Finished automobile CT scanning image internal defect analysis system based on neural network model

Abstract

The invention relates to the technical field of mode recognition of vehicle internal defects, in particular to a whole vehicle CT scanning image internal defect analysis system based on a neural network model, and the system comprises an industrial CT device which is used for carrying out the CT scanning of a to-be-detected vehicle, and generating a three-dimensional scanning image representing each part of the vehicle; the part feature extraction device is used for extracting structural features of each part from the three-dimensional scanning image, and the structural features comprise a surface topological structure of each part and geometric features of each surface; according to the technical scheme, compared with traditional vehicle detection, structural features and position features of all parts in the vehicle are obtained through the CT scanning technology, and therefore the structural features and the position features of all the parts in the vehicle are obtained through the CT scanning technology; the number of the structural features can analyze whether the part is true or not, the structural features can analyze whether the part has cracks and hidden damage or not, and the position features can analyze whether the installation position and the installation direction of the part have errors or not.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of pattern recognition of vehicle internal defects, in particular to a whole vehicle CT scanning image internal defect analysis system based on a neural network model. BACKGROUND

[0002] The content of this section is only to provide background information related to the present application, which may not constitute prior art.

[0003] Automobile manufacturing is a highly developed industry. After the completion of vehicle manufacturing, strict defect detection is required. The current mainstream vehicle detection method mainly includes:

[0004] Actual driving detection: through actual driving test, identify problems that affect the normal use of the vehicle or have significant performance faults.

[0005] External appearance detection: used to find defects on the surface of the vehicle appearance.

[0006] Software data detection: by reading the fault codes feedback by the internal sensors of the vehicle, diagnose the internal electronic or mechanical fault information of the vehicle. The above detection methods essentially focus on finding problems that affect user experience or have obvious fault characteristics. However, they are difficult to effectively identify and judge subtle defects at the internal structural level of the vehicle in practice, for example: whether the various parts, connecting pieces, wire harnesses, engines, etc. that make up the vehicle have misplacement or omission in the installation position; whether the installation angle of the key components meets the design requirements; whether the parts themselves have small cracks or damage; whether the assembly tolerance between the parts exceeds the allowed range; whether some non-core connecting pieces or reinforcing pieces are missing.

[0007] Because the existing detection means cannot cover these deep and structural installation and manufacturing defects, it is difficult to ensure the quality consistency of the vehicle. Some vehicles with internal part misplacement, omission or poor installation are allowed to leave the factory. After a period of use, such hidden dangers often manifest as vehicle body noise, fault rate increase and other problems, seriously damaging user experience and reducing product reliability. SUMMARY

[0008] The content section of the present application is used to introduce the concept in a brief form, which will be described in detail in the specific embodiments section. The content section of the present application is not intended to identify key features or essential features of the claimed technical solution, nor is it intended to limit the scope of the claimed technical solution.

[0009] Some embodiments of the present application propose a whole vehicle CT scanning image internal defect analysis system based on a neural network model to solve the technical problems mentioned in the background section.

[0010] As a first aspect of the present application, some embodiments of the present application provide a neural network model-based vehicle CT scan image internal defect analysis system, comprising:

[0011] An industrial CT device is configured to perform CT scanning on a vehicle to be tested to generate three-dimensional scan images representing parts of the vehicle;

[0012] A part feature extraction device is configured to extract structural features of each part from the three-dimensional scan images, wherein the structural features include surface topological structures of each part and geometric features of each surface;

[0013] A position feature extraction device is configured to generate a plurality of position anchors in the three-dimensional scan images, calculate relative positional relationships between each part and the position anchors, and generate position features of each part;

[0014] An information aggregation device is configured to fuse the position features and the structural features of each part to generate fused features of each part;

[0015] A fault probability extraction device is configured to compare the fused features of each part with standard fused features to generate fault probabilities of installation defects of each part;

[0016] A defect analysis device is configured to extract the fault probabilities of installation defects of each part to generate an internal defect risk level of the vehicle.

[0017] In the technical solution provided by the present application, compared with traditional vehicle detection, CT scanning technology is used to obtain structural features and position features of all parts in the vehicle. The number of structural features can be used to analyze whether the parts are genuine, the structural features themselves can be used to analyze whether the parts have cracks and hidden damage, and the position features can be used to analyze whether the installation position and installation direction of the parts are incorrect. In this way, the accuracy of vehicle internal defect determination is greatly improved. Meanwhile, in the present solution, a plurality of position anchors are used to calculate the position features of each part. Therefore, the relative angle of each part relative to the position anchor can be enlarged by using different position anchors, so that the installation angle of each part can be more accurately determined, the angle error of each part can be enlarged, and the accuracy of risk determination can be improved.

[0018] The industrial CT device comprises:

[0019] A scanning information acquisition device is configured to acquire original scan images of the vehicle to be tested in the CT;

[0020] A part segmentation device is configured to segment the original scan images based on VGStudio MAX to obtain three-dimensional scan images of all parts;

[0021] Part number device, artificial calibration of several marked parts, based on the position relationship of adjacent parts, automatically generate information coding of all parts.

[0022] In the technical scheme provided in the application, the original CT scan image of the complex part cluster is automatically and high-precision segmented by using VGStudio MAX, and the three-dimensional scan image of each independent part is reliably obtained. By manually calibrating a small amount of key marked parts as "seeds", the system can intelligently deduce and automatically generate unique information coding of all adjacent parts, thereby significantly reducing the workload and subjective error risk of manual numbering.

[0023] The existing part defect detection method based on CT scanning generally adopts a surface feature comparison method to determine whether there is a defect, which needs to identify whether the parts are the same part, and then identify whether the faces of the parts correspond. In this way, it is difficult to efficiently and accurately identify the local and subtle defects (such as small cracks and deformation) on the surface of the part, and establish a stable and comparable corresponding relationship between the surface of the part to be detected and the standard part.

[0024] The part feature extraction device comprises:

[0025] The serial number generation module extracts all surfaces of the part from the three-dimensional scan image of the part, and arranges and generates the surface serial numbers according to the areas of the surfaces;

[0026] The topological structure extraction module constructs the adjacent surface topological structure H i of the part i based on whether the surfaces are adjacent; the surface topological structure H i is an adjacency matrix of each surface j of the part i; in H i , the serial numbers in the rows are the surface serial numbers of the part i, in H i , the serial numbers in the columns are the surface serial numbers of the part i, and in H i , the elements are j represents the index of the surface serial number, , and the surface with the surface serial number j is adjacent to the surface with the surface serial number k.

[0027] The geometric feature extraction module extracts the geometric properties of each surface j of the part i , generates subsets of the geometric properties according to the surface serial numbers, and collects all the subsets of the geometric properties to generate the geometric features R i ; the geometric properties include the number of edges, the peripheral perimeter, the area of the surface, the surface curvature, and R i represents the geometric features of the i-th part, represents the subset of the geometric properties of the j-th surface of the i-th part; represents the geometric properties of the j-th surface of the i-th part;

[0028] a feature aggregation module, the surface topology H i and the geometric feature R i are respectively merged as the structural feature HR of the part as a node information i .

[0029] The technical scheme provided in the application innovatively decomposes the surface information of a part into a topology and a geometric feature, two complementary dimensions respectively capturing the connection relationship (adjacency matrix) between surfaces and the morphological parameters (such as area, perimeter, and curvature) of a single surface, so that the system can independently and sensitively detect the topology structure abnormality (such as newly added / missing surface) or geometric feature distortion (such as area reduction and curvature change) caused by defects such as cracks and deformation; the strategy of generating a surface serial number in the order of surface area is adopted, the uniqueness of the area distribution of the surface of a part (few asymmetric surfaces have equal areas) is fully utilized, and stable and efficient "surface-surface" direct correspondence between the surface of a part to be inspected and the surface of a standard part can be realized without complex surface recognition algorithms, greatly simplifying the comparison process. This set of methods significantly enhances the detection capability and judgment accuracy of part surface defects (processing flaws and deformation).

[0030] The existing vehicle part installation position detection method is difficult to accurately determine whether the actual installation position and angle of the part in the vehicle coordinate system deviate from the design requirements due to the lack of a unified and stable spatial reference datum and effective quantitative features, especially the sensitivity to subtle position deviation and angle deviation.

[0031] Further, the position feature extraction device comprises:

[0032] An anchor point position generation module, at least five position anchors are calibrated by a person in a three-dimensional scanning image, and the five position anchors are respectively located at a center position q1 of the vehicle to be measured, a front left wheel axis center position q2 of the vehicle to be measured, a front right wheel axis center position q3 of the vehicle to be measured, a rear left wheel axis center position q4 of the vehicle to be measured, and a rear right wheel axis center position q5 of the vehicle to be measured;

[0033] A part vector generation module, obtaining a three-dimensional scanning image and information code of a part, and extracting a vector between the longest two end points on the part from the three-dimensional scanning image The vector is respectively calculated with the vector The vector The vector The vector The vector and the vector The included angle is calculated to obtain The vector The shortest distance between any point in the vector and the positions q1, q2, q3, q4, and q5 is calculated to obtain Among them, Let O represent the vector connecting the two longest endpoints of part i in the three-dimensional coordinate system, and let O represent the position of the origin in the three-dimensional coordinate system. Representing vectors respectively with vector vector vector vector and vectors The included angle, Let q1, q2, q3, q4, q5 represent the vectors respectively. vector vector vector vector and vectors The shortest distance to any point on the [aspect].

[0034] The location feature generation module will E, as the positional feature of part i i .

[0035] This solution manually calibrates key anchor points (vehicle center and four-wheel axle centers) and extracts the longest end line vector of the part itself as a reference, constructing a composite positional feature that includes the angle between this vector and the vectors of each anchor point, as well as the shortest distance. This method eliminates the dependence on specific identification points of the part, simplifying feature construction. Utilizing a reference network composed of multiple anchor points significantly enhances the sensitivity and quantification accuracy for detecting subtle deviations in the spatial position and installation angle of the part, thereby accurately determining installation deviations.

[0036] When performing defect analysis by fusing structural features (HRi) and location features (Ei) of a part, existing methods fail to dynamically adjust the relative importance of the two types of features based on the part's own physical properties (such as size and shape), resulting in the fused features failing to accurately reflect the differentiated contributions of different parts to the overall vehicle defect risk.

[0037] Existing methods fail to dynamically adjust the relative importance of the two types of features based on the physical properties of the parts themselves, resulting in the fused features failing to accurately reflect the differentiated contributions of different parts to the overall vehicle defect risk.

[0038] Information aggregation devices include:

[0039] The part classification factor identification module generates a weighting factor α based on the 3D dimensions of the 3D scan image of part i. i ;

[0040] Among them, b i The vector representing the shortest line connecting the two endpoints of part i;

[0041] The feature weight generation module generates structural features HR for each part i.i Generate structural feature weights For the positional feature E of each part i i Generate location feature weights

[0042]

[0043] Where c represents the transformation coefficient, c > 0, and tanh() represents the hyperbolic tangent function;

[0044] The fusion feature generation module integrates structural features (HR). i Structural feature weights Location feature E i and location feature weights They are merged into a single set as the fusion feature.

[0045] This innovative approach introduces weighting factors based on the part's three-dimensional dimensions (volume) and aspect ratio (shortest end vector), and utilizes the hyperbolic tangent function to generate structural feature weights and positional feature weights. This allows the fused features to adaptively adjust their weights: for parts with large volumes, structural features are given higher weights; for parts with high aspect ratios, positional features are given higher weights.

[0046] The fault probability extraction device includes:

[0047] The standard feature acquisition module is used to save the fusion features of the various parts standards of the calibrated vehicle;

[0048] The comparison group generation module generates a comparison group by matching the standard fusion features and the fusion features of the vehicle under test one-to-one according to the information encoding.

[0049] Install the defect generation module, calculate the standard fusion features and the fusion features of the vehicle under test in each comparison group, and generate the failure probability of each part's installation defect.

[0050] The installation defect generation module includes:

[0051] The structural failure probability generator calculates the similarity between the structural features in the standard fusion features and the structural features in the fusion features of the vehicle under test, generating the first probability factor. This represents the first probability factor for part i;

[0052] The location fault probability generator calculates the similarity between the location features in the standard fusion features and the location features in the fusion features of the vehicle under test, and generates a second probability factor. The second probability factor represents part i; similarity is calculated based on a neural network model.

[0053] Fault probability calculator, based on structural feature weights Location feature weights First probability factor And the second probability factor Calculate the failure probability PD of component assembly defects i PD i This represents the failure probability of the i-th part in the vehicle under test;

[0054]

[0055] In the technical solution provided in this application, when calculating the probability of whether a part has an installation failure, it does not only rely on the similarity of the fused features, but also comprehensively evaluates the similarity, structural feature weights and position feature weights. In practice, this can reduce the impact of the structure of different parts on installation redundancy, and thus accurately characterize whether there is an installation defect.

[0056] Neural network models are more efficient at calculating similarity than directly calculating cosine similarity. A two-branch network with shared weights maps the standard fused features and the fused features of the test vehicle to the same embedding space, generating corresponding embedding vectors. The output network then outputs the similarity between the two features. During training, objective functions such as contrastive loss or triplet loss are used to force a reduction in the vector distance between similar samples and an increase in the distance between dissimilar samples, ultimately enabling the network to learn to output a semantically correct similarity metric. Therefore, when constructing samples, the standard fused features and the fused features of the test vehicle can be used as samples, and the similarity probability calculated using cosine similarity can be used as labels. By continuously training the neural network model, it can quickly generate similarity probabilities based on the input standard fused features and the fused features of the test vehicle.

[0057] The defect analysis device includes:

[0058] The parts arrangement module arranges all parts i in the vehicle under test according to their volume to form a parts sequence RT;

[0059] The parts extraction module randomly extracts multiple parts judgment groups from the parts sequence RT.

[0060] The risk level calculation module calculates the sum of the failure probabilities of each part judgment group in turn to generate risk level factors, and maps the highest risk level factor to the risk level.

[0061] The technical solution provided in this application does not consider the installation probability of all parts when calculating the risk level, thus reducing information redundancy for a large number of installed parts. By using information extraction, the information dimensionality is reduced, increasing computational efficiency. Furthermore, to ensure the randomness of the extraction and the representativeness of the extracted parts, the parts are arranged by volume, so that a corresponding number of parts can be extracted within different volume ranges.

[0062] Furthermore, the extraction method of the parts extraction module is as follows:

[0063] S1: Preset the number of parts to be sampled in the part judgment group D, preset the number of part judgment groups M, and preset the sampling uniformity index G. min ;

[0064] S2: Divide the sequence of parts RT into D intervals g, g = 1, 2, 3, ... D, and the range of interval g is [s g h g ], s g h represents the lower bound of the range of values ​​for interval g. g This represents the upper bound of the range of values ​​for interval g;

[0065] S3: Divide each interval g into M sampling ranges, with a sampling range length of Δ. g ; The kth extraction range I gk ;

[0066] I gk =[s g +(k-1)Δ g ,h g +kΔ g );

[0067] S4: Randomly select M parts P from each sampling range. gk M part determination groups are generated, and an initial matrix A is obtained.

[0068] P gk ~Uniform(s g +(k-1)Δ g ,h g +kΔ g Uniform indicates that a value is randomly selected.

[0069]

[0070] S5: Randomly arrange each row of the initial matrix A to generate matrix B;

[0071]

[0072] S5: Each column of matrix B is a part determination group, q j This represents the j-th column in matrix B;

[0073]

[0074] S6: Calculate the sampling uniformity index G of matrix B;

[0075] l represents the column index in matrix B;

[0076] S7: Calculate whether G is less than G min If G is less than G min Then, the columns of the M matrices B are used as the part judgment group. If G is greater than G... min Then reapply P gk Random values ​​are selected.

[0077] In the technical solution provided in this application, the number of parts RT is first divided into D intervals, and then each interval is divided into M sampling ranges. This ensures that the final sample is sufficiently uniform. In order to further determine whether the sample is sufficiently uniform, the sampling uniformity index G is calculated. The sampling quality is judged based on the sampling uniformity index G, thereby ensuring that the final generated part judgment group can cover the entire sample range as uniformly as possible.

[0078] Furthermore, based on the extracted component assembly groups, the structural and positional features of the corresponding components are selectively extracted, and the failure probability of the component is calculated.

[0079] In this solution, the structural and positional features of the corresponding parts are selectively extracted, and the failure probability of the parts is calculated, which can reduce the amount of data processing and increase the system processing speed.

[0080] The technical solution of this application embodiment has at least the following advantages and beneficial effects:

[0081] Deep defect detection capability: Utilizing industrial CT scanning technology, the system can acquire detailed 3D structural features (including surface topology and geometry) and precise location characteristics of all parts inside the vehicle. This enables the system to effectively identify deep-seated, structural installation and manufacturing defects that are difficult to access using traditional methods; for example:

[0082] Missing, misaligned, or minor damage (such as cracks or hidden defects) of parts;

[0083] Deviation in the installation angle / direction of key components;

[0084] The assembly tolerances between parts exceed the limits;

[0085] Omission of non-core connectors or reinforcements.

[0086] Precise positioning and analysis:

[0087] Structural feature analysis: directly judge the integrity of the part (whether it exists or is damaged) based on its structural features.

[0088] Position Feature Optimization: An innovative multi-position anchor point mechanism is employed to calculate the position features of the part. This method can:

[0089] Magnified angular deviation: By calculating the positional relationship of the part relative to multiple different anchor points, the sensitivity to subtle errors in the installation angle / direction of the part is significantly enhanced.

[0090] Improved positional accuracy: More accurately determine whether the installation position of parts meets design requirements.

[0091] Comprehensive risk quantification: By integrating structural features with optimized location features and comparing them with standard features, the system can quantify the probability of installation defects in each part and ultimately comprehensively assess the internal defect risk level of the entire vehicle. Attached Figure Description

[0092] Figure 1 This is a schematic diagram of a system for analyzing internal defects in whole-vehicle CT scan images based on a neural network model.

[0093] Figure 2 The original 3D scan image obtained for industrial CT scanning of the vehicle under test.

[0094] Figure 3 A 3D scan image of a single part.

[0095] Figure 4 This is a schematic diagram illustrating the generation of surface topology.

[0096] Figure 5 This is a schematic diagram of the geometric features.

[0097] Figure 6 For a certain part in the vehicle under test and A schematic diagram. Detailed Implementation

[0098] 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.

[0099] 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.

[0100] 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.

[0101] refer to Figure 1 Example 1: A vehicle CT scan image internal defect analysis system based on a neural network model includes: an industrial CT device, a part feature extraction device, a location feature extraction device, an information aggregation device, a fault probability extraction device, and a defect analysis device. The part feature extraction device and the location feature extraction device are respectively connected to the industrial CT device, and are respectively connected to the information aggregation device. The information aggregation device is connected to the fault probability extraction device, and the fault probability extraction device is connected to the defect analysis device.

[0102] Industrial CT equipment is used to perform CT scans on vehicles under test, generating three-dimensional scan images representing various parts of the vehicle. An industrial CT equipment includes: a scan information acquisition device, a parts segmentation device, and a parts numbering device.

[0103] The scanning information acquisition device is used to acquire the raw scan images of the vehicle under test within the CT scanner. The vehicle is transported to the industrial CT scanner via conveyor belt or similar means for scanning. After the industrial CT scan, the raw 3D image of the vehicle is obtained. Figure 2 As shown.

[0104] The parts segmentation device uses VGStudio MAX to segment the original scanned images and acquire 3D scanned images of all parts. The vehicle under test is a whole formed by connecting multiple parts together through riveting, welding, bolting, etc. Many of these connections are not integral. For example, bolted connections and fastening connections. Thus, through automatic segmentation and critical area growth in VGStudio MAX software, all parts are automatically segmented. These segmented parts refer to those parts that are identified as not being a single entity in VGStudio MAX software. Figure 3 As shown.

[0105] The part numbering device allows for the manual marking of several parts, and automatically generates information codes for all parts based on the positional relationships of adjacent parts.

[0106] Identifying all parts using machine learning to feature them is too difficult. Therefore, several parts can be manually labeled, and the connections between adjacent parts can be used to automatically identify all surrounding parts. After identifying all parts, corresponding information codes are generated for these parts according to pre-set rules. These information codes are essentially the part's serial number. Of course, if a specific part is found to be incorrect, the vehicle needs to have that part reinstalled before re-inspection.

[0107] The system includes a component feature extraction device, which extracts the structural features of each component from the 3D scanned image. The structural features include the surface topology of each component and the geometric features of each surface. A position feature extraction device generates several position anchor points in the 3D scanned image, calculates the relative positional relationship between each component and the position anchor points, and generates the position features of each component. An information aggregation device fuses the position features and structural features of the components to generate the fused features of each component. A fault probability extraction device compares the fused features of each component with the standard fused features to generate the fault probability of the installation defects of each component. A defect analysis device extracts the fault probability of the installation defects of each component to generate the risk level of internal defects in the vehicle.

[0108] The principle behind this solution is as follows: the vehicle under test is fed into an industrial CT scanner, which automatically scans the vehicle's internal structure. After scanning, automatic segmentation software is used to segment all the parts. Following both manual and automatic calibration, information codes for all parts are obtained, meaning all parts are identified and assigned information codes. During this process, the segmentation results are checked step-by-step. Parts that were not successfully segmented but were mistakenly joined together are manually re-marked, ensuring that every part in the vehicle under test is accurately identified and that the number of scanned parts matches the pre-set number.

[0109] The vehicle contains a large number of parts, and due to manufacturing errors, it is impossible for every part to be installed flawlessly. Therefore, this solution calculates the risk level of internal defects by comprehensively evaluating the failure probability of each part, reducing the information dimensionality, rather than relying on the installation error of one or two parts to trigger an alarm.

[0110] Example 2: Based on Example 1, Example 2 provides methods for extracting structural and positional features. Specifically:

[0111] The part feature extraction device includes: a serial number generation module, a topology extraction module, a geometric feature extraction module, and a feature aggregation module. The serial number generation module is connected to both the topology extraction module and the geometric feature extraction module, and the feature aggregation module is connected to both the topology extraction module and the geometric feature extraction module.

[0112] The serial number generation module extracts all surfaces of a part from its 3D scanned image and arranges them according to their area to generate a serial number. Each part has many faces; for example, a cube has six faces. The module reads the area of ​​each face and sorts them according to their area size to obtain a serial number for each face. Sorting by area size allows for direct identification of each face.

[0113] The topology extraction module constructs the adjacent surface topology H based on whether the surfaces are adjacent. i Surface topology H i H is the adjacency matrix of each surface j of part i; i The serial number in the middle is the face number of part i, H i The column number in the middle is the face number of part i, H i The elements in are j represents the index of the face number. This indicates that the surface with surface number j is adjacent to the surface with surface number k; for example Figure 4 As shown, Figure 4 This is the adjacency matrix of a cube after all its faces have been unfolded. As can be seen, the adjacency matrix indicates the connection method between all faces.

[0114] like Figure 5 As shown, the geometric feature extraction module extracts the geometric characteristics of each surface j of part i. And generate a subset of geometric properties based on the face number, and collect a subset of all geometric properties. Generate geometric feature R i Geometric properties include: number of edges, perimeter, surface area, surface curvature (mean curvature), and R. i Represents the geometric features of the i-th part. A subset representing the geometric characteristics of the j-th face of the i-th part; This represents the geometric characteristics of the j-th face of the i-th part.

[0115] The feature aggregation module aggregates the surface topology H i and geometric features R i Each node's information is merged into the structural feature HR of the part. i .

[0116] Thus, structural features HR i This includes the surface topology of adjacent faces, as well as the geometric features of each face, namely its perimeter, area, and degree of curvature. By comparing the structural features of the part, it is possible to accurately determine whether the part has deformed.

[0117] The location feature extraction device includes: an anchor point location generation module, a part vector generation module, and a location feature generation module. These modules are connected sequentially.

[0118] The anchor point generation module manually marks at least five anchor points in the 3D scanned image. These five anchor points are located at the center position q1 of the vehicle under test, the center position of the left front wheel axle q2, the center position of the right front wheel axle q3, the center position of the left rear wheel axle q4, and the center position of the right rear wheel axle q5. These five anchor points are the five most accurately calibrated positions on the vehicle. In practice, the number of anchor points can be increased or decreased.

[0119] The part vector generation module acquires the 3D scanned image and information encoding of the part, and extracts the vector of the longest line connecting the two endpoints of the part from the 3D scanned image. vector respectively with vector vector vector vector vector and vectors Calculate the included angle to obtain Calculate vectors The shortest distance between any point in the interval and positions q1, q2, q3, q4, and q5 is obtained. in, Let O represent the vector connecting the two longest endpoints of part i in the three-dimensional coordinate system, and let O represent the position of the origin in the three-dimensional coordinate system. Representing vectors respectively with vector vector vector vector vector and vectors The included angle, Let q1, q2, q3, q4, q5 represent the vectors respectively. vector vector vector vector and vectors The shortest distance to any point on the [aspect].

[0120] The location feature generation module will E, as the positional feature of part i i Positional features are essentially the distance and angle between each part and its anchor point. Based on these positional features, it is possible to determine whether there are any errors in the installation angle and method of the parts.

[0121] Example 3: The key to this application lies in how to extract, aggregate, and reduce the dimensionality of a large number of complex comparison results into a risk level that can reflect the overall vehicle quality. Example 3 is provided for this purpose.

[0122] The information aggregation device aggregates structural and positional features, while the fault probability extraction device extracts feature information characterizing component installation defects from the structural and positional features. The defect analysis device extracts risk level information from the feature information characterizing installation defects.

[0123] Specifically, the information aggregation device includes a part classification factor identification module, a feature weight generation module, and a fusion feature generation module. These modules are connected sequentially. The feature weight generation module is also connected to the feature aggregation module and the location feature generation module.

[0124] like Figure 6 As shown, the part classification factor recognition module generates a weighting factor α based on the 3D dimensions of the 3D scan image of part i. i ; in, The vector representing the shortest line connecting the two endpoints of part i; the feature weight generation module, for each part i's structural features HR i Generate structural feature weights For the positional feature E of each part i i Generate location feature weights Where c represents the transformation coefficient, c > 0, tanh() represents the hyperbolic tangent function, and V i Represents the volume of the i-th part; the fusion feature generation module combines the structural features HR... i Structural feature weights Location feature E i and location feature weights They are merged into a single set as the fusion feature.

[0125] The reason for generating two different structural weights in the technical solution provided in this application is to assign different weight information to each part during comparison. In practice, structural features and positional features consider two different installation defect risks of the parts. Structural features mainly examine whether there are defects on the surface of the part and whether the part is deformed. Positional features mainly examine whether the part is installed in the predetermined position and whether the installation angle of the part is appropriate. Different parts are examined in different ways. For example, some smaller connectors and bolts have a smaller impact on the overall quality of the vehicle, whether it is structural features or positional features, so both structural feature weights and positional feature weights are relatively small. However, the ratio of the longest length to the shortest width of a bolt is relatively large, and the corresponding positional feature weight will increase. That is, smaller parts are more likely to obtain a larger positional feature weight, and vice versa.

[0126] The fault probability extraction device includes: a standard feature acquisition module, a comparison group generation module, and an installation defect generation module. These modules are connected sequentially. The installation defect generation module is then connected to the fusion feature generation module.

[0127] The standard feature acquisition module is used to store the fused features of the various parts of the calibrated vehicle. The fused features of the standard refer to the feature information obtained after the aforementioned processing of a vehicle without installation defects.

[0128] The comparison group generation module generates a comparison group by matching the standard fusion features and the fusion features of the vehicle under test one-to-one according to the information encoding.

[0129] Each part i has an information code. By matching the information codes, each part can be matched one-to-one. Therefore, the two parts in a comparison group are the part on the test vehicle and the part at the same position on the calibration vehicle.

[0130] Install the defect generation module to calculate the standard fusion features and the fusion features of the vehicle under test in each comparison group, and generate the failure probability of each part's installation defect.

[0131] The defect generation module includes: a structural fault probability generator, a location fault probability generator, and a fault probability calculator. The fault probability calculator is connected to both the structural and location fault probability generators. Both the structural and location fault probability generators are connected to the fusion feature generation module. The structural fault probability generator is used to acquire standard fusion features and the fusion features of the vehicle under test. The location fault probability generator is used to acquire both standard fusion features and the fusion features of the vehicle under test.

[0132] The structural failure probability generator calculates the similarity between the structural features in the standard fusion features and the structural features in the fusion features of the vehicle under test, generating the first probability factor. This represents the first probability factor for part i;

[0133] The location fault probability generator calculates the similarity between the location features in the standard fusion features and the location features in the fusion features of the vehicle under test, and generates a second probability factor. The second probability factor represents part i; similarity is calculated based on a neural network model.

[0134] The similarity in this application is calculated using a neural network model. The main reason for using similarity is that it can measure the similarity between two features. A high similarity indicates that the structure and installation position of the parts have a high degree of overlap, and there may be fewer installation defects.

[0135] Fault probability calculator, based on structural feature weights Location feature weights First probability factor And the second probability factor Calculate the failure probability PD of component assembly defects i PD i This represents the failure probability of the i-th part in the vehicle under test;

[0136]

[0137] The failure probability calculator only indicates the probability of an installation defect occurring for that part. In vehicle design, each part's assembly has design redundancy; the existence of a failure probability does not necessarily indicate an error in the part's installation that affects vehicle quality. Only when a large number of parts exhibit high failure probabilities can a high installation risk be indicated. However, given the vast number of parts in a vehicle, comparisons are typically made with a selection of key parts. Therefore, this application provides the following technical solution:

[0138] The defect analysis device includes a parts arrangement module, a parts extraction module, and a risk level calculation module. These modules are connected sequentially. The risk level calculation module is connected to a failure probability calculator.

[0139] The parts arrangement module arranges all parts i in the vehicle under test according to their volume, forming a parts sequence RT. The reason for arranging them by volume is to ensure that each part has the same sequence number during each test, preventing errors. Additionally, sampling a portion of each volume allows for a better evaluation of the overall vehicle's assembly defects.

[0140] The parts extraction module randomly selects multiple parts judgment groups from the parts sequence RT.

[0141] When selecting parts for the decision group, the selected parts need to be evenly distributed within the part sequence RT to ensure uniformity in part selection. Therefore, the part extraction module uses the following extraction method:

[0142] S1: Preset the number of parts to be sampled in the part judgment group D, preset the number of part judgment groups M, and preset the sampling uniformity index G. min ;

[0143] S2: Divide the sequence of parts RT into D intervals g, g = 1, 2, 3, ... D, and the range of interval g is [s g h g ], s g h represents the lower bound of the range of values ​​for interval g. g This represents the upper bound of the range of values ​​for interval g;

[0144] S3: Divide each interval g into M sampling ranges, with a sampling range length of Δ. g ; The kth extraction range I gk ;

[0145] I gk =[s g +(k-1)Δ g ,h g +kΔ g );

[0146] S4: Randomly select M parts P from each sampling range. gk M part determination groups are generated, and an initial matrix A is obtained.

[0147] P gk ~Uniform(s g +(k-1)Δ g ,h g +kΔ gUniform indicates that a value is randomly selected.

[0148]

[0149] S5: Randomly arrange each row of the initial matrix A to generate matrix B;

[0150]

[0151] S5: Each column of matrix B is a part determination group, q j This represents the j-th column in matrix B;

[0152]

[0153] S6: Calculate the sampling uniformity index G of matrix B;

[0154] l represents the column index in matrix B;

[0155] S7: Calculate whether G is less than G min If G is less than G min Then, the columns of the M matrices B are used as the part judgment group. If G is greater than G... min Then reapply P gk Random values ​​are selected.

[0156] Therefore, the above scheme can extract the required component judgment groups as needed. This allows for feature reduction of complex fault information through extraction, thus lowering the feature dimensionality.

[0157] The risk level calculation module calculates the sum of the failure probabilities of each part judgment group in turn to generate risk level factors, and maps the highest risk level factor to the risk level.

[0158] Furthermore, the sum of failure probabilities for the same part determination group is calculated to generate a risk level factor. For example, if there are three part determination groups, and the calculated risk level factors are 15.6, 17.8, and 20.1 respectively, then the final risk level factor of 20.1 is used as the standard to map the risk level.

[0159] The mapping method for risk levels mainly involves setting a corresponding range of risk level factors for each risk level. For example, risk level 1 is 0-3, risk level 2 is 3-4, and so on.

[0160] Furthermore, in practice, after the part determination group is extracted, the structural features and positional features of the corresponding parts can be extracted, and the failure probability of the part can be calculated to reduce the amount of calculation.

[0161] 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 neural network model-based internal defect analysis system for a whole vehicle CT scan image, characterized by, The system comprises: an industrial CT device for CT scanning of a vehicle to be tested to generate three-dimensional scanning images representing parts of the vehicle; a part feature extraction device for extracting structural features of the parts from the three-dimensional scanning images, the structural features including surface topologies of each part and geometric features of each surface; a position feature extraction device for generating a plurality of position anchors in the three-dimensional scanning images, calculating relative positional relationships between the parts and the position anchors, and generating position features of the parts; an information aggregation device for fusing the position features and the structural features of the parts to generate fused features of the parts; a fault probability extraction device for comparing the fused features of the parts with standard fused features to generate fault probabilities of installation defects of the parts; a defect analysis device for extracting the fault probabilities of the installation defects of the parts to generate a risk level of internal defects of the vehicle. 2.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 1, characterized in that, The industrial CT device comprises: a scanning information acquisition device for acquiring original scanning images of the vehicle to be tested in a CT; a part segmentation device for segmenting the original scanning images based on VGStudio MAX to acquire three-dimensional scanning images of all the parts; a part numbering device for manually calibrating a plurality of marker parts, and automatically generating information codes of all the parts based on positional relationships of adjacent parts. 3.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 2, characterized in that, The part feature extraction device comprises: a serial number generation module for extracting all surfaces of a part from a three-dimensional scanning image of the part, and arranging the surfaces according to areas of the surfaces to generate surface serial numbers; a topology extraction module, constructing an adjacent surface topology H based on whether surfaces are adjacent i , the surface topology H i is an adjacency matrix of each surface j of the part i; H i The sequence number in the middle row is the surface sequence number of the part i, H i The sequence number in the column is the surface sequence number of the part i, H i The element in H j represents the index of the surface sequence number, indicates that the surface with surface sequence number j is adjacent to the surface with surface sequence number k; a geometry feature extraction module to extract geometry characteristics of each surface j of part i and generate subsets of geometry characteristics according to surface serial number, and collect all subsets of geometry characteristics generate geometry feature R i geometry characteristics include: number of edges, peripheral perimeter, area of surface, surface curvature, R i geometry characteristics of the i-th part, subsets of geometry characteristics of the j-th surface of the i-th part; geometry characteristics of the j-th surface of the i-th part; a feature aggregation module, which merges the surface topology H i and the geometric feature R i as a node information respectively into the structural feature HR i of the part. 4.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 2, characterized in that, The position feature extraction device comprises: an anchor position generation module for manually calibrating at least five position anchors in the three-dimensional scanning images, the five position anchors being located at a center position q1 of the vehicle to be tested, a left front wheel shaft center position q2 of the vehicle to be tested, a right front wheel shaft center position q3 of the vehicle to be tested, a left rear wheel shaft center position q4 of the vehicle to be tested, and a right rear wheel shaft center position q5 of the vehicle to be tested. The part vector generation module acquires a three-dimensional scanning image and information code of the part, and extracts a vector of a line between two longest end points on the part from the three-dimensional scanning image The vector is respectively perpendicular to the vector The vector is respectively perpendicular to the vector The vector is respectively perpendicular to the vector The vector is respectively perpendicular to the vector The included angle is calculated to obtain The vector is calculated to obtain the shortest distance between any point and the positions q1, q2, q3, q4, q5 Wherein, the vector represents a line between two longest end points on the part i in a three-dimensional coordinate system, and O represents the position of the origin in the three-dimensional coordinate system, The vectors respectively represent the vectors and the vector The vector The vector The vector and the vector The included angle of the vector The vectors respectively represent q1, q2, q3, q4, q5 and the vector The vector The vector The vector The vector and the vector The shortest distance of any point on the vector a position feature generation module, to as a position feature E of the part i i . 5.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 4, characterized in that, The information aggregation device comprises: The part classification factor identification module generates a weight factor a according to the three-dimensional size of the three-dimensional scanning image of the part i i ; where b i denotes the vector of the line between the two shortest end points of the part i; a feature weight generation module for generating a structure feature weight for each part i i generating a structure feature weight a position feature E for each part i i generating a position feature weight where c represents a conversion factor, c > 0, tanh() represents a hyperbolic tangent function, V i V i represents the volume of the i-th part; The fusion feature generation module fuses the structural feature HR i , the structural feature weight , the location feature E i , and the location feature weight into a set as the fusion feature. 6.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 5, characterized in that, The fault probability extraction device comprises: a standard feature acquisition module for saving standard fused features of parts of a calibration vehicle; a comparison group generation module for generating a comparison group by one-to-one correspondence between the standard fused features and the fused features of the vehicle to be tested according to the information codes; an installation defect generation module for calculating the standard fused features and the fused features of the vehicle to be tested in each comparison group to generate fault probabilities of installation defects of the parts.

7. The neural network model-based internal defect analysis system of the whole vehicle CT scanning image according to claim 6, wherein the installation defect generation module comprises: The structure failure probability generator calculates the similarity between the structure feature in the standard fusion feature and the structure feature in the fusion feature of the vehicle to be tested, and generates a first probability factor The first probability factor represents the part i The position fault probability generator calculates the similarity between the position feature in the standard fusion feature and the position feature in the fusion feature of the vehicle to be tested, and generates a second probability factor The second probability factor represents the part i; the similarity is calculated based on a neural network model; Failure probability calculator based on structural feature weights Position feature weights First probability factor And second probability factor Calculate the failure probability PD of a part installation defect i , PD i Indicates the failure probability of the i-th part in the vehicle under test; 8.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 7, characterized in that, the defect analysis device comprises: a part arrangement module for arranging all the parts i in the vehicle to be tested according to part volumes to form a part series RT; a part extraction module for randomly extracting a plurality of part judgment groups from the part series RT; a risk level calculation module for sequentially calculating a sum of fault probabilities of each part judgment group to generate a risk level factor, and mapping the highest risk level factor to a risk level. 9.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 8, characterized in that, The extraction method of the part extraction module is as follows: S1: The number of parts extraction D in the part judgment group is set in advance, the number M of the part judgment group is set in advance, and the sampling uniformity index G is preset min ; S2: Divide the sequence of parts RT into D intervals g, g = 1, 2, 3, ... D, and the range of interval g is [s g h g ], s g h represents the lower bound of the range of values ​​for interval g. g This represents the upper bound of the range of values ​​for interval g; S3: divide each interval g into M extraction ranges, and the length of the extraction range is Δ g ; the kth extraction range I gk ; I gk = [s g +(k-1)Δ g ,h g +kΔ g ); S4: randomly sampling M parts P in each sampling range gk M part judgment groups are generated, and an initial matrix A is obtained. P gk ~Uniform(s g +(k-1)Δ g ,h g +kΔ g ); Uniform denotes random value; S5: independently performing random arrangement on each row of the initial matrix A to generate a matrix B; S5: Each column of matrix B is a component decision group, q j denotes the jth column of matrix B; S6: calculating a sampling uniformity index G of the matrix B; / denotes the index of the column in matrix B; S7: Calculate if G is less than G min If G is less than G min then the columns of the M matrices B are judged as parts, and if G is greater than G min , then P gk is randomly revalued. 10.The neural network model based whole vehicle CT scan image internal defect analysis system according to claim 8, characterized in that, Based on the extracted part mounting group, structural features and position features of the corresponding part are selectively extracted, and the failure probability of the part is calculated.

Images