Automobile welding part precision analysis method and system based on deep learning

Abstract

The invention belongs to the field of precision analysis, and particularly relates to an automobile welding part precision analysis method and system based on deep learning. The method comprises the following steps: firstly, acquiring an electronic three-dimensional model of a welded part, extracting inspection characteristics such as holes, contours and surfaces to construct a library, and constructing a judgment model on a design model according to an error range; whether the part is qualified or not is judged by registering the to-be-evaluated model with the design model and comparing the inspection features with the evaluation model. Meanwhile, inspection features and reasons of unqualified parts are collected to form samples, an error analysis model is constructed by using a deep learning algorithm, and sample training and verification models are divided, so that the unqualified reasons can be output according to the inspection features. The method overcomes the defects of traditional detection, is high in detection precision and good in comprehensiveness, is of great significance for guaranteeing the welding and assembling quality of an automobile body, and effectively improves the automobile manufacturing quality control level.

Description

Technical Field

[0001] The present invention relates to the field of computer-aided design (CAD), and in particular to vehicle-related design; in particular, to a method and system for analyzing the accuracy of automobile welding parts based on deep learning. Background Art

[0002] The body of a car is made up of welded metal sheets; it is divided into different welding assemblies according to different areas of the body, such as the side panel assembly, door assembly, front floor assembly, rear floor assembly, and front cabin assembly.

[0003] During the welding process of the automobile body, the various sub-assemblies and assemblies are welded first, and then the sub-assemblies and assemblies are welded together; when the welding error of a certain welding assembly of the automobile exceeds the required value, the result is that other parts cannot be installed on the body; therefore, it is particularly important to detect the welding error or accuracy of automobile welding parts.

[0004] At present, the traditional quality inspection method of mechanical parts mainly uses contact measurement methods such as calipers, micrometers, and inspection fixtures, which rely on frequent manual point selection, and the measurement results are easily affected by point selection errors. 3D scanning technology, as a non-contact measurement technology that can quickly obtain high-precision, high-density surface sampling points of objects, can effectively solve these problems.

[0005] The present application proposes an automobile welding parts error analysis method integrating three-dimensional scanning. Summary of the invention

[0006] The present invention provides an automobile welding parts accuracy analysis method and system based on deep learning. The technical problem to be solved is: how to analyze the errors of parts based on deep learning.

[0007] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0008] On the one hand, the automobile welding parts precision analysis method based on deep learning includes the following steps:

[0009] S100, obtaining an electronic three-dimensional model of a welded component, wherein the electronic three-dimensional model of the welded component is a three-dimensional model to be evaluated;

[0010] S200, acquiring all inspection features of the component, the inspection features including holes, contours, and profiles of the component, and constructing an inspection feature library with the inspection features;

[0011] S300, obtaining an error range allowed for the inspection feature of a component; based on the error range, constructing a separate evaluation three-dimensional model around each inspection feature on the basis of the component design three-dimensional model; the evaluation three-dimensional model serves as a reference for evaluating the error of the inspection feature of the three-dimensional model to be evaluated, and the evaluation three-dimensional model can envelop the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range;

[0012] S400, collating the three-dimensional model of the component to be evaluated with the designed three-dimensional model;

[0013] S500, after the registration is completed, compare whether the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged; if the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged, the three-dimensional model to be evaluated of the component to be evaluated meets the specified error range and the component is qualified; if the inspection feature of the three-dimensional model to be evaluated is outside the three-dimensional model to be judged, the three-dimensional model to be evaluated of the component to be evaluated does not meet the specified error range and the component is unqualified.

[0014] Further, in step S100, the electronic three-dimensional data is obtained by using a laser scanner to emit a laser beam to the surface of the welding component. The laser is received by the scanner after being reflected from the surface of the component. The distance information of each point on the surface of the component is calculated by measuring the flight time or phase change of the laser, thereby generating high-precision three-dimensional point cloud data.

[0015] Further, the inspection features and the corresponding reasons for failure of the parts in the historical records are collected and combined to obtain sample data;

[0016] Building an error analysis model based on a deep learning algorithm, the error analysis model is used to output the reasons for failure based on the inspection features corresponding to the failed parts;

[0017] Dividing sample data into training samples and verification samples; inputting the training samples into a pre-built error analysis model for training; and then adjusting parameters of the error analysis model using the verification samples;

[0018] When the inspection features corresponding to the unqualified parts are obtained, the inspection features are input into the error analysis model to obtain the reasons for the unqualified parts.

[0019] Further, step S400 includes the following specific steps:

[0020] (2.1) Randomly sample the part design model Q to obtain the point set Q0;

[0021] (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq;

[0022] (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment;

[0023] (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0;

[0024] (2.5) The residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function are used to weight the point pair D0:

[0025]

[0026] Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant;

[0027] (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P;

[0028] (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as

[0029]

[0030] Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

[0031] On the other hand, the automobile welding parts precision analysis system based on deep learning is characterized by comprising a parts three-dimensional model generation module, an inspection feature library construction module, a judgment model construction module, a model registration module, and an evaluation module;

[0032] A component 3D model generation module is used to obtain an electronic 3D model of a component that has been welded, wherein the electronic 3D model of the component that has been welded is the 3D model to be evaluated;

[0033] An inspection feature library building module is used to obtain all inspection features of the component, the inspection features including holes, contours, and surfaces of the component, and to build an inspection feature library with the inspection features;

[0034] A judgment model construction module is used to obtain the error range allowed for the inspection feature of the component, and according to the error range, on the basis of the component design three-dimensional model, a separate judgment three-dimensional model is constructed around each inspection feature, the judgment three-dimensional model is used as a reference for evaluating the error of the inspection feature of the three-dimensional model to be evaluated, and the judgment three-dimensional model can envelop the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range;

[0035] A model registration module, used for registering the three-dimensional model to be evaluated of the component to be evaluated with the designed three-dimensional model;

[0036] The evaluation module is used to compare whether the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged; if the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged, the three-dimensional model to be evaluated of the component to be evaluated meets the specified error range and the component is qualified; if the inspection feature of the three-dimensional model to be evaluated leaks out of the three-dimensional model to be judged, the three-dimensional model to be evaluated of the component to be evaluated does not meet the specified error range and the component is unqualified.

[0037] Furthermore, a sample data construction module is used to collect inspection features and corresponding reasons for failure of parts and components that are unqualified in historical records, and combine them to obtain sample data;

[0038] An error analysis model building module, which builds an error analysis model based on a deep learning algorithm, wherein the error analysis model is used to output the reasons for failure based on the inspection features corresponding to the failed parts;

[0039] The training module is used to divide the sample data into training samples and verification samples; input the training samples into a pre-built error analysis model for training; and then use the verification samples to adjust the parameters of the error analysis model;

[0040] The inspection module is used to obtain the inspection features corresponding to the unqualified parts, input the inspection features into the error analysis model, and obtain the reasons for the unqualified parts.

[0041] Furthermore, the model registration module specifically includes the following:

[0042] (2.1) Randomly sample the part design model Q to obtain the point set Q0;

[0043] (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq;

[0044] (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment;

[0045] (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0;

[0046] (2.5) The residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function are used to weight the point pair D0:

[0047]

[0048] Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant;

[0049] (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P;

[0050] (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as

[0051]

[0052] Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

[0053] The beneficial effects of the present invention are:

[0054] (1) High detection accuracy: By building a separate evaluation 3D model around the inspection features of parts such as holes, contours, and surfaces, and judging based on the precise error range, compared with the traditional method of only using inspection tools, it can evaluate the accuracy of parts more carefully, effectively avoid misjudgment caused by ignoring local key feature errors in overall inspection, accurately identify qualified and unqualified parts, and provide reliable guarantee for the welding quality of automobile bodies. (2) Strong comprehensiveness of detection: Comprehensively consider the various inspection features of parts, build a corresponding inspection feature library, and comprehensively evaluate the error of each feature during the analysis process. It is not limited to the detection of a single or a few features, and fully covers all aspects of possible errors in parts during production and manufacturing, ensuring comprehensive control of part accuracy and helping to improve the overall assembly accuracy of the automobile. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 It is a logical schematic diagram of the present invention. DETAILED DESCRIPTION

[0056] In order to make the contents of the present invention more clearly understood, the technical solutions in the embodiments of the present invention are described clearly and completely.

[0057] Embodiment 1

[0058] The precision analysis method of automobile welding parts based on deep learning includes the following steps:

[0059] S100, obtaining an electronic three-dimensional model of a welded component, wherein the electronic three-dimensional model of the welded component is a three-dimensional model to be evaluated, and the three-dimensional model to be evaluated is in the form of a point cloud, which is point cloud P. In a specific embodiment, the electronic three-dimensional data is obtained by using a laser scanner to emit a laser beam to the surface of the welded component, and the laser is received by the scanner after being reflected on the surface of the component. The distance information of each point on the surface of the component is calculated by measuring the flight time or phase change of the laser, thereby generating high-precision three-dimensional point cloud data. This method can quickly obtain the overall shape information of the component and is suitable for components with complex shapes, but the measurement accuracy may be affected by factors such as the surface material and glossiness of the component, and the data needs to be post-processed, such as denoising and filtering.

[0060] S200, obtaining all inspection features of the component, wherein the inspection features include holes, contours, and profiles of the component, and constructing an inspection feature library with the inspection features.

[0061] S300, obtaining the error range allowed for the inspection feature of the component, and constructing a separate evaluation three-dimensional model around each inspection feature based on the component design three-dimensional model according to the error range, the evaluation three-dimensional model is used as a benchmark for evaluating the error of the inspection feature of the three-dimensional model to be evaluated, and the evaluation three-dimensional model can envelop the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range. Specifically, the design three-dimensional model is the three-dimensional model of the component design stage, and the design three-dimensional model is the benchmark for production and manufacturing; the error range means that each component is allowed to have certain errors in size, shape, etc. during the production and manufacturing process due to the influence of factors such as process and material, which will not affect the error in use and manufacturing. For the evaluation three-dimensional model, for example, the error specified for the hole of the component is an error of ±1mm, that is, a three-dimensional model with an error of ±1mm on the hole boundary of the design three-dimensional model of the component is constructed, and this three-dimensional model is the evaluation three-dimensional model.

[0062] S400: collating the three-dimensional model of the component to be evaluated with the designed three-dimensional model.

[0063] After S500, compare whether the inspection features of the three-dimensional model to be evaluated are located within the judgment three-dimensional model; if the inspection features of the three-dimensional model to be evaluated are located within the judgment three-dimensional model, then the three-dimensional model of the part to be evaluated meets the specified error range and the part is qualified; if the inspection features of the three-dimensional model to be evaluated leak outside the judgment three-dimensional model, then the three-dimensional model of the part to be evaluated does not meet the specified error range and the part is unqualified.

[0064] This method for precision analysis of automotive welded parts based on deep learning has many advantages. (2) High detection accuracy: By constructing a separate judgment three-dimensional model around the inspection features such as holes, contours, and surfaces of the parts, and making judgments based on precise error ranges, it can evaluate the part precision more meticulously compared to the traditional method of only using inspection tools. It effectively avoids misjudgments caused by ignoring local key feature errors in overall detection, accurately identifies qualified and unqualified parts, and provides a reliable guarantee for the welding quality of the automotive body. (2) Strong detection comprehensiveness: Comprehensively consider various inspection features of the parts, construct a corresponding inspection feature library, and comprehensively evaluate the error conditions of each feature during the analysis process. It is not limited to the detection of single or a few features, and completely covers all aspects where errors may occur in the production and manufacturing of parts, ensuring a comprehensive control of the part precision and helping to improve the overall assembly precision of the vehicle.

[0065] Furthermore, collect the inspection features corresponding to the unqualified parts in the historical records and the corresponding reasons for non-conformity, and combine them to obtain sample data. The reason for non-conformity is due to a certain link in the production and manufacturing of the part, such as stamping equipment, welding equipment, fixtures, etc.

[0066] Construct an error analysis model based on the deep learning algorithm, and the error analysis model is used to output the reason for non-conformity based on the inspection features corresponding to the unqualified parts.

[0067] Divide the sample data into training samples and validation samples; input the training samples into the pre-constructed error analysis model for training; then, use the validation samples to adjust the parameters of the error analysis model.

[0068] When obtaining the inspection features corresponding to the unqualified parts, input the inspection features into the error analysis model to obtain the reason for non-conformity.

[0069] Furthermore, step S400 includes the following specific steps:

[0070] (2.1) Randomly sample the part design model Q to obtain a point set Q0;

[0071] (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq;

[0072] (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment;

[0073] (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0;

[0074] (2.5) Use the residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function to weight the point pair D0:

[0075]

[0076] Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant;

[0077] (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P;

[0078] (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as

[0079]

[0080] Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

[0081] Embodiment 2

[0082] There is a corresponding system for the precision analysis method of automotive welding parts based on deep learning.

[0083] The automobile welding parts precision analysis system based on deep learning includes a parts 3D model generation module, an inspection feature library construction module, a judgment model construction module, a model registration module, and an evaluation module;

[0084] The three-dimensional model generation module for parts is used to obtain the electronic three-dimensional model of the welded parts. The electronic three-dimensional model of the welded parts is the three-dimensional model to be evaluated, and this three-dimensional model to be evaluated is in the form of point cloud, which is also the point cloud P. In a specific embodiment, the electronic three-dimensional data is obtained by using a laser scanner to emit a laser beam onto the surface of the welded parts. After the laser is reflected on the surface of the parts, it is received by the scanner. By measuring the flight time or phase change of the laser, the distance information of each point on the surface of the parts is calculated, and then high-precision three-dimensional point cloud data is generated. This method can quickly obtain the overall shape information of the parts and is applicable to parts with complex shapes. However, it may be affected by factors such as the surface material and gloss of the parts, and certain post-processing of the data, such as denoising and filtering, is required.

[0085] The inspection feature library construction module is used to obtain all the inspection features of the parts. The inspection features include the holes, contours, and surfaces of the parts, and construct an inspection feature library with these inspection features.

[0086] The judgment model construction module is used to obtain the allowable error range of the inspection features of the parts. According to the error range, on the basis of the three-dimensional design model of the parts, a separate judgment three-dimensional model is constructed around each inspection feature. The judgment three-dimensional model serves as a benchmark for evaluating the error of the inspection features of the three-dimensional model to be evaluated, and the judgment three-dimensional model can enclose the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range. Specifically, the design three-dimensional model is the three-dimensional model in the design stage of the parts, and the design three-dimensional model is the benchmark for production and manufacturing. The error range means that during the production and manufacturing process of each part, due to the influence of factors such as process and materials, a certain error in dimensions, shapes, etc. is allowed, and this error will not affect the use and manufacturing. For the judgment three-dimensional model, for example, the allowable error for the hole of the part is ±1mm. That is, a three-dimensional model with an error of ±1mm is constructed on the hole boundary of the design three-dimensional model of the part, and this three-dimensional model is the judgment three-dimensional model.

[0087] The model registration module is used to register the three-dimensional model to be evaluated of the parts to be evaluated with the design three-dimensional model.

[0088] The evaluation module is used to compare whether the inspection features of the three-dimensional model to be evaluated are located within the judgment three-dimensional model. If the inspection features of the three-dimensional model to be evaluated are located within the judgment three-dimensional model, then the three-dimensional model to be evaluated of the parts to be evaluated meets the specified error range, and the parts are qualified. If the inspection features of the three-dimensional model to be evaluated leak outside the judgment three-dimensional model, then the three-dimensional model to be evaluated of the parts to be evaluated does not meet the specified error range, and the parts are unqualified.

[0089] This deep learning-based precision analysis method for automobile welding parts has many advantages: (1) High detection accuracy: By building a separate evaluation three-dimensional model around the inspection features of the parts such as holes, contours, and surfaces, and judging based on the precise error range, compared with the traditional method of only using inspection tools, it can more carefully evaluate the accuracy of parts and effectively avoid misjudgments caused by ignoring local key feature errors in overall inspection, accurately identify qualified and unqualified parts, and provide reliable guarantees for the quality of automobile body welding. (2) Strong comprehensiveness of detection: Comprehensively consider the various inspection features of parts and build a corresponding inspection feature library. In the analysis process, the error of each feature is comprehensively evaluated. It is not limited to the detection of a single or a few features, and fully covers all aspects of possible errors in parts and components during production and manufacturing, ensuring comprehensive control of part accuracy and helping to improve the overall assembly accuracy of the automobile.

[0090] Furthermore, the sample data construction module is used to collect the inspection features and corresponding reasons for failure of the parts in the historical records, and combine them to obtain sample data. The reason for failure is a certain link in the production and manufacturing of the parts, such as stamping equipment, welding equipment, fixtures, etc.

[0091] The error analysis model building module builds an error analysis model based on a deep learning algorithm, and the error analysis model is used to output the reasons for failure based on the inspection features corresponding to the failed parts.

[0092] The training module is used to divide the sample data into training samples and verification samples; input the training samples into a pre-built error analysis model for training; and then use the verification samples to adjust the parameters of the error analysis model.

[0093] The inspection module is used to obtain the inspection features corresponding to the unqualified parts, input the inspection features into the error analysis model, and obtain the reasons for the unqualified parts.

[0094] Furthermore, the model registration module specifically includes the following:

[0095] (2.1) Randomly sample the part design model Q to obtain the point set Q0;

[0096] (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq;

[0097] (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment;

[0098] (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0;

[0099] (2.5) The residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function are used to weight the point pair D0:

[0100]

[0101] Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant;

[0102] (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P;

[0103] (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as

[0104]

[0105] Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

[0106] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for analyzing the accuracy of automobile welding parts based on deep learning, characterized in that: The following steps are involved: S100, obtaining an electronic three-dimensional model of a welded component, wherein the electronic three-dimensional model of the welded component is a three-dimensional model to be evaluated; S200, acquiring all inspection features of the component, the inspection features including holes, contours, and profiles of the component, and constructing an inspection feature library with the inspection features; S300, obtaining an error range allowed for the inspection feature of a component; based on the error range, constructing a separate evaluation three-dimensional model around each inspection feature on the basis of the component design three-dimensional model; the evaluation three-dimensional model serves as a reference for evaluating the error of the inspection feature of the three-dimensional model to be evaluated, and the evaluation three-dimensional model can envelop the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range; S400, collating the three-dimensional model of the component to be evaluated with the designed three-dimensional model; S500, after the registration is completed, comparing whether the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be evaluated; if the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be evaluated, the three-dimensional model to be evaluated of the component to be evaluated meets the specified error range, and the component is qualified; If the inspection features of the three-dimensional model to be evaluated are outside the evaluation three-dimensional model, the three-dimensional model to be evaluated of the component to be evaluated does not conform to the prescribed error range, and the component is unqualified.

2. The method for analyzing the accuracy of automobile welding parts based on deep learning according to claim 1 is characterized in that: In step S100, the electronic three-dimensional data is obtained by using a laser scanner to emit a laser beam to the surface of the welding component. The laser is received by the scanner after being reflected from the surface of the component. The distance information of each point on the surface of the component is calculated by measuring the flight time or phase change of the laser, thereby generating high-precision three-dimensional point cloud data.

3. The method for analyzing the accuracy of automobile welding parts based on deep learning according to claim 1 is characterized in that: Collect the inspection features and corresponding reasons for failure of the parts in the historical records, and combine them to obtain sample data; Building an error analysis model based on a deep learning algorithm, the error analysis model is used to output the reasons for failure based on the inspection features corresponding to the failed parts; Divide the sample data into training samples and verification samples; Inputting the training samples into a pre-built error analysis model for training; Afterwards, the error analysis model is parameter adjusted using the verification sample; When the inspection features corresponding to the unqualified parts are obtained, the inspection features are input into the error analysis model to obtain the reasons for the unqualified parts.

4. The method for analyzing the accuracy of automobile welding parts based on deep learning according to claim 1, characterized in that: Step S400 includes the following specific steps: (2.1) Randomly sample the part design model Q to obtain the point set Q0; (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq; (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment; (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0; (2.5) The residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function are used to weight the point pair D0: Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant; (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P; (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

5. The automobile welding parts precision analysis system based on deep learning is characterized by: It includes a component 3D model generation module, an inspection feature library construction module, a judgment model construction module, a model registration module, and an evaluation module; A component 3D model generation module is used to obtain an electronic 3D model of a component that has been welded, wherein the electronic 3D model of the component that has been welded is the 3D model to be evaluated; An inspection feature library building module is used to obtain all inspection features of the component, the inspection features including holes, contours, and surfaces of the component, and to build an inspection feature library with the inspection features; A judgment model construction module is used to obtain the error range allowed for the inspection feature of the component, and according to the error range, on the basis of the component design three-dimensional model, a separate judgment three-dimensional model is constructed around each inspection feature, the judgment three-dimensional model is used as a reference for evaluating the error of the inspection feature of the three-dimensional model to be evaluated, and the judgment three-dimensional model can envelop the inspection features of the three-dimensional model to be evaluated that meet the requirements of the error range; A model registration module, used for registering the three-dimensional model to be evaluated of the component to be evaluated with the designed three-dimensional model; An evaluation module is used to compare whether the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged; if the inspection feature of the three-dimensional model to be evaluated is located within the three-dimensional model to be judged, the three-dimensional model to be evaluated of the component to be evaluated meets the specified error range and the component is qualified; If the inspection features of the three-dimensional model to be evaluated are outside the evaluation three-dimensional model, the three-dimensional model to be evaluated of the component to be evaluated does not conform to the prescribed error range, and the component is unqualified.

6. The automobile welding parts precision analysis system based on deep learning according to claim 5 is characterized in that: A sample data building module is used to collect the inspection features and corresponding reasons for failure of the parts and components that failed in the historical records, and combine them to obtain sample data; An error analysis model building module, which builds an error analysis model based on a deep learning algorithm, wherein the error analysis model is used to output the reasons for failure based on the inspection features corresponding to the failed parts; A training module is used to divide sample data into training samples and verification samples; Inputting the training samples into a pre-built error analysis model for training; Afterwards, the error analysis model is parameter adjusted using the verification sample; The inspection module is used to obtain the inspection features corresponding to the unqualified parts, input the inspection features into the error analysis model, and obtain the reasons for the unqualified parts.

7. The automobile welding parts precision analysis system based on deep learning according to claim 5 is characterized in that: The model registration module specifically includes the following: (2.1) Randomly sample the part design model Q to obtain the point set Q0; (2.2) The covariance matrices of point cloud P and point set Q0 are calculated by principal component analysis, and then the eigenvectors of the corresponding covariance matrices are used as XYZ axes to establish their respective standard posture coordinate systems Cp and Cq; (2.3) Convert Cp and Cq to the same coordinate system through coordinate transformation to complete the rough alignment; (2.4) For the point cloud P that has completed the rough registration, the part design model Q is used as a reference to calculate the closest distance point set Q1=C(P,Q) from the point cloud P to the part design model Q. The point cloud P and the point set Q1 have a one-to-one correspondence, forming an initial corresponding point correspondence relationship D0; (2.5) The residual value of the midpoint of the sampling point cloud P from the rough registration result and the corresponding weighting function are used to weight the point pair D0: Among them, v is the residual, μ is the mean of v, σ is the variance of v, and n is a constant; (2.6) Combine the point pair weights obtained in step (2.5), calculate the rotation matrix R and translation vector qT, and perform coordinate transformation on the scanned point cloud P; (2.7) Repeat steps (2.1) to (2.3) to perform iterative calculations. Each calculation updates the point correspondence to obtain the point pair Dk. The registration error of the kth iteration is calculated as Where di(k) is the distance from P to Q1 in the kth iteration, n is the number of points in the point cloud P, and when |Ek-Ek-1|<ε, ε refers to the iterative convergence threshold in the fine registration.

Images