image, and perform feature extraction to obtain the first broad-spectrum feature data, where the first broad-spectrum feature data at least includes gray-level information entropy, first contrast, energy, differential entropy, second contrast, average value, first-order moment, and second-order moment.

[0092] Perform dimensionality reduction analysis on the first broad-spectrum feature data, extract the main features, and store them as the first target key feature set.

[0093] Traverse the flaw detection target set to obtain the multi-target key feature set.

[0094] Specifically, first, based on the flaw detection target, obtain the corresponding historical flaw detection records, which include historical surface images and historical damage information consistent with the type of flaw detection target. Then, solve the gray-level co-occurrence matrix, gray-level difference, and color moments of the first historical surface image, and perform feature extraction to obtain the first broad-spectrum feature data.

[0095] Optionally, the first broad-spectrum feature data at least includes gray-level information entropy, first contrast, energy, differential entropy, second contrast, average value, first-order moment, and second-order moment. Specifically, gray-level information entropy is an index to measure the complexity of the gray-level distribution of an image. The larger the gray-level information entropy, the richer the texture information of the image; the first contrast is an index to measure the gray-level contrast of an image calculated based on the gray-level co-occurrence matrix. The larger the contrast, the more drastic the gray-level change of the image; energy (ASM) is an index to measure the texture uniformity of an image. The larger the energy, the more uniform the texture of the image. Differential entropy is an index to measure the complexity of the gray-level difference of an image based on the gray-level difference. The larger the differential entropy, the richer the gray-level difference information of the image. The second contrast is an index to measure the gray-level contrast of an image calculated based on the gray-level difference. The average value is an index to measure the average gray-level of an image. The first-order moment is an index representing the mean of the gray-level distribution. The second-order moment is an index to measure the second-order statistical characteristics (such as variance) of the gray-level distribution of an image.

[0096] Furthermore, based on the corresponding flaw detection target, perform dimensionality reduction analysis on the broad-spectrum feature data, extract the main features, and store them as the first target key feature set.

[0097] In some implementation manners, according to the multi-target key feature set, construct multiple classification detection channels, and perform supervised training based on the historical flaw detection records. Integrate the multiple trained classification detection channels to obtain a multi-channel classification detector, and further include:

[0098] Based on the first target key feature set, construct a first classification detection model according to the classification algorithm.

[0099] Using the first historical flaw detection record as the training data set, perform supervised training of the first classification detection model to obtain the first classification detection channel.

[0100] Based on the multi-objective key feature set and the historical flaw detection record, construct and train multiple classification detection channels.

[0101] Obtain the partitioning constraints for adaptive grid partitioning, and based on the partitioning constraints, configure an input allocator, connect multiple classification detection channels to the input allocator, and construct the multi-channel classification detector.

[0102] Optionally, first, use multiple feature indicators in the first objective key feature set as input data classes, and construct a first classification detection model based on a classification algorithm, where the classification algorithm includes support vector machine, decision tree, neural network (such as KNN), etc. Then, use the historical flaw detection record as the training data set, label the damage labels, and train each classification detection model so that it can predict the corresponding flaw detection category according to the input features.

[0103] Furthermore, repeat the above process for each objective key feature set to construct multiple classification detection channels. Each channel is an independent classification detection model trained for a specific objective key feature set. Finally, construct an input allocator to connect all the classification detection channels to form a multi-channel classification detector, which can allocate the input data to the appropriate classification detection channels according to the partitioning constraints of the adaptive grid partitioning. Exemplarily, if the partitioning constraints include the position of the component to be tested and the corresponding multiple flaw detection targets, through the above partitioning constraints, the input allocator will transmit the input data corresponding to the multiple flaw detection targets to the corresponding classification detection channels for flaw detection.

[0104] Through the above method steps, specialized training is carried out for different feature sets, and multiple classification detection channels specifically applicable to a single flaw detection target are constructed. In other words, each classification detection channel is specifically trained for a specific flaw detection target, can detect its specific target more accurately and effectively, and will not be interfered by other targets, which helps to improve the interpretability and detection efficiency of the multi-channel classification detector.

[0105] In addition, the design of the multi-channel classification detector enables it to be deployed more flexibly. Exemplarily, if a new flaw detection target needs to be added, only a new classification detection channel needs to be added, rather than retraining the entire system. At the same time, if a problem occurs in a certain channel, it can be repaired or replaced separately without affecting the operation of other channels.

[0106] Furthermore, for the intelligent flaw detection of the meshed three-dimensional image, it further includes:

[0107] Obtain the flaw detection result, parse to obtain multiple edge damages, and output them as an edge damage set.

[0108] Based on the damage type, damage location, damage direction, and damage size, perform a continuity analysis on the edge damage set to obtain a continuity analysis result, where the continuity analysis result includes multiple edge damage groups.

[0109] Based on the continuity analysis result, perform damage merging processing on the edge damage set to obtain a merged edge damage set.

[0110] Update the merged edge damage set to the flaw detection result.

[0111] Optionally, due to the posteriority of feature recognition and damage detection, when obtaining a grid-based three-dimensional image, there is often a situation where a complete damage image is divided into discrete parts, thus affecting the accuracy of flaw detection analysis and the continuity and integrity of the damages in the analysis results. Therefore, based on the damages located at the grid edges in the grid-based three-dimensional image obtained from the flaw detection result, an edge damage set is obtained to merge connected damages, reduce duplicate and redundant damages, and provide a clearer and more accurate flaw detection result.

[0112] Specifically, first, obtain the flaw detection result from a multi-channel classification detector. The flaw detection result includes a series of damages, and each damage includes its position information, size information, and type information. Then, parse out multiple edge damages from the flaw detection result and output these damages as an edge damage set. Next, based on the damage type, damage location, damage direction, and damage size, perform a continuity analysis on the edge damage set. Exemplarily, it involves checking whether there are multiple damages that are continuous or close in space, direction, or type, and then forming these continuous damages into an edge damage group. Repeat the analysis process to output multiple edge damage groups as the continuity analysis result.

[0113] Further, after performing the continuity analysis, perform damage merging processing on the edge damage set. Specifically, merge each edge damage group in the continuity analysis result into a single damage to obtain a merged edge damage set; finally, update the merged edge damage set into the flaw detection result to replace the edge damage set.

[0114] Through the above steps, a clearer and more continuous damage image can be provided, thereby improving the accuracy and interpretability of flaw detection.

[0115] In some implementation manners, after performing a continuity analysis on the edge damage set based on the damage type, damage location, damage direction, and damage size, it further includes:

[0116] Analyze the results of the continuity analysis, discriminate the type differences of multiple edge damage groups, and extract multiple differential edge damage groups.

[0117] Merge multiple differential edge damage groups to obtain multiple differential damage partitions.

[0118] Activate the multi-channel classification detector to perform updated detection on multiple differential damage partitions to obtain an updated damage type set.

[0119] Combine the updated damage type set to perform damage merging processing on the edge damage set.

[0120] Optionally, due to the incompleteness of the edge damage in the edge damage group, the flaw detection of the edge damage by the multi-channel classification detector may be biased, thus misjudging the damage type of the edge damage. Therefore, multiple differential edge damage groups are respectively merged, that is, multiple differential edge damage groups are regarded as new grids to obtain multiple differential damage partitions. Then, activate the multi-channel classification detector to perform updated detection on multiple differential damage partitions to obtain more accurate damage type information (i.e., the updated damage type set). Finally, combine the updated damage type set to perform damage merging processing on the edge damage set, including marking the damage type of the merged edge damage set with the updated damage type set. Thereby, better identifying and understanding the characteristics of the damage, and more effectively processing and organizing the damage information.

[0121] In summary, the intelligent detection method based on non-destructive flaw detection equipment provided by the present invention has the following technical effects:

[0122] By establishing a twin digital model, which includes a model of the component to be measured and an endoscope model. Collect internal images of the component to be measured through the endoscope to obtain the original surface image. Perform multi-dimensional feature extraction on the original surface image to obtain a multi-dimensional feature set. Based on this multi-dimensional feature set, perform image screening on the original surface image to obtain a standard surface image. Interact with the interactive control component to obtain the motion parameter sequence of the endoscope optical component. Based on these motion parameter sequences, fit the standard surface image to the twin digital model to generate a three-dimensional restored image. Based on the shape characteristics of the component to be measured model, perform adaptive mesh division on the three-dimensional restored image to obtain a meshed three-dimensional image. Combine the data dimensionality reduction method to construct a multi-channel classification detector for intelligent flaw detection of the meshed three-dimensional image. This multi-channel classification detector includes multiple flaw detection channels, and each channel is responsible for performing different types of flaw detection tasks. Thereby achieving the technical effects of reducing the detection cost and the operation difficulty.

[0123] It should be understood that the embodiments and the above descriptions disclosed in the present invention enable those skilled in the art to implement the present invention using the present invention. At the same time, the present invention is not limited to the above-mentioned part of the embodiments. It should be understood that those of ordinary skill in the art can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. An intelligent detection method based on non-destructive testing equipment, characterized in that: The method comprises: Establish a twin digital model, including the model of the component to be tested and the endoscope model; Collect the internal image of the component to be tested through the endoscope to obtain the original surface image; Performing multidimensional feature extraction on the original surface image to obtain a multidimensional feature set, and performing image screening on the original surface image based on the multidimensional feature set to obtain a standard surface image; The interactive control component obtains a motion parameter sequence of the endoscope optical component, and fits the standard surface image to the twin digital model based on the motion parameter sequence to obtain a three-dimensional restored image; Based on the shape characteristics of the model of the component to be tested, adaptively meshing the three-dimensional restored image to obtain a meshed three-dimensional image; In combination with the data dimension reduction method, a multi-channel classification detector is constructed to perform intelligent flaw detection of the gridded three-dimensional image, wherein the multi-channel classification detector includes multiple flaw detection channels for performing different types of flaw detection tasks.

2. The intelligent detection method based on non-destructive testing equipment according to claim 1, characterized in that: The internal image of the component to be tested is collected through the endoscope to obtain the original surface image, including: Based on the model of the component to be tested, defining multi-level acquisition parameters; Perform image acquisition path planning of the component model to be tested in combination with the multi-level acquisition parameters to obtain an internal image acquisition scheme, wherein the internal image acquisition scheme includes a plurality of serialized acquisition points; The endoscope is activated to collect images based on the internal image collection scheme to generate an original surface image.

3. The intelligent detection method based on non-destructive testing equipment according to claim 2 is characterized in that: Extracting multidimensional features of the original surface image to obtain a multidimensional feature set, and performing image screening on the original surface image based on the multidimensional feature set to obtain a standard surface image, including: Graying the original surface image, and solving a gray-level co-occurrence matrix of the gray-level image; Based on the gray level co-occurrence matrix, gray level information entropy and a first contrast ratio are calculated and obtained; Establishing a scatter point coordinate system, using the grayscale information entropy and the first contrast as the first coordinate and the second coordinate of the scatter point coordinate system respectively, and performing scatter point distribution of the original surface image; A cluster analysis is performed on the scatter point distribution results, and the original surface image is screened according to the cluster analysis results to eliminate multiple original images included in abnormal clusters to generate the standard surface image.

4. The intelligent detection method based on non-destructive testing equipment according to claim 3 is characterized in that: Combined with the data dimensionality reduction method, a multi-channel classification detector is constructed, including: Acquire historical flaw detection records of the target environment, wherein the historical flaw detection records include historical surface images and associated historical damage information; The interactive target flaw detection scenario obtains the flaw detection task information and extracts the flaw detection target set; Extracting broad-spectrum feature data of the historical surface image, and grouping and reducing the dimension of the broad-spectrum feature data based on the flaw detection target set to obtain a multi-target key feature set; According to the multi-target key feature set, a plurality of classification detection channels are constructed, and supervised training is performed based on the historical flaw detection records, and the plurality of classification detection channels after training are integrated to obtain a multi-channel classification detector.

5. The intelligent detection method based on non-destructive testing equipment according to claim 4, characterized in that: Extracting the broad-spectrum feature data of the historical surface image, and grouping and reducing the dimension of the broad-spectrum feature data based on the flaw detection target set to obtain a multi-target key feature set, further comprising: Acquire a first flaw detection target based on the flaw detection target set, where the first flaw detection target corresponds to any damage type; According to the first flaw detection target, traverse the historical flaw detection records to extract a first historical flaw detection record, where the first historical flaw detection record includes a first historical surface image and first historical damage information; Solving the grayscale co-occurrence matrix, grayscale difference, and color moment of the first historical surface image, and performing feature extraction to obtain first broad-spectrum feature data, wherein the first broad-spectrum feature data at least includes grayscale information entropy, first contrast, energy, differential entropy, second contrast, average value, first-order moment, and second-order moment; Performing dimensionality reduction analysis on the first broad-spectrum feature data, extracting main features, and storing them as a first target key feature set; The detection target set is traversed to obtain a multi-target key feature set.

6. The intelligent detection method based on non-destructive testing equipment according to claim 4, characterized in that: According to the multi-target key feature set, multiple classification detection channels are constructed, and supervised training is performed based on the historical flaw detection records, and the multiple classification detection channels after training are integrated to obtain a multi-channel classification detector, which also includes: According to the first target key feature set, a first classification detection model is constructed based on a classification algorithm; Using the first historical flaw detection record as a training data set, performing supervised training on the first classification detection model to obtain a first classification detection channel; Based on the multi-target key feature set and the historical flaw detection records, construct and train a plurality of the classification detection channels; A partition constraint of the adaptive grid partition is obtained, and based on the partition constraint, an input distributor is configured, a plurality of the classification detection channels are connected to the input distributor, and the multi-channel classification detector is constructed.

7. The intelligent detection method based on non-destructive testing equipment according to claim 1, characterized in that: The intelligent flaw detection of the gridded three-dimensional image also includes: Obtain the flaw detection results, parse and obtain multiple edge damages, and output them as an edge damage set; Based on the damage type, damage location, damage direction, and damage size, a continuity analysis is performed on the edge damage set to obtain a continuity analysis result, wherein the continuity analysis result includes a plurality of edge damage groups; Based on the continuity analysis result, performing damage merging processing on the edge damage set to obtain a merged edge damage set; The combined edge damage set is updated to the flaw detection result.

8. The intelligent detection method based on non-destructive testing equipment according to claim 7 is characterized in that: Based on the damage type, damage location, damage direction, and damage size, the edge damage set is subjected to continuity analysis, and then, the following is further included: Analyzing the continuity analysis results, distinguishing the types of the plurality of edge injury groups, and extracting a plurality of difference edge injury groups; Merging a plurality of the differential edge damage groups to obtain a plurality of differential damage partitions; activating the multi-channel classification detector to perform update detection on the plurality of differential damage partitions to obtain an updated damage type set; The edge damage set is subjected to damage merging processing in combination with the updated damage type set.

Images