Quality detection method and system for vehicle safety belt production and storage medium

Abstract

The invention discloses a quality detection method and system for vehicle safety belt production and a storage medium, and relates to the field of manufacturing quality control of automobile passive safety parts. The method aims at solving the problems that an existing detection means is weak in tiny defect recognition capacity, insufficient in defect evolution modeling and lack of intelligent feedback regulation and control. According to the method, multiple detection means such as medium-wave infrared thermal imaging, laser confocal microscopy, structured light three-dimensional reconstruction and capacitance sensing are fused, and multi-mode defect sensing of key parts such as a braid, an insertion area and an adhesive layer is achieved; deducing a defect evolution process and generating a high-credibility pseudo label by constructing a thermal-stress-rheology coupling finite element model; the graph neural network and the AI model are combined to complete defect type identification, risk quantification and grading, and automatic optimization of process parameters and production line control response are realized based on the quality grade, so that the detection precision, the response efficiency and the product yield are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of manufacturing quality control of automotive passive safety components, specifically to a quality inspection method, system, and storage medium for the production of vehicle seat belts. Background Technology

[0002] In current vehicle seatbelt production, significant interface complexity exists in areas such as the webbing and sensor bonding, and the metal-plastic joint. Traditional visual or mechanical inspection methods struggle to identify early microscopic defects like blistering, wrinkling, and delamination. Simultaneously, small-scale defects such as microcracks, pinholes, and uneven adhesive layers are often missed by two-dimensional image or surface inspection equipment, leading to delayed exposure of quality risks. Furthermore, existing inspection methods are largely result-oriented, lacking defect evolution modeling based on thermo-mechanical coupling mechanisms, making timely risk prediction difficult. Inconsistent data dimensions and spatial scales among multi-source inspection devices, coupled with a lack of efficient fusion and evaluation mechanisms, reduce identification accuracy. At the quality assessment level, defect classification generally relies on manual experience or static rules, lacking quantitative scoring and risk visualization, easily causing delayed responses to high-risk defects. Moreover, identification results are insufficient to drive production line process control, lacking process optimization and closed-loop feedback mechanisms based on defect types, thus hindering the intelligent development of quality control. Summary of the Invention

[0003] The purpose of this invention is to address the problems existing in the background art described above, and to propose a quality inspection method, system, and storage medium for the production of vehicle seat belts.

[0004] The objective of this invention can be achieved through the following technical solutions:

[0005] In a first aspect, the present invention provides a quality inspection method, system, and storage medium for the production of vehicle seat belts, comprising:

[0006] S01: After assembly, the detection operation is performed in the area where the seat belt webbing fits with the sensor and the metal-plastic insertion area to form a data packet with a uniform structure.

[0007] S02: Obtain the data package of the structure to construct a thermo-stress-rheological coupling model that reflects the evolution of potential defects under real service environment, and output the spatially distributed defect risk level Rsim and the three-dimensional defect risk distribution map based on the finite element method;

[0008] S03: Combine the multimodal detection data from S01 with the potential defect risk labeling map from S02, perform the operation, and output the confidence score Pr for each defect type. vi Defect identification result set F0 and quality grading coefficient Q0;

[0009] S04: Analyze defect grading and quality levels, inputting structured defects and outputting F0 and quality grading coefficients Q0; construct a multi-index fusion scoring function Q = F based on the characteristic parameters of various defects; including: crack defect values, bubble defect values, and wrinkle defect values; normalize the corresponding numerical results of each defect item and merge them into a global defect index Q. q ; calculate Q q Compare with the preset multi-level quality level thresholds; if Q q Less than the quality grade threshold T A If Q, then the defect level is A; if Q q Less than the quality grade threshold T B Greater than the quality grade threshold T A If Q, then the defect level is B; if Q q Greater than the quality grade threshold T B If so, the defect level is C;

[0010] S05: Automatically analyze defect levels and execute controls:

[0011] For grade A, the current process status is recorded as a normal sample, no adjustment is required, and the process proceeds to the next product testing step.

[0012] For a grade of B, activate the automatic process adjustment module to analyze the key process variables that are related to the current defect type and the score, and make adjustments accordingly.

[0013] For level C, the control response mechanism is activated: immediately stop the operation of the current workstation; trigger an audible and visual alarm, notify the manual re-inspection system, mark and isolate the abnormal batch of products, and move them into the quality isolation area; report the current data to the process quality center system to form a traceability log.

[0014] In a preferred embodiment of the present invention, the detection operation includes:

[0015] S101: The temperature distribution matrix T(x,y) is extracted using a mid-wave infrared thermal imager, and the boundary of the detected temperature anomaly is obtained by combining it with the thermal diffusion gradient field, marking the potential anomaly area;

[0016] S102: Use a laser confocal microscope to scan and construct a local surface topography map to obtain a 3D surface profile map S(x,y,z); based on the 3D surface profile map S(x,y,z), extract quantitative features such as roughness index Ra and topography curvature K(x,y);

[0017] S103: Uses a high-frequency near-field capacitive sensing array to identify areas of abnormal thickness;

[0018] S104: Use a structured light projector and a binocular stereo camera to obtain P(x,y,z) point cloud data;

[0019] S105: Perform spatiotemporal alignment and standardization on the data obtained from S101, S102, S103 and S104 respectively to form a data packet with a unified structure.

[0020] As a preferred embodiment of the present invention, the process of outputting the spatially distributed defect risk level Rsim and the three-dimensional defect risk distribution map based on the finite element method is as follows:

[0021] Data packages and production physical parameters of the structure are extracted from the seat belt manufacturing material database. A three-field coupled finite element model is constructed based on S(x,y,z) and P(x,y,z) to simulate the response mechanism of thermo-mechanical-rheological behavior to defect evolution and realize the output of defect risk distribution map R(x,y,z).

[0022] Mesh modeling uses a scanned reconstruction model to generate physical domain boundaries. Based on the defect region information in P(x,y,z), Om is divided into finite element meshes to form a mesh set G{i}. For defective regions or interface stress concentration regions, a local refinement strategy is adopted. Based on the inspection data and product design load database, the actual working condition boundary conditions are reconstructed, including thermal boundary conditions, force boundary conditions, and contact boundaries. Based on the established finite element model and boundary conditions, a coupled set of partial differential equations is constructed. The risk function {R}T is used to calculate the spatially distributed defect risk level Rsim. The defect risk level Rsim is the risk intensity score on each element / system, which is then visualized into a three-dimensional color map to determine and output high-risk areas, forming a potential defect risk label map. The label map fields include local maximum stress, cumulative creep, and degree of thermal gradient concentration.

[0023] As a preferred embodiment of the present invention, the confidence level Pr for each defect type is output. vi The process of determining the defect identification result set F0 and the quality grading coefficient Q0:

[0024] For each defect type v in the structural map Gdefect=(V,E) i Modeled as semantic embedding points, node features are derived from the prior template Mv iThe Rsim field is generated; then, it is aligned with the Dsense data in spatial dimension and used as one of the inputs to the AI ​​model; the AI ​​model includes a perceptual feature extraction module, a simulation weighted attention module, a defect map inference module, and a recognition result output module; feature modeling is performed on different modal data, including: using a convolutional neural network to extract temperature texture features from the heat map T(x,y); using a 3D point cloud network to model the microcrack undulation structure from the surface morphology S(x,y,z); using a UNet structure to extract the local uniformity pattern from the adhesive layer thickness map H(x,y); using a sparse 3D convolution and Transformer structure to extract the chimeric spatial distribution from the interpolated point cloud P(x,y,z); the extracted features are respectively formed into embedding vector groups to form a unified modal feature set F1;

[0025] Fields from the simulated field distribution map Rsim are introduced as attention-guided weights, resulting in a weighted spatial perception vector F0. Based on the structured defect map, graph neural networks are used for inference enhancement, including: generating initial defect node embedding vectors using template matching, propagating defect recognition probabilities within the map structure based on feature similarity, and outputting the confidence score Pr for each defect type. vi ,

[0026] Based on the confidence level Pr of the defect type vi Determine the defect type; if the confidence level Pr vi If the value exceeds the set threshold, it is determined to be the corresponding defect type; the spatial location, type, severity and matching path corresponding to the defect type are integrated to obtain the defect identification result set F0; the quality grading coefficient Q0 is calculated.

[0027] In a preferred embodiment of the present invention, for a grade of B, the automatic process adjustment module is activated to analyze the key process variables associated with the current defect type and score, and then makes adjustments:

[0028] A process parameter optimization module is introduced, and the optimal adjustment is selected based on an online optimization strategy; the objective function is constructed using a Bayesian optimization algorithm. The parameter combination A0 that minimizes the expected value of the defect score;

[0029] Since the smallest parameter combination A0 is a complex and computationally expensive function, Bayesian optimization uses a Gaussian process to fit the function and obtain the mean and uncertainty. Then, the next parameter combination to be tried is determined by the acquisition function, and the EI is used to select the parameters that are most likely to be better than the current optimal solution.

[0030] Apply the selected parameter combination to the production line, record the new defect coefficient, and repeat steps 2 to 4 until Q0 no longer decreases or the maximum number of iterations is reached.

[0031] In another aspect, the present invention provides a quality inspection system for the production of vehicle seat belts, including a defect perception and acquisition module, a modeling and simulation analysis module, a defect identification and quality analysis module, and an intelligent feedback control module.

[0032] The defect perception and acquisition module is deployed at the inspection station of the seat belt production line to acquire the heat map T(x,y), three-dimensional morphology S(x,y,z), adhesive layer thickness map H(x,y), and structural point cloud P(x,y,z) of the webbing and splicing area, respectively, so as to achieve high-resolution and all-round perception of key defects such as wrinkles, bubbles, cracks and uneven adhesive layer.

[0033] The modeling and simulation analysis module standardizes and fuses multimodal sensing data, and combines material properties and actual working conditions to construct a coupled thermal-stress-rheological three-field finite element model to simulate the temperature field, stress field and deformation response of the material under real load. Through solving, it outputs a three-dimensional defect risk map R(x,y,z), predicts the failure probability of high-risk areas, and generates a pseudo-label map to improve the accuracy and reliability of identifying latent defects.

[0034] The defect identification and quality analysis module takes multimodal sensing data and simulated risk maps as input, and uses convolutional neural networks, PointNet++, UNet, and Transformer structures to extract various features. It then fuses defect map knowledge through graph neural networks to achieve multidimensional identification of defect type, location, confidence level, and severity. Furthermore, it constructs a quality scoring model to quantify the impact of various defects and calculate the global defect index Q. q Complete the product quality grade determination;

[0035] The intelligent feedback control module automatically responds based on the identified defect level: for A-level products, the process status is automatically recorded as a normal sample; for B-level products, the parameter optimization module is triggered, and key process parameters are adjusted iteratively using Bayesian optimization to reduce the defect score; for C-level products, the shutdown alarm, anomaly isolation and data reporting mechanism are immediately activated to form a quality closed loop and traceability log, effectively improving the intelligent control capability and production stability of the production line.

[0036] Thirdly, the present invention also provides a computer-readable storage medium storing a computer program thereon. When the program is executed by a computer device (such as a personal computer, server, or network device), it is used to implement all or part of the steps of the quality inspection system for vehicle seat belt production described in the present invention, including: controlling multimodal sensors to collect infrared thermal images, three-dimensional topography, thickness distribution, and structural point cloud data of the webbing and splicing area; fusing and standardizing multi-source data; constructing a thermo-stress-rheological coupled finite element model and generating a defect risk distribution map; identifying defect types and severity through a deep learning model and assessing product quality levels; and performing control operations such as parameter optimization, line stop response, and quality traceability based on the assessment results. The storage medium includes, but is not limited to, media capable of electronically storing program instructions, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0037] Compared with the prior art, the beneficial effects of the present invention are:

[0038] 1. This invention integrates advanced detection technologies such as mid-wave infrared, laser microscopy, structured light, and capacitance array to achieve comprehensive detection of key parts such as webbing, splice area, and adhesive layer. It also combines a thermo-stress-rheology coupled finite element simulation model to deduce the evolution process of potential defects and generate highly reliable pseudo-labels.

[0039] 2. By deeply integrating multimodal perception and graph neural network inference, the accuracy and robustness of defect identification are improved. It also enables quantitative scoring of defect levels and automatic control response, and has high intelligence, high sensitivity and traceability, which significantly improves the quality inspection level and production yield of vehicle seat belt products. Attached Figure Description

[0040] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0041] Figure 1 This is a flowchart illustrating the invention process of this invention;

[0042] Figure 2 This is a schematic diagram of the principle of the present invention. Detailed Implementation

[0043] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0044] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0045] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims refers to any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0046] Please see Figure 1 As shown, the present invention provides a quality inspection method for the production of vehicle seat belts, comprising:

[0047] S01: After assembly, perform the detection operation at the area where the seatbelt webbing fits against the sensor and the metal-plastic connector area.

[0048] S101: Using a mid-wave infrared thermal imager, when the surface of the webbing is irradiated with heat, the stress concentration, local bubbles or wrinkles inside the material will cause differences in thermal conductivity, which will be a non-uniform temperature distribution area on the infrared thermal image. The temperature distribution matrix T(x,y) is extracted and combined with the thermal diffusion gradient field nablaT to obtain the detection temperature anomaly boundary and mark the potential anomaly area.

[0049] S102: A local surface morphology map is scanned and constructed using a laser confocal microscope. By modeling the surface shape based on the difference in reflection intensity, microcracks, pinholes, bubbles, and abrupt changes in surface particles are identified, collectively referred to as the 3D surface profile map S(x,y,z). Based on the 3D surface profile map S(x,y,z), quantitative features such as roughness index Ra and morphology curvature K(x,y) are extracted.

[0050] S103: Using a high-frequency near-field capacitance sensing array, the dielectric gap between non-metallic materials, i.e., the local thickness of the adhesive layer, is calculated by measuring the capacitance change C(x,y) in different regions; a thickness distribution map H(x,y) is constructed based on the local thickness of the adhesive layer, and the mean H is obtained. μ With variance H σ Assess the uniformity of the adhesive layer and identify areas of abnormal thickness;

[0051] S104: Using a structured light projector and binocular stereo camera, a coded grating is formed by structured light projection, and the reflected images are captured by dual cameras. The three-dimensional point cloud model of the interpolation area is reconstructed using the parallax method to obtain P(x,y,z) point cloud data.

[0052] S105: Perform spatiotemporal alignment and standardization processing (based on template matching and region mapping) on ​​the data obtained from S101, S102, S103 and S104 respectively to form a data packet with a unified structure.

[0053] S02: Obtain the data package of the structure to construct a thermo-stress-rheological coupling model that reflects the evolution of potential defects under real service conditions, and output a three-dimensional defect risk distribution map based on the finite element method (FEM) as an auxiliary supervision data source (pseudo-label) for the AI ​​model, specifically:

[0054] S201: Extract the data package of the structure and production physical parameters from the seat belt manufacturing material database, wherein the production physical parameters include, but are not limited to: Young's modulus, Poisson's ratio, coefficient of thermal expansion, temperature-dependent viscosity, thermal conductivity, density, specific heat capacity and ultimate stress;

[0055] S202: Based on S(x,y,z) and P(x,y,z), a three-field coupled finite element model is constructed to simulate the response mechanism of thermo-mechanical-rheological behavior to defect evolution and realize the output of defect risk distribution map R(x,y,z);

[0056] Mesh modeling is performed by generating physical domain boundaries using a scan-reconstructed model. Based on the defect region information in P(x,y,z), Om is divided into finite element meshes to form a mesh set G{i}. For defect regions or interface stress concentration regions, a local refinement strategy is adopted to improve mesh accuracy and ensure simulation accuracy.

[0057] Based on the test data and product design load database, the actual working condition boundary conditions are reconstructed, including thermal boundary conditions, force boundary conditions and contact boundaries.

[0058] Thermal boundary condition: Apply heat flux boundary qn to an external heat source or temperature difference region, the value of which is derived from the T(x,y) distribution;

[0059] Force boundary condition: Apply an axial insertion force Fz to the insertion joint;

[0060] Contact boundary: Define the frictional contact boundary between metal and plastic, and set the friction factor m. u The Penalty method is used to consider nonlinear contact behavior; if interface defects exist, an additional initial contact gap condition delta0 is introduced.

[0061] Based on the established finite element model and boundary conditions, the following coupled partial differential equation system is constructed: Where κ is the thermal conductivity tensor, Q is the volumetric heat source term, γ is the rheological internal variable, and C is the thermal conductivity tensor. p Let be the specific heat capacity, σ be the stress tensor, u be the displacement vector field, ε(u) be the strain tensor, C be the fourth-order elastic stiffness tensor, and f(T,σ) be the rheological evolution function. After the solution is completed, the temperature field T(x,y,z), stress field σ(x,y,z), and rheological variable γ(x,y,z) are extracted. The spatially distributed defect risk level Rsim is calculated using the risk function {R}T. The defect risk level Rsim is the risk intensity score on each unit / system, with a numerical range of [0,1]. It represents the estimated probability of failure of the material under complex stress-thermal-rheological coupling. The result is visualized as a three-dimensional color map. High-risk areas are identified and output to form a potential defect risk label map. The label map fields include local maximum stress, cumulative creep, and degree of thermal gradient concentration.

[0062] S03: Combine the multimodal detection data from S01 with the potential defect risk labeling map from S02; Perform the following operation:

[0063] S301: Each defect type v in the structural map Gdefect=(V,E) i Modeled as semantic embedding points, node features are derived from the prior template Mv i Generation; Propagation rules of graph neural networks:

[0064] in, Let N(i) be the feature vector of node i at layer l+1, and let N(i) be the set of neighboring nodes of node i; W is the learnable weight matrix, and σ(·) is the non-linear activation function.

[0065] The Rsim field is then aligned with the Dsense data in spatial dimension and used as one of the input items for the AI ​​model.

[0066] S302: The AI ​​model includes a perceptual feature extraction module, a simulation weighted attention module, a defect map reasoning module, and a recognition result output module;

[0067] The module performs feature modeling on different modal data, including: extracting temperature texture features from the heat map T(x,y) using a convolutional neural network (CNN); modeling the microcrack undulation structure from the surface morphology S(x,y,z) using a 3D point cloud network (PointNet++); extracting local uniformity patterns from the adhesive layer thickness map H(x,y) using a UNet structure; and extracting the splicing spatial distribution from the interpolated point cloud P(x,y,z) using sparse 3D convolution and a Transformer structure. The extracted features are then formed into embedding vector groups to constitute a unified modal feature set F1.

[0068] S303: To improve the model's recognition sensitivity in high-risk areas, fields from the simulation field distribution map Rsim are also introduced as attention-guided weights:

[0069] Where β is a learnable weighting coefficient; The output is the weighted spatial perception vector F0, which is the result of normalizing the simulation fields.

[0070] Based on the structured defect map, inference enhancement is performed using a graph neural network (GNN), including: generating initial defect node embedding vectors using template matching, propagating defect recognition probabilities in the map structure based on feature similarity, and outputting the confidence score Pr for each defect type. vi ,

[0071] S304: Confidence level Pr based on defect type vi Determine the defect type; if the confidence level Pr vi If the value exceeds the set threshold, it is determined to be the corresponding defect type; the spatial location, type, severity and matching path corresponding to the defect type are integrated to obtain the defect identification result set F0;

[0072] Based on the obtained defect identification result set F0, the quality grading coefficient Q0 is calculated as follows: Q0 = w1 × r risk +w2×Pr vi +w3×severity index, where the weights w1, w2, and w3 can be automatically tuned by the model or set manually; the coefficient range is 0 to 100, with higher values ​​indicating more severe quality problems; r risk Fields in the defect risk distribution map output by the finite element multiphysics coupling model. The severity index is a defect severity indicator, set by the defect map library.

[0073] S04: Defect classification and quality level analysis, input defect structured output F0 and quality classification coefficient Q0;

[0074] Based on the characteristic parameters of various defects, a multi-index fusion scoring function Q = F is constructed; including: crack defect value: Q LW =a1×L1+a2×Lθ+a3×Pr LW The sum of weights a1, a2, and a3 is 1, and their specific values ​​can be automatically adjusted by the model or manually defined. L1 is the crack length, Lθ is the crack angle, and Pr... LWTo identify crack confidence levels, similarly, values ​​for bubble-type defects (calculated by weighting the bubble density per unit area and the maximum bubble area) and wrinkling-type defects (calculated by weighting the maximum wrinkling angle and the difference in surface undulation height) are obtained. Then, the corresponding numerical results for each defect item are normalized and combined into a global defect index Q. q : Where, γ qi The impact coefficient of each type of defect on the whole;

[0075] The calculated Q q Compare with the preset multi-level quality level thresholds; if Q q Less than the quality grade threshold T A If Q, then the defect level is A; if Q q Less than the quality grade threshold T B Greater than the quality grade threshold T A If Q, then the defect level is B; if Q q Greater than the quality grade threshold T B If so, the defect level is C;

[0076] The defect classification results are bound to the information to generate data packet G. z The data package includes the workpiece number, batch number, process number, defect information structure, and global defect index Q. q With rating labels (A / B / C).

[0077] S05: Automatically analyze defect levels and execute controls:

[0078] For grade A, the current process status is recorded as a normal sample, no adjustment is required, and the process proceeds to the next product testing step.

[0079] For a grade of B, activate the automatic process adjustment module to analyze the key process variables that are related to the current defect type and the score, and make adjustments accordingly.

[0080] For level C, activate the control response mechanism: immediately stop the current workstation operation (or suspend the automated process); trigger an audible and visual alarm, notify the manual re-inspection system, mark and isolate the abnormal batch of products, and move them to the quality isolation area; report the current data to the process quality center system to form a traceability log.

[0081] S501: For automatic adjustment of Class B defects, a process parameter optimization module is introduced to select the optimal adjustment based on an online optimization strategy;

[0082] Construct the objective function using the Bayesian optimization algorithm: The parameter combination A0 that minimizes the expected value of the defect score;

[0083] Since the smallest parameter combination A0 is a complex and computationally expensive function (e.g., requiring actual production testing), Bayesian optimization uses a Gaussian process (GPR) to fit the data, obtaining the mean and uncertainty (variance); then, the next parameter combination to be tried is determined by the sampling function, and EI (Expected Improvement) is used to select the parameters most likely to be better than the current optimal solution;

[0084] Apply the selected parameter combination to the production line, record the new defect coefficient, and repeat steps 2 to 4 until Q0 no longer decreases or the maximum number of iterations is reached.

[0085] Please see Figure 2 As shown, another aspect of the present invention provides a quality inspection system for the production of vehicle seat belts, comprising: a defect perception and acquisition module, a modeling and simulation analysis module, a defect identification and quality analysis module, and an intelligent feedback control module;

[0086] The defect perception and acquisition module is deployed at the inspection station in the seat belt production line. It is responsible for acquiring multimodal defects in key areas such as webbing and splicing. It integrates various non-contact sensors, including an infrared thermal imager, a laser confocal microscope, a capacitance array, and a structured light stereo imaging device. Through these devices, it acquires thermal images T(x,y), three-dimensional morphology S(x,y,z), adhesive layer thickness map H(x,y), and structural point cloud P(x,y,z), respectively, to achieve comprehensive perception of various defects such as wrinkles, bubbles, cracks, and uneven thickness, providing basic data for subsequent analysis.

[0087] The modeling and simulation analysis module fuses, standardizes, and structures multi-source sensing data, and then constructs a coupled finite element model of thermal-stress-rheological fields by combining material properties and actual operating conditions. The module simulates and solves for the temperature field, stress field, and material deformation response under actual loading, generating a three-dimensional defect risk map R(x,y,z). This map reflects the potential failure probability of local materials under service conditions and serves as an auxiliary supervisory pseudo-label for AI identification, improving the system's accuracy in identifying latent defects and its physical interpretation capabilities.

[0088] The defect identification and quality analysis module takes both perceived data and simulation results as input, and uses deep learning algorithms to identify and classify various defects. It introduces convolutional neural networks, point cloud networks, and Transformer structures to extract multimodal features, and uses graph neural networks to fuse defect map knowledge for reasoning. The identification results include defect type, location, severity, and confidence level. It further calculates the global defect value Qq based on the defect risk function and level analysis standard, and determines the product quality level (A / B / C), outputting a complete defect analysis and quality assessment result set.

[0089] The intelligent feedback control module executes an automated response mechanism based on the identified defect level. For Grade A products, the system automatically records the current process status as a normal sample; for Grade B, the parameter optimization module is activated, using a Bayesian optimization algorithm to iteratively adjust key process variables to reduce the defect score; for Grade C, the system immediately triggers a shutdown, alarm, and anomaly isolation process, and generates a defect traceability log which is uploaded to the quality control platform.

[0090] This invention also provides a computer-readable storage medium storing a computer program. When the program is executed by a computer device (such as a personal computer, server, or network device), it implements all or part of the steps of the quality inspection system for vehicle seat belt production described in this invention, including: controlling multimodal sensors to collect infrared thermal images, three-dimensional topography, thickness distribution, and structural point cloud data of the webbing and splicing area; fusing and standardizing multi-source data; constructing a thermo-stress-rheological coupled finite element model and generating a defect risk distribution map; identifying defect types and severity through a deep learning model and assessing product quality levels; and performing control operations such as parameter optimization, line stop response, and quality traceability based on the assessment results. The storage medium includes, but is not limited to, media capable of electronically storing program instructions, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0091] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to any specific implementation. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A quality inspection method, system, and storage medium for vehicle seat belt production, characterized in that, include: S01: The area where the seat belt webbing fits with the sensor and the metal-plastic insertion area after assembly; Perform detection operations to generate data packets with a unified structure; S02: Obtain the data package of the structure to construct a thermo-stress-rheological coupling model that reflects the evolution of potential defects under real service environment, and output the spatially distributed defect risk level and three-dimensional defect risk distribution map based on the finite element method; S03: Perform operations on the multimodal detection data and the potential defect risk label map, and output the confidence score, defect identification result set and quality grading coefficient for each defect type; S04: Analyze defect grading and quality levels, inputting structured defect output and quality grading coefficients; construct a multi-index fusion scoring function based on the characteristic parameters of various defects, including: crack defect values, bubble defect values, and wrinkle defect values; normalize the corresponding numerical results of each defect item and merge them into a global defect index; compare the global defect index with preset multi-level quality level thresholds; if the global defect index is less than the quality level threshold T... A If the global defect index is less than the quality level threshold T, then the defect level is A; B Greater than the quality grade threshold T A If the global defect index is greater than the quality level threshold T, then the defect level is B; B If so, the defect level is C; S05: Automatically analyze and control the defect level. Specifically: For level A, record the current process status as a normal sample, no adjustment is needed, and proceed to the next product inspection process; For level B, activate the automatic process adjustment module, analyze the key process variables related to the current defect type and score, and make adjustments; For level C, activate the control response mechanism: immediately stop the current workstation; trigger an audible and visual alarm, notify the manual re-inspection system, mark and isolate the abnormal batch of products, and move them to the quality isolation area; report the current data to the process quality center system to form a traceability log.

2. The quality inspection method for vehicle seat belt production according to claim 1, characterized in that, Perform the detection operation, including: S101: The temperature distribution matrix T(x,y) is extracted using a mid-wave infrared thermal imager, and the boundary of the detected temperature anomaly is obtained by combining it with the thermal diffusion gradient field, marking the potential anomaly area; S102: Use a laser confocal microscope to scan and construct a local surface topography map to obtain a 3D surface profile map S(x,y,z); based on the 3D surface profile map S(x,y,z), extract the roughness index Ra and the topography curvature K(x,y) quantitative features. S103: Uses a high-frequency near-field capacitive sensing array to identify areas of abnormal thickness; S104: Use a structured light projector and a binocular stereo camera to obtain P(x,y,z) point cloud data; S105: Perform spatiotemporal alignment and standardization on the data obtained in steps S101-S104 to form a data packet with a unified structure.

3. The quality inspection method for vehicle seat belt production according to claim 2, characterized in that, The process of outputting spatially distributed defect risk levels and three-dimensional defect risk distribution maps based on the finite element method: Data packages and production physical parameters of the structure are extracted from the seat belt manufacturing material database. A three-field coupled finite element model is constructed based on S(x,y,z) and P(x,y,z) to simulate the response mechanism of thermo-mechanical-rheological behavior to defect evolution and realize the output of defect risk distribution map R(x,y,z). Mesh modeling uses a scan-reconstructed model to generate physical domain boundaries. Based on the defect region information in P(x,y,z), Om is divided into finite element meshes to form a mesh set G{i}. For defect-containing regions or interface stress concentration regions, a local meshing strategy is adopted. Based on the inspection data and product design load database, the actual working condition boundary conditions are reconstructed. Based on the established finite element model and boundary conditions, a coupled set of partial differential equations is constructed. The risk function {R}T is used to calculate the spatially distributed defect risk level. The defect risk level is the risk intensity score on each unit / system, which is then visualized into a three-dimensional color map to identify and output high-risk areas, forming a potential defect risk label map.

4. The quality inspection method for vehicle seat belt production according to claim 3, characterized in that, The process of outputting the confidence score, defect identification result set, and quality grading coefficient for each defect type: For each defect type v in the structural map Gdefect=(V,E) i Modeled as semantic embedding points, node features are derived from the prior template Mv i The Rsim field is generated; then, it is aligned with the Dsense data in spatial dimension and used as one of the inputs to the AI ​​model; the AI ​​model includes a perceptual feature extraction module, a simulation weighted attention module, a defect map inference module, and a recognition result output module; feature modeling is performed on different modal data, including: extracting temperature texture features from the heat map T(x,y) using a convolutional neural network; modeling the microcrack undulation structure from the surface morphology S(x,y,z) using a 3D point cloud network; extracting local uniformity patterns from the adhesive layer thickness map H(x,y) using a UNet structure; and extracting the splicing spatial distribution from the interpolated point cloud P(x,y,z) using sparse 3D convolution and a Transformer structure; the extracted features are then formed into embedding vector groups to constitute a unified modal feature set; Fields from the simulated field distribution map are introduced as attention-guided weights, and the output result is a weighted spatial perception vector. Based on the structured defect map, a graph neural network is used for inference enhancement, including: generating initial defect node embedding vectors using template matching, propagating defect recognition probabilities in the map structure based on feature similarity, and outputting the confidence score of each defect type. Based on the confidence level of the defect type, the defect type is determined. If the confidence level is greater than the set threshold, it is determined to be the corresponding defect type. The spatial location, type, severity and matching path corresponding to the determined defect type are integrated to obtain the defect identification result set. The quality grading coefficient is calculated.

5. The quality inspection method for vehicle seat belt production according to claim 4, characterized in that, For a grade of B, the automatic process adjustment module is activated to analyze the key process variables related to the current defect type and score, and then make adjustments accordingly. A process parameter optimization module is introduced to select the optimal adjustment based on an online optimization strategy; an objective function is constructed using a Bayesian optimization algorithm; and the parameter combination that minimizes the expected value of the defect score is identified. Since the minimum parameter combination is a complex and computationally expensive function, Bayesian optimization uses a Gaussian process to fit the data, obtaining the mean and uncertainty. Then, the sampling function determines the next parameter combination to try, and EI is used to select the parameters that are most likely to be better than the current optimal solution. Apply the selected parameter combination to the production line, record the new defect coefficient, and repeat steps two to four until it stops decreasing or the maximum number of iterations is reached.

6. A quality inspection system for vehicle seat belt production, characterized in that, It includes a defect perception and acquisition module, a modeling and simulation analysis module, a defect identification and quality analysis module, and an intelligent feedback control module; The defect perception and acquisition module is deployed at the inspection station of the seat belt production line to acquire the heat map T(x,y), three-dimensional morphology S(x,y,z), adhesive layer thickness map H(x,y), and structural point cloud P(x,y,z) of the webbing and splicing area, respectively, so as to achieve high-resolution and all-round perception of key defects such as wrinkles, bubbles, cracks and uneven adhesive layer. The modeling and simulation analysis module standardizes and fuses multimodal sensing data, and combines material property parameters and actual working conditions to construct a finite element model that couples the thermal, stress, and rheological fields, simulating the temperature field, stress field, and deformation response of the material under real loads. By solving the problem, a three-dimensional defect risk map R(x,y,z) is output, which predicts the probability of damage to high-risk areas and generates a pseudo-label map, thereby improving the accuracy and reliability of identifying latent defects. The defect identification and quality analysis module takes multimodal perception data and simulation risk maps as input, and uses convolutional neural networks, PointNet++, UNet and Transformer structures to extract various features. It integrates defect map knowledge through graph neural networks to achieve multidimensional identification of defect type, location, confidence and severity. It further constructs a quality scoring model to quantify the impact of various defects and calculate global defect indicators to complete the product quality level determination. The intelligent feedback control module automatically responds based on the identified defect level: for A-level products, the process status is automatically recorded as a normal sample; for B-level products, the parameter optimization module is triggered, and key process parameters are adjusted iteratively using Bayesian optimization to reduce the defect score; for C-level products, the shutdown alarm, anomaly isolation and data reporting mechanism are immediately activated to form a quality closed loop and traceability log, effectively improving the intelligent control capability and production stability of the production line.

7. A computer-readable storage medium storing instructions thereon, characterized in that, When the instructions are executed by the processor, they implement the quality inspection method for vehicle seat belt production as described in any one of claims 1-5.

Images