Vehicle assembly defect online monitoring method and monitoring system based on CT image

Abstract

The application relates to the fields of pre-diagnosis and health management, industrial vision and image sensor technology, in particular to a vehicle assembly defect online monitoring method and monitoring system based on CT images, which comprises the following steps: step S1, collecting CT projection data of a vehicle passing through each station, establishing a CT process file of the vehicle, performing assembly structure-oriented reconstruction on the projection data to obtain a three-dimensional CT image, and converting the three-dimensional CT image obtained through reconstruction into an assembly relationship diagram; the application quantifies defect evidence into executable risk information and graded alarms through structure reliability rendering and working condition probability evaluation combined with an assembly sequence, so that reliable detection and risk early warning of vehicle internal assembly defects can be realized under the online beat of an assembly line, and the problems in the prior art that shielding compensation is unreliable, a fixed template generates false defects, image difference lacks assembly semantics, and detection results cannot quantify risks and drive early warning are solved.

Description

Technical Field

[0001] This application relates to the fields of fault detection and CT image technology, and more specifically, to a method and system for online monitoring of vehicle assembly defects based on CT images. Background Technology

[0002] The content in this section provides only background information related to this application and may not constitute prior art.

[0003] Vehicles are complex structures assembled from thousands of parts through multiple processes. In scenarios such as vehicle assembly lines, powertrain assembly lines, battery pack wiring lines, and axle and chassis assembly lines, numerous assembly defects (such as missing bolts, loose clips, improperly pressed seals, misaligned bearings, loose wiring harness connectors, foreign matter residue in cavities, and weak welds) occur internally or are obscured by subsequently assembled components. These defects are difficult to detect through visual inspection before leaving the factory, but can lead to abnormal noises, leaks, loosening, and even safety accidents during vehicle use. Traditional methods such as manual visual inspection, tightening curve monitoring, and two-dimensional X-ray transmission inspection can only cover the surface and visible area, or only obtain overlapping projection information, failing to reliably assess the internal three-dimensional assembly status.

[0004] Therefore, industrial CT is used for inspection. This involves registering the measured CT image with a fixed standard CT image or CAD model and then directly calculating the difference. The difference area is judged as a defect. This solution has a high degree of automation and does not require manual layer-by-layer image reading.

[0005] However, for areas obscured by large metal components, the projection information obtained from a single scan is incomplete, resulting in stripe artifacts and missing data in the reconstruction results, making it easy to miss detections. For standard template comparison, each vehicle objectively has legitimate manufacturing tolerances, clamping posture deviations, and material batch differences. When directly subtracting from a fixed template, these individual differences that do not constitute defects also manifest as image differences, leading to a high false alarm rate. Furthermore, ordinary interpolation is typically used to fill in areas with metal artifacts, scattering, and occlusion. The interpolated content lacks physical basis, which may either mask real defects or fabricate non-existent structures.

[0006] Therefore, existing industrial CT scanners cannot accurately determine whether a vehicle has assembly defects. Summary of the Invention

[0007] The summary section of this application is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description section below. This summary section is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0008] Some embodiments of this application propose an online monitoring method and system for vehicle assembly defects based on CT images to address the technical problems mentioned in the background section above.

[0009] As a first aspect of this application, this application provides a method for online monitoring of vehicle assembly defects based on CT images, comprising the following steps: Step S1: Collect CT projection data of the vehicle passing through each workstation, establish a CT process file for the vehicle, reconstruct the projection data for the assembly structure to obtain a three-dimensional CT image, and convert the reconstructed three-dimensional CT image into an assembly relationship diagram. Step S2: Generate a fitted qualified CT image of the current assembly state based on the historical qualified vehicle sample library; Step S3: Perform bidirectional residual calculation between the measured CT image and the fitted qualified CT image, and propagate the residual along the assembly relationship diagram to obtain the defect confidence of each structural node; Step S4: Determine the structural reliability of each structural node based on the defect reliability of each structural node, generate a structural reliability rendering map based on the structural reliability, generate the probability of occurrence of several assembly error conditions in combination with the assembly sequence, and generate an assembly error alarm based on the probability of occurrence.

[0010] This application uses vehicle identification as the primary key, organizing scans of the same vehicle at various workstations into a continuous assembly process image chain. During the reconstruction phase, it introduces three priors: preceding structure, component geometry, and qualified samples. A credibility field is established for each voxel in the reconstruction result, converting the 3D CT image into an assembly relationship graph with components as nodes and assembly relationships as edges. Based on a historical database of qualified vehicle samples, a fitted qualified CT image, preserving individual differences, is generated for the tested vehicle as a comparison benchmark. Bidirectional residual calculation is performed on this benchmark and propagated along the assembly relationship graph, imbuing image differences with assembly semantics. Furthermore, through structural credibility rendering and combined with operational probability assessment based on assembly sequence, defect evidence is quantified into actionable risk information and graded alarms. This enables reliable detection and risk warning of internal vehicle assembly defects under online assembly line conditions, overcoming the problems of unreliable occlusion compensation, false defects generated by fixed templates, lack of assembly semantics in image differences, and inability to quantify risks and drive warnings using detection results in existing solutions.

[0011] Furthermore, step S1 includes the following steps: Step S11: When the vehicle moves along the assembly line to the t-th inspection station, collect the projection data of that station. It also obtains the vehicle identification number, the current workstation number t, and the assembly list for the current workstation. Assembly sequence information; Step S12: Analyze and reconstruct the projection data of each workstation, identify occlusion and artifact regions, and obtain the occlusion weight matrix W; Step S13: Determine the currently visible structure and corresponding parts based on the occlusion weight matrix W. Input the structural information reconstructed from the previous workstation projection data, the geometric priors of the currently visible parts, and qualified samples into the reconstruction process in the form of regularization terms, together with the measured projection data of the current workstation, to construct and solve the reconstruction objective function to obtain the three-dimensional CT image. Step S14: For each voxel position x in the 3D CT image, calculate the confidence level c(x) to establish a confidence field; Step S15: For the 3D CT image with credibility information, treat each component as a structural node, extract the edges between the components, and construct the assembly relationship diagram for the t-th workstation. , Represents a set of structural nodes. Represents the set of relation edges.

[0012] This application uses vehicle identification as the primary key to organize the projection data of the same vehicle at each workstation into a continuous assembly process image chain, enabling any workstation inspection to access the images and structural information of all preceding workstations for that vehicle. By using an occlusion weight matrix to distinguish reliable data from contaminated data at the ray level, and by introducing three types of priors in the reconstruction—preceding structures, component geometric priors, and qualified samples—the compensation content for occluded areas has a physical and technological basis, avoiding the problem of ordinary interpolation filling masking real defects or fabricating non-existent structures. By establishing a credibility field for each voxel, reliable and unreliable regions in the reconstruction results are explicitly distinguished at the source. By converting 3D CT images into an assembly relationship graph with components as nodes and assembly relationships as edges, a structured foundation is provided for subsequently assigning assembly semantics to image differences.

[0013] Furthermore, the occlusion weight matrix W is for the projected data. Each ray in the process is assigned a weight to form the ray.

[0014] Furthermore, the reconstruction objective function is: ; Where V represents the three-dimensional CT image to be reconstructed, and the value V(x) of the three-dimensional CT image at spatial location x is the linear attenuation coefficient of that voxel; This represents the optimal 3D CT image obtained from the reconstruction; A represents the forward projection matrix of the CT system, which describes the physical mapping from the 3D CT image to the projection data. represents the projection data vector obtained from the actual measurement at station t; W represents the occlusion weight matrix obtained in step S12. This indicates element-wise multiplication. This represents the square of the second norm of a vector; , , These are the weight coefficients of the three regularization terms, all of which are preset non-negative real numbers used to balance the influence of the data fidelity term and each prior term; This is a regularization term for the consistency of the preceding workstation structure. For the geometric prior regularization terms of the components; This is a regularization term for qualified samples.

[0015] This application uses an occlusion weight matrix to weight data fidelity terms element-wise, automatically reducing the weight of rays affected by stripe artifacts, scattering contamination, or missing data during reconstruction, thus suppressing the contamination of reconstruction results by unreliable projection data. Furthermore, it imposes constraints on the reconstruction process through preceding station structure consistency regularization, component geometry prior regularization, and qualified sample regularization, ensuring that the reconstructed content of the occluded area is simultaneously limited by projection evidence, the vehicle's own preceding structure, and the prior knowledge of the qualified group. Individual differences within the manufacturing tolerance range are not penalized, thereby improving the reliability of defect detection in occluded areas and reducing false alarms while ensuring the physical reliability of the reconstruction results.

[0016] Furthermore, step S2 includes the following steps: Step S21: Using vehicle identification v and current workstation number t as indexes, retrieve K qualified samples from the historical qualified vehicle sample database Q(m). Each qualified sample contains its three-dimensional CT image at workstation t and the corresponding assembly relationship diagram. Step S22: Assemble according to the current workstation's assembly list The 3D CT image of each qualified sample is decomposed into three parts: the basic structure of the vehicle body, the newly added structure at the current workstation, and the connecting transition area. Step S23: Select a set of stable structures from the 3D CT images of the tested vehicle. As a positioning framework; Step S24: Based on the positioning skeleton, for each qualified sample k, solve the spatial transformation that maps it to the measured vehicle coordinate system, and register the qualified sample to the measured 3D CT image according to the spatial transformation. The measured vehicle coordinate system refers to the vehicle coordinate system of the 3D CT image of the vehicle being measured at the current workstation; Step S25: Take the median density at each voxel position of the K registered qualified samples to obtain the qualified population reference value; compare the measured 3D CT image with the qualified population reference value voxel by voxel, and identify samples with deviations exceeding the tolerance band. And the credibility c(x) is not lower than The area is marked as a suspected abnormal area; Step S26: Using the positioning skeleton as the positioning basis, migrate the newly added structure and connecting transition area of ​​the current station of the qualified sample to the actual vehicle coordinates, and fuse it with the actual image to generate a fitted qualified CT image.

[0017] This application uses a stable structure as the positioning framework to register multiple qualified samples to the coordinate system of the measured vehicle, and takes the median density at each voxel position as the reference value for the qualified group. This makes the comparison benchmark represent the common state of the qualified group rather than the individual characteristics of any single sample. By replacing the contents of qualified samples with the contents of suspected abnormal areas only, and retaining the measured contents of the remaining areas, the legitimate manufacturing tolerances, clamping posture deviations and material batch differences of the tested vehicle are retained and do not enter the subsequent residuals. This fundamentally eliminates the false defects caused by directly subtracting from the fixed standard template and reduces the false alarm rate.

[0018] Furthermore, in step S26, for stable structural regions and regions not marked as suspected anomalies, the content of the measured CT image is retained; only for regions marked as suspected anomalies, the corresponding content of the registered qualified samples is replaced by weighted fusion weights; the fusion weights are determined by normalizing the similarity between each qualified sample and the measured vehicle in the stable structural region, and the higher the similarity, the greater the weight.

[0019] Furthermore, step S3 includes the following steps: Step S31: Remove the contents of the fitted qualified CT image from the measured CT image to obtain the positive residual; remove the contents of the measured CT image from the fitted qualified CT image to obtain the inverse residual; Step S32: For each structural node u, perform volume integrals and normalize the forward and reverse residuals within its spatial range to obtain the node image residual energy. For each relation edge e, calculate the edge change. ; Step S33: Propagate the positive and negative residuals along the relationship between the components to obtain the defect confidence level of each structural node.

[0020] This application uses bidirectional calculation of forward and reverse residuals to explicitly characterize the extra material (corresponding to misassembly, foreign object residue, etc.) and missing material (corresponding to omission, incomplete assembly, etc.) in the measured image, avoiding confusion and missed detection of anomaly types caused by unidirectional subtraction. By propagating the residuals along the relational edges of the assembly relationship diagram, the grayscale differences in the image are given assembly semantics: differences that are isolated and do not disrupt any assembly relationship are downweighted, while differences that disrupt assembly relationships such as contact, coaxiality, sealing, or fastening force transmission are upweighted, thereby transforming ordinary image differences into interpretable evidence of assembly defects and improving the accuracy of defect judgment.

[0021] Furthermore, step S4 includes the following steps: Step S41: Obtain the node defect confidence level of all structural nodes and generate a confidence level distribution map; Step S42: Obtain the assembly sequence of structural nodes, and generate the probability of occurrence of several corresponding assembly error conditions based on the defect confidence of each structural node. Step S43: Generate an assembly error alarm based on the probability of the assembly error occurring.

[0022] This application uses a structural reliability rendering diagram to visually present the assembly reliability of each component in the form of a three-dimensional color image, enabling production line personnel to directly locate low-reliability components within the production line cycle time without having to read CT slices layer by layer. By combining the assembly sequence to assess the probability of occurrence of various assembly error conditions, the associated risks of abnormalities in the first-assembled baseline structure to the subsequently assembled components are reflected, and defect evidence is quantified into actionable risk information. By generating graded alarms according to the probability of occurrence, differentiated handling of line stoppage, re-inspection, and recording is achieved, ensuring reliable defect interception while avoiding unnecessary interference with the production line cycle time.

[0023] Furthermore, the construction of the historical qualified vehicle sample library includes the following steps: Step A1: Acquire the CT process file of each vehicle on the production line in real time and obtain multiple candidate samples; Step A2: Organize candidate samples using workstation number t as the index key; each candidate sample record contains the 3D CT image of the vehicle at the corresponding workstation, the confidence field c(x), and the assembly relationship diagram. and assembly list .

[0024] Step A3: Register all samples to the unified reference coordinate system of the vehicle model using stable structures (vehicle longitudinal beams, main positioning holes, etc.); check the confidence field of each sample and remove samples or local data blocks whose regional average confidence is lower than the preset threshold. Step A4: The sample library is updated on a rolling basis: new qualified vehicle samples are continuously added to the library, and samples that have exceeded the preset retention period are deactivated.

[0025] As a second aspect of this application, this application provides an online monitoring system for vehicle assembly defects based on CT images, comprising: The CT acquisition device is installed at the inspection station on the assembly line. Data storage server, housing process archives and historical qualified vehicle sample database; An algorithm server is used to run the online vehicle assembly defect monitoring method based on CT images. Attached Figure Description

[0026] Figure 1 This is a flowchart of a method for online monitoring of vehicle assembly defects based on CT images.

[0027] Figure 2 This is a schematic diagram for calculating two-way residuals. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments. The same reference numerals in the accompanying drawings represent the same components. It should be noted that the described embodiments are only some, not all, of the embodiments of this application. All other embodiments obtained by those skilled in the art based on the described embodiments of this application without creative effort are within the scope of protection of this application.

[0029] Compared to the embodiments shown in the accompanying drawings, feasible embodiments within the scope of this application may have fewer components, other components not shown in the drawings, different components, differently arranged components, or components with different connections, etc. Furthermore, two or more components in the drawings may be implemented in a single component, or a single component shown in the drawings may be implemented as multiple separate components.

[0030] Unless otherwise defined, the technical or scientific terms used herein shall have the ordinary meaning as understood by one of ordinary skill in the art to which this application pertains. The terms “first,” “second,” and similar terms used in this specification and claims do not indicate any order, quantity, or importance, but are merely used to distinguish different components. Similarly, the terms “an” or “a” and similar terms do not necessarily indicate a quantity limitation. Terms such as “upper” and “lower” are used only to indicate relative positional relationships, and these relative positional relationships may change accordingly when the absolute position of the described object changes.

[0031] The strong attenuation of X-rays by large metal components in the vehicle body causes streaks, scattering contamination, and projection loss. The CT information of the obscured areas is inherently incomplete. If ordinary interpolation is used to fill in the gaps, the filled content has no physical basis, which may cover up real defects or fabricate non-existent structures. Moreover, each vehicle has legitimate manufacturing tolerances, clamping posture deviations, and material batch differences. If the measured images are directly subtracted from a fixed standard template, these individual differences will also appear as image differences, which will drown out the real defect signals. Even if the image differences are obtained, the differences themselves are only gray-level information and cannot answer "what part is missing, what connection relationship is wrong, and how much risk does it bring to the production line?" Furthermore, a single isolated scan is out of sync with the assembly sequence, making risk quantification and early warning impossible.

[0032] To this end, this application uses vehicle identity as the primary key to organize the scans of the same vehicle at various workstations into an assembly process image chain, enabling each inspection to access the structural information of all preceding workstations for that vehicle. Then, during the reconstruction phase, three types of prior knowledge are introduced: preceding structure, component geometry, and qualified samples. The occlusion compensation content is limited by triple constraints of projection evidence, preceding structure, and assembly list. Simultaneously, each voxel is labeled with a credibility level, allowing "reliable" and "unreliable" to be distinguished at the source. Next, the 3D CT image is converted into an assembly relationship diagram with components as nodes and relationships such as contact, coaxiality, and sealing as edges. A fitted qualified CT image, retaining its individual differences, is generated specifically for the vehicle as a comparison benchmark. On this benchmark, forward and reverse bidirectional residuals are generated and propagated along the relationship edges, giving assembly meaning to image differences. Finally, a structural credibility rendering diagram is generated based on the defect credibility of each structural node. The probability of various assembly error conditions is assessed in conjunction with the assembly sequence, and graded alarms are generated according to the probability.

[0033] In this way, individual differences no longer produce false defects, occlusion compensation no longer fabricates structures, image differences become interpretable assembly evidence, and the risks of various assembly errors are quantified as probabilities of occurrence and directly drive production line early warning.

[0034] The implementation scenarios for the following embodiments are set as follows: A certain vehicle assembly line has T CT online inspection stations set up along the process flow, for example, T = 6, which are respectively arranged after the body-in-white rolls off the line, after the chassis is assembled, after the powertrain is installed, after the wiring harness and pipeline are laid, after the interior parts are installed, and before the vehicle rolls off the line. The vehicle is moved between the stations by AGV or skid transport system.

[0035] Each inspection station is equipped with a surround-type CT acquisition device. The X-ray source and flat panel detector are installed on a circular track, or two collaborative robots hold the X-ray source and detector respectively and move synchronously around the vehicle. After the vehicle is in place, the position is locked by the positioning pin and the visual positioning unit, and then the cone-beam scan is triggered to acquire projection data from multiple angles.

[0036] The X-ray source, flat panel detector, and motion control of the cone-beam CT scanning device are existing technologies, and their specific structures will not be further described in this application; the key to this application lies in the following processing procedure performed on the acquired projection data and reconstructed images.

[0037] refer to Figure 1 Example 1: Online monitoring method for vehicle assembly defects based on CT images.

[0038] The method in this embodiment includes the following steps: Step S1: Collect CT projection data of the vehicle passing through each workstation, establish a CT process file for the vehicle, reconstruct the projection data for the assembly structure to obtain a three-dimensional CT image, and convert the reconstructed three-dimensional CT image into an assembly relationship diagram. Step S2: Generate a fitted qualified CT image of the current assembly state based on the historical qualified vehicle sample library; Step S3: Perform bidirectional residual calculation between the measured CT image and the fitted qualified CT image, and propagate the residual along the assembly relationship diagram to obtain the defect confidence of each structural node; Step S4: Determine the structural reliability of each structural node based on the defect reliability of each structural node, generate a structural reliability rendering map based on the structural reliability, generate the probability of occurrence of several assembly error conditions in combination with the assembly sequence, and generate an assembly error alarm based on the probability of occurrence.

[0039] The above outlines the general steps of this method and the general function of each step. Specific implementation details for each step are explained below: Step S1: Collect CT projection data of the vehicle passing through each workstation, establish a CT process file for the vehicle, reconstruct the projection data for the assembly structure to obtain a three-dimensional CT image, and convert the reconstructed three-dimensional CT image into an assembly relationship diagram.

[0040] Step S1 includes the following steps: Step S11: When the vehicle or assembly to be inspected moves along the assembly line to the t-th inspection station, the projection data of that station is collected. It also obtains the vehicle identification number, the current workstation number t, and the assembly list for the current workstation. Assembly sequence information.

[0041] Specifically, projection data This is a collection of two-dimensional attenuated images acquired by a CT acquisition device from multiple angles; vehicle identification, current workstation number t, and assembly list for the current workstation. Assembly sequence information and current tooling status are read from the vehicle's manufacturing execution system.

[0042] The vehicle identification number (v) represents the vehicle's identification code, which can be used to identify the target vehicle and its corresponding model code (m).

[0043] The current workstation number t indicates the current process.

[0044] Assembly list That is, the set of parts and their material codes that should be newly installed at station t according to the process specifications.

[0045] Assembly sequence information is the order in which each component is assembled.

[0046] This method does not treat each CT projection data as an isolated sample. Instead, it uses the vehicle identification v as the primary key and organizes the projection data of the same vehicle at different workstations into a continuous assembly process image chain according to the assembly process sequence. It also includes the current workstation number t and the assembly list. Assembly sequence information and current tooling status are written into the process archive.

[0047] In this way, the inspection of any vehicle at any station can access the images and structural information of all its preceding stations, which is the data basis for the subsequent reconstruction of prior knowledge and sequence correction.

[0048] Step S12: Analyze and reconstruct the projection data of each workstation, identify occlusion and artifact regions, and obtain the occlusion weight matrix W.

[0049] Specifically: the projection data is initially analyzed and reconstructed to segment out high-attenuation metal regions and project them forward to obtain metal tracks; at the same time, the consistency of adjacent angle projections is checked, and rays with local projection residuals exceeding the threshold are judged as contaminated rays; accordingly, each ray is assigned an occlusion weight to form an occlusion weight matrix W, whose elements take values ​​in the range of [0,1]. The weight of reliable rays approaches 1, and the weight of rays contaminated by stripes, scattering, or missing rays approaches 0.

[0050] The occlusion weight matrix W is for the projection data. Each ray (i.e., a measurement by each detector pixel at each scanning angle) is assigned a weight. Let the acquisition scale of a single cone-beam scan at position t be: The projection angle is (Number of angles around the vehicle collected by the X-ray source and detector); Number of pixel rows for the flat panel detector is... The number of pixel columns is (That is, the detector outputs one image at each angle) (Two-dimensional attenuated image).

[0051] The total number of rays (measurements) in a single scan is: ; Where M represents the projection data The total length after being expanded into a vector is the total number of occlusion weights; each ray corresponds to one occlusion weight. ( ), The weight of reliable rays approaches 1, while the weight of rays contaminated by streak artifacts, scattering, or missing data approaches 0.

[0052] The preliminary analytical reconstruction can be achieved using analytical reconstruction algorithms commonly used in cone-beam CT. Such algorithms are existing technologies, and this application will not further describe their specific processes.

[0053] Step S13: Determine the currently visible structure and corresponding components based on the occlusion weight matrix W. Input the structural information reconstructed from the previous workstation projection data, the geometric priors of the currently visible components, and qualified samples into the reconstruction process in the form of regularization terms, together with the measured projection data of the current workstation, to construct and solve the reconstruction objective function to obtain the three-dimensional CT image.

[0054] Specifically, steps S11 and S12 obtain projection data. Projection data is generated when an X-ray source illuminates the vehicle from a certain angle; the rays are attenuated after passing through the vehicle body and then strike a flat panel detector, which records a two-dimensional "attenuated image." Therefore, after obtaining the projection data, it is necessary to reconstruct it to obtain a three-dimensional CT image. How to reconstruct a three-dimensional CT image using complete projection data is a current technology.

[0055] However, there are occluded areas in the projection data (see step S12). To reconstruct these occluded areas, this application utilizes prior information such as the structural information of the preceding workstations for constraint. Therefore, regularization is introduced into the reconstruction objective function. Given the objective function, how to iteratively solve for the optimal 3D CT image is an existing technique, and this application will not further describe the reconstruction solution process.

[0056] The currently visible structure is determined as follows: the degree to which each spatial region is covered by reliable rays (i.e., projection support) is calculated by the occlusion weight matrix W. The structural region with projection support not lower than a preset threshold is the currently visible structure, and its corresponding component is the currently visible component.

[0057] The reconstruction objective function is shown in equation (1): ; Where V represents the three-dimensional CT image to be reconstructed, and the value V(x) of the three-dimensional CT image at spatial location x is the linear attenuation coefficient of that voxel; This represents the optimal 3D CT image obtained from the reconstruction; A represents the forward projection matrix of the CT system, which describes the physical mapping from the 3D CT image to the projection data. represents the projection data vector obtained from the actual measurement at station t; W represents the occlusion weight matrix obtained in step S12. This indicates element-wise multiplication. This represents the square of the second norm of a vector; , , These are the weight coefficients of the three regularization terms, all of which are preset non-negative real numbers used to balance the influence of the data fidelity term and each prior term; This is a regularization term for the consistency of the preceding workstation structure. For the geometric prior regularization terms of the components; This is a regularization term for qualified samples.

[0058] The 3D CT image of the vehicle from the previous workstation is rigidly registered and locally elastically registered and mapped to the coordinate system of the current workstation. This is then used to align the image according to the assembly list. The stable structural region that should not change is constrained to maintain consistency between the current reconstruction result and the mapped preceding structure. The larger the deviation, the larger the value of this term. Among them, the rigid registration and local elastic registration algorithms are existing technologies, and their specific processes will not be further described in this application.

[0059] For visible parts at the current workstation, call their symbolic distance field to constrain the deviation between the reconstructed part surface and the CAD geometric prior to be within the allowable manufacturing tolerance range. The symbolic distance field is a way to convert a three-dimensional shape into a mathematical expression that is easy to calculate. That is, it records the signed shortest distance from each point in space to the standard surface of the part, which can be generated from the CAD model of the part.

[0060] For areas where the current angle is obscured, instead of performing ordinary interpolation, we first search for the still visible parent structure in the image chain of the vehicle's previous workstations, that is, the structure to which the obscured area is attached that was not obscured in the previous workstation; then we retrieve local volume data blocks with the same connection relationship as the parent structure from the historical qualified vehicle sample library as candidate compensation content. The reconstruction results of the constrained occluded area are made closer to the candidate compensation content.

[0061] Specifically: ; in, The voxel domain of the current workstation's 3D CT image is represented; x represents... The voxel position in the image; V(x) represents the value of the 3D CT image to be reconstructed at position x; For the stable region mask function, the value range is [0,1]: for the assembly list according to the current workstation There should be no assembly changes, and the reliability of the preceding reconstruction results should not be lower than the reliability threshold. The region is set to 1, and the newly added assembly region and the previous low-confidence region of the current workstation are set to 0, so that the boundary can be smoothly transitioned. This represents the 3D CT image reconstructed from the previous workstation (workstation t-1) for this vehicle; This represents the coordinate mapping obtained by rigid registration and local elastic registration, which maps the current position x in the coordinate system to the corresponding position in the previous coordinate system. That is, the value of the preceding structure mapped to the current position x.

[0062] ; in, This represents the set of indices of visible components at the current workstation, i.e., components that have been installed at the current workstation and are not completely obscured in the current scan; i represents... Component index; Let V represent the set of surface voxels of the i-th component extracted from the 3D CT image V; The value of the symbolic distance field generated by the CAD model of the i-th component at position x represents the shortest distance from x to the CAD standard surface of the component. The value of x is positive when it is outside the standard surface, negative when it is inside the standard surface, and 0 when it is exactly on the surface. This represents the absolute value of the distance; Let be the half-width of the manufacturing tolerance band for the i-th component, which is a preset value; This indicates the operation of taking the non-negative part. When the reconstructed surface point falls within the tolerance zone, the penalty for that point is zero, meaning that individual differences within the tolerance range are not penalized; a penalty is only applied when the surface deviation exceeds the tolerance zone.

[0063] ; ; in, The mask function for the occlusion area has a value range of [0,1]: 1 for the area occluded at the current angle and 0 for the area with sufficient projection support; the occlusion area can be determined by the projection support obtained by the weighted statistics of the occlusion weights of the ray passing through position x (corresponding elements in the occlusion weight matrix W) being lower than a preset threshold, and the boundary can be smoothly transitioned; The value of the candidate compensation content at position x is obtained by weighted fusion of multiple qualified sample local volume data blocks according to formula (1-4); The value r represents the number of candidate compensation data blocks retained after triple constraint filtering, and is a positive integer; r represents the candidate data block index. The r-th candidate compensation data block is derived from the following: first, the parent structure to which the occluded area is attached is determined in the image chain of the preceding workstation of the vehicle; then, local volume data blocks with the same connection relationship as the parent structure are retrieved from the historical qualified vehicle sample library, and only data blocks that simultaneously satisfy the projection evidence constraint, the preceding structure constraint, and the assembly list constraint are retained. The triple constraints are: projection evidence constraint, that is, the error between the forward projection of the candidate data block and the measured projection of the uncontaminated current workstation is less than a preset threshold; preceding structure constraint, that is, the structure mapped from the preceding workstation of the vehicle is continuous at the boundary; and assembly list constraint, that is, the parts to which the candidate data block belongs are included in the cumulative assembly list up to the current workstation. This represents the registration transformation that maps the current position x in the coordinate system to the coordinate system of the r-th candidate data block, which is obtained by the registration between the parent structure to which the data block is attached and the measured parent structure. Let be the fusion weight of the r-th candidate data block, determined by normalization based on the similarity between the parent structure to which the data block is attached and the measured parent structure. Higher similarity results in a larger weight. The sum of the weights is 1.

[0064] A 3D CT image can be reconstructed based on the projection data of each workstation. However, it is necessary to introduce a reliability measure to determine which locations in the 3D CT image are reliable and which are not.

[0065] Step S14: For each voxel position x in the 3D CT image, calculate the confidence level c(x) to establish a confidence field.

[0066] The formula for calculating credibility is shown in equation (2):

[0067] Where c(x) represents the confidence level of the reconstruction result at voxel position x, and its value ranges from [0,1]. The projection support represents the degree to which position x is covered by uncontaminated rays, and is obtained by weighted statistics of the rays passing through the voxel. The preceding structure support represents the degree of consistency between the reconstruction result at position x and the preceding workstation mapping structure. Historical sample support indicates the degree to which the compensation content at position x is supported by multiple qualified samples. , , These are the fusion weights for the three support metrics, which are preset non-negative real numbers whose sum is 1. For c(x) below a preset confidence threshold... In areas where uncertainty is identified, subsequent steps will not directly use image differences in those areas as evidence of defects.

[0068] Thus, the reconstruction output is not a simple 3D CT image, but a 3D CT image with credibility information. "Which areas can be trusted and which areas cannot be fully trusted" is explicitly expressed from the source.

[0069] Step S15: For the 3D CT image with credibility information, treat each component as a structural node, extract the edges between the components, and construct the assembly relationship diagram for the t-th workstation. , Represents a set of structural nodes. Represents the set of relation edges.

[0070] The segmentation and extraction of structural regions can be guided by assembly CAD drawings: first, the spatial range of each component is initially located using CAD drawings, and then the boundaries are refined within this range according to density and morphological characteristics; the segmentation algorithm itself is existing technology, and its specific process will not be further described in this application.

[0071] In practice, each component is represented as a structural node. Node attributes include: component category Material code, centroid location ,attitude Density statistics (Mean and variance of attenuation coefficients within the region), volume, and regional average reliability.

[0072] The physical and functional connections between components are represented as relation edges, and the types of edges include seven categories: contact, fitting, penetration, clamping, coaxial, sealing, and electrical connection.

[0073] Examples of edge determination rules are as follows: Contact edge: The minimum distance between the surfaces of the two nodes is less than a threshold. Furthermore, the contact surfaces are normally opposite; the coaxial sides have an angle between their axes that is less than the angle threshold. And the eccentricity of the axis is less than the eccentricity threshold. The compaction edge exhibits a density gradient compaction feature along the normal direction at the contact interface, and the contact area is greater than a threshold. The sealing edge maintains continuous contact between the sealing element node and both sides of the structure along the sealing path. The electrical connection edge has male and female connector terminals that are interlocked, and the metal terminal density is continuous. The types of edge relationships can be added or removed according to actual conditions; this embodiment uses the above five edge relationships as examples. For example, it can include mating edges and through edges.

[0074] From the set of structural nodes With relation edge set Assembly relationship diagram constituting station t .

[0075] For components with incomplete boundaries in CT images, such as fasteners whose boundaries are eroded by metal artifacts, their spatial extent can be determined by combining their relationship with the surrounding structure. .

[0076] Whether the component boundary is intact is determined based on the occlusion area: if the projection support of the area where the component is located... If the value is below a preset threshold, meaning the area is within an occluded region, then its boundary is determined to be incomplete.

[0077] The reverse determination of the spatial range is shown in equation (3):

[0078] in, This represents the candidate space range for the i-th component, i.e., the voxel set; This indicates the optimal spatial range obtained by reverse determination; This represents the set of adjacent node indices that have an assembly relationship with the i-th component; This represents the j-th adjacent node; For relational compatibility functions, used to measure the candidate range With adjacent nodes The degree to which the edges satisfy a given relationship type (contact, coaxial, clamping, etc.) are connected; the higher the degree of satisfaction, the larger the value. This is a morphological regularization function used to penalize candidate ranges that deviate from the standard geometry of the component (given by the CAD drawing); is the weighting coefficient for morphological regularization, and is a preset non-negative real number.

[0079] Taking fasteners as an example, even if the boundary of the bolt shank is contaminated by artifacts, this method can still reverse the position and range of the bolt based on the continuous relationship combination of "hole position-washer-nut-bearing surface" (hole position and shank are coaxial, washer and nut end face are in contact, nut and bearing surface are pressed together), thereby reducing the missegmentation caused by metal artifacts; if only contour segmentation is relied upon, the missing boundary at the artifact erosion point will directly lead to the fastener being missed or the range being incorrect.

[0080] Step S2: Generate a fitted qualified CT image of the current assembly state based on the historical qualified vehicle sample library.

[0081] The purpose of this step is to construct a qualified benchmark that is consistent with the individual conditions of the vehicle being tested, that is, "the qualified CT state that the vehicle should present while maintaining its own individual differences", rather than directly subtracting from a fixed standard image.

[0082] Step S2 includes the following steps: Step S21: Using vehicle identification v and current workstation number t as indexes, retrieve K qualified samples from the historical qualified vehicle sample database Q(m). Each qualified sample contains its 3D CT image at workstation t and the corresponding assembly relationship diagram.

[0083] Wherein, the historical qualified vehicle sample library Q(m) represents the qualified sample library corresponding to model m, and its construction method is shown in Example 2. K is a preset positive integer, for example, it can be selected within this range according to the sample library size and work station cycle requirements: the larger K is, the more robust the qualified group reference is, but the computational amount of retrieval and fusion increases accordingly.

[0084] Step S22: Assemble according to the current workstation's assembly list The 3D CT image of each qualified sample is decomposed into three parts: the basic structure of the vehicle body, the newly added structure at the current workstation, and the connecting transition area. The basic body structure refers to the part that has been assembled in the previous workstation. The current workstation has a newly added structure, namely The areas corresponding to the listed components; The connecting transition region, i.e., the outward expansion radius of the contact area between the new structure and the parent structure. The area formed The preset length is, for example, 5 to 15 millimeters. The transition area is used to ensure a natural connection between the new structure and the parent structure after migration; if no transition area is set, abrupt changes in grayscale and geometry will appear at the contact interface between the migrated content and the measured content, and this seam will appear as a pseudo-difference in the subsequent residual calculation.

[0085] Step S23: Select a set of stable structures from the 3D CT images of the tested vehicle. As a positioning framework.

[0086] A stable structure refers to a structure where no assembly changes occur at the current workstation and the average reliability of the region is higher than the threshold. The structure includes components such as longitudinal beams, main positioning holes, and suspension brackets.

[0087] Step S24: Based on the positioning skeleton, for each qualified sample k, solve the spatial transformation that maps it to the measured vehicle coordinate system, and register the qualified sample to the measured 3D CT image according to the spatial transformation. The measured vehicle coordinate system refers to the vehicle coordinate system of the 3D CT image of the vehicle being measured at the current workstation.

[0088] The spatial transformation is obtained as shown in equation (4): ; in, The spatial transformation from the k-th qualified sample to the measured vehicle coordinate system is represented by a layered approach, consisting of first rigid transformation and then local elastic transformation using B-spline. The solution represents the optimal transformation obtained; x represents the set of measured vehicle stability structures. Spatial points in; This represents the structural point with the same name as x in the kth qualified sample. The correspondence is determined by feature matching of stable structures. Represents the spatial gradient of the transformed field. Used to constrain the smoothness of the transformed field and prevent drastic non-physical deformation; The smoothing regularization weight coefficient is a preset non-negative real number.

[0089] After obtaining the spatial transformation from the k-th qualified sample to the measured vehicle coordinate system, the qualified sample can be transformed back to the measured vehicle coordinate system. In this way, all qualified samples can be spatially aligned with the measured vehicle.

[0090] Step S25: Take the median density at each voxel position of the K registered qualified samples to obtain the qualified population reference value; compare the measured 3D CT image with the qualified population reference value voxel by voxel, and identify samples with deviations exceeding the tolerance band. And the credibility c(x) is not lower than The area is marked as a suspected abnormal area.

[0091] in, This is the allowable deviation magnitude determined jointly by the manufacturing tolerance and imaging noise level of the region. The median density is used instead of any single sample because the median is not sensitive to the individual characteristics of individual samples and can represent the common state of the qualified population.

[0092] Step S26: Using the positioning skeleton as the positioning basis, migrate the newly added structure and connecting transition area of ​​the current station of the qualified sample to the actual vehicle coordinates, and fuse it with the actual image to generate a fitted qualified CT image. In step S26, for stable structural regions and regions not marked as suspected abnormalities, the contents of the measured CT images are retained; For regions marked as suspected anomalies, the corresponding content of the registered qualified samples is used to replace them according to the fusion weight.

[0093] The migration process is shown in equation (5): ; in, This represents the value at position x of a properly fitted CT image; The value of the measured CT image at position x is represented; M(x) is the preservation mask function, which takes the value of 0 or 1: For stable structural regions and regions not judged as suspected anomalies, M(x) = 1, that is, the measured content is completely preserved, thereby preserving the vehicle's own manufacturing tolerances, attitude deviations and material differences; only for local regions suspected of being abnormal, M(x) = 0, which is replaced by qualified sample content; Represents the three-dimensional CT image of the k-th qualified sample; This means that the transformation obtained according to equation (4) maps position x to the corresponding position after the kth qualified sample coordinate system; The fusion weight of the kth qualified sample is determined by normalizing the similarity between the sample and the measured vehicle in the stable structural region. The higher the similarity, the greater the weight, and the sum of the K weights is 1.

[0094] Since the replacement content is supported by multiple qualified samples rather than given by a single template, the problem of individual characteristics of a single sample being mistakenly taken as the standard is avoided; and since all measured content is retained in non-abnormal areas, the vehicle's legitimate tolerances and attitude differences will not enter the subsequent residuals, fundamentally eliminating the pseudo-defects caused by using a fixed template for subtraction.

[0095] refer to Figure 2 Step S3: Perform bidirectional residual calculation and propagate along the assembly relationship diagram to obtain the defect confidence level of each structural node.

[0096] Specifically, step S3 includes the following steps: Step S31: Remove the contents of the fitted qualified CT image from the measured CT image to obtain the positive residual; remove the contents of the measured CT image from the fitted qualified CT image to obtain the negative residual, as shown in Equations (6) and (7) respectively: ; ; in, The positive residual at position x reflects the extra material or structure in the measured image. Its clustered area may correspond to anomalies such as misassembly, foreign objects, material accumulation, and extra gap filling. The reverse residual at position x reflects the missing material or structure in the measured image that is in a relatively qualified state. Its clustered area may correspond to anomalies such as missing parts, missing materials, incomplete assembly, and interrupted connection. The operation represents taking the non-negative part; c(x) is the confidence level defined by equation (2), used to suppress the residuals in uncertain regions, so that image differences in low confidence regions are not directly regarded as evidence of defects; This represents the value at position x of a properly fitted CT image; This represents the value at position x in the measured CT image.

[0097] Step S32: For each structural node u, perform volume integrals and normalize the forward and reverse residuals within its spatial range to obtain the node image residual energy. For each relation edge e, calculate the edge change. .

[0098] Specifically: Contact edge interruption record Coaxial edges are normalized according to the ratio of axial offset to eccentricity threshold; compaction edges are normalized according to the degree of weakening of compaction characteristics; sealing edges are normalized according to the proportion of sealing path fracture length; electrical connection edges are normalized according to the amount of insufficient terminal mating depth; edges that remain unchanged. .

[0099] Step S33: Propagate the positive and negative residuals along the relationship between the components to obtain the defect confidence level of each structural node; Specifically, if a residual exists only in a single voxel aggregation region and does not disrupt any assembly relationship, that is, the change in all relation edges of the node to which it is located is below the threshold, then its defect confidence is reduced; if the residual simultaneously causes contact relationship interruption, coaxial relationship offset, sealing path breakage, or fastening force transmission path abnormality, then its defect confidence is increased.

[0100] The defect confidence level of structural node u is calculated according to equation (8): ; Where D(u) represents the defect confidence level of structural node u; The image residual energy of node u is obtained by aggregation in step S32 and normalized to [0,1]; E(u) represents the set of relation edges connected to node u; This represents the change in relation edge e, with a value range of [0,1]. The importance weight of the relation edge e is indicated by the preset weights based on the edge type. For example, the weights of fastening force transmission paths and sealing paths are higher than those of general contact paths, and the weights of each edge at each node are normalized. , , respectively, are the fusion coefficients of image evidence and relational evidence, and are preset non-negative real numbers whose sum is 1. When D(u) is greater than or equal to the defect determination threshold... When node u and its residual region are identified as defect candidates.

[0101] This process transforms ordinary image grayscale differences into evidence of defects with assembly implications: the significance of a difference depends on whether it disrupts the assembly relationship; for example, a grayscale difference of the same volume, if isolated in a non-functional area, has its defect confidence reduced, but if it happens to break a sealing edge, its defect confidence is significantly increased.

[0102] Step S4: Determine the structural reliability based on the defect reliability of each structural node, generate a structural reliability rendering map, generate the probability of assembly error conditions based on the assembly sequence, and generate an alarm for assembly errors.

[0103] Specifically, step S4 includes the following steps: Step S41: Obtain the node defect confidence level of all structural nodes and generate a confidence level distribution map.

[0104] The structural reliability of structural node u is calculated according to equation (9): ; Where S(u) represents the structural reliability of structural node u, with a value range of [0,1], representing the reliability of the normal assembly state of the component corresponding to the node; D(u) is the node defect reliability obtained from equation (8). The higher the defect reliability of a node, the lower its structural reliability.

[0105] The credibility distribution map is generated according to equation (10): ; in, This represents the value of the confidence distribution plot at voxel position x; This represents the spatial extent of structural node u, i.e., the set of voxels occupied by the component. Let be the set of structural nodes in the assembly relationship diagram at station t; for background voxels that do not belong to any structural nodes. Take 1. (The rest of the text appears to be a list of steps or instructions and doesn't translate directly.) The structure confidence rendering map is obtained by overlaying and rendering the preset color mapping (e.g., the color changes from green to red corresponding to the structure confidence level from high to low) onto the 3D CT image of the actual CT.

[0106] In this way, the assembly reliability of each component is presented intuitively in the form of a three-dimensional color map, and production line personnel can directly locate low-reliability components without reading CT slices layer by layer; if only a list of values ​​is output, personnel need to compare material codes and spatial positions item by item, and it is impossible to complete the assessment within the cycle time.

[0107] Step S42: Obtain the assembly sequence of structural nodes, and generate the probability of occurrence of several corresponding assembly error conditions based on the defect confidence of each structural node.

[0108] For each structural node u, determine its predecessor node set Pre(u), which is the set of structural nodes that are installed before node u in the assembly order and have a relationship edge with node u.

[0109] The confidence level of the defect after sequential correction is calculated according to formula (11): ; in, The defect confidence level of node u after the assembly sequence is corrected is represented by the value range [0,1]; v represents the structural node in the set of preceding nodes Pre(u); D(u) and D(v) are the defect confidence levels of nodes u and v obtained according to equation (8), respectively. The sequence propagation coefficient has a value range of [0,1] and is a preset real number used to control the propagation intensity of defects from preceding nodes to subsequent nodes. This represents a chain multiplication operation; when Pre(u) is an empty set, the chain multiplication term takes the value 1, i.e. .

[0110] If we consider the defect confidence level of each node as the probability of its existence as an assembly anomaly, then any anomaly in node u itself, or any anomaly in any of its preceding nodes that propagates through the assembly process, will increase the defect confidence level of node u after correction. This is because assembly is performed sequentially, level by level. If the first installed reference structure is abnormal (e.g., misaligned reference surface or misaligned positioning holes), it will affect the assembly quality of subsequently installed components. Without sequential correction, this risk propagating along the assembly sequence cannot be reflected, and the risk of later-installed components will be underestimated.

[0111] Then calculate the probability of each assembly error condition. Pre-set Q types of assembly error conditions, such as omission, incorrect assembly, incomplete assembly, abnormal tightening, welding defects, sealing defects, and foreign object residue. The types and number of conditions can be set according to the actual situation of the production line.

[0112] For each working condition q, determine the set of associated nodes according to its evidence pattern. Association weight : For example, the set of structural nodes whose evidence pattern matches the working condition q, such as nodes whose inverse residuals associated with missing parts cover the entire range of the parts and whose corresponding relationship edges are missing, nodes whose inverse residuals associated with incomplete assembly show a local wedge-shaped distribution and whose coaxial edges or clamping edges are offset, and nodes whose positive residuals associated with foreign matter residue show an isolated high-density block and do not belong to any parts in the assembly list. The degree of matching between the evidence pattern of node u and the working condition q is calculated by a preset rule according to the combination pattern of positive residual, negative residual and relation edge change, with a value range of [0,1].

[0113] The probability of condition q occurring is calculated according to equation (12): ; in, Let q represent the probability of the q-th assembly error condition occurring, with a value range of [0,1]; q = 1,2,...,Q; Represents the set of structural nodes associated with operating condition q; This represents the association weight of node u with operating condition q; The defect confidence level is obtained after order correction by equation (11).

[0114] Equation (12) is also an "or" type probability synthesis: any node related to working condition q with high defect confidence will significantly increase the probability of the occurrence of the working condition, and the evidence of multiple nodes accumulates monotonically; the probability of occurrence of each working condition is calculated independently, allowing the same vehicle to report multiple working conditions at the same time, avoiding the underreporting caused by mutual exclusion between different working conditions.

[0115] Step S43: Generate an assembly error alarm based on the probability of the assembly error occurring.

[0116] Alarm levels are determined according to formula (13): ; Where A represents the alert level; This represents the maximum probability of occurring among the Q assembly error conditions; , To preset alarm thresholds, satisfy ; This is a stoppage-level alarm, triggering a production line shutdown and notifying a repair request. The vehicle is entering the re-inspection queue as a re-inspection alert is triggered. At the record level, only the results are written to the quality record without interrupting the production line.

[0117] Example 2: Construction of a historical qualified vehicle sample database.

[0118] This embodiment illustrates the construction method of the historical qualified vehicle sample library Q(m) in Embodiment 1. The samples in the sample library provide prior sources for reconstruction compensation (Equations (1-3), (1-4)) and the generation of fitted qualified CT images (step S2). Their quality directly determines the reliability of the qualified benchmark. Therefore, the entry, organization, and maintenance of the samples are implemented according to the following steps: Step A1: Acquire the CT process file of each vehicle on the production line in real time and obtain multiple candidate samples.

[0119] For each candidate sample, the following condition must be met: (1) No defect candidates were detected in the monitoring of all T inspection stations for the vehicle, that is, the defect confidence D(u) of all structural nodes at each station was lower than the defect judgment threshold. ; (2) The vehicle passed the final inspection after it rolled off the production line; (3) The vehicle was manually checked and confirmed to have no assembly abnormalities according to the preset sampling ratio.

[0120] Step A2: Organize candidate samples using workstation number t as the index key; each candidate sample record contains the 3D CT image of the vehicle at the corresponding workstation, the confidence field c(x), and the assembly relationship diagram. and assembly list .

[0121] Step A3: Register all samples to the unified reference coordinate system of the vehicle model using stable structures (vehicle longitudinal beams, main positioning holes, etc.); check the credibility field of each sample and remove samples or local data blocks whose regional average credibility is lower than the preset threshold.

[0122] Unified coordinates allow the K samples retrieved in step S21 to directly enter the hierarchical registration of equation (4) without needing to perform coarse alignment one by one; quality control prevents samples with severe artifacts from entering the fusion process and avoids low-quality prior contamination of the qualified benchmark.

[0123] Step A4: The sample library is updated on a rolling basis: new qualified vehicle samples are continuously added to the library, and samples that have exceeded the preset retention period are deactivated.

[0124] Furthermore, when process changes or component design changes occur (e.g., component replacement), the samples under the affected index keys are deactivated and re-accumulated. Before the sample size is restored, the search quantity K can be reduced and the proportion of manual review increased as a transitional measure. When samples are insufficient in the early stages of production line commissioning, an initial sample library can be established using vehicles that have undergone offline high-precision CT scanning during the trial production phase and have been manually verified as qualified. In this way, the sample library evolves synchronously with the production line process status, ensuring the timeliness of prior information. Without change management, old samples before the process change will mistakenly identify the normal new structure after the change as "abnormal," causing batch false alarms.

[0125] Example 3: Online monitoring system for vehicle assembly defects based on CT images.

[0126] This embodiment provides a monitoring system for implementing the above method, including: The CT acquisition device is arranged at the inspection station of the assembly line. The CT acquisition device includes an X-ray source, a flat panel detector, a circular track or a dual-robot cooperative motion mechanism, a triggering unit synchronized with the conveyor line, and a radiation protection chamber. The X-ray source and the flat panel detector are existing technologies, and their specific structures will not be further described in this application. Data storage server, housing process archives and historical qualified vehicle sample database; An algorithm server is used to run the online vehicle assembly defect monitoring method based on CT images.

[0127] The above are merely preferred embodiments of this application and are not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

Claims

1. A method for online monitoring of vehicle assembly defects based on CT images, characterized in that, Includes the following steps: Step S1: Collect CT projection data of the vehicle passing through each workstation, establish a CT process file for the vehicle, reconstruct the projection data for the assembly structure to obtain a three-dimensional CT image, and convert the reconstructed three-dimensional CT image into an assembly relationship diagram. Step S2: Generate a fitted qualified CT image of the current assembly state based on the historical qualified vehicle sample library; Step S3: Perform bidirectional residual calculation between the measured CT image and the fitted qualified CT image, and propagate the residual along the assembly relationship diagram to obtain the defect confidence of each structural node; Step S4: Determine the structural reliability of each structural node based on the defect reliability of each structural node, generate a structural reliability rendering map based on the structural reliability, generate the probability of occurrence of several assembly error conditions in combination with the assembly sequence, and generate an assembly error alarm based on the probability of occurrence.

2. The online monitoring method for vehicle assembly defects based on CT images according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: When the vehicle moves along the assembly line to the t-th inspection station, collect the projection data of that station. It also obtains the vehicle identification number, the current workstation number t, and the assembly list for the current workstation. Assembly sequence information; Step S12: Analyze and reconstruct the projection data of each workstation, identify occlusion and artifact regions, and obtain the occlusion weight matrix W; Step S13: Determine the currently visible structure and corresponding parts based on the occlusion weight matrix W. Input the structural information reconstructed from the previous workstation projection data, the geometric priors of the currently visible parts, and qualified samples into the reconstruction process in the form of regularization terms, together with the measured projection data of the current workstation, to construct and solve the reconstruction objective function to obtain the three-dimensional CT image. Step S14: For each voxel position x in the 3D CT image, calculate the confidence level c(x) to establish a confidence field; Step S15: For the 3D CT image with credibility information, treat each component as a structural node, extract the edges between the components, and construct the assembly relationship diagram for the t-th workstation. , Represents a set of structural nodes. Represents the set of relation edges.

3. The online monitoring method for vehicle assembly defects based on CT images according to claim 2, characterized in that, The occlusion weight matrix W is for the projection data. Each ray in the process is assigned a weight to form the ray.

4. The online monitoring method for vehicle assembly defects based on CT images according to claim 2, characterized in that, The reconstruction objective function is: ； Where V represents the three-dimensional CT image to be reconstructed, and the value V(x) of the three-dimensional CT image at spatial location x is the linear attenuation coefficient of that voxel; This represents the optimal 3D CT image obtained from the reconstruction; A represents the forward projection matrix of the CT system, which describes the physical mapping from the 3D CT image to the projection data. represents the projection data vector obtained from the actual measurement at station t; W represents the occlusion weight matrix obtained in step S12. This indicates element-wise multiplication. This represents the square of the second norm of a vector; , , These are the weight coefficients of the three regularization terms, all of which are preset non-negative real numbers used to balance the influence of the data fidelity term and each prior term; This is a regularization term for the consistency of the preceding workstation structure. For the geometric prior regularization terms of the components; This is a regularization term for qualified samples.

5. The online monitoring method for vehicle assembly defects based on CT images according to claim 1, characterized in that, Step S2 includes the following steps: Step S21: Using vehicle identification v and current workstation number t as indexes, retrieve K qualified samples from the historical qualified vehicle sample database Q(m). Each qualified sample contains its three-dimensional CT image at workstation t and the corresponding assembly relationship diagram. Step S22: Assemble according to the current workstation's assembly list The 3D CT image of each qualified sample is decomposed into three parts: the basic structure of the vehicle body, the newly added structure at the current workstation, and the connecting transition area. Step S23: Select a set of stable structures from the 3D CT images of the tested vehicle. As a positioning framework; Step S24: Based on the positioning skeleton, for each qualified sample k, solve the spatial transformation that maps it to the measured vehicle coordinate system, and register the qualified sample to the measured 3D CT image according to the spatial transformation. The measured vehicle coordinate system refers to the vehicle coordinate system of the 3D CT image of the vehicle being measured at the current workstation; Step S25: Take the median density at each voxel position of the K registered qualified samples to obtain the qualified population reference value; compare the measured 3D CT image with the qualified population reference value voxel by voxel, and identify samples with deviations exceeding the tolerance band. And the credibility c(x) is not lower than The area is marked as a suspected abnormal area; Step S26: Using the positioning skeleton as the positioning basis, migrate the newly added structure and connecting transition area of ​​the current station of the qualified sample to the actual vehicle coordinates, and fuse it with the actual image to generate a fitted qualified CT image.

6. The online monitoring method for vehicle assembly defects based on CT images according to claim 5, characterized in that, In step S26, for stable structural regions and regions not marked as suspected abnormalities, the contents of the measured CT images are retained; For regions marked as suspected anomalies, the corresponding content of the registered qualified samples is used to replace them according to the fusion weight.

7. The online monitoring method for vehicle assembly defects based on CT images according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Remove the contents of the fitted qualified CT image from the measured CT image to obtain the positive residual; remove the contents of the measured CT image from the fitted qualified CT image to obtain the inverse residual; Step S32: For each structural node u, perform volume integrals and normalize the forward and reverse residuals within its spatial range to obtain the node image residual energy. For each relation edge e, calculate the edge change. ; Step S33: Propagate the positive and negative residuals along the relationship between the components to obtain the defect confidence level of each structural node.

8. The online monitoring method for vehicle assembly defects based on CT images according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Obtain the node defect confidence level of all structural nodes and generate a confidence level distribution map; Step S42: Obtain the assembly sequence of structural nodes, and generate the probability of occurrence of several corresponding assembly error conditions based on the defect confidence of each structural node. Step S43: Generate an assembly error alarm based on the probability of the assembly error occurring.

9. The online monitoring method for vehicle assembly defects based on CT images according to claim 1, characterized in that, The construction of a historical qualified vehicle sample database includes the following steps: Step A1: Acquire the CT process file of each vehicle on the production line in real time and obtain multiple candidate samples; Step A2: Organize candidate samples using workstation number t as the index key; each candidate sample record contains the 3D CT image of the vehicle at the corresponding workstation, the confidence field c(x), and the assembly relationship diagram. and assembly list ; Step A3: Register all samples to the unified reference coordinate system of the vehicle model with a stable structure; check the confidence field of each sample and remove samples or local data blocks with an average regional confidence level lower than the preset threshold. Step A4: The sample library is updated on a rolling basis: new qualified vehicle samples are continuously added to the library, and samples that have exceeded the preset retention period are deactivated.

10. An online monitoring system for vehicle assembly defects based on CT images, characterized in that, include: The CT acquisition device is installed at the inspection station on the assembly line. Data storage server, housing process archives and historical qualified vehicle sample database; An algorithm server is used to run the online vehicle assembly defect monitoring method based on CT images as described in any one of claims 1 to 9.

Images